1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
70 EXPORT_SYMBOL(memory_cgrp_subsys
);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly
;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata
= 1;
83 static int really_do_swap_account __initdata
= 0;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names
[] = {
100 enum mem_cgroup_events_index
{
101 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS
,
108 static const char * const mem_cgroup_events_names
[] = {
115 static const char * const mem_cgroup_lru_names
[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target
{
130 MEM_CGROUP_TARGET_THRESH
,
131 MEM_CGROUP_TARGET_SOFTLIMIT
,
132 MEM_CGROUP_TARGET_NUMAINFO
,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu
{
140 long count
[MEM_CGROUP_STAT_NSTATS
];
141 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
142 unsigned long nr_page_events
;
143 unsigned long targets
[MEM_CGROUP_NTARGETS
];
146 struct mem_cgroup_reclaim_iter
{
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup
*last_visited
;
154 /* scan generation, increased every round-trip */
155 unsigned int generation
;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone
{
162 struct lruvec lruvec
;
163 unsigned long lru_size
[NR_LRU_LISTS
];
165 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
167 struct rb_node tree_node
; /* RB tree node */
168 unsigned long long usage_in_excess
;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node
{
176 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone
{
185 struct rb_root rb_root
;
189 struct mem_cgroup_tree_per_node
{
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
193 struct mem_cgroup_tree
{
194 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
199 struct mem_cgroup_threshold
{
200 struct eventfd_ctx
*eventfd
;
205 struct mem_cgroup_threshold_ary
{
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold
;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries
[0];
214 struct mem_cgroup_thresholds
{
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary
*primary
;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary
*spare
;
226 struct mem_cgroup_eventfd_list
{
227 struct list_head list
;
228 struct eventfd_ctx
*eventfd
;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event
{
236 * memcg which the event belongs to.
238 struct mem_cgroup
*memcg
;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx
*eventfd
;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list
;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event
)(struct mem_cgroup
*memcg
,
253 struct eventfd_ctx
*eventfd
, const char *args
);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event
)(struct mem_cgroup
*memcg
,
260 struct eventfd_ctx
*eventfd
);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t
*wqh
;
268 struct work_struct remove
;
271 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
272 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css
;
288 * the counter to account for memory usage
290 struct res_counter res
;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure
;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw
;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem
;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups
;
315 /* OOM-Killer disable */
316 int oom_kill_disable
;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum
;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock
;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds
;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds
;
330 /* For oom notifier event fd */
331 struct list_head oom_notify
;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate
;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account
;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock
;
347 struct mem_cgroup_stat_cpu __percpu
*stat
;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base
;
353 spinlock_t pcp_counter_lock
;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem
;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list. per-memcg */
361 struct list_head memcg_slab_caches
;
362 /* Not a spinlock, we can take a lot of time walking the list */
363 struct mutex slab_caches_mutex
;
364 /* Index in the kmem_cache->memcg_params->memcg_caches array */
368 int last_scanned_node
;
370 nodemask_t scan_nodes
;
371 atomic_t numainfo_events
;
372 atomic_t numainfo_updating
;
375 /* List of events which userspace want to receive */
376 struct list_head event_list
;
377 spinlock_t event_list_lock
;
379 struct mem_cgroup_per_node
*nodeinfo
[0];
380 /* WARNING: nodeinfo must be the last member here */
383 /* internal only representation about the status of kmem accounting. */
385 KMEM_ACCOUNTED_ACTIVE
, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
392 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
395 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
400 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
403 * Our caller must use css_get() first, because memcg_uncharge_kmem()
404 * will call css_put() if it sees the memcg is dead.
407 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
408 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
411 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
413 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
414 &memcg
->kmem_account_flags
);
418 /* Stuffs for move charges at task migration. */
420 * Types of charges to be moved. "move_charge_at_immitgrate" and
421 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
424 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
425 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
429 /* "mc" and its members are protected by cgroup_mutex */
430 static struct move_charge_struct
{
431 spinlock_t lock
; /* for from, to */
432 struct mem_cgroup
*from
;
433 struct mem_cgroup
*to
;
434 unsigned long immigrate_flags
;
435 unsigned long precharge
;
436 unsigned long moved_charge
;
437 unsigned long moved_swap
;
438 struct task_struct
*moving_task
; /* a task moving charges */
439 wait_queue_head_t waitq
; /* a waitq for other context */
441 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
442 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
445 static bool move_anon(void)
447 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
450 static bool move_file(void)
452 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
456 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
457 * limit reclaim to prevent infinite loops, if they ever occur.
459 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
460 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
463 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
464 MEM_CGROUP_CHARGE_TYPE_ANON
,
465 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
466 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
470 /* for encoding cft->private value on file */
478 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
479 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
480 #define MEMFILE_ATTR(val) ((val) & 0xffff)
481 /* Used for OOM nofiier */
482 #define OOM_CONTROL (0)
485 * Reclaim flags for mem_cgroup_hierarchical_reclaim
487 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
488 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
489 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
490 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
493 * The memcg_create_mutex will be held whenever a new cgroup is created.
494 * As a consequence, any change that needs to protect against new child cgroups
495 * appearing has to hold it as well.
497 static DEFINE_MUTEX(memcg_create_mutex
);
499 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
501 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
504 /* Some nice accessors for the vmpressure. */
505 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
508 memcg
= root_mem_cgroup
;
509 return &memcg
->vmpressure
;
512 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
514 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
517 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
519 return (memcg
== root_mem_cgroup
);
523 * We restrict the id in the range of [1, 65535], so it can fit into
526 #define MEM_CGROUP_ID_MAX USHRT_MAX
528 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
531 * The ID of the root cgroup is 0, but memcg treat 0 as an
532 * invalid ID, so we return (cgroup_id + 1).
534 return memcg
->css
.cgroup
->id
+ 1;
537 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
539 struct cgroup_subsys_state
*css
;
541 css
= css_from_id(id
- 1, &memory_cgrp_subsys
);
542 return mem_cgroup_from_css(css
);
545 /* Writing them here to avoid exposing memcg's inner layout */
546 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
548 void sock_update_memcg(struct sock
*sk
)
550 if (mem_cgroup_sockets_enabled
) {
551 struct mem_cgroup
*memcg
;
552 struct cg_proto
*cg_proto
;
554 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
556 /* Socket cloning can throw us here with sk_cgrp already
557 * filled. It won't however, necessarily happen from
558 * process context. So the test for root memcg given
559 * the current task's memcg won't help us in this case.
561 * Respecting the original socket's memcg is a better
562 * decision in this case.
565 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
566 css_get(&sk
->sk_cgrp
->memcg
->css
);
571 memcg
= mem_cgroup_from_task(current
);
572 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
573 if (!mem_cgroup_is_root(memcg
) &&
574 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
575 sk
->sk_cgrp
= cg_proto
;
580 EXPORT_SYMBOL(sock_update_memcg
);
582 void sock_release_memcg(struct sock
*sk
)
584 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
585 struct mem_cgroup
*memcg
;
586 WARN_ON(!sk
->sk_cgrp
->memcg
);
587 memcg
= sk
->sk_cgrp
->memcg
;
588 css_put(&sk
->sk_cgrp
->memcg
->css
);
592 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
594 if (!memcg
|| mem_cgroup_is_root(memcg
))
597 return &memcg
->tcp_mem
;
599 EXPORT_SYMBOL(tcp_proto_cgroup
);
601 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
603 if (!memcg_proto_activated(&memcg
->tcp_mem
))
605 static_key_slow_dec(&memcg_socket_limit_enabled
);
608 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
613 #ifdef CONFIG_MEMCG_KMEM
615 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
616 * The main reason for not using cgroup id for this:
617 * this works better in sparse environments, where we have a lot of memcgs,
618 * but only a few kmem-limited. Or also, if we have, for instance, 200
619 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
620 * 200 entry array for that.
622 * The current size of the caches array is stored in
623 * memcg_limited_groups_array_size. It will double each time we have to
626 static DEFINE_IDA(kmem_limited_groups
);
627 int memcg_limited_groups_array_size
;
630 * MIN_SIZE is different than 1, because we would like to avoid going through
631 * the alloc/free process all the time. In a small machine, 4 kmem-limited
632 * cgroups is a reasonable guess. In the future, it could be a parameter or
633 * tunable, but that is strictly not necessary.
635 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
636 * this constant directly from cgroup, but it is understandable that this is
637 * better kept as an internal representation in cgroup.c. In any case, the
638 * cgrp_id space is not getting any smaller, and we don't have to necessarily
639 * increase ours as well if it increases.
641 #define MEMCG_CACHES_MIN_SIZE 4
642 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
645 * A lot of the calls to the cache allocation functions are expected to be
646 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
647 * conditional to this static branch, we'll have to allow modules that does
648 * kmem_cache_alloc and the such to see this symbol as well
650 struct static_key memcg_kmem_enabled_key
;
651 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
653 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
655 if (memcg_kmem_is_active(memcg
)) {
656 static_key_slow_dec(&memcg_kmem_enabled_key
);
657 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
660 * This check can't live in kmem destruction function,
661 * since the charges will outlive the cgroup
663 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
666 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
669 #endif /* CONFIG_MEMCG_KMEM */
671 static void disarm_static_keys(struct mem_cgroup
*memcg
)
673 disarm_sock_keys(memcg
);
674 disarm_kmem_keys(memcg
);
677 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
679 static struct mem_cgroup_per_zone
*
680 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
682 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
683 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
686 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
691 static struct mem_cgroup_per_zone
*
692 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
694 int nid
= page_to_nid(page
);
695 int zid
= page_zonenum(page
);
697 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
700 static struct mem_cgroup_tree_per_zone
*
701 soft_limit_tree_node_zone(int nid
, int zid
)
703 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
706 static struct mem_cgroup_tree_per_zone
*
707 soft_limit_tree_from_page(struct page
*page
)
709 int nid
= page_to_nid(page
);
710 int zid
= page_zonenum(page
);
712 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
716 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
717 struct mem_cgroup_per_zone
*mz
,
718 struct mem_cgroup_tree_per_zone
*mctz
,
719 unsigned long long new_usage_in_excess
)
721 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
722 struct rb_node
*parent
= NULL
;
723 struct mem_cgroup_per_zone
*mz_node
;
728 mz
->usage_in_excess
= new_usage_in_excess
;
729 if (!mz
->usage_in_excess
)
733 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
735 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
738 * We can't avoid mem cgroups that are over their soft
739 * limit by the same amount
741 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
744 rb_link_node(&mz
->tree_node
, parent
, p
);
745 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
750 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
751 struct mem_cgroup_per_zone
*mz
,
752 struct mem_cgroup_tree_per_zone
*mctz
)
756 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
761 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
762 struct mem_cgroup_per_zone
*mz
,
763 struct mem_cgroup_tree_per_zone
*mctz
)
765 spin_lock(&mctz
->lock
);
766 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
767 spin_unlock(&mctz
->lock
);
771 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
773 unsigned long long excess
;
774 struct mem_cgroup_per_zone
*mz
;
775 struct mem_cgroup_tree_per_zone
*mctz
;
776 int nid
= page_to_nid(page
);
777 int zid
= page_zonenum(page
);
778 mctz
= soft_limit_tree_from_page(page
);
781 * Necessary to update all ancestors when hierarchy is used.
782 * because their event counter is not touched.
784 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
785 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
786 excess
= res_counter_soft_limit_excess(&memcg
->res
);
788 * We have to update the tree if mz is on RB-tree or
789 * mem is over its softlimit.
791 if (excess
|| mz
->on_tree
) {
792 spin_lock(&mctz
->lock
);
793 /* if on-tree, remove it */
795 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
797 * Insert again. mz->usage_in_excess will be updated.
798 * If excess is 0, no tree ops.
800 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
801 spin_unlock(&mctz
->lock
);
806 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
809 struct mem_cgroup_per_zone
*mz
;
810 struct mem_cgroup_tree_per_zone
*mctz
;
812 for_each_node(node
) {
813 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
814 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
815 mctz
= soft_limit_tree_node_zone(node
, zone
);
816 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
821 static struct mem_cgroup_per_zone
*
822 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
824 struct rb_node
*rightmost
= NULL
;
825 struct mem_cgroup_per_zone
*mz
;
829 rightmost
= rb_last(&mctz
->rb_root
);
831 goto done
; /* Nothing to reclaim from */
833 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
835 * Remove the node now but someone else can add it back,
836 * we will to add it back at the end of reclaim to its correct
837 * position in the tree.
839 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
840 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
841 !css_tryget(&mz
->memcg
->css
))
847 static struct mem_cgroup_per_zone
*
848 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
850 struct mem_cgroup_per_zone
*mz
;
852 spin_lock(&mctz
->lock
);
853 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
854 spin_unlock(&mctz
->lock
);
859 * Implementation Note: reading percpu statistics for memcg.
861 * Both of vmstat[] and percpu_counter has threshold and do periodic
862 * synchronization to implement "quick" read. There are trade-off between
863 * reading cost and precision of value. Then, we may have a chance to implement
864 * a periodic synchronizion of counter in memcg's counter.
866 * But this _read() function is used for user interface now. The user accounts
867 * memory usage by memory cgroup and he _always_ requires exact value because
868 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
869 * have to visit all online cpus and make sum. So, for now, unnecessary
870 * synchronization is not implemented. (just implemented for cpu hotplug)
872 * If there are kernel internal actions which can make use of some not-exact
873 * value, and reading all cpu value can be performance bottleneck in some
874 * common workload, threashold and synchonization as vmstat[] should be
877 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
878 enum mem_cgroup_stat_index idx
)
884 for_each_online_cpu(cpu
)
885 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
886 #ifdef CONFIG_HOTPLUG_CPU
887 spin_lock(&memcg
->pcp_counter_lock
);
888 val
+= memcg
->nocpu_base
.count
[idx
];
889 spin_unlock(&memcg
->pcp_counter_lock
);
895 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
898 int val
= (charge
) ? 1 : -1;
899 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
902 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
903 enum mem_cgroup_events_index idx
)
905 unsigned long val
= 0;
909 for_each_online_cpu(cpu
)
910 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
911 #ifdef CONFIG_HOTPLUG_CPU
912 spin_lock(&memcg
->pcp_counter_lock
);
913 val
+= memcg
->nocpu_base
.events
[idx
];
914 spin_unlock(&memcg
->pcp_counter_lock
);
920 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
922 bool anon
, int nr_pages
)
925 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
926 * counted as CACHE even if it's on ANON LRU.
929 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
932 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
935 if (PageTransHuge(page
))
936 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
939 /* pagein of a big page is an event. So, ignore page size */
941 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
943 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
944 nr_pages
= -nr_pages
; /* for event */
947 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
951 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
953 struct mem_cgroup_per_zone
*mz
;
955 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
956 return mz
->lru_size
[lru
];
960 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
961 unsigned int lru_mask
)
963 struct mem_cgroup_per_zone
*mz
;
965 unsigned long ret
= 0;
967 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
970 if (BIT(lru
) & lru_mask
)
971 ret
+= mz
->lru_size
[lru
];
977 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
978 int nid
, unsigned int lru_mask
)
983 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
984 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
990 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
991 unsigned int lru_mask
)
996 for_each_node_state(nid
, N_MEMORY
)
997 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
1001 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
1002 enum mem_cgroup_events_target target
)
1004 unsigned long val
, next
;
1006 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
1007 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
1008 /* from time_after() in jiffies.h */
1009 if ((long)next
- (long)val
< 0) {
1011 case MEM_CGROUP_TARGET_THRESH
:
1012 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
1014 case MEM_CGROUP_TARGET_SOFTLIMIT
:
1015 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1017 case MEM_CGROUP_TARGET_NUMAINFO
:
1018 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1023 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1030 * Check events in order.
1033 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1036 /* threshold event is triggered in finer grain than soft limit */
1037 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1038 MEM_CGROUP_TARGET_THRESH
))) {
1040 bool do_numainfo __maybe_unused
;
1042 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1043 MEM_CGROUP_TARGET_SOFTLIMIT
);
1044 #if MAX_NUMNODES > 1
1045 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1046 MEM_CGROUP_TARGET_NUMAINFO
);
1050 mem_cgroup_threshold(memcg
);
1051 if (unlikely(do_softlimit
))
1052 mem_cgroup_update_tree(memcg
, page
);
1053 #if MAX_NUMNODES > 1
1054 if (unlikely(do_numainfo
))
1055 atomic_inc(&memcg
->numainfo_events
);
1061 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1064 * mm_update_next_owner() may clear mm->owner to NULL
1065 * if it races with swapoff, page migration, etc.
1066 * So this can be called with p == NULL.
1071 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1074 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1076 struct mem_cgroup
*memcg
= NULL
;
1080 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1081 if (unlikely(!memcg
))
1083 } while (!css_tryget(&memcg
->css
));
1089 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1090 * ref. count) or NULL if the whole root's subtree has been visited.
1092 * helper function to be used by mem_cgroup_iter
1094 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1095 struct mem_cgroup
*last_visited
)
1097 struct cgroup_subsys_state
*prev_css
, *next_css
;
1099 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1101 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1104 * Even if we found a group we have to make sure it is
1105 * alive. css && !memcg means that the groups should be
1106 * skipped and we should continue the tree walk.
1107 * last_visited css is safe to use because it is
1108 * protected by css_get and the tree walk is rcu safe.
1110 * We do not take a reference on the root of the tree walk
1111 * because we might race with the root removal when it would
1112 * be the only node in the iterated hierarchy and mem_cgroup_iter
1113 * would end up in an endless loop because it expects that at
1114 * least one valid node will be returned. Root cannot disappear
1115 * because caller of the iterator should hold it already so
1116 * skipping css reference should be safe.
1119 if ((next_css
== &root
->css
) ||
1120 ((next_css
->flags
& CSS_ONLINE
) && css_tryget(next_css
)))
1121 return mem_cgroup_from_css(next_css
);
1123 prev_css
= next_css
;
1130 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1133 * When a group in the hierarchy below root is destroyed, the
1134 * hierarchy iterator can no longer be trusted since it might
1135 * have pointed to the destroyed group. Invalidate it.
1137 atomic_inc(&root
->dead_count
);
1140 static struct mem_cgroup
*
1141 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1142 struct mem_cgroup
*root
,
1145 struct mem_cgroup
*position
= NULL
;
1147 * A cgroup destruction happens in two stages: offlining and
1148 * release. They are separated by a RCU grace period.
1150 * If the iterator is valid, we may still race with an
1151 * offlining. The RCU lock ensures the object won't be
1152 * released, tryget will fail if we lost the race.
1154 *sequence
= atomic_read(&root
->dead_count
);
1155 if (iter
->last_dead_count
== *sequence
) {
1157 position
= iter
->last_visited
;
1160 * We cannot take a reference to root because we might race
1161 * with root removal and returning NULL would end up in
1162 * an endless loop on the iterator user level when root
1163 * would be returned all the time.
1165 if (position
&& position
!= root
&&
1166 !css_tryget(&position
->css
))
1172 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1173 struct mem_cgroup
*last_visited
,
1174 struct mem_cgroup
*new_position
,
1175 struct mem_cgroup
*root
,
1178 /* root reference counting symmetric to mem_cgroup_iter_load */
1179 if (last_visited
&& last_visited
!= root
)
1180 css_put(&last_visited
->css
);
1182 * We store the sequence count from the time @last_visited was
1183 * loaded successfully instead of rereading it here so that we
1184 * don't lose destruction events in between. We could have
1185 * raced with the destruction of @new_position after all.
