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
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 struct cgroup_subsys_state
*mem_cgroup_root_css __read_mostly
;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 int do_swap_account __read_mostly
;
86 #define do_swap_account 0
89 static const char * const mem_cgroup_stat_names
[] = {
99 static const char * const mem_cgroup_events_names
[] = {
106 static const char * const mem_cgroup_lru_names
[] = {
115 * Per memcg event counter is incremented at every pagein/pageout. With THP,
116 * it will be incremated by the number of pages. This counter is used for
117 * for trigger some periodic events. This is straightforward and better
118 * than using jiffies etc. to handle periodic memcg event.
120 enum mem_cgroup_events_target
{
121 MEM_CGROUP_TARGET_THRESH
,
122 MEM_CGROUP_TARGET_SOFTLIMIT
,
123 MEM_CGROUP_TARGET_NUMAINFO
,
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
130 struct mem_cgroup_stat_cpu
{
131 long count
[MEM_CGROUP_STAT_NSTATS
];
132 unsigned long events
[MEMCG_NR_EVENTS
];
133 unsigned long nr_page_events
;
134 unsigned long targets
[MEM_CGROUP_NTARGETS
];
137 struct reclaim_iter
{
138 struct mem_cgroup
*position
;
139 /* scan generation, increased every round-trip */
140 unsigned int generation
;
144 * per-zone information in memory controller.
146 struct mem_cgroup_per_zone
{
147 struct lruvec lruvec
;
148 unsigned long lru_size
[NR_LRU_LISTS
];
150 struct reclaim_iter iter
[DEF_PRIORITY
+ 1];
152 struct rb_node tree_node
; /* RB tree node */
153 unsigned long usage_in_excess
;/* Set to the value by which */
154 /* the soft limit is exceeded*/
156 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
157 /* use container_of */
160 struct mem_cgroup_per_node
{
161 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
165 * Cgroups above their limits are maintained in a RB-Tree, independent of
166 * their hierarchy representation
169 struct mem_cgroup_tree_per_zone
{
170 struct rb_root rb_root
;
174 struct mem_cgroup_tree_per_node
{
175 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
178 struct mem_cgroup_tree
{
179 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
182 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
184 struct mem_cgroup_threshold
{
185 struct eventfd_ctx
*eventfd
;
186 unsigned long threshold
;
190 struct mem_cgroup_threshold_ary
{
191 /* An array index points to threshold just below or equal to usage. */
192 int current_threshold
;
193 /* Size of entries[] */
195 /* Array of thresholds */
196 struct mem_cgroup_threshold entries
[0];
199 struct mem_cgroup_thresholds
{
200 /* Primary thresholds array */
201 struct mem_cgroup_threshold_ary
*primary
;
203 * Spare threshold array.
204 * This is needed to make mem_cgroup_unregister_event() "never fail".
205 * It must be able to store at least primary->size - 1 entries.
207 struct mem_cgroup_threshold_ary
*spare
;
211 struct mem_cgroup_eventfd_list
{
212 struct list_head list
;
213 struct eventfd_ctx
*eventfd
;
217 * cgroup_event represents events which userspace want to receive.
219 struct mem_cgroup_event
{
221 * memcg which the event belongs to.
223 struct mem_cgroup
*memcg
;
225 * eventfd to signal userspace about the event.
227 struct eventfd_ctx
*eventfd
;
229 * Each of these stored in a list by the cgroup.
231 struct list_head list
;
233 * register_event() callback will be used to add new userspace
234 * waiter for changes related to this event. Use eventfd_signal()
235 * on eventfd to send notification to userspace.
237 int (*register_event
)(struct mem_cgroup
*memcg
,
238 struct eventfd_ctx
*eventfd
, const char *args
);
240 * unregister_event() callback will be called when userspace closes
241 * the eventfd or on cgroup removing. This callback must be set,
242 * if you want provide notification functionality.
244 void (*unregister_event
)(struct mem_cgroup
*memcg
,
245 struct eventfd_ctx
*eventfd
);
247 * All fields below needed to unregister event when
248 * userspace closes eventfd.
251 wait_queue_head_t
*wqh
;
253 struct work_struct remove
;
256 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
257 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
260 * The memory controller data structure. The memory controller controls both
261 * page cache and RSS per cgroup. We would eventually like to provide
262 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
263 * to help the administrator determine what knobs to tune.
266 struct cgroup_subsys_state css
;
268 /* Accounted resources */
269 struct page_counter memory
;
270 struct page_counter memsw
;
271 struct page_counter kmem
;
273 /* Normal memory consumption range */
277 unsigned long soft_limit
;
279 /* vmpressure notifications */
280 struct vmpressure vmpressure
;
282 /* css_online() has been completed */
286 * Should the accounting and control be hierarchical, per subtree?
290 /* protected by memcg_oom_lock */
295 /* OOM-Killer disable */
296 int oom_kill_disable
;
298 /* protect arrays of thresholds */
299 struct mutex thresholds_lock
;
301 /* thresholds for memory usage. RCU-protected */
302 struct mem_cgroup_thresholds thresholds
;
304 /* thresholds for mem+swap usage. RCU-protected */
305 struct mem_cgroup_thresholds memsw_thresholds
;
307 /* For oom notifier event fd */
308 struct list_head oom_notify
;
311 * Should we move charges of a task when a task is moved into this
312 * mem_cgroup ? And what type of charges should we move ?
314 unsigned long move_charge_at_immigrate
;
316 * set > 0 if pages under this cgroup are moving to other cgroup.
318 atomic_t moving_account
;
319 /* taken only while moving_account > 0 */
320 spinlock_t move_lock
;
321 struct task_struct
*move_lock_task
;
322 unsigned long move_lock_flags
;
326 struct mem_cgroup_stat_cpu __percpu
*stat
;
327 spinlock_t pcp_counter_lock
;
329 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
330 struct cg_proto tcp_mem
;
332 #if defined(CONFIG_MEMCG_KMEM)
333 /* Index in the kmem_cache->memcg_params.memcg_caches array */
335 bool kmem_acct_activated
;
336 bool kmem_acct_active
;
339 int last_scanned_node
;
341 nodemask_t scan_nodes
;
342 atomic_t numainfo_events
;
343 atomic_t numainfo_updating
;
346 #ifdef CONFIG_CGROUP_WRITEBACK
347 struct list_head cgwb_list
;
348 struct wb_domain cgwb_domain
;
351 /* List of events which userspace want to receive */
352 struct list_head event_list
;
353 spinlock_t event_list_lock
;
355 struct mem_cgroup_per_node
*nodeinfo
[0];
356 /* WARNING: nodeinfo must be the last member here */
359 #ifdef CONFIG_MEMCG_KMEM
360 bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
362 return memcg
->kmem_acct_active
;
366 /* Stuffs for move charges at task migration. */
368 * Types of charges to be moved.
370 #define MOVE_ANON 0x1U
371 #define MOVE_FILE 0x2U
372 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
374 /* "mc" and its members are protected by cgroup_mutex */
375 static struct move_charge_struct
{
376 spinlock_t lock
; /* for from, to */
377 struct mem_cgroup
*from
;
378 struct mem_cgroup
*to
;
380 unsigned long precharge
;
381 unsigned long moved_charge
;
382 unsigned long moved_swap
;
383 struct task_struct
*moving_task
; /* a task moving charges */
384 wait_queue_head_t waitq
; /* a waitq for other context */
386 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
387 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
391 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
392 * limit reclaim to prevent infinite loops, if they ever occur.
394 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
395 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
398 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
399 MEM_CGROUP_CHARGE_TYPE_ANON
,
400 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
401 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
405 /* for encoding cft->private value on file */
413 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
414 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
415 #define MEMFILE_ATTR(val) ((val) & 0xffff)
416 /* Used for OOM nofiier */
417 #define OOM_CONTROL (0)
420 * The memcg_create_mutex will be held whenever a new cgroup is created.
421 * As a consequence, any change that needs to protect against new child cgroups
422 * appearing has to hold it as well.
424 static DEFINE_MUTEX(memcg_create_mutex
);
426 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
428 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
431 /* Some nice accessors for the vmpressure. */
432 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
435 memcg
= root_mem_cgroup
;
436 return &memcg
->vmpressure
;
439 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
441 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
444 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
446 return (memcg
== root_mem_cgroup
);
450 * We restrict the id in the range of [1, 65535], so it can fit into
453 #define MEM_CGROUP_ID_MAX USHRT_MAX
455 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
457 return memcg
->css
.id
;
461 * A helper function to get mem_cgroup from ID. must be called under
462 * rcu_read_lock(). The caller is responsible for calling
463 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
464 * refcnt from swap can be called against removed memcg.)
466 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
468 struct cgroup_subsys_state
*css
;
470 css
= css_from_id(id
, &memory_cgrp_subsys
);
471 return mem_cgroup_from_css(css
);
474 /* Writing them here to avoid exposing memcg's inner layout */
475 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
477 void sock_update_memcg(struct sock
*sk
)
479 if (mem_cgroup_sockets_enabled
) {
480 struct mem_cgroup
*memcg
;
481 struct cg_proto
*cg_proto
;
483 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
485 /* Socket cloning can throw us here with sk_cgrp already
486 * filled. It won't however, necessarily happen from
487 * process context. So the test for root memcg given
488 * the current task's memcg won't help us in this case.
490 * Respecting the original socket's memcg is a better
491 * decision in this case.
494 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
495 css_get(&sk
->sk_cgrp
->memcg
->css
);
500 memcg
= mem_cgroup_from_task(current
);
501 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
502 if (!mem_cgroup_is_root(memcg
) &&
503 memcg_proto_active(cg_proto
) &&
504 css_tryget_online(&memcg
->css
)) {
505 sk
->sk_cgrp
= cg_proto
;
510 EXPORT_SYMBOL(sock_update_memcg
);
512 void sock_release_memcg(struct sock
*sk
)
514 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
515 struct mem_cgroup
*memcg
;
516 WARN_ON(!sk
->sk_cgrp
->memcg
);
517 memcg
= sk
->sk_cgrp
->memcg
;
518 css_put(&sk
->sk_cgrp
->memcg
->css
);
522 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
524 if (!memcg
|| mem_cgroup_is_root(memcg
))
527 return &memcg
->tcp_mem
;
529 EXPORT_SYMBOL(tcp_proto_cgroup
);
533 #ifdef CONFIG_MEMCG_KMEM
535 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
536 * The main reason for not using cgroup id for this:
537 * this works better in sparse environments, where we have a lot of memcgs,
538 * but only a few kmem-limited. Or also, if we have, for instance, 200
539 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
540 * 200 entry array for that.
542 * The current size of the caches array is stored in memcg_nr_cache_ids. It
543 * will double each time we have to increase it.
545 static DEFINE_IDA(memcg_cache_ida
);
546 int memcg_nr_cache_ids
;
548 /* Protects memcg_nr_cache_ids */
549 static DECLARE_RWSEM(memcg_cache_ids_sem
);
551 void memcg_get_cache_ids(void)
553 down_read(&memcg_cache_ids_sem
);
556 void memcg_put_cache_ids(void)
558 up_read(&memcg_cache_ids_sem
);
562 * MIN_SIZE is different than 1, because we would like to avoid going through
563 * the alloc/free process all the time. In a small machine, 4 kmem-limited
564 * cgroups is a reasonable guess. In the future, it could be a parameter or
565 * tunable, but that is strictly not necessary.
567 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
568 * this constant directly from cgroup, but it is understandable that this is
569 * better kept as an internal representation in cgroup.c. In any case, the
570 * cgrp_id space is not getting any smaller, and we don't have to necessarily
571 * increase ours as well if it increases.
573 #define MEMCG_CACHES_MIN_SIZE 4
574 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
577 * A lot of the calls to the cache allocation functions are expected to be
578 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
579 * conditional to this static branch, we'll have to allow modules that does
580 * kmem_cache_alloc and the such to see this symbol as well
582 struct static_key memcg_kmem_enabled_key
;
583 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
585 #endif /* CONFIG_MEMCG_KMEM */
587 static struct mem_cgroup_per_zone
*
588 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
590 int nid
= zone_to_nid(zone
);
591 int zid
= zone_idx(zone
);
593 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
596 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
602 * mem_cgroup_css_from_page - css of the memcg associated with a page
603 * @page: page of interest
605 * If memcg is bound to the default hierarchy, css of the memcg associated
606 * with @page is returned. The returned css remains associated with @page
607 * until it is released.
609 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
612 * XXX: The above description of behavior on the default hierarchy isn't
613 * strictly true yet as replace_page_cache_page() can modify the
614 * association before @page is released even on the default hierarchy;
615 * however, the current and planned usages don't mix the the two functions
616 * and replace_page_cache_page() will soon be updated to make the invariant
619 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
621 struct mem_cgroup
*memcg
;
625 memcg
= page
->mem_cgroup
;
627 if (!memcg
|| !cgroup_on_dfl(memcg
->css
.cgroup
))
628 memcg
= root_mem_cgroup
;
634 static struct mem_cgroup_per_zone
*
635 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
637 int nid
= page_to_nid(page
);
638 int zid
= page_zonenum(page
);
640 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
643 static struct mem_cgroup_tree_per_zone
*
644 soft_limit_tree_node_zone(int nid
, int zid
)
646 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
649 static struct mem_cgroup_tree_per_zone
*
650 soft_limit_tree_from_page(struct page
*page
)
652 int nid
= page_to_nid(page
);
653 int zid
= page_zonenum(page
);
655 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
658 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
659 struct mem_cgroup_tree_per_zone
*mctz
,
660 unsigned long new_usage_in_excess
)
662 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
663 struct rb_node
*parent
= NULL
;
664 struct mem_cgroup_per_zone
*mz_node
;
669 mz
->usage_in_excess
= new_usage_in_excess
;
670 if (!mz
->usage_in_excess
)
674 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
676 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
679 * We can't avoid mem cgroups that are over their soft
680 * limit by the same amount
682 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
685 rb_link_node(&mz
->tree_node
, parent
, p
);
686 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
690 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
691 struct mem_cgroup_tree_per_zone
*mctz
)
695 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
699 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
700 struct mem_cgroup_tree_per_zone
*mctz
)
704 spin_lock_irqsave(&mctz
->lock
, flags
);
705 __mem_cgroup_remove_exceeded(mz
, mctz
);
706 spin_unlock_irqrestore(&mctz
->lock
, flags
);
709 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
711 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
712 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
713 unsigned long excess
= 0;
715 if (nr_pages
> soft_limit
)
716 excess
= nr_pages
- soft_limit
;
721 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
723 unsigned long excess
;
724 struct mem_cgroup_per_zone
*mz
;
725 struct mem_cgroup_tree_per_zone
*mctz
;
727 mctz
= soft_limit_tree_from_page(page
);
729 * Necessary to update all ancestors when hierarchy is used.
730 * because their event counter is not touched.
732 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
733 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
734 excess
= soft_limit_excess(memcg
);
736 * We have to update the tree if mz is on RB-tree or
737 * mem is over its softlimit.
739 if (excess
|| mz
->on_tree
) {
742 spin_lock_irqsave(&mctz
->lock
, flags
);
743 /* if on-tree, remove it */
745 __mem_cgroup_remove_exceeded(mz
, mctz
);
747 * Insert again. mz->usage_in_excess will be updated.
748 * If excess is 0, no tree ops.
750 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
751 spin_unlock_irqrestore(&mctz
->lock
, flags
);
756 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
758 struct mem_cgroup_tree_per_zone
*mctz
;
759 struct mem_cgroup_per_zone
*mz
;
763 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
764 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
765 mctz
= soft_limit_tree_node_zone(nid
, zid
);
766 mem_cgroup_remove_exceeded(mz
, mctz
);
771 static struct mem_cgroup_per_zone
*
772 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
774 struct rb_node
*rightmost
= NULL
;
775 struct mem_cgroup_per_zone
*mz
;
779 rightmost
= rb_last(&mctz
->rb_root
);
781 goto done
; /* Nothing to reclaim from */
783 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
785 * Remove the node now but someone else can add it back,
786 * we will to add it back at the end of reclaim to its correct
787 * position in the tree.
