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
];
190 struct mem_cgroup_lru_info
{
191 struct mem_cgroup_per_node
*nodeinfo
[0];
195 * Cgroups above their limits are maintained in a RB-Tree, independent of
196 * their hierarchy representation
199 struct mem_cgroup_tree_per_zone
{
200 struct rb_root rb_root
;
204 struct mem_cgroup_tree_per_node
{
205 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
208 struct mem_cgroup_tree
{
209 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
212 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
214 struct mem_cgroup_threshold
{
215 struct eventfd_ctx
*eventfd
;
220 struct mem_cgroup_threshold_ary
{
221 /* An array index points to threshold just below or equal to usage. */
222 int current_threshold
;
223 /* Size of entries[] */
225 /* Array of thresholds */
226 struct mem_cgroup_threshold entries
[0];
229 struct mem_cgroup_thresholds
{
230 /* Primary thresholds array */
231 struct mem_cgroup_threshold_ary
*primary
;
233 * Spare threshold array.
234 * This is needed to make mem_cgroup_unregister_event() "never fail".
235 * It must be able to store at least primary->size - 1 entries.
237 struct mem_cgroup_threshold_ary
*spare
;
241 struct mem_cgroup_eventfd_list
{
242 struct list_head list
;
243 struct eventfd_ctx
*eventfd
;
246 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
247 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
250 * The memory controller data structure. The memory controller controls both
251 * page cache and RSS per cgroup. We would eventually like to provide
252 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
253 * to help the administrator determine what knobs to tune.
255 * TODO: Add a water mark for the memory controller. Reclaim will begin when
256 * we hit the water mark. May be even add a low water mark, such that
257 * no reclaim occurs from a cgroup at it's low water mark, this is
258 * a feature that will be implemented much later in the future.
261 struct cgroup_subsys_state css
;
263 * the counter to account for memory usage
265 struct res_counter res
;
267 /* vmpressure notifications */
268 struct vmpressure vmpressure
;
272 * the counter to account for mem+swap usage.
274 struct res_counter memsw
;
277 * rcu_freeing is used only when freeing struct mem_cgroup,
278 * so put it into a union to avoid wasting more memory.
279 * It must be disjoint from the css field. It could be
280 * in a union with the res field, but res plays a much
281 * larger part in mem_cgroup life than memsw, and might
282 * be of interest, even at time of free, when debugging.
283 * So share rcu_head with the less interesting memsw.
285 struct rcu_head rcu_freeing
;
287 * We also need some space for a worker in deferred freeing.
288 * By the time we call it, rcu_freeing is no longer in use.
290 struct work_struct work_freeing
;
294 * the counter to account for kernel memory usage.
296 struct res_counter kmem
;
298 * Should the accounting and control be hierarchical, per subtree?
301 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
305 atomic_t oom_wakeups
;
310 /* OOM-Killer disable */
311 int oom_kill_disable
;
313 /* set when res.limit == memsw.limit */
314 bool memsw_is_minimum
;
316 /* protect arrays of thresholds */
317 struct mutex thresholds_lock
;
319 /* thresholds for memory usage. RCU-protected */
320 struct mem_cgroup_thresholds thresholds
;
322 /* thresholds for mem+swap usage. RCU-protected */
323 struct mem_cgroup_thresholds memsw_thresholds
;
325 /* For oom notifier event fd */
326 struct list_head oom_notify
;
329 * Should we move charges of a task when a task is moved into this
330 * mem_cgroup ? And what type of charges should we move ?
332 unsigned long move_charge_at_immigrate
;
334 * set > 0 if pages under this cgroup are moving to other cgroup.
336 atomic_t moving_account
;
337 /* taken only while moving_account > 0 */
338 spinlock_t move_lock
;
342 struct mem_cgroup_stat_cpu __percpu
*stat
;
344 * used when a cpu is offlined or other synchronizations
345 * See mem_cgroup_read_stat().
347 struct mem_cgroup_stat_cpu nocpu_base
;
348 spinlock_t pcp_counter_lock
;
351 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
352 struct tcp_memcontrol tcp_mem
;
354 #if defined(CONFIG_MEMCG_KMEM)
355 /* analogous to slab_common's slab_caches list. per-memcg */
356 struct list_head memcg_slab_caches
;
357 /* Not a spinlock, we can take a lot of time walking the list */
358 struct mutex slab_caches_mutex
;
359 /* Index in the kmem_cache->memcg_params->memcg_caches array */
363 int last_scanned_node
;
365 nodemask_t scan_nodes
;
366 atomic_t numainfo_events
;
367 atomic_t numainfo_updating
;
371 * Per cgroup active and inactive list, similar to the
372 * per zone LRU lists.
374 * WARNING: This has to be the last element of the struct. Don't
375 * add new fields after this point.
377 struct mem_cgroup_lru_info info
;
380 static size_t memcg_size(void)
382 return sizeof(struct mem_cgroup
) +
383 nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
386 /* internal only representation about the status of kmem accounting. */
388 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
389 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
390 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
393 /* We account when limit is on, but only after call sites are patched */
394 #define KMEM_ACCOUNTED_MASK \
395 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
397 #ifdef CONFIG_MEMCG_KMEM
398 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
400 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
403 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
405 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
408 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
410 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
413 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
415 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
418 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
420 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
421 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
424 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
426 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
427 &memcg
->kmem_account_flags
);
431 /* Stuffs for move charges at task migration. */
433 * Types of charges to be moved. "move_charge_at_immitgrate" and
434 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
437 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
438 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
442 /* "mc" and its members are protected by cgroup_mutex */
443 static struct move_charge_struct
{
444 spinlock_t lock
; /* for from, to */
445 struct mem_cgroup
*from
;
446 struct mem_cgroup
*to
;
447 unsigned long immigrate_flags
;
448 unsigned long precharge
;
449 unsigned long moved_charge
;
450 unsigned long moved_swap
;
451 struct task_struct
*moving_task
; /* a task moving charges */
452 wait_queue_head_t waitq
; /* a waitq for other context */
454 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
455 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
458 static bool move_anon(void)
460 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
463 static bool move_file(void)
465 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
469 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
470 * limit reclaim to prevent infinite loops, if they ever occur.
472 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
473 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
476 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
477 MEM_CGROUP_CHARGE_TYPE_ANON
,
478 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
479 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
483 /* for encoding cft->private value on file */
491 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
492 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
493 #define MEMFILE_ATTR(val) ((val) & 0xffff)
494 /* Used for OOM nofiier */
495 #define OOM_CONTROL (0)
498 * Reclaim flags for mem_cgroup_hierarchical_reclaim
500 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
501 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
502 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
503 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
506 * The memcg_create_mutex will be held whenever a new cgroup is created.
507 * As a consequence, any change that needs to protect against new child cgroups
508 * appearing has to hold it as well.
510 static DEFINE_MUTEX(memcg_create_mutex
);
512 static void mem_cgroup_get(struct mem_cgroup
*memcg
);
513 static void mem_cgroup_put(struct mem_cgroup
*memcg
);
516 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
518 return container_of(s
, struct mem_cgroup
, css
);
521 /* Some nice accessors for the vmpressure. */
522 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
525 memcg
= root_mem_cgroup
;
526 return &memcg
->vmpressure
;
529 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
531 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
534 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
536 return &mem_cgroup_from_css(css
)->vmpressure
;
539 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
541 return (memcg
== root_mem_cgroup
);
545 /* add_to_swap -> get_swap_page_by_memcg -> .. */
546 bool memcg_is_root(struct page
*page
)
548 struct page_cgroup
*pc
;
550 if (mem_cgroup_disabled())
553 pc
= lookup_page_cgroup(page
);
555 return mem_cgroup_is_root(pc
->mem_cgroup
);
559 /* Writing them here to avoid exposing memcg's inner layout */
560 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
562 void sock_update_memcg(struct sock
*sk
)
564 if (mem_cgroup_sockets_enabled
) {
565 struct mem_cgroup
*memcg
;
566 struct cg_proto
*cg_proto
;
568 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
570 /* Socket cloning can throw us here with sk_cgrp already
571 * filled. It won't however, necessarily happen from
572 * process context. So the test for root memcg given
573 * the current task's memcg won't help us in this case.
575 * Respecting the original socket's memcg is a better
576 * decision in this case.
579 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
580 mem_cgroup_get(sk
->sk_cgrp
->memcg
);
585 memcg
= mem_cgroup_from_task(current
);
586 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
587 if (!mem_cgroup_is_root(memcg
) && memcg_proto_active(cg_proto
)) {
588 mem_cgroup_get(memcg
);
589 sk
->sk_cgrp
= cg_proto
;
594 EXPORT_SYMBOL(sock_update_memcg
);
596 void sock_release_memcg(struct sock
*sk
)
598 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
599 struct mem_cgroup
*memcg
;
600 WARN_ON(!sk
->sk_cgrp
->memcg
);
601 memcg
= sk
->sk_cgrp
->memcg
;
602 mem_cgroup_put(memcg
);
606 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
608 if (!memcg
|| mem_cgroup_is_root(memcg
))
611 return &memcg
->tcp_mem
.cg_proto
;
613 EXPORT_SYMBOL(tcp_proto_cgroup
);
615 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
617 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
619 static_key_slow_dec(&memcg_socket_limit_enabled
);
622 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
627 #ifdef CONFIG_MEMCG_KMEM
629 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
630 * There are two main reasons for not using the css_id for this:
631 * 1) this works better in sparse environments, where we have a lot of memcgs,
632 * but only a few kmem-limited. Or also, if we have, for instance, 200
633 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
634 * 200 entry array for that.
636 * 2) In order not to violate the cgroup API, we would like to do all memory
637 * allocation in ->create(). At that point, we haven't yet allocated the
638 * css_id. Having a separate index prevents us from messing with the cgroup
641 * The current size of the caches array is stored in
642 * memcg_limited_groups_array_size. It will double each time we have to
645 static DEFINE_IDA(kmem_limited_groups
);
646 int memcg_limited_groups_array_size
;
649 * MIN_SIZE is different than 1, because we would like to avoid going through
650 * the alloc/free process all the time. In a small machine, 4 kmem-limited
651 * cgroups is a reasonable guess. In the future, it could be a parameter or
652 * tunable, but that is strictly not necessary.
654 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
655 * this constant directly from cgroup, but it is understandable that this is
656 * better kept as an internal representation in cgroup.c. In any case, the
657 * css_id space is not getting any smaller, and we don't have to necessarily
658 * increase ours as well if it increases.
660 #define MEMCG_CACHES_MIN_SIZE 4
661 #define MEMCG_CACHES_MAX_SIZE 65535
664 * A lot of the calls to the cache allocation functions are expected to be
665 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
666 * conditional to this static branch, we'll have to allow modules that does
667 * kmem_cache_alloc and the such to see this symbol as well
669 struct static_key memcg_kmem_enabled_key
;
670 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
672 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
674 if (memcg_kmem_is_active(memcg
)) {
675 static_key_slow_dec(&memcg_kmem_enabled_key
);
676 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
679 * This check can't live in kmem destruction function,
680 * since the charges will outlive the cgroup
682 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
685 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
688 #endif /* CONFIG_MEMCG_KMEM */
690 static void disarm_static_keys(struct mem_cgroup
*memcg
)
692 disarm_sock_keys(memcg
);
693 disarm_kmem_keys(memcg
);
696 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
698 static struct mem_cgroup_per_zone
*
699 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
701 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
702 return &memcg
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
705 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
710 static struct mem_cgroup_per_zone
*
711 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
713 int nid
= page_to_nid(page
);
714 int zid
= page_zonenum(page
);
716 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
719 static struct mem_cgroup_tree_per_zone
*
720 soft_limit_tree_node_zone(int nid
, int zid
)
722 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
725 static struct mem_cgroup_tree_per_zone
*
726 soft_limit_tree_from_page(struct page
*page
)
728 int nid
= page_to_nid(page
);
729 int zid
= page_zonenum(page
);
731 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
735 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
736 struct mem_cgroup_per_zone
*mz
,
737 struct mem_cgroup_tree_per_zone
*mctz
,
738 unsigned long long new_usage_in_excess
)
740 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
741 struct rb_node
*parent
= NULL
;
742 struct mem_cgroup_per_zone
*mz_node
;
747 mz
->usage_in_excess
= new_usage_in_excess
;
748 if (!mz
->usage_in_excess
)
752 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
754 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
757 * We can't avoid mem cgroups that are over their soft
758 * limit by the same amount
760 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
763 rb_link_node(&mz
->tree_node
, parent
, p
);
764 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
769 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
770 struct mem_cgroup_per_zone
*mz
,
771 struct mem_cgroup_tree_per_zone
*mctz
)
775 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
780 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
781 struct mem_cgroup_per_zone
*mz
,
782 struct mem_cgroup_tree_per_zone
*mctz
)
784 spin_lock(&mctz
->lock
);
785 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
786 spin_unlock(&mctz
->lock
);
790 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
792 unsigned long long excess
;
793 struct mem_cgroup_per_zone
*mz
;
794 struct mem_cgroup_tree_per_zone
*mctz
;
795 int nid
= page_to_nid(page
);
796 int zid
= page_zonenum(page
);
797 mctz
= soft_limit_tree_from_page(page
);
800 * Necessary to update all ancestors when hierarchy is used.
801 * because their event counter is not touched.
803 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
804 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
805 excess
= res_counter_soft_limit_excess(&memcg
->res
);
807 * We have to update the tree if mz is on RB-tree or
808 * mem is over its softlimit.
810 if (excess
|| mz
->on_tree
) {
811 spin_lock(&mctz
->lock
);
812 /* if on-tree, remove it */
814 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
816 * Insert again. mz->usage_in_excess will be updated.
817 * If excess is 0, no tree ops.
819 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
820 spin_unlock(&mctz
->lock
);
825 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
828 struct mem_cgroup_per_zone
*mz
;
829 struct mem_cgroup_tree_per_zone
*mctz
;
831 for_each_node(node
) {
832 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
833 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
834 mctz
= soft_limit_tree_node_zone(node
, zone
);
835 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
840 static struct mem_cgroup_per_zone
*
841 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
843 struct rb_node
*rightmost
= NULL
;
844 struct mem_cgroup_per_zone
*mz
;
848 rightmost
= rb_last(&mctz
->rb_root
);
850 goto done
; /* Nothing to reclaim from */
852 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
854 * Remove the node now but someone else can add it back,
855 * we will to add it back at the end of reclaim to its correct
856 * position in the tree.
858 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
859 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
860 !css_tryget(&mz
->memcg
->css
))
866 static struct mem_cgroup_per_zone
*
867 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
869 struct mem_cgroup_per_zone
*mz
;
871 spin_lock(&mctz
->lock
);
872 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
873 spin_unlock(&mctz
->lock
);
878 * Implementation Note: reading percpu statistics for memcg.
880 * Both of vmstat[] and percpu_counter has threshold and do periodic
881 * synchronization to implement "quick" read. There are trade-off between
882 * reading cost and precision of value. Then, we may have a chance to implement
883 * a periodic synchronizion of counter in memcg's counter.
885 * But this _read() function is used for user interface now. The user accounts
886 * memory usage by memory cgroup and he _always_ requires exact value because
887 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
888 * have to visit all online cpus and make sum. So, for now, unnecessary
889 * synchronization is not implemented. (just implemented for cpu hotplug)
891 * If there are kernel internal actions which can make use of some not-exact
892 * value, and reading all cpu value can be performance bottleneck in some
893 * common workload, threashold and synchonization as vmstat[] should be
896 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
897 enum mem_cgroup_stat_index idx
)
903 for_each_online_cpu(cpu
)
904 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
905 #ifdef CONFIG_HOTPLUG_CPU
906 spin_lock(&memcg
->pcp_counter_lock
);
907 val
+= memcg
->nocpu_base
.count
[idx
];
908 spin_unlock(&memcg
->pcp_counter_lock
);
914 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
917 int val
= (charge
) ? 1 : -1;
918 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
921 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
922 enum mem_cgroup_events_index idx
)
924 unsigned long val
= 0;
927 for_each_online_cpu(cpu
)
928 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
929 #ifdef CONFIG_HOTPLUG_CPU
930 spin_lock(&memcg
->pcp_counter_lock
);
931 val
+= memcg
->nocpu_base
.events
[idx
];
932 spin_unlock(&memcg
->pcp_counter_lock
);
937 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
939 bool anon
, int nr_pages
)
944 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
945 * counted as CACHE even if it's on ANON LRU.
948 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
951 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
954 if (PageTransHuge(page
))
955 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
958 /* pagein of a big page is an event. So, ignore page size */
960 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
962 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
963 nr_pages
= -nr_pages
; /* for event */
966 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
972 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
974 struct mem_cgroup_per_zone
*mz
;
976 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
977 return mz
->lru_size
[lru
];
981 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
982 unsigned int lru_mask
)
984 struct mem_cgroup_per_zone
*mz
;
986 unsigned long ret
= 0;
988 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
991 if (BIT(lru
) & lru_mask
)
992 ret
+= mz
->lru_size
[lru
];
998 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
999 int nid
, unsigned int lru_mask
)
1004 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
1005 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
1006 nid
, zid
, lru_mask
);
1011 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
1012 unsigned int lru_mask
)
1017 for_each_node_state(nid
, N_MEMORY
)
1018 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
1022 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
1023 enum mem_cgroup_events_target target
)
1025 unsigned long val
, next
;
1027 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
1028 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
1029 /* from time_after() in jiffies.h */
1030 if ((long)next
- (long)val
< 0) {
1032 case MEM_CGROUP_TARGET_THRESH
:
1033 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
1035 case MEM_CGROUP_TARGET_SOFTLIMIT
:
1036 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1038 case MEM_CGROUP_TARGET_NUMAINFO
:
1039 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1044 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1051 * Check events in order.
1054 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1057 /* threshold event is triggered in finer grain than soft limit */
1058 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1059 MEM_CGROUP_TARGET_THRESH
))) {
1061 bool do_numainfo __maybe_unused
;
1063 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1064 MEM_CGROUP_TARGET_SOFTLIMIT
);
1065 #if MAX_NUMNODES > 1
1066 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1067 MEM_CGROUP_TARGET_NUMAINFO
);
1071 mem_cgroup_threshold(memcg
);
1072 if (unlikely(do_softlimit
))
1073 mem_cgroup_update_tree(memcg
, page
);
1074 #if MAX_NUMNODES > 1
1075 if (unlikely(do_numainfo
))
1076 atomic_inc(&memcg
->numainfo_events
);
1082 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
1084 return mem_cgroup_from_css(
1085 cgroup_subsys_state(cont
, mem_cgroup_subsys_id
));
1088 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1091 * mm_update_next_owner() may clear mm->owner to NULL
1092 * if it races with swapoff, page migration, etc.