1187 iter
->last_visited
= new_position
;
1189 iter
->last_dead_count
= sequence
;
1193 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1194 * @root: hierarchy root
1195 * @prev: previously returned memcg, NULL on first invocation
1196 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1198 * Returns references to children of the hierarchy below @root, or
1199 * @root itself, or %NULL after a full round-trip.
1201 * Caller must pass the return value in @prev on subsequent
1202 * invocations for reference counting, or use mem_cgroup_iter_break()
1203 * to cancel a hierarchy walk before the round-trip is complete.
1205 * Reclaimers can specify a zone and a priority level in @reclaim to
1206 * divide up the memcgs in the hierarchy among all concurrent
1207 * reclaimers operating on the same zone and priority.
1209 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1210 struct mem_cgroup
*prev
,
1211 struct mem_cgroup_reclaim_cookie
*reclaim
)
1213 struct mem_cgroup
*memcg
= NULL
;
1214 struct mem_cgroup
*last_visited
= NULL
;
1216 if (mem_cgroup_disabled())
1220 root
= root_mem_cgroup
;
1222 if (prev
&& !reclaim
)
1223 last_visited
= prev
;
1225 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1233 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1234 int uninitialized_var(seq
);
1237 int nid
= zone_to_nid(reclaim
->zone
);
1238 int zid
= zone_idx(reclaim
->zone
);
1239 struct mem_cgroup_per_zone
*mz
;
1241 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1242 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1243 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1244 iter
->last_visited
= NULL
;
1248 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1251 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1254 mem_cgroup_iter_update(iter
, last_visited
, memcg
, root
,
1259 else if (!prev
&& memcg
)
1260 reclaim
->generation
= iter
->generation
;
1269 if (prev
&& prev
!= root
)
1270 css_put(&prev
->css
);
1276 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1277 * @root: hierarchy root
1278 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1280 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1281 struct mem_cgroup
*prev
)
1284 root
= root_mem_cgroup
;
1285 if (prev
&& prev
!= root
)
1286 css_put(&prev
->css
);
1290 * Iteration constructs for visiting all cgroups (under a tree). If
1291 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1292 * be used for reference counting.
1294 #define for_each_mem_cgroup_tree(iter, root) \
1295 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1297 iter = mem_cgroup_iter(root, iter, NULL))
1299 #define for_each_mem_cgroup(iter) \
1300 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1302 iter = mem_cgroup_iter(NULL, iter, NULL))
1304 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1306 struct mem_cgroup
*memcg
;
1309 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1310 if (unlikely(!memcg
))
1315 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1318 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1326 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1329 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1330 * @zone: zone of the wanted lruvec
1331 * @memcg: memcg of the wanted lruvec
1333 * Returns the lru list vector holding pages for the given @zone and
1334 * @mem. This can be the global zone lruvec, if the memory controller
1337 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1338 struct mem_cgroup
*memcg
)
1340 struct mem_cgroup_per_zone
*mz
;
1341 struct lruvec
*lruvec
;
1343 if (mem_cgroup_disabled()) {
1344 lruvec
= &zone
->lruvec
;
1348 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1349 lruvec
= &mz
->lruvec
;
1352 * Since a node can be onlined after the mem_cgroup was created,
1353 * we have to be prepared to initialize lruvec->zone here;
1354 * and if offlined then reonlined, we need to reinitialize it.
1356 if (unlikely(lruvec
->zone
!= zone
))
1357 lruvec
->zone
= zone
;
1362 * Following LRU functions are allowed to be used without PCG_LOCK.
1363 * Operations are called by routine of global LRU independently from memcg.
1364 * What we have to take care of here is validness of pc->mem_cgroup.
1366 * Changes to pc->mem_cgroup happens when
1369 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1370 * It is added to LRU before charge.
1371 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1372 * When moving account, the page is not on LRU. It's isolated.
1376 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1378 * @zone: zone of the page
1380 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1382 struct mem_cgroup_per_zone
*mz
;
1383 struct mem_cgroup
*memcg
;
1384 struct page_cgroup
*pc
;
1385 struct lruvec
*lruvec
;
1387 if (mem_cgroup_disabled()) {
1388 lruvec
= &zone
->lruvec
;
1392 pc
= lookup_page_cgroup(page
);
1393 memcg
= pc
->mem_cgroup
;
1396 * Surreptitiously switch any uncharged offlist page to root:
1397 * an uncharged page off lru does nothing to secure
1398 * its former mem_cgroup from sudden removal.
1400 * Our caller holds lru_lock, and PageCgroupUsed is updated
1401 * under page_cgroup lock: between them, they make all uses
1402 * of pc->mem_cgroup safe.
1404 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1405 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1407 mz
= page_cgroup_zoneinfo(memcg
, page
);
1408 lruvec
= &mz
->lruvec
;
1411 * Since a node can be onlined after the mem_cgroup was created,
1412 * we have to be prepared to initialize lruvec->zone here;
1413 * and if offlined then reonlined, we need to reinitialize it.
1415 if (unlikely(lruvec
->zone
!= zone
))
1416 lruvec
->zone
= zone
;
1421 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1422 * @lruvec: mem_cgroup per zone lru vector
1423 * @lru: index of lru list the page is sitting on
1424 * @nr_pages: positive when adding or negative when removing
1426 * This function must be called when a page is added to or removed from an
1429 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1432 struct mem_cgroup_per_zone
*mz
;
1433 unsigned long *lru_size
;
1435 if (mem_cgroup_disabled())
1438 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1439 lru_size
= mz
->lru_size
+ lru
;
1440 *lru_size
+= nr_pages
;
1441 VM_BUG_ON((long)(*lru_size
) < 0);
1445 * Checks whether given mem is same or in the root_mem_cgroup's
1448 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1449 struct mem_cgroup
*memcg
)
1451 if (root_memcg
== memcg
)
1453 if (!root_memcg
->use_hierarchy
|| !memcg
)
1455 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1458 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1459 struct mem_cgroup
*memcg
)
1464 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1469 bool task_in_mem_cgroup(struct task_struct
*task
,
1470 const struct mem_cgroup
*memcg
)
1472 struct mem_cgroup
*curr
= NULL
;
1473 struct task_struct
*p
;
1476 p
= find_lock_task_mm(task
);
1478 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1482 * All threads may have already detached their mm's, but the oom
1483 * killer still needs to detect if they have already been oom
1484 * killed to prevent needlessly killing additional tasks.
1487 curr
= mem_cgroup_from_task(task
);
1489 css_get(&curr
->css
);
1495 * We should check use_hierarchy of "memcg" not "curr". Because checking
1496 * use_hierarchy of "curr" here make this function true if hierarchy is
1497 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1498 * hierarchy(even if use_hierarchy is disabled in "memcg").
1500 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1501 css_put(&curr
->css
);
1505 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1507 unsigned long inactive_ratio
;
1508 unsigned long inactive
;
1509 unsigned long active
;
1512 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1513 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1515 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1517 inactive_ratio
= int_sqrt(10 * gb
);
1521 return inactive
* inactive_ratio
< active
;
1524 #define mem_cgroup_from_res_counter(counter, member) \
1525 container_of(counter, struct mem_cgroup, member)
1528 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1529 * @memcg: the memory cgroup
1531 * Returns the maximum amount of memory @mem can be charged with, in
1534 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1536 unsigned long long margin
;
1538 margin
= res_counter_margin(&memcg
->res
);
1539 if (do_swap_account
)
1540 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1541 return margin
>> PAGE_SHIFT
;
1544 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1547 if (!css_parent(&memcg
->css
))
1548 return vm_swappiness
;
1550 return memcg
->swappiness
;
1554 * memcg->moving_account is used for checking possibility that some thread is
1555 * calling move_account(). When a thread on CPU-A starts moving pages under
1556 * a memcg, other threads should check memcg->moving_account under
1557 * rcu_read_lock(), like this:
1561 * memcg->moving_account+1 if (memcg->mocing_account)
1563 * synchronize_rcu() update something.
1568 /* for quick checking without looking up memcg */
1569 atomic_t memcg_moving __read_mostly
;
1571 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1573 atomic_inc(&memcg_moving
);
1574 atomic_inc(&memcg
->moving_account
);
1578 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1581 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1582 * We check NULL in callee rather than caller.
1585 atomic_dec(&memcg_moving
);
1586 atomic_dec(&memcg
->moving_account
);
1591 * 2 routines for checking "mem" is under move_account() or not.
1593 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1594 * is used for avoiding races in accounting. If true,
1595 * pc->mem_cgroup may be overwritten.
1597 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1598 * under hierarchy of moving cgroups. This is for
1599 * waiting at hith-memory prressure caused by "move".
1602 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1604 VM_BUG_ON(!rcu_read_lock_held());
1605 return atomic_read(&memcg
->moving_account
) > 0;
1608 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1610 struct mem_cgroup
*from
;
1611 struct mem_cgroup
*to
;
1614 * Unlike task_move routines, we access mc.to, mc.from not under
1615 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1617 spin_lock(&mc
.lock
);
1623 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1624 || mem_cgroup_same_or_subtree(memcg
, to
);
1626 spin_unlock(&mc
.lock
);
1630 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1632 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1633 if (mem_cgroup_under_move(memcg
)) {
1635 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1636 /* moving charge context might have finished. */
1639 finish_wait(&mc
.waitq
, &wait
);
1647 * Take this lock when
1648 * - a code tries to modify page's memcg while it's USED.
1649 * - a code tries to modify page state accounting in a memcg.
1650 * see mem_cgroup_stolen(), too.
1652 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1653 unsigned long *flags
)
1655 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1658 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1659 unsigned long *flags
)
1661 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1664 #define K(x) ((x) << (PAGE_SHIFT-10))
1666 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1667 * @memcg: The memory cgroup that went over limit
1668 * @p: Task that is going to be killed
1670 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1673 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1675 /* oom_info_lock ensures that parallel ooms do not interleave */
1676 static DEFINE_MUTEX(oom_info_lock
);
1677 struct mem_cgroup
*iter
;
1683 mutex_lock(&oom_info_lock
);
1686 pr_info("Task in ");
1687 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1688 pr_info(" killed as a result of limit of ");
1689 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1694 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1695 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1696 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1697 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1698 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1699 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1700 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1701 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1702 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1703 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1704 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1705 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1707 for_each_mem_cgroup_tree(iter
, memcg
) {
1708 pr_info("Memory cgroup stats for ");
1709 pr_cont_cgroup_path(iter
->css
.cgroup
);
1712 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1713 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1715 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1716 K(mem_cgroup_read_stat(iter
, i
)));
1719 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1720 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1721 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1725 mutex_unlock(&oom_info_lock
);
1729 * This function returns the number of memcg under hierarchy tree. Returns
1730 * 1(self count) if no children.
1732 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1735 struct mem_cgroup
*iter
;
1737 for_each_mem_cgroup_tree(iter
, memcg
)
1743 * Return the memory (and swap, if configured) limit for a memcg.
1745 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1749 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1752 * Do not consider swap space if we cannot swap due to swappiness
1754 if (mem_cgroup_swappiness(memcg
)) {
1757 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1758 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1761 * If memsw is finite and limits the amount of swap space
1762 * available to this memcg, return that limit.
1764 limit
= min(limit
, memsw
);
1770 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1773 struct mem_cgroup
*iter
;
1774 unsigned long chosen_points
= 0;
1775 unsigned long totalpages
;
1776 unsigned int points
= 0;
1777 struct task_struct
*chosen
= NULL
;
1780 * If current has a pending SIGKILL or is exiting, then automatically
1781 * select it. The goal is to allow it to allocate so that it may
1782 * quickly exit and free its memory.
1784 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1785 set_thread_flag(TIF_MEMDIE
);
1789 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1790 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1791 for_each_mem_cgroup_tree(iter
, memcg
) {
1792 struct css_task_iter it
;
1793 struct task_struct
*task
;
1795 css_task_iter_start(&iter
->css
, &it
);
1796 while ((task
= css_task_iter_next(&it
))) {
1797 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1799 case OOM_SCAN_SELECT
:
1801 put_task_struct(chosen
);
1803 chosen_points
= ULONG_MAX
;
1804 get_task_struct(chosen
);
1806 case OOM_SCAN_CONTINUE
:
1808 case OOM_SCAN_ABORT
:
1809 css_task_iter_end(&it
);
1810 mem_cgroup_iter_break(memcg
, iter
);
1812 put_task_struct(chosen
);
1817 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1818 if (!points
|| points
< chosen_points
)
1820 /* Prefer thread group leaders for display purposes */
1821 if (points
== chosen_points
&&
1822 thread_group_leader(chosen
))
1826 put_task_struct(chosen
);
1828 chosen_points
= points
;
1829 get_task_struct(chosen
);
1831 css_task_iter_end(&it
);
1836 points
= chosen_points
* 1000 / totalpages
;
1837 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1838 NULL
, "Memory cgroup out of memory");
1841 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1843 unsigned long flags
)
1845 unsigned long total
= 0;
1846 bool noswap
= false;
1849 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1851 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1854 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1856 drain_all_stock_async(memcg
);
1857 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1859 * Allow limit shrinkers, which are triggered directly
1860 * by userspace, to catch signals and stop reclaim
1861 * after minimal progress, regardless of the margin.
1863 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1865 if (mem_cgroup_margin(memcg
))
1868 * If nothing was reclaimed after two attempts, there
1869 * may be no reclaimable pages in this hierarchy.
1878 * test_mem_cgroup_node_reclaimable
1879 * @memcg: the target memcg
1880 * @nid: the node ID to be checked.
1881 * @noswap : specify true here if the user wants flle only information.
1883 * This function returns whether the specified memcg contains any
1884 * reclaimable pages on a node. Returns true if there are any reclaimable
1885 * pages in the node.
1887 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1888 int nid
, bool noswap
)
1890 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1892 if (noswap
|| !total_swap_pages
)
1894 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1899 #if MAX_NUMNODES > 1
1902 * Always updating the nodemask is not very good - even if we have an empty
1903 * list or the wrong list here, we can start from some node and traverse all
1904 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1907 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1911 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1912 * pagein/pageout changes since the last update.
1914 if (!atomic_read(&memcg
->numainfo_events
))
1916 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1919 /* make a nodemask where this memcg uses memory from */
1920 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1922 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1924 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1925 node_clear(nid
, memcg
->scan_nodes
);
1928 atomic_set(&memcg
->numainfo_events
, 0);
1929 atomic_set(&memcg
->numainfo_updating
, 0);
1933 * Selecting a node where we start reclaim from. Because what we need is just
1934 * reducing usage counter, start from anywhere is O,K. Considering
1935 * memory reclaim from current node, there are pros. and cons.
1937 * Freeing memory from current node means freeing memory from a node which
1938 * we'll use or we've used. So, it may make LRU bad. And if several threads
1939 * hit limits, it will see a contention on a node. But freeing from remote
1940 * node means more costs for memory reclaim because of memory latency.
1942 * Now, we use round-robin. Better algorithm is welcomed.
1944 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1948 mem_cgroup_may_update_nodemask(memcg
);
1949 node
= memcg
->last_scanned_node
;
1951 node
= next_node(node
, memcg
->scan_nodes
);
1952 if (node
== MAX_NUMNODES
)
1953 node
= first_node(memcg
->scan_nodes
);
1955 * We call this when we hit limit, not when pages are added to LRU.
1956 * No LRU may hold pages because all pages are UNEVICTABLE or
1957 * memcg is too small and all pages are not on LRU. In that case,
1958 * we use curret node.
1960 if (unlikely(node
== MAX_NUMNODES
))
1961 node
= numa_node_id();
1963 memcg
->last_scanned_node
= node
;
1968 * Check all nodes whether it contains reclaimable pages or not.
1969 * For quick scan, we make use of scan_nodes. This will allow us to skip
1970 * unused nodes. But scan_nodes is lazily updated and may not cotain
1971 * enough new information. We need to do double check.
1973 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1978 * quick check...making use of scan_node.
1979 * We can skip unused nodes.
1981 if (!nodes_empty(memcg
->scan_nodes
)) {
1982 for (nid
= first_node(memcg
->scan_nodes
);
1984 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1986 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1991 * Check rest of nodes.
1993 for_each_node_state(nid
, N_MEMORY
) {
1994 if (node_isset(nid
, memcg
->scan_nodes
))
1996 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2003 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2008 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2010 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2014 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2017 unsigned long *total_scanned
)
2019 struct mem_cgroup
*victim
= NULL
;
2022 unsigned long excess
;
2023 unsigned long nr_scanned
;
2024 struct mem_cgroup_reclaim_cookie reclaim
= {
2029 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2032 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2037 * If we have not been able to reclaim
2038 * anything, it might because there are
2039 * no reclaimable pages under this hierarchy
2044 * We want to do more targeted reclaim.
2045 * excess >> 2 is not to excessive so as to
2046 * reclaim too much, nor too less that we keep
2047 * coming back to reclaim from this cgroup
2049 if (total
>= (excess
>> 2) ||
2050 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2055 if (!mem_cgroup_reclaimable(victim
, false))
2057 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2059 *total_scanned
+= nr_scanned
;
2060 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2063 mem_cgroup_iter_break(root_memcg
, victim
);
2067 #ifdef CONFIG_LOCKDEP
2068 static struct lockdep_map memcg_oom_lock_dep_map
= {
2069 .name
= "memcg_oom_lock",
2073 static DEFINE_SPINLOCK(memcg_oom_lock
);
2076 * Check OOM-Killer is already running under our hierarchy.
2077 * If someone is running, return false.
2079 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2081 struct mem_cgroup
*iter
, *failed
= NULL
;
2083 spin_lock(&memcg_oom_lock
);
2085 for_each_mem_cgroup_tree(iter
, memcg
) {
2086 if (iter
->oom_lock
) {
2088 * this subtree of our hierarchy is already locked
2089 * so we cannot give a lock.
2092 mem_cgroup_iter_break(memcg
, iter
);
2095 iter
->oom_lock
= true;
2100 * OK, we failed to lock the whole subtree so we have
2101 * to clean up what we set up to the failing subtree
2103 for_each_mem_cgroup_tree(iter
, memcg
) {
2104 if (iter
== failed
) {
2105 mem_cgroup_iter_break(memcg
, iter
);
2108 iter
->oom_lock
= false;
2111 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2113 spin_unlock(&memcg_oom_lock
);
2118 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2120 struct mem_cgroup
*iter
;
2122 spin_lock(&memcg_oom_lock
);
2123 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2124 for_each_mem_cgroup_tree(iter
, memcg
)
2125 iter
->oom_lock
= false;
2126 spin_unlock(&memcg_oom_lock
);
2129 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2131 struct mem_cgroup
*iter
;
2133 for_each_mem_cgroup_tree(iter
, memcg
)
2134 atomic_inc(&iter
->under_oom
);
2137 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2139 struct mem_cgroup
*iter
;
2142 * When a new child is created while the hierarchy is under oom,
2143 * mem_cgroup_oom_lock() may not be called. We have to use
2144 * atomic_add_unless() here.
