1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
70 EXPORT_SYMBOL(mem_cgroup_subsys
);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly
;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata
= 1;
83 static int really_do_swap_account __initdata
= 0;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names
[] = {
100 enum mem_cgroup_events_index
{
101 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS
,
108 static const char * const mem_cgroup_events_names
[] = {
115 static const char * const mem_cgroup_lru_names
[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target
{
130 MEM_CGROUP_TARGET_THRESH
,
131 MEM_CGROUP_TARGET_SOFTLIMIT
,
132 MEM_CGROUP_TARGET_NUMAINFO
,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu
{
140 long count
[MEM_CGROUP_STAT_NSTATS
];
141 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
142 unsigned long nr_page_events
;
143 unsigned long targets
[MEM_CGROUP_NTARGETS
];
146 struct mem_cgroup_reclaim_iter
{
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup
*last_visited
;
152 unsigned long last_dead_count
;
154 /* scan generation, increased every round-trip */
155 unsigned int generation
;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone
{
162 struct lruvec lruvec
;
163 unsigned long lru_size
[NR_LRU_LISTS
];
165 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
167 struct rb_node tree_node
; /* RB tree node */
168 unsigned long long usage_in_excess
;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node
{
176 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone
{
185 struct rb_root rb_root
;
189 struct mem_cgroup_tree_per_node
{
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
193 struct mem_cgroup_tree
{
194 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
199 struct mem_cgroup_threshold
{
200 struct eventfd_ctx
*eventfd
;
205 struct mem_cgroup_threshold_ary
{
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold
;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries
[0];
214 struct mem_cgroup_thresholds
{
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary
*primary
;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary
*spare
;
226 struct mem_cgroup_eventfd_list
{
227 struct list_head list
;
228 struct eventfd_ctx
*eventfd
;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event
{
236 * memcg which the event belongs to.
238 struct mem_cgroup
*memcg
;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx
*eventfd
;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list
;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event
)(struct mem_cgroup
*memcg
,
253 struct eventfd_ctx
*eventfd
, const char *args
);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event
)(struct mem_cgroup
*memcg
,
260 struct eventfd_ctx
*eventfd
);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t
*wqh
;
268 struct work_struct remove
;
271 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
272 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css
;
288 * the counter to account for memory usage
290 struct res_counter res
;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure
;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw
;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem
;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups
;
315 /* OOM-Killer disable */
316 int oom_kill_disable
;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum
;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock
;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds
;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds
;
330 /* For oom notifier event fd */
331 struct list_head oom_notify
;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate
;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account
;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock
;
347 struct mem_cgroup_stat_cpu __percpu
*stat
;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base
;
353 spinlock_t pcp_counter_lock
;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem
;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list. per-memcg */
361 struct list_head memcg_slab_caches
;
362 /* Not a spinlock, we can take a lot of time walking the list */
363 struct mutex slab_caches_mutex
;
364 /* Index in the kmem_cache->memcg_params->memcg_caches array */
368 int last_scanned_node
;
370 nodemask_t scan_nodes
;
371 atomic_t numainfo_events
;
372 atomic_t numainfo_updating
;
375 /* List of events which userspace want to receive */
376 struct list_head event_list
;
377 spinlock_t event_list_lock
;
379 struct mem_cgroup_per_node
*nodeinfo
[0];
380 /* WARNING: nodeinfo must be the last member here */
383 /* internal only representation about the status of kmem accounting. */
385 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
387 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
390 /* We account when limit is on, but only after call sites are patched */
391 #define KMEM_ACCOUNTED_MASK \
392 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
394 #ifdef CONFIG_MEMCG_KMEM
395 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
397 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
400 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
402 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
405 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
407 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
410 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
412 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
415 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
418 * Our caller must use css_get() first, because memcg_uncharge_kmem()
419 * will call css_put() if it sees the memcg is dead.
422 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
423 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
426 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
428 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
429 &memcg
->kmem_account_flags
);
433 /* Stuffs for move charges at task migration. */
435 * Types of charges to be moved. "move_charge_at_immitgrate" and
436 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
439 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
440 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
444 /* "mc" and its members are protected by cgroup_mutex */
445 static struct move_charge_struct
{
446 spinlock_t lock
; /* for from, to */
447 struct mem_cgroup
*from
;
448 struct mem_cgroup
*to
;
449 unsigned long immigrate_flags
;
450 unsigned long precharge
;
451 unsigned long moved_charge
;
452 unsigned long moved_swap
;
453 struct task_struct
*moving_task
; /* a task moving charges */
454 wait_queue_head_t waitq
; /* a waitq for other context */
456 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
457 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
460 static bool move_anon(void)
462 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
465 static bool move_file(void)
467 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
471 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
472 * limit reclaim to prevent infinite loops, if they ever occur.
474 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
475 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
478 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
479 MEM_CGROUP_CHARGE_TYPE_ANON
,
480 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
481 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
485 /* for encoding cft->private value on file */
493 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
494 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
495 #define MEMFILE_ATTR(val) ((val) & 0xffff)
496 /* Used for OOM nofiier */
497 #define OOM_CONTROL (0)
500 * Reclaim flags for mem_cgroup_hierarchical_reclaim
502 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
503 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
504 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
505 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
508 * The memcg_create_mutex will be held whenever a new cgroup is created.
509 * As a consequence, any change that needs to protect against new child cgroups
510 * appearing has to hold it as well.
512 static DEFINE_MUTEX(memcg_create_mutex
);
514 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
516 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
519 /* Some nice accessors for the vmpressure. */
520 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
523 memcg
= root_mem_cgroup
;
524 return &memcg
->vmpressure
;
527 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
529 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
532 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
534 return (memcg
== root_mem_cgroup
);
538 * We restrict the id in the range of [1, 65535], so it can fit into
541 #define MEM_CGROUP_ID_MAX USHRT_MAX
543 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
546 * The ID of the root cgroup is 0, but memcg treat 0 as an
547 * invalid ID, so we return (cgroup_id + 1).
549 return memcg
->css
.cgroup
->id
+ 1;
552 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
554 struct cgroup_subsys_state
*css
;
556 css
= css_from_id(id
- 1, &mem_cgroup_subsys
);
557 return mem_cgroup_from_css(css
);
560 /* Writing them here to avoid exposing memcg's inner layout */
561 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
563 void sock_update_memcg(struct sock
*sk
)
565 if (mem_cgroup_sockets_enabled
) {
566 struct mem_cgroup
*memcg
;
567 struct cg_proto
*cg_proto
;
569 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
571 /* Socket cloning can throw us here with sk_cgrp already
572 * filled. It won't however, necessarily happen from
573 * process context. So the test for root memcg given
574 * the current task's memcg won't help us in this case.
576 * Respecting the original socket's memcg is a better
577 * decision in this case.
580 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
581 css_get(&sk
->sk_cgrp
->memcg
->css
);
586 memcg
= mem_cgroup_from_task(current
);
587 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
588 if (!mem_cgroup_is_root(memcg
) &&
589 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
590 sk
->sk_cgrp
= cg_proto
;
595 EXPORT_SYMBOL(sock_update_memcg
);
597 void sock_release_memcg(struct sock
*sk
)
599 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
600 struct mem_cgroup
*memcg
;
601 WARN_ON(!sk
->sk_cgrp
->memcg
);
602 memcg
= sk
->sk_cgrp
->memcg
;
603 css_put(&sk
->sk_cgrp
->memcg
->css
);
607 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
609 if (!memcg
|| mem_cgroup_is_root(memcg
))
612 return &memcg
->tcp_mem
;
614 EXPORT_SYMBOL(tcp_proto_cgroup
);
616 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
618 if (!memcg_proto_activated(&memcg
->tcp_mem
))
620 static_key_slow_dec(&memcg_socket_limit_enabled
);
623 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
628 #ifdef CONFIG_MEMCG_KMEM
630 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
631 * The main reason for not using cgroup id for this:
632 * this works better in sparse environments, where we have a lot of memcgs,
633 * but only a few kmem-limited. Or also, if we have, for instance, 200
634 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
635 * 200 entry array for that.
637 * The current size of the caches array is stored in
638 * memcg_limited_groups_array_size. It will double each time we have to
641 static DEFINE_IDA(kmem_limited_groups
);
642 int memcg_limited_groups_array_size
;
645 * MIN_SIZE is different than 1, because we would like to avoid going through
646 * the alloc/free process all the time. In a small machine, 4 kmem-limited
647 * cgroups is a reasonable guess. In the future, it could be a parameter or
648 * tunable, but that is strictly not necessary.
650 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
651 * this constant directly from cgroup, but it is understandable that this is
652 * better kept as an internal representation in cgroup.c. In any case, the
653 * cgrp_id space is not getting any smaller, and we don't have to necessarily
654 * increase ours as well if it increases.
656 #define MEMCG_CACHES_MIN_SIZE 4
657 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
660 * A lot of the calls to the cache allocation functions are expected to be
661 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
662 * conditional to this static branch, we'll have to allow modules that does
663 * kmem_cache_alloc and the such to see this symbol as well
665 struct static_key memcg_kmem_enabled_key
;
666 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
668 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
670 if (memcg_kmem_is_active(memcg
)) {
671 static_key_slow_dec(&memcg_kmem_enabled_key
);
672 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
675 * This check can't live in kmem destruction function,
676 * since the charges will outlive the cgroup
678 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
681 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
684 #endif /* CONFIG_MEMCG_KMEM */
686 static void disarm_static_keys(struct mem_cgroup
*memcg
)
688 disarm_sock_keys(memcg
);
689 disarm_kmem_keys(memcg
);
692 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
694 static struct mem_cgroup_per_zone
*
695 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
697 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
698 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
701 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
706 static struct mem_cgroup_per_zone
*
707 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
709 int nid
= page_to_nid(page
);
710 int zid
= page_zonenum(page
);
712 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
715 static struct mem_cgroup_tree_per_zone
*
716 soft_limit_tree_node_zone(int nid
, int zid
)
718 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
721 static struct mem_cgroup_tree_per_zone
*
722 soft_limit_tree_from_page(struct page
*page
)
724 int nid
= page_to_nid(page
);
725 int zid
= page_zonenum(page
);
727 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
731 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
732 struct mem_cgroup_per_zone
*mz
,
733 struct mem_cgroup_tree_per_zone
*mctz
,
734 unsigned long long new_usage_in_excess
)
736 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
737 struct rb_node
*parent
= NULL
;
738 struct mem_cgroup_per_zone
*mz_node
;
743 mz
->usage_in_excess
= new_usage_in_excess
;
744 if (!mz
->usage_in_excess
)
748 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
750 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
753 * We can't avoid mem cgroups that are over their soft
754 * limit by the same amount
756 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
759 rb_link_node(&mz
->tree_node
, parent
, p
);
760 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
765 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
766 struct mem_cgroup_per_zone
*mz
,
767 struct mem_cgroup_tree_per_zone
*mctz
)
771 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
776 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
777 struct mem_cgroup_per_zone
*mz
,
778 struct mem_cgroup_tree_per_zone
*mctz
)
780 spin_lock(&mctz
->lock
);
781 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
782 spin_unlock(&mctz
->lock
);
786 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
788 unsigned long long excess
;
789 struct mem_cgroup_per_zone
*mz
;
790 struct mem_cgroup_tree_per_zone
*mctz
;
791 int nid
= page_to_nid(page
);
792 int zid
= page_zonenum(page
);
793 mctz
= soft_limit_tree_from_page(page
);
796 * Necessary to update all ancestors when hierarchy is used.
797 * because their event counter is not touched.
799 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
800 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
801 excess
= res_counter_soft_limit_excess(&memcg
->res
);
803 * We have to update the tree if mz is on RB-tree or
804 * mem is over its softlimit.
806 if (excess
|| mz
->on_tree
) {
807 spin_lock(&mctz
->lock
);
808 /* if on-tree, remove it */
810 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
812 * Insert again. mz->usage_in_excess will be updated.
813 * If excess is 0, no tree ops.
815 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
816 spin_unlock(&mctz
->lock
);
821 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
824 struct mem_cgroup_per_zone
*mz
;
825 struct mem_cgroup_tree_per_zone
*mctz
;
827 for_each_node(node
) {
828 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
829 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
830 mctz
= soft_limit_tree_node_zone(node
, zone
);
831 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
836 static struct mem_cgroup_per_zone
*
837 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
839 struct rb_node
*rightmost
= NULL
;
840 struct mem_cgroup_per_zone
*mz
;
844 rightmost
= rb_last(&mctz
->rb_root
);
846 goto done
; /* Nothing to reclaim from */
848 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
850 * Remove the node now but someone else can add it back,
851 * we will to add it back at the end of reclaim to its correct
852 * position in the tree.
854 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
855 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
856 !css_tryget(&mz
->memcg
->css
))
862 static struct mem_cgroup_per_zone
*
863 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
865 struct mem_cgroup_per_zone
*mz
;
867 spin_lock(&mctz
->lock
);
868 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
869 spin_unlock(&mctz
->lock
);
874 * Implementation Note: reading percpu statistics for memcg.
876 * Both of vmstat[] and percpu_counter has threshold and do periodic
877 * synchronization to implement "quick" read. There are trade-off between
878 * reading cost and precision of value. Then, we may have a chance to implement
879 * a periodic synchronizion of counter in memcg's counter.
881 * But this _read() function is used for user interface now. The user accounts
882 * memory usage by memory cgroup and he _always_ requires exact value because
883 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
884 * have to visit all online cpus and make sum. So, for now, unnecessary
885 * synchronization is not implemented. (just implemented for cpu hotplug)
887 * If there are kernel internal actions which can make use of some not-exact
888 * value, and reading all cpu value can be performance bottleneck in some
889 * common workload, threashold and synchonization as vmstat[] should be
892 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
893 enum mem_cgroup_stat_index idx
)
899 for_each_online_cpu(cpu
)
900 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
901 #ifdef CONFIG_HOTPLUG_CPU
902 spin_lock(&memcg
->pcp_counter_lock
);
903 val
+= memcg
->nocpu_base
.count
[idx
];
904 spin_unlock(&memcg
->pcp_counter_lock
);
910 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
913 int val
= (charge
) ? 1 : -1;
914 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
917 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
918 enum mem_cgroup_events_index idx
)
920 unsigned long val
= 0;
924 for_each_online_cpu(cpu
)
925 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
926 #ifdef CONFIG_HOTPLUG_CPU
927 spin_lock(&memcg
->pcp_counter_lock
);
928 val
+= memcg
->nocpu_base
.events
[idx
];
929 spin_unlock(&memcg
->pcp_counter_lock
);
935 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
937 bool anon
, int nr_pages
)
942 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
943 * counted as CACHE even if it's on ANON LRU.
946 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
949 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
952 if (PageTransHuge(page
))
953 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
956 /* pagein of a big page is an event. So, ignore page size */
958 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
960 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
961 nr_pages
= -nr_pages
; /* for event */
964 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
970 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
972 struct mem_cgroup_per_zone
*mz
;
974 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
975 return mz
->lru_size
[lru
];
979 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
980 unsigned int lru_mask
)
982 struct mem_cgroup_per_zone
*mz
;
984 unsigned long ret
= 0;
986 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
989 if (BIT(lru
) & lru_mask
)
990 ret
+= mz
->lru_size
[lru
];
996 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
997 int nid
, unsigned int lru_mask
)
1002 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
1003 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
1004 nid
, zid
, lru_mask
);
1009 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
1010 unsigned int lru_mask
)
1015 for_each_node_state(nid
, N_MEMORY
)
1016 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
1020 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
1021 enum mem_cgroup_events_target target
)
1023 unsigned long val
, next
;
1025 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
1026 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
1027 /* from time_after() in jiffies.h */
1028 if ((long)next
- (long)val
< 0) {
1030 case MEM_CGROUP_TARGET_THRESH
:
1031 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
1033 case MEM_CGROUP_TARGET_SOFTLIMIT
:
1034 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1036 case MEM_CGROUP_TARGET_NUMAINFO
:
1037 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1042 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1049 * Check events in order.
1052 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1055 /* threshold event is triggered in finer grain than soft limit */
1056 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1057 MEM_CGROUP_TARGET_THRESH
))) {
1059 bool do_numainfo __maybe_unused
;
1061 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1062 MEM_CGROUP_TARGET_SOFTLIMIT
);
1063 #if MAX_NUMNODES > 1
1064 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1065 MEM_CGROUP_TARGET_NUMAINFO
);
1069 mem_cgroup_threshold(memcg
);
1070 if (unlikely(do_softlimit
))
1071 mem_cgroup_update_tree(memcg
, page
);
1072 #if MAX_NUMNODES > 1
1073 if (unlikely(do_numainfo
))
1074 atomic_inc(&memcg
->numainfo_events
);
1080 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1083 * mm_update_next_owner() may clear mm->owner to NULL
1084 * if it races with swapoff, page migration, etc.
1085 * So this can be called with p == NULL.
1090 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
1093 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1095 struct mem_cgroup
*memcg
= NULL
;
1100 * Because we have no locks, mm->owner's may be being moved to other
1101 * cgroup. We use css_tryget() here even if this looks
1102 * pessimistic (rather than adding locks here).
1106 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1107 if (unlikely(!memcg
))
1109 } while (!css_tryget(&memcg
->css
));
1115 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1116 * ref. count) or NULL if the whole root's subtree has been visited.
1118 * helper function to be used by mem_cgroup_iter
1120 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1121 struct mem_cgroup
*last_visited
)
1123 struct cgroup_subsys_state
*prev_css
, *next_css
;
1125 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1127 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1130 * Even if we found a group we have to make sure it is
1131 * alive. css && !memcg means that the groups should be
1132 * skipped and we should continue the tree walk.
1133 * last_visited css is safe to use because it is
1134 * protected by css_get and the tree walk is rcu safe.
1137 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1139 if (css_tryget(&mem
->css
))
1142 prev_css
= next_css
;
1150 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1153 * When a group in the hierarchy below root is destroyed, the
1154 * hierarchy iterator can no longer be trusted since it might
1155 * have pointed to the destroyed group. Invalidate it.
1157 atomic_inc(&root
->dead_count
);
1160 static struct mem_cgroup
*
1161 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1162 struct mem_cgroup
*root
,
1165 struct mem_cgroup
*position
= NULL
;
1167 * A cgroup destruction happens in two stages: offlining and
1168 * release. They are separated by a RCU grace period.
1170 * If the iterator is valid, we may still race with an
1171 * offlining. The RCU lock ensures the object won't be
1172 * released, tryget will fail if we lost the race.
1174 *sequence
= atomic_read(&root
->dead_count
);
1175 if (iter
->last_dead_count
== *sequence
) {
1177 position
= iter
->last_visited
;
1178 if (position
&& !css_tryget(&position
->css
))
1184 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1185 struct mem_cgroup
*last_visited
,
1186 struct mem_cgroup
*new_position
,
1190 css_put(&last_visited
->css
);
1192 * We store the sequence count from the time @last_visited was
1193 * loaded successfully instead of rereading it here so that we
1194 * don't lose destruction events in between. We could have
1195 * raced with the destruction of @new_position after all.