1093 * So this can be called with p == NULL.
1098 return mem_cgroup_from_css(task_subsys_state(p
, mem_cgroup_subsys_id
));
1101 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1103 struct mem_cgroup
*memcg
= NULL
;
1108 * Because we have no locks, mm->owner's may be being moved to other
1109 * cgroup. We use css_tryget() here even if this looks
1110 * pessimistic (rather than adding locks here).
1114 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1115 if (unlikely(!memcg
))
1117 } while (!css_tryget(&memcg
->css
));
1123 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1124 * ref. count) or NULL if the whole root's subtree has been visited.
1126 * helper function to be used by mem_cgroup_iter
1128 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1129 struct mem_cgroup
*last_visited
)
1131 struct cgroup
*prev_cgroup
, *next_cgroup
;
1134 * Root is not visited by cgroup iterators so it needs an
1140 prev_cgroup
= (last_visited
== root
) ? NULL
1141 : last_visited
->css
.cgroup
;
1143 next_cgroup
= cgroup_next_descendant_pre(
1144 prev_cgroup
, root
->css
.cgroup
);
1147 * Even if we found a group we have to make sure it is
1148 * alive. css && !memcg means that the groups should be
1149 * skipped and we should continue the tree walk.
1150 * last_visited css is safe to use because it is
1151 * protected by css_get and the tree walk is rcu safe.
1154 struct mem_cgroup
*mem
= mem_cgroup_from_cont(
1156 if (css_tryget(&mem
->css
))
1159 prev_cgroup
= next_cgroup
;
1168 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1169 * @root: hierarchy root
1170 * @prev: previously returned memcg, NULL on first invocation
1171 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1173 * Returns references to children of the hierarchy below @root, or
1174 * @root itself, or %NULL after a full round-trip.
1176 * Caller must pass the return value in @prev on subsequent
1177 * invocations for reference counting, or use mem_cgroup_iter_break()
1178 * to cancel a hierarchy walk before the round-trip is complete.
1180 * Reclaimers can specify a zone and a priority level in @reclaim to
1181 * divide up the memcgs in the hierarchy among all concurrent
1182 * reclaimers operating on the same zone and priority.
1184 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1185 struct mem_cgroup
*prev
,
1186 struct mem_cgroup_reclaim_cookie
*reclaim
)
1188 struct mem_cgroup
*memcg
= NULL
;
1189 struct mem_cgroup
*last_visited
= NULL
;
1190 unsigned long uninitialized_var(dead_count
);
1192 if (mem_cgroup_disabled())
1196 root
= root_mem_cgroup
;
1198 if (prev
&& !reclaim
)
1199 last_visited
= prev
;
1201 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1209 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1212 int nid
= zone_to_nid(reclaim
->zone
);
1213 int zid
= zone_idx(reclaim
->zone
);
1214 struct mem_cgroup_per_zone
*mz
;
1216 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1217 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1218 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1219 iter
->last_visited
= NULL
;
1224 * If the dead_count mismatches, a destruction
1225 * has happened or is happening concurrently.
1226 * If the dead_count matches, a destruction
1227 * might still happen concurrently, but since
1228 * we checked under RCU, that destruction
1229 * won't free the object until we release the
1230 * RCU reader lock. Thus, the dead_count
1231 * check verifies the pointer is still valid,
1232 * css_tryget() verifies the cgroup pointed to
1235 dead_count
= atomic_read(&root
->dead_count
);
1236 if (dead_count
== iter
->last_dead_count
) {
1238 last_visited
= iter
->last_visited
;
1239 if (last_visited
&& last_visited
!= root
&&
1240 !css_tryget(&last_visited
->css
))
1241 last_visited
= NULL
;
1245 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1248 if (last_visited
&& last_visited
!= root
)
1249 css_put(&last_visited
->css
);
1251 iter
->last_visited
= memcg
;
1253 iter
->last_dead_count
= dead_count
;
1257 else if (!prev
&& memcg
)
1258 reclaim
->generation
= iter
->generation
;
1267 if (prev
&& prev
!= root
)
1268 css_put(&prev
->css
);
1274 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1275 * @root: hierarchy root
1276 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1278 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1279 struct mem_cgroup
*prev
)
1282 root
= root_mem_cgroup
;
1283 if (prev
&& prev
!= root
)
1284 css_put(&prev
->css
);
1288 * Iteration constructs for visiting all cgroups (under a tree). If
1289 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1290 * be used for reference counting.
1292 #define for_each_mem_cgroup_tree(iter, root) \
1293 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1295 iter = mem_cgroup_iter(root, iter, NULL))
1297 #define for_each_mem_cgroup(iter) \
1298 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1300 iter = mem_cgroup_iter(NULL, iter, NULL))
1302 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1304 struct mem_cgroup
*memcg
;
1307 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1308 if (unlikely(!memcg
))
1313 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1316 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1324 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1327 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1328 * @zone: zone of the wanted lruvec
1329 * @memcg: memcg of the wanted lruvec
1331 * Returns the lru list vector holding pages for the given @zone and
1332 * @mem. This can be the global zone lruvec, if the memory controller
1335 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1336 struct mem_cgroup
*memcg
)
1338 struct mem_cgroup_per_zone
*mz
;
1339 struct lruvec
*lruvec
;
1341 if (mem_cgroup_disabled()) {
1342 lruvec
= &zone
->lruvec
;
1346 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1347 lruvec
= &mz
->lruvec
;
1350 * Since a node can be onlined after the mem_cgroup was created,
1351 * we have to be prepared to initialize lruvec->zone here;
1352 * and if offlined then reonlined, we need to reinitialize it.
1354 if (unlikely(lruvec
->zone
!= zone
))
1355 lruvec
->zone
= zone
;
1360 * Following LRU functions are allowed to be used without PCG_LOCK.
1361 * Operations are called by routine of global LRU independently from memcg.
1362 * What we have to take care of here is validness of pc->mem_cgroup.
1364 * Changes to pc->mem_cgroup happens when
1367 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1368 * It is added to LRU before charge.
1369 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1370 * When moving account, the page is not on LRU. It's isolated.
1374 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1376 * @zone: zone of the page
1378 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1380 struct mem_cgroup_per_zone
*mz
;
1381 struct mem_cgroup
*memcg
;
1382 struct page_cgroup
*pc
;
1383 struct lruvec
*lruvec
;
1385 if (mem_cgroup_disabled()) {
1386 lruvec
= &zone
->lruvec
;
1390 pc
= lookup_page_cgroup(page
);
1391 memcg
= pc
->mem_cgroup
;
1394 * Surreptitiously switch any uncharged offlist page to root:
1395 * an uncharged page off lru does nothing to secure
1396 * its former mem_cgroup from sudden removal.
1398 * Our caller holds lru_lock, and PageCgroupUsed is updated
1399 * under page_cgroup lock: between them, they make all uses
1400 * of pc->mem_cgroup safe.
1402 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1403 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1405 mz
= page_cgroup_zoneinfo(memcg
, page
);
1406 lruvec
= &mz
->lruvec
;
1409 * Since a node can be onlined after the mem_cgroup was created,
1410 * we have to be prepared to initialize lruvec->zone here;
1411 * and if offlined then reonlined, we need to reinitialize it.
1413 if (unlikely(lruvec
->zone
!= zone
))
1414 lruvec
->zone
= zone
;
1419 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1420 * @lruvec: mem_cgroup per zone lru vector
1421 * @lru: index of lru list the page is sitting on
1422 * @nr_pages: positive when adding or negative when removing
1424 * This function must be called when a page is added to or removed from an
1427 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1430 struct mem_cgroup_per_zone
*mz
;
1431 unsigned long *lru_size
;
1433 if (mem_cgroup_disabled())
1436 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1437 lru_size
= mz
->lru_size
+ lru
;
1438 *lru_size
+= nr_pages
;
1439 VM_BUG_ON((long)(*lru_size
) < 0);
1443 * Checks whether given mem is same or in the root_mem_cgroup's
1446 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1447 struct mem_cgroup
*memcg
)
1449 if (root_memcg
== memcg
)
1451 if (!root_memcg
->use_hierarchy
|| !memcg
)
1453 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1456 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1457 struct mem_cgroup
*memcg
)
1462 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1467 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*memcg
)
1470 struct mem_cgroup
*curr
= NULL
;
1471 struct task_struct
*p
;
1473 p
= find_lock_task_mm(task
);
1475 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1479 * All threads may have already detached their mm's, but the oom
1480 * killer still needs to detect if they have already been oom
1481 * killed to prevent needlessly killing additional tasks.
1484 curr
= mem_cgroup_from_task(task
);
1486 css_get(&curr
->css
);
1492 * We should check use_hierarchy of "memcg" not "curr". Because checking
1493 * use_hierarchy of "curr" here make this function true if hierarchy is
1494 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1495 * hierarchy(even if use_hierarchy is disabled in "memcg").
1497 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1498 css_put(&curr
->css
);
1502 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1504 unsigned long inactive_ratio
;
1505 unsigned long inactive
;
1506 unsigned long active
;
1509 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1510 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1512 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1514 inactive_ratio
= int_sqrt(10 * gb
);
1518 return inactive
* inactive_ratio
< active
;
1521 #define mem_cgroup_from_res_counter(counter, member) \
1522 container_of(counter, struct mem_cgroup, member)
1525 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1526 * @memcg: the memory cgroup
1528 * Returns the maximum amount of memory @mem can be charged with, in
1531 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1533 unsigned long long margin
;
1535 margin
= res_counter_margin(&memcg
->res
);
1536 if (do_swap_account
)
1537 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1538 return margin
>> PAGE_SHIFT
;
1541 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1543 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1546 if (cgrp
->parent
== NULL
)
1547 return vm_swappiness
;
1549 return memcg
->swappiness
;
1553 * memcg->moving_account is used for checking possibility that some thread is
1554 * calling move_account(). When a thread on CPU-A starts moving pages under
1555 * a memcg, other threads should check memcg->moving_account under
1556 * rcu_read_lock(), like this:
1560 * memcg->moving_account+1 if (memcg->mocing_account)
1562 * synchronize_rcu() update something.
1567 /* for quick checking without looking up memcg */
1568 atomic_t memcg_moving __read_mostly
;
1570 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1572 atomic_inc(&memcg_moving
);
1573 atomic_inc(&memcg
->moving_account
);
1577 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1580 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1581 * We check NULL in callee rather than caller.
1584 atomic_dec(&memcg_moving
);
1585 atomic_dec(&memcg
->moving_account
);
1590 * 2 routines for checking "mem" is under move_account() or not.
1592 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1593 * is used for avoiding races in accounting. If true,
1594 * pc->mem_cgroup may be overwritten.
1596 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1597 * under hierarchy of moving cgroups. This is for
1598 * waiting at hith-memory prressure caused by "move".
1601 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1603 VM_BUG_ON(!rcu_read_lock_held());
1604 return atomic_read(&memcg
->moving_account
) > 0;
1607 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1609 struct mem_cgroup
*from
;
1610 struct mem_cgroup
*to
;
1613 * Unlike task_move routines, we access mc.to, mc.from not under
1614 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1616 spin_lock(&mc
.lock
);
1622 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1623 || mem_cgroup_same_or_subtree(memcg
, to
);
1625 spin_unlock(&mc
.lock
);
1629 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1631 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1632 if (mem_cgroup_under_move(memcg
)) {
1634 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1635 /* moving charge context might have finished. */
1638 finish_wait(&mc
.waitq
, &wait
);
1646 * Take this lock when
1647 * - a code tries to modify page's memcg while it's USED.
1648 * - a code tries to modify page state accounting in a memcg.
1649 * see mem_cgroup_stolen(), too.
1651 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1652 unsigned long *flags
)
1654 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1657 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1658 unsigned long *flags
)
1660 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1663 #define K(x) ((x) << (PAGE_SHIFT-10))
1665 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1666 * @memcg: The memory cgroup that went over limit
1667 * @p: Task that is going to be killed
1669 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1672 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1674 struct cgroup
*task_cgrp
;
1675 struct cgroup
*mem_cgrp
;
1677 * Need a buffer in BSS, can't rely on allocations. The code relies
1678 * on the assumption that OOM is serialized for memory controller.
1679 * If this assumption is broken, revisit this code.
1681 static char memcg_name
[PATH_MAX
];
1683 struct mem_cgroup
*iter
;
1691 mem_cgrp
= memcg
->css
.cgroup
;
1692 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1694 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1697 * Unfortunately, we are unable to convert to a useful name
1698 * But we'll still print out the usage information
1705 pr_info("Task in %s killed", memcg_name
);
1708 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1716 * Continues from above, so we don't need an KERN_ level
1718 pr_cont(" as a result of limit of %s\n", memcg_name
);
1721 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1722 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1723 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1724 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1725 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1726 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1727 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1728 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1729 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1730 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1731 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1732 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1734 for_each_mem_cgroup_tree(iter
, memcg
) {
1735 pr_info("Memory cgroup stats");
1738 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1740 pr_cont(" for %s", memcg_name
);
1744 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1745 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1747 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1748 K(mem_cgroup_read_stat(iter
, i
)));
1751 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1752 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1753 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1760 * This function returns the number of memcg under hierarchy tree. Returns
1761 * 1(self count) if no children.
1763 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1766 struct mem_cgroup
*iter
;
1768 for_each_mem_cgroup_tree(iter
, memcg
)
1774 * Return the memory (and swap, if configured) limit for a memcg.
1776 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1780 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1783 * Do not consider swap space if we cannot swap due to swappiness
1785 if (mem_cgroup_swappiness(memcg
)) {
1788 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1789 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1792 * If memsw is finite and limits the amount of swap space
1793 * available to this memcg, return that limit.
1795 limit
= min(limit
, memsw
);
1801 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1804 struct mem_cgroup
*iter
;
1805 unsigned long chosen_points
= 0;
1806 unsigned long totalpages
;
1807 unsigned int points
= 0;
1808 struct task_struct
*chosen
= NULL
;
1811 * If current has a pending SIGKILL or is exiting, then automatically
1812 * select it. The goal is to allow it to allocate so that it may
1813 * quickly exit and free its memory.
1815 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1816 set_thread_flag(TIF_MEMDIE
);
1820 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1821 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1822 for_each_mem_cgroup_tree(iter
, memcg
) {
1823 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1824 struct cgroup_iter it
;
1825 struct task_struct
*task
;
1827 cgroup_iter_start(cgroup
, &it
);
1828 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1829 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1831 case OOM_SCAN_SELECT
:
1833 put_task_struct(chosen
);
1835 chosen_points
= ULONG_MAX
;
1836 get_task_struct(chosen
);
1838 case OOM_SCAN_CONTINUE
:
1840 case OOM_SCAN_ABORT
:
1841 cgroup_iter_end(cgroup
, &it
);
1842 mem_cgroup_iter_break(memcg
, iter
);
1844 put_task_struct(chosen
);
1849 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1850 if (points
> chosen_points
) {
1852 put_task_struct(chosen
);
1854 chosen_points
= points
;
1855 get_task_struct(chosen
);
1858 cgroup_iter_end(cgroup
, &it
);
1863 points
= chosen_points
* 1000 / totalpages
;
1864 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1865 NULL
, "Memory cgroup out of memory");
1868 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1870 unsigned long flags
)
1872 unsigned long total
= 0;
1873 bool noswap
= false;
1876 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1878 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1881 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1883 drain_all_stock_async(memcg
);
1884 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1886 * Allow limit shrinkers, which are triggered directly
1887 * by userspace, to catch signals and stop reclaim
1888 * after minimal progress, regardless of the margin.
1890 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1892 if (mem_cgroup_margin(memcg
))
1895 * If nothing was reclaimed after two attempts, there
1896 * may be no reclaimable pages in this hierarchy.
1905 * test_mem_cgroup_node_reclaimable
1906 * @memcg: the target memcg
1907 * @nid: the node ID to be checked.
1908 * @noswap : specify true here if the user wants flle only information.
1910 * This function returns whether the specified memcg contains any
1911 * reclaimable pages on a node. Returns true if there are any reclaimable
1912 * pages in the node.
1914 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1915 int nid
, bool noswap
)
1917 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1919 if (noswap
|| !total_swap_pages
)
1921 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1926 #if MAX_NUMNODES > 1
1929 * Always updating the nodemask is not very good - even if we have an empty
1930 * list or the wrong list here, we can start from some node and traverse all
1931 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1934 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1938 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1939 * pagein/pageout changes since the last update.
1941 if (!atomic_read(&memcg
->numainfo_events
))
1943 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1946 /* make a nodemask where this memcg uses memory from */
1947 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1949 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1951 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1952 node_clear(nid
, memcg
->scan_nodes
);
1955 atomic_set(&memcg
->numainfo_events
, 0);
1956 atomic_set(&memcg
->numainfo_updating
, 0);
1960 * Selecting a node where we start reclaim from. Because what we need is just
1961 * reducing usage counter, start from anywhere is O,K. Considering
1962 * memory reclaim from current node, there are pros. and cons.
1964 * Freeing memory from current node means freeing memory from a node which
1965 * we'll use or we've used. So, it may make LRU bad. And if several threads
1966 * hit limits, it will see a contention on a node. But freeing from remote
1967 * node means more costs for memory reclaim because of memory latency.
1969 * Now, we use round-robin. Better algorithm is welcomed.
1971 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1975 mem_cgroup_may_update_nodemask(memcg
);
1976 node
= memcg
->last_scanned_node
;
1978 node
= next_node(node
, memcg
->scan_nodes
);
1979 if (node
== MAX_NUMNODES
)
1980 node
= first_node(memcg
->scan_nodes
);
1982 * We call this when we hit limit, not when pages are added to LRU.
1983 * No LRU may hold pages because all pages are UNEVICTABLE or
1984 * memcg is too small and all pages are not on LRU. In that case,
1985 * we use curret node.
1987 if (unlikely(node
== MAX_NUMNODES
))
1988 node
= numa_node_id();
1990 memcg
->last_scanned_node
= node
;
1995 * Check all nodes whether it contains reclaimable pages or not.
1996 * For quick scan, we make use of scan_nodes. This will allow us to skip
1997 * unused nodes. But scan_nodes is lazily updated and may not cotain
1998 * enough new information. We need to do double check.
2000 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2005 * quick check...making use of scan_node.
2006 * We can skip unused nodes.
2008 if (!nodes_empty(memcg
->scan_nodes
)) {
2009 for (nid
= first_node(memcg
->scan_nodes
);
2011 nid
= next_node(nid
, memcg
->scan_nodes
)) {
2013 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2018 * Check rest of nodes.