2146 for_each_mem_cgroup_tree(iter
, memcg
)
2147 atomic_add_unless(&iter
->under_oom
, -1, 0);
2150 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2152 struct oom_wait_info
{
2153 struct mem_cgroup
*memcg
;
2157 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2158 unsigned mode
, int sync
, void *arg
)
2160 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2161 struct mem_cgroup
*oom_wait_memcg
;
2162 struct oom_wait_info
*oom_wait_info
;
2164 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2165 oom_wait_memcg
= oom_wait_info
->memcg
;
2168 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2169 * Then we can use css_is_ancestor without taking care of RCU.
2171 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2172 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2174 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2177 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2179 atomic_inc(&memcg
->oom_wakeups
);
2180 /* for filtering, pass "memcg" as argument. */
2181 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2184 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2186 if (memcg
&& atomic_read(&memcg
->under_oom
))
2187 memcg_wakeup_oom(memcg
);
2190 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2192 if (!current
->memcg_oom
.may_oom
)
2195 * We are in the middle of the charge context here, so we
2196 * don't want to block when potentially sitting on a callstack
2197 * that holds all kinds of filesystem and mm locks.
2199 * Also, the caller may handle a failed allocation gracefully
2200 * (like optional page cache readahead) and so an OOM killer
2201 * invocation might not even be necessary.
2203 * That's why we don't do anything here except remember the
2204 * OOM context and then deal with it at the end of the page
2205 * fault when the stack is unwound, the locks are released,
2206 * and when we know whether the fault was overall successful.
2208 css_get(&memcg
->css
);
2209 current
->memcg_oom
.memcg
= memcg
;
2210 current
->memcg_oom
.gfp_mask
= mask
;
2211 current
->memcg_oom
.order
= order
;
2215 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2216 * @handle: actually kill/wait or just clean up the OOM state
2218 * This has to be called at the end of a page fault if the memcg OOM
2219 * handler was enabled.
2221 * Memcg supports userspace OOM handling where failed allocations must
2222 * sleep on a waitqueue until the userspace task resolves the
2223 * situation. Sleeping directly in the charge context with all kinds
2224 * of locks held is not a good idea, instead we remember an OOM state
2225 * in the task and mem_cgroup_oom_synchronize() has to be called at
2226 * the end of the page fault to complete the OOM handling.
2228 * Returns %true if an ongoing memcg OOM situation was detected and
2229 * completed, %false otherwise.
2231 bool mem_cgroup_oom_synchronize(bool handle
)
2233 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2234 struct oom_wait_info owait
;
2237 /* OOM is global, do not handle */
2244 owait
.memcg
= memcg
;
2245 owait
.wait
.flags
= 0;
2246 owait
.wait
.func
= memcg_oom_wake_function
;
2247 owait
.wait
.private = current
;
2248 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2250 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2251 mem_cgroup_mark_under_oom(memcg
);
2253 locked
= mem_cgroup_oom_trylock(memcg
);
2256 mem_cgroup_oom_notify(memcg
);
2258 if (locked
&& !memcg
->oom_kill_disable
) {
2259 mem_cgroup_unmark_under_oom(memcg
);
2260 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2261 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2262 current
->memcg_oom
.order
);
2265 mem_cgroup_unmark_under_oom(memcg
);
2266 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2270 mem_cgroup_oom_unlock(memcg
);
2272 * There is no guarantee that an OOM-lock contender
2273 * sees the wakeups triggered by the OOM kill
2274 * uncharges. Wake any sleepers explicitely.
2276 memcg_oom_recover(memcg
);
2279 current
->memcg_oom
.memcg
= NULL
;
2280 css_put(&memcg
->css
);
2285 * Currently used to update mapped file statistics, but the routine can be
2286 * generalized to update other statistics as well.
2288 * Notes: Race condition
2290 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2291 * it tends to be costly. But considering some conditions, we doesn't need
2292 * to do so _always_.
2294 * Considering "charge", lock_page_cgroup() is not required because all
2295 * file-stat operations happen after a page is attached to radix-tree. There
2296 * are no race with "charge".
2298 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2299 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2300 * if there are race with "uncharge". Statistics itself is properly handled
2303 * Considering "move", this is an only case we see a race. To make the race
2304 * small, we check mm->moving_account and detect there are possibility of race
2305 * If there is, we take a lock.
2308 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2309 bool *locked
, unsigned long *flags
)
2311 struct mem_cgroup
*memcg
;
2312 struct page_cgroup
*pc
;
2314 pc
= lookup_page_cgroup(page
);
2316 memcg
= pc
->mem_cgroup
;
2317 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2320 * If this memory cgroup is not under account moving, we don't
2321 * need to take move_lock_mem_cgroup(). Because we already hold
2322 * rcu_read_lock(), any calls to move_account will be delayed until
2323 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2325 if (!mem_cgroup_stolen(memcg
))
2328 move_lock_mem_cgroup(memcg
, flags
);
2329 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2330 move_unlock_mem_cgroup(memcg
, flags
);
2336 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2338 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2341 * It's guaranteed that pc->mem_cgroup never changes while
2342 * lock is held because a routine modifies pc->mem_cgroup
2343 * should take move_lock_mem_cgroup().
2345 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2348 void mem_cgroup_update_page_stat(struct page
*page
,
2349 enum mem_cgroup_stat_index idx
, int val
)
2351 struct mem_cgroup
*memcg
;
2352 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2353 unsigned long uninitialized_var(flags
);
2355 if (mem_cgroup_disabled())
2358 VM_BUG_ON(!rcu_read_lock_held());
2359 memcg
= pc
->mem_cgroup
;
2360 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2363 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2367 * size of first charge trial. "32" comes from vmscan.c's magic value.
2368 * TODO: maybe necessary to use big numbers in big irons.
2370 #define CHARGE_BATCH 32U
2371 struct memcg_stock_pcp
{
2372 struct mem_cgroup
*cached
; /* this never be root cgroup */
2373 unsigned int nr_pages
;
2374 struct work_struct work
;
2375 unsigned long flags
;
2376 #define FLUSHING_CACHED_CHARGE 0
2378 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2379 static DEFINE_MUTEX(percpu_charge_mutex
);
2382 * consume_stock: Try to consume stocked charge on this cpu.
2383 * @memcg: memcg to consume from.
2384 * @nr_pages: how many pages to charge.
2386 * The charges will only happen if @memcg matches the current cpu's memcg
2387 * stock, and at least @nr_pages are available in that stock. Failure to
2388 * service an allocation will refill the stock.
2390 * returns true if successful, false otherwise.
2392 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2394 struct memcg_stock_pcp
*stock
;
2397 if (nr_pages
> CHARGE_BATCH
)
2400 stock
= &get_cpu_var(memcg_stock
);
2401 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2402 stock
->nr_pages
-= nr_pages
;
2403 else /* need to call res_counter_charge */
2405 put_cpu_var(memcg_stock
);
2410 * Returns stocks cached in percpu to res_counter and reset cached information.
2412 static void drain_stock(struct memcg_stock_pcp
*stock
)
2414 struct mem_cgroup
*old
= stock
->cached
;
2416 if (stock
->nr_pages
) {
2417 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2419 res_counter_uncharge(&old
->res
, bytes
);
2420 if (do_swap_account
)
2421 res_counter_uncharge(&old
->memsw
, bytes
);
2422 stock
->nr_pages
= 0;
2424 stock
->cached
= NULL
;
2428 * This must be called under preempt disabled or must be called by
2429 * a thread which is pinned to local cpu.
2431 static void drain_local_stock(struct work_struct
*dummy
)
2433 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2435 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2438 static void __init
memcg_stock_init(void)
2442 for_each_possible_cpu(cpu
) {
2443 struct memcg_stock_pcp
*stock
=
2444 &per_cpu(memcg_stock
, cpu
);
2445 INIT_WORK(&stock
->work
, drain_local_stock
);
2450 * Cache charges(val) which is from res_counter, to local per_cpu area.
2451 * This will be consumed by consume_stock() function, later.
2453 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2455 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2457 if (stock
->cached
!= memcg
) { /* reset if necessary */
2459 stock
->cached
= memcg
;
2461 stock
->nr_pages
+= nr_pages
;
2462 put_cpu_var(memcg_stock
);
2466 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2467 * of the hierarchy under it. sync flag says whether we should block
2468 * until the work is done.
2470 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2474 /* Notify other cpus that system-wide "drain" is running */
2477 for_each_online_cpu(cpu
) {
2478 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2479 struct mem_cgroup
*memcg
;
2481 memcg
= stock
->cached
;
2482 if (!memcg
|| !stock
->nr_pages
)
2484 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2486 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2488 drain_local_stock(&stock
->work
);
2490 schedule_work_on(cpu
, &stock
->work
);
2498 for_each_online_cpu(cpu
) {
2499 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2500 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2501 flush_work(&stock
->work
);
2508 * Tries to drain stocked charges in other cpus. This function is asynchronous
2509 * and just put a work per cpu for draining localy on each cpu. Caller can
2510 * expects some charges will be back to res_counter later but cannot wait for
2513 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2516 * If someone calls draining, avoid adding more kworker runs.
2518 if (!mutex_trylock(&percpu_charge_mutex
))
2520 drain_all_stock(root_memcg
, false);
2521 mutex_unlock(&percpu_charge_mutex
);
2524 /* This is a synchronous drain interface. */
2525 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2527 /* called when force_empty is called */
2528 mutex_lock(&percpu_charge_mutex
);
2529 drain_all_stock(root_memcg
, true);
2530 mutex_unlock(&percpu_charge_mutex
);
2534 * This function drains percpu counter value from DEAD cpu and
2535 * move it to local cpu. Note that this function can be preempted.
2537 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2541 spin_lock(&memcg
->pcp_counter_lock
);
2542 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2543 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2545 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2546 memcg
->nocpu_base
.count
[i
] += x
;
2548 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2549 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2551 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2552 memcg
->nocpu_base
.events
[i
] += x
;
2554 spin_unlock(&memcg
->pcp_counter_lock
);
2557 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2558 unsigned long action
,
2561 int cpu
= (unsigned long)hcpu
;
2562 struct memcg_stock_pcp
*stock
;
2563 struct mem_cgroup
*iter
;
2565 if (action
== CPU_ONLINE
)
2568 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2571 for_each_mem_cgroup(iter
)
2572 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2574 stock
= &per_cpu(memcg_stock
, cpu
);
2580 /* See __mem_cgroup_try_charge() for details */
2582 CHARGE_OK
, /* success */
2583 CHARGE_RETRY
, /* need to retry but retry is not bad */
2584 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2585 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2588 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2589 unsigned int nr_pages
, unsigned int min_pages
,
2592 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2593 struct mem_cgroup
*mem_over_limit
;
2594 struct res_counter
*fail_res
;
2595 unsigned long flags
= 0;
2598 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2601 if (!do_swap_account
)
2603 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2607 res_counter_uncharge(&memcg
->res
, csize
);
2608 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2609 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2611 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2613 * Never reclaim on behalf of optional batching, retry with a
2614 * single page instead.
2616 if (nr_pages
> min_pages
)
2617 return CHARGE_RETRY
;
2619 if (!(gfp_mask
& __GFP_WAIT
))
2620 return CHARGE_WOULDBLOCK
;
2622 if (gfp_mask
& __GFP_NORETRY
)
2623 return CHARGE_NOMEM
;
2625 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2626 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2627 return CHARGE_RETRY
;
2629 * Even though the limit is exceeded at this point, reclaim
2630 * may have been able to free some pages. Retry the charge
2631 * before killing the task.
2633 * Only for regular pages, though: huge pages are rather
2634 * unlikely to succeed so close to the limit, and we fall back
2635 * to regular pages anyway in case of failure.
2637 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2638 return CHARGE_RETRY
;
2641 * At task move, charge accounts can be doubly counted. So, it's
2642 * better to wait until the end of task_move if something is going on.
2644 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2645 return CHARGE_RETRY
;
2648 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2650 return CHARGE_NOMEM
;
2654 * __mem_cgroup_try_charge() does
2655 * 1. detect memcg to be charged against from passed *mm and *ptr,
2656 * 2. update res_counter
2657 * 3. call memory reclaim if necessary.
2659 * In some special case, if the task is fatal, fatal_signal_pending() or
2660 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2661 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2662 * as possible without any hazards. 2: all pages should have a valid
2663 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2664 * pointer, that is treated as a charge to root_mem_cgroup.
2666 * So __mem_cgroup_try_charge() will return
2667 * 0 ... on success, filling *ptr with a valid memcg pointer.
2668 * -ENOMEM ... charge failure because of resource limits.
2669 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2671 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2672 * the oom-killer can be invoked.
2674 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2676 unsigned int nr_pages
,
2677 struct mem_cgroup
**ptr
,
2680 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2681 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2682 struct mem_cgroup
*memcg
= NULL
;
2686 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2687 * in system level. So, allow to go ahead dying process in addition to
2690 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2691 || fatal_signal_pending(current
)))
2694 if (unlikely(task_in_memcg_oom(current
)))
2697 if (gfp_mask
& __GFP_NOFAIL
)
2700 if (*ptr
) { /* css should be a valid one */
2702 if (mem_cgroup_is_root(memcg
))
2704 if (consume_stock(memcg
, nr_pages
))
2706 css_get(&memcg
->css
);
2708 struct task_struct
*p
;
2711 p
= rcu_dereference(mm
->owner
);
2713 * Because we don't have task_lock(), "p" can exit.
2714 * In that case, "memcg" can point to root or p can be NULL with
2715 * race with swapoff. Then, we have small risk of mis-accouning.
2716 * But such kind of mis-account by race always happens because
2717 * we don't have cgroup_mutex(). It's overkill and we allo that
2719 * (*) swapoff at el will charge against mm-struct not against
2720 * task-struct. So, mm->owner can be NULL.
2722 memcg
= mem_cgroup_from_task(p
);
2724 memcg
= root_mem_cgroup
;
2725 if (mem_cgroup_is_root(memcg
)) {
2729 if (consume_stock(memcg
, nr_pages
)) {
2731 * It seems dagerous to access memcg without css_get().
2732 * But considering how consume_stok works, it's not
2733 * necessary. If consume_stock success, some charges
2734 * from this memcg are cached on this cpu. So, we
2735 * don't need to call css_get()/css_tryget() before
2736 * calling consume_stock().
2741 /* after here, we may be blocked. we need to get refcnt */
2742 if (!css_tryget(&memcg
->css
)) {
2750 bool invoke_oom
= oom
&& !nr_oom_retries
;
2752 /* If killed, bypass charge */
2753 if (fatal_signal_pending(current
)) {
2754 css_put(&memcg
->css
);
2758 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2759 nr_pages
, invoke_oom
);
2763 case CHARGE_RETRY
: /* not in OOM situation but retry */
2765 css_put(&memcg
->css
);
2768 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2769 css_put(&memcg
->css
);
2771 case CHARGE_NOMEM
: /* OOM routine works */
2772 if (!oom
|| invoke_oom
) {
2773 css_put(&memcg
->css
);
2779 } while (ret
!= CHARGE_OK
);
2781 if (batch
> nr_pages
)
2782 refill_stock(memcg
, batch
- nr_pages
);
2783 css_put(&memcg
->css
);
2788 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2793 *ptr
= root_mem_cgroup
;
2798 * Somemtimes we have to undo a charge we got by try_charge().
2799 * This function is for that and do uncharge, put css's refcnt.
2800 * gotten by try_charge().
2802 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2803 unsigned int nr_pages
)
2805 if (!mem_cgroup_is_root(memcg
)) {
2806 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2808 res_counter_uncharge(&memcg
->res
, bytes
);
2809 if (do_swap_account
)
2810 res_counter_uncharge(&memcg
->memsw
, bytes
);
2815 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2816 * This is useful when moving usage to parent cgroup.
2818 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2819 unsigned int nr_pages
)
2821 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2823 if (mem_cgroup_is_root(memcg
))
2826 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2827 if (do_swap_account
)
2828 res_counter_uncharge_until(&memcg
->memsw
,
2829 memcg
->memsw
.parent
, bytes
);
2833 * A helper function to get mem_cgroup from ID. must be called under
2834 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2835 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2836 * called against removed memcg.)
2838 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2840 /* ID 0 is unused ID */
2843 return mem_cgroup_from_id(id
);
2846 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2848 struct mem_cgroup
*memcg
= NULL
;
2849 struct page_cgroup
*pc
;
2853 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2855 pc
= lookup_page_cgroup(page
);
2856 lock_page_cgroup(pc
);
2857 if (PageCgroupUsed(pc
)) {
2858 memcg
= pc
->mem_cgroup
;
2859 if (memcg
&& !css_tryget(&memcg
->css
))
2861 } else if (PageSwapCache(page
)) {
2862 ent
.val
= page_private(page
);
2863 id
= lookup_swap_cgroup_id(ent
);
2865 memcg
= mem_cgroup_lookup(id
);
2866 if (memcg
&& !css_tryget(&memcg
->css
))
2870 unlock_page_cgroup(pc
);
2874 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2876 unsigned int nr_pages
,
2877 enum charge_type ctype
,
2880 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2881 struct zone
*uninitialized_var(zone
);
2882 struct lruvec
*lruvec
;
2883 bool was_on_lru
= false;
2886 lock_page_cgroup(pc
);
2887 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2889 * we don't need page_cgroup_lock about tail pages, becase they are not
2890 * accessed by any other context at this point.
2894 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2895 * may already be on some other mem_cgroup's LRU. Take care of it.
2898 zone
= page_zone(page
);
2899 spin_lock_irq(&zone
->lru_lock
);
2900 if (PageLRU(page
)) {
2901 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2903 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2908 pc
->mem_cgroup
= memcg
;
2910 * We access a page_cgroup asynchronously without lock_page_cgroup().
2911 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2912 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2913 * before USED bit, we need memory barrier here.
2914 * See mem_cgroup_add_lru_list(), etc.
2917 SetPageCgroupUsed(pc
);
2921 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2922 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2924 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2926 spin_unlock_irq(&zone
->lru_lock
);
2929 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2934 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2935 unlock_page_cgroup(pc
);
2938 * "charge_statistics" updated event counter. Then, check it.
2939 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2940 * if they exceeds softlimit.
2942 memcg_check_events(memcg
, page
);
2945 static DEFINE_MUTEX(set_limit_mutex
);
2947 #ifdef CONFIG_MEMCG_KMEM
2948 static DEFINE_MUTEX(activate_kmem_mutex
);
2950 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2952 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2953 memcg_kmem_is_active(memcg
);
2957 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2958 * in the memcg_cache_params struct.
