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/page_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 reclaim_iter
{
147 struct mem_cgroup
*position
;
148 /* scan generation, increased every round-trip */
149 unsigned int generation
;
153 * per-zone information in memory controller.
155 struct mem_cgroup_per_zone
{
156 struct lruvec lruvec
;
157 unsigned long lru_size
[NR_LRU_LISTS
];
159 struct reclaim_iter iter
[DEF_PRIORITY
+ 1];
161 struct rb_node tree_node
; /* RB tree node */
162 unsigned long usage_in_excess
;/* Set to the value by which */
163 /* the soft limit is exceeded*/
165 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
166 /* use container_of */
169 struct mem_cgroup_per_node
{
170 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
174 * Cgroups above their limits are maintained in a RB-Tree, independent of
175 * their hierarchy representation
178 struct mem_cgroup_tree_per_zone
{
179 struct rb_root rb_root
;
183 struct mem_cgroup_tree_per_node
{
184 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
187 struct mem_cgroup_tree
{
188 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
191 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
193 struct mem_cgroup_threshold
{
194 struct eventfd_ctx
*eventfd
;
195 unsigned long threshold
;
199 struct mem_cgroup_threshold_ary
{
200 /* An array index points to threshold just below or equal to usage. */
201 int current_threshold
;
202 /* Size of entries[] */
204 /* Array of thresholds */
205 struct mem_cgroup_threshold entries
[0];
208 struct mem_cgroup_thresholds
{
209 /* Primary thresholds array */
210 struct mem_cgroup_threshold_ary
*primary
;
212 * Spare threshold array.
213 * This is needed to make mem_cgroup_unregister_event() "never fail".
214 * It must be able to store at least primary->size - 1 entries.
216 struct mem_cgroup_threshold_ary
*spare
;
220 struct mem_cgroup_eventfd_list
{
221 struct list_head list
;
222 struct eventfd_ctx
*eventfd
;
226 * cgroup_event represents events which userspace want to receive.
228 struct mem_cgroup_event
{
230 * memcg which the event belongs to.
232 struct mem_cgroup
*memcg
;
234 * eventfd to signal userspace about the event.
236 struct eventfd_ctx
*eventfd
;
238 * Each of these stored in a list by the cgroup.
240 struct list_head list
;
242 * register_event() callback will be used to add new userspace
243 * waiter for changes related to this event. Use eventfd_signal()
244 * on eventfd to send notification to userspace.
246 int (*register_event
)(struct mem_cgroup
*memcg
,
247 struct eventfd_ctx
*eventfd
, const char *args
);
249 * unregister_event() callback will be called when userspace closes
250 * the eventfd or on cgroup removing. This callback must be set,
251 * if you want provide notification functionality.
253 void (*unregister_event
)(struct mem_cgroup
*memcg
,
254 struct eventfd_ctx
*eventfd
);
256 * All fields below needed to unregister event when
257 * userspace closes eventfd.
260 wait_queue_head_t
*wqh
;
262 struct work_struct remove
;
265 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
266 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
269 * The memory controller data structure. The memory controller controls both
270 * page cache and RSS per cgroup. We would eventually like to provide
271 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
272 * to help the administrator determine what knobs to tune.
274 * TODO: Add a water mark for the memory controller. Reclaim will begin when
275 * we hit the water mark. May be even add a low water mark, such that
276 * no reclaim occurs from a cgroup at it's low water mark, this is
277 * a feature that will be implemented much later in the future.
280 struct cgroup_subsys_state css
;
282 /* Accounted resources */
283 struct page_counter memory
;
284 struct page_counter memsw
;
285 struct page_counter kmem
;
287 unsigned long soft_limit
;
289 /* vmpressure notifications */
290 struct vmpressure vmpressure
;
292 /* css_online() has been completed */
296 * Should the accounting and control be hierarchical, per subtree?
299 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
303 atomic_t oom_wakeups
;
306 /* OOM-Killer disable */
307 int oom_kill_disable
;
309 /* protect arrays of thresholds */
310 struct mutex thresholds_lock
;
312 /* thresholds for memory usage. RCU-protected */
313 struct mem_cgroup_thresholds thresholds
;
315 /* thresholds for mem+swap usage. RCU-protected */
316 struct mem_cgroup_thresholds memsw_thresholds
;
318 /* For oom notifier event fd */
319 struct list_head oom_notify
;
322 * Should we move charges of a task when a task is moved into this
323 * mem_cgroup ? And what type of charges should we move ?
325 unsigned long move_charge_at_immigrate
;
327 * set > 0 if pages under this cgroup are moving to other cgroup.
329 atomic_t moving_account
;
330 /* taken only while moving_account > 0 */
331 spinlock_t move_lock
;
335 struct mem_cgroup_stat_cpu __percpu
*stat
;
337 * used when a cpu is offlined or other synchronizations
338 * See mem_cgroup_read_stat().
340 struct mem_cgroup_stat_cpu nocpu_base
;
341 spinlock_t pcp_counter_lock
;
343 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
344 struct cg_proto tcp_mem
;
346 #if defined(CONFIG_MEMCG_KMEM)
347 /* analogous to slab_common's slab_caches list, but per-memcg;
348 * protected by memcg_slab_mutex */
349 struct list_head memcg_slab_caches
;
350 /* Index in the kmem_cache->memcg_params->memcg_caches array */
354 int last_scanned_node
;
356 nodemask_t scan_nodes
;
357 atomic_t numainfo_events
;
358 atomic_t numainfo_updating
;
361 /* List of events which userspace want to receive */
362 struct list_head event_list
;
363 spinlock_t event_list_lock
;
365 struct mem_cgroup_per_node
*nodeinfo
[0];
366 /* WARNING: nodeinfo must be the last member here */
369 /* internal only representation about the status of kmem accounting. */
371 KMEM_ACCOUNTED_ACTIVE
, /* accounted by this cgroup itself */
374 #ifdef CONFIG_MEMCG_KMEM
375 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
377 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
380 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
382 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
387 /* Stuffs for move charges at task migration. */
389 * Types of charges to be moved. "move_charge_at_immitgrate" and
390 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
393 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
394 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
398 /* "mc" and its members are protected by cgroup_mutex */
399 static struct move_charge_struct
{
400 spinlock_t lock
; /* for from, to */
401 struct mem_cgroup
*from
;
402 struct mem_cgroup
*to
;
403 unsigned long immigrate_flags
;
404 unsigned long precharge
;
405 unsigned long moved_charge
;
406 unsigned long moved_swap
;
407 struct task_struct
*moving_task
; /* a task moving charges */
408 wait_queue_head_t waitq
; /* a waitq for other context */
410 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
411 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
414 static bool move_anon(void)
416 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
419 static bool move_file(void)
421 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
425 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
426 * limit reclaim to prevent infinite loops, if they ever occur.
428 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
429 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
432 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
433 MEM_CGROUP_CHARGE_TYPE_ANON
,
434 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
435 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
439 /* for encoding cft->private value on file */
447 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
448 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
449 #define MEMFILE_ATTR(val) ((val) & 0xffff)
450 /* Used for OOM nofiier */
451 #define OOM_CONTROL (0)
454 * The memcg_create_mutex will be held whenever a new cgroup is created.
455 * As a consequence, any change that needs to protect against new child cgroups
456 * appearing has to hold it as well.
458 static DEFINE_MUTEX(memcg_create_mutex
);
460 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
462 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
465 /* Some nice accessors for the vmpressure. */
466 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
469 memcg
= root_mem_cgroup
;
470 return &memcg
->vmpressure
;
473 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
475 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
478 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
480 return (memcg
== root_mem_cgroup
);
484 * We restrict the id in the range of [1, 65535], so it can fit into
487 #define MEM_CGROUP_ID_MAX USHRT_MAX
489 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
491 return memcg
->css
.id
;
494 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
496 struct cgroup_subsys_state
*css
;
498 css
= css_from_id(id
, &memory_cgrp_subsys
);
499 return mem_cgroup_from_css(css
);
502 /* Writing them here to avoid exposing memcg's inner layout */
503 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
505 void sock_update_memcg(struct sock
*sk
)
507 if (mem_cgroup_sockets_enabled
) {
508 struct mem_cgroup
*memcg
;
509 struct cg_proto
*cg_proto
;
511 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
513 /* Socket cloning can throw us here with sk_cgrp already
514 * filled. It won't however, necessarily happen from
515 * process context. So the test for root memcg given
516 * the current task's memcg won't help us in this case.
518 * Respecting the original socket's memcg is a better
519 * decision in this case.
522 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
523 css_get(&sk
->sk_cgrp
->memcg
->css
);
528 memcg
= mem_cgroup_from_task(current
);
529 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
530 if (!mem_cgroup_is_root(memcg
) &&
531 memcg_proto_active(cg_proto
) &&
532 css_tryget_online(&memcg
->css
)) {
533 sk
->sk_cgrp
= cg_proto
;
538 EXPORT_SYMBOL(sock_update_memcg
);
540 void sock_release_memcg(struct sock
*sk
)
542 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
543 struct mem_cgroup
*memcg
;
544 WARN_ON(!sk
->sk_cgrp
->memcg
);
545 memcg
= sk
->sk_cgrp
->memcg
;
546 css_put(&sk
->sk_cgrp
->memcg
->css
);
550 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
552 if (!memcg
|| mem_cgroup_is_root(memcg
))
555 return &memcg
->tcp_mem
;
557 EXPORT_SYMBOL(tcp_proto_cgroup
);
559 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
561 if (!memcg_proto_activated(&memcg
->tcp_mem
))
563 static_key_slow_dec(&memcg_socket_limit_enabled
);
566 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
571 #ifdef CONFIG_MEMCG_KMEM
573 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
574 * The main reason for not using cgroup id for this:
575 * this works better in sparse environments, where we have a lot of memcgs,
576 * but only a few kmem-limited. Or also, if we have, for instance, 200
577 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
578 * 200 entry array for that.
580 * The current size of the caches array is stored in
581 * memcg_limited_groups_array_size. It will double each time we have to
584 static DEFINE_IDA(kmem_limited_groups
);
585 int memcg_limited_groups_array_size
;
588 * MIN_SIZE is different than 1, because we would like to avoid going through
589 * the alloc/free process all the time. In a small machine, 4 kmem-limited
590 * cgroups is a reasonable guess. In the future, it could be a parameter or
591 * tunable, but that is strictly not necessary.
593 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
594 * this constant directly from cgroup, but it is understandable that this is
595 * better kept as an internal representation in cgroup.c. In any case, the
596 * cgrp_id space is not getting any smaller, and we don't have to necessarily
597 * increase ours as well if it increases.
599 #define MEMCG_CACHES_MIN_SIZE 4
600 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
603 * A lot of the calls to the cache allocation functions are expected to be
604 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
605 * conditional to this static branch, we'll have to allow modules that does
606 * kmem_cache_alloc and the such to see this symbol as well
608 struct static_key memcg_kmem_enabled_key
;
609 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
611 static void memcg_free_cache_id(int id
);
613 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
615 if (memcg_kmem_is_active(memcg
)) {
616 static_key_slow_dec(&memcg_kmem_enabled_key
);
617 memcg_free_cache_id(memcg
->kmemcg_id
);
620 * This check can't live in kmem destruction function,
621 * since the charges will outlive the cgroup
623 WARN_ON(page_counter_read(&memcg
->kmem
));
626 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
629 #endif /* CONFIG_MEMCG_KMEM */
631 static void disarm_static_keys(struct mem_cgroup
*memcg
)
633 disarm_sock_keys(memcg
);
634 disarm_kmem_keys(memcg
);
637 static struct mem_cgroup_per_zone
*
638 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
640 int nid
= zone_to_nid(zone
);
641 int zid
= zone_idx(zone
);
643 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
646 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
651 static struct mem_cgroup_per_zone
*
652 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
654 int nid
= page_to_nid(page
);
655 int zid
= page_zonenum(page
);
657 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
660 static struct mem_cgroup_tree_per_zone
*
661 soft_limit_tree_node_zone(int nid
, int zid
)
663 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
666 static struct mem_cgroup_tree_per_zone
*
667 soft_limit_tree_from_page(struct page
*page
)
669 int nid
= page_to_nid(page
);
670 int zid
= page_zonenum(page
);
672 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
675 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
676 struct mem_cgroup_tree_per_zone
*mctz
,
677 unsigned long new_usage_in_excess
)
679 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
680 struct rb_node
*parent
= NULL
;
681 struct mem_cgroup_per_zone
*mz_node
;
686 mz
->usage_in_excess
= new_usage_in_excess
;
687 if (!mz
->usage_in_excess
)
691 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
693 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
696 * We can't avoid mem cgroups that are over their soft
697 * limit by the same amount
699 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
702 rb_link_node(&mz
->tree_node
, parent
, p
);
703 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
707 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
708 struct mem_cgroup_tree_per_zone
*mctz
)
712 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
716 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
717 struct mem_cgroup_tree_per_zone
*mctz
)
721 spin_lock_irqsave(&mctz
->lock
, flags
);
722 __mem_cgroup_remove_exceeded(mz
, mctz
);
723 spin_unlock_irqrestore(&mctz
->lock
, flags
);
726 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
728 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
729 unsigned long soft_limit
= ACCESS_ONCE(memcg
->soft_limit
);
730 unsigned long excess
= 0;
732 if (nr_pages
> soft_limit
)
733 excess
= nr_pages
- soft_limit
;
738 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
740 unsigned long excess
;
741 struct mem_cgroup_per_zone
*mz
;
742 struct mem_cgroup_tree_per_zone
*mctz
;
744 mctz
= soft_limit_tree_from_page(page
);
746 * Necessary to update all ancestors when hierarchy is used.
747 * because their event counter is not touched.
749 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
750 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
751 excess
= soft_limit_excess(memcg
);
753 * We have to update the tree if mz is on RB-tree or
754 * mem is over its softlimit.
756 if (excess
|| mz
->on_tree
) {
759 spin_lock_irqsave(&mctz
->lock
, flags
);
760 /* if on-tree, remove it */
762 __mem_cgroup_remove_exceeded(mz
, mctz
);
764 * Insert again. mz->usage_in_excess will be updated.
765 * If excess is 0, no tree ops.
767 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
768 spin_unlock_irqrestore(&mctz
->lock
, flags
);
773 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
775 struct mem_cgroup_tree_per_zone
*mctz
;
776 struct mem_cgroup_per_zone
*mz
;
780 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
781 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
782 mctz
= soft_limit_tree_node_zone(nid
, zid
);
783 mem_cgroup_remove_exceeded(mz
, mctz
);
788 static struct mem_cgroup_per_zone
*
789 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
791 struct rb_node
*rightmost
= NULL
;
792 struct mem_cgroup_per_zone
*mz
;
796 rightmost
= rb_last(&mctz
->rb_root
);
798 goto done
; /* Nothing to reclaim from */
800 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
802 * Remove the node now but someone else can add it back,
803 * we will to add it back at the end of reclaim to its correct
804 * position in the tree.
806 __mem_cgroup_remove_exceeded(mz
, mctz
);
807 if (!soft_limit_excess(mz
->memcg
) ||
808 !css_tryget_online(&mz
->memcg
->css
))
814 static struct mem_cgroup_per_zone
*
815 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
817 struct mem_cgroup_per_zone
*mz
;
819 spin_lock_irq(&mctz
->lock
);
820 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
821 spin_unlock_irq(&mctz
->lock
);
826 * Implementation Note: reading percpu statistics for memcg.
828 * Both of vmstat[] and percpu_counter has threshold and do periodic
829 * synchronization to implement "quick" read. There are trade-off between
830 * reading cost and precision of value. Then, we may have a chance to implement
831 * a periodic synchronizion of counter in memcg's counter.