1197 iter
->last_visited
= new_position
;
1199 iter
->last_dead_count
= sequence
;
1203 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1204 * @root: hierarchy root
1205 * @prev: previously returned memcg, NULL on first invocation
1206 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1208 * Returns references to children of the hierarchy below @root, or
1209 * @root itself, or %NULL after a full round-trip.
1211 * Caller must pass the return value in @prev on subsequent
1212 * invocations for reference counting, or use mem_cgroup_iter_break()
1213 * to cancel a hierarchy walk before the round-trip is complete.
1215 * Reclaimers can specify a zone and a priority level in @reclaim to
1216 * divide up the memcgs in the hierarchy among all concurrent
1217 * reclaimers operating on the same zone and priority.
1219 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1220 struct mem_cgroup
*prev
,
1221 struct mem_cgroup_reclaim_cookie
*reclaim
)
1223 struct mem_cgroup
*memcg
= NULL
;
1224 struct mem_cgroup
*last_visited
= NULL
;
1226 if (mem_cgroup_disabled())
1230 root
= root_mem_cgroup
;
1232 if (prev
&& !reclaim
)
1233 last_visited
= prev
;
1235 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1243 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1244 int uninitialized_var(seq
);
1247 int nid
= zone_to_nid(reclaim
->zone
);
1248 int zid
= zone_idx(reclaim
->zone
);
1249 struct mem_cgroup_per_zone
*mz
;
1251 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1252 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1253 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1254 iter
->last_visited
= NULL
;
1258 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1261 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1264 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1268 else if (!prev
&& memcg
)
1269 reclaim
->generation
= iter
->generation
;
1278 if (prev
&& prev
!= root
)
1279 css_put(&prev
->css
);
1285 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1286 * @root: hierarchy root
1287 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1289 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1290 struct mem_cgroup
*prev
)
1293 root
= root_mem_cgroup
;
1294 if (prev
&& prev
!= root
)
1295 css_put(&prev
->css
);
1299 * Iteration constructs for visiting all cgroups (under a tree). If
1300 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1301 * be used for reference counting.
1303 #define for_each_mem_cgroup_tree(iter, root) \
1304 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1306 iter = mem_cgroup_iter(root, iter, NULL))
1308 #define for_each_mem_cgroup(iter) \
1309 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1311 iter = mem_cgroup_iter(NULL, iter, NULL))
1313 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1315 struct mem_cgroup
*memcg
;
1318 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1319 if (unlikely(!memcg
))
1324 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1327 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1335 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1338 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1339 * @zone: zone of the wanted lruvec
1340 * @memcg: memcg of the wanted lruvec
1342 * Returns the lru list vector holding pages for the given @zone and
1343 * @mem. This can be the global zone lruvec, if the memory controller
1346 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1347 struct mem_cgroup
*memcg
)
1349 struct mem_cgroup_per_zone
*mz
;
1350 struct lruvec
*lruvec
;
1352 if (mem_cgroup_disabled()) {
1353 lruvec
= &zone
->lruvec
;
1357 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1358 lruvec
= &mz
->lruvec
;
1361 * Since a node can be onlined after the mem_cgroup was created,
1362 * we have to be prepared to initialize lruvec->zone here;
1363 * and if offlined then reonlined, we need to reinitialize it.
1365 if (unlikely(lruvec
->zone
!= zone
))
1366 lruvec
->zone
= zone
;
1371 * Following LRU functions are allowed to be used without PCG_LOCK.
1372 * Operations are called by routine of global LRU independently from memcg.
1373 * What we have to take care of here is validness of pc->mem_cgroup.
1375 * Changes to pc->mem_cgroup happens when
1378 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1379 * It is added to LRU before charge.
1380 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1381 * When moving account, the page is not on LRU. It's isolated.
1385 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1387 * @zone: zone of the page
1389 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1391 struct mem_cgroup_per_zone
*mz
;
1392 struct mem_cgroup
*memcg
;
1393 struct page_cgroup
*pc
;
1394 struct lruvec
*lruvec
;
1396 if (mem_cgroup_disabled()) {
1397 lruvec
= &zone
->lruvec
;
1401 pc
= lookup_page_cgroup(page
);
1402 memcg
= pc
->mem_cgroup
;
1405 * Surreptitiously switch any uncharged offlist page to root:
1406 * an uncharged page off lru does nothing to secure
1407 * its former mem_cgroup from sudden removal.
1409 * Our caller holds lru_lock, and PageCgroupUsed is updated
1410 * under page_cgroup lock: between them, they make all uses
1411 * of pc->mem_cgroup safe.
1413 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1414 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1416 mz
= page_cgroup_zoneinfo(memcg
, page
);
1417 lruvec
= &mz
->lruvec
;
1420 * Since a node can be onlined after the mem_cgroup was created,
1421 * we have to be prepared to initialize lruvec->zone here;
1422 * and if offlined then reonlined, we need to reinitialize it.
1424 if (unlikely(lruvec
->zone
!= zone
))
1425 lruvec
->zone
= zone
;
1430 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1431 * @lruvec: mem_cgroup per zone lru vector
1432 * @lru: index of lru list the page is sitting on
1433 * @nr_pages: positive when adding or negative when removing
1435 * This function must be called when a page is added to or removed from an
1438 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1441 struct mem_cgroup_per_zone
*mz
;
1442 unsigned long *lru_size
;
1444 if (mem_cgroup_disabled())
1447 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1448 lru_size
= mz
->lru_size
+ lru
;
1449 *lru_size
+= nr_pages
;
1450 VM_BUG_ON((long)(*lru_size
) < 0);
1454 * Checks whether given mem is same or in the root_mem_cgroup's
1457 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1458 struct mem_cgroup
*memcg
)
1460 if (root_memcg
== memcg
)
1462 if (!root_memcg
->use_hierarchy
|| !memcg
)
1464 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1467 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1468 struct mem_cgroup
*memcg
)
1473 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1478 bool task_in_mem_cgroup(struct task_struct
*task
,
1479 const struct mem_cgroup
*memcg
)
1481 struct mem_cgroup
*curr
= NULL
;
1482 struct task_struct
*p
;
1485 p
= find_lock_task_mm(task
);
1487 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1491 * All threads may have already detached their mm's, but the oom
1492 * killer still needs to detect if they have already been oom
1493 * killed to prevent needlessly killing additional tasks.
1496 curr
= mem_cgroup_from_task(task
);
1498 css_get(&curr
->css
);
1504 * We should check use_hierarchy of "memcg" not "curr". Because checking
1505 * use_hierarchy of "curr" here make this function true if hierarchy is
1506 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1507 * hierarchy(even if use_hierarchy is disabled in "memcg").
1509 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1510 css_put(&curr
->css
);
1514 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1516 unsigned long inactive_ratio
;
1517 unsigned long inactive
;
1518 unsigned long active
;
1521 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1522 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1524 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1526 inactive_ratio
= int_sqrt(10 * gb
);
1530 return inactive
* inactive_ratio
< active
;
1533 #define mem_cgroup_from_res_counter(counter, member) \
1534 container_of(counter, struct mem_cgroup, member)
1537 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1538 * @memcg: the memory cgroup
1540 * Returns the maximum amount of memory @mem can be charged with, in
1543 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1545 unsigned long long margin
;
1547 margin
= res_counter_margin(&memcg
->res
);
1548 if (do_swap_account
)
1549 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1550 return margin
>> PAGE_SHIFT
;
1553 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1556 if (!css_parent(&memcg
->css
))
1557 return vm_swappiness
;
1559 return memcg
->swappiness
;
1563 * memcg->moving_account is used for checking possibility that some thread is
1564 * calling move_account(). When a thread on CPU-A starts moving pages under
1565 * a memcg, other threads should check memcg->moving_account under
1566 * rcu_read_lock(), like this:
1570 * memcg->moving_account+1 if (memcg->mocing_account)
1572 * synchronize_rcu() update something.
1577 /* for quick checking without looking up memcg */
1578 atomic_t memcg_moving __read_mostly
;
1580 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1582 atomic_inc(&memcg_moving
);
1583 atomic_inc(&memcg
->moving_account
);
1587 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1590 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1591 * We check NULL in callee rather than caller.
1594 atomic_dec(&memcg_moving
);
1595 atomic_dec(&memcg
->moving_account
);
1600 * 2 routines for checking "mem" is under move_account() or not.
1602 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1603 * is used for avoiding races in accounting. If true,
1604 * pc->mem_cgroup may be overwritten.
1606 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1607 * under hierarchy of moving cgroups. This is for
1608 * waiting at hith-memory prressure caused by "move".
1611 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1613 VM_BUG_ON(!rcu_read_lock_held());
1614 return atomic_read(&memcg
->moving_account
) > 0;
1617 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1619 struct mem_cgroup
*from
;
1620 struct mem_cgroup
*to
;
1623 * Unlike task_move routines, we access mc.to, mc.from not under
1624 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1626 spin_lock(&mc
.lock
);
1632 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1633 || mem_cgroup_same_or_subtree(memcg
, to
);
1635 spin_unlock(&mc
.lock
);
1639 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1641 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1642 if (mem_cgroup_under_move(memcg
)) {
1644 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1645 /* moving charge context might have finished. */
1648 finish_wait(&mc
.waitq
, &wait
);
1656 * Take this lock when
1657 * - a code tries to modify page's memcg while it's USED.
1658 * - a code tries to modify page state accounting in a memcg.
1659 * see mem_cgroup_stolen(), too.
1661 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1662 unsigned long *flags
)
1664 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1667 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1668 unsigned long *flags
)
1670 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1673 #define K(x) ((x) << (PAGE_SHIFT-10))
1675 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1676 * @memcg: The memory cgroup that went over limit
1677 * @p: Task that is going to be killed
1679 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1682 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1685 * protects memcg_name and makes sure that parallel ooms do not
1688 static DEFINE_SPINLOCK(oom_info_lock
);
1689 struct cgroup
*task_cgrp
;
1690 struct cgroup
*mem_cgrp
;
1691 static char memcg_name
[PATH_MAX
];
1693 struct mem_cgroup
*iter
;
1699 spin_lock(&oom_info_lock
);
1702 mem_cgrp
= memcg
->css
.cgroup
;
1703 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1705 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1708 * Unfortunately, we are unable to convert to a useful name
1709 * But we'll still print out the usage information
1716 pr_info("Task in %s killed", memcg_name
);
1719 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1727 * Continues from above, so we don't need an KERN_ level
1729 pr_cont(" as a result of limit of %s\n", memcg_name
);
1732 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1733 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1734 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1735 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1736 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1737 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1738 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1739 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1740 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1741 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1742 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1743 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1745 for_each_mem_cgroup_tree(iter
, memcg
) {
1746 pr_info("Memory cgroup stats");
1749 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1751 pr_cont(" for %s", memcg_name
);
1755 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1756 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1758 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1759 K(mem_cgroup_read_stat(iter
, i
)));
1762 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1763 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1764 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1768 spin_unlock(&oom_info_lock
);
1772 * This function returns the number of memcg under hierarchy tree. Returns
1773 * 1(self count) if no children.
1775 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1778 struct mem_cgroup
*iter
;
1780 for_each_mem_cgroup_tree(iter
, memcg
)
1786 * Return the memory (and swap, if configured) limit for a memcg.
1788 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1792 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1795 * Do not consider swap space if we cannot swap due to swappiness
1797 if (mem_cgroup_swappiness(memcg
)) {
1800 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1801 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1804 * If memsw is finite and limits the amount of swap space
1805 * available to this memcg, return that limit.
1807 limit
= min(limit
, memsw
);
1813 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1816 struct mem_cgroup
*iter
;
1817 unsigned long chosen_points
= 0;
1818 unsigned long totalpages
;
1819 unsigned int points
= 0;
1820 struct task_struct
*chosen
= NULL
;
1823 * If current has a pending SIGKILL or is exiting, then automatically
1824 * select it. The goal is to allow it to allocate so that it may
1825 * quickly exit and free its memory.
1827 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1828 set_thread_flag(TIF_MEMDIE
);
1832 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1833 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1834 for_each_mem_cgroup_tree(iter
, memcg
) {
1835 struct css_task_iter it
;
1836 struct task_struct
*task
;
1838 css_task_iter_start(&iter
->css
, &it
);
1839 while ((task
= css_task_iter_next(&it
))) {
1840 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1842 case OOM_SCAN_SELECT
:
1844 put_task_struct(chosen
);
1846 chosen_points
= ULONG_MAX
;
1847 get_task_struct(chosen
);
1849 case OOM_SCAN_CONTINUE
:
1851 case OOM_SCAN_ABORT
:
1852 css_task_iter_end(&it
);
1853 mem_cgroup_iter_break(memcg
, iter
);
1855 put_task_struct(chosen
);
1860 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1861 if (points
> chosen_points
) {
1863 put_task_struct(chosen
);
1865 chosen_points
= points
;
1866 get_task_struct(chosen
);
1869 css_task_iter_end(&it
);
1874 points
= chosen_points
* 1000 / totalpages
;
1875 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1876 NULL
, "Memory cgroup out of memory");
1879 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1881 unsigned long flags
)
1883 unsigned long total
= 0;
1884 bool noswap
= false;
1887 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1889 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1892 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1894 drain_all_stock_async(memcg
);
1895 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1897 * Allow limit shrinkers, which are triggered directly
1898 * by userspace, to catch signals and stop reclaim
1899 * after minimal progress, regardless of the margin.
1901 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1903 if (mem_cgroup_margin(memcg
))
1906 * If nothing was reclaimed after two attempts, there
1907 * may be no reclaimable pages in this hierarchy.
1916 * test_mem_cgroup_node_reclaimable
1917 * @memcg: the target memcg
1918 * @nid: the node ID to be checked.
1919 * @noswap : specify true here if the user wants flle only information.
1921 * This function returns whether the specified memcg contains any
1922 * reclaimable pages on a node. Returns true if there are any reclaimable
1923 * pages in the node.
1925 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1926 int nid
, bool noswap
)
1928 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1930 if (noswap
|| !total_swap_pages
)
1932 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1937 #if MAX_NUMNODES > 1
1940 * Always updating the nodemask is not very good - even if we have an empty
1941 * list or the wrong list here, we can start from some node and traverse all
1942 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1945 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1949 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1950 * pagein/pageout changes since the last update.
1952 if (!atomic_read(&memcg
->numainfo_events
))
1954 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1957 /* make a nodemask where this memcg uses memory from */
1958 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1960 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1962 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1963 node_clear(nid
, memcg
->scan_nodes
);
1966 atomic_set(&memcg
->numainfo_events
, 0);
1967 atomic_set(&memcg
->numainfo_updating
, 0);
1971 * Selecting a node where we start reclaim from. Because what we need is just
1972 * reducing usage counter, start from anywhere is O,K. Considering
1973 * memory reclaim from current node, there are pros. and cons.
1975 * Freeing memory from current node means freeing memory from a node which
1976 * we'll use or we've used. So, it may make LRU bad. And if several threads
1977 * hit limits, it will see a contention on a node. But freeing from remote
1978 * node means more costs for memory reclaim because of memory latency.
1980 * Now, we use round-robin. Better algorithm is welcomed.
1982 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1986 mem_cgroup_may_update_nodemask(memcg
);
1987 node
= memcg
->last_scanned_node
;
1989 node
= next_node(node
, memcg
->scan_nodes
);
1990 if (node
== MAX_NUMNODES
)
1991 node
= first_node(memcg
->scan_nodes
);
1993 * We call this when we hit limit, not when pages are added to LRU.
1994 * No LRU may hold pages because all pages are UNEVICTABLE or
1995 * memcg is too small and all pages are not on LRU. In that case,
1996 * we use curret node.
1998 if (unlikely(node
== MAX_NUMNODES
))
1999 node
= numa_node_id();
2001 memcg
->last_scanned_node
= node
;
2006 * Check all nodes whether it contains reclaimable pages or not.
2007 * For quick scan, we make use of scan_nodes. This will allow us to skip
2008 * unused nodes. But scan_nodes is lazily updated and may not cotain
2009 * enough new information. We need to do double check.
2011 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2016 * quick check...making use of scan_node.
2017 * We can skip unused nodes.
2019 if (!nodes_empty(memcg
->scan_nodes
)) {
2020 for (nid
= first_node(memcg
->scan_nodes
);
2022 nid
= next_node(nid
, memcg
->scan_nodes
)) {
2024 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2029 * Check rest of nodes.
2031 for_each_node_state(nid
, N_MEMORY
) {
2032 if (node_isset(nid
, memcg
->scan_nodes
))
2034 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2041 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2046 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2048 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2052 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2055 unsigned long *total_scanned
)
2057 struct mem_cgroup
*victim
= NULL
;
2060 unsigned long excess
;
2061 unsigned long nr_scanned
;
2062 struct mem_cgroup_reclaim_cookie reclaim
= {
2067 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2070 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2075 * If we have not been able to reclaim
2076 * anything, it might because there are
2077 * no reclaimable pages under this hierarchy
2082 * We want to do more targeted reclaim.
2083 * excess >> 2 is not to excessive so as to
2084 * reclaim too much, nor too less that we keep
2085 * coming back to reclaim from this cgroup
2087 if (total
>= (excess
>> 2) ||
2088 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2093 if (!mem_cgroup_reclaimable(victim
, false))
2095 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2097 *total_scanned
+= nr_scanned
;
2098 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2101 mem_cgroup_iter_break(root_memcg
, victim
);
2105 #ifdef CONFIG_LOCKDEP
2106 static struct lockdep_map memcg_oom_lock_dep_map
= {
2107 .name
= "memcg_oom_lock",
2111 static DEFINE_SPINLOCK(memcg_oom_lock
);
2114 * Check OOM-Killer is already running under our hierarchy.
2115 * If someone is running, return false.
2117 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2119 struct mem_cgroup
*iter
, *failed
= NULL
;
2121 spin_lock(&memcg_oom_lock
);
2123 for_each_mem_cgroup_tree(iter
, memcg
) {
2124 if (iter
->oom_lock
) {
2126 * this subtree of our hierarchy is already locked
2127 * so we cannot give a lock.
2130 mem_cgroup_iter_break(memcg
, iter
);
2133 iter
->oom_lock
= true;
2138 * OK, we failed to lock the whole subtree so we have
2139 * to clean up what we set up to the failing subtree
2141 for_each_mem_cgroup_tree(iter
, memcg
) {
2142 if (iter
== failed
) {
2143 mem_cgroup_iter_break(memcg
, iter
);
2146 iter
->oom_lock
= false;
2149 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2151 spin_unlock(&memcg_oom_lock
);
2156 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2158 struct mem_cgroup
*iter
;
2160 spin_lock(&memcg_oom_lock
);
2161 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2162 for_each_mem_cgroup_tree(iter
, memcg
)
2163 iter
->oom_lock
= false;
2164 spin_unlock(&memcg_oom_lock
);
2167 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2169 struct mem_cgroup
*iter
;
2171 for_each_mem_cgroup_tree(iter
, memcg
)
2172 atomic_inc(&iter
->under_oom
);
2175 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2177 struct mem_cgroup
*iter
;
2180 * When a new child is created while the hierarchy is under oom,
2181 * mem_cgroup_oom_lock() may not be called. We have to use
2182 * atomic_add_unless() here.