2020 for_each_node_state(nid
, N_MEMORY
) {
2021 if (node_isset(nid
, memcg
->scan_nodes
))
2023 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2030 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2035 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2037 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2041 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2044 unsigned long *total_scanned
)
2046 struct mem_cgroup
*victim
= NULL
;
2049 unsigned long excess
;
2050 unsigned long nr_scanned
;
2051 struct mem_cgroup_reclaim_cookie reclaim
= {
2056 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2059 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2064 * If we have not been able to reclaim
2065 * anything, it might because there are
2066 * no reclaimable pages under this hierarchy
2071 * We want to do more targeted reclaim.
2072 * excess >> 2 is not to excessive so as to
2073 * reclaim too much, nor too less that we keep
2074 * coming back to reclaim from this cgroup
2076 if (total
>= (excess
>> 2) ||
2077 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2082 if (!mem_cgroup_reclaimable(victim
, false))
2084 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2086 *total_scanned
+= nr_scanned
;
2087 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2090 mem_cgroup_iter_break(root_memcg
, victim
);
2094 static DEFINE_SPINLOCK(memcg_oom_lock
);
2097 * Check OOM-Killer is already running under our hierarchy.
2098 * If someone is running, return false.
2100 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2102 struct mem_cgroup
*iter
, *failed
= NULL
;
2104 spin_lock(&memcg_oom_lock
);
2106 for_each_mem_cgroup_tree(iter
, memcg
) {
2107 if (iter
->oom_lock
) {
2109 * this subtree of our hierarchy is already locked
2110 * so we cannot give a lock.
2113 mem_cgroup_iter_break(memcg
, iter
);
2116 iter
->oom_lock
= true;
2121 * OK, we failed to lock the whole subtree so we have
2122 * to clean up what we set up to the failing subtree
2124 for_each_mem_cgroup_tree(iter
, memcg
) {
2125 if (iter
== failed
) {
2126 mem_cgroup_iter_break(memcg
, iter
);
2129 iter
->oom_lock
= false;
2133 spin_unlock(&memcg_oom_lock
);
2138 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2140 struct mem_cgroup
*iter
;
2142 spin_lock(&memcg_oom_lock
);
2143 for_each_mem_cgroup_tree(iter
, memcg
)
2144 iter
->oom_lock
= false;
2145 spin_unlock(&memcg_oom_lock
);
2148 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2150 struct mem_cgroup
*iter
;
2152 for_each_mem_cgroup_tree(iter
, memcg
)
2153 atomic_inc(&iter
->under_oom
);
2156 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2158 struct mem_cgroup
*iter
;
2161 * When a new child is created while the hierarchy is under oom,
2162 * mem_cgroup_oom_lock() may not be called. We have to use
2163 * atomic_add_unless() here.
2165 for_each_mem_cgroup_tree(iter
, memcg
)
2166 atomic_add_unless(&iter
->under_oom
, -1, 0);
2169 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2171 struct oom_wait_info
{
2172 struct mem_cgroup
*memcg
;
2176 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2177 unsigned mode
, int sync
, void *arg
)
2179 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2180 struct mem_cgroup
*oom_wait_memcg
;
2181 struct oom_wait_info
*oom_wait_info
;
2183 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2184 oom_wait_memcg
= oom_wait_info
->memcg
;
2187 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2188 * Then we can use css_is_ancestor without taking care of RCU.
2190 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2191 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2193 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2196 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2198 atomic_inc(&memcg
->oom_wakeups
);
2199 /* for filtering, pass "memcg" as argument. */
2200 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2203 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2205 if (memcg
&& atomic_read(&memcg
->under_oom
))
2206 memcg_wakeup_oom(memcg
);
2209 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2211 if (!current
->memcg_oom
.may_oom
)
2214 * We are in the middle of the charge context here, so we
2215 * don't want to block when potentially sitting on a callstack
2216 * that holds all kinds of filesystem and mm locks.
2218 * Also, the caller may handle a failed allocation gracefully
2219 * (like optional page cache readahead) and so an OOM killer
2220 * invocation might not even be necessary.
2222 * That's why we don't do anything here except remember the
2223 * OOM context and then deal with it at the end of the page
2224 * fault when the stack is unwound, the locks are released,
2225 * and when we know whether the fault was overall successful.
2227 css_get(&memcg
->css
);
2228 current
->memcg_oom
.memcg
= memcg
;
2229 current
->memcg_oom
.gfp_mask
= mask
;
2230 current
->memcg_oom
.order
= order
;
2234 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2235 * @handle: actually kill/wait or just clean up the OOM state
2237 * This has to be called at the end of a page fault if the memcg OOM
2238 * handler was enabled.
2240 * Memcg supports userspace OOM handling where failed allocations must
2241 * sleep on a waitqueue until the userspace task resolves the
2242 * situation. Sleeping directly in the charge context with all kinds
2243 * of locks held is not a good idea, instead we remember an OOM state
2244 * in the task and mem_cgroup_oom_synchronize() has to be called at
2245 * the end of the page fault to complete the OOM handling.
2247 * Returns %true if an ongoing memcg OOM situation was detected and
2248 * completed, %false otherwise.
2250 bool mem_cgroup_oom_synchronize(bool handle
)
2252 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2253 struct oom_wait_info owait
;
2256 /* OOM is global, do not handle */
2263 owait
.memcg
= memcg
;
2264 owait
.wait
.flags
= 0;
2265 owait
.wait
.func
= memcg_oom_wake_function
;
2266 owait
.wait
.private = current
;
2267 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2269 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2270 mem_cgroup_mark_under_oom(memcg
);
2272 locked
= mem_cgroup_oom_trylock(memcg
);
2275 mem_cgroup_oom_notify(memcg
);
2277 if (locked
&& !memcg
->oom_kill_disable
) {
2278 mem_cgroup_unmark_under_oom(memcg
);
2279 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2280 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2281 current
->memcg_oom
.order
);
2284 mem_cgroup_unmark_under_oom(memcg
);
2285 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2289 mem_cgroup_oom_unlock(memcg
);
2291 * There is no guarantee that an OOM-lock contender
2292 * sees the wakeups triggered by the OOM kill
2293 * uncharges. Wake any sleepers explicitely.
2295 memcg_oom_recover(memcg
);
2298 current
->memcg_oom
.memcg
= NULL
;
2299 css_put(&memcg
->css
);
2304 * Currently used to update mapped file statistics, but the routine can be
2305 * generalized to update other statistics as well.
2307 * Notes: Race condition
2309 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2310 * it tends to be costly. But considering some conditions, we doesn't need
2311 * to do so _always_.
2313 * Considering "charge", lock_page_cgroup() is not required because all
2314 * file-stat operations happen after a page is attached to radix-tree. There
2315 * are no race with "charge".
2317 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2318 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2319 * if there are race with "uncharge". Statistics itself is properly handled
2322 * Considering "move", this is an only case we see a race. To make the race
2323 * small, we check mm->moving_account and detect there are possibility of race
2324 * If there is, we take a lock.
2327 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2328 bool *locked
, unsigned long *flags
)
2330 struct mem_cgroup
*memcg
;
2331 struct page_cgroup
*pc
;
2333 pc
= lookup_page_cgroup(page
);
2335 memcg
= pc
->mem_cgroup
;
2336 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2339 * If this memory cgroup is not under account moving, we don't
2340 * need to take move_lock_mem_cgroup(). Because we already hold
2341 * rcu_read_lock(), any calls to move_account will be delayed until
2342 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2344 if (!mem_cgroup_stolen(memcg
))
2347 move_lock_mem_cgroup(memcg
, flags
);
2348 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2349 move_unlock_mem_cgroup(memcg
, flags
);
2355 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2357 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2360 * It's guaranteed that pc->mem_cgroup never changes while
2361 * lock is held because a routine modifies pc->mem_cgroup
2362 * should take move_lock_mem_cgroup().
2364 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2367 void mem_cgroup_update_page_stat(struct page
*page
,
2368 enum mem_cgroup_page_stat_item idx
, int val
)
2370 struct mem_cgroup
*memcg
;
2371 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2372 unsigned long uninitialized_var(flags
);
2374 if (mem_cgroup_disabled())
2377 memcg
= pc
->mem_cgroup
;
2378 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2382 case MEMCG_NR_FILE_MAPPED
:
2383 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2389 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2393 * size of first charge trial. "32" comes from vmscan.c's magic value.
2394 * TODO: maybe necessary to use big numbers in big irons.
2396 #define CHARGE_BATCH 32U
2397 struct memcg_stock_pcp
{
2398 struct mem_cgroup
*cached
; /* this never be root cgroup */
2399 unsigned int nr_pages
;
2400 struct work_struct work
;
2401 unsigned long flags
;
2402 #define FLUSHING_CACHED_CHARGE 0
2404 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2405 static DEFINE_MUTEX(percpu_charge_mutex
);
2408 * consume_stock: Try to consume stocked charge on this cpu.
2409 * @memcg: memcg to consume from.
2410 * @nr_pages: how many pages to charge.
2412 * The charges will only happen if @memcg matches the current cpu's memcg
2413 * stock, and at least @nr_pages are available in that stock. Failure to
2414 * service an allocation will refill the stock.
2416 * returns true if successful, false otherwise.
2418 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2420 struct memcg_stock_pcp
*stock
;
2423 if (nr_pages
> CHARGE_BATCH
)
2426 stock
= &get_cpu_var(memcg_stock
);
2427 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2428 stock
->nr_pages
-= nr_pages
;
2429 else /* need to call res_counter_charge */
2431 put_cpu_var(memcg_stock
);
2436 * Returns stocks cached in percpu to res_counter and reset cached information.
2438 static void drain_stock(struct memcg_stock_pcp
*stock
)
2440 struct mem_cgroup
*old
= stock
->cached
;
2442 if (stock
->nr_pages
) {
2443 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2445 res_counter_uncharge(&old
->res
, bytes
);
2446 if (do_swap_account
)
2447 res_counter_uncharge(&old
->memsw
, bytes
);
2448 stock
->nr_pages
= 0;
2450 stock
->cached
= NULL
;
2454 * This must be called under preempt disabled or must be called by
2455 * a thread which is pinned to local cpu.
2457 static void drain_local_stock(struct work_struct
*dummy
)
2459 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2461 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2464 static void __init
memcg_stock_init(void)
2468 for_each_possible_cpu(cpu
) {
2469 struct memcg_stock_pcp
*stock
=
2470 &per_cpu(memcg_stock
, cpu
);
2471 INIT_WORK(&stock
->work
, drain_local_stock
);
2476 * Cache charges(val) which is from res_counter, to local per_cpu area.
2477 * This will be consumed by consume_stock() function, later.
2479 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2481 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2483 if (stock
->cached
!= memcg
) { /* reset if necessary */
2485 stock
->cached
= memcg
;
2487 stock
->nr_pages
+= nr_pages
;
2488 put_cpu_var(memcg_stock
);
2492 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2493 * of the hierarchy under it. sync flag says whether we should block
2494 * until the work is done.
2496 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2500 /* Notify other cpus that system-wide "drain" is running */
2503 for_each_online_cpu(cpu
) {
2504 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2505 struct mem_cgroup
*memcg
;
2507 memcg
= stock
->cached
;
2508 if (!memcg
|| !stock
->nr_pages
)
2510 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2512 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2514 drain_local_stock(&stock
->work
);
2516 schedule_work_on(cpu
, &stock
->work
);
2524 for_each_online_cpu(cpu
) {
2525 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2526 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2527 flush_work(&stock
->work
);
2534 * Tries to drain stocked charges in other cpus. This function is asynchronous
2535 * and just put a work per cpu for draining localy on each cpu. Caller can
2536 * expects some charges will be back to res_counter later but cannot wait for
2539 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2542 * If someone calls draining, avoid adding more kworker runs.
2544 if (!mutex_trylock(&percpu_charge_mutex
))
2546 drain_all_stock(root_memcg
, false);
2547 mutex_unlock(&percpu_charge_mutex
);
2550 /* This is a synchronous drain interface. */
2551 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2553 /* called when force_empty is called */
2554 mutex_lock(&percpu_charge_mutex
);
2555 drain_all_stock(root_memcg
, true);
2556 mutex_unlock(&percpu_charge_mutex
);
2560 * This function drains percpu counter value from DEAD cpu and
2561 * move it to local cpu. Note that this function can be preempted.
2563 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2567 spin_lock(&memcg
->pcp_counter_lock
);
2568 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2569 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2571 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2572 memcg
->nocpu_base
.count
[i
] += x
;
2574 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2575 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2577 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2578 memcg
->nocpu_base
.events
[i
] += x
;
2580 spin_unlock(&memcg
->pcp_counter_lock
);
2583 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2584 unsigned long action
,
2587 int cpu
= (unsigned long)hcpu
;
2588 struct memcg_stock_pcp
*stock
;
2589 struct mem_cgroup
*iter
;
2591 if (action
== CPU_ONLINE
)
2594 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2597 for_each_mem_cgroup(iter
)
2598 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2600 stock
= &per_cpu(memcg_stock
, cpu
);
2606 /* See __mem_cgroup_try_charge() for details */
2608 CHARGE_OK
, /* success */
2609 CHARGE_RETRY
, /* need to retry but retry is not bad */
2610 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2611 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2614 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2615 unsigned int nr_pages
, unsigned int min_pages
,
2618 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2619 struct mem_cgroup
*mem_over_limit
;
2620 struct res_counter
*fail_res
;
2621 unsigned long flags
= 0;
2624 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2627 if (!do_swap_account
)
2629 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2633 res_counter_uncharge(&memcg
->res
, csize
);
2634 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2635 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2637 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2639 * Never reclaim on behalf of optional batching, retry with a
2640 * single page instead.
2642 if (nr_pages
> min_pages
)
2643 return CHARGE_RETRY
;
2645 if (!(gfp_mask
& __GFP_WAIT
))
2646 return CHARGE_WOULDBLOCK
;
2648 if (gfp_mask
& __GFP_NORETRY
)
2649 return CHARGE_NOMEM
;
2651 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2652 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2653 return CHARGE_RETRY
;
2655 * Even though the limit is exceeded at this point, reclaim
2656 * may have been able to free some pages. Retry the charge
2657 * before killing the task.
2659 * Only for regular pages, though: huge pages are rather
2660 * unlikely to succeed so close to the limit, and we fall back
2661 * to regular pages anyway in case of failure.
2663 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2664 return CHARGE_RETRY
;
2667 * At task move, charge accounts can be doubly counted. So, it's
2668 * better to wait until the end of task_move if something is going on.
2670 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2671 return CHARGE_RETRY
;
2674 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2676 return CHARGE_NOMEM
;
2680 * __mem_cgroup_try_charge() does
2681 * 1. detect memcg to be charged against from passed *mm and *ptr,
2682 * 2. update res_counter
2683 * 3. call memory reclaim if necessary.
2685 * In some special case, if the task is fatal, fatal_signal_pending() or
2686 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2687 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2688 * as possible without any hazards. 2: all pages should have a valid
2689 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2690 * pointer, that is treated as a charge to root_mem_cgroup.
2692 * So __mem_cgroup_try_charge() will return
2693 * 0 ... on success, filling *ptr with a valid memcg pointer.
2694 * -ENOMEM ... charge failure because of resource limits.
2695 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2697 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2698 * the oom-killer can be invoked.
2700 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2702 unsigned int nr_pages
,
2703 struct mem_cgroup
**ptr
,
2706 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2707 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2708 struct mem_cgroup
*memcg
= NULL
;
2712 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2713 * in system level. So, allow to go ahead dying process in addition to
2716 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2717 || fatal_signal_pending(current
)))
2720 if (unlikely(task_in_memcg_oom(current
)))
2724 * We always charge the cgroup the mm_struct belongs to.
2725 * The mm_struct's mem_cgroup changes on task migration if the
2726 * thread group leader migrates. It's possible that mm is not
2727 * set, if so charge the root memcg (happens for pagecache usage).
2730 *ptr
= root_mem_cgroup
;
2732 if (*ptr
) { /* css should be a valid one */
2734 if (mem_cgroup_is_root(memcg
))
2736 if (consume_stock(memcg
, nr_pages
))
2738 css_get(&memcg
->css
);
2740 struct task_struct
*p
;
2743 p
= rcu_dereference(mm
->owner
);
2745 * Because we don't have task_lock(), "p" can exit.
2746 * In that case, "memcg" can point to root or p can be NULL with
2747 * race with swapoff. Then, we have small risk of mis-accouning.
2748 * But such kind of mis-account by race always happens because
2749 * we don't have cgroup_mutex(). It's overkill and we allo that
2751 * (*) swapoff at el will charge against mm-struct not against
2752 * task-struct. So, mm->owner can be NULL.
2754 memcg
= mem_cgroup_from_task(p
);
2756 memcg
= root_mem_cgroup
;
2757 if (mem_cgroup_is_root(memcg
)) {
2761 if (consume_stock(memcg
, nr_pages
)) {
2763 * It seems dagerous to access memcg without css_get().
2764 * But considering how consume_stok works, it's not
2765 * necessary. If consume_stock success, some charges
2766 * from this memcg are cached on this cpu. So, we
2767 * don't need to call css_get()/css_tryget() before
2768 * calling consume_stock().
2773 /* after here, we may be blocked. we need to get refcnt */
2774 if (!css_tryget(&memcg
->css
)) {
2782 bool invoke_oom
= oom
&& !nr_oom_retries
;
2784 /* If killed, bypass charge */
2785 if (fatal_signal_pending(current
)) {
2786 css_put(&memcg
->css
);
2790 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2791 nr_pages
, invoke_oom
);
2795 case CHARGE_RETRY
: /* not in OOM situation but retry */
2797 css_put(&memcg
->css
);
2800 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2801 css_put(&memcg
->css
);
2803 case CHARGE_NOMEM
: /* OOM routine works */
2804 if (!oom
|| invoke_oom
) {
2805 css_put(&memcg
->css
);
2811 } while (ret
!= CHARGE_OK
);
2813 if (batch
> nr_pages
)
2814 refill_stock(memcg
, batch
- nr_pages
);
2815 css_put(&memcg
->css
);
2823 *ptr
= root_mem_cgroup
;
2828 * Somemtimes we have to undo a charge we got by try_charge().
2829 * This function is for that and do uncharge, put css's refcnt.
2830 * gotten by try_charge().
2832 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2833 unsigned int nr_pages
)
2835 if (!mem_cgroup_is_root(memcg
)) {
2836 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2838 res_counter_uncharge(&memcg
->res
, bytes
);
2839 if (do_swap_account
)
2840 res_counter_uncharge(&memcg
->memsw
, bytes
);
2845 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2846 * This is useful when moving usage to parent cgroup.