833 * But this _read() function is used for user interface now. The user accounts
834 * memory usage by memory cgroup and he _always_ requires exact value because
835 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
836 * have to visit all online cpus and make sum. So, for now, unnecessary
837 * synchronization is not implemented. (just implemented for cpu hotplug)
839 * If there are kernel internal actions which can make use of some not-exact
840 * value, and reading all cpu value can be performance bottleneck in some
841 * common workload, threashold and synchonization as vmstat[] should be
844 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
845 enum mem_cgroup_stat_index idx
)
851 for_each_online_cpu(cpu
)
852 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
853 #ifdef CONFIG_HOTPLUG_CPU
854 spin_lock(&memcg
->pcp_counter_lock
);
855 val
+= memcg
->nocpu_base
.count
[idx
];
856 spin_unlock(&memcg
->pcp_counter_lock
);
862 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
863 enum mem_cgroup_events_index idx
)
865 unsigned long val
= 0;
869 for_each_online_cpu(cpu
)
870 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
871 #ifdef CONFIG_HOTPLUG_CPU
872 spin_lock(&memcg
->pcp_counter_lock
);
873 val
+= memcg
->nocpu_base
.events
[idx
];
874 spin_unlock(&memcg
->pcp_counter_lock
);
880 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
885 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
886 * counted as CACHE even if it's on ANON LRU.
889 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
892 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
895 if (PageTransHuge(page
))
896 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
899 /* pagein of a big page is an event. So, ignore page size */
901 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
903 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
904 nr_pages
= -nr_pages
; /* for event */
907 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
910 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
912 struct mem_cgroup_per_zone
*mz
;
914 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
915 return mz
->lru_size
[lru
];
918 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
920 unsigned int lru_mask
)
922 unsigned long nr
= 0;
925 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
927 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
928 struct mem_cgroup_per_zone
*mz
;
932 if (!(BIT(lru
) & lru_mask
))
934 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
935 nr
+= mz
->lru_size
[lru
];
941 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
942 unsigned int lru_mask
)
944 unsigned long nr
= 0;
947 for_each_node_state(nid
, N_MEMORY
)
948 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
952 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
953 enum mem_cgroup_events_target target
)
955 unsigned long val
, next
;
957 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
958 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
959 /* from time_after() in jiffies.h */
960 if ((long)next
- (long)val
< 0) {
962 case MEM_CGROUP_TARGET_THRESH
:
963 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
965 case MEM_CGROUP_TARGET_SOFTLIMIT
:
966 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
968 case MEM_CGROUP_TARGET_NUMAINFO
:
969 next
= val
+ NUMAINFO_EVENTS_TARGET
;
974 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
981 * Check events in order.
984 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
986 /* threshold event is triggered in finer grain than soft limit */
987 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
988 MEM_CGROUP_TARGET_THRESH
))) {
990 bool do_numainfo __maybe_unused
;
992 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
993 MEM_CGROUP_TARGET_SOFTLIMIT
);
995 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
996 MEM_CGROUP_TARGET_NUMAINFO
);
998 mem_cgroup_threshold(memcg
);
999 if (unlikely(do_softlimit
))
1000 mem_cgroup_update_tree(memcg
, page
);
1001 #if MAX_NUMNODES > 1
1002 if (unlikely(do_numainfo
))
1003 atomic_inc(&memcg
->numainfo_events
);
1008 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1011 * mm_update_next_owner() may clear mm->owner to NULL
1012 * if it races with swapoff, page migration, etc.
1013 * So this can be called with p == NULL.
1018 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1021 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1023 struct mem_cgroup
*memcg
= NULL
;
1028 * Page cache insertions can happen withou an
1029 * actual mm context, e.g. during disk probing
1030 * on boot, loopback IO, acct() writes etc.
1033 memcg
= root_mem_cgroup
;
1035 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1036 if (unlikely(!memcg
))
1037 memcg
= root_mem_cgroup
;
1039 } while (!css_tryget_online(&memcg
->css
));
1045 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1046 * @root: hierarchy root
1047 * @prev: previously returned memcg, NULL on first invocation
1048 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1050 * Returns references to children of the hierarchy below @root, or
1051 * @root itself, or %NULL after a full round-trip.
1053 * Caller must pass the return value in @prev on subsequent
1054 * invocations for reference counting, or use mem_cgroup_iter_break()
1055 * to cancel a hierarchy walk before the round-trip is complete.
1057 * Reclaimers can specify a zone and a priority level in @reclaim to
1058 * divide up the memcgs in the hierarchy among all concurrent
1059 * reclaimers operating on the same zone and priority.
1061 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1062 struct mem_cgroup
*prev
,
1063 struct mem_cgroup_reclaim_cookie
*reclaim
)
1065 struct reclaim_iter
*uninitialized_var(iter
);
1066 struct cgroup_subsys_state
*css
= NULL
;
1067 struct mem_cgroup
*memcg
= NULL
;
1068 struct mem_cgroup
*pos
= NULL
;
1070 if (mem_cgroup_disabled())
1074 root
= root_mem_cgroup
;
1076 if (prev
&& !reclaim
)
1079 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1088 struct mem_cgroup_per_zone
*mz
;
1090 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1091 iter
= &mz
->iter
[reclaim
->priority
];
1093 if (prev
&& reclaim
->generation
!= iter
->generation
)
1097 pos
= ACCESS_ONCE(iter
->position
);
1099 * A racing update may change the position and
1100 * put the last reference, hence css_tryget(),
1101 * or retry to see the updated position.
1103 } while (pos
&& !css_tryget(&pos
->css
));
1110 css
= css_next_descendant_pre(css
, &root
->css
);
1113 * Reclaimers share the hierarchy walk, and a
1114 * new one might jump in right at the end of
1115 * the hierarchy - make sure they see at least
1116 * one group and restart from the beginning.
1124 * Verify the css and acquire a reference. The root
1125 * is provided by the caller, so we know it's alive
1126 * and kicking, and don't take an extra reference.
1128 memcg
= mem_cgroup_from_css(css
);
1130 if (css
== &root
->css
)
1133 if (css_tryget(css
)) {
1135 * Make sure the memcg is initialized:
1136 * mem_cgroup_css_online() orders the the
1137 * initialization against setting the flag.
1139 if (smp_load_acquire(&memcg
->initialized
))
1149 if (cmpxchg(&iter
->position
, pos
, memcg
) == pos
) {
1151 css_get(&memcg
->css
);
1157 * pairs with css_tryget when dereferencing iter->position
1166 reclaim
->generation
= iter
->generation
;
1172 if (prev
&& prev
!= root
)
1173 css_put(&prev
->css
);
1179 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1180 * @root: hierarchy root
1181 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1183 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1184 struct mem_cgroup
*prev
)
1187 root
= root_mem_cgroup
;
1188 if (prev
&& prev
!= root
)
1189 css_put(&prev
->css
);
1193 * Iteration constructs for visiting all cgroups (under a tree). If
1194 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1195 * be used for reference counting.
1197 #define for_each_mem_cgroup_tree(iter, root) \
1198 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1200 iter = mem_cgroup_iter(root, iter, NULL))
1202 #define for_each_mem_cgroup(iter) \
1203 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1205 iter = mem_cgroup_iter(NULL, iter, NULL))
1207 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1209 struct mem_cgroup
*memcg
;
1212 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1213 if (unlikely(!memcg
))
1218 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1221 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1229 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1232 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1233 * @zone: zone of the wanted lruvec
1234 * @memcg: memcg of the wanted lruvec
1236 * Returns the lru list vector holding pages for the given @zone and
1237 * @mem. This can be the global zone lruvec, if the memory controller
1240 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1241 struct mem_cgroup
*memcg
)
1243 struct mem_cgroup_per_zone
*mz
;
1244 struct lruvec
*lruvec
;
1246 if (mem_cgroup_disabled()) {
1247 lruvec
= &zone
->lruvec
;
1251 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1252 lruvec
= &mz
->lruvec
;
1255 * Since a node can be onlined after the mem_cgroup was created,
1256 * we have to be prepared to initialize lruvec->zone here;
1257 * and if offlined then reonlined, we need to reinitialize it.
1259 if (unlikely(lruvec
->zone
!= zone
))
1260 lruvec
->zone
= zone
;
1265 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1267 * @zone: zone of the page
1269 * This function is only safe when following the LRU page isolation
1270 * and putback protocol: the LRU lock must be held, and the page must
1271 * either be PageLRU() or the caller must have isolated/allocated it.
1273 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1275 struct mem_cgroup_per_zone
*mz
;
1276 struct mem_cgroup
*memcg
;
1277 struct page_cgroup
*pc
;
1278 struct lruvec
*lruvec
;
1280 if (mem_cgroup_disabled()) {
1281 lruvec
= &zone
->lruvec
;
1285 pc
= lookup_page_cgroup(page
);
1286 memcg
= pc
->mem_cgroup
;
1288 * Swapcache readahead pages are added to the LRU - and
1289 * possibly migrated - before they are charged.
1292 memcg
= root_mem_cgroup
;
1294 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1295 lruvec
= &mz
->lruvec
;
1298 * Since a node can be onlined after the mem_cgroup was created,
1299 * we have to be prepared to initialize lruvec->zone here;
1300 * and if offlined then reonlined, we need to reinitialize it.
1302 if (unlikely(lruvec
->zone
!= zone
))
1303 lruvec
->zone
= zone
;
1308 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1309 * @lruvec: mem_cgroup per zone lru vector
1310 * @lru: index of lru list the page is sitting on
1311 * @nr_pages: positive when adding or negative when removing
1313 * This function must be called when a page is added to or removed from an
1316 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1319 struct mem_cgroup_per_zone
*mz
;
1320 unsigned long *lru_size
;
1322 if (mem_cgroup_disabled())
1325 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1326 lru_size
= mz
->lru_size
+ lru
;
1327 *lru_size
+= nr_pages
;
1328 VM_BUG_ON((long)(*lru_size
) < 0);
1332 * Checks whether given mem is same or in the root_mem_cgroup's
1335 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1336 struct mem_cgroup
*memcg
)
1338 if (root_memcg
== memcg
)
1340 if (!root_memcg
->use_hierarchy
)
1342 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1345 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1346 struct mem_cgroup
*memcg
)
1351 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1356 bool task_in_mem_cgroup(struct task_struct
*task
,
1357 const struct mem_cgroup
*memcg
)
1359 struct mem_cgroup
*curr
;
1360 struct task_struct
*p
;
1363 p
= find_lock_task_mm(task
);
1365 curr
= get_mem_cgroup_from_mm(p
->mm
);
1369 * All threads may have already detached their mm's, but the oom
1370 * killer still needs to detect if they have already been oom
1371 * killed to prevent needlessly killing additional tasks.
1374 curr
= mem_cgroup_from_task(task
);
1375 css_get(&curr
->css
);
1379 * We should check use_hierarchy of "memcg" not "curr". Because checking
1380 * use_hierarchy of "curr" here make this function true if hierarchy is
1381 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1382 * hierarchy(even if use_hierarchy is disabled in "memcg").
1384 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1385 css_put(&curr
->css
);
1389 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1391 unsigned long inactive_ratio
;
1392 unsigned long inactive
;
1393 unsigned long active
;
1396 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1397 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1399 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1401 inactive_ratio
= int_sqrt(10 * gb
);
1405 return inactive
* inactive_ratio
< active
;
1408 #define mem_cgroup_from_counter(counter, member) \
1409 container_of(counter, struct mem_cgroup, member)
1412 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1413 * @memcg: the memory cgroup
1415 * Returns the maximum amount of memory @mem can be charged with, in
1418 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1420 unsigned long margin
= 0;
1421 unsigned long count
;
1422 unsigned long limit
;
1424 count
= page_counter_read(&memcg
->memory
);
1425 limit
= ACCESS_ONCE(memcg
->memory
.limit
);
1427 margin
= limit
- count
;
1429 if (do_swap_account
) {
1430 count
= page_counter_read(&memcg
->memsw
);
1431 limit
= ACCESS_ONCE(memcg
->memsw
.limit
);
1433 margin
= min(margin
, limit
- count
);
1439 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1442 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1443 return vm_swappiness
;
1445 return memcg
->swappiness
;
1449 * A routine for checking "mem" is under move_account() or not.
1451 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1452 * moving cgroups. This is for waiting at high-memory pressure
1455 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1457 struct mem_cgroup
*from
;
1458 struct mem_cgroup
*to
;
1461 * Unlike task_move routines, we access mc.to, mc.from not under
1462 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1464 spin_lock(&mc
.lock
);
1470 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1471 || mem_cgroup_same_or_subtree(memcg
, to
);
1473 spin_unlock(&mc
.lock
);
1477 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1479 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1480 if (mem_cgroup_under_move(memcg
)) {
1482 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1483 /* moving charge context might have finished. */
1486 finish_wait(&mc
.waitq
, &wait
);
1493 #define K(x) ((x) << (PAGE_SHIFT-10))
1495 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1496 * @memcg: The memory cgroup that went over limit
1497 * @p: Task that is going to be killed
1499 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1502 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1504 /* oom_info_lock ensures that parallel ooms do not interleave */
1505 static DEFINE_MUTEX(oom_info_lock
);
1506 struct mem_cgroup
*iter
;
1512 mutex_lock(&oom_info_lock
);
1515 pr_info("Task in ");
1516 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1517 pr_info(" killed as a result of limit of ");
1518 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1523 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1524 K((u64
)page_counter_read(&memcg
->memory
)),
1525 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1526 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1527 K((u64
)page_counter_read(&memcg
->memsw
)),
1528 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1529 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1530 K((u64
)page_counter_read(&memcg
->kmem
)),
1531 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1533 for_each_mem_cgroup_tree(iter
, memcg
) {
1534 pr_info("Memory cgroup stats for ");
1535 pr_cont_cgroup_path(iter
->css
.cgroup
);
1538 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1539 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1541 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1542 K(mem_cgroup_read_stat(iter
, i
)));
1545 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1546 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1547 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1551 mutex_unlock(&oom_info_lock
);
1555 * This function returns the number of memcg under hierarchy tree. Returns
1556 * 1(self count) if no children.
1558 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1561 struct mem_cgroup
*iter
;
1563 for_each_mem_cgroup_tree(iter
, memcg
)
1569 * Return the memory (and swap, if configured) limit for a memcg.
1571 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1573 unsigned long limit
;
1575 limit
= memcg
->memory
.limit
;
1576 if (mem_cgroup_swappiness(memcg
)) {
1577 unsigned long memsw_limit
;
1579 memsw_limit
= memcg
->memsw
.limit
;
1580 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1585 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1588 struct mem_cgroup
*iter
;
1589 unsigned long chosen_points
= 0;
1590 unsigned long totalpages
;
1591 unsigned int points
= 0;
1592 struct task_struct
*chosen
= NULL
;
1595 * If current has a pending SIGKILL or is exiting, then automatically
1596 * select it. The goal is to allow it to allocate so that it may
1597 * quickly exit and free its memory.
1599 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1600 set_thread_flag(TIF_MEMDIE
);
1604 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1605 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1606 for_each_mem_cgroup_tree(iter
, memcg
) {
1607 struct css_task_iter it
;
1608 struct task_struct
*task
;
1610 css_task_iter_start(&iter
->css
, &it
);
1611 while ((task
= css_task_iter_next(&it
))) {
1612 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1614 case OOM_SCAN_SELECT
:
1616 put_task_struct(chosen
);
1618 chosen_points
= ULONG_MAX
;
1619 get_task_struct(chosen
);
1621 case OOM_SCAN_CONTINUE
:
1623 case OOM_SCAN_ABORT
:
1624 css_task_iter_end(&it
);
1625 mem_cgroup_iter_break(memcg
, iter
);
1627 put_task_struct(chosen
);
1632 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1633 if (!points
|| points
< chosen_points
)
1635 /* Prefer thread group leaders for display purposes */
1636 if (points
== chosen_points
&&
1637 thread_group_leader(chosen
))
1641 put_task_struct(chosen
);
1643 chosen_points
= points
;
1644 get_task_struct(chosen
);
1646 css_task_iter_end(&it
);
1651 points
= chosen_points
* 1000 / totalpages
;
1652 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1653 NULL
, "Memory cgroup out of memory");
1657 * test_mem_cgroup_node_reclaimable
1658 * @memcg: the target memcg
1659 * @nid: the node ID to be checked.
1660 * @noswap : specify true here if the user wants flle only information.
1662 * This function returns whether the specified memcg contains any
1663 * reclaimable pages on a node. Returns true if there are any reclaimable
1664 * pages in the node.
