1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
70 EXPORT_SYMBOL(memory_cgrp_subsys
);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly
;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata
= 1;
83 static int really_do_swap_account __initdata
;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names
[] = {
100 enum mem_cgroup_events_index
{
101 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS
,
108 static const char * const mem_cgroup_events_names
[] = {
115 static const char * const mem_cgroup_lru_names
[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target
{
130 MEM_CGROUP_TARGET_THRESH
,
131 MEM_CGROUP_TARGET_SOFTLIMIT
,
132 MEM_CGROUP_TARGET_NUMAINFO
,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu
{
140 long count
[MEM_CGROUP_STAT_NSTATS
];
141 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
142 unsigned long nr_page_events
;
143 unsigned long targets
[MEM_CGROUP_NTARGETS
];
146 struct mem_cgroup_reclaim_iter
{
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup
*last_visited
;
154 /* scan generation, increased every round-trip */
155 unsigned int generation
;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone
{
162 struct lruvec lruvec
;
163 unsigned long lru_size
[NR_LRU_LISTS
];
165 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
167 struct rb_node tree_node
; /* RB tree node */
168 unsigned long long usage_in_excess
;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node
{
176 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone
{
185 struct rb_root rb_root
;
189 struct mem_cgroup_tree_per_node
{
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
193 struct mem_cgroup_tree
{
194 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
199 struct mem_cgroup_threshold
{
200 struct eventfd_ctx
*eventfd
;
205 struct mem_cgroup_threshold_ary
{
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold
;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries
[0];
214 struct mem_cgroup_thresholds
{
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary
*primary
;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary
*spare
;
226 struct mem_cgroup_eventfd_list
{
227 struct list_head list
;
228 struct eventfd_ctx
*eventfd
;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event
{
236 * memcg which the event belongs to.
238 struct mem_cgroup
*memcg
;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx
*eventfd
;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list
;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event
)(struct mem_cgroup
*memcg
,
253 struct eventfd_ctx
*eventfd
, const char *args
);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event
)(struct mem_cgroup
*memcg
,
260 struct eventfd_ctx
*eventfd
);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t
*wqh
;
268 struct work_struct remove
;
271 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
272 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css
;
288 * the counter to account for memory usage
290 struct res_counter res
;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure
;
295 /* css_online() has been completed */
299 * the counter to account for mem+swap usage.
301 struct res_counter memsw
;
304 * the counter to account for kernel memory usage.
306 struct res_counter kmem
;
308 * Should the accounting and control be hierarchical, per subtree?
311 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
315 atomic_t oom_wakeups
;
318 /* OOM-Killer disable */
319 int oom_kill_disable
;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock
;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds
;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds
;
330 /* For oom notifier event fd */
331 struct list_head oom_notify
;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate
;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account
;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock
;
347 struct mem_cgroup_stat_cpu __percpu
*stat
;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base
;
353 spinlock_t pcp_counter_lock
;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem
;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list, but per-memcg;
361 * protected by memcg_slab_mutex */
362 struct list_head memcg_slab_caches
;
363 /* Index in the kmem_cache->memcg_params->memcg_caches array */
367 int last_scanned_node
;
369 nodemask_t scan_nodes
;
370 atomic_t numainfo_events
;
371 atomic_t numainfo_updating
;
374 /* List of events which userspace want to receive */
375 struct list_head event_list
;
376 spinlock_t event_list_lock
;
378 struct mem_cgroup_per_node
*nodeinfo
[0];
379 /* WARNING: nodeinfo must be the last member here */
382 /* internal only representation about the status of kmem accounting. */
384 KMEM_ACCOUNTED_ACTIVE
, /* accounted by this cgroup itself */
385 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
388 #ifdef CONFIG_MEMCG_KMEM
389 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
391 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
394 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
396 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
399 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
402 * Our caller must use css_get() first, because memcg_uncharge_kmem()
403 * will call css_put() if it sees the memcg is dead.
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
407 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
413 &memcg
->kmem_account_flags
);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct
{
430 spinlock_t lock
; /* for from, to */
431 struct mem_cgroup
*from
;
432 struct mem_cgroup
*to
;
433 unsigned long immigrate_flags
;
434 unsigned long precharge
;
435 unsigned long moved_charge
;
436 unsigned long moved_swap
;
437 struct task_struct
*moving_task
; /* a task moving charges */
438 wait_queue_head_t waitq
; /* a waitq for other context */
440 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
441 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON
,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * The memcg_create_mutex will be held whenever a new cgroup is created.
485 * As a consequence, any change that needs to protect against new child cgroups
486 * appearing has to hold it as well.
488 static DEFINE_MUTEX(memcg_create_mutex
);
490 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
492 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
495 /* Some nice accessors for the vmpressure. */
496 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
499 memcg
= root_mem_cgroup
;
500 return &memcg
->vmpressure
;
503 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
505 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
508 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
510 return (memcg
== root_mem_cgroup
);
514 * We restrict the id in the range of [1, 65535], so it can fit into
517 #define MEM_CGROUP_ID_MAX USHRT_MAX
519 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
521 return memcg
->css
.id
;
524 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
526 struct cgroup_subsys_state
*css
;
528 css
= css_from_id(id
, &memory_cgrp_subsys
);
529 return mem_cgroup_from_css(css
);
532 /* Writing them here to avoid exposing memcg's inner layout */
533 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
535 void sock_update_memcg(struct sock
*sk
)
537 if (mem_cgroup_sockets_enabled
) {
538 struct mem_cgroup
*memcg
;
539 struct cg_proto
*cg_proto
;
541 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
543 /* Socket cloning can throw us here with sk_cgrp already
544 * filled. It won't however, necessarily happen from
545 * process context. So the test for root memcg given
546 * the current task's memcg won't help us in this case.
548 * Respecting the original socket's memcg is a better
549 * decision in this case.
552 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
553 css_get(&sk
->sk_cgrp
->memcg
->css
);
558 memcg
= mem_cgroup_from_task(current
);
559 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
560 if (!mem_cgroup_is_root(memcg
) &&
561 memcg_proto_active(cg_proto
) &&
562 css_tryget_online(&memcg
->css
)) {
563 sk
->sk_cgrp
= cg_proto
;
568 EXPORT_SYMBOL(sock_update_memcg
);
570 void sock_release_memcg(struct sock
*sk
)
572 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
573 struct mem_cgroup
*memcg
;
574 WARN_ON(!sk
->sk_cgrp
->memcg
);
575 memcg
= sk
->sk_cgrp
->memcg
;
576 css_put(&sk
->sk_cgrp
->memcg
->css
);
580 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
582 if (!memcg
|| mem_cgroup_is_root(memcg
))
585 return &memcg
->tcp_mem
;
587 EXPORT_SYMBOL(tcp_proto_cgroup
);
589 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
591 if (!memcg_proto_activated(&memcg
->tcp_mem
))
593 static_key_slow_dec(&memcg_socket_limit_enabled
);
596 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
601 #ifdef CONFIG_MEMCG_KMEM
603 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
604 * The main reason for not using cgroup id for this:
605 * this works better in sparse environments, where we have a lot of memcgs,
606 * but only a few kmem-limited. Or also, if we have, for instance, 200
607 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
608 * 200 entry array for that.
610 * The current size of the caches array is stored in
611 * memcg_limited_groups_array_size. It will double each time we have to
614 static DEFINE_IDA(kmem_limited_groups
);
615 int memcg_limited_groups_array_size
;
618 * MIN_SIZE is different than 1, because we would like to avoid going through
619 * the alloc/free process all the time. In a small machine, 4 kmem-limited
620 * cgroups is a reasonable guess. In the future, it could be a parameter or
621 * tunable, but that is strictly not necessary.
623 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
624 * this constant directly from cgroup, but it is understandable that this is
625 * better kept as an internal representation in cgroup.c. In any case, the
626 * cgrp_id space is not getting any smaller, and we don't have to necessarily
627 * increase ours as well if it increases.
629 #define MEMCG_CACHES_MIN_SIZE 4
630 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
633 * A lot of the calls to the cache allocation functions are expected to be
634 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
635 * conditional to this static branch, we'll have to allow modules that does
636 * kmem_cache_alloc and the such to see this symbol as well
638 struct static_key memcg_kmem_enabled_key
;
639 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
641 static void memcg_free_cache_id(int id
);
643 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
645 if (memcg_kmem_is_active(memcg
)) {
646 static_key_slow_dec(&memcg_kmem_enabled_key
);
647 memcg_free_cache_id(memcg
->kmemcg_id
);
650 * This check can't live in kmem destruction function,
651 * since the charges will outlive the cgroup
653 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
656 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
659 #endif /* CONFIG_MEMCG_KMEM */
661 static void disarm_static_keys(struct mem_cgroup
*memcg
)
663 disarm_sock_keys(memcg
);
664 disarm_kmem_keys(memcg
);
667 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
669 static struct mem_cgroup_per_zone
*
670 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
672 int nid
= zone_to_nid(zone
);
673 int zid
= zone_idx(zone
);
675 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
678 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
683 static struct mem_cgroup_per_zone
*
684 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
686 int nid
= page_to_nid(page
);
687 int zid
= page_zonenum(page
);
689 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
692 static struct mem_cgroup_tree_per_zone
*
693 soft_limit_tree_node_zone(int nid
, int zid
)
695 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
698 static struct mem_cgroup_tree_per_zone
*
699 soft_limit_tree_from_page(struct page
*page
)
701 int nid
= page_to_nid(page
);
702 int zid
= page_zonenum(page
);
704 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
707 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
708 struct mem_cgroup_tree_per_zone
*mctz
,
709 unsigned long long new_usage_in_excess
)
711 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
712 struct rb_node
*parent
= NULL
;
713 struct mem_cgroup_per_zone
*mz_node
;
718 mz
->usage_in_excess
= new_usage_in_excess
;
719 if (!mz
->usage_in_excess
)
723 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
725 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
728 * We can't avoid mem cgroups that are over their soft
729 * limit by the same amount
731 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
734 rb_link_node(&mz
->tree_node
, parent
, p
);
735 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
739 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
740 struct mem_cgroup_tree_per_zone
*mctz
)
744 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
748 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
749 struct mem_cgroup_tree_per_zone
*mctz
)
753 spin_lock_irqsave(&mctz
->lock
, flags
);
754 __mem_cgroup_remove_exceeded(mz
, mctz
);
755 spin_unlock_irqrestore(&mctz
->lock
, flags
);
759 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
761 unsigned long long excess
;
762 struct mem_cgroup_per_zone
*mz
;
763 struct mem_cgroup_tree_per_zone
*mctz
;
765 mctz
= soft_limit_tree_from_page(page
);
767 * Necessary to update all ancestors when hierarchy is used.
768 * because their event counter is not touched.
770 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
771 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
772 excess
= res_counter_soft_limit_excess(&memcg
->res
);
774 * We have to update the tree if mz is on RB-tree or
775 * mem is over its softlimit.
777 if (excess
|| mz
->on_tree
) {
780 spin_lock_irqsave(&mctz
->lock
, flags
);
781 /* if on-tree, remove it */
783 __mem_cgroup_remove_exceeded(mz
, mctz
);
785 * Insert again. mz->usage_in_excess will be updated.
786 * If excess is 0, no tree ops.
788 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
789 spin_unlock_irqrestore(&mctz
->lock
, flags
);
794 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
796 struct mem_cgroup_tree_per_zone
*mctz
;
797 struct mem_cgroup_per_zone
*mz
;
801 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
802 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
803 mctz
= soft_limit_tree_node_zone(nid
, zid
);
804 mem_cgroup_remove_exceeded(mz
, mctz
);
809 static struct mem_cgroup_per_zone
*
810 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
812 struct rb_node
*rightmost
= NULL
;
813 struct mem_cgroup_per_zone
*mz
;
817 rightmost
= rb_last(&mctz
->rb_root
);
819 goto done
; /* Nothing to reclaim from */
821 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
823 * Remove the node now but someone else can add it back,
824 * we will to add it back at the end of reclaim to its correct
825 * position in the tree.
827 __mem_cgroup_remove_exceeded(mz
, mctz
);
828 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
829 !css_tryget_online(&mz
->memcg
->css
))
835 static struct mem_cgroup_per_zone
*
836 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
838 struct mem_cgroup_per_zone
*mz
;
840 spin_lock_irq(&mctz
->lock
);
841 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
842 spin_unlock_irq(&mctz
->lock
);
847 * Implementation Note: reading percpu statistics for memcg.
849 * Both of vmstat[] and percpu_counter has threshold and do periodic
850 * synchronization to implement "quick" read. There are trade-off between
851 * reading cost and precision of value. Then, we may have a chance to implement
852 * a periodic synchronizion of counter in memcg's counter.
854 * But this _read() function is used for user interface now. The user accounts
855 * memory usage by memory cgroup and he _always_ requires exact value because
856 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
857 * have to visit all online cpus and make sum. So, for now, unnecessary
858 * synchronization is not implemented. (just implemented for cpu hotplug)
860 * If there are kernel internal actions which can make use of some not-exact
861 * value, and reading all cpu value can be performance bottleneck in some
862 * common workload, threashold and synchonization as vmstat[] should be
865 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
866 enum mem_cgroup_stat_index idx
)
872 for_each_online_cpu(cpu
)
873 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
874 #ifdef CONFIG_HOTPLUG_CPU
875 spin_lock(&memcg
->pcp_counter_lock
);
876 val
+= memcg
->nocpu_base
.count
[idx
];
877 spin_unlock(&memcg
->pcp_counter_lock
);
883 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
884 enum mem_cgroup_events_index idx
)
886 unsigned long val
= 0;
890 for_each_online_cpu(cpu
)
891 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
892 #ifdef CONFIG_HOTPLUG_CPU
893 spin_lock(&memcg
->pcp_counter_lock
);
894 val
+= memcg
->nocpu_base
.events
[idx
];
895 spin_unlock(&memcg
->pcp_counter_lock
);
901 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
906 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
907 * counted as CACHE even if it's on ANON LRU.
910 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
913 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
916 if (PageTransHuge(page
))
917 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
920 /* pagein of a big page is an event. So, ignore page size */
922 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
924 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
925 nr_pages
= -nr_pages
; /* for event */
928 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
931 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
933 struct mem_cgroup_per_zone
*mz
;
935 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
936 return mz
->lru_size
[lru
];
939 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
941 unsigned int lru_mask
)
943 unsigned long nr
= 0;
946 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
948 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
949 struct mem_cgroup_per_zone
*mz
;
953 if (!(BIT(lru
) & lru_mask
))
955 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
956 nr
+= mz
->lru_size
[lru
];
962 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
963 unsigned int lru_mask
)
965 unsigned long nr
= 0;
968 for_each_node_state(nid
, N_MEMORY
)
969 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
973 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
974 enum mem_cgroup_events_target target
)
976 unsigned long val
, next
;
978 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
979 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
980 /* from time_after() in jiffies.h */
981 if ((long)next
- (long)val
< 0) {
983 case MEM_CGROUP_TARGET_THRESH
:
984 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
986 case MEM_CGROUP_TARGET_SOFTLIMIT
:
987 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
989 case MEM_CGROUP_TARGET_NUMAINFO
:
990 next
= val
+ NUMAINFO_EVENTS_TARGET
;
995 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1002 * Check events in order.
1005 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1007 /* threshold event is triggered in finer grain than soft limit */
1008 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1009 MEM_CGROUP_TARGET_THRESH
))) {
1011 bool do_numainfo __maybe_unused
;
1013 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1014 MEM_CGROUP_TARGET_SOFTLIMIT
);
1015 #if MAX_NUMNODES > 1
1016 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1017 MEM_CGROUP_TARGET_NUMAINFO
);
1019 mem_cgroup_threshold(memcg
);
1020 if (unlikely(do_softlimit
))
1021 mem_cgroup_update_tree(memcg
, page
);
1022 #if MAX_NUMNODES > 1
1023 if (unlikely(do_numainfo
))
1024 atomic_inc(&memcg
->numainfo_events
);
1029 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1032 * mm_update_next_owner() may clear mm->owner to NULL
1033 * if it races with swapoff, page migration, etc.
1034 * So this can be called with p == NULL.
1039 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1042 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1044 struct mem_cgroup
*memcg
= NULL
;
1049 * Page cache insertions can happen withou an
1050 * actual mm context, e.g. during disk probing
1051 * on boot, loopback IO, acct() writes etc.
1054 memcg
= root_mem_cgroup
;
1056 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1057 if (unlikely(!memcg
))
1058 memcg
= root_mem_cgroup
;
1060 } while (!css_tryget_online(&memcg
->css
));
1066 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1067 * ref. count) or NULL if the whole root's subtree has been visited.
1069 * helper function to be used by mem_cgroup_iter
1071 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1072 struct mem_cgroup
*last_visited
)
1074 struct cgroup_subsys_state
*prev_css
, *next_css
;
1076 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1078 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1081 * Even if we found a group we have to make sure it is
1082 * alive. css && !memcg means that the groups should be
1083 * skipped and we should continue the tree walk.
1084 * last_visited css is safe to use because it is
1085 * protected by css_get and the tree walk is rcu safe.
1087 * We do not take a reference on the root of the tree walk
1088 * because we might race with the root removal when it would
1089 * be the only node in the iterated hierarchy and mem_cgroup_iter
1090 * would end up in an endless loop because it expects that at
1091 * least one valid node will be returned. Root cannot disappear
1092 * because caller of the iterator should hold it already so
1093 * skipping css reference should be safe.