2848 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2849 unsigned int nr_pages
)
2851 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2853 if (mem_cgroup_is_root(memcg
))
2856 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2857 if (do_swap_account
)
2858 res_counter_uncharge_until(&memcg
->memsw
,
2859 memcg
->memsw
.parent
, bytes
);
2863 * A helper function to get mem_cgroup from ID. must be called under
2864 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2865 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2866 * called against removed memcg.)
2868 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2870 struct cgroup_subsys_state
*css
;
2872 /* ID 0 is unused ID */
2875 css
= css_lookup(&mem_cgroup_subsys
, id
);
2878 return mem_cgroup_from_css(css
);
2881 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2883 struct mem_cgroup
*memcg
= NULL
;
2884 struct page_cgroup
*pc
;
2888 VM_BUG_ON(!PageLocked(page
));
2890 pc
= lookup_page_cgroup(page
);
2891 lock_page_cgroup(pc
);
2892 if (PageCgroupUsed(pc
)) {
2893 memcg
= pc
->mem_cgroup
;
2894 if (memcg
&& !css_tryget(&memcg
->css
))
2896 } else if (PageSwapCache(page
)) {
2897 ent
.val
= page_private(page
);
2898 id
= lookup_swap_cgroup_id(ent
);
2900 memcg
= mem_cgroup_lookup(id
);
2901 if (memcg
&& !css_tryget(&memcg
->css
))
2905 unlock_page_cgroup(pc
);
2909 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2911 unsigned int nr_pages
,
2912 enum charge_type ctype
,
2915 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2916 struct zone
*uninitialized_var(zone
);
2917 struct lruvec
*lruvec
;
2918 bool was_on_lru
= false;
2921 lock_page_cgroup(pc
);
2922 VM_BUG_ON(PageCgroupUsed(pc
));
2924 * we don't need page_cgroup_lock about tail pages, becase they are not
2925 * accessed by any other context at this point.
2929 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2930 * may already be on some other mem_cgroup's LRU. Take care of it.
2933 zone
= page_zone(page
);
2934 spin_lock_irq(&zone
->lru_lock
);
2935 if (PageLRU(page
)) {
2936 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2938 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2943 pc
->mem_cgroup
= memcg
;
2945 * We access a page_cgroup asynchronously without lock_page_cgroup().
2946 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2947 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2948 * before USED bit, we need memory barrier here.
2949 * See mem_cgroup_add_lru_list(), etc.
2952 SetPageCgroupUsed(pc
);
2956 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2957 VM_BUG_ON(PageLRU(page
));
2959 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2961 spin_unlock_irq(&zone
->lru_lock
);
2964 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2969 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2970 unlock_page_cgroup(pc
);
2973 * "charge_statistics" updated event counter. Then, check it.
2974 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2975 * if they exceeds softlimit.
2977 memcg_check_events(memcg
, page
);
2980 static DEFINE_MUTEX(set_limit_mutex
);
2982 #ifdef CONFIG_MEMCG_KMEM
2983 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2985 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2986 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2990 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2991 * in the memcg_cache_params struct.
2993 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2995 struct kmem_cache
*cachep
;
2997 VM_BUG_ON(p
->is_root_cache
);
2998 cachep
= p
->root_cache
;
2999 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
3002 #ifdef CONFIG_SLABINFO
3003 static int mem_cgroup_slabinfo_read(struct cgroup
*cont
, struct cftype
*cft
,
3006 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
3007 struct memcg_cache_params
*params
;
3009 if (!memcg_can_account_kmem(memcg
))
3012 print_slabinfo_header(m
);
3014 mutex_lock(&memcg
->slab_caches_mutex
);
3015 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
3016 cache_show(memcg_params_to_cache(params
), m
);
3017 mutex_unlock(&memcg
->slab_caches_mutex
);
3023 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
3025 struct res_counter
*fail_res
;
3026 struct mem_cgroup
*_memcg
;
3030 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
3035 * Conditions under which we can wait for the oom_killer. Those are
3036 * the same conditions tested by the core page allocator
3038 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
3041 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
3044 if (ret
== -EINTR
) {
3046 * __mem_cgroup_try_charge() chosed to bypass to root due to
3047 * OOM kill or fatal signal. Since our only options are to
3048 * either fail the allocation or charge it to this cgroup, do
3049 * it as a temporary condition. But we can't fail. From a
3050 * kmem/slab perspective, the cache has already been selected,
3051 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3054 * This condition will only trigger if the task entered
3055 * memcg_charge_kmem in a sane state, but was OOM-killed during
3056 * __mem_cgroup_try_charge() above. Tasks that were already
3057 * dying when the allocation triggers should have been already
3058 * directed to the root cgroup in memcontrol.h
3060 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3061 if (do_swap_account
)
3062 res_counter_charge_nofail(&memcg
->memsw
, size
,
3066 res_counter_uncharge(&memcg
->kmem
, size
);
3071 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3073 res_counter_uncharge(&memcg
->res
, size
);
3074 if (do_swap_account
)
3075 res_counter_uncharge(&memcg
->memsw
, size
);
3078 if (res_counter_uncharge(&memcg
->kmem
, size
))
3081 if (memcg_kmem_test_and_clear_dead(memcg
))
3082 mem_cgroup_put(memcg
);
3085 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3090 mutex_lock(&memcg
->slab_caches_mutex
);
3091 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3092 mutex_unlock(&memcg
->slab_caches_mutex
);
3096 * helper for acessing a memcg's index. It will be used as an index in the
3097 * child cache array in kmem_cache, and also to derive its name. This function
3098 * will return -1 when this is not a kmem-limited memcg.
3100 int memcg_cache_id(struct mem_cgroup
*memcg
)
3102 return memcg
? memcg
->kmemcg_id
: -1;
3106 * This ends up being protected by the set_limit mutex, during normal
3107 * operation, because that is its main call site.
3109 * But when we create a new cache, we can call this as well if its parent
3110 * is kmem-limited. That will have to hold set_limit_mutex as well.
3112 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3116 num
= ida_simple_get(&kmem_limited_groups
,
3117 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3121 * After this point, kmem_accounted (that we test atomically in
3122 * the beginning of this conditional), is no longer 0. This
3123 * guarantees only one process will set the following boolean
3124 * to true. We don't need test_and_set because we're protected
3125 * by the set_limit_mutex anyway.
3127 memcg_kmem_set_activated(memcg
);
3129 ret
= memcg_update_all_caches(num
+1);
3131 ida_simple_remove(&kmem_limited_groups
, num
);
3132 memcg_kmem_clear_activated(memcg
);
3136 memcg
->kmemcg_id
= num
;
3137 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3138 mutex_init(&memcg
->slab_caches_mutex
);
3142 static size_t memcg_caches_array_size(int num_groups
)
3145 if (num_groups
<= 0)
3148 size
= 2 * num_groups
;
3149 if (size
< MEMCG_CACHES_MIN_SIZE
)
3150 size
= MEMCG_CACHES_MIN_SIZE
;
3151 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3152 size
= MEMCG_CACHES_MAX_SIZE
;
3158 * We should update the current array size iff all caches updates succeed. This
3159 * can only be done from the slab side. The slab mutex needs to be held when
3162 void memcg_update_array_size(int num
)
3164 if (num
> memcg_limited_groups_array_size
)
3165 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3168 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3170 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3172 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3174 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3176 if (num_groups
> memcg_limited_groups_array_size
) {
3178 ssize_t size
= memcg_caches_array_size(num_groups
);
3180 size
*= sizeof(void *);
3181 size
+= sizeof(struct memcg_cache_params
);
3183 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3184 if (!s
->memcg_params
) {
3185 s
->memcg_params
= cur_params
;
3189 s
->memcg_params
->is_root_cache
= true;
3192 * There is the chance it will be bigger than
3193 * memcg_limited_groups_array_size, if we failed an allocation
3194 * in a cache, in which case all caches updated before it, will
3195 * have a bigger array.
3197 * But if that is the case, the data after
3198 * memcg_limited_groups_array_size is certainly unused
3200 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3201 if (!cur_params
->memcg_caches
[i
])
3203 s
->memcg_params
->memcg_caches
[i
] =
3204 cur_params
->memcg_caches
[i
];
3208 * Ideally, we would wait until all caches succeed, and only
3209 * then free the old one. But this is not worth the extra
3210 * pointer per-cache we'd have to have for this.
3212 * It is not a big deal if some caches are left with a size
3213 * bigger than the others. And all updates will reset this
3221 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3222 struct kmem_cache
*root_cache
)
3224 size_t size
= sizeof(struct memcg_cache_params
);
3226 if (!memcg_kmem_enabled())
3230 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3232 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3233 if (!s
->memcg_params
)
3237 s
->memcg_params
->memcg
= memcg
;
3238 s
->memcg_params
->root_cache
= root_cache
;
3239 INIT_WORK(&s
->memcg_params
->destroy
,
3240 kmem_cache_destroy_work_func
);
3242 s
->memcg_params
->is_root_cache
= true;
3247 void memcg_release_cache(struct kmem_cache
*s
)
3249 struct kmem_cache
*root
;
3250 struct mem_cgroup
*memcg
;
3254 * This happens, for instance, when a root cache goes away before we
3257 if (!s
->memcg_params
)
3260 if (s
->memcg_params
->is_root_cache
)
3263 memcg
= s
->memcg_params
->memcg
;
3264 id
= memcg_cache_id(memcg
);
3266 root
= s
->memcg_params
->root_cache
;
3267 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3269 mutex_lock(&memcg
->slab_caches_mutex
);
3270 list_del(&s
->memcg_params
->list
);
3271 mutex_unlock(&memcg
->slab_caches_mutex
);
3273 mem_cgroup_put(memcg
);
3275 kfree(s
->memcg_params
);
3279 * During the creation a new cache, we need to disable our accounting mechanism
3280 * altogether. This is true even if we are not creating, but rather just
3281 * enqueing new caches to be created.
3283 * This is because that process will trigger allocations; some visible, like
3284 * explicit kmallocs to auxiliary data structures, name strings and internal
3285 * cache structures; some well concealed, like INIT_WORK() that can allocate
3286 * objects during debug.
3288 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3289 * to it. This may not be a bounded recursion: since the first cache creation
3290 * failed to complete (waiting on the allocation), we'll just try to create the
3291 * cache again, failing at the same point.
3293 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3294 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3295 * inside the following two functions.
3297 static inline void memcg_stop_kmem_account(void)
3299 VM_BUG_ON(!current
->mm
);
3300 current
->memcg_kmem_skip_account
++;
3303 static inline void memcg_resume_kmem_account(void)
3305 VM_BUG_ON(!current
->mm
);
3306 current
->memcg_kmem_skip_account
--;
3309 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3311 struct kmem_cache
*cachep
;
3312 struct memcg_cache_params
*p
;
3314 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3316 cachep
= memcg_params_to_cache(p
);
3319 * If we get down to 0 after shrink, we could delete right away.
3320 * However, memcg_release_pages() already puts us back in the workqueue
3321 * in that case. If we proceed deleting, we'll get a dangling
3322 * reference, and removing the object from the workqueue in that case
3323 * is unnecessary complication. We are not a fast path.
3325 * Note that this case is fundamentally different from racing with
3326 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3327 * kmem_cache_shrink, not only we would be reinserting a dead cache
3328 * into the queue, but doing so from inside the worker racing to
3331 * So if we aren't down to zero, we'll just schedule a worker and try
3334 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3335 kmem_cache_shrink(cachep
);
3336 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3339 kmem_cache_destroy(cachep
);
3342 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3344 if (!cachep
->memcg_params
->dead
)
3348 * There are many ways in which we can get here.
3350 * We can get to a memory-pressure situation while the delayed work is
3351 * still pending to run. The vmscan shrinkers can then release all
3352 * cache memory and get us to destruction. If this is the case, we'll
3353 * be executed twice, which is a bug (the second time will execute over
3354 * bogus data). In this case, cancelling the work should be fine.
3356 * But we can also get here from the worker itself, if
3357 * kmem_cache_shrink is enough to shake all the remaining objects and
3358 * get the page count to 0. In this case, we'll deadlock if we try to
3359 * cancel the work (the worker runs with an internal lock held, which
3360 * is the same lock we would hold for cancel_work_sync().)
3362 * Since we can't possibly know who got us here, just refrain from
3363 * running if there is already work pending
3365 if (work_pending(&cachep
->memcg_params
->destroy
))
3368 * We have to defer the actual destroying to a workqueue, because
3369 * we might currently be in a context that cannot sleep.
3371 schedule_work(&cachep
->memcg_params
->destroy
);
3375 * This lock protects updaters, not readers. We want readers to be as fast as
3376 * they can, and they will either see NULL or a valid cache value. Our model
3377 * allow them to see NULL, in which case the root memcg will be selected.
3379 * We need this lock because multiple allocations to the same cache from a non
3380 * will span more than one worker. Only one of them can create the cache.
3382 static DEFINE_MUTEX(memcg_cache_mutex
);
3385 * Called with memcg_cache_mutex held
3387 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3388 struct kmem_cache
*s
)
3390 struct kmem_cache
*new;
3391 static char *tmp_name
= NULL
;
3393 lockdep_assert_held(&memcg_cache_mutex
);
3396 * kmem_cache_create_memcg duplicates the given name and
3397 * cgroup_name for this name requires RCU context.
3398 * This static temporary buffer is used to prevent from
3399 * pointless shortliving allocation.
3402 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3408 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3409 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3412 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3413 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3416 new->allocflags
|= __GFP_KMEMCG
;
3421 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3422 struct kmem_cache
*cachep
)
3424 struct kmem_cache
*new_cachep
;
3427 BUG_ON(!memcg_can_account_kmem(memcg
));
3429 idx
= memcg_cache_id(memcg
);
3431 mutex_lock(&memcg_cache_mutex
);
3432 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3436 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3437 if (new_cachep
== NULL
) {
3438 new_cachep
= cachep
;
3442 mem_cgroup_get(memcg
);
3443 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3445 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3447 * the readers won't lock, make sure everybody sees the updated value,
3448 * so they won't put stuff in the queue again for no reason
3452 mutex_unlock(&memcg_cache_mutex
);
3456 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3458 struct kmem_cache
*c
;
3461 if (!s
->memcg_params
)
3463 if (!s
->memcg_params
->is_root_cache
)
3467 * If the cache is being destroyed, we trust that there is no one else
3468 * requesting objects from it. Even if there are, the sanity checks in
3469 * kmem_cache_destroy should caught this ill-case.
3471 * Still, we don't want anyone else freeing memcg_caches under our
3472 * noses, which can happen if a new memcg comes to life. As usual,
3473 * we'll take the set_limit_mutex to protect ourselves against this.
3475 mutex_lock(&set_limit_mutex
);
3476 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3477 c
= s
->memcg_params
->memcg_caches
[i
];
3482 * We will now manually delete the caches, so to avoid races
3483 * we need to cancel all pending destruction workers and
3484 * proceed with destruction ourselves.
3486 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3487 * and that could spawn the workers again: it is likely that
3488 * the cache still have active pages until this very moment.
3489 * This would lead us back to mem_cgroup_destroy_cache.
3491 * But that will not execute at all if the "dead" flag is not
3492 * set, so flip it down to guarantee we are in control.
3494 c
->memcg_params
->dead
= false;
3495 cancel_work_sync(&c
->memcg_params
->destroy
);
3496 kmem_cache_destroy(c
);
3498 mutex_unlock(&set_limit_mutex
);
3501 struct create_work
{
3502 struct mem_cgroup
*memcg
;
3503 struct kmem_cache
*cachep
;
3504 struct work_struct work
;
3507 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3509 struct kmem_cache
*cachep
;
3510 struct memcg_cache_params
*params
;
3512 if (!memcg_kmem_is_active(memcg
))
3515 mutex_lock(&memcg
->slab_caches_mutex
);
3516 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3517 cachep
= memcg_params_to_cache(params
);
3518 cachep
->memcg_params
->dead
= true;
3519 schedule_work(&cachep
->memcg_params
->destroy
);
3521 mutex_unlock(&memcg
->slab_caches_mutex
);
3524 static void memcg_create_cache_work_func(struct work_struct
*w
)
3526 struct create_work
*cw
;
3528 cw
= container_of(w
, struct create_work
, work
);
3529 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3530 /* Drop the reference gotten when we enqueued. */
3531 css_put(&cw
->memcg
->css
);
3536 * Enqueue the creation of a per-memcg kmem_cache.
3538 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3539 struct kmem_cache
*cachep
)
3541 struct create_work
*cw
;
3543 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3545 css_put(&memcg
->css
);
3550 cw
->cachep
= cachep
;
3552 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3553 schedule_work(&cw
->work
);
3556 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3557 struct kmem_cache
*cachep
)
3560 * We need to stop accounting when we kmalloc, because if the
3561 * corresponding kmalloc cache is not yet created, the first allocation
3562 * in __memcg_create_cache_enqueue will recurse.
3564 * However, it is better to enclose the whole function. Depending on
3565 * the debugging options enabled, INIT_WORK(), for instance, can
3566 * trigger an allocation. This too, will make us recurse. Because at
3567 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3568 * the safest choice is to do it like this, wrapping the whole function.
3570 memcg_stop_kmem_account();
3571 __memcg_create_cache_enqueue(memcg
, cachep
);
3572 memcg_resume_kmem_account();
3575 * Return the kmem_cache we're supposed to use for a slab allocation.
3576 * We try to use the current memcg's version of the cache.
3578 * If the cache does not exist yet, if we are the first user of it,
3579 * we either create it immediately, if possible, or create it asynchronously
3581 * In the latter case, we will let the current allocation go through with
3582 * the original cache.
3584 * Can't be called in interrupt context or from kernel threads.
3585 * This function needs to be called with rcu_read_lock() held.
3587 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3590 struct mem_cgroup
*memcg
;
3593 VM_BUG_ON(!cachep
->memcg_params
);
3594 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3596 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3600 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3602 if (!memcg_can_account_kmem(memcg
))
3605 idx
= memcg_cache_id(memcg
);
3608 * barrier to mare sure we're always seeing the up to date value. The
3609 * code updating memcg_caches will issue a write barrier to match this.
3611 read_barrier_depends();
3612 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3613 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3617 /* The corresponding put will be done in the workqueue. */
3618 if (!css_tryget(&memcg
->css
))
3623 * If we are in a safe context (can wait, and not in interrupt
3624 * context), we could be be predictable and return right away.
3625 * This would guarantee that the allocation being performed
3626 * already belongs in the new cache.
3628 * However, there are some clashes that can arrive from locking.
3629 * For instance, because we acquire the slab_mutex while doing
3630 * kmem_cache_dup, this means no further allocation could happen
3631 * with the slab_mutex held.