1666 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1667 int nid
, bool noswap
)
1669 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1671 if (noswap
|| !total_swap_pages
)
1673 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1678 #if MAX_NUMNODES > 1
1681 * Always updating the nodemask is not very good - even if we have an empty
1682 * list or the wrong list here, we can start from some node and traverse all
1683 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1686 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1690 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1691 * pagein/pageout changes since the last update.
1693 if (!atomic_read(&memcg
->numainfo_events
))
1695 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1698 /* make a nodemask where this memcg uses memory from */
1699 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1701 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1703 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1704 node_clear(nid
, memcg
->scan_nodes
);
1707 atomic_set(&memcg
->numainfo_events
, 0);
1708 atomic_set(&memcg
->numainfo_updating
, 0);
1712 * Selecting a node where we start reclaim from. Because what we need is just
1713 * reducing usage counter, start from anywhere is O,K. Considering
1714 * memory reclaim from current node, there are pros. and cons.
1716 * Freeing memory from current node means freeing memory from a node which
1717 * we'll use or we've used. So, it may make LRU bad. And if several threads
1718 * hit limits, it will see a contention on a node. But freeing from remote
1719 * node means more costs for memory reclaim because of memory latency.
1721 * Now, we use round-robin. Better algorithm is welcomed.
1723 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1727 mem_cgroup_may_update_nodemask(memcg
);
1728 node
= memcg
->last_scanned_node
;
1730 node
= next_node(node
, memcg
->scan_nodes
);
1731 if (node
== MAX_NUMNODES
)
1732 node
= first_node(memcg
->scan_nodes
);
1734 * We call this when we hit limit, not when pages are added to LRU.
1735 * No LRU may hold pages because all pages are UNEVICTABLE or
1736 * memcg is too small and all pages are not on LRU. In that case,
1737 * we use curret node.
1739 if (unlikely(node
== MAX_NUMNODES
))
1740 node
= numa_node_id();
1742 memcg
->last_scanned_node
= node
;
1746 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1752 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1755 unsigned long *total_scanned
)
1757 struct mem_cgroup
*victim
= NULL
;
1760 unsigned long excess
;
1761 unsigned long nr_scanned
;
1762 struct mem_cgroup_reclaim_cookie reclaim
= {
1767 excess
= soft_limit_excess(root_memcg
);
1770 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1775 * If we have not been able to reclaim
1776 * anything, it might because there are
1777 * no reclaimable pages under this hierarchy
1782 * We want to do more targeted reclaim.
1783 * excess >> 2 is not to excessive so as to
1784 * reclaim too much, nor too less that we keep
1785 * coming back to reclaim from this cgroup
1787 if (total
>= (excess
>> 2) ||
1788 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1793 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1795 *total_scanned
+= nr_scanned
;
1796 if (!soft_limit_excess(root_memcg
))
1799 mem_cgroup_iter_break(root_memcg
, victim
);
1803 #ifdef CONFIG_LOCKDEP
1804 static struct lockdep_map memcg_oom_lock_dep_map
= {
1805 .name
= "memcg_oom_lock",
1809 static DEFINE_SPINLOCK(memcg_oom_lock
);
1812 * Check OOM-Killer is already running under our hierarchy.
1813 * If someone is running, return false.
1815 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1817 struct mem_cgroup
*iter
, *failed
= NULL
;
1819 spin_lock(&memcg_oom_lock
);
1821 for_each_mem_cgroup_tree(iter
, memcg
) {
1822 if (iter
->oom_lock
) {
1824 * this subtree of our hierarchy is already locked
1825 * so we cannot give a lock.
1828 mem_cgroup_iter_break(memcg
, iter
);
1831 iter
->oom_lock
= true;
1836 * OK, we failed to lock the whole subtree so we have
1837 * to clean up what we set up to the failing subtree
1839 for_each_mem_cgroup_tree(iter
, memcg
) {
1840 if (iter
== failed
) {
1841 mem_cgroup_iter_break(memcg
, iter
);
1844 iter
->oom_lock
= false;
1847 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1849 spin_unlock(&memcg_oom_lock
);
1854 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1856 struct mem_cgroup
*iter
;
1858 spin_lock(&memcg_oom_lock
);
1859 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1860 for_each_mem_cgroup_tree(iter
, memcg
)
1861 iter
->oom_lock
= false;
1862 spin_unlock(&memcg_oom_lock
);
1865 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1867 struct mem_cgroup
*iter
;
1869 for_each_mem_cgroup_tree(iter
, memcg
)
1870 atomic_inc(&iter
->under_oom
);
1873 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1875 struct mem_cgroup
*iter
;
1878 * When a new child is created while the hierarchy is under oom,
1879 * mem_cgroup_oom_lock() may not be called. We have to use
1880 * atomic_add_unless() here.
1882 for_each_mem_cgroup_tree(iter
, memcg
)
1883 atomic_add_unless(&iter
->under_oom
, -1, 0);
1886 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1888 struct oom_wait_info
{
1889 struct mem_cgroup
*memcg
;
1893 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1894 unsigned mode
, int sync
, void *arg
)
1896 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1897 struct mem_cgroup
*oom_wait_memcg
;
1898 struct oom_wait_info
*oom_wait_info
;
1900 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1901 oom_wait_memcg
= oom_wait_info
->memcg
;
1904 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1905 * Then we can use css_is_ancestor without taking care of RCU.
1907 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
1908 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
1910 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1913 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
1915 atomic_inc(&memcg
->oom_wakeups
);
1916 /* for filtering, pass "memcg" as argument. */
1917 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1920 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1922 if (memcg
&& atomic_read(&memcg
->under_oom
))
1923 memcg_wakeup_oom(memcg
);
1926 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1928 if (!current
->memcg_oom
.may_oom
)
1931 * We are in the middle of the charge context here, so we
1932 * don't want to block when potentially sitting on a callstack
1933 * that holds all kinds of filesystem and mm locks.
1935 * Also, the caller may handle a failed allocation gracefully
1936 * (like optional page cache readahead) and so an OOM killer
1937 * invocation might not even be necessary.
1939 * That's why we don't do anything here except remember the
1940 * OOM context and then deal with it at the end of the page
1941 * fault when the stack is unwound, the locks are released,
1942 * and when we know whether the fault was overall successful.
1944 css_get(&memcg
->css
);
1945 current
->memcg_oom
.memcg
= memcg
;
1946 current
->memcg_oom
.gfp_mask
= mask
;
1947 current
->memcg_oom
.order
= order
;
1951 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1952 * @handle: actually kill/wait or just clean up the OOM state
1954 * This has to be called at the end of a page fault if the memcg OOM
1955 * handler was enabled.
1957 * Memcg supports userspace OOM handling where failed allocations must
1958 * sleep on a waitqueue until the userspace task resolves the
1959 * situation. Sleeping directly in the charge context with all kinds
1960 * of locks held is not a good idea, instead we remember an OOM state
1961 * in the task and mem_cgroup_oom_synchronize() has to be called at
1962 * the end of the page fault to complete the OOM handling.
1964 * Returns %true if an ongoing memcg OOM situation was detected and
1965 * completed, %false otherwise.
1967 bool mem_cgroup_oom_synchronize(bool handle
)
1969 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
1970 struct oom_wait_info owait
;
1973 /* OOM is global, do not handle */
1980 owait
.memcg
= memcg
;
1981 owait
.wait
.flags
= 0;
1982 owait
.wait
.func
= memcg_oom_wake_function
;
1983 owait
.wait
.private = current
;
1984 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1986 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1987 mem_cgroup_mark_under_oom(memcg
);
1989 locked
= mem_cgroup_oom_trylock(memcg
);
1992 mem_cgroup_oom_notify(memcg
);
1994 if (locked
&& !memcg
->oom_kill_disable
) {
1995 mem_cgroup_unmark_under_oom(memcg
);
1996 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1997 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
1998 current
->memcg_oom
.order
);
2001 mem_cgroup_unmark_under_oom(memcg
);
2002 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2006 mem_cgroup_oom_unlock(memcg
);
2008 * There is no guarantee that an OOM-lock contender
2009 * sees the wakeups triggered by the OOM kill
2010 * uncharges. Wake any sleepers explicitely.
2012 memcg_oom_recover(memcg
);
2015 current
->memcg_oom
.memcg
= NULL
;
2016 css_put(&memcg
->css
);
2021 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
2022 * @page: page that is going to change accounted state
2023 * @locked: &memcg->move_lock slowpath was taken
2024 * @flags: IRQ-state flags for &memcg->move_lock
2026 * This function must mark the beginning of an accounted page state
2027 * change to prevent double accounting when the page is concurrently
2028 * being moved to another memcg:
2030 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
2031 * if (TestClearPageState(page))
2032 * mem_cgroup_update_page_stat(memcg, state, -1);
2033 * mem_cgroup_end_page_stat(memcg, locked, flags);
2035 * The RCU lock is held throughout the transaction. The fast path can
2036 * get away without acquiring the memcg->move_lock (@locked is false)
2037 * because page moving starts with an RCU grace period.
2039 * The RCU lock also protects the memcg from being freed when the page
2040 * state that is going to change is the only thing preventing the page
2041 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2042 * which allows migration to go ahead and uncharge the page before the
2043 * account transaction might be complete.
2045 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
,
2047 unsigned long *flags
)
2049 struct mem_cgroup
*memcg
;
2050 struct page_cgroup
*pc
;
2054 if (mem_cgroup_disabled())
2057 pc
= lookup_page_cgroup(page
);
2059 memcg
= pc
->mem_cgroup
;
2060 if (unlikely(!memcg
))
2064 if (atomic_read(&memcg
->moving_account
) <= 0)
2067 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
2068 if (memcg
!= pc
->mem_cgroup
) {
2069 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
2078 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2079 * @memcg: the memcg that was accounted against
2080 * @locked: value received from mem_cgroup_begin_page_stat()
2081 * @flags: value received from mem_cgroup_begin_page_stat()
2083 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
, bool locked
,
2084 unsigned long flags
)
2086 if (memcg
&& locked
)
2087 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2093 * mem_cgroup_update_page_stat - update page state statistics
2094 * @memcg: memcg to account against
2095 * @idx: page state item to account
2096 * @val: number of pages (positive or negative)
2098 * See mem_cgroup_begin_page_stat() for locking requirements.
2100 void mem_cgroup_update_page_stat(struct mem_cgroup
*memcg
,
2101 enum mem_cgroup_stat_index idx
, int val
)
2103 VM_BUG_ON(!rcu_read_lock_held());
2106 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2110 * size of first charge trial. "32" comes from vmscan.c's magic value.
2111 * TODO: maybe necessary to use big numbers in big irons.
2113 #define CHARGE_BATCH 32U
2114 struct memcg_stock_pcp
{
2115 struct mem_cgroup
*cached
; /* this never be root cgroup */
2116 unsigned int nr_pages
;
2117 struct work_struct work
;
2118 unsigned long flags
;
2119 #define FLUSHING_CACHED_CHARGE 0
2121 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2122 static DEFINE_MUTEX(percpu_charge_mutex
);
2125 * consume_stock: Try to consume stocked charge on this cpu.
2126 * @memcg: memcg to consume from.
2127 * @nr_pages: how many pages to charge.
2129 * The charges will only happen if @memcg matches the current cpu's memcg
2130 * stock, and at least @nr_pages are available in that stock. Failure to
2131 * service an allocation will refill the stock.
2133 * returns true if successful, false otherwise.
2135 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2137 struct memcg_stock_pcp
*stock
;
2140 if (nr_pages
> CHARGE_BATCH
)
2143 stock
= &get_cpu_var(memcg_stock
);
2144 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2145 stock
->nr_pages
-= nr_pages
;
2148 put_cpu_var(memcg_stock
);
2153 * Returns stocks cached in percpu and reset cached information.
2155 static void drain_stock(struct memcg_stock_pcp
*stock
)
2157 struct mem_cgroup
*old
= stock
->cached
;
2159 if (stock
->nr_pages
) {
2160 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2161 if (do_swap_account
)
2162 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2163 css_put_many(&old
->css
, stock
->nr_pages
);
2164 stock
->nr_pages
= 0;
2166 stock
->cached
= NULL
;
2170 * This must be called under preempt disabled or must be called by
2171 * a thread which is pinned to local cpu.
2173 static void drain_local_stock(struct work_struct
*dummy
)
2175 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2177 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2180 static void __init
memcg_stock_init(void)
2184 for_each_possible_cpu(cpu
) {
2185 struct memcg_stock_pcp
*stock
=
2186 &per_cpu(memcg_stock
, cpu
);
2187 INIT_WORK(&stock
->work
, drain_local_stock
);
2192 * Cache charges(val) to local per_cpu area.
2193 * This will be consumed by consume_stock() function, later.
2195 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2197 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2199 if (stock
->cached
!= memcg
) { /* reset if necessary */
2201 stock
->cached
= memcg
;
2203 stock
->nr_pages
+= nr_pages
;
2204 put_cpu_var(memcg_stock
);
2208 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2209 * of the hierarchy under it.
2211 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2215 /* If someone's already draining, avoid adding running more workers. */
2216 if (!mutex_trylock(&percpu_charge_mutex
))
2218 /* Notify other cpus that system-wide "drain" is running */
2221 for_each_online_cpu(cpu
) {
2222 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2223 struct mem_cgroup
*memcg
;
2225 memcg
= stock
->cached
;
2226 if (!memcg
|| !stock
->nr_pages
)
2228 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2230 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2232 drain_local_stock(&stock
->work
);
2234 schedule_work_on(cpu
, &stock
->work
);
2239 mutex_unlock(&percpu_charge_mutex
);
2243 * This function drains percpu counter value from DEAD cpu and
2244 * move it to local cpu. Note that this function can be preempted.
2246 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2250 spin_lock(&memcg
->pcp_counter_lock
);
2251 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2252 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2254 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2255 memcg
->nocpu_base
.count
[i
] += x
;
2257 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2258 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2260 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2261 memcg
->nocpu_base
.events
[i
] += x
;
2263 spin_unlock(&memcg
->pcp_counter_lock
);
2266 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2267 unsigned long action
,
2270 int cpu
= (unsigned long)hcpu
;
2271 struct memcg_stock_pcp
*stock
;
2272 struct mem_cgroup
*iter
;
2274 if (action
== CPU_ONLINE
)
2277 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2280 for_each_mem_cgroup(iter
)
2281 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2283 stock
= &per_cpu(memcg_stock
, cpu
);
2288 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2289 unsigned int nr_pages
)
2291 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2292 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2293 struct mem_cgroup
*mem_over_limit
;
2294 struct page_counter
*counter
;
2295 unsigned long nr_reclaimed
;
2296 bool may_swap
= true;
2297 bool drained
= false;
2300 if (mem_cgroup_is_root(memcg
))
2303 if (consume_stock(memcg
, nr_pages
))
2306 if (!do_swap_account
||
2307 !page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2308 if (!page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2310 if (do_swap_account
)
2311 page_counter_uncharge(&memcg
->memsw
, batch
);
2312 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2314 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2318 if (batch
> nr_pages
) {
2324 * Unlike in global OOM situations, memcg is not in a physical
2325 * memory shortage. Allow dying and OOM-killed tasks to
2326 * bypass the last charges so that they can exit quickly and
2327 * free their memory.
2329 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2330 fatal_signal_pending(current
) ||
2331 current
->flags
& PF_EXITING
))
2334 if (unlikely(task_in_memcg_oom(current
)))
2337 if (!(gfp_mask
& __GFP_WAIT
))
2340 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2341 gfp_mask
, may_swap
);
2343 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2347 drain_all_stock(mem_over_limit
);
2352 if (gfp_mask
& __GFP_NORETRY
)
2355 * Even though the limit is exceeded at this point, reclaim
2356 * may have been able to free some pages. Retry the charge
2357 * before killing the task.
2359 * Only for regular pages, though: huge pages are rather
2360 * unlikely to succeed so close to the limit, and we fall back
2361 * to regular pages anyway in case of failure.
2363 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2366 * At task move, charge accounts can be doubly counted. So, it's
2367 * better to wait until the end of task_move if something is going on.
