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 /* Whether the swap controller is active */
76 #ifdef CONFIG_MEMCG_SWAP
77 int do_swap_account __read_mostly
;
79 #define do_swap_account 0
82 static const char * const mem_cgroup_stat_names
[] = {
91 static const char * const mem_cgroup_events_names
[] = {
98 static const char * const mem_cgroup_lru_names
[] = {
107 * Per memcg event counter is incremented at every pagein/pageout. With THP,
108 * it will be incremated by the number of pages. This counter is used for
109 * for trigger some periodic events. This is straightforward and better
110 * than using jiffies etc. to handle periodic memcg event.
112 enum mem_cgroup_events_target
{
113 MEM_CGROUP_TARGET_THRESH
,
114 MEM_CGROUP_TARGET_SOFTLIMIT
,
115 MEM_CGROUP_TARGET_NUMAINFO
,
118 #define THRESHOLDS_EVENTS_TARGET 128
119 #define SOFTLIMIT_EVENTS_TARGET 1024
120 #define NUMAINFO_EVENTS_TARGET 1024
122 struct mem_cgroup_stat_cpu
{
123 long count
[MEM_CGROUP_STAT_NSTATS
];
124 unsigned long events
[MEMCG_NR_EVENTS
];
125 unsigned long nr_page_events
;
126 unsigned long targets
[MEM_CGROUP_NTARGETS
];
129 struct reclaim_iter
{
130 struct mem_cgroup
*position
;
131 /* scan generation, increased every round-trip */
132 unsigned int generation
;
136 * per-zone information in memory controller.
138 struct mem_cgroup_per_zone
{
139 struct lruvec lruvec
;
140 unsigned long lru_size
[NR_LRU_LISTS
];
142 struct reclaim_iter iter
[DEF_PRIORITY
+ 1];
144 struct rb_node tree_node
; /* RB tree node */
145 unsigned long usage_in_excess
;/* Set to the value by which */
146 /* the soft limit is exceeded*/
148 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
149 /* use container_of */
152 struct mem_cgroup_per_node
{
153 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
157 * Cgroups above their limits are maintained in a RB-Tree, independent of
158 * their hierarchy representation
161 struct mem_cgroup_tree_per_zone
{
162 struct rb_root rb_root
;
166 struct mem_cgroup_tree_per_node
{
167 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
170 struct mem_cgroup_tree
{
171 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
174 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
176 struct mem_cgroup_threshold
{
177 struct eventfd_ctx
*eventfd
;
178 unsigned long threshold
;
182 struct mem_cgroup_threshold_ary
{
183 /* An array index points to threshold just below or equal to usage. */
184 int current_threshold
;
185 /* Size of entries[] */
187 /* Array of thresholds */
188 struct mem_cgroup_threshold entries
[0];
191 struct mem_cgroup_thresholds
{
192 /* Primary thresholds array */
193 struct mem_cgroup_threshold_ary
*primary
;
195 * Spare threshold array.
196 * This is needed to make mem_cgroup_unregister_event() "never fail".
197 * It must be able to store at least primary->size - 1 entries.
199 struct mem_cgroup_threshold_ary
*spare
;
203 struct mem_cgroup_eventfd_list
{
204 struct list_head list
;
205 struct eventfd_ctx
*eventfd
;
209 * cgroup_event represents events which userspace want to receive.
211 struct mem_cgroup_event
{
213 * memcg which the event belongs to.
215 struct mem_cgroup
*memcg
;
217 * eventfd to signal userspace about the event.
219 struct eventfd_ctx
*eventfd
;
221 * Each of these stored in a list by the cgroup.
223 struct list_head list
;
225 * register_event() callback will be used to add new userspace
226 * waiter for changes related to this event. Use eventfd_signal()
227 * on eventfd to send notification to userspace.
229 int (*register_event
)(struct mem_cgroup
*memcg
,
230 struct eventfd_ctx
*eventfd
, const char *args
);
232 * unregister_event() callback will be called when userspace closes
233 * the eventfd or on cgroup removing. This callback must be set,
234 * if you want provide notification functionality.
236 void (*unregister_event
)(struct mem_cgroup
*memcg
,
237 struct eventfd_ctx
*eventfd
);
239 * All fields below needed to unregister event when
240 * userspace closes eventfd.
243 wait_queue_head_t
*wqh
;
245 struct work_struct remove
;
248 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
249 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
252 * The memory controller data structure. The memory controller controls both
253 * page cache and RSS per cgroup. We would eventually like to provide
254 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
255 * to help the administrator determine what knobs to tune.
257 * TODO: Add a water mark for the memory controller. Reclaim will begin when
258 * we hit the water mark. May be even add a low water mark, such that
259 * no reclaim occurs from a cgroup at it's low water mark, this is
260 * a feature that will be implemented much later in the future.
263 struct cgroup_subsys_state css
;
265 /* Accounted resources */
266 struct page_counter memory
;
267 struct page_counter memsw
;
268 struct page_counter kmem
;
270 /* Normal memory consumption range */
274 unsigned long soft_limit
;
276 /* vmpressure notifications */
277 struct vmpressure vmpressure
;
279 /* css_online() has been completed */
283 * Should the accounting and control be hierarchical, per subtree?
289 atomic_t oom_wakeups
;
292 /* OOM-Killer disable */
293 int oom_kill_disable
;
295 /* protect arrays of thresholds */
296 struct mutex thresholds_lock
;
298 /* thresholds for memory usage. RCU-protected */
299 struct mem_cgroup_thresholds thresholds
;
301 /* thresholds for mem+swap usage. RCU-protected */
302 struct mem_cgroup_thresholds memsw_thresholds
;
304 /* For oom notifier event fd */
305 struct list_head oom_notify
;
308 * Should we move charges of a task when a task is moved into this
309 * mem_cgroup ? And what type of charges should we move ?
311 unsigned long move_charge_at_immigrate
;
313 * set > 0 if pages under this cgroup are moving to other cgroup.
315 atomic_t moving_account
;
316 /* taken only while moving_account > 0 */
317 spinlock_t move_lock
;
318 struct task_struct
*move_lock_task
;
319 unsigned long move_lock_flags
;
323 struct mem_cgroup_stat_cpu __percpu
*stat
;
325 * used when a cpu is offlined or other synchronizations
326 * See mem_cgroup_read_stat().
328 struct mem_cgroup_stat_cpu nocpu_base
;
329 spinlock_t pcp_counter_lock
;
331 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
332 struct cg_proto tcp_mem
;
334 #if defined(CONFIG_MEMCG_KMEM)
335 /* Index in the kmem_cache->memcg_params.memcg_caches array */
337 bool kmem_acct_activated
;
338 bool kmem_acct_active
;
341 int last_scanned_node
;
343 nodemask_t scan_nodes
;
344 atomic_t numainfo_events
;
345 atomic_t numainfo_updating
;
348 /* List of events which userspace want to receive */
349 struct list_head event_list
;
350 spinlock_t event_list_lock
;
352 struct mem_cgroup_per_node
*nodeinfo
[0];
353 /* WARNING: nodeinfo must be the last member here */
356 #ifdef CONFIG_MEMCG_KMEM
357 bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
359 return memcg
->kmem_acct_active
;
363 /* Stuffs for move charges at task migration. */
365 * Types of charges to be moved.
367 #define MOVE_ANON 0x1U
368 #define MOVE_FILE 0x2U
369 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
371 /* "mc" and its members are protected by cgroup_mutex */
372 static struct move_charge_struct
{
373 spinlock_t lock
; /* for from, to */
374 struct mem_cgroup
*from
;
375 struct mem_cgroup
*to
;
377 unsigned long precharge
;
378 unsigned long moved_charge
;
379 unsigned long moved_swap
;
380 struct task_struct
*moving_task
; /* a task moving charges */
381 wait_queue_head_t waitq
; /* a waitq for other context */
383 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
384 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
388 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
389 * limit reclaim to prevent infinite loops, if they ever occur.
391 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
392 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
395 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
396 MEM_CGROUP_CHARGE_TYPE_ANON
,
397 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
398 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
402 /* for encoding cft->private value on file */
410 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
411 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
412 #define MEMFILE_ATTR(val) ((val) & 0xffff)
413 /* Used for OOM nofiier */
414 #define OOM_CONTROL (0)
417 * The memcg_create_mutex will be held whenever a new cgroup is created.
418 * As a consequence, any change that needs to protect against new child cgroups
419 * appearing has to hold it as well.
421 static DEFINE_MUTEX(memcg_create_mutex
);
423 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
425 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
428 /* Some nice accessors for the vmpressure. */
429 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
432 memcg
= root_mem_cgroup
;
433 return &memcg
->vmpressure
;
436 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
438 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
441 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
443 return (memcg
== root_mem_cgroup
);
447 * We restrict the id in the range of [1, 65535], so it can fit into
450 #define MEM_CGROUP_ID_MAX USHRT_MAX
452 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
454 return memcg
->css
.id
;
457 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
459 struct cgroup_subsys_state
*css
;
461 css
= css_from_id(id
, &memory_cgrp_subsys
);
462 return mem_cgroup_from_css(css
);
465 /* Writing them here to avoid exposing memcg's inner layout */
466 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
468 void sock_update_memcg(struct sock
*sk
)
470 if (mem_cgroup_sockets_enabled
) {
471 struct mem_cgroup
*memcg
;
472 struct cg_proto
*cg_proto
;
474 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
476 /* Socket cloning can throw us here with sk_cgrp already
477 * filled. It won't however, necessarily happen from
478 * process context. So the test for root memcg given
479 * the current task's memcg won't help us in this case.
481 * Respecting the original socket's memcg is a better
482 * decision in this case.
485 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
486 css_get(&sk
->sk_cgrp
->memcg
->css
);
491 memcg
= mem_cgroup_from_task(current
);
492 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
493 if (!mem_cgroup_is_root(memcg
) &&
494 memcg_proto_active(cg_proto
) &&
495 css_tryget_online(&memcg
->css
)) {
496 sk
->sk_cgrp
= cg_proto
;
501 EXPORT_SYMBOL(sock_update_memcg
);
503 void sock_release_memcg(struct sock
*sk
)
505 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
506 struct mem_cgroup
*memcg
;
507 WARN_ON(!sk
->sk_cgrp
->memcg
);
508 memcg
= sk
->sk_cgrp
->memcg
;
509 css_put(&sk
->sk_cgrp
->memcg
->css
);
513 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
515 if (!memcg
|| mem_cgroup_is_root(memcg
))
518 return &memcg
->tcp_mem
;
520 EXPORT_SYMBOL(tcp_proto_cgroup
);
524 #ifdef CONFIG_MEMCG_KMEM
526 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
527 * The main reason for not using cgroup id for this:
528 * this works better in sparse environments, where we have a lot of memcgs,
529 * but only a few kmem-limited. Or also, if we have, for instance, 200
530 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
531 * 200 entry array for that.
533 * The current size of the caches array is stored in memcg_nr_cache_ids. It
534 * will double each time we have to increase it.
536 static DEFINE_IDA(memcg_cache_ida
);
537 int memcg_nr_cache_ids
;
539 /* Protects memcg_nr_cache_ids */
540 static DECLARE_RWSEM(memcg_cache_ids_sem
);
542 void memcg_get_cache_ids(void)
544 down_read(&memcg_cache_ids_sem
);
547 void memcg_put_cache_ids(void)
549 up_read(&memcg_cache_ids_sem
);
553 * MIN_SIZE is different than 1, because we would like to avoid going through
554 * the alloc/free process all the time. In a small machine, 4 kmem-limited
555 * cgroups is a reasonable guess. In the future, it could be a parameter or
556 * tunable, but that is strictly not necessary.
558 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
559 * this constant directly from cgroup, but it is understandable that this is
560 * better kept as an internal representation in cgroup.c. In any case, the
561 * cgrp_id space is not getting any smaller, and we don't have to necessarily
562 * increase ours as well if it increases.
564 #define MEMCG_CACHES_MIN_SIZE 4
565 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
568 * A lot of the calls to the cache allocation functions are expected to be
569 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
570 * conditional to this static branch, we'll have to allow modules that does
571 * kmem_cache_alloc and the such to see this symbol as well
573 struct static_key memcg_kmem_enabled_key
;
574 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
576 #endif /* CONFIG_MEMCG_KMEM */
578 static struct mem_cgroup_per_zone
*
579 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
581 int nid
= zone_to_nid(zone
);
582 int zid
= zone_idx(zone
);
584 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
587 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
592 static struct mem_cgroup_per_zone
*
593 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
595 int nid
= page_to_nid(page
);
596 int zid
= page_zonenum(page
);
598 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
601 static struct mem_cgroup_tree_per_zone
*
602 soft_limit_tree_node_zone(int nid
, int zid
)
604 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
607 static struct mem_cgroup_tree_per_zone
*
608 soft_limit_tree_from_page(struct page
*page
)
610 int nid
= page_to_nid(page
);
611 int zid
= page_zonenum(page
);
613 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
616 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
617 struct mem_cgroup_tree_per_zone
*mctz
,
618 unsigned long new_usage_in_excess
)
620 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
621 struct rb_node
*parent
= NULL
;
622 struct mem_cgroup_per_zone
*mz_node
;
627 mz
->usage_in_excess
= new_usage_in_excess
;
628 if (!mz
->usage_in_excess
)
632 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
634 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
637 * We can't avoid mem cgroups that are over their soft
638 * limit by the same amount
640 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
643 rb_link_node(&mz
->tree_node
, parent
, p
);
644 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
648 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
649 struct mem_cgroup_tree_per_zone
*mctz
)
653 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
657 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
658 struct mem_cgroup_tree_per_zone
*mctz
)
662 spin_lock_irqsave(&mctz
->lock
, flags
);
663 __mem_cgroup_remove_exceeded(mz
, mctz
);
664 spin_unlock_irqrestore(&mctz
->lock
, flags
);
667 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
669 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
670 unsigned long soft_limit
= ACCESS_ONCE(memcg
->soft_limit
);
671 unsigned long excess
= 0;
673 if (nr_pages
> soft_limit
)
674 excess
= nr_pages
- soft_limit
;
679 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
681 unsigned long excess
;
682 struct mem_cgroup_per_zone
*mz
;
683 struct mem_cgroup_tree_per_zone
*mctz
;
685 mctz
= soft_limit_tree_from_page(page
);
687 * Necessary to update all ancestors when hierarchy is used.
688 * because their event counter is not touched.
690 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
691 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
692 excess
= soft_limit_excess(memcg
);
694 * We have to update the tree if mz is on RB-tree or
695 * mem is over its softlimit.
697 if (excess
|| mz
->on_tree
) {
700 spin_lock_irqsave(&mctz
->lock
, flags
);
701 /* if on-tree, remove it */
703 __mem_cgroup_remove_exceeded(mz
, mctz
);
705 * Insert again. mz->usage_in_excess will be updated.
706 * If excess is 0, no tree ops.
708 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
709 spin_unlock_irqrestore(&mctz
->lock
, flags
);
714 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
716 struct mem_cgroup_tree_per_zone
*mctz
;
717 struct mem_cgroup_per_zone
*mz
;
721 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
722 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
723 mctz
= soft_limit_tree_node_zone(nid
, zid
);
724 mem_cgroup_remove_exceeded(mz
, mctz
);
729 static struct mem_cgroup_per_zone
*
730 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
732 struct rb_node
*rightmost
= NULL
;
733 struct mem_cgroup_per_zone
*mz
;
737 rightmost
= rb_last(&mctz
->rb_root
);
739 goto done
; /* Nothing to reclaim from */
741 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
743 * Remove the node now but someone else can add it back,
744 * we will to add it back at the end of reclaim to its correct
745 * position in the tree.
