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/vmalloc.h>
53 #include <linux/vmpressure.h>
54 #include <linux/mm_inline.h>
55 #include <linux/page_cgroup.h>
56 #include <linux/cpu.h>
57 #include <linux/oom.h>
58 #include <linux/lockdep.h>
59 #include <linux/file.h>
63 #include <net/tcp_memcontrol.h>
66 #include <asm/uaccess.h>
68 #include <trace/events/vmscan.h>
70 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
71 EXPORT_SYMBOL(mem_cgroup_subsys
);
73 #define MEM_CGROUP_RECLAIM_RETRIES 5
74 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
76 #ifdef CONFIG_MEMCG_SWAP
77 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
78 int do_swap_account __read_mostly
;
80 /* for remember boot option*/
81 #ifdef CONFIG_MEMCG_SWAP_ENABLED
82 static int really_do_swap_account __initdata
= 1;
84 static int really_do_swap_account __initdata
= 0;
88 #define do_swap_account 0
92 static const char * const mem_cgroup_stat_names
[] = {
101 enum mem_cgroup_events_index
{
102 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
103 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
104 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
105 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
106 MEM_CGROUP_EVENTS_NSTATS
,
109 static const char * const mem_cgroup_events_names
[] = {
116 static const char * const mem_cgroup_lru_names
[] = {
125 * Per memcg event counter is incremented at every pagein/pageout. With THP,
126 * it will be incremated by the number of pages. This counter is used for
127 * for trigger some periodic events. This is straightforward and better
128 * than using jiffies etc. to handle periodic memcg event.
130 enum mem_cgroup_events_target
{
131 MEM_CGROUP_TARGET_THRESH
,
132 MEM_CGROUP_TARGET_SOFTLIMIT
,
133 MEM_CGROUP_TARGET_NUMAINFO
,
136 #define THRESHOLDS_EVENTS_TARGET 128
137 #define SOFTLIMIT_EVENTS_TARGET 1024
138 #define NUMAINFO_EVENTS_TARGET 1024
140 struct mem_cgroup_stat_cpu
{
141 long count
[MEM_CGROUP_STAT_NSTATS
];
142 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
143 unsigned long nr_page_events
;
144 unsigned long targets
[MEM_CGROUP_NTARGETS
];
147 struct mem_cgroup_reclaim_iter
{
149 * last scanned hierarchy member. Valid only if last_dead_count
150 * matches memcg->dead_count of the hierarchy root group.
152 struct mem_cgroup
*last_visited
;
153 unsigned long last_dead_count
;
155 /* scan generation, increased every round-trip */
156 unsigned int generation
;
160 * per-zone information in memory controller.
162 struct mem_cgroup_per_zone
{
163 struct lruvec lruvec
;
164 unsigned long lru_size
[NR_LRU_LISTS
];
166 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
168 struct rb_node tree_node
; /* RB tree node */
169 unsigned long long usage_in_excess
;/* Set to the value by which */
170 /* the soft limit is exceeded*/
172 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
173 /* use container_of */
176 struct mem_cgroup_per_node
{
177 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
181 * Cgroups above their limits are maintained in a RB-Tree, independent of
182 * their hierarchy representation
185 struct mem_cgroup_tree_per_zone
{
186 struct rb_root rb_root
;
190 struct mem_cgroup_tree_per_node
{
191 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
194 struct mem_cgroup_tree
{
195 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
198 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
200 struct mem_cgroup_threshold
{
201 struct eventfd_ctx
*eventfd
;
206 struct mem_cgroup_threshold_ary
{
207 /* An array index points to threshold just below or equal to usage. */
208 int current_threshold
;
209 /* Size of entries[] */
211 /* Array of thresholds */
212 struct mem_cgroup_threshold entries
[0];
215 struct mem_cgroup_thresholds
{
216 /* Primary thresholds array */
217 struct mem_cgroup_threshold_ary
*primary
;
219 * Spare threshold array.
220 * This is needed to make mem_cgroup_unregister_event() "never fail".
221 * It must be able to store at least primary->size - 1 entries.
223 struct mem_cgroup_threshold_ary
*spare
;
227 struct mem_cgroup_eventfd_list
{
228 struct list_head list
;
229 struct eventfd_ctx
*eventfd
;
233 * cgroup_event represents events which userspace want to receive.
235 struct mem_cgroup_event
{
237 * memcg which the event belongs to.
239 struct mem_cgroup
*memcg
;
241 * eventfd to signal userspace about the event.
243 struct eventfd_ctx
*eventfd
;
245 * Each of these stored in a list by the cgroup.
247 struct list_head list
;
249 * register_event() callback will be used to add new userspace
250 * waiter for changes related to this event. Use eventfd_signal()
251 * on eventfd to send notification to userspace.
253 int (*register_event
)(struct mem_cgroup
*memcg
,
254 struct eventfd_ctx
*eventfd
, const char *args
);
256 * unregister_event() callback will be called when userspace closes
257 * the eventfd or on cgroup removing. This callback must be set,
258 * if you want provide notification functionality.
260 void (*unregister_event
)(struct mem_cgroup
*memcg
,
261 struct eventfd_ctx
*eventfd
);
263 * All fields below needed to unregister event when
264 * userspace closes eventfd.
267 wait_queue_head_t
*wqh
;
269 struct work_struct remove
;
272 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
273 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
276 * The memory controller data structure. The memory controller controls both
277 * page cache and RSS per cgroup. We would eventually like to provide
278 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
279 * to help the administrator determine what knobs to tune.
281 * TODO: Add a water mark for the memory controller. Reclaim will begin when
282 * we hit the water mark. May be even add a low water mark, such that
283 * no reclaim occurs from a cgroup at it's low water mark, this is
284 * a feature that will be implemented much later in the future.
287 struct cgroup_subsys_state css
;
289 * the counter to account for memory usage
291 struct res_counter res
;
293 /* vmpressure notifications */
294 struct vmpressure vmpressure
;
297 * the counter to account for mem+swap usage.
299 struct res_counter memsw
;
302 * the counter to account for kernel memory usage.
304 struct res_counter kmem
;
306 * Should the accounting and control be hierarchical, per subtree?
309 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
313 atomic_t oom_wakeups
;
316 /* OOM-Killer disable */
317 int oom_kill_disable
;
319 /* set when res.limit == memsw.limit */
320 bool memsw_is_minimum
;
322 /* protect arrays of thresholds */
323 struct mutex thresholds_lock
;
325 /* thresholds for memory usage. RCU-protected */
326 struct mem_cgroup_thresholds thresholds
;
328 /* thresholds for mem+swap usage. RCU-protected */
329 struct mem_cgroup_thresholds memsw_thresholds
;
331 /* For oom notifier event fd */
332 struct list_head oom_notify
;
335 * Should we move charges of a task when a task is moved into this
336 * mem_cgroup ? And what type of charges should we move ?
338 unsigned long move_charge_at_immigrate
;
340 * set > 0 if pages under this cgroup are moving to other cgroup.
342 atomic_t moving_account
;
343 /* taken only while moving_account > 0 */
344 spinlock_t move_lock
;
348 struct mem_cgroup_stat_cpu __percpu
*stat
;
350 * used when a cpu is offlined or other synchronizations
351 * See mem_cgroup_read_stat().
353 struct mem_cgroup_stat_cpu nocpu_base
;
354 spinlock_t pcp_counter_lock
;
357 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
358 struct cg_proto tcp_mem
;
360 #if defined(CONFIG_MEMCG_KMEM)
361 /* analogous to slab_common's slab_caches list. per-memcg */
362 struct list_head memcg_slab_caches
;
363 /* Not a spinlock, we can take a lot of time walking the list */
364 struct mutex slab_caches_mutex
;
365 /* Index in the kmem_cache->memcg_params->memcg_caches array */
369 int last_scanned_node
;
371 nodemask_t scan_nodes
;
372 atomic_t numainfo_events
;
373 atomic_t numainfo_updating
;
376 /* List of events which userspace want to receive */
377 struct list_head event_list
;
378 spinlock_t event_list_lock
;
380 struct mem_cgroup_per_node
*nodeinfo
[0];
381 /* WARNING: nodeinfo must be the last member here */
384 static size_t memcg_size(void)
386 return sizeof(struct mem_cgroup
) +
387 nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
390 /* internal only representation about the status of kmem accounting. */
392 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
393 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
394 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
397 /* We account when limit is on, but only after call sites are patched */
398 #define KMEM_ACCOUNTED_MASK \
399 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
401 #ifdef CONFIG_MEMCG_KMEM
402 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
404 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
407 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
409 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
412 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
414 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
417 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
419 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
422 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
425 * Our caller must use css_get() first, because memcg_uncharge_kmem()
426 * will call css_put() if it sees the memcg is dead.
429 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
430 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
433 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
435 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
436 &memcg
->kmem_account_flags
);
440 /* Stuffs for move charges at task migration. */
442 * Types of charges to be moved. "move_charge_at_immitgrate" and
443 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
446 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
447 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
451 /* "mc" and its members are protected by cgroup_mutex */
452 static struct move_charge_struct
{
453 spinlock_t lock
; /* for from, to */
454 struct mem_cgroup
*from
;
455 struct mem_cgroup
*to
;
456 unsigned long immigrate_flags
;
457 unsigned long precharge
;
458 unsigned long moved_charge
;
459 unsigned long moved_swap
;
460 struct task_struct
*moving_task
; /* a task moving charges */
461 wait_queue_head_t waitq
; /* a waitq for other context */
463 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
464 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
467 static bool move_anon(void)
469 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
472 static bool move_file(void)
474 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
478 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
479 * limit reclaim to prevent infinite loops, if they ever occur.
481 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
482 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
485 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
486 MEM_CGROUP_CHARGE_TYPE_ANON
,
487 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
488 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
492 /* for encoding cft->private value on file */
500 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
501 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
502 #define MEMFILE_ATTR(val) ((val) & 0xffff)
503 /* Used for OOM nofiier */
504 #define OOM_CONTROL (0)
507 * Reclaim flags for mem_cgroup_hierarchical_reclaim
509 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
510 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
511 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
512 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
515 * The memcg_create_mutex will be held whenever a new cgroup is created.
516 * As a consequence, any change that needs to protect against new child cgroups
517 * appearing has to hold it as well.
519 static DEFINE_MUTEX(memcg_create_mutex
);
521 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
523 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
526 /* Some nice accessors for the vmpressure. */
527 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
530 memcg
= root_mem_cgroup
;
531 return &memcg
->vmpressure
;
534 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
536 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
539 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
541 return (memcg
== root_mem_cgroup
);
545 * We restrict the id in the range of [1, 65535], so it can fit into
548 #define MEM_CGROUP_ID_MAX USHRT_MAX
550 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
553 * The ID of the root cgroup is 0, but memcg treat 0 as an
554 * invalid ID, so we return (cgroup_id + 1).
556 return memcg
->css
.cgroup
->id
+ 1;
559 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
561 struct cgroup_subsys_state
*css
;
563 css
= css_from_id(id
- 1, &mem_cgroup_subsys
);
564 return mem_cgroup_from_css(css
);
567 /* Writing them here to avoid exposing memcg's inner layout */
568 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
570 void sock_update_memcg(struct sock
*sk
)
572 if (mem_cgroup_sockets_enabled
) {
573 struct mem_cgroup
*memcg
;
574 struct cg_proto
*cg_proto
;
576 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
578 /* Socket cloning can throw us here with sk_cgrp already
579 * filled. It won't however, necessarily happen from
580 * process context. So the test for root memcg given
581 * the current task's memcg won't help us in this case.
583 * Respecting the original socket's memcg is a better
584 * decision in this case.
587 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
588 css_get(&sk
->sk_cgrp
->memcg
->css
);
593 memcg
= mem_cgroup_from_task(current
);
594 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
595 if (!mem_cgroup_is_root(memcg
) &&
596 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
597 sk
->sk_cgrp
= cg_proto
;
602 EXPORT_SYMBOL(sock_update_memcg
);
604 void sock_release_memcg(struct sock
*sk
)
606 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
607 struct mem_cgroup
*memcg
;
608 WARN_ON(!sk
->sk_cgrp
->memcg
);
609 memcg
= sk
->sk_cgrp
->memcg
;
610 css_put(&sk
->sk_cgrp
->memcg
->css
);
614 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
616 if (!memcg
|| mem_cgroup_is_root(memcg
))
619 return &memcg
->tcp_mem
;
621 EXPORT_SYMBOL(tcp_proto_cgroup
);
623 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
625 if (!memcg_proto_activated(&memcg
->tcp_mem
))
627 static_key_slow_dec(&memcg_socket_limit_enabled
);
630 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
635 #ifdef CONFIG_MEMCG_KMEM
637 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
638 * The main reason for not using cgroup id for this:
639 * this works better in sparse environments, where we have a lot of memcgs,
640 * but only a few kmem-limited. Or also, if we have, for instance, 200
641 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
642 * 200 entry array for that.
644 * The current size of the caches array is stored in
645 * memcg_limited_groups_array_size. It will double each time we have to
648 static DEFINE_IDA(kmem_limited_groups
);
649 int memcg_limited_groups_array_size
;
652 * MIN_SIZE is different than 1, because we would like to avoid going through
653 * the alloc/free process all the time. In a small machine, 4 kmem-limited
654 * cgroups is a reasonable guess. In the future, it could be a parameter or
655 * tunable, but that is strictly not necessary.
657 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
658 * this constant directly from cgroup, but it is understandable that this is
659 * better kept as an internal representation in cgroup.c. In any case, the
660 * cgrp_id space is not getting any smaller, and we don't have to necessarily
661 * increase ours as well if it increases.
663 #define MEMCG_CACHES_MIN_SIZE 4
664 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
667 * A lot of the calls to the cache allocation functions are expected to be
668 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
669 * conditional to this static branch, we'll have to allow modules that does
670 * kmem_cache_alloc and the such to see this symbol as well
672 struct static_key memcg_kmem_enabled_key
;
673 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
675 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
677 if (memcg_kmem_is_active(memcg
)) {
678 static_key_slow_dec(&memcg_kmem_enabled_key
);
679 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
682 * This check can't live in kmem destruction function,
683 * since the charges will outlive the cgroup
685 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
688 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
691 #endif /* CONFIG_MEMCG_KMEM */
693 static void disarm_static_keys(struct mem_cgroup
*memcg
)
695 disarm_sock_keys(memcg
);
696 disarm_kmem_keys(memcg
);
699 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
701 static struct mem_cgroup_per_zone
*
702 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
704 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
705 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
708 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
713 static struct mem_cgroup_per_zone
*
714 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
716 int nid
= page_to_nid(page
);
717 int zid
= page_zonenum(page
);
719 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
722 static struct mem_cgroup_tree_per_zone
*
723 soft_limit_tree_node_zone(int nid
, int zid
)
725 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
728 static struct mem_cgroup_tree_per_zone
*
729 soft_limit_tree_from_page(struct page
*page
)
731 int nid
= page_to_nid(page
);
732 int zid
= page_zonenum(page
);
734 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
738 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
739 struct mem_cgroup_per_zone
*mz
,
740 struct mem_cgroup_tree_per_zone
*mctz
,
741 unsigned long long new_usage_in_excess
)
743 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
744 struct rb_node
*parent
= NULL
;
745 struct mem_cgroup_per_zone
*mz_node
;
750 mz
->usage_in_excess
= new_usage_in_excess
;
751 if (!mz
->usage_in_excess
)
755 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
757 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
760 * We can't avoid mem cgroups that are over their soft
761 * limit by the same amount
763 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
766 rb_link_node(&mz
->tree_node
, parent
, p
);
767 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
772 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
773 struct mem_cgroup_per_zone
*mz
,
774 struct mem_cgroup_tree_per_zone
*mctz
)
778 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
783 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
784 struct mem_cgroup_per_zone
*mz
,
785 struct mem_cgroup_tree_per_zone
*mctz
)
787 spin_lock(&mctz
->lock
);
788 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
789 spin_unlock(&mctz
->lock
);
793 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
795 unsigned long long excess
;
796 struct mem_cgroup_per_zone
*mz
;
797 struct mem_cgroup_tree_per_zone
*mctz
;
798 int nid
= page_to_nid(page
);
799 int zid
= page_zonenum(page
);
800 mctz
= soft_limit_tree_from_page(page
);
803 * Necessary to update all ancestors when hierarchy is used.
804 * because their event counter is not touched.
806 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
807 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
808 excess
= res_counter_soft_limit_excess(&memcg
->res
);
810 * We have to update the tree if mz is on RB-tree or
811 * mem is over its softlimit.
813 if (excess
|| mz
->on_tree
) {
814 spin_lock(&mctz
->lock
);
815 /* if on-tree, remove it */
817 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
819 * Insert again. mz->usage_in_excess will be updated.
820 * If excess is 0, no tree ops.
822 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
823 spin_unlock(&mctz
->lock
);
828 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
831 struct mem_cgroup_per_zone
*mz
;
832 struct mem_cgroup_tree_per_zone
*mctz
;
834 for_each_node(node
) {
835 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
836 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
837 mctz
= soft_limit_tree_node_zone(node
, zone
);
838 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
843 static struct mem_cgroup_per_zone
*
844 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
846 struct rb_node
*rightmost
= NULL
;
847 struct mem_cgroup_per_zone
*mz
;
851 rightmost
= rb_last(&mctz
->rb_root
);
853 goto done
; /* Nothing to reclaim from */
855 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
857 * Remove the node now but someone else can add it back,
858 * we will to add it back at the end of reclaim to its correct
859 * position in the tree.
861 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
862 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
863 !css_tryget(&mz
->memcg
->css
))
869 static struct mem_cgroup_per_zone
*
870 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
872 struct mem_cgroup_per_zone
*mz
;
874 spin_lock(&mctz
->lock
);
875 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
876 spin_unlock(&mctz
->lock
);
881 * Implementation Note: reading percpu statistics for memcg.
883 * Both of vmstat[] and percpu_counter has threshold and do periodic
884 * synchronization to implement "quick" read. There are trade-off between
885 * reading cost and precision of value. Then, we may have a chance to implement
886 * a periodic synchronizion of counter in memcg's counter.
888 * But this _read() function is used for user interface now. The user accounts
889 * memory usage by memory cgroup and he _always_ requires exact value because
890 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
891 * have to visit all online cpus and make sum. So, for now, unnecessary
892 * synchronization is not implemented. (just implemented for cpu hotplug)
894 * If there are kernel internal actions which can make use of some not-exact
895 * value, and reading all cpu value can be performance bottleneck in some
896 * common workload, threashold and synchonization as vmstat[] should be
899 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
900 enum mem_cgroup_stat_index idx
)
906 for_each_online_cpu(cpu
)
907 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
908 #ifdef CONFIG_HOTPLUG_CPU
909 spin_lock(&memcg
->pcp_counter_lock
);
910 val
+= memcg
->nocpu_base
.count
[idx
];
911 spin_unlock(&memcg
->pcp_counter_lock
);
917 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
920 int val
= (charge
) ? 1 : -1;
921 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
924 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
925 enum mem_cgroup_events_index idx
)
927 unsigned long val
= 0;
931 for_each_online_cpu(cpu
)
932 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
933 #ifdef CONFIG_HOTPLUG_CPU
934 spin_lock(&memcg
->pcp_counter_lock
);
935 val
+= memcg
->nocpu_base
.events
[idx
];
936 spin_unlock(&memcg
->pcp_counter_lock
);
942 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
944 bool anon
, int nr_pages
)
949 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
950 * counted as CACHE even if it's on ANON LRU.