2369 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2375 if (gfp_mask
& __GFP_NOFAIL
)
2378 if (fatal_signal_pending(current
))
2381 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2383 if (!(gfp_mask
& __GFP_NOFAIL
))
2389 css_get_many(&memcg
->css
, batch
);
2390 if (batch
> nr_pages
)
2391 refill_stock(memcg
, batch
- nr_pages
);
2396 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2398 if (mem_cgroup_is_root(memcg
))
2401 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2402 if (do_swap_account
)
2403 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2405 css_put_many(&memcg
->css
, nr_pages
);
2409 * A helper function to get mem_cgroup from ID. must be called under
2410 * rcu_read_lock(). The caller is responsible for calling
2411 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2412 * refcnt from swap can be called against removed memcg.)
2414 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2416 /* ID 0 is unused ID */
2419 return mem_cgroup_from_id(id
);
2423 * try_get_mem_cgroup_from_page - look up page's memcg association
2426 * Look up, get a css reference, and return the memcg that owns @page.
2428 * The page must be locked to prevent racing with swap-in and page
2429 * cache charges. If coming from an unlocked page table, the caller
2430 * must ensure the page is on the LRU or this can race with charging.
2432 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2434 struct mem_cgroup
*memcg
;
2435 struct page_cgroup
*pc
;
2439 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2441 pc
= lookup_page_cgroup(page
);
2442 memcg
= pc
->mem_cgroup
;
2445 if (!css_tryget_online(&memcg
->css
))
2447 } else if (PageSwapCache(page
)) {
2448 ent
.val
= page_private(page
);
2449 id
= lookup_swap_cgroup_id(ent
);
2451 memcg
= mem_cgroup_lookup(id
);
2452 if (memcg
&& !css_tryget_online(&memcg
->css
))
2459 static void lock_page_lru(struct page
*page
, int *isolated
)
2461 struct zone
*zone
= page_zone(page
);
2463 spin_lock_irq(&zone
->lru_lock
);
2464 if (PageLRU(page
)) {
2465 struct lruvec
*lruvec
;
2467 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2469 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2475 static void unlock_page_lru(struct page
*page
, int isolated
)
2477 struct zone
*zone
= page_zone(page
);
2480 struct lruvec
*lruvec
;
2482 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2483 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2485 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2487 spin_unlock_irq(&zone
->lru_lock
);
2490 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2493 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2496 VM_BUG_ON_PAGE(pc
->mem_cgroup
, page
);
2498 * we don't need page_cgroup_lock about tail pages, becase they are not
2499 * accessed by any other context at this point.
2503 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2504 * may already be on some other mem_cgroup's LRU. Take care of it.
2507 lock_page_lru(page
, &isolated
);
2510 * Nobody should be changing or seriously looking at
2511 * pc->mem_cgroup at this point:
2513 * - the page is uncharged
2515 * - the page is off-LRU
2517 * - an anonymous fault has exclusive page access, except for
2518 * a locked page table
2520 * - a page cache insertion, a swapin fault, or a migration
2521 * have the page locked
2523 pc
->mem_cgroup
= memcg
;
2526 unlock_page_lru(page
, isolated
);
2529 #ifdef CONFIG_MEMCG_KMEM
2531 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2532 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2534 static DEFINE_MUTEX(memcg_slab_mutex
);
2537 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2538 * in the memcg_cache_params struct.
2540 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2542 struct kmem_cache
*cachep
;
2544 VM_BUG_ON(p
->is_root_cache
);
2545 cachep
= p
->root_cache
;
2546 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2549 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
,
2550 unsigned long nr_pages
)
2552 struct page_counter
*counter
;
2555 ret
= page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
);
2559 ret
= try_charge(memcg
, gfp
, nr_pages
);
2560 if (ret
== -EINTR
) {
2562 * try_charge() chose to bypass to root due to OOM kill or
2563 * fatal signal. Since our only options are to either fail
2564 * the allocation or charge it to this cgroup, do it as a
2565 * temporary condition. But we can't fail. From a kmem/slab
2566 * perspective, the cache has already been selected, by
2567 * mem_cgroup_kmem_get_cache(), so it is too late to change
2570 * This condition will only trigger if the task entered
2571 * memcg_charge_kmem in a sane state, but was OOM-killed
2572 * during try_charge() above. Tasks that were already dying
2573 * when the allocation triggers should have been already
2574 * directed to the root cgroup in memcontrol.h
2576 page_counter_charge(&memcg
->memory
, nr_pages
);
2577 if (do_swap_account
)
2578 page_counter_charge(&memcg
->memsw
, nr_pages
);
2579 css_get_many(&memcg
->css
, nr_pages
);
2582 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2587 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
,
2588 unsigned long nr_pages
)
2590 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2591 if (do_swap_account
)
2592 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2594 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2596 css_put_many(&memcg
->css
, nr_pages
);
2600 * helper for acessing a memcg's index. It will be used as an index in the
2601 * child cache array in kmem_cache, and also to derive its name. This function
2602 * will return -1 when this is not a kmem-limited memcg.
2604 int memcg_cache_id(struct mem_cgroup
*memcg
)
2606 return memcg
? memcg
->kmemcg_id
: -1;
2609 static int memcg_alloc_cache_id(void)
2614 id
= ida_simple_get(&kmem_limited_groups
,
2615 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2619 if (id
< memcg_limited_groups_array_size
)
2623 * There's no space for the new id in memcg_caches arrays,
2624 * so we have to grow them.
2627 size
= 2 * (id
+ 1);
2628 if (size
< MEMCG_CACHES_MIN_SIZE
)
2629 size
= MEMCG_CACHES_MIN_SIZE
;
2630 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2631 size
= MEMCG_CACHES_MAX_SIZE
;
2633 mutex_lock(&memcg_slab_mutex
);
2634 err
= memcg_update_all_caches(size
);
2635 mutex_unlock(&memcg_slab_mutex
);
2638 ida_simple_remove(&kmem_limited_groups
, id
);
2644 static void memcg_free_cache_id(int id
)
2646 ida_simple_remove(&kmem_limited_groups
, id
);
2650 * We should update the current array size iff all caches updates succeed. This
2651 * can only be done from the slab side. The slab mutex needs to be held when
2654 void memcg_update_array_size(int num
)
2656 memcg_limited_groups_array_size
= num
;
2659 static void memcg_register_cache(struct mem_cgroup
*memcg
,
2660 struct kmem_cache
*root_cache
)
2662 static char memcg_name_buf
[NAME_MAX
+ 1]; /* protected by
2664 struct kmem_cache
*cachep
;
2667 lockdep_assert_held(&memcg_slab_mutex
);
2669 id
= memcg_cache_id(memcg
);
2672 * Since per-memcg caches are created asynchronously on first
2673 * allocation (see memcg_kmem_get_cache()), several threads can try to
2674 * create the same cache, but only one of them may succeed.
2676 if (cache_from_memcg_idx(root_cache
, id
))
2679 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
2680 cachep
= memcg_create_kmem_cache(memcg
, root_cache
, memcg_name_buf
);
2682 * If we could not create a memcg cache, do not complain, because
2683 * that's not critical at all as we can always proceed with the root
2689 css_get(&memcg
->css
);
2690 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2693 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2694 * barrier here to ensure nobody will see the kmem_cache partially
2699 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
2700 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
2703 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
2705 struct kmem_cache
*root_cache
;
2706 struct mem_cgroup
*memcg
;
2709 lockdep_assert_held(&memcg_slab_mutex
);
2711 BUG_ON(is_root_cache(cachep
));
2713 root_cache
= cachep
->memcg_params
->root_cache
;
2714 memcg
= cachep
->memcg_params
->memcg
;
2715 id
= memcg_cache_id(memcg
);
2717 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
2718 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
2720 list_del(&cachep
->memcg_params
->list
);
2722 kmem_cache_destroy(cachep
);
2724 /* drop the reference taken in memcg_register_cache */
2725 css_put(&memcg
->css
);
2729 * During the creation a new cache, we need to disable our accounting mechanism
2730 * altogether. This is true even if we are not creating, but rather just
2731 * enqueing new caches to be created.
2733 * This is because that process will trigger allocations; some visible, like
2734 * explicit kmallocs to auxiliary data structures, name strings and internal
2735 * cache structures; some well concealed, like INIT_WORK() that can allocate
2736 * objects during debug.
2738 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2739 * to it. This may not be a bounded recursion: since the first cache creation
2740 * failed to complete (waiting on the allocation), we'll just try to create the
2741 * cache again, failing at the same point.
2743 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2744 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2745 * inside the following two functions.
2747 static inline void memcg_stop_kmem_account(void)
2749 VM_BUG_ON(!current
->mm
);
2750 current
->memcg_kmem_skip_account
++;
2753 static inline void memcg_resume_kmem_account(void)
2755 VM_BUG_ON(!current
->mm
);
2756 current
->memcg_kmem_skip_account
--;
2759 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
2761 struct kmem_cache
*c
;
2764 mutex_lock(&memcg_slab_mutex
);
2765 for_each_memcg_cache_index(i
) {
2766 c
= cache_from_memcg_idx(s
, i
);
2770 memcg_unregister_cache(c
);
2772 if (cache_from_memcg_idx(s
, i
))
2775 mutex_unlock(&memcg_slab_mutex
);
2779 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
2781 struct kmem_cache
*cachep
;
2782 struct memcg_cache_params
*params
, *tmp
;
2784 if (!memcg_kmem_is_active(memcg
))
2787 mutex_lock(&memcg_slab_mutex
);
2788 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
2789 cachep
= memcg_params_to_cache(params
);
2790 kmem_cache_shrink(cachep
);
2791 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
2792 memcg_unregister_cache(cachep
);
2794 mutex_unlock(&memcg_slab_mutex
);
2797 struct memcg_register_cache_work
{
2798 struct mem_cgroup
*memcg
;
2799 struct kmem_cache
*cachep
;
2800 struct work_struct work
;
2803 static void memcg_register_cache_func(struct work_struct
*w
)
2805 struct memcg_register_cache_work
*cw
=
2806 container_of(w
, struct memcg_register_cache_work
, work
);
2807 struct mem_cgroup
*memcg
= cw
->memcg
;
2808 struct kmem_cache
*cachep
= cw
->cachep
;
2810 mutex_lock(&memcg_slab_mutex
);
2811 memcg_register_cache(memcg
, cachep
);
2812 mutex_unlock(&memcg_slab_mutex
);
2814 css_put(&memcg
->css
);
2819 * Enqueue the creation of a per-memcg kmem_cache.
2821 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
2822 struct kmem_cache
*cachep
)
2824 struct memcg_register_cache_work
*cw
;
2826 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2828 css_put(&memcg
->css
);
2833 cw
->cachep
= cachep
;
2835 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
2836 schedule_work(&cw
->work
);
2839 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
2840 struct kmem_cache
*cachep
)
2843 * We need to stop accounting when we kmalloc, because if the
2844 * corresponding kmalloc cache is not yet created, the first allocation
2845 * in __memcg_schedule_register_cache will recurse.
2847 * However, it is better to enclose the whole function. Depending on
2848 * the debugging options enabled, INIT_WORK(), for instance, can
2849 * trigger an allocation. This too, will make us recurse. Because at
2850 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2851 * the safest choice is to do it like this, wrapping the whole function.
2853 memcg_stop_kmem_account();
2854 __memcg_schedule_register_cache(memcg
, cachep
);
2855 memcg_resume_kmem_account();
2858 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
2860 unsigned int nr_pages
= 1 << order
;
2863 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
, nr_pages
);
2865 atomic_add(nr_pages
, &cachep
->memcg_params
->nr_pages
);
2869 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
2871 unsigned int nr_pages
= 1 << order
;
2873 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, nr_pages
);
2874 atomic_sub(nr_pages
, &cachep
->memcg_params
->nr_pages
);
2878 * Return the kmem_cache we're supposed to use for a slab allocation.
2879 * We try to use the current memcg's version of the cache.
2881 * If the cache does not exist yet, if we are the first user of it,
2882 * we either create it immediately, if possible, or create it asynchronously
2884 * In the latter case, we will let the current allocation go through with
2885 * the original cache.
2887 * Can't be called in interrupt context or from kernel threads.
2888 * This function needs to be called with rcu_read_lock() held.
2890 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
2893 struct mem_cgroup
*memcg
;
2894 struct kmem_cache
*memcg_cachep
;
2896 VM_BUG_ON(!cachep
->memcg_params
);
2897 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
2899 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
2903 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
2905 if (!memcg_kmem_is_active(memcg
))
2908 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
2909 if (likely(memcg_cachep
)) {
2910 cachep
= memcg_cachep
;
2914 /* The corresponding put will be done in the workqueue. */
2915 if (!css_tryget_online(&memcg
->css
))
2920 * If we are in a safe context (can wait, and not in interrupt
2921 * context), we could be be predictable and return right away.
2922 * This would guarantee that the allocation being performed
2923 * already belongs in the new cache.
2925 * However, there are some clashes that can arrive from locking.
2926 * For instance, because we acquire the slab_mutex while doing
2927 * memcg_create_kmem_cache, this means no further allocation
2928 * could happen with the slab_mutex held. So it's better to
2931 memcg_schedule_register_cache(memcg
, cachep
);
2939 * We need to verify if the allocation against current->mm->owner's memcg is
2940 * possible for the given order. But the page is not allocated yet, so we'll
2941 * need a further commit step to do the final arrangements.
2943 * It is possible for the task to switch cgroups in this mean time, so at
2944 * commit time, we can't rely on task conversion any longer. We'll then use
2945 * the handle argument to return to the caller which cgroup we should commit
2946 * against. We could also return the memcg directly and avoid the pointer
2947 * passing, but a boolean return value gives better semantics considering
2948 * the compiled-out case as well.
2950 * Returning true means the allocation is possible.
2953 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
2955 struct mem_cgroup
*memcg
;
2961 * Disabling accounting is only relevant for some specific memcg
2962 * internal allocations. Therefore we would initially not have such
2963 * check here, since direct calls to the page allocator that are
2964 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
2965 * outside memcg core. We are mostly concerned with cache allocations,
2966 * and by having this test at memcg_kmem_get_cache, we are already able
2967 * to relay the allocation to the root cache and bypass the memcg cache
2970 * There is one exception, though: the SLUB allocator does not create
2971 * large order caches, but rather service large kmallocs directly from
2972 * the page allocator. Therefore, the following sequence when backed by
2973 * the SLUB allocator:
2975 * memcg_stop_kmem_account();
2976 * kmalloc(<large_number>)
2977 * memcg_resume_kmem_account();
2979 * would effectively ignore the fact that we should skip accounting,
2980 * since it will drive us directly to this function without passing
2981 * through the cache selector memcg_kmem_get_cache. Such large
2982 * allocations are extremely rare but can happen, for instance, for the
2983 * cache arrays. We bring this test here.
2985 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
2988 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2990 if (!memcg_kmem_is_active(memcg
)) {
2991 css_put(&memcg
->css
);
2995 ret
= memcg_charge_kmem(memcg
, gfp
, 1 << order
);
2999 css_put(&memcg
->css
);
3003 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3006 struct page_cgroup
*pc
;
3008 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3010 /* The page allocation failed. Revert */
3012 memcg_uncharge_kmem(memcg
, 1 << order
);
3015 pc
= lookup_page_cgroup(page
);
3016 pc
->mem_cgroup
= memcg
;
3019 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3021 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
3022 struct mem_cgroup
*memcg
= pc
->mem_cgroup
;
3027 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3029 memcg_uncharge_kmem(memcg
, 1 << order
);
3030 pc
->mem_cgroup
= NULL
;
3033 static inline void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3036 #endif /* CONFIG_MEMCG_KMEM */
3038 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3041 * Because tail pages are not marked as "used", set it. We're under
3042 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3043 * charge/uncharge will be never happen and move_account() is done under
3044 * compound_lock(), so we don't have to take care of races.
