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/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
66 EXPORT_SYMBOL(mem_cgroup_subsys
);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly
;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata
= 1;
79 static int really_do_swap_account __initdata
= 0;
83 #define do_swap_account 0
88 * Statistics for memory cgroup.
90 enum mem_cgroup_stat_index
{
92 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
95 MEM_CGROUP_STAT_RSS
, /* # of pages charged as anon rss */
96 MEM_CGROUP_STAT_FILE_MAPPED
, /* # of pages charged as file rss */
97 MEM_CGROUP_STAT_SWAP
, /* # of pages, swapped out */
98 MEM_CGROUP_STAT_NSTATS
,
101 static const char * const mem_cgroup_stat_names
[] = {
108 enum mem_cgroup_events_index
{
109 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
110 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
111 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
112 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
113 MEM_CGROUP_EVENTS_NSTATS
,
116 static const char * const mem_cgroup_events_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
{
147 /* css_id of the last scanned hierarchy member */
149 /* scan generation, increased every round-trip */
150 unsigned int generation
;
154 * per-zone information in memory controller.
156 struct mem_cgroup_per_zone
{
157 struct lruvec lruvec
;
158 unsigned long lru_size
[NR_LRU_LISTS
];
160 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
162 struct rb_node tree_node
; /* RB tree node */
163 unsigned long long usage_in_excess
;/* Set to the value by which */
164 /* the soft limit is exceeded*/
166 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
167 /* use container_of */
170 struct mem_cgroup_per_node
{
171 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
174 struct mem_cgroup_lru_info
{
175 struct mem_cgroup_per_node
*nodeinfo
[MAX_NUMNODES
];
179 * Cgroups above their limits are maintained in a RB-Tree, independent of
180 * their hierarchy representation
183 struct mem_cgroup_tree_per_zone
{
184 struct rb_root rb_root
;
188 struct mem_cgroup_tree_per_node
{
189 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
192 struct mem_cgroup_tree
{
193 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
196 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
198 struct mem_cgroup_threshold
{
199 struct eventfd_ctx
*eventfd
;
204 struct mem_cgroup_threshold_ary
{
205 /* An array index points to threshold just below or equal to usage. */
206 int current_threshold
;
207 /* Size of entries[] */
209 /* Array of thresholds */
210 struct mem_cgroup_threshold entries
[0];
213 struct mem_cgroup_thresholds
{
214 /* Primary thresholds array */
215 struct mem_cgroup_threshold_ary
*primary
;
217 * Spare threshold array.
218 * This is needed to make mem_cgroup_unregister_event() "never fail".
219 * It must be able to store at least primary->size - 1 entries.
221 struct mem_cgroup_threshold_ary
*spare
;
225 struct mem_cgroup_eventfd_list
{
226 struct list_head list
;
227 struct eventfd_ctx
*eventfd
;
230 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
231 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
234 * The memory controller data structure. The memory controller controls both
235 * page cache and RSS per cgroup. We would eventually like to provide
236 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
237 * to help the administrator determine what knobs to tune.
239 * TODO: Add a water mark for the memory controller. Reclaim will begin when
240 * we hit the water mark. May be even add a low water mark, such that
241 * no reclaim occurs from a cgroup at it's low water mark, this is
242 * a feature that will be implemented much later in the future.
245 struct cgroup_subsys_state css
;
247 * the counter to account for memory usage
249 struct res_counter res
;
253 * the counter to account for mem+swap usage.
255 struct res_counter memsw
;
258 * rcu_freeing is used only when freeing struct mem_cgroup,
259 * so put it into a union to avoid wasting more memory.
260 * It must be disjoint from the css field. It could be
261 * in a union with the res field, but res plays a much
262 * larger part in mem_cgroup life than memsw, and might
263 * be of interest, even at time of free, when debugging.
264 * So share rcu_head with the less interesting memsw.
266 struct rcu_head rcu_freeing
;
268 * We also need some space for a worker in deferred freeing.
269 * By the time we call it, rcu_freeing is no longer in use.
271 struct work_struct work_freeing
;
275 * the counter to account for kernel memory usage.
277 struct res_counter kmem
;
279 * Per cgroup active and inactive list, similar to the
280 * per zone LRU lists.