953 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
956 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
959 if (PageTransHuge(page
))
960 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
963 /* pagein of a big page is an event. So, ignore page size */
965 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
967 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
968 nr_pages
= -nr_pages
; /* for event */
971 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
977 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
979 struct mem_cgroup_per_zone
*mz
;
981 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
982 return mz
->lru_size
[lru
];
986 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
987 unsigned int lru_mask
)
989 struct mem_cgroup_per_zone
*mz
;
991 unsigned long ret
= 0;
993 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
996 if (BIT(lru
) & lru_mask
)
997 ret
+= mz
->lru_size
[lru
];
1002 static unsigned long
1003 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
1004 int nid
, unsigned int lru_mask
)
1009 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
1010 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
1011 nid
, zid
, lru_mask
);
1016 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
1017 unsigned int lru_mask
)
1022 for_each_node_state(nid
, N_MEMORY
)
1023 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
1027 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
1028 enum mem_cgroup_events_target target
)
1030 unsigned long val
, next
;
1032 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
1033 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
1034 /* from time_after() in jiffies.h */
1035 if ((long)next
- (long)val
< 0) {
1037 case MEM_CGROUP_TARGET_THRESH
:
1038 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
1040 case MEM_CGROUP_TARGET_SOFTLIMIT
:
1041 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1043 case MEM_CGROUP_TARGET_NUMAINFO
:
1044 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1049 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1056 * Check events in order.
1059 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1062 /* threshold event is triggered in finer grain than soft limit */
1063 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1064 MEM_CGROUP_TARGET_THRESH
))) {
1066 bool do_numainfo __maybe_unused
;
1068 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1069 MEM_CGROUP_TARGET_SOFTLIMIT
);
1070 #if MAX_NUMNODES > 1
1071 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1072 MEM_CGROUP_TARGET_NUMAINFO
);
1076 mem_cgroup_threshold(memcg
);
1077 if (unlikely(do_softlimit
))
1078 mem_cgroup_update_tree(memcg
, page
);
1079 #if MAX_NUMNODES > 1
1080 if (unlikely(do_numainfo
))
1081 atomic_inc(&memcg
->numainfo_events
);
1087 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1090 * mm_update_next_owner() may clear mm->owner to NULL
1091 * if it races with swapoff, page migration, etc.
1092 * So this can be called with p == NULL.
1097 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
1100 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1102 struct mem_cgroup
*memcg
= NULL
;
1107 * Because we have no locks, mm->owner's may be being moved to other
1108 * cgroup. We use css_tryget() here even if this looks
1109 * pessimistic (rather than adding locks here).
1113 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1114 if (unlikely(!memcg
))
1116 } while (!css_tryget(&memcg
->css
));
1122 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1123 * ref. count) or NULL if the whole root's subtree has been visited.
1125 * helper function to be used by mem_cgroup_iter
1127 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1128 struct mem_cgroup
*last_visited
)
1130 struct cgroup_subsys_state
*prev_css
, *next_css
;
1132 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1134 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1137 * Even if we found a group we have to make sure it is
1138 * alive. css && !memcg means that the groups should be
1139 * skipped and we should continue the tree walk.
1140 * last_visited css is safe to use because it is
1141 * protected by css_get and the tree walk is rcu safe.
1144 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1146 if (css_tryget(&mem
->css
))
1149 prev_css
= next_css
;
1157 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1160 * When a group in the hierarchy below root is destroyed, the
1161 * hierarchy iterator can no longer be trusted since it might
1162 * have pointed to the destroyed group. Invalidate it.
1164 atomic_inc(&root
->dead_count
);
1167 static struct mem_cgroup
*
1168 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1169 struct mem_cgroup
*root
,
1172 struct mem_cgroup
*position
= NULL
;
1174 * A cgroup destruction happens in two stages: offlining and
1175 * release. They are separated by a RCU grace period.
1177 * If the iterator is valid, we may still race with an
1178 * offlining. The RCU lock ensures the object won't be
1179 * released, tryget will fail if we lost the race.
1181 *sequence
= atomic_read(&root
->dead_count
);
1182 if (iter
->last_dead_count
== *sequence
) {
1184 position
= iter
->last_visited
;
1185 if (position
&& !css_tryget(&position
->css
))
1191 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1192 struct mem_cgroup
*last_visited
,
1193 struct mem_cgroup
*new_position
,
1197 css_put(&last_visited
->css
);
1199 * We store the sequence count from the time @last_visited was
1200 * loaded successfully instead of rereading it here so that we
1201 * don't lose destruction events in between. We could have
1202 * raced with the destruction of @new_position after all.
1204 iter
->last_visited
= new_position
;
1206 iter
->last_dead_count
= sequence
;
1210 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1211 * @root: hierarchy root
1212 * @prev: previously returned memcg, NULL on first invocation
1213 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1215 * Returns references to children of the hierarchy below @root, or
1216 * @root itself, or %NULL after a full round-trip.
1218 * Caller must pass the return value in @prev on subsequent
1219 * invocations for reference counting, or use mem_cgroup_iter_break()
1220 * to cancel a hierarchy walk before the round-trip is complete.
1222 * Reclaimers can specify a zone and a priority level in @reclaim to
1223 * divide up the memcgs in the hierarchy among all concurrent
1224 * reclaimers operating on the same zone and priority.
1226 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1227 struct mem_cgroup
*prev
,
1228 struct mem_cgroup_reclaim_cookie
*reclaim
)
1230 struct mem_cgroup
*memcg
= NULL
;
1231 struct mem_cgroup
*last_visited
= NULL
;
1233 if (mem_cgroup_disabled())
1237 root
= root_mem_cgroup
;
1239 if (prev
&& !reclaim
)
1240 last_visited
= prev
;
1242 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1250 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1251 int uninitialized_var(seq
);
1254 int nid
= zone_to_nid(reclaim
->zone
);
1255 int zid
= zone_idx(reclaim
->zone
);
1256 struct mem_cgroup_per_zone
*mz
;
1258 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1259 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1260 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1261 iter
->last_visited
= NULL
;
1265 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1268 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1271 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1275 else if (!prev
&& memcg
)
1276 reclaim
->generation
= iter
->generation
;
1285 if (prev
&& prev
!= root
)
1286 css_put(&prev
->css
);
1292 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1293 * @root: hierarchy root
1294 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1296 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1297 struct mem_cgroup
*prev
)
1300 root
= root_mem_cgroup
;
1301 if (prev
&& prev
!= root
)
1302 css_put(&prev
->css
);
1306 * Iteration constructs for visiting all cgroups (under a tree). If
1307 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1308 * be used for reference counting.
1310 #define for_each_mem_cgroup_tree(iter, root) \
1311 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1313 iter = mem_cgroup_iter(root, iter, NULL))
1315 #define for_each_mem_cgroup(iter) \
1316 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1318 iter = mem_cgroup_iter(NULL, iter, NULL))
1320 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1322 struct mem_cgroup
*memcg
;
1325 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1326 if (unlikely(!memcg
))
1331 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1334 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1342 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1345 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1346 * @zone: zone of the wanted lruvec
1347 * @memcg: memcg of the wanted lruvec
1349 * Returns the lru list vector holding pages for the given @zone and
1350 * @mem. This can be the global zone lruvec, if the memory controller
1353 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1354 struct mem_cgroup
*memcg
)
1356 struct mem_cgroup_per_zone
*mz
;
1357 struct lruvec
*lruvec
;
1359 if (mem_cgroup_disabled()) {
1360 lruvec
= &zone
->lruvec
;
1364 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1365 lruvec
= &mz
->lruvec
;
1368 * Since a node can be onlined after the mem_cgroup was created,
1369 * we have to be prepared to initialize lruvec->zone here;
1370 * and if offlined then reonlined, we need to reinitialize it.
1372 if (unlikely(lruvec
->zone
!= zone
))
1373 lruvec
->zone
= zone
;
1378 * Following LRU functions are allowed to be used without PCG_LOCK.
1379 * Operations are called by routine of global LRU independently from memcg.
1380 * What we have to take care of here is validness of pc->mem_cgroup.
1382 * Changes to pc->mem_cgroup happens when
1385 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1386 * It is added to LRU before charge.
1387 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1388 * When moving account, the page is not on LRU. It's isolated.
1392 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1394 * @zone: zone of the page
1396 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1398 struct mem_cgroup_per_zone
*mz
;
1399 struct mem_cgroup
*memcg
;
1400 struct page_cgroup
*pc
;
1401 struct lruvec
*lruvec
;
1403 if (mem_cgroup_disabled()) {
1404 lruvec
= &zone
->lruvec
;
1408 pc
= lookup_page_cgroup(page
);
1409 memcg
= pc
->mem_cgroup
;
1412 * Surreptitiously switch any uncharged offlist page to root:
1413 * an uncharged page off lru does nothing to secure
1414 * its former mem_cgroup from sudden removal.
1416 * Our caller holds lru_lock, and PageCgroupUsed is updated
1417 * under page_cgroup lock: between them, they make all uses
1418 * of pc->mem_cgroup safe.
1420 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1421 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1423 mz
= page_cgroup_zoneinfo(memcg
, page
);
1424 lruvec
= &mz
->lruvec
;
1427 * Since a node can be onlined after the mem_cgroup was created,
1428 * we have to be prepared to initialize lruvec->zone here;
1429 * and if offlined then reonlined, we need to reinitialize it.
1431 if (unlikely(lruvec
->zone
!= zone
))
1432 lruvec
->zone
= zone
;
1437 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1438 * @lruvec: mem_cgroup per zone lru vector
1439 * @lru: index of lru list the page is sitting on
1440 * @nr_pages: positive when adding or negative when removing
1442 * This function must be called when a page is added to or removed from an
1445 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1448 struct mem_cgroup_per_zone
*mz
;
1449 unsigned long *lru_size
;
1451 if (mem_cgroup_disabled())
1454 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1455 lru_size
= mz
->lru_size
+ lru
;
1456 *lru_size
+= nr_pages
;
1457 VM_BUG_ON((long)(*lru_size
) < 0);
1461 * Checks whether given mem is same or in the root_mem_cgroup's
1464 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1465 struct mem_cgroup
*memcg
)
1467 if (root_memcg
== memcg
)
1469 if (!root_memcg
->use_hierarchy
|| !memcg
)
1471 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1474 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1475 struct mem_cgroup
*memcg
)
1480 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1485 bool task_in_mem_cgroup(struct task_struct
*task
,
1486 const struct mem_cgroup
*memcg
)
1488 struct mem_cgroup
*curr
= NULL
;
1489 struct task_struct
*p
;
1492 p
= find_lock_task_mm(task
);
1494 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1498 * All threads may have already detached their mm's, but the oom
1499 * killer still needs to detect if they have already been oom
1500 * killed to prevent needlessly killing additional tasks.
1503 curr
= mem_cgroup_from_task(task
);
1505 css_get(&curr
->css
);
1511 * We should check use_hierarchy of "memcg" not "curr". Because checking
1512 * use_hierarchy of "curr" here make this function true if hierarchy is
1513 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1514 * hierarchy(even if use_hierarchy is disabled in "memcg").
1516 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1517 css_put(&curr
->css
);
1521 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1523 unsigned long inactive_ratio
;
1524 unsigned long inactive
;
1525 unsigned long active
;
1528 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1529 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1531 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1533 inactive_ratio
= int_sqrt(10 * gb
);
1537 return inactive
* inactive_ratio
< active
;
1540 #define mem_cgroup_from_res_counter(counter, member) \
1541 container_of(counter, struct mem_cgroup, member)
1544 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1545 * @memcg: the memory cgroup
1547 * Returns the maximum amount of memory @mem can be charged with, in
1550 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1552 unsigned long long margin
;
1554 margin
= res_counter_margin(&memcg
->res
);
1555 if (do_swap_account
)
1556 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1557 return margin
>> PAGE_SHIFT
;
1560 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1563 if (!css_parent(&memcg
->css
))
1564 return vm_swappiness
;
1566 return memcg
->swappiness
;
1570 * memcg->moving_account is used for checking possibility that some thread is
1571 * calling move_account(). When a thread on CPU-A starts moving pages under
1572 * a memcg, other threads should check memcg->moving_account under
1573 * rcu_read_lock(), like this:
1577 * memcg->moving_account+1 if (memcg->mocing_account)
1579 * synchronize_rcu() update something.
1584 /* for quick checking without looking up memcg */
1585 atomic_t memcg_moving __read_mostly
;
1587 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1589 atomic_inc(&memcg_moving
);
1590 atomic_inc(&memcg
->moving_account
);
1594 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1597 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1598 * We check NULL in callee rather than caller.
1601 atomic_dec(&memcg_moving
);
1602 atomic_dec(&memcg
->moving_account
);
1607 * 2 routines for checking "mem" is under move_account() or not.
1609 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1610 * is used for avoiding races in accounting. If true,
1611 * pc->mem_cgroup may be overwritten.
1613 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1614 * under hierarchy of moving cgroups. This is for
1615 * waiting at hith-memory prressure caused by "move".
1618 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1620 VM_BUG_ON(!rcu_read_lock_held());
1621 return atomic_read(&memcg
->moving_account
) > 0;
1624 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1626 struct mem_cgroup
*from
;
1627 struct mem_cgroup
*to
;
1630 * Unlike task_move routines, we access mc.to, mc.from not under
1631 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1633 spin_lock(&mc
.lock
);
1639 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1640 || mem_cgroup_same_or_subtree(memcg
, to
);
1642 spin_unlock(&mc
.lock
);
1646 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1648 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1649 if (mem_cgroup_under_move(memcg
)) {
1651 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1652 /* moving charge context might have finished. */
1655 finish_wait(&mc
.waitq
, &wait
);
1663 * Take this lock when
1664 * - a code tries to modify page's memcg while it's USED.
1665 * - a code tries to modify page state accounting in a memcg.
1666 * see mem_cgroup_stolen(), too.
1668 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1669 unsigned long *flags
)
1671 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1674 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1675 unsigned long *flags
)
1677 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1680 #define K(x) ((x) << (PAGE_SHIFT-10))
1682 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1683 * @memcg: The memory cgroup that went over limit
1684 * @p: Task that is going to be killed
1686 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1689 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1692 * protects memcg_name and makes sure that parallel ooms do not
1695 static DEFINE_SPINLOCK(oom_info_lock
);
1696 struct cgroup
*task_cgrp
;
1697 struct cgroup
*mem_cgrp
;
1698 static char memcg_name
[PATH_MAX
];
1700 struct mem_cgroup
*iter
;
1706 spin_lock(&oom_info_lock
);
1709 mem_cgrp
= memcg
->css
.cgroup
;
1710 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1712 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1715 * Unfortunately, we are unable to convert to a useful name
1716 * But we'll still print out the usage information
1723 pr_info("Task in %s killed", memcg_name
);
1726 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1734 * Continues from above, so we don't need an KERN_ level
1736 pr_cont(" as a result of limit of %s\n", memcg_name
);
1739 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1740 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1741 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1742 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1743 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1744 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1745 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1746 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1747 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1748 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1749 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1750 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1752 for_each_mem_cgroup_tree(iter
, memcg
) {
1753 pr_info("Memory cgroup stats");
1756 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1758 pr_cont(" for %s", memcg_name
);
1762 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1763 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1765 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1766 K(mem_cgroup_read_stat(iter
, i
)));
1769 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1770 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1771 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1775 spin_unlock(&oom_info_lock
);
1779 * This function returns the number of memcg under hierarchy tree. Returns
1780 * 1(self count) if no children.
1782 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1785 struct mem_cgroup
*iter
;
1787 for_each_mem_cgroup_tree(iter
, memcg
)
1793 * Return the memory (and swap, if configured) limit for a memcg.
1795 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1799 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1802 * Do not consider swap space if we cannot swap due to swappiness
1804 if (mem_cgroup_swappiness(memcg
)) {
1807 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1808 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1811 * If memsw is finite and limits the amount of swap space
1812 * available to this memcg, return that limit.
1814 limit
= min(limit
, memsw
);
1820 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1823 struct mem_cgroup
*iter
;
1824 unsigned long chosen_points
= 0;
1825 unsigned long totalpages
;
1826 unsigned int points
= 0;
1827 struct task_struct
*chosen
= NULL
;
1830 * If current has a pending SIGKILL or is exiting, then automatically
1831 * select it. The goal is to allow it to allocate so that it may
1832 * quickly exit and free its memory.
1834 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1835 set_thread_flag(TIF_MEMDIE
);
1839 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1840 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1841 for_each_mem_cgroup_tree(iter
, memcg
) {
1842 struct css_task_iter it
;
1843 struct task_struct
*task
;
1845 css_task_iter_start(&iter
->css
, &it
);
1846 while ((task
= css_task_iter_next(&it
))) {
1847 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1849 case OOM_SCAN_SELECT
:
1851 put_task_struct(chosen
);
1853 chosen_points
= ULONG_MAX
;
1854 get_task_struct(chosen
);
1856 case OOM_SCAN_CONTINUE
:
1858 case OOM_SCAN_ABORT
:
1859 css_task_iter_end(&it
);
1860 mem_cgroup_iter_break(memcg
, iter
);
1862 put_task_struct(chosen
);
1867 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1868 if (points
> chosen_points
) {
1870 put_task_struct(chosen
);
1872 chosen_points
= points
;
1873 get_task_struct(chosen
);
1876 css_task_iter_end(&it
);
1881 points
= chosen_points
* 1000 / totalpages
;
1882 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1883 NULL
, "Memory cgroup out of memory");
1886 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1888 unsigned long flags
)
1890 unsigned long total
= 0;
1891 bool noswap
= false;
1894 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1896 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1899 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1901 drain_all_stock_async(memcg
);
1902 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1904 * Allow limit shrinkers, which are triggered directly
1905 * by userspace, to catch signals and stop reclaim
1906 * after minimal progress, regardless of the margin.
1908 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1910 if (mem_cgroup_margin(memcg
))
1913 * If nothing was reclaimed after two attempts, there
1914 * may be no reclaimable pages in this hierarchy.
1923 * test_mem_cgroup_node_reclaimable
1924 * @memcg: the target memcg
1925 * @nid: the node ID to be checked.
1926 * @noswap : specify true here if the user wants flle only information.
1928 * This function returns whether the specified memcg contains any
1929 * reclaimable pages on a node. Returns true if there are any reclaimable
1930 * pages in the node.
1932 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1933 int nid
, bool noswap
)
1935 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1937 if (noswap
|| !total_swap_pages
)
1939 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1944 #if MAX_NUMNODES > 1
1947 * Always updating the nodemask is not very good - even if we have an empty
1948 * list or the wrong list here, we can start from some node and traverse all
1949 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1952 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1956 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1957 * pagein/pageout changes since the last update.
1959 if (!atomic_read(&memcg
->numainfo_events
))
1961 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1964 /* make a nodemask where this memcg uses memory from */
1965 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1967 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1969 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1970 node_clear(nid
, memcg
->scan_nodes
);
1973 atomic_set(&memcg
->numainfo_events
, 0);
1974 atomic_set(&memcg
->numainfo_updating
, 0);
1978 * Selecting a node where we start reclaim from. Because what we need is just
1979 * reducing usage counter, start from anywhere is O,K. Considering
1980 * memory reclaim from current node, there are pros. and cons.
1982 * Freeing memory from current node means freeing memory from a node which
1983 * we'll use or we've used. So, it may make LRU bad. And if several threads
1984 * hit limits, it will see a contention on a node. But freeing from remote
1985 * node means more costs for memory reclaim because of memory latency.
1987 * Now, we use round-robin. Better algorithm is welcomed.
1989 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1993 mem_cgroup_may_update_nodemask(memcg
);
1994 node
= memcg
->last_scanned_node
;
1996 node
= next_node(node
, memcg
->scan_nodes
);
1997 if (node
== MAX_NUMNODES
)
1998 node
= first_node(memcg
->scan_nodes
);
2000 * We call this when we hit limit, not when pages are added to LRU.
2001 * No LRU may hold pages because all pages are UNEVICTABLE or
2002 * memcg is too small and all pages are not on LRU. In that case,
2003 * we use curret node.
2005 if (unlikely(node
== MAX_NUMNODES
))
2006 node
= numa_node_id();
2008 memcg
->last_scanned_node
= node
;
2013 * Check all nodes whether it contains reclaimable pages or not.