1096 struct mem_cgroup
*memcg
= mem_cgroup_from_css(next_css
);
1098 if (next_css
== &root
->css
)
1101 if (css_tryget_online(next_css
)) {
1103 * Make sure the memcg is initialized:
1104 * mem_cgroup_css_online() orders the the
1105 * initialization against setting the flag.
1107 if (smp_load_acquire(&memcg
->initialized
))
1112 prev_css
= next_css
;
1119 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1122 * When a group in the hierarchy below root is destroyed, the
1123 * hierarchy iterator can no longer be trusted since it might
1124 * have pointed to the destroyed group. Invalidate it.
1126 atomic_inc(&root
->dead_count
);
1129 static struct mem_cgroup
*
1130 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1131 struct mem_cgroup
*root
,
1134 struct mem_cgroup
*position
= NULL
;
1136 * A cgroup destruction happens in two stages: offlining and
1137 * release. They are separated by a RCU grace period.
1139 * If the iterator is valid, we may still race with an
1140 * offlining. The RCU lock ensures the object won't be
1141 * released, tryget will fail if we lost the race.
1143 *sequence
= atomic_read(&root
->dead_count
);
1144 if (iter
->last_dead_count
== *sequence
) {
1146 position
= iter
->last_visited
;
1149 * We cannot take a reference to root because we might race
1150 * with root removal and returning NULL would end up in
1151 * an endless loop on the iterator user level when root
1152 * would be returned all the time.
1154 if (position
&& position
!= root
&&
1155 !css_tryget_online(&position
->css
))
1161 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1162 struct mem_cgroup
*last_visited
,
1163 struct mem_cgroup
*new_position
,
1164 struct mem_cgroup
*root
,
1167 /* root reference counting symmetric to mem_cgroup_iter_load */
1168 if (last_visited
&& last_visited
!= root
)
1169 css_put(&last_visited
->css
);
1171 * We store the sequence count from the time @last_visited was
1172 * loaded successfully instead of rereading it here so that we
1173 * don't lose destruction events in between. We could have
1174 * raced with the destruction of @new_position after all.
1176 iter
->last_visited
= new_position
;
1178 iter
->last_dead_count
= sequence
;
1182 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1183 * @root: hierarchy root
1184 * @prev: previously returned memcg, NULL on first invocation
1185 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1187 * Returns references to children of the hierarchy below @root, or
1188 * @root itself, or %NULL after a full round-trip.
1190 * Caller must pass the return value in @prev on subsequent
1191 * invocations for reference counting, or use mem_cgroup_iter_break()
1192 * to cancel a hierarchy walk before the round-trip is complete.
1194 * Reclaimers can specify a zone and a priority level in @reclaim to
1195 * divide up the memcgs in the hierarchy among all concurrent
1196 * reclaimers operating on the same zone and priority.
1198 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1199 struct mem_cgroup
*prev
,
1200 struct mem_cgroup_reclaim_cookie
*reclaim
)
1202 struct mem_cgroup
*memcg
= NULL
;
1203 struct mem_cgroup
*last_visited
= NULL
;
1205 if (mem_cgroup_disabled())
1209 root
= root_mem_cgroup
;
1211 if (prev
&& !reclaim
)
1212 last_visited
= prev
;
1214 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1222 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1223 int uninitialized_var(seq
);
1226 struct mem_cgroup_per_zone
*mz
;
1228 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1229 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1230 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1231 iter
->last_visited
= NULL
;
1235 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1238 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1241 mem_cgroup_iter_update(iter
, last_visited
, memcg
, root
,
1246 else if (!prev
&& memcg
)
1247 reclaim
->generation
= iter
->generation
;
1256 if (prev
&& prev
!= root
)
1257 css_put(&prev
->css
);
1263 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1264 * @root: hierarchy root
1265 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1267 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1268 struct mem_cgroup
*prev
)
1271 root
= root_mem_cgroup
;
1272 if (prev
&& prev
!= root
)
1273 css_put(&prev
->css
);
1277 * Iteration constructs for visiting all cgroups (under a tree). If
1278 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1279 * be used for reference counting.
1281 #define for_each_mem_cgroup_tree(iter, root) \
1282 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1284 iter = mem_cgroup_iter(root, iter, NULL))
1286 #define for_each_mem_cgroup(iter) \
1287 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1289 iter = mem_cgroup_iter(NULL, iter, NULL))
1291 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1293 struct mem_cgroup
*memcg
;
1296 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1297 if (unlikely(!memcg
))
1302 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1305 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1313 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1316 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1317 * @zone: zone of the wanted lruvec
1318 * @memcg: memcg of the wanted lruvec
1320 * Returns the lru list vector holding pages for the given @zone and
1321 * @mem. This can be the global zone lruvec, if the memory controller
1324 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1325 struct mem_cgroup
*memcg
)
1327 struct mem_cgroup_per_zone
*mz
;
1328 struct lruvec
*lruvec
;
1330 if (mem_cgroup_disabled()) {
1331 lruvec
= &zone
->lruvec
;
1335 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1336 lruvec
= &mz
->lruvec
;
1339 * Since a node can be onlined after the mem_cgroup was created,
1340 * we have to be prepared to initialize lruvec->zone here;
1341 * and if offlined then reonlined, we need to reinitialize it.
1343 if (unlikely(lruvec
->zone
!= zone
))
1344 lruvec
->zone
= zone
;
1349 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1351 * @zone: zone of the page
1353 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1355 struct mem_cgroup_per_zone
*mz
;
1356 struct mem_cgroup
*memcg
;
1357 struct page_cgroup
*pc
;
1358 struct lruvec
*lruvec
;
1360 if (mem_cgroup_disabled()) {
1361 lruvec
= &zone
->lruvec
;
1365 pc
= lookup_page_cgroup(page
);
1366 memcg
= pc
->mem_cgroup
;
1369 * Surreptitiously switch any uncharged offlist page to root:
1370 * an uncharged page off lru does nothing to secure
1371 * its former mem_cgroup from sudden removal.
1373 * Our caller holds lru_lock, and PageCgroupUsed is updated
1374 * under page_cgroup lock: between them, they make all uses
1375 * of pc->mem_cgroup safe.
1377 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1378 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1380 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1381 lruvec
= &mz
->lruvec
;
1384 * Since a node can be onlined after the mem_cgroup was created,
1385 * we have to be prepared to initialize lruvec->zone here;
1386 * and if offlined then reonlined, we need to reinitialize it.
1388 if (unlikely(lruvec
->zone
!= zone
))
1389 lruvec
->zone
= zone
;
1394 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1395 * @lruvec: mem_cgroup per zone lru vector
1396 * @lru: index of lru list the page is sitting on
1397 * @nr_pages: positive when adding or negative when removing
1399 * This function must be called when a page is added to or removed from an
1402 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1405 struct mem_cgroup_per_zone
*mz
;
1406 unsigned long *lru_size
;
1408 if (mem_cgroup_disabled())
1411 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1412 lru_size
= mz
->lru_size
+ lru
;
1413 *lru_size
+= nr_pages
;
1414 VM_BUG_ON((long)(*lru_size
) < 0);
1418 * Checks whether given mem is same or in the root_mem_cgroup's
1421 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1422 struct mem_cgroup
*memcg
)
1424 if (root_memcg
== memcg
)
1426 if (!root_memcg
->use_hierarchy
|| !memcg
)
1428 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1431 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1432 struct mem_cgroup
*memcg
)
1437 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1442 bool task_in_mem_cgroup(struct task_struct
*task
,
1443 const struct mem_cgroup
*memcg
)
1445 struct mem_cgroup
*curr
= NULL
;
1446 struct task_struct
*p
;
1449 p
= find_lock_task_mm(task
);
1451 curr
= get_mem_cgroup_from_mm(p
->mm
);
1455 * All threads may have already detached their mm's, but the oom
1456 * killer still needs to detect if they have already been oom
1457 * killed to prevent needlessly killing additional tasks.
1460 curr
= mem_cgroup_from_task(task
);
1462 css_get(&curr
->css
);
1466 * We should check use_hierarchy of "memcg" not "curr". Because checking
1467 * use_hierarchy of "curr" here make this function true if hierarchy is
1468 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1469 * hierarchy(even if use_hierarchy is disabled in "memcg").
1471 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1472 css_put(&curr
->css
);
1476 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1478 unsigned long inactive_ratio
;
1479 unsigned long inactive
;
1480 unsigned long active
;
1483 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1484 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1486 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1488 inactive_ratio
= int_sqrt(10 * gb
);
1492 return inactive
* inactive_ratio
< active
;
1495 #define mem_cgroup_from_res_counter(counter, member) \
1496 container_of(counter, struct mem_cgroup, member)
1499 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1500 * @memcg: the memory cgroup
1502 * Returns the maximum amount of memory @mem can be charged with, in
1505 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1507 unsigned long long margin
;
1509 margin
= res_counter_margin(&memcg
->res
);
1510 if (do_swap_account
)
1511 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1512 return margin
>> PAGE_SHIFT
;
1515 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1518 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1519 return vm_swappiness
;
1521 return memcg
->swappiness
;
1525 * memcg->moving_account is used for checking possibility that some thread is
1526 * calling move_account(). When a thread on CPU-A starts moving pages under
1527 * a memcg, other threads should check memcg->moving_account under
1528 * rcu_read_lock(), like this:
1532 * memcg->moving_account+1 if (memcg->mocing_account)
1534 * synchronize_rcu() update something.
1539 /* for quick checking without looking up memcg */
1540 atomic_t memcg_moving __read_mostly
;
1542 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1544 atomic_inc(&memcg_moving
);
1545 atomic_inc(&memcg
->moving_account
);
1549 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1552 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1553 * We check NULL in callee rather than caller.
1556 atomic_dec(&memcg_moving
);
1557 atomic_dec(&memcg
->moving_account
);
1562 * A routine for checking "mem" is under move_account() or not.
1564 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1565 * moving cgroups. This is for waiting at high-memory pressure
1568 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1570 struct mem_cgroup
*from
;
1571 struct mem_cgroup
*to
;
1574 * Unlike task_move routines, we access mc.to, mc.from not under
1575 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1577 spin_lock(&mc
.lock
);
1583 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1584 || mem_cgroup_same_or_subtree(memcg
, to
);
1586 spin_unlock(&mc
.lock
);
1590 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1592 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1593 if (mem_cgroup_under_move(memcg
)) {
1595 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1596 /* moving charge context might have finished. */
1599 finish_wait(&mc
.waitq
, &wait
);
1607 * Take this lock when
1608 * - a code tries to modify page's memcg while it's USED.
1609 * - a code tries to modify page state accounting in a memcg.
1611 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1612 unsigned long *flags
)
1614 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1617 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1618 unsigned long *flags
)
1620 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1623 #define K(x) ((x) << (PAGE_SHIFT-10))
1625 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1626 * @memcg: The memory cgroup that went over limit
1627 * @p: Task that is going to be killed
1629 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1632 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1634 /* oom_info_lock ensures that parallel ooms do not interleave */
1635 static DEFINE_MUTEX(oom_info_lock
);
1636 struct mem_cgroup
*iter
;
1642 mutex_lock(&oom_info_lock
);
1645 pr_info("Task in ");
1646 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1647 pr_info(" killed as a result of limit of ");
1648 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1653 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1654 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1655 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1656 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1657 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1658 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1659 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1660 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1661 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1662 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1663 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1664 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1666 for_each_mem_cgroup_tree(iter
, memcg
) {
1667 pr_info("Memory cgroup stats for ");
1668 pr_cont_cgroup_path(iter
->css
.cgroup
);
1671 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1672 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1674 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1675 K(mem_cgroup_read_stat(iter
, i
)));
1678 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1679 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1680 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1684 mutex_unlock(&oom_info_lock
);
1688 * This function returns the number of memcg under hierarchy tree. Returns
1689 * 1(self count) if no children.
1691 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1694 struct mem_cgroup
*iter
;
1696 for_each_mem_cgroup_tree(iter
, memcg
)
1702 * Return the memory (and swap, if configured) limit for a memcg.
1704 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1708 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1711 * Do not consider swap space if we cannot swap due to swappiness
1713 if (mem_cgroup_swappiness(memcg
)) {
1716 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1717 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1720 * If memsw is finite and limits the amount of swap space
1721 * available to this memcg, return that limit.
1723 limit
= min(limit
, memsw
);
1729 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1732 struct mem_cgroup
*iter
;
1733 unsigned long chosen_points
= 0;
1734 unsigned long totalpages
;
1735 unsigned int points
= 0;
1736 struct task_struct
*chosen
= NULL
;
1739 * If current has a pending SIGKILL or is exiting, then automatically
1740 * select it. The goal is to allow it to allocate so that it may
1741 * quickly exit and free its memory.
1743 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1744 set_thread_flag(TIF_MEMDIE
);
1748 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1749 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1750 for_each_mem_cgroup_tree(iter
, memcg
) {
1751 struct css_task_iter it
;
1752 struct task_struct
*task
;
1754 css_task_iter_start(&iter
->css
, &it
);
1755 while ((task
= css_task_iter_next(&it
))) {
1756 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1758 case OOM_SCAN_SELECT
:
1760 put_task_struct(chosen
);
1762 chosen_points
= ULONG_MAX
;
1763 get_task_struct(chosen
);
1765 case OOM_SCAN_CONTINUE
:
1767 case OOM_SCAN_ABORT
:
1768 css_task_iter_end(&it
);
1769 mem_cgroup_iter_break(memcg
, iter
);
1771 put_task_struct(chosen
);
1776 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1777 if (!points
|| points
< chosen_points
)
1779 /* Prefer thread group leaders for display purposes */
1780 if (points
== chosen_points
&&
1781 thread_group_leader(chosen
))
1785 put_task_struct(chosen
);
1787 chosen_points
= points
;
1788 get_task_struct(chosen
);
1790 css_task_iter_end(&it
);
1795 points
= chosen_points
* 1000 / totalpages
;
1796 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1797 NULL
, "Memory cgroup out of memory");
1801 * test_mem_cgroup_node_reclaimable
1802 * @memcg: the target memcg
1803 * @nid: the node ID to be checked.
1804 * @noswap : specify true here if the user wants flle only information.
1806 * This function returns whether the specified memcg contains any
1807 * reclaimable pages on a node. Returns true if there are any reclaimable
1808 * pages in the node.
1810 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1811 int nid
, bool noswap
)
1813 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1815 if (noswap
|| !total_swap_pages
)
1817 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1822 #if MAX_NUMNODES > 1
1825 * Always updating the nodemask is not very good - even if we have an empty
1826 * list or the wrong list here, we can start from some node and traverse all
1827 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1830 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1834 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1835 * pagein/pageout changes since the last update.
1837 if (!atomic_read(&memcg
->numainfo_events
))
1839 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1842 /* make a nodemask where this memcg uses memory from */
1843 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1845 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1847 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1848 node_clear(nid
, memcg
->scan_nodes
);
1851 atomic_set(&memcg
->numainfo_events
, 0);
1852 atomic_set(&memcg
->numainfo_updating
, 0);
1856 * Selecting a node where we start reclaim from. Because what we need is just
1857 * reducing usage counter, start from anywhere is O,K. Considering
1858 * memory reclaim from current node, there are pros. and cons.
1860 * Freeing memory from current node means freeing memory from a node which
1861 * we'll use or we've used. So, it may make LRU bad. And if several threads
1862 * hit limits, it will see a contention on a node. But freeing from remote
1863 * node means more costs for memory reclaim because of memory latency.
1865 * Now, we use round-robin. Better algorithm is welcomed.
1867 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1871 mem_cgroup_may_update_nodemask(memcg
);
1872 node
= memcg
->last_scanned_node
;
1874 node
= next_node(node
, memcg
->scan_nodes
);
1875 if (node
== MAX_NUMNODES
)
1876 node
= first_node(memcg
->scan_nodes
);
1878 * We call this when we hit limit, not when pages are added to LRU.
1879 * No LRU may hold pages because all pages are UNEVICTABLE or
1880 * memcg is too small and all pages are not on LRU. In that case,
1881 * we use curret node.
1883 if (unlikely(node
== MAX_NUMNODES
))
1884 node
= numa_node_id();
1886 memcg
->last_scanned_node
= node
;
1891 * Check all nodes whether it contains reclaimable pages or not.
1892 * For quick scan, we make use of scan_nodes. This will allow us to skip
1893 * unused nodes. But scan_nodes is lazily updated and may not cotain
1894 * enough new information. We need to do double check.
1896 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1901 * quick check...making use of scan_node.
1902 * We can skip unused nodes.
1904 if (!nodes_empty(memcg
->scan_nodes
)) {
1905 for (nid
= first_node(memcg
->scan_nodes
);
1907 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1909 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1914 * Check rest of nodes.
1916 for_each_node_state(nid
, N_MEMORY
) {
1917 if (node_isset(nid
, memcg
->scan_nodes
))
1919 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1926 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1931 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1933 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1937 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1940 unsigned long *total_scanned
)
1942 struct mem_cgroup
*victim
= NULL
;
1945 unsigned long excess
;
1946 unsigned long nr_scanned
;
1947 struct mem_cgroup_reclaim_cookie reclaim
= {
1952 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
1955 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1960 * If we have not been able to reclaim
1961 * anything, it might because there are
1962 * no reclaimable pages under this hierarchy
1967 * We want to do more targeted reclaim.