2184 for_each_mem_cgroup_tree(iter
, memcg
)
2185 atomic_add_unless(&iter
->under_oom
, -1, 0);
2188 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2190 struct oom_wait_info
{
2191 struct mem_cgroup
*memcg
;
2195 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2196 unsigned mode
, int sync
, void *arg
)
2198 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2199 struct mem_cgroup
*oom_wait_memcg
;
2200 struct oom_wait_info
*oom_wait_info
;
2202 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2203 oom_wait_memcg
= oom_wait_info
->memcg
;
2206 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2207 * Then we can use css_is_ancestor without taking care of RCU.
2209 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2210 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2212 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2215 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2217 atomic_inc(&memcg
->oom_wakeups
);
2218 /* for filtering, pass "memcg" as argument. */
2219 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2222 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2224 if (memcg
&& atomic_read(&memcg
->under_oom
))
2225 memcg_wakeup_oom(memcg
);
2228 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2230 if (!current
->memcg_oom
.may_oom
)
2233 * We are in the middle of the charge context here, so we
2234 * don't want to block when potentially sitting on a callstack
2235 * that holds all kinds of filesystem and mm locks.
2237 * Also, the caller may handle a failed allocation gracefully
2238 * (like optional page cache readahead) and so an OOM killer
2239 * invocation might not even be necessary.
2241 * That's why we don't do anything here except remember the
2242 * OOM context and then deal with it at the end of the page
2243 * fault when the stack is unwound, the locks are released,
2244 * and when we know whether the fault was overall successful.
2246 css_get(&memcg
->css
);
2247 current
->memcg_oom
.memcg
= memcg
;
2248 current
->memcg_oom
.gfp_mask
= mask
;
2249 current
->memcg_oom
.order
= order
;
2253 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2254 * @handle: actually kill/wait or just clean up the OOM state
2256 * This has to be called at the end of a page fault if the memcg OOM
2257 * handler was enabled.
2259 * Memcg supports userspace OOM handling where failed allocations must
2260 * sleep on a waitqueue until the userspace task resolves the
2261 * situation. Sleeping directly in the charge context with all kinds
2262 * of locks held is not a good idea, instead we remember an OOM state
2263 * in the task and mem_cgroup_oom_synchronize() has to be called at
2264 * the end of the page fault to complete the OOM handling.
2266 * Returns %true if an ongoing memcg OOM situation was detected and
2267 * completed, %false otherwise.
2269 bool mem_cgroup_oom_synchronize(bool handle
)
2271 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2272 struct oom_wait_info owait
;
2275 /* OOM is global, do not handle */
2282 owait
.memcg
= memcg
;
2283 owait
.wait
.flags
= 0;
2284 owait
.wait
.func
= memcg_oom_wake_function
;
2285 owait
.wait
.private = current
;
2286 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2288 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2289 mem_cgroup_mark_under_oom(memcg
);
2291 locked
= mem_cgroup_oom_trylock(memcg
);
2294 mem_cgroup_oom_notify(memcg
);
2296 if (locked
&& !memcg
->oom_kill_disable
) {
2297 mem_cgroup_unmark_under_oom(memcg
);
2298 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2299 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2300 current
->memcg_oom
.order
);
2303 mem_cgroup_unmark_under_oom(memcg
);
2304 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2308 mem_cgroup_oom_unlock(memcg
);
2310 * There is no guarantee that an OOM-lock contender
2311 * sees the wakeups triggered by the OOM kill
2312 * uncharges. Wake any sleepers explicitely.
2314 memcg_oom_recover(memcg
);
2317 current
->memcg_oom
.memcg
= NULL
;
2318 css_put(&memcg
->css
);
2323 * Currently used to update mapped file statistics, but the routine can be
2324 * generalized to update other statistics as well.
2326 * Notes: Race condition
2328 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2329 * it tends to be costly. But considering some conditions, we doesn't need
2330 * to do so _always_.
2332 * Considering "charge", lock_page_cgroup() is not required because all
2333 * file-stat operations happen after a page is attached to radix-tree. There
2334 * are no race with "charge".
2336 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2337 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2338 * if there are race with "uncharge". Statistics itself is properly handled
2341 * Considering "move", this is an only case we see a race. To make the race
2342 * small, we check mm->moving_account and detect there are possibility of race
2343 * If there is, we take a lock.
2346 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2347 bool *locked
, unsigned long *flags
)
2349 struct mem_cgroup
*memcg
;
2350 struct page_cgroup
*pc
;
2352 pc
= lookup_page_cgroup(page
);
2354 memcg
= pc
->mem_cgroup
;
2355 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2358 * If this memory cgroup is not under account moving, we don't
2359 * need to take move_lock_mem_cgroup(). Because we already hold
2360 * rcu_read_lock(), any calls to move_account will be delayed until
2361 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2363 if (!mem_cgroup_stolen(memcg
))
2366 move_lock_mem_cgroup(memcg
, flags
);
2367 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2368 move_unlock_mem_cgroup(memcg
, flags
);
2374 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2376 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2379 * It's guaranteed that pc->mem_cgroup never changes while
2380 * lock is held because a routine modifies pc->mem_cgroup
2381 * should take move_lock_mem_cgroup().
2383 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2386 void mem_cgroup_update_page_stat(struct page
*page
,
2387 enum mem_cgroup_stat_index idx
, int val
)
2389 struct mem_cgroup
*memcg
;
2390 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2391 unsigned long uninitialized_var(flags
);
2393 if (mem_cgroup_disabled())
2396 VM_BUG_ON(!rcu_read_lock_held());
2397 memcg
= pc
->mem_cgroup
;
2398 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2401 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2405 * size of first charge trial. "32" comes from vmscan.c's magic value.
2406 * TODO: maybe necessary to use big numbers in big irons.
2408 #define CHARGE_BATCH 32U
2409 struct memcg_stock_pcp
{
2410 struct mem_cgroup
*cached
; /* this never be root cgroup */
2411 unsigned int nr_pages
;
2412 struct work_struct work
;
2413 unsigned long flags
;
2414 #define FLUSHING_CACHED_CHARGE 0
2416 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2417 static DEFINE_MUTEX(percpu_charge_mutex
);
2420 * consume_stock: Try to consume stocked charge on this cpu.
2421 * @memcg: memcg to consume from.
2422 * @nr_pages: how many pages to charge.
2424 * The charges will only happen if @memcg matches the current cpu's memcg
2425 * stock, and at least @nr_pages are available in that stock. Failure to
2426 * service an allocation will refill the stock.
2428 * returns true if successful, false otherwise.
2430 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2432 struct memcg_stock_pcp
*stock
;
2435 if (nr_pages
> CHARGE_BATCH
)
2438 stock
= &get_cpu_var(memcg_stock
);
2439 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2440 stock
->nr_pages
-= nr_pages
;
2441 else /* need to call res_counter_charge */
2443 put_cpu_var(memcg_stock
);
2448 * Returns stocks cached in percpu to res_counter and reset cached information.
2450 static void drain_stock(struct memcg_stock_pcp
*stock
)
2452 struct mem_cgroup
*old
= stock
->cached
;
2454 if (stock
->nr_pages
) {
2455 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2457 res_counter_uncharge(&old
->res
, bytes
);
2458 if (do_swap_account
)
2459 res_counter_uncharge(&old
->memsw
, bytes
);
2460 stock
->nr_pages
= 0;
2462 stock
->cached
= NULL
;
2466 * This must be called under preempt disabled or must be called by
2467 * a thread which is pinned to local cpu.
2469 static void drain_local_stock(struct work_struct
*dummy
)
2471 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2473 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2476 static void __init
memcg_stock_init(void)
2480 for_each_possible_cpu(cpu
) {
2481 struct memcg_stock_pcp
*stock
=
2482 &per_cpu(memcg_stock
, cpu
);
2483 INIT_WORK(&stock
->work
, drain_local_stock
);
2488 * Cache charges(val) which is from res_counter, to local per_cpu area.
2489 * This will be consumed by consume_stock() function, later.
2491 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2493 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2495 if (stock
->cached
!= memcg
) { /* reset if necessary */
2497 stock
->cached
= memcg
;
2499 stock
->nr_pages
+= nr_pages
;
2500 put_cpu_var(memcg_stock
);
2504 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2505 * of the hierarchy under it. sync flag says whether we should block
2506 * until the work is done.
2508 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2512 /* Notify other cpus that system-wide "drain" is running */
2515 for_each_online_cpu(cpu
) {
2516 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2517 struct mem_cgroup
*memcg
;
2519 memcg
= stock
->cached
;
2520 if (!memcg
|| !stock
->nr_pages
)
2522 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2524 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2526 drain_local_stock(&stock
->work
);
2528 schedule_work_on(cpu
, &stock
->work
);
2536 for_each_online_cpu(cpu
) {
2537 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2538 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2539 flush_work(&stock
->work
);
2546 * Tries to drain stocked charges in other cpus. This function is asynchronous
2547 * and just put a work per cpu for draining localy on each cpu. Caller can
2548 * expects some charges will be back to res_counter later but cannot wait for
2551 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2554 * If someone calls draining, avoid adding more kworker runs.
2556 if (!mutex_trylock(&percpu_charge_mutex
))
2558 drain_all_stock(root_memcg
, false);
2559 mutex_unlock(&percpu_charge_mutex
);
2562 /* This is a synchronous drain interface. */
2563 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2565 /* called when force_empty is called */
2566 mutex_lock(&percpu_charge_mutex
);
2567 drain_all_stock(root_memcg
, true);
2568 mutex_unlock(&percpu_charge_mutex
);
2572 * This function drains percpu counter value from DEAD cpu and
2573 * move it to local cpu. Note that this function can be preempted.
2575 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2579 spin_lock(&memcg
->pcp_counter_lock
);
2580 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2581 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2583 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2584 memcg
->nocpu_base
.count
[i
] += x
;
2586 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2587 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2589 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2590 memcg
->nocpu_base
.events
[i
] += x
;
2592 spin_unlock(&memcg
->pcp_counter_lock
);
2595 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2596 unsigned long action
,
2599 int cpu
= (unsigned long)hcpu
;
2600 struct memcg_stock_pcp
*stock
;
2601 struct mem_cgroup
*iter
;
2603 if (action
== CPU_ONLINE
)
2606 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2609 for_each_mem_cgroup(iter
)
2610 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2612 stock
= &per_cpu(memcg_stock
, cpu
);
2618 /* See __mem_cgroup_try_charge() for details */
2620 CHARGE_OK
, /* success */
2621 CHARGE_RETRY
, /* need to retry but retry is not bad */
2622 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2623 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2626 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2627 unsigned int nr_pages
, unsigned int min_pages
,
2630 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2631 struct mem_cgroup
*mem_over_limit
;
2632 struct res_counter
*fail_res
;
2633 unsigned long flags
= 0;
2636 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2639 if (!do_swap_account
)
2641 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2645 res_counter_uncharge(&memcg
->res
, csize
);
2646 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2647 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2649 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2651 * Never reclaim on behalf of optional batching, retry with a
2652 * single page instead.
2654 if (nr_pages
> min_pages
)
2655 return CHARGE_RETRY
;
2657 if (!(gfp_mask
& __GFP_WAIT
))
2658 return CHARGE_WOULDBLOCK
;
2660 if (gfp_mask
& __GFP_NORETRY
)
2661 return CHARGE_NOMEM
;
2663 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2664 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2665 return CHARGE_RETRY
;
2667 * Even though the limit is exceeded at this point, reclaim
2668 * may have been able to free some pages. Retry the charge
2669 * before killing the task.
2671 * Only for regular pages, though: huge pages are rather
2672 * unlikely to succeed so close to the limit, and we fall back
2673 * to regular pages anyway in case of failure.
2675 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2676 return CHARGE_RETRY
;
2679 * At task move, charge accounts can be doubly counted. So, it's
2680 * better to wait until the end of task_move if something is going on.
2682 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2683 return CHARGE_RETRY
;
2686 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2688 return CHARGE_NOMEM
;
2692 * __mem_cgroup_try_charge() does
2693 * 1. detect memcg to be charged against from passed *mm and *ptr,
2694 * 2. update res_counter
2695 * 3. call memory reclaim if necessary.
2697 * In some special case, if the task is fatal, fatal_signal_pending() or
2698 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2699 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2700 * as possible without any hazards. 2: all pages should have a valid
2701 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2702 * pointer, that is treated as a charge to root_mem_cgroup.
2704 * So __mem_cgroup_try_charge() will return
2705 * 0 ... on success, filling *ptr with a valid memcg pointer.
2706 * -ENOMEM ... charge failure because of resource limits.
2707 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2709 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2710 * the oom-killer can be invoked.
2712 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2714 unsigned int nr_pages
,
2715 struct mem_cgroup
**ptr
,
2718 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2719 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2720 struct mem_cgroup
*memcg
= NULL
;
2724 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2725 * in system level. So, allow to go ahead dying process in addition to
2728 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2729 || fatal_signal_pending(current
)))
2732 if (unlikely(task_in_memcg_oom(current
)))
2735 if (gfp_mask
& __GFP_NOFAIL
)
2739 * We always charge the cgroup the mm_struct belongs to.
2740 * The mm_struct's mem_cgroup changes on task migration if the
2741 * thread group leader migrates. It's possible that mm is not
2742 * set, if so charge the root memcg (happens for pagecache usage).
2745 *ptr
= root_mem_cgroup
;
2747 if (*ptr
) { /* css should be a valid one */
2749 if (mem_cgroup_is_root(memcg
))
2751 if (consume_stock(memcg
, nr_pages
))
2753 css_get(&memcg
->css
);
2755 struct task_struct
*p
;
2758 p
= rcu_dereference(mm
->owner
);
2760 * Because we don't have task_lock(), "p" can exit.
2761 * In that case, "memcg" can point to root or p can be NULL with
2762 * race with swapoff. Then, we have small risk of mis-accouning.
2763 * But such kind of mis-account by race always happens because
2764 * we don't have cgroup_mutex(). It's overkill and we allo that
2766 * (*) swapoff at el will charge against mm-struct not against
2767 * task-struct. So, mm->owner can be NULL.
2769 memcg
= mem_cgroup_from_task(p
);
2771 memcg
= root_mem_cgroup
;
2772 if (mem_cgroup_is_root(memcg
)) {
2776 if (consume_stock(memcg
, nr_pages
)) {
2778 * It seems dagerous to access memcg without css_get().
2779 * But considering how consume_stok works, it's not
2780 * necessary. If consume_stock success, some charges
2781 * from this memcg are cached on this cpu. So, we
2782 * don't need to call css_get()/css_tryget() before
2783 * calling consume_stock().
2788 /* after here, we may be blocked. we need to get refcnt */
2789 if (!css_tryget(&memcg
->css
)) {
2797 bool invoke_oom
= oom
&& !nr_oom_retries
;
2799 /* If killed, bypass charge */
2800 if (fatal_signal_pending(current
)) {
2801 css_put(&memcg
->css
);
2805 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2806 nr_pages
, invoke_oom
);
2810 case CHARGE_RETRY
: /* not in OOM situation but retry */
2812 css_put(&memcg
->css
);
2815 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2816 css_put(&memcg
->css
);
2818 case CHARGE_NOMEM
: /* OOM routine works */
2819 if (!oom
|| invoke_oom
) {
2820 css_put(&memcg
->css
);
2826 } while (ret
!= CHARGE_OK
);
2828 if (batch
> nr_pages
)
2829 refill_stock(memcg
, batch
- nr_pages
);
2830 css_put(&memcg
->css
);
2835 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2840 *ptr
= root_mem_cgroup
;
2845 * Somemtimes we have to undo a charge we got by try_charge().
2846 * This function is for that and do uncharge, put css's refcnt.
2847 * gotten by try_charge().
2849 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2850 unsigned int nr_pages
)
2852 if (!mem_cgroup_is_root(memcg
)) {
2853 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2855 res_counter_uncharge(&memcg
->res
, bytes
);
2856 if (do_swap_account
)
2857 res_counter_uncharge(&memcg
->memsw
, bytes
);
2862 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2863 * This is useful when moving usage to parent cgroup.
2865 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2866 unsigned int nr_pages
)
2868 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2870 if (mem_cgroup_is_root(memcg
))
2873 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2874 if (do_swap_account
)
2875 res_counter_uncharge_until(&memcg
->memsw
,
2876 memcg
->memsw
.parent
, bytes
);
2880 * A helper function to get mem_cgroup from ID. must be called under
2881 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2882 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2883 * called against removed memcg.)
2885 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2887 /* ID 0 is unused ID */
2890 return mem_cgroup_from_id(id
);
2893 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2895 struct mem_cgroup
*memcg
= NULL
;
2896 struct page_cgroup
*pc
;
2900 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2902 pc
= lookup_page_cgroup(page
);
2903 lock_page_cgroup(pc
);
2904 if (PageCgroupUsed(pc
)) {
2905 memcg
= pc
->mem_cgroup
;
2906 if (memcg
&& !css_tryget(&memcg
->css
))
2908 } else if (PageSwapCache(page
)) {
2909 ent
.val
= page_private(page
);
2910 id
= lookup_swap_cgroup_id(ent
);
2912 memcg
= mem_cgroup_lookup(id
);
2913 if (memcg
&& !css_tryget(&memcg
->css
))
2917 unlock_page_cgroup(pc
);
2921 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2923 unsigned int nr_pages
,
2924 enum charge_type ctype
,
2927 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2928 struct zone
*uninitialized_var(zone
);
2929 struct lruvec
*lruvec
;
2930 bool was_on_lru
= false;
2933 lock_page_cgroup(pc
);
2934 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2936 * we don't need page_cgroup_lock about tail pages, becase they are not
2937 * accessed by any other context at this point.
2941 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2942 * may already be on some other mem_cgroup's LRU. Take care of it.
2945 zone
= page_zone(page
);
2946 spin_lock_irq(&zone
->lru_lock
);
2947 if (PageLRU(page
)) {
2948 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2950 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2955 pc
->mem_cgroup
= memcg
;
2957 * We access a page_cgroup asynchronously without lock_page_cgroup().
2958 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2959 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2960 * before USED bit, we need memory barrier here.
2961 * See mem_cgroup_add_lru_list(), etc.
2964 SetPageCgroupUsed(pc
);
2968 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2969 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2971 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2973 spin_unlock_irq(&zone
->lru_lock
);
2976 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2981 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2982 unlock_page_cgroup(pc
);
2985 * "charge_statistics" updated event counter. Then, check it.
2986 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2987 * if they exceeds softlimit.
2989 memcg_check_events(memcg
, page
);
2992 static DEFINE_MUTEX(set_limit_mutex
);
2994 #ifdef CONFIG_MEMCG_KMEM
2995 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2997 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2998 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
) ==
2999 KMEM_ACCOUNTED_MASK
;
3003 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
3004 * in the memcg_cache_params struct.
