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/swap_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 lruvec
*lruvec
;
1279 if (mem_cgroup_disabled()) {
1280 lruvec
= &zone
->lruvec
;
1284 memcg
= page
->mem_cgroup
;
1286 * Swapcache readahead pages are added to the LRU - and
1287 * possibly migrated - before they are charged.
1290 memcg
= root_mem_cgroup
;
1292 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1293 lruvec
= &mz
->lruvec
;
1296 * Since a node can be onlined after the mem_cgroup was created,
1297 * we have to be prepared to initialize lruvec->zone here;
1298 * and if offlined then reonlined, we need to reinitialize it.
1300 if (unlikely(lruvec
->zone
!= zone
))
1301 lruvec
->zone
= zone
;
1306 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1307 * @lruvec: mem_cgroup per zone lru vector
1308 * @lru: index of lru list the page is sitting on
1309 * @nr_pages: positive when adding or negative when removing
1311 * This function must be called when a page is added to or removed from an
1314 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1317 struct mem_cgroup_per_zone
*mz
;
1318 unsigned long *lru_size
;
1320 if (mem_cgroup_disabled())
1323 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1324 lru_size
= mz
->lru_size
+ lru
;
1325 *lru_size
+= nr_pages
;
1326 VM_BUG_ON((long)(*lru_size
) < 0);
1329 bool mem_cgroup_is_descendant(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
)
1333 if (!root
->use_hierarchy
)
1335 return cgroup_is_descendant(memcg
->css
.cgroup
, root
->css
.cgroup
);
1338 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1340 struct mem_cgroup
*task_memcg
;
1341 struct task_struct
*p
;
1344 p
= find_lock_task_mm(task
);
1346 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1350 * All threads may have already detached their mm's, but the oom
1351 * killer still needs to detect if they have already been oom
1352 * killed to prevent needlessly killing additional tasks.
1355 task_memcg
= mem_cgroup_from_task(task
);
1356 css_get(&task_memcg
->css
);
1359 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1360 css_put(&task_memcg
->css
);
1364 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1366 unsigned long inactive_ratio
;
1367 unsigned long inactive
;
1368 unsigned long active
;
1371 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1372 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1374 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1376 inactive_ratio
= int_sqrt(10 * gb
);
1380 return inactive
* inactive_ratio
< active
;
1383 #define mem_cgroup_from_counter(counter, member) \
1384 container_of(counter, struct mem_cgroup, member)
1387 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1388 * @memcg: the memory cgroup
1390 * Returns the maximum amount of memory @mem can be charged with, in
1393 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1395 unsigned long margin
= 0;
1396 unsigned long count
;
1397 unsigned long limit
;
1399 count
= page_counter_read(&memcg
->memory
);
1400 limit
= ACCESS_ONCE(memcg
->memory
.limit
);
1402 margin
= limit
- count
;
1404 if (do_swap_account
) {
1405 count
= page_counter_read(&memcg
->memsw
);
1406 limit
= ACCESS_ONCE(memcg
->memsw
.limit
);
1408 margin
= min(margin
, limit
- count
);
1414 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1417 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1418 return vm_swappiness
;
1420 return memcg
->swappiness
;
1424 * A routine for checking "mem" is under move_account() or not.
1426 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1427 * moving cgroups. This is for waiting at high-memory pressure
1430 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1432 struct mem_cgroup
*from
;
1433 struct mem_cgroup
*to
;
1436 * Unlike task_move routines, we access mc.to, mc.from not under
1437 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1439 spin_lock(&mc
.lock
);
1445 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1446 mem_cgroup_is_descendant(to
, memcg
);
1448 spin_unlock(&mc
.lock
);
1452 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1454 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1455 if (mem_cgroup_under_move(memcg
)) {
1457 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1458 /* moving charge context might have finished. */
1461 finish_wait(&mc
.waitq
, &wait
);
1468 #define K(x) ((x) << (PAGE_SHIFT-10))
1470 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1471 * @memcg: The memory cgroup that went over limit
1472 * @p: Task that is going to be killed
1474 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1477 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1479 /* oom_info_lock ensures that parallel ooms do not interleave */
1480 static DEFINE_MUTEX(oom_info_lock
);
1481 struct mem_cgroup
*iter
;
1487 mutex_lock(&oom_info_lock
);
1490 pr_info("Task in ");
1491 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1492 pr_info(" killed as a result of limit of ");
1493 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1498 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1499 K((u64
)page_counter_read(&memcg
->memory
)),
1500 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1501 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1502 K((u64
)page_counter_read(&memcg
->memsw
)),
1503 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1504 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1505 K((u64
)page_counter_read(&memcg
->kmem
)),
1506 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1508 for_each_mem_cgroup_tree(iter
, memcg
) {
1509 pr_info("Memory cgroup stats for ");
1510 pr_cont_cgroup_path(iter
->css
.cgroup
);
1513 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1514 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1516 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1517 K(mem_cgroup_read_stat(iter
, i
)));
1520 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1521 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1522 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1526 mutex_unlock(&oom_info_lock
);
1530 * This function returns the number of memcg under hierarchy tree. Returns
1531 * 1(self count) if no children.
1533 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1536 struct mem_cgroup
*iter
;
1538 for_each_mem_cgroup_tree(iter
, memcg
)
1544 * Return the memory (and swap, if configured) limit for a memcg.
1546 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1548 unsigned long limit
;
1550 limit
= memcg
->memory
.limit
;
1551 if (mem_cgroup_swappiness(memcg
)) {
1552 unsigned long memsw_limit
;
1554 memsw_limit
= memcg
->memsw
.limit
;
1555 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1560 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1563 struct mem_cgroup
*iter
;
1564 unsigned long chosen_points
= 0;
1565 unsigned long totalpages
;
1566 unsigned int points
= 0;
1567 struct task_struct
*chosen
= NULL
;
1570 * If current has a pending SIGKILL or is exiting, then automatically
1571 * select it. The goal is to allow it to allocate so that it may
1572 * quickly exit and free its memory.
1574 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1575 set_thread_flag(TIF_MEMDIE
);
1579 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1580 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1581 for_each_mem_cgroup_tree(iter
, memcg
) {
1582 struct css_task_iter it
;
1583 struct task_struct
*task
;
1585 css_task_iter_start(&iter
->css
, &it
);
1586 while ((task
= css_task_iter_next(&it
))) {
1587 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1589 case OOM_SCAN_SELECT
:
1591 put_task_struct(chosen
);
1593 chosen_points
= ULONG_MAX
;
1594 get_task_struct(chosen
);
1596 case OOM_SCAN_CONTINUE
:
1598 case OOM_SCAN_ABORT
:
1599 css_task_iter_end(&it
);
1600 mem_cgroup_iter_break(memcg
, iter
);
1602 put_task_struct(chosen
);
1607 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1608 if (!points
|| points
< chosen_points
)
1610 /* Prefer thread group leaders for display purposes */
1611 if (points
== chosen_points
&&
1612 thread_group_leader(chosen
))
1616 put_task_struct(chosen
);
1618 chosen_points
= points
;
1619 get_task_struct(chosen
);
1621 css_task_iter_end(&it
);
1626 points
= chosen_points
* 1000 / totalpages
;
1627 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1628 NULL
, "Memory cgroup out of memory");
1632 * test_mem_cgroup_node_reclaimable
1633 * @memcg: the target memcg
1634 * @nid: the node ID to be checked.
1635 * @noswap : specify true here if the user wants flle only information.
1637 * This function returns whether the specified memcg contains any
1638 * reclaimable pages on a node. Returns true if there are any reclaimable
1639 * pages in the node.
1641 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1642 int nid
, bool noswap
)
1644 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1646 if (noswap
|| !total_swap_pages
)
1648 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1653 #if MAX_NUMNODES > 1
1656 * Always updating the nodemask is not very good - even if we have an empty
1657 * list or the wrong list here, we can start from some node and traverse all
1658 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1661 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1665 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1666 * pagein/pageout changes since the last update.
1668 if (!atomic_read(&memcg
->numainfo_events
))
1670 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1673 /* make a nodemask where this memcg uses memory from */
1674 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1676 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1678 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1679 node_clear(nid
, memcg
->scan_nodes
);
1682 atomic_set(&memcg
->numainfo_events
, 0);
1683 atomic_set(&memcg
->numainfo_updating
, 0);
1687 * Selecting a node where we start reclaim from. Because what we need is just
1688 * reducing usage counter, start from anywhere is O,K. Considering
1689 * memory reclaim from current node, there are pros. and cons.
1691 * Freeing memory from current node means freeing memory from a node which
1692 * we'll use or we've used. So, it may make LRU bad. And if several threads
1693 * hit limits, it will see a contention on a node. But freeing from remote
1694 * node means more costs for memory reclaim because of memory latency.
1696 * Now, we use round-robin. Better algorithm is welcomed.
1698 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1702 mem_cgroup_may_update_nodemask(memcg
);
1703 node
= memcg
->last_scanned_node
;
1705 node
= next_node(node
, memcg
->scan_nodes
);
1706 if (node
== MAX_NUMNODES
)
1707 node
= first_node(memcg
->scan_nodes
);
1709 * We call this when we hit limit, not when pages are added to LRU.
1710 * No LRU may hold pages because all pages are UNEVICTABLE or
1711 * memcg is too small and all pages are not on LRU. In that case,
1712 * we use curret node.
1714 if (unlikely(node
== MAX_NUMNODES
))
1715 node
= numa_node_id();
1717 memcg
->last_scanned_node
= node
;
1721 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1727 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1730 unsigned long *total_scanned
)
1732 struct mem_cgroup
*victim
= NULL
;
1735 unsigned long excess
;
1736 unsigned long nr_scanned
;
1737 struct mem_cgroup_reclaim_cookie reclaim
= {
1742 excess
= soft_limit_excess(root_memcg
);
1745 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1750 * If we have not been able to reclaim
1751 * anything, it might because there are
1752 * no reclaimable pages under this hierarchy
1757 * We want to do more targeted reclaim.
1758 * excess >> 2 is not to excessive so as to
1759 * reclaim too much, nor too less that we keep
1760 * coming back to reclaim from this cgroup
1762 if (total
>= (excess
>> 2) ||
1763 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1768 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1770 *total_scanned
+= nr_scanned
;
1771 if (!soft_limit_excess(root_memcg
))
1774 mem_cgroup_iter_break(root_memcg
, victim
);
1778 #ifdef CONFIG_LOCKDEP
1779 static struct lockdep_map memcg_oom_lock_dep_map
= {
1780 .name
= "memcg_oom_lock",
1784 static DEFINE_SPINLOCK(memcg_oom_lock
);
1787 * Check OOM-Killer is already running under our hierarchy.
1788 * If someone is running, return false.
1790 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1792 struct mem_cgroup
*iter
, *failed
= NULL
;
1794 spin_lock(&memcg_oom_lock
);
1796 for_each_mem_cgroup_tree(iter
, memcg
) {
1797 if (iter
->oom_lock
) {
1799 * this subtree of our hierarchy is already locked
1800 * so we cannot give a lock.
1803 mem_cgroup_iter_break(memcg
, iter
);
1806 iter
->oom_lock
= true;
1811 * OK, we failed to lock the whole subtree so we have
1812 * to clean up what we set up to the failing subtree
1814 for_each_mem_cgroup_tree(iter
, memcg
) {
1815 if (iter
== failed
) {
1816 mem_cgroup_iter_break(memcg
, iter
);
1819 iter
->oom_lock
= false;
1822 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1824 spin_unlock(&memcg_oom_lock
);
1829 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1831 struct mem_cgroup
*iter
;
1833 spin_lock(&memcg_oom_lock
);
1834 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1835 for_each_mem_cgroup_tree(iter
, memcg
)
1836 iter
->oom_lock
= false;
1837 spin_unlock(&memcg_oom_lock
);
1840 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1842 struct mem_cgroup
*iter
;
1844 for_each_mem_cgroup_tree(iter
, memcg
)
1845 atomic_inc(&iter
->under_oom
);
1848 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1850 struct mem_cgroup
*iter
;
1853 * When a new child is created while the hierarchy is under oom,
1854 * mem_cgroup_oom_lock() may not be called. We have to use
1855 * atomic_add_unless() here.