1968 * excess >> 2 is not to excessive so as to
1969 * reclaim too much, nor too less that we keep
1970 * coming back to reclaim from this cgroup
1972 if (total
>= (excess
>> 2) ||
1973 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1978 if (!mem_cgroup_reclaimable(victim
, false))
1980 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1982 *total_scanned
+= nr_scanned
;
1983 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
1986 mem_cgroup_iter_break(root_memcg
, victim
);
1990 #ifdef CONFIG_LOCKDEP
1991 static struct lockdep_map memcg_oom_lock_dep_map
= {
1992 .name
= "memcg_oom_lock",
1996 static DEFINE_SPINLOCK(memcg_oom_lock
);
1999 * Check OOM-Killer is already running under our hierarchy.
2000 * If someone is running, return false.
2002 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2004 struct mem_cgroup
*iter
, *failed
= NULL
;
2006 spin_lock(&memcg_oom_lock
);
2008 for_each_mem_cgroup_tree(iter
, memcg
) {
2009 if (iter
->oom_lock
) {
2011 * this subtree of our hierarchy is already locked
2012 * so we cannot give a lock.
2015 mem_cgroup_iter_break(memcg
, iter
);
2018 iter
->oom_lock
= true;
2023 * OK, we failed to lock the whole subtree so we have
2024 * to clean up what we set up to the failing subtree
2026 for_each_mem_cgroup_tree(iter
, memcg
) {
2027 if (iter
== failed
) {
2028 mem_cgroup_iter_break(memcg
, iter
);
2031 iter
->oom_lock
= false;
2034 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2036 spin_unlock(&memcg_oom_lock
);
2041 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2043 struct mem_cgroup
*iter
;
2045 spin_lock(&memcg_oom_lock
);
2046 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2047 for_each_mem_cgroup_tree(iter
, memcg
)
2048 iter
->oom_lock
= false;
2049 spin_unlock(&memcg_oom_lock
);
2052 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2054 struct mem_cgroup
*iter
;
2056 for_each_mem_cgroup_tree(iter
, memcg
)
2057 atomic_inc(&iter
->under_oom
);
2060 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2062 struct mem_cgroup
*iter
;
2065 * When a new child is created while the hierarchy is under oom,
2066 * mem_cgroup_oom_lock() may not be called. We have to use
2067 * atomic_add_unless() here.
2069 for_each_mem_cgroup_tree(iter
, memcg
)
2070 atomic_add_unless(&iter
->under_oom
, -1, 0);
2073 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2075 struct oom_wait_info
{
2076 struct mem_cgroup
*memcg
;
2080 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2081 unsigned mode
, int sync
, void *arg
)
2083 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2084 struct mem_cgroup
*oom_wait_memcg
;
2085 struct oom_wait_info
*oom_wait_info
;
2087 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2088 oom_wait_memcg
= oom_wait_info
->memcg
;
2091 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2092 * Then we can use css_is_ancestor without taking care of RCU.
2094 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2095 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2097 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2100 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2102 atomic_inc(&memcg
->oom_wakeups
);
2103 /* for filtering, pass "memcg" as argument. */
2104 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2107 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2109 if (memcg
&& atomic_read(&memcg
->under_oom
))
2110 memcg_wakeup_oom(memcg
);
2113 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2115 if (!current
->memcg_oom
.may_oom
)
2118 * We are in the middle of the charge context here, so we
2119 * don't want to block when potentially sitting on a callstack
2120 * that holds all kinds of filesystem and mm locks.
2122 * Also, the caller may handle a failed allocation gracefully
2123 * (like optional page cache readahead) and so an OOM killer
2124 * invocation might not even be necessary.
2126 * That's why we don't do anything here except remember the
2127 * OOM context and then deal with it at the end of the page
2128 * fault when the stack is unwound, the locks are released,
2129 * and when we know whether the fault was overall successful.
2131 css_get(&memcg
->css
);
2132 current
->memcg_oom
.memcg
= memcg
;
2133 current
->memcg_oom
.gfp_mask
= mask
;
2134 current
->memcg_oom
.order
= order
;
2138 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2139 * @handle: actually kill/wait or just clean up the OOM state
2141 * This has to be called at the end of a page fault if the memcg OOM
2142 * handler was enabled.
2144 * Memcg supports userspace OOM handling where failed allocations must
2145 * sleep on a waitqueue until the userspace task resolves the
2146 * situation. Sleeping directly in the charge context with all kinds
2147 * of locks held is not a good idea, instead we remember an OOM state
2148 * in the task and mem_cgroup_oom_synchronize() has to be called at
2149 * the end of the page fault to complete the OOM handling.
2151 * Returns %true if an ongoing memcg OOM situation was detected and
2152 * completed, %false otherwise.
2154 bool mem_cgroup_oom_synchronize(bool handle
)
2156 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2157 struct oom_wait_info owait
;
2160 /* OOM is global, do not handle */
2167 owait
.memcg
= memcg
;
2168 owait
.wait
.flags
= 0;
2169 owait
.wait
.func
= memcg_oom_wake_function
;
2170 owait
.wait
.private = current
;
2171 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2173 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2174 mem_cgroup_mark_under_oom(memcg
);
2176 locked
= mem_cgroup_oom_trylock(memcg
);
2179 mem_cgroup_oom_notify(memcg
);
2181 if (locked
&& !memcg
->oom_kill_disable
) {
2182 mem_cgroup_unmark_under_oom(memcg
);
2183 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2184 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2185 current
->memcg_oom
.order
);
2188 mem_cgroup_unmark_under_oom(memcg
);
2189 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2193 mem_cgroup_oom_unlock(memcg
);
2195 * There is no guarantee that an OOM-lock contender
2196 * sees the wakeups triggered by the OOM kill
2197 * uncharges. Wake any sleepers explicitely.
2199 memcg_oom_recover(memcg
);
2202 current
->memcg_oom
.memcg
= NULL
;
2203 css_put(&memcg
->css
);
2208 * Used to update mapped file or writeback or other statistics.
2210 * Notes: Race condition
2212 * Charging occurs during page instantiation, while the page is
2213 * unmapped and locked in page migration, or while the page table is
2214 * locked in THP migration. No race is possible.
2216 * Uncharge happens to pages with zero references, no race possible.
2218 * Charge moving between groups is protected by checking mm->moving
2219 * account and taking the move_lock in the slowpath.
2222 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2223 bool *locked
, unsigned long *flags
)
2225 struct mem_cgroup
*memcg
;
2226 struct page_cgroup
*pc
;
2228 pc
= lookup_page_cgroup(page
);
2230 memcg
= pc
->mem_cgroup
;
2231 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2234 * If this memory cgroup is not under account moving, we don't
2235 * need to take move_lock_mem_cgroup(). Because we already hold
2236 * rcu_read_lock(), any calls to move_account will be delayed until
2237 * rcu_read_unlock().
2239 VM_BUG_ON(!rcu_read_lock_held());
2240 if (atomic_read(&memcg
->moving_account
) <= 0)
2243 move_lock_mem_cgroup(memcg
, flags
);
2244 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2245 move_unlock_mem_cgroup(memcg
, flags
);
2251 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2253 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2256 * It's guaranteed that pc->mem_cgroup never changes while
2257 * lock is held because a routine modifies pc->mem_cgroup
2258 * should take move_lock_mem_cgroup().
2260 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2263 void mem_cgroup_update_page_stat(struct page
*page
,
2264 enum mem_cgroup_stat_index idx
, int val
)
2266 struct mem_cgroup
*memcg
;
2267 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2268 unsigned long uninitialized_var(flags
);
2270 if (mem_cgroup_disabled())
2273 VM_BUG_ON(!rcu_read_lock_held());
2274 memcg
= pc
->mem_cgroup
;
2275 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2278 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2282 * size of first charge trial. "32" comes from vmscan.c's magic value.
2283 * TODO: maybe necessary to use big numbers in big irons.
2285 #define CHARGE_BATCH 32U
2286 struct memcg_stock_pcp
{
2287 struct mem_cgroup
*cached
; /* this never be root cgroup */
2288 unsigned int nr_pages
;
2289 struct work_struct work
;
2290 unsigned long flags
;
2291 #define FLUSHING_CACHED_CHARGE 0
2293 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2294 static DEFINE_MUTEX(percpu_charge_mutex
);
2297 * consume_stock: Try to consume stocked charge on this cpu.
2298 * @memcg: memcg to consume from.
2299 * @nr_pages: how many pages to charge.
2301 * The charges will only happen if @memcg matches the current cpu's memcg
2302 * stock, and at least @nr_pages are available in that stock. Failure to
2303 * service an allocation will refill the stock.
2305 * returns true if successful, false otherwise.
2307 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2309 struct memcg_stock_pcp
*stock
;
2312 if (nr_pages
> CHARGE_BATCH
)
2315 stock
= &get_cpu_var(memcg_stock
);
2316 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2317 stock
->nr_pages
-= nr_pages
;
2318 else /* need to call res_counter_charge */
2320 put_cpu_var(memcg_stock
);
2325 * Returns stocks cached in percpu to res_counter and reset cached information.
2327 static void drain_stock(struct memcg_stock_pcp
*stock
)
2329 struct mem_cgroup
*old
= stock
->cached
;
2331 if (stock
->nr_pages
) {
2332 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2334 res_counter_uncharge(&old
->res
, bytes
);
2335 if (do_swap_account
)
2336 res_counter_uncharge(&old
->memsw
, bytes
);
2337 stock
->nr_pages
= 0;
2339 stock
->cached
= NULL
;
2343 * This must be called under preempt disabled or must be called by
2344 * a thread which is pinned to local cpu.
2346 static void drain_local_stock(struct work_struct
*dummy
)
2348 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2350 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2353 static void __init
memcg_stock_init(void)
2357 for_each_possible_cpu(cpu
) {
2358 struct memcg_stock_pcp
*stock
=
2359 &per_cpu(memcg_stock
, cpu
);
2360 INIT_WORK(&stock
->work
, drain_local_stock
);
2365 * Cache charges(val) which is from res_counter, to local per_cpu area.
2366 * This will be consumed by consume_stock() function, later.
2368 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2370 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2372 if (stock
->cached
!= memcg
) { /* reset if necessary */
2374 stock
->cached
= memcg
;
2376 stock
->nr_pages
+= nr_pages
;
2377 put_cpu_var(memcg_stock
);
2381 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2382 * of the hierarchy under it. sync flag says whether we should block
2383 * until the work is done.
2385 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2389 /* Notify other cpus that system-wide "drain" is running */
2392 for_each_online_cpu(cpu
) {
2393 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2394 struct mem_cgroup
*memcg
;
2396 memcg
= stock
->cached
;
2397 if (!memcg
|| !stock
->nr_pages
)
2399 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2401 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2403 drain_local_stock(&stock
->work
);
2405 schedule_work_on(cpu
, &stock
->work
);
2413 for_each_online_cpu(cpu
) {
2414 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2415 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2416 flush_work(&stock
->work
);
2423 * Tries to drain stocked charges in other cpus. This function is asynchronous
2424 * and just put a work per cpu for draining localy on each cpu. Caller can
2425 * expects some charges will be back to res_counter later but cannot wait for
2428 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2431 * If someone calls draining, avoid adding more kworker runs.
2433 if (!mutex_trylock(&percpu_charge_mutex
))
2435 drain_all_stock(root_memcg
, false);
2436 mutex_unlock(&percpu_charge_mutex
);
2439 /* This is a synchronous drain interface. */
2440 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2442 /* called when force_empty is called */
2443 mutex_lock(&percpu_charge_mutex
);
2444 drain_all_stock(root_memcg
, true);
2445 mutex_unlock(&percpu_charge_mutex
);
2449 * This function drains percpu counter value from DEAD cpu and
2450 * move it to local cpu. Note that this function can be preempted.
2452 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2456 spin_lock(&memcg
->pcp_counter_lock
);
2457 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2458 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2460 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2461 memcg
->nocpu_base
.count
[i
] += x
;
2463 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2464 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2466 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2467 memcg
->nocpu_base
.events
[i
] += x
;
2469 spin_unlock(&memcg
->pcp_counter_lock
);
2472 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2473 unsigned long action
,
2476 int cpu
= (unsigned long)hcpu
;
2477 struct memcg_stock_pcp
*stock
;
2478 struct mem_cgroup
*iter
;
2480 if (action
== CPU_ONLINE
)
2483 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2486 for_each_mem_cgroup(iter
)
2487 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2489 stock
= &per_cpu(memcg_stock
, cpu
);
2494 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2495 unsigned int nr_pages
)
2497 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2498 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2499 struct mem_cgroup
*mem_over_limit
;
2500 struct res_counter
*fail_res
;
2501 unsigned long nr_reclaimed
;
2502 unsigned long long size
;
2503 bool may_swap
= true;
2504 bool drained
= false;
2507 if (mem_cgroup_is_root(memcg
))
2510 if (consume_stock(memcg
, nr_pages
))
2513 size
= batch
* PAGE_SIZE
;
2514 if (!do_swap_account
||
2515 !res_counter_charge(&memcg
->memsw
, size
, &fail_res
)) {
2516 if (!res_counter_charge(&memcg
->res
, size
, &fail_res
))
2518 if (do_swap_account
)
2519 res_counter_uncharge(&memcg
->memsw
, size
);
2520 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2522 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2526 if (batch
> nr_pages
) {
2532 * Unlike in global OOM situations, memcg is not in a physical
2533 * memory shortage. Allow dying and OOM-killed tasks to
2534 * bypass the last charges so that they can exit quickly and
2535 * free their memory.
2537 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2538 fatal_signal_pending(current
) ||
2539 current
->flags
& PF_EXITING
))
2542 if (unlikely(task_in_memcg_oom(current
)))
2545 if (!(gfp_mask
& __GFP_WAIT
))
2548 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2549 gfp_mask
, may_swap
);
2551 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2555 drain_all_stock_async(mem_over_limit
);
2560 if (gfp_mask
& __GFP_NORETRY
)
2563 * Even though the limit is exceeded at this point, reclaim
2564 * may have been able to free some pages. Retry the charge
2565 * before killing the task.
2567 * Only for regular pages, though: huge pages are rather
2568 * unlikely to succeed so close to the limit, and we fall back
2569 * to regular pages anyway in case of failure.
2571 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2574 * At task move, charge accounts can be doubly counted. So, it's
2575 * better to wait until the end of task_move if something is going on.
2577 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2583 if (gfp_mask
& __GFP_NOFAIL
)
2586 if (fatal_signal_pending(current
))
2589 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2591 if (!(gfp_mask
& __GFP_NOFAIL
))
2597 if (batch
> nr_pages
)
2598 refill_stock(memcg
, batch
- nr_pages
);
2603 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2605 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2607 if (mem_cgroup_is_root(memcg
))
2610 res_counter_uncharge(&memcg
->res
, bytes
);
2611 if (do_swap_account
)
2612 res_counter_uncharge(&memcg
->memsw
, bytes
);
2616 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2617 * This is useful when moving usage to parent cgroup.
2619 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2620 unsigned int nr_pages
)
2622 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2624 if (mem_cgroup_is_root(memcg
))
2627 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2628 if (do_swap_account
)
2629 res_counter_uncharge_until(&memcg
->memsw
,
2630 memcg
->memsw
.parent
, bytes
);
2634 * A helper function to get mem_cgroup from ID. must be called under
2635 * rcu_read_lock(). The caller is responsible for calling
2636 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2637 * refcnt from swap can be called against removed memcg.)
2639 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2641 /* ID 0 is unused ID */
2644 return mem_cgroup_from_id(id
);
2648 * try_get_mem_cgroup_from_page - look up page's memcg association
2651 * Look up, get a css reference, and return the memcg that owns @page.
2653 * The page must be locked to prevent racing with swap-in and page
2654 * cache charges. If coming from an unlocked page table, the caller
2655 * must ensure the page is on the LRU or this can race with charging.
2657 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2659 struct mem_cgroup
*memcg
= NULL
;
2660 struct page_cgroup
*pc
;
2664 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2666 pc
= lookup_page_cgroup(page
);
2667 if (PageCgroupUsed(pc
)) {
2668 memcg
= pc
->mem_cgroup
;
2669 if (memcg
&& !css_tryget_online(&memcg
->css
))
2671 } else if (PageSwapCache(page
)) {
2672 ent
.val
= page_private(page
);
2673 id
= lookup_swap_cgroup_id(ent
);
2675 memcg
= mem_cgroup_lookup(id
);
2676 if (memcg
&& !css_tryget_online(&memcg
->css
))
2683 static void lock_page_lru(struct page
*page
, int *isolated
)
2685 struct zone
*zone
= page_zone(page
);
2687 spin_lock_irq(&zone
->lru_lock
);
2688 if (PageLRU(page
)) {
2689 struct lruvec
*lruvec
;
2691 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2693 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2699 static void unlock_page_lru(struct page
*page
, int isolated
)
2701 struct zone
*zone
= page_zone(page
);
2704 struct lruvec
*lruvec
;
2706 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2707 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2709 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2711 spin_unlock_irq(&zone
->lru_lock
);
2714 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2717 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2720 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2722 * we don't need page_cgroup_lock about tail pages, becase they are not
2723 * accessed by any other context at this point.
2727 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2728 * may already be on some other mem_cgroup's LRU. Take care of it.
2731 lock_page_lru(page
, &isolated
);
2734 * Nobody should be changing or seriously looking at
2735 * pc->mem_cgroup and pc->flags at this point:
2737 * - the page is uncharged
2739 * - the page is off-LRU
2741 * - an anonymous fault has exclusive page access, except for
2742 * a locked page table
2744 * - a page cache insertion, a swapin fault, or a migration
2745 * have the page locked
2747 pc
->mem_cgroup
= memcg
;
2748 pc
->flags
= PCG_USED
| PCG_MEM
| (do_swap_account
? PCG_MEMSW
: 0);
2751 unlock_page_lru(page
, isolated
);
2754 static DEFINE_MUTEX(set_limit_mutex
);
2756 #ifdef CONFIG_MEMCG_KMEM
2758 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2759 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2761 static DEFINE_MUTEX(memcg_slab_mutex
);
2763 static DEFINE_MUTEX(activate_kmem_mutex
);
2766 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2767 * in the memcg_cache_params struct.