3633 * Also, because cache creation issue get_online_cpus(), this
3634 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3635 * that ends up reversed during cpu hotplug. (cpuset allocates
3636 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3637 * better to defer everything.
3639 memcg_create_cache_enqueue(memcg
, cachep
);
3645 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3648 * We need to verify if the allocation against current->mm->owner's memcg is
3649 * possible for the given order. But the page is not allocated yet, so we'll
3650 * need a further commit step to do the final arrangements.
3652 * It is possible for the task to switch cgroups in this mean time, so at
3653 * commit time, we can't rely on task conversion any longer. We'll then use
3654 * the handle argument to return to the caller which cgroup we should commit
3655 * against. We could also return the memcg directly and avoid the pointer
3656 * passing, but a boolean return value gives better semantics considering
3657 * the compiled-out case as well.
3659 * Returning true means the allocation is possible.
3662 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3664 struct mem_cgroup
*memcg
;
3668 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3671 * very rare case described in mem_cgroup_from_task. Unfortunately there
3672 * isn't much we can do without complicating this too much, and it would
3673 * be gfp-dependent anyway. Just let it go
3675 if (unlikely(!memcg
))
3678 if (!memcg_can_account_kmem(memcg
)) {
3679 css_put(&memcg
->css
);
3683 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3687 css_put(&memcg
->css
);
3691 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3694 struct page_cgroup
*pc
;
3696 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3698 /* The page allocation failed. Revert */
3700 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3704 pc
= lookup_page_cgroup(page
);
3705 lock_page_cgroup(pc
);
3706 pc
->mem_cgroup
= memcg
;
3707 SetPageCgroupUsed(pc
);
3708 unlock_page_cgroup(pc
);
3711 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3713 struct mem_cgroup
*memcg
= NULL
;
3714 struct page_cgroup
*pc
;
3717 pc
= lookup_page_cgroup(page
);
3719 * Fast unlocked return. Theoretically might have changed, have to
3720 * check again after locking.
3722 if (!PageCgroupUsed(pc
))
3725 lock_page_cgroup(pc
);
3726 if (PageCgroupUsed(pc
)) {
3727 memcg
= pc
->mem_cgroup
;
3728 ClearPageCgroupUsed(pc
);
3730 unlock_page_cgroup(pc
);
3733 * We trust that only if there is a memcg associated with the page, it
3734 * is a valid allocation
3739 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3740 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3743 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3746 #endif /* CONFIG_MEMCG_KMEM */
3748 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3750 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3752 * Because tail pages are not marked as "used", set it. We're under
3753 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3754 * charge/uncharge will be never happen and move_account() is done under
3755 * compound_lock(), so we don't have to take care of races.
3757 void mem_cgroup_split_huge_fixup(struct page
*head
)
3759 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3760 struct page_cgroup
*pc
;
3761 struct mem_cgroup
*memcg
;
3764 if (mem_cgroup_disabled())
3767 memcg
= head_pc
->mem_cgroup
;
3768 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3770 pc
->mem_cgroup
= memcg
;
3771 smp_wmb();/* see __commit_charge() */
3772 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3774 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3777 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3780 * mem_cgroup_move_account - move account of the page
3782 * @nr_pages: number of regular pages (>1 for huge pages)
3783 * @pc: page_cgroup of the page.
3784 * @from: mem_cgroup which the page is moved from.
3785 * @to: mem_cgroup which the page is moved to. @from != @to.
3787 * The caller must confirm following.
3788 * - page is not on LRU (isolate_page() is useful.)
3789 * - compound_lock is held when nr_pages > 1
3791 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3794 static int mem_cgroup_move_account(struct page
*page
,
3795 unsigned int nr_pages
,
3796 struct page_cgroup
*pc
,
3797 struct mem_cgroup
*from
,
3798 struct mem_cgroup
*to
)
3800 unsigned long flags
;
3802 bool anon
= PageAnon(page
);
3804 VM_BUG_ON(from
== to
);
3805 VM_BUG_ON(PageLRU(page
));
3807 * The page is isolated from LRU. So, collapse function
3808 * will not handle this page. But page splitting can happen.
3809 * Do this check under compound_page_lock(). The caller should
3813 if (nr_pages
> 1 && !PageTransHuge(page
))
3816 lock_page_cgroup(pc
);
3819 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3822 move_lock_mem_cgroup(from
, &flags
);
3824 if (!anon
&& page_mapped(page
)) {
3825 /* Update mapped_file data for mem_cgroup */
3827 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3828 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3831 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3833 /* caller should have done css_get */
3834 pc
->mem_cgroup
= to
;
3835 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3836 move_unlock_mem_cgroup(from
, &flags
);
3839 unlock_page_cgroup(pc
);
3843 memcg_check_events(to
, page
);
3844 memcg_check_events(from
, page
);
3850 * mem_cgroup_move_parent - moves page to the parent group
3851 * @page: the page to move
3852 * @pc: page_cgroup of the page
3853 * @child: page's cgroup
3855 * move charges to its parent or the root cgroup if the group has no
3856 * parent (aka use_hierarchy==0).
3857 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3858 * mem_cgroup_move_account fails) the failure is always temporary and
3859 * it signals a race with a page removal/uncharge or migration. In the
3860 * first case the page is on the way out and it will vanish from the LRU
3861 * on the next attempt and the call should be retried later.
3862 * Isolation from the LRU fails only if page has been isolated from
3863 * the LRU since we looked at it and that usually means either global
3864 * reclaim or migration going on. The page will either get back to the
3866 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3867 * (!PageCgroupUsed) or moved to a different group. The page will
3868 * disappear in the next attempt.
3870 static int mem_cgroup_move_parent(struct page
*page
,
3871 struct page_cgroup
*pc
,
3872 struct mem_cgroup
*child
)
3874 struct mem_cgroup
*parent
;
3875 unsigned int nr_pages
;
3876 unsigned long uninitialized_var(flags
);
3879 VM_BUG_ON(mem_cgroup_is_root(child
));
3882 if (!get_page_unless_zero(page
))
3884 if (isolate_lru_page(page
))
3887 nr_pages
= hpage_nr_pages(page
);
3889 parent
= parent_mem_cgroup(child
);
3891 * If no parent, move charges to root cgroup.
3894 parent
= root_mem_cgroup
;
3897 VM_BUG_ON(!PageTransHuge(page
));
3898 flags
= compound_lock_irqsave(page
);
3901 ret
= mem_cgroup_move_account(page
, nr_pages
,
3904 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3907 compound_unlock_irqrestore(page
, flags
);
3908 putback_lru_page(page
);
3916 * Charge the memory controller for page usage.
3918 * 0 if the charge was successful
3919 * < 0 if the cgroup is over its limit
3921 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3922 gfp_t gfp_mask
, enum charge_type ctype
)
3924 struct mem_cgroup
*memcg
= NULL
;
3925 unsigned int nr_pages
= 1;
3929 if (PageTransHuge(page
)) {
3930 nr_pages
<<= compound_order(page
);
3931 VM_BUG_ON(!PageTransHuge(page
));
3933 * Never OOM-kill a process for a huge page. The
3934 * fault handler will fall back to regular pages.
3939 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3942 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3946 int mem_cgroup_newpage_charge(struct page
*page
,
3947 struct mm_struct
*mm
, gfp_t gfp_mask
)
3949 if (mem_cgroup_disabled())
3951 VM_BUG_ON(page_mapped(page
));
3952 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3954 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3955 MEM_CGROUP_CHARGE_TYPE_ANON
);
3959 * While swap-in, try_charge -> commit or cancel, the page is locked.
3960 * And when try_charge() successfully returns, one refcnt to memcg without
3961 * struct page_cgroup is acquired. This refcnt will be consumed by
3962 * "commit()" or removed by "cancel()"
3964 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3967 struct mem_cgroup
**memcgp
)
3969 struct mem_cgroup
*memcg
;
3970 struct page_cgroup
*pc
;
3973 pc
= lookup_page_cgroup(page
);
3975 * Every swap fault against a single page tries to charge the
3976 * page, bail as early as possible. shmem_unuse() encounters
3977 * already charged pages, too. The USED bit is protected by
3978 * the page lock, which serializes swap cache removal, which
3979 * in turn serializes uncharging.
3981 if (PageCgroupUsed(pc
))
3983 if (!do_swap_account
)
3985 memcg
= try_get_mem_cgroup_from_page(page
);
3989 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3990 css_put(&memcg
->css
);
3995 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
4001 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
4002 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
4005 if (mem_cgroup_disabled())
4008 * A racing thread's fault, or swapoff, may have already
4009 * updated the pte, and even removed page from swap cache: in
4010 * those cases unuse_pte()'s pte_same() test will fail; but
4011 * there's also a KSM case which does need to charge the page.
4013 if (!PageSwapCache(page
)) {
4016 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4021 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4024 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4026 if (mem_cgroup_disabled())
4030 __mem_cgroup_cancel_charge(memcg
, 1);
4034 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4035 enum charge_type ctype
)
4037 if (mem_cgroup_disabled())
4042 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4044 * Now swap is on-memory. This means this page may be
4045 * counted both as mem and swap....double count.
4046 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4047 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4048 * may call delete_from_swap_cache() before reach here.
4050 if (do_swap_account
&& PageSwapCache(page
)) {
4051 swp_entry_t ent
= {.val
= page_private(page
)};
4052 mem_cgroup_uncharge_swap(ent
);
4056 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4057 struct mem_cgroup
*memcg
)
4059 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4060 MEM_CGROUP_CHARGE_TYPE_ANON
);
4063 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4066 struct mem_cgroup
*memcg
= NULL
;
4067 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4070 if (mem_cgroup_disabled())
4072 if (PageCompound(page
))
4075 if (!PageSwapCache(page
))
4076 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4077 else { /* page is swapcache/shmem */
4078 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4081 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4086 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4087 unsigned int nr_pages
,
4088 const enum charge_type ctype
)
4090 struct memcg_batch_info
*batch
= NULL
;
4091 bool uncharge_memsw
= true;
4093 /* If swapout, usage of swap doesn't decrease */
4094 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4095 uncharge_memsw
= false;
4097 batch
= ¤t
->memcg_batch
;
4099 * In usual, we do css_get() when we remember memcg pointer.
4100 * But in this case, we keep res->usage until end of a series of
4101 * uncharges. Then, it's ok to ignore memcg's refcnt.
4104 batch
->memcg
= memcg
;
4106 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4107 * In those cases, all pages freed continuously can be expected to be in
4108 * the same cgroup and we have chance to coalesce uncharges.
4109 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4110 * because we want to do uncharge as soon as possible.
4113 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4114 goto direct_uncharge
;
4117 goto direct_uncharge
;
4120 * In typical case, batch->memcg == mem. This means we can
4121 * merge a series of uncharges to an uncharge of res_counter.
4122 * If not, we uncharge res_counter ony by one.
4124 if (batch
->memcg
!= memcg
)
4125 goto direct_uncharge
;
4126 /* remember freed charge and uncharge it later */
4129 batch
->memsw_nr_pages
++;
4132 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4134 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4135 if (unlikely(batch
->memcg
!= memcg
))
4136 memcg_oom_recover(memcg
);
4140 * uncharge if !page_mapped(page)
4142 static struct mem_cgroup
*
4143 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4146 struct mem_cgroup
*memcg
= NULL
;
4147 unsigned int nr_pages
= 1;
4148 struct page_cgroup
*pc
;
4151 if (mem_cgroup_disabled())
4154 if (PageTransHuge(page
)) {
4155 nr_pages
<<= compound_order(page
);
4156 VM_BUG_ON(!PageTransHuge(page
));
4159 * Check if our page_cgroup is valid
4161 pc
= lookup_page_cgroup(page
);
4162 if (unlikely(!PageCgroupUsed(pc
)))
4165 lock_page_cgroup(pc
);
4167 memcg
= pc
->mem_cgroup
;
4169 if (!PageCgroupUsed(pc
))
4172 anon
= PageAnon(page
);
4175 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4177 * Generally PageAnon tells if it's the anon statistics to be
4178 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4179 * used before page reached the stage of being marked PageAnon.
4183 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4184 /* See mem_cgroup_prepare_migration() */
4185 if (page_mapped(page
))
4188 * Pages under migration may not be uncharged. But
4189 * end_migration() /must/ be the one uncharging the
4190 * unused post-migration page and so it has to call
4191 * here with the migration bit still set. See the
4192 * res_counter handling below.
4194 if (!end_migration
&& PageCgroupMigration(pc
))
4197 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4198 if (!PageAnon(page
)) { /* Shared memory */
4199 if (page
->mapping
&& !page_is_file_cache(page
))
4201 } else if (page_mapped(page
)) /* Anon */
4208 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4210 ClearPageCgroupUsed(pc
);
4212 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4213 * freed from LRU. This is safe because uncharged page is expected not
4214 * to be reused (freed soon). Exception is SwapCache, it's handled by
4215 * special functions.
4218 unlock_page_cgroup(pc
);
4220 * even after unlock, we have memcg->res.usage here and this memcg
4221 * will never be freed.
4223 memcg_check_events(memcg
, page
);
4224 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4225 mem_cgroup_swap_statistics(memcg
, true);
4226 mem_cgroup_get(memcg
);
4229 * Migration does not charge the res_counter for the
4230 * replacement page, so leave it alone when phasing out the
4231 * page that is unused after the migration.
4233 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4234 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4239 unlock_page_cgroup(pc
);
4243 void mem_cgroup_uncharge_page(struct page
*page
)
4246 if (page_mapped(page
))
4248 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4250 * If the page is in swap cache, uncharge should be deferred
4251 * to the swap path, which also properly accounts swap usage
4252 * and handles memcg lifetime.
4254 * Note that this check is not stable and reclaim may add the
4255 * page to swap cache at any time after this. However, if the
4256 * page is not in swap cache by the time page->mapcount hits
4257 * 0, there won't be any page table references to the swap
4258 * slot, and reclaim will free it and not actually write the
4261 if (PageSwapCache(page
))
4263 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4266 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4268 VM_BUG_ON(page_mapped(page
));
4269 VM_BUG_ON(page
->mapping
);
4270 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4274 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4275 * In that cases, pages are freed continuously and we can expect pages
4276 * are in the same memcg. All these calls itself limits the number of
4277 * pages freed at once, then uncharge_start/end() is called properly.
4278 * This may be called prural(2) times in a context,
4281 void mem_cgroup_uncharge_start(void)
4283 current
->memcg_batch
.do_batch
++;
4284 /* We can do nest. */
4285 if (current
->memcg_batch
.do_batch
== 1) {
4286 current
->memcg_batch
.memcg
= NULL
;
4287 current
->memcg_batch
.nr_pages
= 0;
4288 current
->memcg_batch
.memsw_nr_pages
= 0;
4292 void mem_cgroup_uncharge_end(void)
4294 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4296 if (!batch
->do_batch
)
4300 if (batch
->do_batch
) /* If stacked, do nothing. */
4306 * This "batch->memcg" is valid without any css_get/put etc...
4307 * bacause we hide charges behind us.
4309 if (batch
->nr_pages
)
4310 res_counter_uncharge(&batch
->memcg
->res
,
4311 batch
->nr_pages
* PAGE_SIZE
);
4312 if (batch
->memsw_nr_pages
)
4313 res_counter_uncharge(&batch
->memcg
->memsw
,
4314 batch
->memsw_nr_pages
* PAGE_SIZE
);
4315 memcg_oom_recover(batch
->memcg
);
4316 /* forget this pointer (for sanity check) */
4317 batch
->memcg
= NULL
;
4322 * called after __delete_from_swap_cache() and drop "page" account.
4323 * memcg information is recorded to swap_cgroup of "ent"
4326 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4328 struct mem_cgroup
*memcg
;
4329 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4331 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4332 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4334 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4337 * record memcg information, if swapout && memcg != NULL,
4338 * mem_cgroup_get() was called in uncharge().
4340 if (do_swap_account
&& swapout
&& memcg
)
4341 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4345 #ifdef CONFIG_MEMCG_SWAP
4347 * called from swap_entry_free(). remove record in swap_cgroup and
4348 * uncharge "memsw" account.
4350 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4352 struct mem_cgroup
*memcg
;
4355 if (!do_swap_account
)
4358 id
= swap_cgroup_record(ent
, 0);
4360 memcg
= mem_cgroup_lookup(id
);
4363 * We uncharge this because swap is freed.
4364 * This memcg can be obsolete one. We avoid calling css_tryget
4366 if (!mem_cgroup_is_root(memcg
))
4367 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4368 mem_cgroup_swap_statistics(memcg
, false);
4369 mem_cgroup_put(memcg
);
4375 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4376 * @entry: swap entry to be moved
4377 * @from: mem_cgroup which the entry is moved from
4378 * @to: mem_cgroup which the entry is moved to
4380 * It succeeds only when the swap_cgroup's record for this entry is the same
4381 * as the mem_cgroup's id of @from.
4383 * Returns 0 on success, -EINVAL on failure.
4385 * The caller must have charged to @to, IOW, called res_counter_charge() about
4386 * both res and memsw, and called css_get().
4388 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4389 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4391 unsigned short old_id
, new_id
;
4393 old_id
= css_id(&from
->css
);
4394 new_id
= css_id(&to
->css
);
4396 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4397 mem_cgroup_swap_statistics(from
, false);
4398 mem_cgroup_swap_statistics(to
, true);
4400 * This function is only called from task migration context now.
4401 * It postpones res_counter and refcount handling till the end
4402 * of task migration(mem_cgroup_clear_mc()) for performance
4403 * improvement. But we cannot postpone mem_cgroup_get(to)
4404 * because if the process that has been moved to @to does
4405 * swap-in, the refcount of @to might be decreased to 0.
4413 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4414 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4421 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4424 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4425 struct mem_cgroup
**memcgp
)
4427 struct mem_cgroup
*memcg
= NULL
;
4428 unsigned int nr_pages
= 1;
4429 struct page_cgroup
*pc
;
4430 enum charge_type ctype
;
4434 if (mem_cgroup_disabled())
4437 if (PageTransHuge(page
))
4438 nr_pages
<<= compound_order(page
);
4440 pc
= lookup_page_cgroup(page
);
4441 lock_page_cgroup(pc
);
4442 if (PageCgroupUsed(pc
)) {
4443 memcg
= pc
->mem_cgroup
;
4444 css_get(&memcg
->css
);
4446 * At migrating an anonymous page, its mapcount goes down
4447 * to 0 and uncharge() will be called. But, even if it's fully
4448 * unmapped, migration may fail and this page has to be
4449 * charged again. We set MIGRATION flag here and delay uncharge
4450 * until end_migration() is called
4452 * Corner Case Thinking
4454 * When the old page was mapped as Anon and it's unmap-and-freed
4455 * while migration was ongoing.
