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?
302 atomic_t oom_wakeups
;
305 /* OOM-Killer disable */
306 int oom_kill_disable
;
308 /* protect arrays of thresholds */
309 struct mutex thresholds_lock
;
311 /* thresholds for memory usage. RCU-protected */
312 struct mem_cgroup_thresholds thresholds
;
314 /* thresholds for mem+swap usage. RCU-protected */
315 struct mem_cgroup_thresholds memsw_thresholds
;
317 /* For oom notifier event fd */
318 struct list_head oom_notify
;
321 * Should we move charges of a task when a task is moved into this
322 * mem_cgroup ? And what type of charges should we move ?
324 unsigned long move_charge_at_immigrate
;
326 * set > 0 if pages under this cgroup are moving to other cgroup.
328 atomic_t moving_account
;
329 /* taken only while moving_account > 0 */
330 spinlock_t move_lock
;
334 struct mem_cgroup_stat_cpu __percpu
*stat
;
336 * used when a cpu is offlined or other synchronizations
337 * See mem_cgroup_read_stat().
339 struct mem_cgroup_stat_cpu nocpu_base
;
340 spinlock_t pcp_counter_lock
;
342 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
343 struct cg_proto tcp_mem
;
345 #if defined(CONFIG_MEMCG_KMEM)
346 /* analogous to slab_common's slab_caches list, but per-memcg;
347 * protected by memcg_slab_mutex */
348 struct list_head memcg_slab_caches
;
349 /* Index in the kmem_cache->memcg_params->memcg_caches array */
353 int last_scanned_node
;
355 nodemask_t scan_nodes
;
356 atomic_t numainfo_events
;
357 atomic_t numainfo_updating
;
360 /* List of events which userspace want to receive */
361 struct list_head event_list
;
362 spinlock_t event_list_lock
;
364 struct mem_cgroup_per_node
*nodeinfo
[0];
365 /* WARNING: nodeinfo must be the last member here */
368 #ifdef CONFIG_MEMCG_KMEM
369 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
371 return memcg
->kmemcg_id
>= 0;
375 /* Stuffs for move charges at task migration. */
377 * Types of charges to be moved. "move_charge_at_immitgrate" and
378 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
381 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
382 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
386 /* "mc" and its members are protected by cgroup_mutex */
387 static struct move_charge_struct
{
388 spinlock_t lock
; /* for from, to */
389 struct mem_cgroup
*from
;
390 struct mem_cgroup
*to
;
391 unsigned long immigrate_flags
;
392 unsigned long precharge
;
393 unsigned long moved_charge
;
394 unsigned long moved_swap
;
395 struct task_struct
*moving_task
; /* a task moving charges */
396 wait_queue_head_t waitq
; /* a waitq for other context */
398 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
399 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
402 static bool move_anon(void)
404 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
407 static bool move_file(void)
409 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
413 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
414 * limit reclaim to prevent infinite loops, if they ever occur.
416 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
417 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
420 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
421 MEM_CGROUP_CHARGE_TYPE_ANON
,
422 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
423 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
427 /* for encoding cft->private value on file */
435 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
436 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
437 #define MEMFILE_ATTR(val) ((val) & 0xffff)
438 /* Used for OOM nofiier */
439 #define OOM_CONTROL (0)
442 * The memcg_create_mutex will be held whenever a new cgroup is created.
443 * As a consequence, any change that needs to protect against new child cgroups
444 * appearing has to hold it as well.
446 static DEFINE_MUTEX(memcg_create_mutex
);
448 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
450 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
453 /* Some nice accessors for the vmpressure. */
454 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
457 memcg
= root_mem_cgroup
;
458 return &memcg
->vmpressure
;
461 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
463 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
466 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
468 return (memcg
== root_mem_cgroup
);
472 * We restrict the id in the range of [1, 65535], so it can fit into
475 #define MEM_CGROUP_ID_MAX USHRT_MAX
477 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
479 return memcg
->css
.id
;
482 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
484 struct cgroup_subsys_state
*css
;
486 css
= css_from_id(id
, &memory_cgrp_subsys
);
487 return mem_cgroup_from_css(css
);
490 /* Writing them here to avoid exposing memcg's inner layout */
491 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
493 void sock_update_memcg(struct sock
*sk
)
495 if (mem_cgroup_sockets_enabled
) {
496 struct mem_cgroup
*memcg
;
497 struct cg_proto
*cg_proto
;
499 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
501 /* Socket cloning can throw us here with sk_cgrp already
502 * filled. It won't however, necessarily happen from
503 * process context. So the test for root memcg given
504 * the current task's memcg won't help us in this case.
506 * Respecting the original socket's memcg is a better
507 * decision in this case.
510 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
511 css_get(&sk
->sk_cgrp
->memcg
->css
);
516 memcg
= mem_cgroup_from_task(current
);
517 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
518 if (!mem_cgroup_is_root(memcg
) &&
519 memcg_proto_active(cg_proto
) &&
520 css_tryget_online(&memcg
->css
)) {
521 sk
->sk_cgrp
= cg_proto
;
526 EXPORT_SYMBOL(sock_update_memcg
);
528 void sock_release_memcg(struct sock
*sk
)
530 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
531 struct mem_cgroup
*memcg
;
532 WARN_ON(!sk
->sk_cgrp
->memcg
);
533 memcg
= sk
->sk_cgrp
->memcg
;
534 css_put(&sk
->sk_cgrp
->memcg
->css
);
538 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
540 if (!memcg
|| mem_cgroup_is_root(memcg
))
543 return &memcg
->tcp_mem
;
545 EXPORT_SYMBOL(tcp_proto_cgroup
);
547 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
549 if (!memcg_proto_activated(&memcg
->tcp_mem
))
551 static_key_slow_dec(&memcg_socket_limit_enabled
);
554 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
559 #ifdef CONFIG_MEMCG_KMEM
561 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
562 * The main reason for not using cgroup id for this:
563 * this works better in sparse environments, where we have a lot of memcgs,
564 * but only a few kmem-limited. Or also, if we have, for instance, 200
565 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
566 * 200 entry array for that.
568 * The current size of the caches array is stored in
569 * memcg_limited_groups_array_size. It will double each time we have to
572 static DEFINE_IDA(kmem_limited_groups
);
573 int memcg_limited_groups_array_size
;
576 * MIN_SIZE is different than 1, because we would like to avoid going through
577 * the alloc/free process all the time. In a small machine, 4 kmem-limited
578 * cgroups is a reasonable guess. In the future, it could be a parameter or
579 * tunable, but that is strictly not necessary.
581 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
582 * this constant directly from cgroup, but it is understandable that this is
583 * better kept as an internal representation in cgroup.c. In any case, the
584 * cgrp_id space is not getting any smaller, and we don't have to necessarily
585 * increase ours as well if it increases.
587 #define MEMCG_CACHES_MIN_SIZE 4
588 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
591 * A lot of the calls to the cache allocation functions are expected to be
592 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
593 * conditional to this static branch, we'll have to allow modules that does
594 * kmem_cache_alloc and the such to see this symbol as well
596 struct static_key memcg_kmem_enabled_key
;
597 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
599 static void memcg_free_cache_id(int id
);
601 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
603 if (memcg_kmem_is_active(memcg
)) {
604 static_key_slow_dec(&memcg_kmem_enabled_key
);
605 memcg_free_cache_id(memcg
->kmemcg_id
);
608 * This check can't live in kmem destruction function,
609 * since the charges will outlive the cgroup
611 WARN_ON(page_counter_read(&memcg
->kmem
));
614 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
617 #endif /* CONFIG_MEMCG_KMEM */
619 static void disarm_static_keys(struct mem_cgroup
*memcg
)
621 disarm_sock_keys(memcg
);
622 disarm_kmem_keys(memcg
);
625 static struct mem_cgroup_per_zone
*
626 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
628 int nid
= zone_to_nid(zone
);
629 int zid
= zone_idx(zone
);
631 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
634 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
639 static struct mem_cgroup_per_zone
*
640 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
642 int nid
= page_to_nid(page
);
643 int zid
= page_zonenum(page
);
645 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
648 static struct mem_cgroup_tree_per_zone
*
649 soft_limit_tree_node_zone(int nid
, int zid
)
651 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
654 static struct mem_cgroup_tree_per_zone
*
655 soft_limit_tree_from_page(struct page
*page
)
657 int nid
= page_to_nid(page
);
658 int zid
= page_zonenum(page
);
660 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
663 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
664 struct mem_cgroup_tree_per_zone
*mctz
,
665 unsigned long new_usage_in_excess
)
667 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
668 struct rb_node
*parent
= NULL
;
669 struct mem_cgroup_per_zone
*mz_node
;
674 mz
->usage_in_excess
= new_usage_in_excess
;
675 if (!mz
->usage_in_excess
)
679 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
681 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
684 * We can't avoid mem cgroups that are over their soft
685 * limit by the same amount
687 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
690 rb_link_node(&mz
->tree_node
, parent
, p
);
691 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
695 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
696 struct mem_cgroup_tree_per_zone
*mctz
)
700 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
704 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
705 struct mem_cgroup_tree_per_zone
*mctz
)
709 spin_lock_irqsave(&mctz
->lock
, flags
);
710 __mem_cgroup_remove_exceeded(mz
, mctz
);
711 spin_unlock_irqrestore(&mctz
->lock
, flags
);
714 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
716 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
717 unsigned long soft_limit
= ACCESS_ONCE(memcg
->soft_limit
);
718 unsigned long excess
= 0;
720 if (nr_pages
> soft_limit
)
721 excess
= nr_pages
- soft_limit
;
726 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
728 unsigned long excess
;
729 struct mem_cgroup_per_zone
*mz
;
730 struct mem_cgroup_tree_per_zone
*mctz
;
732 mctz
= soft_limit_tree_from_page(page
);
734 * Necessary to update all ancestors when hierarchy is used.
735 * because their event counter is not touched.
737 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
738 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
739 excess
= soft_limit_excess(memcg
);
741 * We have to update the tree if mz is on RB-tree or
742 * mem is over its softlimit.
744 if (excess
|| mz
->on_tree
) {
747 spin_lock_irqsave(&mctz
->lock
, flags
);
748 /* if on-tree, remove it */
750 __mem_cgroup_remove_exceeded(mz
, mctz
);
752 * Insert again. mz->usage_in_excess will be updated.
753 * If excess is 0, no tree ops.
755 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
756 spin_unlock_irqrestore(&mctz
->lock
, flags
);
761 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
763 struct mem_cgroup_tree_per_zone
*mctz
;
764 struct mem_cgroup_per_zone
*mz
;
768 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
769 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
770 mctz
= soft_limit_tree_node_zone(nid
, zid
);
771 mem_cgroup_remove_exceeded(mz
, mctz
);
776 static struct mem_cgroup_per_zone
*
777 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
779 struct rb_node
*rightmost
= NULL
;
780 struct mem_cgroup_per_zone
*mz
;
784 rightmost
= rb_last(&mctz
->rb_root
);
786 goto done
; /* Nothing to reclaim from */
788 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
790 * Remove the node now but someone else can add it back,
791 * we will to add it back at the end of reclaim to its correct
792 * position in the tree.
794 __mem_cgroup_remove_exceeded(mz
, mctz
);
795 if (!soft_limit_excess(mz
->memcg
) ||
796 !css_tryget_online(&mz
->memcg
->css
))
802 static struct mem_cgroup_per_zone
*
803 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
805 struct mem_cgroup_per_zone
*mz
;
807 spin_lock_irq(&mctz
->lock
);
808 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
809 spin_unlock_irq(&mctz
->lock
);
814 * Implementation Note: reading percpu statistics for memcg.
816 * Both of vmstat[] and percpu_counter has threshold and do periodic
817 * synchronization to implement "quick" read. There are trade-off between
818 * reading cost and precision of value. Then, we may have a chance to implement
819 * a periodic synchronizion of counter in memcg's counter.
821 * But this _read() function is used for user interface now. The user accounts
822 * memory usage by memory cgroup and he _always_ requires exact value because
823 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
824 * have to visit all online cpus and make sum. So, for now, unnecessary
825 * synchronization is not implemented. (just implemented for cpu hotplug)
827 * If there are kernel internal actions which can make use of some not-exact
828 * value, and reading all cpu value can be performance bottleneck in some
829 * common workload, threashold and synchonization as vmstat[] should be
832 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
833 enum mem_cgroup_stat_index idx
)
839 for_each_online_cpu(cpu
)
840 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
841 #ifdef CONFIG_HOTPLUG_CPU
842 spin_lock(&memcg
->pcp_counter_lock
);
843 val
+= memcg
->nocpu_base
.count
[idx
];
844 spin_unlock(&memcg
->pcp_counter_lock
);
850 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
851 enum mem_cgroup_events_index idx
)
853 unsigned long val
= 0;
857 for_each_online_cpu(cpu
)
858 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
859 #ifdef CONFIG_HOTPLUG_CPU
860 spin_lock(&memcg
->pcp_counter_lock
);
861 val
+= memcg
->nocpu_base
.events
[idx
];
862 spin_unlock(&memcg
->pcp_counter_lock
);
868 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
873 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
874 * counted as CACHE even if it's on ANON LRU.
877 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
880 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
883 if (PageTransHuge(page
))
884 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
887 /* pagein of a big page is an event. So, ignore page size */
889 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
891 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
892 nr_pages
= -nr_pages
; /* for event */
895 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
898 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
900 struct mem_cgroup_per_zone
*mz
;
902 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
903 return mz
->lru_size
[lru
];
906 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
908 unsigned int lru_mask
)
910 unsigned long nr
= 0;
913 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
915 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
916 struct mem_cgroup_per_zone
*mz
;
920 if (!(BIT(lru
) & lru_mask
))
922 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
923 nr
+= mz
->lru_size
[lru
];
929 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
930 unsigned int lru_mask
)
932 unsigned long nr
= 0;
935 for_each_node_state(nid
, N_MEMORY
)
936 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
940 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
941 enum mem_cgroup_events_target target
)
943 unsigned long val
, next
;
945 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
946 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
947 /* from time_after() in jiffies.h */
948 if ((long)next
- (long)val
< 0) {
950 case MEM_CGROUP_TARGET_THRESH
:
951 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
953 case MEM_CGROUP_TARGET_SOFTLIMIT
:
954 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
956 case MEM_CGROUP_TARGET_NUMAINFO
:
957 next
= val
+ NUMAINFO_EVENTS_TARGET
;
962 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
969 * Check events in order.
