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
;
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
;
295 /* css_online() has been completed */
299 * the counter to account for mem+swap usage.
301 struct res_counter memsw
;
304 * the counter to account for kernel memory usage.
306 struct res_counter kmem
;
308 * Should the accounting and control be hierarchical, per subtree?
311 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
315 atomic_t oom_wakeups
;
318 /* OOM-Killer disable */
319 int oom_kill_disable
;
321 /* set when res.limit == memsw.limit */
322 bool memsw_is_minimum
;
324 /* protect arrays of thresholds */
325 struct mutex thresholds_lock
;
327 /* thresholds for memory usage. RCU-protected */
328 struct mem_cgroup_thresholds thresholds
;
330 /* thresholds for mem+swap usage. RCU-protected */
331 struct mem_cgroup_thresholds memsw_thresholds
;
333 /* For oom notifier event fd */
334 struct list_head oom_notify
;
337 * Should we move charges of a task when a task is moved into this
338 * mem_cgroup ? And what type of charges should we move ?
340 unsigned long move_charge_at_immigrate
;
342 * set > 0 if pages under this cgroup are moving to other cgroup.
344 atomic_t moving_account
;
345 /* taken only while moving_account > 0 */
346 spinlock_t move_lock
;
350 struct mem_cgroup_stat_cpu __percpu
*stat
;
352 * used when a cpu is offlined or other synchronizations
353 * See mem_cgroup_read_stat().
355 struct mem_cgroup_stat_cpu nocpu_base
;
356 spinlock_t pcp_counter_lock
;
359 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
360 struct cg_proto tcp_mem
;
362 #if defined(CONFIG_MEMCG_KMEM)
363 /* analogous to slab_common's slab_caches list, but per-memcg;
364 * protected by memcg_slab_mutex */
365 struct list_head memcg_slab_caches
;
366 /* Index in the kmem_cache->memcg_params->memcg_caches array */
370 int last_scanned_node
;
372 nodemask_t scan_nodes
;
373 atomic_t numainfo_events
;
374 atomic_t numainfo_updating
;
377 /* List of events which userspace want to receive */
378 struct list_head event_list
;
379 spinlock_t event_list_lock
;
381 struct mem_cgroup_per_node
*nodeinfo
[0];
382 /* WARNING: nodeinfo must be the last member here */
385 /* internal only representation about the status of kmem accounting. */
387 KMEM_ACCOUNTED_ACTIVE
, /* accounted by this cgroup itself */
388 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
391 #ifdef CONFIG_MEMCG_KMEM
392 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
394 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
397 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
399 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
402 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
405 * Our caller must use css_get() first, because memcg_uncharge_kmem()
406 * will call css_put() if it sees the memcg is dead.
409 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
410 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
413 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
415 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
416 &memcg
->kmem_account_flags
);
420 /* Stuffs for move charges at task migration. */
422 * Types of charges to be moved. "move_charge_at_immitgrate" and
423 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
426 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
427 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
431 /* "mc" and its members are protected by cgroup_mutex */
432 static struct move_charge_struct
{
433 spinlock_t lock
; /* for from, to */
434 struct mem_cgroup
*from
;
435 struct mem_cgroup
*to
;
436 unsigned long immigrate_flags
;
437 unsigned long precharge
;
438 unsigned long moved_charge
;
439 unsigned long moved_swap
;
440 struct task_struct
*moving_task
; /* a task moving charges */
441 wait_queue_head_t waitq
; /* a waitq for other context */
443 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
444 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
447 static bool move_anon(void)
449 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
452 static bool move_file(void)
454 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
458 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
459 * limit reclaim to prevent infinite loops, if they ever occur.
461 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
462 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
465 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
466 MEM_CGROUP_CHARGE_TYPE_ANON
,
467 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
468 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
472 /* for encoding cft->private value on file */
480 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
481 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
482 #define MEMFILE_ATTR(val) ((val) & 0xffff)
483 /* Used for OOM nofiier */
484 #define OOM_CONTROL (0)
487 * Reclaim flags for mem_cgroup_hierarchical_reclaim
489 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
490 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
491 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
492 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
495 * The memcg_create_mutex will be held whenever a new cgroup is created.
496 * As a consequence, any change that needs to protect against new child cgroups
497 * appearing has to hold it as well.
499 static DEFINE_MUTEX(memcg_create_mutex
);
501 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
503 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
506 /* Some nice accessors for the vmpressure. */
507 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
510 memcg
= root_mem_cgroup
;
511 return &memcg
->vmpressure
;
514 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
516 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
519 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
521 return (memcg
== root_mem_cgroup
);
525 * We restrict the id in the range of [1, 65535], so it can fit into
528 #define MEM_CGROUP_ID_MAX USHRT_MAX
530 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
532 return memcg
->css
.id
;
535 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
537 struct cgroup_subsys_state
*css
;
539 css
= css_from_id(id
, &memory_cgrp_subsys
);
540 return mem_cgroup_from_css(css
);
543 /* Writing them here to avoid exposing memcg's inner layout */
544 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
546 void sock_update_memcg(struct sock
*sk
)
548 if (mem_cgroup_sockets_enabled
) {
549 struct mem_cgroup
*memcg
;
550 struct cg_proto
*cg_proto
;
552 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
554 /* Socket cloning can throw us here with sk_cgrp already
555 * filled. It won't however, necessarily happen from
556 * process context. So the test for root memcg given
557 * the current task's memcg won't help us in this case.
559 * Respecting the original socket's memcg is a better
560 * decision in this case.
563 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
564 css_get(&sk
->sk_cgrp
->memcg
->css
);
569 memcg
= mem_cgroup_from_task(current
);
570 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
571 if (!mem_cgroup_is_root(memcg
) &&
572 memcg_proto_active(cg_proto
) &&
573 css_tryget_online(&memcg
->css
)) {
574 sk
->sk_cgrp
= cg_proto
;
579 EXPORT_SYMBOL(sock_update_memcg
);
581 void sock_release_memcg(struct sock
*sk
)
583 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
584 struct mem_cgroup
*memcg
;
585 WARN_ON(!sk
->sk_cgrp
->memcg
);
586 memcg
= sk
->sk_cgrp
->memcg
;
587 css_put(&sk
->sk_cgrp
->memcg
->css
);
591 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
593 if (!memcg
|| mem_cgroup_is_root(memcg
))
596 return &memcg
->tcp_mem
;
598 EXPORT_SYMBOL(tcp_proto_cgroup
);
600 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
602 if (!memcg_proto_activated(&memcg
->tcp_mem
))
604 static_key_slow_dec(&memcg_socket_limit_enabled
);
607 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
612 #ifdef CONFIG_MEMCG_KMEM
614 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
615 * The main reason for not using cgroup id for this:
616 * this works better in sparse environments, where we have a lot of memcgs,
617 * but only a few kmem-limited. Or also, if we have, for instance, 200
618 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
619 * 200 entry array for that.
621 * The current size of the caches array is stored in
622 * memcg_limited_groups_array_size. It will double each time we have to
625 static DEFINE_IDA(kmem_limited_groups
);
626 int memcg_limited_groups_array_size
;
629 * MIN_SIZE is different than 1, because we would like to avoid going through
630 * the alloc/free process all the time. In a small machine, 4 kmem-limited
631 * cgroups is a reasonable guess. In the future, it could be a parameter or
632 * tunable, but that is strictly not necessary.
634 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
635 * this constant directly from cgroup, but it is understandable that this is
636 * better kept as an internal representation in cgroup.c. In any case, the
637 * cgrp_id space is not getting any smaller, and we don't have to necessarily
638 * increase ours as well if it increases.
640 #define MEMCG_CACHES_MIN_SIZE 4
641 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
644 * A lot of the calls to the cache allocation functions are expected to be
645 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
646 * conditional to this static branch, we'll have to allow modules that does
647 * kmem_cache_alloc and the such to see this symbol as well
649 struct static_key memcg_kmem_enabled_key
;
650 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
652 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
654 if (memcg_kmem_is_active(memcg
)) {
655 static_key_slow_dec(&memcg_kmem_enabled_key
);
656 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
659 * This check can't live in kmem destruction function,
660 * since the charges will outlive the cgroup
662 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
665 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
668 #endif /* CONFIG_MEMCG_KMEM */
670 static void disarm_static_keys(struct mem_cgroup
*memcg
)
672 disarm_sock_keys(memcg
);
673 disarm_kmem_keys(memcg
);
676 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
678 static struct mem_cgroup_per_zone
*
679 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
681 int nid
= zone_to_nid(zone
);
682 int zid
= zone_idx(zone
);
684 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
687 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
692 static struct mem_cgroup_per_zone
*
693 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
695 int nid
= page_to_nid(page
);
696 int zid
= page_zonenum(page
);
698 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
701 static struct mem_cgroup_tree_per_zone
*
702 soft_limit_tree_node_zone(int nid
, int zid
)
704 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
707 static struct mem_cgroup_tree_per_zone
*
708 soft_limit_tree_from_page(struct page
*page
)
710 int nid
= page_to_nid(page
);
711 int zid
= page_zonenum(page
);
713 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
716 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
717 struct mem_cgroup_tree_per_zone
*mctz
,
718 unsigned long long new_usage_in_excess
)
720 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
721 struct rb_node
*parent
= NULL
;
722 struct mem_cgroup_per_zone
*mz_node
;
727 mz
->usage_in_excess
= new_usage_in_excess
;
728 if (!mz
->usage_in_excess
)
732 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
734 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
737 * We can't avoid mem cgroups that are over their soft
738 * limit by the same amount
740 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
743 rb_link_node(&mz
->tree_node
, parent
, p
);
744 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
748 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
749 struct mem_cgroup_tree_per_zone
*mctz
)
753 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
757 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
758 struct mem_cgroup_tree_per_zone
*mctz
)
762 spin_lock_irqsave(&mctz
->lock
, flags
);
763 __mem_cgroup_remove_exceeded(mz
, mctz
);
764 spin_unlock_irqrestore(&mctz
->lock
, flags
);
768 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
770 unsigned long long excess
;
771 struct mem_cgroup_per_zone
*mz
;
772 struct mem_cgroup_tree_per_zone
*mctz
;
774 mctz
= soft_limit_tree_from_page(page
);
776 * Necessary to update all ancestors when hierarchy is used.
777 * because their event counter is not touched.
779 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
780 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
781 excess
= res_counter_soft_limit_excess(&memcg
->res
);
783 * We have to update the tree if mz is on RB-tree or
784 * mem is over its softlimit.
786 if (excess
|| mz
->on_tree
) {
789 spin_lock_irqsave(&mctz
->lock
, flags
);
790 /* if on-tree, remove it */
792 __mem_cgroup_remove_exceeded(mz
, mctz
);
794 * Insert again. mz->usage_in_excess will be updated.
795 * If excess is 0, no tree ops.
797 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
798 spin_unlock_irqrestore(&mctz
->lock
, flags
);
803 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
805 struct mem_cgroup_tree_per_zone
*mctz
;
806 struct mem_cgroup_per_zone
*mz
;
810 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
811 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
812 mctz
= soft_limit_tree_node_zone(nid
, zid
);
813 mem_cgroup_remove_exceeded(mz
, mctz
);
818 static struct mem_cgroup_per_zone
*
819 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
821 struct rb_node
*rightmost
= NULL
;
822 struct mem_cgroup_per_zone
*mz
;
826 rightmost
= rb_last(&mctz
->rb_root
);
828 goto done
; /* Nothing to reclaim from */
830 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
832 * Remove the node now but someone else can add it back,
833 * we will to add it back at the end of reclaim to its correct
834 * position in the tree.
836 __mem_cgroup_remove_exceeded(mz
, mctz
);
837 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
838 !css_tryget_online(&mz
->memcg
->css
))
844 static struct mem_cgroup_per_zone
*
845 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
847 struct mem_cgroup_per_zone
*mz
;
849 spin_lock_irq(&mctz
->lock
);
850 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
851 spin_unlock_irq(&mctz
->lock
);
856 * Implementation Note: reading percpu statistics for memcg.
858 * Both of vmstat[] and percpu_counter has threshold and do periodic
859 * synchronization to implement "quick" read. There are trade-off between
860 * reading cost and precision of value. Then, we may have a chance to implement
861 * a periodic synchronizion of counter in memcg's counter.
863 * But this _read() function is used for user interface now. The user accounts
864 * memory usage by memory cgroup and he _always_ requires exact value because
865 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
866 * have to visit all online cpus and make sum. So, for now, unnecessary
867 * synchronization is not implemented. (just implemented for cpu hotplug)
869 * If there are kernel internal actions which can make use of some not-exact
870 * value, and reading all cpu value can be performance bottleneck in some
871 * common workload, threashold and synchonization as vmstat[] should be
874 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
875 enum mem_cgroup_stat_index idx
)
881 for_each_online_cpu(cpu
)
882 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
883 #ifdef CONFIG_HOTPLUG_CPU
884 spin_lock(&memcg
->pcp_counter_lock
);
885 val
+= memcg
->nocpu_base
.count
[idx
];
886 spin_unlock(&memcg
->pcp_counter_lock
);
892 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
893 enum mem_cgroup_events_index idx
)
895 unsigned long val
= 0;
899 for_each_online_cpu(cpu
)
900 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
901 #ifdef CONFIG_HOTPLUG_CPU
902 spin_lock(&memcg
->pcp_counter_lock
);
903 val
+= memcg
->nocpu_base
.events
[idx
];
904 spin_unlock(&memcg
->pcp_counter_lock
);
910 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
915 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
916 * counted as CACHE even if it's on ANON LRU.
919 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
922 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
925 if (PageTransHuge(page
))
926 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
929 /* pagein of a big page is an event. So, ignore page size */
931 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
933 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
934 nr_pages
= -nr_pages
; /* for event */
937 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
940 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
942 struct mem_cgroup_per_zone
*mz
;
944 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
945 return mz
->lru_size
[lru
];
948 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
950 unsigned int lru_mask
)
952 unsigned long nr
= 0;
955 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
957 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
958 struct mem_cgroup_per_zone
*mz
;
962 if (!(BIT(lru
) & lru_mask
))
964 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
965 nr
+= mz
->lru_size
[lru
];
971 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
972 unsigned int lru_mask
)
974 unsigned long nr
= 0;
977 for_each_node_state(nid
, N_MEMORY
)
978 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
982 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
983 enum mem_cgroup_events_target target
)
985 unsigned long val
, next
;
987 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
988 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
989 /* from time_after() in jiffies.h */
990 if ((long)next
- (long)val
< 0) {
992 case MEM_CGROUP_TARGET_THRESH
:
993 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
995 case MEM_CGROUP_TARGET_SOFTLIMIT
:
996 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
998 case MEM_CGROUP_TARGET_NUMAINFO
:
999 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1004 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1011 * Check events in order.
1014 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1016 /* threshold event is triggered in finer grain than soft limit */
1017 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1018 MEM_CGROUP_TARGET_THRESH
))) {
1020 bool do_numainfo __maybe_unused
;
1022 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1023 MEM_CGROUP_TARGET_SOFTLIMIT
);
1024 #if MAX_NUMNODES > 1
1025 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1026 MEM_CGROUP_TARGET_NUMAINFO
);
1028 mem_cgroup_threshold(memcg
);
1029 if (unlikely(do_softlimit
))
1030 mem_cgroup_update_tree(memcg
, page
);
1031 #if MAX_NUMNODES > 1
1032 if (unlikely(do_numainfo
))
1033 atomic_inc(&memcg
->numainfo_events
);
1038 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1041 * mm_update_next_owner() may clear mm->owner to NULL
1042 * if it races with swapoff, page migration, etc.
1043 * So this can be called with p == NULL.
1048 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1051 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1053 struct mem_cgroup
*memcg
= NULL
;
1058 * Page cache insertions can happen withou an
1059 * actual mm context, e.g. during disk probing
1060 * on boot, loopback IO, acct() writes etc.
1063 memcg
= root_mem_cgroup
;
1065 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1066 if (unlikely(!memcg
))
1067 memcg
= root_mem_cgroup
;
1069 } while (!css_tryget_online(&memcg
->css
));
1075 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1076 * ref. count) or NULL if the whole root's subtree has been visited.
1078 * helper function to be used by mem_cgroup_iter
1080 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1081 struct mem_cgroup
*last_visited
)
1083 struct cgroup_subsys_state
*prev_css
, *next_css
;
1085 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1087 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1090 * Even if we found a group we have to make sure it is
1091 * alive. css && !memcg means that the groups should be
1092 * skipped and we should continue the tree walk.
1093 * last_visited css is safe to use because it is
1094 * protected by css_get and the tree walk is rcu safe.
1096 * We do not take a reference on the root of the tree walk
1097 * because we might race with the root removal when it would
1098 * be the only node in the iterated hierarchy and mem_cgroup_iter
1099 * would end up in an endless loop because it expects that at
1100 * least one valid node will be returned. Root cannot disappear
1101 * because caller of the iterator should hold it already so
1102 * skipping css reference should be safe.
1105 struct mem_cgroup
*memcg
= mem_cgroup_from_css(next_css
);
1107 if (next_css
== &root
->css
)
1110 if (css_tryget_online(next_css
)) {
1112 * Make sure the memcg is initialized:
1113 * mem_cgroup_css_online() orders the the
1114 * initialization against setting the flag.
1116 if (smp_load_acquire(&memcg
->initialized
))
1121 prev_css
= next_css
;
1128 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1131 * When a group in the hierarchy below root is destroyed, the
1132 * hierarchy iterator can no longer be trusted since it might
1133 * have pointed to the destroyed group. Invalidate it.
1135 atomic_inc(&root
->dead_count
);
1138 static struct mem_cgroup
*
1139 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1140 struct mem_cgroup
*root
,
1143 struct mem_cgroup
*position
= NULL
;
1145 * A cgroup destruction happens in two stages: offlining and
1146 * release. They are separated by a RCU grace period.