4456 * If unmap finds the old page, uncharge() of it will be delayed
4457 * until end_migration(). If unmap finds a new page, it's
4458 * uncharged when it make mapcount to be 1->0. If unmap code
4459 * finds swap_migration_entry, the new page will not be mapped
4460 * and end_migration() will find it(mapcount==0).
4463 * When the old page was mapped but migraion fails, the kernel
4464 * remaps it. A charge for it is kept by MIGRATION flag even
4465 * if mapcount goes down to 0. We can do remap successfully
4466 * without charging it again.
4469 * The "old" page is under lock_page() until the end of
4470 * migration, so, the old page itself will not be swapped-out.
4471 * If the new page is swapped out before end_migraton, our
4472 * hook to usual swap-out path will catch the event.
4475 SetPageCgroupMigration(pc
);
4477 unlock_page_cgroup(pc
);
4479 * If the page is not charged at this point,
4487 * We charge new page before it's used/mapped. So, even if unlock_page()
4488 * is called before end_migration, we can catch all events on this new
4489 * page. In the case new page is migrated but not remapped, new page's
4490 * mapcount will be finally 0 and we call uncharge in end_migration().
4493 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4495 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4497 * The page is committed to the memcg, but it's not actually
4498 * charged to the res_counter since we plan on replacing the
4499 * old one and only one page is going to be left afterwards.
4501 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4504 /* remove redundant charge if migration failed*/
4505 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4506 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4508 struct page
*used
, *unused
;
4509 struct page_cgroup
*pc
;
4515 if (!migration_ok
) {
4522 anon
= PageAnon(used
);
4523 __mem_cgroup_uncharge_common(unused
,
4524 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4525 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4527 css_put(&memcg
->css
);
4529 * We disallowed uncharge of pages under migration because mapcount
4530 * of the page goes down to zero, temporarly.
4531 * Clear the flag and check the page should be charged.
4533 pc
= lookup_page_cgroup(oldpage
);
4534 lock_page_cgroup(pc
);
4535 ClearPageCgroupMigration(pc
);
4536 unlock_page_cgroup(pc
);
4539 * If a page is a file cache, radix-tree replacement is very atomic
4540 * and we can skip this check. When it was an Anon page, its mapcount
4541 * goes down to 0. But because we added MIGRATION flage, it's not
4542 * uncharged yet. There are several case but page->mapcount check
4543 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4544 * check. (see prepare_charge() also)
4547 mem_cgroup_uncharge_page(used
);
4551 * At replace page cache, newpage is not under any memcg but it's on
4552 * LRU. So, this function doesn't touch res_counter but handles LRU
4553 * in correct way. Both pages are locked so we cannot race with uncharge.
4555 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4556 struct page
*newpage
)
4558 struct mem_cgroup
*memcg
= NULL
;
4559 struct page_cgroup
*pc
;
4560 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4562 if (mem_cgroup_disabled())
4565 pc
= lookup_page_cgroup(oldpage
);
4566 /* fix accounting on old pages */
4567 lock_page_cgroup(pc
);
4568 if (PageCgroupUsed(pc
)) {
4569 memcg
= pc
->mem_cgroup
;
4570 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4571 ClearPageCgroupUsed(pc
);
4573 unlock_page_cgroup(pc
);
4576 * When called from shmem_replace_page(), in some cases the
4577 * oldpage has already been charged, and in some cases not.
4582 * Even if newpage->mapping was NULL before starting replacement,
4583 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4584 * LRU while we overwrite pc->mem_cgroup.
4586 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4589 #ifdef CONFIG_DEBUG_VM
4590 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4592 struct page_cgroup
*pc
;
4594 pc
= lookup_page_cgroup(page
);
4596 * Can be NULL while feeding pages into the page allocator for
4597 * the first time, i.e. during boot or memory hotplug;
4598 * or when mem_cgroup_disabled().
4600 if (likely(pc
) && PageCgroupUsed(pc
))
4605 bool mem_cgroup_bad_page_check(struct page
*page
)
4607 if (mem_cgroup_disabled())
4610 return lookup_page_cgroup_used(page
) != NULL
;
4613 void mem_cgroup_print_bad_page(struct page
*page
)
4615 struct page_cgroup
*pc
;
4617 pc
= lookup_page_cgroup_used(page
);
4619 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4620 pc
, pc
->flags
, pc
->mem_cgroup
);
4625 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4626 unsigned long long val
)
4629 u64 memswlimit
, memlimit
;
4631 int children
= mem_cgroup_count_children(memcg
);
4632 u64 curusage
, oldusage
;
4636 * For keeping hierarchical_reclaim simple, how long we should retry
4637 * is depends on callers. We set our retry-count to be function
4638 * of # of children which we should visit in this loop.
4640 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4642 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4645 while (retry_count
) {
4646 if (signal_pending(current
)) {
4651 * Rather than hide all in some function, I do this in
4652 * open coded manner. You see what this really does.
4653 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4655 mutex_lock(&set_limit_mutex
);
4656 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4657 if (memswlimit
< val
) {
4659 mutex_unlock(&set_limit_mutex
);
4663 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4667 ret
= res_counter_set_limit(&memcg
->res
, val
);
4669 if (memswlimit
== val
)
4670 memcg
->memsw_is_minimum
= true;
4672 memcg
->memsw_is_minimum
= false;
4674 mutex_unlock(&set_limit_mutex
);
4679 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4680 MEM_CGROUP_RECLAIM_SHRINK
);
4681 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4682 /* Usage is reduced ? */
4683 if (curusage
>= oldusage
)
4686 oldusage
= curusage
;
4688 if (!ret
&& enlarge
)
4689 memcg_oom_recover(memcg
);
4694 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4695 unsigned long long val
)
4698 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4699 int children
= mem_cgroup_count_children(memcg
);
4703 /* see mem_cgroup_resize_res_limit */
4704 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4705 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4706 while (retry_count
) {
4707 if (signal_pending(current
)) {
4712 * Rather than hide all in some function, I do this in
4713 * open coded manner. You see what this really does.
4714 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4716 mutex_lock(&set_limit_mutex
);
4717 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4718 if (memlimit
> val
) {
4720 mutex_unlock(&set_limit_mutex
);
4723 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4724 if (memswlimit
< val
)
4726 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4728 if (memlimit
== val
)
4729 memcg
->memsw_is_minimum
= true;
4731 memcg
->memsw_is_minimum
= false;
4733 mutex_unlock(&set_limit_mutex
);
4738 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4739 MEM_CGROUP_RECLAIM_NOSWAP
|
4740 MEM_CGROUP_RECLAIM_SHRINK
);
4741 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4742 /* Usage is reduced ? */
4743 if (curusage
>= oldusage
)
4746 oldusage
= curusage
;
4748 if (!ret
&& enlarge
)
4749 memcg_oom_recover(memcg
);
4753 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4755 unsigned long *total_scanned
)
4757 unsigned long nr_reclaimed
= 0;
4758 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4759 unsigned long reclaimed
;
4761 struct mem_cgroup_tree_per_zone
*mctz
;
4762 unsigned long long excess
;
4763 unsigned long nr_scanned
;
4768 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4770 * This loop can run a while, specially if mem_cgroup's continuously
4771 * keep exceeding their soft limit and putting the system under
4778 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4783 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4784 gfp_mask
, &nr_scanned
);
4785 nr_reclaimed
+= reclaimed
;
4786 *total_scanned
+= nr_scanned
;
4787 spin_lock(&mctz
->lock
);
4790 * If we failed to reclaim anything from this memory cgroup
4791 * it is time to move on to the next cgroup
4797 * Loop until we find yet another one.
4799 * By the time we get the soft_limit lock
4800 * again, someone might have aded the
4801 * group back on the RB tree. Iterate to
4802 * make sure we get a different mem.
4803 * mem_cgroup_largest_soft_limit_node returns
4804 * NULL if no other cgroup is present on
4808 __mem_cgroup_largest_soft_limit_node(mctz
);
4810 css_put(&next_mz
->memcg
->css
);
4811 else /* next_mz == NULL or other memcg */
4815 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4816 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4818 * One school of thought says that we should not add
4819 * back the node to the tree if reclaim returns 0.
4820 * But our reclaim could return 0, simply because due
4821 * to priority we are exposing a smaller subset of
4822 * memory to reclaim from. Consider this as a longer
4825 /* If excess == 0, no tree ops */
4826 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4827 spin_unlock(&mctz
->lock
);
4828 css_put(&mz
->memcg
->css
);
4831 * Could not reclaim anything and there are no more
4832 * mem cgroups to try or we seem to be looping without
4833 * reclaiming anything.
4835 if (!nr_reclaimed
&&
4837 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4839 } while (!nr_reclaimed
);
4841 css_put(&next_mz
->memcg
->css
);
4842 return nr_reclaimed
;
4846 * mem_cgroup_force_empty_list - clears LRU of a group
4847 * @memcg: group to clear
4850 * @lru: lru to to clear
4852 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4853 * reclaim the pages page themselves - pages are moved to the parent (or root)
4856 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4857 int node
, int zid
, enum lru_list lru
)
4859 struct lruvec
*lruvec
;
4860 unsigned long flags
;
4861 struct list_head
*list
;
4865 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4866 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4867 list
= &lruvec
->lists
[lru
];
4871 struct page_cgroup
*pc
;
4874 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4875 if (list_empty(list
)) {
4876 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4879 page
= list_entry(list
->prev
, struct page
, lru
);
4881 list_move(&page
->lru
, list
);
4883 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4886 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4888 pc
= lookup_page_cgroup(page
);
4890 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4891 /* found lock contention or "pc" is obsolete. */
4896 } while (!list_empty(list
));
4900 * make mem_cgroup's charge to be 0 if there is no task by moving
4901 * all the charges and pages to the parent.
4902 * This enables deleting this mem_cgroup.
4904 * Caller is responsible for holding css reference on the memcg.
4906 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4912 /* This is for making all *used* pages to be on LRU. */
4913 lru_add_drain_all();
4914 drain_all_stock_sync(memcg
);
4915 mem_cgroup_start_move(memcg
);
4916 for_each_node_state(node
, N_MEMORY
) {
4917 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4920 mem_cgroup_force_empty_list(memcg
,
4925 mem_cgroup_end_move(memcg
);
4926 memcg_oom_recover(memcg
);
4930 * Kernel memory may not necessarily be trackable to a specific
4931 * process. So they are not migrated, and therefore we can't
4932 * expect their value to drop to 0 here.
4933 * Having res filled up with kmem only is enough.
4935 * This is a safety check because mem_cgroup_force_empty_list
4936 * could have raced with mem_cgroup_replace_page_cache callers
4937 * so the lru seemed empty but the page could have been added
4938 * right after the check. RES_USAGE should be safe as we always
4939 * charge before adding to the LRU.
4941 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4942 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4943 } while (usage
> 0);
4947 * This mainly exists for tests during the setting of set of use_hierarchy.
4948 * Since this is the very setting we are changing, the current hierarchy value
4951 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4955 /* bounce at first found */
4956 cgroup_for_each_child(pos
, memcg
->css
.cgroup
)
4962 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4963 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4964 * from mem_cgroup_count_children(), in the sense that we don't really care how
4965 * many children we have; we only need to know if we have any. It also counts
4966 * any memcg without hierarchy as infertile.
4968 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4970 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4974 * Reclaims as many pages from the given memcg as possible and moves
4975 * the rest to the parent.
4977 * Caller is responsible for holding css reference for memcg.
4979 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4981 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4982 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4984 /* returns EBUSY if there is a task or if we come here twice. */
4985 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4988 /* we call try-to-free pages for make this cgroup empty */
4989 lru_add_drain_all();
4990 /* try to free all pages in this cgroup */
4991 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4994 if (signal_pending(current
))
4997 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
5001 /* maybe some writeback is necessary */
5002 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
5007 mem_cgroup_reparent_charges(memcg
);
5012 static int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
5014 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5017 if (mem_cgroup_is_root(memcg
))
5019 css_get(&memcg
->css
);
5020 ret
= mem_cgroup_force_empty(memcg
);
5021 css_put(&memcg
->css
);
5027 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
5029 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
5032 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
5036 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5037 struct cgroup
*parent
= cont
->parent
;
5038 struct mem_cgroup
*parent_memcg
= NULL
;
5041 parent_memcg
= mem_cgroup_from_cont(parent
);
5043 mutex_lock(&memcg_create_mutex
);
5045 if (memcg
->use_hierarchy
== val
)
5049 * If parent's use_hierarchy is set, we can't make any modifications
5050 * in the child subtrees. If it is unset, then the change can
5051 * occur, provided the current cgroup has no children.
5053 * For the root cgroup, parent_mem is NULL, we allow value to be
5054 * set if there are no children.
5056 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5057 (val
== 1 || val
== 0)) {
5058 if (!__memcg_has_children(memcg
))
5059 memcg
->use_hierarchy
= val
;
5066 mutex_unlock(&memcg_create_mutex
);
5072 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5073 enum mem_cgroup_stat_index idx
)
5075 struct mem_cgroup
*iter
;
5078 /* Per-cpu values can be negative, use a signed accumulator */
5079 for_each_mem_cgroup_tree(iter
, memcg
)
5080 val
+= mem_cgroup_read_stat(iter
, idx
);
5082 if (val
< 0) /* race ? */
5087 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5091 if (!mem_cgroup_is_root(memcg
)) {
5093 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5095 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5099 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5100 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5102 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5103 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5106 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5108 return val
<< PAGE_SHIFT
;
5111 static ssize_t
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
,
5112 struct file
*file
, char __user
*buf
,
5113 size_t nbytes
, loff_t
*ppos
)
5115 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5121 type
= MEMFILE_TYPE(cft
->private);
5122 name
= MEMFILE_ATTR(cft
->private);
5126 if (name
== RES_USAGE
)
5127 val
= mem_cgroup_usage(memcg
, false);
5129 val
= res_counter_read_u64(&memcg
->res
, name
);
5132 if (name
== RES_USAGE
)
5133 val
= mem_cgroup_usage(memcg
, true);
5135 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5138 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5144 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5145 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5148 static int memcg_update_kmem_limit(struct cgroup
*cont
, u64 val
)
5151 #ifdef CONFIG_MEMCG_KMEM
5152 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5154 * For simplicity, we won't allow this to be disabled. It also can't
5155 * be changed if the cgroup has children already, or if tasks had
5158 * If tasks join before we set the limit, a person looking at
5159 * kmem.usage_in_bytes will have no way to determine when it took
5160 * place, which makes the value quite meaningless.
5162 * After it first became limited, changes in the value of the limit are
5163 * of course permitted.
5165 mutex_lock(&memcg_create_mutex
);
5166 mutex_lock(&set_limit_mutex
);
5167 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
5168 if (cgroup_task_count(cont
) || memcg_has_children(memcg
)) {
5172 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5175 ret
= memcg_update_cache_sizes(memcg
);
5177 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
5180 static_key_slow_inc(&memcg_kmem_enabled_key
);
5182 * setting the active bit after the inc will guarantee no one
5183 * starts accounting before all call sites are patched
5185 memcg_kmem_set_active(memcg
);
5188 * kmem charges can outlive the cgroup. In the case of slab
5189 * pages, for instance, a page contain objects from various
5190 * processes, so it is unfeasible to migrate them away. We
5191 * need to reference count the memcg because of that.
5193 mem_cgroup_get(memcg
);
5195 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5197 mutex_unlock(&set_limit_mutex
);
5198 mutex_unlock(&memcg_create_mutex
);
5203 #ifdef CONFIG_MEMCG_KMEM
5204 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5207 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5211 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5213 * When that happen, we need to disable the static branch only on those
5214 * memcgs that enabled it. To achieve this, we would be forced to
5215 * complicate the code by keeping track of which memcgs were the ones
5216 * that actually enabled limits, and which ones got it from its
5219 * It is a lot simpler just to do static_key_slow_inc() on every child
5220 * that is accounted.
5222 if (!memcg_kmem_is_active(memcg
))
5226 * destroy(), called if we fail, will issue static_key_slow_inc() and
5227 * mem_cgroup_put() if kmem is enabled. We have to either call them
5228 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5229 * this more consistent, since it always leads to the same destroy path
5231 mem_cgroup_get(memcg
);
5232 static_key_slow_inc(&memcg_kmem_enabled_key
);
5234 mutex_lock(&set_limit_mutex
);
5235 ret
= memcg_update_cache_sizes(memcg
);
5236 mutex_unlock(&set_limit_mutex
);
5240 #endif /* CONFIG_MEMCG_KMEM */
5243 * The user of this function is...
5246 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
5249 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5252 unsigned long long val
;
5255 type
= MEMFILE_TYPE(cft
->private);
5256 name
= MEMFILE_ATTR(cft
->private);
5260 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5264 /* This function does all necessary parse...reuse it */
5265 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5269 ret
= mem_cgroup_resize_limit(memcg
, val
);
5270 else if (type
== _MEMSWAP
)
5271 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5272 else if (type
== _KMEM
)
5273 ret
= memcg_update_kmem_limit(cont
, val
);
5277 case RES_SOFT_LIMIT
:
5278 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5282 * For memsw, soft limits are hard to implement in terms
5283 * of semantics, for now, we support soft limits for
5284 * control without swap
5287 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5292 ret
= -EINVAL
; /* should be BUG() ? */
5298 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5299 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5301 struct cgroup
*cgroup
;
5302 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5304 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5305 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5306 cgroup
= memcg
->css
.cgroup
;
5307 if (!memcg
->use_hierarchy
)
5310 while (cgroup
->parent
) {
5311 cgroup
= cgroup
->parent
;
5312 memcg
= mem_cgroup_from_cont(cgroup
);
5313 if (!memcg
->use_hierarchy
)
5315 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5316 min_limit
= min(min_limit
, tmp
);
5317 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5318 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5321 *mem_limit
= min_limit
;
5322 *memsw_limit
= min_memsw_limit
;
5325 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
5327 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5331 type
= MEMFILE_TYPE(event
);
5332 name
= MEMFILE_ATTR(event
);
5337 res_counter_reset_max(&memcg
->res
);
5338 else if (type
== _MEMSWAP
)
5339 res_counter_reset_max(&memcg
->memsw
);
5340 else if (type
== _KMEM
)
5341 res_counter_reset_max(&memcg
->kmem
);
5347 res_counter_reset_failcnt(&memcg
->res
);
5348 else if (type
== _MEMSWAP
)
5349 res_counter_reset_failcnt(&memcg
->memsw
);
5350 else if (type
== _KMEM
)
5351 res_counter_reset_failcnt(&memcg
->kmem
);
5360 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
5363 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
5367 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5368 struct cftype
*cft
, u64 val
)
5370 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5372 if (val
>= (1 << NR_MOVE_TYPE
))
5376 * No kind of locking is needed in here, because ->can_attach() will
5377 * check this value once in the beginning of the process, and then carry
5378 * on with stale data. This means that changes to this value will only
5379 * affect task migrations starting after the change.