789 __mem_cgroup_remove_exceeded(mz
, mctz
);
790 if (!soft_limit_excess(mz
->memcg
) ||
791 !css_tryget_online(&mz
->memcg
->css
))
797 static struct mem_cgroup_per_zone
*
798 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
800 struct mem_cgroup_per_zone
*mz
;
802 spin_lock_irq(&mctz
->lock
);
803 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
804 spin_unlock_irq(&mctz
->lock
);
809 * Implementation Note: reading percpu statistics for memcg.
811 * Both of vmstat[] and percpu_counter has threshold and do periodic
812 * synchronization to implement "quick" read. There are trade-off between
813 * reading cost and precision of value. Then, we may have a chance to implement
814 * a periodic synchronizion of counter in memcg's counter.
816 * But this _read() function is used for user interface now. The user accounts
817 * memory usage by memory cgroup and he _always_ requires exact value because
818 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
819 * have to visit all online cpus and make sum. So, for now, unnecessary
820 * synchronization is not implemented. (just implemented for cpu hotplug)
822 * If there are kernel internal actions which can make use of some not-exact
823 * value, and reading all cpu value can be performance bottleneck in some
824 * common workload, threashold and synchonization as vmstat[] should be
827 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
828 enum mem_cgroup_stat_index idx
)
833 for_each_possible_cpu(cpu
)
834 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
838 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
839 enum mem_cgroup_events_index idx
)
841 unsigned long val
= 0;
844 for_each_possible_cpu(cpu
)
845 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
849 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
854 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
855 * counted as CACHE even if it's on ANON LRU.
858 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
861 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
864 if (PageTransHuge(page
))
865 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
868 /* pagein of a big page is an event. So, ignore page size */
870 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
872 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
873 nr_pages
= -nr_pages
; /* for event */
876 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
879 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
881 struct mem_cgroup_per_zone
*mz
;
883 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
884 return mz
->lru_size
[lru
];
887 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
889 unsigned int lru_mask
)
891 unsigned long nr
= 0;
894 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
896 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
897 struct mem_cgroup_per_zone
*mz
;
901 if (!(BIT(lru
) & lru_mask
))
903 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
904 nr
+= mz
->lru_size
[lru
];
910 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
911 unsigned int lru_mask
)
913 unsigned long nr
= 0;
916 for_each_node_state(nid
, N_MEMORY
)
917 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
921 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
922 enum mem_cgroup_events_target target
)
924 unsigned long val
, next
;
926 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
927 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
928 /* from time_after() in jiffies.h */
929 if ((long)next
- (long)val
< 0) {
931 case MEM_CGROUP_TARGET_THRESH
:
932 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
934 case MEM_CGROUP_TARGET_SOFTLIMIT
:
935 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
937 case MEM_CGROUP_TARGET_NUMAINFO
:
938 next
= val
+ NUMAINFO_EVENTS_TARGET
;
943 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
950 * Check events in order.
953 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
955 /* threshold event is triggered in finer grain than soft limit */
956 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
957 MEM_CGROUP_TARGET_THRESH
))) {
959 bool do_numainfo __maybe_unused
;
961 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
962 MEM_CGROUP_TARGET_SOFTLIMIT
);
964 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
965 MEM_CGROUP_TARGET_NUMAINFO
);
967 mem_cgroup_threshold(memcg
);
968 if (unlikely(do_softlimit
))
969 mem_cgroup_update_tree(memcg
, page
);
971 if (unlikely(do_numainfo
))
972 atomic_inc(&memcg
->numainfo_events
);
977 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
980 * mm_update_next_owner() may clear mm->owner to NULL
981 * if it races with swapoff, page migration, etc.
982 * So this can be called with p == NULL.
987 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
990 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
992 struct mem_cgroup
*memcg
= NULL
;
997 * Page cache insertions can happen withou an
998 * actual mm context, e.g. during disk probing
999 * on boot, loopback IO, acct() writes etc.
1002 memcg
= root_mem_cgroup
;
1004 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1005 if (unlikely(!memcg
))
1006 memcg
= root_mem_cgroup
;
1008 } while (!css_tryget_online(&memcg
->css
));
1014 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1015 * @root: hierarchy root
1016 * @prev: previously returned memcg, NULL on first invocation
1017 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1019 * Returns references to children of the hierarchy below @root, or
1020 * @root itself, or %NULL after a full round-trip.
1022 * Caller must pass the return value in @prev on subsequent
1023 * invocations for reference counting, or use mem_cgroup_iter_break()
1024 * to cancel a hierarchy walk before the round-trip is complete.
1026 * Reclaimers can specify a zone and a priority level in @reclaim to
1027 * divide up the memcgs in the hierarchy among all concurrent
1028 * reclaimers operating on the same zone and priority.
1030 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1031 struct mem_cgroup
*prev
,
1032 struct mem_cgroup_reclaim_cookie
*reclaim
)
1034 struct reclaim_iter
*uninitialized_var(iter
);
1035 struct cgroup_subsys_state
*css
= NULL
;
1036 struct mem_cgroup
*memcg
= NULL
;
1037 struct mem_cgroup
*pos
= NULL
;
1039 if (mem_cgroup_disabled())
1043 root
= root_mem_cgroup
;
1045 if (prev
&& !reclaim
)
1048 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1057 struct mem_cgroup_per_zone
*mz
;
1059 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1060 iter
= &mz
->iter
[reclaim
->priority
];
1062 if (prev
&& reclaim
->generation
!= iter
->generation
)
1066 pos
= READ_ONCE(iter
->position
);
1068 * A racing update may change the position and
1069 * put the last reference, hence css_tryget(),
1070 * or retry to see the updated position.
1072 } while (pos
&& !css_tryget(&pos
->css
));
1079 css
= css_next_descendant_pre(css
, &root
->css
);
1082 * Reclaimers share the hierarchy walk, and a
1083 * new one might jump in right at the end of
1084 * the hierarchy - make sure they see at least
1085 * one group and restart from the beginning.
1093 * Verify the css and acquire a reference. The root
1094 * is provided by the caller, so we know it's alive
1095 * and kicking, and don't take an extra reference.
1097 memcg
= mem_cgroup_from_css(css
);
1099 if (css
== &root
->css
)
1102 if (css_tryget(css
)) {
1104 * Make sure the memcg is initialized:
1105 * mem_cgroup_css_online() orders the the
1106 * initialization against setting the flag.
1108 if (smp_load_acquire(&memcg
->initialized
))
1118 if (cmpxchg(&iter
->position
, pos
, memcg
) == pos
) {
1120 css_get(&memcg
->css
);
1126 * pairs with css_tryget when dereferencing iter->position
1135 reclaim
->generation
= iter
->generation
;
1141 if (prev
&& prev
!= root
)
1142 css_put(&prev
->css
);
1148 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1149 * @root: hierarchy root
1150 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1152 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1153 struct mem_cgroup
*prev
)
1156 root
= root_mem_cgroup
;
1157 if (prev
&& prev
!= root
)
1158 css_put(&prev
->css
);
1162 * Iteration constructs for visiting all cgroups (under a tree). If
1163 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1164 * be used for reference counting.
1166 #define for_each_mem_cgroup_tree(iter, root) \
1167 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1169 iter = mem_cgroup_iter(root, iter, NULL))
1171 #define for_each_mem_cgroup(iter) \
1172 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1174 iter = mem_cgroup_iter(NULL, iter, NULL))
1176 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1178 struct mem_cgroup
*memcg
;
1181 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1182 if (unlikely(!memcg
))
1187 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1190 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1198 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1201 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1202 * @zone: zone of the wanted lruvec
1203 * @memcg: memcg of the wanted lruvec
1205 * Returns the lru list vector holding pages for the given @zone and
1206 * @mem. This can be the global zone lruvec, if the memory controller
1209 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1210 struct mem_cgroup
*memcg
)
1212 struct mem_cgroup_per_zone
*mz
;
1213 struct lruvec
*lruvec
;
1215 if (mem_cgroup_disabled()) {
1216 lruvec
= &zone
->lruvec
;
1220 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1221 lruvec
= &mz
->lruvec
;
1224 * Since a node can be onlined after the mem_cgroup was created,
1225 * we have to be prepared to initialize lruvec->zone here;
1226 * and if offlined then reonlined, we need to reinitialize it.
1228 if (unlikely(lruvec
->zone
!= zone
))
1229 lruvec
->zone
= zone
;
1234 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1236 * @zone: zone of the page
1238 * This function is only safe when following the LRU page isolation
1239 * and putback protocol: the LRU lock must be held, and the page must
1240 * either be PageLRU() or the caller must have isolated/allocated it.
1242 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1244 struct mem_cgroup_per_zone
*mz
;
1245 struct mem_cgroup
*memcg
;
1246 struct lruvec
*lruvec
;
1248 if (mem_cgroup_disabled()) {
1249 lruvec
= &zone
->lruvec
;
1253 memcg
= page
->mem_cgroup
;
1255 * Swapcache readahead pages are added to the LRU - and
1256 * possibly migrated - before they are charged.
1259 memcg
= root_mem_cgroup
;
1261 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1262 lruvec
= &mz
->lruvec
;
1265 * Since a node can be onlined after the mem_cgroup was created,
1266 * we have to be prepared to initialize lruvec->zone here;
1267 * and if offlined then reonlined, we need to reinitialize it.
1269 if (unlikely(lruvec
->zone
!= zone
))
1270 lruvec
->zone
= zone
;
1275 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1276 * @lruvec: mem_cgroup per zone lru vector
1277 * @lru: index of lru list the page is sitting on
1278 * @nr_pages: positive when adding or negative when removing
1280 * This function must be called when a page is added to or removed from an
1283 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1286 struct mem_cgroup_per_zone
*mz
;
1287 unsigned long *lru_size
;
1289 if (mem_cgroup_disabled())
1292 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1293 lru_size
= mz
->lru_size
+ lru
;
1294 *lru_size
+= nr_pages
;
1295 VM_BUG_ON((long)(*lru_size
) < 0);
1298 bool mem_cgroup_is_descendant(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
)
1302 if (!root
->use_hierarchy
)
1304 return cgroup_is_descendant(memcg
->css
.cgroup
, root
->css
.cgroup
);
1307 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1309 struct mem_cgroup
*task_memcg
;
1310 struct task_struct
*p
;
1313 p
= find_lock_task_mm(task
);
1315 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1319 * All threads may have already detached their mm's, but the oom
1320 * killer still needs to detect if they have already been oom
1321 * killed to prevent needlessly killing additional tasks.
1324 task_memcg
= mem_cgroup_from_task(task
);
1325 css_get(&task_memcg
->css
);
1328 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1329 css_put(&task_memcg
->css
);
1333 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1335 unsigned long inactive_ratio
;
1336 unsigned long inactive
;
1337 unsigned long active
;
1340 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1341 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1343 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1345 inactive_ratio
= int_sqrt(10 * gb
);
1349 return inactive
* inactive_ratio
< active
;
1352 bool mem_cgroup_lruvec_online(struct lruvec
*lruvec
)
1354 struct mem_cgroup_per_zone
*mz
;
1355 struct mem_cgroup
*memcg
;
1357 if (mem_cgroup_disabled())
1360 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1363 return !!(memcg
->css
.flags
& CSS_ONLINE
);
1366 #define mem_cgroup_from_counter(counter, member) \
1367 container_of(counter, struct mem_cgroup, member)
1370 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1371 * @memcg: the memory cgroup
1373 * Returns the maximum amount of memory @mem can be charged with, in
1376 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1378 unsigned long margin
= 0;
1379 unsigned long count
;
1380 unsigned long limit
;
1382 count
= page_counter_read(&memcg
->memory
);
1383 limit
= READ_ONCE(memcg
->memory
.limit
);
1385 margin
= limit
- count
;
1387 if (do_swap_account
) {
1388 count
= page_counter_read(&memcg
->memsw
);
1389 limit
= READ_ONCE(memcg
->memsw
.limit
);
1391 margin
= min(margin
, limit
- count
);
1397 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1400 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1401 return vm_swappiness
;
1403 return memcg
->swappiness
;
1407 * A routine for checking "mem" is under move_account() or not.
1409 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1410 * moving cgroups. This is for waiting at high-memory pressure
1413 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1415 struct mem_cgroup
*from
;
1416 struct mem_cgroup
*to
;
1419 * Unlike task_move routines, we access mc.to, mc.from not under
1420 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1422 spin_lock(&mc
.lock
);
1428 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1429 mem_cgroup_is_descendant(to
, memcg
);
1431 spin_unlock(&mc
.lock
);
1435 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1437 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1438 if (mem_cgroup_under_move(memcg
)) {
1440 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1441 /* moving charge context might have finished. */
1444 finish_wait(&mc
.waitq
, &wait
);
1451 #define K(x) ((x) << (PAGE_SHIFT-10))
1453 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1454 * @memcg: The memory cgroup that went over limit
1455 * @p: Task that is going to be killed
1457 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1460 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1462 /* oom_info_lock ensures that parallel ooms do not interleave */
1463 static DEFINE_MUTEX(oom_info_lock
);
1464 struct mem_cgroup
*iter
;
1467 mutex_lock(&oom_info_lock
);
1471 pr_info("Task in ");
1472 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1473 pr_cont(" killed as a result of limit of ");
1475 pr_info("Memory limit reached of cgroup ");
1478 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1483 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1484 K((u64
)page_counter_read(&memcg
->memory
)),
1485 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1486 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1487 K((u64
)page_counter_read(&memcg
->memsw
)),
1488 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1489 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1490 K((u64
)page_counter_read(&memcg
->kmem
)),
1491 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1493 for_each_mem_cgroup_tree(iter
, memcg
) {
1494 pr_info("Memory cgroup stats for ");
1495 pr_cont_cgroup_path(iter
->css
.cgroup
);
1498 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1499 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1501 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1502 K(mem_cgroup_read_stat(iter
, i
)));
1505 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1506 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1507 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1511 mutex_unlock(&oom_info_lock
);
1515 * This function returns the number of memcg under hierarchy tree. Returns
1516 * 1(self count) if no children.
1518 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1521 struct mem_cgroup
*iter
;
1523 for_each_mem_cgroup_tree(iter
, memcg
)
1529 * Return the memory (and swap, if configured) limit for a memcg.
1531 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1533 unsigned long limit
;
1535 limit
= memcg
->memory
.limit
;
1536 if (mem_cgroup_swappiness(memcg
)) {
1537 unsigned long memsw_limit
;
1539 memsw_limit
= memcg
->memsw
.limit
;
1540 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1545 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1548 struct oom_control oc
= {
1551 .gfp_mask
= gfp_mask
,
1554 struct mem_cgroup
*iter
;
1555 unsigned long chosen_points
= 0;
1556 unsigned long totalpages
;
1557 unsigned int points
= 0;
1558 struct task_struct
*chosen
= NULL
;
1560 mutex_lock(&oom_lock
);
1563 * If current has a pending SIGKILL or is exiting, then automatically
1564 * select it. The goal is to allow it to allocate so that it may
1565 * quickly exit and free its memory.
1567 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1568 mark_oom_victim(current
);
1572 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1573 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1574 for_each_mem_cgroup_tree(iter
, memcg
) {
1575 struct css_task_iter it
;
1576 struct task_struct
*task
;
1578 css_task_iter_start(&iter
->css
, &it
);
1579 while ((task
= css_task_iter_next(&it
))) {
1580 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1581 case OOM_SCAN_SELECT
:
1583 put_task_struct(chosen
);
1585 chosen_points
= ULONG_MAX
;
1586 get_task_struct(chosen
);
1588 case OOM_SCAN_CONTINUE
:
1590 case OOM_SCAN_ABORT
:
1591 css_task_iter_end(&it
);
1592 mem_cgroup_iter_break(memcg
, iter
);
1594 put_task_struct(chosen
);
1599 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1600 if (!points
|| points
< chosen_points
)
1602 /* Prefer thread group leaders for display purposes */
1603 if (points
== chosen_points
&&
1604 thread_group_leader(chosen
))
1608 put_task_struct(chosen
);
1610 chosen_points
= points
;
1611 get_task_struct(chosen
);
1613 css_task_iter_end(&it
);
1617 points
= chosen_points
* 1000 / totalpages
;
1618 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1619 "Memory cgroup out of memory");
1622 mutex_unlock(&oom_lock
);
1625 #if MAX_NUMNODES > 1
1628 * test_mem_cgroup_node_reclaimable
1629 * @memcg: the target memcg
1630 * @nid: the node ID to be checked.