1148 * If the iterator is valid, we may still race with an
1149 * offlining. The RCU lock ensures the object won't be
1150 * released, tryget will fail if we lost the race.
1152 *sequence
= atomic_read(&root
->dead_count
);
1153 if (iter
->last_dead_count
== *sequence
) {
1155 position
= iter
->last_visited
;
1158 * We cannot take a reference to root because we might race
1159 * with root removal and returning NULL would end up in
1160 * an endless loop on the iterator user level when root
1161 * would be returned all the time.
1163 if (position
&& position
!= root
&&
1164 !css_tryget_online(&position
->css
))
1170 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1171 struct mem_cgroup
*last_visited
,
1172 struct mem_cgroup
*new_position
,
1173 struct mem_cgroup
*root
,
1176 /* root reference counting symmetric to mem_cgroup_iter_load */
1177 if (last_visited
&& last_visited
!= root
)
1178 css_put(&last_visited
->css
);
1180 * We store the sequence count from the time @last_visited was
1181 * loaded successfully instead of rereading it here so that we
1182 * don't lose destruction events in between. We could have
1183 * raced with the destruction of @new_position after all.
1185 iter
->last_visited
= new_position
;
1187 iter
->last_dead_count
= sequence
;
1191 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1192 * @root: hierarchy root
1193 * @prev: previously returned memcg, NULL on first invocation
1194 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1196 * Returns references to children of the hierarchy below @root, or
1197 * @root itself, or %NULL after a full round-trip.
1199 * Caller must pass the return value in @prev on subsequent
1200 * invocations for reference counting, or use mem_cgroup_iter_break()
1201 * to cancel a hierarchy walk before the round-trip is complete.
1203 * Reclaimers can specify a zone and a priority level in @reclaim to
1204 * divide up the memcgs in the hierarchy among all concurrent
1205 * reclaimers operating on the same zone and priority.
1207 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1208 struct mem_cgroup
*prev
,
1209 struct mem_cgroup_reclaim_cookie
*reclaim
)
1211 struct mem_cgroup
*memcg
= NULL
;
1212 struct mem_cgroup
*last_visited
= NULL
;
1214 if (mem_cgroup_disabled())
1218 root
= root_mem_cgroup
;
1220 if (prev
&& !reclaim
)
1221 last_visited
= prev
;
1223 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1231 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1232 int uninitialized_var(seq
);
1235 struct mem_cgroup_per_zone
*mz
;
1237 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1238 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1239 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1240 iter
->last_visited
= NULL
;
1244 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1247 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1250 mem_cgroup_iter_update(iter
, last_visited
, memcg
, root
,
1255 else if (!prev
&& memcg
)
1256 reclaim
->generation
= iter
->generation
;
1265 if (prev
&& prev
!= root
)
1266 css_put(&prev
->css
);
1272 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1273 * @root: hierarchy root
1274 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1276 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1277 struct mem_cgroup
*prev
)
1280 root
= root_mem_cgroup
;
1281 if (prev
&& prev
!= root
)
1282 css_put(&prev
->css
);
1286 * Iteration constructs for visiting all cgroups (under a tree). If
1287 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1288 * be used for reference counting.
1290 #define for_each_mem_cgroup_tree(iter, root) \
1291 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1293 iter = mem_cgroup_iter(root, iter, NULL))
1295 #define for_each_mem_cgroup(iter) \
1296 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1298 iter = mem_cgroup_iter(NULL, iter, NULL))
1300 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1302 struct mem_cgroup
*memcg
;
1305 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1306 if (unlikely(!memcg
))
1311 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1314 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1322 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1325 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1326 * @zone: zone of the wanted lruvec
1327 * @memcg: memcg of the wanted lruvec
1329 * Returns the lru list vector holding pages for the given @zone and
1330 * @mem. This can be the global zone lruvec, if the memory controller
1333 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1334 struct mem_cgroup
*memcg
)
1336 struct mem_cgroup_per_zone
*mz
;
1337 struct lruvec
*lruvec
;
1339 if (mem_cgroup_disabled()) {
1340 lruvec
= &zone
->lruvec
;
1344 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1345 lruvec
= &mz
->lruvec
;
1348 * Since a node can be onlined after the mem_cgroup was created,
1349 * we have to be prepared to initialize lruvec->zone here;
1350 * and if offlined then reonlined, we need to reinitialize it.
1352 if (unlikely(lruvec
->zone
!= zone
))
1353 lruvec
->zone
= zone
;
1358 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1360 * @zone: zone of the page
1362 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1364 struct mem_cgroup_per_zone
*mz
;
1365 struct mem_cgroup
*memcg
;
1366 struct page_cgroup
*pc
;
1367 struct lruvec
*lruvec
;
1369 if (mem_cgroup_disabled()) {
1370 lruvec
= &zone
->lruvec
;
1374 pc
= lookup_page_cgroup(page
);
1375 memcg
= pc
->mem_cgroup
;
1378 * Surreptitiously switch any uncharged offlist page to root:
1379 * an uncharged page off lru does nothing to secure
1380 * its former mem_cgroup from sudden removal.
1382 * Our caller holds lru_lock, and PageCgroupUsed is updated
1383 * under page_cgroup lock: between them, they make all uses
1384 * of pc->mem_cgroup safe.
1386 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1387 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1389 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1390 lruvec
= &mz
->lruvec
;
1393 * Since a node can be onlined after the mem_cgroup was created,
1394 * we have to be prepared to initialize lruvec->zone here;
1395 * and if offlined then reonlined, we need to reinitialize it.
1397 if (unlikely(lruvec
->zone
!= zone
))
1398 lruvec
->zone
= zone
;
1403 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1404 * @lruvec: mem_cgroup per zone lru vector
1405 * @lru: index of lru list the page is sitting on
1406 * @nr_pages: positive when adding or negative when removing
1408 * This function must be called when a page is added to or removed from an
1411 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1414 struct mem_cgroup_per_zone
*mz
;
1415 unsigned long *lru_size
;
1417 if (mem_cgroup_disabled())
1420 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1421 lru_size
= mz
->lru_size
+ lru
;
1422 *lru_size
+= nr_pages
;
1423 VM_BUG_ON((long)(*lru_size
) < 0);
1427 * Checks whether given mem is same or in the root_mem_cgroup's
1430 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1431 struct mem_cgroup
*memcg
)
1433 if (root_memcg
== memcg
)
1435 if (!root_memcg
->use_hierarchy
|| !memcg
)
1437 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1440 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1441 struct mem_cgroup
*memcg
)
1446 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1451 bool task_in_mem_cgroup(struct task_struct
*task
,
1452 const struct mem_cgroup
*memcg
)
1454 struct mem_cgroup
*curr
= NULL
;
1455 struct task_struct
*p
;
1458 p
= find_lock_task_mm(task
);
1460 curr
= get_mem_cgroup_from_mm(p
->mm
);
1464 * All threads may have already detached their mm's, but the oom
1465 * killer still needs to detect if they have already been oom
1466 * killed to prevent needlessly killing additional tasks.
1469 curr
= mem_cgroup_from_task(task
);
1471 css_get(&curr
->css
);
1475 * We should check use_hierarchy of "memcg" not "curr". Because checking
1476 * use_hierarchy of "curr" here make this function true if hierarchy is
1477 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1478 * hierarchy(even if use_hierarchy is disabled in "memcg").
1480 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1481 css_put(&curr
->css
);
1485 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1487 unsigned long inactive_ratio
;
1488 unsigned long inactive
;
1489 unsigned long active
;
1492 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1493 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1495 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1497 inactive_ratio
= int_sqrt(10 * gb
);
1501 return inactive
* inactive_ratio
< active
;
1504 #define mem_cgroup_from_res_counter(counter, member) \
1505 container_of(counter, struct mem_cgroup, member)
1508 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1509 * @memcg: the memory cgroup
1511 * Returns the maximum amount of memory @mem can be charged with, in
1514 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1516 unsigned long long margin
;
1518 margin
= res_counter_margin(&memcg
->res
);
1519 if (do_swap_account
)
1520 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1521 return margin
>> PAGE_SHIFT
;
1524 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1527 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1528 return vm_swappiness
;
1530 return memcg
->swappiness
;
1534 * memcg->moving_account is used for checking possibility that some thread is
1535 * calling move_account(). When a thread on CPU-A starts moving pages under
1536 * a memcg, other threads should check memcg->moving_account under
1537 * rcu_read_lock(), like this:
1541 * memcg->moving_account+1 if (memcg->mocing_account)
1543 * synchronize_rcu() update something.
1548 /* for quick checking without looking up memcg */
1549 atomic_t memcg_moving __read_mostly
;
1551 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1553 atomic_inc(&memcg_moving
);
1554 atomic_inc(&memcg
->moving_account
);
1558 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1561 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1562 * We check NULL in callee rather than caller.
1565 atomic_dec(&memcg_moving
);
1566 atomic_dec(&memcg
->moving_account
);
1571 * A routine for checking "mem" is under move_account() or not.
1573 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1574 * moving cgroups. This is for waiting at high-memory pressure
1577 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1579 struct mem_cgroup
*from
;
1580 struct mem_cgroup
*to
;
1583 * Unlike task_move routines, we access mc.to, mc.from not under
1584 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1586 spin_lock(&mc
.lock
);
1592 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1593 || mem_cgroup_same_or_subtree(memcg
, to
);
1595 spin_unlock(&mc
.lock
);
1599 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1601 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1602 if (mem_cgroup_under_move(memcg
)) {
1604 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1605 /* moving charge context might have finished. */
1608 finish_wait(&mc
.waitq
, &wait
);
1616 * Take this lock when
1617 * - a code tries to modify page's memcg while it's USED.
1618 * - a code tries to modify page state accounting in a memcg.
1620 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1621 unsigned long *flags
)
1623 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1626 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1627 unsigned long *flags
)
1629 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1632 #define K(x) ((x) << (PAGE_SHIFT-10))
1634 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1635 * @memcg: The memory cgroup that went over limit
1636 * @p: Task that is going to be killed
1638 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1641 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1643 /* oom_info_lock ensures that parallel ooms do not interleave */
1644 static DEFINE_MUTEX(oom_info_lock
);
1645 struct mem_cgroup
*iter
;
1651 mutex_lock(&oom_info_lock
);
1654 pr_info("Task in ");
1655 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1656 pr_info(" killed as a result of limit of ");
1657 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1662 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1663 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1664 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1665 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1666 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1667 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1668 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1669 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1670 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1671 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1672 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1673 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1675 for_each_mem_cgroup_tree(iter
, memcg
) {
1676 pr_info("Memory cgroup stats for ");
1677 pr_cont_cgroup_path(iter
->css
.cgroup
);
1680 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1681 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1683 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1684 K(mem_cgroup_read_stat(iter
, i
)));
1687 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1688 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1689 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1693 mutex_unlock(&oom_info_lock
);
1697 * This function returns the number of memcg under hierarchy tree. Returns
1698 * 1(self count) if no children.
1700 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1703 struct mem_cgroup
*iter
;
1705 for_each_mem_cgroup_tree(iter
, memcg
)
1711 * Return the memory (and swap, if configured) limit for a memcg.
1713 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1717 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1720 * Do not consider swap space if we cannot swap due to swappiness
1722 if (mem_cgroup_swappiness(memcg
)) {
1725 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1726 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1729 * If memsw is finite and limits the amount of swap space
1730 * available to this memcg, return that limit.
1732 limit
= min(limit
, memsw
);
1738 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1741 struct mem_cgroup
*iter
;
1742 unsigned long chosen_points
= 0;
1743 unsigned long totalpages
;
1744 unsigned int points
= 0;
1745 struct task_struct
*chosen
= NULL
;
1748 * If current has a pending SIGKILL or is exiting, then automatically
1749 * select it. The goal is to allow it to allocate so that it may
1750 * quickly exit and free its memory.
1752 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1753 set_thread_flag(TIF_MEMDIE
);
1757 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1758 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1759 for_each_mem_cgroup_tree(iter
, memcg
) {
1760 struct css_task_iter it
;
1761 struct task_struct
*task
;
1763 css_task_iter_start(&iter
->css
, &it
);
1764 while ((task
= css_task_iter_next(&it
))) {
1765 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1767 case OOM_SCAN_SELECT
:
1769 put_task_struct(chosen
);
1771 chosen_points
= ULONG_MAX
;
1772 get_task_struct(chosen
);
1774 case OOM_SCAN_CONTINUE
:
1776 case OOM_SCAN_ABORT
:
1777 css_task_iter_end(&it
);
1778 mem_cgroup_iter_break(memcg
, iter
);
1780 put_task_struct(chosen
);
1785 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1786 if (!points
|| points
< chosen_points
)
1788 /* Prefer thread group leaders for display purposes */
1789 if (points
== chosen_points
&&
1790 thread_group_leader(chosen
))
1794 put_task_struct(chosen
);
1796 chosen_points
= points
;
1797 get_task_struct(chosen
);
1799 css_task_iter_end(&it
);
1804 points
= chosen_points
* 1000 / totalpages
;
1805 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1806 NULL
, "Memory cgroup out of memory");
1809 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1811 unsigned long flags
)
1813 unsigned long total
= 0;
1814 bool noswap
= false;
1817 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1819 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1822 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1824 drain_all_stock_async(memcg
);
1825 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1827 * Allow limit shrinkers, which are triggered directly
1828 * by userspace, to catch signals and stop reclaim
1829 * after minimal progress, regardless of the margin.
1831 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1833 if (mem_cgroup_margin(memcg
))
1836 * If nothing was reclaimed after two attempts, there
1837 * may be no reclaimable pages in this hierarchy.
1846 * test_mem_cgroup_node_reclaimable
1847 * @memcg: the target memcg
1848 * @nid: the node ID to be checked.
1849 * @noswap : specify true here if the user wants flle only information.
1851 * This function returns whether the specified memcg contains any
1852 * reclaimable pages on a node. Returns true if there are any reclaimable
1853 * pages in the node.
1855 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1856 int nid
, bool noswap
)
1858 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1860 if (noswap
|| !total_swap_pages
)
1862 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1867 #if MAX_NUMNODES > 1
1870 * Always updating the nodemask is not very good - even if we have an empty
1871 * list or the wrong list here, we can start from some node and traverse all
1872 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1875 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1879 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1880 * pagein/pageout changes since the last update.
1882 if (!atomic_read(&memcg
->numainfo_events
))
1884 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1887 /* make a nodemask where this memcg uses memory from */
1888 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1890 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1892 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1893 node_clear(nid
, memcg
->scan_nodes
);
1896 atomic_set(&memcg
->numainfo_events
, 0);
1897 atomic_set(&memcg
->numainfo_updating
, 0);
1901 * Selecting a node where we start reclaim from. Because what we need is just
1902 * reducing usage counter, start from anywhere is O,K. Considering
1903 * memory reclaim from current node, there are pros. and cons.
1905 * Freeing memory from current node means freeing memory from a node which
1906 * we'll use or we've used. So, it may make LRU bad. And if several threads
1907 * hit limits, it will see a contention on a node. But freeing from remote
1908 * node means more costs for memory reclaim because of memory latency.
1910 * Now, we use round-robin. Better algorithm is welcomed.
1912 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1916 mem_cgroup_may_update_nodemask(memcg
);
1917 node
= memcg
->last_scanned_node
;
1919 node
= next_node(node
, memcg
->scan_nodes
);
1920 if (node
== MAX_NUMNODES
)
1921 node
= first_node(memcg
->scan_nodes
);
1923 * We call this when we hit limit, not when pages are added to LRU.
1924 * No LRU may hold pages because all pages are UNEVICTABLE or
1925 * memcg is too small and all pages are not on LRU. In that case,
1926 * we use curret node.
1928 if (unlikely(node
== MAX_NUMNODES
))
1929 node
= numa_node_id();
1931 memcg
->last_scanned_node
= node
;
1936 * Check all nodes whether it contains reclaimable pages or not.
1937 * For quick scan, we make use of scan_nodes. This will allow us to skip
1938 * unused nodes. But scan_nodes is lazily updated and may not cotain
1939 * enough new information. We need to do double check.
1941 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1946 * quick check...making use of scan_node.
1947 * We can skip unused nodes.
1949 if (!nodes_empty(memcg
->scan_nodes
)) {
1950 for (nid
= first_node(memcg
->scan_nodes
);
1952 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1954 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1959 * Check rest of nodes.
1961 for_each_node_state(nid
, N_MEMORY
) {
1962 if (node_isset(nid
, memcg
->scan_nodes
))
1964 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1971 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1976 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1978 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1982 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1985 unsigned long *total_scanned
)
1987 struct mem_cgroup
*victim
= NULL
;
1990 unsigned long excess
;
1991 unsigned long nr_scanned
;
1992 struct mem_cgroup_reclaim_cookie reclaim
= {
1997 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2000 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2005 * If we have not been able to reclaim
2006 * anything, it might because there are
2007 * no reclaimable pages under this hierarchy
2012 * We want to do more targeted reclaim.
2013 * excess >> 2 is not to excessive so as to
2014 * reclaim too much, nor too less that we keep
2015 * coming back to reclaim from this cgroup
2017 if (total
>= (excess
>> 2) ||
2018 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2023 if (!mem_cgroup_reclaimable(victim
, false))
2025 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2027 *total_scanned
+= nr_scanned
;
2028 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2031 mem_cgroup_iter_break(root_memcg
, victim
);
2035 #ifdef CONFIG_LOCKDEP
2036 static struct lockdep_map memcg_oom_lock_dep_map
= {
2037 .name
= "memcg_oom_lock",
2041 static DEFINE_SPINLOCK(memcg_oom_lock
);
2044 * Check OOM-Killer is already running under our hierarchy.
2045 * If someone is running, return false.