282 struct mem_cgroup_lru_info info
;
283 int last_scanned_node
;
285 nodemask_t scan_nodes
;
286 atomic_t numainfo_events
;
287 atomic_t numainfo_updating
;
290 * Should the accounting and control be hierarchical, per subtree?
293 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
301 /* OOM-Killer disable */
302 int oom_kill_disable
;
304 /* set when res.limit == memsw.limit */
305 bool memsw_is_minimum
;
307 /* protect arrays of thresholds */
308 struct mutex thresholds_lock
;
310 /* thresholds for memory usage. RCU-protected */
311 struct mem_cgroup_thresholds thresholds
;
313 /* thresholds for mem+swap usage. RCU-protected */
314 struct mem_cgroup_thresholds memsw_thresholds
;
316 /* For oom notifier event fd */
317 struct list_head oom_notify
;
320 * Should we move charges of a task when a task is moved into this
321 * mem_cgroup ? And what type of charges should we move ?
323 unsigned long move_charge_at_immigrate
;
325 * set > 0 if pages under this cgroup are moving to other cgroup.
327 atomic_t moving_account
;
328 /* taken only while moving_account > 0 */
329 spinlock_t move_lock
;
333 struct mem_cgroup_stat_cpu __percpu
*stat
;
335 * used when a cpu is offlined or other synchronizations
336 * See mem_cgroup_read_stat().
338 struct mem_cgroup_stat_cpu nocpu_base
;
339 spinlock_t pcp_counter_lock
;
341 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
342 struct tcp_memcontrol tcp_mem
;
344 #if defined(CONFIG_MEMCG_KMEM)
345 /* analogous to slab_common's slab_caches list. per-memcg */
346 struct list_head memcg_slab_caches
;
347 /* Not a spinlock, we can take a lot of time walking the list */
348 struct mutex slab_caches_mutex
;
349 /* Index in the kmem_cache->memcg_params->memcg_caches array */
354 /* internal only representation about the status of kmem accounting. */
356 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
357 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
358 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
361 /* We account when limit is on, but only after call sites are patched */
362 #define KMEM_ACCOUNTED_MASK \
363 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
365 #ifdef CONFIG_MEMCG_KMEM
366 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
368 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
371 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
373 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
376 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
378 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
381 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
383 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
386 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
388 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
389 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
392 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
394 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
395 &memcg
->kmem_account_flags
);
399 /* Stuffs for move charges at task migration. */
401 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
402 * left-shifted bitmap of these types.
405 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
406 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
410 /* "mc" and its members are protected by cgroup_mutex */
411 static struct move_charge_struct
{
412 spinlock_t lock
; /* for from, to */
413 struct mem_cgroup
*from
;
414 struct mem_cgroup
*to
;
415 unsigned long precharge
;
416 unsigned long moved_charge
;
417 unsigned long moved_swap
;
418 struct task_struct
*moving_task
; /* a task moving charges */
419 wait_queue_head_t waitq
; /* a waitq for other context */
421 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
422 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
425 static bool move_anon(void)
427 return test_bit(MOVE_CHARGE_TYPE_ANON
,
428 &mc
.to
->move_charge_at_immigrate
);
431 static bool move_file(void)
433 return test_bit(MOVE_CHARGE_TYPE_FILE
,
434 &mc
.to
->move_charge_at_immigrate
);
438 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
439 * limit reclaim to prevent infinite loops, if they ever occur.
441 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
442 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
445 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
446 MEM_CGROUP_CHARGE_TYPE_ANON
,
447 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
448 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
452 /* for encoding cft->private value on file */
460 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
461 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
462 #define MEMFILE_ATTR(val) ((val) & 0xffff)
463 /* Used for OOM nofiier */
464 #define OOM_CONTROL (0)
467 * Reclaim flags for mem_cgroup_hierarchical_reclaim
469 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
470 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
471 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
472 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
474 static void mem_cgroup_get(struct mem_cgroup
*memcg
);
475 static void mem_cgroup_put(struct mem_cgroup
*memcg
);
478 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
480 return container_of(s
, struct mem_cgroup
, css
);
483 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
485 return (memcg
== root_mem_cgroup
);
488 /* Writing them here to avoid exposing memcg's inner layout */
489 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
491 void sock_update_memcg(struct sock
*sk
)
493 if (mem_cgroup_sockets_enabled
) {
494 struct mem_cgroup
*memcg
;
495 struct cg_proto
*cg_proto
;
497 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
499 /* Socket cloning can throw us here with sk_cgrp already
500 * filled. It won't however, necessarily happen from
501 * process context. So the test for root memcg given
502 * the current task's memcg won't help us in this case.