1857 for_each_mem_cgroup_tree(iter
, memcg
)
1858 atomic_add_unless(&iter
->under_oom
, -1, 0);
1861 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1863 struct oom_wait_info
{
1864 struct mem_cgroup
*memcg
;
1868 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1869 unsigned mode
, int sync
, void *arg
)
1871 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1872 struct mem_cgroup
*oom_wait_memcg
;
1873 struct oom_wait_info
*oom_wait_info
;
1875 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1876 oom_wait_memcg
= oom_wait_info
->memcg
;
1878 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1879 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1881 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1884 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
1886 atomic_inc(&memcg
->oom_wakeups
);
1887 /* for filtering, pass "memcg" as argument. */
1888 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1891 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1893 if (memcg
&& atomic_read(&memcg
->under_oom
))
1894 memcg_wakeup_oom(memcg
);
1897 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1899 if (!current
->memcg_oom
.may_oom
)
1902 * We are in the middle of the charge context here, so we
1903 * don't want to block when potentially sitting on a callstack
1904 * that holds all kinds of filesystem and mm locks.
1906 * Also, the caller may handle a failed allocation gracefully
1907 * (like optional page cache readahead) and so an OOM killer
1908 * invocation might not even be necessary.
1910 * That's why we don't do anything here except remember the
1911 * OOM context and then deal with it at the end of the page
1912 * fault when the stack is unwound, the locks are released,
1913 * and when we know whether the fault was overall successful.
1915 css_get(&memcg
->css
);
1916 current
->memcg_oom
.memcg
= memcg
;
1917 current
->memcg_oom
.gfp_mask
= mask
;
1918 current
->memcg_oom
.order
= order
;
1922 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1923 * @handle: actually kill/wait or just clean up the OOM state
1925 * This has to be called at the end of a page fault if the memcg OOM
1926 * handler was enabled.
1928 * Memcg supports userspace OOM handling where failed allocations must
1929 * sleep on a waitqueue until the userspace task resolves the
1930 * situation. Sleeping directly in the charge context with all kinds
1931 * of locks held is not a good idea, instead we remember an OOM state
1932 * in the task and mem_cgroup_oom_synchronize() has to be called at
1933 * the end of the page fault to complete the OOM handling.
1935 * Returns %true if an ongoing memcg OOM situation was detected and
1936 * completed, %false otherwise.
1938 bool mem_cgroup_oom_synchronize(bool handle
)
1940 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
1941 struct oom_wait_info owait
;
1944 /* OOM is global, do not handle */
1951 owait
.memcg
= memcg
;
1952 owait
.wait
.flags
= 0;
1953 owait
.wait
.func
= memcg_oom_wake_function
;
1954 owait
.wait
.private = current
;
1955 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1957 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1958 mem_cgroup_mark_under_oom(memcg
);
1960 locked
= mem_cgroup_oom_trylock(memcg
);
1963 mem_cgroup_oom_notify(memcg
);
1965 if (locked
&& !memcg
->oom_kill_disable
) {
1966 mem_cgroup_unmark_under_oom(memcg
);
1967 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1968 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
1969 current
->memcg_oom
.order
);
1972 mem_cgroup_unmark_under_oom(memcg
);
1973 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1977 mem_cgroup_oom_unlock(memcg
);
1979 * There is no guarantee that an OOM-lock contender
1980 * sees the wakeups triggered by the OOM kill
1981 * uncharges. Wake any sleepers explicitely.
1983 memcg_oom_recover(memcg
);
1986 current
->memcg_oom
.memcg
= NULL
;
1987 css_put(&memcg
->css
);
1992 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1993 * @page: page that is going to change accounted state
1994 * @locked: &memcg->move_lock slowpath was taken
1995 * @flags: IRQ-state flags for &memcg->move_lock
1997 * This function must mark the beginning of an accounted page state
1998 * change to prevent double accounting when the page is concurrently
1999 * being moved to another memcg:
2001 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
2002 * if (TestClearPageState(page))
2003 * mem_cgroup_update_page_stat(memcg, state, -1);
2004 * mem_cgroup_end_page_stat(memcg, locked, flags);
2006 * The RCU lock is held throughout the transaction. The fast path can
2007 * get away without acquiring the memcg->move_lock (@locked is false)
2008 * because page moving starts with an RCU grace period.
2010 * The RCU lock also protects the memcg from being freed when the page
2011 * state that is going to change is the only thing preventing the page
2012 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2013 * which allows migration to go ahead and uncharge the page before the
2014 * account transaction might be complete.
2016 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
,
2018 unsigned long *flags
)
2020 struct mem_cgroup
*memcg
;
2024 if (mem_cgroup_disabled())
2027 memcg
= page
->mem_cgroup
;
2028 if (unlikely(!memcg
))
2032 if (atomic_read(&memcg
->moving_account
) <= 0)
2035 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
2036 if (memcg
!= page
->mem_cgroup
) {
2037 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
2046 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2047 * @memcg: the memcg that was accounted against
2048 * @locked: value received from mem_cgroup_begin_page_stat()
2049 * @flags: value received from mem_cgroup_begin_page_stat()
2051 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
, bool *locked
,
2052 unsigned long *flags
)
2054 if (memcg
&& *locked
)
2055 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
2061 * mem_cgroup_update_page_stat - update page state statistics
2062 * @memcg: memcg to account against
2063 * @idx: page state item to account
2064 * @val: number of pages (positive or negative)
2066 * See mem_cgroup_begin_page_stat() for locking requirements.
2068 void mem_cgroup_update_page_stat(struct mem_cgroup
*memcg
,
2069 enum mem_cgroup_stat_index idx
, int val
)
2071 VM_BUG_ON(!rcu_read_lock_held());
2074 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2078 * size of first charge trial. "32" comes from vmscan.c's magic value.
2079 * TODO: maybe necessary to use big numbers in big irons.
2081 #define CHARGE_BATCH 32U
2082 struct memcg_stock_pcp
{
2083 struct mem_cgroup
*cached
; /* this never be root cgroup */
2084 unsigned int nr_pages
;
2085 struct work_struct work
;
2086 unsigned long flags
;
2087 #define FLUSHING_CACHED_CHARGE 0
2089 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2090 static DEFINE_MUTEX(percpu_charge_mutex
);
2093 * consume_stock: Try to consume stocked charge on this cpu.
2094 * @memcg: memcg to consume from.
2095 * @nr_pages: how many pages to charge.
2097 * The charges will only happen if @memcg matches the current cpu's memcg
2098 * stock, and at least @nr_pages are available in that stock. Failure to
2099 * service an allocation will refill the stock.
2101 * returns true if successful, false otherwise.
2103 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2105 struct memcg_stock_pcp
*stock
;
2108 if (nr_pages
> CHARGE_BATCH
)
2111 stock
= &get_cpu_var(memcg_stock
);
2112 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2113 stock
->nr_pages
-= nr_pages
;
2116 put_cpu_var(memcg_stock
);
2121 * Returns stocks cached in percpu and reset cached information.
2123 static void drain_stock(struct memcg_stock_pcp
*stock
)
2125 struct mem_cgroup
*old
= stock
->cached
;
2127 if (stock
->nr_pages
) {
2128 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2129 if (do_swap_account
)
2130 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2131 css_put_many(&old
->css
, stock
->nr_pages
);
2132 stock
->nr_pages
= 0;
2134 stock
->cached
= NULL
;
2138 * This must be called under preempt disabled or must be called by
2139 * a thread which is pinned to local cpu.
2141 static void drain_local_stock(struct work_struct
*dummy
)
2143 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2145 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2148 static void __init
memcg_stock_init(void)
2152 for_each_possible_cpu(cpu
) {
2153 struct memcg_stock_pcp
*stock
=
2154 &per_cpu(memcg_stock
, cpu
);
2155 INIT_WORK(&stock
->work
, drain_local_stock
);
2160 * Cache charges(val) to local per_cpu area.
2161 * This will be consumed by consume_stock() function, later.
2163 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2165 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2167 if (stock
->cached
!= memcg
) { /* reset if necessary */
2169 stock
->cached
= memcg
;
2171 stock
->nr_pages
+= nr_pages
;
2172 put_cpu_var(memcg_stock
);
2176 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2177 * of the hierarchy under it.
2179 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2183 /* If someone's already draining, avoid adding running more workers. */
2184 if (!mutex_trylock(&percpu_charge_mutex
))
2186 /* Notify other cpus that system-wide "drain" is running */
2189 for_each_online_cpu(cpu
) {
2190 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2191 struct mem_cgroup
*memcg
;
2193 memcg
= stock
->cached
;
2194 if (!memcg
|| !stock
->nr_pages
)
2196 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
2198 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2200 drain_local_stock(&stock
->work
);
2202 schedule_work_on(cpu
, &stock
->work
);
2207 mutex_unlock(&percpu_charge_mutex
);
2211 * This function drains percpu counter value from DEAD cpu and
2212 * move it to local cpu. Note that this function can be preempted.
2214 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2218 spin_lock(&memcg
->pcp_counter_lock
);
2219 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2220 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2222 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2223 memcg
->nocpu_base
.count
[i
] += x
;
2225 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2226 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2228 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2229 memcg
->nocpu_base
.events
[i
] += x
;
2231 spin_unlock(&memcg
->pcp_counter_lock
);
2234 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2235 unsigned long action
,
2238 int cpu
= (unsigned long)hcpu
;
2239 struct memcg_stock_pcp
*stock
;
2240 struct mem_cgroup
*iter
;
2242 if (action
== CPU_ONLINE
)
2245 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2248 for_each_mem_cgroup(iter
)
2249 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2251 stock
= &per_cpu(memcg_stock
, cpu
);
2256 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2257 unsigned int nr_pages
)
2259 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2260 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2261 struct mem_cgroup
*mem_over_limit
;
2262 struct page_counter
*counter
;
2263 unsigned long nr_reclaimed
;
2264 bool may_swap
= true;
2265 bool drained
= false;
2268 if (mem_cgroup_is_root(memcg
))
2271 if (consume_stock(memcg
, nr_pages
))
2274 if (!do_swap_account
||
2275 !page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2276 if (!page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2278 if (do_swap_account
)
2279 page_counter_uncharge(&memcg
->memsw
, batch
);
2280 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2282 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2286 if (batch
> nr_pages
) {
2292 * Unlike in global OOM situations, memcg is not in a physical
2293 * memory shortage. Allow dying and OOM-killed tasks to
2294 * bypass the last charges so that they can exit quickly and
2295 * free their memory.
2297 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2298 fatal_signal_pending(current
) ||
2299 current
->flags
& PF_EXITING
))
2302 if (unlikely(task_in_memcg_oom(current
)))
2305 if (!(gfp_mask
& __GFP_WAIT
))
2308 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2309 gfp_mask
, may_swap
);
2311 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2315 drain_all_stock(mem_over_limit
);
2320 if (gfp_mask
& __GFP_NORETRY
)
2323 * Even though the limit is exceeded at this point, reclaim
2324 * may have been able to free some pages. Retry the charge
2325 * before killing the task.
2327 * Only for regular pages, though: huge pages are rather
2328 * unlikely to succeed so close to the limit, and we fall back
2329 * to regular pages anyway in case of failure.
2331 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2334 * At task move, charge accounts can be doubly counted. So, it's
2335 * better to wait until the end of task_move if something is going on.
2337 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2343 if (gfp_mask
& __GFP_NOFAIL
)
2346 if (fatal_signal_pending(current
))
2349 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2351 if (!(gfp_mask
& __GFP_NOFAIL
))
2357 css_get_many(&memcg
->css
, batch
);
2358 if (batch
> nr_pages
)
2359 refill_stock(memcg
, batch
- nr_pages
);
2364 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2366 if (mem_cgroup_is_root(memcg
))
2369 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2370 if (do_swap_account
)
2371 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2373 css_put_many(&memcg
->css
, nr_pages
);
2377 * A helper function to get mem_cgroup from ID. must be called under
2378 * rcu_read_lock(). The caller is responsible for calling
2379 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2380 * refcnt from swap can be called against removed memcg.)
2382 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2384 /* ID 0 is unused ID */
2387 return mem_cgroup_from_id(id
);
2391 * try_get_mem_cgroup_from_page - look up page's memcg association
2394 * Look up, get a css reference, and return the memcg that owns @page.
2396 * The page must be locked to prevent racing with swap-in and page
2397 * cache charges. If coming from an unlocked page table, the caller
2398 * must ensure the page is on the LRU or this can race with charging.
2400 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2402 struct mem_cgroup
*memcg
;
2406 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2408 memcg
= page
->mem_cgroup
;
2410 if (!css_tryget_online(&memcg
->css
))
2412 } else if (PageSwapCache(page
)) {
2413 ent
.val
= page_private(page
);
2414 id
= lookup_swap_cgroup_id(ent
);
2416 memcg
= mem_cgroup_lookup(id
);
2417 if (memcg
&& !css_tryget_online(&memcg
->css
))
2424 static void lock_page_lru(struct page
*page
, int *isolated
)
2426 struct zone
*zone
= page_zone(page
);
2428 spin_lock_irq(&zone
->lru_lock
);
2429 if (PageLRU(page
)) {
2430 struct lruvec
*lruvec
;
2432 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2434 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2440 static void unlock_page_lru(struct page
*page
, int isolated
)
2442 struct zone
*zone
= page_zone(page
);
2445 struct lruvec
*lruvec
;
2447 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2448 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2450 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2452 spin_unlock_irq(&zone
->lru_lock
);
2455 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2460 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2463 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2464 * may already be on some other mem_cgroup's LRU. Take care of it.