2047 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2049 struct mem_cgroup
*iter
, *failed
= NULL
;
2051 spin_lock(&memcg_oom_lock
);
2053 for_each_mem_cgroup_tree(iter
, memcg
) {
2054 if (iter
->oom_lock
) {
2056 * this subtree of our hierarchy is already locked
2057 * so we cannot give a lock.
2060 mem_cgroup_iter_break(memcg
, iter
);
2063 iter
->oom_lock
= true;
2068 * OK, we failed to lock the whole subtree so we have
2069 * to clean up what we set up to the failing subtree
2071 for_each_mem_cgroup_tree(iter
, memcg
) {
2072 if (iter
== failed
) {
2073 mem_cgroup_iter_break(memcg
, iter
);
2076 iter
->oom_lock
= false;
2079 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2081 spin_unlock(&memcg_oom_lock
);
2086 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2088 struct mem_cgroup
*iter
;
2090 spin_lock(&memcg_oom_lock
);
2091 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2092 for_each_mem_cgroup_tree(iter
, memcg
)
2093 iter
->oom_lock
= false;
2094 spin_unlock(&memcg_oom_lock
);
2097 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2099 struct mem_cgroup
*iter
;
2101 for_each_mem_cgroup_tree(iter
, memcg
)
2102 atomic_inc(&iter
->under_oom
);
2105 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2107 struct mem_cgroup
*iter
;
2110 * When a new child is created while the hierarchy is under oom,
2111 * mem_cgroup_oom_lock() may not be called. We have to use
2112 * atomic_add_unless() here.
2114 for_each_mem_cgroup_tree(iter
, memcg
)
2115 atomic_add_unless(&iter
->under_oom
, -1, 0);
2118 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2120 struct oom_wait_info
{
2121 struct mem_cgroup
*memcg
;
2125 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2126 unsigned mode
, int sync
, void *arg
)
2128 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2129 struct mem_cgroup
*oom_wait_memcg
;
2130 struct oom_wait_info
*oom_wait_info
;
2132 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2133 oom_wait_memcg
= oom_wait_info
->memcg
;
2136 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2137 * Then we can use css_is_ancestor without taking care of RCU.
2139 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2140 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2142 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2145 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2147 atomic_inc(&memcg
->oom_wakeups
);
2148 /* for filtering, pass "memcg" as argument. */
2149 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2152 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2154 if (memcg
&& atomic_read(&memcg
->under_oom
))
2155 memcg_wakeup_oom(memcg
);
2158 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2160 if (!current
->memcg_oom
.may_oom
)
2163 * We are in the middle of the charge context here, so we
2164 * don't want to block when potentially sitting on a callstack
2165 * that holds all kinds of filesystem and mm locks.
2167 * Also, the caller may handle a failed allocation gracefully
2168 * (like optional page cache readahead) and so an OOM killer
2169 * invocation might not even be necessary.
2171 * That's why we don't do anything here except remember the
2172 * OOM context and then deal with it at the end of the page
2173 * fault when the stack is unwound, the locks are released,
2174 * and when we know whether the fault was overall successful.
2176 css_get(&memcg
->css
);
2177 current
->memcg_oom
.memcg
= memcg
;
2178 current
->memcg_oom
.gfp_mask
= mask
;
2179 current
->memcg_oom
.order
= order
;
2183 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2184 * @handle: actually kill/wait or just clean up the OOM state
2186 * This has to be called at the end of a page fault if the memcg OOM
2187 * handler was enabled.
2189 * Memcg supports userspace OOM handling where failed allocations must
2190 * sleep on a waitqueue until the userspace task resolves the
2191 * situation. Sleeping directly in the charge context with all kinds
2192 * of locks held is not a good idea, instead we remember an OOM state
2193 * in the task and mem_cgroup_oom_synchronize() has to be called at
2194 * the end of the page fault to complete the OOM handling.
2196 * Returns %true if an ongoing memcg OOM situation was detected and
2197 * completed, %false otherwise.
2199 bool mem_cgroup_oom_synchronize(bool handle
)
2201 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2202 struct oom_wait_info owait
;
2205 /* OOM is global, do not handle */
2212 owait
.memcg
= memcg
;
2213 owait
.wait
.flags
= 0;
2214 owait
.wait
.func
= memcg_oom_wake_function
;
2215 owait
.wait
.private = current
;
2216 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2218 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2219 mem_cgroup_mark_under_oom(memcg
);
2221 locked
= mem_cgroup_oom_trylock(memcg
);
2224 mem_cgroup_oom_notify(memcg
);
2226 if (locked
&& !memcg
->oom_kill_disable
) {
2227 mem_cgroup_unmark_under_oom(memcg
);
2228 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2229 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2230 current
->memcg_oom
.order
);
2233 mem_cgroup_unmark_under_oom(memcg
);
2234 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2238 mem_cgroup_oom_unlock(memcg
);
2240 * There is no guarantee that an OOM-lock contender
2241 * sees the wakeups triggered by the OOM kill
2242 * uncharges. Wake any sleepers explicitely.
2244 memcg_oom_recover(memcg
);
2247 current
->memcg_oom
.memcg
= NULL
;
2248 css_put(&memcg
->css
);
2253 * Used to update mapped file or writeback or other statistics.
2255 * Notes: Race condition
2257 * Charging occurs during page instantiation, while the page is
2258 * unmapped and locked in page migration, or while the page table is
2259 * locked in THP migration. No race is possible.
2261 * Uncharge happens to pages with zero references, no race possible.
2263 * Charge moving between groups is protected by checking mm->moving
2264 * account and taking the move_lock in the slowpath.
2267 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2268 bool *locked
, unsigned long *flags
)
2270 struct mem_cgroup
*memcg
;
2271 struct page_cgroup
*pc
;
2273 pc
= lookup_page_cgroup(page
);
2275 memcg
= pc
->mem_cgroup
;
2276 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2279 * If this memory cgroup is not under account moving, we don't
2280 * need to take move_lock_mem_cgroup(). Because we already hold
2281 * rcu_read_lock(), any calls to move_account will be delayed until
2282 * rcu_read_unlock().
2284 VM_BUG_ON(!rcu_read_lock_held());
2285 if (atomic_read(&memcg
->moving_account
) <= 0)
2288 move_lock_mem_cgroup(memcg
, flags
);
2289 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2290 move_unlock_mem_cgroup(memcg
, flags
);
2296 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2298 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2301 * It's guaranteed that pc->mem_cgroup never changes while
2302 * lock is held because a routine modifies pc->mem_cgroup
2303 * should take move_lock_mem_cgroup().
2305 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2308 void mem_cgroup_update_page_stat(struct page
*page
,
2309 enum mem_cgroup_stat_index idx
, int val
)
2311 struct mem_cgroup
*memcg
;
2312 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2313 unsigned long uninitialized_var(flags
);
2315 if (mem_cgroup_disabled())
2318 VM_BUG_ON(!rcu_read_lock_held());
2319 memcg
= pc
->mem_cgroup
;
2320 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2323 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2327 * size of first charge trial. "32" comes from vmscan.c's magic value.
2328 * TODO: maybe necessary to use big numbers in big irons.
2330 #define CHARGE_BATCH 32U
2331 struct memcg_stock_pcp
{
2332 struct mem_cgroup
*cached
; /* this never be root cgroup */
2333 unsigned int nr_pages
;
2334 struct work_struct work
;
2335 unsigned long flags
;
2336 #define FLUSHING_CACHED_CHARGE 0
2338 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2339 static DEFINE_MUTEX(percpu_charge_mutex
);
2342 * consume_stock: Try to consume stocked charge on this cpu.
2343 * @memcg: memcg to consume from.
2344 * @nr_pages: how many pages to charge.
2346 * The charges will only happen if @memcg matches the current cpu's memcg
2347 * stock, and at least @nr_pages are available in that stock. Failure to
2348 * service an allocation will refill the stock.
2350 * returns true if successful, false otherwise.
2352 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2354 struct memcg_stock_pcp
*stock
;
2357 if (nr_pages
> CHARGE_BATCH
)
2360 stock
= &get_cpu_var(memcg_stock
);
2361 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2362 stock
->nr_pages
-= nr_pages
;
2363 else /* need to call res_counter_charge */
2365 put_cpu_var(memcg_stock
);
2370 * Returns stocks cached in percpu to res_counter and reset cached information.
2372 static void drain_stock(struct memcg_stock_pcp
*stock
)
2374 struct mem_cgroup
*old
= stock
->cached
;
2376 if (stock
->nr_pages
) {
2377 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2379 res_counter_uncharge(&old
->res
, bytes
);
2380 if (do_swap_account
)
2381 res_counter_uncharge(&old
->memsw
, bytes
);
2382 stock
->nr_pages
= 0;
2384 stock
->cached
= NULL
;
2388 * This must be called under preempt disabled or must be called by
2389 * a thread which is pinned to local cpu.
2391 static void drain_local_stock(struct work_struct
*dummy
)
2393 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2395 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2398 static void __init
memcg_stock_init(void)
2402 for_each_possible_cpu(cpu
) {
2403 struct memcg_stock_pcp
*stock
=
2404 &per_cpu(memcg_stock
, cpu
);
2405 INIT_WORK(&stock
->work
, drain_local_stock
);
2410 * Cache charges(val) which is from res_counter, to local per_cpu area.
2411 * This will be consumed by consume_stock() function, later.
2413 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2415 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2417 if (stock
->cached
!= memcg
) { /* reset if necessary */
2419 stock
->cached
= memcg
;
2421 stock
->nr_pages
+= nr_pages
;
2422 put_cpu_var(memcg_stock
);
2426 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2427 * of the hierarchy under it. sync flag says whether we should block
2428 * until the work is done.
2430 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2434 /* Notify other cpus that system-wide "drain" is running */
2437 for_each_online_cpu(cpu
) {
2438 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2439 struct mem_cgroup
*memcg
;
2441 memcg
= stock
->cached
;
2442 if (!memcg
|| !stock
->nr_pages
)
2444 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2446 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2448 drain_local_stock(&stock
->work
);
2450 schedule_work_on(cpu
, &stock
->work
);
2458 for_each_online_cpu(cpu
) {
2459 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2460 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2461 flush_work(&stock
->work
);
2468 * Tries to drain stocked charges in other cpus. This function is asynchronous
2469 * and just put a work per cpu for draining localy on each cpu. Caller can
2470 * expects some charges will be back to res_counter later but cannot wait for
2473 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2476 * If someone calls draining, avoid adding more kworker runs.
2478 if (!mutex_trylock(&percpu_charge_mutex
))
2480 drain_all_stock(root_memcg
, false);
2481 mutex_unlock(&percpu_charge_mutex
);
2484 /* This is a synchronous drain interface. */
2485 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2487 /* called when force_empty is called */
2488 mutex_lock(&percpu_charge_mutex
);
2489 drain_all_stock(root_memcg
, true);
2490 mutex_unlock(&percpu_charge_mutex
);
2494 * This function drains percpu counter value from DEAD cpu and
2495 * move it to local cpu. Note that this function can be preempted.
2497 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2501 spin_lock(&memcg
->pcp_counter_lock
);
2502 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2503 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2505 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2506 memcg
->nocpu_base
.count
[i
] += x
;
2508 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2509 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2511 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2512 memcg
->nocpu_base
.events
[i
] += x
;
2514 spin_unlock(&memcg
->pcp_counter_lock
);
2517 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2518 unsigned long action
,
2521 int cpu
= (unsigned long)hcpu
;
2522 struct memcg_stock_pcp
*stock
;
2523 struct mem_cgroup
*iter
;
2525 if (action
== CPU_ONLINE
)
2528 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2531 for_each_mem_cgroup(iter
)
2532 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2534 stock
= &per_cpu(memcg_stock
, cpu
);
2539 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2540 unsigned int nr_pages
)
2542 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2543 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2544 struct mem_cgroup
*mem_over_limit
;
2545 struct res_counter
*fail_res
;
2546 unsigned long nr_reclaimed
;
2547 unsigned long flags
= 0;
2548 unsigned long long size
;
2551 if (mem_cgroup_is_root(memcg
))
2554 if (consume_stock(memcg
, nr_pages
))
2557 size
= batch
* PAGE_SIZE
;
2558 if (!res_counter_charge(&memcg
->res
, size
, &fail_res
)) {
2559 if (!do_swap_account
)
2561 if (!res_counter_charge(&memcg
->memsw
, size
, &fail_res
))
2563 res_counter_uncharge(&memcg
->res
, size
);
2564 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2565 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2567 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2569 if (batch
> nr_pages
) {
2575 * Unlike in global OOM situations, memcg is not in a physical
2576 * memory shortage. Allow dying and OOM-killed tasks to
2577 * bypass the last charges so that they can exit quickly and
2578 * free their memory.
2580 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2581 fatal_signal_pending(current
) ||
2582 current
->flags
& PF_EXITING
))
2585 if (unlikely(task_in_memcg_oom(current
)))
2588 if (!(gfp_mask
& __GFP_WAIT
))
2591 nr_reclaimed
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2593 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2596 if (gfp_mask
& __GFP_NORETRY
)
2599 * Even though the limit is exceeded at this point, reclaim
2600 * may have been able to free some pages. Retry the charge
2601 * before killing the task.
2603 * Only for regular pages, though: huge pages are rather
2604 * unlikely to succeed so close to the limit, and we fall back
2605 * to regular pages anyway in case of failure.
2607 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2610 * At task move, charge accounts can be doubly counted. So, it's
2611 * better to wait until the end of task_move if something is going on.
2613 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2619 if (gfp_mask
& __GFP_NOFAIL
)
2622 if (fatal_signal_pending(current
))
2625 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2627 if (!(gfp_mask
& __GFP_NOFAIL
))
2633 if (batch
> nr_pages
)
2634 refill_stock(memcg
, batch
- nr_pages
);
2639 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2641 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2643 if (mem_cgroup_is_root(memcg
))
2646 res_counter_uncharge(&memcg
->res
, bytes
);
2647 if (do_swap_account
)
2648 res_counter_uncharge(&memcg
->memsw
, bytes
);
2652 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2653 * This is useful when moving usage to parent cgroup.
2655 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2656 unsigned int nr_pages
)
2658 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2660 if (mem_cgroup_is_root(memcg
))
2663 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2664 if (do_swap_account
)
2665 res_counter_uncharge_until(&memcg
->memsw
,
2666 memcg
->memsw
.parent
, bytes
);
2670 * A helper function to get mem_cgroup from ID. must be called under
2671 * rcu_read_lock(). The caller is responsible for calling
2672 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2673 * refcnt from swap can be called against removed memcg.)
2675 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2677 /* ID 0 is unused ID */
2680 return mem_cgroup_from_id(id
);
2684 * try_get_mem_cgroup_from_page - look up page's memcg association
2687 * Look up, get a css reference, and return the memcg that owns @page.
2689 * The page must be locked to prevent racing with swap-in and page
2690 * cache charges. If coming from an unlocked page table, the caller
2691 * must ensure the page is on the LRU or this can race with charging.
2693 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2695 struct mem_cgroup
*memcg
= NULL
;
2696 struct page_cgroup
*pc
;
2700 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2702 pc
= lookup_page_cgroup(page
);
2703 if (PageCgroupUsed(pc
)) {
2704 memcg
= pc
->mem_cgroup
;
2705 if (memcg
&& !css_tryget_online(&memcg
->css
))
2707 } else if (PageSwapCache(page
)) {
2708 ent
.val
= page_private(page
);
2709 id
= lookup_swap_cgroup_id(ent
);
2711 memcg
= mem_cgroup_lookup(id
);
2712 if (memcg
&& !css_tryget_online(&memcg
->css
))
2719 static void lock_page_lru(struct page
*page
, int *isolated
)
2721 struct zone
*zone
= page_zone(page
);
2723 spin_lock_irq(&zone
->lru_lock
);
2724 if (PageLRU(page
)) {
2725 struct lruvec
*lruvec
;
2727 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2729 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2735 static void unlock_page_lru(struct page
*page
, int isolated
)
2737 struct zone
*zone
= page_zone(page
);
2740 struct lruvec
*lruvec
;
2742 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2743 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2745 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2747 spin_unlock_irq(&zone
->lru_lock
);
2750 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2753 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2756 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2758 * we don't need page_cgroup_lock about tail pages, becase they are not
2759 * accessed by any other context at this point.
2763 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2764 * may already be on some other mem_cgroup's LRU. Take care of it.
2767 lock_page_lru(page
, &isolated
);
2770 * Nobody should be changing or seriously looking at
2771 * pc->mem_cgroup and pc->flags at this point:
2773 * - the page is uncharged
2775 * - the page is off-LRU
2777 * - an anonymous fault has exclusive page access, except for
2778 * a locked page table
2780 * - a page cache insertion, a swapin fault, or a migration
2781 * have the page locked
2783 pc
->mem_cgroup
= memcg
;
2784 pc
->flags
= PCG_USED
| PCG_MEM
| (do_swap_account
? PCG_MEMSW
: 0);
2787 unlock_page_lru(page
, isolated
);
2790 static DEFINE_MUTEX(set_limit_mutex
);
2792 #ifdef CONFIG_MEMCG_KMEM
2794 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2795 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2797 static DEFINE_MUTEX(memcg_slab_mutex
);
2799 static DEFINE_MUTEX(activate_kmem_mutex
);
2801 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2803 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2804 memcg_kmem_is_active(memcg
);
2808 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2809 * in the memcg_cache_params struct.