3046 void mem_cgroup_split_huge_fixup(struct page
*head
)
3048 struct page_cgroup
*pc
= lookup_page_cgroup(head
);
3051 if (mem_cgroup_disabled())
3054 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
3055 pc
[i
].mem_cgroup
= pc
[0].mem_cgroup
;
3057 __this_cpu_sub(pc
[0].mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3060 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3063 * mem_cgroup_move_account - move account of the page
3065 * @nr_pages: number of regular pages (>1 for huge pages)
3066 * @pc: page_cgroup of the page.
3067 * @from: mem_cgroup which the page is moved from.
3068 * @to: mem_cgroup which the page is moved to. @from != @to.
3070 * The caller must confirm following.
3071 * - page is not on LRU (isolate_page() is useful.)
3072 * - compound_lock is held when nr_pages > 1
3074 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3077 static int mem_cgroup_move_account(struct page
*page
,
3078 unsigned int nr_pages
,
3079 struct page_cgroup
*pc
,
3080 struct mem_cgroup
*from
,
3081 struct mem_cgroup
*to
)
3083 unsigned long flags
;
3086 VM_BUG_ON(from
== to
);
3087 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3089 * The page is isolated from LRU. So, collapse function
3090 * will not handle this page. But page splitting can happen.
3091 * Do this check under compound_page_lock(). The caller should
3095 if (nr_pages
> 1 && !PageTransHuge(page
))
3099 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3100 * of its source page while we change it: page migration takes
3101 * both pages off the LRU, but page cache replacement doesn't.
3103 if (!trylock_page(page
))
3107 if (pc
->mem_cgroup
!= from
)
3110 spin_lock_irqsave(&from
->move_lock
, flags
);
3112 if (!PageAnon(page
) && page_mapped(page
)) {
3113 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3115 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3119 if (PageWriteback(page
)) {
3120 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3122 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3127 * It is safe to change pc->mem_cgroup here because the page
3128 * is referenced, charged, and isolated - we can't race with
3129 * uncharging, charging, migration, or LRU putback.
3132 /* caller should have done css_get */
3133 pc
->mem_cgroup
= to
;
3134 spin_unlock_irqrestore(&from
->move_lock
, flags
);
3138 local_irq_disable();
3139 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
3140 memcg_check_events(to
, page
);
3141 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
3142 memcg_check_events(from
, page
);
3150 #ifdef CONFIG_MEMCG_SWAP
3151 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
3154 int val
= (charge
) ? 1 : -1;
3155 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
3159 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3160 * @entry: swap entry to be moved
3161 * @from: mem_cgroup which the entry is moved from
3162 * @to: mem_cgroup which the entry is moved to
3164 * It succeeds only when the swap_cgroup's record for this entry is the same
3165 * as the mem_cgroup's id of @from.
3167 * Returns 0 on success, -EINVAL on failure.
3169 * The caller must have charged to @to, IOW, called page_counter_charge() about
3170 * both res and memsw, and called css_get().
3172 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3173 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3175 unsigned short old_id
, new_id
;
3177 old_id
= mem_cgroup_id(from
);
3178 new_id
= mem_cgroup_id(to
);
3180 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3181 mem_cgroup_swap_statistics(from
, false);
3182 mem_cgroup_swap_statistics(to
, true);
3184 * This function is only called from task migration context now.
3185 * It postpones page_counter and refcount handling till the end
3186 * of task migration(mem_cgroup_clear_mc()) for performance
3187 * improvement. But we cannot postpone css_get(to) because if
3188 * the process that has been moved to @to does swap-in, the
3189 * refcount of @to might be decreased to 0.
3191 * We are in attach() phase, so the cgroup is guaranteed to be
3192 * alive, so we can just call css_get().
3200 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3201 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3207 #ifdef CONFIG_DEBUG_VM
3208 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
3210 struct page_cgroup
*pc
;
3212 pc
= lookup_page_cgroup(page
);
3214 * Can be NULL while feeding pages into the page allocator for
3215 * the first time, i.e. during boot or memory hotplug;
3216 * or when mem_cgroup_disabled().
3218 if (likely(pc
) && pc
->mem_cgroup
)
3223 bool mem_cgroup_bad_page_check(struct page
*page
)
3225 if (mem_cgroup_disabled())
3228 return lookup_page_cgroup_used(page
) != NULL
;
3231 void mem_cgroup_print_bad_page(struct page
*page
)
3233 struct page_cgroup
*pc
;
3235 pc
= lookup_page_cgroup_used(page
);
3237 pr_alert("pc:%p pc->mem_cgroup:%p\n", pc
, pc
->mem_cgroup
);
3241 static DEFINE_MUTEX(memcg_limit_mutex
);
3243 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3244 unsigned long limit
)
3246 unsigned long curusage
;
3247 unsigned long oldusage
;
3248 bool enlarge
= false;
3253 * For keeping hierarchical_reclaim simple, how long we should retry
3254 * is depends on callers. We set our retry-count to be function
3255 * of # of children which we should visit in this loop.
3257 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
3258 mem_cgroup_count_children(memcg
);
3260 oldusage
= page_counter_read(&memcg
->memory
);
3263 if (signal_pending(current
)) {
3268 mutex_lock(&memcg_limit_mutex
);
3269 if (limit
> memcg
->memsw
.limit
) {
3270 mutex_unlock(&memcg_limit_mutex
);
3274 if (limit
> memcg
->memory
.limit
)
3276 ret
= page_counter_limit(&memcg
->memory
, limit
);
3277 mutex_unlock(&memcg_limit_mutex
);
3282 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
3284 curusage
= page_counter_read(&memcg
->memory
);
3285 /* Usage is reduced ? */
3286 if (curusage
>= oldusage
)
3289 oldusage
= curusage
;
3290 } while (retry_count
);
3292 if (!ret
&& enlarge
)
3293 memcg_oom_recover(memcg
);
3298 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3299 unsigned long limit
)
3301 unsigned long curusage
;
3302 unsigned long oldusage
;
3303 bool enlarge
= false;
3307 /* see mem_cgroup_resize_res_limit */
3308 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
3309 mem_cgroup_count_children(memcg
);
3311 oldusage
= page_counter_read(&memcg
->memsw
);
3314 if (signal_pending(current
)) {
3319 mutex_lock(&memcg_limit_mutex
);
3320 if (limit
< memcg
->memory
.limit
) {
3321 mutex_unlock(&memcg_limit_mutex
);
3325 if (limit
> memcg
->memsw
.limit
)
3327 ret
= page_counter_limit(&memcg
->memsw
, limit
);
3328 mutex_unlock(&memcg_limit_mutex
);
3333 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
3335 curusage
= page_counter_read(&memcg
->memsw
);
3336 /* Usage is reduced ? */
3337 if (curusage
>= oldusage
)
3340 oldusage
= curusage
;
3341 } while (retry_count
);
3343 if (!ret
&& enlarge
)
3344 memcg_oom_recover(memcg
);
3349 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3351 unsigned long *total_scanned
)
3353 unsigned long nr_reclaimed
= 0;
3354 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3355 unsigned long reclaimed
;
3357 struct mem_cgroup_tree_per_zone
*mctz
;
3358 unsigned long excess
;
3359 unsigned long nr_scanned
;
3364 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3366 * This loop can run a while, specially if mem_cgroup's continuously
3367 * keep exceeding their soft limit and putting the system under
3374 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3379 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
3380 gfp_mask
, &nr_scanned
);
3381 nr_reclaimed
+= reclaimed
;
3382 *total_scanned
+= nr_scanned
;
3383 spin_lock_irq(&mctz
->lock
);
3384 __mem_cgroup_remove_exceeded(mz
, mctz
);
3387 * If we failed to reclaim anything from this memory cgroup
3388 * it is time to move on to the next cgroup
3392 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3394 excess
= soft_limit_excess(mz
->memcg
);
3396 * One school of thought says that we should not add
3397 * back the node to the tree if reclaim returns 0.
3398 * But our reclaim could return 0, simply because due
3399 * to priority we are exposing a smaller subset of
3400 * memory to reclaim from. Consider this as a longer
3403 /* If excess == 0, no tree ops */
3404 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3405 spin_unlock_irq(&mctz
->lock
);
3406 css_put(&mz
->memcg
->css
);
3409 * Could not reclaim anything and there are no more
3410 * mem cgroups to try or we seem to be looping without
3411 * reclaiming anything.
3413 if (!nr_reclaimed
&&
3415 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3417 } while (!nr_reclaimed
);
3419 css_put(&next_mz
->memcg
->css
);
3420 return nr_reclaimed
;
3424 * Test whether @memcg has children, dead or alive. Note that this
3425 * function doesn't care whether @memcg has use_hierarchy enabled and
3426 * returns %true if there are child csses according to the cgroup
3427 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3429 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3434 * The lock does not prevent addition or deletion of children, but
3435 * it prevents a new child from being initialized based on this
3436 * parent in css_online(), so it's enough to decide whether
3437 * hierarchically inherited attributes can still be changed or not.
3439 lockdep_assert_held(&memcg_create_mutex
);
3442 ret
= css_next_child(NULL
, &memcg
->css
);
3448 * Reclaims as many pages from the given memcg as possible and moves
3449 * the rest to the parent.
3451 * Caller is responsible for holding css reference for memcg.
3453 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3455 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3457 /* we call try-to-free pages for make this cgroup empty */
3458 lru_add_drain_all();
3459 /* try to free all pages in this cgroup */
3460 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3463 if (signal_pending(current
))
3466 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3470 /* maybe some writeback is necessary */
3471 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3479 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3480 char *buf
, size_t nbytes
,
3483 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3485 if (mem_cgroup_is_root(memcg
))
3487 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3490 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3493 return mem_cgroup_from_css(css
)->use_hierarchy
;
3496 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3497 struct cftype
*cft
, u64 val
)
3500 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3501 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3503 mutex_lock(&memcg_create_mutex
);
3505 if (memcg
->use_hierarchy
== val
)
3509 * If parent's use_hierarchy is set, we can't make any modifications
3510 * in the child subtrees. If it is unset, then the change can
3511 * occur, provided the current cgroup has no children.
3513 * For the root cgroup, parent_mem is NULL, we allow value to be
3514 * set if there are no children.
3516 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3517 (val
== 1 || val
== 0)) {
3518 if (!memcg_has_children(memcg
))
3519 memcg
->use_hierarchy
= val
;
3526 mutex_unlock(&memcg_create_mutex
);
3531 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
3532 enum mem_cgroup_stat_index idx
)
3534 struct mem_cgroup
*iter
;
3537 /* Per-cpu values can be negative, use a signed accumulator */
3538 for_each_mem_cgroup_tree(iter
, memcg
)
3539 val
+= mem_cgroup_read_stat(iter
, idx
);
3541 if (val
< 0) /* race ? */
3546 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3550 if (mem_cgroup_is_root(memcg
)) {
3551 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
3552 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
3554 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
3557 val
= page_counter_read(&memcg
->memory
);
3559 val
= page_counter_read(&memcg
->memsw
);
3561 return val
<< PAGE_SHIFT
;
3572 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3575 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3576 struct page_counter
*counter
;
3578 switch (MEMFILE_TYPE(cft
->private)) {
3580 counter
= &memcg
->memory
;
3583 counter
= &memcg
->memsw
;
3586 counter
= &memcg
->kmem
;
3592 switch (MEMFILE_ATTR(cft
->private)) {
3594 if (counter
== &memcg
->memory
)
3595 return mem_cgroup_usage(memcg
, false);
3596 if (counter
== &memcg
->memsw
)
3597 return mem_cgroup_usage(memcg
, true);
3598 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3600 return (u64
)counter
->limit
* PAGE_SIZE
;
3602 return (u64
)counter
->watermark
* PAGE_SIZE
;
3604 return counter
->failcnt
;
3605 case RES_SOFT_LIMIT
:
3606 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3612 #ifdef CONFIG_MEMCG_KMEM
3613 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
3614 unsigned long nr_pages
)
3619 if (memcg_kmem_is_active(memcg
))
3623 * We are going to allocate memory for data shared by all memory
3624 * cgroups so let's stop accounting here.
3626 memcg_stop_kmem_account();
3629 * For simplicity, we won't allow this to be disabled. It also can't
3630 * be changed if the cgroup has children already, or if tasks had
3633 * If tasks join before we set the limit, a person looking at
3634 * kmem.usage_in_bytes will have no way to determine when it took
3635 * place, which makes the value quite meaningless.
3637 * After it first became limited, changes in the value of the limit are
3638 * of course permitted.
3640 mutex_lock(&memcg_create_mutex
);
3641 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
3642 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
3644 mutex_unlock(&memcg_create_mutex
);
3648 memcg_id
= memcg_alloc_cache_id();
3654 memcg
->kmemcg_id
= memcg_id
;
3655 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3658 * We couldn't have accounted to this cgroup, because it hasn't got the
3659 * active bit set yet, so this should succeed.
3661 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
3664 static_key_slow_inc(&memcg_kmem_enabled_key
);
3666 * Setting the active bit after enabling static branching will
3667 * guarantee no one starts accounting before all call sites are
3670 memcg_kmem_set_active(memcg
);
3672 memcg_resume_kmem_account();
3676 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3677 unsigned long limit
)
3681 mutex_lock(&memcg_limit_mutex
);
3682 if (!memcg_kmem_is_active(memcg
))
3683 ret
= memcg_activate_kmem(memcg
, limit
);
3685 ret
= page_counter_limit(&memcg
->kmem
, limit
);
3686 mutex_unlock(&memcg_limit_mutex
);
3690 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
3693 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
3698 mutex_lock(&memcg_limit_mutex
);
3700 * If the parent cgroup is not kmem-active now, it cannot be activated
3701 * after this point, because it has at least one child already.
3703 if (memcg_kmem_is_active(parent
))
3704 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
3705 mutex_unlock(&memcg_limit_mutex
);
3709 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3710 unsigned long limit
)
3714 #endif /* CONFIG_MEMCG_KMEM */
3717 * The user of this function is...
3720 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3721 char *buf
, size_t nbytes
, loff_t off
)
3723 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3724 unsigned long nr_pages
;
3727 buf
= strstrip(buf
);
3728 ret
= page_counter_memparse(buf
, &nr_pages
);
3732 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3734 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3738 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3740 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3743 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3746 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3750 case RES_SOFT_LIMIT
:
3751 memcg
->soft_limit
= nr_pages
;
3755 return ret
?: nbytes
;
3758 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3759 size_t nbytes
, loff_t off
)
3761 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3762 struct page_counter
*counter
;
3764 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3766 counter
= &memcg
->memory
;
3769 counter
= &memcg
->memsw
;
3772 counter
= &memcg
->kmem
;
3778 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3780 page_counter_reset_watermark(counter
);
3783 counter
->failcnt
= 0;
3792 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3795 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3799 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3800 struct cftype
*cft
, u64 val
)
3802 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3804 if (val
>= (1 << NR_MOVE_TYPE
))
3808 * No kind of locking is needed in here, because ->can_attach() will
3809 * check this value once in the beginning of the process, and then carry
3810 * on with stale data. This means that changes to this value will only
3811 * affect task migrations starting after the change.