2769 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2771 struct kmem_cache
*cachep
;
2773 VM_BUG_ON(p
->is_root_cache
);
2774 cachep
= p
->root_cache
;
2775 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2778 #ifdef CONFIG_SLABINFO
2779 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
2781 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
2782 struct memcg_cache_params
*params
;
2784 if (!memcg_kmem_is_active(memcg
))
2787 print_slabinfo_header(m
);
2789 mutex_lock(&memcg_slab_mutex
);
2790 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2791 cache_show(memcg_params_to_cache(params
), m
);
2792 mutex_unlock(&memcg_slab_mutex
);
2798 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2800 struct res_counter
*fail_res
;
2803 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2807 ret
= try_charge(memcg
, gfp
, size
>> PAGE_SHIFT
);
2808 if (ret
== -EINTR
) {
2810 * try_charge() chose to bypass to root due to OOM kill or
2811 * fatal signal. Since our only options are to either fail
2812 * the allocation or charge it to this cgroup, do it as a
2813 * temporary condition. But we can't fail. From a kmem/slab
2814 * perspective, the cache has already been selected, by
2815 * mem_cgroup_kmem_get_cache(), so it is too late to change
2818 * This condition will only trigger if the task entered
2819 * memcg_charge_kmem in a sane state, but was OOM-killed
2820 * during try_charge() above. Tasks that were already dying
2821 * when the allocation triggers should have been already
2822 * directed to the root cgroup in memcontrol.h
2824 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2825 if (do_swap_account
)
2826 res_counter_charge_nofail(&memcg
->memsw
, size
,
2830 res_counter_uncharge(&memcg
->kmem
, size
);
2835 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2837 res_counter_uncharge(&memcg
->res
, size
);
2838 if (do_swap_account
)
2839 res_counter_uncharge(&memcg
->memsw
, size
);
2842 if (res_counter_uncharge(&memcg
->kmem
, size
))
2846 * Releases a reference taken in kmem_cgroup_css_offline in case
2847 * this last uncharge is racing with the offlining code or it is
2848 * outliving the memcg existence.
2850 * The memory barrier imposed by test&clear is paired with the
2851 * explicit one in memcg_kmem_mark_dead().
2853 if (memcg_kmem_test_and_clear_dead(memcg
))
2854 css_put(&memcg
->css
);
2858 * helper for acessing a memcg's index. It will be used as an index in the
2859 * child cache array in kmem_cache, and also to derive its name. This function
2860 * will return -1 when this is not a kmem-limited memcg.
2862 int memcg_cache_id(struct mem_cgroup
*memcg
)
2864 return memcg
? memcg
->kmemcg_id
: -1;
2867 static int memcg_alloc_cache_id(void)
2872 id
= ida_simple_get(&kmem_limited_groups
,
2873 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2877 if (id
< memcg_limited_groups_array_size
)
2881 * There's no space for the new id in memcg_caches arrays,
2882 * so we have to grow them.
2885 size
= 2 * (id
+ 1);
2886 if (size
< MEMCG_CACHES_MIN_SIZE
)
2887 size
= MEMCG_CACHES_MIN_SIZE
;
2888 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2889 size
= MEMCG_CACHES_MAX_SIZE
;
2891 mutex_lock(&memcg_slab_mutex
);
2892 err
= memcg_update_all_caches(size
);
2893 mutex_unlock(&memcg_slab_mutex
);
2896 ida_simple_remove(&kmem_limited_groups
, id
);
2902 static void memcg_free_cache_id(int id
)
2904 ida_simple_remove(&kmem_limited_groups
, id
);
2908 * We should update the current array size iff all caches updates succeed. This
2909 * can only be done from the slab side. The slab mutex needs to be held when
2912 void memcg_update_array_size(int num
)
2914 memcg_limited_groups_array_size
= num
;
2917 static void memcg_register_cache(struct mem_cgroup
*memcg
,
2918 struct kmem_cache
*root_cache
)
2920 static char memcg_name_buf
[NAME_MAX
+ 1]; /* protected by
2922 struct kmem_cache
*cachep
;
2925 lockdep_assert_held(&memcg_slab_mutex
);
2927 id
= memcg_cache_id(memcg
);
2930 * Since per-memcg caches are created asynchronously on first
2931 * allocation (see memcg_kmem_get_cache()), several threads can try to
2932 * create the same cache, but only one of them may succeed.
2934 if (cache_from_memcg_idx(root_cache
, id
))
2937 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
2938 cachep
= memcg_create_kmem_cache(memcg
, root_cache
, memcg_name_buf
);
2940 * If we could not create a memcg cache, do not complain, because
2941 * that's not critical at all as we can always proceed with the root
2947 css_get(&memcg
->css
);
2948 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2951 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2952 * barrier here to ensure nobody will see the kmem_cache partially
2957 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
2958 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
2961 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
2963 struct kmem_cache
*root_cache
;
2964 struct mem_cgroup
*memcg
;
2967 lockdep_assert_held(&memcg_slab_mutex
);
2969 BUG_ON(is_root_cache(cachep
));
2971 root_cache
= cachep
->memcg_params
->root_cache
;
2972 memcg
= cachep
->memcg_params
->memcg
;
2973 id
= memcg_cache_id(memcg
);
2975 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
2976 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
2978 list_del(&cachep
->memcg_params
->list
);
2980 kmem_cache_destroy(cachep
);
2982 /* drop the reference taken in memcg_register_cache */
2983 css_put(&memcg
->css
);
2987 * During the creation a new cache, we need to disable our accounting mechanism
2988 * altogether. This is true even if we are not creating, but rather just
2989 * enqueing new caches to be created.
2991 * This is because that process will trigger allocations; some visible, like
2992 * explicit kmallocs to auxiliary data structures, name strings and internal
2993 * cache structures; some well concealed, like INIT_WORK() that can allocate
2994 * objects during debug.
2996 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2997 * to it. This may not be a bounded recursion: since the first cache creation
2998 * failed to complete (waiting on the allocation), we'll just try to create the
2999 * cache again, failing at the same point.
3001 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3002 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3003 * inside the following two functions.
3005 static inline void memcg_stop_kmem_account(void)
3007 VM_BUG_ON(!current
->mm
);
3008 current
->memcg_kmem_skip_account
++;
3011 static inline void memcg_resume_kmem_account(void)
3013 VM_BUG_ON(!current
->mm
);
3014 current
->memcg_kmem_skip_account
--;
3017 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
3019 struct kmem_cache
*c
;
3022 mutex_lock(&memcg_slab_mutex
);
3023 for_each_memcg_cache_index(i
) {
3024 c
= cache_from_memcg_idx(s
, i
);
3028 memcg_unregister_cache(c
);
3030 if (cache_from_memcg_idx(s
, i
))
3033 mutex_unlock(&memcg_slab_mutex
);
3037 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3039 struct kmem_cache
*cachep
;
3040 struct memcg_cache_params
*params
, *tmp
;
3042 if (!memcg_kmem_is_active(memcg
))
3045 mutex_lock(&memcg_slab_mutex
);
3046 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
3047 cachep
= memcg_params_to_cache(params
);
3048 kmem_cache_shrink(cachep
);
3049 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3050 memcg_unregister_cache(cachep
);
3052 mutex_unlock(&memcg_slab_mutex
);
3055 struct memcg_register_cache_work
{
3056 struct mem_cgroup
*memcg
;
3057 struct kmem_cache
*cachep
;
3058 struct work_struct work
;
3061 static void memcg_register_cache_func(struct work_struct
*w
)
3063 struct memcg_register_cache_work
*cw
=
3064 container_of(w
, struct memcg_register_cache_work
, work
);
3065 struct mem_cgroup
*memcg
= cw
->memcg
;
3066 struct kmem_cache
*cachep
= cw
->cachep
;
3068 mutex_lock(&memcg_slab_mutex
);
3069 memcg_register_cache(memcg
, cachep
);
3070 mutex_unlock(&memcg_slab_mutex
);
3072 css_put(&memcg
->css
);
3077 * Enqueue the creation of a per-memcg kmem_cache.
3079 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3080 struct kmem_cache
*cachep
)
3082 struct memcg_register_cache_work
*cw
;
3084 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
3086 css_put(&memcg
->css
);
3091 cw
->cachep
= cachep
;
3093 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
3094 schedule_work(&cw
->work
);
3097 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3098 struct kmem_cache
*cachep
)
3101 * We need to stop accounting when we kmalloc, because if the
3102 * corresponding kmalloc cache is not yet created, the first allocation
3103 * in __memcg_schedule_register_cache will recurse.
3105 * However, it is better to enclose the whole function. Depending on
3106 * the debugging options enabled, INIT_WORK(), for instance, can
3107 * trigger an allocation. This too, will make us recurse. Because at
3108 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3109 * the safest choice is to do it like this, wrapping the whole function.
3111 memcg_stop_kmem_account();
3112 __memcg_schedule_register_cache(memcg
, cachep
);
3113 memcg_resume_kmem_account();
3116 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
3120 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
,
3121 PAGE_SIZE
<< order
);
3123 atomic_add(1 << order
, &cachep
->memcg_params
->nr_pages
);
3127 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
3129 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, PAGE_SIZE
<< order
);
3130 atomic_sub(1 << order
, &cachep
->memcg_params
->nr_pages
);
3134 * Return the kmem_cache we're supposed to use for a slab allocation.
3135 * We try to use the current memcg's version of the cache.
3137 * If the cache does not exist yet, if we are the first user of it,
3138 * we either create it immediately, if possible, or create it asynchronously
3140 * In the latter case, we will let the current allocation go through with
3141 * the original cache.
3143 * Can't be called in interrupt context or from kernel threads.
3144 * This function needs to be called with rcu_read_lock() held.
3146 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3149 struct mem_cgroup
*memcg
;
3150 struct kmem_cache
*memcg_cachep
;
3152 VM_BUG_ON(!cachep
->memcg_params
);
3153 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3155 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3159 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3161 if (!memcg_kmem_is_active(memcg
))
3164 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3165 if (likely(memcg_cachep
)) {
3166 cachep
= memcg_cachep
;
3170 /* The corresponding put will be done in the workqueue. */
3171 if (!css_tryget_online(&memcg
->css
))
3176 * If we are in a safe context (can wait, and not in interrupt
3177 * context), we could be be predictable and return right away.
3178 * This would guarantee that the allocation being performed
3179 * already belongs in the new cache.
3181 * However, there are some clashes that can arrive from locking.
3182 * For instance, because we acquire the slab_mutex while doing
3183 * memcg_create_kmem_cache, this means no further allocation
3184 * could happen with the slab_mutex held. So it's better to
3187 memcg_schedule_register_cache(memcg
, cachep
);
3195 * We need to verify if the allocation against current->mm->owner's memcg is
3196 * possible for the given order. But the page is not allocated yet, so we'll
3197 * need a further commit step to do the final arrangements.
3199 * It is possible for the task to switch cgroups in this mean time, so at
3200 * commit time, we can't rely on task conversion any longer. We'll then use
3201 * the handle argument to return to the caller which cgroup we should commit
3202 * against. We could also return the memcg directly and avoid the pointer
3203 * passing, but a boolean return value gives better semantics considering
3204 * the compiled-out case as well.
3206 * Returning true means the allocation is possible.
3209 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3211 struct mem_cgroup
*memcg
;
3217 * Disabling accounting is only relevant for some specific memcg
3218 * internal allocations. Therefore we would initially not have such
3219 * check here, since direct calls to the page allocator that are
3220 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3221 * outside memcg core. We are mostly concerned with cache allocations,
3222 * and by having this test at memcg_kmem_get_cache, we are already able
3223 * to relay the allocation to the root cache and bypass the memcg cache
3226 * There is one exception, though: the SLUB allocator does not create
3227 * large order caches, but rather service large kmallocs directly from
3228 * the page allocator. Therefore, the following sequence when backed by
3229 * the SLUB allocator:
3231 * memcg_stop_kmem_account();
3232 * kmalloc(<large_number>)
3233 * memcg_resume_kmem_account();
3235 * would effectively ignore the fact that we should skip accounting,
3236 * since it will drive us directly to this function without passing
3237 * through the cache selector memcg_kmem_get_cache. Such large
3238 * allocations are extremely rare but can happen, for instance, for the
3239 * cache arrays. We bring this test here.
3241 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3244 memcg
= get_mem_cgroup_from_mm(current
->mm
);
3246 if (!memcg_kmem_is_active(memcg
)) {
3247 css_put(&memcg
->css
);
3251 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3255 css_put(&memcg
->css
);
3259 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3262 struct page_cgroup
*pc
;
3264 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3266 /* The page allocation failed. Revert */
3268 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3272 * The page is freshly allocated and not visible to any
3273 * outside callers yet. Set up pc non-atomically.
3275 pc
= lookup_page_cgroup(page
);
3276 pc
->mem_cgroup
= memcg
;
3277 pc
->flags
= PCG_USED
;
3280 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3282 struct mem_cgroup
*memcg
= NULL
;
3283 struct page_cgroup
*pc
;
3286 pc
= lookup_page_cgroup(page
);
3287 if (!PageCgroupUsed(pc
))
3290 memcg
= pc
->mem_cgroup
;
3294 * We trust that only if there is a memcg associated with the page, it
3295 * is a valid allocation
3300 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3301 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3304 static inline void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3307 #endif /* CONFIG_MEMCG_KMEM */
3309 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3312 * Because tail pages are not marked as "used", set it. We're under
3313 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3314 * charge/uncharge will be never happen and move_account() is done under
3315 * compound_lock(), so we don't have to take care of races.
3317 void mem_cgroup_split_huge_fixup(struct page
*head
)
3319 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3320 struct page_cgroup
*pc
;
3321 struct mem_cgroup
*memcg
;
3324 if (mem_cgroup_disabled())
3327 memcg
= head_pc
->mem_cgroup
;
3328 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3330 pc
->mem_cgroup
= memcg
;
3331 pc
->flags
= head_pc
->flags
;
3333 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3336 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3339 * mem_cgroup_move_account - move account of the page
3341 * @nr_pages: number of regular pages (>1 for huge pages)
3342 * @pc: page_cgroup of the page.
3343 * @from: mem_cgroup which the page is moved from.
3344 * @to: mem_cgroup which the page is moved to. @from != @to.
3346 * The caller must confirm following.
3347 * - page is not on LRU (isolate_page() is useful.)
3348 * - compound_lock is held when nr_pages > 1
3350 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3353 static int mem_cgroup_move_account(struct page
*page
,
3354 unsigned int nr_pages
,
3355 struct page_cgroup
*pc
,
3356 struct mem_cgroup
*from
,
3357 struct mem_cgroup
*to
)
3359 unsigned long flags
;
3362 VM_BUG_ON(from
== to
);
3363 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3365 * The page is isolated from LRU. So, collapse function
3366 * will not handle this page. But page splitting can happen.
3367 * Do this check under compound_page_lock(). The caller should
3371 if (nr_pages
> 1 && !PageTransHuge(page
))
3375 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3376 * of its source page while we change it: page migration takes
3377 * both pages off the LRU, but page cache replacement doesn't.
3379 if (!trylock_page(page
))
3383 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3386 move_lock_mem_cgroup(from
, &flags
);
3388 if (!PageAnon(page
) && page_mapped(page
)) {
3389 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3391 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3395 if (PageWriteback(page
)) {
3396 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3398 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3403 * It is safe to change pc->mem_cgroup here because the page
3404 * is referenced, charged, and isolated - we can't race with
3405 * uncharging, charging, migration, or LRU putback.
3408 /* caller should have done css_get */
3409 pc
->mem_cgroup
= to
;
3410 move_unlock_mem_cgroup(from
, &flags
);
3413 local_irq_disable();
3414 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
3415 memcg_check_events(to
, page
);
3416 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
3417 memcg_check_events(from
, page
);
3426 * mem_cgroup_move_parent - moves page to the parent group
3427 * @page: the page to move
3428 * @pc: page_cgroup of the page
3429 * @child: page's cgroup
3431 * move charges to its parent or the root cgroup if the group has no
3432 * parent (aka use_hierarchy==0).
3433 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3434 * mem_cgroup_move_account fails) the failure is always temporary and
3435 * it signals a race with a page removal/uncharge or migration. In the
3436 * first case the page is on the way out and it will vanish from the LRU
3437 * on the next attempt and the call should be retried later.
3438 * Isolation from the LRU fails only if page has been isolated from
3439 * the LRU since we looked at it and that usually means either global
3440 * reclaim or migration going on. The page will either get back to the
3442 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3443 * (!PageCgroupUsed) or moved to a different group. The page will
3444 * disappear in the next attempt.
3446 static int mem_cgroup_move_parent(struct page
*page
,
3447 struct page_cgroup
*pc
,
3448 struct mem_cgroup
*child
)
3450 struct mem_cgroup
*parent
;
3451 unsigned int nr_pages
;
3452 unsigned long uninitialized_var(flags
);
3455 VM_BUG_ON(mem_cgroup_is_root(child
));
3458 if (!get_page_unless_zero(page
))
3460 if (isolate_lru_page(page
))
3463 nr_pages
= hpage_nr_pages(page
);
3465 parent
= parent_mem_cgroup(child
);
3467 * If no parent, move charges to root cgroup.