972 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
974 /* threshold event is triggered in finer grain than soft limit */
975 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
976 MEM_CGROUP_TARGET_THRESH
))) {
978 bool do_numainfo __maybe_unused
;
980 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
981 MEM_CGROUP_TARGET_SOFTLIMIT
);
983 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
984 MEM_CGROUP_TARGET_NUMAINFO
);
986 mem_cgroup_threshold(memcg
);
987 if (unlikely(do_softlimit
))
988 mem_cgroup_update_tree(memcg
, page
);
990 if (unlikely(do_numainfo
))
991 atomic_inc(&memcg
->numainfo_events
);
996 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
999 * mm_update_next_owner() may clear mm->owner to NULL
1000 * if it races with swapoff, page migration, etc.
1001 * So this can be called with p == NULL.
1006 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1009 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1011 struct mem_cgroup
*memcg
= NULL
;
1016 * Page cache insertions can happen withou an
1017 * actual mm context, e.g. during disk probing
1018 * on boot, loopback IO, acct() writes etc.
1021 memcg
= root_mem_cgroup
;
1023 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1024 if (unlikely(!memcg
))
1025 memcg
= root_mem_cgroup
;
1027 } while (!css_tryget_online(&memcg
->css
));
1033 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1034 * @root: hierarchy root
1035 * @prev: previously returned memcg, NULL on first invocation
1036 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1038 * Returns references to children of the hierarchy below @root, or
1039 * @root itself, or %NULL after a full round-trip.
1041 * Caller must pass the return value in @prev on subsequent
1042 * invocations for reference counting, or use mem_cgroup_iter_break()
1043 * to cancel a hierarchy walk before the round-trip is complete.
1045 * Reclaimers can specify a zone and a priority level in @reclaim to
1046 * divide up the memcgs in the hierarchy among all concurrent
1047 * reclaimers operating on the same zone and priority.
1049 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1050 struct mem_cgroup
*prev
,
1051 struct mem_cgroup_reclaim_cookie
*reclaim
)
1053 struct reclaim_iter
*uninitialized_var(iter
);
1054 struct cgroup_subsys_state
*css
= NULL
;
1055 struct mem_cgroup
*memcg
= NULL
;
1056 struct mem_cgroup
*pos
= NULL
;
1058 if (mem_cgroup_disabled())
1062 root
= root_mem_cgroup
;
1064 if (prev
&& !reclaim
)
1067 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1076 struct mem_cgroup_per_zone
*mz
;
1078 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1079 iter
= &mz
->iter
[reclaim
->priority
];
1081 if (prev
&& reclaim
->generation
!= iter
->generation
)
1085 pos
= ACCESS_ONCE(iter
->position
);
1087 * A racing update may change the position and
1088 * put the last reference, hence css_tryget(),
1089 * or retry to see the updated position.
1091 } while (pos
&& !css_tryget(&pos
->css
));
1098 css
= css_next_descendant_pre(css
, &root
->css
);
1101 * Reclaimers share the hierarchy walk, and a
1102 * new one might jump in right at the end of
1103 * the hierarchy - make sure they see at least
1104 * one group and restart from the beginning.
1112 * Verify the css and acquire a reference. The root
1113 * is provided by the caller, so we know it's alive
1114 * and kicking, and don't take an extra reference.
1116 memcg
= mem_cgroup_from_css(css
);
1118 if (css
== &root
->css
)
1121 if (css_tryget(css
)) {
1123 * Make sure the memcg is initialized:
1124 * mem_cgroup_css_online() orders the the
1125 * initialization against setting the flag.
1127 if (smp_load_acquire(&memcg
->initialized
))
1137 if (cmpxchg(&iter
->position
, pos
, memcg
) == pos
) {
1139 css_get(&memcg
->css
);
1145 * pairs with css_tryget when dereferencing iter->position
1154 reclaim
->generation
= iter
->generation
;
1160 if (prev
&& prev
!= root
)
1161 css_put(&prev
->css
);
1167 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1168 * @root: hierarchy root
1169 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1171 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1172 struct mem_cgroup
*prev
)
1175 root
= root_mem_cgroup
;
1176 if (prev
&& prev
!= root
)
1177 css_put(&prev
->css
);
1181 * Iteration constructs for visiting all cgroups (under a tree). If
1182 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1183 * be used for reference counting.
1185 #define for_each_mem_cgroup_tree(iter, root) \
1186 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1188 iter = mem_cgroup_iter(root, iter, NULL))
1190 #define for_each_mem_cgroup(iter) \
1191 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1193 iter = mem_cgroup_iter(NULL, iter, NULL))
1195 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1197 struct mem_cgroup
*memcg
;
1200 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1201 if (unlikely(!memcg
))
1206 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1209 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1217 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1220 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1221 * @zone: zone of the wanted lruvec
1222 * @memcg: memcg of the wanted lruvec
1224 * Returns the lru list vector holding pages for the given @zone and
1225 * @mem. This can be the global zone lruvec, if the memory controller
1228 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1229 struct mem_cgroup
*memcg
)
1231 struct mem_cgroup_per_zone
*mz
;
1232 struct lruvec
*lruvec
;
1234 if (mem_cgroup_disabled()) {
1235 lruvec
= &zone
->lruvec
;
1239 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1240 lruvec
= &mz
->lruvec
;
1243 * Since a node can be onlined after the mem_cgroup was created,
1244 * we have to be prepared to initialize lruvec->zone here;
1245 * and if offlined then reonlined, we need to reinitialize it.
1247 if (unlikely(lruvec
->zone
!= zone
))
1248 lruvec
->zone
= zone
;
1253 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1255 * @zone: zone of the page
1257 * This function is only safe when following the LRU page isolation
1258 * and putback protocol: the LRU lock must be held, and the page must
1259 * either be PageLRU() or the caller must have isolated/allocated it.
1261 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1263 struct mem_cgroup_per_zone
*mz
;
1264 struct mem_cgroup
*memcg
;
1265 struct lruvec
*lruvec
;
1267 if (mem_cgroup_disabled()) {
1268 lruvec
= &zone
->lruvec
;
1272 memcg
= page
->mem_cgroup
;
1274 * Swapcache readahead pages are added to the LRU - and
1275 * possibly migrated - before they are charged.
1278 memcg
= root_mem_cgroup
;
1280 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1281 lruvec
= &mz
->lruvec
;
1284 * Since a node can be onlined after the mem_cgroup was created,
1285 * we have to be prepared to initialize lruvec->zone here;
1286 * and if offlined then reonlined, we need to reinitialize it.
1288 if (unlikely(lruvec
->zone
!= zone
))
1289 lruvec
->zone
= zone
;
1294 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1295 * @lruvec: mem_cgroup per zone lru vector
1296 * @lru: index of lru list the page is sitting on
1297 * @nr_pages: positive when adding or negative when removing
1299 * This function must be called when a page is added to or removed from an
1302 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1305 struct mem_cgroup_per_zone
*mz
;
1306 unsigned long *lru_size
;
1308 if (mem_cgroup_disabled())
1311 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1312 lru_size
= mz
->lru_size
+ lru
;
1313 *lru_size
+= nr_pages
;
1314 VM_BUG_ON((long)(*lru_size
) < 0);
1317 bool mem_cgroup_is_descendant(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
)
1321 if (!root
->use_hierarchy
)
1323 return cgroup_is_descendant(memcg
->css
.cgroup
, root
->css
.cgroup
);
1326 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1328 struct mem_cgroup
*task_memcg
;
1329 struct task_struct
*p
;
1332 p
= find_lock_task_mm(task
);
1334 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1338 * All threads may have already detached their mm's, but the oom
1339 * killer still needs to detect if they have already been oom
1340 * killed to prevent needlessly killing additional tasks.
1343 task_memcg
= mem_cgroup_from_task(task
);
1344 css_get(&task_memcg
->css
);
1347 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1348 css_put(&task_memcg
->css
);
1352 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1354 unsigned long inactive_ratio
;
1355 unsigned long inactive
;
1356 unsigned long active
;
1359 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1360 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1362 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1364 inactive_ratio
= int_sqrt(10 * gb
);
1368 return inactive
* inactive_ratio
< active
;
1371 #define mem_cgroup_from_counter(counter, member) \
1372 container_of(counter, struct mem_cgroup, member)
1375 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1376 * @memcg: the memory cgroup
1378 * Returns the maximum amount of memory @mem can be charged with, in
1381 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1383 unsigned long margin
= 0;
1384 unsigned long count
;
1385 unsigned long limit
;
1387 count
= page_counter_read(&memcg
->memory
);
1388 limit
= ACCESS_ONCE(memcg
->memory
.limit
);
1390 margin
= limit
- count
;
1392 if (do_swap_account
) {
1393 count
= page_counter_read(&memcg
->memsw
);
1394 limit
= ACCESS_ONCE(memcg
->memsw
.limit
);
1396 margin
= min(margin
, limit
- count
);
1402 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1405 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1406 return vm_swappiness
;
1408 return memcg
->swappiness
;
1412 * A routine for checking "mem" is under move_account() or not.
1414 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1415 * moving cgroups. This is for waiting at high-memory pressure
1418 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1420 struct mem_cgroup
*from
;
1421 struct mem_cgroup
*to
;
1424 * Unlike task_move routines, we access mc.to, mc.from not under
1425 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1427 spin_lock(&mc
.lock
);
1433 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1434 mem_cgroup_is_descendant(to
, memcg
);
1436 spin_unlock(&mc
.lock
);
1440 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1442 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1443 if (mem_cgroup_under_move(memcg
)) {
1445 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1446 /* moving charge context might have finished. */
1449 finish_wait(&mc
.waitq
, &wait
);
1456 #define K(x) ((x) << (PAGE_SHIFT-10))
1458 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1459 * @memcg: The memory cgroup that went over limit
1460 * @p: Task that is going to be killed
1462 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1465 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1467 /* oom_info_lock ensures that parallel ooms do not interleave */
1468 static DEFINE_MUTEX(oom_info_lock
);
1469 struct mem_cgroup
*iter
;
1475 mutex_lock(&oom_info_lock
);
1478 pr_info("Task in ");
1479 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1480 pr_cont(" killed as a result of limit of ");
1481 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1486 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1487 K((u64
)page_counter_read(&memcg
->memory
)),
1488 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1489 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1490 K((u64
)page_counter_read(&memcg
->memsw
)),
1491 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1492 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1493 K((u64
)page_counter_read(&memcg
->kmem
)),
1494 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1496 for_each_mem_cgroup_tree(iter
, memcg
) {
1497 pr_info("Memory cgroup stats for ");
1498 pr_cont_cgroup_path(iter
->css
.cgroup
);
1501 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1502 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1504 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1505 K(mem_cgroup_read_stat(iter
, i
)));
1508 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1509 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1510 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1514 mutex_unlock(&oom_info_lock
);
1518 * This function returns the number of memcg under hierarchy tree. Returns
1519 * 1(self count) if no children.
1521 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1524 struct mem_cgroup
*iter
;
1526 for_each_mem_cgroup_tree(iter
, memcg
)
1532 * Return the memory (and swap, if configured) limit for a memcg.
1534 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1536 unsigned long limit
;
1538 limit
= memcg
->memory
.limit
;
1539 if (mem_cgroup_swappiness(memcg
)) {
1540 unsigned long memsw_limit
;
1542 memsw_limit
= memcg
->memsw
.limit
;
1543 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1548 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1551 struct mem_cgroup
*iter
;
1552 unsigned long chosen_points
= 0;
1553 unsigned long totalpages
;
1554 unsigned int points
= 0;
1555 struct task_struct
*chosen
= NULL
;
1558 * If current has a pending SIGKILL or is exiting, then automatically
1559 * select it. The goal is to allow it to allocate so that it may
1560 * quickly exit and free its memory.
1562 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1563 set_thread_flag(TIF_MEMDIE
);
1567 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1568 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1569 for_each_mem_cgroup_tree(iter
, memcg
) {
1570 struct css_task_iter it
;
1571 struct task_struct
*task
;
1573 css_task_iter_start(&iter
->css
, &it
);
1574 while ((task
= css_task_iter_next(&it
))) {
1575 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1577 case OOM_SCAN_SELECT
:
1579 put_task_struct(chosen
);
1581 chosen_points
= ULONG_MAX
;
1582 get_task_struct(chosen
);
1584 case OOM_SCAN_CONTINUE
:
1586 case OOM_SCAN_ABORT
:
1587 css_task_iter_end(&it
);
1588 mem_cgroup_iter_break(memcg
, iter
);
1590 put_task_struct(chosen
);
1595 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1596 if (!points
|| points
< chosen_points
)
1598 /* Prefer thread group leaders for display purposes */
1599 if (points
== chosen_points
&&
1600 thread_group_leader(chosen
))
1604 put_task_struct(chosen
);
1606 chosen_points
= points
;
1607 get_task_struct(chosen
);
1609 css_task_iter_end(&it
);
1614 points
= chosen_points
* 1000 / totalpages
;
1615 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1616 NULL
, "Memory cgroup out of memory");
1619 #if MAX_NUMNODES > 1
1622 * test_mem_cgroup_node_reclaimable
1623 * @memcg: the target memcg
1624 * @nid: the node ID to be checked.
1625 * @noswap : specify true here if the user wants flle only information.
1627 * This function returns whether the specified memcg contains any
1628 * reclaimable pages on a node. Returns true if there are any reclaimable
1629 * pages in the node.
1631 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1632 int nid
, bool noswap
)
1634 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1636 if (noswap
|| !total_swap_pages
)
1638 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1645 * Always updating the nodemask is not very good - even if we have an empty
1646 * list or the wrong list here, we can start from some node and traverse all
1647 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1650 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1654 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1655 * pagein/pageout changes since the last update.
1657 if (!atomic_read(&memcg
->numainfo_events
))
1659 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1662 /* make a nodemask where this memcg uses memory from */
1663 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1665 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1667 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1668 node_clear(nid
, memcg
->scan_nodes
);
1671 atomic_set(&memcg
->numainfo_events
, 0);
1672 atomic_set(&memcg
->numainfo_updating
, 0);
1676 * Selecting a node where we start reclaim from. Because what we need is just
1677 * reducing usage counter, start from anywhere is O,K. Considering
1678 * memory reclaim from current node, there are pros. and cons.
1680 * Freeing memory from current node means freeing memory from a node which
1681 * we'll use or we've used. So, it may make LRU bad. And if several threads
1682 * hit limits, it will see a contention on a node. But freeing from remote
1683 * node means more costs for memory reclaim because of memory latency.
1685 * Now, we use round-robin. Better algorithm is welcomed.
1687 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1691 mem_cgroup_may_update_nodemask(memcg
);
1692 node
= memcg
->last_scanned_node
;
1694 node
= next_node(node
, memcg
->scan_nodes
);
1695 if (node
== MAX_NUMNODES
)
1696 node
= first_node(memcg
->scan_nodes
);
1698 * We call this when we hit limit, not when pages are added to LRU.
1699 * No LRU may hold pages because all pages are UNEVICTABLE or
1700 * memcg is too small and all pages are not on LRU. In that case,
1701 * we use curret node.