1631 * @noswap : specify true here if the user wants flle only information.
1633 * This function returns whether the specified memcg contains any
1634 * reclaimable pages on a node. Returns true if there are any reclaimable
1635 * pages in the node.
1637 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1638 int nid
, bool noswap
)
1640 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1642 if (noswap
|| !total_swap_pages
)
1644 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1651 * Always updating the nodemask is not very good - even if we have an empty
1652 * list or the wrong list here, we can start from some node and traverse all
1653 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1656 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1660 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1661 * pagein/pageout changes since the last update.
1663 if (!atomic_read(&memcg
->numainfo_events
))
1665 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1668 /* make a nodemask where this memcg uses memory from */
1669 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1671 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1673 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1674 node_clear(nid
, memcg
->scan_nodes
);
1677 atomic_set(&memcg
->numainfo_events
, 0);
1678 atomic_set(&memcg
->numainfo_updating
, 0);
1682 * Selecting a node where we start reclaim from. Because what we need is just
1683 * reducing usage counter, start from anywhere is O,K. Considering
1684 * memory reclaim from current node, there are pros. and cons.
1686 * Freeing memory from current node means freeing memory from a node which
1687 * we'll use or we've used. So, it may make LRU bad. And if several threads
1688 * hit limits, it will see a contention on a node. But freeing from remote
1689 * node means more costs for memory reclaim because of memory latency.
1691 * Now, we use round-robin. Better algorithm is welcomed.
1693 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1697 mem_cgroup_may_update_nodemask(memcg
);
1698 node
= memcg
->last_scanned_node
;
1700 node
= next_node(node
, memcg
->scan_nodes
);
1701 if (node
== MAX_NUMNODES
)
1702 node
= first_node(memcg
->scan_nodes
);
1704 * We call this when we hit limit, not when pages are added to LRU.
1705 * No LRU may hold pages because all pages are UNEVICTABLE or
1706 * memcg is too small and all pages are not on LRU. In that case,
1707 * we use curret node.
1709 if (unlikely(node
== MAX_NUMNODES
))
1710 node
= numa_node_id();
1712 memcg
->last_scanned_node
= node
;
1716 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1722 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1725 unsigned long *total_scanned
)
1727 struct mem_cgroup
*victim
= NULL
;
1730 unsigned long excess
;
1731 unsigned long nr_scanned
;
1732 struct mem_cgroup_reclaim_cookie reclaim
= {
1737 excess
= soft_limit_excess(root_memcg
);
1740 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1745 * If we have not been able to reclaim
1746 * anything, it might because there are
1747 * no reclaimable pages under this hierarchy
1752 * We want to do more targeted reclaim.
1753 * excess >> 2 is not to excessive so as to
1754 * reclaim too much, nor too less that we keep
1755 * coming back to reclaim from this cgroup
1757 if (total
>= (excess
>> 2) ||
1758 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1763 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1765 *total_scanned
+= nr_scanned
;
1766 if (!soft_limit_excess(root_memcg
))
1769 mem_cgroup_iter_break(root_memcg
, victim
);
1773 #ifdef CONFIG_LOCKDEP
1774 static struct lockdep_map memcg_oom_lock_dep_map
= {
1775 .name
= "memcg_oom_lock",
1779 static DEFINE_SPINLOCK(memcg_oom_lock
);
1782 * Check OOM-Killer is already running under our hierarchy.
1783 * If someone is running, return false.
1785 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1787 struct mem_cgroup
*iter
, *failed
= NULL
;
1789 spin_lock(&memcg_oom_lock
);
1791 for_each_mem_cgroup_tree(iter
, memcg
) {
1792 if (iter
->oom_lock
) {
1794 * this subtree of our hierarchy is already locked
1795 * so we cannot give a lock.
1798 mem_cgroup_iter_break(memcg
, iter
);
1801 iter
->oom_lock
= true;
1806 * OK, we failed to lock the whole subtree so we have
1807 * to clean up what we set up to the failing subtree
1809 for_each_mem_cgroup_tree(iter
, memcg
) {
1810 if (iter
== failed
) {
1811 mem_cgroup_iter_break(memcg
, iter
);
1814 iter
->oom_lock
= false;
1817 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1819 spin_unlock(&memcg_oom_lock
);
1824 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1826 struct mem_cgroup
*iter
;
1828 spin_lock(&memcg_oom_lock
);
1829 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1830 for_each_mem_cgroup_tree(iter
, memcg
)
1831 iter
->oom_lock
= false;
1832 spin_unlock(&memcg_oom_lock
);
1835 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1837 struct mem_cgroup
*iter
;
1839 spin_lock(&memcg_oom_lock
);
1840 for_each_mem_cgroup_tree(iter
, memcg
)
1842 spin_unlock(&memcg_oom_lock
);
1845 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1847 struct mem_cgroup
*iter
;
1850 * When a new child is created while the hierarchy is under oom,
1851 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1853 spin_lock(&memcg_oom_lock
);
1854 for_each_mem_cgroup_tree(iter
, memcg
)
1855 if (iter
->under_oom
> 0)
1857 spin_unlock(&memcg_oom_lock
);
1860 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1862 struct oom_wait_info
{
1863 struct mem_cgroup
*memcg
;
1867 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1868 unsigned mode
, int sync
, void *arg
)
1870 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1871 struct mem_cgroup
*oom_wait_memcg
;
1872 struct oom_wait_info
*oom_wait_info
;
1874 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1875 oom_wait_memcg
= oom_wait_info
->memcg
;
1877 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1878 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1880 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1883 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1886 * For the following lockless ->under_oom test, the only required
1887 * guarantee is that it must see the state asserted by an OOM when
1888 * this function is called as a result of userland actions
1889 * triggered by the notification of the OOM. This is trivially
1890 * achieved by invoking mem_cgroup_mark_under_oom() before
1891 * triggering notification.
1893 if (memcg
&& memcg
->under_oom
)
1894 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1897 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1899 if (!current
->memcg_oom
.may_oom
)
1902 * We are in the middle of the charge context here, so we
1903 * don't want to block when potentially sitting on a callstack
1904 * that holds all kinds of filesystem and mm locks.
1906 * Also, the caller may handle a failed allocation gracefully
1907 * (like optional page cache readahead) and so an OOM killer
1908 * invocation might not even be necessary.
1910 * That's why we don't do anything here except remember the
1911 * OOM context and then deal with it at the end of the page
1912 * fault when the stack is unwound, the locks are released,
1913 * and when we know whether the fault was overall successful.
1915 css_get(&memcg
->css
);
1916 current
->memcg_oom
.memcg
= memcg
;
1917 current
->memcg_oom
.gfp_mask
= mask
;
1918 current
->memcg_oom
.order
= order
;
1922 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1923 * @handle: actually kill/wait or just clean up the OOM state
1925 * This has to be called at the end of a page fault if the memcg OOM
1926 * handler was enabled.
1928 * Memcg supports userspace OOM handling where failed allocations must
1929 * sleep on a waitqueue until the userspace task resolves the
1930 * situation. Sleeping directly in the charge context with all kinds
1931 * of locks held is not a good idea, instead we remember an OOM state
1932 * in the task and mem_cgroup_oom_synchronize() has to be called at
1933 * the end of the page fault to complete the OOM handling.
1935 * Returns %true if an ongoing memcg OOM situation was detected and
1936 * completed, %false otherwise.
1938 bool mem_cgroup_oom_synchronize(bool handle
)
1940 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
1941 struct oom_wait_info owait
;
1944 /* OOM is global, do not handle */
1948 if (!handle
|| oom_killer_disabled
)
1951 owait
.memcg
= memcg
;
1952 owait
.wait
.flags
= 0;
1953 owait
.wait
.func
= memcg_oom_wake_function
;
1954 owait
.wait
.private = current
;
1955 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1957 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1958 mem_cgroup_mark_under_oom(memcg
);
1960 locked
= mem_cgroup_oom_trylock(memcg
);
1963 mem_cgroup_oom_notify(memcg
);
1965 if (locked
&& !memcg
->oom_kill_disable
) {
1966 mem_cgroup_unmark_under_oom(memcg
);
1967 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1968 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
1969 current
->memcg_oom
.order
);
1972 mem_cgroup_unmark_under_oom(memcg
);
1973 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1977 mem_cgroup_oom_unlock(memcg
);
1979 * There is no guarantee that an OOM-lock contender
1980 * sees the wakeups triggered by the OOM kill
1981 * uncharges. Wake any sleepers explicitely.
1983 memcg_oom_recover(memcg
);
1986 current
->memcg_oom
.memcg
= NULL
;
1987 css_put(&memcg
->css
);
1992 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1993 * @page: page that is going to change accounted state
1995 * This function must mark the beginning of an accounted page state
1996 * change to prevent double accounting when the page is concurrently
1997 * being moved to another memcg:
1999 * memcg = mem_cgroup_begin_page_stat(page);
2000 * if (TestClearPageState(page))
2001 * mem_cgroup_update_page_stat(memcg, state, -1);
2002 * mem_cgroup_end_page_stat(memcg);
2004 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
2006 struct mem_cgroup
*memcg
;
2007 unsigned long flags
;
2010 * The RCU lock is held throughout the transaction. The fast
2011 * path can get away without acquiring the memcg->move_lock
2012 * because page moving starts with an RCU grace period.
2014 * The RCU lock also protects the memcg from being freed when
2015 * the page state that is going to change is the only thing
2016 * preventing the page from being uncharged.
2017 * E.g. end-writeback clearing PageWriteback(), which allows
2018 * migration to go ahead and uncharge the page before the
2019 * account transaction might be complete.
2023 if (mem_cgroup_disabled())
2026 memcg
= page
->mem_cgroup
;
2027 if (unlikely(!memcg
))
2030 if (atomic_read(&memcg
->moving_account
) <= 0)
2033 spin_lock_irqsave(&memcg
->move_lock
, flags
);
2034 if (memcg
!= page
->mem_cgroup
) {
2035 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2040 * When charge migration first begins, we can have locked and
2041 * unlocked page stat updates happening concurrently. Track
2042 * the task who has the lock for mem_cgroup_end_page_stat().
2044 memcg
->move_lock_task
= current
;
2045 memcg
->move_lock_flags
= flags
;
2049 EXPORT_SYMBOL(mem_cgroup_begin_page_stat
);
2052 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2053 * @memcg: the memcg that was accounted against
2055 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
2057 if (memcg
&& memcg
->move_lock_task
== current
) {
2058 unsigned long flags
= memcg
->move_lock_flags
;
2060 memcg
->move_lock_task
= NULL
;
2061 memcg
->move_lock_flags
= 0;
2063 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
2068 EXPORT_SYMBOL(mem_cgroup_end_page_stat
);
2071 * mem_cgroup_update_page_stat - update page state statistics
2072 * @memcg: memcg to account against
2073 * @idx: page state item to account
2074 * @val: number of pages (positive or negative)
2076 * See mem_cgroup_begin_page_stat() for locking requirements.
2078 void mem_cgroup_update_page_stat(struct mem_cgroup
*memcg
,
2079 enum mem_cgroup_stat_index idx
, int val
)
2081 VM_BUG_ON(!rcu_read_lock_held());
2084 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2088 * size of first charge trial. "32" comes from vmscan.c's magic value.
2089 * TODO: maybe necessary to use big numbers in big irons.
2091 #define CHARGE_BATCH 32U
2092 struct memcg_stock_pcp
{
2093 struct mem_cgroup
*cached
; /* this never be root cgroup */
2094 unsigned int nr_pages
;
2095 struct work_struct work
;
2096 unsigned long flags
;
2097 #define FLUSHING_CACHED_CHARGE 0
2099 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2100 static DEFINE_MUTEX(percpu_charge_mutex
);
2103 * consume_stock: Try to consume stocked charge on this cpu.
2104 * @memcg: memcg to consume from.
2105 * @nr_pages: how many pages to charge.
2107 * The charges will only happen if @memcg matches the current cpu's memcg
2108 * stock, and at least @nr_pages are available in that stock. Failure to
2109 * service an allocation will refill the stock.
2111 * returns true if successful, false otherwise.
2113 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2115 struct memcg_stock_pcp
*stock
;
2118 if (nr_pages
> CHARGE_BATCH
)
2121 stock
= &get_cpu_var(memcg_stock
);
2122 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
2123 stock
->nr_pages
-= nr_pages
;
2126 put_cpu_var(memcg_stock
);
2131 * Returns stocks cached in percpu and reset cached information.
2133 static void drain_stock(struct memcg_stock_pcp
*stock
)
2135 struct mem_cgroup
*old
= stock
->cached
;
2137 if (stock
->nr_pages
) {
2138 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
2139 if (do_swap_account
)
2140 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
2141 css_put_many(&old
->css
, stock
->nr_pages
);
2142 stock
->nr_pages
= 0;
2144 stock
->cached
= NULL
;
2148 * This must be called under preempt disabled or must be called by
2149 * a thread which is pinned to local cpu.
2151 static void drain_local_stock(struct work_struct
*dummy
)
2153 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2155 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2159 * Cache charges(val) to local per_cpu area.
2160 * This will be consumed by consume_stock() function, later.
2162 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2164 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2166 if (stock
->cached
!= memcg
) { /* reset if necessary */
2168 stock
->cached
= memcg
;
2170 stock
->nr_pages
+= nr_pages
;
2171 put_cpu_var(memcg_stock
);
2175 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2176 * of the hierarchy under it.
2178 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
2182 /* If someone's already draining, avoid adding running more workers. */
2183 if (!mutex_trylock(&percpu_charge_mutex
))
2185 /* Notify other cpus that system-wide "drain" is running */
2188 for_each_online_cpu(cpu
) {
2189 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2190 struct mem_cgroup
*memcg
;
2192 memcg
= stock
->cached
;
2193 if (!memcg
|| !stock
->nr_pages
)
2195 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
2197 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2199 drain_local_stock(&stock
->work
);
2201 schedule_work_on(cpu
, &stock
->work
);
2206 mutex_unlock(&percpu_charge_mutex
);
2209 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2210 unsigned long action
,
2213 int cpu
= (unsigned long)hcpu
;
2214 struct memcg_stock_pcp
*stock
;
2216 if (action
== CPU_ONLINE
)
2219 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2222 stock
= &per_cpu(memcg_stock
, cpu
);
2227 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2228 unsigned int nr_pages
)
2230 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2231 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2232 struct mem_cgroup
*mem_over_limit
;
2233 struct page_counter
*counter
;
2234 unsigned long nr_reclaimed
;
2235 bool may_swap
= true;
2236 bool drained
= false;
2239 if (mem_cgroup_is_root(memcg
))
2242 if (consume_stock(memcg
, nr_pages
))
2245 if (!do_swap_account
||
2246 !page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
2247 if (!page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
2249 if (do_swap_account
)
2250 page_counter_uncharge(&memcg
->memsw
, batch
);
2251 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
2253 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
2257 if (batch
> nr_pages
) {
2263 * Unlike in global OOM situations, memcg is not in a physical
2264 * memory shortage. Allow dying and OOM-killed tasks to
2265 * bypass the last charges so that they can exit quickly and
2266 * free their memory.
2268 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2269 fatal_signal_pending(current
) ||
2270 current
->flags
& PF_EXITING
))
2273 if (unlikely(task_in_memcg_oom(current
)))
2276 if (!(gfp_mask
& __GFP_WAIT
))
2279 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2281 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2282 gfp_mask
, may_swap
);
2284 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2288 drain_all_stock(mem_over_limit
);
2293 if (gfp_mask
& __GFP_NORETRY
)
2296 * Even though the limit is exceeded at this point, reclaim
2297 * may have been able to free some pages. Retry the charge
2298 * before killing the task.