747 __mem_cgroup_remove_exceeded(mz
, mctz
);
748 if (!soft_limit_excess(mz
->memcg
) ||
749 !css_tryget_online(&mz
->memcg
->css
))
755 static struct mem_cgroup_per_zone
*
756 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
758 struct mem_cgroup_per_zone
*mz
;
760 spin_lock_irq(&mctz
->lock
);
761 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
762 spin_unlock_irq(&mctz
->lock
);
767 * Implementation Note: reading percpu statistics for memcg.
769 * Both of vmstat[] and percpu_counter has threshold and do periodic
770 * synchronization to implement "quick" read. There are trade-off between
771 * reading cost and precision of value. Then, we may have a chance to implement
772 * a periodic synchronizion of counter in memcg's counter.
774 * But this _read() function is used for user interface now. The user accounts
775 * memory usage by memory cgroup and he _always_ requires exact value because
776 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
777 * have to visit all online cpus and make sum. So, for now, unnecessary
778 * synchronization is not implemented. (just implemented for cpu hotplug)
780 * If there are kernel internal actions which can make use of some not-exact
781 * value, and reading all cpu value can be performance bottleneck in some
782 * common workload, threashold and synchonization as vmstat[] should be
785 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
786 enum mem_cgroup_stat_index idx
)
792 for_each_online_cpu(cpu
)
793 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
794 #ifdef CONFIG_HOTPLUG_CPU
795 spin_lock(&memcg
->pcp_counter_lock
);
796 val
+= memcg
->nocpu_base
.count
[idx
];
797 spin_unlock(&memcg
->pcp_counter_lock
);
803 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
804 enum mem_cgroup_events_index idx
)
806 unsigned long val
= 0;
810 for_each_online_cpu(cpu
)
811 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
812 #ifdef CONFIG_HOTPLUG_CPU
813 spin_lock(&memcg
->pcp_counter_lock
);
814 val
+= memcg
->nocpu_base
.events
[idx
];
815 spin_unlock(&memcg
->pcp_counter_lock
);
821 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
826 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
827 * counted as CACHE even if it's on ANON LRU.
830 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
833 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
836 if (PageTransHuge(page
))
837 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
840 /* pagein of a big page is an event. So, ignore page size */
842 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
844 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
845 nr_pages
= -nr_pages
; /* for event */
848 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
851 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
853 struct mem_cgroup_per_zone
*mz
;
855 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
856 return mz
->lru_size
[lru
];
859 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
861 unsigned int lru_mask
)
863 unsigned long nr
= 0;
866 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
868 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
869 struct mem_cgroup_per_zone
*mz
;
873 if (!(BIT(lru
) & lru_mask
))
875 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
876 nr
+= mz
->lru_size
[lru
];
882 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
883 unsigned int lru_mask
)
885 unsigned long nr
= 0;
888 for_each_node_state(nid
, N_MEMORY
)
889 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
893 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
894 enum mem_cgroup_events_target target
)
896 unsigned long val
, next
;
898 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
899 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
900 /* from time_after() in jiffies.h */
901 if ((long)next
- (long)val
< 0) {
903 case MEM_CGROUP_TARGET_THRESH
:
904 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
906 case MEM_CGROUP_TARGET_SOFTLIMIT
:
907 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
909 case MEM_CGROUP_TARGET_NUMAINFO
:
910 next
= val
+ NUMAINFO_EVENTS_TARGET
;
915 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
922 * Check events in order.
925 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
927 /* threshold event is triggered in finer grain than soft limit */
928 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
929 MEM_CGROUP_TARGET_THRESH
))) {
931 bool do_numainfo __maybe_unused
;
933 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
934 MEM_CGROUP_TARGET_SOFTLIMIT
);
936 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
937 MEM_CGROUP_TARGET_NUMAINFO
);
939 mem_cgroup_threshold(memcg
);
940 if (unlikely(do_softlimit
))
941 mem_cgroup_update_tree(memcg
, page
);
943 if (unlikely(do_numainfo
))
944 atomic_inc(&memcg
->numainfo_events
);
949 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
952 * mm_update_next_owner() may clear mm->owner to NULL
953 * if it races with swapoff, page migration, etc.
954 * So this can be called with p == NULL.
959 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
962 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
964 struct mem_cgroup
*memcg
= NULL
;
969 * Page cache insertions can happen withou an
970 * actual mm context, e.g. during disk probing
971 * on boot, loopback IO, acct() writes etc.
974 memcg
= root_mem_cgroup
;
976 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
977 if (unlikely(!memcg
))
978 memcg
= root_mem_cgroup
;
980 } while (!css_tryget_online(&memcg
->css
));
986 * mem_cgroup_iter - iterate over memory cgroup hierarchy
987 * @root: hierarchy root
988 * @prev: previously returned memcg, NULL on first invocation
989 * @reclaim: cookie for shared reclaim walks, NULL for full walks
991 * Returns references to children of the hierarchy below @root, or
992 * @root itself, or %NULL after a full round-trip.
994 * Caller must pass the return value in @prev on subsequent
995 * invocations for reference counting, or use mem_cgroup_iter_break()
996 * to cancel a hierarchy walk before the round-trip is complete.
998 * Reclaimers can specify a zone and a priority level in @reclaim to
999 * divide up the memcgs in the hierarchy among all concurrent
1000 * reclaimers operating on the same zone and priority.
1002 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1003 struct mem_cgroup
*prev
,
1004 struct mem_cgroup_reclaim_cookie
*reclaim
)
1006 struct reclaim_iter
*uninitialized_var(iter
);
1007 struct cgroup_subsys_state
*css
= NULL
;
1008 struct mem_cgroup
*memcg
= NULL
;
1009 struct mem_cgroup
*pos
= NULL
;
1011 if (mem_cgroup_disabled())
1015 root
= root_mem_cgroup
;
1017 if (prev
&& !reclaim
)
1020 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1029 struct mem_cgroup_per_zone
*mz
;
1031 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1032 iter
= &mz
->iter
[reclaim
->priority
];
1034 if (prev
&& reclaim
->generation
!= iter
->generation
)
1038 pos
= ACCESS_ONCE(iter
->position
);
1040 * A racing update may change the position and
1041 * put the last reference, hence css_tryget(),
1042 * or retry to see the updated position.
1044 } while (pos
&& !css_tryget(&pos
->css
));
1051 css
= css_next_descendant_pre(css
, &root
->css
);
1054 * Reclaimers share the hierarchy walk, and a
1055 * new one might jump in right at the end of
1056 * the hierarchy - make sure they see at least
1057 * one group and restart from the beginning.
1065 * Verify the css and acquire a reference. The root
1066 * is provided by the caller, so we know it's alive
1067 * and kicking, and don't take an extra reference.
1069 memcg
= mem_cgroup_from_css(css
);
1071 if (css
== &root
->css
)
1074 if (css_tryget(css
)) {
1076 * Make sure the memcg is initialized:
1077 * mem_cgroup_css_online() orders the the
1078 * initialization against setting the flag.
1080 if (smp_load_acquire(&memcg
->initialized
))
1090 if (cmpxchg(&iter
->position
, pos
, memcg
) == pos
) {
1092 css_get(&memcg
->css
);
1098 * pairs with css_tryget when dereferencing iter->position
1107 reclaim
->generation
= iter
->generation
;
1113 if (prev
&& prev
!= root
)
1114 css_put(&prev
->css
);
1120 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1121 * @root: hierarchy root
1122 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1124 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1125 struct mem_cgroup
*prev
)
1128 root
= root_mem_cgroup
;
1129 if (prev
&& prev
!= root
)
1130 css_put(&prev
->css
);
1134 * Iteration constructs for visiting all cgroups (under a tree). If
1135 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1136 * be used for reference counting.
1138 #define for_each_mem_cgroup_tree(iter, root) \
1139 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1141 iter = mem_cgroup_iter(root, iter, NULL))
1143 #define for_each_mem_cgroup(iter) \
1144 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1146 iter = mem_cgroup_iter(NULL, iter, NULL))
1148 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1150 struct mem_cgroup
*memcg
;
1153 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1154 if (unlikely(!memcg
))
1159 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1162 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1170 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1173 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1174 * @zone: zone of the wanted lruvec
1175 * @memcg: memcg of the wanted lruvec
1177 * Returns the lru list vector holding pages for the given @zone and
1178 * @mem. This can be the global zone lruvec, if the memory controller
1181 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1182 struct mem_cgroup
*memcg
)
1184 struct mem_cgroup_per_zone
*mz
;
1185 struct lruvec
*lruvec
;
1187 if (mem_cgroup_disabled()) {
1188 lruvec
= &zone
->lruvec
;
1192 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1193 lruvec
= &mz
->lruvec
;
1196 * Since a node can be onlined after the mem_cgroup was created,
1197 * we have to be prepared to initialize lruvec->zone here;
1198 * and if offlined then reonlined, we need to reinitialize it.
1200 if (unlikely(lruvec
->zone
!= zone
))
1201 lruvec
->zone
= zone
;
1206 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1208 * @zone: zone of the page
1210 * This function is only safe when following the LRU page isolation
1211 * and putback protocol: the LRU lock must be held, and the page must
1212 * either be PageLRU() or the caller must have isolated/allocated it.
1214 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1216 struct mem_cgroup_per_zone
*mz
;
1217 struct mem_cgroup
*memcg
;
1218 struct lruvec
*lruvec
;
1220 if (mem_cgroup_disabled()) {
1221 lruvec
= &zone
->lruvec
;
1225 memcg
= page
->mem_cgroup
;
1227 * Swapcache readahead pages are added to the LRU - and
1228 * possibly migrated - before they are charged.
1231 memcg
= root_mem_cgroup
;
1233 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1234 lruvec
= &mz
->lruvec
;
1237 * Since a node can be onlined after the mem_cgroup was created,
1238 * we have to be prepared to initialize lruvec->zone here;
1239 * and if offlined then reonlined, we need to reinitialize it.
1241 if (unlikely(lruvec
->zone
!= zone
))
1242 lruvec
->zone
= zone
;
1247 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1248 * @lruvec: mem_cgroup per zone lru vector
1249 * @lru: index of lru list the page is sitting on
1250 * @nr_pages: positive when adding or negative when removing
1252 * This function must be called when a page is added to or removed from an
1255 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1258 struct mem_cgroup_per_zone
*mz
;
1259 unsigned long *lru_size
;
1261 if (mem_cgroup_disabled())
1264 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1265 lru_size
= mz
->lru_size
+ lru
;
1266 *lru_size
+= nr_pages
;
1267 VM_BUG_ON((long)(*lru_size
) < 0);
1270 bool mem_cgroup_is_descendant(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
)
1274 if (!root
->use_hierarchy
)
1276 return cgroup_is_descendant(memcg
->css
.cgroup
, root
->css
.cgroup
);
1279 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1281 struct mem_cgroup
*task_memcg
;
1282 struct task_struct
*p
;
1285 p
= find_lock_task_mm(task
);
1287 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1291 * All threads may have already detached their mm's, but the oom
1292 * killer still needs to detect if they have already been oom
1293 * killed to prevent needlessly killing additional tasks.
1296 task_memcg
= mem_cgroup_from_task(task
);
1297 css_get(&task_memcg
->css
);
1300 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1301 css_put(&task_memcg
->css
);
1305 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1307 unsigned long inactive_ratio
;
1308 unsigned long inactive
;
1309 unsigned long active
;
1312 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1313 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1315 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1317 inactive_ratio
= int_sqrt(10 * gb
);
1321 return inactive
* inactive_ratio
< active
;
1324 bool mem_cgroup_lruvec_online(struct lruvec
*lruvec
)
1326 struct mem_cgroup_per_zone
*mz
;
1327 struct mem_cgroup
*memcg
;
1329 if (mem_cgroup_disabled())
1332 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1335 return !!(memcg
->css
.flags
& CSS_ONLINE
);
1338 #define mem_cgroup_from_counter(counter, member) \
1339 container_of(counter, struct mem_cgroup, member)
1342 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1343 * @memcg: the memory cgroup
1345 * Returns the maximum amount of memory @mem can be charged with, in
1348 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1350 unsigned long margin
= 0;
1351 unsigned long count
;
1352 unsigned long limit
;
1354 count
= page_counter_read(&memcg
->memory
);
1355 limit
= ACCESS_ONCE(memcg
->memory
.limit
);
1357 margin
= limit
- count
;
1359 if (do_swap_account
) {
1360 count
= page_counter_read(&memcg
->memsw
);
1361 limit
= ACCESS_ONCE(memcg
->memsw
.limit
);
1363 margin
= min(margin
, limit
- count
);
1369 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1372 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1373 return vm_swappiness
;
1375 return memcg
->swappiness
;
1379 * A routine for checking "mem" is under move_account() or not.
1381 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1382 * moving cgroups. This is for waiting at high-memory pressure
1385 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1387 struct mem_cgroup
*from
;
1388 struct mem_cgroup
*to
;
1391 * Unlike task_move routines, we access mc.to, mc.from not under
1392 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1394 spin_lock(&mc
.lock
);
1400 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1401 mem_cgroup_is_descendant(to
, memcg
);
1403 spin_unlock(&mc
.lock
);
1407 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1409 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1410 if (mem_cgroup_under_move(memcg
)) {
1412 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1413 /* moving charge context might have finished. */
1416 finish_wait(&mc
.waitq
, &wait
);
1423 #define K(x) ((x) << (PAGE_SHIFT-10))
1425 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1426 * @memcg: The memory cgroup that went over limit
1427 * @p: Task that is going to be killed
1429 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1432 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1434 /* oom_info_lock ensures that parallel ooms do not interleave */
1435 static DEFINE_MUTEX(oom_info_lock
);
1436 struct mem_cgroup
*iter
;
1442 mutex_lock(&oom_info_lock
);
1445 pr_info("Task in ");
1446 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1447 pr_cont(" killed as a result of limit of ");
1448 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1453 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1454 K((u64
)page_counter_read(&memcg
->memory
)),
1455 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1456 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1457 K((u64
)page_counter_read(&memcg
->memsw
)),
1458 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1459 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1460 K((u64
)page_counter_read(&memcg
->kmem
)),
1461 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1463 for_each_mem_cgroup_tree(iter
, memcg
) {
1464 pr_info("Memory cgroup stats for ");
1465 pr_cont_cgroup_path(iter
->css
.cgroup
);
1468 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1469 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1471 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1472 K(mem_cgroup_read_stat(iter
, i
)));
1475 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1476 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1477 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1481 mutex_unlock(&oom_info_lock
);
1485 * This function returns the number of memcg under hierarchy tree. Returns
1486 * 1(self count) if no children.
1488 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1491 struct mem_cgroup
*iter
;
1493 for_each_mem_cgroup_tree(iter
, memcg
)
1499 * Return the memory (and swap, if configured) limit for a memcg.
1501 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1503 unsigned long limit
;
1505 limit
= memcg
->memory
.limit
;
1506 if (mem_cgroup_swappiness(memcg
)) {
1507 unsigned long memsw_limit
;
1509 memsw_limit
= memcg
->memsw
.limit
;
1510 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1515 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1518 struct mem_cgroup
*iter
;
1519 unsigned long chosen_points
= 0;
1520 unsigned long totalpages
;
1521 unsigned int points
= 0;
1522 struct task_struct
*chosen
= NULL
;
1525 * If current has a pending SIGKILL or is exiting, then automatically
1526 * select it. The goal is to allow it to allocate so that it may
1527 * quickly exit and free its memory.