2811 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2813 struct kmem_cache
*cachep
;
2815 VM_BUG_ON(p
->is_root_cache
);
2816 cachep
= p
->root_cache
;
2817 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2820 #ifdef CONFIG_SLABINFO
2821 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
2823 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
2824 struct memcg_cache_params
*params
;
2826 if (!memcg_can_account_kmem(memcg
))
2829 print_slabinfo_header(m
);
2831 mutex_lock(&memcg_slab_mutex
);
2832 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2833 cache_show(memcg_params_to_cache(params
), m
);
2834 mutex_unlock(&memcg_slab_mutex
);
2840 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2842 struct res_counter
*fail_res
;
2845 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2849 ret
= try_charge(memcg
, gfp
, size
>> PAGE_SHIFT
);
2850 if (ret
== -EINTR
) {
2852 * try_charge() chose to bypass to root due to OOM kill or
2853 * fatal signal. Since our only options are to either fail
2854 * the allocation or charge it to this cgroup, do it as a
2855 * temporary condition. But we can't fail. From a kmem/slab
2856 * perspective, the cache has already been selected, by
2857 * mem_cgroup_kmem_get_cache(), so it is too late to change
2860 * This condition will only trigger if the task entered
2861 * memcg_charge_kmem in a sane state, but was OOM-killed
2862 * during try_charge() above. Tasks that were already dying
2863 * when the allocation triggers should have been already
2864 * directed to the root cgroup in memcontrol.h
2866 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2867 if (do_swap_account
)
2868 res_counter_charge_nofail(&memcg
->memsw
, size
,
2872 res_counter_uncharge(&memcg
->kmem
, size
);
2877 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2879 res_counter_uncharge(&memcg
->res
, size
);
2880 if (do_swap_account
)
2881 res_counter_uncharge(&memcg
->memsw
, size
);
2884 if (res_counter_uncharge(&memcg
->kmem
, size
))
2888 * Releases a reference taken in kmem_cgroup_css_offline in case
2889 * this last uncharge is racing with the offlining code or it is
2890 * outliving the memcg existence.
2892 * The memory barrier imposed by test&clear is paired with the
2893 * explicit one in memcg_kmem_mark_dead().
2895 if (memcg_kmem_test_and_clear_dead(memcg
))
2896 css_put(&memcg
->css
);
2900 * helper for acessing a memcg's index. It will be used as an index in the
2901 * child cache array in kmem_cache, and also to derive its name. This function
2902 * will return -1 when this is not a kmem-limited memcg.
2904 int memcg_cache_id(struct mem_cgroup
*memcg
)
2906 return memcg
? memcg
->kmemcg_id
: -1;
2909 static size_t memcg_caches_array_size(int num_groups
)
2912 if (num_groups
<= 0)
2915 size
= 2 * num_groups
;
2916 if (size
< MEMCG_CACHES_MIN_SIZE
)
2917 size
= MEMCG_CACHES_MIN_SIZE
;
2918 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2919 size
= MEMCG_CACHES_MAX_SIZE
;
2925 * We should update the current array size iff all caches updates succeed. This
2926 * can only be done from the slab side. The slab mutex needs to be held when
2929 void memcg_update_array_size(int num
)
2931 if (num
> memcg_limited_groups_array_size
)
2932 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2935 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2937 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2939 VM_BUG_ON(!is_root_cache(s
));
2941 if (num_groups
> memcg_limited_groups_array_size
) {
2943 struct memcg_cache_params
*new_params
;
2944 ssize_t size
= memcg_caches_array_size(num_groups
);
2946 size
*= sizeof(void *);
2947 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
2949 new_params
= kzalloc(size
, GFP_KERNEL
);
2953 new_params
->is_root_cache
= true;
2956 * There is the chance it will be bigger than
2957 * memcg_limited_groups_array_size, if we failed an allocation
2958 * in a cache, in which case all caches updated before it, will
2959 * have a bigger array.
2961 * But if that is the case, the data after
2962 * memcg_limited_groups_array_size is certainly unused
2964 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
2965 if (!cur_params
->memcg_caches
[i
])
2967 new_params
->memcg_caches
[i
] =
2968 cur_params
->memcg_caches
[i
];
2972 * Ideally, we would wait until all caches succeed, and only
2973 * then free the old one. But this is not worth the extra
2974 * pointer per-cache we'd have to have for this.
2976 * It is not a big deal if some caches are left with a size
2977 * bigger than the others. And all updates will reset this
2980 rcu_assign_pointer(s
->memcg_params
, new_params
);
2982 kfree_rcu(cur_params
, rcu_head
);
2987 int memcg_alloc_cache_params(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
2988 struct kmem_cache
*root_cache
)
2992 if (!memcg_kmem_enabled())
2996 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
2997 size
+= memcg_limited_groups_array_size
* sizeof(void *);
2999 size
= sizeof(struct memcg_cache_params
);
3001 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3002 if (!s
->memcg_params
)
3006 s
->memcg_params
->memcg
= memcg
;
3007 s
->memcg_params
->root_cache
= root_cache
;
3008 css_get(&memcg
->css
);
3010 s
->memcg_params
->is_root_cache
= true;
3015 void memcg_free_cache_params(struct kmem_cache
*s
)
3017 if (!s
->memcg_params
)
3019 if (!s
->memcg_params
->is_root_cache
)
3020 css_put(&s
->memcg_params
->memcg
->css
);
3021 kfree(s
->memcg_params
);
3024 static void memcg_register_cache(struct mem_cgroup
*memcg
,
3025 struct kmem_cache
*root_cache
)
3027 static char memcg_name_buf
[NAME_MAX
+ 1]; /* protected by
3029 struct kmem_cache
*cachep
;
3032 lockdep_assert_held(&memcg_slab_mutex
);
3034 id
= memcg_cache_id(memcg
);
3037 * Since per-memcg caches are created asynchronously on first
3038 * allocation (see memcg_kmem_get_cache()), several threads can try to
3039 * create the same cache, but only one of them may succeed.
3041 if (cache_from_memcg_idx(root_cache
, id
))
3044 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
3045 cachep
= memcg_create_kmem_cache(memcg
, root_cache
, memcg_name_buf
);
3047 * If we could not create a memcg cache, do not complain, because
3048 * that's not critical at all as we can always proceed with the root
3054 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3057 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3058 * barrier here to ensure nobody will see the kmem_cache partially
3063 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
3064 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
3067 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
3069 struct kmem_cache
*root_cache
;
3070 struct mem_cgroup
*memcg
;
3073 lockdep_assert_held(&memcg_slab_mutex
);
3075 BUG_ON(is_root_cache(cachep
));
3077 root_cache
= cachep
->memcg_params
->root_cache
;
3078 memcg
= cachep
->memcg_params
->memcg
;
3079 id
= memcg_cache_id(memcg
);
3081 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
3082 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
3084 list_del(&cachep
->memcg_params
->list
);
3086 kmem_cache_destroy(cachep
);
3090 * During the creation a new cache, we need to disable our accounting mechanism
3091 * altogether. This is true even if we are not creating, but rather just
3092 * enqueing new caches to be created.
3094 * This is because that process will trigger allocations; some visible, like
3095 * explicit kmallocs to auxiliary data structures, name strings and internal
3096 * cache structures; some well concealed, like INIT_WORK() that can allocate
3097 * objects during debug.
3099 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3100 * to it. This may not be a bounded recursion: since the first cache creation
3101 * failed to complete (waiting on the allocation), we'll just try to create the
3102 * cache again, failing at the same point.
3104 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3105 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3106 * inside the following two functions.
3108 static inline void memcg_stop_kmem_account(void)
3110 VM_BUG_ON(!current
->mm
);
3111 current
->memcg_kmem_skip_account
++;
3114 static inline void memcg_resume_kmem_account(void)
3116 VM_BUG_ON(!current
->mm
);
3117 current
->memcg_kmem_skip_account
--;
3120 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
3122 struct kmem_cache
*c
;
3125 mutex_lock(&memcg_slab_mutex
);
3126 for_each_memcg_cache_index(i
) {
3127 c
= cache_from_memcg_idx(s
, i
);
3131 memcg_unregister_cache(c
);
3133 if (cache_from_memcg_idx(s
, i
))
3136 mutex_unlock(&memcg_slab_mutex
);
3140 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3142 struct kmem_cache
*cachep
;
3143 struct memcg_cache_params
*params
, *tmp
;
3145 if (!memcg_kmem_is_active(memcg
))
3148 mutex_lock(&memcg_slab_mutex
);
3149 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
3150 cachep
= memcg_params_to_cache(params
);
3151 kmem_cache_shrink(cachep
);
3152 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3153 memcg_unregister_cache(cachep
);
3155 mutex_unlock(&memcg_slab_mutex
);
3158 struct memcg_register_cache_work
{
3159 struct mem_cgroup
*memcg
;
3160 struct kmem_cache
*cachep
;
3161 struct work_struct work
;
3164 static void memcg_register_cache_func(struct work_struct
*w
)
3166 struct memcg_register_cache_work
*cw
=
3167 container_of(w
, struct memcg_register_cache_work
, work
);
3168 struct mem_cgroup
*memcg
= cw
->memcg
;
3169 struct kmem_cache
*cachep
= cw
->cachep
;
3171 mutex_lock(&memcg_slab_mutex
);
3172 memcg_register_cache(memcg
, cachep
);
3173 mutex_unlock(&memcg_slab_mutex
);
3175 css_put(&memcg
->css
);
3180 * Enqueue the creation of a per-memcg kmem_cache.
3182 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3183 struct kmem_cache
*cachep
)
3185 struct memcg_register_cache_work
*cw
;
3187 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
3189 css_put(&memcg
->css
);
3194 cw
->cachep
= cachep
;
3196 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
3197 schedule_work(&cw
->work
);
3200 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3201 struct kmem_cache
*cachep
)
3204 * We need to stop accounting when we kmalloc, because if the
3205 * corresponding kmalloc cache is not yet created, the first allocation
3206 * in __memcg_schedule_register_cache will recurse.
3208 * However, it is better to enclose the whole function. Depending on
3209 * the debugging options enabled, INIT_WORK(), for instance, can
3210 * trigger an allocation. This too, will make us recurse. Because at
3211 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3212 * the safest choice is to do it like this, wrapping the whole function.
3214 memcg_stop_kmem_account();
3215 __memcg_schedule_register_cache(memcg
, cachep
);
3216 memcg_resume_kmem_account();
3219 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
3223 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
,
3224 PAGE_SIZE
<< order
);
3226 atomic_add(1 << order
, &cachep
->memcg_params
->nr_pages
);
3230 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
3232 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, PAGE_SIZE
<< order
);
3233 atomic_sub(1 << order
, &cachep
->memcg_params
->nr_pages
);
3237 * Return the kmem_cache we're supposed to use for a slab allocation.
3238 * We try to use the current memcg's version of the cache.
3240 * If the cache does not exist yet, if we are the first user of it,
3241 * we either create it immediately, if possible, or create it asynchronously
3243 * In the latter case, we will let the current allocation go through with
3244 * the original cache.
3246 * Can't be called in interrupt context or from kernel threads.
3247 * This function needs to be called with rcu_read_lock() held.
3249 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3252 struct mem_cgroup
*memcg
;
3253 struct kmem_cache
*memcg_cachep
;
3255 VM_BUG_ON(!cachep
->memcg_params
);
3256 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3258 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3262 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3264 if (!memcg_can_account_kmem(memcg
))
3267 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3268 if (likely(memcg_cachep
)) {
3269 cachep
= memcg_cachep
;
3273 /* The corresponding put will be done in the workqueue. */
3274 if (!css_tryget_online(&memcg
->css
))
3279 * If we are in a safe context (can wait, and not in interrupt
3280 * context), we could be be predictable and return right away.
3281 * This would guarantee that the allocation being performed
3282 * already belongs in the new cache.
3284 * However, there are some clashes that can arrive from locking.
3285 * For instance, because we acquire the slab_mutex while doing
3286 * memcg_create_kmem_cache, this means no further allocation
3287 * could happen with the slab_mutex held. So it's better to
3290 memcg_schedule_register_cache(memcg
, cachep
);
3298 * We need to verify if the allocation against current->mm->owner's memcg is
3299 * possible for the given order. But the page is not allocated yet, so we'll
3300 * need a further commit step to do the final arrangements.
3302 * It is possible for the task to switch cgroups in this mean time, so at
3303 * commit time, we can't rely on task conversion any longer. We'll then use
3304 * the handle argument to return to the caller which cgroup we should commit
3305 * against. We could also return the memcg directly and avoid the pointer
3306 * passing, but a boolean return value gives better semantics considering
3307 * the compiled-out case as well.
3309 * Returning true means the allocation is possible.
3312 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3314 struct mem_cgroup
*memcg
;
3320 * Disabling accounting is only relevant for some specific memcg
3321 * internal allocations. Therefore we would initially not have such
3322 * check here, since direct calls to the page allocator that are
3323 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3324 * outside memcg core. We are mostly concerned with cache allocations,
3325 * and by having this test at memcg_kmem_get_cache, we are already able
3326 * to relay the allocation to the root cache and bypass the memcg cache
3329 * There is one exception, though: the SLUB allocator does not create
3330 * large order caches, but rather service large kmallocs directly from
3331 * the page allocator. Therefore, the following sequence when backed by
3332 * the SLUB allocator:
3334 * memcg_stop_kmem_account();
3335 * kmalloc(<large_number>)
3336 * memcg_resume_kmem_account();
3338 * would effectively ignore the fact that we should skip accounting,
3339 * since it will drive us directly to this function without passing
3340 * through the cache selector memcg_kmem_get_cache. Such large
3341 * allocations are extremely rare but can happen, for instance, for the
3342 * cache arrays. We bring this test here.
3344 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3347 memcg
= get_mem_cgroup_from_mm(current
->mm
);
3349 if (!memcg_can_account_kmem(memcg
)) {
3350 css_put(&memcg
->css
);
3354 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3358 css_put(&memcg
->css
);
3362 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3365 struct page_cgroup
*pc
;
3367 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3369 /* The page allocation failed. Revert */
3371 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3375 * The page is freshly allocated and not visible to any
3376 * outside callers yet. Set up pc non-atomically.
3378 pc
= lookup_page_cgroup(page
);
3379 pc
->mem_cgroup
= memcg
;
3380 pc
->flags
= PCG_USED
;
3383 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3385 struct mem_cgroup
*memcg
= NULL
;
3386 struct page_cgroup
*pc
;
3389 pc
= lookup_page_cgroup(page
);
3390 if (!PageCgroupUsed(pc
))
3393 memcg
= pc
->mem_cgroup
;
3397 * We trust that only if there is a memcg associated with the page, it
3398 * is a valid allocation
3403 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3404 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3407 static inline void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3410 #endif /* CONFIG_MEMCG_KMEM */
3412 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3415 * Because tail pages are not marked as "used", set it. We're under
3416 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3417 * charge/uncharge will be never happen and move_account() is done under
3418 * compound_lock(), so we don't have to take care of races.
3420 void mem_cgroup_split_huge_fixup(struct page
*head
)
3422 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3423 struct page_cgroup
*pc
;
3424 struct mem_cgroup
*memcg
;
3427 if (mem_cgroup_disabled())
3430 memcg
= head_pc
->mem_cgroup
;
3431 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3433 pc
->mem_cgroup
= memcg
;
3434 pc
->flags
= head_pc
->flags
;
3436 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3439 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3442 * mem_cgroup_move_account - move account of the page
3444 * @nr_pages: number of regular pages (>1 for huge pages)
3445 * @pc: page_cgroup of the page.
3446 * @from: mem_cgroup which the page is moved from.
3447 * @to: mem_cgroup which the page is moved to. @from != @to.
3449 * The caller must confirm following.
3450 * - page is not on LRU (isolate_page() is useful.)
3451 * - compound_lock is held when nr_pages > 1
3453 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3456 static int mem_cgroup_move_account(struct page
*page
,
3457 unsigned int nr_pages
,
3458 struct page_cgroup
*pc
,
3459 struct mem_cgroup
*from
,
3460 struct mem_cgroup
*to
)
3462 unsigned long flags
;
3465 VM_BUG_ON(from
== to
);
3466 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3468 * The page is isolated from LRU. So, collapse function
3469 * will not handle this page. But page splitting can happen.
3470 * Do this check under compound_page_lock(). The caller should
3474 if (nr_pages
> 1 && !PageTransHuge(page
))
3478 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3479 * of its source page while we change it: page migration takes
3480 * both pages off the LRU, but page cache replacement doesn't.
3482 if (!trylock_page(page
))
3486 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3489 move_lock_mem_cgroup(from
, &flags
);
3491 if (!PageAnon(page
) && page_mapped(page
)) {
3492 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3494 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3498 if (PageWriteback(page
)) {
3499 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3501 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3506 * It is safe to change pc->mem_cgroup here because the page
3507 * is referenced, charged, and isolated - we can't race with
3508 * uncharging, charging, migration, or LRU putback.
3511 /* caller should have done css_get */
3512 pc
->mem_cgroup
= to
;
3513 move_unlock_mem_cgroup(from
, &flags
);
3516 local_irq_disable();
3517 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
3518 memcg_check_events(to
, page
);
3519 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
3520 memcg_check_events(from
, page
);
3529 * mem_cgroup_move_parent - moves page to the parent group
3530 * @page: the page to move
3531 * @pc: page_cgroup of the page
3532 * @child: page's cgroup
3534 * move charges to its parent or the root cgroup if the group has no
3535 * parent (aka use_hierarchy==0).
3536 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3537 * mem_cgroup_move_account fails) the failure is always temporary and
3538 * it signals a race with a page removal/uncharge or migration. In the
3539 * first case the page is on the way out and it will vanish from the LRU
3540 * on the next attempt and the call should be retried later.
3541 * Isolation from the LRU fails only if page has been isolated from
3542 * the LRU since we looked at it and that usually means either global
3543 * reclaim or migration going on. The page will either get back to the
3545 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3546 * (!PageCgroupUsed) or moved to a different group. The page will
3547 * disappear in the next attempt.