2960 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2962 struct kmem_cache
*cachep
;
2964 VM_BUG_ON(p
->is_root_cache
);
2965 cachep
= p
->root_cache
;
2966 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2969 #ifdef CONFIG_SLABINFO
2970 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
2972 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
2973 struct memcg_cache_params
*params
;
2975 if (!memcg_can_account_kmem(memcg
))
2978 print_slabinfo_header(m
);
2980 mutex_lock(&memcg
->slab_caches_mutex
);
2981 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2982 cache_show(memcg_params_to_cache(params
), m
);
2983 mutex_unlock(&memcg
->slab_caches_mutex
);
2989 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2991 struct res_counter
*fail_res
;
2992 struct mem_cgroup
*_memcg
;
2995 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
3000 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
3001 &_memcg
, oom_gfp_allowed(gfp
));
3003 if (ret
== -EINTR
) {
3005 * __mem_cgroup_try_charge() chosed to bypass to root due to
3006 * OOM kill or fatal signal. Since our only options are to
3007 * either fail the allocation or charge it to this cgroup, do
3008 * it as a temporary condition. But we can't fail. From a
3009 * kmem/slab perspective, the cache has already been selected,
3010 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3013 * This condition will only trigger if the task entered
3014 * memcg_charge_kmem in a sane state, but was OOM-killed during
3015 * __mem_cgroup_try_charge() above. Tasks that were already
3016 * dying when the allocation triggers should have been already
3017 * directed to the root cgroup in memcontrol.h
3019 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3020 if (do_swap_account
)
3021 res_counter_charge_nofail(&memcg
->memsw
, size
,
3025 res_counter_uncharge(&memcg
->kmem
, size
);
3030 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3032 res_counter_uncharge(&memcg
->res
, size
);
3033 if (do_swap_account
)
3034 res_counter_uncharge(&memcg
->memsw
, size
);
3037 if (res_counter_uncharge(&memcg
->kmem
, size
))
3041 * Releases a reference taken in kmem_cgroup_css_offline in case
3042 * this last uncharge is racing with the offlining code or it is
3043 * outliving the memcg existence.
3045 * The memory barrier imposed by test&clear is paired with the
3046 * explicit one in memcg_kmem_mark_dead().
3048 if (memcg_kmem_test_and_clear_dead(memcg
))
3049 css_put(&memcg
->css
);
3053 * helper for acessing a memcg's index. It will be used as an index in the
3054 * child cache array in kmem_cache, and also to derive its name. This function
3055 * will return -1 when this is not a kmem-limited memcg.
3057 int memcg_cache_id(struct mem_cgroup
*memcg
)
3059 return memcg
? memcg
->kmemcg_id
: -1;
3062 static size_t memcg_caches_array_size(int num_groups
)
3065 if (num_groups
<= 0)
3068 size
= 2 * num_groups
;
3069 if (size
< MEMCG_CACHES_MIN_SIZE
)
3070 size
= MEMCG_CACHES_MIN_SIZE
;
3071 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3072 size
= MEMCG_CACHES_MAX_SIZE
;
3078 * We should update the current array size iff all caches updates succeed. This
3079 * can only be done from the slab side. The slab mutex needs to be held when
3082 void memcg_update_array_size(int num
)
3084 if (num
> memcg_limited_groups_array_size
)
3085 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3088 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3090 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3092 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3094 VM_BUG_ON(!is_root_cache(s
));
3096 if (num_groups
> memcg_limited_groups_array_size
) {
3098 struct memcg_cache_params
*new_params
;
3099 ssize_t size
= memcg_caches_array_size(num_groups
);
3101 size
*= sizeof(void *);
3102 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3104 new_params
= kzalloc(size
, GFP_KERNEL
);
3108 new_params
->is_root_cache
= true;
3111 * There is the chance it will be bigger than
3112 * memcg_limited_groups_array_size, if we failed an allocation
3113 * in a cache, in which case all caches updated before it, will
3114 * have a bigger array.
3116 * But if that is the case, the data after
3117 * memcg_limited_groups_array_size is certainly unused
3119 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3120 if (!cur_params
->memcg_caches
[i
])
3122 new_params
->memcg_caches
[i
] =
3123 cur_params
->memcg_caches
[i
];
3127 * Ideally, we would wait until all caches succeed, and only
3128 * then free the old one. But this is not worth the extra
3129 * pointer per-cache we'd have to have for this.
3131 * It is not a big deal if some caches are left with a size
3132 * bigger than the others. And all updates will reset this
3135 rcu_assign_pointer(s
->memcg_params
, new_params
);
3137 kfree_rcu(cur_params
, rcu_head
);
3142 int memcg_alloc_cache_params(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3143 struct kmem_cache
*root_cache
)
3147 if (!memcg_kmem_enabled())
3151 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3152 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3154 size
= sizeof(struct memcg_cache_params
);
3156 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3157 if (!s
->memcg_params
)
3161 s
->memcg_params
->memcg
= memcg
;
3162 s
->memcg_params
->root_cache
= root_cache
;
3163 INIT_WORK(&s
->memcg_params
->destroy
,
3164 kmem_cache_destroy_work_func
);
3166 s
->memcg_params
->is_root_cache
= true;
3171 void memcg_free_cache_params(struct kmem_cache
*s
)
3173 kfree(s
->memcg_params
);
3176 void memcg_register_cache(struct kmem_cache
*s
)
3178 struct kmem_cache
*root
;
3179 struct mem_cgroup
*memcg
;
3182 if (is_root_cache(s
))
3186 * Holding the slab_mutex assures nobody will touch the memcg_caches
3187 * array while we are modifying it.
3189 lockdep_assert_held(&slab_mutex
);
3191 root
= s
->memcg_params
->root_cache
;
3192 memcg
= s
->memcg_params
->memcg
;
3193 id
= memcg_cache_id(memcg
);
3195 css_get(&memcg
->css
);
3199 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3200 * barrier here to ensure nobody will see the kmem_cache partially
3206 * Initialize the pointer to this cache in its parent's memcg_params
3207 * before adding it to the memcg_slab_caches list, otherwise we can
3208 * fail to convert memcg_params_to_cache() while traversing the list.
3210 VM_BUG_ON(root
->memcg_params
->memcg_caches
[id
]);
3211 root
->memcg_params
->memcg_caches
[id
] = s
;
3213 mutex_lock(&memcg
->slab_caches_mutex
);
3214 list_add(&s
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3215 mutex_unlock(&memcg
->slab_caches_mutex
);
3218 void memcg_unregister_cache(struct kmem_cache
*s
)
3220 struct kmem_cache
*root
;
3221 struct mem_cgroup
*memcg
;
3224 if (is_root_cache(s
))
3228 * Holding the slab_mutex assures nobody will touch the memcg_caches
3229 * array while we are modifying it.
3231 lockdep_assert_held(&slab_mutex
);
3233 root
= s
->memcg_params
->root_cache
;
3234 memcg
= s
->memcg_params
->memcg
;
3235 id
= memcg_cache_id(memcg
);
3237 mutex_lock(&memcg
->slab_caches_mutex
);
3238 list_del(&s
->memcg_params
->list
);
3239 mutex_unlock(&memcg
->slab_caches_mutex
);
3242 * Clear the pointer to this cache in its parent's memcg_params only
3243 * after removing it from the memcg_slab_caches list, otherwise we can
3244 * fail to convert memcg_params_to_cache() while traversing the list.
3246 VM_BUG_ON(!root
->memcg_params
->memcg_caches
[id
]);
3247 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3249 css_put(&memcg
->css
);
3253 * During the creation a new cache, we need to disable our accounting mechanism
3254 * altogether. This is true even if we are not creating, but rather just
3255 * enqueing new caches to be created.
3257 * This is because that process will trigger allocations; some visible, like
3258 * explicit kmallocs to auxiliary data structures, name strings and internal
3259 * cache structures; some well concealed, like INIT_WORK() that can allocate
3260 * objects during debug.
3262 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3263 * to it. This may not be a bounded recursion: since the first cache creation
3264 * failed to complete (waiting on the allocation), we'll just try to create the
3265 * cache again, failing at the same point.
3267 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3268 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3269 * inside the following two functions.
3271 static inline void memcg_stop_kmem_account(void)
3273 VM_BUG_ON(!current
->mm
);
3274 current
->memcg_kmem_skip_account
++;
3277 static inline void memcg_resume_kmem_account(void)
3279 VM_BUG_ON(!current
->mm
);
3280 current
->memcg_kmem_skip_account
--;
3283 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3285 struct kmem_cache
*cachep
;
3286 struct memcg_cache_params
*p
;
3288 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3290 cachep
= memcg_params_to_cache(p
);
3293 * If we get down to 0 after shrink, we could delete right away.
3294 * However, memcg_release_pages() already puts us back in the workqueue
3295 * in that case. If we proceed deleting, we'll get a dangling
3296 * reference, and removing the object from the workqueue in that case
3297 * is unnecessary complication. We are not a fast path.
3299 * Note that this case is fundamentally different from racing with
3300 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3301 * kmem_cache_shrink, not only we would be reinserting a dead cache
3302 * into the queue, but doing so from inside the worker racing to
3305 * So if we aren't down to zero, we'll just schedule a worker and try
3308 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0)
3309 kmem_cache_shrink(cachep
);
3311 kmem_cache_destroy(cachep
);
3314 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3316 if (!cachep
->memcg_params
->dead
)
3320 * There are many ways in which we can get here.
3322 * We can get to a memory-pressure situation while the delayed work is
3323 * still pending to run. The vmscan shrinkers can then release all
3324 * cache memory and get us to destruction. If this is the case, we'll
3325 * be executed twice, which is a bug (the second time will execute over
3326 * bogus data). In this case, cancelling the work should be fine.
3328 * But we can also get here from the worker itself, if
3329 * kmem_cache_shrink is enough to shake all the remaining objects and
3330 * get the page count to 0. In this case, we'll deadlock if we try to
3331 * cancel the work (the worker runs with an internal lock held, which
3332 * is the same lock we would hold for cancel_work_sync().)
3334 * Since we can't possibly know who got us here, just refrain from
3335 * running if there is already work pending
3337 if (work_pending(&cachep
->memcg_params
->destroy
))
3340 * We have to defer the actual destroying to a workqueue, because
3341 * we might currently be in a context that cannot sleep.
3343 schedule_work(&cachep
->memcg_params
->destroy
);
3346 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3347 struct kmem_cache
*s
)
3349 struct kmem_cache
*new = NULL
;
3350 static char *tmp_path
= NULL
, *tmp_name
= NULL
;
3351 static DEFINE_MUTEX(mutex
); /* protects tmp_name */
3353 BUG_ON(!memcg_can_account_kmem(memcg
));
3357 * kmem_cache_create_memcg duplicates the given name and
3358 * cgroup_name for this name requires RCU context.
3359 * This static temporary buffer is used to prevent from
3360 * pointless shortliving allocation.
3362 if (!tmp_path
|| !tmp_name
) {
3364 tmp_path
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3366 tmp_name
= kmalloc(NAME_MAX
+ 1, GFP_KERNEL
);
3367 if (!tmp_path
|| !tmp_name
)
3371 cgroup_name(memcg
->css
.cgroup
, tmp_name
, NAME_MAX
+ 1);
3372 snprintf(tmp_path
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3373 memcg_cache_id(memcg
), tmp_name
);
3375 new = kmem_cache_create_memcg(memcg
, tmp_path
, s
->object_size
, s
->align
,
3376 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3378 new->allocflags
|= __GFP_KMEMCG
;
3382 mutex_unlock(&mutex
);
3386 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3388 struct kmem_cache
*c
;
3391 if (!s
->memcg_params
)
3393 if (!s
->memcg_params
->is_root_cache
)
3397 * If the cache is being destroyed, we trust that there is no one else
3398 * requesting objects from it. Even if there are, the sanity checks in
3399 * kmem_cache_destroy should caught this ill-case.
3401 * Still, we don't want anyone else freeing memcg_caches under our
3402 * noses, which can happen if a new memcg comes to life. As usual,
3403 * we'll take the activate_kmem_mutex to protect ourselves against
3406 mutex_lock(&activate_kmem_mutex
);
3407 for_each_memcg_cache_index(i
) {
3408 c
= cache_from_memcg_idx(s
, i
);
3413 * We will now manually delete the caches, so to avoid races
3414 * we need to cancel all pending destruction workers and
3415 * proceed with destruction ourselves.
3417 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3418 * and that could spawn the workers again: it is likely that
3419 * the cache still have active pages until this very moment.
3420 * This would lead us back to mem_cgroup_destroy_cache.
3422 * But that will not execute at all if the "dead" flag is not
3423 * set, so flip it down to guarantee we are in control.
3425 c
->memcg_params
->dead
= false;
3426 cancel_work_sync(&c
->memcg_params
->destroy
);
3427 kmem_cache_destroy(c
);
3429 mutex_unlock(&activate_kmem_mutex
);
3432 struct create_work
{
3433 struct mem_cgroup
*memcg
;
3434 struct kmem_cache
*cachep
;
3435 struct work_struct work
;
3438 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3440 struct kmem_cache
*cachep
;
3441 struct memcg_cache_params
*params
;
3443 if (!memcg_kmem_is_active(memcg
))
3446 mutex_lock(&memcg
->slab_caches_mutex
);
3447 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3448 cachep
= memcg_params_to_cache(params
);
3449 cachep
->memcg_params
->dead
= true;
3450 schedule_work(&cachep
->memcg_params
->destroy
);
3452 mutex_unlock(&memcg
->slab_caches_mutex
);
3455 static void memcg_create_cache_work_func(struct work_struct
*w
)
3457 struct create_work
*cw
;
3459 cw
= container_of(w
, struct create_work
, work
);
3460 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3461 css_put(&cw
->memcg
->css
);
3466 * Enqueue the creation of a per-memcg kmem_cache.
3468 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3469 struct kmem_cache
*cachep
)
3471 struct create_work
*cw
;
3473 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3475 css_put(&memcg
->css
);
3480 cw
->cachep
= cachep
;
3482 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3483 schedule_work(&cw
->work
);
3486 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3487 struct kmem_cache
*cachep
)
3490 * We need to stop accounting when we kmalloc, because if the
3491 * corresponding kmalloc cache is not yet created, the first allocation
3492 * in __memcg_create_cache_enqueue will recurse.
3494 * However, it is better to enclose the whole function. Depending on
3495 * the debugging options enabled, INIT_WORK(), for instance, can
3496 * trigger an allocation. This too, will make us recurse. Because at
3497 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3498 * the safest choice is to do it like this, wrapping the whole function.
3500 memcg_stop_kmem_account();
3501 __memcg_create_cache_enqueue(memcg
, cachep
);
3502 memcg_resume_kmem_account();
3505 * Return the kmem_cache we're supposed to use for a slab allocation.
3506 * We try to use the current memcg's version of the cache.
3508 * If the cache does not exist yet, if we are the first user of it,
3509 * we either create it immediately, if possible, or create it asynchronously
3511 * In the latter case, we will let the current allocation go through with
3512 * the original cache.
3514 * Can't be called in interrupt context or from kernel threads.
3515 * This function needs to be called with rcu_read_lock() held.
3517 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3520 struct mem_cgroup
*memcg
;
3521 struct kmem_cache
*memcg_cachep
;
3523 VM_BUG_ON(!cachep
->memcg_params
);
3524 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3526 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3530 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3532 if (!memcg_can_account_kmem(memcg
))
3535 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3536 if (likely(memcg_cachep
)) {
3537 cachep
= memcg_cachep
;
3541 /* The corresponding put will be done in the workqueue. */
3542 if (!css_tryget(&memcg
->css
))
3547 * If we are in a safe context (can wait, and not in interrupt
3548 * context), we could be be predictable and return right away.
3549 * This would guarantee that the allocation being performed
3550 * already belongs in the new cache.
3552 * However, there are some clashes that can arrive from locking.
3553 * For instance, because we acquire the slab_mutex while doing
3554 * kmem_cache_dup, this means no further allocation could happen
3555 * with the slab_mutex held.
3557 * Also, because cache creation issue get_online_cpus(), this
3558 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3559 * that ends up reversed during cpu hotplug. (cpuset allocates
3560 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3561 * better to defer everything.
3563 memcg_create_cache_enqueue(memcg
, cachep
);
3569 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3572 * We need to verify if the allocation against current->mm->owner's memcg is
3573 * possible for the given order. But the page is not allocated yet, so we'll
3574 * need a further commit step to do the final arrangements.
3576 * It is possible for the task to switch cgroups in this mean time, so at
3577 * commit time, we can't rely on task conversion any longer. We'll then use
3578 * the handle argument to return to the caller which cgroup we should commit
3579 * against. We could also return the memcg directly and avoid the pointer
3580 * passing, but a boolean return value gives better semantics considering
3581 * the compiled-out case as well.
3583 * Returning true means the allocation is possible.
3586 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3588 struct mem_cgroup
*memcg
;
3594 * Disabling accounting is only relevant for some specific memcg
3595 * internal allocations. Therefore we would initially not have such
3596 * check here, since direct calls to the page allocator that are marked
3597 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3598 * concerned with cache allocations, and by having this test at
3599 * memcg_kmem_get_cache, we are already able to relay the allocation to
3600 * the root cache and bypass the memcg cache altogether.
3602 * There is one exception, though: the SLUB allocator does not create
3603 * large order caches, but rather service large kmallocs directly from
3604 * the page allocator. Therefore, the following sequence when backed by
3605 * the SLUB allocator:
3607 * memcg_stop_kmem_account();
3608 * kmalloc(<large_number>)
3609 * memcg_resume_kmem_account();
3611 * would effectively ignore the fact that we should skip accounting,
3612 * since it will drive us directly to this function without passing
3613 * through the cache selector memcg_kmem_get_cache. Such large
3614 * allocations are extremely rare but can happen, for instance, for the
3615 * cache arrays. We bring this test here.
3617 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3620 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3623 * very rare case described in mem_cgroup_from_task. Unfortunately there
3624 * isn't much we can do without complicating this too much, and it would
3625 * be gfp-dependent anyway. Just let it go
3627 if (unlikely(!memcg
))
3630 if (!memcg_can_account_kmem(memcg
)) {
3631 css_put(&memcg
->css
);
3635 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3639 css_put(&memcg
->css
);
3643 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3646 struct page_cgroup
*pc
;
3648 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3650 /* The page allocation failed. Revert */
3652 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3656 pc
= lookup_page_cgroup(page
);
3657 lock_page_cgroup(pc
);
3658 pc
->mem_cgroup
= memcg
;
3659 SetPageCgroupUsed(pc
);
3660 unlock_page_cgroup(pc
);
3663 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3665 struct mem_cgroup
*memcg
= NULL
;
3666 struct page_cgroup
*pc
;
3669 pc
= lookup_page_cgroup(page
);
3671 * Fast unlocked return. Theoretically might have changed, have to
3672 * check again after locking.
3674 if (!PageCgroupUsed(pc
))
3677 lock_page_cgroup(pc
);
3678 if (PageCgroupUsed(pc
)) {
3679 memcg
= pc
->mem_cgroup
;
3680 ClearPageCgroupUsed(pc
);
3682 unlock_page_cgroup(pc
);
3685 * We trust that only if there is a memcg associated with the page, it
3686 * is a valid allocation
3691 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3692 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3695 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3698 #endif /* CONFIG_MEMCG_KMEM */
3700 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3702 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3704 * Because tail pages are not marked as "used", set it. We're under
3705 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3706 * charge/uncharge will be never happen and move_account() is done under
3707 * compound_lock(), so we don't have to take care of races.