1529 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1530 mark_tsk_oom_victim(current
);
1534 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1535 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1536 for_each_mem_cgroup_tree(iter
, memcg
) {
1537 struct css_task_iter it
;
1538 struct task_struct
*task
;
1540 css_task_iter_start(&iter
->css
, &it
);
1541 while ((task
= css_task_iter_next(&it
))) {
1542 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1544 case OOM_SCAN_SELECT
:
1546 put_task_struct(chosen
);
1548 chosen_points
= ULONG_MAX
;
1549 get_task_struct(chosen
);
1551 case OOM_SCAN_CONTINUE
:
1553 case OOM_SCAN_ABORT
:
1554 css_task_iter_end(&it
);
1555 mem_cgroup_iter_break(memcg
, iter
);
1557 put_task_struct(chosen
);
1562 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1563 if (!points
|| points
< chosen_points
)
1565 /* Prefer thread group leaders for display purposes */
1566 if (points
== chosen_points
&&
1567 thread_group_leader(chosen
))
1571 put_task_struct(chosen
);
1573 chosen_points
= points
;
1574 get_task_struct(chosen
);
1576 css_task_iter_end(&it
);
1581 points
= chosen_points
* 1000 / totalpages
;
1582 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1583 NULL
, "Memory cgroup out of memory");
1586 #if MAX_NUMNODES > 1
1589 * test_mem_cgroup_node_reclaimable
1590 * @memcg: the target memcg
1591 * @nid: the node ID to be checked.
1592 * @noswap : specify true here if the user wants flle only information.
1594 * This function returns whether the specified memcg contains any
1595 * reclaimable pages on a node. Returns true if there are any reclaimable
1596 * pages in the node.
1598 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1599 int nid
, bool noswap
)
1601 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1603 if (noswap
|| !total_swap_pages
)
1605 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1612 * Always updating the nodemask is not very good - even if we have an empty
1613 * list or the wrong list here, we can start from some node and traverse all
1614 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1617 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1621 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1622 * pagein/pageout changes since the last update.
1624 if (!atomic_read(&memcg
->numainfo_events
))
1626 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1629 /* make a nodemask where this memcg uses memory from */
1630 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1632 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1634 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1635 node_clear(nid
, memcg
->scan_nodes
);
1638 atomic_set(&memcg
->numainfo_events
, 0);
1639 atomic_set(&memcg
->numainfo_updating
, 0);
1643 * Selecting a node where we start reclaim from. Because what we need is just
1644 * reducing usage counter, start from anywhere is O,K. Considering
1645 * memory reclaim from current node, there are pros. and cons.
1647 * Freeing memory from current node means freeing memory from a node which
1648 * we'll use or we've used. So, it may make LRU bad. And if several threads
1649 * hit limits, it will see a contention on a node. But freeing from remote
1650 * node means more costs for memory reclaim because of memory latency.
1652 * Now, we use round-robin. Better algorithm is welcomed.
1654 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1658 mem_cgroup_may_update_nodemask(memcg
);
1659 node
= memcg
->last_scanned_node
;
1661 node
= next_node(node
, memcg
->scan_nodes
);
1662 if (node
== MAX_NUMNODES
)
1663 node
= first_node(memcg
->scan_nodes
);
1665 * We call this when we hit limit, not when pages are added to LRU.
1666 * No LRU may hold pages because all pages are UNEVICTABLE or
1667 * memcg is too small and all pages are not on LRU. In that case,
1668 * we use curret node.
1670 if (unlikely(node
== MAX_NUMNODES
))
1671 node
= numa_node_id();
1673 memcg
->last_scanned_node
= node
;
1677 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1683 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1686 unsigned long *total_scanned
)
1688 struct mem_cgroup
*victim
= NULL
;
1691 unsigned long excess
;
1692 unsigned long nr_scanned
;
1693 struct mem_cgroup_reclaim_cookie reclaim
= {
1698 excess
= soft_limit_excess(root_memcg
);
1701 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1706 * If we have not been able to reclaim
1707 * anything, it might because there are
1708 * no reclaimable pages under this hierarchy
1713 * We want to do more targeted reclaim.
1714 * excess >> 2 is not to excessive so as to
1715 * reclaim too much, nor too less that we keep
1716 * coming back to reclaim from this cgroup
1718 if (total
>= (excess
>> 2) ||
1719 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1724 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1726 *total_scanned
+= nr_scanned
;
1727 if (!soft_limit_excess(root_memcg
))
1730 mem_cgroup_iter_break(root_memcg
, victim
);
1734 #ifdef CONFIG_LOCKDEP
1735 static struct lockdep_map memcg_oom_lock_dep_map
= {
1736 .name
= "memcg_oom_lock",
1740 static DEFINE_SPINLOCK(memcg_oom_lock
);
1743 * Check OOM-Killer is already running under our hierarchy.
1744 * If someone is running, return false.
1746 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1748 struct mem_cgroup
*iter
, *failed
= NULL
;
1750 spin_lock(&memcg_oom_lock
);
1752 for_each_mem_cgroup_tree(iter
, memcg
) {
1753 if (iter
->oom_lock
) {
1755 * this subtree of our hierarchy is already locked
1756 * so we cannot give a lock.
1759 mem_cgroup_iter_break(memcg
, iter
);
1762 iter
->oom_lock
= true;
1767 * OK, we failed to lock the whole subtree so we have
1768 * to clean up what we set up to the failing subtree
1770 for_each_mem_cgroup_tree(iter
, memcg
) {
1771 if (iter
== failed
) {
1772 mem_cgroup_iter_break(memcg
, iter
);
1775 iter
->oom_lock
= false;
1778 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1780 spin_unlock(&memcg_oom_lock
);
1785 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1787 struct mem_cgroup
*iter
;
1789 spin_lock(&memcg_oom_lock
);
1790 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1791 for_each_mem_cgroup_tree(iter
, memcg
)
1792 iter
->oom_lock
= false;
1793 spin_unlock(&memcg_oom_lock
);
1796 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1798 struct mem_cgroup
*iter
;
1800 for_each_mem_cgroup_tree(iter
, memcg
)
1801 atomic_inc(&iter
->under_oom
);
1804 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1806 struct mem_cgroup
*iter
;
1809 * When a new child is created while the hierarchy is under oom,
1810 * mem_cgroup_oom_lock() may not be called. We have to use
1811 * atomic_add_unless() here.
1813 for_each_mem_cgroup_tree(iter
, memcg
)
1814 atomic_add_unless(&iter
->under_oom
, -1, 0);
1817 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1819 struct oom_wait_info
{
1820 struct mem_cgroup
*memcg
;
1824 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1825 unsigned mode
, int sync
, void *arg
)
1827 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1828 struct mem_cgroup
*oom_wait_memcg
;
1829 struct oom_wait_info
*oom_wait_info
;
1831 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1832 oom_wait_memcg
= oom_wait_info
->memcg
;
1834 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1835 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1837 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1840 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
1842 atomic_inc(&memcg
->oom_wakeups
);
1843 /* for filtering, pass "memcg" as argument. */
1844 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1847 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1849 if (memcg
&& atomic_read(&memcg
->under_oom
))
1850 memcg_wakeup_oom(memcg
);
1853 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1855 if (!current
->memcg_oom
.may_oom
)
1858 * We are in the middle of the charge context here, so we
1859 * don't want to block when potentially sitting on a callstack
1860 * that holds all kinds of filesystem and mm locks.
1862 * Also, the caller may handle a failed allocation gracefully
1863 * (like optional page cache readahead) and so an OOM killer
1864 * invocation might not even be necessary.
1866 * That's why we don't do anything here except remember the
1867 * OOM context and then deal with it at the end of the page
1868 * fault when the stack is unwound, the locks are released,
1869 * and when we know whether the fault was overall successful.
1871 css_get(&memcg
->css
);
1872 current
->memcg_oom
.memcg
= memcg
;
1873 current
->memcg_oom
.gfp_mask
= mask
;
1874 current
->memcg_oom
.order
= order
;
1878 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1879 * @handle: actually kill/wait or just clean up the OOM state
1881 * This has to be called at the end of a page fault if the memcg OOM
1882 * handler was enabled.
1884 * Memcg supports userspace OOM handling where failed allocations must
1885 * sleep on a waitqueue until the userspace task resolves the
1886 * situation. Sleeping directly in the charge context with all kinds
1887 * of locks held is not a good idea, instead we remember an OOM state
1888 * in the task and mem_cgroup_oom_synchronize() has to be called at
1889 * the end of the page fault to complete the OOM handling.
1891 * Returns %true if an ongoing memcg OOM situation was detected and
1892 * completed, %false otherwise.
1894 bool mem_cgroup_oom_synchronize(bool handle
)
1896 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
1897 struct oom_wait_info owait
;
1900 /* OOM is global, do not handle */
1904 if (!handle
|| oom_killer_disabled
)
1907 owait
.memcg
= memcg
;
1908 owait
.wait
.flags
= 0;
1909 owait
.wait
.func
= memcg_oom_wake_function
;
1910 owait
.wait
.private = current
;
1911 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1913 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1914 mem_cgroup_mark_under_oom(memcg
);
1916 locked
= mem_cgroup_oom_trylock(memcg
);
1919 mem_cgroup_oom_notify(memcg
);
1921 if (locked
&& !memcg
->oom_kill_disable
) {
1922 mem_cgroup_unmark_under_oom(memcg
);
1923 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1924 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
1925 current
->memcg_oom
.order
);
1928 mem_cgroup_unmark_under_oom(memcg
);
1929 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1933 mem_cgroup_oom_unlock(memcg
);
1935 * There is no guarantee that an OOM-lock contender
1936 * sees the wakeups triggered by the OOM kill
1937 * uncharges. Wake any sleepers explicitely.
1939 memcg_oom_recover(memcg
);
1942 current
->memcg_oom
.memcg
= NULL
;
1943 css_put(&memcg
->css
);
1948 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1949 * @page: page that is going to change accounted state
1951 * This function must mark the beginning of an accounted page state
1952 * change to prevent double accounting when the page is concurrently
1953 * being moved to another memcg:
1955 * memcg = mem_cgroup_begin_page_stat(page);
1956 * if (TestClearPageState(page))
1957 * mem_cgroup_update_page_stat(memcg, state, -1);
1958 * mem_cgroup_end_page_stat(memcg);
1960 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
1962 struct mem_cgroup
*memcg
;
1963 unsigned long flags
;
1966 * The RCU lock is held throughout the transaction. The fast
1967 * path can get away without acquiring the memcg->move_lock
1968 * because page moving starts with an RCU grace period.
1970 * The RCU lock also protects the memcg from being freed when
1971 * the page state that is going to change is the only thing
1972 * preventing the page from being uncharged.
1973 * E.g. end-writeback clearing PageWriteback(), which allows
1974 * migration to go ahead and uncharge the page before the
1975 * account transaction might be complete.
1979 if (mem_cgroup_disabled())
1982 memcg
= page
->mem_cgroup
;
1983 if (unlikely(!memcg
))
1986 if (atomic_read(&memcg
->moving_account
) <= 0)
1989 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1990 if (memcg
!= page
->mem_cgroup
) {
1991 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1996 * When charge migration first begins, we can have locked and
1997 * unlocked page stat updates happening concurrently. Track
1998 * the task who has the lock for mem_cgroup_end_page_stat().
2000 memcg
->move_lock_task
= current
;
2001 memcg
->move_lock_flags
= flags
;
2007 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2008 * @memcg: the memcg that was accounted against
2010 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
2012 if (memcg
&& memcg
->move_lock_task
== current
) {
2013 unsigned long flags
= memcg
->move_lock_flags
;
2015 memcg
->move_lock_task
= NULL
;
2016 memcg
->move_lock_flags
= 0;
2018 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2025 * mem_cgroup_update_page_stat - update page state statistics
2026 * @memcg: memcg to account against
2027 * @idx: page state item to account
2028 * @val: number of pages (positive or negative)
2030 * See mem_cgroup_begin_page_stat() for locking requirements.
2032 void mem_cgroup_update_page_stat(struct mem_cgroup
*memcg
,
2033 enum mem_cgroup_stat_index idx
, int val
)
2035 VM_BUG_ON(!rcu_read_lock_held());
2038 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2042 * size of first charge trial. "32" comes from vmscan.c's magic value.
2043 * TODO: maybe necessary to use big numbers in big irons.
2045 #define CHARGE_BATCH 32U
2046 struct memcg_stock_pcp
{
2047 struct mem_cgroup
*cached
; /* this never be root cgroup */
2048 unsigned int nr_pages
;
2049 struct work_struct work
;
2050 unsigned long flags
;
2051 #define FLUSHING_CACHED_CHARGE 0
2053 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2054 static DEFINE_MUTEX(percpu_charge_mutex
);
2057 * consume_stock: Try to consume stocked charge on this cpu.
2058 * @memcg: memcg to consume from.
2059 * @nr_pages: how many pages to charge.
2061 * The charges will only happen if @memcg matches the current cpu's memcg
2062 * stock, and at least @nr_pages are available in that stock. Failure to
2063 * service an allocation will refill the stock.
2065 * returns true if successful, false otherwise.
2067 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2069 struct memcg_stock_pcp
*stock
;
2072 if (nr_pages
> CHARGE_BATCH
)
2075 stock
= &get_cpu_var(memcg_stock
);
2076 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2077 stock
->nr_pages
-= nr_pages
;
2080 put_cpu_var(memcg_stock
);
2085 * Returns stocks cached in percpu and reset cached information.
2087 static void drain_stock(struct memcg_stock_pcp
*stock
)
2089 struct mem_cgroup
*old
= stock
->cached
;
2091 if (stock
->nr_pages
) {
2092 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2093 if (do_swap_account
)
2094 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2095 css_put_many(&old
->css
, stock
->nr_pages
);
2096 stock
->nr_pages
= 0;
2098 stock
->cached
= NULL
;
2102 * This must be called under preempt disabled or must be called by
2103 * a thread which is pinned to local cpu.
2105 static void drain_local_stock(struct work_struct
*dummy
)
2107 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2109 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2113 * Cache charges(val) to local per_cpu area.
2114 * This will be consumed by consume_stock() function, later.
2116 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2118 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2120 if (stock
->cached
!= memcg
) { /* reset if necessary */
2122 stock
->cached
= memcg
;
2124 stock
->nr_pages
+= nr_pages
;
2125 put_cpu_var(memcg_stock
);
2129 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2130 * of the hierarchy under it.
2132 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2136 /* If someone's already draining, avoid adding running more workers. */
2137 if (!mutex_trylock(&percpu_charge_mutex
))
2139 /* Notify other cpus that system-wide "drain" is running */
2142 for_each_online_cpu(cpu
) {
2143 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2144 struct mem_cgroup
*memcg
;
2146 memcg
= stock
->cached
;
2147 if (!memcg
|| !stock
->nr_pages
)
2149 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
2151 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2153 drain_local_stock(&stock
->work
);
2155 schedule_work_on(cpu
, &stock
->work
);
2160 mutex_unlock(&percpu_charge_mutex
);
2164 * This function drains percpu counter value from DEAD cpu and
2165 * move it to local cpu. Note that this function can be preempted.
2167 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2171 spin_lock(&memcg
->pcp_counter_lock
);
2172 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2173 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2175 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2176 memcg
->nocpu_base
.count
[i
] += x
;
2178 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2179 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2181 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2182 memcg
->nocpu_base
.events
[i
] += x
;
2184 spin_unlock(&memcg
->pcp_counter_lock
);
2187 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2188 unsigned long action
,
2191 int cpu
= (unsigned long)hcpu
;
2192 struct memcg_stock_pcp
*stock
;
2193 struct mem_cgroup
*iter
;
2195 if (action
== CPU_ONLINE
)
2198 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2201 for_each_mem_cgroup(iter
)
2202 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2204 stock
= &per_cpu(memcg_stock
, cpu
);
2209 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2210 unsigned int nr_pages
)
2212 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2213 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2214 struct mem_cgroup
*mem_over_limit
;
2215 struct page_counter
*counter
;
2216 unsigned long nr_reclaimed
;
2217 bool may_swap
= true;
2218 bool drained
= false;
2221 if (mem_cgroup_is_root(memcg
))
2224 if (consume_stock(memcg
, nr_pages
))
2227 if (!do_swap_account
||
2228 !page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2229 if (!page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2231 if (do_swap_account
)
2232 page_counter_uncharge(&memcg
->memsw
, batch
);
2233 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2235 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2239 if (batch
> nr_pages
) {
2245 * Unlike in global OOM situations, memcg is not in a physical
2246 * memory shortage. Allow dying and OOM-killed tasks to
2247 * bypass the last charges so that they can exit quickly and
2248 * free their memory.