3549 static int mem_cgroup_move_parent(struct page
*page
,
3550 struct page_cgroup
*pc
,
3551 struct mem_cgroup
*child
)
3553 struct mem_cgroup
*parent
;
3554 unsigned int nr_pages
;
3555 unsigned long uninitialized_var(flags
);
3558 VM_BUG_ON(mem_cgroup_is_root(child
));
3561 if (!get_page_unless_zero(page
))
3563 if (isolate_lru_page(page
))
3566 nr_pages
= hpage_nr_pages(page
);
3568 parent
= parent_mem_cgroup(child
);
3570 * If no parent, move charges to root cgroup.
3573 parent
= root_mem_cgroup
;
3576 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3577 flags
= compound_lock_irqsave(page
);
3580 ret
= mem_cgroup_move_account(page
, nr_pages
,
3583 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3586 compound_unlock_irqrestore(page
, flags
);
3587 putback_lru_page(page
);
3594 #ifdef CONFIG_MEMCG_SWAP
3595 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
3598 int val
= (charge
) ? 1 : -1;
3599 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
3603 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3604 * @entry: swap entry to be moved
3605 * @from: mem_cgroup which the entry is moved from
3606 * @to: mem_cgroup which the entry is moved to
3608 * It succeeds only when the swap_cgroup's record for this entry is the same
3609 * as the mem_cgroup's id of @from.
3611 * Returns 0 on success, -EINVAL on failure.
3613 * The caller must have charged to @to, IOW, called res_counter_charge() about
3614 * both res and memsw, and called css_get().
3616 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3617 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3619 unsigned short old_id
, new_id
;
3621 old_id
= mem_cgroup_id(from
);
3622 new_id
= mem_cgroup_id(to
);
3624 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3625 mem_cgroup_swap_statistics(from
, false);
3626 mem_cgroup_swap_statistics(to
, true);
3628 * This function is only called from task migration context now.
3629 * It postpones res_counter and refcount handling till the end
3630 * of task migration(mem_cgroup_clear_mc()) for performance
3631 * improvement. But we cannot postpone css_get(to) because if
3632 * the process that has been moved to @to does swap-in, the
3633 * refcount of @to might be decreased to 0.
3635 * We are in attach() phase, so the cgroup is guaranteed to be
3636 * alive, so we can just call css_get().
3644 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3645 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3651 #ifdef CONFIG_DEBUG_VM
3652 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
3654 struct page_cgroup
*pc
;
3656 pc
= lookup_page_cgroup(page
);
3658 * Can be NULL while feeding pages into the page allocator for
3659 * the first time, i.e. during boot or memory hotplug;
3660 * or when mem_cgroup_disabled().
3662 if (likely(pc
) && PageCgroupUsed(pc
))
3667 bool mem_cgroup_bad_page_check(struct page
*page
)
3669 if (mem_cgroup_disabled())
3672 return lookup_page_cgroup_used(page
) != NULL
;
3675 void mem_cgroup_print_bad_page(struct page
*page
)
3677 struct page_cgroup
*pc
;
3679 pc
= lookup_page_cgroup_used(page
);
3681 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3682 pc
, pc
->flags
, pc
->mem_cgroup
);
3687 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3688 unsigned long long val
)
3691 u64 memswlimit
, memlimit
;
3693 int children
= mem_cgroup_count_children(memcg
);
3694 u64 curusage
, oldusage
;
3698 * For keeping hierarchical_reclaim simple, how long we should retry
3699 * is depends on callers. We set our retry-count to be function
3700 * of # of children which we should visit in this loop.
3702 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
3704 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3707 while (retry_count
) {
3708 if (signal_pending(current
)) {
3713 * Rather than hide all in some function, I do this in
3714 * open coded manner. You see what this really does.
3715 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3717 mutex_lock(&set_limit_mutex
);
3718 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3719 if (memswlimit
< val
) {
3721 mutex_unlock(&set_limit_mutex
);
3725 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3729 ret
= res_counter_set_limit(&memcg
->res
, val
);
3731 if (memswlimit
== val
)
3732 memcg
->memsw_is_minimum
= true;
3734 memcg
->memsw_is_minimum
= false;
3736 mutex_unlock(&set_limit_mutex
);
3741 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
3742 MEM_CGROUP_RECLAIM_SHRINK
);
3743 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3744 /* Usage is reduced ? */
3745 if (curusage
>= oldusage
)
3748 oldusage
= curusage
;
3750 if (!ret
&& enlarge
)
3751 memcg_oom_recover(memcg
);
3756 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3757 unsigned long long val
)
3760 u64 memlimit
, memswlimit
, oldusage
, curusage
;
3761 int children
= mem_cgroup_count_children(memcg
);
3765 /* see mem_cgroup_resize_res_limit */
3766 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
3767 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3768 while (retry_count
) {
3769 if (signal_pending(current
)) {
3774 * Rather than hide all in some function, I do this in
3775 * open coded manner. You see what this really does.
3776 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3778 mutex_lock(&set_limit_mutex
);
3779 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3780 if (memlimit
> val
) {
3782 mutex_unlock(&set_limit_mutex
);
3785 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3786 if (memswlimit
< val
)
3788 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
3790 if (memlimit
== val
)
3791 memcg
->memsw_is_minimum
= true;
3793 memcg
->memsw_is_minimum
= false;
3795 mutex_unlock(&set_limit_mutex
);
3800 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
3801 MEM_CGROUP_RECLAIM_NOSWAP
|
3802 MEM_CGROUP_RECLAIM_SHRINK
);
3803 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3804 /* Usage is reduced ? */
3805 if (curusage
>= oldusage
)
3808 oldusage
= curusage
;
3810 if (!ret
&& enlarge
)
3811 memcg_oom_recover(memcg
);
3815 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3817 unsigned long *total_scanned
)
3819 unsigned long nr_reclaimed
= 0;
3820 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3821 unsigned long reclaimed
;
3823 struct mem_cgroup_tree_per_zone
*mctz
;
3824 unsigned long long excess
;
3825 unsigned long nr_scanned
;
3830 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3832 * This loop can run a while, specially if mem_cgroup's continuously
3833 * keep exceeding their soft limit and putting the system under
3840 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3845 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
3846 gfp_mask
, &nr_scanned
);
3847 nr_reclaimed
+= reclaimed
;
3848 *total_scanned
+= nr_scanned
;
3849 spin_lock_irq(&mctz
->lock
);
3852 * If we failed to reclaim anything from this memory cgroup
3853 * it is time to move on to the next cgroup
3859 * Loop until we find yet another one.
3861 * By the time we get the soft_limit lock
3862 * again, someone might have aded the
3863 * group back on the RB tree. Iterate to
3864 * make sure we get a different mem.
3865 * mem_cgroup_largest_soft_limit_node returns
3866 * NULL if no other cgroup is present on
3870 __mem_cgroup_largest_soft_limit_node(mctz
);
3872 css_put(&next_mz
->memcg
->css
);
3873 else /* next_mz == NULL or other memcg */
3877 __mem_cgroup_remove_exceeded(mz
, mctz
);
3878 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
3880 * One school of thought says that we should not add
3881 * back the node to the tree if reclaim returns 0.
3882 * But our reclaim could return 0, simply because due
3883 * to priority we are exposing a smaller subset of
3884 * memory to reclaim from. Consider this as a longer
3887 /* If excess == 0, no tree ops */
3888 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3889 spin_unlock_irq(&mctz
->lock
);
3890 css_put(&mz
->memcg
->css
);
3893 * Could not reclaim anything and there are no more
3894 * mem cgroups to try or we seem to be looping without
3895 * reclaiming anything.
3897 if (!nr_reclaimed
&&
3899 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3901 } while (!nr_reclaimed
);
3903 css_put(&next_mz
->memcg
->css
);
3904 return nr_reclaimed
;
3908 * mem_cgroup_force_empty_list - clears LRU of a group
3909 * @memcg: group to clear
3912 * @lru: lru to to clear
3914 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3915 * reclaim the pages page themselves - pages are moved to the parent (or root)
3918 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
3919 int node
, int zid
, enum lru_list lru
)
3921 struct lruvec
*lruvec
;
3922 unsigned long flags
;
3923 struct list_head
*list
;
3927 zone
= &NODE_DATA(node
)->node_zones
[zid
];
3928 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
3929 list
= &lruvec
->lists
[lru
];
3933 struct page_cgroup
*pc
;
3936 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3937 if (list_empty(list
)) {
3938 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3941 page
= list_entry(list
->prev
, struct page
, lru
);
3943 list_move(&page
->lru
, list
);
3945 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3948 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3950 pc
= lookup_page_cgroup(page
);
3952 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
3953 /* found lock contention or "pc" is obsolete. */
3958 } while (!list_empty(list
));
3962 * make mem_cgroup's charge to be 0 if there is no task by moving
3963 * all the charges and pages to the parent.
3964 * This enables deleting this mem_cgroup.
3966 * Caller is responsible for holding css reference on the memcg.
3968 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
3974 /* This is for making all *used* pages to be on LRU. */
3975 lru_add_drain_all();
3976 drain_all_stock_sync(memcg
);
3977 mem_cgroup_start_move(memcg
);
3978 for_each_node_state(node
, N_MEMORY
) {
3979 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3982 mem_cgroup_force_empty_list(memcg
,
3987 mem_cgroup_end_move(memcg
);
3988 memcg_oom_recover(memcg
);
3992 * Kernel memory may not necessarily be trackable to a specific
3993 * process. So they are not migrated, and therefore we can't
3994 * expect their value to drop to 0 here.
3995 * Having res filled up with kmem only is enough.
3997 * This is a safety check because mem_cgroup_force_empty_list
3998 * could have raced with mem_cgroup_replace_page_cache callers
3999 * so the lru seemed empty but the page could have been added
4000 * right after the check. RES_USAGE should be safe as we always
4001 * charge before adding to the LRU.
4003 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4004 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4005 } while (usage
> 0);
4009 * Test whether @memcg has children, dead or alive. Note that this
4010 * function doesn't care whether @memcg has use_hierarchy enabled and
4011 * returns %true if there are child csses according to the cgroup
4012 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
4014 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4019 * The lock does not prevent addition or deletion of children, but
4020 * it prevents a new child from being initialized based on this
4021 * parent in css_online(), so it's enough to decide whether
4022 * hierarchically inherited attributes can still be changed or not.
4024 lockdep_assert_held(&memcg_create_mutex
);
4027 ret
= css_next_child(NULL
, &memcg
->css
);
4033 * Reclaims as many pages from the given memcg as possible and moves
4034 * the rest to the parent.
4036 * Caller is responsible for holding css reference for memcg.
4038 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4040 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4042 /* we call try-to-free pages for make this cgroup empty */
4043 lru_add_drain_all();
4044 /* try to free all pages in this cgroup */
4045 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4048 if (signal_pending(current
))
4051 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4055 /* maybe some writeback is necessary */
4056 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4064 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
4065 char *buf
, size_t nbytes
,
4068 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4070 if (mem_cgroup_is_root(memcg
))
4072 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
4075 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4078 return mem_cgroup_from_css(css
)->use_hierarchy
;
4081 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4082 struct cftype
*cft
, u64 val
)
4085 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4086 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
4088 mutex_lock(&memcg_create_mutex
);
4090 if (memcg
->use_hierarchy
== val
)
4094 * If parent's use_hierarchy is set, we can't make any modifications
4095 * in the child subtrees. If it is unset, then the change can
4096 * occur, provided the current cgroup has no children.
4098 * For the root cgroup, parent_mem is NULL, we allow value to be
4099 * set if there are no children.
4101 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4102 (val
== 1 || val
== 0)) {
4103 if (!memcg_has_children(memcg
))
4104 memcg
->use_hierarchy
= val
;
4111 mutex_unlock(&memcg_create_mutex
);
4116 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4117 enum mem_cgroup_stat_index idx
)
4119 struct mem_cgroup
*iter
;
4122 /* Per-cpu values can be negative, use a signed accumulator */
4123 for_each_mem_cgroup_tree(iter
, memcg
)
4124 val
+= mem_cgroup_read_stat(iter
, idx
);
4126 if (val
< 0) /* race ? */
4131 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4135 if (!mem_cgroup_is_root(memcg
)) {
4137 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4139 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4143 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4144 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4146 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4147 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4150 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4152 return val
<< PAGE_SHIFT
;
4156 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
4159 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4160 enum res_type type
= MEMFILE_TYPE(cft
->private);
4161 int name
= MEMFILE_ATTR(cft
->private);
4165 if (name
== RES_USAGE
)
4166 return mem_cgroup_usage(memcg
, false);
4167 return res_counter_read_u64(&memcg
->res
, name
);
4169 if (name
== RES_USAGE
)
4170 return mem_cgroup_usage(memcg
, true);
4171 return res_counter_read_u64(&memcg
->memsw
, name
);
4173 return res_counter_read_u64(&memcg
->kmem
, name
);
4180 #ifdef CONFIG_MEMCG_KMEM
4181 /* should be called with activate_kmem_mutex held */
4182 static int __memcg_activate_kmem(struct mem_cgroup
*memcg
,
4183 unsigned long long limit
)
4188 if (memcg_kmem_is_active(memcg
))
4192 * We are going to allocate memory for data shared by all memory
4193 * cgroups so let's stop accounting here.
4195 memcg_stop_kmem_account();
4198 * For simplicity, we won't allow this to be disabled. It also can't
4199 * be changed if the cgroup has children already, or if tasks had
4202 * If tasks join before we set the limit, a person looking at
4203 * kmem.usage_in_bytes will have no way to determine when it took
4204 * place, which makes the value quite meaningless.
4206 * After it first became limited, changes in the value of the limit are
4207 * of course permitted.
4209 mutex_lock(&memcg_create_mutex
);
4210 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
4211 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
4213 mutex_unlock(&memcg_create_mutex
);
4217 memcg_id
= ida_simple_get(&kmem_limited_groups
,
4218 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
4225 * Make sure we have enough space for this cgroup in each root cache's
4228 mutex_lock(&memcg_slab_mutex
);
4229 err
= memcg_update_all_caches(memcg_id
+ 1);
4230 mutex_unlock(&memcg_slab_mutex
);
4234 memcg
->kmemcg_id
= memcg_id
;
4235 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
4238 * We couldn't have accounted to this cgroup, because it hasn't got the
4239 * active bit set yet, so this should succeed.
4241 err
= res_counter_set_limit(&memcg
->kmem
, limit
);
4244 static_key_slow_inc(&memcg_kmem_enabled_key
);
4246 * Setting the active bit after enabling static branching will
4247 * guarantee no one starts accounting before all call sites are
4250 memcg_kmem_set_active(memcg
);
4252 memcg_resume_kmem_account();
4256 ida_simple_remove(&kmem_limited_groups
, memcg_id
);
4260 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
4261 unsigned long long limit
)
4265 mutex_lock(&activate_kmem_mutex
);
4266 ret
= __memcg_activate_kmem(memcg
, limit
);
4267 mutex_unlock(&activate_kmem_mutex
);
4271 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4272 unsigned long long val
)
4276 if (!memcg_kmem_is_active(memcg
))
4277 ret
= memcg_activate_kmem(memcg
, val
);
4279 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4283 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4286 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4291 mutex_lock(&activate_kmem_mutex
);
4293 * If the parent cgroup is not kmem-active now, it cannot be activated
4294 * after this point, because it has at least one child already.
4296 if (memcg_kmem_is_active(parent
))
4297 ret
= __memcg_activate_kmem(memcg
, RES_COUNTER_MAX
);
4298 mutex_unlock(&activate_kmem_mutex
);
4302 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4303 unsigned long long val
)
4307 #endif /* CONFIG_MEMCG_KMEM */
4310 * The user of this function is...
4313 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
4314 char *buf
, size_t nbytes
, loff_t off
)
4316 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4319 unsigned long long val
;
4322 buf
= strstrip(buf
);
4323 type
= MEMFILE_TYPE(of_cft(of
)->private);
4324 name
= MEMFILE_ATTR(of_cft(of
)->private);
4328 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4332 /* This function does all necessary parse...reuse it */
4333 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4337 ret
= mem_cgroup_resize_limit(memcg
, val
);
4338 else if (type
== _MEMSWAP
)
4339 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
4340 else if (type
== _KMEM
)
4341 ret
= memcg_update_kmem_limit(memcg
, val
);
4345 case RES_SOFT_LIMIT
:
4346 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4350 * For memsw, soft limits are hard to implement in terms
4351 * of semantics, for now, we support soft limits for
4352 * control without swap
4355 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
4360 ret
= -EINVAL
; /* should be BUG() ? */
4363 return ret
?: nbytes
;
4366 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
4367 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
4369 unsigned long long min_limit
, min_memsw_limit
, tmp
;
4371 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4372 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4373 if (!memcg
->use_hierarchy
)
4376 while (memcg
->css
.parent
) {
4377 memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
4378 if (!memcg
->use_hierarchy
)
4380 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4381 min_limit
= min(min_limit
, tmp
);
4382 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4383 min_memsw_limit
= min(min_memsw_limit
, tmp
);
4386 *mem_limit
= min_limit
;
4387 *memsw_limit
= min_memsw_limit
;
4390 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
4391 size_t nbytes
, loff_t off
)
4393 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4397 type
= MEMFILE_TYPE(of_cft(of
)->private);
4398 name
= MEMFILE_ATTR(of_cft(of
)->private);
4403 res_counter_reset_max(&memcg
->res
);
4404 else if (type
== _MEMSWAP
)
4405 res_counter_reset_max(&memcg
->memsw
);
4406 else if (type
== _KMEM
)
4407 res_counter_reset_max(&memcg
->kmem
);
4413 res_counter_reset_failcnt(&memcg
->res
);
4414 else if (type
== _MEMSWAP
)
4415 res_counter_reset_failcnt(&memcg
->memsw
);
4416 else if (type
== _KMEM
)
4417 res_counter_reset_failcnt(&memcg
->kmem
);
4426 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
4429 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
4433 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4434 struct cftype
*cft
, u64 val
)
4436 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4438 if (val
>= (1 << NR_MOVE_TYPE
))
4442 * No kind of locking is needed in here, because ->can_attach() will
4443 * check this value once in the beginning of the process, and then carry
4444 * on with stale data. This means that changes to this value will only
4445 * affect task migrations starting after the change.