3709 void mem_cgroup_split_huge_fixup(struct page
*head
)
3711 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3712 struct page_cgroup
*pc
;
3713 struct mem_cgroup
*memcg
;
3716 if (mem_cgroup_disabled())
3719 memcg
= head_pc
->mem_cgroup
;
3720 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3722 pc
->mem_cgroup
= memcg
;
3723 smp_wmb();/* see __commit_charge() */
3724 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3726 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3729 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3732 * mem_cgroup_move_account - move account of the page
3734 * @nr_pages: number of regular pages (>1 for huge pages)
3735 * @pc: page_cgroup of the page.
3736 * @from: mem_cgroup which the page is moved from.
3737 * @to: mem_cgroup which the page is moved to. @from != @to.
3739 * The caller must confirm following.
3740 * - page is not on LRU (isolate_page() is useful.)
3741 * - compound_lock is held when nr_pages > 1
3743 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3746 static int mem_cgroup_move_account(struct page
*page
,
3747 unsigned int nr_pages
,
3748 struct page_cgroup
*pc
,
3749 struct mem_cgroup
*from
,
3750 struct mem_cgroup
*to
)
3752 unsigned long flags
;
3754 bool anon
= PageAnon(page
);
3756 VM_BUG_ON(from
== to
);
3757 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3759 * The page is isolated from LRU. So, collapse function
3760 * will not handle this page. But page splitting can happen.
3761 * Do this check under compound_page_lock(). The caller should
3765 if (nr_pages
> 1 && !PageTransHuge(page
))
3768 lock_page_cgroup(pc
);
3771 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3774 move_lock_mem_cgroup(from
, &flags
);
3776 if (!anon
&& page_mapped(page
)) {
3777 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3779 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3783 if (PageWriteback(page
)) {
3784 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3786 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3790 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3792 /* caller should have done css_get */
3793 pc
->mem_cgroup
= to
;
3794 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3795 move_unlock_mem_cgroup(from
, &flags
);
3798 unlock_page_cgroup(pc
);
3802 memcg_check_events(to
, page
);
3803 memcg_check_events(from
, page
);
3809 * mem_cgroup_move_parent - moves page to the parent group
3810 * @page: the page to move
3811 * @pc: page_cgroup of the page
3812 * @child: page's cgroup
3814 * move charges to its parent or the root cgroup if the group has no
3815 * parent (aka use_hierarchy==0).
3816 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3817 * mem_cgroup_move_account fails) the failure is always temporary and
3818 * it signals a race with a page removal/uncharge or migration. In the
3819 * first case the page is on the way out and it will vanish from the LRU
3820 * on the next attempt and the call should be retried later.
3821 * Isolation from the LRU fails only if page has been isolated from
3822 * the LRU since we looked at it and that usually means either global
3823 * reclaim or migration going on. The page will either get back to the
3825 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3826 * (!PageCgroupUsed) or moved to a different group. The page will
3827 * disappear in the next attempt.
3829 static int mem_cgroup_move_parent(struct page
*page
,
3830 struct page_cgroup
*pc
,
3831 struct mem_cgroup
*child
)
3833 struct mem_cgroup
*parent
;
3834 unsigned int nr_pages
;
3835 unsigned long uninitialized_var(flags
);
3838 VM_BUG_ON(mem_cgroup_is_root(child
));
3841 if (!get_page_unless_zero(page
))
3843 if (isolate_lru_page(page
))
3846 nr_pages
= hpage_nr_pages(page
);
3848 parent
= parent_mem_cgroup(child
);
3850 * If no parent, move charges to root cgroup.
3853 parent
= root_mem_cgroup
;
3856 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3857 flags
= compound_lock_irqsave(page
);
3860 ret
= mem_cgroup_move_account(page
, nr_pages
,
3863 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3866 compound_unlock_irqrestore(page
, flags
);
3867 putback_lru_page(page
);
3874 int mem_cgroup_newpage_charge(struct page
*page
,
3875 struct mm_struct
*mm
, gfp_t gfp_mask
)
3877 struct mem_cgroup
*memcg
= NULL
;
3878 unsigned int nr_pages
= 1;
3882 if (mem_cgroup_disabled())
3885 VM_BUG_ON_PAGE(page_mapped(page
), page
);
3886 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
3889 if (PageTransHuge(page
)) {
3890 nr_pages
<<= compound_order(page
);
3891 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3893 * Never OOM-kill a process for a huge page. The
3894 * fault handler will fall back to regular pages.
3899 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3902 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
,
3903 MEM_CGROUP_CHARGE_TYPE_ANON
, false);
3908 * While swap-in, try_charge -> commit or cancel, the page is locked.
3909 * And when try_charge() successfully returns, one refcnt to memcg without
3910 * struct page_cgroup is acquired. This refcnt will be consumed by
3911 * "commit()" or removed by "cancel()"
3913 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3916 struct mem_cgroup
**memcgp
)
3918 struct mem_cgroup
*memcg
;
3919 struct page_cgroup
*pc
;
3922 pc
= lookup_page_cgroup(page
);
3924 * Every swap fault against a single page tries to charge the
3925 * page, bail as early as possible. shmem_unuse() encounters
3926 * already charged pages, too. The USED bit is protected by
3927 * the page lock, which serializes swap cache removal, which
3928 * in turn serializes uncharging.
3930 if (PageCgroupUsed(pc
))
3932 if (!do_swap_account
)
3934 memcg
= try_get_mem_cgroup_from_page(page
);
3938 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3939 css_put(&memcg
->css
);
3944 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3950 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3951 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3954 if (mem_cgroup_disabled())
3957 * A racing thread's fault, or swapoff, may have already
3958 * updated the pte, and even removed page from swap cache: in
3959 * those cases unuse_pte()'s pte_same() test will fail; but
3960 * there's also a KSM case which does need to charge the page.
3962 if (!PageSwapCache(page
)) {
3965 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3970 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3973 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3975 if (mem_cgroup_disabled())
3979 __mem_cgroup_cancel_charge(memcg
, 1);
3983 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3984 enum charge_type ctype
)
3986 if (mem_cgroup_disabled())
3991 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3993 * Now swap is on-memory. This means this page may be
3994 * counted both as mem and swap....double count.
3995 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3996 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3997 * may call delete_from_swap_cache() before reach here.
3999 if (do_swap_account
&& PageSwapCache(page
)) {
4000 swp_entry_t ent
= {.val
= page_private(page
)};
4001 mem_cgroup_uncharge_swap(ent
);
4005 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4006 struct mem_cgroup
*memcg
)
4008 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4009 MEM_CGROUP_CHARGE_TYPE_ANON
);
4012 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4015 struct mem_cgroup
*memcg
= NULL
;
4016 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4019 if (mem_cgroup_disabled())
4021 if (PageCompound(page
))
4024 if (!PageSwapCache(page
)) {
4026 * Page cache insertions can happen without an actual
4027 * task context, e.g. during disk probing on boot.
4030 memcg
= root_mem_cgroup
;
4031 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, &memcg
, true);
4033 __mem_cgroup_commit_charge(memcg
, page
, 1, type
, false);
4034 } else { /* page is swapcache/shmem */
4035 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4038 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4043 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4044 unsigned int nr_pages
,
4045 const enum charge_type ctype
)
4047 struct memcg_batch_info
*batch
= NULL
;
4048 bool uncharge_memsw
= true;
4050 /* If swapout, usage of swap doesn't decrease */
4051 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4052 uncharge_memsw
= false;
4054 batch
= ¤t
->memcg_batch
;
4056 * In usual, we do css_get() when we remember memcg pointer.
4057 * But in this case, we keep res->usage until end of a series of
4058 * uncharges. Then, it's ok to ignore memcg's refcnt.
4061 batch
->memcg
= memcg
;
4063 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4064 * In those cases, all pages freed continuously can be expected to be in
4065 * the same cgroup and we have chance to coalesce uncharges.
4066 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4067 * because we want to do uncharge as soon as possible.
4070 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4071 goto direct_uncharge
;
4074 goto direct_uncharge
;
4077 * In typical case, batch->memcg == mem. This means we can
4078 * merge a series of uncharges to an uncharge of res_counter.
4079 * If not, we uncharge res_counter ony by one.
4081 if (batch
->memcg
!= memcg
)
4082 goto direct_uncharge
;
4083 /* remember freed charge and uncharge it later */
4086 batch
->memsw_nr_pages
++;
4089 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4091 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4092 if (unlikely(batch
->memcg
!= memcg
))
4093 memcg_oom_recover(memcg
);
4097 * uncharge if !page_mapped(page)
4099 static struct mem_cgroup
*
4100 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4103 struct mem_cgroup
*memcg
= NULL
;
4104 unsigned int nr_pages
= 1;
4105 struct page_cgroup
*pc
;
4108 if (mem_cgroup_disabled())
4111 if (PageTransHuge(page
)) {
4112 nr_pages
<<= compound_order(page
);
4113 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
4116 * Check if our page_cgroup is valid
4118 pc
= lookup_page_cgroup(page
);
4119 if (unlikely(!PageCgroupUsed(pc
)))
4122 lock_page_cgroup(pc
);
4124 memcg
= pc
->mem_cgroup
;
4126 if (!PageCgroupUsed(pc
))
4129 anon
= PageAnon(page
);
4132 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4134 * Generally PageAnon tells if it's the anon statistics to be
4135 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4136 * used before page reached the stage of being marked PageAnon.
4140 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4141 /* See mem_cgroup_prepare_migration() */
4142 if (page_mapped(page
))
4145 * Pages under migration may not be uncharged. But
4146 * end_migration() /must/ be the one uncharging the
4147 * unused post-migration page and so it has to call
4148 * here with the migration bit still set. See the
4149 * res_counter handling below.
4151 if (!end_migration
&& PageCgroupMigration(pc
))
4154 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4155 if (!PageAnon(page
)) { /* Shared memory */
4156 if (page
->mapping
&& !page_is_file_cache(page
))
4158 } else if (page_mapped(page
)) /* Anon */
4165 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4167 ClearPageCgroupUsed(pc
);
4169 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4170 * freed from LRU. This is safe because uncharged page is expected not
4171 * to be reused (freed soon). Exception is SwapCache, it's handled by
4172 * special functions.
4175 unlock_page_cgroup(pc
);
4177 * even after unlock, we have memcg->res.usage here and this memcg
4178 * will never be freed, so it's safe to call css_get().
4180 memcg_check_events(memcg
, page
);
4181 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4182 mem_cgroup_swap_statistics(memcg
, true);
4183 css_get(&memcg
->css
);
4186 * Migration does not charge the res_counter for the
4187 * replacement page, so leave it alone when phasing out the
4188 * page that is unused after the migration.
4190 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4191 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4196 unlock_page_cgroup(pc
);
4200 void mem_cgroup_uncharge_page(struct page
*page
)
4203 if (page_mapped(page
))
4205 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
4207 * If the page is in swap cache, uncharge should be deferred
4208 * to the swap path, which also properly accounts swap usage
4209 * and handles memcg lifetime.
4211 * Note that this check is not stable and reclaim may add the
4212 * page to swap cache at any time after this. However, if the
4213 * page is not in swap cache by the time page->mapcount hits
4214 * 0, there won't be any page table references to the swap
4215 * slot, and reclaim will free it and not actually write the
4218 if (PageSwapCache(page
))
4220 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4223 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4225 VM_BUG_ON_PAGE(page_mapped(page
), page
);
4226 VM_BUG_ON_PAGE(page
->mapping
, page
);
4227 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4231 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4232 * In that cases, pages are freed continuously and we can expect pages
4233 * are in the same memcg. All these calls itself limits the number of
4234 * pages freed at once, then uncharge_start/end() is called properly.
4235 * This may be called prural(2) times in a context,
4238 void mem_cgroup_uncharge_start(void)
4240 current
->memcg_batch
.do_batch
++;
4241 /* We can do nest. */
4242 if (current
->memcg_batch
.do_batch
== 1) {
4243 current
->memcg_batch
.memcg
= NULL
;
4244 current
->memcg_batch
.nr_pages
= 0;
4245 current
->memcg_batch
.memsw_nr_pages
= 0;
4249 void mem_cgroup_uncharge_end(void)
4251 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4253 if (!batch
->do_batch
)
4257 if (batch
->do_batch
) /* If stacked, do nothing. */
4263 * This "batch->memcg" is valid without any css_get/put etc...
4264 * bacause we hide charges behind us.
4266 if (batch
->nr_pages
)
4267 res_counter_uncharge(&batch
->memcg
->res
,
4268 batch
->nr_pages
* PAGE_SIZE
);
4269 if (batch
->memsw_nr_pages
)
4270 res_counter_uncharge(&batch
->memcg
->memsw
,
4271 batch
->memsw_nr_pages
* PAGE_SIZE
);
4272 memcg_oom_recover(batch
->memcg
);
4273 /* forget this pointer (for sanity check) */
4274 batch
->memcg
= NULL
;
4279 * called after __delete_from_swap_cache() and drop "page" account.
4280 * memcg information is recorded to swap_cgroup of "ent"
4283 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4285 struct mem_cgroup
*memcg
;
4286 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4288 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4289 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4291 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4294 * record memcg information, if swapout && memcg != NULL,
4295 * css_get() was called in uncharge().
4297 if (do_swap_account
&& swapout
&& memcg
)
4298 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4302 #ifdef CONFIG_MEMCG_SWAP
4304 * called from swap_entry_free(). remove record in swap_cgroup and
4305 * uncharge "memsw" account.
4307 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4309 struct mem_cgroup
*memcg
;
4312 if (!do_swap_account
)
4315 id
= swap_cgroup_record(ent
, 0);
4317 memcg
= mem_cgroup_lookup(id
);
4320 * We uncharge this because swap is freed.
4321 * This memcg can be obsolete one. We avoid calling css_tryget
4323 if (!mem_cgroup_is_root(memcg
))
4324 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4325 mem_cgroup_swap_statistics(memcg
, false);
4326 css_put(&memcg
->css
);
4332 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4333 * @entry: swap entry to be moved
4334 * @from: mem_cgroup which the entry is moved from
4335 * @to: mem_cgroup which the entry is moved to
4337 * It succeeds only when the swap_cgroup's record for this entry is the same
4338 * as the mem_cgroup's id of @from.
4340 * Returns 0 on success, -EINVAL on failure.
4342 * The caller must have charged to @to, IOW, called res_counter_charge() about
4343 * both res and memsw, and called css_get().
4345 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4346 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4348 unsigned short old_id
, new_id
;
4350 old_id
= mem_cgroup_id(from
);
4351 new_id
= mem_cgroup_id(to
);
4353 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4354 mem_cgroup_swap_statistics(from
, false);
4355 mem_cgroup_swap_statistics(to
, true);
4357 * This function is only called from task migration context now.
4358 * It postpones res_counter and refcount handling till the end
4359 * of task migration(mem_cgroup_clear_mc()) for performance
4360 * improvement. But we cannot postpone css_get(to) because if
4361 * the process that has been moved to @to does swap-in, the
4362 * refcount of @to might be decreased to 0.
4364 * We are in attach() phase, so the cgroup is guaranteed to be
4365 * alive, so we can just call css_get().
4373 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4374 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4381 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4384 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4385 struct mem_cgroup
**memcgp
)
4387 struct mem_cgroup
*memcg
= NULL
;
4388 unsigned int nr_pages
= 1;
4389 struct page_cgroup
*pc
;
4390 enum charge_type ctype
;
4394 if (mem_cgroup_disabled())
4397 if (PageTransHuge(page
))
4398 nr_pages
<<= compound_order(page
);
4400 pc
= lookup_page_cgroup(page
);
4401 lock_page_cgroup(pc
);
4402 if (PageCgroupUsed(pc
)) {
4403 memcg
= pc
->mem_cgroup
;
4404 css_get(&memcg
->css
);
4406 * At migrating an anonymous page, its mapcount goes down
4407 * to 0 and uncharge() will be called. But, even if it's fully
4408 * unmapped, migration may fail and this page has to be
4409 * charged again. We set MIGRATION flag here and delay uncharge
4410 * until end_migration() is called
4412 * Corner Case Thinking
4414 * When the old page was mapped as Anon and it's unmap-and-freed
4415 * while migration was ongoing.
4416 * If unmap finds the old page, uncharge() of it will be delayed
4417 * until end_migration(). If unmap finds a new page, it's
4418 * uncharged when it make mapcount to be 1->0. If unmap code
4419 * finds swap_migration_entry, the new page will not be mapped
4420 * and end_migration() will find it(mapcount==0).
4423 * When the old page was mapped but migraion fails, the kernel
4424 * remaps it. A charge for it is kept by MIGRATION flag even
4425 * if mapcount goes down to 0. We can do remap successfully
4426 * without charging it again.
4429 * The "old" page is under lock_page() until the end of
4430 * migration, so, the old page itself will not be swapped-out.
4431 * If the new page is swapped out before end_migraton, our
4432 * hook to usual swap-out path will catch the event.
4435 SetPageCgroupMigration(pc
);
4437 unlock_page_cgroup(pc
);
4439 * If the page is not charged at this point,
4447 * We charge new page before it's used/mapped. So, even if unlock_page()
4448 * is called before end_migration, we can catch all events on this new
4449 * page. In the case new page is migrated but not remapped, new page's
4450 * mapcount will be finally 0 and we call uncharge in end_migration().
4453 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4455 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4457 * The page is committed to the memcg, but it's not actually
4458 * charged to the res_counter since we plan on replacing the
4459 * old one and only one page is going to be left afterwards.
4461 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4464 /* remove redundant charge if migration failed*/
4465 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4466 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4468 struct page
*used
, *unused
;
4469 struct page_cgroup
*pc
;
4475 if (!migration_ok
) {
4482 anon
= PageAnon(used
);
4483 __mem_cgroup_uncharge_common(unused
,
4484 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4485 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4487 css_put(&memcg
->css
);
4489 * We disallowed uncharge of pages under migration because mapcount
4490 * of the page goes down to zero, temporarly.
4491 * Clear the flag and check the page should be charged.
4493 pc
= lookup_page_cgroup(oldpage
);
4494 lock_page_cgroup(pc
);
4495 ClearPageCgroupMigration(pc
);
4496 unlock_page_cgroup(pc
);
4499 * If a page is a file cache, radix-tree replacement is very atomic
4500 * and we can skip this check. When it was an Anon page, its mapcount
4501 * goes down to 0. But because we added MIGRATION flage, it's not
4502 * uncharged yet. There are several case but page->mapcount check
4503 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4504 * check. (see prepare_charge() also)
4507 mem_cgroup_uncharge_page(used
);
4511 * At replace page cache, newpage is not under any memcg but it's on
4512 * LRU. So, this function doesn't touch res_counter but handles LRU
4513 * in correct way. Both pages are locked so we cannot race with uncharge.
4515 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4516 struct page
*newpage
)
4518 struct mem_cgroup
*memcg
= NULL
;
4519 struct page_cgroup
*pc
;
4520 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4522 if (mem_cgroup_disabled())
4525 pc
= lookup_page_cgroup(oldpage
);
4526 /* fix accounting on old pages */
4527 lock_page_cgroup(pc
);
4528 if (PageCgroupUsed(pc
)) {
4529 memcg
= pc
->mem_cgroup
;
4530 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4531 ClearPageCgroupUsed(pc
);
4533 unlock_page_cgroup(pc
);
4536 * When called from shmem_replace_page(), in some cases the
4537 * oldpage has already been charged, and in some cases not.
4542 * Even if newpage->mapping was NULL before starting replacement,
4543 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4544 * LRU while we overwrite pc->mem_cgroup.