3813 memcg
->move_charge_at_immigrate
= val
;
3817 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3818 struct cftype
*cft
, u64 val
)
3825 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3829 unsigned int lru_mask
;
3832 static const struct numa_stat stats
[] = {
3833 { "total", LRU_ALL
},
3834 { "file", LRU_ALL_FILE
},
3835 { "anon", LRU_ALL_ANON
},
3836 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3838 const struct numa_stat
*stat
;
3841 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3843 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3844 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3845 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3846 for_each_node_state(nid
, N_MEMORY
) {
3847 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3849 seq_printf(m
, " N%d=%lu", nid
, nr
);
3854 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3855 struct mem_cgroup
*iter
;
3858 for_each_mem_cgroup_tree(iter
, memcg
)
3859 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3860 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3861 for_each_node_state(nid
, N_MEMORY
) {
3863 for_each_mem_cgroup_tree(iter
, memcg
)
3864 nr
+= mem_cgroup_node_nr_lru_pages(
3865 iter
, nid
, stat
->lru_mask
);
3866 seq_printf(m
, " N%d=%lu", nid
, nr
);
3873 #endif /* CONFIG_NUMA */
3875 static inline void mem_cgroup_lru_names_not_uptodate(void)
3877 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3880 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3882 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3883 unsigned long memory
, memsw
;
3884 struct mem_cgroup
*mi
;
3887 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3888 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3890 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
3891 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3894 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3895 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3896 mem_cgroup_read_events(memcg
, i
));
3898 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3899 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3900 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3902 /* Hierarchical information */
3903 memory
= memsw
= PAGE_COUNTER_MAX
;
3904 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3905 memory
= min(memory
, mi
->memory
.limit
);
3906 memsw
= min(memsw
, mi
->memsw
.limit
);
3908 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3909 (u64
)memory
* PAGE_SIZE
);
3910 if (do_swap_account
)
3911 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3912 (u64
)memsw
* PAGE_SIZE
);
3914 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3917 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3919 for_each_mem_cgroup_tree(mi
, memcg
)
3920 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3921 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
3924 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3925 unsigned long long val
= 0;
3927 for_each_mem_cgroup_tree(mi
, memcg
)
3928 val
+= mem_cgroup_read_events(mi
, i
);
3929 seq_printf(m
, "total_%s %llu\n",
3930 mem_cgroup_events_names
[i
], val
);
3933 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3934 unsigned long long val
= 0;
3936 for_each_mem_cgroup_tree(mi
, memcg
)
3937 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3938 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3941 #ifdef CONFIG_DEBUG_VM
3944 struct mem_cgroup_per_zone
*mz
;
3945 struct zone_reclaim_stat
*rstat
;
3946 unsigned long recent_rotated
[2] = {0, 0};
3947 unsigned long recent_scanned
[2] = {0, 0};
3949 for_each_online_node(nid
)
3950 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3951 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3952 rstat
= &mz
->lruvec
.reclaim_stat
;
3954 recent_rotated
[0] += rstat
->recent_rotated
[0];
3955 recent_rotated
[1] += rstat
->recent_rotated
[1];
3956 recent_scanned
[0] += rstat
->recent_scanned
[0];
3957 recent_scanned
[1] += rstat
->recent_scanned
[1];
3959 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3960 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3961 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3962 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3969 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3972 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3974 return mem_cgroup_swappiness(memcg
);
3977 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3978 struct cftype
*cft
, u64 val
)
3980 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3986 memcg
->swappiness
= val
;
3988 vm_swappiness
= val
;
3993 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3995 struct mem_cgroup_threshold_ary
*t
;
3996 unsigned long usage
;
4001 t
= rcu_dereference(memcg
->thresholds
.primary
);
4003 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4008 usage
= mem_cgroup_usage(memcg
, swap
);
4011 * current_threshold points to threshold just below or equal to usage.
4012 * If it's not true, a threshold was crossed after last
4013 * call of __mem_cgroup_threshold().
4015 i
= t
->current_threshold
;
4018 * Iterate backward over array of thresholds starting from
4019 * current_threshold and check if a threshold is crossed.
4020 * If none of thresholds below usage is crossed, we read
4021 * only one element of the array here.
4023 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4024 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4026 /* i = current_threshold + 1 */
4030 * Iterate forward over array of thresholds starting from
4031 * current_threshold+1 and check if a threshold is crossed.
4032 * If none of thresholds above usage is crossed, we read
4033 * only one element of the array here.
4035 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4036 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4038 /* Update current_threshold */
4039 t
->current_threshold
= i
- 1;
4044 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4047 __mem_cgroup_threshold(memcg
, false);
4048 if (do_swap_account
)
4049 __mem_cgroup_threshold(memcg
, true);
4051 memcg
= parent_mem_cgroup(memcg
);
4055 static int compare_thresholds(const void *a
, const void *b
)
4057 const struct mem_cgroup_threshold
*_a
= a
;
4058 const struct mem_cgroup_threshold
*_b
= b
;
4060 if (_a
->threshold
> _b
->threshold
)
4063 if (_a
->threshold
< _b
->threshold
)
4069 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4071 struct mem_cgroup_eventfd_list
*ev
;
4073 spin_lock(&memcg_oom_lock
);
4075 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4076 eventfd_signal(ev
->eventfd
, 1);
4078 spin_unlock(&memcg_oom_lock
);
4082 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4084 struct mem_cgroup
*iter
;
4086 for_each_mem_cgroup_tree(iter
, memcg
)
4087 mem_cgroup_oom_notify_cb(iter
);
4090 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4091 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4093 struct mem_cgroup_thresholds
*thresholds
;
4094 struct mem_cgroup_threshold_ary
*new;
4095 unsigned long threshold
;
4096 unsigned long usage
;
4099 ret
= page_counter_memparse(args
, &threshold
);
4103 mutex_lock(&memcg
->thresholds_lock
);
4106 thresholds
= &memcg
->thresholds
;
4107 usage
= mem_cgroup_usage(memcg
, false);
4108 } else if (type
== _MEMSWAP
) {
4109 thresholds
= &memcg
->memsw_thresholds
;
4110 usage
= mem_cgroup_usage(memcg
, true);
4114 /* Check if a threshold crossed before adding a new one */
4115 if (thresholds
->primary
)
4116 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4118 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4120 /* Allocate memory for new array of thresholds */
4121 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
4129 /* Copy thresholds (if any) to new array */
4130 if (thresholds
->primary
) {
4131 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4132 sizeof(struct mem_cgroup_threshold
));
4135 /* Add new threshold */
4136 new->entries
[size
- 1].eventfd
= eventfd
;
4137 new->entries
[size
- 1].threshold
= threshold
;
4139 /* Sort thresholds. Registering of new threshold isn't time-critical */
4140 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4141 compare_thresholds
, NULL
);
4143 /* Find current threshold */
4144 new->current_threshold
= -1;
4145 for (i
= 0; i
< size
; i
++) {
4146 if (new->entries
[i
].threshold
<= usage
) {
4148 * new->current_threshold will not be used until
4149 * rcu_assign_pointer(), so it's safe to increment
4152 ++new->current_threshold
;
4157 /* Free old spare buffer and save old primary buffer as spare */
4158 kfree(thresholds
->spare
);
4159 thresholds
->spare
= thresholds
->primary
;
4161 rcu_assign_pointer(thresholds
->primary
, new);
4163 /* To be sure that nobody uses thresholds */
4167 mutex_unlock(&memcg
->thresholds_lock
);
4172 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4173 struct eventfd_ctx
*eventfd
, const char *args
)
4175 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4178 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4179 struct eventfd_ctx
*eventfd
, const char *args
)
4181 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4184 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4185 struct eventfd_ctx
*eventfd
, enum res_type type
)
4187 struct mem_cgroup_thresholds
*thresholds
;
4188 struct mem_cgroup_threshold_ary
*new;
4189 unsigned long usage
;
4192 mutex_lock(&memcg
->thresholds_lock
);
4195 thresholds
= &memcg
->thresholds
;
4196 usage
= mem_cgroup_usage(memcg
, false);
4197 } else if (type
== _MEMSWAP
) {
4198 thresholds
= &memcg
->memsw_thresholds
;
4199 usage
= mem_cgroup_usage(memcg
, true);
4203 if (!thresholds
->primary
)
4206 /* Check if a threshold crossed before removing */
4207 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4209 /* Calculate new number of threshold */
4211 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4212 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4216 new = thresholds
->spare
;
4218 /* Set thresholds array to NULL if we don't have thresholds */
4227 /* Copy thresholds and find current threshold */
4228 new->current_threshold
= -1;
4229 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4230 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4233 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4234 if (new->entries
[j
].threshold
<= usage
) {
4236 * new->current_threshold will not be used
4237 * until rcu_assign_pointer(), so it's safe to increment
4240 ++new->current_threshold
;
4246 /* Swap primary and spare array */
4247 thresholds
->spare
= thresholds
->primary
;
4248 /* If all events are unregistered, free the spare array */
4250 kfree(thresholds
->spare
);
4251 thresholds
->spare
= NULL
;
4254 rcu_assign_pointer(thresholds
->primary
, new);
4256 /* To be sure that nobody uses thresholds */
4259 mutex_unlock(&memcg
->thresholds_lock
);
4262 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4263 struct eventfd_ctx
*eventfd
)
4265 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4268 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4269 struct eventfd_ctx
*eventfd
)
4271 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4274 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4275 struct eventfd_ctx
*eventfd
, const char *args
)
4277 struct mem_cgroup_eventfd_list
*event
;
4279 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4283 spin_lock(&memcg_oom_lock
);
4285 event
->eventfd
= eventfd
;
4286 list_add(&event
->list
, &memcg
->oom_notify
);
4288 /* already in OOM ? */
4289 if (atomic_read(&memcg
->under_oom
))
4290 eventfd_signal(eventfd
, 1);
4291 spin_unlock(&memcg_oom_lock
);
4296 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4297 struct eventfd_ctx
*eventfd
)
4299 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4301 spin_lock(&memcg_oom_lock
);
4303 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4304 if (ev
->eventfd
== eventfd
) {
4305 list_del(&ev
->list
);
4310 spin_unlock(&memcg_oom_lock
);
4313 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4315 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
4317 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4318 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
4322 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4323 struct cftype
*cft
, u64 val
)
4325 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4327 /* cannot set to root cgroup and only 0 and 1 are allowed */
4328 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4331 memcg
->oom_kill_disable
= val
;
4333 memcg_oom_recover(memcg
);
4338 #ifdef CONFIG_MEMCG_KMEM
4339 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4343 memcg
->kmemcg_id
= -1;
4344 ret
= memcg_propagate_kmem(memcg
);
4348 return mem_cgroup_sockets_init(memcg
, ss
);
4351 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4353 mem_cgroup_sockets_destroy(memcg
);
4356 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4361 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4367 * DO NOT USE IN NEW FILES.
4369 * "cgroup.event_control" implementation.
4371 * This is way over-engineered. It tries to support fully configurable
4372 * events for each user. Such level of flexibility is completely
4373 * unnecessary especially in the light of the planned unified hierarchy.
4375 * Please deprecate this and replace with something simpler if at all
4380 * Unregister event and free resources.
4382 * Gets called from workqueue.
4384 static void memcg_event_remove(struct work_struct
*work
)
4386 struct mem_cgroup_event
*event
=
4387 container_of(work
, struct mem_cgroup_event
, remove
);
4388 struct mem_cgroup
*memcg
= event
->memcg
;
4390 remove_wait_queue(event
->wqh
, &event
->wait
);
4392 event
->unregister_event(memcg
, event
->eventfd
);
4394 /* Notify userspace the event is going away. */
4395 eventfd_signal(event
->eventfd
, 1);
4397 eventfd_ctx_put(event
->eventfd
);
4399 css_put(&memcg
->css
);
4403 * Gets called on POLLHUP on eventfd when user closes it.
4405 * Called with wqh->lock held and interrupts disabled.
4407 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
4408 int sync
, void *key
)
4410 struct mem_cgroup_event
*event
=
4411 container_of(wait
, struct mem_cgroup_event
, wait
);
4412 struct mem_cgroup
*memcg
= event
->memcg
;
4413 unsigned long flags
= (unsigned long)key
;
4415 if (flags
& POLLHUP
) {
4417 * If the event has been detached at cgroup removal, we
4418 * can simply return knowing the other side will cleanup
4421 * We can't race against event freeing since the other
4422 * side will require wqh->lock via remove_wait_queue(),
4425 spin_lock(&memcg
->event_list_lock
);
4426 if (!list_empty(&event
->list
)) {
4427 list_del_init(&event
->list
);
4429 * We are in atomic context, but cgroup_event_remove()
4430 * may sleep, so we have to call it in workqueue.
4432 schedule_work(&event
->remove
);
4434 spin_unlock(&memcg
->event_list_lock
);
4440 static void memcg_event_ptable_queue_proc(struct file
*file
,
4441 wait_queue_head_t
*wqh
, poll_table
*pt
)
4443 struct mem_cgroup_event
*event
=
4444 container_of(pt
, struct mem_cgroup_event
, pt
);
4447 add_wait_queue(wqh
, &event
->wait
);
4451 * DO NOT USE IN NEW FILES.
4453 * Parse input and register new cgroup event handler.
4455 * Input must be in format '<event_fd> <control_fd> <args>'.
4456 * Interpretation of args is defined by control file implementation.
4458 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4459 char *buf
, size_t nbytes
, loff_t off
)
4461 struct cgroup_subsys_state
*css
= of_css(of
);
4462 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4463 struct mem_cgroup_event
*event
;
4464 struct cgroup_subsys_state
*cfile_css
;
4465 unsigned int efd
, cfd
;
4472 buf
= strstrip(buf
);
4474 efd
= simple_strtoul(buf
, &endp
, 10);
4479 cfd
= simple_strtoul(buf
, &endp
, 10);
4480 if ((*endp
!= ' ') && (*endp
!= '\0'))
4484 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4488 event
->memcg
= memcg
;
4489 INIT_LIST_HEAD(&event
->list
);
4490 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4491 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4492 INIT_WORK(&event
->remove
, memcg_event_remove
);
4500 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4501 if (IS_ERR(event
->eventfd
)) {
4502 ret
= PTR_ERR(event
->eventfd
);
4509 goto out_put_eventfd
;
4512 /* the process need read permission on control file */
4513 /* AV: shouldn't we check that it's been opened for read instead? */
4514 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4519 * Determine the event callbacks and set them in @event. This used
4520 * to be done via struct cftype but cgroup core no longer knows
4521 * about these events. The following is crude but the whole thing
4522 * is for compatibility anyway.
4524 * DO NOT ADD NEW FILES.
4526 name
= cfile
.file
->f_dentry
->d_name
.name
;
4528 if (!strcmp(name
, "memory.usage_in_bytes")) {
4529 event
->register_event
= mem_cgroup_usage_register_event
;
4530 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4531 } else if (!strcmp(name
, "memory.oom_control")) {
4532 event
->register_event
= mem_cgroup_oom_register_event
;
4533 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4534 } else if (!strcmp(name
, "memory.pressure_level")) {
4535 event
->register_event
= vmpressure_register_event
;
4536 event
->unregister_event
= vmpressure_unregister_event
;
4537 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4538 event
->register_event
= memsw_cgroup_usage_register_event
;
4539 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4546 * Verify @cfile should belong to @css. Also, remaining events are
4547 * automatically removed on cgroup destruction but the removal is
4548 * asynchronous, so take an extra ref on @css.