4447 memcg
->move_charge_at_immigrate
= val
;
4451 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4452 struct cftype
*cft
, u64 val
)
4459 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
4463 unsigned int lru_mask
;
4466 static const struct numa_stat stats
[] = {
4467 { "total", LRU_ALL
},
4468 { "file", LRU_ALL_FILE
},
4469 { "anon", LRU_ALL_ANON
},
4470 { "unevictable", BIT(LRU_UNEVICTABLE
) },
4472 const struct numa_stat
*stat
;
4475 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4477 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4478 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
4479 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
4480 for_each_node_state(nid
, N_MEMORY
) {
4481 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4483 seq_printf(m
, " N%d=%lu", nid
, nr
);
4488 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4489 struct mem_cgroup
*iter
;
4492 for_each_mem_cgroup_tree(iter
, memcg
)
4493 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
4494 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
4495 for_each_node_state(nid
, N_MEMORY
) {
4497 for_each_mem_cgroup_tree(iter
, memcg
)
4498 nr
+= mem_cgroup_node_nr_lru_pages(
4499 iter
, nid
, stat
->lru_mask
);
4500 seq_printf(m
, " N%d=%lu", nid
, nr
);
4507 #endif /* CONFIG_NUMA */
4509 static inline void mem_cgroup_lru_names_not_uptodate(void)
4511 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
4514 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4516 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4517 struct mem_cgroup
*mi
;
4520 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
4521 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
4523 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
4524 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
4527 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
4528 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
4529 mem_cgroup_read_events(memcg
, i
));
4531 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4532 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
4533 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
4535 /* Hierarchical information */
4537 unsigned long long limit
, memsw_limit
;
4538 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
4539 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
4540 if (do_swap_account
)
4541 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4545 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
4548 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
4550 for_each_mem_cgroup_tree(mi
, memcg
)
4551 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
4552 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
4555 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
4556 unsigned long long val
= 0;
4558 for_each_mem_cgroup_tree(mi
, memcg
)
4559 val
+= mem_cgroup_read_events(mi
, i
);
4560 seq_printf(m
, "total_%s %llu\n",
4561 mem_cgroup_events_names
[i
], val
);
4564 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
4565 unsigned long long val
= 0;
4567 for_each_mem_cgroup_tree(mi
, memcg
)
4568 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
4569 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
4572 #ifdef CONFIG_DEBUG_VM
4575 struct mem_cgroup_per_zone
*mz
;
4576 struct zone_reclaim_stat
*rstat
;
4577 unsigned long recent_rotated
[2] = {0, 0};
4578 unsigned long recent_scanned
[2] = {0, 0};
4580 for_each_online_node(nid
)
4581 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4582 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
4583 rstat
= &mz
->lruvec
.reclaim_stat
;
4585 recent_rotated
[0] += rstat
->recent_rotated
[0];
4586 recent_rotated
[1] += rstat
->recent_rotated
[1];
4587 recent_scanned
[0] += rstat
->recent_scanned
[0];
4588 recent_scanned
[1] += rstat
->recent_scanned
[1];
4590 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
4591 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
4592 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
4593 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
4600 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4603 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4605 return mem_cgroup_swappiness(memcg
);
4608 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4609 struct cftype
*cft
, u64 val
)
4611 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4617 memcg
->swappiness
= val
;
4619 vm_swappiness
= val
;
4624 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4626 struct mem_cgroup_threshold_ary
*t
;
4632 t
= rcu_dereference(memcg
->thresholds
.primary
);
4634 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4639 usage
= mem_cgroup_usage(memcg
, swap
);
4642 * current_threshold points to threshold just below or equal to usage.
4643 * If it's not true, a threshold was crossed after last
4644 * call of __mem_cgroup_threshold().
4646 i
= t
->current_threshold
;
4649 * Iterate backward over array of thresholds starting from
4650 * current_threshold and check if a threshold is crossed.
4651 * If none of thresholds below usage is crossed, we read
4652 * only one element of the array here.
4654 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4655 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4657 /* i = current_threshold + 1 */
4661 * Iterate forward over array of thresholds starting from
4662 * current_threshold+1 and check if a threshold is crossed.
4663 * If none of thresholds above usage is crossed, we read
4664 * only one element of the array here.
4666 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4667 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4669 /* Update current_threshold */
4670 t
->current_threshold
= i
- 1;
4675 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4678 __mem_cgroup_threshold(memcg
, false);
4679 if (do_swap_account
)
4680 __mem_cgroup_threshold(memcg
, true);
4682 memcg
= parent_mem_cgroup(memcg
);
4686 static int compare_thresholds(const void *a
, const void *b
)
4688 const struct mem_cgroup_threshold
*_a
= a
;
4689 const struct mem_cgroup_threshold
*_b
= b
;
4691 if (_a
->threshold
> _b
->threshold
)
4694 if (_a
->threshold
< _b
->threshold
)
4700 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4702 struct mem_cgroup_eventfd_list
*ev
;
4704 spin_lock(&memcg_oom_lock
);
4706 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4707 eventfd_signal(ev
->eventfd
, 1);
4709 spin_unlock(&memcg_oom_lock
);
4713 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4715 struct mem_cgroup
*iter
;
4717 for_each_mem_cgroup_tree(iter
, memcg
)
4718 mem_cgroup_oom_notify_cb(iter
);
4721 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4722 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4724 struct mem_cgroup_thresholds
*thresholds
;
4725 struct mem_cgroup_threshold_ary
*new;
4726 u64 threshold
, usage
;
4729 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
4733 mutex_lock(&memcg
->thresholds_lock
);
4736 thresholds
= &memcg
->thresholds
;
4737 usage
= mem_cgroup_usage(memcg
, false);
4738 } else if (type
== _MEMSWAP
) {
4739 thresholds
= &memcg
->memsw_thresholds
;
4740 usage
= mem_cgroup_usage(memcg
, true);
4744 /* Check if a threshold crossed before adding a new one */
4745 if (thresholds
->primary
)
4746 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4748 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4750 /* Allocate memory for new array of thresholds */
4751 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
4759 /* Copy thresholds (if any) to new array */
4760 if (thresholds
->primary
) {
4761 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4762 sizeof(struct mem_cgroup_threshold
));
4765 /* Add new threshold */
4766 new->entries
[size
- 1].eventfd
= eventfd
;
4767 new->entries
[size
- 1].threshold
= threshold
;
4769 /* Sort thresholds. Registering of new threshold isn't time-critical */
4770 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4771 compare_thresholds
, NULL
);
4773 /* Find current threshold */
4774 new->current_threshold
= -1;
4775 for (i
= 0; i
< size
; i
++) {
4776 if (new->entries
[i
].threshold
<= usage
) {
4778 * new->current_threshold will not be used until
4779 * rcu_assign_pointer(), so it's safe to increment
4782 ++new->current_threshold
;
4787 /* Free old spare buffer and save old primary buffer as spare */
4788 kfree(thresholds
->spare
);
4789 thresholds
->spare
= thresholds
->primary
;
4791 rcu_assign_pointer(thresholds
->primary
, new);
4793 /* To be sure that nobody uses thresholds */
4797 mutex_unlock(&memcg
->thresholds_lock
);
4802 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4803 struct eventfd_ctx
*eventfd
, const char *args
)
4805 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4808 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4809 struct eventfd_ctx
*eventfd
, const char *args
)
4811 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4814 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4815 struct eventfd_ctx
*eventfd
, enum res_type type
)
4817 struct mem_cgroup_thresholds
*thresholds
;
4818 struct mem_cgroup_threshold_ary
*new;
4822 mutex_lock(&memcg
->thresholds_lock
);
4825 thresholds
= &memcg
->thresholds
;
4826 usage
= mem_cgroup_usage(memcg
, false);
4827 } else if (type
== _MEMSWAP
) {
4828 thresholds
= &memcg
->memsw_thresholds
;
4829 usage
= mem_cgroup_usage(memcg
, true);
4833 if (!thresholds
->primary
)
4836 /* Check if a threshold crossed before removing */
4837 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4839 /* Calculate new number of threshold */
4841 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4842 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4846 new = thresholds
->spare
;
4848 /* Set thresholds array to NULL if we don't have thresholds */
4857 /* Copy thresholds and find current threshold */
4858 new->current_threshold
= -1;
4859 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4860 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4863 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4864 if (new->entries
[j
].threshold
<= usage
) {
4866 * new->current_threshold will not be used
4867 * until rcu_assign_pointer(), so it's safe to increment
4870 ++new->current_threshold
;
4876 /* Swap primary and spare array */
4877 thresholds
->spare
= thresholds
->primary
;
4878 /* If all events are unregistered, free the spare array */
4880 kfree(thresholds
->spare
);
4881 thresholds
->spare
= NULL
;
4884 rcu_assign_pointer(thresholds
->primary
, new);
4886 /* To be sure that nobody uses thresholds */
4889 mutex_unlock(&memcg
->thresholds_lock
);
4892 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4893 struct eventfd_ctx
*eventfd
)
4895 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4898 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4899 struct eventfd_ctx
*eventfd
)
4901 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4904 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4905 struct eventfd_ctx
*eventfd
, const char *args
)
4907 struct mem_cgroup_eventfd_list
*event
;
4909 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4913 spin_lock(&memcg_oom_lock
);
4915 event
->eventfd
= eventfd
;
4916 list_add(&event
->list
, &memcg
->oom_notify
);
4918 /* already in OOM ? */
4919 if (atomic_read(&memcg
->under_oom
))
4920 eventfd_signal(eventfd
, 1);
4921 spin_unlock(&memcg_oom_lock
);
4926 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4927 struct eventfd_ctx
*eventfd
)
4929 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4931 spin_lock(&memcg_oom_lock
);
4933 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4934 if (ev
->eventfd
== eventfd
) {
4935 list_del(&ev
->list
);
4940 spin_unlock(&memcg_oom_lock
);
4943 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4945 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
4947 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4948 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
4952 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4953 struct cftype
*cft
, u64 val
)
4955 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4957 /* cannot set to root cgroup and only 0 and 1 are allowed */
4958 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4961 memcg
->oom_kill_disable
= val
;
4963 memcg_oom_recover(memcg
);
4968 #ifdef CONFIG_MEMCG_KMEM
4969 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4973 memcg
->kmemcg_id
= -1;
4974 ret
= memcg_propagate_kmem(memcg
);
4978 return mem_cgroup_sockets_init(memcg
, ss
);
4981 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4983 mem_cgroup_sockets_destroy(memcg
);
4986 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
4988 if (!memcg_kmem_is_active(memcg
))
4992 * kmem charges can outlive the cgroup. In the case of slab
4993 * pages, for instance, a page contain objects from various
4994 * processes. As we prevent from taking a reference for every
4995 * such allocation we have to be careful when doing uncharge
4996 * (see memcg_uncharge_kmem) and here during offlining.
4998 * The idea is that that only the _last_ uncharge which sees
4999 * the dead memcg will drop the last reference. An additional
5000 * reference is taken here before the group is marked dead
5001 * which is then paired with css_put during uncharge resp. here.
5003 * Although this might sound strange as this path is called from
5004 * css_offline() when the referencemight have dropped down to 0 and
5005 * shouldn't be incremented anymore (css_tryget_online() would
5006 * fail) we do not have other options because of the kmem
5007 * allocations lifetime.
5009 css_get(&memcg
->css
);
5011 memcg_kmem_mark_dead(memcg
);
5013 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5016 if (memcg_kmem_test_and_clear_dead(memcg
))
5017 css_put(&memcg
->css
);
5020 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5025 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5029 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5035 * DO NOT USE IN NEW FILES.
5037 * "cgroup.event_control" implementation.
5039 * This is way over-engineered. It tries to support fully configurable
5040 * events for each user. Such level of flexibility is completely
5041 * unnecessary especially in the light of the planned unified hierarchy.
5043 * Please deprecate this and replace with something simpler if at all
5048 * Unregister event and free resources.
5050 * Gets called from workqueue.
5052 static void memcg_event_remove(struct work_struct
*work
)
5054 struct mem_cgroup_event
*event
=
5055 container_of(work
, struct mem_cgroup_event
, remove
);
5056 struct mem_cgroup
*memcg
= event
->memcg
;
5058 remove_wait_queue(event
->wqh
, &event
->wait
);
5060 event
->unregister_event(memcg
, event
->eventfd
);
5062 /* Notify userspace the event is going away. */
5063 eventfd_signal(event
->eventfd
, 1);
5065 eventfd_ctx_put(event
->eventfd
);
5067 css_put(&memcg
->css
);
5071 * Gets called on POLLHUP on eventfd when user closes it.
5073 * Called with wqh->lock held and interrupts disabled.
5075 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
5076 int sync
, void *key
)
5078 struct mem_cgroup_event
*event
=
5079 container_of(wait
, struct mem_cgroup_event
, wait
);
5080 struct mem_cgroup
*memcg
= event
->memcg
;
5081 unsigned long flags
= (unsigned long)key
;
5083 if (flags
& POLLHUP
) {
5085 * If the event has been detached at cgroup removal, we
5086 * can simply return knowing the other side will cleanup
5089 * We can't race against event freeing since the other
5090 * side will require wqh->lock via remove_wait_queue(),
5093 spin_lock(&memcg
->event_list_lock
);
5094 if (!list_empty(&event
->list
)) {
5095 list_del_init(&event
->list
);
5097 * We are in atomic context, but cgroup_event_remove()
5098 * may sleep, so we have to call it in workqueue.
5100 schedule_work(&event
->remove
);
5102 spin_unlock(&memcg
->event_list_lock
);
5108 static void memcg_event_ptable_queue_proc(struct file
*file
,
5109 wait_queue_head_t
*wqh
, poll_table
*pt
)
5111 struct mem_cgroup_event
*event
=
5112 container_of(pt
, struct mem_cgroup_event
, pt
);
5115 add_wait_queue(wqh
, &event
->wait
);
5119 * DO NOT USE IN NEW FILES.
5121 * Parse input and register new cgroup event handler.
5123 * Input must be in format '<event_fd> <control_fd> <args>'.
5124 * Interpretation of args is defined by control file implementation.
5126 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
5127 char *buf
, size_t nbytes
, loff_t off
)
5129 struct cgroup_subsys_state
*css
= of_css(of
);
5130 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5131 struct mem_cgroup_event
*event
;
5132 struct cgroup_subsys_state
*cfile_css
;
5133 unsigned int efd
, cfd
;
5140 buf
= strstrip(buf
);
5142 efd
= simple_strtoul(buf
, &endp
, 10);
5147 cfd
= simple_strtoul(buf
, &endp
, 10);
5148 if ((*endp
!= ' ') && (*endp
!= '\0'))
5152 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5156 event
->memcg
= memcg
;
5157 INIT_LIST_HEAD(&event
->list
);
5158 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
5159 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
5160 INIT_WORK(&event
->remove
, memcg_event_remove
);
5168 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
5169 if (IS_ERR(event
->eventfd
)) {
5170 ret
= PTR_ERR(event
->eventfd
);
5177 goto out_put_eventfd
;
5180 /* the process need read permission on control file */
5181 /* AV: shouldn't we check that it's been opened for read instead? */
5182 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
5187 * Determine the event callbacks and set them in @event. This used
5188 * to be done via struct cftype but cgroup core no longer knows
5189 * about these events. The following is crude but the whole thing
5190 * is for compatibility anyway.
5192 * DO NOT ADD NEW FILES.
5194 name
= cfile
.file
->f_dentry
->d_name
.name
;
5196 if (!strcmp(name
, "memory.usage_in_bytes")) {
5197 event
->register_event
= mem_cgroup_usage_register_event
;
5198 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
5199 } else if (!strcmp(name
, "memory.oom_control")) {
5200 event
->register_event
= mem_cgroup_oom_register_event
;
5201 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
5202 } else if (!strcmp(name
, "memory.pressure_level")) {
5203 event
->register_event
= vmpressure_register_event
;
5204 event
->unregister_event
= vmpressure_unregister_event
;
5205 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
5206 event
->register_event
= memsw_cgroup_usage_register_event
;
5207 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
5214 * Verify @cfile should belong to @css. Also, remaining events are
5215 * automatically removed on cgroup destruction but the removal is
5216 * asynchronous, so take an extra ref on @css.