4546 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4549 #ifdef CONFIG_DEBUG_VM
4550 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4552 struct page_cgroup
*pc
;
4554 pc
= lookup_page_cgroup(page
);
4556 * Can be NULL while feeding pages into the page allocator for
4557 * the first time, i.e. during boot or memory hotplug;
4558 * or when mem_cgroup_disabled().
4560 if (likely(pc
) && PageCgroupUsed(pc
))
4565 bool mem_cgroup_bad_page_check(struct page
*page
)
4567 if (mem_cgroup_disabled())
4570 return lookup_page_cgroup_used(page
) != NULL
;
4573 void mem_cgroup_print_bad_page(struct page
*page
)
4575 struct page_cgroup
*pc
;
4577 pc
= lookup_page_cgroup_used(page
);
4579 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4580 pc
, pc
->flags
, pc
->mem_cgroup
);
4585 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4586 unsigned long long val
)
4589 u64 memswlimit
, memlimit
;
4591 int children
= mem_cgroup_count_children(memcg
);
4592 u64 curusage
, oldusage
;
4596 * For keeping hierarchical_reclaim simple, how long we should retry
4597 * is depends on callers. We set our retry-count to be function
4598 * of # of children which we should visit in this loop.
4600 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4602 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4605 while (retry_count
) {
4606 if (signal_pending(current
)) {
4611 * Rather than hide all in some function, I do this in
4612 * open coded manner. You see what this really does.
4613 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4615 mutex_lock(&set_limit_mutex
);
4616 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4617 if (memswlimit
< val
) {
4619 mutex_unlock(&set_limit_mutex
);
4623 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4627 ret
= res_counter_set_limit(&memcg
->res
, val
);
4629 if (memswlimit
== val
)
4630 memcg
->memsw_is_minimum
= true;
4632 memcg
->memsw_is_minimum
= false;
4634 mutex_unlock(&set_limit_mutex
);
4639 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4640 MEM_CGROUP_RECLAIM_SHRINK
);
4641 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4642 /* Usage is reduced ? */
4643 if (curusage
>= oldusage
)
4646 oldusage
= curusage
;
4648 if (!ret
&& enlarge
)
4649 memcg_oom_recover(memcg
);
4654 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4655 unsigned long long val
)
4658 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4659 int children
= mem_cgroup_count_children(memcg
);
4663 /* see mem_cgroup_resize_res_limit */
4664 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4665 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4666 while (retry_count
) {
4667 if (signal_pending(current
)) {
4672 * Rather than hide all in some function, I do this in
4673 * open coded manner. You see what this really does.
4674 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4676 mutex_lock(&set_limit_mutex
);
4677 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4678 if (memlimit
> val
) {
4680 mutex_unlock(&set_limit_mutex
);
4683 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4684 if (memswlimit
< val
)
4686 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4688 if (memlimit
== val
)
4689 memcg
->memsw_is_minimum
= true;
4691 memcg
->memsw_is_minimum
= false;
4693 mutex_unlock(&set_limit_mutex
);
4698 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4699 MEM_CGROUP_RECLAIM_NOSWAP
|
4700 MEM_CGROUP_RECLAIM_SHRINK
);
4701 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4702 /* Usage is reduced ? */
4703 if (curusage
>= oldusage
)
4706 oldusage
= curusage
;
4708 if (!ret
&& enlarge
)
4709 memcg_oom_recover(memcg
);
4713 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4715 unsigned long *total_scanned
)
4717 unsigned long nr_reclaimed
= 0;
4718 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4719 unsigned long reclaimed
;
4721 struct mem_cgroup_tree_per_zone
*mctz
;
4722 unsigned long long excess
;
4723 unsigned long nr_scanned
;
4728 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4730 * This loop can run a while, specially if mem_cgroup's continuously
4731 * keep exceeding their soft limit and putting the system under
4738 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4743 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4744 gfp_mask
, &nr_scanned
);
4745 nr_reclaimed
+= reclaimed
;
4746 *total_scanned
+= nr_scanned
;
4747 spin_lock(&mctz
->lock
);
4750 * If we failed to reclaim anything from this memory cgroup
4751 * it is time to move on to the next cgroup
4757 * Loop until we find yet another one.
4759 * By the time we get the soft_limit lock
4760 * again, someone might have aded the
4761 * group back on the RB tree. Iterate to
4762 * make sure we get a different mem.
4763 * mem_cgroup_largest_soft_limit_node returns
4764 * NULL if no other cgroup is present on
4768 __mem_cgroup_largest_soft_limit_node(mctz
);
4770 css_put(&next_mz
->memcg
->css
);
4771 else /* next_mz == NULL or other memcg */
4775 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4776 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4778 * One school of thought says that we should not add
4779 * back the node to the tree if reclaim returns 0.
4780 * But our reclaim could return 0, simply because due
4781 * to priority we are exposing a smaller subset of
4782 * memory to reclaim from. Consider this as a longer
4785 /* If excess == 0, no tree ops */
4786 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4787 spin_unlock(&mctz
->lock
);
4788 css_put(&mz
->memcg
->css
);
4791 * Could not reclaim anything and there are no more
4792 * mem cgroups to try or we seem to be looping without
4793 * reclaiming anything.
4795 if (!nr_reclaimed
&&
4797 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4799 } while (!nr_reclaimed
);
4801 css_put(&next_mz
->memcg
->css
);
4802 return nr_reclaimed
;
4806 * mem_cgroup_force_empty_list - clears LRU of a group
4807 * @memcg: group to clear
4810 * @lru: lru to to clear
4812 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4813 * reclaim the pages page themselves - pages are moved to the parent (or root)
4816 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4817 int node
, int zid
, enum lru_list lru
)
4819 struct lruvec
*lruvec
;
4820 unsigned long flags
;
4821 struct list_head
*list
;
4825 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4826 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4827 list
= &lruvec
->lists
[lru
];
4831 struct page_cgroup
*pc
;
4834 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4835 if (list_empty(list
)) {
4836 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4839 page
= list_entry(list
->prev
, struct page
, lru
);
4841 list_move(&page
->lru
, list
);
4843 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4846 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4848 pc
= lookup_page_cgroup(page
);
4850 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4851 /* found lock contention or "pc" is obsolete. */
4856 } while (!list_empty(list
));
4860 * make mem_cgroup's charge to be 0 if there is no task by moving
4861 * all the charges and pages to the parent.
4862 * This enables deleting this mem_cgroup.
4864 * Caller is responsible for holding css reference on the memcg.
4866 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4872 /* This is for making all *used* pages to be on LRU. */
4873 lru_add_drain_all();
4874 drain_all_stock_sync(memcg
);
4875 mem_cgroup_start_move(memcg
);
4876 for_each_node_state(node
, N_MEMORY
) {
4877 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4880 mem_cgroup_force_empty_list(memcg
,
4885 mem_cgroup_end_move(memcg
);
4886 memcg_oom_recover(memcg
);
4890 * Kernel memory may not necessarily be trackable to a specific
4891 * process. So they are not migrated, and therefore we can't
4892 * expect their value to drop to 0 here.
4893 * Having res filled up with kmem only is enough.
4895 * This is a safety check because mem_cgroup_force_empty_list
4896 * could have raced with mem_cgroup_replace_page_cache callers
4897 * so the lru seemed empty but the page could have been added
4898 * right after the check. RES_USAGE should be safe as we always
4899 * charge before adding to the LRU.
4901 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4902 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4903 } while (usage
> 0);
4906 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4908 lockdep_assert_held(&memcg_create_mutex
);
4910 * The lock does not prevent addition or deletion to the list
4911 * of children, but it prevents a new child from being
4912 * initialized based on this parent in css_online(), so it's
4913 * enough to decide whether hierarchically inherited
4914 * attributes can still be changed or not.
4916 return memcg
->use_hierarchy
&&
4917 !list_empty(&memcg
->css
.cgroup
->children
);
4921 * Reclaims as many pages from the given memcg as possible and moves
4922 * the rest to the parent.
4924 * Caller is responsible for holding css reference for memcg.
4926 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4928 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4929 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4931 /* returns EBUSY if there is a task or if we come here twice. */
4932 if (cgroup_has_tasks(cgrp
) || !list_empty(&cgrp
->children
))
4935 /* we call try-to-free pages for make this cgroup empty */
4936 lru_add_drain_all();
4937 /* try to free all pages in this cgroup */
4938 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4941 if (signal_pending(current
))
4944 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4948 /* maybe some writeback is necessary */
4949 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4954 mem_cgroup_reparent_charges(memcg
);
4959 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4962 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4964 if (mem_cgroup_is_root(memcg
))
4966 return mem_cgroup_force_empty(memcg
);
4969 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4972 return mem_cgroup_from_css(css
)->use_hierarchy
;
4975 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4976 struct cftype
*cft
, u64 val
)
4979 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4980 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4982 mutex_lock(&memcg_create_mutex
);
4984 if (memcg
->use_hierarchy
== val
)
4988 * If parent's use_hierarchy is set, we can't make any modifications
4989 * in the child subtrees. If it is unset, then the change can
4990 * occur, provided the current cgroup has no children.
4992 * For the root cgroup, parent_mem is NULL, we allow value to be
4993 * set if there are no children.
4995 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4996 (val
== 1 || val
== 0)) {
4997 if (list_empty(&memcg
->css
.cgroup
->children
))
4998 memcg
->use_hierarchy
= val
;
5005 mutex_unlock(&memcg_create_mutex
);
5011 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5012 enum mem_cgroup_stat_index idx
)
5014 struct mem_cgroup
*iter
;
5017 /* Per-cpu values can be negative, use a signed accumulator */
5018 for_each_mem_cgroup_tree(iter
, memcg
)
5019 val
+= mem_cgroup_read_stat(iter
, idx
);
5021 if (val
< 0) /* race ? */
5026 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5030 if (!mem_cgroup_is_root(memcg
)) {
5032 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5034 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5038 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5039 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5041 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5042 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5045 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5047 return val
<< PAGE_SHIFT
;
5050 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
5053 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5058 type
= MEMFILE_TYPE(cft
->private);
5059 name
= MEMFILE_ATTR(cft
->private);
5063 if (name
== RES_USAGE
)
5064 val
= mem_cgroup_usage(memcg
, false);
5066 val
= res_counter_read_u64(&memcg
->res
, name
);
5069 if (name
== RES_USAGE
)
5070 val
= mem_cgroup_usage(memcg
, true);
5072 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5075 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5084 #ifdef CONFIG_MEMCG_KMEM
5085 /* should be called with activate_kmem_mutex held */
5086 static int __memcg_activate_kmem(struct mem_cgroup
*memcg
,
5087 unsigned long long limit
)
5092 if (memcg_kmem_is_active(memcg
))
5096 * We are going to allocate memory for data shared by all memory
5097 * cgroups so let's stop accounting here.
5099 memcg_stop_kmem_account();
5102 * For simplicity, we won't allow this to be disabled. It also can't
5103 * be changed if the cgroup has children already, or if tasks had
5106 * If tasks join before we set the limit, a person looking at
5107 * kmem.usage_in_bytes will have no way to determine when it took
5108 * place, which makes the value quite meaningless.
5110 * After it first became limited, changes in the value of the limit are
5111 * of course permitted.
5113 mutex_lock(&memcg_create_mutex
);
5114 if (cgroup_has_tasks(memcg
->css
.cgroup
) || memcg_has_children(memcg
))
5116 mutex_unlock(&memcg_create_mutex
);
5120 memcg_id
= ida_simple_get(&kmem_limited_groups
,
5121 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
5128 * Make sure we have enough space for this cgroup in each root cache's
5131 err
= memcg_update_all_caches(memcg_id
+ 1);
5135 memcg
->kmemcg_id
= memcg_id
;
5136 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
5137 mutex_init(&memcg
->slab_caches_mutex
);
5140 * We couldn't have accounted to this cgroup, because it hasn't got the
5141 * active bit set yet, so this should succeed.
5143 err
= res_counter_set_limit(&memcg
->kmem
, limit
);
5146 static_key_slow_inc(&memcg_kmem_enabled_key
);
5148 * Setting the active bit after enabling static branching will
5149 * guarantee no one starts accounting before all call sites are
5152 memcg_kmem_set_active(memcg
);
5154 memcg_resume_kmem_account();
5158 ida_simple_remove(&kmem_limited_groups
, memcg_id
);
5162 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
5163 unsigned long long limit
)
5167 mutex_lock(&activate_kmem_mutex
);
5168 ret
= __memcg_activate_kmem(memcg
, limit
);
5169 mutex_unlock(&activate_kmem_mutex
);
5173 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
5174 unsigned long long val
)
5178 if (!memcg_kmem_is_active(memcg
))
5179 ret
= memcg_activate_kmem(memcg
, val
);
5181 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5185 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5188 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5193 mutex_lock(&activate_kmem_mutex
);
5195 * If the parent cgroup is not kmem-active now, it cannot be activated
5196 * after this point, because it has at least one child already.
5198 if (memcg_kmem_is_active(parent
))
5199 ret
= __memcg_activate_kmem(memcg
, RES_COUNTER_MAX
);
5200 mutex_unlock(&activate_kmem_mutex
);
5204 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
5205 unsigned long long val
)
5209 #endif /* CONFIG_MEMCG_KMEM */
5212 * The user of this function is...
5215 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5218 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5221 unsigned long long val
;
5224 type
= MEMFILE_TYPE(cft
->private);
5225 name
= MEMFILE_ATTR(cft
->private);
5229 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5233 /* This function does all necessary parse...reuse it */
5234 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5238 ret
= mem_cgroup_resize_limit(memcg
, val
);
5239 else if (type
== _MEMSWAP
)
5240 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5241 else if (type
== _KMEM
)
5242 ret
= memcg_update_kmem_limit(memcg
, val
);
5246 case RES_SOFT_LIMIT
:
5247 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5251 * For memsw, soft limits are hard to implement in terms
5252 * of semantics, for now, we support soft limits for
5253 * control without swap
5256 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5261 ret
= -EINVAL
; /* should be BUG() ? */
5267 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5268 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5270 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5272 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5273 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5274 if (!memcg
->use_hierarchy
)
5277 while (css_parent(&memcg
->css
)) {
5278 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5279 if (!memcg
->use_hierarchy
)
5281 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5282 min_limit
= min(min_limit
, tmp
);
5283 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5284 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5287 *mem_limit
= min_limit
;
5288 *memsw_limit
= min_memsw_limit
;
5291 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5293 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5297 type
= MEMFILE_TYPE(event
);
5298 name
= MEMFILE_ATTR(event
);
5303 res_counter_reset_max(&memcg
->res
);
5304 else if (type
== _MEMSWAP
)
5305 res_counter_reset_max(&memcg
->memsw
);
5306 else if (type
== _KMEM
)
5307 res_counter_reset_max(&memcg
->kmem
);
5313 res_counter_reset_failcnt(&memcg
->res
);
5314 else if (type
== _MEMSWAP
)
5315 res_counter_reset_failcnt(&memcg
->memsw
);
5316 else if (type
== _KMEM
)
5317 res_counter_reset_failcnt(&memcg
->kmem
);
5326 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5329 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5333 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5334 struct cftype
*cft
, u64 val
)
5336 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5338 if (val
>= (1 << NR_MOVE_TYPE
))
5342 * No kind of locking is needed in here, because ->can_attach() will
5343 * check this value once in the beginning of the process, and then carry
5344 * on with stale data. This means that changes to this value will only
5345 * affect task migrations starting after the change.
5347 memcg
->move_charge_at_immigrate
= val
;
5351 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5352 struct cftype
*cft
, u64 val
)
5359 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
5363 unsigned int lru_mask
;
5366 static const struct numa_stat stats
[] = {
5367 { "total", LRU_ALL
},
5368 { "file", LRU_ALL_FILE
},
5369 { "anon", LRU_ALL_ANON
},
5370 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5372 const struct numa_stat
*stat
;
5375 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5377 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5378 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5379 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5380 for_each_node_state(nid
, N_MEMORY
) {
5381 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5383 seq_printf(m
, " N%d=%lu", nid
, nr
);
5388 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5389 struct mem_cgroup
*iter
;
5392 for_each_mem_cgroup_tree(iter
, memcg
)
5393 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5394 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5395 for_each_node_state(nid
, N_MEMORY
) {
5397 for_each_mem_cgroup_tree(iter
, memcg
)
5398 nr
+= mem_cgroup_node_nr_lru_pages(
5399 iter
, nid
, stat
->lru_mask
);
5400 seq_printf(m
, " N%d=%lu", nid
, nr
);
5407 #endif /* CONFIG_NUMA */
5409 static inline void mem_cgroup_lru_names_not_uptodate(void)
5411 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5414 static int memcg_stat_show(struct seq_file
*m
, void *v
)
5416 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5417 struct mem_cgroup
*mi
;
5420 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5421 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5423 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5424 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5427 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5428 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5429 mem_cgroup_read_events(memcg
, i
));
5431 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5432 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5433 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5435 /* Hierarchical information */
5437 unsigned long long limit
, memsw_limit
;
5438 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5439 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5440 if (do_swap_account
)
5441 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5445 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5448 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5450 for_each_mem_cgroup_tree(mi
, memcg
)
5451 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5452 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5455 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5456 unsigned long long val
= 0;
5458 for_each_mem_cgroup_tree(mi
, memcg
)
5459 val
+= mem_cgroup_read_events(mi
, i
);
5460 seq_printf(m
, "total_%s %llu\n",
5461 mem_cgroup_events_names
[i
], val
);
5464 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5465 unsigned long long val
= 0;
5467 for_each_mem_cgroup_tree(mi
, memcg
)
5468 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5469 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5472 #ifdef CONFIG_DEBUG_VM
5475 struct mem_cgroup_per_zone
*mz
;
5476 struct zone_reclaim_stat
*rstat
;
5477 unsigned long recent_rotated
[2] = {0, 0};
5478 unsigned long recent_scanned
[2] = {0, 0};
5480 for_each_online_node(nid
)
5481 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5482 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5483 rstat
= &mz
->lruvec
.reclaim_stat
;
5485 recent_rotated
[0] += rstat
->recent_rotated
[0];
5486 recent_rotated
[1] += rstat
->recent_rotated
[1];
5487 recent_scanned
[0] += rstat
->recent_scanned
[0];
5488 recent_scanned
[1] += rstat
->recent_scanned
[1];
5490 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5491 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5492 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5493 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5500 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5503 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5505 return mem_cgroup_swappiness(memcg
);
5508 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5509 struct cftype
*cft
, u64 val
)
5511 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5512 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5514 if (val
> 100 || !parent
)
5517 mutex_lock(&memcg_create_mutex
);
5519 /* If under hierarchy, only empty-root can set this value */
5520 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5521 mutex_unlock(&memcg_create_mutex
);
5525 memcg
->swappiness
= val
;
5527 mutex_unlock(&memcg_create_mutex
);
5532 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5534 struct mem_cgroup_threshold_ary
*t
;
5540 t
= rcu_dereference(memcg
->thresholds
.primary
);
5542 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5547 usage
= mem_cgroup_usage(memcg
, swap
);
5550 * current_threshold points to threshold just below or equal to usage.