2467 lock_page_lru(page
, &isolated
);
2470 * Nobody should be changing or seriously looking at
2471 * page->mem_cgroup at this point:
2473 * - the page is uncharged
2475 * - the page is off-LRU
2477 * - an anonymous fault has exclusive page access, except for
2478 * a locked page table
2480 * - a page cache insertion, a swapin fault, or a migration
2481 * have the page locked
2483 page
->mem_cgroup
= memcg
;
2486 unlock_page_lru(page
, isolated
);
2489 #ifdef CONFIG_MEMCG_KMEM
2491 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2492 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2494 static DEFINE_MUTEX(memcg_slab_mutex
);
2497 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2498 * in the memcg_cache_params struct.
2500 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2502 struct kmem_cache
*cachep
;
2504 VM_BUG_ON(p
->is_root_cache
);
2505 cachep
= p
->root_cache
;
2506 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2509 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
,
2510 unsigned long nr_pages
)
2512 struct page_counter
*counter
;
2515 ret
= page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
);
2519 ret
= try_charge(memcg
, gfp
, nr_pages
);
2520 if (ret
== -EINTR
) {
2522 * try_charge() chose to bypass to root due to OOM kill or
2523 * fatal signal. Since our only options are to either fail
2524 * the allocation or charge it to this cgroup, do it as a
2525 * temporary condition. But we can't fail. From a kmem/slab
2526 * perspective, the cache has already been selected, by
2527 * mem_cgroup_kmem_get_cache(), so it is too late to change
2530 * This condition will only trigger if the task entered
2531 * memcg_charge_kmem in a sane state, but was OOM-killed
2532 * during try_charge() above. Tasks that were already dying
2533 * when the allocation triggers should have been already
2534 * directed to the root cgroup in memcontrol.h
2536 page_counter_charge(&memcg
->memory
, nr_pages
);
2537 if (do_swap_account
)
2538 page_counter_charge(&memcg
->memsw
, nr_pages
);
2539 css_get_many(&memcg
->css
, nr_pages
);
2542 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2547 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
,
2548 unsigned long nr_pages
)
2550 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2551 if (do_swap_account
)
2552 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2554 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2556 css_put_many(&memcg
->css
, nr_pages
);
2560 * helper for acessing a memcg's index. It will be used as an index in the
2561 * child cache array in kmem_cache, and also to derive its name. This function
2562 * will return -1 when this is not a kmem-limited memcg.
2564 int memcg_cache_id(struct mem_cgroup
*memcg
)
2566 return memcg
? memcg
->kmemcg_id
: -1;
2569 static int memcg_alloc_cache_id(void)
2574 id
= ida_simple_get(&kmem_limited_groups
,
2575 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2579 if (id
< memcg_limited_groups_array_size
)
2583 * There's no space for the new id in memcg_caches arrays,
2584 * so we have to grow them.
2587 size
= 2 * (id
+ 1);
2588 if (size
< MEMCG_CACHES_MIN_SIZE
)
2589 size
= MEMCG_CACHES_MIN_SIZE
;
2590 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2591 size
= MEMCG_CACHES_MAX_SIZE
;
2593 mutex_lock(&memcg_slab_mutex
);
2594 err
= memcg_update_all_caches(size
);
2595 mutex_unlock(&memcg_slab_mutex
);
2598 ida_simple_remove(&kmem_limited_groups
, id
);
2604 static void memcg_free_cache_id(int id
)
2606 ida_simple_remove(&kmem_limited_groups
, id
);
2610 * We should update the current array size iff all caches updates succeed. This
2611 * can only be done from the slab side. The slab mutex needs to be held when
2614 void memcg_update_array_size(int num
)
2616 memcg_limited_groups_array_size
= num
;
2619 static void memcg_register_cache(struct mem_cgroup
*memcg
,
2620 struct kmem_cache
*root_cache
)
2622 static char memcg_name_buf
[NAME_MAX
+ 1]; /* protected by
2624 struct kmem_cache
*cachep
;
2627 lockdep_assert_held(&memcg_slab_mutex
);
2629 id
= memcg_cache_id(memcg
);
2632 * Since per-memcg caches are created asynchronously on first
2633 * allocation (see memcg_kmem_get_cache()), several threads can try to
2634 * create the same cache, but only one of them may succeed.
2636 if (cache_from_memcg_idx(root_cache
, id
))
2639 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
2640 cachep
= memcg_create_kmem_cache(memcg
, root_cache
, memcg_name_buf
);
2642 * If we could not create a memcg cache, do not complain, because
2643 * that's not critical at all as we can always proceed with the root
2649 css_get(&memcg
->css
);
2650 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2653 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2654 * barrier here to ensure nobody will see the kmem_cache partially
2659 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
2660 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
2663 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
2665 struct kmem_cache
*root_cache
;
2666 struct mem_cgroup
*memcg
;
2669 lockdep_assert_held(&memcg_slab_mutex
);
2671 BUG_ON(is_root_cache(cachep
));
2673 root_cache
= cachep
->memcg_params
->root_cache
;
2674 memcg
= cachep
->memcg_params
->memcg
;
2675 id
= memcg_cache_id(memcg
);
2677 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
2678 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
2680 list_del(&cachep
->memcg_params
->list
);
2682 kmem_cache_destroy(cachep
);
2684 /* drop the reference taken in memcg_register_cache */
2685 css_put(&memcg
->css
);
2689 * During the creation a new cache, we need to disable our accounting mechanism
2690 * altogether. This is true even if we are not creating, but rather just
2691 * enqueing new caches to be created.
2693 * This is because that process will trigger allocations; some visible, like
2694 * explicit kmallocs to auxiliary data structures, name strings and internal
2695 * cache structures; some well concealed, like INIT_WORK() that can allocate
2696 * objects during debug.
2698 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2699 * to it. This may not be a bounded recursion: since the first cache creation
2700 * failed to complete (waiting on the allocation), we'll just try to create the
2701 * cache again, failing at the same point.
2703 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2704 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2705 * inside the following two functions.
2707 static inline void memcg_stop_kmem_account(void)
2709 VM_BUG_ON(!current
->mm
);
2710 current
->memcg_kmem_skip_account
++;
2713 static inline void memcg_resume_kmem_account(void)
2715 VM_BUG_ON(!current
->mm
);
2716 current
->memcg_kmem_skip_account
--;
2719 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
2721 struct kmem_cache
*c
;
2724 mutex_lock(&memcg_slab_mutex
);
2725 for_each_memcg_cache_index(i
) {
2726 c
= cache_from_memcg_idx(s
, i
);
2730 memcg_unregister_cache(c
);
2732 if (cache_from_memcg_idx(s
, i
))
2735 mutex_unlock(&memcg_slab_mutex
);
2739 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
2741 struct kmem_cache
*cachep
;
2742 struct memcg_cache_params
*params
, *tmp
;
2744 if (!memcg_kmem_is_active(memcg
))
2747 mutex_lock(&memcg_slab_mutex
);
2748 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
2749 cachep
= memcg_params_to_cache(params
);
2750 kmem_cache_shrink(cachep
);
2751 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
2752 memcg_unregister_cache(cachep
);
2754 mutex_unlock(&memcg_slab_mutex
);
2757 struct memcg_register_cache_work
{
2758 struct mem_cgroup
*memcg
;
2759 struct kmem_cache
*cachep
;
2760 struct work_struct work
;
2763 static void memcg_register_cache_func(struct work_struct
*w
)
2765 struct memcg_register_cache_work
*cw
=
2766 container_of(w
, struct memcg_register_cache_work
, work
);
2767 struct mem_cgroup
*memcg
= cw
->memcg
;
2768 struct kmem_cache
*cachep
= cw
->cachep
;
2770 mutex_lock(&memcg_slab_mutex
);
2771 memcg_register_cache(memcg
, cachep
);
2772 mutex_unlock(&memcg_slab_mutex
);
2774 css_put(&memcg
->css
);
2779 * Enqueue the creation of a per-memcg kmem_cache.
2781 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
2782 struct kmem_cache
*cachep
)
2784 struct memcg_register_cache_work
*cw
;
2786 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2788 css_put(&memcg
->css
);
2793 cw
->cachep
= cachep
;
2795 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
2796 schedule_work(&cw
->work
);
2799 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
2800 struct kmem_cache
*cachep
)
2803 * We need to stop accounting when we kmalloc, because if the
2804 * corresponding kmalloc cache is not yet created, the first allocation
2805 * in __memcg_schedule_register_cache will recurse.
2807 * However, it is better to enclose the whole function. Depending on
2808 * the debugging options enabled, INIT_WORK(), for instance, can
2809 * trigger an allocation. This too, will make us recurse. Because at
2810 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2811 * the safest choice is to do it like this, wrapping the whole function.
2813 memcg_stop_kmem_account();
2814 __memcg_schedule_register_cache(memcg
, cachep
);
2815 memcg_resume_kmem_account();
2818 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
2820 unsigned int nr_pages
= 1 << order
;
2823 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
, nr_pages
);
2825 atomic_add(nr_pages
, &cachep
->memcg_params
->nr_pages
);
2829 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
2831 unsigned int nr_pages
= 1 << order
;
2833 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, nr_pages
);
2834 atomic_sub(nr_pages
, &cachep
->memcg_params
->nr_pages
);
2838 * Return the kmem_cache we're supposed to use for a slab allocation.
2839 * We try to use the current memcg's version of the cache.
2841 * If the cache does not exist yet, if we are the first user of it,
2842 * we either create it immediately, if possible, or create it asynchronously
2844 * In the latter case, we will let the current allocation go through with
2845 * the original cache.
2847 * Can't be called in interrupt context or from kernel threads.
2848 * This function needs to be called with rcu_read_lock() held.
2850 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
2853 struct mem_cgroup
*memcg
;
2854 struct kmem_cache
*memcg_cachep
;
2856 VM_BUG_ON(!cachep
->memcg_params
);
2857 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
2859 if (current
->memcg_kmem_skip_account
)
2863 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
2865 if (!memcg_kmem_is_active(memcg
))
2868 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
2869 if (likely(memcg_cachep
)) {
2870 cachep
= memcg_cachep
;
2874 /* The corresponding put will be done in the workqueue. */
2875 if (!css_tryget_online(&memcg
->css
))
2880 * If we are in a safe context (can wait, and not in interrupt
2881 * context), we could be be predictable and return right away.
2882 * This would guarantee that the allocation being performed
2883 * already belongs in the new cache.
2885 * However, there are some clashes that can arrive from locking.
2886 * For instance, because we acquire the slab_mutex while doing
2887 * memcg_create_kmem_cache, this means no further allocation
2888 * could happen with the slab_mutex held. So it's better to
2891 memcg_schedule_register_cache(memcg
, cachep
);
2899 * We need to verify if the allocation against current->mm->owner's memcg is
2900 * possible for the given order. But the page is not allocated yet, so we'll
2901 * need a further commit step to do the final arrangements.
2903 * It is possible for the task to switch cgroups in this mean time, so at
2904 * commit time, we can't rely on task conversion any longer. We'll then use
2905 * the handle argument to return to the caller which cgroup we should commit
2906 * against. We could also return the memcg directly and avoid the pointer
2907 * passing, but a boolean return value gives better semantics considering
2908 * the compiled-out case as well.
2910 * Returning true means the allocation is possible.
2913 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
2915 struct mem_cgroup
*memcg
;
2921 * Disabling accounting is only relevant for some specific memcg
2922 * internal allocations. Therefore we would initially not have such
2923 * check here, since direct calls to the page allocator that are
2924 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
2925 * outside memcg core. We are mostly concerned with cache allocations,
2926 * and by having this test at memcg_kmem_get_cache, we are already able
2927 * to relay the allocation to the root cache and bypass the memcg cache
2930 * There is one exception, though: the SLUB allocator does not create
2931 * large order caches, but rather service large kmallocs directly from
2932 * the page allocator. Therefore, the following sequence when backed by
2933 * the SLUB allocator:
2935 * memcg_stop_kmem_account();
2936 * kmalloc(<large_number>)
2937 * memcg_resume_kmem_account();
2939 * would effectively ignore the fact that we should skip accounting,
2940 * since it will drive us directly to this function without passing
2941 * through the cache selector memcg_kmem_get_cache. Such large
2942 * allocations are extremely rare but can happen, for instance, for the
2943 * cache arrays. We bring this test here.