2014 * For quick scan, we make use of scan_nodes. This will allow us to skip
2015 * unused nodes. But scan_nodes is lazily updated and may not cotain
2016 * enough new information. We need to do double check.
2018 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2023 * quick check...making use of scan_node.
2024 * We can skip unused nodes.
2026 if (!nodes_empty(memcg
->scan_nodes
)) {
2027 for (nid
= first_node(memcg
->scan_nodes
);
2029 nid
= next_node(nid
, memcg
->scan_nodes
)) {
2031 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2036 * Check rest of nodes.
2038 for_each_node_state(nid
, N_MEMORY
) {
2039 if (node_isset(nid
, memcg
->scan_nodes
))
2041 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2048 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2053 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2055 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2059 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2062 unsigned long *total_scanned
)
2064 struct mem_cgroup
*victim
= NULL
;
2067 unsigned long excess
;
2068 unsigned long nr_scanned
;
2069 struct mem_cgroup_reclaim_cookie reclaim
= {
2074 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2077 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2082 * If we have not been able to reclaim
2083 * anything, it might because there are
2084 * no reclaimable pages under this hierarchy
2089 * We want to do more targeted reclaim.
2090 * excess >> 2 is not to excessive so as to
2091 * reclaim too much, nor too less that we keep
2092 * coming back to reclaim from this cgroup
2094 if (total
>= (excess
>> 2) ||
2095 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2100 if (!mem_cgroup_reclaimable(victim
, false))
2102 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2104 *total_scanned
+= nr_scanned
;
2105 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2108 mem_cgroup_iter_break(root_memcg
, victim
);
2112 #ifdef CONFIG_LOCKDEP
2113 static struct lockdep_map memcg_oom_lock_dep_map
= {
2114 .name
= "memcg_oom_lock",
2118 static DEFINE_SPINLOCK(memcg_oom_lock
);
2121 * Check OOM-Killer is already running under our hierarchy.
2122 * If someone is running, return false.
2124 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2126 struct mem_cgroup
*iter
, *failed
= NULL
;
2128 spin_lock(&memcg_oom_lock
);
2130 for_each_mem_cgroup_tree(iter
, memcg
) {
2131 if (iter
->oom_lock
) {
2133 * this subtree of our hierarchy is already locked
2134 * so we cannot give a lock.
2137 mem_cgroup_iter_break(memcg
, iter
);
2140 iter
->oom_lock
= true;
2145 * OK, we failed to lock the whole subtree so we have
2146 * to clean up what we set up to the failing subtree
2148 for_each_mem_cgroup_tree(iter
, memcg
) {
2149 if (iter
== failed
) {
2150 mem_cgroup_iter_break(memcg
, iter
);
2153 iter
->oom_lock
= false;
2156 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2158 spin_unlock(&memcg_oom_lock
);
2163 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2165 struct mem_cgroup
*iter
;
2167 spin_lock(&memcg_oom_lock
);
2168 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2169 for_each_mem_cgroup_tree(iter
, memcg
)
2170 iter
->oom_lock
= false;
2171 spin_unlock(&memcg_oom_lock
);
2174 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2176 struct mem_cgroup
*iter
;
2178 for_each_mem_cgroup_tree(iter
, memcg
)
2179 atomic_inc(&iter
->under_oom
);
2182 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2184 struct mem_cgroup
*iter
;
2187 * When a new child is created while the hierarchy is under oom,
2188 * mem_cgroup_oom_lock() may not be called. We have to use
2189 * atomic_add_unless() here.
2191 for_each_mem_cgroup_tree(iter
, memcg
)
2192 atomic_add_unless(&iter
->under_oom
, -1, 0);
2195 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2197 struct oom_wait_info
{
2198 struct mem_cgroup
*memcg
;
2202 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2203 unsigned mode
, int sync
, void *arg
)
2205 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2206 struct mem_cgroup
*oom_wait_memcg
;
2207 struct oom_wait_info
*oom_wait_info
;
2209 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2210 oom_wait_memcg
= oom_wait_info
->memcg
;
2213 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2214 * Then we can use css_is_ancestor without taking care of RCU.
2216 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2217 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2219 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2222 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2224 atomic_inc(&memcg
->oom_wakeups
);
2225 /* for filtering, pass "memcg" as argument. */
2226 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2229 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2231 if (memcg
&& atomic_read(&memcg
->under_oom
))
2232 memcg_wakeup_oom(memcg
);
2235 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2237 if (!current
->memcg_oom
.may_oom
)
2240 * We are in the middle of the charge context here, so we
2241 * don't want to block when potentially sitting on a callstack
2242 * that holds all kinds of filesystem and mm locks.
2244 * Also, the caller may handle a failed allocation gracefully
2245 * (like optional page cache readahead) and so an OOM killer
2246 * invocation might not even be necessary.
2248 * That's why we don't do anything here except remember the
2249 * OOM context and then deal with it at the end of the page
2250 * fault when the stack is unwound, the locks are released,
2251 * and when we know whether the fault was overall successful.
2253 css_get(&memcg
->css
);
2254 current
->memcg_oom
.memcg
= memcg
;
2255 current
->memcg_oom
.gfp_mask
= mask
;
2256 current
->memcg_oom
.order
= order
;
2260 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2261 * @handle: actually kill/wait or just clean up the OOM state
2263 * This has to be called at the end of a page fault if the memcg OOM
2264 * handler was enabled.
2266 * Memcg supports userspace OOM handling where failed allocations must
2267 * sleep on a waitqueue until the userspace task resolves the
2268 * situation. Sleeping directly in the charge context with all kinds
2269 * of locks held is not a good idea, instead we remember an OOM state
2270 * in the task and mem_cgroup_oom_synchronize() has to be called at
2271 * the end of the page fault to complete the OOM handling.
2273 * Returns %true if an ongoing memcg OOM situation was detected and
2274 * completed, %false otherwise.
2276 bool mem_cgroup_oom_synchronize(bool handle
)
2278 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2279 struct oom_wait_info owait
;
2282 /* OOM is global, do not handle */
2289 owait
.memcg
= memcg
;
2290 owait
.wait
.flags
= 0;
2291 owait
.wait
.func
= memcg_oom_wake_function
;
2292 owait
.wait
.private = current
;
2293 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2295 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2296 mem_cgroup_mark_under_oom(memcg
);
2298 locked
= mem_cgroup_oom_trylock(memcg
);
2301 mem_cgroup_oom_notify(memcg
);
2303 if (locked
&& !memcg
->oom_kill_disable
) {
2304 mem_cgroup_unmark_under_oom(memcg
);
2305 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2306 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2307 current
->memcg_oom
.order
);
2310 mem_cgroup_unmark_under_oom(memcg
);
2311 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2315 mem_cgroup_oom_unlock(memcg
);
2317 * There is no guarantee that an OOM-lock contender
2318 * sees the wakeups triggered by the OOM kill
2319 * uncharges. Wake any sleepers explicitely.
2321 memcg_oom_recover(memcg
);
2324 current
->memcg_oom
.memcg
= NULL
;
2325 css_put(&memcg
->css
);
2330 * Currently used to update mapped file statistics, but the routine can be
2331 * generalized to update other statistics as well.
2333 * Notes: Race condition
2335 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2336 * it tends to be costly. But considering some conditions, we doesn't need
2337 * to do so _always_.
2339 * Considering "charge", lock_page_cgroup() is not required because all
2340 * file-stat operations happen after a page is attached to radix-tree. There
2341 * are no race with "charge".
2343 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2344 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2345 * if there are race with "uncharge". Statistics itself is properly handled
2348 * Considering "move", this is an only case we see a race. To make the race
2349 * small, we check mm->moving_account and detect there are possibility of race
2350 * If there is, we take a lock.
2353 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2354 bool *locked
, unsigned long *flags
)
2356 struct mem_cgroup
*memcg
;
2357 struct page_cgroup
*pc
;
2359 pc
= lookup_page_cgroup(page
);
2361 memcg
= pc
->mem_cgroup
;
2362 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2365 * If this memory cgroup is not under account moving, we don't
2366 * need to take move_lock_mem_cgroup(). Because we already hold
2367 * rcu_read_lock(), any calls to move_account will be delayed until
2368 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2370 if (!mem_cgroup_stolen(memcg
))
2373 move_lock_mem_cgroup(memcg
, flags
);
2374 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2375 move_unlock_mem_cgroup(memcg
, flags
);
2381 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2383 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2386 * It's guaranteed that pc->mem_cgroup never changes while
2387 * lock is held because a routine modifies pc->mem_cgroup
2388 * should take move_lock_mem_cgroup().
2390 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2393 void mem_cgroup_update_page_stat(struct page
*page
,
2394 enum mem_cgroup_stat_index idx
, int val
)
2396 struct mem_cgroup
*memcg
;
2397 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2398 unsigned long uninitialized_var(flags
);
2400 if (mem_cgroup_disabled())
2403 VM_BUG_ON(!rcu_read_lock_held());
2404 memcg
= pc
->mem_cgroup
;
2405 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2408 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2412 * size of first charge trial. "32" comes from vmscan.c's magic value.
2413 * TODO: maybe necessary to use big numbers in big irons.
2415 #define CHARGE_BATCH 32U
2416 struct memcg_stock_pcp
{
2417 struct mem_cgroup
*cached
; /* this never be root cgroup */
2418 unsigned int nr_pages
;
2419 struct work_struct work
;
2420 unsigned long flags
;
2421 #define FLUSHING_CACHED_CHARGE 0
2423 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2424 static DEFINE_MUTEX(percpu_charge_mutex
);
2427 * consume_stock: Try to consume stocked charge on this cpu.
2428 * @memcg: memcg to consume from.
2429 * @nr_pages: how many pages to charge.
2431 * The charges will only happen if @memcg matches the current cpu's memcg
2432 * stock, and at least @nr_pages are available in that stock. Failure to
2433 * service an allocation will refill the stock.
2435 * returns true if successful, false otherwise.
2437 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2439 struct memcg_stock_pcp
*stock
;
2442 if (nr_pages
> CHARGE_BATCH
)
2445 stock
= &get_cpu_var(memcg_stock
);
2446 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2447 stock
->nr_pages
-= nr_pages
;
2448 else /* need to call res_counter_charge */
2450 put_cpu_var(memcg_stock
);
2455 * Returns stocks cached in percpu to res_counter and reset cached information.
2457 static void drain_stock(struct memcg_stock_pcp
*stock
)
2459 struct mem_cgroup
*old
= stock
->cached
;
2461 if (stock
->nr_pages
) {
2462 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2464 res_counter_uncharge(&old
->res
, bytes
);
2465 if (do_swap_account
)
2466 res_counter_uncharge(&old
->memsw
, bytes
);
2467 stock
->nr_pages
= 0;
2469 stock
->cached
= NULL
;
2473 * This must be called under preempt disabled or must be called by
2474 * a thread which is pinned to local cpu.
2476 static void drain_local_stock(struct work_struct
*dummy
)
2478 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2480 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2483 static void __init
memcg_stock_init(void)
2487 for_each_possible_cpu(cpu
) {
2488 struct memcg_stock_pcp
*stock
=
2489 &per_cpu(memcg_stock
, cpu
);
2490 INIT_WORK(&stock
->work
, drain_local_stock
);
2495 * Cache charges(val) which is from res_counter, to local per_cpu area.
2496 * This will be consumed by consume_stock() function, later.
2498 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2500 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2502 if (stock
->cached
!= memcg
) { /* reset if necessary */
2504 stock
->cached
= memcg
;
2506 stock
->nr_pages
+= nr_pages
;
2507 put_cpu_var(memcg_stock
);
2511 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2512 * of the hierarchy under it. sync flag says whether we should block
2513 * until the work is done.
2515 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2519 /* Notify other cpus that system-wide "drain" is running */
2522 for_each_online_cpu(cpu
) {
2523 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2524 struct mem_cgroup
*memcg
;
2526 memcg
= stock
->cached
;
2527 if (!memcg
|| !stock
->nr_pages
)
2529 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2531 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2533 drain_local_stock(&stock
->work
);
2535 schedule_work_on(cpu
, &stock
->work
);
2543 for_each_online_cpu(cpu
) {
2544 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2545 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2546 flush_work(&stock
->work
);
2553 * Tries to drain stocked charges in other cpus. This function is asynchronous
2554 * and just put a work per cpu for draining localy on each cpu. Caller can
2555 * expects some charges will be back to res_counter later but cannot wait for
2558 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2561 * If someone calls draining, avoid adding more kworker runs.
2563 if (!mutex_trylock(&percpu_charge_mutex
))
2565 drain_all_stock(root_memcg
, false);
2566 mutex_unlock(&percpu_charge_mutex
);
2569 /* This is a synchronous drain interface. */
2570 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2572 /* called when force_empty is called */
2573 mutex_lock(&percpu_charge_mutex
);
2574 drain_all_stock(root_memcg
, true);
2575 mutex_unlock(&percpu_charge_mutex
);
2579 * This function drains percpu counter value from DEAD cpu and
2580 * move it to local cpu. Note that this function can be preempted.
2582 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2586 spin_lock(&memcg
->pcp_counter_lock
);
2587 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2588 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2590 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2591 memcg
->nocpu_base
.count
[i
] += x
;
2593 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2594 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2596 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2597 memcg
->nocpu_base
.events
[i
] += x
;
2599 spin_unlock(&memcg
->pcp_counter_lock
);
2602 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2603 unsigned long action
,
2606 int cpu
= (unsigned long)hcpu
;
2607 struct memcg_stock_pcp
*stock
;
2608 struct mem_cgroup
*iter
;
2610 if (action
== CPU_ONLINE
)
2613 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2616 for_each_mem_cgroup(iter
)
2617 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2619 stock
= &per_cpu(memcg_stock
, cpu
);
2625 /* See __mem_cgroup_try_charge() for details */
2627 CHARGE_OK
, /* success */
2628 CHARGE_RETRY
, /* need to retry but retry is not bad */
2629 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2630 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2633 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2634 unsigned int nr_pages
, unsigned int min_pages
,
2637 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2638 struct mem_cgroup
*mem_over_limit
;
2639 struct res_counter
*fail_res
;
2640 unsigned long flags
= 0;
2643 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2646 if (!do_swap_account
)
2648 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2652 res_counter_uncharge(&memcg
->res
, csize
);
2653 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2654 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2656 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2658 * Never reclaim on behalf of optional batching, retry with a
2659 * single page instead.
2661 if (nr_pages
> min_pages
)
2662 return CHARGE_RETRY
;
2664 if (!(gfp_mask
& __GFP_WAIT
))
2665 return CHARGE_WOULDBLOCK
;
2667 if (gfp_mask
& __GFP_NORETRY
)
2668 return CHARGE_NOMEM
;
2670 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2671 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2672 return CHARGE_RETRY
;
2674 * Even though the limit is exceeded at this point, reclaim
2675 * may have been able to free some pages. Retry the charge
2676 * before killing the task.
2678 * Only for regular pages, though: huge pages are rather
2679 * unlikely to succeed so close to the limit, and we fall back
2680 * to regular pages anyway in case of failure.
2682 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2683 return CHARGE_RETRY
;
2686 * At task move, charge accounts can be doubly counted. So, it's
2687 * better to wait until the end of task_move if something is going on.
2689 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2690 return CHARGE_RETRY
;
2693 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2695 return CHARGE_NOMEM
;
2699 * __mem_cgroup_try_charge() does
2700 * 1. detect memcg to be charged against from passed *mm and *ptr,
2701 * 2. update res_counter
2702 * 3. call memory reclaim if necessary.
2704 * In some special case, if the task is fatal, fatal_signal_pending() or
2705 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2706 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2707 * as possible without any hazards. 2: all pages should have a valid
2708 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2709 * pointer, that is treated as a charge to root_mem_cgroup.
2711 * So __mem_cgroup_try_charge() will return
2712 * 0 ... on success, filling *ptr with a valid memcg pointer.
2713 * -ENOMEM ... charge failure because of resource limits.
2714 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2716 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2717 * the oom-killer can be invoked.
2719 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2721 unsigned int nr_pages
,
2722 struct mem_cgroup
**ptr
,
2725 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2726 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2727 struct mem_cgroup
*memcg
= NULL
;
2731 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2732 * in system level. So, allow to go ahead dying process in addition to
2735 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2736 || fatal_signal_pending(current
)))
2739 if (unlikely(task_in_memcg_oom(current
)))
2742 if (gfp_mask
& __GFP_NOFAIL
)
2746 * We always charge the cgroup the mm_struct belongs to.
2747 * The mm_struct's mem_cgroup changes on task migration if the
2748 * thread group leader migrates. It's possible that mm is not
2749 * set, if so charge the root memcg (happens for pagecache usage).
2752 *ptr
= root_mem_cgroup
;
2754 if (*ptr
) { /* css should be a valid one */
2756 if (mem_cgroup_is_root(memcg
))
2758 if (consume_stock(memcg
, nr_pages
))
2760 css_get(&memcg
->css
);
2762 struct task_struct
*p
;
2765 p
= rcu_dereference(mm
->owner
);
2767 * Because we don't have task_lock(), "p" can exit.
2768 * In that case, "memcg" can point to root or p can be NULL with
2769 * race with swapoff. Then, we have small risk of mis-accouning.
2770 * But such kind of mis-account by race always happens because
2771 * we don't have cgroup_mutex(). It's overkill and we allo that
2773 * (*) swapoff at el will charge against mm-struct not against
2774 * task-struct. So, mm->owner can be NULL.
2776 memcg
= mem_cgroup_from_task(p
);
2778 memcg
= root_mem_cgroup
;
2779 if (mem_cgroup_is_root(memcg
)) {
2783 if (consume_stock(memcg
, nr_pages
)) {
2785 * It seems dagerous to access memcg without css_get().
2786 * But considering how consume_stok works, it's not
2787 * necessary. If consume_stock success, some charges
2788 * from this memcg are cached on this cpu. So, we
2789 * don't need to call css_get()/css_tryget() before
2790 * calling consume_stock().
2795 /* after here, we may be blocked. we need to get refcnt */
2796 if (!css_tryget(&memcg
->css
)) {
2804 bool invoke_oom
= oom
&& !nr_oom_retries
;
2806 /* If killed, bypass charge */
2807 if (fatal_signal_pending(current
)) {
2808 css_put(&memcg
->css
);
2812 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2813 nr_pages
, invoke_oom
);
2817 case CHARGE_RETRY
: /* not in OOM situation but retry */
2819 css_put(&memcg
->css
);
2822 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2823 css_put(&memcg
->css
);
2825 case CHARGE_NOMEM
: /* OOM routine works */
2826 if (!oom
|| invoke_oom
) {
2827 css_put(&memcg
->css
);
2833 } while (ret
!= CHARGE_OK
);
2835 if (batch
> nr_pages
)
2836 refill_stock(memcg
, batch
- nr_pages
);
2837 css_put(&memcg
->css
);
2842 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2847 *ptr
= root_mem_cgroup
;
2852 * Somemtimes we have to undo a charge we got by try_charge().
2853 * This function is for that and do uncharge, put css's refcnt.
2854 * gotten by try_charge().
2856 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2857 unsigned int nr_pages
)
2859 if (!mem_cgroup_is_root(memcg
)) {
2860 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2862 res_counter_uncharge(&memcg
->res
, bytes
);
2863 if (do_swap_account
)
2864 res_counter_uncharge(&memcg
->memsw
, bytes
);
2869 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2870 * This is useful when moving usage to parent cgroup.
2872 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2873 unsigned int nr_pages
)
2875 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2877 if (mem_cgroup_is_root(memcg
))
2880 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2881 if (do_swap_account
)
2882 res_counter_uncharge_until(&memcg
->memsw
,
2883 memcg
->memsw
.parent
, bytes
);
2887 * A helper function to get mem_cgroup from ID. must be called under
2888 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2889 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2890 * called against removed memcg.)