5381 memcg
->move_charge_at_immigrate
= val
;
5385 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5386 struct cftype
*cft
, u64 val
)
5393 static int memcg_numa_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5397 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5398 unsigned long node_nr
;
5399 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5401 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5402 seq_printf(m
, "total=%lu", total_nr
);
5403 for_each_node_state(nid
, N_MEMORY
) {
5404 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5405 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5409 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5410 seq_printf(m
, "file=%lu", file_nr
);
5411 for_each_node_state(nid
, N_MEMORY
) {
5412 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5414 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5418 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5419 seq_printf(m
, "anon=%lu", anon_nr
);
5420 for_each_node_state(nid
, N_MEMORY
) {
5421 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5423 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5427 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5428 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5429 for_each_node_state(nid
, N_MEMORY
) {
5430 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5431 BIT(LRU_UNEVICTABLE
));
5432 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5437 #endif /* CONFIG_NUMA */
5439 static inline void mem_cgroup_lru_names_not_uptodate(void)
5441 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5444 static int memcg_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5447 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5448 struct mem_cgroup
*mi
;
5451 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5452 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5454 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5455 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5458 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5459 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5460 mem_cgroup_read_events(memcg
, i
));
5462 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5463 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5464 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5466 /* Hierarchical information */
5468 unsigned long long limit
, memsw_limit
;
5469 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5470 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5471 if (do_swap_account
)
5472 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5476 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5479 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5481 for_each_mem_cgroup_tree(mi
, memcg
)
5482 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5483 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5486 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5487 unsigned long long val
= 0;
5489 for_each_mem_cgroup_tree(mi
, memcg
)
5490 val
+= mem_cgroup_read_events(mi
, i
);
5491 seq_printf(m
, "total_%s %llu\n",
5492 mem_cgroup_events_names
[i
], val
);
5495 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5496 unsigned long long val
= 0;
5498 for_each_mem_cgroup_tree(mi
, memcg
)
5499 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5500 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5503 #ifdef CONFIG_DEBUG_VM
5506 struct mem_cgroup_per_zone
*mz
;
5507 struct zone_reclaim_stat
*rstat
;
5508 unsigned long recent_rotated
[2] = {0, 0};
5509 unsigned long recent_scanned
[2] = {0, 0};
5511 for_each_online_node(nid
)
5512 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5513 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5514 rstat
= &mz
->lruvec
.reclaim_stat
;
5516 recent_rotated
[0] += rstat
->recent_rotated
[0];
5517 recent_rotated
[1] += rstat
->recent_rotated
[1];
5518 recent_scanned
[0] += rstat
->recent_scanned
[0];
5519 recent_scanned
[1] += rstat
->recent_scanned
[1];
5521 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5522 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5523 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5524 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5531 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
5533 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5535 return mem_cgroup_swappiness(memcg
);
5538 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
5541 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5542 struct mem_cgroup
*parent
;
5547 if (cgrp
->parent
== NULL
)
5550 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5552 mutex_lock(&memcg_create_mutex
);
5554 /* If under hierarchy, only empty-root can set this value */
5555 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5556 mutex_unlock(&memcg_create_mutex
);
5560 memcg
->swappiness
= val
;
5562 mutex_unlock(&memcg_create_mutex
);
5567 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5569 struct mem_cgroup_threshold_ary
*t
;
5575 t
= rcu_dereference(memcg
->thresholds
.primary
);
5577 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5582 usage
= mem_cgroup_usage(memcg
, swap
);
5585 * current_threshold points to threshold just below or equal to usage.
5586 * If it's not true, a threshold was crossed after last
5587 * call of __mem_cgroup_threshold().
5589 i
= t
->current_threshold
;
5592 * Iterate backward over array of thresholds starting from
5593 * current_threshold and check if a threshold is crossed.
5594 * If none of thresholds below usage is crossed, we read
5595 * only one element of the array here.
5597 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5598 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5600 /* i = current_threshold + 1 */
5604 * Iterate forward over array of thresholds starting from
5605 * current_threshold+1 and check if a threshold is crossed.
5606 * If none of thresholds above usage is crossed, we read
5607 * only one element of the array here.
5609 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5610 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5612 /* Update current_threshold */
5613 t
->current_threshold
= i
- 1;
5618 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5621 __mem_cgroup_threshold(memcg
, false);
5622 if (do_swap_account
)
5623 __mem_cgroup_threshold(memcg
, true);
5625 memcg
= parent_mem_cgroup(memcg
);
5629 static int compare_thresholds(const void *a
, const void *b
)
5631 const struct mem_cgroup_threshold
*_a
= a
;
5632 const struct mem_cgroup_threshold
*_b
= b
;
5634 if (_a
->threshold
> _b
->threshold
)
5637 if (_a
->threshold
< _b
->threshold
)
5643 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5645 struct mem_cgroup_eventfd_list
*ev
;
5647 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5648 eventfd_signal(ev
->eventfd
, 1);
5652 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5654 struct mem_cgroup
*iter
;
5656 for_each_mem_cgroup_tree(iter
, memcg
)
5657 mem_cgroup_oom_notify_cb(iter
);
5660 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5661 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5663 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5664 struct mem_cgroup_thresholds
*thresholds
;
5665 struct mem_cgroup_threshold_ary
*new;
5666 enum res_type type
= MEMFILE_TYPE(cft
->private);
5667 u64 threshold
, usage
;
5670 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5674 mutex_lock(&memcg
->thresholds_lock
);
5677 thresholds
= &memcg
->thresholds
;
5678 else if (type
== _MEMSWAP
)
5679 thresholds
= &memcg
->memsw_thresholds
;
5683 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5685 /* Check if a threshold crossed before adding a new one */
5686 if (thresholds
->primary
)
5687 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5689 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5691 /* Allocate memory for new array of thresholds */
5692 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5700 /* Copy thresholds (if any) to new array */
5701 if (thresholds
->primary
) {
5702 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5703 sizeof(struct mem_cgroup_threshold
));
5706 /* Add new threshold */
5707 new->entries
[size
- 1].eventfd
= eventfd
;
5708 new->entries
[size
- 1].threshold
= threshold
;
5710 /* Sort thresholds. Registering of new threshold isn't time-critical */
5711 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5712 compare_thresholds
, NULL
);
5714 /* Find current threshold */
5715 new->current_threshold
= -1;
5716 for (i
= 0; i
< size
; i
++) {
5717 if (new->entries
[i
].threshold
<= usage
) {
5719 * new->current_threshold will not be used until
5720 * rcu_assign_pointer(), so it's safe to increment
5723 ++new->current_threshold
;
5728 /* Free old spare buffer and save old primary buffer as spare */
5729 kfree(thresholds
->spare
);
5730 thresholds
->spare
= thresholds
->primary
;
5732 rcu_assign_pointer(thresholds
->primary
, new);
5734 /* To be sure that nobody uses thresholds */
5738 mutex_unlock(&memcg
->thresholds_lock
);
5743 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5744 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5746 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5747 struct mem_cgroup_thresholds
*thresholds
;
5748 struct mem_cgroup_threshold_ary
*new;
5749 enum res_type type
= MEMFILE_TYPE(cft
->private);
5753 mutex_lock(&memcg
->thresholds_lock
);
5755 thresholds
= &memcg
->thresholds
;
5756 else if (type
== _MEMSWAP
)
5757 thresholds
= &memcg
->memsw_thresholds
;
5761 if (!thresholds
->primary
)
5764 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5766 /* Check if a threshold crossed before removing */
5767 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5769 /* Calculate new number of threshold */
5771 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5772 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5776 new = thresholds
->spare
;
5778 /* Set thresholds array to NULL if we don't have thresholds */
5787 /* Copy thresholds and find current threshold */
5788 new->current_threshold
= -1;
5789 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5790 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5793 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5794 if (new->entries
[j
].threshold
<= usage
) {
5796 * new->current_threshold will not be used
5797 * until rcu_assign_pointer(), so it's safe to increment
5800 ++new->current_threshold
;
5806 /* Swap primary and spare array */
5807 thresholds
->spare
= thresholds
->primary
;
5808 /* If all events are unregistered, free the spare array */
5810 kfree(thresholds
->spare
);
5811 thresholds
->spare
= NULL
;
5814 rcu_assign_pointer(thresholds
->primary
, new);
5816 /* To be sure that nobody uses thresholds */
5819 mutex_unlock(&memcg
->thresholds_lock
);
5822 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5823 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5825 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5826 struct mem_cgroup_eventfd_list
*event
;
5827 enum res_type type
= MEMFILE_TYPE(cft
->private);
5829 BUG_ON(type
!= _OOM_TYPE
);
5830 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5834 spin_lock(&memcg_oom_lock
);
5836 event
->eventfd
= eventfd
;
5837 list_add(&event
->list
, &memcg
->oom_notify
);
5839 /* already in OOM ? */
5840 if (atomic_read(&memcg
->under_oom
))
5841 eventfd_signal(eventfd
, 1);
5842 spin_unlock(&memcg_oom_lock
);
5847 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5848 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5850 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5851 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5852 enum res_type type
= MEMFILE_TYPE(cft
->private);
5854 BUG_ON(type
!= _OOM_TYPE
);
5856 spin_lock(&memcg_oom_lock
);
5858 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5859 if (ev
->eventfd
== eventfd
) {
5860 list_del(&ev
->list
);
5865 spin_unlock(&memcg_oom_lock
);
5868 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
5869 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5871 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5873 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5875 if (atomic_read(&memcg
->under_oom
))
5876 cb
->fill(cb
, "under_oom", 1);
5878 cb
->fill(cb
, "under_oom", 0);
5882 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
5883 struct cftype
*cft
, u64 val
)
5885 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5886 struct mem_cgroup
*parent
;
5888 /* cannot set to root cgroup and only 0 and 1 are allowed */
5889 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
5892 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5894 mutex_lock(&memcg_create_mutex
);
5895 /* oom-kill-disable is a flag for subhierarchy. */
5896 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5897 mutex_unlock(&memcg_create_mutex
);
5900 memcg
->oom_kill_disable
= val
;
5902 memcg_oom_recover(memcg
);
5903 mutex_unlock(&memcg_create_mutex
);
5907 #ifdef CONFIG_MEMCG_KMEM
5908 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5912 memcg
->kmemcg_id
= -1;
5913 ret
= memcg_propagate_kmem(memcg
);
5917 return mem_cgroup_sockets_init(memcg
, ss
);
5920 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5922 mem_cgroup_sockets_destroy(memcg
);
5924 memcg_kmem_mark_dead(memcg
);
5926 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5930 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5931 * path here, being careful not to race with memcg_uncharge_kmem: it is
5932 * possible that the charges went down to 0 between mark_dead and the
5933 * res_counter read, so in that case, we don't need the put
5935 if (memcg_kmem_test_and_clear_dead(memcg
))
5936 mem_cgroup_put(memcg
);
5939 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5944 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5949 static struct cftype mem_cgroup_files
[] = {
5951 .name
= "usage_in_bytes",
5952 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5953 .read
= mem_cgroup_read
,
5954 .register_event
= mem_cgroup_usage_register_event
,
5955 .unregister_event
= mem_cgroup_usage_unregister_event
,
5958 .name
= "max_usage_in_bytes",
5959 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5960 .trigger
= mem_cgroup_reset
,
5961 .read
= mem_cgroup_read
,
5964 .name
= "limit_in_bytes",
5965 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5966 .write_string
= mem_cgroup_write
,
5967 .read
= mem_cgroup_read
,
5970 .name
= "soft_limit_in_bytes",
5971 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5972 .write_string
= mem_cgroup_write
,
5973 .read
= mem_cgroup_read
,
5977 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5978 .trigger
= mem_cgroup_reset
,
5979 .read
= mem_cgroup_read
,
5983 .read_seq_string
= memcg_stat_show
,
5986 .name
= "force_empty",
5987 .trigger
= mem_cgroup_force_empty_write
,
5990 .name
= "use_hierarchy",
5991 .flags
= CFTYPE_INSANE
,
5992 .write_u64
= mem_cgroup_hierarchy_write
,
5993 .read_u64
= mem_cgroup_hierarchy_read
,
5996 .name
= "swappiness",
5997 .read_u64
= mem_cgroup_swappiness_read
,
5998 .write_u64
= mem_cgroup_swappiness_write
,
6001 .name
= "move_charge_at_immigrate",
6002 .read_u64
= mem_cgroup_move_charge_read
,
6003 .write_u64
= mem_cgroup_move_charge_write
,
6006 .name
= "oom_control",
6007 .read_map
= mem_cgroup_oom_control_read
,
6008 .write_u64
= mem_cgroup_oom_control_write
,
6009 .register_event
= mem_cgroup_oom_register_event
,
6010 .unregister_event
= mem_cgroup_oom_unregister_event
,
6011 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6014 .name
= "pressure_level",
6015 .register_event
= vmpressure_register_event
,
6016 .unregister_event
= vmpressure_unregister_event
,
6020 .name
= "numa_stat",
6021 .read_seq_string
= memcg_numa_stat_show
,
6024 #ifdef CONFIG_MEMCG_KMEM
6026 .name
= "kmem.limit_in_bytes",
6027 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6028 .write_string
= mem_cgroup_write
,
6029 .read
= mem_cgroup_read
,
6032 .name
= "kmem.usage_in_bytes",
6033 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6034 .read
= mem_cgroup_read
,
6037 .name
= "kmem.failcnt",
6038 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6039 .trigger
= mem_cgroup_reset
,
6040 .read
= mem_cgroup_read
,
6043 .name
= "kmem.max_usage_in_bytes",
6044 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6045 .trigger
= mem_cgroup_reset
,
6046 .read
= mem_cgroup_read
,
6048 #ifdef CONFIG_SLABINFO
6050 .name
= "kmem.slabinfo",
6051 .read_seq_string
= mem_cgroup_slabinfo_read
,
6055 { }, /* terminate */
6058 #ifdef CONFIG_MEMCG_SWAP
6059 static struct cftype memsw_cgroup_files
[] = {
6061 .name
= "memsw.usage_in_bytes",
6062 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6063 .read
= mem_cgroup_read
,
6064 .register_event
= mem_cgroup_usage_register_event
,
6065 .unregister_event
= mem_cgroup_usage_unregister_event
,
6068 .name
= "memsw.max_usage_in_bytes",
6069 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6070 .trigger
= mem_cgroup_reset
,
6071 .read
= mem_cgroup_read
,
6074 .name
= "memsw.limit_in_bytes",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6076 .write_string
= mem_cgroup_write
,
6077 .read
= mem_cgroup_read
,
6080 .name
= "memsw.failcnt",
6081 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6082 .trigger
= mem_cgroup_reset
,
6083 .read
= mem_cgroup_read
,
6085 { }, /* terminate */
6088 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6090 struct mem_cgroup_per_node
*pn
;
6091 struct mem_cgroup_per_zone
*mz
;
6092 int zone
, tmp
= node
;
6094 * This routine is called against possible nodes.
6095 * But it's BUG to call kmalloc() against offline node.
6097 * TODO: this routine can waste much memory for nodes which will
6098 * never be onlined. It's better to use memory hotplug callback
6101 if (!node_state(node
, N_NORMAL_MEMORY
))
6103 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6107 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6108 mz
= &pn
->zoneinfo
[zone
];
6109 lruvec_init(&mz
->lruvec
);
6110 mz
->usage_in_excess
= 0;
6111 mz
->on_tree
= false;
6114 memcg
->info
.nodeinfo
[node
] = pn
;
6118 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6120 kfree(memcg
->info
.nodeinfo
[node
]);
6123 static struct mem_cgroup
*mem_cgroup_alloc(void)
6125 struct mem_cgroup
*memcg
;
6126 size_t size
= memcg_size();
6128 /* Can be very big if nr_node_ids is very big */
6129 if (size
< PAGE_SIZE
)
6130 memcg
= kzalloc(size
, GFP_KERNEL
);
6132 memcg
= vzalloc(size
);
6137 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6140 spin_lock_init(&memcg
->pcp_counter_lock
);
6144 if (size
< PAGE_SIZE
)
6152 * At destroying mem_cgroup, references from swap_cgroup can remain.
6153 * (scanning all at force_empty is too costly...)
6155 * Instead of clearing all references at force_empty, we remember
6156 * the number of reference from swap_cgroup and free mem_cgroup when
6157 * it goes down to 0.
6159 * Removal of cgroup itself succeeds regardless of refs from swap.
6162 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6165 size_t size
= memcg_size();
6167 mem_cgroup_remove_from_trees(memcg
);
6168 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
6171 free_mem_cgroup_per_zone_info(memcg
, node
);
6173 free_percpu(memcg
->stat
);
6176 * We need to make sure that (at least for now), the jump label
6177 * destruction code runs outside of the cgroup lock. This is because
6178 * get_online_cpus(), which is called from the static_branch update,
6179 * can't be called inside the cgroup_lock. cpusets are the ones
6180 * enforcing this dependency, so if they ever change, we might as well.
6182 * schedule_work() will guarantee this happens. Be careful if you need
6183 * to move this code around, and make sure it is outside
6186 disarm_static_keys(memcg
);
6187 if (size
< PAGE_SIZE
)
6195 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6196 * but in process context. The work_freeing structure is overlaid
6197 * on the rcu_freeing structure, which itself is overlaid on memsw.