3006 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
3008 struct kmem_cache
*cachep
;
3010 VM_BUG_ON(p
->is_root_cache
);
3011 cachep
= p
->root_cache
;
3012 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
3015 #ifdef CONFIG_SLABINFO
3016 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
3018 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3019 struct memcg_cache_params
*params
;
3021 if (!memcg_can_account_kmem(memcg
))
3024 print_slabinfo_header(m
);
3026 mutex_lock(&memcg
->slab_caches_mutex
);
3027 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
3028 cache_show(memcg_params_to_cache(params
), m
);
3029 mutex_unlock(&memcg
->slab_caches_mutex
);
3035 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
3037 struct res_counter
*fail_res
;
3038 struct mem_cgroup
*_memcg
;
3041 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
3046 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
3047 &_memcg
, oom_gfp_allowed(gfp
));
3049 if (ret
== -EINTR
) {
3051 * __mem_cgroup_try_charge() chosed to bypass to root due to
3052 * OOM kill or fatal signal. Since our only options are to
3053 * either fail the allocation or charge it to this cgroup, do
3054 * it as a temporary condition. But we can't fail. From a
3055 * kmem/slab perspective, the cache has already been selected,
3056 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3059 * This condition will only trigger if the task entered
3060 * memcg_charge_kmem in a sane state, but was OOM-killed during
3061 * __mem_cgroup_try_charge() above. Tasks that were already
3062 * dying when the allocation triggers should have been already
3063 * directed to the root cgroup in memcontrol.h
3065 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3066 if (do_swap_account
)
3067 res_counter_charge_nofail(&memcg
->memsw
, size
,
3071 res_counter_uncharge(&memcg
->kmem
, size
);
3076 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3078 res_counter_uncharge(&memcg
->res
, size
);
3079 if (do_swap_account
)
3080 res_counter_uncharge(&memcg
->memsw
, size
);
3083 if (res_counter_uncharge(&memcg
->kmem
, size
))
3087 * Releases a reference taken in kmem_cgroup_css_offline in case
3088 * this last uncharge is racing with the offlining code or it is
3089 * outliving the memcg existence.
3091 * The memory barrier imposed by test&clear is paired with the
3092 * explicit one in memcg_kmem_mark_dead().
3094 if (memcg_kmem_test_and_clear_dead(memcg
))
3095 css_put(&memcg
->css
);
3099 * helper for acessing a memcg's index. It will be used as an index in the
3100 * child cache array in kmem_cache, and also to derive its name. This function
3101 * will return -1 when this is not a kmem-limited memcg.
3103 int memcg_cache_id(struct mem_cgroup
*memcg
)
3105 return memcg
? memcg
->kmemcg_id
: -1;
3109 * This ends up being protected by the set_limit mutex, during normal
3110 * operation, because that is its main call site.
3112 * But when we create a new cache, we can call this as well if its parent
3113 * is kmem-limited. That will have to hold set_limit_mutex as well.
3115 static int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3119 num
= ida_simple_get(&kmem_limited_groups
,
3120 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3124 * After this point, kmem_accounted (that we test atomically in
3125 * the beginning of this conditional), is no longer 0. This
3126 * guarantees only one process will set the following boolean
3127 * to true. We don't need test_and_set because we're protected
3128 * by the set_limit_mutex anyway.
3130 memcg_kmem_set_activated(memcg
);
3132 ret
= memcg_update_all_caches(num
+1);
3134 ida_simple_remove(&kmem_limited_groups
, num
);
3135 memcg_kmem_clear_activated(memcg
);
3139 memcg
->kmemcg_id
= num
;
3140 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3141 mutex_init(&memcg
->slab_caches_mutex
);
3145 static size_t memcg_caches_array_size(int num_groups
)
3148 if (num_groups
<= 0)
3151 size
= 2 * num_groups
;
3152 if (size
< MEMCG_CACHES_MIN_SIZE
)
3153 size
= MEMCG_CACHES_MIN_SIZE
;
3154 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3155 size
= MEMCG_CACHES_MAX_SIZE
;
3161 * We should update the current array size iff all caches updates succeed. This
3162 * can only be done from the slab side. The slab mutex needs to be held when
3165 void memcg_update_array_size(int num
)
3167 if (num
> memcg_limited_groups_array_size
)
3168 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3171 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3173 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3175 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3177 VM_BUG_ON(!is_root_cache(s
));
3179 if (num_groups
> memcg_limited_groups_array_size
) {
3181 ssize_t size
= memcg_caches_array_size(num_groups
);
3183 size
*= sizeof(void *);
3184 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3186 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3187 if (!s
->memcg_params
) {
3188 s
->memcg_params
= cur_params
;
3192 s
->memcg_params
->is_root_cache
= true;
3195 * There is the chance it will be bigger than
3196 * memcg_limited_groups_array_size, if we failed an allocation
3197 * in a cache, in which case all caches updated before it, will
3198 * have a bigger array.
3200 * But if that is the case, the data after
3201 * memcg_limited_groups_array_size is certainly unused
3203 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3204 if (!cur_params
->memcg_caches
[i
])
3206 s
->memcg_params
->memcg_caches
[i
] =
3207 cur_params
->memcg_caches
[i
];
3211 * Ideally, we would wait until all caches succeed, and only
3212 * then free the old one. But this is not worth the extra
3213 * pointer per-cache we'd have to have for this.
3215 * It is not a big deal if some caches are left with a size
3216 * bigger than the others. And all updates will reset this
3224 int memcg_alloc_cache_params(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3225 struct kmem_cache
*root_cache
)
3229 if (!memcg_kmem_enabled())
3233 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3234 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3236 size
= sizeof(struct memcg_cache_params
);
3238 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3239 if (!s
->memcg_params
)
3243 s
->memcg_params
->memcg
= memcg
;
3244 s
->memcg_params
->root_cache
= root_cache
;
3245 INIT_WORK(&s
->memcg_params
->destroy
,
3246 kmem_cache_destroy_work_func
);
3248 s
->memcg_params
->is_root_cache
= true;
3253 void memcg_free_cache_params(struct kmem_cache
*s
)
3255 kfree(s
->memcg_params
);
3258 void memcg_register_cache(struct kmem_cache
*s
)
3260 struct kmem_cache
*root
;
3261 struct mem_cgroup
*memcg
;
3264 if (is_root_cache(s
))
3267 root
= s
->memcg_params
->root_cache
;
3268 memcg
= s
->memcg_params
->memcg
;
3269 id
= memcg_cache_id(memcg
);
3271 css_get(&memcg
->css
);
3273 mutex_lock(&memcg
->slab_caches_mutex
);
3274 list_add(&s
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3275 mutex_unlock(&memcg
->slab_caches_mutex
);
3278 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3279 * barrier here to ensure nobody will see the kmem_cache partially
3284 root
->memcg_params
->memcg_caches
[id
] = s
;
3287 void memcg_unregister_cache(struct kmem_cache
*s
)
3289 struct kmem_cache
*root
;
3290 struct mem_cgroup
*memcg
;
3293 if (is_root_cache(s
))
3296 memcg
= s
->memcg_params
->memcg
;
3297 id
= memcg_cache_id(memcg
);
3299 root
= s
->memcg_params
->root_cache
;
3300 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3302 mutex_lock(&memcg
->slab_caches_mutex
);
3303 list_del(&s
->memcg_params
->list
);
3304 mutex_unlock(&memcg
->slab_caches_mutex
);
3306 css_put(&memcg
->css
);
3310 * During the creation a new cache, we need to disable our accounting mechanism
3311 * altogether. This is true even if we are not creating, but rather just
3312 * enqueing new caches to be created.
3314 * This is because that process will trigger allocations; some visible, like
3315 * explicit kmallocs to auxiliary data structures, name strings and internal
3316 * cache structures; some well concealed, like INIT_WORK() that can allocate
3317 * objects during debug.
3319 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3320 * to it. This may not be a bounded recursion: since the first cache creation
3321 * failed to complete (waiting on the allocation), we'll just try to create the
3322 * cache again, failing at the same point.
3324 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3325 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3326 * inside the following two functions.
3328 static inline void memcg_stop_kmem_account(void)
3330 VM_BUG_ON(!current
->mm
);
3331 current
->memcg_kmem_skip_account
++;
3334 static inline void memcg_resume_kmem_account(void)
3336 VM_BUG_ON(!current
->mm
);
3337 current
->memcg_kmem_skip_account
--;
3340 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3342 struct kmem_cache
*cachep
;
3343 struct memcg_cache_params
*p
;
3345 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3347 cachep
= memcg_params_to_cache(p
);
3350 * If we get down to 0 after shrink, we could delete right away.
3351 * However, memcg_release_pages() already puts us back in the workqueue
3352 * in that case. If we proceed deleting, we'll get a dangling
3353 * reference, and removing the object from the workqueue in that case
3354 * is unnecessary complication. We are not a fast path.
3356 * Note that this case is fundamentally different from racing with
3357 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3358 * kmem_cache_shrink, not only we would be reinserting a dead cache
3359 * into the queue, but doing so from inside the worker racing to
3362 * So if we aren't down to zero, we'll just schedule a worker and try
3365 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3366 kmem_cache_shrink(cachep
);
3367 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3370 kmem_cache_destroy(cachep
);
3373 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3375 if (!cachep
->memcg_params
->dead
)
3379 * There are many ways in which we can get here.
3381 * We can get to a memory-pressure situation while the delayed work is
3382 * still pending to run. The vmscan shrinkers can then release all
3383 * cache memory and get us to destruction. If this is the case, we'll
3384 * be executed twice, which is a bug (the second time will execute over
3385 * bogus data). In this case, cancelling the work should be fine.
3387 * But we can also get here from the worker itself, if
3388 * kmem_cache_shrink is enough to shake all the remaining objects and
3389 * get the page count to 0. In this case, we'll deadlock if we try to
3390 * cancel the work (the worker runs with an internal lock held, which
3391 * is the same lock we would hold for cancel_work_sync().)
3393 * Since we can't possibly know who got us here, just refrain from
3394 * running if there is already work pending
3396 if (work_pending(&cachep
->memcg_params
->destroy
))
3399 * We have to defer the actual destroying to a workqueue, because
3400 * we might currently be in a context that cannot sleep.
3402 schedule_work(&cachep
->memcg_params
->destroy
);
3406 * This lock protects updaters, not readers. We want readers to be as fast as
3407 * they can, and they will either see NULL or a valid cache value. Our model
3408 * allow them to see NULL, in which case the root memcg will be selected.
3410 * We need this lock because multiple allocations to the same cache from a non
3411 * will span more than one worker. Only one of them can create the cache.
3413 static DEFINE_MUTEX(memcg_cache_mutex
);
3416 * Called with memcg_cache_mutex held
3418 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3419 struct kmem_cache
*s
)
3421 struct kmem_cache
*new;
3422 static char *tmp_name
= NULL
;
3424 lockdep_assert_held(&memcg_cache_mutex
);
3427 * kmem_cache_create_memcg duplicates the given name and
3428 * cgroup_name for this name requires RCU context.
3429 * This static temporary buffer is used to prevent from
3430 * pointless shortliving allocation.
3433 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3439 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3440 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3443 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3444 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3447 new->allocflags
|= __GFP_KMEMCG
;
3452 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3453 struct kmem_cache
*cachep
)
3455 struct kmem_cache
*new_cachep
;
3458 BUG_ON(!memcg_can_account_kmem(memcg
));
3460 idx
= memcg_cache_id(memcg
);
3462 mutex_lock(&memcg_cache_mutex
);
3463 new_cachep
= cache_from_memcg_idx(cachep
, idx
);
3467 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3468 if (new_cachep
== NULL
)
3469 new_cachep
= cachep
;
3472 mutex_unlock(&memcg_cache_mutex
);
3476 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3478 struct kmem_cache
*c
;
3481 if (!s
->memcg_params
)
3483 if (!s
->memcg_params
->is_root_cache
)
3487 * If the cache is being destroyed, we trust that there is no one else
3488 * requesting objects from it. Even if there are, the sanity checks in
3489 * kmem_cache_destroy should caught this ill-case.
3491 * Still, we don't want anyone else freeing memcg_caches under our
3492 * noses, which can happen if a new memcg comes to life. As usual,
3493 * we'll take the set_limit_mutex to protect ourselves against this.
3495 mutex_lock(&set_limit_mutex
);
3496 for_each_memcg_cache_index(i
) {
3497 c
= cache_from_memcg_idx(s
, i
);
3502 * We will now manually delete the caches, so to avoid races
3503 * we need to cancel all pending destruction workers and
3504 * proceed with destruction ourselves.
3506 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3507 * and that could spawn the workers again: it is likely that
3508 * the cache still have active pages until this very moment.
3509 * This would lead us back to mem_cgroup_destroy_cache.
3511 * But that will not execute at all if the "dead" flag is not
3512 * set, so flip it down to guarantee we are in control.
3514 c
->memcg_params
->dead
= false;
3515 cancel_work_sync(&c
->memcg_params
->destroy
);
3516 kmem_cache_destroy(c
);
3518 mutex_unlock(&set_limit_mutex
);
3521 struct create_work
{
3522 struct mem_cgroup
*memcg
;
3523 struct kmem_cache
*cachep
;
3524 struct work_struct work
;
3527 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3529 struct kmem_cache
*cachep
;
3530 struct memcg_cache_params
*params
;
3532 if (!memcg_kmem_is_active(memcg
))
3535 mutex_lock(&memcg
->slab_caches_mutex
);
3536 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3537 cachep
= memcg_params_to_cache(params
);
3538 cachep
->memcg_params
->dead
= true;
3539 schedule_work(&cachep
->memcg_params
->destroy
);
3541 mutex_unlock(&memcg
->slab_caches_mutex
);
3544 static void memcg_create_cache_work_func(struct work_struct
*w
)
3546 struct create_work
*cw
;
3548 cw
= container_of(w
, struct create_work
, work
);
3549 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3550 css_put(&cw
->memcg
->css
);
3555 * Enqueue the creation of a per-memcg kmem_cache.
3557 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3558 struct kmem_cache
*cachep
)
3560 struct create_work
*cw
;
3562 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3564 css_put(&memcg
->css
);
3569 cw
->cachep
= cachep
;
3571 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3572 schedule_work(&cw
->work
);
3575 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3576 struct kmem_cache
*cachep
)
3579 * We need to stop accounting when we kmalloc, because if the
3580 * corresponding kmalloc cache is not yet created, the first allocation
3581 * in __memcg_create_cache_enqueue will recurse.
3583 * However, it is better to enclose the whole function. Depending on
3584 * the debugging options enabled, INIT_WORK(), for instance, can
3585 * trigger an allocation. This too, will make us recurse. Because at
3586 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3587 * the safest choice is to do it like this, wrapping the whole function.
3589 memcg_stop_kmem_account();
3590 __memcg_create_cache_enqueue(memcg
, cachep
);
3591 memcg_resume_kmem_account();
3594 * Return the kmem_cache we're supposed to use for a slab allocation.
3595 * We try to use the current memcg's version of the cache.
3597 * If the cache does not exist yet, if we are the first user of it,
3598 * we either create it immediately, if possible, or create it asynchronously
3600 * In the latter case, we will let the current allocation go through with
3601 * the original cache.
3603 * Can't be called in interrupt context or from kernel threads.
3604 * This function needs to be called with rcu_read_lock() held.
3606 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3609 struct mem_cgroup
*memcg
;
3610 struct kmem_cache
*memcg_cachep
;
3612 VM_BUG_ON(!cachep
->memcg_params
);
3613 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3615 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3619 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3621 if (!memcg_can_account_kmem(memcg
))
3624 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3625 if (likely(memcg_cachep
)) {
3626 cachep
= memcg_cachep
;
3630 /* The corresponding put will be done in the workqueue. */
3631 if (!css_tryget(&memcg
->css
))
3636 * If we are in a safe context (can wait, and not in interrupt
3637 * context), we could be be predictable and return right away.
3638 * This would guarantee that the allocation being performed
3639 * already belongs in the new cache.
3641 * However, there are some clashes that can arrive from locking.
3642 * For instance, because we acquire the slab_mutex while doing
3643 * kmem_cache_dup, this means no further allocation could happen
3644 * with the slab_mutex held.
3646 * Also, because cache creation issue get_online_cpus(), this
3647 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3648 * that ends up reversed during cpu hotplug. (cpuset allocates
3649 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3650 * better to defer everything.
3652 memcg_create_cache_enqueue(memcg
, cachep
);
3658 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3661 * We need to verify if the allocation against current->mm->owner's memcg is
3662 * possible for the given order. But the page is not allocated yet, so we'll
3663 * need a further commit step to do the final arrangements.
3665 * It is possible for the task to switch cgroups in this mean time, so at
3666 * commit time, we can't rely on task conversion any longer. We'll then use
3667 * the handle argument to return to the caller which cgroup we should commit
3668 * against. We could also return the memcg directly and avoid the pointer
3669 * passing, but a boolean return value gives better semantics considering
3670 * the compiled-out case as well.
3672 * Returning true means the allocation is possible.
3675 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3677 struct mem_cgroup
*memcg
;
3683 * Disabling accounting is only relevant for some specific memcg
3684 * internal allocations. Therefore we would initially not have such
3685 * check here, since direct calls to the page allocator that are marked
3686 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3687 * concerned with cache allocations, and by having this test at
3688 * memcg_kmem_get_cache, we are already able to relay the allocation to
3689 * the root cache and bypass the memcg cache altogether.
3691 * There is one exception, though: the SLUB allocator does not create
3692 * large order caches, but rather service large kmallocs directly from
3693 * the page allocator. Therefore, the following sequence when backed by
3694 * the SLUB allocator:
3696 * memcg_stop_kmem_account();
3697 * kmalloc(<large_number>)
3698 * memcg_resume_kmem_account();
3700 * would effectively ignore the fact that we should skip accounting,
3701 * since it will drive us directly to this function without passing
3702 * through the cache selector memcg_kmem_get_cache. Such large
3703 * allocations are extremely rare but can happen, for instance, for the
3704 * cache arrays. We bring this test here.
3706 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3709 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3712 * very rare case described in mem_cgroup_from_task. Unfortunately there
3713 * isn't much we can do without complicating this too much, and it would
3714 * be gfp-dependent anyway. Just let it go
3716 if (unlikely(!memcg
))
3719 if (!memcg_can_account_kmem(memcg
)) {
3720 css_put(&memcg
->css
);
3724 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3728 css_put(&memcg
->css
);
3732 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3735 struct page_cgroup
*pc
;
3737 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3739 /* The page allocation failed. Revert */
3741 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3745 pc
= lookup_page_cgroup(page
);
3746 lock_page_cgroup(pc
);
3747 pc
->mem_cgroup
= memcg
;
3748 SetPageCgroupUsed(pc
);
3749 unlock_page_cgroup(pc
);
3752 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3754 struct mem_cgroup
*memcg
= NULL
;
3755 struct page_cgroup
*pc
;
3758 pc
= lookup_page_cgroup(page
);
3760 * Fast unlocked return. Theoretically might have changed, have to
3761 * check again after locking.
3763 if (!PageCgroupUsed(pc
))
3766 lock_page_cgroup(pc
);
3767 if (PageCgroupUsed(pc
)) {
3768 memcg
= pc
->mem_cgroup
;
3769 ClearPageCgroupUsed(pc
);
3771 unlock_page_cgroup(pc
);
3774 * We trust that only if there is a memcg associated with the page, it
3775 * is a valid allocation
3780 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3781 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3784 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3787 #endif /* CONFIG_MEMCG_KMEM */
3789 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3791 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3793 * Because tail pages are not marked as "used", set it. We're under
3794 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3795 * charge/uncharge will be never happen and move_account() is done under
3796 * compound_lock(), so we don't have to take care of races.