2892 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2894 /* ID 0 is unused ID */
2897 return mem_cgroup_from_id(id
);
2900 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2902 struct mem_cgroup
*memcg
= NULL
;
2903 struct page_cgroup
*pc
;
2907 VM_BUG_ON(!PageLocked(page
));
2909 pc
= lookup_page_cgroup(page
);
2910 lock_page_cgroup(pc
);
2911 if (PageCgroupUsed(pc
)) {
2912 memcg
= pc
->mem_cgroup
;
2913 if (memcg
&& !css_tryget(&memcg
->css
))
2915 } else if (PageSwapCache(page
)) {
2916 ent
.val
= page_private(page
);
2917 id
= lookup_swap_cgroup_id(ent
);
2919 memcg
= mem_cgroup_lookup(id
);
2920 if (memcg
&& !css_tryget(&memcg
->css
))
2924 unlock_page_cgroup(pc
);
2928 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2930 unsigned int nr_pages
,
2931 enum charge_type ctype
,
2934 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2935 struct zone
*uninitialized_var(zone
);
2936 struct lruvec
*lruvec
;
2937 bool was_on_lru
= false;
2940 lock_page_cgroup(pc
);
2941 VM_BUG_ON(PageCgroupUsed(pc
));
2943 * we don't need page_cgroup_lock about tail pages, becase they are not
2944 * accessed by any other context at this point.
2948 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2949 * may already be on some other mem_cgroup's LRU. Take care of it.
2952 zone
= page_zone(page
);
2953 spin_lock_irq(&zone
->lru_lock
);
2954 if (PageLRU(page
)) {
2955 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2957 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2962 pc
->mem_cgroup
= memcg
;
2964 * We access a page_cgroup asynchronously without lock_page_cgroup().
2965 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2966 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2967 * before USED bit, we need memory barrier here.
2968 * See mem_cgroup_add_lru_list(), etc.
2971 SetPageCgroupUsed(pc
);
2975 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2976 VM_BUG_ON(PageLRU(page
));
2978 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2980 spin_unlock_irq(&zone
->lru_lock
);
2983 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2988 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2989 unlock_page_cgroup(pc
);
2992 * "charge_statistics" updated event counter. Then, check it.
2993 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2994 * if they exceeds softlimit.
2996 memcg_check_events(memcg
, page
);
2999 static DEFINE_MUTEX(set_limit_mutex
);
3001 #ifdef CONFIG_MEMCG_KMEM
3002 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
3004 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
3005 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
) ==
3006 KMEM_ACCOUNTED_MASK
;
3010 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
3011 * in the memcg_cache_params struct.
3013 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
3015 struct kmem_cache
*cachep
;
3017 VM_BUG_ON(p
->is_root_cache
);
3018 cachep
= p
->root_cache
;
3019 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
3022 #ifdef CONFIG_SLABINFO
3023 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
3025 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3026 struct memcg_cache_params
*params
;
3028 if (!memcg_can_account_kmem(memcg
))
3031 print_slabinfo_header(m
);
3033 mutex_lock(&memcg
->slab_caches_mutex
);
3034 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
3035 cache_show(memcg_params_to_cache(params
), m
);
3036 mutex_unlock(&memcg
->slab_caches_mutex
);
3042 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
3044 struct res_counter
*fail_res
;
3045 struct mem_cgroup
*_memcg
;
3048 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
3053 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
3054 &_memcg
, oom_gfp_allowed(gfp
));
3056 if (ret
== -EINTR
) {
3058 * __mem_cgroup_try_charge() chosed to bypass to root due to
3059 * OOM kill or fatal signal. Since our only options are to
3060 * either fail the allocation or charge it to this cgroup, do
3061 * it as a temporary condition. But we can't fail. From a
3062 * kmem/slab perspective, the cache has already been selected,
3063 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3066 * This condition will only trigger if the task entered
3067 * memcg_charge_kmem in a sane state, but was OOM-killed during
3068 * __mem_cgroup_try_charge() above. Tasks that were already
3069 * dying when the allocation triggers should have been already
3070 * directed to the root cgroup in memcontrol.h
3072 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3073 if (do_swap_account
)
3074 res_counter_charge_nofail(&memcg
->memsw
, size
,
3078 res_counter_uncharge(&memcg
->kmem
, size
);
3083 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3085 res_counter_uncharge(&memcg
->res
, size
);
3086 if (do_swap_account
)
3087 res_counter_uncharge(&memcg
->memsw
, size
);
3090 if (res_counter_uncharge(&memcg
->kmem
, size
))
3094 * Releases a reference taken in kmem_cgroup_css_offline in case
3095 * this last uncharge is racing with the offlining code or it is
3096 * outliving the memcg existence.
3098 * The memory barrier imposed by test&clear is paired with the
3099 * explicit one in memcg_kmem_mark_dead().
3101 if (memcg_kmem_test_and_clear_dead(memcg
))
3102 css_put(&memcg
->css
);
3105 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3110 mutex_lock(&memcg
->slab_caches_mutex
);
3111 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3112 mutex_unlock(&memcg
->slab_caches_mutex
);
3116 * helper for acessing a memcg's index. It will be used as an index in the
3117 * child cache array in kmem_cache, and also to derive its name. This function
3118 * will return -1 when this is not a kmem-limited memcg.
3120 int memcg_cache_id(struct mem_cgroup
*memcg
)
3122 return memcg
? memcg
->kmemcg_id
: -1;
3126 * This ends up being protected by the set_limit mutex, during normal
3127 * operation, because that is its main call site.
3129 * But when we create a new cache, we can call this as well if its parent
3130 * is kmem-limited. That will have to hold set_limit_mutex as well.
3132 static int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3136 num
= ida_simple_get(&kmem_limited_groups
,
3137 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3141 * After this point, kmem_accounted (that we test atomically in
3142 * the beginning of this conditional), is no longer 0. This
3143 * guarantees only one process will set the following boolean
3144 * to true. We don't need test_and_set because we're protected
3145 * by the set_limit_mutex anyway.
3147 memcg_kmem_set_activated(memcg
);
3149 ret
= memcg_update_all_caches(num
+1);
3151 ida_simple_remove(&kmem_limited_groups
, num
);
3152 memcg_kmem_clear_activated(memcg
);
3156 memcg
->kmemcg_id
= num
;
3157 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3158 mutex_init(&memcg
->slab_caches_mutex
);
3162 static size_t memcg_caches_array_size(int num_groups
)
3165 if (num_groups
<= 0)
3168 size
= 2 * num_groups
;
3169 if (size
< MEMCG_CACHES_MIN_SIZE
)
3170 size
= MEMCG_CACHES_MIN_SIZE
;
3171 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3172 size
= MEMCG_CACHES_MAX_SIZE
;
3178 * We should update the current array size iff all caches updates succeed. This
3179 * can only be done from the slab side. The slab mutex needs to be held when
3182 void memcg_update_array_size(int num
)
3184 if (num
> memcg_limited_groups_array_size
)
3185 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3188 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3190 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3192 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3194 VM_BUG_ON(!is_root_cache(s
));
3196 if (num_groups
> memcg_limited_groups_array_size
) {
3198 ssize_t size
= memcg_caches_array_size(num_groups
);
3200 size
*= sizeof(void *);
3201 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3203 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3204 if (!s
->memcg_params
) {
3205 s
->memcg_params
= cur_params
;
3209 s
->memcg_params
->is_root_cache
= true;
3212 * There is the chance it will be bigger than
3213 * memcg_limited_groups_array_size, if we failed an allocation
3214 * in a cache, in which case all caches updated before it, will
3215 * have a bigger array.
3217 * But if that is the case, the data after
3218 * memcg_limited_groups_array_size is certainly unused
3220 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3221 if (!cur_params
->memcg_caches
[i
])
3223 s
->memcg_params
->memcg_caches
[i
] =
3224 cur_params
->memcg_caches
[i
];
3228 * Ideally, we would wait until all caches succeed, and only
3229 * then free the old one. But this is not worth the extra
3230 * pointer per-cache we'd have to have for this.
3232 * It is not a big deal if some caches are left with a size
3233 * bigger than the others. And all updates will reset this
3241 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3242 struct kmem_cache
*root_cache
)
3246 if (!memcg_kmem_enabled())
3250 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3251 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3253 size
= sizeof(struct memcg_cache_params
);
3255 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3256 if (!s
->memcg_params
)
3260 s
->memcg_params
->memcg
= memcg
;
3261 s
->memcg_params
->root_cache
= root_cache
;
3262 INIT_WORK(&s
->memcg_params
->destroy
,
3263 kmem_cache_destroy_work_func
);
3265 s
->memcg_params
->is_root_cache
= true;
3270 void memcg_release_cache(struct kmem_cache
*s
)
3272 struct kmem_cache
*root
;
3273 struct mem_cgroup
*memcg
;
3277 * This happens, for instance, when a root cache goes away before we
3280 if (!s
->memcg_params
)
3283 if (s
->memcg_params
->is_root_cache
)
3286 memcg
= s
->memcg_params
->memcg
;
3287 id
= memcg_cache_id(memcg
);
3289 root
= s
->memcg_params
->root_cache
;
3290 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3292 mutex_lock(&memcg
->slab_caches_mutex
);
3293 list_del(&s
->memcg_params
->list
);
3294 mutex_unlock(&memcg
->slab_caches_mutex
);
3296 css_put(&memcg
->css
);
3298 kfree(s
->memcg_params
);
3302 * During the creation a new cache, we need to disable our accounting mechanism
3303 * altogether. This is true even if we are not creating, but rather just
3304 * enqueing new caches to be created.
3306 * This is because that process will trigger allocations; some visible, like
3307 * explicit kmallocs to auxiliary data structures, name strings and internal
3308 * cache structures; some well concealed, like INIT_WORK() that can allocate
3309 * objects during debug.
3311 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3312 * to it. This may not be a bounded recursion: since the first cache creation
3313 * failed to complete (waiting on the allocation), we'll just try to create the
3314 * cache again, failing at the same point.
3316 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3317 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3318 * inside the following two functions.
3320 static inline void memcg_stop_kmem_account(void)
3322 VM_BUG_ON(!current
->mm
);
3323 current
->memcg_kmem_skip_account
++;
3326 static inline void memcg_resume_kmem_account(void)
3328 VM_BUG_ON(!current
->mm
);
3329 current
->memcg_kmem_skip_account
--;
3332 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3334 struct kmem_cache
*cachep
;
3335 struct memcg_cache_params
*p
;
3337 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3339 cachep
= memcg_params_to_cache(p
);
3342 * If we get down to 0 after shrink, we could delete right away.
3343 * However, memcg_release_pages() already puts us back in the workqueue
3344 * in that case. If we proceed deleting, we'll get a dangling
3345 * reference, and removing the object from the workqueue in that case
3346 * is unnecessary complication. We are not a fast path.
3348 * Note that this case is fundamentally different from racing with
3349 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3350 * kmem_cache_shrink, not only we would be reinserting a dead cache
3351 * into the queue, but doing so from inside the worker racing to
3354 * So if we aren't down to zero, we'll just schedule a worker and try
3357 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3358 kmem_cache_shrink(cachep
);
3359 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3362 kmem_cache_destroy(cachep
);
3365 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3367 if (!cachep
->memcg_params
->dead
)
3371 * There are many ways in which we can get here.
3373 * We can get to a memory-pressure situation while the delayed work is
3374 * still pending to run. The vmscan shrinkers can then release all
3375 * cache memory and get us to destruction. If this is the case, we'll
3376 * be executed twice, which is a bug (the second time will execute over
3377 * bogus data). In this case, cancelling the work should be fine.
3379 * But we can also get here from the worker itself, if
3380 * kmem_cache_shrink is enough to shake all the remaining objects and
3381 * get the page count to 0. In this case, we'll deadlock if we try to
3382 * cancel the work (the worker runs with an internal lock held, which
3383 * is the same lock we would hold for cancel_work_sync().)
3385 * Since we can't possibly know who got us here, just refrain from
3386 * running if there is already work pending
3388 if (work_pending(&cachep
->memcg_params
->destroy
))
3391 * We have to defer the actual destroying to a workqueue, because
3392 * we might currently be in a context that cannot sleep.
3394 schedule_work(&cachep
->memcg_params
->destroy
);
3398 * This lock protects updaters, not readers. We want readers to be as fast as
3399 * they can, and they will either see NULL or a valid cache value. Our model
3400 * allow them to see NULL, in which case the root memcg will be selected.
3402 * We need this lock because multiple allocations to the same cache from a non
3403 * will span more than one worker. Only one of them can create the cache.
3405 static DEFINE_MUTEX(memcg_cache_mutex
);
3408 * Called with memcg_cache_mutex held
3410 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3411 struct kmem_cache
*s
)
3413 struct kmem_cache
*new;
3414 static char *tmp_name
= NULL
;
3416 lockdep_assert_held(&memcg_cache_mutex
);
3419 * kmem_cache_create_memcg duplicates the given name and
3420 * cgroup_name for this name requires RCU context.
3421 * This static temporary buffer is used to prevent from
3422 * pointless shortliving allocation.
3425 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3431 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3432 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3435 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3436 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3439 new->allocflags
|= __GFP_KMEMCG
;
3444 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3445 struct kmem_cache
*cachep
)
3447 struct kmem_cache
*new_cachep
;
3450 BUG_ON(!memcg_can_account_kmem(memcg
));
3452 idx
= memcg_cache_id(memcg
);
3454 mutex_lock(&memcg_cache_mutex
);
3455 new_cachep
= cache_from_memcg_idx(cachep
, idx
);
3457 css_put(&memcg
->css
);
3461 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3462 if (new_cachep
== NULL
) {
3463 new_cachep
= cachep
;
3464 css_put(&memcg
->css
);
3468 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3470 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3472 * the readers won't lock, make sure everybody sees the updated value,
3473 * so they won't put stuff in the queue again for no reason
3477 mutex_unlock(&memcg_cache_mutex
);
3481 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3483 struct kmem_cache
*c
;
3486 if (!s
->memcg_params
)
3488 if (!s
->memcg_params
->is_root_cache
)
3492 * If the cache is being destroyed, we trust that there is no one else
3493 * requesting objects from it. Even if there are, the sanity checks in
3494 * kmem_cache_destroy should caught this ill-case.
3496 * Still, we don't want anyone else freeing memcg_caches under our
3497 * noses, which can happen if a new memcg comes to life. As usual,
3498 * we'll take the set_limit_mutex to protect ourselves against this.
3500 mutex_lock(&set_limit_mutex
);
3501 for_each_memcg_cache_index(i
) {
3502 c
= cache_from_memcg_idx(s
, i
);
3507 * We will now manually delete the caches, so to avoid races
3508 * we need to cancel all pending destruction workers and
3509 * proceed with destruction ourselves.
3511 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3512 * and that could spawn the workers again: it is likely that
3513 * the cache still have active pages until this very moment.
3514 * This would lead us back to mem_cgroup_destroy_cache.
3516 * But that will not execute at all if the "dead" flag is not
3517 * set, so flip it down to guarantee we are in control.
3519 c
->memcg_params
->dead
= false;
3520 cancel_work_sync(&c
->memcg_params
->destroy
);
3521 kmem_cache_destroy(c
);
3523 mutex_unlock(&set_limit_mutex
);
3526 struct create_work
{
3527 struct mem_cgroup
*memcg
;
3528 struct kmem_cache
*cachep
;
3529 struct work_struct work
;
3532 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3534 struct kmem_cache
*cachep
;
3535 struct memcg_cache_params
*params
;
3537 if (!memcg_kmem_is_active(memcg
))
3540 mutex_lock(&memcg
->slab_caches_mutex
);
3541 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3542 cachep
= memcg_params_to_cache(params
);
3543 cachep
->memcg_params
->dead
= true;
3544 schedule_work(&cachep
->memcg_params
->destroy
);
3546 mutex_unlock(&memcg
->slab_caches_mutex
);
3549 static void memcg_create_cache_work_func(struct work_struct
*w
)
3551 struct create_work
*cw
;
3553 cw
= container_of(w
, struct create_work
, work
);
3554 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3559 * Enqueue the creation of a per-memcg kmem_cache.
3561 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3562 struct kmem_cache
*cachep
)
3564 struct create_work
*cw
;
3566 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3568 css_put(&memcg
->css
);
3573 cw
->cachep
= cachep
;
3575 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3576 schedule_work(&cw
->work
);
3579 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3580 struct kmem_cache
*cachep
)
3583 * We need to stop accounting when we kmalloc, because if the
3584 * corresponding kmalloc cache is not yet created, the first allocation
3585 * in __memcg_create_cache_enqueue will recurse.
3587 * However, it is better to enclose the whole function. Depending on
3588 * the debugging options enabled, INIT_WORK(), for instance, can
3589 * trigger an allocation. This too, will make us recurse. Because at
3590 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3591 * the safest choice is to do it like this, wrapping the whole function.
3593 memcg_stop_kmem_account();
3594 __memcg_create_cache_enqueue(memcg
, cachep
);
3595 memcg_resume_kmem_account();
3598 * Return the kmem_cache we're supposed to use for a slab allocation.
3599 * We try to use the current memcg's version of the cache.
3601 * If the cache does not exist yet, if we are the first user of it,
3602 * we either create it immediately, if possible, or create it asynchronously
3604 * In the latter case, we will let the current allocation go through with
3605 * the original cache.
3607 * Can't be called in interrupt context or from kernel threads.
3608 * This function needs to be called with rcu_read_lock() held.
3610 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3613 struct mem_cgroup
*memcg
;
3616 VM_BUG_ON(!cachep
->memcg_params
);
3617 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3619 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3623 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3625 if (!memcg_can_account_kmem(memcg
))
3628 idx
= memcg_cache_id(memcg
);
3631 * barrier to mare sure we're always seeing the up to date value. The
3632 * code updating memcg_caches will issue a write barrier to match this.
3634 read_barrier_depends();
3635 if (likely(cache_from_memcg_idx(cachep
, idx
))) {
3636 cachep
= cache_from_memcg_idx(cachep
, idx
);
3640 /* The corresponding put will be done in the workqueue. */
3641 if (!css_tryget(&memcg
->css
))
3646 * If we are in a safe context (can wait, and not in interrupt
3647 * context), we could be be predictable and return right away.
3648 * This would guarantee that the allocation being performed
3649 * already belongs in the new cache.
3651 * However, there are some clashes that can arrive from locking.
3652 * For instance, because we acquire the slab_mutex while doing
3653 * kmem_cache_dup, this means no further allocation could happen
3654 * with the slab_mutex held.
3656 * Also, because cache creation issue get_online_cpus(), this
3657 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3658 * that ends up reversed during cpu hotplug. (cpuset allocates
3659 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3660 * better to defer everything.
3662 memcg_create_cache_enqueue(memcg
, cachep
);
3668 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3671 * We need to verify if the allocation against current->mm->owner's memcg is
3672 * possible for the given order. But the page is not allocated yet, so we'll
3673 * need a further commit step to do the final arrangements.
3675 * It is possible for the task to switch cgroups in this mean time, so at
3676 * commit time, we can't rely on task conversion any longer. We'll then use
3677 * the handle argument to return to the caller which cgroup we should commit
3678 * against. We could also return the memcg directly and avoid the pointer
3679 * passing, but a boolean return value gives better semantics considering
3680 * the compiled-out case as well.
3682 * Returning true means the allocation is possible.
3685 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3687 struct mem_cgroup
*memcg
;
3693 * Disabling accounting is only relevant for some specific memcg
3694 * internal allocations. Therefore we would initially not have such
3695 * check here, since direct calls to the page allocator that are marked
3696 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3697 * concerned with cache allocations, and by having this test at
3698 * memcg_kmem_get_cache, we are already able to relay the allocation to
3699 * the root cache and bypass the memcg cache altogether.