3470 parent
= root_mem_cgroup
;
3473 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3474 flags
= compound_lock_irqsave(page
);
3477 ret
= mem_cgroup_move_account(page
, nr_pages
,
3480 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3483 compound_unlock_irqrestore(page
, flags
);
3484 putback_lru_page(page
);
3491 #ifdef CONFIG_MEMCG_SWAP
3492 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
3495 int val
= (charge
) ? 1 : -1;
3496 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
3500 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3501 * @entry: swap entry to be moved
3502 * @from: mem_cgroup which the entry is moved from
3503 * @to: mem_cgroup which the entry is moved to
3505 * It succeeds only when the swap_cgroup's record for this entry is the same
3506 * as the mem_cgroup's id of @from.
3508 * Returns 0 on success, -EINVAL on failure.
3510 * The caller must have charged to @to, IOW, called res_counter_charge() about
3511 * both res and memsw, and called css_get().
3513 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3514 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3516 unsigned short old_id
, new_id
;
3518 old_id
= mem_cgroup_id(from
);
3519 new_id
= mem_cgroup_id(to
);
3521 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3522 mem_cgroup_swap_statistics(from
, false);
3523 mem_cgroup_swap_statistics(to
, true);
3525 * This function is only called from task migration context now.
3526 * It postpones res_counter and refcount handling till the end
3527 * of task migration(mem_cgroup_clear_mc()) for performance
3528 * improvement. But we cannot postpone css_get(to) because if
3529 * the process that has been moved to @to does swap-in, the
3530 * refcount of @to might be decreased to 0.
3532 * We are in attach() phase, so the cgroup is guaranteed to be
3533 * alive, so we can just call css_get().
3541 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3542 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3548 #ifdef CONFIG_DEBUG_VM
3549 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
3551 struct page_cgroup
*pc
;
3553 pc
= lookup_page_cgroup(page
);
3555 * Can be NULL while feeding pages into the page allocator for
3556 * the first time, i.e. during boot or memory hotplug;
3557 * or when mem_cgroup_disabled().
3559 if (likely(pc
) && PageCgroupUsed(pc
))
3564 bool mem_cgroup_bad_page_check(struct page
*page
)
3566 if (mem_cgroup_disabled())
3569 return lookup_page_cgroup_used(page
) != NULL
;
3572 void mem_cgroup_print_bad_page(struct page
*page
)
3574 struct page_cgroup
*pc
;
3576 pc
= lookup_page_cgroup_used(page
);
3578 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3579 pc
, pc
->flags
, pc
->mem_cgroup
);
3584 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3585 unsigned long long val
)
3589 int children
= mem_cgroup_count_children(memcg
);
3590 u64 curusage
, oldusage
;
3594 * For keeping hierarchical_reclaim simple, how long we should retry
3595 * is depends on callers. We set our retry-count to be function
3596 * of # of children which we should visit in this loop.
3598 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
3600 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3603 while (retry_count
) {
3604 if (signal_pending(current
)) {
3609 * Rather than hide all in some function, I do this in
3610 * open coded manner. You see what this really does.
3611 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3613 mutex_lock(&set_limit_mutex
);
3614 if (res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) < val
) {
3616 mutex_unlock(&set_limit_mutex
);
3620 if (res_counter_read_u64(&memcg
->res
, RES_LIMIT
) < val
)
3623 ret
= res_counter_set_limit(&memcg
->res
, val
);
3624 mutex_unlock(&set_limit_mutex
);
3629 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
3631 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3632 /* Usage is reduced ? */
3633 if (curusage
>= oldusage
)
3636 oldusage
= curusage
;
3638 if (!ret
&& enlarge
)
3639 memcg_oom_recover(memcg
);
3644 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3645 unsigned long long val
)
3648 u64 oldusage
, curusage
;
3649 int children
= mem_cgroup_count_children(memcg
);
3653 /* see mem_cgroup_resize_res_limit */
3654 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
3655 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3656 while (retry_count
) {
3657 if (signal_pending(current
)) {
3662 * Rather than hide all in some function, I do this in
3663 * open coded manner. You see what this really does.
3664 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3666 mutex_lock(&set_limit_mutex
);
3667 if (res_counter_read_u64(&memcg
->res
, RES_LIMIT
) > val
) {
3669 mutex_unlock(&set_limit_mutex
);
3672 if (res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) < val
)
3674 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
3675 mutex_unlock(&set_limit_mutex
);
3680 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
3682 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3683 /* Usage is reduced ? */
3684 if (curusage
>= oldusage
)
3687 oldusage
= curusage
;
3689 if (!ret
&& enlarge
)
3690 memcg_oom_recover(memcg
);
3694 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3696 unsigned long *total_scanned
)
3698 unsigned long nr_reclaimed
= 0;
3699 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3700 unsigned long reclaimed
;
3702 struct mem_cgroup_tree_per_zone
*mctz
;
3703 unsigned long long excess
;
3704 unsigned long nr_scanned
;
3709 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3711 * This loop can run a while, specially if mem_cgroup's continuously
3712 * keep exceeding their soft limit and putting the system under
3719 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3724 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
3725 gfp_mask
, &nr_scanned
);
3726 nr_reclaimed
+= reclaimed
;
3727 *total_scanned
+= nr_scanned
;
3728 spin_lock_irq(&mctz
->lock
);
3731 * If we failed to reclaim anything from this memory cgroup
3732 * it is time to move on to the next cgroup
3738 * Loop until we find yet another one.
3740 * By the time we get the soft_limit lock
3741 * again, someone might have aded the
3742 * group back on the RB tree. Iterate to
3743 * make sure we get a different mem.
3744 * mem_cgroup_largest_soft_limit_node returns
3745 * NULL if no other cgroup is present on
3749 __mem_cgroup_largest_soft_limit_node(mctz
);
3751 css_put(&next_mz
->memcg
->css
);
3752 else /* next_mz == NULL or other memcg */
3756 __mem_cgroup_remove_exceeded(mz
, mctz
);
3757 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
3759 * One school of thought says that we should not add
3760 * back the node to the tree if reclaim returns 0.
3761 * But our reclaim could return 0, simply because due
3762 * to priority we are exposing a smaller subset of
3763 * memory to reclaim from. Consider this as a longer
3766 /* If excess == 0, no tree ops */
3767 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3768 spin_unlock_irq(&mctz
->lock
);
3769 css_put(&mz
->memcg
->css
);
3772 * Could not reclaim anything and there are no more
3773 * mem cgroups to try or we seem to be looping without
3774 * reclaiming anything.
3776 if (!nr_reclaimed
&&
3778 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3780 } while (!nr_reclaimed
);
3782 css_put(&next_mz
->memcg
->css
);
3783 return nr_reclaimed
;
3787 * mem_cgroup_force_empty_list - clears LRU of a group
3788 * @memcg: group to clear
3791 * @lru: lru to to clear
3793 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3794 * reclaim the pages page themselves - pages are moved to the parent (or root)
3797 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
3798 int node
, int zid
, enum lru_list lru
)
3800 struct lruvec
*lruvec
;
3801 unsigned long flags
;
3802 struct list_head
*list
;
3806 zone
= &NODE_DATA(node
)->node_zones
[zid
];
3807 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
3808 list
= &lruvec
->lists
[lru
];
3812 struct page_cgroup
*pc
;
3815 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3816 if (list_empty(list
)) {
3817 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3820 page
= list_entry(list
->prev
, struct page
, lru
);
3822 list_move(&page
->lru
, list
);
3824 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3827 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3829 pc
= lookup_page_cgroup(page
);
3831 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
3832 /* found lock contention or "pc" is obsolete. */
3837 } while (!list_empty(list
));
3841 * make mem_cgroup's charge to be 0 if there is no task by moving
3842 * all the charges and pages to the parent.
3843 * This enables deleting this mem_cgroup.
3845 * Caller is responsible for holding css reference on the memcg.
3847 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
3853 /* This is for making all *used* pages to be on LRU. */
3854 lru_add_drain_all();
3855 drain_all_stock_sync(memcg
);
3856 mem_cgroup_start_move(memcg
);
3857 for_each_node_state(node
, N_MEMORY
) {
3858 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3861 mem_cgroup_force_empty_list(memcg
,
3866 mem_cgroup_end_move(memcg
);
3867 memcg_oom_recover(memcg
);
3871 * Kernel memory may not necessarily be trackable to a specific
3872 * process. So they are not migrated, and therefore we can't
3873 * expect their value to drop to 0 here.
3874 * Having res filled up with kmem only is enough.
3876 * This is a safety check because mem_cgroup_force_empty_list
3877 * could have raced with mem_cgroup_replace_page_cache callers
3878 * so the lru seemed empty but the page could have been added
3879 * right after the check. RES_USAGE should be safe as we always
3880 * charge before adding to the LRU.
3882 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
3883 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
3884 } while (usage
> 0);
3888 * Test whether @memcg has children, dead or alive. Note that this
3889 * function doesn't care whether @memcg has use_hierarchy enabled and
3890 * returns %true if there are child csses according to the cgroup
3891 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3893 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3898 * The lock does not prevent addition or deletion of children, but
3899 * it prevents a new child from being initialized based on this
3900 * parent in css_online(), so it's enough to decide whether
3901 * hierarchically inherited attributes can still be changed or not.
3903 lockdep_assert_held(&memcg_create_mutex
);
3906 ret
= css_next_child(NULL
, &memcg
->css
);
3912 * Reclaims as many pages from the given memcg as possible and moves
3913 * the rest to the parent.
3915 * Caller is responsible for holding css reference for memcg.
3917 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3919 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3921 /* we call try-to-free pages for make this cgroup empty */
3922 lru_add_drain_all();
3923 /* try to free all pages in this cgroup */
3924 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
3927 if (signal_pending(current
))
3930 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3934 /* maybe some writeback is necessary */
3935 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3943 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3944 char *buf
, size_t nbytes
,
3947 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3949 if (mem_cgroup_is_root(memcg
))
3951 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3954 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3957 return mem_cgroup_from_css(css
)->use_hierarchy
;
3960 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3961 struct cftype
*cft
, u64 val
)
3964 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3965 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3967 mutex_lock(&memcg_create_mutex
);
3969 if (memcg
->use_hierarchy
== val
)
3973 * If parent's use_hierarchy is set, we can't make any modifications
3974 * in the child subtrees. If it is unset, then the change can
3975 * occur, provided the current cgroup has no children.
3977 * For the root cgroup, parent_mem is NULL, we allow value to be
3978 * set if there are no children.
3980 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3981 (val
== 1 || val
== 0)) {
3982 if (!memcg_has_children(memcg
))
3983 memcg
->use_hierarchy
= val
;
3990 mutex_unlock(&memcg_create_mutex
);
3995 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
3996 enum mem_cgroup_stat_index idx
)
3998 struct mem_cgroup
*iter
;
4001 /* Per-cpu values can be negative, use a signed accumulator */
4002 for_each_mem_cgroup_tree(iter
, memcg
)
4003 val
+= mem_cgroup_read_stat(iter
, idx
);
4005 if (val
< 0) /* race ? */
4010 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4014 if (!mem_cgroup_is_root(memcg
)) {
4016 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4018 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4022 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4023 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4025 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4026 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4029 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4031 return val
<< PAGE_SHIFT
;
4035 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
4038 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4039 enum res_type type
= MEMFILE_TYPE(cft
->private);
4040 int name
= MEMFILE_ATTR(cft
->private);
4044 if (name
== RES_USAGE
)
4045 return mem_cgroup_usage(memcg
, false);
4046 return res_counter_read_u64(&memcg
->res
, name
);
4048 if (name
== RES_USAGE
)
4049 return mem_cgroup_usage(memcg
, true);
4050 return res_counter_read_u64(&memcg
->memsw
, name
);
4052 return res_counter_read_u64(&memcg
->kmem
, name
);
4059 #ifdef CONFIG_MEMCG_KMEM
4060 /* should be called with activate_kmem_mutex held */
4061 static int __memcg_activate_kmem(struct mem_cgroup
*memcg
,
4062 unsigned long long limit
)
4067 if (memcg_kmem_is_active(memcg
))
4071 * We are going to allocate memory for data shared by all memory
4072 * cgroups so let's stop accounting here.
4074 memcg_stop_kmem_account();
4077 * For simplicity, we won't allow this to be disabled. It also can't
4078 * be changed if the cgroup has children already, or if tasks had
4081 * If tasks join before we set the limit, a person looking at
4082 * kmem.usage_in_bytes will have no way to determine when it took
4083 * place, which makes the value quite meaningless.
4085 * After it first became limited, changes in the value of the limit are
4086 * of course permitted.
4088 mutex_lock(&memcg_create_mutex
);
4089 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
4090 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
4092 mutex_unlock(&memcg_create_mutex
);
4096 memcg_id
= memcg_alloc_cache_id();
4102 memcg
->kmemcg_id
= memcg_id
;
4103 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
4106 * We couldn't have accounted to this cgroup, because it hasn't got the
4107 * active bit set yet, so this should succeed.
4109 err
= res_counter_set_limit(&memcg
->kmem
, limit
);
4112 static_key_slow_inc(&memcg_kmem_enabled_key
);
4114 * Setting the active bit after enabling static branching will
4115 * guarantee no one starts accounting before all call sites are
4118 memcg_kmem_set_active(memcg
);
4120 memcg_resume_kmem_account();
4124 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
4125 unsigned long long limit
)
4129 mutex_lock(&activate_kmem_mutex
);
4130 ret
= __memcg_activate_kmem(memcg
, limit
);
4131 mutex_unlock(&activate_kmem_mutex
);
4135 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4136 unsigned long long val
)
4140 if (!memcg_kmem_is_active(memcg
))
4141 ret
= memcg_activate_kmem(memcg
, val
);
4143 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4147 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4150 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4155 mutex_lock(&activate_kmem_mutex
);
4157 * If the parent cgroup is not kmem-active now, it cannot be activated
4158 * after this point, because it has at least one child already.
4160 if (memcg_kmem_is_active(parent
))
4161 ret
= __memcg_activate_kmem(memcg
, RES_COUNTER_MAX
);
4162 mutex_unlock(&activate_kmem_mutex
);
4166 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4167 unsigned long long val
)
4171 #endif /* CONFIG_MEMCG_KMEM */
4174 * The user of this function is...
4177 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
4178 char *buf
, size_t nbytes
, loff_t off
)
4180 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4183 unsigned long long val
;
4186 buf
= strstrip(buf
);
4187 type
= MEMFILE_TYPE(of_cft(of
)->private);
4188 name
= MEMFILE_ATTR(of_cft(of
)->private);
4192 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4196 /* This function does all necessary parse...reuse it */
4197 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4201 ret
= mem_cgroup_resize_limit(memcg
, val
);
4202 else if (type
== _MEMSWAP
)
4203 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
4204 else if (type
== _KMEM
)
4205 ret
= memcg_update_kmem_limit(memcg
, val
);
4209 case RES_SOFT_LIMIT
:
4210 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4214 * For memsw, soft limits are hard to implement in terms
4215 * of semantics, for now, we support soft limits for
4216 * control without swap
4219 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
4224 ret
= -EINVAL
; /* should be BUG() ? */
4227 return ret
?: nbytes
;
4230 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
4231 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
4233 unsigned long long min_limit
, min_memsw_limit
, tmp
;
4235 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4236 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4237 if (!memcg
->use_hierarchy
)
4240 while (memcg
->css
.parent
) {
4241 memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
4242 if (!memcg
->use_hierarchy
)
4244 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4245 min_limit
= min(min_limit
, tmp
);
4246 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4247 min_memsw_limit
= min(min_memsw_limit
, tmp
);
4250 *mem_limit
= min_limit
;
4251 *memsw_limit
= min_memsw_limit
;
4254 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
4255 size_t nbytes
, loff_t off
)
4257 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4261 type
= MEMFILE_TYPE(of_cft(of
)->private);
4262 name
= MEMFILE_ATTR(of_cft(of
)->private);
4267 res_counter_reset_max(&memcg
->res
);
4268 else if (type
== _MEMSWAP
)
4269 res_counter_reset_max(&memcg
->memsw
);
4270 else if (type
== _KMEM
)
4271 res_counter_reset_max(&memcg
->kmem
);
4277 res_counter_reset_failcnt(&memcg
->res
);
4278 else if (type
== _MEMSWAP
)
4279 res_counter_reset_failcnt(&memcg
->memsw
);
4280 else if (type
== _KMEM
)
4281 res_counter_reset_failcnt(&memcg
->kmem
);
4290 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
4293 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
4297 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4298 struct cftype
*cft
, u64 val
)
4300 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4302 if (val
>= (1 << NR_MOVE_TYPE
))
4306 * No kind of locking is needed in here, because ->can_attach() will
4307 * check this value once in the beginning of the process, and then carry
4308 * on with stale data. This means that changes to this value will only
4309 * affect task migrations starting after the change.