2250 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2251 fatal_signal_pending(current
) ||
2252 current
->flags
& PF_EXITING
))
2255 if (unlikely(task_in_memcg_oom(current
)))
2258 if (!(gfp_mask
& __GFP_WAIT
))
2261 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2263 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2264 gfp_mask
, may_swap
);
2266 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2270 drain_all_stock(mem_over_limit
);
2275 if (gfp_mask
& __GFP_NORETRY
)
2278 * Even though the limit is exceeded at this point, reclaim
2279 * may have been able to free some pages. Retry the charge
2280 * before killing the task.
2282 * Only for regular pages, though: huge pages are rather
2283 * unlikely to succeed so close to the limit, and we fall back
2284 * to regular pages anyway in case of failure.
2286 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2289 * At task move, charge accounts can be doubly counted. So, it's
2290 * better to wait until the end of task_move if something is going on.
2292 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2298 if (gfp_mask
& __GFP_NOFAIL
)
2301 if (fatal_signal_pending(current
))
2304 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2306 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2308 if (!(gfp_mask
& __GFP_NOFAIL
))
2314 css_get_many(&memcg
->css
, batch
);
2315 if (batch
> nr_pages
)
2316 refill_stock(memcg
, batch
- nr_pages
);
2318 * If the hierarchy is above the normal consumption range,
2319 * make the charging task trim their excess contribution.
2322 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2324 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
2325 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2326 } while ((memcg
= parent_mem_cgroup(memcg
)));
2331 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2333 if (mem_cgroup_is_root(memcg
))
2336 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2337 if (do_swap_account
)
2338 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2340 css_put_many(&memcg
->css
, nr_pages
);
2344 * A helper function to get mem_cgroup from ID. must be called under
2345 * rcu_read_lock(). The caller is responsible for calling
2346 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2347 * refcnt from swap can be called against removed memcg.)
2349 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2351 /* ID 0 is unused ID */
2354 return mem_cgroup_from_id(id
);
2358 * try_get_mem_cgroup_from_page - look up page's memcg association
2361 * Look up, get a css reference, and return the memcg that owns @page.
2363 * The page must be locked to prevent racing with swap-in and page
2364 * cache charges. If coming from an unlocked page table, the caller
2365 * must ensure the page is on the LRU or this can race with charging.
2367 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2369 struct mem_cgroup
*memcg
;
2373 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2375 memcg
= page
->mem_cgroup
;
2377 if (!css_tryget_online(&memcg
->css
))
2379 } else if (PageSwapCache(page
)) {
2380 ent
.val
= page_private(page
);
2381 id
= lookup_swap_cgroup_id(ent
);
2383 memcg
= mem_cgroup_lookup(id
);
2384 if (memcg
&& !css_tryget_online(&memcg
->css
))
2391 static void lock_page_lru(struct page
*page
, int *isolated
)
2393 struct zone
*zone
= page_zone(page
);
2395 spin_lock_irq(&zone
->lru_lock
);
2396 if (PageLRU(page
)) {
2397 struct lruvec
*lruvec
;
2399 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2401 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2407 static void unlock_page_lru(struct page
*page
, int isolated
)
2409 struct zone
*zone
= page_zone(page
);
2412 struct lruvec
*lruvec
;
2414 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2415 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2417 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2419 spin_unlock_irq(&zone
->lru_lock
);
2422 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2427 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2430 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2431 * may already be on some other mem_cgroup's LRU. Take care of it.
2434 lock_page_lru(page
, &isolated
);
2437 * Nobody should be changing or seriously looking at
2438 * page->mem_cgroup at this point:
2440 * - the page is uncharged
2442 * - the page is off-LRU
2444 * - an anonymous fault has exclusive page access, except for
2445 * a locked page table
2447 * - a page cache insertion, a swapin fault, or a migration
2448 * have the page locked
2450 page
->mem_cgroup
= memcg
;
2453 unlock_page_lru(page
, isolated
);
2456 #ifdef CONFIG_MEMCG_KMEM
2457 int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
,
2458 unsigned long nr_pages
)
2460 struct page_counter
*counter
;
2463 ret
= page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
);
2467 ret
= try_charge(memcg
, gfp
, nr_pages
);
2468 if (ret
== -EINTR
) {
2470 * try_charge() chose to bypass to root due to OOM kill or
2471 * fatal signal. Since our only options are to either fail
2472 * the allocation or charge it to this cgroup, do it as a
2473 * temporary condition. But we can't fail. From a kmem/slab
2474 * perspective, the cache has already been selected, by
2475 * mem_cgroup_kmem_get_cache(), so it is too late to change
2478 * This condition will only trigger if the task entered
2479 * memcg_charge_kmem in a sane state, but was OOM-killed
2480 * during try_charge() above. Tasks that were already dying
2481 * when the allocation triggers should have been already
2482 * directed to the root cgroup in memcontrol.h
2484 page_counter_charge(&memcg
->memory
, nr_pages
);
2485 if (do_swap_account
)
2486 page_counter_charge(&memcg
->memsw
, nr_pages
);
2487 css_get_many(&memcg
->css
, nr_pages
);
2490 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2495 void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, unsigned long nr_pages
)
2497 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2498 if (do_swap_account
)
2499 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2501 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2503 css_put_many(&memcg
->css
, nr_pages
);
2507 * helper for acessing a memcg's index. It will be used as an index in the
2508 * child cache array in kmem_cache, and also to derive its name. This function
2509 * will return -1 when this is not a kmem-limited memcg.
2511 int memcg_cache_id(struct mem_cgroup
*memcg
)
2513 return memcg
? memcg
->kmemcg_id
: -1;
2516 static int memcg_alloc_cache_id(void)
2521 id
= ida_simple_get(&memcg_cache_ida
,
2522 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2526 if (id
< memcg_nr_cache_ids
)
2530 * There's no space for the new id in memcg_caches arrays,
2531 * so we have to grow them.
2533 down_write(&memcg_cache_ids_sem
);
2535 size
= 2 * (id
+ 1);
2536 if (size
< MEMCG_CACHES_MIN_SIZE
)
2537 size
= MEMCG_CACHES_MIN_SIZE
;
2538 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2539 size
= MEMCG_CACHES_MAX_SIZE
;
2541 err
= memcg_update_all_caches(size
);
2543 err
= memcg_update_all_list_lrus(size
);
2545 memcg_nr_cache_ids
= size
;
2547 up_write(&memcg_cache_ids_sem
);
2550 ida_simple_remove(&memcg_cache_ida
, id
);
2556 static void memcg_free_cache_id(int id
)
2558 ida_simple_remove(&memcg_cache_ida
, id
);
2561 struct memcg_kmem_cache_create_work
{
2562 struct mem_cgroup
*memcg
;
2563 struct kmem_cache
*cachep
;
2564 struct work_struct work
;
2567 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2569 struct memcg_kmem_cache_create_work
*cw
=
2570 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2571 struct mem_cgroup
*memcg
= cw
->memcg
;
2572 struct kmem_cache
*cachep
= cw
->cachep
;
2574 memcg_create_kmem_cache(memcg
, cachep
);
2576 css_put(&memcg
->css
);
2581 * Enqueue the creation of a per-memcg kmem_cache.
2583 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2584 struct kmem_cache
*cachep
)
2586 struct memcg_kmem_cache_create_work
*cw
;
2588 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2592 css_get(&memcg
->css
);
2595 cw
->cachep
= cachep
;
2596 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2598 schedule_work(&cw
->work
);
2601 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2602 struct kmem_cache
*cachep
)
2605 * We need to stop accounting when we kmalloc, because if the
2606 * corresponding kmalloc cache is not yet created, the first allocation
2607 * in __memcg_schedule_kmem_cache_create will recurse.
2609 * However, it is better to enclose the whole function. Depending on
2610 * the debugging options enabled, INIT_WORK(), for instance, can
2611 * trigger an allocation. This too, will make us recurse. Because at
2612 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2613 * the safest choice is to do it like this, wrapping the whole function.
2615 current
->memcg_kmem_skip_account
= 1;
2616 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2617 current
->memcg_kmem_skip_account
= 0;
2621 * Return the kmem_cache we're supposed to use for a slab allocation.
2622 * We try to use the current memcg's version of the cache.
2624 * If the cache does not exist yet, if we are the first user of it,
2625 * we either create it immediately, if possible, or create it asynchronously
2627 * In the latter case, we will let the current allocation go through with
2628 * the original cache.
2630 * Can't be called in interrupt context or from kernel threads.
2631 * This function needs to be called with rcu_read_lock() held.
2633 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2635 struct mem_cgroup
*memcg
;
2636 struct kmem_cache
*memcg_cachep
;
2639 VM_BUG_ON(!is_root_cache(cachep
));
2641 if (current
->memcg_kmem_skip_account
)
2644 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2645 kmemcg_id
= ACCESS_ONCE(memcg
->kmemcg_id
);
2649 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2650 if (likely(memcg_cachep
))
2651 return memcg_cachep
;
2654 * If we are in a safe context (can wait, and not in interrupt
2655 * context), we could be be predictable and return right away.
2656 * This would guarantee that the allocation being performed
2657 * already belongs in the new cache.
2659 * However, there are some clashes that can arrive from locking.
2660 * For instance, because we acquire the slab_mutex while doing
2661 * memcg_create_kmem_cache, this means no further allocation
2662 * could happen with the slab_mutex held. So it's better to
2665 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2667 css_put(&memcg
->css
);
2671 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2673 if (!is_root_cache(cachep
))
2674 css_put(&cachep
->memcg_params
.memcg
->css
);
2678 * We need to verify if the allocation against current->mm->owner's memcg is
2679 * possible for the given order. But the page is not allocated yet, so we'll
2680 * need a further commit step to do the final arrangements.
2682 * It is possible for the task to switch cgroups in this mean time, so at
2683 * commit time, we can't rely on task conversion any longer. We'll then use
2684 * the handle argument to return to the caller which cgroup we should commit
2685 * against. We could also return the memcg directly and avoid the pointer
2686 * passing, but a boolean return value gives better semantics considering
2687 * the compiled-out case as well.
2689 * Returning true means the allocation is possible.
2692 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
2694 struct mem_cgroup
*memcg
;
2699 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2701 if (!memcg_kmem_is_active(memcg
)) {
2702 css_put(&memcg
->css
);
2706 ret
= memcg_charge_kmem(memcg
, gfp
, 1 << order
);
2710 css_put(&memcg
->css
);
2714 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2717 VM_BUG_ON(mem_cgroup_is_root(memcg
));
2719 /* The page allocation failed. Revert */
2721 memcg_uncharge_kmem(memcg
, 1 << order
);
2724 page
->mem_cgroup
= memcg
;
2727 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
2729 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2734 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2736 memcg_uncharge_kmem(memcg
, 1 << order
);
2737 page
->mem_cgroup
= NULL
;
2740 struct mem_cgroup
*__mem_cgroup_from_kmem(void *ptr
)
2742 struct mem_cgroup
*memcg
= NULL
;
2743 struct kmem_cache
*cachep
;
2746 page
= virt_to_head_page(ptr
);
2747 if (PageSlab(page
)) {
2748 cachep
= page
->slab_cache
;
2749 if (!is_root_cache(cachep
))
2750 memcg
= cachep
->memcg_params
.memcg
;
2752 /* page allocated by alloc_kmem_pages */
2753 memcg
= page
->mem_cgroup
;
2757 #endif /* CONFIG_MEMCG_KMEM */
2759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2762 * Because tail pages are not marked as "used", set it. We're under
2763 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2764 * charge/uncharge will be never happen and move_account() is done under
2765 * compound_lock(), so we don't have to take care of races.
2767 void mem_cgroup_split_huge_fixup(struct page
*head
)
2771 if (mem_cgroup_disabled())
2774 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2775 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2777 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2780 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2783 * mem_cgroup_move_account - move account of the page
2785 * @nr_pages: number of regular pages (>1 for huge pages)
2786 * @from: mem_cgroup which the page is moved from.
2787 * @to: mem_cgroup which the page is moved to. @from != @to.
2789 * The caller must confirm following.
2790 * - page is not on LRU (isolate_page() is useful.)
2791 * - compound_lock is held when nr_pages > 1
2793 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2796 static int mem_cgroup_move_account(struct page
*page
,
2797 unsigned int nr_pages
,
2798 struct mem_cgroup
*from
,
2799 struct mem_cgroup
*to
)
2801 unsigned long flags
;
2804 VM_BUG_ON(from
== to
);
2805 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2807 * The page is isolated from LRU. So, collapse function
2808 * will not handle this page. But page splitting can happen.
2809 * Do this check under compound_page_lock(). The caller should
2813 if (nr_pages
> 1 && !PageTransHuge(page
))
2817 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2818 * of its source page while we change it: page migration takes
2819 * both pages off the LRU, but page cache replacement doesn't.
2821 if (!trylock_page(page
))
2825 if (page
->mem_cgroup
!= from
)
2828 spin_lock_irqsave(&from
->move_lock
, flags
);
2830 if (!PageAnon(page
) && page_mapped(page
)) {
2831 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
2833 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
2837 if (PageWriteback(page
)) {
2838 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
2840 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
2845 * It is safe to change page->mem_cgroup here because the page
2846 * is referenced, charged, and isolated - we can't race with
2847 * uncharging, charging, migration, or LRU putback.
2850 /* caller should have done css_get */
2851 page
->mem_cgroup
= to
;
2852 spin_unlock_irqrestore(&from
->move_lock
, flags
);
2856 local_irq_disable();
2857 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
2858 memcg_check_events(to
, page
);
2859 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
2860 memcg_check_events(from
, page
);
2868 #ifdef CONFIG_MEMCG_SWAP
2869 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2872 int val
= (charge
) ? 1 : -1;
2873 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2877 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2878 * @entry: swap entry to be moved
2879 * @from: mem_cgroup which the entry is moved from
2880 * @to: mem_cgroup which the entry is moved to
2882 * It succeeds only when the swap_cgroup's record for this entry is the same
2883 * as the mem_cgroup's id of @from.
2885 * Returns 0 on success, -EINVAL on failure.
2887 * The caller must have charged to @to, IOW, called page_counter_charge() about
2888 * both res and memsw, and called css_get().
2890 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2891 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2893 unsigned short old_id
, new_id
;
2895 old_id
= mem_cgroup_id(from
);
2896 new_id
= mem_cgroup_id(to
);
2898 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2899 mem_cgroup_swap_statistics(from
, false);
2900 mem_cgroup_swap_statistics(to
, true);
2906 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2907 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2913 static DEFINE_MUTEX(memcg_limit_mutex
);
2915 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2916 unsigned long limit
)
2918 unsigned long curusage
;
2919 unsigned long oldusage
;
2920 bool enlarge
= false;
2925 * For keeping hierarchical_reclaim simple, how long we should retry
2926 * is depends on callers. We set our retry-count to be function
2927 * of # of children which we should visit in this loop.