3701 * There is one exception, though: the SLUB allocator does not create
3702 * large order caches, but rather service large kmallocs directly from
3703 * the page allocator. Therefore, the following sequence when backed by
3704 * the SLUB allocator:
3706 * memcg_stop_kmem_account();
3707 * kmalloc(<large_number>)
3708 * memcg_resume_kmem_account();
3710 * would effectively ignore the fact that we should skip accounting,
3711 * since it will drive us directly to this function without passing
3712 * through the cache selector memcg_kmem_get_cache. Such large
3713 * allocations are extremely rare but can happen, for instance, for the
3714 * cache arrays. We bring this test here.
3716 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3719 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3722 * very rare case described in mem_cgroup_from_task. Unfortunately there
3723 * isn't much we can do without complicating this too much, and it would
3724 * be gfp-dependent anyway. Just let it go
3726 if (unlikely(!memcg
))
3729 if (!memcg_can_account_kmem(memcg
)) {
3730 css_put(&memcg
->css
);
3734 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3738 css_put(&memcg
->css
);
3742 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3745 struct page_cgroup
*pc
;
3747 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3749 /* The page allocation failed. Revert */
3751 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3755 pc
= lookup_page_cgroup(page
);
3756 lock_page_cgroup(pc
);
3757 pc
->mem_cgroup
= memcg
;
3758 SetPageCgroupUsed(pc
);
3759 unlock_page_cgroup(pc
);
3762 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3764 struct mem_cgroup
*memcg
= NULL
;
3765 struct page_cgroup
*pc
;
3768 pc
= lookup_page_cgroup(page
);
3770 * Fast unlocked return. Theoretically might have changed, have to
3771 * check again after locking.
3773 if (!PageCgroupUsed(pc
))
3776 lock_page_cgroup(pc
);
3777 if (PageCgroupUsed(pc
)) {
3778 memcg
= pc
->mem_cgroup
;
3779 ClearPageCgroupUsed(pc
);
3781 unlock_page_cgroup(pc
);
3784 * We trust that only if there is a memcg associated with the page, it
3785 * is a valid allocation
3790 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3791 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3794 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3797 #endif /* CONFIG_MEMCG_KMEM */
3799 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3801 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3803 * Because tail pages are not marked as "used", set it. We're under
3804 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3805 * charge/uncharge will be never happen and move_account() is done under
3806 * compound_lock(), so we don't have to take care of races.
3808 void mem_cgroup_split_huge_fixup(struct page
*head
)
3810 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3811 struct page_cgroup
*pc
;
3812 struct mem_cgroup
*memcg
;
3815 if (mem_cgroup_disabled())
3818 memcg
= head_pc
->mem_cgroup
;
3819 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3821 pc
->mem_cgroup
= memcg
;
3822 smp_wmb();/* see __commit_charge() */
3823 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3825 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3828 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3831 void mem_cgroup_move_account_page_stat(struct mem_cgroup
*from
,
3832 struct mem_cgroup
*to
,
3833 unsigned int nr_pages
,
3834 enum mem_cgroup_stat_index idx
)
3836 /* Update stat data for mem_cgroup */
3838 __this_cpu_sub(from
->stat
->count
[idx
], nr_pages
);
3839 __this_cpu_add(to
->stat
->count
[idx
], nr_pages
);
3844 * mem_cgroup_move_account - move account of the page
3846 * @nr_pages: number of regular pages (>1 for huge pages)
3847 * @pc: page_cgroup of the page.
3848 * @from: mem_cgroup which the page is moved from.
3849 * @to: mem_cgroup which the page is moved to. @from != @to.
3851 * The caller must confirm following.
3852 * - page is not on LRU (isolate_page() is useful.)
3853 * - compound_lock is held when nr_pages > 1
3855 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3858 static int mem_cgroup_move_account(struct page
*page
,
3859 unsigned int nr_pages
,
3860 struct page_cgroup
*pc
,
3861 struct mem_cgroup
*from
,
3862 struct mem_cgroup
*to
)
3864 unsigned long flags
;
3866 bool anon
= PageAnon(page
);
3868 VM_BUG_ON(from
== to
);
3869 VM_BUG_ON(PageLRU(page
));
3871 * The page is isolated from LRU. So, collapse function
3872 * will not handle this page. But page splitting can happen.
3873 * Do this check under compound_page_lock(). The caller should
3877 if (nr_pages
> 1 && !PageTransHuge(page
))
3880 lock_page_cgroup(pc
);
3883 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3886 move_lock_mem_cgroup(from
, &flags
);
3888 if (!anon
&& page_mapped(page
))
3889 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3890 MEM_CGROUP_STAT_FILE_MAPPED
);
3892 if (PageWriteback(page
))
3893 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3894 MEM_CGROUP_STAT_WRITEBACK
);
3896 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3898 /* caller should have done css_get */
3899 pc
->mem_cgroup
= to
;
3900 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3901 move_unlock_mem_cgroup(from
, &flags
);
3904 unlock_page_cgroup(pc
);
3908 memcg_check_events(to
, page
);
3909 memcg_check_events(from
, page
);
3915 * mem_cgroup_move_parent - moves page to the parent group
3916 * @page: the page to move
3917 * @pc: page_cgroup of the page
3918 * @child: page's cgroup
3920 * move charges to its parent or the root cgroup if the group has no
3921 * parent (aka use_hierarchy==0).
3922 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3923 * mem_cgroup_move_account fails) the failure is always temporary and
3924 * it signals a race with a page removal/uncharge or migration. In the
3925 * first case the page is on the way out and it will vanish from the LRU
3926 * on the next attempt and the call should be retried later.
3927 * Isolation from the LRU fails only if page has been isolated from
3928 * the LRU since we looked at it and that usually means either global
3929 * reclaim or migration going on. The page will either get back to the
3931 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3932 * (!PageCgroupUsed) or moved to a different group. The page will
3933 * disappear in the next attempt.
3935 static int mem_cgroup_move_parent(struct page
*page
,
3936 struct page_cgroup
*pc
,
3937 struct mem_cgroup
*child
)
3939 struct mem_cgroup
*parent
;
3940 unsigned int nr_pages
;
3941 unsigned long uninitialized_var(flags
);
3944 VM_BUG_ON(mem_cgroup_is_root(child
));
3947 if (!get_page_unless_zero(page
))
3949 if (isolate_lru_page(page
))
3952 nr_pages
= hpage_nr_pages(page
);
3954 parent
= parent_mem_cgroup(child
);
3956 * If no parent, move charges to root cgroup.
3959 parent
= root_mem_cgroup
;
3962 VM_BUG_ON(!PageTransHuge(page
));
3963 flags
= compound_lock_irqsave(page
);
3966 ret
= mem_cgroup_move_account(page
, nr_pages
,
3969 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3972 compound_unlock_irqrestore(page
, flags
);
3973 putback_lru_page(page
);
3981 * Charge the memory controller for page usage.
3983 * 0 if the charge was successful
3984 * < 0 if the cgroup is over its limit
3986 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3987 gfp_t gfp_mask
, enum charge_type ctype
)
3989 struct mem_cgroup
*memcg
= NULL
;
3990 unsigned int nr_pages
= 1;
3994 if (PageTransHuge(page
)) {
3995 nr_pages
<<= compound_order(page
);
3996 VM_BUG_ON(!PageTransHuge(page
));
3998 * Never OOM-kill a process for a huge page. The
3999 * fault handler will fall back to regular pages.
4004 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
4007 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
4011 int mem_cgroup_newpage_charge(struct page
*page
,
4012 struct mm_struct
*mm
, gfp_t gfp_mask
)
4014 if (mem_cgroup_disabled())
4016 VM_BUG_ON(page_mapped(page
));
4017 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4019 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
4020 MEM_CGROUP_CHARGE_TYPE_ANON
);
4024 * While swap-in, try_charge -> commit or cancel, the page is locked.
4025 * And when try_charge() successfully returns, one refcnt to memcg without
4026 * struct page_cgroup is acquired. This refcnt will be consumed by
4027 * "commit()" or removed by "cancel()"
4029 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
4032 struct mem_cgroup
**memcgp
)
4034 struct mem_cgroup
*memcg
;
4035 struct page_cgroup
*pc
;
4038 pc
= lookup_page_cgroup(page
);
4040 * Every swap fault against a single page tries to charge the
4041 * page, bail as early as possible. shmem_unuse() encounters
4042 * already charged pages, too. The USED bit is protected by
4043 * the page lock, which serializes swap cache removal, which
4044 * in turn serializes uncharging.
4046 if (PageCgroupUsed(pc
))
4048 if (!do_swap_account
)
4050 memcg
= try_get_mem_cgroup_from_page(page
);
4054 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
4055 css_put(&memcg
->css
);
4060 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
4066 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
4067 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
4070 if (mem_cgroup_disabled())
4073 * A racing thread's fault, or swapoff, may have already
4074 * updated the pte, and even removed page from swap cache: in
4075 * those cases unuse_pte()'s pte_same() test will fail; but
4076 * there's also a KSM case which does need to charge the page.
4078 if (!PageSwapCache(page
)) {
4081 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4086 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4089 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4091 if (mem_cgroup_disabled())
4095 __mem_cgroup_cancel_charge(memcg
, 1);
4099 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4100 enum charge_type ctype
)
4102 if (mem_cgroup_disabled())
4107 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4109 * Now swap is on-memory. This means this page may be
4110 * counted both as mem and swap....double count.
4111 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4112 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4113 * may call delete_from_swap_cache() before reach here.
4115 if (do_swap_account
&& PageSwapCache(page
)) {
4116 swp_entry_t ent
= {.val
= page_private(page
)};
4117 mem_cgroup_uncharge_swap(ent
);
4121 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4122 struct mem_cgroup
*memcg
)
4124 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4125 MEM_CGROUP_CHARGE_TYPE_ANON
);
4128 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4131 struct mem_cgroup
*memcg
= NULL
;
4132 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4135 if (mem_cgroup_disabled())
4137 if (PageCompound(page
))
4140 if (!PageSwapCache(page
))
4141 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4142 else { /* page is swapcache/shmem */
4143 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4146 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4151 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4152 unsigned int nr_pages
,
4153 const enum charge_type ctype
)
4155 struct memcg_batch_info
*batch
= NULL
;
4156 bool uncharge_memsw
= true;
4158 /* If swapout, usage of swap doesn't decrease */
4159 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4160 uncharge_memsw
= false;
4162 batch
= ¤t
->memcg_batch
;
4164 * In usual, we do css_get() when we remember memcg pointer.
4165 * But in this case, we keep res->usage until end of a series of
4166 * uncharges. Then, it's ok to ignore memcg's refcnt.
4169 batch
->memcg
= memcg
;
4171 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4172 * In those cases, all pages freed continuously can be expected to be in
4173 * the same cgroup and we have chance to coalesce uncharges.
4174 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4175 * because we want to do uncharge as soon as possible.
4178 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4179 goto direct_uncharge
;
4182 goto direct_uncharge
;
4185 * In typical case, batch->memcg == mem. This means we can
4186 * merge a series of uncharges to an uncharge of res_counter.
4187 * If not, we uncharge res_counter ony by one.
4189 if (batch
->memcg
!= memcg
)
4190 goto direct_uncharge
;
4191 /* remember freed charge and uncharge it later */
4194 batch
->memsw_nr_pages
++;
4197 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4199 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4200 if (unlikely(batch
->memcg
!= memcg
))
4201 memcg_oom_recover(memcg
);
4205 * uncharge if !page_mapped(page)
4207 static struct mem_cgroup
*
4208 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4211 struct mem_cgroup
*memcg
= NULL
;
4212 unsigned int nr_pages
= 1;
4213 struct page_cgroup
*pc
;
4216 if (mem_cgroup_disabled())
4219 if (PageTransHuge(page
)) {
4220 nr_pages
<<= compound_order(page
);
4221 VM_BUG_ON(!PageTransHuge(page
));
4224 * Check if our page_cgroup is valid
4226 pc
= lookup_page_cgroup(page
);
4227 if (unlikely(!PageCgroupUsed(pc
)))
4230 lock_page_cgroup(pc
);
4232 memcg
= pc
->mem_cgroup
;
4234 if (!PageCgroupUsed(pc
))
4237 anon
= PageAnon(page
);
4240 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4242 * Generally PageAnon tells if it's the anon statistics to be
4243 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4244 * used before page reached the stage of being marked PageAnon.
4248 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4249 /* See mem_cgroup_prepare_migration() */
4250 if (page_mapped(page
))
4253 * Pages under migration may not be uncharged. But
4254 * end_migration() /must/ be the one uncharging the
4255 * unused post-migration page and so it has to call
4256 * here with the migration bit still set. See the
4257 * res_counter handling below.
4259 if (!end_migration
&& PageCgroupMigration(pc
))
4262 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4263 if (!PageAnon(page
)) { /* Shared memory */
4264 if (page
->mapping
&& !page_is_file_cache(page
))
4266 } else if (page_mapped(page
)) /* Anon */
4273 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4275 ClearPageCgroupUsed(pc
);
4277 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4278 * freed from LRU. This is safe because uncharged page is expected not
4279 * to be reused (freed soon). Exception is SwapCache, it's handled by
4280 * special functions.
4283 unlock_page_cgroup(pc
);
4285 * even after unlock, we have memcg->res.usage here and this memcg
4286 * will never be freed, so it's safe to call css_get().
4288 memcg_check_events(memcg
, page
);
4289 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4290 mem_cgroup_swap_statistics(memcg
, true);
4291 css_get(&memcg
->css
);
4294 * Migration does not charge the res_counter for the
4295 * replacement page, so leave it alone when phasing out the
4296 * page that is unused after the migration.
4298 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4299 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4304 unlock_page_cgroup(pc
);
4308 void mem_cgroup_uncharge_page(struct page
*page
)
4311 if (page_mapped(page
))
4313 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4315 * If the page is in swap cache, uncharge should be deferred
4316 * to the swap path, which also properly accounts swap usage
4317 * and handles memcg lifetime.
4319 * Note that this check is not stable and reclaim may add the
4320 * page to swap cache at any time after this. However, if the
4321 * page is not in swap cache by the time page->mapcount hits
4322 * 0, there won't be any page table references to the swap
4323 * slot, and reclaim will free it and not actually write the
4326 if (PageSwapCache(page
))
4328 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4331 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4333 VM_BUG_ON(page_mapped(page
));
4334 VM_BUG_ON(page
->mapping
);
4335 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4339 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4340 * In that cases, pages are freed continuously and we can expect pages
4341 * are in the same memcg. All these calls itself limits the number of
4342 * pages freed at once, then uncharge_start/end() is called properly.
4343 * This may be called prural(2) times in a context,
4346 void mem_cgroup_uncharge_start(void)
4348 current
->memcg_batch
.do_batch
++;
4349 /* We can do nest. */
4350 if (current
->memcg_batch
.do_batch
== 1) {
4351 current
->memcg_batch
.memcg
= NULL
;
4352 current
->memcg_batch
.nr_pages
= 0;
4353 current
->memcg_batch
.memsw_nr_pages
= 0;
4357 void mem_cgroup_uncharge_end(void)
4359 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4361 if (!batch
->do_batch
)
4365 if (batch
->do_batch
) /* If stacked, do nothing. */
4371 * This "batch->memcg" is valid without any css_get/put etc...
4372 * bacause we hide charges behind us.
4374 if (batch
->nr_pages
)
4375 res_counter_uncharge(&batch
->memcg
->res
,
4376 batch
->nr_pages
* PAGE_SIZE
);
4377 if (batch
->memsw_nr_pages
)
4378 res_counter_uncharge(&batch
->memcg
->memsw
,
4379 batch
->memsw_nr_pages
* PAGE_SIZE
);
4380 memcg_oom_recover(batch
->memcg
);
4381 /* forget this pointer (for sanity check) */
4382 batch
->memcg
= NULL
;
4387 * called after __delete_from_swap_cache() and drop "page" account.
4388 * memcg information is recorded to swap_cgroup of "ent"
4391 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4393 struct mem_cgroup
*memcg
;
4394 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4396 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4397 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4399 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4402 * record memcg information, if swapout && memcg != NULL,
4403 * css_get() was called in uncharge().
4405 if (do_swap_account
&& swapout
&& memcg
)
4406 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4410 #ifdef CONFIG_MEMCG_SWAP
4412 * called from swap_entry_free(). remove record in swap_cgroup and
4413 * uncharge "memsw" account.
4415 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4417 struct mem_cgroup
*memcg
;
4420 if (!do_swap_account
)
4423 id
= swap_cgroup_record(ent
, 0);
4425 memcg
= mem_cgroup_lookup(id
);
4428 * We uncharge this because swap is freed.
4429 * This memcg can be obsolete one. We avoid calling css_tryget
4431 if (!mem_cgroup_is_root(memcg
))
4432 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4433 mem_cgroup_swap_statistics(memcg
, false);
4434 css_put(&memcg
->css
);
4440 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4441 * @entry: swap entry to be moved
4442 * @from: mem_cgroup which the entry is moved from
4443 * @to: mem_cgroup which the entry is moved to
4445 * It succeeds only when the swap_cgroup's record for this entry is the same
4446 * as the mem_cgroup's id of @from.
4448 * Returns 0 on success, -EINVAL on failure.
4450 * The caller must have charged to @to, IOW, called res_counter_charge() about
4451 * both res and memsw, and called css_get().
4453 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4454 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4456 unsigned short old_id
, new_id
;
4458 old_id
= mem_cgroup_id(from
);
4459 new_id
= mem_cgroup_id(to
);
4461 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4462 mem_cgroup_swap_statistics(from
, false);
4463 mem_cgroup_swap_statistics(to
, true);
4465 * This function is only called from task migration context now.
4466 * It postpones res_counter and refcount handling till the end
4467 * of task migration(mem_cgroup_clear_mc()) for performance
4468 * improvement. But we cannot postpone css_get(to) because if
4469 * the process that has been moved to @to does swap-in, the
4470 * refcount of @to might be decreased to 0.
4472 * We are in attach() phase, so the cgroup is guaranteed to be
4473 * alive, so we can just call css_get().
4481 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4482 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4489 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4492 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4493 struct mem_cgroup
**memcgp
)
4495 struct mem_cgroup
*memcg
= NULL
;
4496 unsigned int nr_pages
= 1;
4497 struct page_cgroup
*pc
;
4498 enum charge_type ctype
;
4502 if (mem_cgroup_disabled())
4505 if (PageTransHuge(page
))
4506 nr_pages
<<= compound_order(page
);
4508 pc
= lookup_page_cgroup(page
);
4509 lock_page_cgroup(pc
);
4510 if (PageCgroupUsed(pc
)) {
4511 memcg
= pc
->mem_cgroup
;
4512 css_get(&memcg
->css
);
4514 * At migrating an anonymous page, its mapcount goes down
4515 * to 0 and uncharge() will be called. But, even if it's fully
4516 * unmapped, migration may fail and this page has to be
4517 * charged again. We set MIGRATION flag here and delay uncharge
4518 * until end_migration() is called
4520 * Corner Case Thinking
4522 * When the old page was mapped as Anon and it's unmap-and-freed
4523 * while migration was ongoing.
4524 * If unmap finds the old page, uncharge() of it will be delayed
4525 * until end_migration(). If unmap finds a new page, it's
4526 * uncharged when it make mapcount to be 1->0. If unmap code
4527 * finds swap_migration_entry, the new page will not be mapped
4528 * and end_migration() will find it(mapcount==0).
4531 * When the old page was mapped but migraion fails, the kernel
4532 * remaps it. A charge for it is kept by MIGRATION flag even
4533 * if mapcount goes down to 0. We can do remap successfully
4534 * without charging it again.
4537 * The "old" page is under lock_page() until the end of
4538 * migration, so, the old page itself will not be swapped-out.
4539 * If the new page is swapped out before end_migraton, our
4540 * hook to usual swap-out path will catch the event.