2300 * Only for regular pages, though: huge pages are rather
2301 * unlikely to succeed so close to the limit, and we fall back
2302 * to regular pages anyway in case of failure.
2304 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2307 * At task move, charge accounts can be doubly counted. So, it's
2308 * better to wait until the end of task_move if something is going on.
2310 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2316 if (gfp_mask
& __GFP_NOFAIL
)
2319 if (fatal_signal_pending(current
))
2322 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2324 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2326 if (!(gfp_mask
& __GFP_NOFAIL
))
2332 css_get_many(&memcg
->css
, batch
);
2333 if (batch
> nr_pages
)
2334 refill_stock(memcg
, batch
- nr_pages
);
2335 if (!(gfp_mask
& __GFP_WAIT
))
2338 * If the hierarchy is above the normal consumption range,
2339 * make the charging task trim their excess contribution.
2342 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2344 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
2345 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2346 } while ((memcg
= parent_mem_cgroup(memcg
)));
2351 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2353 if (mem_cgroup_is_root(memcg
))
2356 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2357 if (do_swap_account
)
2358 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2360 css_put_many(&memcg
->css
, nr_pages
);
2364 * try_get_mem_cgroup_from_page - look up page's memcg association
2367 * Look up, get a css reference, and return the memcg that owns @page.
2369 * The page must be locked to prevent racing with swap-in and page
2370 * cache charges. If coming from an unlocked page table, the caller
2371 * must ensure the page is on the LRU or this can race with charging.
2373 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2375 struct mem_cgroup
*memcg
;
2379 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2381 memcg
= page
->mem_cgroup
;
2383 if (!css_tryget_online(&memcg
->css
))
2385 } else if (PageSwapCache(page
)) {
2386 ent
.val
= page_private(page
);
2387 id
= lookup_swap_cgroup_id(ent
);
2389 memcg
= mem_cgroup_from_id(id
);
2390 if (memcg
&& !css_tryget_online(&memcg
->css
))
2397 static void lock_page_lru(struct page
*page
, int *isolated
)
2399 struct zone
*zone
= page_zone(page
);
2401 spin_lock_irq(&zone
->lru_lock
);
2402 if (PageLRU(page
)) {
2403 struct lruvec
*lruvec
;
2405 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2407 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2413 static void unlock_page_lru(struct page
*page
, int isolated
)
2415 struct zone
*zone
= page_zone(page
);
2418 struct lruvec
*lruvec
;
2420 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2421 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2423 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2425 spin_unlock_irq(&zone
->lru_lock
);
2428 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2433 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2436 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2437 * may already be on some other mem_cgroup's LRU. Take care of it.
2440 lock_page_lru(page
, &isolated
);
2443 * Nobody should be changing or seriously looking at
2444 * page->mem_cgroup at this point:
2446 * - the page is uncharged
2448 * - the page is off-LRU
2450 * - an anonymous fault has exclusive page access, except for
2451 * a locked page table
2453 * - a page cache insertion, a swapin fault, or a migration
2454 * have the page locked
2456 page
->mem_cgroup
= memcg
;
2459 unlock_page_lru(page
, isolated
);
2462 #ifdef CONFIG_MEMCG_KMEM
2463 int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
,
2464 unsigned long nr_pages
)
2466 struct page_counter
*counter
;
2469 ret
= page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
);
2473 ret
= try_charge(memcg
, gfp
, nr_pages
);
2474 if (ret
== -EINTR
) {
2476 * try_charge() chose to bypass to root due to OOM kill or
2477 * fatal signal. Since our only options are to either fail
2478 * the allocation or charge it to this cgroup, do it as a
2479 * temporary condition. But we can't fail. From a kmem/slab
2480 * perspective, the cache has already been selected, by
2481 * mem_cgroup_kmem_get_cache(), so it is too late to change
2484 * This condition will only trigger if the task entered
2485 * memcg_charge_kmem in a sane state, but was OOM-killed
2486 * during try_charge() above. Tasks that were already dying
2487 * when the allocation triggers should have been already
2488 * directed to the root cgroup in memcontrol.h
2490 page_counter_charge(&memcg
->memory
, nr_pages
);
2491 if (do_swap_account
)
2492 page_counter_charge(&memcg
->memsw
, nr_pages
);
2493 css_get_many(&memcg
->css
, nr_pages
);
2496 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2501 void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, unsigned long nr_pages
)
2503 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2504 if (do_swap_account
)
2505 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2507 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2509 css_put_many(&memcg
->css
, nr_pages
);
2513 * helper for acessing a memcg's index. It will be used as an index in the
2514 * child cache array in kmem_cache, and also to derive its name. This function
2515 * will return -1 when this is not a kmem-limited memcg.
2517 int memcg_cache_id(struct mem_cgroup
*memcg
)
2519 return memcg
? memcg
->kmemcg_id
: -1;
2522 static int memcg_alloc_cache_id(void)
2527 id
= ida_simple_get(&memcg_cache_ida
,
2528 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2532 if (id
< memcg_nr_cache_ids
)
2536 * There's no space for the new id in memcg_caches arrays,
2537 * so we have to grow them.
2539 down_write(&memcg_cache_ids_sem
);
2541 size
= 2 * (id
+ 1);
2542 if (size
< MEMCG_CACHES_MIN_SIZE
)
2543 size
= MEMCG_CACHES_MIN_SIZE
;
2544 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2545 size
= MEMCG_CACHES_MAX_SIZE
;
2547 err
= memcg_update_all_caches(size
);
2549 err
= memcg_update_all_list_lrus(size
);
2551 memcg_nr_cache_ids
= size
;
2553 up_write(&memcg_cache_ids_sem
);
2556 ida_simple_remove(&memcg_cache_ida
, id
);
2562 static void memcg_free_cache_id(int id
)
2564 ida_simple_remove(&memcg_cache_ida
, id
);
2567 struct memcg_kmem_cache_create_work
{
2568 struct mem_cgroup
*memcg
;
2569 struct kmem_cache
*cachep
;
2570 struct work_struct work
;
2573 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2575 struct memcg_kmem_cache_create_work
*cw
=
2576 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2577 struct mem_cgroup
*memcg
= cw
->memcg
;
2578 struct kmem_cache
*cachep
= cw
->cachep
;
2580 memcg_create_kmem_cache(memcg
, cachep
);
2582 css_put(&memcg
->css
);
2587 * Enqueue the creation of a per-memcg kmem_cache.
2589 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2590 struct kmem_cache
*cachep
)
2592 struct memcg_kmem_cache_create_work
*cw
;
2594 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2598 css_get(&memcg
->css
);
2601 cw
->cachep
= cachep
;
2602 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2604 schedule_work(&cw
->work
);
2607 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2608 struct kmem_cache
*cachep
)
2611 * We need to stop accounting when we kmalloc, because if the
2612 * corresponding kmalloc cache is not yet created, the first allocation
2613 * in __memcg_schedule_kmem_cache_create will recurse.
2615 * However, it is better to enclose the whole function. Depending on
2616 * the debugging options enabled, INIT_WORK(), for instance, can
2617 * trigger an allocation. This too, will make us recurse. Because at
2618 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2619 * the safest choice is to do it like this, wrapping the whole function.
2621 current
->memcg_kmem_skip_account
= 1;
2622 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2623 current
->memcg_kmem_skip_account
= 0;
2627 * Return the kmem_cache we're supposed to use for a slab allocation.
2628 * We try to use the current memcg's version of the cache.
2630 * If the cache does not exist yet, if we are the first user of it,
2631 * we either create it immediately, if possible, or create it asynchronously
2633 * In the latter case, we will let the current allocation go through with
2634 * the original cache.
2636 * Can't be called in interrupt context or from kernel threads.
2637 * This function needs to be called with rcu_read_lock() held.
2639 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2641 struct mem_cgroup
*memcg
;
2642 struct kmem_cache
*memcg_cachep
;
2645 VM_BUG_ON(!is_root_cache(cachep
));
2647 if (current
->memcg_kmem_skip_account
)
2650 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2651 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2655 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2656 if (likely(memcg_cachep
))
2657 return memcg_cachep
;
2660 * If we are in a safe context (can wait, and not in interrupt
2661 * context), we could be be predictable and return right away.
2662 * This would guarantee that the allocation being performed
2663 * already belongs in the new cache.
2665 * However, there are some clashes that can arrive from locking.
2666 * For instance, because we acquire the slab_mutex while doing
2667 * memcg_create_kmem_cache, this means no further allocation
2668 * could happen with the slab_mutex held. So it's better to
2671 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2673 css_put(&memcg
->css
);
2677 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2679 if (!is_root_cache(cachep
))
2680 css_put(&cachep
->memcg_params
.memcg
->css
);
2684 * We need to verify if the allocation against current->mm->owner's memcg is
2685 * possible for the given order. But the page is not allocated yet, so we'll
2686 * need a further commit step to do the final arrangements.
2688 * It is possible for the task to switch cgroups in this mean time, so at
2689 * commit time, we can't rely on task conversion any longer. We'll then use
2690 * the handle argument to return to the caller which cgroup we should commit
2691 * against. We could also return the memcg directly and avoid the pointer
2692 * passing, but a boolean return value gives better semantics considering
2693 * the compiled-out case as well.
2695 * Returning true means the allocation is possible.
2698 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
2700 struct mem_cgroup
*memcg
;
2705 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2707 if (!memcg_kmem_is_active(memcg
)) {
2708 css_put(&memcg
->css
);
2712 ret
= memcg_charge_kmem(memcg
, gfp
, 1 << order
);
2716 css_put(&memcg
->css
);
2720 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2723 VM_BUG_ON(mem_cgroup_is_root(memcg
));
2725 /* The page allocation failed. Revert */
2727 memcg_uncharge_kmem(memcg
, 1 << order
);
2730 page
->mem_cgroup
= memcg
;
2733 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
2735 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2740 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2742 memcg_uncharge_kmem(memcg
, 1 << order
);
2743 page
->mem_cgroup
= NULL
;
2746 struct mem_cgroup
*__mem_cgroup_from_kmem(void *ptr
)
2748 struct mem_cgroup
*memcg
= NULL
;
2749 struct kmem_cache
*cachep
;
2752 page
= virt_to_head_page(ptr
);
2753 if (PageSlab(page
)) {
2754 cachep
= page
->slab_cache
;
2755 if (!is_root_cache(cachep
))
2756 memcg
= cachep
->memcg_params
.memcg
;
2758 /* page allocated by alloc_kmem_pages */
2759 memcg
= page
->mem_cgroup
;
2763 #endif /* CONFIG_MEMCG_KMEM */
2765 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2768 * Because tail pages are not marked as "used", set it. We're under
2769 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2770 * charge/uncharge will be never happen and move_account() is done under
2771 * compound_lock(), so we don't have to take care of races.
2773 void mem_cgroup_split_huge_fixup(struct page
*head
)
2777 if (mem_cgroup_disabled())
2780 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2781 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2783 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2786 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2788 #ifdef CONFIG_MEMCG_SWAP
2789 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2792 int val
= (charge
) ? 1 : -1;
2793 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2797 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2798 * @entry: swap entry to be moved
2799 * @from: mem_cgroup which the entry is moved from
2800 * @to: mem_cgroup which the entry is moved to
2802 * It succeeds only when the swap_cgroup's record for this entry is the same
2803 * as the mem_cgroup's id of @from.
2805 * Returns 0 on success, -EINVAL on failure.
2807 * The caller must have charged to @to, IOW, called page_counter_charge() about
2808 * both res and memsw, and called css_get().
2810 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2811 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2813 unsigned short old_id
, new_id
;
2815 old_id
= mem_cgroup_id(from
);
2816 new_id
= mem_cgroup_id(to
);
2818 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2819 mem_cgroup_swap_statistics(from
, false);
2820 mem_cgroup_swap_statistics(to
, true);
2826 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2827 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2833 static DEFINE_MUTEX(memcg_limit_mutex
);
2835 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2836 unsigned long limit
)
2838 unsigned long curusage
;
2839 unsigned long oldusage
;
2840 bool enlarge
= false;
2845 * For keeping hierarchical_reclaim simple, how long we should retry
2846 * is depends on callers. We set our retry-count to be function
2847 * of # of children which we should visit in this loop.
2849 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2850 mem_cgroup_count_children(memcg
);
2852 oldusage
= page_counter_read(&memcg
->memory
);
2855 if (signal_pending(current
)) {
2860 mutex_lock(&memcg_limit_mutex
);
2861 if (limit
> memcg
->memsw
.limit
) {
2862 mutex_unlock(&memcg_limit_mutex
);
2866 if (limit
> memcg
->memory
.limit
)
2868 ret
= page_counter_limit(&memcg
->memory
, limit
);
2869 mutex_unlock(&memcg_limit_mutex
);
2874 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2876 curusage
= page_counter_read(&memcg
->memory
);
2877 /* Usage is reduced ? */
2878 if (curusage
>= oldusage
)
2881 oldusage
= curusage
;
2882 } while (retry_count
);
2884 if (!ret
&& enlarge
)
2885 memcg_oom_recover(memcg
);
2890 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2891 unsigned long limit
)
2893 unsigned long curusage
;
2894 unsigned long oldusage
;
2895 bool enlarge
= false;
2899 /* see mem_cgroup_resize_res_limit */
2900 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2901 mem_cgroup_count_children(memcg
);
2903 oldusage
= page_counter_read(&memcg
->memsw
);
2906 if (signal_pending(current
)) {
2911 mutex_lock(&memcg_limit_mutex
);
2912 if (limit
< memcg
->memory
.limit
) {
2913 mutex_unlock(&memcg_limit_mutex
);
2917 if (limit
> memcg
->memsw
.limit
)
2919 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2920 mutex_unlock(&memcg_limit_mutex
);
2925 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2927 curusage
= page_counter_read(&memcg
->memsw
);
2928 /* Usage is reduced ? */
2929 if (curusage
>= oldusage
)
2932 oldusage
= curusage
;
2933 } while (retry_count
);
2935 if (!ret
&& enlarge
)
2936 memcg_oom_recover(memcg
);
2941 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2943 unsigned long *total_scanned
)
2945 unsigned long nr_reclaimed
= 0;
2946 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2947 unsigned long reclaimed
;
2949 struct mem_cgroup_tree_per_zone
*mctz
;
2950 unsigned long excess
;
2951 unsigned long nr_scanned
;
2956 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2958 * This loop can run a while, specially if mem_cgroup's continuously
2959 * keep exceeding their soft limit and putting the system under
2966 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2971 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2972 gfp_mask
, &nr_scanned
);
2973 nr_reclaimed
+= reclaimed
;
2974 *total_scanned
+= nr_scanned
;
2975 spin_lock_irq(&mctz
->lock
);
2976 __mem_cgroup_remove_exceeded(mz
, mctz
);
2979 * If we failed to reclaim anything from this memory cgroup
2980 * it is time to move on to the next cgroup
2984 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2986 excess
= soft_limit_excess(mz
->memcg
);
2988 * One school of thought says that we should not add
2989 * back the node to the tree if reclaim returns 0.
2990 * But our reclaim could return 0, simply because due
2991 * to priority we are exposing a smaller subset of
2992 * memory to reclaim from. Consider this as a longer
2995 /* If excess == 0, no tree ops */
2996 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2997 spin_unlock_irq(&mctz
->lock
);
2998 css_put(&mz
->memcg
->css
);
3001 * Could not reclaim anything and there are no more
3002 * mem cgroups to try or we seem to be looping without
3003 * reclaiming anything.
3005 if (!nr_reclaimed
&&
3007 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3009 } while (!nr_reclaimed
);
3011 css_put(&next_mz
->memcg
->css
);
3012 return nr_reclaimed
;
3016 * Test whether @memcg has children, dead or alive. Note that this
3017 * function doesn't care whether @memcg has use_hierarchy enabled and
3018 * returns %true if there are child csses according to the cgroup
3019 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3021 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3026 * The lock does not prevent addition or deletion of children, but
3027 * it prevents a new child from being initialized based on this
3028 * parent in css_online(), so it's enough to decide whether
3029 * hierarchically inherited attributes can still be changed or not.