2945 if (current
->memcg_kmem_skip_account
)
2948 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2950 if (!memcg_kmem_is_active(memcg
)) {
2951 css_put(&memcg
->css
);
2955 ret
= memcg_charge_kmem(memcg
, gfp
, 1 << order
);
2959 css_put(&memcg
->css
);
2963 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2966 VM_BUG_ON(mem_cgroup_is_root(memcg
));
2968 /* The page allocation failed. Revert */
2970 memcg_uncharge_kmem(memcg
, 1 << order
);
2973 page
->mem_cgroup
= memcg
;
2976 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
2978 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2983 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2985 memcg_uncharge_kmem(memcg
, 1 << order
);
2986 page
->mem_cgroup
= NULL
;
2989 static inline void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
2992 #endif /* CONFIG_MEMCG_KMEM */
2994 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2997 * Because tail pages are not marked as "used", set it. We're under
2998 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2999 * charge/uncharge will be never happen and move_account() is done under
3000 * compound_lock(), so we don't have to take care of races.
3002 void mem_cgroup_split_huge_fixup(struct page
*head
)
3006 if (mem_cgroup_disabled())
3009 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
3010 head
[i
].mem_cgroup
= head
->mem_cgroup
;
3012 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3015 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3018 * mem_cgroup_move_account - move account of the page
3020 * @nr_pages: number of regular pages (>1 for huge pages)
3021 * @from: mem_cgroup which the page is moved from.
3022 * @to: mem_cgroup which the page is moved to. @from != @to.
3024 * The caller must confirm following.
3025 * - page is not on LRU (isolate_page() is useful.)
3026 * - compound_lock is held when nr_pages > 1
3028 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3031 static int mem_cgroup_move_account(struct page
*page
,
3032 unsigned int nr_pages
,
3033 struct mem_cgroup
*from
,
3034 struct mem_cgroup
*to
)
3036 unsigned long flags
;
3039 VM_BUG_ON(from
== to
);
3040 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3042 * The page is isolated from LRU. So, collapse function
3043 * will not handle this page. But page splitting can happen.
3044 * Do this check under compound_page_lock(). The caller should
3048 if (nr_pages
> 1 && !PageTransHuge(page
))
3052 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
3053 * of its source page while we change it: page migration takes
3054 * both pages off the LRU, but page cache replacement doesn't.
3056 if (!trylock_page(page
))
3060 if (page
->mem_cgroup
!= from
)
3063 spin_lock_irqsave(&from
->move_lock
, flags
);
3065 if (!PageAnon(page
) && page_mapped(page
)) {
3066 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3068 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3072 if (PageWriteback(page
)) {
3073 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3075 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3080 * It is safe to change page->mem_cgroup here because the page
3081 * is referenced, charged, and isolated - we can't race with
3082 * uncharging, charging, migration, or LRU putback.
3085 /* caller should have done css_get */
3086 page
->mem_cgroup
= to
;
3087 spin_unlock_irqrestore(&from
->move_lock
, flags
);
3091 local_irq_disable();
3092 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
3093 memcg_check_events(to
, page
);
3094 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
3095 memcg_check_events(from
, page
);
3103 #ifdef CONFIG_MEMCG_SWAP
3104 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
3107 int val
= (charge
) ? 1 : -1;
3108 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
3112 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3113 * @entry: swap entry to be moved
3114 * @from: mem_cgroup which the entry is moved from
3115 * @to: mem_cgroup which the entry is moved to
3117 * It succeeds only when the swap_cgroup's record for this entry is the same
3118 * as the mem_cgroup's id of @from.
3120 * Returns 0 on success, -EINVAL on failure.
3122 * The caller must have charged to @to, IOW, called page_counter_charge() about
3123 * both res and memsw, and called css_get().
3125 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3126 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3128 unsigned short old_id
, new_id
;
3130 old_id
= mem_cgroup_id(from
);
3131 new_id
= mem_cgroup_id(to
);
3133 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3134 mem_cgroup_swap_statistics(from
, false);
3135 mem_cgroup_swap_statistics(to
, true);
3137 * This function is only called from task migration context now.
3138 * It postpones page_counter and refcount handling till the end
3139 * of task migration(mem_cgroup_clear_mc()) for performance
3140 * improvement. But we cannot postpone css_get(to) because if
3141 * the process that has been moved to @to does swap-in, the
3142 * refcount of @to might be decreased to 0.
3144 * We are in attach() phase, so the cgroup is guaranteed to be
3145 * alive, so we can just call css_get().
3153 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3154 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3160 static DEFINE_MUTEX(memcg_limit_mutex
);
3162 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3163 unsigned long limit
)
3165 unsigned long curusage
;
3166 unsigned long oldusage
;
3167 bool enlarge
= false;
3172 * For keeping hierarchical_reclaim simple, how long we should retry
3173 * is depends on callers. We set our retry-count to be function
3174 * of # of children which we should visit in this loop.
3176 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
3177 mem_cgroup_count_children(memcg
);
3179 oldusage
= page_counter_read(&memcg
->memory
);
3182 if (signal_pending(current
)) {
3187 mutex_lock(&memcg_limit_mutex
);
3188 if (limit
> memcg
->memsw
.limit
) {
3189 mutex_unlock(&memcg_limit_mutex
);
3193 if (limit
> memcg
->memory
.limit
)
3195 ret
= page_counter_limit(&memcg
->memory
, limit
);
3196 mutex_unlock(&memcg_limit_mutex
);
3201 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
3203 curusage
= page_counter_read(&memcg
->memory
);
3204 /* Usage is reduced ? */
3205 if (curusage
>= oldusage
)
3208 oldusage
= curusage
;
3209 } while (retry_count
);
3211 if (!ret
&& enlarge
)
3212 memcg_oom_recover(memcg
);
3217 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3218 unsigned long limit
)
3220 unsigned long curusage
;
3221 unsigned long oldusage
;
3222 bool enlarge
= false;
3226 /* see mem_cgroup_resize_res_limit */
3227 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
3228 mem_cgroup_count_children(memcg
);
3230 oldusage
= page_counter_read(&memcg
->memsw
);
3233 if (signal_pending(current
)) {
3238 mutex_lock(&memcg_limit_mutex
);
3239 if (limit
< memcg
->memory
.limit
) {
3240 mutex_unlock(&memcg_limit_mutex
);
3244 if (limit
> memcg
->memsw
.limit
)
3246 ret
= page_counter_limit(&memcg
->memsw
, limit
);
3247 mutex_unlock(&memcg_limit_mutex
);
3252 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
3254 curusage
= page_counter_read(&memcg
->memsw
);
3255 /* Usage is reduced ? */
3256 if (curusage
>= oldusage
)
3259 oldusage
= curusage
;
3260 } while (retry_count
);
3262 if (!ret
&& enlarge
)
3263 memcg_oom_recover(memcg
);
3268 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3270 unsigned long *total_scanned
)
3272 unsigned long nr_reclaimed
= 0;
3273 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3274 unsigned long reclaimed
;
3276 struct mem_cgroup_tree_per_zone
*mctz
;
3277 unsigned long excess
;
3278 unsigned long nr_scanned
;
3283 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3285 * This loop can run a while, specially if mem_cgroup's continuously
3286 * keep exceeding their soft limit and putting the system under
3293 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3298 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
3299 gfp_mask
, &nr_scanned
);
3300 nr_reclaimed
+= reclaimed
;
3301 *total_scanned
+= nr_scanned
;
3302 spin_lock_irq(&mctz
->lock
);
3303 __mem_cgroup_remove_exceeded(mz
, mctz
);
3306 * If we failed to reclaim anything from this memory cgroup
3307 * it is time to move on to the next cgroup
3311 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3313 excess
= soft_limit_excess(mz
->memcg
);
3315 * One school of thought says that we should not add
3316 * back the node to the tree if reclaim returns 0.
3317 * But our reclaim could return 0, simply because due
3318 * to priority we are exposing a smaller subset of
3319 * memory to reclaim from. Consider this as a longer
3322 /* If excess == 0, no tree ops */
3323 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3324 spin_unlock_irq(&mctz
->lock
);
3325 css_put(&mz
->memcg
->css
);
3328 * Could not reclaim anything and there are no more
3329 * mem cgroups to try or we seem to be looping without
3330 * reclaiming anything.
3332 if (!nr_reclaimed
&&
3334 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3336 } while (!nr_reclaimed
);
3338 css_put(&next_mz
->memcg
->css
);
3339 return nr_reclaimed
;
3343 * Test whether @memcg has children, dead or alive. Note that this
3344 * function doesn't care whether @memcg has use_hierarchy enabled and
3345 * returns %true if there are child csses according to the cgroup
3346 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3348 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3353 * The lock does not prevent addition or deletion of children, but
3354 * it prevents a new child from being initialized based on this
3355 * parent in css_online(), so it's enough to decide whether
3356 * hierarchically inherited attributes can still be changed or not.
3358 lockdep_assert_held(&memcg_create_mutex
);
3361 ret
= css_next_child(NULL
, &memcg
->css
);
3367 * Reclaims as many pages from the given memcg as possible and moves
3368 * the rest to the parent.
3370 * Caller is responsible for holding css reference for memcg.
3372 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3374 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3376 /* we call try-to-free pages for make this cgroup empty */
3377 lru_add_drain_all();
3378 /* try to free all pages in this cgroup */
3379 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3382 if (signal_pending(current
))
3385 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3389 /* maybe some writeback is necessary */
3390 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3398 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3399 char *buf
, size_t nbytes
,
3402 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3404 if (mem_cgroup_is_root(memcg
))
3406 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3409 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3412 return mem_cgroup_from_css(css
)->use_hierarchy
;
3415 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3416 struct cftype
*cft
, u64 val
)
3419 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3420 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3422 mutex_lock(&memcg_create_mutex
);
3424 if (memcg
->use_hierarchy
== val
)
3428 * If parent's use_hierarchy is set, we can't make any modifications
3429 * in the child subtrees. If it is unset, then the change can
3430 * occur, provided the current cgroup has no children.
3432 * For the root cgroup, parent_mem is NULL, we allow value to be
3433 * set if there are no children.
3435 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3436 (val
== 1 || val
== 0)) {
3437 if (!memcg_has_children(memcg
))
3438 memcg
->use_hierarchy
= val
;
3445 mutex_unlock(&memcg_create_mutex
);
3450 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
3451 enum mem_cgroup_stat_index idx
)
3453 struct mem_cgroup
*iter
;
3456 /* Per-cpu values can be negative, use a signed accumulator */
3457 for_each_mem_cgroup_tree(iter
, memcg
)
3458 val
+= mem_cgroup_read_stat(iter
, idx
);
3460 if (val
< 0) /* race ? */
3465 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3469 if (mem_cgroup_is_root(memcg
)) {
3470 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
3471 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
3473 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
3476 val
= page_counter_read(&memcg
->memory
);
3478 val
= page_counter_read(&memcg
->memsw
);
3480 return val
<< PAGE_SHIFT
;
3491 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3494 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3495 struct page_counter
*counter
;
3497 switch (MEMFILE_TYPE(cft
->private)) {
3499 counter
= &memcg
->memory
;
3502 counter
= &memcg
->memsw
;
3505 counter
= &memcg
->kmem
;
3511 switch (MEMFILE_ATTR(cft
->private)) {
3513 if (counter
== &memcg
->memory
)
3514 return mem_cgroup_usage(memcg
, false);
3515 if (counter
== &memcg
->memsw
)
3516 return mem_cgroup_usage(memcg
, true);
3517 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3519 return (u64
)counter
->limit
* PAGE_SIZE
;
3521 return (u64
)counter
->watermark
* PAGE_SIZE
;
3523 return counter
->failcnt
;
3524 case RES_SOFT_LIMIT
:
3525 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3531 #ifdef CONFIG_MEMCG_KMEM
3532 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
3533 unsigned long nr_pages
)
3538 if (memcg_kmem_is_active(memcg
))
3542 * For simplicity, we won't allow this to be disabled. It also can't
3543 * be changed if the cgroup has children already, or if tasks had
3546 * If tasks join before we set the limit, a person looking at
3547 * kmem.usage_in_bytes will have no way to determine when it took
3548 * place, which makes the value quite meaningless.
3550 * After it first became limited, changes in the value of the limit are
3551 * of course permitted.
3553 mutex_lock(&memcg_create_mutex
);
3554 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
3555 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
3557 mutex_unlock(&memcg_create_mutex
);
3561 memcg_id
= memcg_alloc_cache_id();
3567 memcg
->kmemcg_id
= memcg_id
;
3568 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3571 * We couldn't have accounted to this cgroup, because it hasn't got the
3572 * active bit set yet, so this should succeed.