4543 SetPageCgroupMigration(pc
);
4545 unlock_page_cgroup(pc
);
4547 * If the page is not charged at this point,
4555 * We charge new page before it's used/mapped. So, even if unlock_page()
4556 * is called before end_migration, we can catch all events on this new
4557 * page. In the case new page is migrated but not remapped, new page's
4558 * mapcount will be finally 0 and we call uncharge in end_migration().
4561 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4563 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4565 * The page is committed to the memcg, but it's not actually
4566 * charged to the res_counter since we plan on replacing the
4567 * old one and only one page is going to be left afterwards.
4569 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4572 /* remove redundant charge if migration failed*/
4573 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4574 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4576 struct page
*used
, *unused
;
4577 struct page_cgroup
*pc
;
4583 if (!migration_ok
) {
4590 anon
= PageAnon(used
);
4591 __mem_cgroup_uncharge_common(unused
,
4592 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4593 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4595 css_put(&memcg
->css
);
4597 * We disallowed uncharge of pages under migration because mapcount
4598 * of the page goes down to zero, temporarly.
4599 * Clear the flag and check the page should be charged.
4601 pc
= lookup_page_cgroup(oldpage
);
4602 lock_page_cgroup(pc
);
4603 ClearPageCgroupMigration(pc
);
4604 unlock_page_cgroup(pc
);
4607 * If a page is a file cache, radix-tree replacement is very atomic
4608 * and we can skip this check. When it was an Anon page, its mapcount
4609 * goes down to 0. But because we added MIGRATION flage, it's not
4610 * uncharged yet. There are several case but page->mapcount check
4611 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4612 * check. (see prepare_charge() also)
4615 mem_cgroup_uncharge_page(used
);
4619 * At replace page cache, newpage is not under any memcg but it's on
4620 * LRU. So, this function doesn't touch res_counter but handles LRU
4621 * in correct way. Both pages are locked so we cannot race with uncharge.
4623 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4624 struct page
*newpage
)
4626 struct mem_cgroup
*memcg
= NULL
;
4627 struct page_cgroup
*pc
;
4628 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4630 if (mem_cgroup_disabled())
4633 pc
= lookup_page_cgroup(oldpage
);
4634 /* fix accounting on old pages */
4635 lock_page_cgroup(pc
);
4636 if (PageCgroupUsed(pc
)) {
4637 memcg
= pc
->mem_cgroup
;
4638 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4639 ClearPageCgroupUsed(pc
);
4641 unlock_page_cgroup(pc
);
4644 * When called from shmem_replace_page(), in some cases the
4645 * oldpage has already been charged, and in some cases not.
4650 * Even if newpage->mapping was NULL before starting replacement,
4651 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4652 * LRU while we overwrite pc->mem_cgroup.
4654 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4657 #ifdef CONFIG_DEBUG_VM
4658 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4660 struct page_cgroup
*pc
;
4662 pc
= lookup_page_cgroup(page
);
4664 * Can be NULL while feeding pages into the page allocator for
4665 * the first time, i.e. during boot or memory hotplug;
4666 * or when mem_cgroup_disabled().
4668 if (likely(pc
) && PageCgroupUsed(pc
))
4673 bool mem_cgroup_bad_page_check(struct page
*page
)
4675 if (mem_cgroup_disabled())
4678 return lookup_page_cgroup_used(page
) != NULL
;
4681 void mem_cgroup_print_bad_page(struct page
*page
)
4683 struct page_cgroup
*pc
;
4685 pc
= lookup_page_cgroup_used(page
);
4687 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4688 pc
, pc
->flags
, pc
->mem_cgroup
);
4693 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4694 unsigned long long val
)
4697 u64 memswlimit
, memlimit
;
4699 int children
= mem_cgroup_count_children(memcg
);
4700 u64 curusage
, oldusage
;
4704 * For keeping hierarchical_reclaim simple, how long we should retry
4705 * is depends on callers. We set our retry-count to be function
4706 * of # of children which we should visit in this loop.
4708 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4710 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4713 while (retry_count
) {
4714 if (signal_pending(current
)) {
4719 * Rather than hide all in some function, I do this in
4720 * open coded manner. You see what this really does.
4721 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4723 mutex_lock(&set_limit_mutex
);
4724 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4725 if (memswlimit
< val
) {
4727 mutex_unlock(&set_limit_mutex
);
4731 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4735 ret
= res_counter_set_limit(&memcg
->res
, val
);
4737 if (memswlimit
== val
)
4738 memcg
->memsw_is_minimum
= true;
4740 memcg
->memsw_is_minimum
= false;
4742 mutex_unlock(&set_limit_mutex
);
4747 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4748 MEM_CGROUP_RECLAIM_SHRINK
);
4749 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4750 /* Usage is reduced ? */
4751 if (curusage
>= oldusage
)
4754 oldusage
= curusage
;
4756 if (!ret
&& enlarge
)
4757 memcg_oom_recover(memcg
);
4762 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4763 unsigned long long val
)
4766 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4767 int children
= mem_cgroup_count_children(memcg
);
4771 /* see mem_cgroup_resize_res_limit */
4772 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4773 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4774 while (retry_count
) {
4775 if (signal_pending(current
)) {
4780 * Rather than hide all in some function, I do this in
4781 * open coded manner. You see what this really does.
4782 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4784 mutex_lock(&set_limit_mutex
);
4785 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4786 if (memlimit
> val
) {
4788 mutex_unlock(&set_limit_mutex
);
4791 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4792 if (memswlimit
< val
)
4794 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4796 if (memlimit
== val
)
4797 memcg
->memsw_is_minimum
= true;
4799 memcg
->memsw_is_minimum
= false;
4801 mutex_unlock(&set_limit_mutex
);
4806 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4807 MEM_CGROUP_RECLAIM_NOSWAP
|
4808 MEM_CGROUP_RECLAIM_SHRINK
);
4809 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4810 /* Usage is reduced ? */
4811 if (curusage
>= oldusage
)
4814 oldusage
= curusage
;
4816 if (!ret
&& enlarge
)
4817 memcg_oom_recover(memcg
);
4821 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4823 unsigned long *total_scanned
)
4825 unsigned long nr_reclaimed
= 0;
4826 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4827 unsigned long reclaimed
;
4829 struct mem_cgroup_tree_per_zone
*mctz
;
4830 unsigned long long excess
;
4831 unsigned long nr_scanned
;
4836 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4838 * This loop can run a while, specially if mem_cgroup's continuously
4839 * keep exceeding their soft limit and putting the system under
4846 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4851 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4852 gfp_mask
, &nr_scanned
);
4853 nr_reclaimed
+= reclaimed
;
4854 *total_scanned
+= nr_scanned
;
4855 spin_lock(&mctz
->lock
);
4858 * If we failed to reclaim anything from this memory cgroup
4859 * it is time to move on to the next cgroup
4865 * Loop until we find yet another one.
4867 * By the time we get the soft_limit lock
4868 * again, someone might have aded the
4869 * group back on the RB tree. Iterate to
4870 * make sure we get a different mem.
4871 * mem_cgroup_largest_soft_limit_node returns
4872 * NULL if no other cgroup is present on
4876 __mem_cgroup_largest_soft_limit_node(mctz
);
4878 css_put(&next_mz
->memcg
->css
);
4879 else /* next_mz == NULL or other memcg */
4883 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4884 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4886 * One school of thought says that we should not add
4887 * back the node to the tree if reclaim returns 0.
4888 * But our reclaim could return 0, simply because due
4889 * to priority we are exposing a smaller subset of
4890 * memory to reclaim from. Consider this as a longer
4893 /* If excess == 0, no tree ops */
4894 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4895 spin_unlock(&mctz
->lock
);
4896 css_put(&mz
->memcg
->css
);
4899 * Could not reclaim anything and there are no more
4900 * mem cgroups to try or we seem to be looping without
4901 * reclaiming anything.
4903 if (!nr_reclaimed
&&
4905 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4907 } while (!nr_reclaimed
);
4909 css_put(&next_mz
->memcg
->css
);
4910 return nr_reclaimed
;
4914 * mem_cgroup_force_empty_list - clears LRU of a group
4915 * @memcg: group to clear
4918 * @lru: lru to to clear
4920 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4921 * reclaim the pages page themselves - pages are moved to the parent (or root)
4924 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4925 int node
, int zid
, enum lru_list lru
)
4927 struct lruvec
*lruvec
;
4928 unsigned long flags
;
4929 struct list_head
*list
;
4933 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4934 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4935 list
= &lruvec
->lists
[lru
];
4939 struct page_cgroup
*pc
;
4942 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4943 if (list_empty(list
)) {
4944 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4947 page
= list_entry(list
->prev
, struct page
, lru
);
4949 list_move(&page
->lru
, list
);
4951 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4954 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4956 pc
= lookup_page_cgroup(page
);
4958 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4959 /* found lock contention or "pc" is obsolete. */
4964 } while (!list_empty(list
));
4968 * make mem_cgroup's charge to be 0 if there is no task by moving
4969 * all the charges and pages to the parent.
4970 * This enables deleting this mem_cgroup.
4972 * Caller is responsible for holding css reference on the memcg.
4974 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4980 /* This is for making all *used* pages to be on LRU. */
4981 lru_add_drain_all();
4982 drain_all_stock_sync(memcg
);
4983 mem_cgroup_start_move(memcg
);
4984 for_each_node_state(node
, N_MEMORY
) {
4985 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4988 mem_cgroup_force_empty_list(memcg
,
4993 mem_cgroup_end_move(memcg
);
4994 memcg_oom_recover(memcg
);
4998 * Kernel memory may not necessarily be trackable to a specific
4999 * process. So they are not migrated, and therefore we can't
5000 * expect their value to drop to 0 here.
5001 * Having res filled up with kmem only is enough.
5003 * This is a safety check because mem_cgroup_force_empty_list
5004 * could have raced with mem_cgroup_replace_page_cache callers
5005 * so the lru seemed empty but the page could have been added
5006 * right after the check. RES_USAGE should be safe as we always
5007 * charge before adding to the LRU.
5009 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
5010 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
5011 } while (usage
> 0);
5014 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
5016 lockdep_assert_held(&memcg_create_mutex
);
5018 * The lock does not prevent addition or deletion to the list
5019 * of children, but it prevents a new child from being
5020 * initialized based on this parent in css_online(), so it's
5021 * enough to decide whether hierarchically inherited
5022 * attributes can still be changed or not.
5024 return memcg
->use_hierarchy
&&
5025 !list_empty(&memcg
->css
.cgroup
->children
);
5029 * Reclaims as many pages from the given memcg as possible and moves
5030 * the rest to the parent.
5032 * Caller is responsible for holding css reference for memcg.
5034 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
5036 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
5037 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
5039 /* returns EBUSY if there is a task or if we come here twice. */
5040 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
5043 /* we call try-to-free pages for make this cgroup empty */
5044 lru_add_drain_all();
5045 /* try to free all pages in this cgroup */
5046 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
5049 if (signal_pending(current
))
5052 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
5056 /* maybe some writeback is necessary */
5057 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
5062 mem_cgroup_reparent_charges(memcg
);
5067 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5070 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5072 if (mem_cgroup_is_root(memcg
))
5074 return mem_cgroup_force_empty(memcg
);
5077 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5080 return mem_cgroup_from_css(css
)->use_hierarchy
;
5083 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5084 struct cftype
*cft
, u64 val
)
5087 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5088 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5090 mutex_lock(&memcg_create_mutex
);
5092 if (memcg
->use_hierarchy
== val
)
5096 * If parent's use_hierarchy is set, we can't make any modifications
5097 * in the child subtrees. If it is unset, then the change can
5098 * occur, provided the current cgroup has no children.
5100 * For the root cgroup, parent_mem is NULL, we allow value to be
5101 * set if there are no children.
5103 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5104 (val
== 1 || val
== 0)) {
5105 if (list_empty(&memcg
->css
.cgroup
->children
))
5106 memcg
->use_hierarchy
= val
;
5113 mutex_unlock(&memcg_create_mutex
);
5119 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5120 enum mem_cgroup_stat_index idx
)
5122 struct mem_cgroup
*iter
;
5125 /* Per-cpu values can be negative, use a signed accumulator */
5126 for_each_mem_cgroup_tree(iter
, memcg
)
5127 val
+= mem_cgroup_read_stat(iter
, idx
);
5129 if (val
< 0) /* race ? */
5134 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5138 if (!mem_cgroup_is_root(memcg
)) {
5140 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5142 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5146 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5147 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5149 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5150 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5153 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5155 return val
<< PAGE_SHIFT
;
5158 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
5161 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5166 type
= MEMFILE_TYPE(cft
->private);
5167 name
= MEMFILE_ATTR(cft
->private);
5171 if (name
== RES_USAGE
)
5172 val
= mem_cgroup_usage(memcg
, false);
5174 val
= res_counter_read_u64(&memcg
->res
, name
);
5177 if (name
== RES_USAGE
)
5178 val
= mem_cgroup_usage(memcg
, true);
5180 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5183 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5192 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5195 #ifdef CONFIG_MEMCG_KMEM
5196 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5198 * For simplicity, we won't allow this to be disabled. It also can't
5199 * be changed if the cgroup has children already, or if tasks had
5202 * If tasks join before we set the limit, a person looking at
5203 * kmem.usage_in_bytes will have no way to determine when it took
5204 * place, which makes the value quite meaningless.
5206 * After it first became limited, changes in the value of the limit are
5207 * of course permitted.
5209 mutex_lock(&memcg_create_mutex
);
5210 mutex_lock(&set_limit_mutex
);
5211 if (!memcg
->kmem_account_flags
&& val
!= RES_COUNTER_MAX
) {
5212 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5216 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5219 ret
= memcg_update_cache_sizes(memcg
);
5221 res_counter_set_limit(&memcg
->kmem
, RES_COUNTER_MAX
);
5224 static_key_slow_inc(&memcg_kmem_enabled_key
);
5226 * setting the active bit after the inc will guarantee no one
5227 * starts accounting before all call sites are patched
5229 memcg_kmem_set_active(memcg
);
5231 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5233 mutex_unlock(&set_limit_mutex
);
5234 mutex_unlock(&memcg_create_mutex
);
5239 #ifdef CONFIG_MEMCG_KMEM
5240 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5243 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5247 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5249 * When that happen, we need to disable the static branch only on those
5250 * memcgs that enabled it. To achieve this, we would be forced to
5251 * complicate the code by keeping track of which memcgs were the ones
5252 * that actually enabled limits, and which ones got it from its
5255 * It is a lot simpler just to do static_key_slow_inc() on every child
5256 * that is accounted.
5258 if (!memcg_kmem_is_active(memcg
))
5262 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5263 * memcg is active already. If the later initialization fails then the
5264 * cgroup core triggers the cleanup so we do not have to do it here.
5266 static_key_slow_inc(&memcg_kmem_enabled_key
);
5268 mutex_lock(&set_limit_mutex
);
5269 memcg_stop_kmem_account();
5270 ret
= memcg_update_cache_sizes(memcg
);
5271 memcg_resume_kmem_account();
5272 mutex_unlock(&set_limit_mutex
);
5276 #endif /* CONFIG_MEMCG_KMEM */
5279 * The user of this function is...
5282 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5285 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5288 unsigned long long val
;
5291 type
= MEMFILE_TYPE(cft
->private);
5292 name
= MEMFILE_ATTR(cft
->private);
5296 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5300 /* This function does all necessary parse...reuse it */
5301 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5305 ret
= mem_cgroup_resize_limit(memcg
, val
);
5306 else if (type
== _MEMSWAP
)
5307 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5308 else if (type
== _KMEM
)
5309 ret
= memcg_update_kmem_limit(css
, val
);
5313 case RES_SOFT_LIMIT
:
5314 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5318 * For memsw, soft limits are hard to implement in terms
5319 * of semantics, for now, we support soft limits for
5320 * control without swap
5323 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5328 ret
= -EINVAL
; /* should be BUG() ? */
5334 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5335 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5337 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5339 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5340 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5341 if (!memcg
->use_hierarchy
)
5344 while (css_parent(&memcg
->css
)) {
5345 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5346 if (!memcg
->use_hierarchy
)
5348 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5349 min_limit
= min(min_limit
, tmp
);
5350 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5351 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5354 *mem_limit
= min_limit
;
5355 *memsw_limit
= min_memsw_limit
;
5358 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5360 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5364 type
= MEMFILE_TYPE(event
);
5365 name
= MEMFILE_ATTR(event
);
5370 res_counter_reset_max(&memcg
->res
);
5371 else if (type
== _MEMSWAP
)
5372 res_counter_reset_max(&memcg
->memsw
);
5373 else if (type
== _KMEM
)
5374 res_counter_reset_max(&memcg
->kmem
);
5380 res_counter_reset_failcnt(&memcg
->res
);
5381 else if (type
== _MEMSWAP
)
5382 res_counter_reset_failcnt(&memcg
->memsw
);
5383 else if (type
== _KMEM
)
5384 res_counter_reset_failcnt(&memcg
->kmem
);
5393 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5396 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5400 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5401 struct cftype
*cft
, u64 val
)
5403 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5405 if (val
>= (1 << NR_MOVE_TYPE
))
5409 * No kind of locking is needed in here, because ->can_attach() will
5410 * check this value once in the beginning of the process, and then carry
5411 * on with stale data. This means that changes to this value will only
5412 * affect task migrations starting after the change.
5414 memcg
->move_charge_at_immigrate
= val
;
5418 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5419 struct cftype
*cft
, u64 val
)
5426 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
5430 unsigned int lru_mask
;
5433 static const struct numa_stat stats
[] = {
5434 { "total", LRU_ALL
},
5435 { "file", LRU_ALL_FILE
},
5436 { "anon", LRU_ALL_ANON
},
5437 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5439 const struct numa_stat
*stat
;
5442 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5444 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5445 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5446 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5447 for_each_node_state(nid
, N_MEMORY
) {
5448 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5450 seq_printf(m
, " N%d=%lu", nid
, nr
);
5455 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5456 struct mem_cgroup
*iter
;
5459 for_each_mem_cgroup_tree(iter
, memcg
)
5460 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5461 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5462 for_each_node_state(nid
, N_MEMORY
) {
5464 for_each_mem_cgroup_tree(iter
, memcg
)
5465 nr
+= mem_cgroup_node_nr_lru_pages(
5466 iter
, nid
, stat
->lru_mask
);
5467 seq_printf(m
, " N%d=%lu", nid
, nr
);
5474 #endif /* CONFIG_NUMA */
5476 static inline void mem_cgroup_lru_names_not_uptodate(void)
5478 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5481 static int memcg_stat_show(struct seq_file
*m
, void *v
)
5483 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5484 struct mem_cgroup
*mi
;
5487 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5488 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5490 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5491 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5494 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5495 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5496 mem_cgroup_read_events(memcg
, i
));
5498 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5499 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5500 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5502 /* Hierarchical information */
5504 unsigned long long limit
, memsw_limit
;
5505 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5506 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5507 if (do_swap_account
)
5508 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5512 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5515 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5517 for_each_mem_cgroup_tree(mi
, memcg
)
5518 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5519 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5522 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5523 unsigned long long val
= 0;
5525 for_each_mem_cgroup_tree(mi
, memcg
)
5526 val
+= mem_cgroup_read_events(mi
, i
);
5527 seq_printf(m
, "total_%s %llu\n",
5528 mem_cgroup_events_names
[i
], val
);
5531 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5532 unsigned long long val
= 0;
5534 for_each_mem_cgroup_tree(mi
, memcg
)
5535 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5536 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5539 #ifdef CONFIG_DEBUG_VM
5542 struct mem_cgroup_per_zone
*mz
;
5543 struct zone_reclaim_stat
*rstat
;
5544 unsigned long recent_rotated
[2] = {0, 0};
5545 unsigned long recent_scanned
[2] = {0, 0};
5547 for_each_online_node(nid
)
5548 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5549 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5550 rstat
= &mz
->lruvec
.reclaim_stat
;
5552 recent_rotated
[0] += rstat
->recent_rotated
[0];
5553 recent_rotated
[1] += rstat
->recent_rotated
[1];
5554 recent_scanned
[0] += rstat
->recent_scanned
[0];
5555 recent_scanned
[1] += rstat
->recent_scanned
[1];
5557 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5558 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5559 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5560 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5567 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5570 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5572 return mem_cgroup_swappiness(memcg
);
5575 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5576 struct cftype
*cft
, u64 val
)
5578 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5579 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5581 if (val
> 100 || !parent
)
5584 mutex_lock(&memcg_create_mutex
);
5586 /* If under hierarchy, only empty-root can set this value */
5587 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5588 mutex_unlock(&memcg_create_mutex
);
5592 memcg
->swappiness
= val
;
5594 mutex_unlock(&memcg_create_mutex
);
5599 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5601 struct mem_cgroup_threshold_ary
*t
;
5607 t
= rcu_dereference(memcg
->thresholds
.primary
);
5609 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5614 usage
= mem_cgroup_usage(memcg
, swap
);
5617 * current_threshold points to threshold just below or equal to usage.