1703 if (unlikely(node
== MAX_NUMNODES
))
1704 node
= numa_node_id();
1706 memcg
->last_scanned_node
= node
;
1710 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1716 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1719 unsigned long *total_scanned
)
1721 struct mem_cgroup
*victim
= NULL
;
1724 unsigned long excess
;
1725 unsigned long nr_scanned
;
1726 struct mem_cgroup_reclaim_cookie reclaim
= {
1731 excess
= soft_limit_excess(root_memcg
);
1734 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1739 * If we have not been able to reclaim
1740 * anything, it might because there are
1741 * no reclaimable pages under this hierarchy
1746 * We want to do more targeted reclaim.
1747 * excess >> 2 is not to excessive so as to
1748 * reclaim too much, nor too less that we keep
1749 * coming back to reclaim from this cgroup
1751 if (total
>= (excess
>> 2) ||
1752 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1757 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1759 *total_scanned
+= nr_scanned
;
1760 if (!soft_limit_excess(root_memcg
))
1763 mem_cgroup_iter_break(root_memcg
, victim
);
1767 #ifdef CONFIG_LOCKDEP
1768 static struct lockdep_map memcg_oom_lock_dep_map
= {
1769 .name
= "memcg_oom_lock",
1773 static DEFINE_SPINLOCK(memcg_oom_lock
);
1776 * Check OOM-Killer is already running under our hierarchy.
1777 * If someone is running, return false.
1779 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1781 struct mem_cgroup
*iter
, *failed
= NULL
;
1783 spin_lock(&memcg_oom_lock
);
1785 for_each_mem_cgroup_tree(iter
, memcg
) {
1786 if (iter
->oom_lock
) {
1788 * this subtree of our hierarchy is already locked
1789 * so we cannot give a lock.
1792 mem_cgroup_iter_break(memcg
, iter
);
1795 iter
->oom_lock
= true;
1800 * OK, we failed to lock the whole subtree so we have
1801 * to clean up what we set up to the failing subtree
1803 for_each_mem_cgroup_tree(iter
, memcg
) {
1804 if (iter
== failed
) {
1805 mem_cgroup_iter_break(memcg
, iter
);
1808 iter
->oom_lock
= false;
1811 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1813 spin_unlock(&memcg_oom_lock
);
1818 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1820 struct mem_cgroup
*iter
;
1822 spin_lock(&memcg_oom_lock
);
1823 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1824 for_each_mem_cgroup_tree(iter
, memcg
)
1825 iter
->oom_lock
= false;
1826 spin_unlock(&memcg_oom_lock
);
1829 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1831 struct mem_cgroup
*iter
;
1833 for_each_mem_cgroup_tree(iter
, memcg
)
1834 atomic_inc(&iter
->under_oom
);
1837 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1839 struct mem_cgroup
*iter
;
1842 * When a new child is created while the hierarchy is under oom,
1843 * mem_cgroup_oom_lock() may not be called. We have to use
1844 * atomic_add_unless() here.
1846 for_each_mem_cgroup_tree(iter
, memcg
)
1847 atomic_add_unless(&iter
->under_oom
, -1, 0);
1850 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1852 struct oom_wait_info
{
1853 struct mem_cgroup
*memcg
;
1857 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1858 unsigned mode
, int sync
, void *arg
)
1860 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1861 struct mem_cgroup
*oom_wait_memcg
;
1862 struct oom_wait_info
*oom_wait_info
;
1864 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1865 oom_wait_memcg
= oom_wait_info
->memcg
;
1867 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1868 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1870 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1873 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
1875 atomic_inc(&memcg
->oom_wakeups
);
1876 /* for filtering, pass "memcg" as argument. */
1877 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1880 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1882 if (memcg
&& atomic_read(&memcg
->under_oom
))
1883 memcg_wakeup_oom(memcg
);
1886 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1888 if (!current
->memcg_oom
.may_oom
)
1891 * We are in the middle of the charge context here, so we
1892 * don't want to block when potentially sitting on a callstack
1893 * that holds all kinds of filesystem and mm locks.
1895 * Also, the caller may handle a failed allocation gracefully
1896 * (like optional page cache readahead) and so an OOM killer
1897 * invocation might not even be necessary.
1899 * That's why we don't do anything here except remember the
1900 * OOM context and then deal with it at the end of the page
1901 * fault when the stack is unwound, the locks are released,
1902 * and when we know whether the fault was overall successful.
1904 css_get(&memcg
->css
);
1905 current
->memcg_oom
.memcg
= memcg
;
1906 current
->memcg_oom
.gfp_mask
= mask
;
1907 current
->memcg_oom
.order
= order
;
1911 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1912 * @handle: actually kill/wait or just clean up the OOM state
1914 * This has to be called at the end of a page fault if the memcg OOM
1915 * handler was enabled.
1917 * Memcg supports userspace OOM handling where failed allocations must
1918 * sleep on a waitqueue until the userspace task resolves the
1919 * situation. Sleeping directly in the charge context with all kinds
1920 * of locks held is not a good idea, instead we remember an OOM state
1921 * in the task and mem_cgroup_oom_synchronize() has to be called at
1922 * the end of the page fault to complete the OOM handling.
1924 * Returns %true if an ongoing memcg OOM situation was detected and
1925 * completed, %false otherwise.
1927 bool mem_cgroup_oom_synchronize(bool handle
)
1929 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
1930 struct oom_wait_info owait
;
1933 /* OOM is global, do not handle */
1940 owait
.memcg
= memcg
;
1941 owait
.wait
.flags
= 0;
1942 owait
.wait
.func
= memcg_oom_wake_function
;
1943 owait
.wait
.private = current
;
1944 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1946 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1947 mem_cgroup_mark_under_oom(memcg
);
1949 locked
= mem_cgroup_oom_trylock(memcg
);
1952 mem_cgroup_oom_notify(memcg
);
1954 if (locked
&& !memcg
->oom_kill_disable
) {
1955 mem_cgroup_unmark_under_oom(memcg
);
1956 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1957 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
1958 current
->memcg_oom
.order
);
1961 mem_cgroup_unmark_under_oom(memcg
);
1962 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1966 mem_cgroup_oom_unlock(memcg
);
1968 * There is no guarantee that an OOM-lock contender
1969 * sees the wakeups triggered by the OOM kill
1970 * uncharges. Wake any sleepers explicitely.
1972 memcg_oom_recover(memcg
);
1975 current
->memcg_oom
.memcg
= NULL
;
1976 css_put(&memcg
->css
);
1981 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1982 * @page: page that is going to change accounted state
1983 * @locked: &memcg->move_lock slowpath was taken
1984 * @flags: IRQ-state flags for &memcg->move_lock
1986 * This function must mark the beginning of an accounted page state
1987 * change to prevent double accounting when the page is concurrently
1988 * being moved to another memcg:
1990 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
1991 * if (TestClearPageState(page))
1992 * mem_cgroup_update_page_stat(memcg, state, -1);
1993 * mem_cgroup_end_page_stat(memcg, locked, flags);
1995 * The RCU lock is held throughout the transaction. The fast path can
1996 * get away without acquiring the memcg->move_lock (@locked is false)
1997 * because page moving starts with an RCU grace period.
1999 * The RCU lock also protects the memcg from being freed when the page
2000 * state that is going to change is the only thing preventing the page
2001 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2002 * which allows migration to go ahead and uncharge the page before the
2003 * account transaction might be complete.
2005 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
,
2007 unsigned long *flags
)
2009 struct mem_cgroup
*memcg
;
2013 if (mem_cgroup_disabled())
2016 memcg
= page
->mem_cgroup
;
2017 if (unlikely(!memcg
))
2021 if (atomic_read(&memcg
->moving_account
) <= 0)
2024 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
2025 if (memcg
!= page
->mem_cgroup
) {
2026 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
2035 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2036 * @memcg: the memcg that was accounted against
2037 * @locked: value received from mem_cgroup_begin_page_stat()
2038 * @flags: value received from mem_cgroup_begin_page_stat()
2040 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
, bool *locked
,
2041 unsigned long *flags
)
2043 if (memcg
&& *locked
)
2044 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
2050 * mem_cgroup_update_page_stat - update page state statistics
2051 * @memcg: memcg to account against
2052 * @idx: page state item to account
2053 * @val: number of pages (positive or negative)
2055 * See mem_cgroup_begin_page_stat() for locking requirements.
2057 void mem_cgroup_update_page_stat(struct mem_cgroup
*memcg
,
2058 enum mem_cgroup_stat_index idx
, int val
)
2060 VM_BUG_ON(!rcu_read_lock_held());
2063 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2067 * size of first charge trial. "32" comes from vmscan.c's magic value.
2068 * TODO: maybe necessary to use big numbers in big irons.
2070 #define CHARGE_BATCH 32U
2071 struct memcg_stock_pcp
{
2072 struct mem_cgroup
*cached
; /* this never be root cgroup */
2073 unsigned int nr_pages
;
2074 struct work_struct work
;
2075 unsigned long flags
;
2076 #define FLUSHING_CACHED_CHARGE 0
2078 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2079 static DEFINE_MUTEX(percpu_charge_mutex
);
2082 * consume_stock: Try to consume stocked charge on this cpu.
2083 * @memcg: memcg to consume from.
2084 * @nr_pages: how many pages to charge.
2086 * The charges will only happen if @memcg matches the current cpu's memcg
2087 * stock, and at least @nr_pages are available in that stock. Failure to
2088 * service an allocation will refill the stock.
2090 * returns true if successful, false otherwise.
2092 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2094 struct memcg_stock_pcp
*stock
;
2097 if (nr_pages
> CHARGE_BATCH
)
2100 stock
= &get_cpu_var(memcg_stock
);
2101 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2102 stock
->nr_pages
-= nr_pages
;
2105 put_cpu_var(memcg_stock
);
2110 * Returns stocks cached in percpu and reset cached information.
2112 static void drain_stock(struct memcg_stock_pcp
*stock
)
2114 struct mem_cgroup
*old
= stock
->cached
;
2116 if (stock
->nr_pages
) {
2117 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2118 if (do_swap_account
)
2119 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2120 css_put_many(&old
->css
, stock
->nr_pages
);
2121 stock
->nr_pages
= 0;
2123 stock
->cached
= NULL
;
2127 * This must be called under preempt disabled or must be called by
2128 * a thread which is pinned to local cpu.
2130 static void drain_local_stock(struct work_struct
*dummy
)
2132 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2134 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2137 static void __init
memcg_stock_init(void)
2141 for_each_possible_cpu(cpu
) {
2142 struct memcg_stock_pcp
*stock
=
2143 &per_cpu(memcg_stock
, cpu
);
2144 INIT_WORK(&stock
->work
, drain_local_stock
);
2149 * Cache charges(val) to local per_cpu area.
2150 * This will be consumed by consume_stock() function, later.
2152 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2154 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2156 if (stock
->cached
!= memcg
) { /* reset if necessary */
2158 stock
->cached
= memcg
;
2160 stock
->nr_pages
+= nr_pages
;
2161 put_cpu_var(memcg_stock
);
2165 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2166 * of the hierarchy under it.
2168 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2172 /* If someone's already draining, avoid adding running more workers. */
2173 if (!mutex_trylock(&percpu_charge_mutex
))
2175 /* Notify other cpus that system-wide "drain" is running */
2178 for_each_online_cpu(cpu
) {
2179 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2180 struct mem_cgroup
*memcg
;
2182 memcg
= stock
->cached
;
2183 if (!memcg
|| !stock
->nr_pages
)
2185 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
2187 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2189 drain_local_stock(&stock
->work
);
2191 schedule_work_on(cpu
, &stock
->work
);
2196 mutex_unlock(&percpu_charge_mutex
);
2200 * This function drains percpu counter value from DEAD cpu and
2201 * move it to local cpu. Note that this function can be preempted.
2203 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2207 spin_lock(&memcg
->pcp_counter_lock
);
2208 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2209 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2211 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2212 memcg
->nocpu_base
.count
[i
] += x
;
2214 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2215 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2217 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2218 memcg
->nocpu_base
.events
[i
] += x
;
2220 spin_unlock(&memcg
->pcp_counter_lock
);
2223 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2224 unsigned long action
,
2227 int cpu
= (unsigned long)hcpu
;
2228 struct memcg_stock_pcp
*stock
;
2229 struct mem_cgroup
*iter
;
2231 if (action
== CPU_ONLINE
)
2234 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2237 for_each_mem_cgroup(iter
)
2238 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2240 stock
= &per_cpu(memcg_stock
, cpu
);
2245 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2246 unsigned int nr_pages
)
2248 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2249 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2250 struct mem_cgroup
*mem_over_limit
;
2251 struct page_counter
*counter
;
2252 unsigned long nr_reclaimed
;
2253 bool may_swap
= true;
2254 bool drained
= false;
2257 if (mem_cgroup_is_root(memcg
))
2260 if (consume_stock(memcg
, nr_pages
))
2263 if (!do_swap_account
||
2264 !page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2265 if (!page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2267 if (do_swap_account
)
2268 page_counter_uncharge(&memcg
->memsw
, batch
);
2269 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2271 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2275 if (batch
> nr_pages
) {
2281 * Unlike in global OOM situations, memcg is not in a physical
2282 * memory shortage. Allow dying and OOM-killed tasks to
2283 * bypass the last charges so that they can exit quickly and
2284 * free their memory.
2286 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2287 fatal_signal_pending(current
) ||
2288 current
->flags
& PF_EXITING
))
2291 if (unlikely(task_in_memcg_oom(current
)))
2294 if (!(gfp_mask
& __GFP_WAIT
))
2297 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2298 gfp_mask
, may_swap
);
2300 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2304 drain_all_stock(mem_over_limit
);
2309 if (gfp_mask
& __GFP_NORETRY
)
2312 * Even though the limit is exceeded at this point, reclaim
2313 * may have been able to free some pages. Retry the charge
2314 * before killing the task.
2316 * Only for regular pages, though: huge pages are rather
2317 * unlikely to succeed so close to the limit, and we fall back
2318 * to regular pages anyway in case of failure.
2320 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2323 * At task move, charge accounts can be doubly counted. So, it's
2324 * better to wait until the end of task_move if something is going on.
2326 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2332 if (gfp_mask
& __GFP_NOFAIL
)
2335 if (fatal_signal_pending(current
))
2338 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2340 if (!(gfp_mask
& __GFP_NOFAIL
))
2346 css_get_many(&memcg
->css
, batch
);
2347 if (batch
> nr_pages
)
2348 refill_stock(memcg
, batch
- nr_pages
);
2353 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2355 if (mem_cgroup_is_root(memcg
))
2358 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2359 if (do_swap_account
)
2360 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2362 css_put_many(&memcg
->css
, nr_pages
);
2366 * A helper function to get mem_cgroup from ID. must be called under
2367 * rcu_read_lock(). The caller is responsible for calling
2368 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2369 * refcnt from swap can be called against removed memcg.)