3798 void mem_cgroup_split_huge_fixup(struct page
*head
)
3800 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3801 struct page_cgroup
*pc
;
3802 struct mem_cgroup
*memcg
;
3805 if (mem_cgroup_disabled())
3808 memcg
= head_pc
->mem_cgroup
;
3809 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3811 pc
->mem_cgroup
= memcg
;
3812 smp_wmb();/* see __commit_charge() */
3813 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3815 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3818 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3821 void mem_cgroup_move_account_page_stat(struct mem_cgroup
*from
,
3822 struct mem_cgroup
*to
,
3823 unsigned int nr_pages
,
3824 enum mem_cgroup_stat_index idx
)
3826 /* Update stat data for mem_cgroup */
3828 __this_cpu_sub(from
->stat
->count
[idx
], nr_pages
);
3829 __this_cpu_add(to
->stat
->count
[idx
], nr_pages
);
3834 * mem_cgroup_move_account - move account of the page
3836 * @nr_pages: number of regular pages (>1 for huge pages)
3837 * @pc: page_cgroup of the page.
3838 * @from: mem_cgroup which the page is moved from.
3839 * @to: mem_cgroup which the page is moved to. @from != @to.
3841 * The caller must confirm following.
3842 * - page is not on LRU (isolate_page() is useful.)
3843 * - compound_lock is held when nr_pages > 1
3845 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3848 static int mem_cgroup_move_account(struct page
*page
,
3849 unsigned int nr_pages
,
3850 struct page_cgroup
*pc
,
3851 struct mem_cgroup
*from
,
3852 struct mem_cgroup
*to
)
3854 unsigned long flags
;
3856 bool anon
= PageAnon(page
);
3858 VM_BUG_ON(from
== to
);
3859 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3861 * The page is isolated from LRU. So, collapse function
3862 * will not handle this page. But page splitting can happen.
3863 * Do this check under compound_page_lock(). The caller should
3867 if (nr_pages
> 1 && !PageTransHuge(page
))
3870 lock_page_cgroup(pc
);
3873 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3876 move_lock_mem_cgroup(from
, &flags
);
3878 if (!anon
&& page_mapped(page
))
3879 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3880 MEM_CGROUP_STAT_FILE_MAPPED
);
3882 if (PageWriteback(page
))
3883 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3884 MEM_CGROUP_STAT_WRITEBACK
);
3886 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3888 /* caller should have done css_get */
3889 pc
->mem_cgroup
= to
;
3890 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3891 move_unlock_mem_cgroup(from
, &flags
);
3894 unlock_page_cgroup(pc
);
3898 memcg_check_events(to
, page
);
3899 memcg_check_events(from
, page
);
3905 * mem_cgroup_move_parent - moves page to the parent group
3906 * @page: the page to move
3907 * @pc: page_cgroup of the page
3908 * @child: page's cgroup
3910 * move charges to its parent or the root cgroup if the group has no
3911 * parent (aka use_hierarchy==0).
3912 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3913 * mem_cgroup_move_account fails) the failure is always temporary and
3914 * it signals a race with a page removal/uncharge or migration. In the
3915 * first case the page is on the way out and it will vanish from the LRU
3916 * on the next attempt and the call should be retried later.
3917 * Isolation from the LRU fails only if page has been isolated from
3918 * the LRU since we looked at it and that usually means either global
3919 * reclaim or migration going on. The page will either get back to the
3921 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3922 * (!PageCgroupUsed) or moved to a different group. The page will
3923 * disappear in the next attempt.
3925 static int mem_cgroup_move_parent(struct page
*page
,
3926 struct page_cgroup
*pc
,
3927 struct mem_cgroup
*child
)
3929 struct mem_cgroup
*parent
;
3930 unsigned int nr_pages
;
3931 unsigned long uninitialized_var(flags
);
3934 VM_BUG_ON(mem_cgroup_is_root(child
));
3937 if (!get_page_unless_zero(page
))
3939 if (isolate_lru_page(page
))
3942 nr_pages
= hpage_nr_pages(page
);
3944 parent
= parent_mem_cgroup(child
);
3946 * If no parent, move charges to root cgroup.
3949 parent
= root_mem_cgroup
;
3952 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3953 flags
= compound_lock_irqsave(page
);
3956 ret
= mem_cgroup_move_account(page
, nr_pages
,
3959 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3962 compound_unlock_irqrestore(page
, flags
);
3963 putback_lru_page(page
);
3971 * Charge the memory controller for page usage.
3973 * 0 if the charge was successful
3974 * < 0 if the cgroup is over its limit
3976 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3977 gfp_t gfp_mask
, enum charge_type ctype
)
3979 struct mem_cgroup
*memcg
= NULL
;
3980 unsigned int nr_pages
= 1;
3984 if (PageTransHuge(page
)) {
3985 nr_pages
<<= compound_order(page
);
3986 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3988 * Never OOM-kill a process for a huge page. The
3989 * fault handler will fall back to regular pages.
3994 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3997 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
4001 int mem_cgroup_newpage_charge(struct page
*page
,
4002 struct mm_struct
*mm
, gfp_t gfp_mask
)
4004 if (mem_cgroup_disabled())
4006 VM_BUG_ON_PAGE(page_mapped(page
), page
);
4007 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
4009 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
4010 MEM_CGROUP_CHARGE_TYPE_ANON
);
4014 * While swap-in, try_charge -> commit or cancel, the page is locked.
4015 * And when try_charge() successfully returns, one refcnt to memcg without
4016 * struct page_cgroup is acquired. This refcnt will be consumed by
4017 * "commit()" or removed by "cancel()"
4019 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
4022 struct mem_cgroup
**memcgp
)
4024 struct mem_cgroup
*memcg
;
4025 struct page_cgroup
*pc
;
4028 pc
= lookup_page_cgroup(page
);
4030 * Every swap fault against a single page tries to charge the
4031 * page, bail as early as possible. shmem_unuse() encounters
4032 * already charged pages, too. The USED bit is protected by
4033 * the page lock, which serializes swap cache removal, which
4034 * in turn serializes uncharging.
4036 if (PageCgroupUsed(pc
))
4038 if (!do_swap_account
)
4040 memcg
= try_get_mem_cgroup_from_page(page
);
4044 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
4045 css_put(&memcg
->css
);
4050 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
4056 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
4057 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
4060 if (mem_cgroup_disabled())
4063 * A racing thread's fault, or swapoff, may have already
4064 * updated the pte, and even removed page from swap cache: in
4065 * those cases unuse_pte()'s pte_same() test will fail; but
4066 * there's also a KSM case which does need to charge the page.
4068 if (!PageSwapCache(page
)) {
4071 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4076 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4079 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4081 if (mem_cgroup_disabled())
4085 __mem_cgroup_cancel_charge(memcg
, 1);
4089 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4090 enum charge_type ctype
)
4092 if (mem_cgroup_disabled())
4097 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4099 * Now swap is on-memory. This means this page may be
4100 * counted both as mem and swap....double count.
4101 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4102 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4103 * may call delete_from_swap_cache() before reach here.
4105 if (do_swap_account
&& PageSwapCache(page
)) {
4106 swp_entry_t ent
= {.val
= page_private(page
)};
4107 mem_cgroup_uncharge_swap(ent
);
4111 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4112 struct mem_cgroup
*memcg
)
4114 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4115 MEM_CGROUP_CHARGE_TYPE_ANON
);
4118 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4121 struct mem_cgroup
*memcg
= NULL
;
4122 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4125 if (mem_cgroup_disabled())
4127 if (PageCompound(page
))
4130 if (!PageSwapCache(page
))
4131 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4132 else { /* page is swapcache/shmem */
4133 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4136 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4141 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4142 unsigned int nr_pages
,
4143 const enum charge_type ctype
)
4145 struct memcg_batch_info
*batch
= NULL
;
4146 bool uncharge_memsw
= true;
4148 /* If swapout, usage of swap doesn't decrease */
4149 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4150 uncharge_memsw
= false;
4152 batch
= ¤t
->memcg_batch
;
4154 * In usual, we do css_get() when we remember memcg pointer.
4155 * But in this case, we keep res->usage until end of a series of
4156 * uncharges. Then, it's ok to ignore memcg's refcnt.
4159 batch
->memcg
= memcg
;
4161 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4162 * In those cases, all pages freed continuously can be expected to be in
4163 * the same cgroup and we have chance to coalesce uncharges.
4164 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4165 * because we want to do uncharge as soon as possible.
4168 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4169 goto direct_uncharge
;
4172 goto direct_uncharge
;
4175 * In typical case, batch->memcg == mem. This means we can
4176 * merge a series of uncharges to an uncharge of res_counter.
4177 * If not, we uncharge res_counter ony by one.
4179 if (batch
->memcg
!= memcg
)
4180 goto direct_uncharge
;
4181 /* remember freed charge and uncharge it later */
4184 batch
->memsw_nr_pages
++;
4187 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4189 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4190 if (unlikely(batch
->memcg
!= memcg
))
4191 memcg_oom_recover(memcg
);
4195 * uncharge if !page_mapped(page)
4197 static struct mem_cgroup
*
4198 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4201 struct mem_cgroup
*memcg
= NULL
;
4202 unsigned int nr_pages
= 1;
4203 struct page_cgroup
*pc
;
4206 if (mem_cgroup_disabled())
4209 if (PageTransHuge(page
)) {
4210 nr_pages
<<= compound_order(page
);
4211 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
4214 * Check if our page_cgroup is valid
4216 pc
= lookup_page_cgroup(page
);
4217 if (unlikely(!PageCgroupUsed(pc
)))
4220 lock_page_cgroup(pc
);
4222 memcg
= pc
->mem_cgroup
;
4224 if (!PageCgroupUsed(pc
))
4227 anon
= PageAnon(page
);
4230 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4232 * Generally PageAnon tells if it's the anon statistics to be
4233 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4234 * used before page reached the stage of being marked PageAnon.
4238 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4239 /* See mem_cgroup_prepare_migration() */
4240 if (page_mapped(page
))
4243 * Pages under migration may not be uncharged. But
4244 * end_migration() /must/ be the one uncharging the
4245 * unused post-migration page and so it has to call
4246 * here with the migration bit still set. See the
4247 * res_counter handling below.
4249 if (!end_migration
&& PageCgroupMigration(pc
))
4252 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4253 if (!PageAnon(page
)) { /* Shared memory */
4254 if (page
->mapping
&& !page_is_file_cache(page
))
4256 } else if (page_mapped(page
)) /* Anon */
4263 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4265 ClearPageCgroupUsed(pc
);
4267 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4268 * freed from LRU. This is safe because uncharged page is expected not
4269 * to be reused (freed soon). Exception is SwapCache, it's handled by
4270 * special functions.
4273 unlock_page_cgroup(pc
);
4275 * even after unlock, we have memcg->res.usage here and this memcg
4276 * will never be freed, so it's safe to call css_get().
4278 memcg_check_events(memcg
, page
);
4279 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4280 mem_cgroup_swap_statistics(memcg
, true);
4281 css_get(&memcg
->css
);
4284 * Migration does not charge the res_counter for the
4285 * replacement page, so leave it alone when phasing out the
4286 * page that is unused after the migration.
4288 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4289 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4294 unlock_page_cgroup(pc
);
4298 void mem_cgroup_uncharge_page(struct page
*page
)
4301 if (page_mapped(page
))
4303 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
4305 * If the page is in swap cache, uncharge should be deferred
4306 * to the swap path, which also properly accounts swap usage
4307 * and handles memcg lifetime.
4309 * Note that this check is not stable and reclaim may add the
4310 * page to swap cache at any time after this. However, if the
4311 * page is not in swap cache by the time page->mapcount hits
4312 * 0, there won't be any page table references to the swap
4313 * slot, and reclaim will free it and not actually write the
4316 if (PageSwapCache(page
))
4318 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4321 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4323 VM_BUG_ON_PAGE(page_mapped(page
), page
);
4324 VM_BUG_ON_PAGE(page
->mapping
, page
);
4325 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4329 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4330 * In that cases, pages are freed continuously and we can expect pages
4331 * are in the same memcg. All these calls itself limits the number of
4332 * pages freed at once, then uncharge_start/end() is called properly.
4333 * This may be called prural(2) times in a context,
4336 void mem_cgroup_uncharge_start(void)
4338 current
->memcg_batch
.do_batch
++;
4339 /* We can do nest. */
4340 if (current
->memcg_batch
.do_batch
== 1) {
4341 current
->memcg_batch
.memcg
= NULL
;
4342 current
->memcg_batch
.nr_pages
= 0;
4343 current
->memcg_batch
.memsw_nr_pages
= 0;
4347 void mem_cgroup_uncharge_end(void)
4349 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4351 if (!batch
->do_batch
)
4355 if (batch
->do_batch
) /* If stacked, do nothing. */
4361 * This "batch->memcg" is valid without any css_get/put etc...
4362 * bacause we hide charges behind us.
4364 if (batch
->nr_pages
)
4365 res_counter_uncharge(&batch
->memcg
->res
,
4366 batch
->nr_pages
* PAGE_SIZE
);
4367 if (batch
->memsw_nr_pages
)
4368 res_counter_uncharge(&batch
->memcg
->memsw
,
4369 batch
->memsw_nr_pages
* PAGE_SIZE
);
4370 memcg_oom_recover(batch
->memcg
);
4371 /* forget this pointer (for sanity check) */
4372 batch
->memcg
= NULL
;
4377 * called after __delete_from_swap_cache() and drop "page" account.
4378 * memcg information is recorded to swap_cgroup of "ent"
4381 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4383 struct mem_cgroup
*memcg
;
4384 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4386 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4387 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4389 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4392 * record memcg information, if swapout && memcg != NULL,
4393 * css_get() was called in uncharge().
4395 if (do_swap_account
&& swapout
&& memcg
)
4396 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4400 #ifdef CONFIG_MEMCG_SWAP
4402 * called from swap_entry_free(). remove record in swap_cgroup and
4403 * uncharge "memsw" account.
4405 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4407 struct mem_cgroup
*memcg
;
4410 if (!do_swap_account
)
4413 id
= swap_cgroup_record(ent
, 0);
4415 memcg
= mem_cgroup_lookup(id
);
4418 * We uncharge this because swap is freed.
4419 * This memcg can be obsolete one. We avoid calling css_tryget
4421 if (!mem_cgroup_is_root(memcg
))
4422 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4423 mem_cgroup_swap_statistics(memcg
, false);
4424 css_put(&memcg
->css
);
4430 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4431 * @entry: swap entry to be moved
4432 * @from: mem_cgroup which the entry is moved from
4433 * @to: mem_cgroup which the entry is moved to
4435 * It succeeds only when the swap_cgroup's record for this entry is the same
4436 * as the mem_cgroup's id of @from.
4438 * Returns 0 on success, -EINVAL on failure.
4440 * The caller must have charged to @to, IOW, called res_counter_charge() about
4441 * both res and memsw, and called css_get().
4443 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4444 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4446 unsigned short old_id
, new_id
;
4448 old_id
= mem_cgroup_id(from
);
4449 new_id
= mem_cgroup_id(to
);
4451 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4452 mem_cgroup_swap_statistics(from
, false);
4453 mem_cgroup_swap_statistics(to
, true);
4455 * This function is only called from task migration context now.
4456 * It postpones res_counter and refcount handling till the end
4457 * of task migration(mem_cgroup_clear_mc()) for performance
4458 * improvement. But we cannot postpone css_get(to) because if
4459 * the process that has been moved to @to does swap-in, the
4460 * refcount of @to might be decreased to 0.
4462 * We are in attach() phase, so the cgroup is guaranteed to be
4463 * alive, so we can just call css_get().
4471 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4472 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4479 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4482 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4483 struct mem_cgroup
**memcgp
)
4485 struct mem_cgroup
*memcg
= NULL
;
4486 unsigned int nr_pages
= 1;
4487 struct page_cgroup
*pc
;
4488 enum charge_type ctype
;
4492 if (mem_cgroup_disabled())
4495 if (PageTransHuge(page
))
4496 nr_pages
<<= compound_order(page
);
4498 pc
= lookup_page_cgroup(page
);
4499 lock_page_cgroup(pc
);
4500 if (PageCgroupUsed(pc
)) {
4501 memcg
= pc
->mem_cgroup
;
4502 css_get(&memcg
->css
);
4504 * At migrating an anonymous page, its mapcount goes down
4505 * to 0 and uncharge() will be called. But, even if it's fully
4506 * unmapped, migration may fail and this page has to be
4507 * charged again. We set MIGRATION flag here and delay uncharge
4508 * until end_migration() is called
4510 * Corner Case Thinking
4512 * When the old page was mapped as Anon and it's unmap-and-freed
4513 * while migration was ongoing.
4514 * If unmap finds the old page, uncharge() of it will be delayed
4515 * until end_migration(). If unmap finds a new page, it's
4516 * uncharged when it make mapcount to be 1->0. If unmap code
4517 * finds swap_migration_entry, the new page will not be mapped
4518 * and end_migration() will find it(mapcount==0).
4521 * When the old page was mapped but migraion fails, the kernel
4522 * remaps it. A charge for it is kept by MIGRATION flag even
4523 * if mapcount goes down to 0. We can do remap successfully
4524 * without charging it again.
4527 * The "old" page is under lock_page() until the end of
4528 * migration, so, the old page itself will not be swapped-out.
4529 * If the new page is swapped out before end_migraton, our
4530 * hook to usual swap-out path will catch the event.
4533 SetPageCgroupMigration(pc
);
4535 unlock_page_cgroup(pc
);
4537 * If the page is not charged at this point,
4545 * We charge new page before it's used/mapped. So, even if unlock_page()
4546 * is called before end_migration, we can catch all events on this new
4547 * page. In the case new page is migrated but not remapped, new page's
4548 * mapcount will be finally 0 and we call uncharge in end_migration().
4551 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4553 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4555 * The page is committed to the memcg, but it's not actually
4556 * charged to the res_counter since we plan on replacing the
4557 * old one and only one page is going to be left afterwards.
4559 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4562 /* remove redundant charge if migration failed*/
4563 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4564 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4566 struct page
*used
, *unused
;
4567 struct page_cgroup
*pc
;
4573 if (!migration_ok
) {
4580 anon
= PageAnon(used
);
4581 __mem_cgroup_uncharge_common(unused
,
4582 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4583 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4585 css_put(&memcg
->css
);
4587 * We disallowed uncharge of pages under migration because mapcount
4588 * of the page goes down to zero, temporarly.
4589 * Clear the flag and check the page should be charged.
4591 pc
= lookup_page_cgroup(oldpage
);
4592 lock_page_cgroup(pc
);
4593 ClearPageCgroupMigration(pc
);
4594 unlock_page_cgroup(pc
);
4597 * If a page is a file cache, radix-tree replacement is very atomic
4598 * and we can skip this check. When it was an Anon page, its mapcount
4599 * goes down to 0. But because we added MIGRATION flage, it's not
4600 * uncharged yet. There are several case but page->mapcount check
4601 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4602 * check. (see prepare_charge() also)
4605 mem_cgroup_uncharge_page(used
);
4609 * At replace page cache, newpage is not under any memcg but it's on
4610 * LRU. So, this function doesn't touch res_counter but handles LRU
4611 * in correct way. Both pages are locked so we cannot race with uncharge.