5618 * If it's not true, a threshold was crossed after last
5619 * call of __mem_cgroup_threshold().
5621 i
= t
->current_threshold
;
5624 * Iterate backward over array of thresholds starting from
5625 * current_threshold and check if a threshold is crossed.
5626 * If none of thresholds below usage is crossed, we read
5627 * only one element of the array here.
5629 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5630 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5632 /* i = current_threshold + 1 */
5636 * Iterate forward over array of thresholds starting from
5637 * current_threshold+1 and check if a threshold is crossed.
5638 * If none of thresholds above usage is crossed, we read
5639 * only one element of the array here.
5641 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5642 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5644 /* Update current_threshold */
5645 t
->current_threshold
= i
- 1;
5650 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5653 __mem_cgroup_threshold(memcg
, false);
5654 if (do_swap_account
)
5655 __mem_cgroup_threshold(memcg
, true);
5657 memcg
= parent_mem_cgroup(memcg
);
5661 static int compare_thresholds(const void *a
, const void *b
)
5663 const struct mem_cgroup_threshold
*_a
= a
;
5664 const struct mem_cgroup_threshold
*_b
= b
;
5666 if (_a
->threshold
> _b
->threshold
)
5669 if (_a
->threshold
< _b
->threshold
)
5675 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5677 struct mem_cgroup_eventfd_list
*ev
;
5679 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5680 eventfd_signal(ev
->eventfd
, 1);
5684 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5686 struct mem_cgroup
*iter
;
5688 for_each_mem_cgroup_tree(iter
, memcg
)
5689 mem_cgroup_oom_notify_cb(iter
);
5692 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5693 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
5695 struct mem_cgroup_thresholds
*thresholds
;
5696 struct mem_cgroup_threshold_ary
*new;
5697 u64 threshold
, usage
;
5700 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5704 mutex_lock(&memcg
->thresholds_lock
);
5707 thresholds
= &memcg
->thresholds
;
5708 else if (type
== _MEMSWAP
)
5709 thresholds
= &memcg
->memsw_thresholds
;
5713 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5715 /* Check if a threshold crossed before adding a new one */
5716 if (thresholds
->primary
)
5717 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5719 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5721 /* Allocate memory for new array of thresholds */
5722 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5730 /* Copy thresholds (if any) to new array */
5731 if (thresholds
->primary
) {
5732 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5733 sizeof(struct mem_cgroup_threshold
));
5736 /* Add new threshold */
5737 new->entries
[size
- 1].eventfd
= eventfd
;
5738 new->entries
[size
- 1].threshold
= threshold
;
5740 /* Sort thresholds. Registering of new threshold isn't time-critical */
5741 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5742 compare_thresholds
, NULL
);
5744 /* Find current threshold */
5745 new->current_threshold
= -1;
5746 for (i
= 0; i
< size
; i
++) {
5747 if (new->entries
[i
].threshold
<= usage
) {
5749 * new->current_threshold will not be used until
5750 * rcu_assign_pointer(), so it's safe to increment
5753 ++new->current_threshold
;
5758 /* Free old spare buffer and save old primary buffer as spare */
5759 kfree(thresholds
->spare
);
5760 thresholds
->spare
= thresholds
->primary
;
5762 rcu_assign_pointer(thresholds
->primary
, new);
5764 /* To be sure that nobody uses thresholds */
5768 mutex_unlock(&memcg
->thresholds_lock
);
5773 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5774 struct eventfd_ctx
*eventfd
, const char *args
)
5776 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
5779 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5780 struct eventfd_ctx
*eventfd
, const char *args
)
5782 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
5785 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5786 struct eventfd_ctx
*eventfd
, enum res_type type
)
5788 struct mem_cgroup_thresholds
*thresholds
;
5789 struct mem_cgroup_threshold_ary
*new;
5793 mutex_lock(&memcg
->thresholds_lock
);
5795 thresholds
= &memcg
->thresholds
;
5796 else if (type
== _MEMSWAP
)
5797 thresholds
= &memcg
->memsw_thresholds
;
5801 if (!thresholds
->primary
)
5804 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5806 /* Check if a threshold crossed before removing */
5807 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5809 /* Calculate new number of threshold */
5811 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5812 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5816 new = thresholds
->spare
;
5818 /* Set thresholds array to NULL if we don't have thresholds */
5827 /* Copy thresholds and find current threshold */
5828 new->current_threshold
= -1;
5829 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5830 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5833 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5834 if (new->entries
[j
].threshold
<= usage
) {
5836 * new->current_threshold will not be used
5837 * until rcu_assign_pointer(), so it's safe to increment
5840 ++new->current_threshold
;
5846 /* Swap primary and spare array */
5847 thresholds
->spare
= thresholds
->primary
;
5848 /* If all events are unregistered, free the spare array */
5850 kfree(thresholds
->spare
);
5851 thresholds
->spare
= NULL
;
5854 rcu_assign_pointer(thresholds
->primary
, new);
5856 /* To be sure that nobody uses thresholds */
5859 mutex_unlock(&memcg
->thresholds_lock
);
5862 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5863 struct eventfd_ctx
*eventfd
)
5865 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
5868 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5869 struct eventfd_ctx
*eventfd
)
5871 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
5874 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
5875 struct eventfd_ctx
*eventfd
, const char *args
)
5877 struct mem_cgroup_eventfd_list
*event
;
5879 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5883 spin_lock(&memcg_oom_lock
);
5885 event
->eventfd
= eventfd
;
5886 list_add(&event
->list
, &memcg
->oom_notify
);
5888 /* already in OOM ? */
5889 if (atomic_read(&memcg
->under_oom
))
5890 eventfd_signal(eventfd
, 1);
5891 spin_unlock(&memcg_oom_lock
);
5896 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
5897 struct eventfd_ctx
*eventfd
)
5899 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5901 spin_lock(&memcg_oom_lock
);
5903 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5904 if (ev
->eventfd
== eventfd
) {
5905 list_del(&ev
->list
);
5910 spin_unlock(&memcg_oom_lock
);
5913 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
5915 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
5917 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
5918 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
5922 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5923 struct cftype
*cft
, u64 val
)
5925 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5926 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5928 /* cannot set to root cgroup and only 0 and 1 are allowed */
5929 if (!parent
|| !((val
== 0) || (val
== 1)))
5932 mutex_lock(&memcg_create_mutex
);
5933 /* oom-kill-disable is a flag for subhierarchy. */
5934 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5935 mutex_unlock(&memcg_create_mutex
);
5938 memcg
->oom_kill_disable
= val
;
5940 memcg_oom_recover(memcg
);
5941 mutex_unlock(&memcg_create_mutex
);
5945 #ifdef CONFIG_MEMCG_KMEM
5946 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5950 memcg
->kmemcg_id
= -1;
5951 ret
= memcg_propagate_kmem(memcg
);
5955 return mem_cgroup_sockets_init(memcg
, ss
);
5958 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5960 mem_cgroup_sockets_destroy(memcg
);
5963 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5965 if (!memcg_kmem_is_active(memcg
))
5969 * kmem charges can outlive the cgroup. In the case of slab
5970 * pages, for instance, a page contain objects from various
5971 * processes. As we prevent from taking a reference for every
5972 * such allocation we have to be careful when doing uncharge
5973 * (see memcg_uncharge_kmem) and here during offlining.
5975 * The idea is that that only the _last_ uncharge which sees
5976 * the dead memcg will drop the last reference. An additional
5977 * reference is taken here before the group is marked dead
5978 * which is then paired with css_put during uncharge resp. here.
5980 * Although this might sound strange as this path is called from
5981 * css_offline() when the referencemight have dropped down to 0
5982 * and shouldn't be incremented anymore (css_tryget would fail)
5983 * we do not have other options because of the kmem allocations
5986 css_get(&memcg
->css
);
5988 memcg_kmem_mark_dead(memcg
);
5990 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5993 if (memcg_kmem_test_and_clear_dead(memcg
))
5994 css_put(&memcg
->css
);
5997 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
6002 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
6006 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
6012 * DO NOT USE IN NEW FILES.
6014 * "cgroup.event_control" implementation.
6016 * This is way over-engineered. It tries to support fully configurable
6017 * events for each user. Such level of flexibility is completely
6018 * unnecessary especially in the light of the planned unified hierarchy.
6020 * Please deprecate this and replace with something simpler if at all
6025 * Unregister event and free resources.
6027 * Gets called from workqueue.
6029 static void memcg_event_remove(struct work_struct
*work
)
6031 struct mem_cgroup_event
*event
=
6032 container_of(work
, struct mem_cgroup_event
, remove
);
6033 struct mem_cgroup
*memcg
= event
->memcg
;
6035 remove_wait_queue(event
->wqh
, &event
->wait
);
6037 event
->unregister_event(memcg
, event
->eventfd
);
6039 /* Notify userspace the event is going away. */
6040 eventfd_signal(event
->eventfd
, 1);
6042 eventfd_ctx_put(event
->eventfd
);
6044 css_put(&memcg
->css
);
6048 * Gets called on POLLHUP on eventfd when user closes it.
6050 * Called with wqh->lock held and interrupts disabled.
6052 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
6053 int sync
, void *key
)
6055 struct mem_cgroup_event
*event
=
6056 container_of(wait
, struct mem_cgroup_event
, wait
);
6057 struct mem_cgroup
*memcg
= event
->memcg
;
6058 unsigned long flags
= (unsigned long)key
;
6060 if (flags
& POLLHUP
) {
6062 * If the event has been detached at cgroup removal, we
6063 * can simply return knowing the other side will cleanup
6066 * We can't race against event freeing since the other
6067 * side will require wqh->lock via remove_wait_queue(),
6070 spin_lock(&memcg
->event_list_lock
);
6071 if (!list_empty(&event
->list
)) {
6072 list_del_init(&event
->list
);
6074 * We are in atomic context, but cgroup_event_remove()
6075 * may sleep, so we have to call it in workqueue.
6077 schedule_work(&event
->remove
);
6079 spin_unlock(&memcg
->event_list_lock
);
6085 static void memcg_event_ptable_queue_proc(struct file
*file
,
6086 wait_queue_head_t
*wqh
, poll_table
*pt
)
6088 struct mem_cgroup_event
*event
=
6089 container_of(pt
, struct mem_cgroup_event
, pt
);
6092 add_wait_queue(wqh
, &event
->wait
);
6096 * DO NOT USE IN NEW FILES.
6098 * Parse input and register new cgroup event handler.
6100 * Input must be in format '<event_fd> <control_fd> <args>'.
6101 * Interpretation of args is defined by control file implementation.
6103 static int memcg_write_event_control(struct cgroup_subsys_state
*css
,
6104 struct cftype
*cft
, const char *buffer
)
6106 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6107 struct mem_cgroup_event
*event
;
6108 struct cgroup_subsys_state
*cfile_css
;
6109 unsigned int efd
, cfd
;
6116 efd
= simple_strtoul(buffer
, &endp
, 10);
6121 cfd
= simple_strtoul(buffer
, &endp
, 10);
6122 if ((*endp
!= ' ') && (*endp
!= '\0'))
6126 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6130 event
->memcg
= memcg
;
6131 INIT_LIST_HEAD(&event
->list
);
6132 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
6133 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
6134 INIT_WORK(&event
->remove
, memcg_event_remove
);
6142 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
6143 if (IS_ERR(event
->eventfd
)) {
6144 ret
= PTR_ERR(event
->eventfd
);
6151 goto out_put_eventfd
;
6154 /* the process need read permission on control file */
6155 /* AV: shouldn't we check that it's been opened for read instead? */
6156 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
6161 * Determine the event callbacks and set them in @event. This used
6162 * to be done via struct cftype but cgroup core no longer knows
6163 * about these events. The following is crude but the whole thing
6164 * is for compatibility anyway.
6166 * DO NOT ADD NEW FILES.
6168 name
= cfile
.file
->f_dentry
->d_name
.name
;
6170 if (!strcmp(name
, "memory.usage_in_bytes")) {
6171 event
->register_event
= mem_cgroup_usage_register_event
;
6172 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
6173 } else if (!strcmp(name
, "memory.oom_control")) {
6174 event
->register_event
= mem_cgroup_oom_register_event
;
6175 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
6176 } else if (!strcmp(name
, "memory.pressure_level")) {
6177 event
->register_event
= vmpressure_register_event
;
6178 event
->unregister_event
= vmpressure_unregister_event
;
6179 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
6180 event
->register_event
= memsw_cgroup_usage_register_event
;
6181 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
6188 * Verify @cfile should belong to @css. Also, remaining events are
6189 * automatically removed on cgroup destruction but the removal is
6190 * asynchronous, so take an extra ref on @css.
6195 cfile_css
= css_from_dir(cfile
.file
->f_dentry
->d_parent
,
6196 &mem_cgroup_subsys
);
6197 if (cfile_css
== css
&& css_tryget(css
))
6204 ret
= event
->register_event(memcg
, event
->eventfd
, buffer
);
6208 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
6210 spin_lock(&memcg
->event_list_lock
);
6211 list_add(&event
->list
, &memcg
->event_list
);
6212 spin_unlock(&memcg
->event_list_lock
);
6224 eventfd_ctx_put(event
->eventfd
);
6233 static struct cftype mem_cgroup_files
[] = {
6235 .name
= "usage_in_bytes",
6236 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
6237 .read_u64
= mem_cgroup_read_u64
,
6240 .name
= "max_usage_in_bytes",
6241 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
6242 .trigger
= mem_cgroup_reset
,
6243 .read_u64
= mem_cgroup_read_u64
,
6246 .name
= "limit_in_bytes",
6247 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
6248 .write_string
= mem_cgroup_write
,
6249 .read_u64
= mem_cgroup_read_u64
,
6252 .name
= "soft_limit_in_bytes",
6253 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
6254 .write_string
= mem_cgroup_write
,
6255 .read_u64
= mem_cgroup_read_u64
,
6259 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
6260 .trigger
= mem_cgroup_reset
,
6261 .read_u64
= mem_cgroup_read_u64
,
6265 .seq_show
= memcg_stat_show
,
6268 .name
= "force_empty",
6269 .trigger
= mem_cgroup_force_empty_write
,
6272 .name
= "use_hierarchy",
6273 .flags
= CFTYPE_INSANE
,
6274 .write_u64
= mem_cgroup_hierarchy_write
,
6275 .read_u64
= mem_cgroup_hierarchy_read
,
6278 .name
= "cgroup.event_control", /* XXX: for compat */
6279 .write_string
= memcg_write_event_control
,
6280 .flags
= CFTYPE_NO_PREFIX
,
6284 .name
= "swappiness",
6285 .read_u64
= mem_cgroup_swappiness_read
,
6286 .write_u64
= mem_cgroup_swappiness_write
,
6289 .name
= "move_charge_at_immigrate",
6290 .read_u64
= mem_cgroup_move_charge_read
,
6291 .write_u64
= mem_cgroup_move_charge_write
,
6294 .name
= "oom_control",
6295 .seq_show
= mem_cgroup_oom_control_read
,
6296 .write_u64
= mem_cgroup_oom_control_write
,
6297 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6300 .name
= "pressure_level",
6304 .name
= "numa_stat",
6305 .seq_show
= memcg_numa_stat_show
,
6308 #ifdef CONFIG_MEMCG_KMEM
6310 .name
= "kmem.limit_in_bytes",
6311 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6312 .write_string
= mem_cgroup_write
,
6313 .read_u64
= mem_cgroup_read_u64
,
6316 .name
= "kmem.usage_in_bytes",
6317 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6318 .read_u64
= mem_cgroup_read_u64
,
6321 .name
= "kmem.failcnt",
6322 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6323 .trigger
= mem_cgroup_reset
,
6324 .read_u64
= mem_cgroup_read_u64
,
6327 .name
= "kmem.max_usage_in_bytes",
6328 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6329 .trigger
= mem_cgroup_reset
,
6330 .read_u64
= mem_cgroup_read_u64
,
6332 #ifdef CONFIG_SLABINFO
6334 .name
= "kmem.slabinfo",
6335 .seq_show
= mem_cgroup_slabinfo_read
,
6339 { }, /* terminate */
6342 #ifdef CONFIG_MEMCG_SWAP
6343 static struct cftype memsw_cgroup_files
[] = {
6345 .name
= "memsw.usage_in_bytes",
6346 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6347 .read_u64
= mem_cgroup_read_u64
,
6350 .name
= "memsw.max_usage_in_bytes",
6351 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6352 .trigger
= mem_cgroup_reset
,
6353 .read_u64
= mem_cgroup_read_u64
,
6356 .name
= "memsw.limit_in_bytes",
6357 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6358 .write_string
= mem_cgroup_write
,
6359 .read_u64
= mem_cgroup_read_u64
,
6362 .name
= "memsw.failcnt",
6363 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6364 .trigger
= mem_cgroup_reset
,
6365 .read_u64
= mem_cgroup_read_u64
,
6367 { }, /* terminate */
6370 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6372 struct mem_cgroup_per_node
*pn
;
6373 struct mem_cgroup_per_zone
*mz
;
6374 int zone
, tmp
= node
;
6376 * This routine is called against possible nodes.
6377 * But it's BUG to call kmalloc() against offline node.
6379 * TODO: this routine can waste much memory for nodes which will
6380 * never be onlined. It's better to use memory hotplug callback
6383 if (!node_state(node
, N_NORMAL_MEMORY
))
6385 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6389 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6390 mz
= &pn
->zoneinfo
[zone
];
6391 lruvec_init(&mz
->lruvec
);
6392 mz
->usage_in_excess
= 0;
6393 mz
->on_tree
= false;
6396 memcg
->nodeinfo
[node
] = pn
;
6400 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6402 kfree(memcg
->nodeinfo
[node
]);
6405 static struct mem_cgroup
*mem_cgroup_alloc(void)
6407 struct mem_cgroup
*memcg
;
6408 size_t size
= memcg_size();
6410 /* Can be very big if nr_node_ids is very big */
6411 if (size
< PAGE_SIZE
)
6412 memcg
= kzalloc(size
, GFP_KERNEL
);
6414 memcg
= vzalloc(size
);
6419 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6422 spin_lock_init(&memcg
->pcp_counter_lock
);
6426 if (size
< PAGE_SIZE
)
6434 * At destroying mem_cgroup, references from swap_cgroup can remain.
6435 * (scanning all at force_empty is too costly...)
6437 * Instead of clearing all references at force_empty, we remember
6438 * the number of reference from swap_cgroup and free mem_cgroup when
6439 * it goes down to 0.
6441 * Removal of cgroup itself succeeds regardless of refs from swap.