2929 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2930 mem_cgroup_count_children(memcg
);
2932 oldusage
= page_counter_read(&memcg
->memory
);
2935 if (signal_pending(current
)) {
2940 mutex_lock(&memcg_limit_mutex
);
2941 if (limit
> memcg
->memsw
.limit
) {
2942 mutex_unlock(&memcg_limit_mutex
);
2946 if (limit
> memcg
->memory
.limit
)
2948 ret
= page_counter_limit(&memcg
->memory
, limit
);
2949 mutex_unlock(&memcg_limit_mutex
);
2954 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2956 curusage
= page_counter_read(&memcg
->memory
);
2957 /* Usage is reduced ? */
2958 if (curusage
>= oldusage
)
2961 oldusage
= curusage
;
2962 } while (retry_count
);
2964 if (!ret
&& enlarge
)
2965 memcg_oom_recover(memcg
);
2970 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2971 unsigned long limit
)
2973 unsigned long curusage
;
2974 unsigned long oldusage
;
2975 bool enlarge
= false;
2979 /* see mem_cgroup_resize_res_limit */
2980 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2981 mem_cgroup_count_children(memcg
);
2983 oldusage
= page_counter_read(&memcg
->memsw
);
2986 if (signal_pending(current
)) {
2991 mutex_lock(&memcg_limit_mutex
);
2992 if (limit
< memcg
->memory
.limit
) {
2993 mutex_unlock(&memcg_limit_mutex
);
2997 if (limit
> memcg
->memsw
.limit
)
2999 ret
= page_counter_limit(&memcg
->memsw
, limit
);
3000 mutex_unlock(&memcg_limit_mutex
);
3005 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
3007 curusage
= page_counter_read(&memcg
->memsw
);
3008 /* Usage is reduced ? */
3009 if (curusage
>= oldusage
)
3012 oldusage
= curusage
;
3013 } while (retry_count
);
3015 if (!ret
&& enlarge
)
3016 memcg_oom_recover(memcg
);
3021 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3023 unsigned long *total_scanned
)
3025 unsigned long nr_reclaimed
= 0;
3026 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3027 unsigned long reclaimed
;
3029 struct mem_cgroup_tree_per_zone
*mctz
;
3030 unsigned long excess
;
3031 unsigned long nr_scanned
;
3036 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3038 * This loop can run a while, specially if mem_cgroup's continuously
3039 * keep exceeding their soft limit and putting the system under
3046 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3051 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
3052 gfp_mask
, &nr_scanned
);
3053 nr_reclaimed
+= reclaimed
;
3054 *total_scanned
+= nr_scanned
;
3055 spin_lock_irq(&mctz
->lock
);
3056 __mem_cgroup_remove_exceeded(mz
, mctz
);
3059 * If we failed to reclaim anything from this memory cgroup
3060 * it is time to move on to the next cgroup
3064 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
3066 excess
= soft_limit_excess(mz
->memcg
);
3068 * One school of thought says that we should not add
3069 * back the node to the tree if reclaim returns 0.
3070 * But our reclaim could return 0, simply because due
3071 * to priority we are exposing a smaller subset of
3072 * memory to reclaim from. Consider this as a longer
3075 /* If excess == 0, no tree ops */
3076 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3077 spin_unlock_irq(&mctz
->lock
);
3078 css_put(&mz
->memcg
->css
);
3081 * Could not reclaim anything and there are no more
3082 * mem cgroups to try or we seem to be looping without
3083 * reclaiming anything.
3085 if (!nr_reclaimed
&&
3087 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3089 } while (!nr_reclaimed
);
3091 css_put(&next_mz
->memcg
->css
);
3092 return nr_reclaimed
;
3096 * Test whether @memcg has children, dead or alive. Note that this
3097 * function doesn't care whether @memcg has use_hierarchy enabled and
3098 * returns %true if there are child csses according to the cgroup
3099 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3101 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3106 * The lock does not prevent addition or deletion of children, but
3107 * it prevents a new child from being initialized based on this
3108 * parent in css_online(), so it's enough to decide whether
3109 * hierarchically inherited attributes can still be changed or not.
3111 lockdep_assert_held(&memcg_create_mutex
);
3114 ret
= css_next_child(NULL
, &memcg
->css
);
3120 * Reclaims as many pages from the given memcg as possible and moves
3121 * the rest to the parent.
3123 * Caller is responsible for holding css reference for memcg.
3125 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3127 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3129 /* we call try-to-free pages for make this cgroup empty */
3130 lru_add_drain_all();
3131 /* try to free all pages in this cgroup */
3132 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3135 if (signal_pending(current
))
3138 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3142 /* maybe some writeback is necessary */
3143 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3151 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3152 char *buf
, size_t nbytes
,
3155 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3157 if (mem_cgroup_is_root(memcg
))
3159 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3162 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3165 return mem_cgroup_from_css(css
)->use_hierarchy
;
3168 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3169 struct cftype
*cft
, u64 val
)
3172 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3173 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3175 mutex_lock(&memcg_create_mutex
);
3177 if (memcg
->use_hierarchy
== val
)
3181 * If parent's use_hierarchy is set, we can't make any modifications
3182 * in the child subtrees. If it is unset, then the change can
3183 * occur, provided the current cgroup has no children.
3185 * For the root cgroup, parent_mem is NULL, we allow value to be
3186 * set if there are no children.
3188 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3189 (val
== 1 || val
== 0)) {
3190 if (!memcg_has_children(memcg
))
3191 memcg
->use_hierarchy
= val
;
3198 mutex_unlock(&memcg_create_mutex
);
3203 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
3204 enum mem_cgroup_stat_index idx
)
3206 struct mem_cgroup
*iter
;
3209 /* Per-cpu values can be negative, use a signed accumulator */
3210 for_each_mem_cgroup_tree(iter
, memcg
)
3211 val
+= mem_cgroup_read_stat(iter
, idx
);
3213 if (val
< 0) /* race ? */
3218 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3222 if (mem_cgroup_is_root(memcg
)) {
3223 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
3224 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
3226 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
3229 val
= page_counter_read(&memcg
->memory
);
3231 val
= page_counter_read(&memcg
->memsw
);
3233 return val
<< PAGE_SHIFT
;
3244 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3247 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3248 struct page_counter
*counter
;
3250 switch (MEMFILE_TYPE(cft
->private)) {
3252 counter
= &memcg
->memory
;
3255 counter
= &memcg
->memsw
;
3258 counter
= &memcg
->kmem
;
3264 switch (MEMFILE_ATTR(cft
->private)) {
3266 if (counter
== &memcg
->memory
)
3267 return mem_cgroup_usage(memcg
, false);
3268 if (counter
== &memcg
->memsw
)
3269 return mem_cgroup_usage(memcg
, true);
3270 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3272 return (u64
)counter
->limit
* PAGE_SIZE
;
3274 return (u64
)counter
->watermark
* PAGE_SIZE
;
3276 return counter
->failcnt
;
3277 case RES_SOFT_LIMIT
:
3278 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3284 #ifdef CONFIG_MEMCG_KMEM
3285 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
3286 unsigned long nr_pages
)
3291 BUG_ON(memcg
->kmemcg_id
>= 0);
3292 BUG_ON(memcg
->kmem_acct_activated
);
3293 BUG_ON(memcg
->kmem_acct_active
);
3296 * For simplicity, we won't allow this to be disabled. It also can't
3297 * be changed if the cgroup has children already, or if tasks had
3300 * If tasks join before we set the limit, a person looking at
3301 * kmem.usage_in_bytes will have no way to determine when it took
3302 * place, which makes the value quite meaningless.
3304 * After it first became limited, changes in the value of the limit are
3305 * of course permitted.
3307 mutex_lock(&memcg_create_mutex
);
3308 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
3309 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
3311 mutex_unlock(&memcg_create_mutex
);
3315 memcg_id
= memcg_alloc_cache_id();
3322 * We couldn't have accounted to this cgroup, because it hasn't got
3323 * activated yet, so this should succeed.
3325 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
3328 static_key_slow_inc(&memcg_kmem_enabled_key
);
3330 * A memory cgroup is considered kmem-active as soon as it gets
3331 * kmemcg_id. Setting the id after enabling static branching will
3332 * guarantee no one starts accounting before all call sites are
3335 memcg
->kmemcg_id
= memcg_id
;
3336 memcg
->kmem_acct_activated
= true;
3337 memcg
->kmem_acct_active
= true;
3342 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3343 unsigned long limit
)
3347 mutex_lock(&memcg_limit_mutex
);
3348 if (!memcg_kmem_is_active(memcg
))
3349 ret
= memcg_activate_kmem(memcg
, limit
);
3351 ret
= page_counter_limit(&memcg
->kmem
, limit
);
3352 mutex_unlock(&memcg_limit_mutex
);
3356 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
3359 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
3364 mutex_lock(&memcg_limit_mutex
);
3366 * If the parent cgroup is not kmem-active now, it cannot be activated
3367 * after this point, because it has at least one child already.
3369 if (memcg_kmem_is_active(parent
))
3370 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
3371 mutex_unlock(&memcg_limit_mutex
);
3375 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3376 unsigned long limit
)
3380 #endif /* CONFIG_MEMCG_KMEM */
3383 * The user of this function is...
3386 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3387 char *buf
, size_t nbytes
, loff_t off
)
3389 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3390 unsigned long nr_pages
;
3393 buf
= strstrip(buf
);
3394 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3398 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3400 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3404 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3406 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3409 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3412 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3416 case RES_SOFT_LIMIT
:
3417 memcg
->soft_limit
= nr_pages
;
3421 return ret
?: nbytes
;
3424 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3425 size_t nbytes
, loff_t off
)
3427 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3428 struct page_counter
*counter
;
3430 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3432 counter
= &memcg
->memory
;
3435 counter
= &memcg
->memsw
;
3438 counter
= &memcg
->kmem
;
3444 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3446 page_counter_reset_watermark(counter
);
3449 counter
->failcnt
= 0;
3458 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3461 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3465 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3466 struct cftype
*cft
, u64 val
)
3468 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3470 if (val
& ~MOVE_MASK
)
3474 * No kind of locking is needed in here, because ->can_attach() will
3475 * check this value once in the beginning of the process, and then carry
3476 * on with stale data. This means that changes to this value will only
3477 * affect task migrations starting after the change.
3479 memcg
->move_charge_at_immigrate
= val
;
3483 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3484 struct cftype
*cft
, u64 val
)
3491 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3495 unsigned int lru_mask
;
3498 static const struct numa_stat stats
[] = {
3499 { "total", LRU_ALL
},
3500 { "file", LRU_ALL_FILE
},
3501 { "anon", LRU_ALL_ANON
},
3502 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3504 const struct numa_stat
*stat
;
3507 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3509 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3510 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3511 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3512 for_each_node_state(nid
, N_MEMORY
) {
3513 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3515 seq_printf(m
, " N%d=%lu", nid
, nr
);
3520 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3521 struct mem_cgroup
*iter
;
3524 for_each_mem_cgroup_tree(iter
, memcg
)
3525 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3526 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3527 for_each_node_state(nid
, N_MEMORY
) {
3529 for_each_mem_cgroup_tree(iter
, memcg
)
3530 nr
+= mem_cgroup_node_nr_lru_pages(
3531 iter
, nid
, stat
->lru_mask
);
3532 seq_printf(m
, " N%d=%lu", nid
, nr
);
3539 #endif /* CONFIG_NUMA */
3541 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3543 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3544 unsigned long memory
, memsw
;
3545 struct mem_cgroup
*mi
;
3548 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3549 MEM_CGROUP_STAT_NSTATS
);
3550 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3551 MEM_CGROUP_EVENTS_NSTATS
);
3552 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3554 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3555 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3557 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
3558 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3561 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3562 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3563 mem_cgroup_read_events(memcg
, i
));
3565 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3566 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3567 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3569 /* Hierarchical information */
3570 memory
= memsw
= PAGE_COUNTER_MAX
;
3571 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3572 memory
= min(memory
, mi
->memory
.limit
);
3573 memsw
= min(memsw
, mi
->memsw
.limit
);
3575 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3576 (u64
)memory
* PAGE_SIZE
);
3577 if (do_swap_account
)
3578 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3579 (u64
)memsw
* PAGE_SIZE
);
3581 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3584 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3586 for_each_mem_cgroup_tree(mi
, memcg
)
3587 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3588 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
3591 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3592 unsigned long long val
= 0;
3594 for_each_mem_cgroup_tree(mi
, memcg
)
3595 val
+= mem_cgroup_read_events(mi
, i
);
3596 seq_printf(m
, "total_%s %llu\n",
3597 mem_cgroup_events_names
[i
], val
);
3600 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3601 unsigned long long val
= 0;
3603 for_each_mem_cgroup_tree(mi
, memcg
)
3604 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3605 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3608 #ifdef CONFIG_DEBUG_VM
3611 struct mem_cgroup_per_zone
*mz
;
3612 struct zone_reclaim_stat
*rstat
;
3613 unsigned long recent_rotated
[2] = {0, 0};
3614 unsigned long recent_scanned
[2] = {0, 0};
3616 for_each_online_node(nid
)
3617 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3618 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3619 rstat
= &mz
->lruvec
.reclaim_stat
;
3621 recent_rotated
[0] += rstat
->recent_rotated
[0];
3622 recent_rotated
[1] += rstat
->recent_rotated
[1];
3623 recent_scanned
[0] += rstat
->recent_scanned
[0];
3624 recent_scanned
[1] += rstat
->recent_scanned
[1];
3626 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3627 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3628 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3629 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3636 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3639 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3641 return mem_cgroup_swappiness(memcg
);
3644 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3645 struct cftype
*cft
, u64 val
)
3647 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3653 memcg
->swappiness
= val
;
3655 vm_swappiness
= val
;
3660 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3662 struct mem_cgroup_threshold_ary
*t
;
3663 unsigned long usage
;
3668 t
= rcu_dereference(memcg
->thresholds
.primary
);
3670 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3675 usage
= mem_cgroup_usage(memcg
, swap
);
3678 * current_threshold points to threshold just below or equal to usage.
3679 * If it's not true, a threshold was crossed after last
3680 * call of __mem_cgroup_threshold().
3682 i
= t
->current_threshold
;
3685 * Iterate backward over array of thresholds starting from
3686 * current_threshold and check if a threshold is crossed.
3687 * If none of thresholds below usage is crossed, we read
3688 * only one element of the array here.
3690 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3691 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3693 /* i = current_threshold + 1 */
3697 * Iterate forward over array of thresholds starting from
3698 * current_threshold+1 and check if a threshold is crossed.
3699 * If none of thresholds above usage is crossed, we read
3700 * only one element of the array here.