4613 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4614 struct page
*newpage
)
4616 struct mem_cgroup
*memcg
= NULL
;
4617 struct page_cgroup
*pc
;
4618 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4620 if (mem_cgroup_disabled())
4623 pc
= lookup_page_cgroup(oldpage
);
4624 /* fix accounting on old pages */
4625 lock_page_cgroup(pc
);
4626 if (PageCgroupUsed(pc
)) {
4627 memcg
= pc
->mem_cgroup
;
4628 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4629 ClearPageCgroupUsed(pc
);
4631 unlock_page_cgroup(pc
);
4634 * When called from shmem_replace_page(), in some cases the
4635 * oldpage has already been charged, and in some cases not.
4640 * Even if newpage->mapping was NULL before starting replacement,
4641 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4642 * LRU while we overwrite pc->mem_cgroup.
4644 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4647 #ifdef CONFIG_DEBUG_VM
4648 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4650 struct page_cgroup
*pc
;
4652 pc
= lookup_page_cgroup(page
);
4654 * Can be NULL while feeding pages into the page allocator for
4655 * the first time, i.e. during boot or memory hotplug;
4656 * or when mem_cgroup_disabled().
4658 if (likely(pc
) && PageCgroupUsed(pc
))
4663 bool mem_cgroup_bad_page_check(struct page
*page
)
4665 if (mem_cgroup_disabled())
4668 return lookup_page_cgroup_used(page
) != NULL
;
4671 void mem_cgroup_print_bad_page(struct page
*page
)
4673 struct page_cgroup
*pc
;
4675 pc
= lookup_page_cgroup_used(page
);
4677 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4678 pc
, pc
->flags
, pc
->mem_cgroup
);
4683 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4684 unsigned long long val
)
4687 u64 memswlimit
, memlimit
;
4689 int children
= mem_cgroup_count_children(memcg
);
4690 u64 curusage
, oldusage
;
4694 * For keeping hierarchical_reclaim simple, how long we should retry
4695 * is depends on callers. We set our retry-count to be function
4696 * of # of children which we should visit in this loop.
4698 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4700 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4703 while (retry_count
) {
4704 if (signal_pending(current
)) {
4709 * Rather than hide all in some function, I do this in
4710 * open coded manner. You see what this really does.
4711 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4713 mutex_lock(&set_limit_mutex
);
4714 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4715 if (memswlimit
< val
) {
4717 mutex_unlock(&set_limit_mutex
);
4721 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4725 ret
= res_counter_set_limit(&memcg
->res
, val
);
4727 if (memswlimit
== val
)
4728 memcg
->memsw_is_minimum
= true;
4730 memcg
->memsw_is_minimum
= false;
4732 mutex_unlock(&set_limit_mutex
);
4737 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4738 MEM_CGROUP_RECLAIM_SHRINK
);
4739 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4740 /* Usage is reduced ? */
4741 if (curusage
>= oldusage
)
4744 oldusage
= curusage
;
4746 if (!ret
&& enlarge
)
4747 memcg_oom_recover(memcg
);
4752 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4753 unsigned long long val
)
4756 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4757 int children
= mem_cgroup_count_children(memcg
);
4761 /* see mem_cgroup_resize_res_limit */
4762 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4763 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4764 while (retry_count
) {
4765 if (signal_pending(current
)) {
4770 * Rather than hide all in some function, I do this in
4771 * open coded manner. You see what this really does.
4772 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4774 mutex_lock(&set_limit_mutex
);
4775 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4776 if (memlimit
> val
) {
4778 mutex_unlock(&set_limit_mutex
);
4781 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4782 if (memswlimit
< val
)
4784 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4786 if (memlimit
== val
)
4787 memcg
->memsw_is_minimum
= true;
4789 memcg
->memsw_is_minimum
= false;
4791 mutex_unlock(&set_limit_mutex
);
4796 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4797 MEM_CGROUP_RECLAIM_NOSWAP
|
4798 MEM_CGROUP_RECLAIM_SHRINK
);
4799 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4800 /* Usage is reduced ? */
4801 if (curusage
>= oldusage
)
4804 oldusage
= curusage
;
4806 if (!ret
&& enlarge
)
4807 memcg_oom_recover(memcg
);
4811 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4813 unsigned long *total_scanned
)
4815 unsigned long nr_reclaimed
= 0;
4816 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4817 unsigned long reclaimed
;
4819 struct mem_cgroup_tree_per_zone
*mctz
;
4820 unsigned long long excess
;
4821 unsigned long nr_scanned
;
4826 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4828 * This loop can run a while, specially if mem_cgroup's continuously
4829 * keep exceeding their soft limit and putting the system under
4836 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4841 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4842 gfp_mask
, &nr_scanned
);
4843 nr_reclaimed
+= reclaimed
;
4844 *total_scanned
+= nr_scanned
;
4845 spin_lock(&mctz
->lock
);
4848 * If we failed to reclaim anything from this memory cgroup
4849 * it is time to move on to the next cgroup
4855 * Loop until we find yet another one.
4857 * By the time we get the soft_limit lock
4858 * again, someone might have aded the
4859 * group back on the RB tree. Iterate to
4860 * make sure we get a different mem.
4861 * mem_cgroup_largest_soft_limit_node returns
4862 * NULL if no other cgroup is present on
4866 __mem_cgroup_largest_soft_limit_node(mctz
);
4868 css_put(&next_mz
->memcg
->css
);
4869 else /* next_mz == NULL or other memcg */
4873 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4874 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4876 * One school of thought says that we should not add
4877 * back the node to the tree if reclaim returns 0.
4878 * But our reclaim could return 0, simply because due
4879 * to priority we are exposing a smaller subset of
4880 * memory to reclaim from. Consider this as a longer
4883 /* If excess == 0, no tree ops */
4884 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4885 spin_unlock(&mctz
->lock
);
4886 css_put(&mz
->memcg
->css
);
4889 * Could not reclaim anything and there are no more
4890 * mem cgroups to try or we seem to be looping without
4891 * reclaiming anything.
4893 if (!nr_reclaimed
&&
4895 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4897 } while (!nr_reclaimed
);
4899 css_put(&next_mz
->memcg
->css
);
4900 return nr_reclaimed
;
4904 * mem_cgroup_force_empty_list - clears LRU of a group
4905 * @memcg: group to clear
4908 * @lru: lru to to clear
4910 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4911 * reclaim the pages page themselves - pages are moved to the parent (or root)
4914 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4915 int node
, int zid
, enum lru_list lru
)
4917 struct lruvec
*lruvec
;
4918 unsigned long flags
;
4919 struct list_head
*list
;
4923 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4924 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4925 list
= &lruvec
->lists
[lru
];
4929 struct page_cgroup
*pc
;
4932 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4933 if (list_empty(list
)) {
4934 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4937 page
= list_entry(list
->prev
, struct page
, lru
);
4939 list_move(&page
->lru
, list
);
4941 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4944 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4946 pc
= lookup_page_cgroup(page
);
4948 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4949 /* found lock contention or "pc" is obsolete. */
4954 } while (!list_empty(list
));
4958 * make mem_cgroup's charge to be 0 if there is no task by moving
4959 * all the charges and pages to the parent.
4960 * This enables deleting this mem_cgroup.
4962 * Caller is responsible for holding css reference on the memcg.
4964 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4970 /* This is for making all *used* pages to be on LRU. */
4971 lru_add_drain_all();
4972 drain_all_stock_sync(memcg
);
4973 mem_cgroup_start_move(memcg
);
4974 for_each_node_state(node
, N_MEMORY
) {
4975 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4978 mem_cgroup_force_empty_list(memcg
,
4983 mem_cgroup_end_move(memcg
);
4984 memcg_oom_recover(memcg
);
4988 * Kernel memory may not necessarily be trackable to a specific
4989 * process. So they are not migrated, and therefore we can't
4990 * expect their value to drop to 0 here.
4991 * Having res filled up with kmem only is enough.
4993 * This is a safety check because mem_cgroup_force_empty_list
4994 * could have raced with mem_cgroup_replace_page_cache callers
4995 * so the lru seemed empty but the page could have been added
4996 * right after the check. RES_USAGE should be safe as we always
4997 * charge before adding to the LRU.
4999 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
5000 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
5001 } while (usage
> 0);
5004 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
5006 lockdep_assert_held(&memcg_create_mutex
);
5008 * The lock does not prevent addition or deletion to the list
5009 * of children, but it prevents a new child from being
5010 * initialized based on this parent in css_online(), so it's
5011 * enough to decide whether hierarchically inherited
5012 * attributes can still be changed or not.
5014 return memcg
->use_hierarchy
&&
5015 !list_empty(&memcg
->css
.cgroup
->children
);
5019 * Reclaims as many pages from the given memcg as possible and moves
5020 * the rest to the parent.
5022 * Caller is responsible for holding css reference for memcg.
5024 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
5026 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
5027 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
5029 /* returns EBUSY if there is a task or if we come here twice. */
5030 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
5033 /* we call try-to-free pages for make this cgroup empty */
5034 lru_add_drain_all();
5035 /* try to free all pages in this cgroup */
5036 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
5039 if (signal_pending(current
))
5042 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
5046 /* maybe some writeback is necessary */
5047 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
5052 mem_cgroup_reparent_charges(memcg
);
5057 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5060 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5062 if (mem_cgroup_is_root(memcg
))
5064 return mem_cgroup_force_empty(memcg
);
5067 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5070 return mem_cgroup_from_css(css
)->use_hierarchy
;
5073 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5074 struct cftype
*cft
, u64 val
)
5077 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5078 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5080 mutex_lock(&memcg_create_mutex
);
5082 if (memcg
->use_hierarchy
== val
)
5086 * If parent's use_hierarchy is set, we can't make any modifications
5087 * in the child subtrees. If it is unset, then the change can
5088 * occur, provided the current cgroup has no children.
5090 * For the root cgroup, parent_mem is NULL, we allow value to be
5091 * set if there are no children.
5093 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5094 (val
== 1 || val
== 0)) {
5095 if (list_empty(&memcg
->css
.cgroup
->children
))
5096 memcg
->use_hierarchy
= val
;
5103 mutex_unlock(&memcg_create_mutex
);
5109 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5110 enum mem_cgroup_stat_index idx
)
5112 struct mem_cgroup
*iter
;
5115 /* Per-cpu values can be negative, use a signed accumulator */
5116 for_each_mem_cgroup_tree(iter
, memcg
)
5117 val
+= mem_cgroup_read_stat(iter
, idx
);
5119 if (val
< 0) /* race ? */
5124 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5128 if (!mem_cgroup_is_root(memcg
)) {
5130 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5132 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5136 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5137 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5139 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5140 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5143 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5145 return val
<< PAGE_SHIFT
;
5148 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
5151 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5156 type
= MEMFILE_TYPE(cft
->private);
5157 name
= MEMFILE_ATTR(cft
->private);
5161 if (name
== RES_USAGE
)
5162 val
= mem_cgroup_usage(memcg
, false);
5164 val
= res_counter_read_u64(&memcg
->res
, name
);
5167 if (name
== RES_USAGE
)
5168 val
= mem_cgroup_usage(memcg
, true);
5170 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5173 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5182 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5185 #ifdef CONFIG_MEMCG_KMEM
5186 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5188 * For simplicity, we won't allow this to be disabled. It also can't
5189 * be changed if the cgroup has children already, or if tasks had
5192 * If tasks join before we set the limit, a person looking at
5193 * kmem.usage_in_bytes will have no way to determine when it took
5194 * place, which makes the value quite meaningless.
5196 * After it first became limited, changes in the value of the limit are
5197 * of course permitted.
5199 mutex_lock(&memcg_create_mutex
);
5200 mutex_lock(&set_limit_mutex
);
5201 if (!memcg
->kmem_account_flags
&& val
!= RES_COUNTER_MAX
) {
5202 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5206 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5209 ret
= memcg_update_cache_sizes(memcg
);
5211 res_counter_set_limit(&memcg
->kmem
, RES_COUNTER_MAX
);
5214 static_key_slow_inc(&memcg_kmem_enabled_key
);
5216 * setting the active bit after the inc will guarantee no one
5217 * starts accounting before all call sites are patched
5219 memcg_kmem_set_active(memcg
);
5221 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5223 mutex_unlock(&set_limit_mutex
);
5224 mutex_unlock(&memcg_create_mutex
);
5229 #ifdef CONFIG_MEMCG_KMEM
5230 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5233 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5237 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5239 * When that happen, we need to disable the static branch only on those
5240 * memcgs that enabled it. To achieve this, we would be forced to
5241 * complicate the code by keeping track of which memcgs were the ones
5242 * that actually enabled limits, and which ones got it from its
5245 * It is a lot simpler just to do static_key_slow_inc() on every child
5246 * that is accounted.
5248 if (!memcg_kmem_is_active(memcg
))
5252 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5253 * memcg is active already. If the later initialization fails then the
5254 * cgroup core triggers the cleanup so we do not have to do it here.
5256 static_key_slow_inc(&memcg_kmem_enabled_key
);
5258 mutex_lock(&set_limit_mutex
);
5259 memcg_stop_kmem_account();
5260 ret
= memcg_update_cache_sizes(memcg
);
5261 memcg_resume_kmem_account();
5262 mutex_unlock(&set_limit_mutex
);
5266 #endif /* CONFIG_MEMCG_KMEM */
5269 * The user of this function is...
5272 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5275 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5278 unsigned long long val
;
5281 type
= MEMFILE_TYPE(cft
->private);
5282 name
= MEMFILE_ATTR(cft
->private);
5286 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5290 /* This function does all necessary parse...reuse it */
5291 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5295 ret
= mem_cgroup_resize_limit(memcg
, val
);
5296 else if (type
== _MEMSWAP
)
5297 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5298 else if (type
== _KMEM
)
5299 ret
= memcg_update_kmem_limit(css
, val
);
5303 case RES_SOFT_LIMIT
:
5304 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5308 * For memsw, soft limits are hard to implement in terms
5309 * of semantics, for now, we support soft limits for
5310 * control without swap
5313 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5318 ret
= -EINVAL
; /* should be BUG() ? */
5324 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5325 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5327 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5329 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5330 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5331 if (!memcg
->use_hierarchy
)
5334 while (css_parent(&memcg
->css
)) {
5335 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5336 if (!memcg
->use_hierarchy
)
5338 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5339 min_limit
= min(min_limit
, tmp
);
5340 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5341 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5344 *mem_limit
= min_limit
;
5345 *memsw_limit
= min_memsw_limit
;
5348 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5350 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5354 type
= MEMFILE_TYPE(event
);
5355 name
= MEMFILE_ATTR(event
);
5360 res_counter_reset_max(&memcg
->res
);
5361 else if (type
== _MEMSWAP
)
5362 res_counter_reset_max(&memcg
->memsw
);
5363 else if (type
== _KMEM
)
5364 res_counter_reset_max(&memcg
->kmem
);
5370 res_counter_reset_failcnt(&memcg
->res
);
5371 else if (type
== _MEMSWAP
)
5372 res_counter_reset_failcnt(&memcg
->memsw
);
5373 else if (type
== _KMEM
)
5374 res_counter_reset_failcnt(&memcg
->kmem
);
5383 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5386 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5390 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5391 struct cftype
*cft
, u64 val
)
5393 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5395 if (val
>= (1 << NR_MOVE_TYPE
))
5399 * No kind of locking is needed in here, because ->can_attach() will
5400 * check this value once in the beginning of the process, and then carry
5401 * on with stale data. This means that changes to this value will only
5402 * affect task migrations starting after the change.
5404 memcg
->move_charge_at_immigrate
= val
;
5408 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5409 struct cftype
*cft
, u64 val
)
5416 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
5420 unsigned int lru_mask
;
5423 static const struct numa_stat stats
[] = {
5424 { "total", LRU_ALL
},
5425 { "file", LRU_ALL_FILE
},
5426 { "anon", LRU_ALL_ANON
},
5427 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5429 const struct numa_stat
*stat
;
5432 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5434 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5435 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5436 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5437 for_each_node_state(nid
, N_MEMORY
) {
5438 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5440 seq_printf(m
, " N%d=%lu", nid
, nr
);
5445 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5446 struct mem_cgroup
*iter
;
5449 for_each_mem_cgroup_tree(iter
, memcg
)
5450 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5451 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5452 for_each_node_state(nid
, N_MEMORY
) {
5454 for_each_mem_cgroup_tree(iter
, memcg
)
5455 nr
+= mem_cgroup_node_nr_lru_pages(
5456 iter
, nid
, stat
->lru_mask
);
5457 seq_printf(m
, " N%d=%lu", nid
, nr
);
5464 #endif /* CONFIG_NUMA */
5466 static inline void mem_cgroup_lru_names_not_uptodate(void)
5468 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5471 static int memcg_stat_show(struct seq_file
*m
, void *v
)
5473 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5474 struct mem_cgroup
*mi
;
5477 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5478 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5480 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5481 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5484 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5485 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5486 mem_cgroup_read_events(memcg
, i
));
5488 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5489 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5490 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5492 /* Hierarchical information */
5494 unsigned long long limit
, memsw_limit
;
5495 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5496 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5497 if (do_swap_account
)
5498 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5502 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5505 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5507 for_each_mem_cgroup_tree(mi
, memcg
)
5508 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5509 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5512 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5513 unsigned long long val
= 0;
5515 for_each_mem_cgroup_tree(mi
, memcg
)
5516 val
+= mem_cgroup_read_events(mi
, i
);
5517 seq_printf(m
, "total_%s %llu\n",
5518 mem_cgroup_events_names
[i
], val
);
5521 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5522 unsigned long long val
= 0;
5524 for_each_mem_cgroup_tree(mi
, memcg
)
5525 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5526 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5529 #ifdef CONFIG_DEBUG_VM
5532 struct mem_cgroup_per_zone
*mz
;
5533 struct zone_reclaim_stat
*rstat
;
5534 unsigned long recent_rotated
[2] = {0, 0};
5535 unsigned long recent_scanned
[2] = {0, 0};
5537 for_each_online_node(nid
)
5538 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5539 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5540 rstat
= &mz
->lruvec
.reclaim_stat
;
5542 recent_rotated
[0] += rstat
->recent_rotated
[0];
5543 recent_rotated
[1] += rstat
->recent_rotated
[1];
5544 recent_scanned
[0] += rstat
->recent_scanned
[0];
5545 recent_scanned
[1] += rstat
->recent_scanned
[1];
5547 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5548 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5549 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5550 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5557 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5560 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5562 return mem_cgroup_swappiness(memcg
);
5565 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5566 struct cftype
*cft
, u64 val
)
5568 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5569 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5571 if (val
> 100 || !parent
)
5574 mutex_lock(&memcg_create_mutex
);
5576 /* If under hierarchy, only empty-root can set this value */
5577 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5578 mutex_unlock(&memcg_create_mutex
);
5582 memcg
->swappiness
= val
;
5584 mutex_unlock(&memcg_create_mutex
);
5589 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5591 struct mem_cgroup_threshold_ary
*t
;
5597 t
= rcu_dereference(memcg
->thresholds
.primary
);
5599 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5604 usage
= mem_cgroup_usage(memcg
, swap
);
5607 * current_threshold points to threshold just below or equal to usage.