3574 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
3577 static_key_slow_inc(&memcg_kmem_enabled_key
);
3579 * Setting the active bit after enabling static branching will
3580 * guarantee no one starts accounting before all call sites are
3583 memcg_kmem_set_active(memcg
);
3588 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3589 unsigned long limit
)
3593 mutex_lock(&memcg_limit_mutex
);
3594 if (!memcg_kmem_is_active(memcg
))
3595 ret
= memcg_activate_kmem(memcg
, limit
);
3597 ret
= page_counter_limit(&memcg
->kmem
, limit
);
3598 mutex_unlock(&memcg_limit_mutex
);
3602 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
3605 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
3610 mutex_lock(&memcg_limit_mutex
);
3612 * If the parent cgroup is not kmem-active now, it cannot be activated
3613 * after this point, because it has at least one child already.
3615 if (memcg_kmem_is_active(parent
))
3616 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
3617 mutex_unlock(&memcg_limit_mutex
);
3621 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3622 unsigned long limit
)
3626 #endif /* CONFIG_MEMCG_KMEM */
3629 * The user of this function is...
3632 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3633 char *buf
, size_t nbytes
, loff_t off
)
3635 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3636 unsigned long nr_pages
;
3639 buf
= strstrip(buf
);
3640 ret
= page_counter_memparse(buf
, &nr_pages
);
3644 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3646 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3650 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3652 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3655 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3658 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3662 case RES_SOFT_LIMIT
:
3663 memcg
->soft_limit
= nr_pages
;
3667 return ret
?: nbytes
;
3670 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3671 size_t nbytes
, loff_t off
)
3673 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3674 struct page_counter
*counter
;
3676 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3678 counter
= &memcg
->memory
;
3681 counter
= &memcg
->memsw
;
3684 counter
= &memcg
->kmem
;
3690 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3692 page_counter_reset_watermark(counter
);
3695 counter
->failcnt
= 0;
3704 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3707 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3711 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3712 struct cftype
*cft
, u64 val
)
3714 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3716 if (val
>= (1 << NR_MOVE_TYPE
))
3720 * No kind of locking is needed in here, because ->can_attach() will
3721 * check this value once in the beginning of the process, and then carry
3722 * on with stale data. This means that changes to this value will only
3723 * affect task migrations starting after the change.
3725 memcg
->move_charge_at_immigrate
= val
;
3729 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3730 struct cftype
*cft
, u64 val
)
3737 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3741 unsigned int lru_mask
;
3744 static const struct numa_stat stats
[] = {
3745 { "total", LRU_ALL
},
3746 { "file", LRU_ALL_FILE
},
3747 { "anon", LRU_ALL_ANON
},
3748 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3750 const struct numa_stat
*stat
;
3753 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3755 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3756 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3757 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3758 for_each_node_state(nid
, N_MEMORY
) {
3759 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3761 seq_printf(m
, " N%d=%lu", nid
, nr
);
3766 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3767 struct mem_cgroup
*iter
;
3770 for_each_mem_cgroup_tree(iter
, memcg
)
3771 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3772 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3773 for_each_node_state(nid
, N_MEMORY
) {
3775 for_each_mem_cgroup_tree(iter
, memcg
)
3776 nr
+= mem_cgroup_node_nr_lru_pages(
3777 iter
, nid
, stat
->lru_mask
);
3778 seq_printf(m
, " N%d=%lu", nid
, nr
);
3785 #endif /* CONFIG_NUMA */
3787 static inline void mem_cgroup_lru_names_not_uptodate(void)
3789 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3792 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3794 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3795 unsigned long memory
, memsw
;
3796 struct mem_cgroup
*mi
;
3799 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3800 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3802 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
3803 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3806 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3807 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3808 mem_cgroup_read_events(memcg
, i
));
3810 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3811 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3812 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3814 /* Hierarchical information */
3815 memory
= memsw
= PAGE_COUNTER_MAX
;
3816 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3817 memory
= min(memory
, mi
->memory
.limit
);
3818 memsw
= min(memsw
, mi
->memsw
.limit
);
3820 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3821 (u64
)memory
* PAGE_SIZE
);
3822 if (do_swap_account
)
3823 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3824 (u64
)memsw
* PAGE_SIZE
);
3826 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3829 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3831 for_each_mem_cgroup_tree(mi
, memcg
)
3832 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3833 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
3836 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3837 unsigned long long val
= 0;
3839 for_each_mem_cgroup_tree(mi
, memcg
)
3840 val
+= mem_cgroup_read_events(mi
, i
);
3841 seq_printf(m
, "total_%s %llu\n",
3842 mem_cgroup_events_names
[i
], val
);
3845 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3846 unsigned long long val
= 0;
3848 for_each_mem_cgroup_tree(mi
, memcg
)
3849 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3850 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3853 #ifdef CONFIG_DEBUG_VM
3856 struct mem_cgroup_per_zone
*mz
;
3857 struct zone_reclaim_stat
*rstat
;
3858 unsigned long recent_rotated
[2] = {0, 0};
3859 unsigned long recent_scanned
[2] = {0, 0};
3861 for_each_online_node(nid
)
3862 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3863 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3864 rstat
= &mz
->lruvec
.reclaim_stat
;
3866 recent_rotated
[0] += rstat
->recent_rotated
[0];
3867 recent_rotated
[1] += rstat
->recent_rotated
[1];
3868 recent_scanned
[0] += rstat
->recent_scanned
[0];
3869 recent_scanned
[1] += rstat
->recent_scanned
[1];
3871 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3872 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3873 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3874 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3881 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3884 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3886 return mem_cgroup_swappiness(memcg
);
3889 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3890 struct cftype
*cft
, u64 val
)
3892 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3898 memcg
->swappiness
= val
;
3900 vm_swappiness
= val
;
3905 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3907 struct mem_cgroup_threshold_ary
*t
;
3908 unsigned long usage
;
3913 t
= rcu_dereference(memcg
->thresholds
.primary
);
3915 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3920 usage
= mem_cgroup_usage(memcg
, swap
);
3923 * current_threshold points to threshold just below or equal to usage.
3924 * If it's not true, a threshold was crossed after last
3925 * call of __mem_cgroup_threshold().
3927 i
= t
->current_threshold
;
3930 * Iterate backward over array of thresholds starting from
3931 * current_threshold and check if a threshold is crossed.
3932 * If none of thresholds below usage is crossed, we read
3933 * only one element of the array here.
3935 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3936 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3938 /* i = current_threshold + 1 */
3942 * Iterate forward over array of thresholds starting from
3943 * current_threshold+1 and check if a threshold is crossed.
3944 * If none of thresholds above usage is crossed, we read
3945 * only one element of the array here.
3947 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3948 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3950 /* Update current_threshold */
3951 t
->current_threshold
= i
- 1;
3956 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3959 __mem_cgroup_threshold(memcg
, false);
3960 if (do_swap_account
)
3961 __mem_cgroup_threshold(memcg
, true);
3963 memcg
= parent_mem_cgroup(memcg
);
3967 static int compare_thresholds(const void *a
, const void *b
)
3969 const struct mem_cgroup_threshold
*_a
= a
;
3970 const struct mem_cgroup_threshold
*_b
= b
;
3972 if (_a
->threshold
> _b
->threshold
)
3975 if (_a
->threshold
< _b
->threshold
)
3981 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3983 struct mem_cgroup_eventfd_list
*ev
;
3985 spin_lock(&memcg_oom_lock
);
3987 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3988 eventfd_signal(ev
->eventfd
, 1);
3990 spin_unlock(&memcg_oom_lock
);
3994 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3996 struct mem_cgroup
*iter
;
3998 for_each_mem_cgroup_tree(iter
, memcg
)
3999 mem_cgroup_oom_notify_cb(iter
);
4002 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4003 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4005 struct mem_cgroup_thresholds
*thresholds
;
4006 struct mem_cgroup_threshold_ary
*new;
4007 unsigned long threshold
;
4008 unsigned long usage
;
4011 ret
= page_counter_memparse(args
, &threshold
);
4015 mutex_lock(&memcg
->thresholds_lock
);
4018 thresholds
= &memcg
->thresholds
;
4019 usage
= mem_cgroup_usage(memcg
, false);
4020 } else if (type
== _MEMSWAP
) {
4021 thresholds
= &memcg
->memsw_thresholds
;
4022 usage
= mem_cgroup_usage(memcg
, true);
4026 /* Check if a threshold crossed before adding a new one */
4027 if (thresholds
->primary
)
4028 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4030 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4032 /* Allocate memory for new array of thresholds */
4033 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
4041 /* Copy thresholds (if any) to new array */
4042 if (thresholds
->primary
) {
4043 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4044 sizeof(struct mem_cgroup_threshold
));
4047 /* Add new threshold */
4048 new->entries
[size
- 1].eventfd
= eventfd
;
4049 new->entries
[size
- 1].threshold
= threshold
;
4051 /* Sort thresholds. Registering of new threshold isn't time-critical */
4052 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4053 compare_thresholds
, NULL
);
4055 /* Find current threshold */
4056 new->current_threshold
= -1;
4057 for (i
= 0; i
< size
; i
++) {
4058 if (new->entries
[i
].threshold
<= usage
) {
4060 * new->current_threshold will not be used until
4061 * rcu_assign_pointer(), so it's safe to increment
4064 ++new->current_threshold
;
4069 /* Free old spare buffer and save old primary buffer as spare */
4070 kfree(thresholds
->spare
);
4071 thresholds
->spare
= thresholds
->primary
;
4073 rcu_assign_pointer(thresholds
->primary
, new);
4075 /* To be sure that nobody uses thresholds */
4079 mutex_unlock(&memcg
->thresholds_lock
);
4084 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4085 struct eventfd_ctx
*eventfd
, const char *args
)
4087 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4090 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4091 struct eventfd_ctx
*eventfd
, const char *args
)
4093 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4096 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4097 struct eventfd_ctx
*eventfd
, enum res_type type
)
4099 struct mem_cgroup_thresholds
*thresholds
;
4100 struct mem_cgroup_threshold_ary
*new;
4101 unsigned long usage
;
4104 mutex_lock(&memcg
->thresholds_lock
);
4107 thresholds
= &memcg
->thresholds
;
4108 usage
= mem_cgroup_usage(memcg
, false);
4109 } else if (type
== _MEMSWAP
) {
4110 thresholds
= &memcg
->memsw_thresholds
;
4111 usage
= mem_cgroup_usage(memcg
, true);
4115 if (!thresholds
->primary
)
4118 /* Check if a threshold crossed before removing */
4119 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4121 /* Calculate new number of threshold */
4123 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4124 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4128 new = thresholds
->spare
;
4130 /* Set thresholds array to NULL if we don't have thresholds */
4139 /* Copy thresholds and find current threshold */
4140 new->current_threshold
= -1;
4141 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4142 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4145 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4146 if (new->entries
[j
].threshold
<= usage
) {
4148 * new->current_threshold will not be used
4149 * until rcu_assign_pointer(), so it's safe to increment
4152 ++new->current_threshold
;
4158 /* Swap primary and spare array */
4159 thresholds
->spare
= thresholds
->primary
;
4160 /* If all events are unregistered, free the spare array */
4162 kfree(thresholds
->spare
);
4163 thresholds
->spare
= NULL
;
4166 rcu_assign_pointer(thresholds
->primary
, new);
4168 /* To be sure that nobody uses thresholds */
4171 mutex_unlock(&memcg
->thresholds_lock
);
4174 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4175 struct eventfd_ctx
*eventfd
)
4177 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4180 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4181 struct eventfd_ctx
*eventfd
)
4183 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4186 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4187 struct eventfd_ctx
*eventfd
, const char *args
)
4189 struct mem_cgroup_eventfd_list
*event
;
4191 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4195 spin_lock(&memcg_oom_lock
);
4197 event
->eventfd
= eventfd
;
4198 list_add(&event
->list
, &memcg
->oom_notify
);
4200 /* already in OOM ? */
4201 if (atomic_read(&memcg
->under_oom
))
4202 eventfd_signal(eventfd
, 1);
4203 spin_unlock(&memcg_oom_lock
);
4208 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4209 struct eventfd_ctx
*eventfd
)
4211 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4213 spin_lock(&memcg_oom_lock
);
4215 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4216 if (ev
->eventfd
== eventfd
) {
4217 list_del(&ev
->list
);
4222 spin_unlock(&memcg_oom_lock
);
4225 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4227 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
4229 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4230 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
4234 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4235 struct cftype
*cft
, u64 val
)
4237 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4239 /* cannot set to root cgroup and only 0 and 1 are allowed */
4240 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4243 memcg
->oom_kill_disable
= val
;
4245 memcg_oom_recover(memcg
);
4250 #ifdef CONFIG_MEMCG_KMEM
4251 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4255 memcg
->kmemcg_id
= -1;
4256 ret
= memcg_propagate_kmem(memcg
);
4260 return mem_cgroup_sockets_init(memcg
, ss
);
4263 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4265 mem_cgroup_sockets_destroy(memcg
);
4268 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4273 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4279 * DO NOT USE IN NEW FILES.