4311 memcg
->move_charge_at_immigrate
= val
;
4315 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4316 struct cftype
*cft
, u64 val
)
4323 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
4327 unsigned int lru_mask
;
4330 static const struct numa_stat stats
[] = {
4331 { "total", LRU_ALL
},
4332 { "file", LRU_ALL_FILE
},
4333 { "anon", LRU_ALL_ANON
},
4334 { "unevictable", BIT(LRU_UNEVICTABLE
) },
4336 const struct numa_stat
*stat
;
4339 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4341 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4342 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
4343 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
4344 for_each_node_state(nid
, N_MEMORY
) {
4345 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4347 seq_printf(m
, " N%d=%lu", nid
, nr
);
4352 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4353 struct mem_cgroup
*iter
;
4356 for_each_mem_cgroup_tree(iter
, memcg
)
4357 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
4358 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
4359 for_each_node_state(nid
, N_MEMORY
) {
4361 for_each_mem_cgroup_tree(iter
, memcg
)
4362 nr
+= mem_cgroup_node_nr_lru_pages(
4363 iter
, nid
, stat
->lru_mask
);
4364 seq_printf(m
, " N%d=%lu", nid
, nr
);
4371 #endif /* CONFIG_NUMA */
4373 static inline void mem_cgroup_lru_names_not_uptodate(void)
4375 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
4378 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4380 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4381 struct mem_cgroup
*mi
;
4384 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
4385 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
4387 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
4388 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
4391 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
4392 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
4393 mem_cgroup_read_events(memcg
, i
));
4395 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4396 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
4397 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
4399 /* Hierarchical information */
4401 unsigned long long limit
, memsw_limit
;
4402 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
4403 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
4404 if (do_swap_account
)
4405 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4409 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
4412 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
4414 for_each_mem_cgroup_tree(mi
, memcg
)
4415 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
4416 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
4419 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
4420 unsigned long long val
= 0;
4422 for_each_mem_cgroup_tree(mi
, memcg
)
4423 val
+= mem_cgroup_read_events(mi
, i
);
4424 seq_printf(m
, "total_%s %llu\n",
4425 mem_cgroup_events_names
[i
], val
);
4428 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
4429 unsigned long long val
= 0;
4431 for_each_mem_cgroup_tree(mi
, memcg
)
4432 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
4433 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
4436 #ifdef CONFIG_DEBUG_VM
4439 struct mem_cgroup_per_zone
*mz
;
4440 struct zone_reclaim_stat
*rstat
;
4441 unsigned long recent_rotated
[2] = {0, 0};
4442 unsigned long recent_scanned
[2] = {0, 0};
4444 for_each_online_node(nid
)
4445 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4446 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
4447 rstat
= &mz
->lruvec
.reclaim_stat
;
4449 recent_rotated
[0] += rstat
->recent_rotated
[0];
4450 recent_rotated
[1] += rstat
->recent_rotated
[1];
4451 recent_scanned
[0] += rstat
->recent_scanned
[0];
4452 recent_scanned
[1] += rstat
->recent_scanned
[1];
4454 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
4455 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
4456 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
4457 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
4464 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4467 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4469 return mem_cgroup_swappiness(memcg
);
4472 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4473 struct cftype
*cft
, u64 val
)
4475 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4481 memcg
->swappiness
= val
;
4483 vm_swappiness
= val
;
4488 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4490 struct mem_cgroup_threshold_ary
*t
;
4496 t
= rcu_dereference(memcg
->thresholds
.primary
);
4498 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4503 usage
= mem_cgroup_usage(memcg
, swap
);
4506 * current_threshold points to threshold just below or equal to usage.
4507 * If it's not true, a threshold was crossed after last
4508 * call of __mem_cgroup_threshold().
4510 i
= t
->current_threshold
;
4513 * Iterate backward over array of thresholds starting from
4514 * current_threshold and check if a threshold is crossed.
4515 * If none of thresholds below usage is crossed, we read
4516 * only one element of the array here.
4518 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4519 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4521 /* i = current_threshold + 1 */
4525 * Iterate forward over array of thresholds starting from
4526 * current_threshold+1 and check if a threshold is crossed.
4527 * If none of thresholds above usage is crossed, we read
4528 * only one element of the array here.
4530 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4531 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4533 /* Update current_threshold */
4534 t
->current_threshold
= i
- 1;
4539 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4542 __mem_cgroup_threshold(memcg
, false);
4543 if (do_swap_account
)
4544 __mem_cgroup_threshold(memcg
, true);
4546 memcg
= parent_mem_cgroup(memcg
);
4550 static int compare_thresholds(const void *a
, const void *b
)
4552 const struct mem_cgroup_threshold
*_a
= a
;
4553 const struct mem_cgroup_threshold
*_b
= b
;
4555 if (_a
->threshold
> _b
->threshold
)
4558 if (_a
->threshold
< _b
->threshold
)
4564 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4566 struct mem_cgroup_eventfd_list
*ev
;
4568 spin_lock(&memcg_oom_lock
);
4570 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4571 eventfd_signal(ev
->eventfd
, 1);
4573 spin_unlock(&memcg_oom_lock
);
4577 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4579 struct mem_cgroup
*iter
;
4581 for_each_mem_cgroup_tree(iter
, memcg
)
4582 mem_cgroup_oom_notify_cb(iter
);
4585 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4586 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4588 struct mem_cgroup_thresholds
*thresholds
;
4589 struct mem_cgroup_threshold_ary
*new;
4590 u64 threshold
, usage
;
4593 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
4597 mutex_lock(&memcg
->thresholds_lock
);
4600 thresholds
= &memcg
->thresholds
;
4601 usage
= mem_cgroup_usage(memcg
, false);
4602 } else if (type
== _MEMSWAP
) {
4603 thresholds
= &memcg
->memsw_thresholds
;
4604 usage
= mem_cgroup_usage(memcg
, true);
4608 /* Check if a threshold crossed before adding a new one */
4609 if (thresholds
->primary
)
4610 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4612 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4614 /* Allocate memory for new array of thresholds */
4615 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
4623 /* Copy thresholds (if any) to new array */
4624 if (thresholds
->primary
) {
4625 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4626 sizeof(struct mem_cgroup_threshold
));
4629 /* Add new threshold */
4630 new->entries
[size
- 1].eventfd
= eventfd
;
4631 new->entries
[size
- 1].threshold
= threshold
;
4633 /* Sort thresholds. Registering of new threshold isn't time-critical */
4634 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4635 compare_thresholds
, NULL
);
4637 /* Find current threshold */
4638 new->current_threshold
= -1;
4639 for (i
= 0; i
< size
; i
++) {
4640 if (new->entries
[i
].threshold
<= usage
) {
4642 * new->current_threshold will not be used until
4643 * rcu_assign_pointer(), so it's safe to increment
4646 ++new->current_threshold
;
4651 /* Free old spare buffer and save old primary buffer as spare */
4652 kfree(thresholds
->spare
);
4653 thresholds
->spare
= thresholds
->primary
;
4655 rcu_assign_pointer(thresholds
->primary
, new);
4657 /* To be sure that nobody uses thresholds */
4661 mutex_unlock(&memcg
->thresholds_lock
);
4666 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4667 struct eventfd_ctx
*eventfd
, const char *args
)
4669 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4672 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4673 struct eventfd_ctx
*eventfd
, const char *args
)
4675 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4678 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4679 struct eventfd_ctx
*eventfd
, enum res_type type
)
4681 struct mem_cgroup_thresholds
*thresholds
;
4682 struct mem_cgroup_threshold_ary
*new;
4686 mutex_lock(&memcg
->thresholds_lock
);
4689 thresholds
= &memcg
->thresholds
;
4690 usage
= mem_cgroup_usage(memcg
, false);
4691 } else if (type
== _MEMSWAP
) {
4692 thresholds
= &memcg
->memsw_thresholds
;
4693 usage
= mem_cgroup_usage(memcg
, true);
4697 if (!thresholds
->primary
)
4700 /* Check if a threshold crossed before removing */
4701 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4703 /* Calculate new number of threshold */
4705 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4706 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4710 new = thresholds
->spare
;
4712 /* Set thresholds array to NULL if we don't have thresholds */
4721 /* Copy thresholds and find current threshold */
4722 new->current_threshold
= -1;
4723 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4724 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4727 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4728 if (new->entries
[j
].threshold
<= usage
) {
4730 * new->current_threshold will not be used
4731 * until rcu_assign_pointer(), so it's safe to increment
4734 ++new->current_threshold
;
4740 /* Swap primary and spare array */
4741 thresholds
->spare
= thresholds
->primary
;
4742 /* If all events are unregistered, free the spare array */
4744 kfree(thresholds
->spare
);
4745 thresholds
->spare
= NULL
;
4748 rcu_assign_pointer(thresholds
->primary
, new);
4750 /* To be sure that nobody uses thresholds */
4753 mutex_unlock(&memcg
->thresholds_lock
);
4756 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4757 struct eventfd_ctx
*eventfd
)
4759 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4762 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4763 struct eventfd_ctx
*eventfd
)
4765 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4768 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4769 struct eventfd_ctx
*eventfd
, const char *args
)
4771 struct mem_cgroup_eventfd_list
*event
;
4773 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4777 spin_lock(&memcg_oom_lock
);
4779 event
->eventfd
= eventfd
;
4780 list_add(&event
->list
, &memcg
->oom_notify
);
4782 /* already in OOM ? */
4783 if (atomic_read(&memcg
->under_oom
))
4784 eventfd_signal(eventfd
, 1);
4785 spin_unlock(&memcg_oom_lock
);
4790 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4791 struct eventfd_ctx
*eventfd
)
4793 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4795 spin_lock(&memcg_oom_lock
);
4797 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4798 if (ev
->eventfd
== eventfd
) {
4799 list_del(&ev
->list
);
4804 spin_unlock(&memcg_oom_lock
);
4807 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4809 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
4811 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4812 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
4816 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4817 struct cftype
*cft
, u64 val
)
4819 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4821 /* cannot set to root cgroup and only 0 and 1 are allowed */
4822 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4825 memcg
->oom_kill_disable
= val
;
4827 memcg_oom_recover(memcg
);
4832 #ifdef CONFIG_MEMCG_KMEM
4833 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4837 memcg
->kmemcg_id
= -1;
4838 ret
= memcg_propagate_kmem(memcg
);
4842 return mem_cgroup_sockets_init(memcg
, ss
);
4845 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4847 mem_cgroup_sockets_destroy(memcg
);
4850 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
4852 if (!memcg_kmem_is_active(memcg
))
4856 * kmem charges can outlive the cgroup. In the case of slab
4857 * pages, for instance, a page contain objects from various
4858 * processes. As we prevent from taking a reference for every
4859 * such allocation we have to be careful when doing uncharge
4860 * (see memcg_uncharge_kmem) and here during offlining.
4862 * The idea is that that only the _last_ uncharge which sees
4863 * the dead memcg will drop the last reference. An additional
4864 * reference is taken here before the group is marked dead
4865 * which is then paired with css_put during uncharge resp. here.
4867 * Although this might sound strange as this path is called from
4868 * css_offline() when the referencemight have dropped down to 0 and
4869 * shouldn't be incremented anymore (css_tryget_online() would
4870 * fail) we do not have other options because of the kmem
4871 * allocations lifetime.
4873 css_get(&memcg
->css
);
4875 memcg_kmem_mark_dead(memcg
);
4877 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
4880 if (memcg_kmem_test_and_clear_dead(memcg
))
4881 css_put(&memcg
->css
);
4884 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4889 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4893 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
4899 * DO NOT USE IN NEW FILES.
4901 * "cgroup.event_control" implementation.
4903 * This is way over-engineered. It tries to support fully configurable
4904 * events for each user. Such level of flexibility is completely
4905 * unnecessary especially in the light of the planned unified hierarchy.
4907 * Please deprecate this and replace with something simpler if at all
4912 * Unregister event and free resources.
4914 * Gets called from workqueue.
4916 static void memcg_event_remove(struct work_struct
*work
)
4918 struct mem_cgroup_event
*event
=
4919 container_of(work
, struct mem_cgroup_event
, remove
);
4920 struct mem_cgroup
*memcg
= event
->memcg
;
4922 remove_wait_queue(event
->wqh
, &event
->wait
);
4924 event
->unregister_event(memcg
, event
->eventfd
);
4926 /* Notify userspace the event is going away. */
4927 eventfd_signal(event
->eventfd
, 1);
4929 eventfd_ctx_put(event
->eventfd
);
4931 css_put(&memcg
->css
);
4935 * Gets called on POLLHUP on eventfd when user closes it.
4937 * Called with wqh->lock held and interrupts disabled.
4939 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
4940 int sync
, void *key
)
4942 struct mem_cgroup_event
*event
=
4943 container_of(wait
, struct mem_cgroup_event
, wait
);
4944 struct mem_cgroup
*memcg
= event
->memcg
;
4945 unsigned long flags
= (unsigned long)key
;
4947 if (flags
& POLLHUP
) {
4949 * If the event has been detached at cgroup removal, we
4950 * can simply return knowing the other side will cleanup
4953 * We can't race against event freeing since the other
4954 * side will require wqh->lock via remove_wait_queue(),
4957 spin_lock(&memcg
->event_list_lock
);
4958 if (!list_empty(&event
->list
)) {
4959 list_del_init(&event
->list
);
4961 * We are in atomic context, but cgroup_event_remove()
4962 * may sleep, so we have to call it in workqueue.
4964 schedule_work(&event
->remove
);
4966 spin_unlock(&memcg
->event_list_lock
);
4972 static void memcg_event_ptable_queue_proc(struct file
*file
,
4973 wait_queue_head_t
*wqh
, poll_table
*pt
)
4975 struct mem_cgroup_event
*event
=
4976 container_of(pt
, struct mem_cgroup_event
, pt
);
4979 add_wait_queue(wqh
, &event
->wait
);
4983 * DO NOT USE IN NEW FILES.
4985 * Parse input and register new cgroup event handler.
4987 * Input must be in format '<event_fd> <control_fd> <args>'.
4988 * Interpretation of args is defined by control file implementation.
4990 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4991 char *buf
, size_t nbytes
, loff_t off
)
4993 struct cgroup_subsys_state
*css
= of_css(of
);
4994 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4995 struct mem_cgroup_event
*event
;
4996 struct cgroup_subsys_state
*cfile_css
;
4997 unsigned int efd
, cfd
;
5004 buf
= strstrip(buf
);
5006 efd
= simple_strtoul(buf
, &endp
, 10);
5011 cfd
= simple_strtoul(buf
, &endp
, 10);
5012 if ((*endp
!= ' ') && (*endp
!= '\0'))
5016 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5020 event
->memcg
= memcg
;
5021 INIT_LIST_HEAD(&event
->list
);
5022 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
5023 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
5024 INIT_WORK(&event
->remove
, memcg_event_remove
);
5032 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
5033 if (IS_ERR(event
->eventfd
)) {
5034 ret
= PTR_ERR(event
->eventfd
);
5041 goto out_put_eventfd
;
5044 /* the process need read permission on control file */
5045 /* AV: shouldn't we check that it's been opened for read instead? */
5046 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
5051 * Determine the event callbacks and set them in @event. This used
5052 * to be done via struct cftype but cgroup core no longer knows
5053 * about these events. The following is crude but the whole thing
5054 * is for compatibility anyway.
5056 * DO NOT ADD NEW FILES.
5058 name
= cfile
.file
->f_dentry
->d_name
.name
;
5060 if (!strcmp(name
, "memory.usage_in_bytes")) {
5061 event
->register_event
= mem_cgroup_usage_register_event
;
5062 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
5063 } else if (!strcmp(name
, "memory.oom_control")) {
5064 event
->register_event
= mem_cgroup_oom_register_event
;
5065 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
5066 } else if (!strcmp(name
, "memory.pressure_level")) {
5067 event
->register_event
= vmpressure_register_event
;
5068 event
->unregister_event
= vmpressure_unregister_event
;
5069 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
5070 event
->register_event
= memsw_cgroup_usage_register_event
;
5071 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
5078 * Verify @cfile should belong to @css. Also, remaining events are
5079 * automatically removed on cgroup destruction but the removal is
5080 * asynchronous, so take an extra ref on @css.