6444 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6447 size_t size
= memcg_size();
6449 mem_cgroup_remove_from_trees(memcg
);
6452 free_mem_cgroup_per_zone_info(memcg
, node
);
6454 free_percpu(memcg
->stat
);
6457 * We need to make sure that (at least for now), the jump label
6458 * destruction code runs outside of the cgroup lock. This is because
6459 * get_online_cpus(), which is called from the static_branch update,
6460 * can't be called inside the cgroup_lock. cpusets are the ones
6461 * enforcing this dependency, so if they ever change, we might as well.
6463 * schedule_work() will guarantee this happens. Be careful if you need
6464 * to move this code around, and make sure it is outside
6467 disarm_static_keys(memcg
);
6468 if (size
< PAGE_SIZE
)
6475 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6477 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6479 if (!memcg
->res
.parent
)
6481 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6483 EXPORT_SYMBOL(parent_mem_cgroup
);
6485 static void __init
mem_cgroup_soft_limit_tree_init(void)
6487 struct mem_cgroup_tree_per_node
*rtpn
;
6488 struct mem_cgroup_tree_per_zone
*rtpz
;
6489 int tmp
, node
, zone
;
6491 for_each_node(node
) {
6493 if (!node_state(node
, N_NORMAL_MEMORY
))
6495 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6498 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6500 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6501 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6502 rtpz
->rb_root
= RB_ROOT
;
6503 spin_lock_init(&rtpz
->lock
);
6508 static struct cgroup_subsys_state
* __ref
6509 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6511 struct mem_cgroup
*memcg
;
6512 long error
= -ENOMEM
;
6515 memcg
= mem_cgroup_alloc();
6517 return ERR_PTR(error
);
6520 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6524 if (parent_css
== NULL
) {
6525 root_mem_cgroup
= memcg
;
6526 res_counter_init(&memcg
->res
, NULL
);
6527 res_counter_init(&memcg
->memsw
, NULL
);
6528 res_counter_init(&memcg
->kmem
, NULL
);
6531 memcg
->last_scanned_node
= MAX_NUMNODES
;
6532 INIT_LIST_HEAD(&memcg
->oom_notify
);
6533 memcg
->move_charge_at_immigrate
= 0;
6534 mutex_init(&memcg
->thresholds_lock
);
6535 spin_lock_init(&memcg
->move_lock
);
6536 vmpressure_init(&memcg
->vmpressure
);
6537 INIT_LIST_HEAD(&memcg
->event_list
);
6538 spin_lock_init(&memcg
->event_list_lock
);
6543 __mem_cgroup_free(memcg
);
6544 return ERR_PTR(error
);
6548 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6550 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6551 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6554 if (css
->cgroup
->id
> MEM_CGROUP_ID_MAX
)
6560 mutex_lock(&memcg_create_mutex
);
6562 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6563 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6564 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6566 if (parent
->use_hierarchy
) {
6567 res_counter_init(&memcg
->res
, &parent
->res
);
6568 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6569 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6572 * No need to take a reference to the parent because cgroup
6573 * core guarantees its existence.
6576 res_counter_init(&memcg
->res
, NULL
);
6577 res_counter_init(&memcg
->memsw
, NULL
);
6578 res_counter_init(&memcg
->kmem
, NULL
);
6580 * Deeper hierachy with use_hierarchy == false doesn't make
6581 * much sense so let cgroup subsystem know about this
6582 * unfortunate state in our controller.
6584 if (parent
!= root_mem_cgroup
)
6585 mem_cgroup_subsys
.broken_hierarchy
= true;
6588 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6589 mutex_unlock(&memcg_create_mutex
);
6594 * Announce all parents that a group from their hierarchy is gone.
6596 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6598 struct mem_cgroup
*parent
= memcg
;
6600 while ((parent
= parent_mem_cgroup(parent
)))
6601 mem_cgroup_iter_invalidate(parent
);
6604 * if the root memcg is not hierarchical we have to check it
6607 if (!root_mem_cgroup
->use_hierarchy
)
6608 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6611 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6613 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6614 struct mem_cgroup_event
*event
, *tmp
;
6617 * Unregister events and notify userspace.
6618 * Notify userspace about cgroup removing only after rmdir of cgroup
6619 * directory to avoid race between userspace and kernelspace.
6621 spin_lock(&memcg
->event_list_lock
);
6622 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
6623 list_del_init(&event
->list
);
6624 schedule_work(&event
->remove
);
6626 spin_unlock(&memcg
->event_list_lock
);
6628 kmem_cgroup_css_offline(memcg
);
6630 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6631 mem_cgroup_reparent_charges(memcg
);
6632 mem_cgroup_destroy_all_caches(memcg
);
6633 vmpressure_cleanup(&memcg
->vmpressure
);
6636 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6638 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6640 * XXX: css_offline() would be where we should reparent all
6641 * memory to prepare the cgroup for destruction. However,
6642 * memcg does not do css_tryget() and res_counter charging
6643 * under the same RCU lock region, which means that charging
6644 * could race with offlining. Offlining only happens to
6645 * cgroups with no tasks in them but charges can show up
6646 * without any tasks from the swapin path when the target
6647 * memcg is looked up from the swapout record and not from the
6648 * current task as it usually is. A race like this can leak
6649 * charges and put pages with stale cgroup pointers into
6653 * lookup_swap_cgroup_id()
6655 * mem_cgroup_lookup()
6658 * disable css_tryget()
6661 * reparent_charges()
6662 * res_counter_charge()
6665 * pc->mem_cgroup = dead memcg
6668 * The bulk of the charges are still moved in offline_css() to
6669 * avoid pinning a lot of pages in case a long-term reference
6670 * like a swapout record is deferring the css_free() to long
6671 * after offlining. But this makes sure we catch any charges
6672 * made after offlining:
6674 mem_cgroup_reparent_charges(memcg
);
6676 memcg_destroy_kmem(memcg
);
6677 __mem_cgroup_free(memcg
);
6681 /* Handlers for move charge at task migration. */
6682 #define PRECHARGE_COUNT_AT_ONCE 256
6683 static int mem_cgroup_do_precharge(unsigned long count
)
6686 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6687 struct mem_cgroup
*memcg
= mc
.to
;
6689 if (mem_cgroup_is_root(memcg
)) {
6690 mc
.precharge
+= count
;
6691 /* we don't need css_get for root */
6694 /* try to charge at once */
6696 struct res_counter
*dummy
;
6698 * "memcg" cannot be under rmdir() because we've already checked
6699 * by cgroup_lock_live_cgroup() that it is not removed and we
6700 * are still under the same cgroup_mutex. So we can postpone
6703 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6705 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6706 PAGE_SIZE
* count
, &dummy
)) {
6707 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6710 mc
.precharge
+= count
;
6714 /* fall back to one by one charge */
6716 if (signal_pending(current
)) {
6720 if (!batch_count
--) {
6721 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6724 ret
= __mem_cgroup_try_charge(NULL
,
6725 GFP_KERNEL
, 1, &memcg
, false);
6727 /* mem_cgroup_clear_mc() will do uncharge later */
6735 * get_mctgt_type - get target type of moving charge
6736 * @vma: the vma the pte to be checked belongs
6737 * @addr: the address corresponding to the pte to be checked
6738 * @ptent: the pte to be checked
6739 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6742 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6743 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6744 * move charge. if @target is not NULL, the page is stored in target->page
6745 * with extra refcnt got(Callers should handle it).
6746 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6747 * target for charge migration. if @target is not NULL, the entry is stored
6750 * Called with pte lock held.
6757 enum mc_target_type
{
6763 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6764 unsigned long addr
, pte_t ptent
)
6766 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6768 if (!page
|| !page_mapped(page
))
6770 if (PageAnon(page
)) {
6771 /* we don't move shared anon */
6774 } else if (!move_file())
6775 /* we ignore mapcount for file pages */
6777 if (!get_page_unless_zero(page
))
6784 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6785 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6787 struct page
*page
= NULL
;
6788 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6790 if (!move_anon() || non_swap_entry(ent
))
6793 * Because lookup_swap_cache() updates some statistics counter,
6794 * we call find_get_page() with swapper_space directly.
6796 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6797 if (do_swap_account
)
6798 entry
->val
= ent
.val
;
6803 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6804 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6810 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6811 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6813 struct page
*page
= NULL
;
6814 struct address_space
*mapping
;
6817 if (!vma
->vm_file
) /* anonymous vma */
6822 mapping
= vma
->vm_file
->f_mapping
;
6823 if (pte_none(ptent
))
6824 pgoff
= linear_page_index(vma
, addr
);
6825 else /* pte_file(ptent) is true */
6826 pgoff
= pte_to_pgoff(ptent
);
6828 /* page is moved even if it's not RSS of this task(page-faulted). */
6829 page
= find_get_page(mapping
, pgoff
);
6832 /* shmem/tmpfs may report page out on swap: account for that too. */
6833 if (radix_tree_exceptional_entry(page
)) {
6834 swp_entry_t swap
= radix_to_swp_entry(page
);
6835 if (do_swap_account
)
6837 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6843 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6844 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6846 struct page
*page
= NULL
;
6847 struct page_cgroup
*pc
;
6848 enum mc_target_type ret
= MC_TARGET_NONE
;
6849 swp_entry_t ent
= { .val
= 0 };
6851 if (pte_present(ptent
))
6852 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6853 else if (is_swap_pte(ptent
))
6854 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6855 else if (pte_none(ptent
) || pte_file(ptent
))
6856 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6858 if (!page
&& !ent
.val
)
6861 pc
= lookup_page_cgroup(page
);
6863 * Do only loose check w/o page_cgroup lock.
6864 * mem_cgroup_move_account() checks the pc is valid or not under
6867 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6868 ret
= MC_TARGET_PAGE
;
6870 target
->page
= page
;
6872 if (!ret
|| !target
)
6875 /* There is a swap entry and a page doesn't exist or isn't charged */
6876 if (ent
.val
&& !ret
&&
6877 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6878 ret
= MC_TARGET_SWAP
;
6885 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6887 * We don't consider swapping or file mapped pages because THP does not
6888 * support them for now.
6889 * Caller should make sure that pmd_trans_huge(pmd) is true.
6891 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6892 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6894 struct page
*page
= NULL
;
6895 struct page_cgroup
*pc
;
6896 enum mc_target_type ret
= MC_TARGET_NONE
;
6898 page
= pmd_page(pmd
);
6899 VM_BUG_ON(!page
|| !PageHead(page
));
6902 pc
= lookup_page_cgroup(page
);
6903 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6904 ret
= MC_TARGET_PAGE
;
6907 target
->page
= page
;
6913 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6914 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6916 return MC_TARGET_NONE
;
6920 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6921 unsigned long addr
, unsigned long end
,
6922 struct mm_walk
*walk
)
6924 struct vm_area_struct
*vma
= walk
->private;
6928 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6929 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6930 mc
.precharge
+= HPAGE_PMD_NR
;
6935 if (pmd_trans_unstable(pmd
))
6937 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6938 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6939 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6940 mc
.precharge
++; /* increment precharge temporarily */
6941 pte_unmap_unlock(pte
- 1, ptl
);
6947 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6949 unsigned long precharge
;
6950 struct vm_area_struct
*vma
;
6952 down_read(&mm
->mmap_sem
);
6953 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6954 struct mm_walk mem_cgroup_count_precharge_walk
= {
6955 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6959 if (is_vm_hugetlb_page(vma
))
6961 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6962 &mem_cgroup_count_precharge_walk
);
6964 up_read(&mm
->mmap_sem
);
6966 precharge
= mc
.precharge
;
6972 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6974 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6976 VM_BUG_ON(mc
.moving_task
);
6977 mc
.moving_task
= current
;
6978 return mem_cgroup_do_precharge(precharge
);
6981 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6982 static void __mem_cgroup_clear_mc(void)
6984 struct mem_cgroup
*from
= mc
.from
;
6985 struct mem_cgroup
*to
= mc
.to
;
6988 /* we must uncharge all the leftover precharges from mc.to */
6990 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6994 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6995 * we must uncharge here.
6997 if (mc
.moved_charge
) {
6998 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6999 mc
.moved_charge
= 0;
7001 /* we must fixup refcnts and charges */
7002 if (mc
.moved_swap
) {
7003 /* uncharge swap account from the old cgroup */
7004 if (!mem_cgroup_is_root(mc
.from
))
7005 res_counter_uncharge(&mc
.from
->memsw
,
7006 PAGE_SIZE
* mc
.moved_swap
);
7008 for (i
= 0; i
< mc
.moved_swap
; i
++)
7009 css_put(&mc
.from
->css
);
7011 if (!mem_cgroup_is_root(mc
.to
)) {
7013 * we charged both to->res and to->memsw, so we should
7016 res_counter_uncharge(&mc
.to
->res
,
7017 PAGE_SIZE
* mc
.moved_swap
);
7019 /* we've already done css_get(mc.to) */
7022 memcg_oom_recover(from
);
7023 memcg_oom_recover(to
);
7024 wake_up_all(&mc
.waitq
);
7027 static void mem_cgroup_clear_mc(void)
7029 struct mem_cgroup
*from
= mc
.from
;
7032 * we must clear moving_task before waking up waiters at the end of
7035 mc
.moving_task
= NULL
;
7036 __mem_cgroup_clear_mc();
7037 spin_lock(&mc
.lock
);
7040 spin_unlock(&mc
.lock
);
7041 mem_cgroup_end_move(from
);
7044 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7045 struct cgroup_taskset
*tset
)
7047 struct task_struct
*p
= cgroup_taskset_first(tset
);
7049 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
7050 unsigned long move_charge_at_immigrate
;
7053 * We are now commited to this value whatever it is. Changes in this
7054 * tunable will only affect upcoming migrations, not the current one.
7055 * So we need to save it, and keep it going.
7057 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
7058 if (move_charge_at_immigrate
) {
7059 struct mm_struct
*mm
;
7060 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
7062 VM_BUG_ON(from
== memcg
);
7064 mm
= get_task_mm(p
);
7067 /* We move charges only when we move a owner of the mm */
7068 if (mm
->owner
== p
) {
7071 VM_BUG_ON(mc
.precharge
);
7072 VM_BUG_ON(mc
.moved_charge
);
7073 VM_BUG_ON(mc
.moved_swap
);
7074 mem_cgroup_start_move(from
);
7075 spin_lock(&mc
.lock
);
7078 mc
.immigrate_flags
= move_charge_at_immigrate
;
7079 spin_unlock(&mc
.lock
);
7080 /* We set mc.moving_task later */
7082 ret
= mem_cgroup_precharge_mc(mm
);
7084 mem_cgroup_clear_mc();
7091 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
7092 struct cgroup_taskset
*tset
)
7094 mem_cgroup_clear_mc();
7097 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
7098 unsigned long addr
, unsigned long end
,
7099 struct mm_walk
*walk
)
7102 struct vm_area_struct
*vma
= walk
->private;
7105 enum mc_target_type target_type
;
7106 union mc_target target
;
7108 struct page_cgroup
*pc
;
7111 * We don't take compound_lock() here but no race with splitting thp
7113 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7114 * under splitting, which means there's no concurrent thp split,
7115 * - if another thread runs into split_huge_page() just after we
7116 * entered this if-block, the thread must wait for page table lock
7117 * to be unlocked in __split_huge_page_splitting(), where the main
7118 * part of thp split is not executed yet.
7120 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
7121 if (mc
.precharge
< HPAGE_PMD_NR
) {
7125 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
7126 if (target_type
== MC_TARGET_PAGE
) {
7128 if (!isolate_lru_page(page
)) {
7129 pc
= lookup_page_cgroup(page
);
7130 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
7131 pc
, mc
.from
, mc
.to
)) {
7132 mc
.precharge
-= HPAGE_PMD_NR
;
7133 mc
.moved_charge
+= HPAGE_PMD_NR
;
7135 putback_lru_page(page
);
7143 if (pmd_trans_unstable(pmd
))
7146 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
7147 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
7148 pte_t ptent
= *(pte
++);
7154 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
7155 case MC_TARGET_PAGE
:
7157 if (isolate_lru_page(page
))
7159 pc
= lookup_page_cgroup(page
);
7160 if (!mem_cgroup_move_account(page
, 1, pc
,
7163 /* we uncharge from mc.from later. */
7166 putback_lru_page(page
);
7167 put
: /* get_mctgt_type() gets the page */
7170 case MC_TARGET_SWAP
:
7172 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
7174 /* we fixup refcnts and charges later. */
7182 pte_unmap_unlock(pte
- 1, ptl
);
7187 * We have consumed all precharges we got in can_attach().
7188 * We try charge one by one, but don't do any additional
7189 * charges to mc.to if we have failed in charge once in attach()
7192 ret
= mem_cgroup_do_precharge(1);
7200 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
7202 struct vm_area_struct
*vma
;
7204 lru_add_drain_all();
7206 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
7208 * Someone who are holding the mmap_sem might be waiting in
7209 * waitq. So we cancel all extra charges, wake up all waiters,
7210 * and retry. Because we cancel precharges, we might not be able
7211 * to move enough charges, but moving charge is a best-effort
7212 * feature anyway, so it wouldn't be a big problem.
7214 __mem_cgroup_clear_mc();
7218 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
7220 struct mm_walk mem_cgroup_move_charge_walk
= {
7221 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
7225 if (is_vm_hugetlb_page(vma
))
7227 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
7228 &mem_cgroup_move_charge_walk
);
7231 * means we have consumed all precharges and failed in
7232 * doing additional charge. Just abandon here.
7236 up_read(&mm
->mmap_sem
);
7239 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
7240 struct cgroup_taskset
*tset
)
7242 struct task_struct
*p
= cgroup_taskset_first(tset
);
7243 struct mm_struct
*mm
= get_task_mm(p
);
7247 mem_cgroup_move_charge(mm
);
7251 mem_cgroup_clear_mc();
7253 #else /* !CONFIG_MMU */
7254 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7255 struct cgroup_taskset
*tset
)
7259 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
7260 struct cgroup_taskset
*tset
)
7263 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
7264 struct cgroup_taskset
*tset
)
7270 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7271 * to verify sane_behavior flag on each mount attempt.
7273 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
7276 * use_hierarchy is forced with sane_behavior. cgroup core
7277 * guarantees that @root doesn't have any children, so turning it
7278 * on for the root memcg is enough.
7280 if (cgroup_sane_behavior(root_css
->cgroup
))
7281 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
7284 struct cgroup_subsys mem_cgroup_subsys
= {
7286 .subsys_id
= mem_cgroup_subsys_id
,
7287 .css_alloc
= mem_cgroup_css_alloc
,
7288 .css_online
= mem_cgroup_css_online
,
7289 .css_offline
= mem_cgroup_css_offline
,
7290 .css_free
= mem_cgroup_css_free
,
7291 .can_attach
= mem_cgroup_can_attach
,
7292 .cancel_attach
= mem_cgroup_cancel_attach
,
7293 .attach
= mem_cgroup_move_task
,
7294 .bind
= mem_cgroup_bind
,
7295 .base_cftypes
= mem_cgroup_files
,
7299 #ifdef CONFIG_MEMCG_SWAP
7300 static int __init
enable_swap_account(char *s
)
7302 if (!strcmp(s
, "1"))
7303 really_do_swap_account
= 1;
7304 else if (!strcmp(s
, "0"))
7305 really_do_swap_account
= 0;
7308 __setup("swapaccount=", enable_swap_account
);
7310 static void __init
memsw_file_init(void)
7312 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
7315 static void __init
enable_swap_cgroup(void)
7317 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7318 do_swap_account
= 1;
7324 static void __init
enable_swap_cgroup(void)
7330 * subsys_initcall() for memory controller.
7332 * Some parts like hotcpu_notifier() have to be initialized from this context
7333 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7334 * everything that doesn't depend on a specific mem_cgroup structure should
7335 * be initialized from here.
7337 static int __init
mem_cgroup_init(void)
7339 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7340 enable_swap_cgroup();
7341 mem_cgroup_soft_limit_tree_init();
7345 subsys_initcall(mem_cgroup_init
);