5608 * If it's not true, a threshold was crossed after last
5609 * call of __mem_cgroup_threshold().
5611 i
= t
->current_threshold
;
5614 * Iterate backward over array of thresholds starting from
5615 * current_threshold and check if a threshold is crossed.
5616 * If none of thresholds below usage is crossed, we read
5617 * only one element of the array here.
5619 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5620 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5622 /* i = current_threshold + 1 */
5626 * Iterate forward over array of thresholds starting from
5627 * current_threshold+1 and check if a threshold is crossed.
5628 * If none of thresholds above usage is crossed, we read
5629 * only one element of the array here.
5631 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5632 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5634 /* Update current_threshold */
5635 t
->current_threshold
= i
- 1;
5640 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5643 __mem_cgroup_threshold(memcg
, false);
5644 if (do_swap_account
)
5645 __mem_cgroup_threshold(memcg
, true);
5647 memcg
= parent_mem_cgroup(memcg
);
5651 static int compare_thresholds(const void *a
, const void *b
)
5653 const struct mem_cgroup_threshold
*_a
= a
;
5654 const struct mem_cgroup_threshold
*_b
= b
;
5656 if (_a
->threshold
> _b
->threshold
)
5659 if (_a
->threshold
< _b
->threshold
)
5665 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5667 struct mem_cgroup_eventfd_list
*ev
;
5669 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5670 eventfd_signal(ev
->eventfd
, 1);
5674 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5676 struct mem_cgroup
*iter
;
5678 for_each_mem_cgroup_tree(iter
, memcg
)
5679 mem_cgroup_oom_notify_cb(iter
);
5682 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5683 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
5685 struct mem_cgroup_thresholds
*thresholds
;
5686 struct mem_cgroup_threshold_ary
*new;
5687 u64 threshold
, usage
;
5690 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5694 mutex_lock(&memcg
->thresholds_lock
);
5697 thresholds
= &memcg
->thresholds
;
5698 else if (type
== _MEMSWAP
)
5699 thresholds
= &memcg
->memsw_thresholds
;
5703 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5705 /* Check if a threshold crossed before adding a new one */
5706 if (thresholds
->primary
)
5707 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5709 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5711 /* Allocate memory for new array of thresholds */
5712 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5720 /* Copy thresholds (if any) to new array */
5721 if (thresholds
->primary
) {
5722 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5723 sizeof(struct mem_cgroup_threshold
));
5726 /* Add new threshold */
5727 new->entries
[size
- 1].eventfd
= eventfd
;
5728 new->entries
[size
- 1].threshold
= threshold
;
5730 /* Sort thresholds. Registering of new threshold isn't time-critical */
5731 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5732 compare_thresholds
, NULL
);
5734 /* Find current threshold */
5735 new->current_threshold
= -1;
5736 for (i
= 0; i
< size
; i
++) {
5737 if (new->entries
[i
].threshold
<= usage
) {
5739 * new->current_threshold will not be used until
5740 * rcu_assign_pointer(), so it's safe to increment
5743 ++new->current_threshold
;
5748 /* Free old spare buffer and save old primary buffer as spare */
5749 kfree(thresholds
->spare
);
5750 thresholds
->spare
= thresholds
->primary
;
5752 rcu_assign_pointer(thresholds
->primary
, new);
5754 /* To be sure that nobody uses thresholds */
5758 mutex_unlock(&memcg
->thresholds_lock
);
5763 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5764 struct eventfd_ctx
*eventfd
, const char *args
)
5766 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
5769 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5770 struct eventfd_ctx
*eventfd
, const char *args
)
5772 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
5775 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5776 struct eventfd_ctx
*eventfd
, enum res_type type
)
5778 struct mem_cgroup_thresholds
*thresholds
;
5779 struct mem_cgroup_threshold_ary
*new;
5783 mutex_lock(&memcg
->thresholds_lock
);
5785 thresholds
= &memcg
->thresholds
;
5786 else if (type
== _MEMSWAP
)
5787 thresholds
= &memcg
->memsw_thresholds
;
5791 if (!thresholds
->primary
)
5794 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5796 /* Check if a threshold crossed before removing */
5797 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5799 /* Calculate new number of threshold */
5801 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5802 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5806 new = thresholds
->spare
;
5808 /* Set thresholds array to NULL if we don't have thresholds */
5817 /* Copy thresholds and find current threshold */
5818 new->current_threshold
= -1;
5819 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5820 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5823 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5824 if (new->entries
[j
].threshold
<= usage
) {
5826 * new->current_threshold will not be used
5827 * until rcu_assign_pointer(), so it's safe to increment
5830 ++new->current_threshold
;
5836 /* Swap primary and spare array */
5837 thresholds
->spare
= thresholds
->primary
;
5838 /* If all events are unregistered, free the spare array */
5840 kfree(thresholds
->spare
);
5841 thresholds
->spare
= NULL
;
5844 rcu_assign_pointer(thresholds
->primary
, new);
5846 /* To be sure that nobody uses thresholds */
5849 mutex_unlock(&memcg
->thresholds_lock
);
5852 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5853 struct eventfd_ctx
*eventfd
)
5855 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
5858 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5859 struct eventfd_ctx
*eventfd
)
5861 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
5864 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
5865 struct eventfd_ctx
*eventfd
, const char *args
)
5867 struct mem_cgroup_eventfd_list
*event
;
5869 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5873 spin_lock(&memcg_oom_lock
);
5875 event
->eventfd
= eventfd
;
5876 list_add(&event
->list
, &memcg
->oom_notify
);
5878 /* already in OOM ? */
5879 if (atomic_read(&memcg
->under_oom
))
5880 eventfd_signal(eventfd
, 1);
5881 spin_unlock(&memcg_oom_lock
);
5886 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
5887 struct eventfd_ctx
*eventfd
)
5889 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5891 spin_lock(&memcg_oom_lock
);
5893 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5894 if (ev
->eventfd
== eventfd
) {
5895 list_del(&ev
->list
);
5900 spin_unlock(&memcg_oom_lock
);
5903 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
5905 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
5907 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
5908 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
5912 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5913 struct cftype
*cft
, u64 val
)
5915 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5916 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5918 /* cannot set to root cgroup and only 0 and 1 are allowed */
5919 if (!parent
|| !((val
== 0) || (val
== 1)))
5922 mutex_lock(&memcg_create_mutex
);
5923 /* oom-kill-disable is a flag for subhierarchy. */
5924 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5925 mutex_unlock(&memcg_create_mutex
);
5928 memcg
->oom_kill_disable
= val
;
5930 memcg_oom_recover(memcg
);
5931 mutex_unlock(&memcg_create_mutex
);
5935 #ifdef CONFIG_MEMCG_KMEM
5936 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5940 memcg
->kmemcg_id
= -1;
5941 ret
= memcg_propagate_kmem(memcg
);
5945 return mem_cgroup_sockets_init(memcg
, ss
);
5948 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5950 mem_cgroup_sockets_destroy(memcg
);
5953 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5955 if (!memcg_kmem_is_active(memcg
))
5959 * kmem charges can outlive the cgroup. In the case of slab
5960 * pages, for instance, a page contain objects from various
5961 * processes. As we prevent from taking a reference for every
5962 * such allocation we have to be careful when doing uncharge
5963 * (see memcg_uncharge_kmem) and here during offlining.
5965 * The idea is that that only the _last_ uncharge which sees
5966 * the dead memcg will drop the last reference. An additional
5967 * reference is taken here before the group is marked dead
5968 * which is then paired with css_put during uncharge resp. here.
5970 * Although this might sound strange as this path is called from
5971 * css_offline() when the referencemight have dropped down to 0
5972 * and shouldn't be incremented anymore (css_tryget would fail)
5973 * we do not have other options because of the kmem allocations
5976 css_get(&memcg
->css
);
5978 memcg_kmem_mark_dead(memcg
);
5980 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5983 if (memcg_kmem_test_and_clear_dead(memcg
))
5984 css_put(&memcg
->css
);
5987 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5992 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5996 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
6002 * DO NOT USE IN NEW FILES.
6004 * "cgroup.event_control" implementation.
6006 * This is way over-engineered. It tries to support fully configurable
6007 * events for each user. Such level of flexibility is completely
6008 * unnecessary especially in the light of the planned unified hierarchy.
6010 * Please deprecate this and replace with something simpler if at all
6015 * Unregister event and free resources.
6017 * Gets called from workqueue.
6019 static void memcg_event_remove(struct work_struct
*work
)
6021 struct mem_cgroup_event
*event
=
6022 container_of(work
, struct mem_cgroup_event
, remove
);
6023 struct mem_cgroup
*memcg
= event
->memcg
;
6025 remove_wait_queue(event
->wqh
, &event
->wait
);
6027 event
->unregister_event(memcg
, event
->eventfd
);
6029 /* Notify userspace the event is going away. */
6030 eventfd_signal(event
->eventfd
, 1);
6032 eventfd_ctx_put(event
->eventfd
);
6034 css_put(&memcg
->css
);
6038 * Gets called on POLLHUP on eventfd when user closes it.
6040 * Called with wqh->lock held and interrupts disabled.
6042 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
6043 int sync
, void *key
)
6045 struct mem_cgroup_event
*event
=
6046 container_of(wait
, struct mem_cgroup_event
, wait
);
6047 struct mem_cgroup
*memcg
= event
->memcg
;
6048 unsigned long flags
= (unsigned long)key
;
6050 if (flags
& POLLHUP
) {
6052 * If the event has been detached at cgroup removal, we
6053 * can simply return knowing the other side will cleanup
6056 * We can't race against event freeing since the other
6057 * side will require wqh->lock via remove_wait_queue(),
6060 spin_lock(&memcg
->event_list_lock
);
6061 if (!list_empty(&event
->list
)) {
6062 list_del_init(&event
->list
);
6064 * We are in atomic context, but cgroup_event_remove()
6065 * may sleep, so we have to call it in workqueue.
6067 schedule_work(&event
->remove
);
6069 spin_unlock(&memcg
->event_list_lock
);
6075 static void memcg_event_ptable_queue_proc(struct file
*file
,
6076 wait_queue_head_t
*wqh
, poll_table
*pt
)
6078 struct mem_cgroup_event
*event
=
6079 container_of(pt
, struct mem_cgroup_event
, pt
);
6082 add_wait_queue(wqh
, &event
->wait
);
6086 * DO NOT USE IN NEW FILES.
6088 * Parse input and register new cgroup event handler.
6090 * Input must be in format '<event_fd> <control_fd> <args>'.
6091 * Interpretation of args is defined by control file implementation.
6093 static int memcg_write_event_control(struct cgroup_subsys_state
*css
,
6094 struct cftype
*cft
, const char *buffer
)
6096 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6097 struct mem_cgroup_event
*event
;
6098 struct cgroup_subsys_state
*cfile_css
;
6099 unsigned int efd
, cfd
;
6106 efd
= simple_strtoul(buffer
, &endp
, 10);
6111 cfd
= simple_strtoul(buffer
, &endp
, 10);
6112 if ((*endp
!= ' ') && (*endp
!= '\0'))
6116 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6120 event
->memcg
= memcg
;
6121 INIT_LIST_HEAD(&event
->list
);
6122 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
6123 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
6124 INIT_WORK(&event
->remove
, memcg_event_remove
);
6132 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
6133 if (IS_ERR(event
->eventfd
)) {
6134 ret
= PTR_ERR(event
->eventfd
);
6141 goto out_put_eventfd
;
6144 /* the process need read permission on control file */
6145 /* AV: shouldn't we check that it's been opened for read instead? */
6146 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
6151 * Determine the event callbacks and set them in @event. This used
6152 * to be done via struct cftype but cgroup core no longer knows
6153 * about these events. The following is crude but the whole thing
6154 * is for compatibility anyway.
6156 * DO NOT ADD NEW FILES.
6158 name
= cfile
.file
->f_dentry
->d_name
.name
;
6160 if (!strcmp(name
, "memory.usage_in_bytes")) {
6161 event
->register_event
= mem_cgroup_usage_register_event
;
6162 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
6163 } else if (!strcmp(name
, "memory.oom_control")) {
6164 event
->register_event
= mem_cgroup_oom_register_event
;
6165 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
6166 } else if (!strcmp(name
, "memory.pressure_level")) {
6167 event
->register_event
= vmpressure_register_event
;
6168 event
->unregister_event
= vmpressure_unregister_event
;
6169 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
6170 event
->register_event
= memsw_cgroup_usage_register_event
;
6171 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
6178 * Verify @cfile should belong to @css. Also, remaining events are
6179 * automatically removed on cgroup destruction but the removal is
6180 * asynchronous, so take an extra ref on @css.
6185 cfile_css
= css_from_dir(cfile
.file
->f_dentry
->d_parent
,
6186 &mem_cgroup_subsys
);
6187 if (cfile_css
== css
&& css_tryget(css
))
6194 ret
= event
->register_event(memcg
, event
->eventfd
, buffer
);
6198 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
6200 spin_lock(&memcg
->event_list_lock
);
6201 list_add(&event
->list
, &memcg
->event_list
);
6202 spin_unlock(&memcg
->event_list_lock
);
6214 eventfd_ctx_put(event
->eventfd
);
6223 static struct cftype mem_cgroup_files
[] = {
6225 .name
= "usage_in_bytes",
6226 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
6227 .read_u64
= mem_cgroup_read_u64
,
6230 .name
= "max_usage_in_bytes",
6231 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
6232 .trigger
= mem_cgroup_reset
,
6233 .read_u64
= mem_cgroup_read_u64
,
6236 .name
= "limit_in_bytes",
6237 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
6238 .write_string
= mem_cgroup_write
,
6239 .read_u64
= mem_cgroup_read_u64
,
6242 .name
= "soft_limit_in_bytes",
6243 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
6244 .write_string
= mem_cgroup_write
,
6245 .read_u64
= mem_cgroup_read_u64
,
6249 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
6250 .trigger
= mem_cgroup_reset
,
6251 .read_u64
= mem_cgroup_read_u64
,
6255 .seq_show
= memcg_stat_show
,
6258 .name
= "force_empty",
6259 .trigger
= mem_cgroup_force_empty_write
,
6262 .name
= "use_hierarchy",
6263 .flags
= CFTYPE_INSANE
,
6264 .write_u64
= mem_cgroup_hierarchy_write
,
6265 .read_u64
= mem_cgroup_hierarchy_read
,
6268 .name
= "cgroup.event_control", /* XXX: for compat */
6269 .write_string
= memcg_write_event_control
,
6270 .flags
= CFTYPE_NO_PREFIX
,
6274 .name
= "swappiness",
6275 .read_u64
= mem_cgroup_swappiness_read
,
6276 .write_u64
= mem_cgroup_swappiness_write
,
6279 .name
= "move_charge_at_immigrate",
6280 .read_u64
= mem_cgroup_move_charge_read
,
6281 .write_u64
= mem_cgroup_move_charge_write
,
6284 .name
= "oom_control",
6285 .seq_show
= mem_cgroup_oom_control_read
,
6286 .write_u64
= mem_cgroup_oom_control_write
,
6287 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6290 .name
= "pressure_level",
6294 .name
= "numa_stat",
6295 .seq_show
= memcg_numa_stat_show
,
6298 #ifdef CONFIG_MEMCG_KMEM
6300 .name
= "kmem.limit_in_bytes",
6301 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6302 .write_string
= mem_cgroup_write
,
6303 .read_u64
= mem_cgroup_read_u64
,
6306 .name
= "kmem.usage_in_bytes",
6307 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6308 .read_u64
= mem_cgroup_read_u64
,
6311 .name
= "kmem.failcnt",
6312 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6313 .trigger
= mem_cgroup_reset
,
6314 .read_u64
= mem_cgroup_read_u64
,
6317 .name
= "kmem.max_usage_in_bytes",
6318 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6319 .trigger
= mem_cgroup_reset
,
6320 .read_u64
= mem_cgroup_read_u64
,
6322 #ifdef CONFIG_SLABINFO
6324 .name
= "kmem.slabinfo",
6325 .seq_show
= mem_cgroup_slabinfo_read
,
6329 { }, /* terminate */
6332 #ifdef CONFIG_MEMCG_SWAP
6333 static struct cftype memsw_cgroup_files
[] = {
6335 .name
= "memsw.usage_in_bytes",
6336 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6337 .read_u64
= mem_cgroup_read_u64
,
6340 .name
= "memsw.max_usage_in_bytes",
6341 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6342 .trigger
= mem_cgroup_reset
,
6343 .read_u64
= mem_cgroup_read_u64
,
6346 .name
= "memsw.limit_in_bytes",
6347 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6348 .write_string
= mem_cgroup_write
,
6349 .read_u64
= mem_cgroup_read_u64
,
6352 .name
= "memsw.failcnt",
6353 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6354 .trigger
= mem_cgroup_reset
,
6355 .read_u64
= mem_cgroup_read_u64
,
6357 { }, /* terminate */
6360 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6362 struct mem_cgroup_per_node
*pn
;
6363 struct mem_cgroup_per_zone
*mz
;
6364 int zone
, tmp
= node
;
6366 * This routine is called against possible nodes.
6367 * But it's BUG to call kmalloc() against offline node.
6369 * TODO: this routine can waste much memory for nodes which will
6370 * never be onlined. It's better to use memory hotplug callback
6373 if (!node_state(node
, N_NORMAL_MEMORY
))
6375 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6379 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6380 mz
= &pn
->zoneinfo
[zone
];
6381 lruvec_init(&mz
->lruvec
);
6382 mz
->usage_in_excess
= 0;
6383 mz
->on_tree
= false;
6386 memcg
->nodeinfo
[node
] = pn
;
6390 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6392 kfree(memcg
->nodeinfo
[node
]);
6395 static struct mem_cgroup
*mem_cgroup_alloc(void)
6397 struct mem_cgroup
*memcg
;
6400 size
= sizeof(struct mem_cgroup
);
6401 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
6403 memcg
= kzalloc(size
, GFP_KERNEL
);
6407 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6410 spin_lock_init(&memcg
->pcp_counter_lock
);
6419 * At destroying mem_cgroup, references from swap_cgroup can remain.
6420 * (scanning all at force_empty is too costly...)
6422 * Instead of clearing all references at force_empty, we remember
6423 * the number of reference from swap_cgroup and free mem_cgroup when
6424 * it goes down to 0.
6426 * Removal of cgroup itself succeeds regardless of refs from swap.
6429 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6433 mem_cgroup_remove_from_trees(memcg
);
6436 free_mem_cgroup_per_zone_info(memcg
, node
);
6438 free_percpu(memcg
->stat
);
6441 * We need to make sure that (at least for now), the jump label
6442 * destruction code runs outside of the cgroup lock. This is because
6443 * get_online_cpus(), which is called from the static_branch update,
6444 * can't be called inside the cgroup_lock. cpusets are the ones
6445 * enforcing this dependency, so if they ever change, we might as well.