3031 lockdep_assert_held(&memcg_create_mutex
);
3034 ret
= css_next_child(NULL
, &memcg
->css
);
3040 * Reclaims as many pages from the given memcg as possible and moves
3041 * the rest to the parent.
3043 * Caller is responsible for holding css reference for memcg.
3045 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
3047 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3049 /* we call try-to-free pages for make this cgroup empty */
3050 lru_add_drain_all();
3051 /* try to free all pages in this cgroup */
3052 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
3055 if (signal_pending(current
))
3058 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
3062 /* maybe some writeback is necessary */
3063 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3071 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
3072 char *buf
, size_t nbytes
,
3075 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3077 if (mem_cgroup_is_root(memcg
))
3079 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
3082 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
3085 return mem_cgroup_from_css(css
)->use_hierarchy
;
3088 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
3089 struct cftype
*cft
, u64 val
)
3092 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3093 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
3095 mutex_lock(&memcg_create_mutex
);
3097 if (memcg
->use_hierarchy
== val
)
3101 * If parent's use_hierarchy is set, we can't make any modifications
3102 * in the child subtrees. If it is unset, then the change can
3103 * occur, provided the current cgroup has no children.
3105 * For the root cgroup, parent_mem is NULL, we allow value to be
3106 * set if there are no children.
3108 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
3109 (val
== 1 || val
== 0)) {
3110 if (!memcg_has_children(memcg
))
3111 memcg
->use_hierarchy
= val
;
3118 mutex_unlock(&memcg_create_mutex
);
3123 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
3124 enum mem_cgroup_stat_index idx
)
3126 struct mem_cgroup
*iter
;
3129 /* Per-cpu values can be negative, use a signed accumulator */
3130 for_each_mem_cgroup_tree(iter
, memcg
)
3131 val
+= mem_cgroup_read_stat(iter
, idx
);
3133 if (val
< 0) /* race ? */
3138 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
3142 if (mem_cgroup_is_root(memcg
)) {
3143 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
3144 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
3146 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
3149 val
= page_counter_read(&memcg
->memory
);
3151 val
= page_counter_read(&memcg
->memsw
);
3153 return val
<< PAGE_SHIFT
;
3164 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
3167 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3168 struct page_counter
*counter
;
3170 switch (MEMFILE_TYPE(cft
->private)) {
3172 counter
= &memcg
->memory
;
3175 counter
= &memcg
->memsw
;
3178 counter
= &memcg
->kmem
;
3184 switch (MEMFILE_ATTR(cft
->private)) {
3186 if (counter
== &memcg
->memory
)
3187 return mem_cgroup_usage(memcg
, false);
3188 if (counter
== &memcg
->memsw
)
3189 return mem_cgroup_usage(memcg
, true);
3190 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
3192 return (u64
)counter
->limit
* PAGE_SIZE
;
3194 return (u64
)counter
->watermark
* PAGE_SIZE
;
3196 return counter
->failcnt
;
3197 case RES_SOFT_LIMIT
:
3198 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
3204 #ifdef CONFIG_MEMCG_KMEM
3205 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
3206 unsigned long nr_pages
)
3211 BUG_ON(memcg
->kmemcg_id
>= 0);
3212 BUG_ON(memcg
->kmem_acct_activated
);
3213 BUG_ON(memcg
->kmem_acct_active
);
3216 * For simplicity, we won't allow this to be disabled. It also can't
3217 * be changed if the cgroup has children already, or if tasks had
3220 * If tasks join before we set the limit, a person looking at
3221 * kmem.usage_in_bytes will have no way to determine when it took
3222 * place, which makes the value quite meaningless.
3224 * After it first became limited, changes in the value of the limit are
3225 * of course permitted.
3227 mutex_lock(&memcg_create_mutex
);
3228 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
3229 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
3231 mutex_unlock(&memcg_create_mutex
);
3235 memcg_id
= memcg_alloc_cache_id();
3242 * We couldn't have accounted to this cgroup, because it hasn't got
3243 * activated yet, so this should succeed.
3245 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
3248 static_key_slow_inc(&memcg_kmem_enabled_key
);
3250 * A memory cgroup is considered kmem-active as soon as it gets
3251 * kmemcg_id. Setting the id after enabling static branching will
3252 * guarantee no one starts accounting before all call sites are
3255 memcg
->kmemcg_id
= memcg_id
;
3256 memcg
->kmem_acct_activated
= true;
3257 memcg
->kmem_acct_active
= true;
3262 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3263 unsigned long limit
)
3267 mutex_lock(&memcg_limit_mutex
);
3268 if (!memcg_kmem_is_active(memcg
))
3269 ret
= memcg_activate_kmem(memcg
, limit
);
3271 ret
= page_counter_limit(&memcg
->kmem
, limit
);
3272 mutex_unlock(&memcg_limit_mutex
);
3276 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
3279 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
3284 mutex_lock(&memcg_limit_mutex
);
3286 * If the parent cgroup is not kmem-active now, it cannot be activated
3287 * after this point, because it has at least one child already.
3289 if (memcg_kmem_is_active(parent
))
3290 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
3291 mutex_unlock(&memcg_limit_mutex
);
3295 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
3296 unsigned long limit
)
3300 #endif /* CONFIG_MEMCG_KMEM */
3303 * The user of this function is...
3306 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3307 char *buf
, size_t nbytes
, loff_t off
)
3309 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3310 unsigned long nr_pages
;
3313 buf
= strstrip(buf
);
3314 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3318 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3320 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3324 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3326 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3329 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3332 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3336 case RES_SOFT_LIMIT
:
3337 memcg
->soft_limit
= nr_pages
;
3341 return ret
?: nbytes
;
3344 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3345 size_t nbytes
, loff_t off
)
3347 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3348 struct page_counter
*counter
;
3350 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3352 counter
= &memcg
->memory
;
3355 counter
= &memcg
->memsw
;
3358 counter
= &memcg
->kmem
;
3364 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3366 page_counter_reset_watermark(counter
);
3369 counter
->failcnt
= 0;
3378 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3381 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3385 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3386 struct cftype
*cft
, u64 val
)
3388 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3390 if (val
& ~MOVE_MASK
)
3394 * No kind of locking is needed in here, because ->can_attach() will
3395 * check this value once in the beginning of the process, and then carry
3396 * on with stale data. This means that changes to this value will only
3397 * affect task migrations starting after the change.
3399 memcg
->move_charge_at_immigrate
= val
;
3403 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3404 struct cftype
*cft
, u64 val
)
3411 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3415 unsigned int lru_mask
;
3418 static const struct numa_stat stats
[] = {
3419 { "total", LRU_ALL
},
3420 { "file", LRU_ALL_FILE
},
3421 { "anon", LRU_ALL_ANON
},
3422 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3424 const struct numa_stat
*stat
;
3427 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3429 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3430 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3431 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3432 for_each_node_state(nid
, N_MEMORY
) {
3433 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3435 seq_printf(m
, " N%d=%lu", nid
, nr
);
3440 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3441 struct mem_cgroup
*iter
;
3444 for_each_mem_cgroup_tree(iter
, memcg
)
3445 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3446 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3447 for_each_node_state(nid
, N_MEMORY
) {
3449 for_each_mem_cgroup_tree(iter
, memcg
)
3450 nr
+= mem_cgroup_node_nr_lru_pages(
3451 iter
, nid
, stat
->lru_mask
);
3452 seq_printf(m
, " N%d=%lu", nid
, nr
);
3459 #endif /* CONFIG_NUMA */
3461 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3463 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3464 unsigned long memory
, memsw
;
3465 struct mem_cgroup
*mi
;
3468 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3469 MEM_CGROUP_STAT_NSTATS
);
3470 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3471 MEM_CGROUP_EVENTS_NSTATS
);
3472 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3474 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3475 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3477 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
3478 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3481 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3482 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3483 mem_cgroup_read_events(memcg
, i
));
3485 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3486 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3487 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3489 /* Hierarchical information */
3490 memory
= memsw
= PAGE_COUNTER_MAX
;
3491 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3492 memory
= min(memory
, mi
->memory
.limit
);
3493 memsw
= min(memsw
, mi
->memsw
.limit
);
3495 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3496 (u64
)memory
* PAGE_SIZE
);
3497 if (do_swap_account
)
3498 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3499 (u64
)memsw
* PAGE_SIZE
);
3501 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3504 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3506 for_each_mem_cgroup_tree(mi
, memcg
)
3507 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3508 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
3511 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3512 unsigned long long val
= 0;
3514 for_each_mem_cgroup_tree(mi
, memcg
)
3515 val
+= mem_cgroup_read_events(mi
, i
);
3516 seq_printf(m
, "total_%s %llu\n",
3517 mem_cgroup_events_names
[i
], val
);
3520 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3521 unsigned long long val
= 0;
3523 for_each_mem_cgroup_tree(mi
, memcg
)
3524 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3525 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3528 #ifdef CONFIG_DEBUG_VM
3531 struct mem_cgroup_per_zone
*mz
;
3532 struct zone_reclaim_stat
*rstat
;
3533 unsigned long recent_rotated
[2] = {0, 0};
3534 unsigned long recent_scanned
[2] = {0, 0};
3536 for_each_online_node(nid
)
3537 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3538 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3539 rstat
= &mz
->lruvec
.reclaim_stat
;
3541 recent_rotated
[0] += rstat
->recent_rotated
[0];
3542 recent_rotated
[1] += rstat
->recent_rotated
[1];
3543 recent_scanned
[0] += rstat
->recent_scanned
[0];
3544 recent_scanned
[1] += rstat
->recent_scanned
[1];
3546 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3547 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3548 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3549 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3556 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3559 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3561 return mem_cgroup_swappiness(memcg
);
3564 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3565 struct cftype
*cft
, u64 val
)
3567 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3573 memcg
->swappiness
= val
;
3575 vm_swappiness
= val
;
3580 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3582 struct mem_cgroup_threshold_ary
*t
;
3583 unsigned long usage
;
3588 t
= rcu_dereference(memcg
->thresholds
.primary
);
3590 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3595 usage
= mem_cgroup_usage(memcg
, swap
);
3598 * current_threshold points to threshold just below or equal to usage.
3599 * If it's not true, a threshold was crossed after last
3600 * call of __mem_cgroup_threshold().
3602 i
= t
->current_threshold
;
3605 * Iterate backward over array of thresholds starting from
3606 * current_threshold and check if a threshold is crossed.
3607 * If none of thresholds below usage is crossed, we read
3608 * only one element of the array here.
3610 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3611 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3613 /* i = current_threshold + 1 */
3617 * Iterate forward over array of thresholds starting from
3618 * current_threshold+1 and check if a threshold is crossed.
3619 * If none of thresholds above usage is crossed, we read
3620 * only one element of the array here.
3622 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3623 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3625 /* Update current_threshold */
3626 t
->current_threshold
= i
- 1;
3631 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3634 __mem_cgroup_threshold(memcg
, false);
3635 if (do_swap_account
)
3636 __mem_cgroup_threshold(memcg
, true);
3638 memcg
= parent_mem_cgroup(memcg
);
3642 static int compare_thresholds(const void *a
, const void *b
)
3644 const struct mem_cgroup_threshold
*_a
= a
;
3645 const struct mem_cgroup_threshold
*_b
= b
;
3647 if (_a
->threshold
> _b
->threshold
)
3650 if (_a
->threshold
< _b
->threshold
)
3656 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3658 struct mem_cgroup_eventfd_list
*ev
;
3660 spin_lock(&memcg_oom_lock
);
3662 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3663 eventfd_signal(ev
->eventfd
, 1);
3665 spin_unlock(&memcg_oom_lock
);
3669 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3671 struct mem_cgroup
*iter
;
3673 for_each_mem_cgroup_tree(iter
, memcg
)
3674 mem_cgroup_oom_notify_cb(iter
);
3677 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3678 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3680 struct mem_cgroup_thresholds
*thresholds
;
3681 struct mem_cgroup_threshold_ary
*new;
3682 unsigned long threshold
;
3683 unsigned long usage
;
3686 ret
= page_counter_memparse(args
, "-1", &threshold
);
3690 mutex_lock(&memcg
->thresholds_lock
);
3693 thresholds
= &memcg
->thresholds
;
3694 usage
= mem_cgroup_usage(memcg
, false);
3695 } else if (type
== _MEMSWAP
) {
3696 thresholds
= &memcg
->memsw_thresholds
;
3697 usage
= mem_cgroup_usage(memcg
, true);
3701 /* Check if a threshold crossed before adding a new one */
3702 if (thresholds
->primary
)
3703 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3705 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3707 /* Allocate memory for new array of thresholds */
3708 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3716 /* Copy thresholds (if any) to new array */
3717 if (thresholds
->primary
) {
3718 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3719 sizeof(struct mem_cgroup_threshold
));
3722 /* Add new threshold */
3723 new->entries
[size
- 1].eventfd
= eventfd
;
3724 new->entries
[size
- 1].threshold
= threshold
;
3726 /* Sort thresholds. Registering of new threshold isn't time-critical */
3727 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3728 compare_thresholds
, NULL
);
3730 /* Find current threshold */
3731 new->current_threshold
= -1;
3732 for (i
= 0; i
< size
; i
++) {
3733 if (new->entries
[i
].threshold
<= usage
) {
3735 * new->current_threshold will not be used until
3736 * rcu_assign_pointer(), so it's safe to increment
3739 ++new->current_threshold
;
3744 /* Free old spare buffer and save old primary buffer as spare */
3745 kfree(thresholds
->spare
);
3746 thresholds
->spare
= thresholds
->primary
;
3748 rcu_assign_pointer(thresholds
->primary
, new);
3750 /* To be sure that nobody uses thresholds */
3754 mutex_unlock(&memcg
->thresholds_lock
);
3759 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3760 struct eventfd_ctx
*eventfd
, const char *args
)
3762 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3765 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3766 struct eventfd_ctx
*eventfd
, const char *args
)
3768 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3771 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3772 struct eventfd_ctx
*eventfd
, enum res_type type
)
3774 struct mem_cgroup_thresholds
*thresholds
;
3775 struct mem_cgroup_threshold_ary
*new;
3776 unsigned long usage
;
3779 mutex_lock(&memcg
->thresholds_lock
);
3782 thresholds
= &memcg
->thresholds
;
3783 usage
= mem_cgroup_usage(memcg
, false);
3784 } else if (type
== _MEMSWAP
) {
3785 thresholds
= &memcg
->memsw_thresholds
;
3786 usage
= mem_cgroup_usage(memcg
, true);
3790 if (!thresholds
->primary
)
3793 /* Check if a threshold crossed before removing */
3794 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3796 /* Calculate new number of threshold */
3798 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3799 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3803 new = thresholds
->spare
;
3805 /* Set thresholds array to NULL if we don't have thresholds */
3814 /* Copy thresholds and find current threshold */
3815 new->current_threshold
= -1;
3816 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3817 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3820 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3821 if (new->entries
[j
].threshold
<= usage
) {
3823 * new->current_threshold will not be used
3824 * until rcu_assign_pointer(), so it's safe to increment
3827 ++new->current_threshold
;
3833 /* Swap primary and spare array */
3834 thresholds
->spare
= thresholds
->primary
;
3835 /* If all events are unregistered, free the spare array */
3837 kfree(thresholds
->spare
);
3838 thresholds
->spare
= NULL
;
3841 rcu_assign_pointer(thresholds
->primary
, new);
3843 /* To be sure that nobody uses thresholds */
3846 mutex_unlock(&memcg
->thresholds_lock
);
3849 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3850 struct eventfd_ctx
*eventfd
)
3852 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3855 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3856 struct eventfd_ctx
*eventfd
)
3858 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3861 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3862 struct eventfd_ctx
*eventfd
, const char *args
)
3864 struct mem_cgroup_eventfd_list
*event
;
3866 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3870 spin_lock(&memcg_oom_lock
);
3872 event
->eventfd
= eventfd
;
3873 list_add(&event
->list
, &memcg
->oom_notify
);
3875 /* already in OOM ? */
3876 if (memcg
->under_oom
)
3877 eventfd_signal(eventfd
, 1);
3878 spin_unlock(&memcg_oom_lock
);
3883 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3884 struct eventfd_ctx
*eventfd
)
3886 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3888 spin_lock(&memcg_oom_lock
);
3890 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3891 if (ev
->eventfd
== eventfd
) {
3892 list_del(&ev
->list
);
3897 spin_unlock(&memcg_oom_lock
);
3900 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3902 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3904 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3905 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3909 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3910 struct cftype
*cft
, u64 val
)
3912 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3914 /* cannot set to root cgroup and only 0 and 1 are allowed */
3915 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3918 memcg
->oom_kill_disable
= val
;
3920 memcg_oom_recover(memcg
);
3925 #ifdef CONFIG_MEMCG_KMEM
3926 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3930 ret
= memcg_propagate_kmem(memcg
);
3934 return mem_cgroup_sockets_init(memcg
, ss
);
3937 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3939 struct cgroup_subsys_state
*css
;
3940 struct mem_cgroup
*parent
, *child
;
3943 if (!memcg
->kmem_acct_active
)
3947 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3948 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3949 * guarantees no cache will be created for this cgroup after we are
3950 * done (see memcg_create_kmem_cache()).