5551 * If it's not true, a threshold was crossed after last
5552 * call of __mem_cgroup_threshold().
5554 i
= t
->current_threshold
;
5557 * Iterate backward over array of thresholds starting from
5558 * current_threshold and check if a threshold is crossed.
5559 * If none of thresholds below usage is crossed, we read
5560 * only one element of the array here.
5562 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5563 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5565 /* i = current_threshold + 1 */
5569 * Iterate forward over array of thresholds starting from
5570 * current_threshold+1 and check if a threshold is crossed.
5571 * If none of thresholds above usage is crossed, we read
5572 * only one element of the array here.
5574 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5575 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5577 /* Update current_threshold */
5578 t
->current_threshold
= i
- 1;
5583 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5586 __mem_cgroup_threshold(memcg
, false);
5587 if (do_swap_account
)
5588 __mem_cgroup_threshold(memcg
, true);
5590 memcg
= parent_mem_cgroup(memcg
);
5594 static int compare_thresholds(const void *a
, const void *b
)
5596 const struct mem_cgroup_threshold
*_a
= a
;
5597 const struct mem_cgroup_threshold
*_b
= b
;
5599 if (_a
->threshold
> _b
->threshold
)
5602 if (_a
->threshold
< _b
->threshold
)
5608 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5610 struct mem_cgroup_eventfd_list
*ev
;
5612 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5613 eventfd_signal(ev
->eventfd
, 1);
5617 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5619 struct mem_cgroup
*iter
;
5621 for_each_mem_cgroup_tree(iter
, memcg
)
5622 mem_cgroup_oom_notify_cb(iter
);
5625 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5626 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
5628 struct mem_cgroup_thresholds
*thresholds
;
5629 struct mem_cgroup_threshold_ary
*new;
5630 u64 threshold
, usage
;
5633 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5637 mutex_lock(&memcg
->thresholds_lock
);
5640 thresholds
= &memcg
->thresholds
;
5641 else if (type
== _MEMSWAP
)
5642 thresholds
= &memcg
->memsw_thresholds
;
5646 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5648 /* Check if a threshold crossed before adding a new one */
5649 if (thresholds
->primary
)
5650 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5652 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5654 /* Allocate memory for new array of thresholds */
5655 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5663 /* Copy thresholds (if any) to new array */
5664 if (thresholds
->primary
) {
5665 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5666 sizeof(struct mem_cgroup_threshold
));
5669 /* Add new threshold */
5670 new->entries
[size
- 1].eventfd
= eventfd
;
5671 new->entries
[size
- 1].threshold
= threshold
;
5673 /* Sort thresholds. Registering of new threshold isn't time-critical */
5674 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5675 compare_thresholds
, NULL
);
5677 /* Find current threshold */
5678 new->current_threshold
= -1;
5679 for (i
= 0; i
< size
; i
++) {
5680 if (new->entries
[i
].threshold
<= usage
) {
5682 * new->current_threshold will not be used until
5683 * rcu_assign_pointer(), so it's safe to increment
5686 ++new->current_threshold
;
5691 /* Free old spare buffer and save old primary buffer as spare */
5692 kfree(thresholds
->spare
);
5693 thresholds
->spare
= thresholds
->primary
;
5695 rcu_assign_pointer(thresholds
->primary
, new);
5697 /* To be sure that nobody uses thresholds */
5701 mutex_unlock(&memcg
->thresholds_lock
);
5706 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5707 struct eventfd_ctx
*eventfd
, const char *args
)
5709 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
5712 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5713 struct eventfd_ctx
*eventfd
, const char *args
)
5715 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
5718 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5719 struct eventfd_ctx
*eventfd
, enum res_type type
)
5721 struct mem_cgroup_thresholds
*thresholds
;
5722 struct mem_cgroup_threshold_ary
*new;
5726 mutex_lock(&memcg
->thresholds_lock
);
5728 thresholds
= &memcg
->thresholds
;
5729 else if (type
== _MEMSWAP
)
5730 thresholds
= &memcg
->memsw_thresholds
;
5734 if (!thresholds
->primary
)
5737 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5739 /* Check if a threshold crossed before removing */
5740 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5742 /* Calculate new number of threshold */
5744 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5745 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5749 new = thresholds
->spare
;
5751 /* Set thresholds array to NULL if we don't have thresholds */
5760 /* Copy thresholds and find current threshold */
5761 new->current_threshold
= -1;
5762 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5763 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5766 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5767 if (new->entries
[j
].threshold
<= usage
) {
5769 * new->current_threshold will not be used
5770 * until rcu_assign_pointer(), so it's safe to increment
5773 ++new->current_threshold
;
5779 /* Swap primary and spare array */
5780 thresholds
->spare
= thresholds
->primary
;
5781 /* If all events are unregistered, free the spare array */
5783 kfree(thresholds
->spare
);
5784 thresholds
->spare
= NULL
;
5787 rcu_assign_pointer(thresholds
->primary
, new);
5789 /* To be sure that nobody uses thresholds */
5792 mutex_unlock(&memcg
->thresholds_lock
);
5795 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5796 struct eventfd_ctx
*eventfd
)
5798 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
5801 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5802 struct eventfd_ctx
*eventfd
)
5804 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
5807 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
5808 struct eventfd_ctx
*eventfd
, const char *args
)
5810 struct mem_cgroup_eventfd_list
*event
;
5812 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5816 spin_lock(&memcg_oom_lock
);
5818 event
->eventfd
= eventfd
;
5819 list_add(&event
->list
, &memcg
->oom_notify
);
5821 /* already in OOM ? */
5822 if (atomic_read(&memcg
->under_oom
))
5823 eventfd_signal(eventfd
, 1);
5824 spin_unlock(&memcg_oom_lock
);
5829 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
5830 struct eventfd_ctx
*eventfd
)
5832 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5834 spin_lock(&memcg_oom_lock
);
5836 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5837 if (ev
->eventfd
== eventfd
) {
5838 list_del(&ev
->list
);
5843 spin_unlock(&memcg_oom_lock
);
5846 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
5848 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
5850 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
5851 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
5855 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5856 struct cftype
*cft
, u64 val
)
5858 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5859 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5861 /* cannot set to root cgroup and only 0 and 1 are allowed */
5862 if (!parent
|| !((val
== 0) || (val
== 1)))
5865 mutex_lock(&memcg_create_mutex
);
5866 /* oom-kill-disable is a flag for subhierarchy. */
5867 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5868 mutex_unlock(&memcg_create_mutex
);
5871 memcg
->oom_kill_disable
= val
;
5873 memcg_oom_recover(memcg
);
5874 mutex_unlock(&memcg_create_mutex
);
5878 #ifdef CONFIG_MEMCG_KMEM
5879 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5883 memcg
->kmemcg_id
= -1;
5884 ret
= memcg_propagate_kmem(memcg
);
5888 return mem_cgroup_sockets_init(memcg
, ss
);
5891 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5893 mem_cgroup_sockets_destroy(memcg
);
5896 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5898 if (!memcg_kmem_is_active(memcg
))
5902 * kmem charges can outlive the cgroup. In the case of slab
5903 * pages, for instance, a page contain objects from various
5904 * processes. As we prevent from taking a reference for every
5905 * such allocation we have to be careful when doing uncharge
5906 * (see memcg_uncharge_kmem) and here during offlining.
5908 * The idea is that that only the _last_ uncharge which sees
5909 * the dead memcg will drop the last reference. An additional
5910 * reference is taken here before the group is marked dead
5911 * which is then paired with css_put during uncharge resp. here.
5913 * Although this might sound strange as this path is called from
5914 * css_offline() when the referencemight have dropped down to 0
5915 * and shouldn't be incremented anymore (css_tryget would fail)
5916 * we do not have other options because of the kmem allocations
5919 css_get(&memcg
->css
);
5921 memcg_kmem_mark_dead(memcg
);
5923 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5926 if (memcg_kmem_test_and_clear_dead(memcg
))
5927 css_put(&memcg
->css
);
5930 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5935 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5939 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5945 * DO NOT USE IN NEW FILES.
5947 * "cgroup.event_control" implementation.
5949 * This is way over-engineered. It tries to support fully configurable
5950 * events for each user. Such level of flexibility is completely
5951 * unnecessary especially in the light of the planned unified hierarchy.
5953 * Please deprecate this and replace with something simpler if at all
5958 * Unregister event and free resources.
5960 * Gets called from workqueue.
5962 static void memcg_event_remove(struct work_struct
*work
)
5964 struct mem_cgroup_event
*event
=
5965 container_of(work
, struct mem_cgroup_event
, remove
);
5966 struct mem_cgroup
*memcg
= event
->memcg
;
5968 remove_wait_queue(event
->wqh
, &event
->wait
);
5970 event
->unregister_event(memcg
, event
->eventfd
);
5972 /* Notify userspace the event is going away. */
5973 eventfd_signal(event
->eventfd
, 1);
5975 eventfd_ctx_put(event
->eventfd
);
5977 css_put(&memcg
->css
);
5981 * Gets called on POLLHUP on eventfd when user closes it.
5983 * Called with wqh->lock held and interrupts disabled.
5985 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
5986 int sync
, void *key
)
5988 struct mem_cgroup_event
*event
=
5989 container_of(wait
, struct mem_cgroup_event
, wait
);
5990 struct mem_cgroup
*memcg
= event
->memcg
;
5991 unsigned long flags
= (unsigned long)key
;
5993 if (flags
& POLLHUP
) {
5995 * If the event has been detached at cgroup removal, we
5996 * can simply return knowing the other side will cleanup
5999 * We can't race against event freeing since the other
6000 * side will require wqh->lock via remove_wait_queue(),
6003 spin_lock(&memcg
->event_list_lock
);
6004 if (!list_empty(&event
->list
)) {
6005 list_del_init(&event
->list
);
6007 * We are in atomic context, but cgroup_event_remove()
6008 * may sleep, so we have to call it in workqueue.
6010 schedule_work(&event
->remove
);
6012 spin_unlock(&memcg
->event_list_lock
);
6018 static void memcg_event_ptable_queue_proc(struct file
*file
,
6019 wait_queue_head_t
*wqh
, poll_table
*pt
)
6021 struct mem_cgroup_event
*event
=
6022 container_of(pt
, struct mem_cgroup_event
, pt
);
6025 add_wait_queue(wqh
, &event
->wait
);
6029 * DO NOT USE IN NEW FILES.
6031 * Parse input and register new cgroup event handler.
6033 * Input must be in format '<event_fd> <control_fd> <args>'.
6034 * Interpretation of args is defined by control file implementation.
6036 static int memcg_write_event_control(struct cgroup_subsys_state
*css
,
6037 struct cftype
*cft
, char *buffer
)
6039 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6040 struct mem_cgroup_event
*event
;
6041 struct cgroup_subsys_state
*cfile_css
;
6042 unsigned int efd
, cfd
;
6049 efd
= simple_strtoul(buffer
, &endp
, 10);
6054 cfd
= simple_strtoul(buffer
, &endp
, 10);
6055 if ((*endp
!= ' ') && (*endp
!= '\0'))
6059 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
6063 event
->memcg
= memcg
;
6064 INIT_LIST_HEAD(&event
->list
);
6065 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
6066 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
6067 INIT_WORK(&event
->remove
, memcg_event_remove
);
6075 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
6076 if (IS_ERR(event
->eventfd
)) {
6077 ret
= PTR_ERR(event
->eventfd
);
6084 goto out_put_eventfd
;
6087 /* the process need read permission on control file */
6088 /* AV: shouldn't we check that it's been opened for read instead? */
6089 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
6094 * Determine the event callbacks and set them in @event. This used
6095 * to be done via struct cftype but cgroup core no longer knows
6096 * about these events. The following is crude but the whole thing
6097 * is for compatibility anyway.
6099 * DO NOT ADD NEW FILES.
6101 name
= cfile
.file
->f_dentry
->d_name
.name
;
6103 if (!strcmp(name
, "memory.usage_in_bytes")) {
6104 event
->register_event
= mem_cgroup_usage_register_event
;
6105 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
6106 } else if (!strcmp(name
, "memory.oom_control")) {
6107 event
->register_event
= mem_cgroup_oom_register_event
;
6108 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
6109 } else if (!strcmp(name
, "memory.pressure_level")) {
6110 event
->register_event
= vmpressure_register_event
;
6111 event
->unregister_event
= vmpressure_unregister_event
;
6112 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
6113 event
->register_event
= memsw_cgroup_usage_register_event
;
6114 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
6121 * Verify @cfile should belong to @css. Also, remaining events are
6122 * automatically removed on cgroup destruction but the removal is
6123 * asynchronous, so take an extra ref on @css.
6125 cfile_css
= css_tryget_from_dir(cfile
.file
->f_dentry
->d_parent
,
6126 &memory_cgrp_subsys
);
6128 if (IS_ERR(cfile_css
))
6130 if (cfile_css
!= css
) {
6135 ret
= event
->register_event(memcg
, event
->eventfd
, buffer
);
6139 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
6141 spin_lock(&memcg
->event_list_lock
);
6142 list_add(&event
->list
, &memcg
->event_list
);
6143 spin_unlock(&memcg
->event_list_lock
);
6155 eventfd_ctx_put(event
->eventfd
);
6164 static struct cftype mem_cgroup_files
[] = {
6166 .name
= "usage_in_bytes",
6167 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
6168 .read_u64
= mem_cgroup_read_u64
,
6171 .name
= "max_usage_in_bytes",
6172 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
6173 .trigger
= mem_cgroup_reset
,
6174 .read_u64
= mem_cgroup_read_u64
,
6177 .name
= "limit_in_bytes",
6178 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
6179 .write_string
= mem_cgroup_write
,
6180 .read_u64
= mem_cgroup_read_u64
,
6183 .name
= "soft_limit_in_bytes",
6184 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
6185 .write_string
= mem_cgroup_write
,
6186 .read_u64
= mem_cgroup_read_u64
,
6190 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
6191 .trigger
= mem_cgroup_reset
,
6192 .read_u64
= mem_cgroup_read_u64
,
6196 .seq_show
= memcg_stat_show
,
6199 .name
= "force_empty",
6200 .trigger
= mem_cgroup_force_empty_write
,
6203 .name
= "use_hierarchy",
6204 .flags
= CFTYPE_INSANE
,
6205 .write_u64
= mem_cgroup_hierarchy_write
,
6206 .read_u64
= mem_cgroup_hierarchy_read
,
6209 .name
= "cgroup.event_control", /* XXX: for compat */
6210 .write_string
= memcg_write_event_control
,
6211 .flags
= CFTYPE_NO_PREFIX
,
6215 .name
= "swappiness",
6216 .read_u64
= mem_cgroup_swappiness_read
,
6217 .write_u64
= mem_cgroup_swappiness_write
,
6220 .name
= "move_charge_at_immigrate",
6221 .read_u64
= mem_cgroup_move_charge_read
,
6222 .write_u64
= mem_cgroup_move_charge_write
,
6225 .name
= "oom_control",
6226 .seq_show
= mem_cgroup_oom_control_read
,
6227 .write_u64
= mem_cgroup_oom_control_write
,
6228 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6231 .name
= "pressure_level",
6235 .name
= "numa_stat",
6236 .seq_show
= memcg_numa_stat_show
,
6239 #ifdef CONFIG_MEMCG_KMEM
6241 .name
= "kmem.limit_in_bytes",
6242 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6243 .write_string
= mem_cgroup_write
,
6244 .read_u64
= mem_cgroup_read_u64
,
6247 .name
= "kmem.usage_in_bytes",
6248 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6249 .read_u64
= mem_cgroup_read_u64
,
6252 .name
= "kmem.failcnt",
6253 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6254 .trigger
= mem_cgroup_reset
,
6255 .read_u64
= mem_cgroup_read_u64
,
6258 .name
= "kmem.max_usage_in_bytes",
6259 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6260 .trigger
= mem_cgroup_reset
,
6261 .read_u64
= mem_cgroup_read_u64
,
6263 #ifdef CONFIG_SLABINFO
6265 .name
= "kmem.slabinfo",
6266 .seq_show
= mem_cgroup_slabinfo_read
,
6270 { }, /* terminate */
6273 #ifdef CONFIG_MEMCG_SWAP
6274 static struct cftype memsw_cgroup_files
[] = {
6276 .name
= "memsw.usage_in_bytes",
6277 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6278 .read_u64
= mem_cgroup_read_u64
,
6281 .name
= "memsw.max_usage_in_bytes",
6282 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6283 .trigger
= mem_cgroup_reset
,
6284 .read_u64
= mem_cgroup_read_u64
,
6287 .name
= "memsw.limit_in_bytes",
6288 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6289 .write_string
= mem_cgroup_write
,
6290 .read_u64
= mem_cgroup_read_u64
,
6293 .name
= "memsw.failcnt",
6294 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6295 .trigger
= mem_cgroup_reset
,
6296 .read_u64
= mem_cgroup_read_u64
,
6298 { }, /* terminate */
6301 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6303 struct mem_cgroup_per_node
*pn
;
6304 struct mem_cgroup_per_zone
*mz
;
6305 int zone
, tmp
= node
;
6307 * This routine is called against possible nodes.
6308 * But it's BUG to call kmalloc() against offline node.
6310 * TODO: this routine can waste much memory for nodes which will
6311 * never be onlined. It's better to use memory hotplug callback
6314 if (!node_state(node
, N_NORMAL_MEMORY
))
6316 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6320 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6321 mz
= &pn
->zoneinfo
[zone
];
6322 lruvec_init(&mz
->lruvec
);
6323 mz
->usage_in_excess
= 0;
6324 mz
->on_tree
= false;
6327 memcg
->nodeinfo
[node
] = pn
;
6331 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6333 kfree(memcg
->nodeinfo
[node
]);
6336 static struct mem_cgroup
*mem_cgroup_alloc(void)
6338 struct mem_cgroup
*memcg
;
6341 size
= sizeof(struct mem_cgroup
);
6342 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
6344 memcg
= kzalloc(size
, GFP_KERNEL
);
6348 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6351 spin_lock_init(&memcg
->pcp_counter_lock
);
6360 * At destroying mem_cgroup, references from swap_cgroup can remain.
6361 * (scanning all at force_empty is too costly...)
6363 * Instead of clearing all references at force_empty, we remember
6364 * the number of reference from swap_cgroup and free mem_cgroup when
6365 * it goes down to 0.
6367 * Removal of cgroup itself succeeds regardless of refs from swap.
6370 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6374 mem_cgroup_remove_from_trees(memcg
);
6377 free_mem_cgroup_per_zone_info(memcg
, node
);
6379 free_percpu(memcg
->stat
);
6382 * We need to make sure that (at least for now), the jump label
6383 * destruction code runs outside of the cgroup lock. This is because
6384 * get_online_cpus(), which is called from the static_branch update,
6385 * can't be called inside the cgroup_lock. cpusets are the ones
6386 * enforcing this dependency, so if they ever change, we might as well.