5218 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_dentry
->d_parent
,
5219 &memory_cgrp_subsys
);
5221 if (IS_ERR(cfile_css
))
5223 if (cfile_css
!= css
) {
5228 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
5232 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
5234 spin_lock(&memcg
->event_list_lock
);
5235 list_add(&event
->list
, &memcg
->event_list
);
5236 spin_unlock(&memcg
->event_list_lock
);
5248 eventfd_ctx_put(event
->eventfd
);
5257 static struct cftype mem_cgroup_files
[] = {
5259 .name
= "usage_in_bytes",
5260 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5261 .read_u64
= mem_cgroup_read_u64
,
5264 .name
= "max_usage_in_bytes",
5265 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5266 .write
= mem_cgroup_reset
,
5267 .read_u64
= mem_cgroup_read_u64
,
5270 .name
= "limit_in_bytes",
5271 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5272 .write
= mem_cgroup_write
,
5273 .read_u64
= mem_cgroup_read_u64
,
5276 .name
= "soft_limit_in_bytes",
5277 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5278 .write
= mem_cgroup_write
,
5279 .read_u64
= mem_cgroup_read_u64
,
5283 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5284 .write
= mem_cgroup_reset
,
5285 .read_u64
= mem_cgroup_read_u64
,
5289 .seq_show
= memcg_stat_show
,
5292 .name
= "force_empty",
5293 .write
= mem_cgroup_force_empty_write
,
5296 .name
= "use_hierarchy",
5297 .write_u64
= mem_cgroup_hierarchy_write
,
5298 .read_u64
= mem_cgroup_hierarchy_read
,
5301 .name
= "cgroup.event_control", /* XXX: for compat */
5302 .write
= memcg_write_event_control
,
5303 .flags
= CFTYPE_NO_PREFIX
,
5307 .name
= "swappiness",
5308 .read_u64
= mem_cgroup_swappiness_read
,
5309 .write_u64
= mem_cgroup_swappiness_write
,
5312 .name
= "move_charge_at_immigrate",
5313 .read_u64
= mem_cgroup_move_charge_read
,
5314 .write_u64
= mem_cgroup_move_charge_write
,
5317 .name
= "oom_control",
5318 .seq_show
= mem_cgroup_oom_control_read
,
5319 .write_u64
= mem_cgroup_oom_control_write
,
5320 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5323 .name
= "pressure_level",
5327 .name
= "numa_stat",
5328 .seq_show
= memcg_numa_stat_show
,
5331 #ifdef CONFIG_MEMCG_KMEM
5333 .name
= "kmem.limit_in_bytes",
5334 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5335 .write
= mem_cgroup_write
,
5336 .read_u64
= mem_cgroup_read_u64
,
5339 .name
= "kmem.usage_in_bytes",
5340 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5341 .read_u64
= mem_cgroup_read_u64
,
5344 .name
= "kmem.failcnt",
5345 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5346 .write
= mem_cgroup_reset
,
5347 .read_u64
= mem_cgroup_read_u64
,
5350 .name
= "kmem.max_usage_in_bytes",
5351 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5352 .write
= mem_cgroup_reset
,
5353 .read_u64
= mem_cgroup_read_u64
,
5355 #ifdef CONFIG_SLABINFO
5357 .name
= "kmem.slabinfo",
5358 .seq_show
= mem_cgroup_slabinfo_read
,
5362 { }, /* terminate */
5365 #ifdef CONFIG_MEMCG_SWAP
5366 static struct cftype memsw_cgroup_files
[] = {
5368 .name
= "memsw.usage_in_bytes",
5369 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5370 .read_u64
= mem_cgroup_read_u64
,
5373 .name
= "memsw.max_usage_in_bytes",
5374 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5375 .write
= mem_cgroup_reset
,
5376 .read_u64
= mem_cgroup_read_u64
,
5379 .name
= "memsw.limit_in_bytes",
5380 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5381 .write
= mem_cgroup_write
,
5382 .read_u64
= mem_cgroup_read_u64
,
5385 .name
= "memsw.failcnt",
5386 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5387 .write
= mem_cgroup_reset
,
5388 .read_u64
= mem_cgroup_read_u64
,
5390 { }, /* terminate */
5393 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5395 struct mem_cgroup_per_node
*pn
;
5396 struct mem_cgroup_per_zone
*mz
;
5397 int zone
, tmp
= node
;
5399 * This routine is called against possible nodes.
5400 * But it's BUG to call kmalloc() against offline node.
5402 * TODO: this routine can waste much memory for nodes which will
5403 * never be onlined. It's better to use memory hotplug callback
5406 if (!node_state(node
, N_NORMAL_MEMORY
))
5408 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5412 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5413 mz
= &pn
->zoneinfo
[zone
];
5414 lruvec_init(&mz
->lruvec
);
5415 mz
->usage_in_excess
= 0;
5416 mz
->on_tree
= false;
5419 memcg
->nodeinfo
[node
] = pn
;
5423 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5425 kfree(memcg
->nodeinfo
[node
]);
5428 static struct mem_cgroup
*mem_cgroup_alloc(void)
5430 struct mem_cgroup
*memcg
;
5433 size
= sizeof(struct mem_cgroup
);
5434 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5436 memcg
= kzalloc(size
, GFP_KERNEL
);
5440 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5443 spin_lock_init(&memcg
->pcp_counter_lock
);
5452 * At destroying mem_cgroup, references from swap_cgroup can remain.
5453 * (scanning all at force_empty is too costly...)
5455 * Instead of clearing all references at force_empty, we remember
5456 * the number of reference from swap_cgroup and free mem_cgroup when
5457 * it goes down to 0.
5459 * Removal of cgroup itself succeeds regardless of refs from swap.
5462 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5466 mem_cgroup_remove_from_trees(memcg
);
5469 free_mem_cgroup_per_zone_info(memcg
, node
);
5471 free_percpu(memcg
->stat
);
5474 * We need to make sure that (at least for now), the jump label
5475 * destruction code runs outside of the cgroup lock. This is because
5476 * get_online_cpus(), which is called from the static_branch update,
5477 * can't be called inside the cgroup_lock. cpusets are the ones
5478 * enforcing this dependency, so if they ever change, we might as well.
5480 * schedule_work() will guarantee this happens. Be careful if you need
5481 * to move this code around, and make sure it is outside
5484 disarm_static_keys(memcg
);
5489 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5491 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
5493 if (!memcg
->res
.parent
)
5495 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
5497 EXPORT_SYMBOL(parent_mem_cgroup
);
5499 static void __init
mem_cgroup_soft_limit_tree_init(void)
5501 struct mem_cgroup_tree_per_node
*rtpn
;
5502 struct mem_cgroup_tree_per_zone
*rtpz
;
5503 int tmp
, node
, zone
;
5505 for_each_node(node
) {
5507 if (!node_state(node
, N_NORMAL_MEMORY
))
5509 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
5512 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5514 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5515 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5516 rtpz
->rb_root
= RB_ROOT
;
5517 spin_lock_init(&rtpz
->lock
);
5522 static struct cgroup_subsys_state
* __ref
5523 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5525 struct mem_cgroup
*memcg
;
5526 long error
= -ENOMEM
;
5529 memcg
= mem_cgroup_alloc();
5531 return ERR_PTR(error
);
5534 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
5538 if (parent_css
== NULL
) {
5539 root_mem_cgroup
= memcg
;
5540 res_counter_init(&memcg
->res
, NULL
);
5541 res_counter_init(&memcg
->memsw
, NULL
);
5542 res_counter_init(&memcg
->kmem
, NULL
);
5545 memcg
->last_scanned_node
= MAX_NUMNODES
;
5546 INIT_LIST_HEAD(&memcg
->oom_notify
);
5547 memcg
->move_charge_at_immigrate
= 0;
5548 mutex_init(&memcg
->thresholds_lock
);
5549 spin_lock_init(&memcg
->move_lock
);
5550 vmpressure_init(&memcg
->vmpressure
);
5551 INIT_LIST_HEAD(&memcg
->event_list
);
5552 spin_lock_init(&memcg
->event_list_lock
);
5557 __mem_cgroup_free(memcg
);
5558 return ERR_PTR(error
);
5562 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5564 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5565 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
5568 if (css
->id
> MEM_CGROUP_ID_MAX
)
5574 mutex_lock(&memcg_create_mutex
);
5576 memcg
->use_hierarchy
= parent
->use_hierarchy
;
5577 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5578 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5580 if (parent
->use_hierarchy
) {
5581 res_counter_init(&memcg
->res
, &parent
->res
);
5582 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
5583 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
5586 * No need to take a reference to the parent because cgroup
5587 * core guarantees its existence.
5590 res_counter_init(&memcg
->res
, NULL
);
5591 res_counter_init(&memcg
->memsw
, NULL
);
5592 res_counter_init(&memcg
->kmem
, NULL
);
5594 * Deeper hierachy with use_hierarchy == false doesn't make
5595 * much sense so let cgroup subsystem know about this
5596 * unfortunate state in our controller.
5598 if (parent
!= root_mem_cgroup
)
5599 memory_cgrp_subsys
.broken_hierarchy
= true;
5601 mutex_unlock(&memcg_create_mutex
);
5603 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
5608 * Make sure the memcg is initialized: mem_cgroup_iter()
5609 * orders reading memcg->initialized against its callers
5610 * reading the memcg members.
5612 smp_store_release(&memcg
->initialized
, 1);
5618 * Announce all parents that a group from their hierarchy is gone.
5620 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
5622 struct mem_cgroup
*parent
= memcg
;
5624 while ((parent
= parent_mem_cgroup(parent
)))
5625 mem_cgroup_iter_invalidate(parent
);
5628 * if the root memcg is not hierarchical we have to check it
5631 if (!root_mem_cgroup
->use_hierarchy
)
5632 mem_cgroup_iter_invalidate(root_mem_cgroup
);
5635 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5637 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5638 struct mem_cgroup_event
*event
, *tmp
;
5639 struct cgroup_subsys_state
*iter
;
5642 * Unregister events and notify userspace.
5643 * Notify userspace about cgroup removing only after rmdir of cgroup
5644 * directory to avoid race between userspace and kernelspace.
5646 spin_lock(&memcg
->event_list_lock
);
5647 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5648 list_del_init(&event
->list
);
5649 schedule_work(&event
->remove
);
5651 spin_unlock(&memcg
->event_list_lock
);
5653 kmem_cgroup_css_offline(memcg
);
5655 mem_cgroup_invalidate_reclaim_iterators(memcg
);
5658 * This requires that offlining is serialized. Right now that is
5659 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5661 css_for_each_descendant_post(iter
, css
)
5662 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter
));
5664 memcg_unregister_all_caches(memcg
);
5665 vmpressure_cleanup(&memcg
->vmpressure
);
5668 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5670 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5672 * XXX: css_offline() would be where we should reparent all
5673 * memory to prepare the cgroup for destruction. However,
5674 * memcg does not do css_tryget_online() and res_counter charging
5675 * under the same RCU lock region, which means that charging
5676 * could race with offlining. Offlining only happens to
5677 * cgroups with no tasks in them but charges can show up
5678 * without any tasks from the swapin path when the target
5679 * memcg is looked up from the swapout record and not from the
5680 * current task as it usually is. A race like this can leak
5681 * charges and put pages with stale cgroup pointers into
5685 * lookup_swap_cgroup_id()
5687 * mem_cgroup_lookup()
5688 * css_tryget_online()
5690 * disable css_tryget_online()
5693 * reparent_charges()
5694 * res_counter_charge()
5697 * pc->mem_cgroup = dead memcg
5700 * The bulk of the charges are still moved in offline_css() to
5701 * avoid pinning a lot of pages in case a long-term reference
5702 * like a swapout record is deferring the css_free() to long
5703 * after offlining. But this makes sure we catch any charges
5704 * made after offlining:
5706 mem_cgroup_reparent_charges(memcg
);
5708 memcg_destroy_kmem(memcg
);
5709 __mem_cgroup_free(memcg
);
5713 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5714 * @css: the target css
5716 * Reset the states of the mem_cgroup associated with @css. This is
5717 * invoked when the userland requests disabling on the default hierarchy
5718 * but the memcg is pinned through dependency. The memcg should stop
5719 * applying policies and should revert to the vanilla state as it may be
5720 * made visible again.
5722 * The current implementation only resets the essential configurations.
5723 * This needs to be expanded to cover all the visible parts.
5725 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5727 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5729 mem_cgroup_resize_limit(memcg
, ULLONG_MAX
);
5730 mem_cgroup_resize_memsw_limit(memcg
, ULLONG_MAX
);
5731 memcg_update_kmem_limit(memcg
, ULLONG_MAX
);
5732 res_counter_set_soft_limit(&memcg
->res
, ULLONG_MAX
);
5736 /* Handlers for move charge at task migration. */
5737 static int mem_cgroup_do_precharge(unsigned long count
)
5741 /* Try a single bulk charge without reclaim first */
5742 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
5744 mc
.precharge
+= count
;
5747 if (ret
== -EINTR
) {
5748 cancel_charge(root_mem_cgroup
, count
);
5752 /* Try charges one by one with reclaim */
5754 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
5756 * In case of failure, any residual charges against
5757 * mc.to will be dropped by mem_cgroup_clear_mc()
5758 * later on. However, cancel any charges that are
5759 * bypassed to root right away or they'll be lost.
5762 cancel_charge(root_mem_cgroup
, 1);
5772 * get_mctgt_type - get target type of moving charge
5773 * @vma: the vma the pte to be checked belongs
5774 * @addr: the address corresponding to the pte to be checked
5775 * @ptent: the pte to be checked
5776 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5779 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5780 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5781 * move charge. if @target is not NULL, the page is stored in target->page
5782 * with extra refcnt got(Callers should handle it).
5783 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5784 * target for charge migration. if @target is not NULL, the entry is stored
5787 * Called with pte lock held.
5794 enum mc_target_type
{
5800 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5801 unsigned long addr
, pte_t ptent
)
5803 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5805 if (!page
|| !page_mapped(page
))
5807 if (PageAnon(page
)) {
5808 /* we don't move shared anon */
5811 } else if (!move_file())
5812 /* we ignore mapcount for file pages */
5814 if (!get_page_unless_zero(page
))
5821 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5822 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5824 struct page
*page
= NULL
;
5825 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5827 if (!move_anon() || non_swap_entry(ent
))
5830 * Because lookup_swap_cache() updates some statistics counter,
5831 * we call find_get_page() with swapper_space directly.
5833 page
= find_get_page(swap_address_space(ent
), ent
.val
);
5834 if (do_swap_account
)
5835 entry
->val
= ent
.val
;
5840 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5841 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5847 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5848 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5850 struct page
*page
= NULL
;
5851 struct address_space
*mapping
;
5854 if (!vma
->vm_file
) /* anonymous vma */
5859 mapping
= vma
->vm_file
->f_mapping
;
5860 if (pte_none(ptent
))
5861 pgoff
= linear_page_index(vma
, addr
);
5862 else /* pte_file(ptent) is true */
5863 pgoff
= pte_to_pgoff(ptent
);
5865 /* page is moved even if it's not RSS of this task(page-faulted). */
5867 /* shmem/tmpfs may report page out on swap: account for that too. */
5868 if (shmem_mapping(mapping
)) {
5869 page
= find_get_entry(mapping
, pgoff
);
5870 if (radix_tree_exceptional_entry(page
)) {
5871 swp_entry_t swp
= radix_to_swp_entry(page
);
5872 if (do_swap_account
)
5874 page
= find_get_page(swap_address_space(swp
), swp
.val
);
5877 page
= find_get_page(mapping
, pgoff
);
5879 page
= find_get_page(mapping
, pgoff
);
5884 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5885 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5887 struct page
*page
= NULL
;
5888 struct page_cgroup
*pc
;
5889 enum mc_target_type ret
= MC_TARGET_NONE
;
5890 swp_entry_t ent
= { .val
= 0 };
5892 if (pte_present(ptent
))
5893 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5894 else if (is_swap_pte(ptent
))
5895 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
5896 else if (pte_none(ptent
) || pte_file(ptent
))
5897 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5899 if (!page
&& !ent
.val
)
5902 pc
= lookup_page_cgroup(page
);
5904 * Do only loose check w/o serialization.
5905 * mem_cgroup_move_account() checks the pc is valid or
5906 * not under LRU exclusion.
5908 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
5909 ret
= MC_TARGET_PAGE
;
5911 target
->page
= page
;
5913 if (!ret
|| !target
)
5916 /* There is a swap entry and a page doesn't exist or isn't charged */
5917 if (ent
.val
&& !ret
&&
5918 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5919 ret
= MC_TARGET_SWAP
;
5926 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5928 * We don't consider swapping or file mapped pages because THP does not
5929 * support them for now.
5930 * Caller should make sure that pmd_trans_huge(pmd) is true.
5932 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5933 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5935 struct page
*page
= NULL
;
5936 struct page_cgroup
*pc
;
5937 enum mc_target_type ret
= MC_TARGET_NONE
;
5939 page
= pmd_page(pmd
);
5940 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5943 pc
= lookup_page_cgroup(page
);
5944 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
5945 ret
= MC_TARGET_PAGE
;
5948 target
->page
= page
;
5954 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5955 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5957 return MC_TARGET_NONE
;
5961 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5962 unsigned long addr
, unsigned long end
,
5963 struct mm_walk
*walk
)
5965 struct vm_area_struct
*vma
= walk
->private;
5969 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5970 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5971 mc
.precharge
+= HPAGE_PMD_NR
;
5976 if (pmd_trans_unstable(pmd
))
5978 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5979 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5980 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5981 mc
.precharge
++; /* increment precharge temporarily */
5982 pte_unmap_unlock(pte
- 1, ptl
);
5988 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5990 unsigned long precharge
;
5991 struct vm_area_struct
*vma
;
5993 down_read(&mm
->mmap_sem
);
5994 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5995 struct mm_walk mem_cgroup_count_precharge_walk
= {
5996 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6000 if (is_vm_hugetlb_page(vma
))
6002 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6003 &mem_cgroup_count_precharge_walk
);
6005 up_read(&mm
->mmap_sem
);
6007 precharge
= mc
.precharge
;
6013 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6015 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6017 VM_BUG_ON(mc
.moving_task
);
6018 mc
.moving_task
= current
;
6019 return mem_cgroup_do_precharge(precharge
);
6022 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6023 static void __mem_cgroup_clear_mc(void)
6025 struct mem_cgroup
*from
= mc
.from
;
6026 struct mem_cgroup
*to
= mc
.to
;
6029 /* we must uncharge all the leftover precharges from mc.to */
6031 cancel_charge(mc
.to
, mc
.precharge
);
6035 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6036 * we must uncharge here.