3702 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3703 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3705 /* Update current_threshold */
3706 t
->current_threshold
= i
- 1;
3711 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3714 __mem_cgroup_threshold(memcg
, false);
3715 if (do_swap_account
)
3716 __mem_cgroup_threshold(memcg
, true);
3718 memcg
= parent_mem_cgroup(memcg
);
3722 static int compare_thresholds(const void *a
, const void *b
)
3724 const struct mem_cgroup_threshold
*_a
= a
;
3725 const struct mem_cgroup_threshold
*_b
= b
;
3727 if (_a
->threshold
> _b
->threshold
)
3730 if (_a
->threshold
< _b
->threshold
)
3736 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3738 struct mem_cgroup_eventfd_list
*ev
;
3740 spin_lock(&memcg_oom_lock
);
3742 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3743 eventfd_signal(ev
->eventfd
, 1);
3745 spin_unlock(&memcg_oom_lock
);
3749 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3751 struct mem_cgroup
*iter
;
3753 for_each_mem_cgroup_tree(iter
, memcg
)
3754 mem_cgroup_oom_notify_cb(iter
);
3757 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3758 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3760 struct mem_cgroup_thresholds
*thresholds
;
3761 struct mem_cgroup_threshold_ary
*new;
3762 unsigned long threshold
;
3763 unsigned long usage
;
3766 ret
= page_counter_memparse(args
, "-1", &threshold
);
3770 mutex_lock(&memcg
->thresholds_lock
);
3773 thresholds
= &memcg
->thresholds
;
3774 usage
= mem_cgroup_usage(memcg
, false);
3775 } else if (type
== _MEMSWAP
) {
3776 thresholds
= &memcg
->memsw_thresholds
;
3777 usage
= mem_cgroup_usage(memcg
, true);
3781 /* Check if a threshold crossed before adding a new one */
3782 if (thresholds
->primary
)
3783 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3785 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3787 /* Allocate memory for new array of thresholds */
3788 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3796 /* Copy thresholds (if any) to new array */
3797 if (thresholds
->primary
) {
3798 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3799 sizeof(struct mem_cgroup_threshold
));
3802 /* Add new threshold */
3803 new->entries
[size
- 1].eventfd
= eventfd
;
3804 new->entries
[size
- 1].threshold
= threshold
;
3806 /* Sort thresholds. Registering of new threshold isn't time-critical */
3807 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3808 compare_thresholds
, NULL
);
3810 /* Find current threshold */
3811 new->current_threshold
= -1;
3812 for (i
= 0; i
< size
; i
++) {
3813 if (new->entries
[i
].threshold
<= usage
) {
3815 * new->current_threshold will not be used until
3816 * rcu_assign_pointer(), so it's safe to increment
3819 ++new->current_threshold
;
3824 /* Free old spare buffer and save old primary buffer as spare */
3825 kfree(thresholds
->spare
);
3826 thresholds
->spare
= thresholds
->primary
;
3828 rcu_assign_pointer(thresholds
->primary
, new);
3830 /* To be sure that nobody uses thresholds */
3834 mutex_unlock(&memcg
->thresholds_lock
);
3839 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3840 struct eventfd_ctx
*eventfd
, const char *args
)
3842 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3845 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3846 struct eventfd_ctx
*eventfd
, const char *args
)
3848 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3851 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3852 struct eventfd_ctx
*eventfd
, enum res_type type
)
3854 struct mem_cgroup_thresholds
*thresholds
;
3855 struct mem_cgroup_threshold_ary
*new;
3856 unsigned long usage
;
3859 mutex_lock(&memcg
->thresholds_lock
);
3862 thresholds
= &memcg
->thresholds
;
3863 usage
= mem_cgroup_usage(memcg
, false);
3864 } else if (type
== _MEMSWAP
) {
3865 thresholds
= &memcg
->memsw_thresholds
;
3866 usage
= mem_cgroup_usage(memcg
, true);
3870 if (!thresholds
->primary
)
3873 /* Check if a threshold crossed before removing */
3874 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3876 /* Calculate new number of threshold */
3878 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3879 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3883 new = thresholds
->spare
;
3885 /* Set thresholds array to NULL if we don't have thresholds */
3894 /* Copy thresholds and find current threshold */
3895 new->current_threshold
= -1;
3896 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3897 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3900 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3901 if (new->entries
[j
].threshold
<= usage
) {
3903 * new->current_threshold will not be used
3904 * until rcu_assign_pointer(), so it's safe to increment
3907 ++new->current_threshold
;
3913 /* Swap primary and spare array */
3914 thresholds
->spare
= thresholds
->primary
;
3915 /* If all events are unregistered, free the spare array */
3917 kfree(thresholds
->spare
);
3918 thresholds
->spare
= NULL
;
3921 rcu_assign_pointer(thresholds
->primary
, new);
3923 /* To be sure that nobody uses thresholds */
3926 mutex_unlock(&memcg
->thresholds_lock
);
3929 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3930 struct eventfd_ctx
*eventfd
)
3932 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3935 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3936 struct eventfd_ctx
*eventfd
)
3938 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3941 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3942 struct eventfd_ctx
*eventfd
, const char *args
)
3944 struct mem_cgroup_eventfd_list
*event
;
3946 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3950 spin_lock(&memcg_oom_lock
);
3952 event
->eventfd
= eventfd
;
3953 list_add(&event
->list
, &memcg
->oom_notify
);
3955 /* already in OOM ? */
3956 if (atomic_read(&memcg
->under_oom
))
3957 eventfd_signal(eventfd
, 1);
3958 spin_unlock(&memcg_oom_lock
);
3963 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3964 struct eventfd_ctx
*eventfd
)
3966 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3968 spin_lock(&memcg_oom_lock
);
3970 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3971 if (ev
->eventfd
== eventfd
) {
3972 list_del(&ev
->list
);
3977 spin_unlock(&memcg_oom_lock
);
3980 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3982 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3984 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3985 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
3989 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3990 struct cftype
*cft
, u64 val
)
3992 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3994 /* cannot set to root cgroup and only 0 and 1 are allowed */
3995 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3998 memcg
->oom_kill_disable
= val
;
4000 memcg_oom_recover(memcg
);
4005 #ifdef CONFIG_MEMCG_KMEM
4006 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4010 ret
= memcg_propagate_kmem(memcg
);
4014 return mem_cgroup_sockets_init(memcg
, ss
);
4017 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
4019 struct cgroup_subsys_state
*css
;
4020 struct mem_cgroup
*parent
, *child
;
4023 if (!memcg
->kmem_acct_active
)
4027 * Clear the 'active' flag before clearing memcg_caches arrays entries.
4028 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
4029 * guarantees no cache will be created for this cgroup after we are
4030 * done (see memcg_create_kmem_cache()).
4032 memcg
->kmem_acct_active
= false;
4034 memcg_deactivate_kmem_caches(memcg
);
4036 kmemcg_id
= memcg
->kmemcg_id
;
4037 BUG_ON(kmemcg_id
< 0);
4039 parent
= parent_mem_cgroup(memcg
);
4041 parent
= root_mem_cgroup
;
4044 * Change kmemcg_id of this cgroup and all its descendants to the
4045 * parent's id, and then move all entries from this cgroup's list_lrus
4046 * to ones of the parent. After we have finished, all list_lrus
4047 * corresponding to this cgroup are guaranteed to remain empty. The
4048 * ordering is imposed by list_lru_node->lock taken by
4049 * memcg_drain_all_list_lrus().
4051 css_for_each_descendant_pre(css
, &memcg
->css
) {
4052 child
= mem_cgroup_from_css(css
);
4053 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
4054 child
->kmemcg_id
= parent
->kmemcg_id
;
4055 if (!memcg
->use_hierarchy
)
4058 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
4060 memcg_free_cache_id(kmemcg_id
);
4063 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4065 if (memcg
->kmem_acct_activated
) {
4066 memcg_destroy_kmem_caches(memcg
);
4067 static_key_slow_dec(&memcg_kmem_enabled_key
);
4068 WARN_ON(page_counter_read(&memcg
->kmem
));
4070 mem_cgroup_sockets_destroy(memcg
);
4073 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4078 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
4082 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4088 * DO NOT USE IN NEW FILES.
4090 * "cgroup.event_control" implementation.
4092 * This is way over-engineered. It tries to support fully configurable
4093 * events for each user. Such level of flexibility is completely
4094 * unnecessary especially in the light of the planned unified hierarchy.
4096 * Please deprecate this and replace with something simpler if at all
4101 * Unregister event and free resources.
4103 * Gets called from workqueue.
4105 static void memcg_event_remove(struct work_struct
*work
)
4107 struct mem_cgroup_event
*event
=
4108 container_of(work
, struct mem_cgroup_event
, remove
);
4109 struct mem_cgroup
*memcg
= event
->memcg
;
4111 remove_wait_queue(event
->wqh
, &event
->wait
);
4113 event
->unregister_event(memcg
, event
->eventfd
);
4115 /* Notify userspace the event is going away. */
4116 eventfd_signal(event
->eventfd
, 1);
4118 eventfd_ctx_put(event
->eventfd
);
4120 css_put(&memcg
->css
);
4124 * Gets called on POLLHUP on eventfd when user closes it.
4126 * Called with wqh->lock held and interrupts disabled.
4128 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
4129 int sync
, void *key
)
4131 struct mem_cgroup_event
*event
=
4132 container_of(wait
, struct mem_cgroup_event
, wait
);
4133 struct mem_cgroup
*memcg
= event
->memcg
;
4134 unsigned long flags
= (unsigned long)key
;
4136 if (flags
& POLLHUP
) {
4138 * If the event has been detached at cgroup removal, we
4139 * can simply return knowing the other side will cleanup
4142 * We can't race against event freeing since the other
4143 * side will require wqh->lock via remove_wait_queue(),
4146 spin_lock(&memcg
->event_list_lock
);
4147 if (!list_empty(&event
->list
)) {
4148 list_del_init(&event
->list
);
4150 * We are in atomic context, but cgroup_event_remove()
4151 * may sleep, so we have to call it in workqueue.
4153 schedule_work(&event
->remove
);
4155 spin_unlock(&memcg
->event_list_lock
);
4161 static void memcg_event_ptable_queue_proc(struct file
*file
,
4162 wait_queue_head_t
*wqh
, poll_table
*pt
)
4164 struct mem_cgroup_event
*event
=
4165 container_of(pt
, struct mem_cgroup_event
, pt
);
4168 add_wait_queue(wqh
, &event
->wait
);
4172 * DO NOT USE IN NEW FILES.
4174 * Parse input and register new cgroup event handler.
4176 * Input must be in format '<event_fd> <control_fd> <args>'.
4177 * Interpretation of args is defined by control file implementation.
4179 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4180 char *buf
, size_t nbytes
, loff_t off
)
4182 struct cgroup_subsys_state
*css
= of_css(of
);
4183 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4184 struct mem_cgroup_event
*event
;
4185 struct cgroup_subsys_state
*cfile_css
;
4186 unsigned int efd
, cfd
;
4193 buf
= strstrip(buf
);
4195 efd
= simple_strtoul(buf
, &endp
, 10);
4200 cfd
= simple_strtoul(buf
, &endp
, 10);
4201 if ((*endp
!= ' ') && (*endp
!= '\0'))
4205 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4209 event
->memcg
= memcg
;
4210 INIT_LIST_HEAD(&event
->list
);
4211 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4212 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4213 INIT_WORK(&event
->remove
, memcg_event_remove
);
4221 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4222 if (IS_ERR(event
->eventfd
)) {
4223 ret
= PTR_ERR(event
->eventfd
);
4230 goto out_put_eventfd
;
4233 /* the process need read permission on control file */
4234 /* AV: shouldn't we check that it's been opened for read instead? */
4235 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4240 * Determine the event callbacks and set them in @event. This used
4241 * to be done via struct cftype but cgroup core no longer knows
4242 * about these events. The following is crude but the whole thing
4243 * is for compatibility anyway.
4245 * DO NOT ADD NEW FILES.
4247 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4249 if (!strcmp(name
, "memory.usage_in_bytes")) {
4250 event
->register_event
= mem_cgroup_usage_register_event
;
4251 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4252 } else if (!strcmp(name
, "memory.oom_control")) {
4253 event
->register_event
= mem_cgroup_oom_register_event
;
4254 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4255 } else if (!strcmp(name
, "memory.pressure_level")) {
4256 event
->register_event
= vmpressure_register_event
;
4257 event
->unregister_event
= vmpressure_unregister_event
;
4258 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4259 event
->register_event
= memsw_cgroup_usage_register_event
;
4260 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4267 * Verify @cfile should belong to @css. Also, remaining events are
4268 * automatically removed on cgroup destruction but the removal is
4269 * asynchronous, so take an extra ref on @css.
4271 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4272 &memory_cgrp_subsys
);
4274 if (IS_ERR(cfile_css
))
4276 if (cfile_css
!= css
) {
4281 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4285 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
4287 spin_lock(&memcg
->event_list_lock
);
4288 list_add(&event
->list
, &memcg
->event_list
);
4289 spin_unlock(&memcg
->event_list_lock
);
4301 eventfd_ctx_put(event
->eventfd
);
4310 static struct cftype mem_cgroup_legacy_files
[] = {
4312 .name
= "usage_in_bytes",
4313 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4314 .read_u64
= mem_cgroup_read_u64
,
4317 .name
= "max_usage_in_bytes",
4318 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4319 .write
= mem_cgroup_reset
,
4320 .read_u64
= mem_cgroup_read_u64
,
4323 .name
= "limit_in_bytes",
4324 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4325 .write
= mem_cgroup_write
,
4326 .read_u64
= mem_cgroup_read_u64
,
4329 .name
= "soft_limit_in_bytes",
4330 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4331 .write
= mem_cgroup_write
,
4332 .read_u64
= mem_cgroup_read_u64
,
4336 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4337 .write
= mem_cgroup_reset
,
4338 .read_u64
= mem_cgroup_read_u64
,
4342 .seq_show
= memcg_stat_show
,
4345 .name
= "force_empty",
4346 .write
= mem_cgroup_force_empty_write
,
4349 .name
= "use_hierarchy",
4350 .write_u64
= mem_cgroup_hierarchy_write
,
4351 .read_u64
= mem_cgroup_hierarchy_read
,
4354 .name
= "cgroup.event_control", /* XXX: for compat */
4355 .write
= memcg_write_event_control
,
4356 .flags
= CFTYPE_NO_PREFIX
,
4360 .name
= "swappiness",
4361 .read_u64
= mem_cgroup_swappiness_read
,
4362 .write_u64
= mem_cgroup_swappiness_write
,
4365 .name
= "move_charge_at_immigrate",
4366 .read_u64
= mem_cgroup_move_charge_read
,
4367 .write_u64
= mem_cgroup_move_charge_write
,
4370 .name
= "oom_control",
4371 .seq_show
= mem_cgroup_oom_control_read
,
4372 .write_u64
= mem_cgroup_oom_control_write
,
4373 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4376 .name
= "pressure_level",
4380 .name
= "numa_stat",
4381 .seq_show
= memcg_numa_stat_show
,
4384 #ifdef CONFIG_MEMCG_KMEM
4386 .name
= "kmem.limit_in_bytes",
4387 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4388 .write
= mem_cgroup_write
,
4389 .read_u64
= mem_cgroup_read_u64
,
4392 .name
= "kmem.usage_in_bytes",
4393 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4394 .read_u64
= mem_cgroup_read_u64
,
4397 .name
= "kmem.failcnt",
4398 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4399 .write
= mem_cgroup_reset
,
4400 .read_u64
= mem_cgroup_read_u64
,
4403 .name
= "kmem.max_usage_in_bytes",
4404 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4405 .write
= mem_cgroup_reset
,
4406 .read_u64
= mem_cgroup_read_u64
,
4408 #ifdef CONFIG_SLABINFO
4410 .name
= "kmem.slabinfo",
4411 .seq_start
= slab_start
,
4412 .seq_next
= slab_next
,
4413 .seq_stop
= slab_stop
,
4414 .seq_show
= memcg_slab_show
,
4418 { }, /* terminate */
4421 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4423 struct mem_cgroup_per_node
*pn
;
4424 struct mem_cgroup_per_zone
*mz
;
4425 int zone
, tmp
= node
;
4427 * This routine is called against possible nodes.
4428 * But it's BUG to call kmalloc() against offline node.
4430 * TODO: this routine can waste much memory for nodes which will
4431 * never be onlined. It's better to use memory hotplug callback
4434 if (!node_state(node
, N_NORMAL_MEMORY
))
4436 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4440 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4441 mz
= &pn
->zoneinfo
[zone
];
4442 lruvec_init(&mz
->lruvec
);
4443 mz
->usage_in_excess
= 0;
4444 mz
->on_tree
= false;
4447 memcg
->nodeinfo
[node
] = pn
;
4451 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4453 kfree(memcg
->nodeinfo
[node
]);
4456 static struct mem_cgroup
*mem_cgroup_alloc(void)
4458 struct mem_cgroup
*memcg
;
4461 size
= sizeof(struct mem_cgroup
);
4462 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4464 memcg
= kzalloc(size
, GFP_KERNEL
);
4468 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4471 spin_lock_init(&memcg
->pcp_counter_lock
);
4480 * At destroying mem_cgroup, references from swap_cgroup can remain.