6388 * schedule_work() will guarantee this happens. Be careful if you need
6389 * to move this code around, and make sure it is outside
6392 disarm_static_keys(memcg
);
6397 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6399 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6401 if (!memcg
->res
.parent
)
6403 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6405 EXPORT_SYMBOL(parent_mem_cgroup
);
6407 static void __init
mem_cgroup_soft_limit_tree_init(void)
6409 struct mem_cgroup_tree_per_node
*rtpn
;
6410 struct mem_cgroup_tree_per_zone
*rtpz
;
6411 int tmp
, node
, zone
;
6413 for_each_node(node
) {
6415 if (!node_state(node
, N_NORMAL_MEMORY
))
6417 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6420 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6422 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6423 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6424 rtpz
->rb_root
= RB_ROOT
;
6425 spin_lock_init(&rtpz
->lock
);
6430 static struct cgroup_subsys_state
* __ref
6431 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6433 struct mem_cgroup
*memcg
;
6434 long error
= -ENOMEM
;
6437 memcg
= mem_cgroup_alloc();
6439 return ERR_PTR(error
);
6442 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6446 if (parent_css
== NULL
) {
6447 root_mem_cgroup
= memcg
;
6448 res_counter_init(&memcg
->res
, NULL
);
6449 res_counter_init(&memcg
->memsw
, NULL
);
6450 res_counter_init(&memcg
->kmem
, NULL
);
6453 memcg
->last_scanned_node
= MAX_NUMNODES
;
6454 INIT_LIST_HEAD(&memcg
->oom_notify
);
6455 memcg
->move_charge_at_immigrate
= 0;
6456 mutex_init(&memcg
->thresholds_lock
);
6457 spin_lock_init(&memcg
->move_lock
);
6458 vmpressure_init(&memcg
->vmpressure
);
6459 INIT_LIST_HEAD(&memcg
->event_list
);
6460 spin_lock_init(&memcg
->event_list_lock
);
6465 __mem_cgroup_free(memcg
);
6466 return ERR_PTR(error
);
6470 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6472 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6473 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6475 if (css
->cgroup
->id
> MEM_CGROUP_ID_MAX
)
6481 mutex_lock(&memcg_create_mutex
);
6483 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6484 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6485 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6487 if (parent
->use_hierarchy
) {
6488 res_counter_init(&memcg
->res
, &parent
->res
);
6489 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6490 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6493 * No need to take a reference to the parent because cgroup
6494 * core guarantees its existence.
6497 res_counter_init(&memcg
->res
, NULL
);
6498 res_counter_init(&memcg
->memsw
, NULL
);
6499 res_counter_init(&memcg
->kmem
, NULL
);
6501 * Deeper hierachy with use_hierarchy == false doesn't make
6502 * much sense so let cgroup subsystem know about this
6503 * unfortunate state in our controller.
6505 if (parent
!= root_mem_cgroup
)
6506 memory_cgrp_subsys
.broken_hierarchy
= true;
6508 mutex_unlock(&memcg_create_mutex
);
6510 return memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
6514 * Announce all parents that a group from their hierarchy is gone.
6516 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6518 struct mem_cgroup
*parent
= memcg
;
6520 while ((parent
= parent_mem_cgroup(parent
)))
6521 mem_cgroup_iter_invalidate(parent
);
6524 * if the root memcg is not hierarchical we have to check it
6527 if (!root_mem_cgroup
->use_hierarchy
)
6528 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6531 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6533 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6534 struct mem_cgroup_event
*event
, *tmp
;
6535 struct cgroup_subsys_state
*iter
;
6538 * Unregister events and notify userspace.
6539 * Notify userspace about cgroup removing only after rmdir of cgroup
6540 * directory to avoid race between userspace and kernelspace.
6542 spin_lock(&memcg
->event_list_lock
);
6543 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
6544 list_del_init(&event
->list
);
6545 schedule_work(&event
->remove
);
6547 spin_unlock(&memcg
->event_list_lock
);
6549 kmem_cgroup_css_offline(memcg
);
6551 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6554 * This requires that offlining is serialized. Right now that is
6555 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6557 css_for_each_descendant_post(iter
, css
)
6558 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter
));
6560 mem_cgroup_destroy_all_caches(memcg
);
6561 vmpressure_cleanup(&memcg
->vmpressure
);
6564 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6566 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6568 * XXX: css_offline() would be where we should reparent all
6569 * memory to prepare the cgroup for destruction. However,
6570 * memcg does not do css_tryget() and res_counter charging
6571 * under the same RCU lock region, which means that charging
6572 * could race with offlining. Offlining only happens to
6573 * cgroups with no tasks in them but charges can show up
6574 * without any tasks from the swapin path when the target
6575 * memcg is looked up from the swapout record and not from the
6576 * current task as it usually is. A race like this can leak
6577 * charges and put pages with stale cgroup pointers into
6581 * lookup_swap_cgroup_id()
6583 * mem_cgroup_lookup()
6586 * disable css_tryget()
6589 * reparent_charges()
6590 * res_counter_charge()
6593 * pc->mem_cgroup = dead memcg
6596 * The bulk of the charges are still moved in offline_css() to
6597 * avoid pinning a lot of pages in case a long-term reference
6598 * like a swapout record is deferring the css_free() to long
6599 * after offlining. But this makes sure we catch any charges
6600 * made after offlining:
6602 mem_cgroup_reparent_charges(memcg
);
6604 memcg_destroy_kmem(memcg
);
6605 __mem_cgroup_free(memcg
);
6609 /* Handlers for move charge at task migration. */
6610 #define PRECHARGE_COUNT_AT_ONCE 256
6611 static int mem_cgroup_do_precharge(unsigned long count
)
6614 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6615 struct mem_cgroup
*memcg
= mc
.to
;
6617 if (mem_cgroup_is_root(memcg
)) {
6618 mc
.precharge
+= count
;
6619 /* we don't need css_get for root */
6622 /* try to charge at once */
6624 struct res_counter
*dummy
;
6626 * "memcg" cannot be under rmdir() because we've already checked
6627 * by cgroup_lock_live_cgroup() that it is not removed and we
6628 * are still under the same cgroup_mutex. So we can postpone
6631 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6633 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6634 PAGE_SIZE
* count
, &dummy
)) {
6635 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6638 mc
.precharge
+= count
;
6642 /* fall back to one by one charge */
6644 if (signal_pending(current
)) {
6648 if (!batch_count
--) {
6649 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6652 ret
= __mem_cgroup_try_charge(NULL
,
6653 GFP_KERNEL
, 1, &memcg
, false);
6655 /* mem_cgroup_clear_mc() will do uncharge later */
6663 * get_mctgt_type - get target type of moving charge
6664 * @vma: the vma the pte to be checked belongs
6665 * @addr: the address corresponding to the pte to be checked
6666 * @ptent: the pte to be checked
6667 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6670 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6671 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6672 * move charge. if @target is not NULL, the page is stored in target->page
6673 * with extra refcnt got(Callers should handle it).
6674 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6675 * target for charge migration. if @target is not NULL, the entry is stored
6678 * Called with pte lock held.
6685 enum mc_target_type
{
6691 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6692 unsigned long addr
, pte_t ptent
)
6694 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6696 if (!page
|| !page_mapped(page
))
6698 if (PageAnon(page
)) {
6699 /* we don't move shared anon */
6702 } else if (!move_file())
6703 /* we ignore mapcount for file pages */
6705 if (!get_page_unless_zero(page
))
6712 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6713 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6715 struct page
*page
= NULL
;
6716 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6718 if (!move_anon() || non_swap_entry(ent
))
6721 * Because lookup_swap_cache() updates some statistics counter,
6722 * we call find_get_page() with swapper_space directly.
6724 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6725 if (do_swap_account
)
6726 entry
->val
= ent
.val
;
6731 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6732 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6738 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6739 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6741 struct page
*page
= NULL
;
6742 struct address_space
*mapping
;
6745 if (!vma
->vm_file
) /* anonymous vma */
6750 mapping
= vma
->vm_file
->f_mapping
;
6751 if (pte_none(ptent
))
6752 pgoff
= linear_page_index(vma
, addr
);
6753 else /* pte_file(ptent) is true */
6754 pgoff
= pte_to_pgoff(ptent
);
6756 /* page is moved even if it's not RSS of this task(page-faulted). */
6757 page
= find_get_page(mapping
, pgoff
);
6760 /* shmem/tmpfs may report page out on swap: account for that too. */
6761 if (radix_tree_exceptional_entry(page
)) {
6762 swp_entry_t swap
= radix_to_swp_entry(page
);
6763 if (do_swap_account
)
6765 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6771 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6772 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6774 struct page
*page
= NULL
;
6775 struct page_cgroup
*pc
;
6776 enum mc_target_type ret
= MC_TARGET_NONE
;
6777 swp_entry_t ent
= { .val
= 0 };
6779 if (pte_present(ptent
))
6780 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6781 else if (is_swap_pte(ptent
))
6782 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6783 else if (pte_none(ptent
) || pte_file(ptent
))
6784 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6786 if (!page
&& !ent
.val
)
6789 pc
= lookup_page_cgroup(page
);
6791 * Do only loose check w/o page_cgroup lock.
6792 * mem_cgroup_move_account() checks the pc is valid or not under
6795 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6796 ret
= MC_TARGET_PAGE
;
6798 target
->page
= page
;
6800 if (!ret
|| !target
)
6803 /* There is a swap entry and a page doesn't exist or isn't charged */
6804 if (ent
.val
&& !ret
&&
6805 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6806 ret
= MC_TARGET_SWAP
;
6813 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6815 * We don't consider swapping or file mapped pages because THP does not
6816 * support them for now.
6817 * Caller should make sure that pmd_trans_huge(pmd) is true.
6819 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6820 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6822 struct page
*page
= NULL
;
6823 struct page_cgroup
*pc
;
6824 enum mc_target_type ret
= MC_TARGET_NONE
;
6826 page
= pmd_page(pmd
);
6827 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
6830 pc
= lookup_page_cgroup(page
);
6831 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6832 ret
= MC_TARGET_PAGE
;
6835 target
->page
= page
;
6841 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6842 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6844 return MC_TARGET_NONE
;
6848 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6849 unsigned long addr
, unsigned long end
,
6850 struct mm_walk
*walk
)
6852 struct vm_area_struct
*vma
= walk
->private;
6856 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6857 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6858 mc
.precharge
+= HPAGE_PMD_NR
;
6863 if (pmd_trans_unstable(pmd
))
6865 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6866 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6867 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6868 mc
.precharge
++; /* increment precharge temporarily */
6869 pte_unmap_unlock(pte
- 1, ptl
);
6875 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6877 unsigned long precharge
;
6878 struct vm_area_struct
*vma
;
6880 down_read(&mm
->mmap_sem
);
6881 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6882 struct mm_walk mem_cgroup_count_precharge_walk
= {
6883 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6887 if (is_vm_hugetlb_page(vma
))
6889 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6890 &mem_cgroup_count_precharge_walk
);
6892 up_read(&mm
->mmap_sem
);
6894 precharge
= mc
.precharge
;
6900 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6902 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6904 VM_BUG_ON(mc
.moving_task
);
6905 mc
.moving_task
= current
;
6906 return mem_cgroup_do_precharge(precharge
);
6909 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6910 static void __mem_cgroup_clear_mc(void)
6912 struct mem_cgroup
*from
= mc
.from
;
6913 struct mem_cgroup
*to
= mc
.to
;
6916 /* we must uncharge all the leftover precharges from mc.to */
6918 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6922 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6923 * we must uncharge here.
6925 if (mc
.moved_charge
) {
6926 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6927 mc
.moved_charge
= 0;
6929 /* we must fixup refcnts and charges */
6930 if (mc
.moved_swap
) {
6931 /* uncharge swap account from the old cgroup */
6932 if (!mem_cgroup_is_root(mc
.from
))
6933 res_counter_uncharge(&mc
.from
->memsw
,
6934 PAGE_SIZE
* mc
.moved_swap
);
6936 for (i
= 0; i
< mc
.moved_swap
; i
++)
6937 css_put(&mc
.from
->css
);
6939 if (!mem_cgroup_is_root(mc
.to
)) {
6941 * we charged both to->res and to->memsw, so we should
6944 res_counter_uncharge(&mc
.to
->res
,
6945 PAGE_SIZE
* mc
.moved_swap
);
6947 /* we've already done css_get(mc.to) */
6950 memcg_oom_recover(from
);
6951 memcg_oom_recover(to
);
6952 wake_up_all(&mc
.waitq
);
6955 static void mem_cgroup_clear_mc(void)
6957 struct mem_cgroup
*from
= mc
.from
;
6960 * we must clear moving_task before waking up waiters at the end of
6963 mc
.moving_task
= NULL
;
6964 __mem_cgroup_clear_mc();
6965 spin_lock(&mc
.lock
);
6968 spin_unlock(&mc
.lock
);
6969 mem_cgroup_end_move(from
);
6972 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6973 struct cgroup_taskset
*tset
)
6975 struct task_struct
*p
= cgroup_taskset_first(tset
);
6977 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6978 unsigned long move_charge_at_immigrate
;
6981 * We are now commited to this value whatever it is. Changes in this
6982 * tunable will only affect upcoming migrations, not the current one.
6983 * So we need to save it, and keep it going.
6985 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6986 if (move_charge_at_immigrate
) {
6987 struct mm_struct
*mm
;
6988 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6990 VM_BUG_ON(from
== memcg
);
6992 mm
= get_task_mm(p
);
6995 /* We move charges only when we move a owner of the mm */
6996 if (mm
->owner
== p
) {
6999 VM_BUG_ON(mc
.precharge
);
7000 VM_BUG_ON(mc
.moved_charge
);
7001 VM_BUG_ON(mc
.moved_swap
);
7002 mem_cgroup_start_move(from
);
7003 spin_lock(&mc
.lock
);
7006 mc
.immigrate_flags
= move_charge_at_immigrate
;
7007 spin_unlock(&mc
.lock
);
7008 /* We set mc.moving_task later */
7010 ret
= mem_cgroup_precharge_mc(mm
);
7012 mem_cgroup_clear_mc();
7019 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
7020 struct cgroup_taskset
*tset
)
7022 mem_cgroup_clear_mc();
7025 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
7026 unsigned long addr
, unsigned long end
,
7027 struct mm_walk
*walk
)
7030 struct vm_area_struct
*vma
= walk
->private;
7033 enum mc_target_type target_type
;
7034 union mc_target target
;
7036 struct page_cgroup
*pc
;
7039 * We don't take compound_lock() here but no race with splitting thp
7041 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7042 * under splitting, which means there's no concurrent thp split,
7043 * - if another thread runs into split_huge_page() just after we
7044 * entered this if-block, the thread must wait for page table lock
7045 * to be unlocked in __split_huge_page_splitting(), where the main
7046 * part of thp split is not executed yet.
7048 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
7049 if (mc
.precharge
< HPAGE_PMD_NR
) {
7053 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
7054 if (target_type
== MC_TARGET_PAGE
) {
7056 if (!isolate_lru_page(page
)) {
7057 pc
= lookup_page_cgroup(page
);
7058 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
7059 pc
, mc
.from
, mc
.to
)) {
7060 mc
.precharge
-= HPAGE_PMD_NR
;
7061 mc
.moved_charge
+= HPAGE_PMD_NR
;
7063 putback_lru_page(page
);
7071 if (pmd_trans_unstable(pmd
))
7074 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
7075 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
7076 pte_t ptent
= *(pte
++);
7082 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
7083 case MC_TARGET_PAGE
:
7085 if (isolate_lru_page(page
))
7087 pc
= lookup_page_cgroup(page
);
7088 if (!mem_cgroup_move_account(page
, 1, pc
,
7091 /* we uncharge from mc.from later. */
7094 putback_lru_page(page
);
7095 put
: /* get_mctgt_type() gets the page */
7098 case MC_TARGET_SWAP
:
7100 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
7102 /* we fixup refcnts and charges later. */
7110 pte_unmap_unlock(pte
- 1, ptl
);
7115 * We have consumed all precharges we got in can_attach().
7116 * We try charge one by one, but don't do any additional
7117 * charges to mc.to if we have failed in charge once in attach()
7120 ret
= mem_cgroup_do_precharge(1);
7128 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
7130 struct vm_area_struct
*vma
;
7132 lru_add_drain_all();
7134 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
7136 * Someone who are holding the mmap_sem might be waiting in
7137 * waitq. So we cancel all extra charges, wake up all waiters,
7138 * and retry. Because we cancel precharges, we might not be able
7139 * to move enough charges, but moving charge is a best-effort
7140 * feature anyway, so it wouldn't be a big problem.
7142 __mem_cgroup_clear_mc();
7146 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
7148 struct mm_walk mem_cgroup_move_charge_walk
= {
7149 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
7153 if (is_vm_hugetlb_page(vma
))
7155 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
7156 &mem_cgroup_move_charge_walk
);
7159 * means we have consumed all precharges and failed in
7160 * doing additional charge. Just abandon here.
7164 up_read(&mm
->mmap_sem
);
7167 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
7168 struct cgroup_taskset
*tset
)
7170 struct task_struct
*p
= cgroup_taskset_first(tset
);
7171 struct mm_struct
*mm
= get_task_mm(p
);
7175 mem_cgroup_move_charge(mm
);
7179 mem_cgroup_clear_mc();
7181 #else /* !CONFIG_MMU */
7182 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
7183 struct cgroup_taskset
*tset
)
7187 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
7188 struct cgroup_taskset
*tset
)
7191 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
7192 struct cgroup_taskset
*tset
)
7198 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7199 * to verify sane_behavior flag on each mount attempt.
7201 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
7204 * use_hierarchy is forced with sane_behavior. cgroup core
7205 * guarantees that @root doesn't have any children, so turning it
7206 * on for the root memcg is enough.
7208 if (cgroup_sane_behavior(root_css
->cgroup
))
7209 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
7212 struct cgroup_subsys memory_cgrp_subsys
= {
7213 .css_alloc
= mem_cgroup_css_alloc
,
7214 .css_online
= mem_cgroup_css_online
,
7215 .css_offline
= mem_cgroup_css_offline
,
7216 .css_free
= mem_cgroup_css_free
,
7217 .can_attach
= mem_cgroup_can_attach
,
7218 .cancel_attach
= mem_cgroup_cancel_attach
,
7219 .attach
= mem_cgroup_move_task
,
7220 .bind
= mem_cgroup_bind
,
7221 .base_cftypes
= mem_cgroup_files
,
7225 #ifdef CONFIG_MEMCG_SWAP
7226 static int __init
enable_swap_account(char *s
)
7228 if (!strcmp(s
, "1"))
7229 really_do_swap_account
= 1;
7230 else if (!strcmp(s
, "0"))
7231 really_do_swap_account
= 0;
7234 __setup("swapaccount=", enable_swap_account
);
7236 static void __init
memsw_file_init(void)
7238 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys
, memsw_cgroup_files
));
7241 static void __init
enable_swap_cgroup(void)
7243 if (!mem_cgroup_disabled() && really_do_swap_account
) {
7244 do_swap_account
= 1;
7250 static void __init
enable_swap_cgroup(void)
7256 * subsys_initcall() for memory controller.
7258 * Some parts like hotcpu_notifier() have to be initialized from this context
7259 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7260 * everything that doesn't depend on a specific mem_cgroup structure should
7261 * be initialized from here.
7263 static int __init
mem_cgroup_init(void)
7265 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7266 enable_swap_cgroup();
7267 mem_cgroup_soft_limit_tree_init();
7271 subsys_initcall(mem_cgroup_init
);