6447 * schedule_work() will guarantee this happens. Be careful if you need
6448 * to move this code around, and make sure it is outside
6451 disarm_static_keys(memcg
);
6456 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6458 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6460 if (!memcg
->res
.parent
)
6462 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6464 EXPORT_SYMBOL(parent_mem_cgroup
);
6466 static void __init
mem_cgroup_soft_limit_tree_init(void)
6468 struct mem_cgroup_tree_per_node
*rtpn
;
6469 struct mem_cgroup_tree_per_zone
*rtpz
;
6470 int tmp
, node
, zone
;
6472 for_each_node(node
) {
6474 if (!node_state(node
, N_NORMAL_MEMORY
))
6476 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6479 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6481 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6482 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6483 rtpz
->rb_root
= RB_ROOT
;
6484 spin_lock_init(&rtpz
->lock
);
6489 static struct cgroup_subsys_state
* __ref
6490 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6492 struct mem_cgroup
*memcg
;
6493 long error
= -ENOMEM
;
6496 memcg
= mem_cgroup_alloc();
6498 return ERR_PTR(error
);
6501 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6505 if (parent_css
== NULL
) {
6506 root_mem_cgroup
= memcg
;
6507 res_counter_init(&memcg
->res
, NULL
);
6508 res_counter_init(&memcg
->memsw
, NULL
);
6509 res_counter_init(&memcg
->kmem
, NULL
);
6512 memcg
->last_scanned_node
= MAX_NUMNODES
;
6513 INIT_LIST_HEAD(&memcg
->oom_notify
);
6514 memcg
->move_charge_at_immigrate
= 0;
6515 mutex_init(&memcg
->thresholds_lock
);
6516 spin_lock_init(&memcg
->move_lock
);
6517 vmpressure_init(&memcg
->vmpressure
);
6518 INIT_LIST_HEAD(&memcg
->event_list
);
6519 spin_lock_init(&memcg
->event_list_lock
);
6524 __mem_cgroup_free(memcg
);
6525 return ERR_PTR(error
);
6529 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6531 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6532 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6535 if (css
->cgroup
->id
> MEM_CGROUP_ID_MAX
)
6541 mutex_lock(&memcg_create_mutex
);
6543 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6544 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6545 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6547 if (parent
->use_hierarchy
) {
6548 res_counter_init(&memcg
->res
, &parent
->res
);
6549 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6550 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6553 * No need to take a reference to the parent because cgroup
6554 * core guarantees its existence.
6557 res_counter_init(&memcg
->res
, NULL
);
6558 res_counter_init(&memcg
->memsw
, NULL
);
6559 res_counter_init(&memcg
->kmem
, NULL
);
6561 * Deeper hierachy with use_hierarchy == false doesn't make
6562 * much sense so let cgroup subsystem know about this
6563 * unfortunate state in our controller.
6565 if (parent
!= root_mem_cgroup
)
6566 mem_cgroup_subsys
.broken_hierarchy
= true;
6569 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6570 mutex_unlock(&memcg_create_mutex
);
6575 * Announce all parents that a group from their hierarchy is gone.
6577 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6579 struct mem_cgroup
*parent
= memcg
;
6581 while ((parent
= parent_mem_cgroup(parent
)))
6582 mem_cgroup_iter_invalidate(parent
);
6585 * if the root memcg is not hierarchical we have to check it
6588 if (!root_mem_cgroup
->use_hierarchy
)
6589 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6592 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6594 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6595 struct mem_cgroup_event
*event
, *tmp
;
6598 * Unregister events and notify userspace.
6599 * Notify userspace about cgroup removing only after rmdir of cgroup
6600 * directory to avoid race between userspace and kernelspace.
6602 spin_lock(&memcg
->event_list_lock
);
6603 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
6604 list_del_init(&event
->list
);
6605 schedule_work(&event
->remove
);
6607 spin_unlock(&memcg
->event_list_lock
);
6609 kmem_cgroup_css_offline(memcg
);
6611 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6612 mem_cgroup_reparent_charges(memcg
);
6613 mem_cgroup_destroy_all_caches(memcg
);
6614 vmpressure_cleanup(&memcg
->vmpressure
);
6617 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6619 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6621 * XXX: css_offline() would be where we should reparent all
6622 * memory to prepare the cgroup for destruction. However,
6623 * memcg does not do css_tryget() and res_counter charging
6624 * under the same RCU lock region, which means that charging
6625 * could race with offlining. Offlining only happens to
6626 * cgroups with no tasks in them but charges can show up
6627 * without any tasks from the swapin path when the target
6628 * memcg is looked up from the swapout record and not from the
6629 * current task as it usually is. A race like this can leak
6630 * charges and put pages with stale cgroup pointers into
6634 * lookup_swap_cgroup_id()
6636 * mem_cgroup_lookup()
6639 * disable css_tryget()
6642 * reparent_charges()
6643 * res_counter_charge()
6646 * pc->mem_cgroup = dead memcg
6649 * The bulk of the charges are still moved in offline_css() to
6650 * avoid pinning a lot of pages in case a long-term reference
6651 * like a swapout record is deferring the css_free() to long
6652 * after offlining. But this makes sure we catch any charges
6653 * made after offlining:
6655 mem_cgroup_reparent_charges(memcg
);
6657 memcg_destroy_kmem(memcg
);
6658 __mem_cgroup_free(memcg
);
6662 /* Handlers for move charge at task migration. */
6663 #define PRECHARGE_COUNT_AT_ONCE 256
6664 static int mem_cgroup_do_precharge(unsigned long count
)
6667 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6668 struct mem_cgroup
*memcg
= mc
.to
;
6670 if (mem_cgroup_is_root(memcg
)) {
6671 mc
.precharge
+= count
;
6672 /* we don't need css_get for root */
6675 /* try to charge at once */
6677 struct res_counter
*dummy
;
6679 * "memcg" cannot be under rmdir() because we've already checked
6680 * by cgroup_lock_live_cgroup() that it is not removed and we
6681 * are still under the same cgroup_mutex. So we can postpone
6684 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6686 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6687 PAGE_SIZE
* count
, &dummy
)) {
6688 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6691 mc
.precharge
+= count
;
6695 /* fall back to one by one charge */
6697 if (signal_pending(current
)) {
6701 if (!batch_count
--) {
6702 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6705 ret
= __mem_cgroup_try_charge(NULL
,
6706 GFP_KERNEL
, 1, &memcg
, false);
6708 /* mem_cgroup_clear_mc() will do uncharge later */
6716 * get_mctgt_type - get target type of moving charge
6717 * @vma: the vma the pte to be checked belongs
6718 * @addr: the address corresponding to the pte to be checked
6719 * @ptent: the pte to be checked
6720 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6723 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6724 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6725 * move charge. if @target is not NULL, the page is stored in target->page
6726 * with extra refcnt got(Callers should handle it).
6727 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6728 * target for charge migration. if @target is not NULL, the entry is stored
6731 * Called with pte lock held.
6738 enum mc_target_type
{
6744 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6745 unsigned long addr
, pte_t ptent
)
6747 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6749 if (!page
|| !page_mapped(page
))
6751 if (PageAnon(page
)) {
6752 /* we don't move shared anon */
6755 } else if (!move_file())
6756 /* we ignore mapcount for file pages */
6758 if (!get_page_unless_zero(page
))
6765 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6766 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6768 struct page
*page
= NULL
;
6769 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6771 if (!move_anon() || non_swap_entry(ent
))
6774 * Because lookup_swap_cache() updates some statistics counter,
6775 * we call find_get_page() with swapper_space directly.
6777 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6778 if (do_swap_account
)
6779 entry
->val
= ent
.val
;
6784 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6785 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6791 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6792 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6794 struct page
*page
= NULL
;
6795 struct address_space
*mapping
;
6798 if (!vma
->vm_file
) /* anonymous vma */
6803 mapping
= vma
->vm_file
->f_mapping
;
6804 if (pte_none(ptent
))
6805 pgoff
= linear_page_index(vma
, addr
);
6806 else /* pte_file(ptent) is true */
6807 pgoff
= pte_to_pgoff(ptent
);
6809 /* page is moved even if it's not RSS of this task(page-faulted). */
6810 page
= find_get_page(mapping
, pgoff
);
6813 /* shmem/tmpfs may report page out on swap: account for that too. */
6814 if (radix_tree_exceptional_entry(page
)) {
6815 swp_entry_t swap
= radix_to_swp_entry(page
);
6816 if (do_swap_account
)
6818 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6824 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6825 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6827 struct page
*page
= NULL
;
6828 struct page_cgroup
*pc
;
6829 enum mc_target_type ret
= MC_TARGET_NONE
;
6830 swp_entry_t ent
= { .val
= 0 };
6832 if (pte_present(ptent
))
6833 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6834 else if (is_swap_pte(ptent
))
6835 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6836 else if (pte_none(ptent
) || pte_file(ptent
))
6837 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6839 if (!page
&& !ent
.val
)
6842 pc
= lookup_page_cgroup(page
);
6844 * Do only loose check w/o page_cgroup lock.
6845 * mem_cgroup_move_account() checks the pc is valid or not under
6848 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6849 ret
= MC_TARGET_PAGE
;
6851 target
->page
= page
;
6853 if (!ret
|| !target
)
6856 /* There is a swap entry and a page doesn't exist or isn't charged */
6857 if (ent
.val
&& !ret
&&
6858 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6859 ret
= MC_TARGET_SWAP
;
6866 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6868 * We don't consider swapping or file mapped pages because THP does not
6869 * support them for now.
6870 * Caller should make sure that pmd_trans_huge(pmd) is true.
6872 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6873 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6875 struct page
*page
= NULL
;
6876 struct page_cgroup
*pc
;
6877 enum mc_target_type ret
= MC_TARGET_NONE
;
6879 page
= pmd_page(pmd
);
6880 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
6883 pc
= lookup_page_cgroup(page
);
6884 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6885 ret
= MC_TARGET_PAGE
;
6888 target
->page
= page
;
6894 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6895 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6897 return MC_TARGET_NONE
;
6901 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6902 unsigned long addr
, unsigned long end
,
6903 struct mm_walk
*walk
)
6905 struct vm_area_struct
*vma
= walk
->private;
6909 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6910 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6911 mc
.precharge
+= HPAGE_PMD_NR
;
6916 if (pmd_trans_unstable(pmd
))
6918 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6919 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6920 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6921 mc
.precharge
++; /* increment precharge temporarily */
6922 pte_unmap_unlock(pte
- 1, ptl
);
6928 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6930 unsigned long precharge
;
6931 struct vm_area_struct
*vma
;
6933 down_read(&mm
->mmap_sem
);
6934 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6935 struct mm_walk mem_cgroup_count_precharge_walk
= {
6936 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6940 if (is_vm_hugetlb_page(vma
))
6942 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6943 &mem_cgroup_count_precharge_walk
);
6945 up_read(&mm
->mmap_sem
);
6947 precharge
= mc
.precharge
;
6953 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6955 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6957 VM_BUG_ON(mc
.moving_task
);
6958 mc
.moving_task
= current
;
6959 return mem_cgroup_do_precharge(precharge
);
6962 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6963 static void __mem_cgroup_clear_mc(void)
6965 struct mem_cgroup
*from
= mc
.from
;
6966 struct mem_cgroup
*to
= mc
.to
;
6969 /* we must uncharge all the leftover precharges from mc.to */
6971 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6975 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6976 * we must uncharge here.
6978 if (mc
.moved_charge
) {
6979 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6980 mc
.moved_charge
= 0;
6982 /* we must fixup refcnts and charges */
6983 if (mc
.moved_swap
) {
6984 /* uncharge swap account from the old cgroup */
6985 if (!mem_cgroup_is_root(mc
.from
))
6986 res_counter_uncharge(&mc
.from
->memsw
,
6987 PAGE_SIZE
* mc
.moved_swap
);
6989 for (i
= 0; i
< mc
.moved_swap
; i
++)
6990 css_put(&mc
.from
->css
);
6992 if (!mem_cgroup_is_root(mc
.to
)) {
6994 * we charged both to->res and to->memsw, so we should
6997 res_counter_uncharge(&mc
.to
->res
,
6998 PAGE_SIZE
* mc
.moved_swap
);
7000 /* we've already done css_get(mc.to) */
7003 memcg_oom_recover(from
);
7004 memcg_oom_recover(to
);
7005 wake_up_all(&mc
.waitq
);
7008 static void mem_cgroup_clear_mc(void)
7010 struct mem_cgroup
*from
= mc
.from
;
7013 * we must clear moving_task before waking up waiters at the end of
7016 mc
.moving_task
= NULL
;
7017 __mem_cgroup_clear_mc();
7018 spin_lock(&mc
.lock
);
7021 spin_unlock(&mc
.lock
);
7022 mem_cgroup_end_move(from
);
7025 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7026 struct cgroup_taskset
*tset
)
7028 struct task_struct
*p
= cgroup_taskset_first(tset
);
7030 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7031 unsigned long move_charge_at_immigrate
;
7034 * We are now commited to this value whatever it is. Changes in this
7035 * tunable will only affect upcoming migrations, not the current one.
7036 * So we need to save it, and keep it going.
7038 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
7039 if (move_charge_at_immigrate
) {
7040 struct mm_struct
*mm
;
7041 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
7043 VM_BUG_ON(from
== memcg
);
7045 mm
= get_task_mm(p
);
7048 /* We move charges only when we move a owner of the mm */
7049 if (mm
->owner
== p
) {
7052 VM_BUG_ON(mc
.precharge
);
7053 VM_BUG_ON(mc
.moved_charge
);
7054 VM_BUG_ON(mc
.moved_swap
);
7055 mem_cgroup_start_move(from
);
7056 spin_lock(&mc
.lock
);
7059 mc
.immigrate_flags
= move_charge_at_immigrate
;
7060 spin_unlock(&mc
.lock
);
7061 /* We set mc.moving_task later */
7063 ret
= mem_cgroup_precharge_mc(mm
);
7065 mem_cgroup_clear_mc();
7072 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
7073 struct cgroup_taskset
*tset
)
7075 mem_cgroup_clear_mc();
7078 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
7079 unsigned long addr
, unsigned long end
,
7080 struct mm_walk
*walk
)
7083 struct vm_area_struct
*vma
= walk
->private;
7086 enum mc_target_type target_type
;
7087 union mc_target target
;
7089 struct page_cgroup
*pc
;
7092 * We don't take compound_lock() here but no race with splitting thp
7094 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7095 * under splitting, which means there's no concurrent thp split,
7096 * - if another thread runs into split_huge_page() just after we
7097 * entered this if-block, the thread must wait for page table lock
7098 * to be unlocked in __split_huge_page_splitting(), where the main
7099 * part of thp split is not executed yet.
7101 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
7102 if (mc
.precharge
< HPAGE_PMD_NR
) {
7106 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
7107 if (target_type
== MC_TARGET_PAGE
) {
7109 if (!isolate_lru_page(page
)) {
7110 pc
= lookup_page_cgroup(page
);
7111 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
7112 pc
, mc
.from
, mc
.to
)) {
7113 mc
.precharge
-= HPAGE_PMD_NR
;
7114 mc
.moved_charge
+= HPAGE_PMD_NR
;
7116 putback_lru_page(page
);
7124 if (pmd_trans_unstable(pmd
))
7127 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
7128 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
7129 pte_t ptent
= *(pte
++);
7135 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
7136 case MC_TARGET_PAGE
:
7138 if (isolate_lru_page(page
))
7140 pc
= lookup_page_cgroup(page
);
7141 if (!mem_cgroup_move_account(page
, 1, pc
,
7144 /* we uncharge from mc.from later. */
7147 putback_lru_page(page
);
7148 put
: /* get_mctgt_type() gets the page */
7151 case MC_TARGET_SWAP
:
7153 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
7155 /* we fixup refcnts and charges later. */
7163 pte_unmap_unlock(pte
- 1, ptl
);
7168 * We have consumed all precharges we got in can_attach().
7169 * We try charge one by one, but don't do any additional
7170 * charges to mc.to if we have failed in charge once in attach()
7173 ret
= mem_cgroup_do_precharge(1);
7181 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
7183 struct vm_area_struct
*vma
;
7185 lru_add_drain_all();
7187 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
7189 * Someone who are holding the mmap_sem might be waiting in
7190 * waitq. So we cancel all extra charges, wake up all waiters,
7191 * and retry. Because we cancel precharges, we might not be able
7192 * to move enough charges, but moving charge is a best-effort
7193 * feature anyway, so it wouldn't be a big problem.
7195 __mem_cgroup_clear_mc();
7199 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
7201 struct mm_walk mem_cgroup_move_charge_walk
= {
7202 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
7206 if (is_vm_hugetlb_page(vma
))
7208 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
7209 &mem_cgroup_move_charge_walk
);
7212 * means we have consumed all precharges and failed in
7213 * doing additional charge. Just abandon here.
7217 up_read(&mm
->mmap_sem
);
7220 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
7221 struct cgroup_taskset
*tset
)
7223 struct task_struct
*p
= cgroup_taskset_first(tset
);
7224 struct mm_struct
*mm
= get_task_mm(p
);
7228 mem_cgroup_move_charge(mm
);
7232 mem_cgroup_clear_mc();
7234 #else /* !CONFIG_MMU */
7235 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7236 struct cgroup_taskset
*tset
)
7240 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
7241 struct cgroup_taskset
*tset
)
7244 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
7245 struct cgroup_taskset
*tset
)
7251 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7252 * to verify sane_behavior flag on each mount attempt.
7254 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
7257 * use_hierarchy is forced with sane_behavior. cgroup core
7258 * guarantees that @root doesn't have any children, so turning it
7259 * on for the root memcg is enough.
7261 if (cgroup_sane_behavior(root_css
->cgroup
))
7262 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
7265 struct cgroup_subsys mem_cgroup_subsys
= {
7267 .subsys_id
= mem_cgroup_subsys_id
,
7268 .css_alloc
= mem_cgroup_css_alloc
,
7269 .css_online
= mem_cgroup_css_online
,
7270 .css_offline
= mem_cgroup_css_offline
,
7271 .css_free
= mem_cgroup_css_free
,
7272 .can_attach
= mem_cgroup_can_attach
,
7273 .cancel_attach
= mem_cgroup_cancel_attach
,
7274 .attach
= mem_cgroup_move_task
,
7275 .bind
= mem_cgroup_bind
,
7276 .base_cftypes
= mem_cgroup_files
,
7280 #ifdef CONFIG_MEMCG_SWAP
7281 static int __init
enable_swap_account(char *s
)
7283 if (!strcmp(s
, "1"))
7284 really_do_swap_account
= 1;
7285 else if (!strcmp(s
, "0"))
7286 really_do_swap_account
= 0;
7289 __setup("swapaccount=", enable_swap_account
);
7291 static void __init
memsw_file_init(void)
7293 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
7296 static void __init
enable_swap_cgroup(void)
7298 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7299 do_swap_account
= 1;
7305 static void __init
enable_swap_cgroup(void)
7311 * subsys_initcall() for memory controller.
7313 * Some parts like hotcpu_notifier() have to be initialized from this context
7314 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7315 * everything that doesn't depend on a specific mem_cgroup structure should
7316 * be initialized from here.
7318 static int __init
mem_cgroup_init(void)
7320 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7321 enable_swap_cgroup();
7322 mem_cgroup_soft_limit_tree_init();
7326 subsys_initcall(mem_cgroup_init
);