5082 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_dentry
->d_parent
,
5083 &memory_cgrp_subsys
);
5085 if (IS_ERR(cfile_css
))
5087 if (cfile_css
!= css
) {
5092 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
5096 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
5098 spin_lock(&memcg
->event_list_lock
);
5099 list_add(&event
->list
, &memcg
->event_list
);
5100 spin_unlock(&memcg
->event_list_lock
);
5112 eventfd_ctx_put(event
->eventfd
);
5121 static struct cftype mem_cgroup_files
[] = {
5123 .name
= "usage_in_bytes",
5124 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5125 .read_u64
= mem_cgroup_read_u64
,
5128 .name
= "max_usage_in_bytes",
5129 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5130 .write
= mem_cgroup_reset
,
5131 .read_u64
= mem_cgroup_read_u64
,
5134 .name
= "limit_in_bytes",
5135 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5136 .write
= mem_cgroup_write
,
5137 .read_u64
= mem_cgroup_read_u64
,
5140 .name
= "soft_limit_in_bytes",
5141 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5142 .write
= mem_cgroup_write
,
5143 .read_u64
= mem_cgroup_read_u64
,
5147 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5148 .write
= mem_cgroup_reset
,
5149 .read_u64
= mem_cgroup_read_u64
,
5153 .seq_show
= memcg_stat_show
,
5156 .name
= "force_empty",
5157 .write
= mem_cgroup_force_empty_write
,
5160 .name
= "use_hierarchy",
5161 .write_u64
= mem_cgroup_hierarchy_write
,
5162 .read_u64
= mem_cgroup_hierarchy_read
,
5165 .name
= "cgroup.event_control", /* XXX: for compat */
5166 .write
= memcg_write_event_control
,
5167 .flags
= CFTYPE_NO_PREFIX
,
5171 .name
= "swappiness",
5172 .read_u64
= mem_cgroup_swappiness_read
,
5173 .write_u64
= mem_cgroup_swappiness_write
,
5176 .name
= "move_charge_at_immigrate",
5177 .read_u64
= mem_cgroup_move_charge_read
,
5178 .write_u64
= mem_cgroup_move_charge_write
,
5181 .name
= "oom_control",
5182 .seq_show
= mem_cgroup_oom_control_read
,
5183 .write_u64
= mem_cgroup_oom_control_write
,
5184 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5187 .name
= "pressure_level",
5191 .name
= "numa_stat",
5192 .seq_show
= memcg_numa_stat_show
,
5195 #ifdef CONFIG_MEMCG_KMEM
5197 .name
= "kmem.limit_in_bytes",
5198 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5199 .write
= mem_cgroup_write
,
5200 .read_u64
= mem_cgroup_read_u64
,
5203 .name
= "kmem.usage_in_bytes",
5204 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5205 .read_u64
= mem_cgroup_read_u64
,
5208 .name
= "kmem.failcnt",
5209 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5210 .write
= mem_cgroup_reset
,
5211 .read_u64
= mem_cgroup_read_u64
,
5214 .name
= "kmem.max_usage_in_bytes",
5215 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5216 .write
= mem_cgroup_reset
,
5217 .read_u64
= mem_cgroup_read_u64
,
5219 #ifdef CONFIG_SLABINFO
5221 .name
= "kmem.slabinfo",
5222 .seq_show
= mem_cgroup_slabinfo_read
,
5226 { }, /* terminate */
5229 #ifdef CONFIG_MEMCG_SWAP
5230 static struct cftype memsw_cgroup_files
[] = {
5232 .name
= "memsw.usage_in_bytes",
5233 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5234 .read_u64
= mem_cgroup_read_u64
,
5237 .name
= "memsw.max_usage_in_bytes",
5238 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5239 .write
= mem_cgroup_reset
,
5240 .read_u64
= mem_cgroup_read_u64
,
5243 .name
= "memsw.limit_in_bytes",
5244 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5245 .write
= mem_cgroup_write
,
5246 .read_u64
= mem_cgroup_read_u64
,
5249 .name
= "memsw.failcnt",
5250 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5251 .write
= mem_cgroup_reset
,
5252 .read_u64
= mem_cgroup_read_u64
,
5254 { }, /* terminate */
5257 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5259 struct mem_cgroup_per_node
*pn
;
5260 struct mem_cgroup_per_zone
*mz
;
5261 int zone
, tmp
= node
;
5263 * This routine is called against possible nodes.
5264 * But it's BUG to call kmalloc() against offline node.
5266 * TODO: this routine can waste much memory for nodes which will
5267 * never be onlined. It's better to use memory hotplug callback
5270 if (!node_state(node
, N_NORMAL_MEMORY
))
5272 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5276 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5277 mz
= &pn
->zoneinfo
[zone
];
5278 lruvec_init(&mz
->lruvec
);
5279 mz
->usage_in_excess
= 0;
5280 mz
->on_tree
= false;
5283 memcg
->nodeinfo
[node
] = pn
;
5287 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5289 kfree(memcg
->nodeinfo
[node
]);
5292 static struct mem_cgroup
*mem_cgroup_alloc(void)
5294 struct mem_cgroup
*memcg
;
5297 size
= sizeof(struct mem_cgroup
);
5298 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5300 memcg
= kzalloc(size
, GFP_KERNEL
);
5304 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5307 spin_lock_init(&memcg
->pcp_counter_lock
);
5316 * At destroying mem_cgroup, references from swap_cgroup can remain.
5317 * (scanning all at force_empty is too costly...)
5319 * Instead of clearing all references at force_empty, we remember
5320 * the number of reference from swap_cgroup and free mem_cgroup when
5321 * it goes down to 0.
5323 * Removal of cgroup itself succeeds regardless of refs from swap.
5326 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5330 mem_cgroup_remove_from_trees(memcg
);
5333 free_mem_cgroup_per_zone_info(memcg
, node
);
5335 free_percpu(memcg
->stat
);
5338 * We need to make sure that (at least for now), the jump label
5339 * destruction code runs outside of the cgroup lock. This is because
5340 * get_online_cpus(), which is called from the static_branch update,
5341 * can't be called inside the cgroup_lock. cpusets are the ones
5342 * enforcing this dependency, so if they ever change, we might as well.
5344 * schedule_work() will guarantee this happens. Be careful if you need
5345 * to move this code around, and make sure it is outside
5348 disarm_static_keys(memcg
);
5353 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5355 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
5357 if (!memcg
->res
.parent
)
5359 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
5361 EXPORT_SYMBOL(parent_mem_cgroup
);
5363 static void __init
mem_cgroup_soft_limit_tree_init(void)
5365 struct mem_cgroup_tree_per_node
*rtpn
;
5366 struct mem_cgroup_tree_per_zone
*rtpz
;
5367 int tmp
, node
, zone
;
5369 for_each_node(node
) {
5371 if (!node_state(node
, N_NORMAL_MEMORY
))
5373 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
5376 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5378 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5379 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5380 rtpz
->rb_root
= RB_ROOT
;
5381 spin_lock_init(&rtpz
->lock
);
5386 static struct cgroup_subsys_state
* __ref
5387 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5389 struct mem_cgroup
*memcg
;
5390 long error
= -ENOMEM
;
5393 memcg
= mem_cgroup_alloc();
5395 return ERR_PTR(error
);
5398 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
5402 if (parent_css
== NULL
) {
5403 root_mem_cgroup
= memcg
;
5404 res_counter_init(&memcg
->res
, NULL
);
5405 res_counter_init(&memcg
->memsw
, NULL
);
5406 res_counter_init(&memcg
->kmem
, NULL
);
5409 memcg
->last_scanned_node
= MAX_NUMNODES
;
5410 INIT_LIST_HEAD(&memcg
->oom_notify
);
5411 memcg
->move_charge_at_immigrate
= 0;
5412 mutex_init(&memcg
->thresholds_lock
);
5413 spin_lock_init(&memcg
->move_lock
);
5414 vmpressure_init(&memcg
->vmpressure
);
5415 INIT_LIST_HEAD(&memcg
->event_list
);
5416 spin_lock_init(&memcg
->event_list_lock
);
5421 __mem_cgroup_free(memcg
);
5422 return ERR_PTR(error
);
5426 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5428 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5429 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
5432 if (css
->id
> MEM_CGROUP_ID_MAX
)
5438 mutex_lock(&memcg_create_mutex
);
5440 memcg
->use_hierarchy
= parent
->use_hierarchy
;
5441 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5442 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5444 if (parent
->use_hierarchy
) {
5445 res_counter_init(&memcg
->res
, &parent
->res
);
5446 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
5447 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
5450 * No need to take a reference to the parent because cgroup
5451 * core guarantees its existence.
5454 res_counter_init(&memcg
->res
, NULL
);
5455 res_counter_init(&memcg
->memsw
, NULL
);
5456 res_counter_init(&memcg
->kmem
, NULL
);
5458 * Deeper hierachy with use_hierarchy == false doesn't make
5459 * much sense so let cgroup subsystem know about this
5460 * unfortunate state in our controller.
5462 if (parent
!= root_mem_cgroup
)
5463 memory_cgrp_subsys
.broken_hierarchy
= true;
5465 mutex_unlock(&memcg_create_mutex
);
5467 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
5472 * Make sure the memcg is initialized: mem_cgroup_iter()
5473 * orders reading memcg->initialized against its callers
5474 * reading the memcg members.
5476 smp_store_release(&memcg
->initialized
, 1);
5482 * Announce all parents that a group from their hierarchy is gone.
5484 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
5486 struct mem_cgroup
*parent
= memcg
;
5488 while ((parent
= parent_mem_cgroup(parent
)))
5489 mem_cgroup_iter_invalidate(parent
);
5492 * if the root memcg is not hierarchical we have to check it
5495 if (!root_mem_cgroup
->use_hierarchy
)
5496 mem_cgroup_iter_invalidate(root_mem_cgroup
);
5499 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5501 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5502 struct mem_cgroup_event
*event
, *tmp
;
5503 struct cgroup_subsys_state
*iter
;
5506 * Unregister events and notify userspace.
5507 * Notify userspace about cgroup removing only after rmdir of cgroup
5508 * directory to avoid race between userspace and kernelspace.
5510 spin_lock(&memcg
->event_list_lock
);
5511 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5512 list_del_init(&event
->list
);
5513 schedule_work(&event
->remove
);
5515 spin_unlock(&memcg
->event_list_lock
);
5517 kmem_cgroup_css_offline(memcg
);
5519 mem_cgroup_invalidate_reclaim_iterators(memcg
);
5522 * This requires that offlining is serialized. Right now that is
5523 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5525 css_for_each_descendant_post(iter
, css
)
5526 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter
));
5528 memcg_unregister_all_caches(memcg
);
5529 vmpressure_cleanup(&memcg
->vmpressure
);
5532 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5534 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5536 * XXX: css_offline() would be where we should reparent all
5537 * memory to prepare the cgroup for destruction. However,
5538 * memcg does not do css_tryget_online() and res_counter charging
5539 * under the same RCU lock region, which means that charging
5540 * could race with offlining. Offlining only happens to
5541 * cgroups with no tasks in them but charges can show up
5542 * without any tasks from the swapin path when the target
5543 * memcg is looked up from the swapout record and not from the
5544 * current task as it usually is. A race like this can leak
5545 * charges and put pages with stale cgroup pointers into
5549 * lookup_swap_cgroup_id()
5551 * mem_cgroup_lookup()
5552 * css_tryget_online()
5554 * disable css_tryget_online()
5557 * reparent_charges()
5558 * res_counter_charge()
5561 * pc->mem_cgroup = dead memcg
5564 * The bulk of the charges are still moved in offline_css() to
5565 * avoid pinning a lot of pages in case a long-term reference
5566 * like a swapout record is deferring the css_free() to long
5567 * after offlining. But this makes sure we catch any charges
5568 * made after offlining:
5570 mem_cgroup_reparent_charges(memcg
);
5572 memcg_destroy_kmem(memcg
);
5573 __mem_cgroup_free(memcg
);
5577 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5578 * @css: the target css
5580 * Reset the states of the mem_cgroup associated with @css. This is
5581 * invoked when the userland requests disabling on the default hierarchy
5582 * but the memcg is pinned through dependency. The memcg should stop
5583 * applying policies and should revert to the vanilla state as it may be
5584 * made visible again.
5586 * The current implementation only resets the essential configurations.
5587 * This needs to be expanded to cover all the visible parts.
5589 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5591 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5593 mem_cgroup_resize_limit(memcg
, ULLONG_MAX
);
5594 mem_cgroup_resize_memsw_limit(memcg
, ULLONG_MAX
);
5595 memcg_update_kmem_limit(memcg
, ULLONG_MAX
);
5596 res_counter_set_soft_limit(&memcg
->res
, ULLONG_MAX
);
5600 /* Handlers for move charge at task migration. */
5601 static int mem_cgroup_do_precharge(unsigned long count
)
5605 /* Try a single bulk charge without reclaim first */
5606 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
5608 mc
.precharge
+= count
;
5611 if (ret
== -EINTR
) {
5612 cancel_charge(root_mem_cgroup
, count
);
5616 /* Try charges one by one with reclaim */
5618 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
5620 * In case of failure, any residual charges against
5621 * mc.to will be dropped by mem_cgroup_clear_mc()
5622 * later on. However, cancel any charges that are
5623 * bypassed to root right away or they'll be lost.
5626 cancel_charge(root_mem_cgroup
, 1);
5636 * get_mctgt_type - get target type of moving charge
5637 * @vma: the vma the pte to be checked belongs
5638 * @addr: the address corresponding to the pte to be checked
5639 * @ptent: the pte to be checked
5640 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5643 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5644 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5645 * move charge. if @target is not NULL, the page is stored in target->page
5646 * with extra refcnt got(Callers should handle it).
5647 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5648 * target for charge migration. if @target is not NULL, the entry is stored
5651 * Called with pte lock held.
5658 enum mc_target_type
{
5664 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5665 unsigned long addr
, pte_t ptent
)
5667 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5669 if (!page
|| !page_mapped(page
))
5671 if (PageAnon(page
)) {
5672 /* we don't move shared anon */
5675 } else if (!move_file())
5676 /* we ignore mapcount for file pages */
5678 if (!get_page_unless_zero(page
))
5685 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5686 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5688 struct page
*page
= NULL
;
5689 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5691 if (!move_anon() || non_swap_entry(ent
))
5694 * Because lookup_swap_cache() updates some statistics counter,
5695 * we call find_get_page() with swapper_space directly.
5697 page
= find_get_page(swap_address_space(ent
), ent
.val
);
5698 if (do_swap_account
)
5699 entry
->val
= ent
.val
;
5704 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5705 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5711 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5712 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5714 struct page
*page
= NULL
;
5715 struct address_space
*mapping
;
5718 if (!vma
->vm_file
) /* anonymous vma */
5723 mapping
= vma
->vm_file
->f_mapping
;
5724 if (pte_none(ptent
))
5725 pgoff
= linear_page_index(vma
, addr
);
5726 else /* pte_file(ptent) is true */
5727 pgoff
= pte_to_pgoff(ptent
);
5729 /* page is moved even if it's not RSS of this task(page-faulted). */
5731 /* shmem/tmpfs may report page out on swap: account for that too. */
5732 if (shmem_mapping(mapping
)) {
5733 page
= find_get_entry(mapping
, pgoff
);
5734 if (radix_tree_exceptional_entry(page
)) {
5735 swp_entry_t swp
= radix_to_swp_entry(page
);
5736 if (do_swap_account
)
5738 page
= find_get_page(swap_address_space(swp
), swp
.val
);
5741 page
= find_get_page(mapping
, pgoff
);
5743 page
= find_get_page(mapping
, pgoff
);
5748 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5749 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5751 struct page
*page
= NULL
;
5752 struct page_cgroup
*pc
;
5753 enum mc_target_type ret
= MC_TARGET_NONE
;
5754 swp_entry_t ent
= { .val
= 0 };
5756 if (pte_present(ptent
))
5757 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5758 else if (is_swap_pte(ptent
))
5759 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
5760 else if (pte_none(ptent
) || pte_file(ptent
))
5761 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5763 if (!page
&& !ent
.val
)
5766 pc
= lookup_page_cgroup(page
);
5768 * Do only loose check w/o serialization.
5769 * mem_cgroup_move_account() checks the pc is valid or
5770 * not under LRU exclusion.
5772 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
5773 ret
= MC_TARGET_PAGE
;
5775 target
->page
= page
;
5777 if (!ret
|| !target
)
5780 /* There is a swap entry and a page doesn't exist or isn't charged */
5781 if (ent
.val
&& !ret
&&
5782 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5783 ret
= MC_TARGET_SWAP
;
5790 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5792 * We don't consider swapping or file mapped pages because THP does not
5793 * support them for now.
5794 * Caller should make sure that pmd_trans_huge(pmd) is true.
5796 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5797 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5799 struct page
*page
= NULL
;
5800 struct page_cgroup
*pc
;
5801 enum mc_target_type ret
= MC_TARGET_NONE
;
5803 page
= pmd_page(pmd
);
5804 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5807 pc
= lookup_page_cgroup(page
);
5808 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
5809 ret
= MC_TARGET_PAGE
;
5812 target
->page
= page
;
5818 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5819 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5821 return MC_TARGET_NONE
;
5825 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5826 unsigned long addr
, unsigned long end
,
5827 struct mm_walk
*walk
)
5829 struct vm_area_struct
*vma
= walk
->private;
5833 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5834 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5835 mc
.precharge
+= HPAGE_PMD_NR
;
5840 if (pmd_trans_unstable(pmd
))
5842 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5843 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5844 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5845 mc
.precharge
++; /* increment precharge temporarily */
5846 pte_unmap_unlock(pte
- 1, ptl
);
5852 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5854 unsigned long precharge
;
5855 struct vm_area_struct
*vma
;
5857 down_read(&mm
->mmap_sem
);
5858 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5859 struct mm_walk mem_cgroup_count_precharge_walk
= {
5860 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5864 if (is_vm_hugetlb_page(vma
))
5866 walk_page_range(vma
->vm_start
, vma
->vm_end
,
5867 &mem_cgroup_count_precharge_walk
);
5869 up_read(&mm
->mmap_sem
);
5871 precharge
= mc
.precharge
;
5877 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5879 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5881 VM_BUG_ON(mc
.moving_task
);
5882 mc
.moving_task
= current
;
5883 return mem_cgroup_do_precharge(precharge
);
5886 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5887 static void __mem_cgroup_clear_mc(void)
5889 struct mem_cgroup
*from
= mc
.from
;
5890 struct mem_cgroup
*to
= mc
.to
;
5893 /* we must uncharge all the leftover precharges from mc.to */
5895 cancel_charge(mc
.to
, mc
.precharge
);
5899 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5900 * we must uncharge here.