2371 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2373 /* ID 0 is unused ID */
2376 return mem_cgroup_from_id(id
);
2380 * try_get_mem_cgroup_from_page - look up page's memcg association
2383 * Look up, get a css reference, and return the memcg that owns @page.
2385 * The page must be locked to prevent racing with swap-in and page
2386 * cache charges. If coming from an unlocked page table, the caller
2387 * must ensure the page is on the LRU or this can race with charging.
2389 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2391 struct mem_cgroup
*memcg
;
2395 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2397 memcg
= page
->mem_cgroup
;
2399 if (!css_tryget_online(&memcg
->css
))
2401 } else if (PageSwapCache(page
)) {
2402 ent
.val
= page_private(page
);
2403 id
= lookup_swap_cgroup_id(ent
);
2405 memcg
= mem_cgroup_lookup(id
);
2406 if (memcg
&& !css_tryget_online(&memcg
->css
))
2413 static void lock_page_lru(struct page
*page
, int *isolated
)
2415 struct zone
*zone
= page_zone(page
);
2417 spin_lock_irq(&zone
->lru_lock
);
2418 if (PageLRU(page
)) {
2419 struct lruvec
*lruvec
;
2421 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2423 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2429 static void unlock_page_lru(struct page
*page
, int isolated
)
2431 struct zone
*zone
= page_zone(page
);
2434 struct lruvec
*lruvec
;
2436 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2437 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2439 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2441 spin_unlock_irq(&zone
->lru_lock
);
2444 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2449 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2452 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2453 * may already be on some other mem_cgroup's LRU. Take care of it.
2456 lock_page_lru(page
, &isolated
);
2459 * Nobody should be changing or seriously looking at
2460 * page->mem_cgroup at this point:
2462 * - the page is uncharged
2464 * - the page is off-LRU
2466 * - an anonymous fault has exclusive page access, except for
2467 * a locked page table
2469 * - a page cache insertion, a swapin fault, or a migration
2470 * have the page locked
2472 page
->mem_cgroup
= memcg
;
2475 unlock_page_lru(page
, isolated
);
2478 #ifdef CONFIG_MEMCG_KMEM
2480 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2481 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2483 static DEFINE_MUTEX(memcg_slab_mutex
);
2486 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2487 * in the memcg_cache_params struct.
2489 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2491 struct kmem_cache
*cachep
;
2493 VM_BUG_ON(p
->is_root_cache
);
2494 cachep
= p
->root_cache
;
2495 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2498 int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
,
2499 unsigned long nr_pages
)
2501 struct page_counter
*counter
;
2504 ret
= page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
);
2508 ret
= try_charge(memcg
, gfp
, nr_pages
);
2509 if (ret
== -EINTR
) {
2511 * try_charge() chose to bypass to root due to OOM kill or
2512 * fatal signal. Since our only options are to either fail
2513 * the allocation or charge it to this cgroup, do it as a
2514 * temporary condition. But we can't fail. From a kmem/slab
2515 * perspective, the cache has already been selected, by
2516 * mem_cgroup_kmem_get_cache(), so it is too late to change
2519 * This condition will only trigger if the task entered
2520 * memcg_charge_kmem in a sane state, but was OOM-killed
2521 * during try_charge() above. Tasks that were already dying
2522 * when the allocation triggers should have been already
2523 * directed to the root cgroup in memcontrol.h
2525 page_counter_charge(&memcg
->memory
, nr_pages
);
2526 if (do_swap_account
)
2527 page_counter_charge(&memcg
->memsw
, nr_pages
);
2528 css_get_many(&memcg
->css
, nr_pages
);
2531 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2536 void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, unsigned long nr_pages
)
2538 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2539 if (do_swap_account
)
2540 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2542 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2544 css_put_many(&memcg
->css
, nr_pages
);
2548 * helper for acessing a memcg's index. It will be used as an index in the
2549 * child cache array in kmem_cache, and also to derive its name. This function
2550 * will return -1 when this is not a kmem-limited memcg.
2552 int memcg_cache_id(struct mem_cgroup
*memcg
)
2554 return memcg
? memcg
->kmemcg_id
: -1;
2557 static int memcg_alloc_cache_id(void)
2562 id
= ida_simple_get(&kmem_limited_groups
,
2563 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2567 if (id
< memcg_limited_groups_array_size
)
2571 * There's no space for the new id in memcg_caches arrays,
2572 * so we have to grow them.
2575 size
= 2 * (id
+ 1);
2576 if (size
< MEMCG_CACHES_MIN_SIZE
)
2577 size
= MEMCG_CACHES_MIN_SIZE
;
2578 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2579 size
= MEMCG_CACHES_MAX_SIZE
;
2581 mutex_lock(&memcg_slab_mutex
);
2582 err
= memcg_update_all_caches(size
);
2583 mutex_unlock(&memcg_slab_mutex
);
2586 ida_simple_remove(&kmem_limited_groups
, id
);
2592 static void memcg_free_cache_id(int id
)
2594 ida_simple_remove(&kmem_limited_groups
, id
);
2598 * We should update the current array size iff all caches updates succeed. This
2599 * can only be done from the slab side. The slab mutex needs to be held when
2602 void memcg_update_array_size(int num
)
2604 memcg_limited_groups_array_size
= num
;
2607 static void memcg_register_cache(struct mem_cgroup
*memcg
,
2608 struct kmem_cache
*root_cache
)
2610 struct kmem_cache
*cachep
;
2613 lockdep_assert_held(&memcg_slab_mutex
);
2615 id
= memcg_cache_id(memcg
);
2618 * Since per-memcg caches are created asynchronously on first
2619 * allocation (see memcg_kmem_get_cache()), several threads can try to
2620 * create the same cache, but only one of them may succeed.
2622 if (cache_from_memcg_idx(root_cache
, id
))
2625 cachep
= memcg_create_kmem_cache(memcg
, root_cache
);
2627 * If we could not create a memcg cache, do not complain, because
2628 * that's not critical at all as we can always proceed with the root
2634 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2637 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2638 * barrier here to ensure nobody will see the kmem_cache partially
2643 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
2644 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
2647 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
2649 struct kmem_cache
*root_cache
;
2650 struct mem_cgroup
*memcg
;
2653 lockdep_assert_held(&memcg_slab_mutex
);
2655 BUG_ON(is_root_cache(cachep
));
2657 root_cache
= cachep
->memcg_params
->root_cache
;
2658 memcg
= cachep
->memcg_params
->memcg
;
2659 id
= memcg_cache_id(memcg
);
2661 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
2662 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
2664 list_del(&cachep
->memcg_params
->list
);
2666 kmem_cache_destroy(cachep
);
2669 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
2671 struct kmem_cache
*c
;
2674 mutex_lock(&memcg_slab_mutex
);
2675 for_each_memcg_cache_index(i
) {
2676 c
= cache_from_memcg_idx(s
, i
);
2680 memcg_unregister_cache(c
);
2682 if (cache_from_memcg_idx(s
, i
))
2685 mutex_unlock(&memcg_slab_mutex
);
2689 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
2691 struct kmem_cache
*cachep
;
2692 struct memcg_cache_params
*params
, *tmp
;
2694 if (!memcg_kmem_is_active(memcg
))
2697 mutex_lock(&memcg_slab_mutex
);
2698 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
2699 cachep
= memcg_params_to_cache(params
);
2700 memcg_unregister_cache(cachep
);
2702 mutex_unlock(&memcg_slab_mutex
);
2705 struct memcg_register_cache_work
{
2706 struct mem_cgroup
*memcg
;
2707 struct kmem_cache
*cachep
;
2708 struct work_struct work
;
2711 static void memcg_register_cache_func(struct work_struct
*w
)
2713 struct memcg_register_cache_work
*cw
=
2714 container_of(w
, struct memcg_register_cache_work
, work
);
2715 struct mem_cgroup
*memcg
= cw
->memcg
;
2716 struct kmem_cache
*cachep
= cw
->cachep
;
2718 mutex_lock(&memcg_slab_mutex
);
2719 memcg_register_cache(memcg
, cachep
);
2720 mutex_unlock(&memcg_slab_mutex
);
2722 css_put(&memcg
->css
);
2727 * Enqueue the creation of a per-memcg kmem_cache.
2729 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
2730 struct kmem_cache
*cachep
)
2732 struct memcg_register_cache_work
*cw
;
2734 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2738 css_get(&memcg
->css
);
2741 cw
->cachep
= cachep
;
2743 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
2744 schedule_work(&cw
->work
);
2747 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
2748 struct kmem_cache
*cachep
)
2751 * We need to stop accounting when we kmalloc, because if the
2752 * corresponding kmalloc cache is not yet created, the first allocation
2753 * in __memcg_schedule_register_cache will recurse.
2755 * However, it is better to enclose the whole function. Depending on
2756 * the debugging options enabled, INIT_WORK(), for instance, can
2757 * trigger an allocation. This too, will make us recurse. Because at
2758 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2759 * the safest choice is to do it like this, wrapping the whole function.
2761 current
->memcg_kmem_skip_account
= 1;
2762 __memcg_schedule_register_cache(memcg
, cachep
);
2763 current
->memcg_kmem_skip_account
= 0;
2767 * Return the kmem_cache we're supposed to use for a slab allocation.
2768 * We try to use the current memcg's version of the cache.
2770 * If the cache does not exist yet, if we are the first user of it,
2771 * we either create it immediately, if possible, or create it asynchronously
2773 * In the latter case, we will let the current allocation go through with
2774 * the original cache.
2776 * Can't be called in interrupt context or from kernel threads.
2777 * This function needs to be called with rcu_read_lock() held.
2779 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2781 struct mem_cgroup
*memcg
;
2782 struct kmem_cache
*memcg_cachep
;
2784 VM_BUG_ON(!cachep
->memcg_params
);
2785 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
2787 if (current
->memcg_kmem_skip_account
)
2790 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2791 if (!memcg_kmem_is_active(memcg
))
2794 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
2795 if (likely(memcg_cachep
))
2796 return memcg_cachep
;
2799 * If we are in a safe context (can wait, and not in interrupt
2800 * context), we could be be predictable and return right away.
2801 * This would guarantee that the allocation being performed
2802 * already belongs in the new cache.
2804 * However, there are some clashes that can arrive from locking.
2805 * For instance, because we acquire the slab_mutex while doing
2806 * memcg_create_kmem_cache, this means no further allocation
2807 * could happen with the slab_mutex held. So it's better to
2810 memcg_schedule_register_cache(memcg
, cachep
);
2812 css_put(&memcg
->css
);
2816 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2818 if (!is_root_cache(cachep
))
2819 css_put(&cachep
->memcg_params
->memcg
->css
);
2823 * We need to verify if the allocation against current->mm->owner's memcg is
2824 * possible for the given order. But the page is not allocated yet, so we'll
2825 * need a further commit step to do the final arrangements.
2827 * It is possible for the task to switch cgroups in this mean time, so at
2828 * commit time, we can't rely on task conversion any longer. We'll then use
2829 * the handle argument to return to the caller which cgroup we should commit
2830 * against. We could also return the memcg directly and avoid the pointer
2831 * passing, but a boolean return value gives better semantics considering
2832 * the compiled-out case as well.
2834 * Returning true means the allocation is possible.
2837 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
2839 struct mem_cgroup
*memcg
;
2844 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2846 if (!memcg_kmem_is_active(memcg
)) {
2847 css_put(&memcg
->css
);
2851 ret
= memcg_charge_kmem(memcg
, gfp
, 1 << order
);
2855 css_put(&memcg
->css
);
2859 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2862 VM_BUG_ON(mem_cgroup_is_root(memcg
));
2864 /* The page allocation failed. Revert */
2866 memcg_uncharge_kmem(memcg
, 1 << order
);
2869 page
->mem_cgroup
= memcg
;
2872 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
2874 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2879 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2881 memcg_uncharge_kmem(memcg
, 1 << order
);
2882 page
->mem_cgroup
= NULL
;
2884 #endif /* CONFIG_MEMCG_KMEM */
2886 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2889 * Because tail pages are not marked as "used", set it. We're under
2890 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2891 * charge/uncharge will be never happen and move_account() is done under
2892 * compound_lock(), so we don't have to take care of races.
2894 void mem_cgroup_split_huge_fixup(struct page
*head
)
2898 if (mem_cgroup_disabled())
2901 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2902 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2904 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2907 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2910 * mem_cgroup_move_account - move account of the page
2912 * @nr_pages: number of regular pages (>1 for huge pages)
2913 * @from: mem_cgroup which the page is moved from.
2914 * @to: mem_cgroup which the page is moved to. @from != @to.
2916 * The caller must confirm following.
2917 * - page is not on LRU (isolate_page() is useful.)
2918 * - compound_lock is held when nr_pages > 1
2920 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2923 static int mem_cgroup_move_account(struct page
*page
,
2924 unsigned int nr_pages
,
2925 struct mem_cgroup
*from
,
2926 struct mem_cgroup
*to
)
2928 unsigned long flags
;
2931 VM_BUG_ON(from
== to
);
2932 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2934 * The page is isolated from LRU. So, collapse function
2935 * will not handle this page. But page splitting can happen.
2936 * Do this check under compound_page_lock(). The caller should
2940 if (nr_pages
> 1 && !PageTransHuge(page
))
2944 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2945 * of its source page while we change it: page migration takes
2946 * both pages off the LRU, but page cache replacement doesn't.
2948 if (!trylock_page(page
))
2952 if (page
->mem_cgroup
!= from
)
2955 spin_lock_irqsave(&from
->move_lock
, flags
);
2957 if (!PageAnon(page
) && page_mapped(page
)) {
2958 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
2960 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
2964 if (PageWriteback(page
)) {
2965 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
2967 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
2972 * It is safe to change page->mem_cgroup here because the page
2973 * is referenced, charged, and isolated - we can't race with
2974 * uncharging, charging, migration, or LRU putback.
2977 /* caller should have done css_get */
2978 page
->mem_cgroup
= to
;
2979 spin_unlock_irqrestore(&from
->move_lock
, flags
);
2983 local_irq_disable();
2984 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
2985 memcg_check_events(to
, page
);
2986 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
2987 memcg_check_events(from
, page
);
2995 #ifdef CONFIG_MEMCG_SWAP
2996 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2999 int val
= (charge
) ? 1 : -1;
3000 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
3004 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3005 * @entry: swap entry to be moved
3006 * @from: mem_cgroup which the entry is moved from
3007 * @to: mem_cgroup which the entry is moved to
3009 * It succeeds only when the swap_cgroup's record for this entry is the same
3010 * as the mem_cgroup's id of @from.