4281 * "cgroup.event_control" implementation.
4283 * This is way over-engineered. It tries to support fully configurable
4284 * events for each user. Such level of flexibility is completely
4285 * unnecessary especially in the light of the planned unified hierarchy.
4287 * Please deprecate this and replace with something simpler if at all
4292 * Unregister event and free resources.
4294 * Gets called from workqueue.
4296 static void memcg_event_remove(struct work_struct
*work
)
4298 struct mem_cgroup_event
*event
=
4299 container_of(work
, struct mem_cgroup_event
, remove
);
4300 struct mem_cgroup
*memcg
= event
->memcg
;
4302 remove_wait_queue(event
->wqh
, &event
->wait
);
4304 event
->unregister_event(memcg
, event
->eventfd
);
4306 /* Notify userspace the event is going away. */
4307 eventfd_signal(event
->eventfd
, 1);
4309 eventfd_ctx_put(event
->eventfd
);
4311 css_put(&memcg
->css
);
4315 * Gets called on POLLHUP on eventfd when user closes it.
4317 * Called with wqh->lock held and interrupts disabled.
4319 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
4320 int sync
, void *key
)
4322 struct mem_cgroup_event
*event
=
4323 container_of(wait
, struct mem_cgroup_event
, wait
);
4324 struct mem_cgroup
*memcg
= event
->memcg
;
4325 unsigned long flags
= (unsigned long)key
;
4327 if (flags
& POLLHUP
) {
4329 * If the event has been detached at cgroup removal, we
4330 * can simply return knowing the other side will cleanup
4333 * We can't race against event freeing since the other
4334 * side will require wqh->lock via remove_wait_queue(),
4337 spin_lock(&memcg
->event_list_lock
);
4338 if (!list_empty(&event
->list
)) {
4339 list_del_init(&event
->list
);
4341 * We are in atomic context, but cgroup_event_remove()
4342 * may sleep, so we have to call it in workqueue.
4344 schedule_work(&event
->remove
);
4346 spin_unlock(&memcg
->event_list_lock
);
4352 static void memcg_event_ptable_queue_proc(struct file
*file
,
4353 wait_queue_head_t
*wqh
, poll_table
*pt
)
4355 struct mem_cgroup_event
*event
=
4356 container_of(pt
, struct mem_cgroup_event
, pt
);
4359 add_wait_queue(wqh
, &event
->wait
);
4363 * DO NOT USE IN NEW FILES.
4365 * Parse input and register new cgroup event handler.
4367 * Input must be in format '<event_fd> <control_fd> <args>'.
4368 * Interpretation of args is defined by control file implementation.
4370 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4371 char *buf
, size_t nbytes
, loff_t off
)
4373 struct cgroup_subsys_state
*css
= of_css(of
);
4374 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4375 struct mem_cgroup_event
*event
;
4376 struct cgroup_subsys_state
*cfile_css
;
4377 unsigned int efd
, cfd
;
4384 buf
= strstrip(buf
);
4386 efd
= simple_strtoul(buf
, &endp
, 10);
4391 cfd
= simple_strtoul(buf
, &endp
, 10);
4392 if ((*endp
!= ' ') && (*endp
!= '\0'))
4396 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4400 event
->memcg
= memcg
;
4401 INIT_LIST_HEAD(&event
->list
);
4402 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4403 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4404 INIT_WORK(&event
->remove
, memcg_event_remove
);
4412 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4413 if (IS_ERR(event
->eventfd
)) {
4414 ret
= PTR_ERR(event
->eventfd
);
4421 goto out_put_eventfd
;
4424 /* the process need read permission on control file */
4425 /* AV: shouldn't we check that it's been opened for read instead? */
4426 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4431 * Determine the event callbacks and set them in @event. This used
4432 * to be done via struct cftype but cgroup core no longer knows
4433 * about these events. The following is crude but the whole thing
4434 * is for compatibility anyway.
4436 * DO NOT ADD NEW FILES.
4438 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4440 if (!strcmp(name
, "memory.usage_in_bytes")) {
4441 event
->register_event
= mem_cgroup_usage_register_event
;
4442 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4443 } else if (!strcmp(name
, "memory.oom_control")) {
4444 event
->register_event
= mem_cgroup_oom_register_event
;
4445 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4446 } else if (!strcmp(name
, "memory.pressure_level")) {
4447 event
->register_event
= vmpressure_register_event
;
4448 event
->unregister_event
= vmpressure_unregister_event
;
4449 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4450 event
->register_event
= memsw_cgroup_usage_register_event
;
4451 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4458 * Verify @cfile should belong to @css. Also, remaining events are
4459 * automatically removed on cgroup destruction but the removal is
4460 * asynchronous, so take an extra ref on @css.
4462 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4463 &memory_cgrp_subsys
);
4465 if (IS_ERR(cfile_css
))
4467 if (cfile_css
!= css
) {
4472 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4476 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
4478 spin_lock(&memcg
->event_list_lock
);
4479 list_add(&event
->list
, &memcg
->event_list
);
4480 spin_unlock(&memcg
->event_list_lock
);
4492 eventfd_ctx_put(event
->eventfd
);
4501 static struct cftype mem_cgroup_files
[] = {
4503 .name
= "usage_in_bytes",
4504 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4505 .read_u64
= mem_cgroup_read_u64
,
4508 .name
= "max_usage_in_bytes",
4509 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4510 .write
= mem_cgroup_reset
,
4511 .read_u64
= mem_cgroup_read_u64
,
4514 .name
= "limit_in_bytes",
4515 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4516 .write
= mem_cgroup_write
,
4517 .read_u64
= mem_cgroup_read_u64
,
4520 .name
= "soft_limit_in_bytes",
4521 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4522 .write
= mem_cgroup_write
,
4523 .read_u64
= mem_cgroup_read_u64
,
4527 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4528 .write
= mem_cgroup_reset
,
4529 .read_u64
= mem_cgroup_read_u64
,
4533 .seq_show
= memcg_stat_show
,
4536 .name
= "force_empty",
4537 .write
= mem_cgroup_force_empty_write
,
4540 .name
= "use_hierarchy",
4541 .write_u64
= mem_cgroup_hierarchy_write
,
4542 .read_u64
= mem_cgroup_hierarchy_read
,
4545 .name
= "cgroup.event_control", /* XXX: for compat */
4546 .write
= memcg_write_event_control
,
4547 .flags
= CFTYPE_NO_PREFIX
,
4551 .name
= "swappiness",
4552 .read_u64
= mem_cgroup_swappiness_read
,
4553 .write_u64
= mem_cgroup_swappiness_write
,
4556 .name
= "move_charge_at_immigrate",
4557 .read_u64
= mem_cgroup_move_charge_read
,
4558 .write_u64
= mem_cgroup_move_charge_write
,
4561 .name
= "oom_control",
4562 .seq_show
= mem_cgroup_oom_control_read
,
4563 .write_u64
= mem_cgroup_oom_control_write
,
4564 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4567 .name
= "pressure_level",
4571 .name
= "numa_stat",
4572 .seq_show
= memcg_numa_stat_show
,
4575 #ifdef CONFIG_MEMCG_KMEM
4577 .name
= "kmem.limit_in_bytes",
4578 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4579 .write
= mem_cgroup_write
,
4580 .read_u64
= mem_cgroup_read_u64
,
4583 .name
= "kmem.usage_in_bytes",
4584 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4585 .read_u64
= mem_cgroup_read_u64
,
4588 .name
= "kmem.failcnt",
4589 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4590 .write
= mem_cgroup_reset
,
4591 .read_u64
= mem_cgroup_read_u64
,
4594 .name
= "kmem.max_usage_in_bytes",
4595 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4596 .write
= mem_cgroup_reset
,
4597 .read_u64
= mem_cgroup_read_u64
,
4599 #ifdef CONFIG_SLABINFO
4601 .name
= "kmem.slabinfo",
4602 .seq_start
= slab_start
,
4603 .seq_next
= slab_next
,
4604 .seq_stop
= slab_stop
,
4605 .seq_show
= memcg_slab_show
,
4609 { }, /* terminate */
4612 #ifdef CONFIG_MEMCG_SWAP
4613 static struct cftype memsw_cgroup_files
[] = {
4615 .name
= "memsw.usage_in_bytes",
4616 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
4617 .read_u64
= mem_cgroup_read_u64
,
4620 .name
= "memsw.max_usage_in_bytes",
4621 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
4622 .write
= mem_cgroup_reset
,
4623 .read_u64
= mem_cgroup_read_u64
,
4626 .name
= "memsw.limit_in_bytes",
4627 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
4628 .write
= mem_cgroup_write
,
4629 .read_u64
= mem_cgroup_read_u64
,
4632 .name
= "memsw.failcnt",
4633 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
4634 .write
= mem_cgroup_reset
,
4635 .read_u64
= mem_cgroup_read_u64
,
4637 { }, /* terminate */
4640 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4642 struct mem_cgroup_per_node
*pn
;
4643 struct mem_cgroup_per_zone
*mz
;
4644 int zone
, tmp
= node
;
4646 * This routine is called against possible nodes.
4647 * But it's BUG to call kmalloc() against offline node.
4649 * TODO: this routine can waste much memory for nodes which will
4650 * never be onlined. It's better to use memory hotplug callback
4653 if (!node_state(node
, N_NORMAL_MEMORY
))
4655 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4659 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4660 mz
= &pn
->zoneinfo
[zone
];
4661 lruvec_init(&mz
->lruvec
);
4662 mz
->usage_in_excess
= 0;
4663 mz
->on_tree
= false;
4666 memcg
->nodeinfo
[node
] = pn
;
4670 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4672 kfree(memcg
->nodeinfo
[node
]);
4675 static struct mem_cgroup
*mem_cgroup_alloc(void)
4677 struct mem_cgroup
*memcg
;
4680 size
= sizeof(struct mem_cgroup
);
4681 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4683 memcg
= kzalloc(size
, GFP_KERNEL
);
4687 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4690 spin_lock_init(&memcg
->pcp_counter_lock
);
4699 * At destroying mem_cgroup, references from swap_cgroup can remain.
4700 * (scanning all at force_empty is too costly...)
4702 * Instead of clearing all references at force_empty, we remember
4703 * the number of reference from swap_cgroup and free mem_cgroup when
4704 * it goes down to 0.
4706 * Removal of cgroup itself succeeds regardless of refs from swap.
4709 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4713 mem_cgroup_remove_from_trees(memcg
);
4716 free_mem_cgroup_per_zone_info(memcg
, node
);
4718 free_percpu(memcg
->stat
);
4720 disarm_static_keys(memcg
);
4725 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4727 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4729 if (!memcg
->memory
.parent
)
4731 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4733 EXPORT_SYMBOL(parent_mem_cgroup
);
4735 static void __init
mem_cgroup_soft_limit_tree_init(void)
4737 struct mem_cgroup_tree_per_node
*rtpn
;
4738 struct mem_cgroup_tree_per_zone
*rtpz
;
4739 int tmp
, node
, zone
;
4741 for_each_node(node
) {
4743 if (!node_state(node
, N_NORMAL_MEMORY
))
4745 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
4748 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
4750 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4751 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
4752 rtpz
->rb_root
= RB_ROOT
;
4753 spin_lock_init(&rtpz
->lock
);
4758 static struct cgroup_subsys_state
* __ref
4759 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4761 struct mem_cgroup
*memcg
;
4762 long error
= -ENOMEM
;
4765 memcg
= mem_cgroup_alloc();
4767 return ERR_PTR(error
);
4770 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4774 if (parent_css
== NULL
) {
4775 root_mem_cgroup
= memcg
;
4776 page_counter_init(&memcg
->memory
, NULL
);
4777 page_counter_init(&memcg
->memsw
, NULL
);
4778 page_counter_init(&memcg
->kmem
, NULL
);
4781 memcg
->last_scanned_node
= MAX_NUMNODES
;
4782 INIT_LIST_HEAD(&memcg
->oom_notify
);
4783 memcg
->move_charge_at_immigrate
= 0;
4784 mutex_init(&memcg
->thresholds_lock
);
4785 spin_lock_init(&memcg
->move_lock
);
4786 vmpressure_init(&memcg
->vmpressure
);
4787 INIT_LIST_HEAD(&memcg
->event_list
);
4788 spin_lock_init(&memcg
->event_list_lock
);
4793 __mem_cgroup_free(memcg
);
4794 return ERR_PTR(error
);
4798 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4800 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4801 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4804 if (css
->id
> MEM_CGROUP_ID_MAX
)
4810 mutex_lock(&memcg_create_mutex
);
4812 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4813 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4814 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4816 if (parent
->use_hierarchy
) {
4817 page_counter_init(&memcg
->memory
, &parent
->memory
);
4818 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4819 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4822 * No need to take a reference to the parent because cgroup
4823 * core guarantees its existence.