4481 * (scanning all at force_empty is too costly...)
4483 * Instead of clearing all references at force_empty, we remember
4484 * the number of reference from swap_cgroup and free mem_cgroup when
4485 * it goes down to 0.
4487 * Removal of cgroup itself succeeds regardless of refs from swap.
4490 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4494 mem_cgroup_remove_from_trees(memcg
);
4497 free_mem_cgroup_per_zone_info(memcg
, node
);
4499 free_percpu(memcg
->stat
);
4504 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4506 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4508 if (!memcg
->memory
.parent
)
4510 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4512 EXPORT_SYMBOL(parent_mem_cgroup
);
4514 static struct cgroup_subsys_state
* __ref
4515 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4517 struct mem_cgroup
*memcg
;
4518 long error
= -ENOMEM
;
4521 memcg
= mem_cgroup_alloc();
4523 return ERR_PTR(error
);
4526 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4530 if (parent_css
== NULL
) {
4531 root_mem_cgroup
= memcg
;
4532 page_counter_init(&memcg
->memory
, NULL
);
4533 memcg
->high
= PAGE_COUNTER_MAX
;
4534 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4535 page_counter_init(&memcg
->memsw
, NULL
);
4536 page_counter_init(&memcg
->kmem
, NULL
);
4539 memcg
->last_scanned_node
= MAX_NUMNODES
;
4540 INIT_LIST_HEAD(&memcg
->oom_notify
);
4541 memcg
->move_charge_at_immigrate
= 0;
4542 mutex_init(&memcg
->thresholds_lock
);
4543 spin_lock_init(&memcg
->move_lock
);
4544 vmpressure_init(&memcg
->vmpressure
);
4545 INIT_LIST_HEAD(&memcg
->event_list
);
4546 spin_lock_init(&memcg
->event_list_lock
);
4547 #ifdef CONFIG_MEMCG_KMEM
4548 memcg
->kmemcg_id
= -1;
4554 __mem_cgroup_free(memcg
);
4555 return ERR_PTR(error
);
4559 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4561 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4562 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4565 if (css
->id
> MEM_CGROUP_ID_MAX
)
4571 mutex_lock(&memcg_create_mutex
);
4573 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4574 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4575 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4577 if (parent
->use_hierarchy
) {
4578 page_counter_init(&memcg
->memory
, &parent
->memory
);
4579 memcg
->high
= PAGE_COUNTER_MAX
;
4580 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4581 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4582 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4585 * No need to take a reference to the parent because cgroup
4586 * core guarantees its existence.
4589 page_counter_init(&memcg
->memory
, NULL
);
4590 memcg
->high
= PAGE_COUNTER_MAX
;
4591 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4592 page_counter_init(&memcg
->memsw
, NULL
);
4593 page_counter_init(&memcg
->kmem
, NULL
);
4595 * Deeper hierachy with use_hierarchy == false doesn't make
4596 * much sense so let cgroup subsystem know about this
4597 * unfortunate state in our controller.
4599 if (parent
!= root_mem_cgroup
)
4600 memory_cgrp_subsys
.broken_hierarchy
= true;
4602 mutex_unlock(&memcg_create_mutex
);
4604 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4609 * Make sure the memcg is initialized: mem_cgroup_iter()
4610 * orders reading memcg->initialized against its callers
4611 * reading the memcg members.
4613 smp_store_release(&memcg
->initialized
, 1);
4618 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4620 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4621 struct mem_cgroup_event
*event
, *tmp
;
4624 * Unregister events and notify userspace.
4625 * Notify userspace about cgroup removing only after rmdir of cgroup
4626 * directory to avoid race between userspace and kernelspace.
4628 spin_lock(&memcg
->event_list_lock
);
4629 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4630 list_del_init(&event
->list
);
4631 schedule_work(&event
->remove
);
4633 spin_unlock(&memcg
->event_list_lock
);
4635 vmpressure_cleanup(&memcg
->vmpressure
);
4637 memcg_deactivate_kmem(memcg
);
4640 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4642 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4644 memcg_destroy_kmem(memcg
);
4645 __mem_cgroup_free(memcg
);
4649 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4650 * @css: the target css
4652 * Reset the states of the mem_cgroup associated with @css. This is
4653 * invoked when the userland requests disabling on the default hierarchy
4654 * but the memcg is pinned through dependency. The memcg should stop
4655 * applying policies and should revert to the vanilla state as it may be
4656 * made visible again.
4658 * The current implementation only resets the essential configurations.
4659 * This needs to be expanded to cover all the visible parts.
4661 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4663 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4665 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4666 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4667 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4669 memcg
->high
= PAGE_COUNTER_MAX
;
4670 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4674 /* Handlers for move charge at task migration. */
4675 static int mem_cgroup_do_precharge(unsigned long count
)
4679 /* Try a single bulk charge without reclaim first */
4680 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
4682 mc
.precharge
+= count
;
4685 if (ret
== -EINTR
) {
4686 cancel_charge(root_mem_cgroup
, count
);
4690 /* Try charges one by one with reclaim */
4692 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4694 * In case of failure, any residual charges against
4695 * mc.to will be dropped by mem_cgroup_clear_mc()
4696 * later on. However, cancel any charges that are
4697 * bypassed to root right away or they'll be lost.
4700 cancel_charge(root_mem_cgroup
, 1);
4710 * get_mctgt_type - get target type of moving charge
4711 * @vma: the vma the pte to be checked belongs
4712 * @addr: the address corresponding to the pte to be checked
4713 * @ptent: the pte to be checked
4714 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4717 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4718 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4719 * move charge. if @target is not NULL, the page is stored in target->page
4720 * with extra refcnt got(Callers should handle it).
4721 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4722 * target for charge migration. if @target is not NULL, the entry is stored
4725 * Called with pte lock held.
4732 enum mc_target_type
{
4738 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4739 unsigned long addr
, pte_t ptent
)
4741 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4743 if (!page
|| !page_mapped(page
))
4745 if (PageAnon(page
)) {
4746 if (!(mc
.flags
& MOVE_ANON
))
4749 if (!(mc
.flags
& MOVE_FILE
))
4752 if (!get_page_unless_zero(page
))
4759 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4760 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4762 struct page
*page
= NULL
;
4763 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4765 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4768 * Because lookup_swap_cache() updates some statistics counter,
4769 * we call find_get_page() with swapper_space directly.
4771 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4772 if (do_swap_account
)
4773 entry
->val
= ent
.val
;
4778 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4779 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4785 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4786 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4788 struct page
*page
= NULL
;
4789 struct address_space
*mapping
;
4792 if (!vma
->vm_file
) /* anonymous vma */
4794 if (!(mc
.flags
& MOVE_FILE
))
4797 mapping
= vma
->vm_file
->f_mapping
;
4798 pgoff
= linear_page_index(vma
, addr
);
4800 /* page is moved even if it's not RSS of this task(page-faulted). */
4802 /* shmem/tmpfs may report page out on swap: account for that too. */
4803 if (shmem_mapping(mapping
)) {
4804 page
= find_get_entry(mapping
, pgoff
);
4805 if (radix_tree_exceptional_entry(page
)) {
4806 swp_entry_t swp
= radix_to_swp_entry(page
);
4807 if (do_swap_account
)
4809 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4812 page
= find_get_page(mapping
, pgoff
);
4814 page
= find_get_page(mapping
, pgoff
);
4819 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4820 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4822 struct page
*page
= NULL
;
4823 enum mc_target_type ret
= MC_TARGET_NONE
;
4824 swp_entry_t ent
= { .val
= 0 };
4826 if (pte_present(ptent
))
4827 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4828 else if (is_swap_pte(ptent
))
4829 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4830 else if (pte_none(ptent
))
4831 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4833 if (!page
&& !ent
.val
)
4837 * Do only loose check w/o serialization.
4838 * mem_cgroup_move_account() checks the page is valid or
4839 * not under LRU exclusion.
4841 if (page
->mem_cgroup
== mc
.from
) {
4842 ret
= MC_TARGET_PAGE
;
4844 target
->page
= page
;
4846 if (!ret
|| !target
)
4849 /* There is a swap entry and a page doesn't exist or isn't charged */
4850 if (ent
.val
&& !ret
&&
4851 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4852 ret
= MC_TARGET_SWAP
;
4859 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4861 * We don't consider swapping or file mapped pages because THP does not
4862 * support them for now.
4863 * Caller should make sure that pmd_trans_huge(pmd) is true.
4865 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4866 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4868 struct page
*page
= NULL
;
4869 enum mc_target_type ret
= MC_TARGET_NONE
;
4871 page
= pmd_page(pmd
);
4872 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4873 if (!(mc
.flags
& MOVE_ANON
))
4875 if (page
->mem_cgroup
== mc
.from
) {
4876 ret
= MC_TARGET_PAGE
;
4879 target
->page
= page
;
4885 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4886 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4888 return MC_TARGET_NONE
;
4892 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4893 unsigned long addr
, unsigned long end
,
4894 struct mm_walk
*walk
)
4896 struct vm_area_struct
*vma
= walk
->vma
;
4900 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4901 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4902 mc
.precharge
+= HPAGE_PMD_NR
;
4907 if (pmd_trans_unstable(pmd
))
4909 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4910 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4911 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4912 mc
.precharge
++; /* increment precharge temporarily */
4913 pte_unmap_unlock(pte
- 1, ptl
);
4919 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4921 unsigned long precharge
;
4923 struct mm_walk mem_cgroup_count_precharge_walk
= {
4924 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4927 down_read(&mm
->mmap_sem
);
4928 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4929 up_read(&mm
->mmap_sem
);
4931 precharge
= mc
.precharge
;
4937 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4939 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4941 VM_BUG_ON(mc
.moving_task
);
4942 mc
.moving_task
= current
;
4943 return mem_cgroup_do_precharge(precharge
);
4946 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4947 static void __mem_cgroup_clear_mc(void)
4949 struct mem_cgroup
*from
= mc
.from
;
4950 struct mem_cgroup
*to
= mc
.to
;
4952 /* we must uncharge all the leftover precharges from mc.to */
4954 cancel_charge(mc
.to
, mc
.precharge
);
4958 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4959 * we must uncharge here.
4961 if (mc
.moved_charge
) {
4962 cancel_charge(mc
.from
, mc
.moved_charge
);
4963 mc
.moved_charge
= 0;
4965 /* we must fixup refcnts and charges */
4966 if (mc
.moved_swap
) {
4967 /* uncharge swap account from the old cgroup */
4968 if (!mem_cgroup_is_root(mc
.from
))
4969 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4972 * we charged both to->memory and to->memsw, so we
4973 * should uncharge to->memory.
4975 if (!mem_cgroup_is_root(mc
.to
))
4976 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4978 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4980 /* we've already done css_get(mc.to) */
4983 memcg_oom_recover(from
);
4984 memcg_oom_recover(to
);
4985 wake_up_all(&mc
.waitq
);
4988 static void mem_cgroup_clear_mc(void)
4991 * we must clear moving_task before waking up waiters at the end of
4994 mc
.moving_task
= NULL
;
4995 __mem_cgroup_clear_mc();
4996 spin_lock(&mc
.lock
);
4999 spin_unlock(&mc
.lock
);
5002 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5003 struct cgroup_taskset
*tset
)
5005 struct task_struct
*p
= cgroup_taskset_first(tset
);
5007 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5008 unsigned long move_flags
;
5011 * We are now commited to this value whatever it is. Changes in this
5012 * tunable will only affect upcoming migrations, not the current one.
5013 * So we need to save it, and keep it going.
5015 move_flags
= ACCESS_ONCE(memcg
->move_charge_at_immigrate
);
5017 struct mm_struct
*mm
;
5018 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
5020 VM_BUG_ON(from
== memcg
);
5022 mm
= get_task_mm(p
);
5025 /* We move charges only when we move a owner of the mm */
5026 if (mm
->owner
== p
) {
5029 VM_BUG_ON(mc
.precharge
);
5030 VM_BUG_ON(mc
.moved_charge
);
5031 VM_BUG_ON(mc
.moved_swap
);
5033 spin_lock(&mc
.lock
);
5036 mc
.flags
= move_flags
;
5037 spin_unlock(&mc
.lock
);
5038 /* We set mc.moving_task later */
5040 ret
= mem_cgroup_precharge_mc(mm
);
5042 mem_cgroup_clear_mc();
5049 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5050 struct cgroup_taskset
*tset
)
5053 mem_cgroup_clear_mc();
5056 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5057 unsigned long addr
, unsigned long end
,
5058 struct mm_walk
*walk
)
5061 struct vm_area_struct
*vma
= walk
->vma
;
5064 enum mc_target_type target_type
;
5065 union mc_target target
;
5069 * We don't take compound_lock() here but no race with splitting thp
5071 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5072 * under splitting, which means there's no concurrent thp split,
5073 * - if another thread runs into split_huge_page() just after we
5074 * entered this if-block, the thread must wait for page table lock
5075 * to be unlocked in __split_huge_page_splitting(), where the main
5076 * part of thp split is not executed yet.
5078 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5079 if (mc
.precharge
< HPAGE_PMD_NR
) {
5083 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5084 if (target_type
== MC_TARGET_PAGE
) {
5086 if (!isolate_lru_page(page
)) {
5087 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
5089 mc
.precharge
-= HPAGE_PMD_NR
;
5090 mc
.moved_charge
+= HPAGE_PMD_NR
;
5092 putback_lru_page(page
);
5100 if (pmd_trans_unstable(pmd
))
5103 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5104 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5105 pte_t ptent
= *(pte
++);
5111 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5112 case MC_TARGET_PAGE
:
5114 if (isolate_lru_page(page
))
5116 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
5118 /* we uncharge from mc.from later. */
5121 putback_lru_page(page
);
5122 put
: /* get_mctgt_type() gets the page */
5125 case MC_TARGET_SWAP
:
5127 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5129 /* we fixup refcnts and charges later. */
5137 pte_unmap_unlock(pte
- 1, ptl
);
5142 * We have consumed all precharges we got in can_attach().
5143 * We try charge one by one, but don't do any additional
5144 * charges to mc.to if we have failed in charge once in attach()
5147 ret
= mem_cgroup_do_precharge(1);
5155 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
5157 struct mm_walk mem_cgroup_move_charge_walk
= {
5158 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5162 lru_add_drain_all();
5164 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5165 * move_lock while we're moving its pages to another memcg.
5166 * Then wait for already started RCU-only updates to finish.
5168 atomic_inc(&mc
.from
->moving_account
);
5171 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5173 * Someone who are holding the mmap_sem might be waiting in
5174 * waitq. So we cancel all extra charges, wake up all waiters,
5175 * and retry. Because we cancel precharges, we might not be able
5176 * to move enough charges, but moving charge is a best-effort
5177 * feature anyway, so it wouldn't be a big problem.
5179 __mem_cgroup_clear_mc();
5184 * When we have consumed all precharges and failed in doing
5185 * additional charge, the page walk just aborts.
5187 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5188 up_read(&mm
->mmap_sem
);
5189 atomic_dec(&mc
.from
->moving_account
);
5192 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5193 struct cgroup_taskset
*tset
)
5195 struct task_struct
*p
= cgroup_taskset_first(tset
);
5196 struct mm_struct
*mm
= get_task_mm(p
);
5200 mem_cgroup_move_charge(mm
);
5204 mem_cgroup_clear_mc();
5206 #else /* !CONFIG_MMU */
5207 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5208 struct cgroup_taskset
*tset
)
5212 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5213 struct cgroup_taskset
*tset
)
5216 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5217 struct cgroup_taskset
*tset
)
5223 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5224 * to verify whether we're attached to the default hierarchy on each mount
5227 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5230 * use_hierarchy is forced on the default hierarchy. cgroup core
5231 * guarantees that @root doesn't have any children, so turning it
5232 * on for the root memcg is enough.