3952 memcg
->kmem_acct_active
= false;
3954 memcg_deactivate_kmem_caches(memcg
);
3956 kmemcg_id
= memcg
->kmemcg_id
;
3957 BUG_ON(kmemcg_id
< 0);
3959 parent
= parent_mem_cgroup(memcg
);
3961 parent
= root_mem_cgroup
;
3964 * Change kmemcg_id of this cgroup and all its descendants to the
3965 * parent's id, and then move all entries from this cgroup's list_lrus
3966 * to ones of the parent. After we have finished, all list_lrus
3967 * corresponding to this cgroup are guaranteed to remain empty. The
3968 * ordering is imposed by list_lru_node->lock taken by
3969 * memcg_drain_all_list_lrus().
3971 css_for_each_descendant_pre(css
, &memcg
->css
) {
3972 child
= mem_cgroup_from_css(css
);
3973 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3974 child
->kmemcg_id
= parent
->kmemcg_id
;
3975 if (!memcg
->use_hierarchy
)
3978 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
3980 memcg_free_cache_id(kmemcg_id
);
3983 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3985 if (memcg
->kmem_acct_activated
) {
3986 memcg_destroy_kmem_caches(memcg
);
3987 static_key_slow_dec(&memcg_kmem_enabled_key
);
3988 WARN_ON(page_counter_read(&memcg
->kmem
));
3990 mem_cgroup_sockets_destroy(memcg
);
3993 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3998 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
4002 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4007 #ifdef CONFIG_CGROUP_WRITEBACK
4009 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
4011 return &memcg
->cgwb_list
;
4014 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4016 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
4019 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4021 wb_domain_exit(&memcg
->cgwb_domain
);
4024 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4026 wb_domain_size_changed(&memcg
->cgwb_domain
);
4029 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
4031 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4033 if (!memcg
->css
.parent
)
4036 return &memcg
->cgwb_domain
;
4040 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4041 * @wb: bdi_writeback in question
4042 * @pavail: out parameter for number of available pages
4043 * @pdirty: out parameter for number of dirty pages
4044 * @pwriteback: out parameter for number of pages under writeback
4046 * Determine the numbers of available, dirty, and writeback pages in @wb's
4047 * memcg. Dirty and writeback are self-explanatory. Available is a bit
4050 * A memcg's headroom is "min(max, high) - used". The available memory is
4051 * calculated as the lowest headroom of itself and the ancestors plus the
4052 * number of pages already being used for file pages. Note that this
4053 * doesn't consider the actual amount of available memory in the system.
4054 * The caller should further cap *@pavail accordingly.
4056 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pavail
,
4057 unsigned long *pdirty
, unsigned long *pwriteback
)
4059 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
4060 struct mem_cgroup
*parent
;
4061 unsigned long head_room
= PAGE_COUNTER_MAX
;
4062 unsigned long file_pages
;
4064 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
4066 /* this should eventually include NR_UNSTABLE_NFS */
4067 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
4069 file_pages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
4070 (1 << LRU_ACTIVE_FILE
));
4071 while ((parent
= parent_mem_cgroup(memcg
))) {
4072 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
4073 unsigned long used
= page_counter_read(&memcg
->memory
);
4075 head_room
= min(head_room
, ceiling
- min(ceiling
, used
));
4079 *pavail
= file_pages
+ head_room
;
4082 #else /* CONFIG_CGROUP_WRITEBACK */
4084 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
4089 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
4093 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
4097 #endif /* CONFIG_CGROUP_WRITEBACK */
4100 * DO NOT USE IN NEW FILES.
4102 * "cgroup.event_control" implementation.
4104 * This is way over-engineered. It tries to support fully configurable
4105 * events for each user. Such level of flexibility is completely
4106 * unnecessary especially in the light of the planned unified hierarchy.
4108 * Please deprecate this and replace with something simpler if at all
4113 * Unregister event and free resources.
4115 * Gets called from workqueue.
4117 static void memcg_event_remove(struct work_struct
*work
)
4119 struct mem_cgroup_event
*event
=
4120 container_of(work
, struct mem_cgroup_event
, remove
);
4121 struct mem_cgroup
*memcg
= event
->memcg
;
4123 remove_wait_queue(event
->wqh
, &event
->wait
);
4125 event
->unregister_event(memcg
, event
->eventfd
);
4127 /* Notify userspace the event is going away. */
4128 eventfd_signal(event
->eventfd
, 1);
4130 eventfd_ctx_put(event
->eventfd
);
4132 css_put(&memcg
->css
);
4136 * Gets called on POLLHUP on eventfd when user closes it.
4138 * Called with wqh->lock held and interrupts disabled.
4140 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
4141 int sync
, void *key
)
4143 struct mem_cgroup_event
*event
=
4144 container_of(wait
, struct mem_cgroup_event
, wait
);
4145 struct mem_cgroup
*memcg
= event
->memcg
;
4146 unsigned long flags
= (unsigned long)key
;
4148 if (flags
& POLLHUP
) {
4150 * If the event has been detached at cgroup removal, we
4151 * can simply return knowing the other side will cleanup
4154 * We can't race against event freeing since the other
4155 * side will require wqh->lock via remove_wait_queue(),
4158 spin_lock(&memcg
->event_list_lock
);
4159 if (!list_empty(&event
->list
)) {
4160 list_del_init(&event
->list
);
4162 * We are in atomic context, but cgroup_event_remove()
4163 * may sleep, so we have to call it in workqueue.
4165 schedule_work(&event
->remove
);
4167 spin_unlock(&memcg
->event_list_lock
);
4173 static void memcg_event_ptable_queue_proc(struct file
*file
,
4174 wait_queue_head_t
*wqh
, poll_table
*pt
)
4176 struct mem_cgroup_event
*event
=
4177 container_of(pt
, struct mem_cgroup_event
, pt
);
4180 add_wait_queue(wqh
, &event
->wait
);
4184 * DO NOT USE IN NEW FILES.
4186 * Parse input and register new cgroup event handler.
4188 * Input must be in format '<event_fd> <control_fd> <args>'.
4189 * Interpretation of args is defined by control file implementation.
4191 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
4192 char *buf
, size_t nbytes
, loff_t off
)
4194 struct cgroup_subsys_state
*css
= of_css(of
);
4195 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4196 struct mem_cgroup_event
*event
;
4197 struct cgroup_subsys_state
*cfile_css
;
4198 unsigned int efd
, cfd
;
4205 buf
= strstrip(buf
);
4207 efd
= simple_strtoul(buf
, &endp
, 10);
4212 cfd
= simple_strtoul(buf
, &endp
, 10);
4213 if ((*endp
!= ' ') && (*endp
!= '\0'))
4217 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
4221 event
->memcg
= memcg
;
4222 INIT_LIST_HEAD(&event
->list
);
4223 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
4224 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
4225 INIT_WORK(&event
->remove
, memcg_event_remove
);
4233 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
4234 if (IS_ERR(event
->eventfd
)) {
4235 ret
= PTR_ERR(event
->eventfd
);
4242 goto out_put_eventfd
;
4245 /* the process need read permission on control file */
4246 /* AV: shouldn't we check that it's been opened for read instead? */
4247 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
4252 * Determine the event callbacks and set them in @event. This used
4253 * to be done via struct cftype but cgroup core no longer knows
4254 * about these events. The following is crude but the whole thing
4255 * is for compatibility anyway.
4257 * DO NOT ADD NEW FILES.
4259 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
4261 if (!strcmp(name
, "memory.usage_in_bytes")) {
4262 event
->register_event
= mem_cgroup_usage_register_event
;
4263 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
4264 } else if (!strcmp(name
, "memory.oom_control")) {
4265 event
->register_event
= mem_cgroup_oom_register_event
;
4266 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
4267 } else if (!strcmp(name
, "memory.pressure_level")) {
4268 event
->register_event
= vmpressure_register_event
;
4269 event
->unregister_event
= vmpressure_unregister_event
;
4270 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
4271 event
->register_event
= memsw_cgroup_usage_register_event
;
4272 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
4279 * Verify @cfile should belong to @css. Also, remaining events are
4280 * automatically removed on cgroup destruction but the removal is
4281 * asynchronous, so take an extra ref on @css.
4283 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
4284 &memory_cgrp_subsys
);
4286 if (IS_ERR(cfile_css
))
4288 if (cfile_css
!= css
) {
4293 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
4297 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
4299 spin_lock(&memcg
->event_list_lock
);
4300 list_add(&event
->list
, &memcg
->event_list
);
4301 spin_unlock(&memcg
->event_list_lock
);
4313 eventfd_ctx_put(event
->eventfd
);
4322 static struct cftype mem_cgroup_legacy_files
[] = {
4324 .name
= "usage_in_bytes",
4325 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4326 .read_u64
= mem_cgroup_read_u64
,
4329 .name
= "max_usage_in_bytes",
4330 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4331 .write
= mem_cgroup_reset
,
4332 .read_u64
= mem_cgroup_read_u64
,
4335 .name
= "limit_in_bytes",
4336 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4337 .write
= mem_cgroup_write
,
4338 .read_u64
= mem_cgroup_read_u64
,
4341 .name
= "soft_limit_in_bytes",
4342 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4343 .write
= mem_cgroup_write
,
4344 .read_u64
= mem_cgroup_read_u64
,
4348 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4349 .write
= mem_cgroup_reset
,
4350 .read_u64
= mem_cgroup_read_u64
,
4354 .seq_show
= memcg_stat_show
,
4357 .name
= "force_empty",
4358 .write
= mem_cgroup_force_empty_write
,
4361 .name
= "use_hierarchy",
4362 .write_u64
= mem_cgroup_hierarchy_write
,
4363 .read_u64
= mem_cgroup_hierarchy_read
,
4366 .name
= "cgroup.event_control", /* XXX: for compat */
4367 .write
= memcg_write_event_control
,
4368 .flags
= CFTYPE_NO_PREFIX
,
4372 .name
= "swappiness",
4373 .read_u64
= mem_cgroup_swappiness_read
,
4374 .write_u64
= mem_cgroup_swappiness_write
,
4377 .name
= "move_charge_at_immigrate",
4378 .read_u64
= mem_cgroup_move_charge_read
,
4379 .write_u64
= mem_cgroup_move_charge_write
,
4382 .name
= "oom_control",
4383 .seq_show
= mem_cgroup_oom_control_read
,
4384 .write_u64
= mem_cgroup_oom_control_write
,
4385 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4388 .name
= "pressure_level",
4392 .name
= "numa_stat",
4393 .seq_show
= memcg_numa_stat_show
,
4396 #ifdef CONFIG_MEMCG_KMEM
4398 .name
= "kmem.limit_in_bytes",
4399 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4400 .write
= mem_cgroup_write
,
4401 .read_u64
= mem_cgroup_read_u64
,
4404 .name
= "kmem.usage_in_bytes",
4405 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4406 .read_u64
= mem_cgroup_read_u64
,
4409 .name
= "kmem.failcnt",
4410 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4411 .write
= mem_cgroup_reset
,
4412 .read_u64
= mem_cgroup_read_u64
,
4415 .name
= "kmem.max_usage_in_bytes",
4416 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4417 .write
= mem_cgroup_reset
,
4418 .read_u64
= mem_cgroup_read_u64
,
4420 #ifdef CONFIG_SLABINFO
4422 .name
= "kmem.slabinfo",
4423 .seq_start
= slab_start
,
4424 .seq_next
= slab_next
,
4425 .seq_stop
= slab_stop
,
4426 .seq_show
= memcg_slab_show
,
4430 { }, /* terminate */
4433 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4435 struct mem_cgroup_per_node
*pn
;
4436 struct mem_cgroup_per_zone
*mz
;
4437 int zone
, tmp
= node
;
4439 * This routine is called against possible nodes.
4440 * But it's BUG to call kmalloc() against offline node.
4442 * TODO: this routine can waste much memory for nodes which will
4443 * never be onlined. It's better to use memory hotplug callback
4446 if (!node_state(node
, N_NORMAL_MEMORY
))
4448 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4452 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4453 mz
= &pn
->zoneinfo
[zone
];
4454 lruvec_init(&mz
->lruvec
);
4455 mz
->usage_in_excess
= 0;
4456 mz
->on_tree
= false;
4459 memcg
->nodeinfo
[node
] = pn
;
4463 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4465 kfree(memcg
->nodeinfo
[node
]);
4468 static struct mem_cgroup
*mem_cgroup_alloc(void)
4470 struct mem_cgroup
*memcg
;
4473 size
= sizeof(struct mem_cgroup
);
4474 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4476 memcg
= kzalloc(size
, GFP_KERNEL
);
4480 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4484 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4487 spin_lock_init(&memcg
->pcp_counter_lock
);
4491 free_percpu(memcg
->stat
);
4498 * At destroying mem_cgroup, references from swap_cgroup can remain.
4499 * (scanning all at force_empty is too costly...)
4501 * Instead of clearing all references at force_empty, we remember
4502 * the number of reference from swap_cgroup and free mem_cgroup when
4503 * it goes down to 0.
4505 * Removal of cgroup itself succeeds regardless of refs from swap.
4508 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4512 mem_cgroup_remove_from_trees(memcg
);
4515 free_mem_cgroup_per_zone_info(memcg
, node
);
4517 free_percpu(memcg
->stat
);
4518 memcg_wb_domain_exit(memcg
);
4523 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4525 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4527 if (!memcg
->memory
.parent
)
4529 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4531 EXPORT_SYMBOL(parent_mem_cgroup
);
4533 static struct cgroup_subsys_state
* __ref
4534 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4536 struct mem_cgroup
*memcg
;
4537 long error
= -ENOMEM
;
4540 memcg
= mem_cgroup_alloc();
4542 return ERR_PTR(error
);
4545 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4549 if (parent_css
== NULL
) {
4550 root_mem_cgroup
= memcg
;
4551 mem_cgroup_root_css
= &memcg
->css
;
4552 page_counter_init(&memcg
->memory
, NULL
);
4553 memcg
->high
= PAGE_COUNTER_MAX
;
4554 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4555 page_counter_init(&memcg
->memsw
, NULL
);
4556 page_counter_init(&memcg
->kmem
, NULL
);
4559 memcg
->last_scanned_node
= MAX_NUMNODES
;
4560 INIT_LIST_HEAD(&memcg
->oom_notify
);
4561 memcg
->move_charge_at_immigrate
= 0;
4562 mutex_init(&memcg
->thresholds_lock
);
4563 spin_lock_init(&memcg
->move_lock
);
4564 vmpressure_init(&memcg
->vmpressure
);
4565 INIT_LIST_HEAD(&memcg
->event_list
);
4566 spin_lock_init(&memcg
->event_list_lock
);
4567 #ifdef CONFIG_MEMCG_KMEM
4568 memcg
->kmemcg_id
= -1;
4570 #ifdef CONFIG_CGROUP_WRITEBACK
4571 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4576 __mem_cgroup_free(memcg
);
4577 return ERR_PTR(error
);
4581 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4583 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4584 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4587 if (css
->id
> MEM_CGROUP_ID_MAX
)
4593 mutex_lock(&memcg_create_mutex
);
4595 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4596 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4597 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4599 if (parent
->use_hierarchy
) {
4600 page_counter_init(&memcg
->memory
, &parent
->memory
);
4601 memcg
->high
= PAGE_COUNTER_MAX
;
4602 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4603 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4604 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4607 * No need to take a reference to the parent because cgroup
4608 * core guarantees its existence.