3012 * Returns 0 on success, -EINVAL on failure.
3014 * The caller must have charged to @to, IOW, called page_counter_charge() about
3015 * both res and memsw, and called css_get().
3017 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3018 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3020 unsigned short old_id
, new_id
;
3022 old_id
= mem_cgroup_id(from
);
3023 new_id
= mem_cgroup_id(to
);
3025 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3026 mem_cgroup_swap_statistics(from
, false);
3027 mem_cgroup_swap_statistics(to
, true);
3033 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3034 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3040 static DEFINE_MUTEX(memcg_limit_mutex
);
3042 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3043 unsigned long limit
)
3045 unsigned long curusage
;
3046 unsigned long oldusage
;
3047 bool enlarge
= false;
3052 * For keeping hierarchical_reclaim simple, how long we should retry
3053 * is depends on callers. We set our retry-count to be function
3054 * of # of children which we should visit in this loop.
3056 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
3057 mem_cgroup_count_children(memcg
);
3059 oldusage
= page_counter_read(&memcg
->memory
);
3062 if (signal_pending(current
)) {
3067 mutex_lock(&memcg_limit_mutex
);
3068 if (limit
> memcg
->memsw
.limit
) {
3069 mutex_unlock(&memcg_limit_mutex
);
3073 if (limit
> memcg
->memory
.limit
)
3075 ret
= page_counter_limit(&memcg
->memory
, limit
);
3076 mutex_unlock(&memcg_limit_mutex
);
3081 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
3083 curusage
= page_counter_read(&memcg
->memory
);
3084 /* Usage is reduced ? */
3085 if (curusage
>= oldusage
)
3088 oldusage
= curusage
;
3089 } while (retry_count
);
3091 if (!ret
&& enlarge
)
3092 memcg_oom_recover(memcg
);
3097 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3098 unsigned long limit
)
3100 unsigned long curusage
;
3101 unsigned long oldusage
;
3102 bool enlarge
= false;
3106 /* see mem_cgroup_resize_res_limit */
3107 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
3108 mem_cgroup_count_children(memcg
);
3110 oldusage
= page_counter_read(&memcg
->memsw
);
3113 if (signal_pending(current
)) {
3118 mutex_lock(&memcg_limit_mutex
);
3119 if (limit
< memcg
->memory
.limit
) {
3120 mutex_unlock(&memcg_limit_mutex
);
3124 if (limit
> memcg
->memsw
.limit
)
3126 ret
= page_counter_limit(&memcg
->memsw
, limit
);
3127 mutex_unlock(&memcg_limit_mutex
);
3132 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
3134 curusage
= page_counter_read(&memcg
->memsw
);
3135 /* Usage is reduced ? */
3136 if (curusage
>= oldusage
)
3139 oldusage
= curusage
;
3140 } while (retry_count
);
3142 if (!ret
&& enlarge
)
3143 memcg_oom_recover(memcg
);
3148 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3150 unsigned long *total_scanned
)
3152 unsigned long nr_reclaimed
= 0;
3153 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3154 unsigned long reclaimed
;
3156 struct mem_cgroup_tree_per_zone
*mctz
;
3157 unsigned long excess
;
3158 unsigned long nr_scanned
;
3163 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3165 * This loop can run a while, specially if mem_cgroup's continuously
3166 * keep exceeding their soft limit and putting the system under
3173 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3178 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
3179 gfp_mask
, &nr_scanned
);
3180 nr_reclaimed
+= reclaimed
;
3181 *total_scanned
+= nr_scanned
;
3182 spin_lock_irq(&mctz
->lock
);
3183 __mem_cgroup_remove_exceeded(mz
, mctz
);
3186 * If we failed to reclaim anything from this memory cgroup
3187 * it is time to move on to the next cgroup
3191 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3193 excess
= soft_limit_excess(mz
->memcg
);
3195 * One school of thought says that we should not add
3196 * back the node to the tree if reclaim returns 0.
3197 * But our reclaim could return 0, simply because due
3198 * to priority we are exposing a smaller subset of
3199 * memory to reclaim from. Consider this as a longer
3202 /* If excess == 0, no tree ops */
3203 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3204 spin_unlock_irq(&mctz
->lock
);
3205 css_put(&mz
->memcg
->css
);
3208 * Could not reclaim anything and there are no more
3209 * mem cgroups to try or we seem to be looping without
3210 * reclaiming anything.
3212 if (!nr_reclaimed
&&
3214 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3216 } while (!nr_reclaimed
);
3218 css_put(&next_mz
->memcg
->css
);
3219 return nr_reclaimed
;
3223 * Test whether @memcg has children, dead or alive. Note that this
3224 * function doesn't care whether @memcg has use_hierarchy enabled and
3225 * returns %true if there are child csses according to the cgroup
3226 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3228 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3233 * The lock does not prevent addition or deletion of children, but
3234 * it prevents a new child from being initialized based on this
3235 * parent in css_online(), so it's enough to decide whether
3236 * hierarchically inherited attributes can still be changed or not.
3238 lockdep_assert_held(&memcg_create_mutex
);
3241 ret
= css_next_child(NULL
, &memcg
->css
);
3247 * Reclaims as many pages from the given memcg as possible and moves
3248 * the rest to the parent.
3250 * Caller is responsible for holding css reference for memcg.
3252 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3254 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3256 /* we call try-to-free pages for make this cgroup empty */
3257 lru_add_drain_all();
3258 /* try to free all pages in this cgroup */
3259 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3262 if (signal_pending(current
))
3265 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3269 /* maybe some writeback is necessary */
3270 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3278 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3279 char *buf
, size_t nbytes
,
3282 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3284 if (mem_cgroup_is_root(memcg
))
3286 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3289 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3292 return mem_cgroup_from_css(css
)->use_hierarchy
;
3295 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3296 struct cftype
*cft
, u64 val
)
3299 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3300 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3302 mutex_lock(&memcg_create_mutex
);
3304 if (memcg
->use_hierarchy
== val
)
3308 * If parent's use_hierarchy is set, we can't make any modifications
3309 * in the child subtrees. If it is unset, then the change can
3310 * occur, provided the current cgroup has no children.
3312 * For the root cgroup, parent_mem is NULL, we allow value to be
3313 * set if there are no children.
3315 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3316 (val
== 1 || val
== 0)) {
3317 if (!memcg_has_children(memcg
))
3318 memcg
->use_hierarchy
= val
;
3325 mutex_unlock(&memcg_create_mutex
);
3330 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
3331 enum mem_cgroup_stat_index idx
)
3333 struct mem_cgroup
*iter
;
3336 /* Per-cpu values can be negative, use a signed accumulator */
3337 for_each_mem_cgroup_tree(iter
, memcg
)
3338 val
+= mem_cgroup_read_stat(iter
, idx
);
3340 if (val
< 0) /* race ? */
3345 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3349 if (mem_cgroup_is_root(memcg
)) {
3350 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
3351 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
3353 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
3356 val
= page_counter_read(&memcg
->memory
);
3358 val
= page_counter_read(&memcg
->memsw
);
3360 return val
<< PAGE_SHIFT
;
3371 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3374 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3375 struct page_counter
*counter
;
3377 switch (MEMFILE_TYPE(cft
->private)) {
3379 counter
= &memcg
->memory
;
3382 counter
= &memcg
->memsw
;
3385 counter
= &memcg
->kmem
;
3391 switch (MEMFILE_ATTR(cft
->private)) {
3393 if (counter
== &memcg
->memory
)
3394 return mem_cgroup_usage(memcg
, false);
3395 if (counter
== &memcg
->memsw
)
3396 return mem_cgroup_usage(memcg
, true);
3397 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3399 return (u64
)counter
->limit
* PAGE_SIZE
;
3401 return (u64
)counter
->watermark
* PAGE_SIZE
;
3403 return counter
->failcnt
;
3404 case RES_SOFT_LIMIT
:
3405 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3411 #ifdef CONFIG_MEMCG_KMEM
3412 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
3413 unsigned long nr_pages
)
3418 if (memcg_kmem_is_active(memcg
))
3422 * For simplicity, we won't allow this to be disabled. It also can't
3423 * be changed if the cgroup has children already, or if tasks had
3426 * If tasks join before we set the limit, a person looking at
3427 * kmem.usage_in_bytes will have no way to determine when it took
3428 * place, which makes the value quite meaningless.
3430 * After it first became limited, changes in the value of the limit are
3431 * of course permitted.
3433 mutex_lock(&memcg_create_mutex
);
3434 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
3435 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
3437 mutex_unlock(&memcg_create_mutex
);
3441 memcg_id
= memcg_alloc_cache_id();
3448 * We couldn't have accounted to this cgroup, because it hasn't got
3449 * activated yet, so this should succeed.
3451 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
3454 static_key_slow_inc(&memcg_kmem_enabled_key
);
3456 * A memory cgroup is considered kmem-active as soon as it gets
3457 * kmemcg_id. Setting the id after enabling static branching will
3458 * guarantee no one starts accounting before all call sites are
3461 memcg
->kmemcg_id
= memcg_id
;
3466 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3467 unsigned long limit
)
3471 mutex_lock(&memcg_limit_mutex
);
3472 if (!memcg_kmem_is_active(memcg
))
3473 ret
= memcg_activate_kmem(memcg
, limit
);
3475 ret
= page_counter_limit(&memcg
->kmem
, limit
);
3476 mutex_unlock(&memcg_limit_mutex
);
3480 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
3483 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
3488 mutex_lock(&memcg_limit_mutex
);
3490 * If the parent cgroup is not kmem-active now, it cannot be activated
3491 * after this point, because it has at least one child already.
3493 if (memcg_kmem_is_active(parent
))
3494 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
3495 mutex_unlock(&memcg_limit_mutex
);
3499 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3500 unsigned long limit
)
3504 #endif /* CONFIG_MEMCG_KMEM */
3507 * The user of this function is...
3510 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3511 char *buf
, size_t nbytes
, loff_t off
)
3513 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3514 unsigned long nr_pages
;
3517 buf
= strstrip(buf
);
3518 ret
= page_counter_memparse(buf
, &nr_pages
);
3522 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3524 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3528 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3530 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3533 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3536 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3540 case RES_SOFT_LIMIT
:
3541 memcg
->soft_limit
= nr_pages
;
3545 return ret
?: nbytes
;
3548 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3549 size_t nbytes
, loff_t off
)
3551 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3552 struct page_counter
*counter
;
3554 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3556 counter
= &memcg
->memory
;
3559 counter
= &memcg
->memsw
;
3562 counter
= &memcg
->kmem
;
3568 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3570 page_counter_reset_watermark(counter
);
3573 counter
->failcnt
= 0;
3582 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3585 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3589 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3590 struct cftype
*cft
, u64 val
)
3592 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3594 if (val
>= (1 << NR_MOVE_TYPE
))
3598 * No kind of locking is needed in here, because ->can_attach() will
3599 * check this value once in the beginning of the process, and then carry
3600 * on with stale data. This means that changes to this value will only
3601 * affect task migrations starting after the change.
3603 memcg
->move_charge_at_immigrate
= val
;
3607 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3608 struct cftype
*cft
, u64 val
)
3615 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3619 unsigned int lru_mask
;
3622 static const struct numa_stat stats
[] = {
3623 { "total", LRU_ALL
},
3624 { "file", LRU_ALL_FILE
},
3625 { "anon", LRU_ALL_ANON
},
3626 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3628 const struct numa_stat
*stat
;
3631 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3633 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3634 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3635 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3636 for_each_node_state(nid
, N_MEMORY
) {
3637 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3639 seq_printf(m
, " N%d=%lu", nid
, nr
);
3644 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3645 struct mem_cgroup
*iter
;
3648 for_each_mem_cgroup_tree(iter
, memcg
)
3649 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3650 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3651 for_each_node_state(nid
, N_MEMORY
) {
3653 for_each_mem_cgroup_tree(iter
, memcg
)
3654 nr
+= mem_cgroup_node_nr_lru_pages(
3655 iter
, nid
, stat
->lru_mask
);
3656 seq_printf(m
, " N%d=%lu", nid
, nr
);
3663 #endif /* CONFIG_NUMA */
3665 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3667 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3668 unsigned long memory
, memsw
;
3669 struct mem_cgroup
*mi
;
3672 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3674 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3675 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3677 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
3678 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3681 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3682 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3683 mem_cgroup_read_events(memcg
, i
));
3685 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3686 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3687 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3689 /* Hierarchical information */
3690 memory
= memsw
= PAGE_COUNTER_MAX
;
3691 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3692 memory
= min(memory
, mi
->memory
.limit
);
3693 memsw
= min(memsw
, mi
->memsw
.limit
);
3695 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3696 (u64
)memory
* PAGE_SIZE
);
3697 if (do_swap_account
)
3698 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3699 (u64
)memsw
* PAGE_SIZE
);
3701 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3704 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3706 for_each_mem_cgroup_tree(mi
, memcg
)
3707 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3708 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
3711 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3712 unsigned long long val
= 0;
3714 for_each_mem_cgroup_tree(mi
, memcg
)
3715 val
+= mem_cgroup_read_events(mi
, i
);
3716 seq_printf(m
, "total_%s %llu\n",
3717 mem_cgroup_events_names
[i
], val
);
3720 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3721 unsigned long long val
= 0;
3723 for_each_mem_cgroup_tree(mi
, memcg
)
3724 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3725 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3728 #ifdef CONFIG_DEBUG_VM
3731 struct mem_cgroup_per_zone
*mz
;
3732 struct zone_reclaim_stat
*rstat
;
3733 unsigned long recent_rotated
[2] = {0, 0};
3734 unsigned long recent_scanned
[2] = {0, 0};
3736 for_each_online_node(nid
)
3737 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3738 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3739 rstat
= &mz
->lruvec
.reclaim_stat
;
3741 recent_rotated
[0] += rstat
->recent_rotated
[0];
3742 recent_rotated
[1] += rstat
->recent_rotated
[1];
3743 recent_scanned
[0] += rstat
->recent_scanned
[0];
3744 recent_scanned
[1] += rstat
->recent_scanned
[1];
3746 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3747 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3748 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3749 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3756 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3759 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3761 return mem_cgroup_swappiness(memcg
);
3764 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3765 struct cftype
*cft
, u64 val
)
3767 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3773 memcg
->swappiness
= val
;
3775 vm_swappiness
= val
;
3780 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3782 struct mem_cgroup_threshold_ary
*t
;
3783 unsigned long usage
;
3788 t
= rcu_dereference(memcg
->thresholds
.primary
);
3790 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3795 usage
= mem_cgroup_usage(memcg
, swap
);
3798 * current_threshold points to threshold just below or equal to usage.
3799 * If it's not true, a threshold was crossed after last
3800 * call of __mem_cgroup_threshold().