6038 if (mc
.moved_charge
) {
6039 cancel_charge(mc
.from
, mc
.moved_charge
);
6040 mc
.moved_charge
= 0;
6042 /* we must fixup refcnts and charges */
6043 if (mc
.moved_swap
) {
6044 /* uncharge swap account from the old cgroup */
6045 if (!mem_cgroup_is_root(mc
.from
))
6046 res_counter_uncharge(&mc
.from
->memsw
,
6047 PAGE_SIZE
* mc
.moved_swap
);
6049 for (i
= 0; i
< mc
.moved_swap
; i
++)
6050 css_put(&mc
.from
->css
);
6053 * we charged both to->res and to->memsw, so we should
6056 if (!mem_cgroup_is_root(mc
.to
))
6057 res_counter_uncharge(&mc
.to
->res
,
6058 PAGE_SIZE
* mc
.moved_swap
);
6059 /* we've already done css_get(mc.to) */
6062 memcg_oom_recover(from
);
6063 memcg_oom_recover(to
);
6064 wake_up_all(&mc
.waitq
);
6067 static void mem_cgroup_clear_mc(void)
6069 struct mem_cgroup
*from
= mc
.from
;
6072 * we must clear moving_task before waking up waiters at the end of
6075 mc
.moving_task
= NULL
;
6076 __mem_cgroup_clear_mc();
6077 spin_lock(&mc
.lock
);
6080 spin_unlock(&mc
.lock
);
6081 mem_cgroup_end_move(from
);
6084 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6085 struct cgroup_taskset
*tset
)
6087 struct task_struct
*p
= cgroup_taskset_first(tset
);
6089 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6090 unsigned long move_charge_at_immigrate
;
6093 * We are now commited to this value whatever it is. Changes in this
6094 * tunable will only affect upcoming migrations, not the current one.
6095 * So we need to save it, and keep it going.
6097 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6098 if (move_charge_at_immigrate
) {
6099 struct mm_struct
*mm
;
6100 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6102 VM_BUG_ON(from
== memcg
);
6104 mm
= get_task_mm(p
);
6107 /* We move charges only when we move a owner of the mm */
6108 if (mm
->owner
== p
) {
6111 VM_BUG_ON(mc
.precharge
);
6112 VM_BUG_ON(mc
.moved_charge
);
6113 VM_BUG_ON(mc
.moved_swap
);
6114 mem_cgroup_start_move(from
);
6115 spin_lock(&mc
.lock
);
6118 mc
.immigrate_flags
= move_charge_at_immigrate
;
6119 spin_unlock(&mc
.lock
);
6120 /* We set mc.moving_task later */
6122 ret
= mem_cgroup_precharge_mc(mm
);
6124 mem_cgroup_clear_mc();
6131 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6132 struct cgroup_taskset
*tset
)
6134 mem_cgroup_clear_mc();
6137 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6138 unsigned long addr
, unsigned long end
,
6139 struct mm_walk
*walk
)
6142 struct vm_area_struct
*vma
= walk
->private;
6145 enum mc_target_type target_type
;
6146 union mc_target target
;
6148 struct page_cgroup
*pc
;
6151 * We don't take compound_lock() here but no race with splitting thp
6153 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6154 * under splitting, which means there's no concurrent thp split,
6155 * - if another thread runs into split_huge_page() just after we
6156 * entered this if-block, the thread must wait for page table lock
6157 * to be unlocked in __split_huge_page_splitting(), where the main
6158 * part of thp split is not executed yet.
6160 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6161 if (mc
.precharge
< HPAGE_PMD_NR
) {
6165 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6166 if (target_type
== MC_TARGET_PAGE
) {
6168 if (!isolate_lru_page(page
)) {
6169 pc
= lookup_page_cgroup(page
);
6170 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6171 pc
, mc
.from
, mc
.to
)) {
6172 mc
.precharge
-= HPAGE_PMD_NR
;
6173 mc
.moved_charge
+= HPAGE_PMD_NR
;
6175 putback_lru_page(page
);
6183 if (pmd_trans_unstable(pmd
))
6186 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6187 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6188 pte_t ptent
= *(pte
++);
6194 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6195 case MC_TARGET_PAGE
:
6197 if (isolate_lru_page(page
))
6199 pc
= lookup_page_cgroup(page
);
6200 if (!mem_cgroup_move_account(page
, 1, pc
,
6203 /* we uncharge from mc.from later. */
6206 putback_lru_page(page
);
6207 put
: /* get_mctgt_type() gets the page */
6210 case MC_TARGET_SWAP
:
6212 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6214 /* we fixup refcnts and charges later. */
6222 pte_unmap_unlock(pte
- 1, ptl
);
6227 * We have consumed all precharges we got in can_attach().
6228 * We try charge one by one, but don't do any additional
6229 * charges to mc.to if we have failed in charge once in attach()
6232 ret
= mem_cgroup_do_precharge(1);
6240 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6242 struct vm_area_struct
*vma
;
6244 lru_add_drain_all();
6246 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6248 * Someone who are holding the mmap_sem might be waiting in
6249 * waitq. So we cancel all extra charges, wake up all waiters,
6250 * and retry. Because we cancel precharges, we might not be able
6251 * to move enough charges, but moving charge is a best-effort
6252 * feature anyway, so it wouldn't be a big problem.
6254 __mem_cgroup_clear_mc();
6258 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6260 struct mm_walk mem_cgroup_move_charge_walk
= {
6261 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6265 if (is_vm_hugetlb_page(vma
))
6267 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6268 &mem_cgroup_move_charge_walk
);
6271 * means we have consumed all precharges and failed in
6272 * doing additional charge. Just abandon here.
6276 up_read(&mm
->mmap_sem
);
6279 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6280 struct cgroup_taskset
*tset
)
6282 struct task_struct
*p
= cgroup_taskset_first(tset
);
6283 struct mm_struct
*mm
= get_task_mm(p
);
6287 mem_cgroup_move_charge(mm
);
6291 mem_cgroup_clear_mc();
6293 #else /* !CONFIG_MMU */
6294 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6295 struct cgroup_taskset
*tset
)
6299 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6300 struct cgroup_taskset
*tset
)
6303 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6304 struct cgroup_taskset
*tset
)
6310 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6311 * to verify whether we're attached to the default hierarchy on each mount
6314 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6317 * use_hierarchy is forced on the default hierarchy. cgroup core
6318 * guarantees that @root doesn't have any children, so turning it
6319 * on for the root memcg is enough.
6321 if (cgroup_on_dfl(root_css
->cgroup
))
6322 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6325 struct cgroup_subsys memory_cgrp_subsys
= {
6326 .css_alloc
= mem_cgroup_css_alloc
,
6327 .css_online
= mem_cgroup_css_online
,
6328 .css_offline
= mem_cgroup_css_offline
,
6329 .css_free
= mem_cgroup_css_free
,
6330 .css_reset
= mem_cgroup_css_reset
,
6331 .can_attach
= mem_cgroup_can_attach
,
6332 .cancel_attach
= mem_cgroup_cancel_attach
,
6333 .attach
= mem_cgroup_move_task
,
6334 .bind
= mem_cgroup_bind
,
6335 .legacy_cftypes
= mem_cgroup_files
,
6339 #ifdef CONFIG_MEMCG_SWAP
6340 static int __init
enable_swap_account(char *s
)
6342 if (!strcmp(s
, "1"))
6343 really_do_swap_account
= 1;
6344 else if (!strcmp(s
, "0"))
6345 really_do_swap_account
= 0;
6348 __setup("swapaccount=", enable_swap_account
);
6350 static void __init
memsw_file_init(void)
6352 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6353 memsw_cgroup_files
));
6356 static void __init
enable_swap_cgroup(void)
6358 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6359 do_swap_account
= 1;
6365 static void __init
enable_swap_cgroup(void)
6370 #ifdef CONFIG_MEMCG_SWAP
6372 * mem_cgroup_swapout - transfer a memsw charge to swap
6373 * @page: page whose memsw charge to transfer
6374 * @entry: swap entry to move the charge to
6376 * Transfer the memsw charge of @page to @entry.
6378 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6380 struct page_cgroup
*pc
;
6381 unsigned short oldid
;
6383 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6384 VM_BUG_ON_PAGE(page_count(page
), page
);
6386 if (!do_swap_account
)
6389 pc
= lookup_page_cgroup(page
);
6391 /* Readahead page, never charged */
6392 if (!PageCgroupUsed(pc
))
6395 VM_BUG_ON_PAGE(!(pc
->flags
& PCG_MEMSW
), page
);
6397 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(pc
->mem_cgroup
));
6398 VM_BUG_ON_PAGE(oldid
, page
);
6400 pc
->flags
&= ~PCG_MEMSW
;
6401 css_get(&pc
->mem_cgroup
->css
);
6402 mem_cgroup_swap_statistics(pc
->mem_cgroup
, true);
6406 * mem_cgroup_uncharge_swap - uncharge a swap entry
6407 * @entry: swap entry to uncharge
6409 * Drop the memsw charge associated with @entry.
6411 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
6413 struct mem_cgroup
*memcg
;
6416 if (!do_swap_account
)
6419 id
= swap_cgroup_record(entry
, 0);
6421 memcg
= mem_cgroup_lookup(id
);
6423 if (!mem_cgroup_is_root(memcg
))
6424 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
6425 mem_cgroup_swap_statistics(memcg
, false);
6426 css_put(&memcg
->css
);
6433 * mem_cgroup_try_charge - try charging a page
6434 * @page: page to charge
6435 * @mm: mm context of the victim
6436 * @gfp_mask: reclaim mode
6437 * @memcgp: charged memcg return
6439 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6440 * pages according to @gfp_mask if necessary.
6442 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6443 * Otherwise, an error code is returned.
6445 * After page->mapping has been set up, the caller must finalize the
6446 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6447 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6449 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
6450 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
6452 struct mem_cgroup
*memcg
= NULL
;
6453 unsigned int nr_pages
= 1;
6456 if (mem_cgroup_disabled())
6459 if (PageSwapCache(page
)) {
6460 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
6462 * Every swap fault against a single page tries to charge the
6463 * page, bail as early as possible. shmem_unuse() encounters
6464 * already charged pages, too. The USED bit is protected by
6465 * the page lock, which serializes swap cache removal, which
6466 * in turn serializes uncharging.
6468 if (PageCgroupUsed(pc
))
6472 if (PageTransHuge(page
)) {
6473 nr_pages
<<= compound_order(page
);
6474 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6477 if (do_swap_account
&& PageSwapCache(page
))
6478 memcg
= try_get_mem_cgroup_from_page(page
);
6480 memcg
= get_mem_cgroup_from_mm(mm
);
6482 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6484 css_put(&memcg
->css
);
6486 if (ret
== -EINTR
) {
6487 memcg
= root_mem_cgroup
;
6496 * mem_cgroup_commit_charge - commit a page charge
6497 * @page: page to charge
6498 * @memcg: memcg to charge the page to
6499 * @lrucare: page might be on LRU already
6501 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6502 * after page->mapping has been set up. This must happen atomically
6503 * as part of the page instantiation, i.e. under the page table lock
6504 * for anonymous pages, under the page lock for page and swap cache.
6506 * In addition, the page must not be on the LRU during the commit, to
6507 * prevent racing with task migration. If it might be, use @lrucare.
6509 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6511 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6514 unsigned int nr_pages
= 1;
6516 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6517 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6519 if (mem_cgroup_disabled())
6522 * Swap faults will attempt to charge the same page multiple
6523 * times. But reuse_swap_page() might have removed the page
6524 * from swapcache already, so we can't check PageSwapCache().
6529 commit_charge(page
, memcg
, lrucare
);
6531 if (PageTransHuge(page
)) {
6532 nr_pages
<<= compound_order(page
);
6533 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6536 local_irq_disable();
6537 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6538 memcg_check_events(memcg
, page
);
6541 if (do_swap_account
&& PageSwapCache(page
)) {
6542 swp_entry_t entry
= { .val
= page_private(page
) };
6544 * The swap entry might not get freed for a long time,
6545 * let's not wait for it. The page already received a
6546 * memory+swap charge, drop the swap entry duplicate.
6548 mem_cgroup_uncharge_swap(entry
);
6553 * mem_cgroup_cancel_charge - cancel a page charge
6554 * @page: page to charge
6555 * @memcg: memcg to charge the page to
6557 * Cancel a charge transaction started by mem_cgroup_try_charge().
6559 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
6561 unsigned int nr_pages
= 1;
6563 if (mem_cgroup_disabled())
6566 * Swap faults will attempt to charge the same page multiple
6567 * times. But reuse_swap_page() might have removed the page
6568 * from swapcache already, so we can't check PageSwapCache().
6573 if (PageTransHuge(page
)) {
6574 nr_pages
<<= compound_order(page
);
6575 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6578 cancel_charge(memcg
, nr_pages
);
6581 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
6582 unsigned long nr_mem
, unsigned long nr_memsw
,
6583 unsigned long nr_anon
, unsigned long nr_file
,
6584 unsigned long nr_huge
, struct page
*dummy_page
)
6586 unsigned long flags
;
6588 if (!mem_cgroup_is_root(memcg
)) {
6590 res_counter_uncharge(&memcg
->res
,
6591 nr_mem
* PAGE_SIZE
);
6593 res_counter_uncharge(&memcg
->memsw
,
6594 nr_memsw
* PAGE_SIZE
);
6595 memcg_oom_recover(memcg
);
6598 local_irq_save(flags
);
6599 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
6600 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
6601 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
6602 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
6603 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_anon
+ nr_file
);
6604 memcg_check_events(memcg
, dummy_page
);
6605 local_irq_restore(flags
);
6608 static void uncharge_list(struct list_head
*page_list
)
6610 struct mem_cgroup
*memcg
= NULL
;
6611 unsigned long nr_memsw
= 0;
6612 unsigned long nr_anon
= 0;
6613 unsigned long nr_file
= 0;
6614 unsigned long nr_huge
= 0;
6615 unsigned long pgpgout
= 0;
6616 unsigned long nr_mem
= 0;
6617 struct list_head
*next
;
6620 next
= page_list
->next
;
6622 unsigned int nr_pages
= 1;
6623 struct page_cgroup
*pc
;
6625 page
= list_entry(next
, struct page
, lru
);
6626 next
= page
->lru
.next
;
6628 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6629 VM_BUG_ON_PAGE(page_count(page
), page
);
6631 pc
= lookup_page_cgroup(page
);
6632 if (!PageCgroupUsed(pc
))
6636 * Nobody should be changing or seriously looking at
6637 * pc->mem_cgroup and pc->flags at this point, we have
6638 * fully exclusive access to the page.
6641 if (memcg
!= pc
->mem_cgroup
) {
6643 uncharge_batch(memcg
, pgpgout
, nr_mem
, nr_memsw
,
6644 nr_anon
, nr_file
, nr_huge
, page
);
6645 pgpgout
= nr_mem
= nr_memsw
= 0;
6646 nr_anon
= nr_file
= nr_huge
= 0;
6648 memcg
= pc
->mem_cgroup
;
6651 if (PageTransHuge(page
)) {
6652 nr_pages
<<= compound_order(page
);
6653 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6654 nr_huge
+= nr_pages
;
6658 nr_anon
+= nr_pages
;
6660 nr_file
+= nr_pages
;
6662 if (pc
->flags
& PCG_MEM
)
6664 if (pc
->flags
& PCG_MEMSW
)
6665 nr_memsw
+= nr_pages
;
6669 } while (next
!= page_list
);
6672 uncharge_batch(memcg
, pgpgout
, nr_mem
, nr_memsw
,
6673 nr_anon
, nr_file
, nr_huge
, page
);
6677 * mem_cgroup_uncharge - uncharge a page
6678 * @page: page to uncharge
6680 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6681 * mem_cgroup_commit_charge().
6683 void mem_cgroup_uncharge(struct page
*page
)
6685 struct page_cgroup
*pc
;
6687 if (mem_cgroup_disabled())
6690 /* Don't touch page->lru of any random page, pre-check: */
6691 pc
= lookup_page_cgroup(page
);
6692 if (!PageCgroupUsed(pc
))
6695 INIT_LIST_HEAD(&page
->lru
);
6696 uncharge_list(&page
->lru
);
6700 * mem_cgroup_uncharge_list - uncharge a list of page
6701 * @page_list: list of pages to uncharge
6703 * Uncharge a list of pages previously charged with
6704 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6706 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6708 if (mem_cgroup_disabled())
6711 if (!list_empty(page_list
))
6712 uncharge_list(page_list
);
6716 * mem_cgroup_migrate - migrate a charge to another page
6717 * @oldpage: currently charged page
6718 * @newpage: page to transfer the charge to
6719 * @lrucare: both pages might be on the LRU already
6721 * Migrate the charge from @oldpage to @newpage.
6723 * Both pages must be locked, @newpage->mapping must be set up.
6725 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
6728 struct page_cgroup
*pc
;
6731 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6732 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6733 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
6734 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
6735 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6736 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6739 if (mem_cgroup_disabled())
6742 /* Page cache replacement: new page already charged? */
6743 pc
= lookup_page_cgroup(newpage
);
6744 if (PageCgroupUsed(pc
))
6747 /* Re-entrant migration: old page already uncharged? */
6748 pc
= lookup_page_cgroup(oldpage
);
6749 if (!PageCgroupUsed(pc
))
6752 VM_BUG_ON_PAGE(!(pc
->flags
& PCG_MEM
), oldpage
);
6753 VM_BUG_ON_PAGE(do_swap_account
&& !(pc
->flags
& PCG_MEMSW
), oldpage
);
6756 lock_page_lru(oldpage
, &isolated
);
6761 unlock_page_lru(oldpage
, isolated
);
6763 commit_charge(newpage
, pc
->mem_cgroup
, lrucare
);
6767 * subsys_initcall() for memory controller.
6769 * Some parts like hotcpu_notifier() have to be initialized from this context
6770 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6771 * everything that doesn't depend on a specific mem_cgroup structure should
6772 * be initialized from here.
6774 static int __init
mem_cgroup_init(void)
6776 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6777 enable_swap_cgroup();
6778 mem_cgroup_soft_limit_tree_init();
6782 subsys_initcall(mem_cgroup_init
);