6199 static void free_work(struct work_struct
*work
)
6201 struct mem_cgroup
*memcg
;
6203 memcg
= container_of(work
, struct mem_cgroup
, work_freeing
);
6204 __mem_cgroup_free(memcg
);
6207 static void free_rcu(struct rcu_head
*rcu_head
)
6209 struct mem_cgroup
*memcg
;
6211 memcg
= container_of(rcu_head
, struct mem_cgroup
, rcu_freeing
);
6212 INIT_WORK(&memcg
->work_freeing
, free_work
);
6213 schedule_work(&memcg
->work_freeing
);
6216 static void mem_cgroup_get(struct mem_cgroup
*memcg
)
6218 atomic_inc(&memcg
->refcnt
);
6221 static void __mem_cgroup_put(struct mem_cgroup
*memcg
, int count
)
6223 if (atomic_sub_and_test(count
, &memcg
->refcnt
)) {
6224 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
6225 call_rcu(&memcg
->rcu_freeing
, free_rcu
);
6227 mem_cgroup_put(parent
);
6231 static void mem_cgroup_put(struct mem_cgroup
*memcg
)
6233 __mem_cgroup_put(memcg
, 1);
6237 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6239 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6241 if (!memcg
->res
.parent
)
6243 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6245 EXPORT_SYMBOL(parent_mem_cgroup
);
6247 static void __init
mem_cgroup_soft_limit_tree_init(void)
6249 struct mem_cgroup_tree_per_node
*rtpn
;
6250 struct mem_cgroup_tree_per_zone
*rtpz
;
6251 int tmp
, node
, zone
;
6253 for_each_node(node
) {
6255 if (!node_state(node
, N_NORMAL_MEMORY
))
6257 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6260 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6262 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6263 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6264 rtpz
->rb_root
= RB_ROOT
;
6265 spin_lock_init(&rtpz
->lock
);
6270 static struct cgroup_subsys_state
* __ref
6271 mem_cgroup_css_alloc(struct cgroup
*cont
)
6273 struct mem_cgroup
*memcg
;
6274 long error
= -ENOMEM
;
6277 memcg
= mem_cgroup_alloc();
6279 return ERR_PTR(error
);
6282 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6286 if (cont
->parent
== NULL
) {
6287 root_mem_cgroup
= memcg
;
6288 res_counter_init(&memcg
->res
, NULL
);
6289 res_counter_init(&memcg
->memsw
, NULL
);
6290 res_counter_init(&memcg
->kmem
, NULL
);
6293 memcg
->last_scanned_node
= MAX_NUMNODES
;
6294 INIT_LIST_HEAD(&memcg
->oom_notify
);
6295 atomic_set(&memcg
->refcnt
, 1);
6296 memcg
->move_charge_at_immigrate
= 0;
6297 mutex_init(&memcg
->thresholds_lock
);
6298 spin_lock_init(&memcg
->move_lock
);
6299 vmpressure_init(&memcg
->vmpressure
);
6304 __mem_cgroup_free(memcg
);
6305 return ERR_PTR(error
);
6309 mem_cgroup_css_online(struct cgroup
*cont
)
6311 struct mem_cgroup
*memcg
, *parent
;
6317 mutex_lock(&memcg_create_mutex
);
6318 memcg
= mem_cgroup_from_cont(cont
);
6319 parent
= mem_cgroup_from_cont(cont
->parent
);
6321 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6322 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6323 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6325 if (parent
->use_hierarchy
) {
6326 res_counter_init(&memcg
->res
, &parent
->res
);
6327 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6328 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6331 * We increment refcnt of the parent to ensure that we can
6332 * safely access it on res_counter_charge/uncharge.
6333 * This refcnt will be decremented when freeing this
6334 * mem_cgroup(see mem_cgroup_put).
6336 mem_cgroup_get(parent
);
6338 res_counter_init(&memcg
->res
, NULL
);
6339 res_counter_init(&memcg
->memsw
, NULL
);
6340 res_counter_init(&memcg
->kmem
, NULL
);
6342 * Deeper hierachy with use_hierarchy == false doesn't make
6343 * much sense so let cgroup subsystem know about this
6344 * unfortunate state in our controller.
6346 if (parent
!= root_mem_cgroup
)
6347 mem_cgroup_subsys
.broken_hierarchy
= true;
6350 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6351 mutex_unlock(&memcg_create_mutex
);
6356 * Announce all parents that a group from their hierarchy is gone.
6358 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6360 struct mem_cgroup
*parent
= memcg
;
6362 while ((parent
= parent_mem_cgroup(parent
)))
6363 atomic_inc(&parent
->dead_count
);
6366 * if the root memcg is not hierarchical we have to check it
6369 if (!root_mem_cgroup
->use_hierarchy
)
6370 atomic_inc(&root_mem_cgroup
->dead_count
);
6373 static void mem_cgroup_css_offline(struct cgroup
*cont
)
6375 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6376 struct cgroup
*iter
;
6378 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6381 * This requires that offlining is serialized. Right now that is
6382 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6385 cgroup_for_each_descendant_post(iter
, cont
) {
6387 mem_cgroup_reparent_charges(mem_cgroup_from_cont(iter
));
6391 mem_cgroup_reparent_charges(memcg
);
6393 mem_cgroup_destroy_all_caches(memcg
);
6396 static void mem_cgroup_css_free(struct cgroup
*cont
)
6398 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6400 kmem_cgroup_destroy(memcg
);
6402 mem_cgroup_put(memcg
);
6406 /* Handlers for move charge at task migration. */
6407 #define PRECHARGE_COUNT_AT_ONCE 256
6408 static int mem_cgroup_do_precharge(unsigned long count
)
6411 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6412 struct mem_cgroup
*memcg
= mc
.to
;
6414 if (mem_cgroup_is_root(memcg
)) {
6415 mc
.precharge
+= count
;
6416 /* we don't need css_get for root */
6419 /* try to charge at once */
6421 struct res_counter
*dummy
;
6423 * "memcg" cannot be under rmdir() because we've already checked
6424 * by cgroup_lock_live_cgroup() that it is not removed and we
6425 * are still under the same cgroup_mutex. So we can postpone
6428 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6430 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6431 PAGE_SIZE
* count
, &dummy
)) {
6432 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6435 mc
.precharge
+= count
;
6439 /* fall back to one by one charge */
6441 if (signal_pending(current
)) {
6445 if (!batch_count
--) {
6446 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6449 ret
= __mem_cgroup_try_charge(NULL
,
6450 GFP_KERNEL
, 1, &memcg
, false);
6452 /* mem_cgroup_clear_mc() will do uncharge later */
6460 * get_mctgt_type - get target type of moving charge
6461 * @vma: the vma the pte to be checked belongs
6462 * @addr: the address corresponding to the pte to be checked
6463 * @ptent: the pte to be checked
6464 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6467 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6468 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6469 * move charge. if @target is not NULL, the page is stored in target->page
6470 * with extra refcnt got(Callers should handle it).
6471 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6472 * target for charge migration. if @target is not NULL, the entry is stored
6475 * Called with pte lock held.
6482 enum mc_target_type
{
6488 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6489 unsigned long addr
, pte_t ptent
)
6491 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6493 if (!page
|| !page_mapped(page
))
6495 if (PageAnon(page
)) {
6496 /* we don't move shared anon */
6499 } else if (!move_file())
6500 /* we ignore mapcount for file pages */
6502 if (!get_page_unless_zero(page
))
6509 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6510 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6512 struct page
*page
= NULL
;
6513 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6515 if (!move_anon() || non_swap_entry(ent
))
6518 * Because lookup_swap_cache() updates some statistics counter,
6519 * we call find_get_page() with swapper_space directly.
6521 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6522 if (do_swap_account
)
6523 entry
->val
= ent
.val
;
6528 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6529 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6535 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6536 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6538 struct page
*page
= NULL
;
6539 struct address_space
*mapping
;
6542 if (!vma
->vm_file
) /* anonymous vma */
6547 mapping
= vma
->vm_file
->f_mapping
;
6548 if (pte_none(ptent
))
6549 pgoff
= linear_page_index(vma
, addr
);
6550 else /* pte_file(ptent) is true */
6551 pgoff
= pte_to_pgoff(ptent
);
6553 /* page is moved even if it's not RSS of this task(page-faulted). */
6554 page
= find_get_page(mapping
, pgoff
);
6557 /* shmem/tmpfs may report page out on swap: account for that too. */
6558 if (radix_tree_exceptional_entry(page
)) {
6559 swp_entry_t swap
= radix_to_swp_entry(page
);
6560 if (do_swap_account
)
6562 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6568 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6569 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6571 struct page
*page
= NULL
;
6572 struct page_cgroup
*pc
;
6573 enum mc_target_type ret
= MC_TARGET_NONE
;
6574 swp_entry_t ent
= { .val
= 0 };
6576 if (pte_present(ptent
))
6577 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6578 else if (is_swap_pte(ptent
))
6579 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6580 else if (pte_none(ptent
) || pte_file(ptent
))
6581 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6583 if (!page
&& !ent
.val
)
6586 pc
= lookup_page_cgroup(page
);
6588 * Do only loose check w/o page_cgroup lock.
6589 * mem_cgroup_move_account() checks the pc is valid or not under
6592 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6593 ret
= MC_TARGET_PAGE
;
6595 target
->page
= page
;
6597 if (!ret
|| !target
)
6600 /* There is a swap entry and a page doesn't exist or isn't charged */
6601 if (ent
.val
&& !ret
&&
6602 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6603 ret
= MC_TARGET_SWAP
;
6610 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6612 * We don't consider swapping or file mapped pages because THP does not
6613 * support them for now.
6614 * Caller should make sure that pmd_trans_huge(pmd) is true.
6616 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6617 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6619 struct page
*page
= NULL
;
6620 struct page_cgroup
*pc
;
6621 enum mc_target_type ret
= MC_TARGET_NONE
;
6623 page
= pmd_page(pmd
);
6624 VM_BUG_ON(!page
|| !PageHead(page
));
6627 pc
= lookup_page_cgroup(page
);
6628 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6629 ret
= MC_TARGET_PAGE
;
6632 target
->page
= page
;
6638 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6639 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6641 return MC_TARGET_NONE
;
6645 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6646 unsigned long addr
, unsigned long end
,
6647 struct mm_walk
*walk
)
6649 struct vm_area_struct
*vma
= walk
->private;
6653 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6654 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6655 mc
.precharge
+= HPAGE_PMD_NR
;
6656 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6660 if (pmd_trans_unstable(pmd
))
6662 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6663 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6664 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6665 mc
.precharge
++; /* increment precharge temporarily */
6666 pte_unmap_unlock(pte
- 1, ptl
);
6672 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6674 unsigned long precharge
;
6675 struct vm_area_struct
*vma
;
6677 down_read(&mm
->mmap_sem
);
6678 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6679 struct mm_walk mem_cgroup_count_precharge_walk
= {
6680 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6684 if (is_vm_hugetlb_page(vma
))
6686 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6687 &mem_cgroup_count_precharge_walk
);
6689 up_read(&mm
->mmap_sem
);
6691 precharge
= mc
.precharge
;
6697 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6699 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6701 VM_BUG_ON(mc
.moving_task
);
6702 mc
.moving_task
= current
;
6703 return mem_cgroup_do_precharge(precharge
);
6706 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6707 static void __mem_cgroup_clear_mc(void)
6709 struct mem_cgroup
*from
= mc
.from
;
6710 struct mem_cgroup
*to
= mc
.to
;
6712 /* we must uncharge all the leftover precharges from mc.to */
6714 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6718 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6719 * we must uncharge here.
6721 if (mc
.moved_charge
) {
6722 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6723 mc
.moved_charge
= 0;
6725 /* we must fixup refcnts and charges */
6726 if (mc
.moved_swap
) {
6727 /* uncharge swap account from the old cgroup */
6728 if (!mem_cgroup_is_root(mc
.from
))
6729 res_counter_uncharge(&mc
.from
->memsw
,
6730 PAGE_SIZE
* mc
.moved_swap
);
6731 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
6733 if (!mem_cgroup_is_root(mc
.to
)) {
6735 * we charged both to->res and to->memsw, so we should
6738 res_counter_uncharge(&mc
.to
->res
,
6739 PAGE_SIZE
* mc
.moved_swap
);
6741 /* we've already done mem_cgroup_get(mc.to) */
6744 memcg_oom_recover(from
);
6745 memcg_oom_recover(to
);
6746 wake_up_all(&mc
.waitq
);
6749 static void mem_cgroup_clear_mc(void)
6751 struct mem_cgroup
*from
= mc
.from
;
6754 * we must clear moving_task before waking up waiters at the end of
6757 mc
.moving_task
= NULL
;
6758 __mem_cgroup_clear_mc();
6759 spin_lock(&mc
.lock
);
6762 spin_unlock(&mc
.lock
);
6763 mem_cgroup_end_move(from
);
6766 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6767 struct cgroup_taskset
*tset
)
6769 struct task_struct
*p
= cgroup_taskset_first(tset
);
6771 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgroup
);
6772 unsigned long move_charge_at_immigrate
;
6775 * We are now commited to this value whatever it is. Changes in this
6776 * tunable will only affect upcoming migrations, not the current one.
6777 * So we need to save it, and keep it going.
6779 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6780 if (move_charge_at_immigrate
) {
6781 struct mm_struct
*mm
;
6782 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6784 VM_BUG_ON(from
== memcg
);
6786 mm
= get_task_mm(p
);
6789 /* We move charges only when we move a owner of the mm */
6790 if (mm
->owner
== p
) {
6793 VM_BUG_ON(mc
.precharge
);
6794 VM_BUG_ON(mc
.moved_charge
);
6795 VM_BUG_ON(mc
.moved_swap
);
6796 mem_cgroup_start_move(from
);
6797 spin_lock(&mc
.lock
);
6800 mc
.immigrate_flags
= move_charge_at_immigrate
;
6801 spin_unlock(&mc
.lock
);
6802 /* We set mc.moving_task later */
6804 ret
= mem_cgroup_precharge_mc(mm
);
6806 mem_cgroup_clear_mc();
6813 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6814 struct cgroup_taskset
*tset
)
6816 mem_cgroup_clear_mc();
6819 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6820 unsigned long addr
, unsigned long end
,
6821 struct mm_walk
*walk
)
6824 struct vm_area_struct
*vma
= walk
->private;
6827 enum mc_target_type target_type
;
6828 union mc_target target
;
6830 struct page_cgroup
*pc
;
6833 * We don't take compound_lock() here but no race with splitting thp
6835 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6836 * under splitting, which means there's no concurrent thp split,
6837 * - if another thread runs into split_huge_page() just after we
6838 * entered this if-block, the thread must wait for page table lock
6839 * to be unlocked in __split_huge_page_splitting(), where the main
6840 * part of thp split is not executed yet.
6842 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6843 if (mc
.precharge
< HPAGE_PMD_NR
) {
6844 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6847 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6848 if (target_type
== MC_TARGET_PAGE
) {
6850 if (!isolate_lru_page(page
)) {
6851 pc
= lookup_page_cgroup(page
);
6852 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6853 pc
, mc
.from
, mc
.to
)) {
6854 mc
.precharge
-= HPAGE_PMD_NR
;
6855 mc
.moved_charge
+= HPAGE_PMD_NR
;
6857 putback_lru_page(page
);
6861 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6865 if (pmd_trans_unstable(pmd
))
6868 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6869 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6870 pte_t ptent
= *(pte
++);
6876 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6877 case MC_TARGET_PAGE
:
6879 if (isolate_lru_page(page
))
6881 pc
= lookup_page_cgroup(page
);
6882 if (!mem_cgroup_move_account(page
, 1, pc
,
6885 /* we uncharge from mc.from later. */
6888 putback_lru_page(page
);
6889 put
: /* get_mctgt_type() gets the page */
6892 case MC_TARGET_SWAP
:
6894 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6896 /* we fixup refcnts and charges later. */
6904 pte_unmap_unlock(pte
- 1, ptl
);
6909 * We have consumed all precharges we got in can_attach().
6910 * We try charge one by one, but don't do any additional
6911 * charges to mc.to if we have failed in charge once in attach()
6914 ret
= mem_cgroup_do_precharge(1);
6922 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6924 struct vm_area_struct
*vma
;
6926 lru_add_drain_all();
6928 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6930 * Someone who are holding the mmap_sem might be waiting in
6931 * waitq. So we cancel all extra charges, wake up all waiters,
6932 * and retry. Because we cancel precharges, we might not be able
6933 * to move enough charges, but moving charge is a best-effort
6934 * feature anyway, so it wouldn't be a big problem.
6936 __mem_cgroup_clear_mc();
6940 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6942 struct mm_walk mem_cgroup_move_charge_walk
= {
6943 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6947 if (is_vm_hugetlb_page(vma
))
6949 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6950 &mem_cgroup_move_charge_walk
);
6953 * means we have consumed all precharges and failed in
6954 * doing additional charge. Just abandon here.
6958 up_read(&mm
->mmap_sem
);
6961 static void mem_cgroup_move_task(struct cgroup
*cont
,
6962 struct cgroup_taskset
*tset
)
6964 struct task_struct
*p
= cgroup_taskset_first(tset
);
6965 struct mm_struct
*mm
= get_task_mm(p
);
6969 mem_cgroup_move_charge(mm
);
6973 mem_cgroup_clear_mc();
6975 #else /* !CONFIG_MMU */
6976 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6977 struct cgroup_taskset
*tset
)
6981 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6982 struct cgroup_taskset
*tset
)
6985 static void mem_cgroup_move_task(struct cgroup
*cont
,
6986 struct cgroup_taskset
*tset
)
6992 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6993 * to verify sane_behavior flag on each mount attempt.
6995 static void mem_cgroup_bind(struct cgroup
*root
)
6998 * use_hierarchy is forced with sane_behavior. cgroup core
6999 * guarantees that @root doesn't have any children, so turning it
7000 * on for the root memcg is enough.
7002 if (cgroup_sane_behavior(root
))
7003 mem_cgroup_from_cont(root
)->use_hierarchy
= true;
7006 struct cgroup_subsys mem_cgroup_subsys
= {
7008 .subsys_id
= mem_cgroup_subsys_id
,
7009 .css_alloc
= mem_cgroup_css_alloc
,
7010 .css_online
= mem_cgroup_css_online
,
7011 .css_offline
= mem_cgroup_css_offline
,
7012 .css_free
= mem_cgroup_css_free
,
7013 .can_attach
= mem_cgroup_can_attach
,
7014 .cancel_attach
= mem_cgroup_cancel_attach
,
7015 .attach
= mem_cgroup_move_task
,
7016 .bind
= mem_cgroup_bind
,
7017 .base_cftypes
= mem_cgroup_files
,
7022 #ifdef CONFIG_MEMCG_SWAP
7023 static int __init
enable_swap_account(char *s
)
7025 /* consider enabled if no parameter or 1 is given */
7026 if (!strcmp(s
, "1"))
7027 really_do_swap_account
= 1;
7028 else if (!strcmp(s
, "0"))
7029 really_do_swap_account
= 0;
7032 __setup("swapaccount=", enable_swap_account
);
7034 static void __init
memsw_file_init(void)
7036 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
7039 static void __init
enable_swap_cgroup(void)
7041 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7042 do_swap_account
= 1;
7048 static void __init
enable_swap_cgroup(void)
7054 * subsys_initcall() for memory controller.
7056 * Some parts like hotcpu_notifier() have to be initialized from this context
7057 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7058 * everything that doesn't depend on a specific mem_cgroup structure should
7059 * be initialized from here.
7061 static int __init
mem_cgroup_init(void)
7063 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7064 enable_swap_cgroup();
7065 mem_cgroup_soft_limit_tree_init();
7069 subsys_initcall(mem_cgroup_init
);