3802 i
= t
->current_threshold
;
3805 * Iterate backward over array of thresholds starting from
3806 * current_threshold and check if a threshold is crossed.
3807 * If none of thresholds below usage is crossed, we read
3808 * only one element of the array here.
3810 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3811 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3813 /* i = current_threshold + 1 */
3817 * Iterate forward over array of thresholds starting from
3818 * current_threshold+1 and check if a threshold is crossed.
3819 * If none of thresholds above usage is crossed, we read
3820 * only one element of the array here.
3822 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3823 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3825 /* Update current_threshold */
3826 t
->current_threshold
= i
- 1;
3831 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3834 __mem_cgroup_threshold(memcg
, false);
3835 if (do_swap_account
)
3836 __mem_cgroup_threshold(memcg
, true);
3838 memcg
= parent_mem_cgroup(memcg
);
3842 static int compare_thresholds(const void *a
, const void *b
)
3844 const struct mem_cgroup_threshold
*_a
= a
;
3845 const struct mem_cgroup_threshold
*_b
= b
;
3847 if (_a
->threshold
> _b
->threshold
)
3850 if (_a
->threshold
< _b
->threshold
)
3856 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3858 struct mem_cgroup_eventfd_list
*ev
;
3860 spin_lock(&memcg_oom_lock
);
3862 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3863 eventfd_signal(ev
->eventfd
, 1);
3865 spin_unlock(&memcg_oom_lock
);
3869 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3871 struct mem_cgroup
*iter
;
3873 for_each_mem_cgroup_tree(iter
, memcg
)
3874 mem_cgroup_oom_notify_cb(iter
);
3877 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3878 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3880 struct mem_cgroup_thresholds
*thresholds
;
3881 struct mem_cgroup_threshold_ary
*new;
3882 unsigned long threshold
;
3883 unsigned long usage
;
3886 ret
= page_counter_memparse(args
, &threshold
);
3890 mutex_lock(&memcg
->thresholds_lock
);
3893 thresholds
= &memcg
->thresholds
;
3894 usage
= mem_cgroup_usage(memcg
, false);
3895 } else if (type
== _MEMSWAP
) {
3896 thresholds
= &memcg
->memsw_thresholds
;
3897 usage
= mem_cgroup_usage(memcg
, true);
3901 /* Check if a threshold crossed before adding a new one */
3902 if (thresholds
->primary
)
3903 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3905 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3907 /* Allocate memory for new array of thresholds */
3908 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3916 /* Copy thresholds (if any) to new array */
3917 if (thresholds
->primary
) {
3918 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3919 sizeof(struct mem_cgroup_threshold
));
3922 /* Add new threshold */
3923 new->entries
[size
- 1].eventfd
= eventfd
;
3924 new->entries
[size
- 1].threshold
= threshold
;
3926 /* Sort thresholds. Registering of new threshold isn't time-critical */
3927 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3928 compare_thresholds
, NULL
);
3930 /* Find current threshold */
3931 new->current_threshold
= -1;
3932 for (i
= 0; i
< size
; i
++) {
3933 if (new->entries
[i
].threshold
<= usage
) {
3935 * new->current_threshold will not be used until
3936 * rcu_assign_pointer(), so it's safe to increment
3939 ++new->current_threshold
;
3944 /* Free old spare buffer and save old primary buffer as spare */
3945 kfree(thresholds
->spare
);
3946 thresholds
->spare
= thresholds
->primary
;
3948 rcu_assign_pointer(thresholds
->primary
, new);
3950 /* To be sure that nobody uses thresholds */
3954 mutex_unlock(&memcg
->thresholds_lock
);
3959 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3960 struct eventfd_ctx
*eventfd
, const char *args
)
3962 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3965 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3966 struct eventfd_ctx
*eventfd
, const char *args
)
3968 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3971 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3972 struct eventfd_ctx
*eventfd
, enum res_type type
)
3974 struct mem_cgroup_thresholds
*thresholds
;
3975 struct mem_cgroup_threshold_ary
*new;
3976 unsigned long usage
;
3979 mutex_lock(&memcg
->thresholds_lock
);
3982 thresholds
= &memcg
->thresholds
;
3983 usage
= mem_cgroup_usage(memcg
, false);
3984 } else if (type
== _MEMSWAP
) {
3985 thresholds
= &memcg
->memsw_thresholds
;
3986 usage
= mem_cgroup_usage(memcg
, true);
3990 if (!thresholds
->primary
)
3993 /* Check if a threshold crossed before removing */
3994 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3996 /* Calculate new number of threshold */
3998 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3999 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4003 new = thresholds
->spare
;
4005 /* Set thresholds array to NULL if we don't have thresholds */
4014 /* Copy thresholds and find current threshold */
4015 new->current_threshold
= -1;
4016 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4017 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4020 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4021 if (new->entries
[j
].threshold
<= usage
) {
4023 * new->current_threshold will not be used
4024 * until rcu_assign_pointer(), so it's safe to increment
4027 ++new->current_threshold
;
4033 /* Swap primary and spare array */
4034 thresholds
->spare
= thresholds
->primary
;
4035 /* If all events are unregistered, free the spare array */
4037 kfree(thresholds
->spare
);
4038 thresholds
->spare
= NULL
;
4041 rcu_assign_pointer(thresholds
->primary
, new);
4043 /* To be sure that nobody uses thresholds */
4046 mutex_unlock(&memcg
->thresholds_lock
);
4049 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4050 struct eventfd_ctx
*eventfd
)
4052 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4055 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4056 struct eventfd_ctx
*eventfd
)
4058 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4061 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4062 struct eventfd_ctx
*eventfd
, const char *args
)
4064 struct mem_cgroup_eventfd_list
*event
;
4066 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4070 spin_lock(&memcg_oom_lock
);
4072 event
->eventfd
= eventfd
;
4073 list_add(&event
->list
, &memcg
->oom_notify
);
4075 /* already in OOM ? */
4076 if (atomic_read(&memcg
->under_oom
))
4077 eventfd_signal(eventfd
, 1);
4078 spin_unlock(&memcg_oom_lock
);
4083 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4084 struct eventfd_ctx
*eventfd
)
4086 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4088 spin_lock(&memcg_oom_lock
);
4090 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4091 if (ev
->eventfd
== eventfd
) {
4092 list_del(&ev
->list
);
4097 spin_unlock(&memcg_oom_lock
);
4100 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4102 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
4104 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4105 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
4109 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4110 struct cftype
*cft
, u64 val
)
4112 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4114 /* cannot set to root cgroup and only 0 and 1 are allowed */
4115 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4118 memcg
->oom_kill_disable
= val
;
4120 memcg_oom_recover(memcg
);
4125 #ifdef CONFIG_MEMCG_KMEM
4126 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4130 ret
= memcg_propagate_kmem(memcg
);
4134 return mem_cgroup_sockets_init(memcg
, ss
);
4137 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4139 memcg_unregister_all_caches(memcg
);
4140 mem_cgroup_sockets_destroy(memcg
);
4143 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4148 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4154 * DO NOT USE IN NEW FILES.
4156 * "cgroup.event_control" implementation.
4158 * This is way over-engineered. It tries to support fully configurable
4159 * events for each user. Such level of flexibility is completely
4160 * unnecessary especially in the light of the planned unified hierarchy.
4162 * Please deprecate this and replace with something simpler if at all
4167 * Unregister event and free resources.
4169 * Gets called from workqueue.
4171 static void memcg_event_remove(struct work_struct
*work
)
4173 struct mem_cgroup_event
*event
=
4174 container_of(work
, struct mem_cgroup_event
, remove
);
4175 struct mem_cgroup
*memcg
= event
->memcg
;
4177 remove_wait_queue(event
->wqh
, &event
->wait
);
4179 event
->unregister_event(memcg
, event
->eventfd
);
4181 /* Notify userspace the event is going away. */
4182 eventfd_signal(event
->eventfd
, 1);
4184 eventfd_ctx_put(event
->eventfd
);
4186 css_put(&memcg
->css
);
4190 * Gets called on POLLHUP on eventfd when user closes it.
4192 * Called with wqh->lock held and interrupts disabled.
4194 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
4195 int sync
, void *key
)
4197 struct mem_cgroup_event
*event
=
4198 container_of(wait
, struct mem_cgroup_event
, wait
);
4199 struct mem_cgroup
*memcg
= event
->memcg
;
4200 unsigned long flags
= (unsigned long)key
;
4202 if (flags
& POLLHUP
) {
4204 * If the event has been detached at cgroup removal, we
4205 * can simply return knowing the other side will cleanup
4208 * We can't race against event freeing since the other
4209 * side will require wqh->lock via remove_wait_queue(),
4212 spin_lock(&memcg
->event_list_lock
);
4213 if (!list_empty(&event
->list
)) {
4214 list_del_init(&event
->list
);
4216 * We are in atomic context, but cgroup_event_remove()
4217 * may sleep, so we have to call it in workqueue.
4219 schedule_work(&event
->remove
);
4221 spin_unlock(&memcg
->event_list_lock
);
4227 static void memcg_event_ptable_queue_proc(struct file
*file
,
4228 wait_queue_head_t
*wqh
, poll_table
*pt
)
4230 struct mem_cgroup_event
*event
=
4231 container_of(pt
, struct mem_cgroup_event
, pt
);
4234 add_wait_queue(wqh
, &event
->wait
);
4238 * DO NOT USE IN NEW FILES.
4240 * Parse input and register new cgroup event handler.
4242 * Input must be in format '<event_fd> <control_fd> <args>'.
4243 * Interpretation of args is defined by control file implementation.
4245 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4246 char *buf
, size_t nbytes
, loff_t off
)
4248 struct cgroup_subsys_state
*css
= of_css(of
);
4249 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4250 struct mem_cgroup_event
*event
;
4251 struct cgroup_subsys_state
*cfile_css
;
4252 unsigned int efd
, cfd
;
4259 buf
= strstrip(buf
);
4261 efd
= simple_strtoul(buf
, &endp
, 10);
4266 cfd
= simple_strtoul(buf
, &endp
, 10);
4267 if ((*endp
!= ' ') && (*endp
!= '\0'))
4271 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4275 event
->memcg
= memcg
;
4276 INIT_LIST_HEAD(&event
->list
);
4277 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4278 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4279 INIT_WORK(&event
->remove
, memcg_event_remove
);
4287 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4288 if (IS_ERR(event
->eventfd
)) {
4289 ret
= PTR_ERR(event
->eventfd
);
4296 goto out_put_eventfd
;
4299 /* the process need read permission on control file */
4300 /* AV: shouldn't we check that it's been opened for read instead? */
4301 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4306 * Determine the event callbacks and set them in @event. This used
4307 * to be done via struct cftype but cgroup core no longer knows
4308 * about these events. The following is crude but the whole thing
4309 * is for compatibility anyway.
4311 * DO NOT ADD NEW FILES.
4313 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4315 if (!strcmp(name
, "memory.usage_in_bytes")) {
4316 event
->register_event
= mem_cgroup_usage_register_event
;
4317 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4318 } else if (!strcmp(name
, "memory.oom_control")) {
4319 event
->register_event
= mem_cgroup_oom_register_event
;
4320 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4321 } else if (!strcmp(name
, "memory.pressure_level")) {
4322 event
->register_event
= vmpressure_register_event
;
4323 event
->unregister_event
= vmpressure_unregister_event
;
4324 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4325 event
->register_event
= memsw_cgroup_usage_register_event
;
4326 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4333 * Verify @cfile should belong to @css. Also, remaining events are
4334 * automatically removed on cgroup destruction but the removal is
4335 * asynchronous, so take an extra ref on @css.
4337 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4338 &memory_cgrp_subsys
);
4340 if (IS_ERR(cfile_css
))
4342 if (cfile_css
!= css
) {
4347 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4351 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
4353 spin_lock(&memcg
->event_list_lock
);
4354 list_add(&event
->list
, &memcg
->event_list
);
4355 spin_unlock(&memcg
->event_list_lock
);
4367 eventfd_ctx_put(event
->eventfd
);
4376 static struct cftype mem_cgroup_files
[] = {
4378 .name
= "usage_in_bytes",
4379 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4380 .read_u64
= mem_cgroup_read_u64
,
4383 .name
= "max_usage_in_bytes",
4384 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4385 .write
= mem_cgroup_reset
,
4386 .read_u64
= mem_cgroup_read_u64
,
4389 .name
= "limit_in_bytes",
4390 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4391 .write
= mem_cgroup_write
,
4392 .read_u64
= mem_cgroup_read_u64
,
4395 .name
= "soft_limit_in_bytes",
4396 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4397 .write
= mem_cgroup_write
,
4398 .read_u64
= mem_cgroup_read_u64
,
4402 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4403 .write
= mem_cgroup_reset
,
4404 .read_u64
= mem_cgroup_read_u64
,
4408 .seq_show
= memcg_stat_show
,
4411 .name
= "force_empty",
4412 .write
= mem_cgroup_force_empty_write
,
4415 .name
= "use_hierarchy",
4416 .write_u64
= mem_cgroup_hierarchy_write
,
4417 .read_u64
= mem_cgroup_hierarchy_read
,
4420 .name
= "cgroup.event_control", /* XXX: for compat */
4421 .write
= memcg_write_event_control
,
4422 .flags
= CFTYPE_NO_PREFIX
,
4426 .name
= "swappiness",
4427 .read_u64
= mem_cgroup_swappiness_read
,
4428 .write_u64
= mem_cgroup_swappiness_write
,
4431 .name
= "move_charge_at_immigrate",
4432 .read_u64
= mem_cgroup_move_charge_read
,
4433 .write_u64
= mem_cgroup_move_charge_write
,
4436 .name
= "oom_control",
4437 .seq_show
= mem_cgroup_oom_control_read
,
4438 .write_u64
= mem_cgroup_oom_control_write
,
4439 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4442 .name
= "pressure_level",
4446 .name
= "numa_stat",
4447 .seq_show
= memcg_numa_stat_show
,
4450 #ifdef CONFIG_MEMCG_KMEM
4452 .name
= "kmem.limit_in_bytes",
4453 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4454 .write
= mem_cgroup_write
,
4455 .read_u64
= mem_cgroup_read_u64
,
4458 .name
= "kmem.usage_in_bytes",
4459 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4460 .read_u64
= mem_cgroup_read_u64
,
4463 .name
= "kmem.failcnt",
4464 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4465 .write
= mem_cgroup_reset
,
4466 .read_u64
= mem_cgroup_read_u64
,
4469 .name
= "kmem.max_usage_in_bytes",
4470 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4471 .write
= mem_cgroup_reset
,
4472 .read_u64
= mem_cgroup_read_u64
,
4474 #ifdef CONFIG_SLABINFO
4476 .name
= "kmem.slabinfo",
4477 .seq_start
= slab_start
,
4478 .seq_next
= slab_next
,
4479 .seq_stop
= slab_stop
,
4480 .seq_show
= memcg_slab_show
,
4484 { }, /* terminate */
4487 #ifdef CONFIG_MEMCG_SWAP
4488 static struct cftype memsw_cgroup_files
[] = {
4490 .name
= "memsw.usage_in_bytes",
4491 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
4492 .read_u64
= mem_cgroup_read_u64
,
4495 .name
= "memsw.max_usage_in_bytes",
4496 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
4497 .write
= mem_cgroup_reset
,
4498 .read_u64
= mem_cgroup_read_u64
,
4501 .name
= "memsw.limit_in_bytes",
4502 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
4503 .write
= mem_cgroup_write
,
4504 .read_u64
= mem_cgroup_read_u64
,
4507 .name
= "memsw.failcnt",
4508 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
4509 .write
= mem_cgroup_reset
,
4510 .read_u64
= mem_cgroup_read_u64
,
4512 { }, /* terminate */
4515 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4517 struct mem_cgroup_per_node
*pn
;
4518 struct mem_cgroup_per_zone
*mz
;
4519 int zone
, tmp
= node
;
4521 * This routine is called against possible nodes.