4550 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_dentry
->d_parent
,
4551 &memory_cgrp_subsys
);
4553 if (IS_ERR(cfile_css
))
4555 if (cfile_css
!= css
) {
4560 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4564 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
4566 spin_lock(&memcg
->event_list_lock
);
4567 list_add(&event
->list
, &memcg
->event_list
);
4568 spin_unlock(&memcg
->event_list_lock
);
4580 eventfd_ctx_put(event
->eventfd
);
4589 static struct cftype mem_cgroup_files
[] = {
4591 .name
= "usage_in_bytes",
4592 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4593 .read_u64
= mem_cgroup_read_u64
,
4596 .name
= "max_usage_in_bytes",
4597 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4598 .write
= mem_cgroup_reset
,
4599 .read_u64
= mem_cgroup_read_u64
,
4602 .name
= "limit_in_bytes",
4603 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4604 .write
= mem_cgroup_write
,
4605 .read_u64
= mem_cgroup_read_u64
,
4608 .name
= "soft_limit_in_bytes",
4609 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4610 .write
= mem_cgroup_write
,
4611 .read_u64
= mem_cgroup_read_u64
,
4615 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4616 .write
= mem_cgroup_reset
,
4617 .read_u64
= mem_cgroup_read_u64
,
4621 .seq_show
= memcg_stat_show
,
4624 .name
= "force_empty",
4625 .write
= mem_cgroup_force_empty_write
,
4628 .name
= "use_hierarchy",
4629 .write_u64
= mem_cgroup_hierarchy_write
,
4630 .read_u64
= mem_cgroup_hierarchy_read
,
4633 .name
= "cgroup.event_control", /* XXX: for compat */
4634 .write
= memcg_write_event_control
,
4635 .flags
= CFTYPE_NO_PREFIX
,
4639 .name
= "swappiness",
4640 .read_u64
= mem_cgroup_swappiness_read
,
4641 .write_u64
= mem_cgroup_swappiness_write
,
4644 .name
= "move_charge_at_immigrate",
4645 .read_u64
= mem_cgroup_move_charge_read
,
4646 .write_u64
= mem_cgroup_move_charge_write
,
4649 .name
= "oom_control",
4650 .seq_show
= mem_cgroup_oom_control_read
,
4651 .write_u64
= mem_cgroup_oom_control_write
,
4652 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4655 .name
= "pressure_level",
4659 .name
= "numa_stat",
4660 .seq_show
= memcg_numa_stat_show
,
4663 #ifdef CONFIG_MEMCG_KMEM
4665 .name
= "kmem.limit_in_bytes",
4666 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4667 .write
= mem_cgroup_write
,
4668 .read_u64
= mem_cgroup_read_u64
,
4671 .name
= "kmem.usage_in_bytes",
4672 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4673 .read_u64
= mem_cgroup_read_u64
,
4676 .name
= "kmem.failcnt",
4677 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4678 .write
= mem_cgroup_reset
,
4679 .read_u64
= mem_cgroup_read_u64
,
4682 .name
= "kmem.max_usage_in_bytes",
4683 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4684 .write
= mem_cgroup_reset
,
4685 .read_u64
= mem_cgroup_read_u64
,
4687 #ifdef CONFIG_SLABINFO
4689 .name
= "kmem.slabinfo",
4690 .seq_start
= slab_start
,
4691 .seq_next
= slab_next
,
4692 .seq_stop
= slab_stop
,
4693 .seq_show
= memcg_slab_show
,
4697 { }, /* terminate */
4700 #ifdef CONFIG_MEMCG_SWAP
4701 static struct cftype memsw_cgroup_files
[] = {
4703 .name
= "memsw.usage_in_bytes",
4704 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
4705 .read_u64
= mem_cgroup_read_u64
,
4708 .name
= "memsw.max_usage_in_bytes",
4709 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
4710 .write
= mem_cgroup_reset
,
4711 .read_u64
= mem_cgroup_read_u64
,
4714 .name
= "memsw.limit_in_bytes",
4715 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
4716 .write
= mem_cgroup_write
,
4717 .read_u64
= mem_cgroup_read_u64
,
4720 .name
= "memsw.failcnt",
4721 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
4722 .write
= mem_cgroup_reset
,
4723 .read_u64
= mem_cgroup_read_u64
,
4725 { }, /* terminate */
4728 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4730 struct mem_cgroup_per_node
*pn
;
4731 struct mem_cgroup_per_zone
*mz
;
4732 int zone
, tmp
= node
;
4734 * This routine is called against possible nodes.
4735 * But it's BUG to call kmalloc() against offline node.
4737 * TODO: this routine can waste much memory for nodes which will
4738 * never be onlined. It's better to use memory hotplug callback
4741 if (!node_state(node
, N_NORMAL_MEMORY
))
4743 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4747 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4748 mz
= &pn
->zoneinfo
[zone
];
4749 lruvec_init(&mz
->lruvec
);
4750 mz
->usage_in_excess
= 0;
4751 mz
->on_tree
= false;
4754 memcg
->nodeinfo
[node
] = pn
;
4758 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4760 kfree(memcg
->nodeinfo
[node
]);
4763 static struct mem_cgroup
*mem_cgroup_alloc(void)
4765 struct mem_cgroup
*memcg
;
4768 size
= sizeof(struct mem_cgroup
);
4769 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4771 memcg
= kzalloc(size
, GFP_KERNEL
);
4775 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4778 spin_lock_init(&memcg
->pcp_counter_lock
);
4787 * At destroying mem_cgroup, references from swap_cgroup can remain.
4788 * (scanning all at force_empty is too costly...)
4790 * Instead of clearing all references at force_empty, we remember
4791 * the number of reference from swap_cgroup and free mem_cgroup when
4792 * it goes down to 0.
4794 * Removal of cgroup itself succeeds regardless of refs from swap.
4797 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4801 mem_cgroup_remove_from_trees(memcg
);
4804 free_mem_cgroup_per_zone_info(memcg
, node
);
4806 free_percpu(memcg
->stat
);
4809 * We need to make sure that (at least for now), the jump label
4810 * destruction code runs outside of the cgroup lock. This is because
4811 * get_online_cpus(), which is called from the static_branch update,
4812 * can't be called inside the cgroup_lock. cpusets are the ones
4813 * enforcing this dependency, so if they ever change, we might as well.
4815 * schedule_work() will guarantee this happens. Be careful if you need
4816 * to move this code around, and make sure it is outside
4819 disarm_static_keys(memcg
);
4824 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4826 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4828 if (!memcg
->memory
.parent
)
4830 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4832 EXPORT_SYMBOL(parent_mem_cgroup
);
4834 static void __init
mem_cgroup_soft_limit_tree_init(void)
4836 struct mem_cgroup_tree_per_node
*rtpn
;
4837 struct mem_cgroup_tree_per_zone
*rtpz
;
4838 int tmp
, node
, zone
;
4840 for_each_node(node
) {
4842 if (!node_state(node
, N_NORMAL_MEMORY
))
4844 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
4847 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
4849 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4850 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
4851 rtpz
->rb_root
= RB_ROOT
;
4852 spin_lock_init(&rtpz
->lock
);
4857 static struct cgroup_subsys_state
* __ref
4858 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4860 struct mem_cgroup
*memcg
;
4861 long error
= -ENOMEM
;
4864 memcg
= mem_cgroup_alloc();
4866 return ERR_PTR(error
);
4869 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4873 if (parent_css
== NULL
) {
4874 root_mem_cgroup
= memcg
;
4875 page_counter_init(&memcg
->memory
, NULL
);
4876 page_counter_init(&memcg
->memsw
, NULL
);
4877 page_counter_init(&memcg
->kmem
, NULL
);
4880 memcg
->last_scanned_node
= MAX_NUMNODES
;
4881 INIT_LIST_HEAD(&memcg
->oom_notify
);
4882 memcg
->move_charge_at_immigrate
= 0;
4883 mutex_init(&memcg
->thresholds_lock
);
4884 spin_lock_init(&memcg
->move_lock
);
4885 vmpressure_init(&memcg
->vmpressure
);
4886 INIT_LIST_HEAD(&memcg
->event_list
);
4887 spin_lock_init(&memcg
->event_list_lock
);
4892 __mem_cgroup_free(memcg
);
4893 return ERR_PTR(error
);
4897 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4899 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4900 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4903 if (css
->id
> MEM_CGROUP_ID_MAX
)
4909 mutex_lock(&memcg_create_mutex
);
4911 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4912 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4913 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4915 if (parent
->use_hierarchy
) {
4916 page_counter_init(&memcg
->memory
, &parent
->memory
);
4917 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4918 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4921 * No need to take a reference to the parent because cgroup
4922 * core guarantees its existence.
4925 page_counter_init(&memcg
->memory
, NULL
);
4926 page_counter_init(&memcg
->memsw
, NULL
);
4927 page_counter_init(&memcg
->kmem
, NULL
);
4929 * Deeper hierachy with use_hierarchy == false doesn't make
4930 * much sense so let cgroup subsystem know about this
4931 * unfortunate state in our controller.
4933 if (parent
!= root_mem_cgroup
)
4934 memory_cgrp_subsys
.broken_hierarchy
= true;
4936 mutex_unlock(&memcg_create_mutex
);
4938 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4943 * Make sure the memcg is initialized: mem_cgroup_iter()
4944 * orders reading memcg->initialized against its callers
4945 * reading the memcg members.
4947 smp_store_release(&memcg
->initialized
, 1);
4952 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4954 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4955 struct mem_cgroup_event
*event
, *tmp
;
4958 * Unregister events and notify userspace.
4959 * Notify userspace about cgroup removing only after rmdir of cgroup
4960 * directory to avoid race between userspace and kernelspace.
4962 spin_lock(&memcg
->event_list_lock
);
4963 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4964 list_del_init(&event
->list
);
4965 schedule_work(&event
->remove
);
4967 spin_unlock(&memcg
->event_list_lock
);
4969 memcg_unregister_all_caches(memcg
);
4970 vmpressure_cleanup(&memcg
->vmpressure
);
4973 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4975 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4977 memcg_destroy_kmem(memcg
);
4978 __mem_cgroup_free(memcg
);
4982 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4983 * @css: the target css
4985 * Reset the states of the mem_cgroup associated with @css. This is
4986 * invoked when the userland requests disabling on the default hierarchy
4987 * but the memcg is pinned through dependency. The memcg should stop
4988 * applying policies and should revert to the vanilla state as it may be
4989 * made visible again.
4991 * The current implementation only resets the essential configurations.
4992 * This needs to be expanded to cover all the visible parts.
4994 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4996 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4998 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4999 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
5000 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
5001 memcg
->soft_limit
= 0;
5005 /* Handlers for move charge at task migration. */
5006 static int mem_cgroup_do_precharge(unsigned long count
)
5010 /* Try a single bulk charge without reclaim first */
5011 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
5013 mc
.precharge
+= count
;
5016 if (ret
== -EINTR
) {
5017 cancel_charge(root_mem_cgroup
, count
);
5021 /* Try charges one by one with reclaim */
5023 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
5025 * In case of failure, any residual charges against
5026 * mc.to will be dropped by mem_cgroup_clear_mc()
5027 * later on. However, cancel any charges that are
5028 * bypassed to root right away or they'll be lost.
5031 cancel_charge(root_mem_cgroup
, 1);
5041 * get_mctgt_type - get target type of moving charge
5042 * @vma: the vma the pte to be checked belongs
5043 * @addr: the address corresponding to the pte to be checked
5044 * @ptent: the pte to be checked
5045 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5048 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5049 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5050 * move charge. if @target is not NULL, the page is stored in target->page
5051 * with extra refcnt got(Callers should handle it).
5052 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5053 * target for charge migration. if @target is not NULL, the entry is stored
5056 * Called with pte lock held.
5063 enum mc_target_type
{
5069 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5070 unsigned long addr
, pte_t ptent
)
5072 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5074 if (!page
|| !page_mapped(page
))
5076 if (PageAnon(page
)) {
5077 /* we don't move shared anon */
5080 } else if (!move_file())
5081 /* we ignore mapcount for file pages */
5083 if (!get_page_unless_zero(page
))
5090 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5091 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5093 struct page
*page
= NULL
;
5094 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5096 if (!move_anon() || non_swap_entry(ent
))
5099 * Because lookup_swap_cache() updates some statistics counter,
5100 * we call find_get_page() with swapper_space directly.
5102 page
= find_get_page(swap_address_space(ent
), ent
.val
);
5103 if (do_swap_account
)
5104 entry
->val
= ent
.val
;
5109 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5110 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5116 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5117 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5119 struct page
*page
= NULL
;
5120 struct address_space
*mapping
;
5123 if (!vma
->vm_file
) /* anonymous vma */
5128 mapping
= vma
->vm_file
->f_mapping
;
5129 if (pte_none(ptent
))
5130 pgoff
= linear_page_index(vma
, addr
);
5131 else /* pte_file(ptent) is true */
5132 pgoff
= pte_to_pgoff(ptent
);
5134 /* page is moved even if it's not RSS of this task(page-faulted). */
5136 /* shmem/tmpfs may report page out on swap: account for that too. */
5137 if (shmem_mapping(mapping
)) {
5138 page
= find_get_entry(mapping
, pgoff
);
5139 if (radix_tree_exceptional_entry(page
)) {
5140 swp_entry_t swp
= radix_to_swp_entry(page
);
5141 if (do_swap_account
)
5143 page
= find_get_page(swap_address_space(swp
), swp
.val
);
5146 page
= find_get_page(mapping
, pgoff
);
5148 page
= find_get_page(mapping
, pgoff
);
5153 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5154 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5156 struct page
*page
= NULL
;
5157 struct page_cgroup
*pc
;
5158 enum mc_target_type ret
= MC_TARGET_NONE
;
5159 swp_entry_t ent
= { .val
= 0 };
5161 if (pte_present(ptent
))
5162 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5163 else if (is_swap_pte(ptent
))
5164 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
5165 else if (pte_none(ptent
) || pte_file(ptent
))
5166 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5168 if (!page
&& !ent
.val
)
5171 pc
= lookup_page_cgroup(page
);
5173 * Do only loose check w/o serialization.
5174 * mem_cgroup_move_account() checks the pc is valid or
5175 * not under LRU exclusion.
5177 if (pc
->mem_cgroup
== mc
.from
) {
5178 ret
= MC_TARGET_PAGE
;
5180 target
->page
= page
;
5182 if (!ret
|| !target
)
5185 /* There is a swap entry and a page doesn't exist or isn't charged */
5186 if (ent
.val
&& !ret
&&
5187 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5188 ret
= MC_TARGET_SWAP
;
5195 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5197 * We don't consider swapping or file mapped pages because THP does not
5198 * support them for now.
5199 * Caller should make sure that pmd_trans_huge(pmd) is true.
5201 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5202 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5204 struct page
*page
= NULL
;
5205 struct page_cgroup
*pc
;
5206 enum mc_target_type ret
= MC_TARGET_NONE
;
5208 page
= pmd_page(pmd
);
5209 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5212 pc
= lookup_page_cgroup(page
);
5213 if (pc
->mem_cgroup
== mc
.from
) {
5214 ret
= MC_TARGET_PAGE
;
5217 target
->page
= page
;
5223 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5224 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5226 return MC_TARGET_NONE
;
5230 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5231 unsigned long addr
, unsigned long end
,
5232 struct mm_walk
*walk
)
5234 struct vm_area_struct
*vma
= walk
->private;
5238 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5239 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5240 mc
.precharge
+= HPAGE_PMD_NR
;
5245 if (pmd_trans_unstable(pmd
))
5247 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5248 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5249 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5250 mc
.precharge
++; /* increment precharge temporarily */
5251 pte_unmap_unlock(pte
- 1, ptl
);
5257 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5259 unsigned long precharge
;
5260 struct vm_area_struct
*vma
;
5262 down_read(&mm
->mmap_sem
);
5263 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5264 struct mm_walk mem_cgroup_count_precharge_walk
= {
5265 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5269 if (is_vm_hugetlb_page(vma
))
5271 walk_page_range(vma
->vm_start
, vma
->vm_end
,
5272 &mem_cgroup_count_precharge_walk
);
5274 up_read(&mm
->mmap_sem
);
5276 precharge
= mc
.precharge
;
5282 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5284 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5286 VM_BUG_ON(mc
.moving_task
);
5287 mc
.moving_task
= current
;
5288 return mem_cgroup_do_precharge(precharge
);
5291 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5292 static void __mem_cgroup_clear_mc(void)
5294 struct mem_cgroup
*from
= mc
.from
;
5295 struct mem_cgroup
*to
= mc
.to
;
5297 /* we must uncharge all the leftover precharges from mc.to */
5299 cancel_charge(mc
.to
, mc
.precharge
);
5303 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5304 * we must uncharge here.
5306 if (mc
.moved_charge
) {
5307 cancel_charge(mc
.from
, mc
.moved_charge
);
5308 mc
.moved_charge
= 0;
5310 /* we must fixup refcnts and charges */
5311 if (mc
.moved_swap
) {
5312 /* uncharge swap account from the old cgroup */
5313 if (!mem_cgroup_is_root(mc
.from
))
5314 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5317 * we charged both to->memory and to->memsw, so we
5318 * should uncharge to->memory.
5320 if (!mem_cgroup_is_root(mc
.to
))
5321 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5323 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
5325 /* we've already done css_get(mc.to) */
5328 memcg_oom_recover(from
);
5329 memcg_oom_recover(to
);
5330 wake_up_all(&mc
.waitq
);
5333 static void mem_cgroup_clear_mc(void)
5336 * we must clear moving_task before waking up waiters at the end of
5339 mc
.moving_task
= NULL
;
5340 __mem_cgroup_clear_mc();
5341 spin_lock(&mc
.lock
);
5344 spin_unlock(&mc
.lock
);
5347 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5348 struct cgroup_taskset
*tset
)
5350 struct task_struct
*p
= cgroup_taskset_first(tset
);
5352 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5353 unsigned long move_charge_at_immigrate
;
5356 * We are now commited to this value whatever it is. Changes in this
5357 * tunable will only affect upcoming migrations, not the current one.