504 * Respecting the original socket's memcg is a better
505 * decision in this case.
508 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
509 mem_cgroup_get(sk
->sk_cgrp
->memcg
);
514 memcg
= mem_cgroup_from_task(current
);
515 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
516 if (!mem_cgroup_is_root(memcg
) && memcg_proto_active(cg_proto
)) {
517 mem_cgroup_get(memcg
);
518 sk
->sk_cgrp
= cg_proto
;
523 EXPORT_SYMBOL(sock_update_memcg
);
525 void sock_release_memcg(struct sock
*sk
)
527 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
528 struct mem_cgroup
*memcg
;
529 WARN_ON(!sk
->sk_cgrp
->memcg
);
530 memcg
= sk
->sk_cgrp
->memcg
;
531 mem_cgroup_put(memcg
);
535 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
537 if (!memcg
|| mem_cgroup_is_root(memcg
))
540 return &memcg
->tcp_mem
.cg_proto
;
542 EXPORT_SYMBOL(tcp_proto_cgroup
);
544 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
546 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
548 static_key_slow_dec(&memcg_socket_limit_enabled
);
551 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
556 #ifdef CONFIG_MEMCG_KMEM
558 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
559 * There are two main reasons for not using the css_id for this:
560 * 1) this works better in sparse environments, where we have a lot of memcgs,
561 * but only a few kmem-limited. Or also, if we have, for instance, 200
562 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
563 * 200 entry array for that.
565 * 2) In order not to violate the cgroup API, we would like to do all memory
566 * allocation in ->create(). At that point, we haven't yet allocated the
567 * css_id. Having a separate index prevents us from messing with the cgroup
570 * The current size of the caches array is stored in
571 * memcg_limited_groups_array_size. It will double each time we have to
574 static DEFINE_IDA(kmem_limited_groups
);
575 int memcg_limited_groups_array_size
;
578 * MIN_SIZE is different than 1, because we would like to avoid going through
579 * the alloc/free process all the time. In a small machine, 4 kmem-limited
580 * cgroups is a reasonable guess. In the future, it could be a parameter or
581 * tunable, but that is strictly not necessary.
583 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
584 * this constant directly from cgroup, but it is understandable that this is
585 * better kept as an internal representation in cgroup.c. In any case, the
586 * css_id space is not getting any smaller, and we don't have to necessarily
587 * increase ours as well if it increases.
589 #define MEMCG_CACHES_MIN_SIZE 4
590 #define MEMCG_CACHES_MAX_SIZE 65535
593 * A lot of the calls to the cache allocation functions are expected to be
594 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
595 * conditional to this static branch, we'll have to allow modules that does
596 * kmem_cache_alloc and the such to see this symbol as well
598 struct static_key memcg_kmem_enabled_key
;
599 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
601 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
603 if (memcg_kmem_is_active(memcg
)) {
604 static_key_slow_dec(&memcg_kmem_enabled_key
);
605 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
608 * This check can't live in kmem destruction function,
609 * since the charges will outlive the cgroup
611 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
614 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
617 #endif /* CONFIG_MEMCG_KMEM */
619 static void disarm_static_keys(struct mem_cgroup
*memcg
)
621 disarm_sock_keys(memcg
);
622 disarm_kmem_keys(memcg
);
625 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
627 static struct mem_cgroup_per_zone
*
628 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
630 return &memcg
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
633 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
638 static struct mem_cgroup_per_zone
*
639 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
641 int nid
= page_to_nid(page
);
642 int zid
= page_zonenum(page
);
644 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
647 static struct mem_cgroup_tree_per_zone
*