4522 * But it's BUG to call kmalloc() against offline node.
4524 * TODO: this routine can waste much memory for nodes which will
4525 * never be onlined. It's better to use memory hotplug callback
4528 if (!node_state(node
, N_NORMAL_MEMORY
))
4530 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4534 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4535 mz
= &pn
->zoneinfo
[zone
];
4536 lruvec_init(&mz
->lruvec
);
4537 mz
->usage_in_excess
= 0;
4538 mz
->on_tree
= false;
4541 memcg
->nodeinfo
[node
] = pn
;
4545 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4547 kfree(memcg
->nodeinfo
[node
]);
4550 static struct mem_cgroup
*mem_cgroup_alloc(void)
4552 struct mem_cgroup
*memcg
;
4555 size
= sizeof(struct mem_cgroup
);
4556 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4558 memcg
= kzalloc(size
, GFP_KERNEL
);
4562 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4565 spin_lock_init(&memcg
->pcp_counter_lock
);
4574 * At destroying mem_cgroup, references from swap_cgroup can remain.
4575 * (scanning all at force_empty is too costly...)
4577 * Instead of clearing all references at force_empty, we remember
4578 * the number of reference from swap_cgroup and free mem_cgroup when
4579 * it goes down to 0.
4581 * Removal of cgroup itself succeeds regardless of refs from swap.
4584 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4588 mem_cgroup_remove_from_trees(memcg
);
4591 free_mem_cgroup_per_zone_info(memcg
, node
);
4593 free_percpu(memcg
->stat
);
4595 disarm_static_keys(memcg
);
4600 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4602 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4604 if (!memcg
->memory
.parent
)
4606 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4608 EXPORT_SYMBOL(parent_mem_cgroup
);
4610 static void __init
mem_cgroup_soft_limit_tree_init(void)
4612 struct mem_cgroup_tree_per_node
*rtpn
;
4613 struct mem_cgroup_tree_per_zone
*rtpz
;
4614 int tmp
, node
, zone
;
4616 for_each_node(node
) {
4618 if (!node_state(node
, N_NORMAL_MEMORY
))
4620 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
4623 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
4625 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4626 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
4627 rtpz
->rb_root
= RB_ROOT
;
4628 spin_lock_init(&rtpz
->lock
);
4633 static struct cgroup_subsys_state
* __ref
4634 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4636 struct mem_cgroup
*memcg
;
4637 long error
= -ENOMEM
;
4640 memcg
= mem_cgroup_alloc();
4642 return ERR_PTR(error
);
4645 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4649 if (parent_css
== NULL
) {
4650 root_mem_cgroup
= memcg
;
4651 page_counter_init(&memcg
->memory
, NULL
);
4652 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4653 page_counter_init(&memcg
->memsw
, NULL
);
4654 page_counter_init(&memcg
->kmem
, NULL
);
4657 memcg
->last_scanned_node
= MAX_NUMNODES
;
4658 INIT_LIST_HEAD(&memcg
->oom_notify
);
4659 memcg
->move_charge_at_immigrate
= 0;
4660 mutex_init(&memcg
->thresholds_lock
);
4661 spin_lock_init(&memcg
->move_lock
);
4662 vmpressure_init(&memcg
->vmpressure
);
4663 INIT_LIST_HEAD(&memcg
->event_list
);
4664 spin_lock_init(&memcg
->event_list_lock
);
4665 #ifdef CONFIG_MEMCG_KMEM
4666 memcg
->kmemcg_id
= -1;
4667 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
4673 __mem_cgroup_free(memcg
);
4674 return ERR_PTR(error
);
4678 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4680 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4681 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4684 if (css
->id
> MEM_CGROUP_ID_MAX
)
4690 mutex_lock(&memcg_create_mutex
);
4692 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4693 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4694 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4696 if (parent
->use_hierarchy
) {
4697 page_counter_init(&memcg
->memory
, &parent
->memory
);
4698 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4699 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4700 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4703 * No need to take a reference to the parent because cgroup
4704 * core guarantees its existence.
4707 page_counter_init(&memcg
->memory
, NULL
);
4708 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4709 page_counter_init(&memcg
->memsw
, NULL
);
4710 page_counter_init(&memcg
->kmem
, NULL
);
4712 * Deeper hierachy with use_hierarchy == false doesn't make
4713 * much sense so let cgroup subsystem know about this
4714 * unfortunate state in our controller.
4716 if (parent
!= root_mem_cgroup
)
4717 memory_cgrp_subsys
.broken_hierarchy
= true;
4719 mutex_unlock(&memcg_create_mutex
);
4721 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4726 * Make sure the memcg is initialized: mem_cgroup_iter()
4727 * orders reading memcg->initialized against its callers
4728 * reading the memcg members.
4730 smp_store_release(&memcg
->initialized
, 1);
4735 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4737 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4738 struct mem_cgroup_event
*event
, *tmp
;
4741 * Unregister events and notify userspace.
4742 * Notify userspace about cgroup removing only after rmdir of cgroup
4743 * directory to avoid race between userspace and kernelspace.
4745 spin_lock(&memcg
->event_list_lock
);
4746 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4747 list_del_init(&event
->list
);
4748 schedule_work(&event
->remove
);
4750 spin_unlock(&memcg
->event_list_lock
);
4752 vmpressure_cleanup(&memcg
->vmpressure
);
4755 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4757 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4759 memcg_destroy_kmem(memcg
);
4760 __mem_cgroup_free(memcg
);
4764 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4765 * @css: the target css
4767 * Reset the states of the mem_cgroup associated with @css. This is
4768 * invoked when the userland requests disabling on the default hierarchy
4769 * but the memcg is pinned through dependency. The memcg should stop
4770 * applying policies and should revert to the vanilla state as it may be
4771 * made visible again.
4773 * The current implementation only resets the essential configurations.
4774 * This needs to be expanded to cover all the visible parts.
4776 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4778 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4780 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4781 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4782 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4783 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4787 /* Handlers for move charge at task migration. */
4788 static int mem_cgroup_do_precharge(unsigned long count
)
4792 /* Try a single bulk charge without reclaim first */
4793 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
4795 mc
.precharge
+= count
;
4798 if (ret
== -EINTR
) {
4799 cancel_charge(root_mem_cgroup
, count
);
4803 /* Try charges one by one with reclaim */
4805 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4807 * In case of failure, any residual charges against
4808 * mc.to will be dropped by mem_cgroup_clear_mc()
4809 * later on. However, cancel any charges that are
4810 * bypassed to root right away or they'll be lost.
4813 cancel_charge(root_mem_cgroup
, 1);
4823 * get_mctgt_type - get target type of moving charge
4824 * @vma: the vma the pte to be checked belongs
4825 * @addr: the address corresponding to the pte to be checked
4826 * @ptent: the pte to be checked
4827 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4830 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4831 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4832 * move charge. if @target is not NULL, the page is stored in target->page
4833 * with extra refcnt got(Callers should handle it).
4834 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4835 * target for charge migration. if @target is not NULL, the entry is stored
4838 * Called with pte lock held.
4845 enum mc_target_type
{
4851 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4852 unsigned long addr
, pte_t ptent
)
4854 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4856 if (!page
|| !page_mapped(page
))
4858 if (PageAnon(page
)) {
4859 /* we don't move shared anon */
4862 } else if (!move_file())
4863 /* we ignore mapcount for file pages */
4865 if (!get_page_unless_zero(page
))
4872 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4873 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4875 struct page
*page
= NULL
;
4876 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4878 if (!move_anon() || non_swap_entry(ent
))
4881 * Because lookup_swap_cache() updates some statistics counter,
4882 * we call find_get_page() with swapper_space directly.
4884 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4885 if (do_swap_account
)
4886 entry
->val
= ent
.val
;
4891 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4892 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4898 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4899 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4901 struct page
*page
= NULL
;
4902 struct address_space
*mapping
;
4905 if (!vma
->vm_file
) /* anonymous vma */
4910 mapping
= vma
->vm_file
->f_mapping
;
4911 pgoff
= linear_page_index(vma
, addr
);
4913 /* page is moved even if it's not RSS of this task(page-faulted). */
4915 /* shmem/tmpfs may report page out on swap: account for that too. */
4916 if (shmem_mapping(mapping
)) {
4917 page
= find_get_entry(mapping
, pgoff
);
4918 if (radix_tree_exceptional_entry(page
)) {
4919 swp_entry_t swp
= radix_to_swp_entry(page
);
4920 if (do_swap_account
)
4922 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4925 page
= find_get_page(mapping
, pgoff
);
4927 page
= find_get_page(mapping
, pgoff
);
4932 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4933 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4935 struct page
*page
= NULL
;
4936 enum mc_target_type ret
= MC_TARGET_NONE
;
4937 swp_entry_t ent
= { .val
= 0 };
4939 if (pte_present(ptent
))
4940 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4941 else if (is_swap_pte(ptent
))
4942 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4943 else if (pte_none(ptent
))
4944 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4946 if (!page
&& !ent
.val
)
4950 * Do only loose check w/o serialization.
4951 * mem_cgroup_move_account() checks the page is valid or
4952 * not under LRU exclusion.
4954 if (page
->mem_cgroup
== mc
.from
) {
4955 ret
= MC_TARGET_PAGE
;
4957 target
->page
= page
;
4959 if (!ret
|| !target
)
4962 /* There is a swap entry and a page doesn't exist or isn't charged */
4963 if (ent
.val
&& !ret
&&
4964 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4965 ret
= MC_TARGET_SWAP
;
4972 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4974 * We don't consider swapping or file mapped pages because THP does not
4975 * support them for now.
4976 * Caller should make sure that pmd_trans_huge(pmd) is true.
4978 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4979 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4981 struct page
*page
= NULL
;
4982 enum mc_target_type ret
= MC_TARGET_NONE
;
4984 page
= pmd_page(pmd
);
4985 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4988 if (page
->mem_cgroup
== mc
.from
) {
4989 ret
= MC_TARGET_PAGE
;
4992 target
->page
= page
;
4998 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4999 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5001 return MC_TARGET_NONE
;
5005 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5006 unsigned long addr
, unsigned long end
,
5007 struct mm_walk
*walk
)
5009 struct vm_area_struct
*vma
= walk
->private;
5013 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5014 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5015 mc
.precharge
+= HPAGE_PMD_NR
;
5020 if (pmd_trans_unstable(pmd
))
5022 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5023 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5024 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5025 mc
.precharge
++; /* increment precharge temporarily */
5026 pte_unmap_unlock(pte
- 1, ptl
);
5032 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5034 unsigned long precharge
;
5035 struct vm_area_struct
*vma
;
5037 down_read(&mm
->mmap_sem
);
5038 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5039 struct mm_walk mem_cgroup_count_precharge_walk
= {
5040 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5044 if (is_vm_hugetlb_page(vma
))
5046 walk_page_range(vma
->vm_start
, vma
->vm_end
,
5047 &mem_cgroup_count_precharge_walk
);
5049 up_read(&mm
->mmap_sem
);
5051 precharge
= mc
.precharge
;
5057 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5059 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5061 VM_BUG_ON(mc
.moving_task
);
5062 mc
.moving_task
= current
;
5063 return mem_cgroup_do_precharge(precharge
);
5066 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5067 static void __mem_cgroup_clear_mc(void)
5069 struct mem_cgroup
*from
= mc
.from
;
5070 struct mem_cgroup
*to
= mc
.to
;
5072 /* we must uncharge all the leftover precharges from mc.to */
5074 cancel_charge(mc
.to
, mc
.precharge
);
5078 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5079 * we must uncharge here.
5081 if (mc
.moved_charge
) {
5082 cancel_charge(mc
.from
, mc
.moved_charge
);
5083 mc
.moved_charge
= 0;
5085 /* we must fixup refcnts and charges */
5086 if (mc
.moved_swap
) {
5087 /* uncharge swap account from the old cgroup */
5088 if (!mem_cgroup_is_root(mc
.from
))
5089 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5092 * we charged both to->memory and to->memsw, so we
5093 * should uncharge to->memory.
5095 if (!mem_cgroup_is_root(mc
.to
))
5096 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5098 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
5100 /* we've already done css_get(mc.to) */
5103 memcg_oom_recover(from
);
5104 memcg_oom_recover(to
);
5105 wake_up_all(&mc
.waitq
);
5108 static void mem_cgroup_clear_mc(void)
5111 * we must clear moving_task before waking up waiters at the end of
5114 mc
.moving_task
= NULL
;
5115 __mem_cgroup_clear_mc();
5116 spin_lock(&mc
.lock
);
5119 spin_unlock(&mc
.lock
);
5122 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5123 struct cgroup_taskset
*tset
)
5125 struct task_struct
*p
= cgroup_taskset_first(tset
);
5127 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5128 unsigned long move_charge_at_immigrate
;
5131 * We are now commited to this value whatever it is. Changes in this
5132 * tunable will only affect upcoming migrations, not the current one.
5133 * So we need to save it, and keep it going.