4611 page_counter_init(&memcg
->memory
, NULL
);
4612 memcg
->high
= PAGE_COUNTER_MAX
;
4613 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4614 page_counter_init(&memcg
->memsw
, NULL
);
4615 page_counter_init(&memcg
->kmem
, NULL
);
4617 * Deeper hierachy with use_hierarchy == false doesn't make
4618 * much sense so let cgroup subsystem know about this
4619 * unfortunate state in our controller.
4621 if (parent
!= root_mem_cgroup
)
4622 memory_cgrp_subsys
.broken_hierarchy
= true;
4624 mutex_unlock(&memcg_create_mutex
);
4626 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4631 * Make sure the memcg is initialized: mem_cgroup_iter()
4632 * orders reading memcg->initialized against its callers
4633 * reading the memcg members.
4635 smp_store_release(&memcg
->initialized
, 1);
4640 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4642 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4643 struct mem_cgroup_event
*event
, *tmp
;
4646 * Unregister events and notify userspace.
4647 * Notify userspace about cgroup removing only after rmdir of cgroup
4648 * directory to avoid race between userspace and kernelspace.
4650 spin_lock(&memcg
->event_list_lock
);
4651 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4652 list_del_init(&event
->list
);
4653 schedule_work(&event
->remove
);
4655 spin_unlock(&memcg
->event_list_lock
);
4657 vmpressure_cleanup(&memcg
->vmpressure
);
4659 memcg_deactivate_kmem(memcg
);
4661 wb_memcg_offline(memcg
);
4664 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4666 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4668 memcg_destroy_kmem(memcg
);
4669 __mem_cgroup_free(memcg
);
4673 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4674 * @css: the target css
4676 * Reset the states of the mem_cgroup associated with @css. This is
4677 * invoked when the userland requests disabling on the default hierarchy
4678 * but the memcg is pinned through dependency. The memcg should stop
4679 * applying policies and should revert to the vanilla state as it may be
4680 * made visible again.
4682 * The current implementation only resets the essential configurations.
4683 * This needs to be expanded to cover all the visible parts.
4685 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4687 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4689 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4690 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4691 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4693 memcg
->high
= PAGE_COUNTER_MAX
;
4694 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4695 memcg_wb_domain_size_changed(memcg
);
4699 /* Handlers for move charge at task migration. */
4700 static int mem_cgroup_do_precharge(unsigned long count
)
4704 /* Try a single bulk charge without reclaim first */
4705 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
4707 mc
.precharge
+= count
;
4710 if (ret
== -EINTR
) {
4711 cancel_charge(root_mem_cgroup
, count
);
4715 /* Try charges one by one with reclaim */
4717 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4719 * In case of failure, any residual charges against
4720 * mc.to will be dropped by mem_cgroup_clear_mc()
4721 * later on. However, cancel any charges that are
4722 * bypassed to root right away or they'll be lost.
4725 cancel_charge(root_mem_cgroup
, 1);
4735 * get_mctgt_type - get target type of moving charge
4736 * @vma: the vma the pte to be checked belongs
4737 * @addr: the address corresponding to the pte to be checked
4738 * @ptent: the pte to be checked
4739 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4742 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4743 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4744 * move charge. if @target is not NULL, the page is stored in target->page
4745 * with extra refcnt got(Callers should handle it).
4746 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4747 * target for charge migration. if @target is not NULL, the entry is stored
4750 * Called with pte lock held.
4757 enum mc_target_type
{
4763 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4764 unsigned long addr
, pte_t ptent
)
4766 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4768 if (!page
|| !page_mapped(page
))
4770 if (PageAnon(page
)) {
4771 if (!(mc
.flags
& MOVE_ANON
))
4774 if (!(mc
.flags
& MOVE_FILE
))
4777 if (!get_page_unless_zero(page
))
4784 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4785 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4787 struct page
*page
= NULL
;
4788 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4790 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4793 * Because lookup_swap_cache() updates some statistics counter,
4794 * we call find_get_page() with swapper_space directly.
4796 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4797 if (do_swap_account
)
4798 entry
->val
= ent
.val
;
4803 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4804 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4810 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4811 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4813 struct page
*page
= NULL
;
4814 struct address_space
*mapping
;
4817 if (!vma
->vm_file
) /* anonymous vma */
4819 if (!(mc
.flags
& MOVE_FILE
))
4822 mapping
= vma
->vm_file
->f_mapping
;
4823 pgoff
= linear_page_index(vma
, addr
);
4825 /* page is moved even if it's not RSS of this task(page-faulted). */
4827 /* shmem/tmpfs may report page out on swap: account for that too. */
4828 if (shmem_mapping(mapping
)) {
4829 page
= find_get_entry(mapping
, pgoff
);
4830 if (radix_tree_exceptional_entry(page
)) {
4831 swp_entry_t swp
= radix_to_swp_entry(page
);
4832 if (do_swap_account
)
4834 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4837 page
= find_get_page(mapping
, pgoff
);
4839 page
= find_get_page(mapping
, pgoff
);
4845 * mem_cgroup_move_account - move account of the page
4847 * @nr_pages: number of regular pages (>1 for huge pages)
4848 * @from: mem_cgroup which the page is moved from.
4849 * @to: mem_cgroup which the page is moved to. @from != @to.
4851 * The caller must confirm following.
4852 * - page is not on LRU (isolate_page() is useful.)
4853 * - compound_lock is held when nr_pages > 1
4855 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4858 static int mem_cgroup_move_account(struct page
*page
,
4859 unsigned int nr_pages
,
4860 struct mem_cgroup
*from
,
4861 struct mem_cgroup
*to
)
4863 unsigned long flags
;
4867 VM_BUG_ON(from
== to
);
4868 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4870 * The page is isolated from LRU. So, collapse function
4871 * will not handle this page. But page splitting can happen.
4872 * Do this check under compound_page_lock(). The caller should
4876 if (nr_pages
> 1 && !PageTransHuge(page
))
4880 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4881 * of its source page while we change it: page migration takes
4882 * both pages off the LRU, but page cache replacement doesn't.
4884 if (!trylock_page(page
))
4888 if (page
->mem_cgroup
!= from
)
4891 anon
= PageAnon(page
);
4893 spin_lock_irqsave(&from
->move_lock
, flags
);
4895 if (!anon
&& page_mapped(page
)) {
4896 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4898 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4903 * move_lock grabbed above and caller set from->moving_account, so
4904 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4905 * So mapping should be stable for dirty pages.
4907 if (!anon
&& PageDirty(page
)) {
4908 struct address_space
*mapping
= page_mapping(page
);
4910 if (mapping_cap_account_dirty(mapping
)) {
4911 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4913 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4918 if (PageWriteback(page
)) {
4919 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4921 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4926 * It is safe to change page->mem_cgroup here because the page
4927 * is referenced, charged, and isolated - we can't race with
4928 * uncharging, charging, migration, or LRU putback.
4931 /* caller should have done css_get */
4932 page
->mem_cgroup
= to
;
4933 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4937 local_irq_disable();
4938 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
4939 memcg_check_events(to
, page
);
4940 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
4941 memcg_check_events(from
, page
);
4949 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4950 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4952 struct page
*page
= NULL
;
4953 enum mc_target_type ret
= MC_TARGET_NONE
;
4954 swp_entry_t ent
= { .val
= 0 };
4956 if (pte_present(ptent
))
4957 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4958 else if (is_swap_pte(ptent
))
4959 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4960 else if (pte_none(ptent
))
4961 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4963 if (!page
&& !ent
.val
)
4967 * Do only loose check w/o serialization.
4968 * mem_cgroup_move_account() checks the page is valid or
4969 * not under LRU exclusion.
4971 if (page
->mem_cgroup
== mc
.from
) {
4972 ret
= MC_TARGET_PAGE
;
4974 target
->page
= page
;
4976 if (!ret
|| !target
)
4979 /* There is a swap entry and a page doesn't exist or isn't charged */
4980 if (ent
.val
&& !ret
&&
4981 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4982 ret
= MC_TARGET_SWAP
;
4989 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4991 * We don't consider swapping or file mapped pages because THP does not
4992 * support them for now.
4993 * Caller should make sure that pmd_trans_huge(pmd) is true.
4995 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4996 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4998 struct page
*page
= NULL
;
4999 enum mc_target_type ret
= MC_TARGET_NONE
;
5001 page
= pmd_page(pmd
);
5002 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5003 if (!(mc
.flags
& MOVE_ANON
))
5005 if (page
->mem_cgroup
== mc
.from
) {
5006 ret
= MC_TARGET_PAGE
;
5009 target
->page
= page
;
5015 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5016 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5018 return MC_TARGET_NONE
;
5022 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5023 unsigned long addr
, unsigned long end
,
5024 struct mm_walk
*walk
)
5026 struct vm_area_struct
*vma
= walk
->vma
;
5030 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5031 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5032 mc
.precharge
+= HPAGE_PMD_NR
;
5037 if (pmd_trans_unstable(pmd
))
5039 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5040 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5041 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5042 mc
.precharge
++; /* increment precharge temporarily */
5043 pte_unmap_unlock(pte
- 1, ptl
);
5049 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5051 unsigned long precharge
;
5053 struct mm_walk mem_cgroup_count_precharge_walk
= {
5054 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5057 down_read(&mm
->mmap_sem
);
5058 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
5059 up_read(&mm
->mmap_sem
);
5061 precharge
= mc
.precharge
;
5067 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5069 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5071 VM_BUG_ON(mc
.moving_task
);
5072 mc
.moving_task
= current
;
5073 return mem_cgroup_do_precharge(precharge
);
5076 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5077 static void __mem_cgroup_clear_mc(void)
5079 struct mem_cgroup
*from
= mc
.from
;
5080 struct mem_cgroup
*to
= mc
.to
;
5082 /* we must uncharge all the leftover precharges from mc.to */
5084 cancel_charge(mc
.to
, mc
.precharge
);
5088 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5089 * we must uncharge here.
5091 if (mc
.moved_charge
) {
5092 cancel_charge(mc
.from
, mc
.moved_charge
);
5093 mc
.moved_charge
= 0;
5095 /* we must fixup refcnts and charges */
5096 if (mc
.moved_swap
) {
5097 /* uncharge swap account from the old cgroup */
5098 if (!mem_cgroup_is_root(mc
.from
))
5099 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
5102 * we charged both to->memory and to->memsw, so we
5103 * should uncharge to->memory.
5105 if (!mem_cgroup_is_root(mc
.to
))
5106 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
5108 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
5110 /* we've already done css_get(mc.to) */
5113 memcg_oom_recover(from
);
5114 memcg_oom_recover(to
);
5115 wake_up_all(&mc
.waitq
);
5118 static void mem_cgroup_clear_mc(void)
5121 * we must clear moving_task before waking up waiters at the end of
5124 mc
.moving_task
= NULL
;
5125 __mem_cgroup_clear_mc();
5126 spin_lock(&mc
.lock
);
5129 spin_unlock(&mc
.lock
);
5132 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5133 struct cgroup_taskset
*tset
)
5135 struct task_struct
*p
= cgroup_taskset_first(tset
);
5137 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5138 unsigned long move_flags
;
5141 * We are now commited to this value whatever it is. Changes in this
5142 * tunable will only affect upcoming migrations, not the current one.
5143 * So we need to save it, and keep it going.
5145 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
5147 struct mm_struct
*mm
;
5148 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
5150 VM_BUG_ON(from
== memcg
);
5152 mm
= get_task_mm(p
);
5155 /* We move charges only when we move a owner of the mm */
5156 if (mm
->owner
== p
) {
5159 VM_BUG_ON(mc
.precharge
);
5160 VM_BUG_ON(mc
.moved_charge
);
5161 VM_BUG_ON(mc
.moved_swap
);
5163 spin_lock(&mc
.lock
);
5166 mc
.flags
= move_flags
;
5167 spin_unlock(&mc
.lock
);
5168 /* We set mc.moving_task later */
5170 ret
= mem_cgroup_precharge_mc(mm
);
5172 mem_cgroup_clear_mc();
5179 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5180 struct cgroup_taskset
*tset
)
5183 mem_cgroup_clear_mc();
5186 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
5187 unsigned long addr
, unsigned long end
,
5188 struct mm_walk
*walk
)
5191 struct vm_area_struct
*vma
= walk
->vma
;
5194 enum mc_target_type target_type
;
5195 union mc_target target
;
5199 * We don't take compound_lock() here but no race with splitting thp
5201 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5202 * under splitting, which means there's no concurrent thp split,
5203 * - if another thread runs into split_huge_page() just after we
5204 * entered this if-block, the thread must wait for page table lock
5205 * to be unlocked in __split_huge_page_splitting(), where the main
5206 * part of thp split is not executed yet.
5208 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5209 if (mc
.precharge
< HPAGE_PMD_NR
) {
5213 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
5214 if (target_type
== MC_TARGET_PAGE
) {
5216 if (!isolate_lru_page(page
)) {
5217 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
5219 mc
.precharge
-= HPAGE_PMD_NR
;
5220 mc
.moved_charge
+= HPAGE_PMD_NR
;
5222 putback_lru_page(page
);
5230 if (pmd_trans_unstable(pmd
))
5233 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5234 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
5235 pte_t ptent
= *(pte
++);
5241 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
5242 case MC_TARGET_PAGE
:
5244 if (isolate_lru_page(page
))
5246 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
5248 /* we uncharge from mc.from later. */
5251 putback_lru_page(page
);
5252 put
: /* get_mctgt_type() gets the page */
5255 case MC_TARGET_SWAP
:
5257 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
5259 /* we fixup refcnts and charges later. */
5267 pte_unmap_unlock(pte
- 1, ptl
);
5272 * We have consumed all precharges we got in can_attach().
5273 * We try charge one by one, but don't do any additional
5274 * charges to mc.to if we have failed in charge once in attach()
5277 ret
= mem_cgroup_do_precharge(1);
5285 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
5287 struct mm_walk mem_cgroup_move_charge_walk
= {
5288 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5292 lru_add_drain_all();
5294 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5295 * move_lock while we're moving its pages to another memcg.
5296 * Then wait for already started RCU-only updates to finish.
5298 atomic_inc(&mc
.from
->moving_account
);
5301 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5303 * Someone who are holding the mmap_sem might be waiting in
5304 * waitq. So we cancel all extra charges, wake up all waiters,
5305 * and retry. Because we cancel precharges, we might not be able
5306 * to move enough charges, but moving charge is a best-effort
5307 * feature anyway, so it wouldn't be a big problem.
5309 __mem_cgroup_clear_mc();
5314 * When we have consumed all precharges and failed in doing
5315 * additional charge, the page walk just aborts.
5317 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5318 up_read(&mm
->mmap_sem
);
5319 atomic_dec(&mc
.from
->moving_account
);
5322 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5323 struct cgroup_taskset
*tset
)
5325 struct task_struct
*p
= cgroup_taskset_first(tset
);
5326 struct mm_struct
*mm
= get_task_mm(p
);
5330 mem_cgroup_move_charge(mm
);
5334 mem_cgroup_clear_mc();
5336 #else /* !CONFIG_MMU */
5337 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5338 struct cgroup_taskset
*tset
)
5342 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5343 struct cgroup_taskset
*tset
)
5346 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5347 struct cgroup_taskset
*tset
)
5353 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5354 * to verify whether we're attached to the default hierarchy on each mount
5357 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5360 * use_hierarchy is forced on the default hierarchy. cgroup core
5361 * guarantees that @root doesn't have any children, so turning it
5362 * on for the root memcg is enough.