648 soft_limit_tree_node_zone(int nid
, int zid
)
650 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
653 static struct mem_cgroup_tree_per_zone
*
654 soft_limit_tree_from_page(struct page
*page
)
656 int nid
= page_to_nid(page
);
657 int zid
= page_zonenum(page
);
659 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
663 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
664 struct mem_cgroup_per_zone
*mz
,
665 struct mem_cgroup_tree_per_zone
*mctz
,
666 unsigned long long new_usage_in_excess
)
668 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
669 struct rb_node
*parent
= NULL
;
670 struct mem_cgroup_per_zone
*mz_node
;
675 mz
->usage_in_excess
= new_usage_in_excess
;
676 if (!mz
->usage_in_excess
)
680 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
682 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
685 * We can't avoid mem cgroups that are over their soft
686 * limit by the same amount
688 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
691 rb_link_node(&mz
->tree_node
, parent
, p
);
692 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
697 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
698 struct mem_cgroup_per_zone
*mz
,
699 struct mem_cgroup_tree_per_zone
*mctz
)
703 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
708 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
709 struct mem_cgroup_per_zone
*mz
,
710 struct mem_cgroup_tree_per_zone
*mctz
)
712 spin_lock(&mctz
->lock
);
713 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
714 spin_unlock(&mctz
->lock
);
718 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
720 unsigned long long excess
;
721 struct mem_cgroup_per_zone
*mz
;
722 struct mem_cgroup_tree_per_zone
*mctz
;
723 int nid
= page_to_nid(page
);
724 int zid
= page_zonenum(page
);
725 mctz
= soft_limit_tree_from_page(page
);
728 * Necessary to update all ancestors when hierarchy is used.
729 * because their event counter is not touched.
731 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
732 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
733 excess
= res_counter_soft_limit_excess(&memcg
->res
);
735 * We have to update the tree if mz is on RB-tree or
736 * mem is over its softlimit.
738 if (excess
|| mz
->on_tree
) {
739 spin_lock(&mctz
->lock
);
740 /* if on-tree, remove it */
742 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
744 * Insert again. mz->usage_in_excess will be updated.
745 * If excess is 0, no tree ops.
747 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
748 spin_unlock(&mctz
->lock
);
753 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
756 struct mem_cgroup_per_zone
*mz
;
757 struct mem_cgroup_tree_per_zone
*mctz
;
759 for_each_node(node
) {
760 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
761 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
762 mctz
= soft_limit_tree_node_zone(node
, zone
);
763 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
768 static struct mem_cgroup_per_zone
*
769 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
771 struct rb_node
*rightmost
= NULL
;
772 struct mem_cgroup_per_zone
*mz
;
776 rightmost
= rb_last(&mctz
->rb_root
);
778 goto done
; /* Nothing to reclaim from */
780 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
782 * Remove the node now but someone else can add it back,
783 * we will to add it back at the end of reclaim to its correct
784 * position in the tree.
786 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
787 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
788 !css_tryget(&mz
->memcg
->css
))
794 static struct mem_cgroup_per_zone
*
795 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
797 struct mem_cgroup_per_zone
*mz
;
799 spin_lock(&mctz
->lock
);
800 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
801 spin_unlock(&mctz
->lock
);
806 * Implementation Note: reading percpu statistics for memcg.
808 * Both of vmstat[] and percpu_counter has threshold and do periodic
809 * synchronization to implement "quick" read. There are trade-off between
810 * reading cost and precision of value. Then, we may have a chance to implement
811 * a periodic synchronizion of counter in memcg's counter.