5135 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
5136 if (move_charge_at_immigrate
) {
5137 struct mm_struct
*mm
;
5138 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
5140 VM_BUG_ON(from
== memcg
);
5142 mm
= get_task_mm(p
);
5145 /* We move charges only when we move a owner of the mm */
5146 if (mm
->owner
== p
) {
5149 VM_BUG_ON(mc
.precharge
);
5150 VM_BUG_ON(mc
.moved_charge
);
5151 VM_BUG_ON(mc
.moved_swap
);
5153 spin_lock(&mc
.lock
);
5156 mc
.immigrate_flags
= move_charge_at_immigrate
;
5157 spin_unlock(&mc
.lock
);
5158 /* We set mc.moving_task later */
5160 ret
= mem_cgroup_precharge_mc(mm
);
5162 mem_cgroup_clear_mc();
5169 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5170 struct cgroup_taskset
*tset
)
5173 mem_cgroup_clear_mc();
5176 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5177 unsigned long addr
, unsigned long end
,
5178 struct mm_walk
*walk
)
5181 struct vm_area_struct
*vma
= walk
->private;
5184 enum mc_target_type target_type
;
5185 union mc_target target
;
5189 * We don't take compound_lock() here but no race with splitting thp
5191 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5192 * under splitting, which means there's no concurrent thp split,
5193 * - if another thread runs into split_huge_page() just after we
5194 * entered this if-block, the thread must wait for page table lock
5195 * to be unlocked in __split_huge_page_splitting(), where the main
5196 * part of thp split is not executed yet.
5198 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5199 if (mc
.precharge
< HPAGE_PMD_NR
) {
5203 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5204 if (target_type
== MC_TARGET_PAGE
) {
5206 if (!isolate_lru_page(page
)) {
5207 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
5209 mc
.precharge
-= HPAGE_PMD_NR
;
5210 mc
.moved_charge
+= HPAGE_PMD_NR
;
5212 putback_lru_page(page
);
5220 if (pmd_trans_unstable(pmd
))
5223 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5224 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5225 pte_t ptent
= *(pte
++);
5231 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5232 case MC_TARGET_PAGE
:
5234 if (isolate_lru_page(page
))
5236 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
5238 /* we uncharge from mc.from later. */
5241 putback_lru_page(page
);
5242 put
: /* get_mctgt_type() gets the page */
5245 case MC_TARGET_SWAP
:
5247 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5249 /* we fixup refcnts and charges later. */
5257 pte_unmap_unlock(pte
- 1, ptl
);
5262 * We have consumed all precharges we got in can_attach().
5263 * We try charge one by one, but don't do any additional
5264 * charges to mc.to if we have failed in charge once in attach()
5267 ret
= mem_cgroup_do_precharge(1);
5275 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
5277 struct vm_area_struct
*vma
;
5279 lru_add_drain_all();
5281 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5282 * move_lock while we're moving its pages to another memcg.
5283 * Then wait for already started RCU-only updates to finish.
5285 atomic_inc(&mc
.from
->moving_account
);
5288 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5290 * Someone who are holding the mmap_sem might be waiting in
5291 * waitq. So we cancel all extra charges, wake up all waiters,
5292 * and retry. Because we cancel precharges, we might not be able
5293 * to move enough charges, but moving charge is a best-effort
5294 * feature anyway, so it wouldn't be a big problem.
5296 __mem_cgroup_clear_mc();
5300 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5302 struct mm_walk mem_cgroup_move_charge_walk
= {
5303 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5307 if (is_vm_hugetlb_page(vma
))
5309 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
5310 &mem_cgroup_move_charge_walk
);
5313 * means we have consumed all precharges and failed in
5314 * doing additional charge. Just abandon here.
5318 up_read(&mm
->mmap_sem
);
5319 atomic_dec(&mc
.from
->moving_account
);
5322 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5323 struct cgroup_taskset
*tset
)
5325 struct task_struct
*p
= cgroup_taskset_first(tset
);
5326 struct mm_struct
*mm
= get_task_mm(p
);
5330 mem_cgroup_move_charge(mm
);
5334 mem_cgroup_clear_mc();
5336 #else /* !CONFIG_MMU */
5337 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5338 struct cgroup_taskset
*tset
)
5342 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5343 struct cgroup_taskset
*tset
)
5346 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5347 struct cgroup_taskset
*tset
)
5353 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5354 * to verify whether we're attached to the default hierarchy on each mount
5357 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5360 * use_hierarchy is forced on the default hierarchy. cgroup core
5361 * guarantees that @root doesn't have any children, so turning it
5362 * on for the root memcg is enough.
5364 if (cgroup_on_dfl(root_css
->cgroup
))
5365 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
5368 struct cgroup_subsys memory_cgrp_subsys
= {
5369 .css_alloc
= mem_cgroup_css_alloc
,
5370 .css_online
= mem_cgroup_css_online
,
5371 .css_offline
= mem_cgroup_css_offline
,
5372 .css_free
= mem_cgroup_css_free
,
5373 .css_reset
= mem_cgroup_css_reset
,
5374 .can_attach
= mem_cgroup_can_attach
,
5375 .cancel_attach
= mem_cgroup_cancel_attach
,
5376 .attach
= mem_cgroup_move_task
,
5377 .bind
= mem_cgroup_bind
,
5378 .legacy_cftypes
= mem_cgroup_files
,
5382 #ifdef CONFIG_MEMCG_SWAP
5383 static int __init
enable_swap_account(char *s
)
5385 if (!strcmp(s
, "1"))
5386 really_do_swap_account
= 1;
5387 else if (!strcmp(s
, "0"))
5388 really_do_swap_account
= 0;
5391 __setup("swapaccount=", enable_swap_account
);
5393 static void __init
memsw_file_init(void)
5395 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5396 memsw_cgroup_files
));
5399 static void __init
enable_swap_cgroup(void)
5401 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5402 do_swap_account
= 1;
5408 static void __init
enable_swap_cgroup(void)
5413 #ifdef CONFIG_MEMCG_SWAP
5415 * mem_cgroup_swapout - transfer a memsw charge to swap
5416 * @page: page whose memsw charge to transfer
5417 * @entry: swap entry to move the charge to
5419 * Transfer the memsw charge of @page to @entry.
5421 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5423 struct mem_cgroup
*memcg
;
5424 unsigned short oldid
;
5426 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5427 VM_BUG_ON_PAGE(page_count(page
), page
);
5429 if (!do_swap_account
)
5432 memcg
= page
->mem_cgroup
;
5434 /* Readahead page, never charged */
5438 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5439 VM_BUG_ON_PAGE(oldid
, page
);
5440 mem_cgroup_swap_statistics(memcg
, true);
5442 page
->mem_cgroup
= NULL
;
5444 if (!mem_cgroup_is_root(memcg
))
5445 page_counter_uncharge(&memcg
->memory
, 1);
5447 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5448 VM_BUG_ON(!irqs_disabled());
5450 mem_cgroup_charge_statistics(memcg
, page
, -1);
5451 memcg_check_events(memcg
, page
);
5455 * mem_cgroup_uncharge_swap - uncharge a swap entry
5456 * @entry: swap entry to uncharge
5458 * Drop the memsw charge associated with @entry.
5460 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5462 struct mem_cgroup
*memcg
;
5465 if (!do_swap_account
)
5468 id
= swap_cgroup_record(entry
, 0);
5470 memcg
= mem_cgroup_lookup(id
);
5472 if (!mem_cgroup_is_root(memcg
))
5473 page_counter_uncharge(&memcg
->memsw
, 1);
5474 mem_cgroup_swap_statistics(memcg
, false);
5475 css_put(&memcg
->css
);
5482 * mem_cgroup_try_charge - try charging a page
5483 * @page: page to charge
5484 * @mm: mm context of the victim
5485 * @gfp_mask: reclaim mode
5486 * @memcgp: charged memcg return
5488 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5489 * pages according to @gfp_mask if necessary.
5491 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5492 * Otherwise, an error code is returned.
5494 * After page->mapping has been set up, the caller must finalize the
5495 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5496 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5498 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5499 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5501 struct mem_cgroup
*memcg
= NULL
;
5502 unsigned int nr_pages
= 1;
5505 if (mem_cgroup_disabled())
5508 if (PageSwapCache(page
)) {
5510 * Every swap fault against a single page tries to charge the
5511 * page, bail as early as possible. shmem_unuse() encounters
5512 * already charged pages, too. The USED bit is protected by
5513 * the page lock, which serializes swap cache removal, which
5514 * in turn serializes uncharging.
5516 if (page
->mem_cgroup
)
5520 if (PageTransHuge(page
)) {
5521 nr_pages
<<= compound_order(page
);
5522 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5525 if (do_swap_account
&& PageSwapCache(page
))
5526 memcg
= try_get_mem_cgroup_from_page(page
);
5528 memcg
= get_mem_cgroup_from_mm(mm
);
5530 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5532 css_put(&memcg
->css
);
5534 if (ret
== -EINTR
) {
5535 memcg
= root_mem_cgroup
;
5544 * mem_cgroup_commit_charge - commit a page charge
5545 * @page: page to charge
5546 * @memcg: memcg to charge the page to
5547 * @lrucare: page might be on LRU already
5549 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5550 * after page->mapping has been set up. This must happen atomically
5551 * as part of the page instantiation, i.e. under the page table lock
5552 * for anonymous pages, under the page lock for page and swap cache.
5554 * In addition, the page must not be on the LRU during the commit, to
5555 * prevent racing with task migration. If it might be, use @lrucare.
5557 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5559 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5562 unsigned int nr_pages
= 1;
5564 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5565 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5567 if (mem_cgroup_disabled())
5570 * Swap faults will attempt to charge the same page multiple
5571 * times. But reuse_swap_page() might have removed the page
5572 * from swapcache already, so we can't check PageSwapCache().
5577 commit_charge(page
, memcg
, lrucare
);
5579 if (PageTransHuge(page
)) {
5580 nr_pages
<<= compound_order(page
);
5581 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5584 local_irq_disable();
5585 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5586 memcg_check_events(memcg
, page
);
5589 if (do_swap_account
&& PageSwapCache(page
)) {
5590 swp_entry_t entry
= { .val
= page_private(page
) };
5592 * The swap entry might not get freed for a long time,
5593 * let's not wait for it. The page already received a
5594 * memory+swap charge, drop the swap entry duplicate.
5596 mem_cgroup_uncharge_swap(entry
);
5601 * mem_cgroup_cancel_charge - cancel a page charge
5602 * @page: page to charge
5603 * @memcg: memcg to charge the page to
5605 * Cancel a charge transaction started by mem_cgroup_try_charge().
5607 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5609 unsigned int nr_pages
= 1;
5611 if (mem_cgroup_disabled())
5614 * Swap faults will attempt to charge the same page multiple
5615 * times. But reuse_swap_page() might have removed the page
5616 * from swapcache already, so we can't check PageSwapCache().
5621 if (PageTransHuge(page
)) {
5622 nr_pages
<<= compound_order(page
);
5623 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5626 cancel_charge(memcg
, nr_pages
);
5629 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5630 unsigned long nr_anon
, unsigned long nr_file
,
5631 unsigned long nr_huge
, struct page
*dummy_page
)
5633 unsigned long nr_pages
= nr_anon
+ nr_file
;
5634 unsigned long flags
;
5636 if (!mem_cgroup_is_root(memcg
)) {
5637 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5638 if (do_swap_account
)
5639 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5640 memcg_oom_recover(memcg
);
5643 local_irq_save(flags
);
5644 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5645 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5646 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5647 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5648 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5649 memcg_check_events(memcg
, dummy_page
);
5650 local_irq_restore(flags
);
5652 if (!mem_cgroup_is_root(memcg
))
5653 css_put_many(&memcg
->css
, nr_pages
);
5656 static void uncharge_list(struct list_head
*page_list
)
5658 struct mem_cgroup
*memcg
= NULL
;
5659 unsigned long nr_anon
= 0;
5660 unsigned long nr_file
= 0;
5661 unsigned long nr_huge
= 0;
5662 unsigned long pgpgout
= 0;
5663 struct list_head
*next
;
5666 next
= page_list
->next
;
5668 unsigned int nr_pages
= 1;
5670 page
= list_entry(next
, struct page
, lru
);
5671 next
= page
->lru
.next
;
5673 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5674 VM_BUG_ON_PAGE(page_count(page
), page
);
5676 if (!page
->mem_cgroup
)
5680 * Nobody should be changing or seriously looking at
5681 * page->mem_cgroup at this point, we have fully
5682 * exclusive access to the page.
5685 if (memcg
!= page
->mem_cgroup
) {
5687 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5689 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5691 memcg
= page
->mem_cgroup
;
5694 if (PageTransHuge(page
)) {
5695 nr_pages
<<= compound_order(page
);
5696 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5697 nr_huge
+= nr_pages
;
5701 nr_anon
+= nr_pages
;
5703 nr_file
+= nr_pages
;
5705 page
->mem_cgroup
= NULL
;
5708 } while (next
!= page_list
);
5711 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5716 * mem_cgroup_uncharge - uncharge a page
5717 * @page: page to uncharge
5719 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5720 * mem_cgroup_commit_charge().
5722 void mem_cgroup_uncharge(struct page
*page
)
5724 if (mem_cgroup_disabled())
5727 /* Don't touch page->lru of any random page, pre-check: */
5728 if (!page
->mem_cgroup
)
5731 INIT_LIST_HEAD(&page
->lru
);
5732 uncharge_list(&page
->lru
);
5736 * mem_cgroup_uncharge_list - uncharge a list of page
5737 * @page_list: list of pages to uncharge
5739 * Uncharge a list of pages previously charged with
5740 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5742 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5744 if (mem_cgroup_disabled())
5747 if (!list_empty(page_list
))
5748 uncharge_list(page_list
);
5752 * mem_cgroup_migrate - migrate a charge to another page
5753 * @oldpage: currently charged page
5754 * @newpage: page to transfer the charge to
5755 * @lrucare: either or both pages might be on the LRU already
5757 * Migrate the charge from @oldpage to @newpage.
5759 * Both pages must be locked, @newpage->mapping must be set up.
5761 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
5764 struct mem_cgroup
*memcg
;
5767 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5768 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5769 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
5770 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
5771 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5772 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5775 if (mem_cgroup_disabled())
5778 /* Page cache replacement: new page already charged? */
5779 if (newpage
->mem_cgroup
)
5783 * Swapcache readahead pages can get migrated before being
5784 * charged, and migration from compaction can happen to an
5785 * uncharged page when the PFN walker finds a page that
5786 * reclaim just put back on the LRU but has not released yet.
5788 memcg
= oldpage
->mem_cgroup
;
5793 lock_page_lru(oldpage
, &isolated
);
5795 oldpage
->mem_cgroup
= NULL
;
5798 unlock_page_lru(oldpage
, isolated
);
5800 commit_charge(newpage
, memcg
, lrucare
);
5804 * subsys_initcall() for memory controller.
5806 * Some parts like hotcpu_notifier() have to be initialized from this context
5807 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5808 * everything that doesn't depend on a specific mem_cgroup structure should
5809 * be initialized from here.
5811 static int __init
mem_cgroup_init(void)
5813 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5814 enable_swap_cgroup();
5815 mem_cgroup_soft_limit_tree_init();
5819 subsys_initcall(mem_cgroup_init
);