4826 page_counter_init(&memcg
->memory
, NULL
);
4827 page_counter_init(&memcg
->memsw
, NULL
);
4828 page_counter_init(&memcg
->kmem
, NULL
);
4830 * Deeper hierachy with use_hierarchy == false doesn't make
4831 * much sense so let cgroup subsystem know about this
4832 * unfortunate state in our controller.
4834 if (parent
!= root_mem_cgroup
)
4835 memory_cgrp_subsys
.broken_hierarchy
= true;
4837 mutex_unlock(&memcg_create_mutex
);
4839 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4844 * Make sure the memcg is initialized: mem_cgroup_iter()
4845 * orders reading memcg->initialized against its callers
4846 * reading the memcg members.
4848 smp_store_release(&memcg
->initialized
, 1);
4853 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4855 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4856 struct mem_cgroup_event
*event
, *tmp
;
4859 * Unregister events and notify userspace.
4860 * Notify userspace about cgroup removing only after rmdir of cgroup
4861 * directory to avoid race between userspace and kernelspace.
4863 spin_lock(&memcg
->event_list_lock
);
4864 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4865 list_del_init(&event
->list
);
4866 schedule_work(&event
->remove
);
4868 spin_unlock(&memcg
->event_list_lock
);
4870 memcg_unregister_all_caches(memcg
);
4871 vmpressure_cleanup(&memcg
->vmpressure
);
4874 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4876 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4878 memcg_destroy_kmem(memcg
);
4879 __mem_cgroup_free(memcg
);
4883 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4884 * @css: the target css
4886 * Reset the states of the mem_cgroup associated with @css. This is
4887 * invoked when the userland requests disabling on the default hierarchy
4888 * but the memcg is pinned through dependency. The memcg should stop
4889 * applying policies and should revert to the vanilla state as it may be
4890 * made visible again.
4892 * The current implementation only resets the essential configurations.
4893 * This needs to be expanded to cover all the visible parts.
4895 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4897 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4899 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4900 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4901 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4902 memcg
->soft_limit
= 0;
4906 /* Handlers for move charge at task migration. */
4907 static int mem_cgroup_do_precharge(unsigned long count
)
4911 /* Try a single bulk charge without reclaim first */
4912 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
4914 mc
.precharge
+= count
;
4917 if (ret
== -EINTR
) {
4918 cancel_charge(root_mem_cgroup
, count
);
4922 /* Try charges one by one with reclaim */
4924 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4926 * In case of failure, any residual charges against
4927 * mc.to will be dropped by mem_cgroup_clear_mc()
4928 * later on. However, cancel any charges that are
4929 * bypassed to root right away or they'll be lost.
4932 cancel_charge(root_mem_cgroup
, 1);
4942 * get_mctgt_type - get target type of moving charge
4943 * @vma: the vma the pte to be checked belongs
4944 * @addr: the address corresponding to the pte to be checked
4945 * @ptent: the pte to be checked
4946 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4949 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4950 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4951 * move charge. if @target is not NULL, the page is stored in target->page
4952 * with extra refcnt got(Callers should handle it).
4953 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4954 * target for charge migration. if @target is not NULL, the entry is stored
4957 * Called with pte lock held.
4964 enum mc_target_type
{
4970 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4971 unsigned long addr
, pte_t ptent
)
4973 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4975 if (!page
|| !page_mapped(page
))
4977 if (PageAnon(page
)) {
4978 /* we don't move shared anon */
4981 } else if (!move_file())
4982 /* we ignore mapcount for file pages */
4984 if (!get_page_unless_zero(page
))
4991 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4992 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4994 struct page
*page
= NULL
;
4995 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4997 if (!move_anon() || non_swap_entry(ent
))
5000 * Because lookup_swap_cache() updates some statistics counter,
5001 * we call find_get_page() with swapper_space directly.
5003 page
= find_get_page(swap_address_space(ent
), ent
.val
);
5004 if (do_swap_account
)
5005 entry
->val
= ent
.val
;
5010 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5011 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5017 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5018 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5020 struct page
*page
= NULL
;
5021 struct address_space
*mapping
;
5024 if (!vma
->vm_file
) /* anonymous vma */
5029 mapping
= vma
->vm_file
->f_mapping
;
5030 if (pte_none(ptent
))
5031 pgoff
= linear_page_index(vma
, addr
);
5032 else /* pte_file(ptent) is true */
5033 pgoff
= pte_to_pgoff(ptent
);
5035 /* page is moved even if it's not RSS of this task(page-faulted). */
5037 /* shmem/tmpfs may report page out on swap: account for that too. */
5038 if (shmem_mapping(mapping
)) {
5039 page
= find_get_entry(mapping
, pgoff
);
5040 if (radix_tree_exceptional_entry(page
)) {
5041 swp_entry_t swp
= radix_to_swp_entry(page
);
5042 if (do_swap_account
)
5044 page
= find_get_page(swap_address_space(swp
), swp
.val
);
5047 page
= find_get_page(mapping
, pgoff
);
5049 page
= find_get_page(mapping
, pgoff
);
5054 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5055 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5057 struct page
*page
= NULL
;
5058 enum mc_target_type ret
= MC_TARGET_NONE
;
5059 swp_entry_t ent
= { .val
= 0 };
5061 if (pte_present(ptent
))
5062 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5063 else if (is_swap_pte(ptent
))
5064 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
5065 else if (pte_none(ptent
) || pte_file(ptent
))
5066 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5068 if (!page
&& !ent
.val
)
5072 * Do only loose check w/o serialization.
5073 * mem_cgroup_move_account() checks the page is valid or
5074 * not under LRU exclusion.
5076 if (page
->mem_cgroup
== mc
.from
) {
5077 ret
= MC_TARGET_PAGE
;
5079 target
->page
= page
;
5081 if (!ret
|| !target
)
5084 /* There is a swap entry and a page doesn't exist or isn't charged */
5085 if (ent
.val
&& !ret
&&
5086 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5087 ret
= MC_TARGET_SWAP
;
5094 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5096 * We don't consider swapping or file mapped pages because THP does not
5097 * support them for now.
5098 * Caller should make sure that pmd_trans_huge(pmd) is true.
5100 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5101 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5103 struct page
*page
= NULL
;
5104 enum mc_target_type ret
= MC_TARGET_NONE
;
5106 page
= pmd_page(pmd
);
5107 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5110 if (page
->mem_cgroup
== mc
.from
) {
5111 ret
= MC_TARGET_PAGE
;
5114 target
->page
= page
;
5120 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5121 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5123 return MC_TARGET_NONE
;
5127 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5128 unsigned long addr
, unsigned long end
,
5129 struct mm_walk
*walk
)
5131 struct vm_area_struct
*vma
= walk
->private;
5135 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5136 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5137 mc
.precharge
+= HPAGE_PMD_NR
;
5142 if (pmd_trans_unstable(pmd
))
5144 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5145 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5146 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5147 mc
.precharge
++; /* increment precharge temporarily */
5148 pte_unmap_unlock(pte
- 1, ptl
);
5154 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5156 unsigned long precharge
;
5157 struct vm_area_struct
*vma
;
5159 down_read(&mm
->mmap_sem
);
5160 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5161 struct mm_walk mem_cgroup_count_precharge_walk
= {
5162 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5166 if (is_vm_hugetlb_page(vma
))
5168 walk_page_range(vma
->vm_start
, vma
->vm_end
,
5169 &mem_cgroup_count_precharge_walk
);
5171 up_read(&mm
->mmap_sem
);
5173 precharge
= mc
.precharge
;
5179 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5181 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5183 VM_BUG_ON(mc
.moving_task
);
5184 mc
.moving_task
= current
;
5185 return mem_cgroup_do_precharge(precharge
);
5188 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5189 static void __mem_cgroup_clear_mc(void)
5191 struct mem_cgroup
*from
= mc
.from
;
5192 struct mem_cgroup
*to
= mc
.to
;
5194 /* we must uncharge all the leftover precharges from mc.to */
5196 cancel_charge(mc
.to
, mc
.precharge
);
5200 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5201 * we must uncharge here.
5203 if (mc
.moved_charge
) {
5204 cancel_charge(mc
.from
, mc
.moved_charge
);
5205 mc
.moved_charge
= 0;
5207 /* we must fixup refcnts and charges */
5208 if (mc
.moved_swap
) {
5209 /* uncharge swap account from the old cgroup */
5210 if (!mem_cgroup_is_root(mc
.from
))
5211 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5214 * we charged both to->memory and to->memsw, so we
5215 * should uncharge to->memory.
5217 if (!mem_cgroup_is_root(mc
.to
))
5218 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5220 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
5222 /* we've already done css_get(mc.to) */
5225 memcg_oom_recover(from
);
5226 memcg_oom_recover(to
);
5227 wake_up_all(&mc
.waitq
);
5230 static void mem_cgroup_clear_mc(void)
5233 * we must clear moving_task before waking up waiters at the end of
5236 mc
.moving_task
= NULL
;
5237 __mem_cgroup_clear_mc();
5238 spin_lock(&mc
.lock
);
5241 spin_unlock(&mc
.lock
);
5244 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5245 struct cgroup_taskset
*tset
)
5247 struct task_struct
*p
= cgroup_taskset_first(tset
);
5249 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5250 unsigned long move_charge_at_immigrate
;
5253 * We are now commited to this value whatever it is. Changes in this
5254 * tunable will only affect upcoming migrations, not the current one.
5255 * So we need to save it, and keep it going.
5257 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
5258 if (move_charge_at_immigrate
) {
5259 struct mm_struct
*mm
;
5260 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
5262 VM_BUG_ON(from
== memcg
);
5264 mm
= get_task_mm(p
);
5267 /* We move charges only when we move a owner of the mm */
5268 if (mm
->owner
== p
) {
5271 VM_BUG_ON(mc
.precharge
);
5272 VM_BUG_ON(mc
.moved_charge
);
5273 VM_BUG_ON(mc
.moved_swap
);
5275 spin_lock(&mc
.lock
);
5278 mc
.immigrate_flags
= move_charge_at_immigrate
;
5279 spin_unlock(&mc
.lock
);
5280 /* We set mc.moving_task later */
5282 ret
= mem_cgroup_precharge_mc(mm
);
5284 mem_cgroup_clear_mc();
5291 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5292 struct cgroup_taskset
*tset
)
5295 mem_cgroup_clear_mc();
5298 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5299 unsigned long addr
, unsigned long end
,
5300 struct mm_walk
*walk
)
5303 struct vm_area_struct
*vma
= walk
->private;
5306 enum mc_target_type target_type
;
5307 union mc_target target
;
5311 * We don't take compound_lock() here but no race with splitting thp
5313 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5314 * under splitting, which means there's no concurrent thp split,
5315 * - if another thread runs into split_huge_page() just after we
5316 * entered this if-block, the thread must wait for page table lock
5317 * to be unlocked in __split_huge_page_splitting(), where the main
5318 * part of thp split is not executed yet.
5320 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5321 if (mc
.precharge
< HPAGE_PMD_NR
) {
5325 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5326 if (target_type
== MC_TARGET_PAGE
) {
5328 if (!isolate_lru_page(page
)) {
5329 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
5331 mc
.precharge
-= HPAGE_PMD_NR
;
5332 mc
.moved_charge
+= HPAGE_PMD_NR
;
5334 putback_lru_page(page
);
5342 if (pmd_trans_unstable(pmd
))
5345 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5346 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5347 pte_t ptent
= *(pte
++);
5353 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5354 case MC_TARGET_PAGE
:
5356 if (isolate_lru_page(page
))
5358 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
5360 /* we uncharge from mc.from later. */
5363 putback_lru_page(page
);
5364 put
: /* get_mctgt_type() gets the page */
5367 case MC_TARGET_SWAP
:
5369 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5371 /* we fixup refcnts and charges later. */
5379 pte_unmap_unlock(pte
- 1, ptl
);
5384 * We have consumed all precharges we got in can_attach().
5385 * We try charge one by one, but don't do any additional
5386 * charges to mc.to if we have failed in charge once in attach()
5389 ret
= mem_cgroup_do_precharge(1);
5397 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
5399 struct vm_area_struct
*vma
;
5401 lru_add_drain_all();
5403 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5404 * move_lock while we're moving its pages to another memcg.