5364 if (cgroup_on_dfl(root_css
->cgroup
))
5365 root_mem_cgroup
->use_hierarchy
= true;
5367 root_mem_cgroup
->use_hierarchy
= false;
5370 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5373 return mem_cgroup_usage(mem_cgroup_from_css(css
), false);
5376 static int memory_low_show(struct seq_file
*m
, void *v
)
5378 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5379 unsigned long low
= READ_ONCE(memcg
->low
);
5381 if (low
== PAGE_COUNTER_MAX
)
5382 seq_puts(m
, "max\n");
5384 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5389 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5390 char *buf
, size_t nbytes
, loff_t off
)
5392 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5396 buf
= strstrip(buf
);
5397 err
= page_counter_memparse(buf
, "max", &low
);
5406 static int memory_high_show(struct seq_file
*m
, void *v
)
5408 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5409 unsigned long high
= READ_ONCE(memcg
->high
);
5411 if (high
== PAGE_COUNTER_MAX
)
5412 seq_puts(m
, "max\n");
5414 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5419 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5420 char *buf
, size_t nbytes
, loff_t off
)
5422 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5426 buf
= strstrip(buf
);
5427 err
= page_counter_memparse(buf
, "max", &high
);
5433 memcg_wb_domain_size_changed(memcg
);
5437 static int memory_max_show(struct seq_file
*m
, void *v
)
5439 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5440 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5442 if (max
== PAGE_COUNTER_MAX
)
5443 seq_puts(m
, "max\n");
5445 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5450 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5451 char *buf
, size_t nbytes
, loff_t off
)
5453 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5457 buf
= strstrip(buf
);
5458 err
= page_counter_memparse(buf
, "max", &max
);
5462 err
= mem_cgroup_resize_limit(memcg
, max
);
5466 memcg_wb_domain_size_changed(memcg
);
5470 static int memory_events_show(struct seq_file
*m
, void *v
)
5472 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5474 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5475 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5476 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5477 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5482 static struct cftype memory_files
[] = {
5485 .read_u64
= memory_current_read
,
5489 .flags
= CFTYPE_NOT_ON_ROOT
,
5490 .seq_show
= memory_low_show
,
5491 .write
= memory_low_write
,
5495 .flags
= CFTYPE_NOT_ON_ROOT
,
5496 .seq_show
= memory_high_show
,
5497 .write
= memory_high_write
,
5501 .flags
= CFTYPE_NOT_ON_ROOT
,
5502 .seq_show
= memory_max_show
,
5503 .write
= memory_max_write
,
5507 .flags
= CFTYPE_NOT_ON_ROOT
,
5508 .seq_show
= memory_events_show
,
5513 struct cgroup_subsys memory_cgrp_subsys
= {
5514 .css_alloc
= mem_cgroup_css_alloc
,
5515 .css_online
= mem_cgroup_css_online
,
5516 .css_offline
= mem_cgroup_css_offline
,
5517 .css_free
= mem_cgroup_css_free
,
5518 .css_reset
= mem_cgroup_css_reset
,
5519 .can_attach
= mem_cgroup_can_attach
,
5520 .cancel_attach
= mem_cgroup_cancel_attach
,
5521 .attach
= mem_cgroup_move_task
,
5522 .bind
= mem_cgroup_bind
,
5523 .dfl_cftypes
= memory_files
,
5524 .legacy_cftypes
= mem_cgroup_legacy_files
,
5529 * mem_cgroup_events - count memory events against a cgroup
5530 * @memcg: the memory cgroup
5531 * @idx: the event index
5532 * @nr: the number of events to account for
5534 void mem_cgroup_events(struct mem_cgroup
*memcg
,
5535 enum mem_cgroup_events_index idx
,
5538 this_cpu_add(memcg
->stat
->events
[idx
], nr
);
5542 * mem_cgroup_low - check if memory consumption is below the normal range
5543 * @root: the highest ancestor to consider
5544 * @memcg: the memory cgroup to check
5546 * Returns %true if memory consumption of @memcg, and that of all
5547 * configurable ancestors up to @root, is below the normal range.
5549 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5551 if (mem_cgroup_disabled())
5555 * The toplevel group doesn't have a configurable range, so
5556 * it's never low when looked at directly, and it is not
5557 * considered an ancestor when assessing the hierarchy.
5560 if (memcg
== root_mem_cgroup
)
5563 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5566 while (memcg
!= root
) {
5567 memcg
= parent_mem_cgroup(memcg
);
5569 if (memcg
== root_mem_cgroup
)
5572 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5579 * mem_cgroup_try_charge - try charging a page
5580 * @page: page to charge
5581 * @mm: mm context of the victim
5582 * @gfp_mask: reclaim mode
5583 * @memcgp: charged memcg return
5585 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5586 * pages according to @gfp_mask if necessary.
5588 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5589 * Otherwise, an error code is returned.
5591 * After page->mapping has been set up, the caller must finalize the
5592 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5593 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5595 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5596 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5598 struct mem_cgroup
*memcg
= NULL
;
5599 unsigned int nr_pages
= 1;
5602 if (mem_cgroup_disabled())
5605 if (PageSwapCache(page
)) {
5607 * Every swap fault against a single page tries to charge the
5608 * page, bail as early as possible. shmem_unuse() encounters
5609 * already charged pages, too. The USED bit is protected by
5610 * the page lock, which serializes swap cache removal, which
5611 * in turn serializes uncharging.
5613 if (page
->mem_cgroup
)
5617 if (PageTransHuge(page
)) {
5618 nr_pages
<<= compound_order(page
);
5619 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5622 if (do_swap_account
&& PageSwapCache(page
))
5623 memcg
= try_get_mem_cgroup_from_page(page
);
5625 memcg
= get_mem_cgroup_from_mm(mm
);
5627 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5629 css_put(&memcg
->css
);
5631 if (ret
== -EINTR
) {
5632 memcg
= root_mem_cgroup
;
5641 * mem_cgroup_commit_charge - commit a page charge
5642 * @page: page to charge
5643 * @memcg: memcg to charge the page to
5644 * @lrucare: page might be on LRU already
5646 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5647 * after page->mapping has been set up. This must happen atomically
5648 * as part of the page instantiation, i.e. under the page table lock
5649 * for anonymous pages, under the page lock for page and swap cache.
5651 * In addition, the page must not be on the LRU during the commit, to
5652 * prevent racing with task migration. If it might be, use @lrucare.
5654 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5656 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5659 unsigned int nr_pages
= 1;
5661 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5662 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5664 if (mem_cgroup_disabled())
5667 * Swap faults will attempt to charge the same page multiple
5668 * times. But reuse_swap_page() might have removed the page
5669 * from swapcache already, so we can't check PageSwapCache().
5674 commit_charge(page
, memcg
, lrucare
);
5676 if (PageTransHuge(page
)) {
5677 nr_pages
<<= compound_order(page
);
5678 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5681 local_irq_disable();
5682 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5683 memcg_check_events(memcg
, page
);
5686 if (do_swap_account
&& PageSwapCache(page
)) {
5687 swp_entry_t entry
= { .val
= page_private(page
) };
5689 * The swap entry might not get freed for a long time,
5690 * let's not wait for it. The page already received a
5691 * memory+swap charge, drop the swap entry duplicate.
5693 mem_cgroup_uncharge_swap(entry
);
5698 * mem_cgroup_cancel_charge - cancel a page charge
5699 * @page: page to charge
5700 * @memcg: memcg to charge the page to
5702 * Cancel a charge transaction started by mem_cgroup_try_charge().
5704 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5706 unsigned int nr_pages
= 1;
5708 if (mem_cgroup_disabled())
5711 * Swap faults will attempt to charge the same page multiple
5712 * times. But reuse_swap_page() might have removed the page
5713 * from swapcache already, so we can't check PageSwapCache().
5718 if (PageTransHuge(page
)) {
5719 nr_pages
<<= compound_order(page
);
5720 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5723 cancel_charge(memcg
, nr_pages
);
5726 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5727 unsigned long nr_anon
, unsigned long nr_file
,
5728 unsigned long nr_huge
, struct page
*dummy_page
)
5730 unsigned long nr_pages
= nr_anon
+ nr_file
;
5731 unsigned long flags
;
5733 if (!mem_cgroup_is_root(memcg
)) {
5734 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5735 if (do_swap_account
)
5736 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5737 memcg_oom_recover(memcg
);
5740 local_irq_save(flags
);
5741 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5742 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5743 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5744 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5745 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5746 memcg_check_events(memcg
, dummy_page
);
5747 local_irq_restore(flags
);
5749 if (!mem_cgroup_is_root(memcg
))
5750 css_put_many(&memcg
->css
, nr_pages
);
5753 static void uncharge_list(struct list_head
*page_list
)
5755 struct mem_cgroup
*memcg
= NULL
;
5756 unsigned long nr_anon
= 0;
5757 unsigned long nr_file
= 0;
5758 unsigned long nr_huge
= 0;
5759 unsigned long pgpgout
= 0;
5760 struct list_head
*next
;
5763 next
= page_list
->next
;
5765 unsigned int nr_pages
= 1;
5767 page
= list_entry(next
, struct page
, lru
);
5768 next
= page
->lru
.next
;
5770 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5771 VM_BUG_ON_PAGE(page_count(page
), page
);
5773 if (!page
->mem_cgroup
)
5777 * Nobody should be changing or seriously looking at
5778 * page->mem_cgroup at this point, we have fully
5779 * exclusive access to the page.
5782 if (memcg
!= page
->mem_cgroup
) {
5784 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5786 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5788 memcg
= page
->mem_cgroup
;
5791 if (PageTransHuge(page
)) {
5792 nr_pages
<<= compound_order(page
);
5793 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5794 nr_huge
+= nr_pages
;
5798 nr_anon
+= nr_pages
;
5800 nr_file
+= nr_pages
;
5802 page
->mem_cgroup
= NULL
;
5805 } while (next
!= page_list
);
5808 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5813 * mem_cgroup_uncharge - uncharge a page
5814 * @page: page to uncharge
5816 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5817 * mem_cgroup_commit_charge().
5819 void mem_cgroup_uncharge(struct page
*page
)
5821 if (mem_cgroup_disabled())
5824 /* Don't touch page->lru of any random page, pre-check: */
5825 if (!page
->mem_cgroup
)
5828 INIT_LIST_HEAD(&page
->lru
);
5829 uncharge_list(&page
->lru
);
5833 * mem_cgroup_uncharge_list - uncharge a list of page
5834 * @page_list: list of pages to uncharge
5836 * Uncharge a list of pages previously charged with
5837 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5839 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5841 if (mem_cgroup_disabled())
5844 if (!list_empty(page_list
))
5845 uncharge_list(page_list
);
5849 * mem_cgroup_migrate - migrate a charge to another page
5850 * @oldpage: currently charged page
5851 * @newpage: page to transfer the charge to
5852 * @lrucare: either or both pages might be on the LRU already
5854 * Migrate the charge from @oldpage to @newpage.
5856 * Both pages must be locked, @newpage->mapping must be set up.
5858 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
5861 struct mem_cgroup
*memcg
;
5864 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5865 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5866 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
5867 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
5868 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5869 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5872 if (mem_cgroup_disabled())
5875 /* Page cache replacement: new page already charged? */
5876 if (newpage
->mem_cgroup
)
5880 * Swapcache readahead pages can get migrated before being
5881 * charged, and migration from compaction can happen to an
5882 * uncharged page when the PFN walker finds a page that
5883 * reclaim just put back on the LRU but has not released yet.
5885 memcg
= oldpage
->mem_cgroup
;
5890 lock_page_lru(oldpage
, &isolated
);
5892 oldpage
->mem_cgroup
= NULL
;
5895 unlock_page_lru(oldpage
, isolated
);
5897 commit_charge(newpage
, memcg
, lrucare
);
5901 * subsys_initcall() for memory controller.
5903 * Some parts like hotcpu_notifier() have to be initialized from this context
5904 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5905 * everything that doesn't depend on a specific mem_cgroup structure should
5906 * be initialized from here.
5908 static int __init
mem_cgroup_init(void)
5912 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5914 for_each_possible_cpu(cpu
)
5915 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5918 for_each_node(node
) {
5919 struct mem_cgroup_tree_per_node
*rtpn
;
5922 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5923 node_online(node
) ? node
: NUMA_NO_NODE
);
5925 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5926 struct mem_cgroup_tree_per_zone
*rtpz
;
5928 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5929 rtpz
->rb_root
= RB_ROOT
;
5930 spin_lock_init(&rtpz
->lock
);
5932 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5937 subsys_initcall(mem_cgroup_init
);
5939 #ifdef CONFIG_MEMCG_SWAP
5941 * mem_cgroup_swapout - transfer a memsw charge to swap
5942 * @page: page whose memsw charge to transfer
5943 * @entry: swap entry to move the charge to
5945 * Transfer the memsw charge of @page to @entry.
5947 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5949 struct mem_cgroup
*memcg
;
5950 unsigned short oldid
;
5952 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5953 VM_BUG_ON_PAGE(page_count(page
), page
);
5955 if (!do_swap_account
)
5958 memcg
= page
->mem_cgroup
;
5960 /* Readahead page, never charged */
5964 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5965 VM_BUG_ON_PAGE(oldid
, page
);
5966 mem_cgroup_swap_statistics(memcg
, true);
5968 page
->mem_cgroup
= NULL
;
5970 if (!mem_cgroup_is_root(memcg
))
5971 page_counter_uncharge(&memcg
->memory
, 1);
5974 * Interrupts should be disabled here because the caller holds the
5975 * mapping->tree_lock lock which is taken with interrupts-off. It is
5976 * important here to have the interrupts disabled because it is the
5977 * only synchronisation we have for udpating the per-CPU variables.
5979 VM_BUG_ON(!irqs_disabled());
5980 mem_cgroup_charge_statistics(memcg
, page
, -1);
5981 memcg_check_events(memcg
, page
);
5985 * mem_cgroup_uncharge_swap - uncharge a swap entry
5986 * @entry: swap entry to uncharge
5988 * Drop the memsw charge associated with @entry.
5990 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5992 struct mem_cgroup
*memcg
;
5995 if (!do_swap_account
)
5998 id
= swap_cgroup_record(entry
, 0);
6000 memcg
= mem_cgroup_from_id(id
);
6002 if (!mem_cgroup_is_root(memcg
))
6003 page_counter_uncharge(&memcg
->memsw
, 1);
6004 mem_cgroup_swap_statistics(memcg
, false);
6005 css_put(&memcg
->css
);
6010 /* for remember boot option*/
6011 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6012 static int really_do_swap_account __initdata
= 1;
6014 static int really_do_swap_account __initdata
;
6017 static int __init
enable_swap_account(char *s
)
6019 if (!strcmp(s
, "1"))
6020 really_do_swap_account
= 1;
6021 else if (!strcmp(s
, "0"))
6022 really_do_swap_account
= 0;
6025 __setup("swapaccount=", enable_swap_account
);
6027 static struct cftype memsw_cgroup_files
[] = {
6029 .name
= "memsw.usage_in_bytes",
6030 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6031 .read_u64
= mem_cgroup_read_u64
,
6034 .name
= "memsw.max_usage_in_bytes",
6035 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6036 .write
= mem_cgroup_reset
,
6037 .read_u64
= mem_cgroup_read_u64
,
6040 .name
= "memsw.limit_in_bytes",
6041 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6042 .write
= mem_cgroup_write
,
6043 .read_u64
= mem_cgroup_read_u64
,
6046 .name
= "memsw.failcnt",
6047 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6048 .write
= mem_cgroup_reset
,
6049 .read_u64
= mem_cgroup_read_u64
,
6051 { }, /* terminate */
6054 static int __init
mem_cgroup_swap_init(void)
6056 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6057 do_swap_account
= 1;
6058 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6059 memsw_cgroup_files
));
6063 subsys_initcall(mem_cgroup_swap_init
);
6065 #endif /* CONFIG_MEMCG_SWAP */