5902 if (mc
.moved_charge
) {
5903 cancel_charge(mc
.from
, mc
.moved_charge
);
5904 mc
.moved_charge
= 0;
5906 /* we must fixup refcnts and charges */
5907 if (mc
.moved_swap
) {
5908 /* uncharge swap account from the old cgroup */
5909 if (!mem_cgroup_is_root(mc
.from
))
5910 res_counter_uncharge(&mc
.from
->memsw
,
5911 PAGE_SIZE
* mc
.moved_swap
);
5913 for (i
= 0; i
< mc
.moved_swap
; i
++)
5914 css_put(&mc
.from
->css
);
5917 * we charged both to->res and to->memsw, so we should
5920 if (!mem_cgroup_is_root(mc
.to
))
5921 res_counter_uncharge(&mc
.to
->res
,
5922 PAGE_SIZE
* mc
.moved_swap
);
5923 /* we've already done css_get(mc.to) */
5926 memcg_oom_recover(from
);
5927 memcg_oom_recover(to
);
5928 wake_up_all(&mc
.waitq
);
5931 static void mem_cgroup_clear_mc(void)
5933 struct mem_cgroup
*from
= mc
.from
;
5936 * we must clear moving_task before waking up waiters at the end of
5939 mc
.moving_task
= NULL
;
5940 __mem_cgroup_clear_mc();
5941 spin_lock(&mc
.lock
);
5944 spin_unlock(&mc
.lock
);
5945 mem_cgroup_end_move(from
);
5948 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5949 struct cgroup_taskset
*tset
)
5951 struct task_struct
*p
= cgroup_taskset_first(tset
);
5953 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5954 unsigned long move_charge_at_immigrate
;
5957 * We are now commited to this value whatever it is. Changes in this
5958 * tunable will only affect upcoming migrations, not the current one.
5959 * So we need to save it, and keep it going.
5961 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
5962 if (move_charge_at_immigrate
) {
5963 struct mm_struct
*mm
;
5964 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
5966 VM_BUG_ON(from
== memcg
);
5968 mm
= get_task_mm(p
);
5971 /* We move charges only when we move a owner of the mm */
5972 if (mm
->owner
== p
) {
5975 VM_BUG_ON(mc
.precharge
);
5976 VM_BUG_ON(mc
.moved_charge
);
5977 VM_BUG_ON(mc
.moved_swap
);
5978 mem_cgroup_start_move(from
);
5979 spin_lock(&mc
.lock
);
5982 mc
.immigrate_flags
= move_charge_at_immigrate
;
5983 spin_unlock(&mc
.lock
);
5984 /* We set mc.moving_task later */
5986 ret
= mem_cgroup_precharge_mc(mm
);
5988 mem_cgroup_clear_mc();
5995 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5996 struct cgroup_taskset
*tset
)
5998 mem_cgroup_clear_mc();
6001 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6002 unsigned long addr
, unsigned long end
,
6003 struct mm_walk
*walk
)
6006 struct vm_area_struct
*vma
= walk
->private;
6009 enum mc_target_type target_type
;
6010 union mc_target target
;
6012 struct page_cgroup
*pc
;
6015 * We don't take compound_lock() here but no race with splitting thp
6017 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6018 * under splitting, which means there's no concurrent thp split,
6019 * - if another thread runs into split_huge_page() just after we
6020 * entered this if-block, the thread must wait for page table lock
6021 * to be unlocked in __split_huge_page_splitting(), where the main
6022 * part of thp split is not executed yet.
6024 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6025 if (mc
.precharge
< HPAGE_PMD_NR
) {
6029 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6030 if (target_type
== MC_TARGET_PAGE
) {
6032 if (!isolate_lru_page(page
)) {
6033 pc
= lookup_page_cgroup(page
);
6034 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6035 pc
, mc
.from
, mc
.to
)) {
6036 mc
.precharge
-= HPAGE_PMD_NR
;
6037 mc
.moved_charge
+= HPAGE_PMD_NR
;
6039 putback_lru_page(page
);
6047 if (pmd_trans_unstable(pmd
))
6050 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6051 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6052 pte_t ptent
= *(pte
++);
6058 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6059 case MC_TARGET_PAGE
:
6061 if (isolate_lru_page(page
))
6063 pc
= lookup_page_cgroup(page
);
6064 if (!mem_cgroup_move_account(page
, 1, pc
,
6067 /* we uncharge from mc.from later. */
6070 putback_lru_page(page
);
6071 put
: /* get_mctgt_type() gets the page */
6074 case MC_TARGET_SWAP
:
6076 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6078 /* we fixup refcnts and charges later. */
6086 pte_unmap_unlock(pte
- 1, ptl
);
6091 * We have consumed all precharges we got in can_attach().
6092 * We try charge one by one, but don't do any additional
6093 * charges to mc.to if we have failed in charge once in attach()
6096 ret
= mem_cgroup_do_precharge(1);
6104 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6106 struct vm_area_struct
*vma
;
6108 lru_add_drain_all();
6110 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6112 * Someone who are holding the mmap_sem might be waiting in
6113 * waitq. So we cancel all extra charges, wake up all waiters,
6114 * and retry. Because we cancel precharges, we might not be able
6115 * to move enough charges, but moving charge is a best-effort
6116 * feature anyway, so it wouldn't be a big problem.
6118 __mem_cgroup_clear_mc();
6122 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6124 struct mm_walk mem_cgroup_move_charge_walk
= {
6125 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6129 if (is_vm_hugetlb_page(vma
))
6131 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6132 &mem_cgroup_move_charge_walk
);
6135 * means we have consumed all precharges and failed in
6136 * doing additional charge. Just abandon here.
6140 up_read(&mm
->mmap_sem
);
6143 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6144 struct cgroup_taskset
*tset
)
6146 struct task_struct
*p
= cgroup_taskset_first(tset
);
6147 struct mm_struct
*mm
= get_task_mm(p
);
6151 mem_cgroup_move_charge(mm
);
6155 mem_cgroup_clear_mc();
6157 #else /* !CONFIG_MMU */
6158 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6159 struct cgroup_taskset
*tset
)
6163 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6164 struct cgroup_taskset
*tset
)
6167 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6168 struct cgroup_taskset
*tset
)
6174 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6175 * to verify whether we're attached to the default hierarchy on each mount
6178 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6181 * use_hierarchy is forced on the default hierarchy. cgroup core
6182 * guarantees that @root doesn't have any children, so turning it
6183 * on for the root memcg is enough.
6185 if (cgroup_on_dfl(root_css
->cgroup
))
6186 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6189 struct cgroup_subsys memory_cgrp_subsys
= {
6190 .css_alloc
= mem_cgroup_css_alloc
,
6191 .css_online
= mem_cgroup_css_online
,
6192 .css_offline
= mem_cgroup_css_offline
,
6193 .css_free
= mem_cgroup_css_free
,
6194 .css_reset
= mem_cgroup_css_reset
,
6195 .can_attach
= mem_cgroup_can_attach
,
6196 .cancel_attach
= mem_cgroup_cancel_attach
,
6197 .attach
= mem_cgroup_move_task
,
6198 .bind
= mem_cgroup_bind
,
6199 .legacy_cftypes
= mem_cgroup_files
,
6203 #ifdef CONFIG_MEMCG_SWAP
6204 static int __init
enable_swap_account(char *s
)
6206 if (!strcmp(s
, "1"))
6207 really_do_swap_account
= 1;
6208 else if (!strcmp(s
, "0"))
6209 really_do_swap_account
= 0;
6212 __setup("swapaccount=", enable_swap_account
);
6214 static void __init
memsw_file_init(void)
6216 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6217 memsw_cgroup_files
));
6220 static void __init
enable_swap_cgroup(void)
6222 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6223 do_swap_account
= 1;
6229 static void __init
enable_swap_cgroup(void)
6234 #ifdef CONFIG_MEMCG_SWAP
6236 * mem_cgroup_swapout - transfer a memsw charge to swap
6237 * @page: page whose memsw charge to transfer
6238 * @entry: swap entry to move the charge to
6240 * Transfer the memsw charge of @page to @entry.
6242 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6244 struct page_cgroup
*pc
;
6245 unsigned short oldid
;
6247 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6248 VM_BUG_ON_PAGE(page_count(page
), page
);
6250 if (!do_swap_account
)
6253 pc
= lookup_page_cgroup(page
);
6255 /* Readahead page, never charged */
6256 if (!PageCgroupUsed(pc
))
6259 VM_BUG_ON_PAGE(!(pc
->flags
& PCG_MEMSW
), page
);
6261 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(pc
->mem_cgroup
));
6262 VM_BUG_ON_PAGE(oldid
, page
);
6264 pc
->flags
&= ~PCG_MEMSW
;
6265 css_get(&pc
->mem_cgroup
->css
);
6266 mem_cgroup_swap_statistics(pc
->mem_cgroup
, true);
6270 * mem_cgroup_uncharge_swap - uncharge a swap entry
6271 * @entry: swap entry to uncharge
6273 * Drop the memsw charge associated with @entry.
6275 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
6277 struct mem_cgroup
*memcg
;
6280 if (!do_swap_account
)
6283 id
= swap_cgroup_record(entry
, 0);
6285 memcg
= mem_cgroup_lookup(id
);
6287 if (!mem_cgroup_is_root(memcg
))
6288 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
6289 mem_cgroup_swap_statistics(memcg
, false);
6290 css_put(&memcg
->css
);
6297 * mem_cgroup_try_charge - try charging a page
6298 * @page: page to charge
6299 * @mm: mm context of the victim
6300 * @gfp_mask: reclaim mode
6301 * @memcgp: charged memcg return
6303 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6304 * pages according to @gfp_mask if necessary.
6306 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6307 * Otherwise, an error code is returned.
6309 * After page->mapping has been set up, the caller must finalize the
6310 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6311 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6313 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
6314 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
6316 struct mem_cgroup
*memcg
= NULL
;
6317 unsigned int nr_pages
= 1;
6320 if (mem_cgroup_disabled())
6323 if (PageSwapCache(page
)) {
6324 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
6326 * Every swap fault against a single page tries to charge the
6327 * page, bail as early as possible. shmem_unuse() encounters
6328 * already charged pages, too. The USED bit is protected by
6329 * the page lock, which serializes swap cache removal, which
6330 * in turn serializes uncharging.
6332 if (PageCgroupUsed(pc
))
6336 if (PageTransHuge(page
)) {
6337 nr_pages
<<= compound_order(page
);
6338 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6341 if (do_swap_account
&& PageSwapCache(page
))
6342 memcg
= try_get_mem_cgroup_from_page(page
);
6344 memcg
= get_mem_cgroup_from_mm(mm
);
6346 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6348 css_put(&memcg
->css
);
6350 if (ret
== -EINTR
) {
6351 memcg
= root_mem_cgroup
;
6360 * mem_cgroup_commit_charge - commit a page charge
6361 * @page: page to charge
6362 * @memcg: memcg to charge the page to
6363 * @lrucare: page might be on LRU already
6365 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6366 * after page->mapping has been set up. This must happen atomically
6367 * as part of the page instantiation, i.e. under the page table lock
6368 * for anonymous pages, under the page lock for page and swap cache.
6370 * In addition, the page must not be on the LRU during the commit, to
6371 * prevent racing with task migration. If it might be, use @lrucare.
6373 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6375 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6378 unsigned int nr_pages
= 1;
6380 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6381 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6383 if (mem_cgroup_disabled())
6386 * Swap faults will attempt to charge the same page multiple
6387 * times. But reuse_swap_page() might have removed the page
6388 * from swapcache already, so we can't check PageSwapCache().
6393 commit_charge(page
, memcg
, lrucare
);
6395 if (PageTransHuge(page
)) {
6396 nr_pages
<<= compound_order(page
);
6397 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6400 local_irq_disable();
6401 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6402 memcg_check_events(memcg
, page
);
6405 if (do_swap_account
&& PageSwapCache(page
)) {
6406 swp_entry_t entry
= { .val
= page_private(page
) };
6408 * The swap entry might not get freed for a long time,
6409 * let's not wait for it. The page already received a
6410 * memory+swap charge, drop the swap entry duplicate.
6412 mem_cgroup_uncharge_swap(entry
);
6417 * mem_cgroup_cancel_charge - cancel a page charge
6418 * @page: page to charge
6419 * @memcg: memcg to charge the page to
6421 * Cancel a charge transaction started by mem_cgroup_try_charge().
6423 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
6425 unsigned int nr_pages
= 1;
6427 if (mem_cgroup_disabled())
6430 * Swap faults will attempt to charge the same page multiple
6431 * times. But reuse_swap_page() might have removed the page
6432 * from swapcache already, so we can't check PageSwapCache().
6437 if (PageTransHuge(page
)) {
6438 nr_pages
<<= compound_order(page
);
6439 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6442 cancel_charge(memcg
, nr_pages
);
6445 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
6446 unsigned long nr_mem
, unsigned long nr_memsw
,
6447 unsigned long nr_anon
, unsigned long nr_file
,
6448 unsigned long nr_huge
, struct page
*dummy_page
)
6450 unsigned long flags
;
6452 if (!mem_cgroup_is_root(memcg
)) {
6454 res_counter_uncharge(&memcg
->res
,
6455 nr_mem
* PAGE_SIZE
);
6457 res_counter_uncharge(&memcg
->memsw
,
6458 nr_memsw
* PAGE_SIZE
);
6459 memcg_oom_recover(memcg
);
6462 local_irq_save(flags
);
6463 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
6464 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
6465 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
6466 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
6467 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_anon
+ nr_file
);
6468 memcg_check_events(memcg
, dummy_page
);
6469 local_irq_restore(flags
);
6472 static void uncharge_list(struct list_head
*page_list
)
6474 struct mem_cgroup
*memcg
= NULL
;
6475 unsigned long nr_memsw
= 0;
6476 unsigned long nr_anon
= 0;
6477 unsigned long nr_file
= 0;
6478 unsigned long nr_huge
= 0;
6479 unsigned long pgpgout
= 0;
6480 unsigned long nr_mem
= 0;
6481 struct list_head
*next
;
6484 next
= page_list
->next
;
6486 unsigned int nr_pages
= 1;
6487 struct page_cgroup
*pc
;
6489 page
= list_entry(next
, struct page
, lru
);
6490 next
= page
->lru
.next
;
6492 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6493 VM_BUG_ON_PAGE(page_count(page
), page
);
6495 pc
= lookup_page_cgroup(page
);
6496 if (!PageCgroupUsed(pc
))
6500 * Nobody should be changing or seriously looking at
6501 * pc->mem_cgroup and pc->flags at this point, we have
6502 * fully exclusive access to the page.
6505 if (memcg
!= pc
->mem_cgroup
) {
6507 uncharge_batch(memcg
, pgpgout
, nr_mem
, nr_memsw
,
6508 nr_anon
, nr_file
, nr_huge
, page
);
6509 pgpgout
= nr_mem
= nr_memsw
= 0;
6510 nr_anon
= nr_file
= nr_huge
= 0;
6512 memcg
= pc
->mem_cgroup
;
6515 if (PageTransHuge(page
)) {
6516 nr_pages
<<= compound_order(page
);
6517 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6518 nr_huge
+= nr_pages
;
6522 nr_anon
+= nr_pages
;
6524 nr_file
+= nr_pages
;
6526 if (pc
->flags
& PCG_MEM
)
6528 if (pc
->flags
& PCG_MEMSW
)
6529 nr_memsw
+= nr_pages
;
6533 } while (next
!= page_list
);
6536 uncharge_batch(memcg
, pgpgout
, nr_mem
, nr_memsw
,
6537 nr_anon
, nr_file
, nr_huge
, page
);
6541 * mem_cgroup_uncharge - uncharge a page
6542 * @page: page to uncharge
6544 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6545 * mem_cgroup_commit_charge().
6547 void mem_cgroup_uncharge(struct page
*page
)
6549 struct page_cgroup
*pc
;
6551 if (mem_cgroup_disabled())
6554 /* Don't touch page->lru of any random page, pre-check: */
6555 pc
= lookup_page_cgroup(page
);
6556 if (!PageCgroupUsed(pc
))
6559 INIT_LIST_HEAD(&page
->lru
);
6560 uncharge_list(&page
->lru
);
6564 * mem_cgroup_uncharge_list - uncharge a list of page
6565 * @page_list: list of pages to uncharge
6567 * Uncharge a list of pages previously charged with
6568 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6570 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6572 if (mem_cgroup_disabled())
6575 if (!list_empty(page_list
))
6576 uncharge_list(page_list
);
6580 * mem_cgroup_migrate - migrate a charge to another page
6581 * @oldpage: currently charged page
6582 * @newpage: page to transfer the charge to
6583 * @lrucare: both pages might be on the LRU already
6585 * Migrate the charge from @oldpage to @newpage.
6587 * Both pages must be locked, @newpage->mapping must be set up.
6589 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
6592 struct page_cgroup
*pc
;
6595 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6596 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6597 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
6598 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
6599 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6600 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6603 if (mem_cgroup_disabled())
6606 /* Page cache replacement: new page already charged? */
6607 pc
= lookup_page_cgroup(newpage
);
6608 if (PageCgroupUsed(pc
))
6611 /* Re-entrant migration: old page already uncharged? */
6612 pc
= lookup_page_cgroup(oldpage
);
6613 if (!PageCgroupUsed(pc
))
6616 VM_BUG_ON_PAGE(!(pc
->flags
& PCG_MEM
), oldpage
);
6617 VM_BUG_ON_PAGE(do_swap_account
&& !(pc
->flags
& PCG_MEMSW
), oldpage
);
6620 lock_page_lru(oldpage
, &isolated
);
6625 unlock_page_lru(oldpage
, isolated
);
6627 commit_charge(newpage
, pc
->mem_cgroup
, lrucare
);
6631 * subsys_initcall() for memory controller.
6633 * Some parts like hotcpu_notifier() have to be initialized from this context
6634 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6635 * everything that doesn't depend on a specific mem_cgroup structure should
6636 * be initialized from here.
6638 static int __init
mem_cgroup_init(void)
6640 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6641 enable_swap_cgroup();
6642 mem_cgroup_soft_limit_tree_init();
6646 subsys_initcall(mem_cgroup_init
);