5234 if (cgroup_on_dfl(root_css
->cgroup
))
5235 root_mem_cgroup
->use_hierarchy
= true;
5237 root_mem_cgroup
->use_hierarchy
= false;
5240 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5243 return mem_cgroup_usage(mem_cgroup_from_css(css
), false);
5246 static int memory_low_show(struct seq_file
*m
, void *v
)
5248 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5249 unsigned long low
= ACCESS_ONCE(memcg
->low
);
5251 if (low
== PAGE_COUNTER_MAX
)
5252 seq_puts(m
, "max\n");
5254 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5259 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5260 char *buf
, size_t nbytes
, loff_t off
)
5262 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5266 buf
= strstrip(buf
);
5267 err
= page_counter_memparse(buf
, "max", &low
);
5276 static int memory_high_show(struct seq_file
*m
, void *v
)
5278 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5279 unsigned long high
= ACCESS_ONCE(memcg
->high
);
5281 if (high
== PAGE_COUNTER_MAX
)
5282 seq_puts(m
, "max\n");
5284 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5289 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5290 char *buf
, size_t nbytes
, loff_t off
)
5292 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5296 buf
= strstrip(buf
);
5297 err
= page_counter_memparse(buf
, "max", &high
);
5306 static int memory_max_show(struct seq_file
*m
, void *v
)
5308 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5309 unsigned long max
= ACCESS_ONCE(memcg
->memory
.limit
);
5311 if (max
== PAGE_COUNTER_MAX
)
5312 seq_puts(m
, "max\n");
5314 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5319 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5320 char *buf
, size_t nbytes
, loff_t off
)
5322 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5326 buf
= strstrip(buf
);
5327 err
= page_counter_memparse(buf
, "max", &max
);
5331 err
= mem_cgroup_resize_limit(memcg
, max
);
5338 static int memory_events_show(struct seq_file
*m
, void *v
)
5340 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5342 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5343 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5344 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5345 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5350 static struct cftype memory_files
[] = {
5353 .read_u64
= memory_current_read
,
5357 .flags
= CFTYPE_NOT_ON_ROOT
,
5358 .seq_show
= memory_low_show
,
5359 .write
= memory_low_write
,
5363 .flags
= CFTYPE_NOT_ON_ROOT
,
5364 .seq_show
= memory_high_show
,
5365 .write
= memory_high_write
,
5369 .flags
= CFTYPE_NOT_ON_ROOT
,
5370 .seq_show
= memory_max_show
,
5371 .write
= memory_max_write
,
5375 .flags
= CFTYPE_NOT_ON_ROOT
,
5376 .seq_show
= memory_events_show
,
5381 struct cgroup_subsys memory_cgrp_subsys
= {
5382 .css_alloc
= mem_cgroup_css_alloc
,
5383 .css_online
= mem_cgroup_css_online
,
5384 .css_offline
= mem_cgroup_css_offline
,
5385 .css_free
= mem_cgroup_css_free
,
5386 .css_reset
= mem_cgroup_css_reset
,
5387 .can_attach
= mem_cgroup_can_attach
,
5388 .cancel_attach
= mem_cgroup_cancel_attach
,
5389 .attach
= mem_cgroup_move_task
,
5390 .bind
= mem_cgroup_bind
,
5391 .dfl_cftypes
= memory_files
,
5392 .legacy_cftypes
= mem_cgroup_legacy_files
,
5397 * mem_cgroup_events - count memory events against a cgroup
5398 * @memcg: the memory cgroup
5399 * @idx: the event index
5400 * @nr: the number of events to account for
5402 void mem_cgroup_events(struct mem_cgroup
*memcg
,
5403 enum mem_cgroup_events_index idx
,
5406 this_cpu_add(memcg
->stat
->events
[idx
], nr
);
5410 * mem_cgroup_low - check if memory consumption is below the normal range
5411 * @root: the highest ancestor to consider
5412 * @memcg: the memory cgroup to check
5414 * Returns %true if memory consumption of @memcg, and that of all
5415 * configurable ancestors up to @root, is below the normal range.
5417 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5419 if (mem_cgroup_disabled())
5423 * The toplevel group doesn't have a configurable range, so
5424 * it's never low when looked at directly, and it is not
5425 * considered an ancestor when assessing the hierarchy.
5428 if (memcg
== root_mem_cgroup
)
5431 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5434 while (memcg
!= root
) {
5435 memcg
= parent_mem_cgroup(memcg
);
5437 if (memcg
== root_mem_cgroup
)
5440 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5447 * mem_cgroup_try_charge - try charging a page
5448 * @page: page to charge
5449 * @mm: mm context of the victim
5450 * @gfp_mask: reclaim mode
5451 * @memcgp: charged memcg return
5453 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5454 * pages according to @gfp_mask if necessary.
5456 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5457 * Otherwise, an error code is returned.
5459 * After page->mapping has been set up, the caller must finalize the
5460 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5461 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5463 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5464 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5466 struct mem_cgroup
*memcg
= NULL
;
5467 unsigned int nr_pages
= 1;
5470 if (mem_cgroup_disabled())
5473 if (PageSwapCache(page
)) {
5475 * Every swap fault against a single page tries to charge the
5476 * page, bail as early as possible. shmem_unuse() encounters
5477 * already charged pages, too. The USED bit is protected by
5478 * the page lock, which serializes swap cache removal, which
5479 * in turn serializes uncharging.
5481 if (page
->mem_cgroup
)
5485 if (PageTransHuge(page
)) {
5486 nr_pages
<<= compound_order(page
);
5487 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5490 if (do_swap_account
&& PageSwapCache(page
))
5491 memcg
= try_get_mem_cgroup_from_page(page
);
5493 memcg
= get_mem_cgroup_from_mm(mm
);
5495 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5497 css_put(&memcg
->css
);
5499 if (ret
== -EINTR
) {
5500 memcg
= root_mem_cgroup
;
5509 * mem_cgroup_commit_charge - commit a page charge
5510 * @page: page to charge
5511 * @memcg: memcg to charge the page to
5512 * @lrucare: page might be on LRU already
5514 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5515 * after page->mapping has been set up. This must happen atomically
5516 * as part of the page instantiation, i.e. under the page table lock
5517 * for anonymous pages, under the page lock for page and swap cache.
5519 * In addition, the page must not be on the LRU during the commit, to
5520 * prevent racing with task migration. If it might be, use @lrucare.
5522 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5524 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5527 unsigned int nr_pages
= 1;
5529 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5530 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5532 if (mem_cgroup_disabled())
5535 * Swap faults will attempt to charge the same page multiple
5536 * times. But reuse_swap_page() might have removed the page
5537 * from swapcache already, so we can't check PageSwapCache().
5542 commit_charge(page
, memcg
, lrucare
);
5544 if (PageTransHuge(page
)) {
5545 nr_pages
<<= compound_order(page
);
5546 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5549 local_irq_disable();
5550 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5551 memcg_check_events(memcg
, page
);
5554 if (do_swap_account
&& PageSwapCache(page
)) {
5555 swp_entry_t entry
= { .val
= page_private(page
) };
5557 * The swap entry might not get freed for a long time,
5558 * let's not wait for it. The page already received a
5559 * memory+swap charge, drop the swap entry duplicate.
5561 mem_cgroup_uncharge_swap(entry
);
5566 * mem_cgroup_cancel_charge - cancel a page charge
5567 * @page: page to charge
5568 * @memcg: memcg to charge the page to
5570 * Cancel a charge transaction started by mem_cgroup_try_charge().
5572 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5574 unsigned int nr_pages
= 1;
5576 if (mem_cgroup_disabled())
5579 * Swap faults will attempt to charge the same page multiple
5580 * times. But reuse_swap_page() might have removed the page
5581 * from swapcache already, so we can't check PageSwapCache().
5586 if (PageTransHuge(page
)) {
5587 nr_pages
<<= compound_order(page
);
5588 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5591 cancel_charge(memcg
, nr_pages
);
5594 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5595 unsigned long nr_anon
, unsigned long nr_file
,
5596 unsigned long nr_huge
, struct page
*dummy_page
)
5598 unsigned long nr_pages
= nr_anon
+ nr_file
;
5599 unsigned long flags
;
5601 if (!mem_cgroup_is_root(memcg
)) {
5602 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5603 if (do_swap_account
)
5604 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5605 memcg_oom_recover(memcg
);
5608 local_irq_save(flags
);
5609 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5610 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5611 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5612 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5613 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5614 memcg_check_events(memcg
, dummy_page
);
5615 local_irq_restore(flags
);
5617 if (!mem_cgroup_is_root(memcg
))
5618 css_put_many(&memcg
->css
, nr_pages
);
5621 static void uncharge_list(struct list_head
*page_list
)
5623 struct mem_cgroup
*memcg
= NULL
;
5624 unsigned long nr_anon
= 0;
5625 unsigned long nr_file
= 0;
5626 unsigned long nr_huge
= 0;
5627 unsigned long pgpgout
= 0;
5628 struct list_head
*next
;
5631 next
= page_list
->next
;
5633 unsigned int nr_pages
= 1;
5635 page
= list_entry(next
, struct page
, lru
);
5636 next
= page
->lru
.next
;
5638 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5639 VM_BUG_ON_PAGE(page_count(page
), page
);
5641 if (!page
->mem_cgroup
)
5645 * Nobody should be changing or seriously looking at
5646 * page->mem_cgroup at this point, we have fully
5647 * exclusive access to the page.
5650 if (memcg
!= page
->mem_cgroup
) {
5652 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5654 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5656 memcg
= page
->mem_cgroup
;
5659 if (PageTransHuge(page
)) {
5660 nr_pages
<<= compound_order(page
);
5661 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5662 nr_huge
+= nr_pages
;
5666 nr_anon
+= nr_pages
;
5668 nr_file
+= nr_pages
;
5670 page
->mem_cgroup
= NULL
;
5673 } while (next
!= page_list
);
5676 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5681 * mem_cgroup_uncharge - uncharge a page
5682 * @page: page to uncharge
5684 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5685 * mem_cgroup_commit_charge().
5687 void mem_cgroup_uncharge(struct page
*page
)
5689 if (mem_cgroup_disabled())
5692 /* Don't touch page->lru of any random page, pre-check: */
5693 if (!page
->mem_cgroup
)
5696 INIT_LIST_HEAD(&page
->lru
);
5697 uncharge_list(&page
->lru
);
5701 * mem_cgroup_uncharge_list - uncharge a list of page
5702 * @page_list: list of pages to uncharge
5704 * Uncharge a list of pages previously charged with
5705 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5707 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5709 if (mem_cgroup_disabled())
5712 if (!list_empty(page_list
))
5713 uncharge_list(page_list
);
5717 * mem_cgroup_migrate - migrate a charge to another page
5718 * @oldpage: currently charged page
5719 * @newpage: page to transfer the charge to
5720 * @lrucare: either or both pages might be on the LRU already
5722 * Migrate the charge from @oldpage to @newpage.
5724 * Both pages must be locked, @newpage->mapping must be set up.
5726 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
5729 struct mem_cgroup
*memcg
;
5732 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5733 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5734 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
5735 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
5736 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5737 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5740 if (mem_cgroup_disabled())
5743 /* Page cache replacement: new page already charged? */
5744 if (newpage
->mem_cgroup
)
5748 * Swapcache readahead pages can get migrated before being
5749 * charged, and migration from compaction can happen to an
5750 * uncharged page when the PFN walker finds a page that
5751 * reclaim just put back on the LRU but has not released yet.
5753 memcg
= oldpage
->mem_cgroup
;
5758 lock_page_lru(oldpage
, &isolated
);
5760 oldpage
->mem_cgroup
= NULL
;
5763 unlock_page_lru(oldpage
, isolated
);
5765 commit_charge(newpage
, memcg
, lrucare
);
5769 * subsys_initcall() for memory controller.
5771 * Some parts like hotcpu_notifier() have to be initialized from this context
5772 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5773 * everything that doesn't depend on a specific mem_cgroup structure should
5774 * be initialized from here.
5776 static int __init
mem_cgroup_init(void)
5780 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5782 for_each_possible_cpu(cpu
)
5783 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5786 for_each_node(node
) {
5787 struct mem_cgroup_tree_per_node
*rtpn
;
5790 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5791 node_online(node
) ? node
: NUMA_NO_NODE
);
5793 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5794 struct mem_cgroup_tree_per_zone
*rtpz
;
5796 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5797 rtpz
->rb_root
= RB_ROOT
;
5798 spin_lock_init(&rtpz
->lock
);
5800 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5805 subsys_initcall(mem_cgroup_init
);
5807 #ifdef CONFIG_MEMCG_SWAP
5809 * mem_cgroup_swapout - transfer a memsw charge to swap
5810 * @page: page whose memsw charge to transfer
5811 * @entry: swap entry to move the charge to
5813 * Transfer the memsw charge of @page to @entry.
5815 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5817 struct mem_cgroup
*memcg
;
5818 unsigned short oldid
;
5820 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5821 VM_BUG_ON_PAGE(page_count(page
), page
);
5823 if (!do_swap_account
)
5826 memcg
= page
->mem_cgroup
;
5828 /* Readahead page, never charged */
5832 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5833 VM_BUG_ON_PAGE(oldid
, page
);
5834 mem_cgroup_swap_statistics(memcg
, true);
5836 page
->mem_cgroup
= NULL
;
5838 if (!mem_cgroup_is_root(memcg
))
5839 page_counter_uncharge(&memcg
->memory
, 1);
5841 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5842 VM_BUG_ON(!irqs_disabled());
5844 mem_cgroup_charge_statistics(memcg
, page
, -1);
5845 memcg_check_events(memcg
, page
);
5849 * mem_cgroup_uncharge_swap - uncharge a swap entry
5850 * @entry: swap entry to uncharge
5852 * Drop the memsw charge associated with @entry.
5854 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5856 struct mem_cgroup
*memcg
;
5859 if (!do_swap_account
)
5862 id
= swap_cgroup_record(entry
, 0);
5864 memcg
= mem_cgroup_lookup(id
);
5866 if (!mem_cgroup_is_root(memcg
))
5867 page_counter_uncharge(&memcg
->memsw
, 1);
5868 mem_cgroup_swap_statistics(memcg
, false);
5869 css_put(&memcg
->css
);
5874 /* for remember boot option*/
5875 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5876 static int really_do_swap_account __initdata
= 1;
5878 static int really_do_swap_account __initdata
;
5881 static int __init
enable_swap_account(char *s
)
5883 if (!strcmp(s
, "1"))
5884 really_do_swap_account
= 1;
5885 else if (!strcmp(s
, "0"))
5886 really_do_swap_account
= 0;
5889 __setup("swapaccount=", enable_swap_account
);
5891 static struct cftype memsw_cgroup_files
[] = {
5893 .name
= "memsw.usage_in_bytes",
5894 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5895 .read_u64
= mem_cgroup_read_u64
,
5898 .name
= "memsw.max_usage_in_bytes",
5899 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5900 .write
= mem_cgroup_reset
,
5901 .read_u64
= mem_cgroup_read_u64
,
5904 .name
= "memsw.limit_in_bytes",
5905 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5906 .write
= mem_cgroup_write
,
5907 .read_u64
= mem_cgroup_read_u64
,
5910 .name
= "memsw.failcnt",
5911 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5912 .write
= mem_cgroup_reset
,
5913 .read_u64
= mem_cgroup_read_u64
,
5915 { }, /* terminate */
5918 static int __init
mem_cgroup_swap_init(void)
5920 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5921 do_swap_account
= 1;
5922 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5923 memsw_cgroup_files
));
5927 subsys_initcall(mem_cgroup_swap_init
);
5929 #endif /* CONFIG_MEMCG_SWAP */