5405 * Then wait for already started RCU-only updates to finish.
5407 atomic_inc(&mc
.from
->moving_account
);
5410 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5412 * Someone who are holding the mmap_sem might be waiting in
5413 * waitq. So we cancel all extra charges, wake up all waiters,
5414 * and retry. Because we cancel precharges, we might not be able
5415 * to move enough charges, but moving charge is a best-effort
5416 * feature anyway, so it wouldn't be a big problem.
5418 __mem_cgroup_clear_mc();
5422 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5424 struct mm_walk mem_cgroup_move_charge_walk
= {
5425 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5429 if (is_vm_hugetlb_page(vma
))
5431 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
5432 &mem_cgroup_move_charge_walk
);
5435 * means we have consumed all precharges and failed in
5436 * doing additional charge. Just abandon here.
5440 up_read(&mm
->mmap_sem
);
5441 atomic_dec(&mc
.from
->moving_account
);
5444 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5445 struct cgroup_taskset
*tset
)
5447 struct task_struct
*p
= cgroup_taskset_first(tset
);
5448 struct mm_struct
*mm
= get_task_mm(p
);
5452 mem_cgroup_move_charge(mm
);
5456 mem_cgroup_clear_mc();
5458 #else /* !CONFIG_MMU */
5459 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5460 struct cgroup_taskset
*tset
)
5464 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5465 struct cgroup_taskset
*tset
)
5468 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5469 struct cgroup_taskset
*tset
)
5475 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5476 * to verify whether we're attached to the default hierarchy on each mount
5479 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5482 * use_hierarchy is forced on the default hierarchy. cgroup core
5483 * guarantees that @root doesn't have any children, so turning it
5484 * on for the root memcg is enough.
5486 if (cgroup_on_dfl(root_css
->cgroup
))
5487 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
5490 struct cgroup_subsys memory_cgrp_subsys
= {
5491 .css_alloc
= mem_cgroup_css_alloc
,
5492 .css_online
= mem_cgroup_css_online
,
5493 .css_offline
= mem_cgroup_css_offline
,
5494 .css_free
= mem_cgroup_css_free
,
5495 .css_reset
= mem_cgroup_css_reset
,
5496 .can_attach
= mem_cgroup_can_attach
,
5497 .cancel_attach
= mem_cgroup_cancel_attach
,
5498 .attach
= mem_cgroup_move_task
,
5499 .bind
= mem_cgroup_bind
,
5500 .legacy_cftypes
= mem_cgroup_files
,
5504 #ifdef CONFIG_MEMCG_SWAP
5505 static int __init
enable_swap_account(char *s
)
5507 if (!strcmp(s
, "1"))
5508 really_do_swap_account
= 1;
5509 else if (!strcmp(s
, "0"))
5510 really_do_swap_account
= 0;
5513 __setup("swapaccount=", enable_swap_account
);
5515 static void __init
memsw_file_init(void)
5517 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5518 memsw_cgroup_files
));
5521 static void __init
enable_swap_cgroup(void)
5523 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5524 do_swap_account
= 1;
5530 static void __init
enable_swap_cgroup(void)
5535 #ifdef CONFIG_MEMCG_SWAP
5537 * mem_cgroup_swapout - transfer a memsw charge to swap
5538 * @page: page whose memsw charge to transfer
5539 * @entry: swap entry to move the charge to
5541 * Transfer the memsw charge of @page to @entry.
5543 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5545 struct mem_cgroup
*memcg
;
5546 unsigned short oldid
;
5548 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5549 VM_BUG_ON_PAGE(page_count(page
), page
);
5551 if (!do_swap_account
)
5554 memcg
= page
->mem_cgroup
;
5556 /* Readahead page, never charged */
5560 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5561 VM_BUG_ON_PAGE(oldid
, page
);
5562 mem_cgroup_swap_statistics(memcg
, true);
5564 page
->mem_cgroup
= NULL
;
5566 if (!mem_cgroup_is_root(memcg
))
5567 page_counter_uncharge(&memcg
->memory
, 1);
5569 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5570 VM_BUG_ON(!irqs_disabled());
5572 mem_cgroup_charge_statistics(memcg
, page
, -1);
5573 memcg_check_events(memcg
, page
);
5577 * mem_cgroup_uncharge_swap - uncharge a swap entry
5578 * @entry: swap entry to uncharge
5580 * Drop the memsw charge associated with @entry.
5582 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5584 struct mem_cgroup
*memcg
;
5587 if (!do_swap_account
)
5590 id
= swap_cgroup_record(entry
, 0);
5592 memcg
= mem_cgroup_lookup(id
);
5594 if (!mem_cgroup_is_root(memcg
))
5595 page_counter_uncharge(&memcg
->memsw
, 1);
5596 mem_cgroup_swap_statistics(memcg
, false);
5597 css_put(&memcg
->css
);
5604 * mem_cgroup_try_charge - try charging a page
5605 * @page: page to charge
5606 * @mm: mm context of the victim
5607 * @gfp_mask: reclaim mode
5608 * @memcgp: charged memcg return
5610 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5611 * pages according to @gfp_mask if necessary.
5613 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5614 * Otherwise, an error code is returned.
5616 * After page->mapping has been set up, the caller must finalize the
5617 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5618 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5620 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5621 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5623 struct mem_cgroup
*memcg
= NULL
;
5624 unsigned int nr_pages
= 1;
5627 if (mem_cgroup_disabled())
5630 if (PageSwapCache(page
)) {
5632 * Every swap fault against a single page tries to charge the
5633 * page, bail as early as possible. shmem_unuse() encounters
5634 * already charged pages, too. The USED bit is protected by
5635 * the page lock, which serializes swap cache removal, which
5636 * in turn serializes uncharging.
5638 if (page
->mem_cgroup
)
5642 if (PageTransHuge(page
)) {
5643 nr_pages
<<= compound_order(page
);
5644 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5647 if (do_swap_account
&& PageSwapCache(page
))
5648 memcg
= try_get_mem_cgroup_from_page(page
);
5650 memcg
= get_mem_cgroup_from_mm(mm
);
5652 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5654 css_put(&memcg
->css
);
5656 if (ret
== -EINTR
) {
5657 memcg
= root_mem_cgroup
;
5666 * mem_cgroup_commit_charge - commit a page charge
5667 * @page: page to charge
5668 * @memcg: memcg to charge the page to
5669 * @lrucare: page might be on LRU already
5671 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5672 * after page->mapping has been set up. This must happen atomically
5673 * as part of the page instantiation, i.e. under the page table lock
5674 * for anonymous pages, under the page lock for page and swap cache.
5676 * In addition, the page must not be on the LRU during the commit, to
5677 * prevent racing with task migration. If it might be, use @lrucare.
5679 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5681 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5684 unsigned int nr_pages
= 1;
5686 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5687 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5689 if (mem_cgroup_disabled())
5692 * Swap faults will attempt to charge the same page multiple
5693 * times. But reuse_swap_page() might have removed the page
5694 * from swapcache already, so we can't check PageSwapCache().
5699 commit_charge(page
, memcg
, lrucare
);
5701 if (PageTransHuge(page
)) {
5702 nr_pages
<<= compound_order(page
);
5703 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5706 local_irq_disable();
5707 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5708 memcg_check_events(memcg
, page
);
5711 if (do_swap_account
&& PageSwapCache(page
)) {
5712 swp_entry_t entry
= { .val
= page_private(page
) };
5714 * The swap entry might not get freed for a long time,
5715 * let's not wait for it. The page already received a
5716 * memory+swap charge, drop the swap entry duplicate.
5718 mem_cgroup_uncharge_swap(entry
);
5723 * mem_cgroup_cancel_charge - cancel a page charge
5724 * @page: page to charge
5725 * @memcg: memcg to charge the page to
5727 * Cancel a charge transaction started by mem_cgroup_try_charge().
5729 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5731 unsigned int nr_pages
= 1;
5733 if (mem_cgroup_disabled())
5736 * Swap faults will attempt to charge the same page multiple
5737 * times. But reuse_swap_page() might have removed the page
5738 * from swapcache already, so we can't check PageSwapCache().
5743 if (PageTransHuge(page
)) {
5744 nr_pages
<<= compound_order(page
);
5745 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5748 cancel_charge(memcg
, nr_pages
);
5751 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5752 unsigned long nr_anon
, unsigned long nr_file
,
5753 unsigned long nr_huge
, struct page
*dummy_page
)
5755 unsigned long nr_pages
= nr_anon
+ nr_file
;
5756 unsigned long flags
;
5758 if (!mem_cgroup_is_root(memcg
)) {
5759 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5760 if (do_swap_account
)
5761 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5762 memcg_oom_recover(memcg
);
5765 local_irq_save(flags
);
5766 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5767 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5768 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5769 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5770 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5771 memcg_check_events(memcg
, dummy_page
);
5772 local_irq_restore(flags
);
5774 if (!mem_cgroup_is_root(memcg
))
5775 css_put_many(&memcg
->css
, nr_pages
);
5778 static void uncharge_list(struct list_head
*page_list
)
5780 struct mem_cgroup
*memcg
= NULL
;
5781 unsigned long nr_anon
= 0;
5782 unsigned long nr_file
= 0;
5783 unsigned long nr_huge
= 0;
5784 unsigned long pgpgout
= 0;
5785 struct list_head
*next
;
5788 next
= page_list
->next
;
5790 unsigned int nr_pages
= 1;
5792 page
= list_entry(next
, struct page
, lru
);
5793 next
= page
->lru
.next
;
5795 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5796 VM_BUG_ON_PAGE(page_count(page
), page
);
5798 if (!page
->mem_cgroup
)
5802 * Nobody should be changing or seriously looking at
5803 * page->mem_cgroup at this point, we have fully
5804 * exclusive access to the page.
5807 if (memcg
!= page
->mem_cgroup
) {
5809 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5811 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5813 memcg
= page
->mem_cgroup
;
5816 if (PageTransHuge(page
)) {
5817 nr_pages
<<= compound_order(page
);
5818 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5819 nr_huge
+= nr_pages
;
5823 nr_anon
+= nr_pages
;
5825 nr_file
+= nr_pages
;
5827 page
->mem_cgroup
= NULL
;
5830 } while (next
!= page_list
);
5833 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5838 * mem_cgroup_uncharge - uncharge a page
5839 * @page: page to uncharge
5841 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5842 * mem_cgroup_commit_charge().
5844 void mem_cgroup_uncharge(struct page
*page
)
5846 if (mem_cgroup_disabled())
5849 /* Don't touch page->lru of any random page, pre-check: */
5850 if (!page
->mem_cgroup
)
5853 INIT_LIST_HEAD(&page
->lru
);
5854 uncharge_list(&page
->lru
);
5858 * mem_cgroup_uncharge_list - uncharge a list of page
5859 * @page_list: list of pages to uncharge
5861 * Uncharge a list of pages previously charged with
5862 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5864 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5866 if (mem_cgroup_disabled())
5869 if (!list_empty(page_list
))
5870 uncharge_list(page_list
);
5874 * mem_cgroup_migrate - migrate a charge to another page
5875 * @oldpage: currently charged page
5876 * @newpage: page to transfer the charge to
5877 * @lrucare: both pages might be on the LRU already
5879 * Migrate the charge from @oldpage to @newpage.
5881 * Both pages must be locked, @newpage->mapping must be set up.
5883 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
5886 struct mem_cgroup
*memcg
;
5889 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5890 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5891 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
5892 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
5893 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5894 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5897 if (mem_cgroup_disabled())
5900 /* Page cache replacement: new page already charged? */
5901 if (newpage
->mem_cgroup
)
5905 * Swapcache readahead pages can get migrated before being
5906 * charged, and migration from compaction can happen to an
5907 * uncharged page when the PFN walker finds a page that
5908 * reclaim just put back on the LRU but has not released yet.
5910 memcg
= oldpage
->mem_cgroup
;
5915 lock_page_lru(oldpage
, &isolated
);
5917 oldpage
->mem_cgroup
= NULL
;
5920 unlock_page_lru(oldpage
, isolated
);
5922 commit_charge(newpage
, memcg
, lrucare
);
5926 * subsys_initcall() for memory controller.
5928 * Some parts like hotcpu_notifier() have to be initialized from this context
5929 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5930 * everything that doesn't depend on a specific mem_cgroup structure should
5931 * be initialized from here.
5933 static int __init
mem_cgroup_init(void)
5935 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5936 enable_swap_cgroup();
5937 mem_cgroup_soft_limit_tree_init();
5941 subsys_initcall(mem_cgroup_init
);