5358 * So we need to save it, and keep it going.
5360 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
5361 if (move_charge_at_immigrate
) {
5362 struct mm_struct
*mm
;
5363 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
5365 VM_BUG_ON(from
== memcg
);
5367 mm
= get_task_mm(p
);
5370 /* We move charges only when we move a owner of the mm */
5371 if (mm
->owner
== p
) {
5374 VM_BUG_ON(mc
.precharge
);
5375 VM_BUG_ON(mc
.moved_charge
);
5376 VM_BUG_ON(mc
.moved_swap
);
5378 spin_lock(&mc
.lock
);
5381 mc
.immigrate_flags
= move_charge_at_immigrate
;
5382 spin_unlock(&mc
.lock
);
5383 /* We set mc.moving_task later */
5385 ret
= mem_cgroup_precharge_mc(mm
);
5387 mem_cgroup_clear_mc();
5394 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5395 struct cgroup_taskset
*tset
)
5398 mem_cgroup_clear_mc();
5401 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5402 unsigned long addr
, unsigned long end
,
5403 struct mm_walk
*walk
)
5406 struct vm_area_struct
*vma
= walk
->private;
5409 enum mc_target_type target_type
;
5410 union mc_target target
;
5412 struct page_cgroup
*pc
;
5415 * We don't take compound_lock() here but no race with splitting thp
5417 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5418 * under splitting, which means there's no concurrent thp split,
5419 * - if another thread runs into split_huge_page() just after we
5420 * entered this if-block, the thread must wait for page table lock
5421 * to be unlocked in __split_huge_page_splitting(), where the main
5422 * part of thp split is not executed yet.
5424 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5425 if (mc
.precharge
< HPAGE_PMD_NR
) {
5429 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5430 if (target_type
== MC_TARGET_PAGE
) {
5432 if (!isolate_lru_page(page
)) {
5433 pc
= lookup_page_cgroup(page
);
5434 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
5435 pc
, mc
.from
, mc
.to
)) {
5436 mc
.precharge
-= HPAGE_PMD_NR
;
5437 mc
.moved_charge
+= HPAGE_PMD_NR
;
5439 putback_lru_page(page
);
5447 if (pmd_trans_unstable(pmd
))
5450 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5451 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5452 pte_t ptent
= *(pte
++);
5458 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5459 case MC_TARGET_PAGE
:
5461 if (isolate_lru_page(page
))
5463 pc
= lookup_page_cgroup(page
);
5464 if (!mem_cgroup_move_account(page
, 1, pc
,
5467 /* we uncharge from mc.from later. */
5470 putback_lru_page(page
);
5471 put
: /* get_mctgt_type() gets the page */
5474 case MC_TARGET_SWAP
:
5476 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5478 /* we fixup refcnts and charges later. */
5486 pte_unmap_unlock(pte
- 1, ptl
);
5491 * We have consumed all precharges we got in can_attach().
5492 * We try charge one by one, but don't do any additional
5493 * charges to mc.to if we have failed in charge once in attach()
5496 ret
= mem_cgroup_do_precharge(1);
5504 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
5506 struct vm_area_struct
*vma
;
5508 lru_add_drain_all();
5510 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5511 * move_lock while we're moving its pages to another memcg.
5512 * Then wait for already started RCU-only updates to finish.
5514 atomic_inc(&mc
.from
->moving_account
);
5517 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5519 * Someone who are holding the mmap_sem might be waiting in
5520 * waitq. So we cancel all extra charges, wake up all waiters,
5521 * and retry. Because we cancel precharges, we might not be able
5522 * to move enough charges, but moving charge is a best-effort
5523 * feature anyway, so it wouldn't be a big problem.
5525 __mem_cgroup_clear_mc();
5529 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5531 struct mm_walk mem_cgroup_move_charge_walk
= {
5532 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5536 if (is_vm_hugetlb_page(vma
))
5538 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
5539 &mem_cgroup_move_charge_walk
);
5542 * means we have consumed all precharges and failed in
5543 * doing additional charge. Just abandon here.
5547 up_read(&mm
->mmap_sem
);
5548 atomic_dec(&mc
.from
->moving_account
);
5551 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5552 struct cgroup_taskset
*tset
)
5554 struct task_struct
*p
= cgroup_taskset_first(tset
);
5555 struct mm_struct
*mm
= get_task_mm(p
);
5559 mem_cgroup_move_charge(mm
);
5563 mem_cgroup_clear_mc();
5565 #else /* !CONFIG_MMU */
5566 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5567 struct cgroup_taskset
*tset
)
5571 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5572 struct cgroup_taskset
*tset
)
5575 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5576 struct cgroup_taskset
*tset
)
5582 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5583 * to verify whether we're attached to the default hierarchy on each mount
5586 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5589 * use_hierarchy is forced on the default hierarchy. cgroup core
5590 * guarantees that @root doesn't have any children, so turning it
5591 * on for the root memcg is enough.
5593 if (cgroup_on_dfl(root_css
->cgroup
))
5594 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
5597 struct cgroup_subsys memory_cgrp_subsys
= {
5598 .css_alloc
= mem_cgroup_css_alloc
,
5599 .css_online
= mem_cgroup_css_online
,
5600 .css_offline
= mem_cgroup_css_offline
,
5601 .css_free
= mem_cgroup_css_free
,
5602 .css_reset
= mem_cgroup_css_reset
,
5603 .can_attach
= mem_cgroup_can_attach
,
5604 .cancel_attach
= mem_cgroup_cancel_attach
,
5605 .attach
= mem_cgroup_move_task
,
5606 .bind
= mem_cgroup_bind
,
5607 .legacy_cftypes
= mem_cgroup_files
,
5611 #ifdef CONFIG_MEMCG_SWAP
5612 static int __init
enable_swap_account(char *s
)
5614 if (!strcmp(s
, "1"))
5615 really_do_swap_account
= 1;
5616 else if (!strcmp(s
, "0"))
5617 really_do_swap_account
= 0;
5620 __setup("swapaccount=", enable_swap_account
);
5622 static void __init
memsw_file_init(void)
5624 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5625 memsw_cgroup_files
));
5628 static void __init
enable_swap_cgroup(void)
5630 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5631 do_swap_account
= 1;
5637 static void __init
enable_swap_cgroup(void)
5642 #ifdef CONFIG_MEMCG_SWAP
5644 * mem_cgroup_swapout - transfer a memsw charge to swap
5645 * @page: page whose memsw charge to transfer
5646 * @entry: swap entry to move the charge to
5648 * Transfer the memsw charge of @page to @entry.
5650 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5652 struct mem_cgroup
*memcg
;
5653 struct page_cgroup
*pc
;
5654 unsigned short oldid
;
5656 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5657 VM_BUG_ON_PAGE(page_count(page
), page
);
5659 if (!do_swap_account
)
5662 pc
= lookup_page_cgroup(page
);
5663 memcg
= pc
->mem_cgroup
;
5665 /* Readahead page, never charged */
5669 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5670 VM_BUG_ON_PAGE(oldid
, page
);
5671 mem_cgroup_swap_statistics(memcg
, true);
5673 pc
->mem_cgroup
= NULL
;
5675 if (!mem_cgroup_is_root(memcg
))
5676 page_counter_uncharge(&memcg
->memory
, 1);
5678 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5679 VM_BUG_ON(!irqs_disabled());
5681 mem_cgroup_charge_statistics(memcg
, page
, -1);
5682 memcg_check_events(memcg
, page
);
5686 * mem_cgroup_uncharge_swap - uncharge a swap entry
5687 * @entry: swap entry to uncharge
5689 * Drop the memsw charge associated with @entry.
5691 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5693 struct mem_cgroup
*memcg
;
5696 if (!do_swap_account
)
5699 id
= swap_cgroup_record(entry
, 0);
5701 memcg
= mem_cgroup_lookup(id
);
5703 if (!mem_cgroup_is_root(memcg
))
5704 page_counter_uncharge(&memcg
->memsw
, 1);
5705 mem_cgroup_swap_statistics(memcg
, false);
5706 css_put(&memcg
->css
);
5713 * mem_cgroup_try_charge - try charging a page
5714 * @page: page to charge
5715 * @mm: mm context of the victim
5716 * @gfp_mask: reclaim mode
5717 * @memcgp: charged memcg return
5719 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5720 * pages according to @gfp_mask if necessary.
5722 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5723 * Otherwise, an error code is returned.
5725 * After page->mapping has been set up, the caller must finalize the
5726 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5727 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5729 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5730 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5732 struct mem_cgroup
*memcg
= NULL
;
5733 unsigned int nr_pages
= 1;
5736 if (mem_cgroup_disabled())
5739 if (PageSwapCache(page
)) {
5740 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
5742 * Every swap fault against a single page tries to charge the
5743 * page, bail as early as possible. shmem_unuse() encounters
5744 * already charged pages, too. The USED bit is protected by
5745 * the page lock, which serializes swap cache removal, which
5746 * in turn serializes uncharging.
5752 if (PageTransHuge(page
)) {
5753 nr_pages
<<= compound_order(page
);
5754 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5757 if (do_swap_account
&& PageSwapCache(page
))
5758 memcg
= try_get_mem_cgroup_from_page(page
);
5760 memcg
= get_mem_cgroup_from_mm(mm
);
5762 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5764 css_put(&memcg
->css
);
5766 if (ret
== -EINTR
) {
5767 memcg
= root_mem_cgroup
;
5776 * mem_cgroup_commit_charge - commit a page charge
5777 * @page: page to charge
5778 * @memcg: memcg to charge the page to
5779 * @lrucare: page might be on LRU already
5781 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5782 * after page->mapping has been set up. This must happen atomically
5783 * as part of the page instantiation, i.e. under the page table lock
5784 * for anonymous pages, under the page lock for page and swap cache.
5786 * In addition, the page must not be on the LRU during the commit, to
5787 * prevent racing with task migration. If it might be, use @lrucare.
5789 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5791 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5794 unsigned int nr_pages
= 1;
5796 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5797 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5799 if (mem_cgroup_disabled())
5802 * Swap faults will attempt to charge the same page multiple
5803 * times. But reuse_swap_page() might have removed the page
5804 * from swapcache already, so we can't check PageSwapCache().
5809 commit_charge(page
, memcg
, lrucare
);
5811 if (PageTransHuge(page
)) {
5812 nr_pages
<<= compound_order(page
);
5813 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5816 local_irq_disable();
5817 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5818 memcg_check_events(memcg
, page
);
5821 if (do_swap_account
&& PageSwapCache(page
)) {
5822 swp_entry_t entry
= { .val
= page_private(page
) };
5824 * The swap entry might not get freed for a long time,
5825 * let's not wait for it. The page already received a
5826 * memory+swap charge, drop the swap entry duplicate.
5828 mem_cgroup_uncharge_swap(entry
);
5833 * mem_cgroup_cancel_charge - cancel a page charge
5834 * @page: page to charge
5835 * @memcg: memcg to charge the page to
5837 * Cancel a charge transaction started by mem_cgroup_try_charge().
5839 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5841 unsigned int nr_pages
= 1;
5843 if (mem_cgroup_disabled())
5846 * Swap faults will attempt to charge the same page multiple
5847 * times. But reuse_swap_page() might have removed the page
5848 * from swapcache already, so we can't check PageSwapCache().
5853 if (PageTransHuge(page
)) {
5854 nr_pages
<<= compound_order(page
);
5855 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5858 cancel_charge(memcg
, nr_pages
);
5861 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5862 unsigned long nr_anon
, unsigned long nr_file
,
5863 unsigned long nr_huge
, struct page
*dummy_page
)
5865 unsigned long nr_pages
= nr_anon
+ nr_file
;
5866 unsigned long flags
;
5868 if (!mem_cgroup_is_root(memcg
)) {
5869 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5870 if (do_swap_account
)
5871 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5872 memcg_oom_recover(memcg
);
5875 local_irq_save(flags
);
5876 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5877 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5878 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5879 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5880 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5881 memcg_check_events(memcg
, dummy_page
);
5882 local_irq_restore(flags
);
5884 if (!mem_cgroup_is_root(memcg
))
5885 css_put_many(&memcg
->css
, nr_pages
);
5888 static void uncharge_list(struct list_head
*page_list
)
5890 struct mem_cgroup
*memcg
= NULL
;
5891 unsigned long nr_anon
= 0;
5892 unsigned long nr_file
= 0;
5893 unsigned long nr_huge
= 0;
5894 unsigned long pgpgout
= 0;
5895 struct list_head
*next
;
5898 next
= page_list
->next
;
5900 unsigned int nr_pages
= 1;
5901 struct page_cgroup
*pc
;
5903 page
= list_entry(next
, struct page
, lru
);
5904 next
= page
->lru
.next
;
5906 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5907 VM_BUG_ON_PAGE(page_count(page
), page
);
5909 pc
= lookup_page_cgroup(page
);
5910 if (!pc
->mem_cgroup
)
5914 * Nobody should be changing or seriously looking at
5915 * pc->mem_cgroup at this point, we have fully
5916 * exclusive access to the page.
5919 if (memcg
!= pc
->mem_cgroup
) {
5921 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5923 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5925 memcg
= pc
->mem_cgroup
;
5928 if (PageTransHuge(page
)) {
5929 nr_pages
<<= compound_order(page
);
5930 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5931 nr_huge
+= nr_pages
;
5935 nr_anon
+= nr_pages
;
5937 nr_file
+= nr_pages
;
5939 pc
->mem_cgroup
= NULL
;
5942 } while (next
!= page_list
);
5945 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5950 * mem_cgroup_uncharge - uncharge a page
5951 * @page: page to uncharge
5953 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5954 * mem_cgroup_commit_charge().
5956 void mem_cgroup_uncharge(struct page
*page
)
5958 struct page_cgroup
*pc
;
5960 if (mem_cgroup_disabled())
5963 /* Don't touch page->lru of any random page, pre-check: */
5964 pc
= lookup_page_cgroup(page
);
5965 if (!pc
->mem_cgroup
)
5968 INIT_LIST_HEAD(&page
->lru
);
5969 uncharge_list(&page
->lru
);
5973 * mem_cgroup_uncharge_list - uncharge a list of page
5974 * @page_list: list of pages to uncharge
5976 * Uncharge a list of pages previously charged with
5977 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5979 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5981 if (mem_cgroup_disabled())
5984 if (!list_empty(page_list
))
5985 uncharge_list(page_list
);
5989 * mem_cgroup_migrate - migrate a charge to another page
5990 * @oldpage: currently charged page
5991 * @newpage: page to transfer the charge to
5992 * @lrucare: both pages might be on the LRU already
5994 * Migrate the charge from @oldpage to @newpage.
5996 * Both pages must be locked, @newpage->mapping must be set up.
5998 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
6001 struct mem_cgroup
*memcg
;
6002 struct page_cgroup
*pc
;
6005 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6006 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6007 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
6008 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
6009 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6010 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6013 if (mem_cgroup_disabled())
6016 /* Page cache replacement: new page already charged? */
6017 pc
= lookup_page_cgroup(newpage
);
6022 * Swapcache readahead pages can get migrated before being
6023 * charged, and migration from compaction can happen to an
6024 * uncharged page when the PFN walker finds a page that
6025 * reclaim just put back on the LRU but has not released yet.
6027 pc
= lookup_page_cgroup(oldpage
);
6028 memcg
= pc
->mem_cgroup
;
6033 lock_page_lru(oldpage
, &isolated
);
6035 pc
->mem_cgroup
= NULL
;
6038 unlock_page_lru(oldpage
, isolated
);
6040 commit_charge(newpage
, memcg
, lrucare
);
6044 * subsys_initcall() for memory controller.
6046 * Some parts like hotcpu_notifier() have to be initialized from this context
6047 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6048 * everything that doesn't depend on a specific mem_cgroup structure should
6049 * be initialized from here.
6051 static int __init
mem_cgroup_init(void)
6053 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6054 enable_swap_cgroup();
6055 mem_cgroup_soft_limit_tree_init();
6059 subsys_initcall(mem_cgroup_init
);