813 * But this _read() function is used for user interface now. The user accounts
814 * memory usage by memory cgroup and he _always_ requires exact value because
815 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
816 * have to visit all online cpus and make sum. So, for now, unnecessary
817 * synchronization is not implemented. (just implemented for cpu hotplug)
819 * If there are kernel internal actions which can make use of some not-exact
820 * value, and reading all cpu value can be performance bottleneck in some
821 * common workload, threashold and synchonization as vmstat[] should be
824 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
825 enum mem_cgroup_stat_index idx
)
831 for_each_online_cpu(cpu
)
832 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
833 #ifdef CONFIG_HOTPLUG_CPU
834 spin_lock(&memcg
->pcp_counter_lock
);
835 val
+= memcg
->nocpu_base
.count
[idx
];
836 spin_unlock(&memcg
->pcp_counter_lock
);
842 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
845 int val
= (charge
) ? 1 : -1;
846 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
849 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
850 enum mem_cgroup_events_index idx
)
852 unsigned long val
= 0;
855 for_each_online_cpu(cpu
)
856 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
857 #ifdef CONFIG_HOTPLUG_CPU
858 spin_lock(&memcg
->pcp_counter_lock
);
859 val
+= memcg
->nocpu_base
.events
[idx
];
860 spin_unlock(&memcg
->pcp_counter_lock
);
865 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
866 bool anon
, int nr_pages
)
871 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
872 * counted as CACHE even if it's on ANON LRU.
875 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
878 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
881 /* pagein of a big page is an event. So, ignore page size */
883 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
885 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
886 nr_pages
= -nr_pages
; /* for event */
889 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
895 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
897 struct mem_cgroup_per_zone
*mz
;
899 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
900 return mz
->lru_size
[lru
];
904 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
905 unsigned int lru_mask
)
907 struct mem_cgroup_per_zone
*mz
;
909 unsigned long ret
= 0;
911 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
914 if (BIT(lru
) & lru_mask
)
915 ret
+= mz
->lru_size
[lru
];
921 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
922 int nid
, unsigned int lru_mask
)
927 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
928 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
934 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
935 unsigned int lru_mask
)
940 for_each_node_state(nid
, N_MEMORY
)
941 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
945 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
946 enum mem_cgroup_events_target target
)
948 unsigned long val
, next
;
950 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
951 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
952 /* from time_after() in jiffies.h */
953 if ((long)next
- (long)val
< 0) {
955 case MEM_CGROUP_TARGET_THRESH
:
956 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
958 case MEM_CGROUP_TARGET_SOFTLIMIT
:
959 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
961 case MEM_CGROUP_TARGET_NUMAINFO
:
962 next
= val
+ NUMAINFO_EVENTS_TARGET
;
967 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
974 * Check events in order.
977 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
980 /* threshold event is triggered in finer grain than soft limit */
981 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
982 MEM_CGROUP_TARGET_THRESH
))) {
984 bool do_numainfo __maybe_unused
;
986 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
987 MEM_CGROUP_TARGET_SOFTLIMIT
);
989 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
990 MEM_CGROUP_TARGET_NUMAINFO
);
994 mem_cgroup_threshold(memcg
);
995 if (unlikely(do_softlimit
))
996 mem_cgroup_update_tree(memcg
, page
);
998 if (unlikely(do_numainfo
))
999 atomic_inc(&memcg
->numainfo_events
);
1005 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
1007 return mem_cgroup_from_css(
1008 cgroup_subsys_state(cont
, mem_cgroup_subsys_id
));
1011 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1014 * mm_update_next_owner() may clear mm->owner to NULL
1015 * if it races with swapoff, page migration, etc.
1016 * So this can be called with p == NULL.
1021 return mem_cgroup_from_css(task_subsys_state(p
, mem_cgroup_subsys_id
));
1024 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1026 struct mem_cgroup
*memcg
= NULL
;
1031 * Because we have no locks, mm->owner's may be being moved to other
1032 * cgroup. We use css_tryget() here even if this looks
1033 * pessimistic (rather than adding locks here).
1037 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1038 if (unlikely(!memcg
))
1040 } while (!css_tryget(&memcg
->css
));
1046 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1047 * @root: hierarchy root
1048 * @prev: previously returned memcg, NULL on first invocation
1049 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1051 * Returns references to children of the hierarchy below @root, or
1052 * @root itself, or %NULL after a full round-trip.
1054 * Caller must pass the return value in @prev on subsequent
1055 * invocations for reference counting, or use mem_cgroup_iter_break()
1056 * to cancel a hierarchy walk before the round-trip is complete.
1058 * Reclaimers can specify a zone and a priority level in @reclaim to
1059 * divide up the memcgs in the hierarchy among all concurrent
1060 * reclaimers operating on the same zone and priority.
1062 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1063 struct mem_cgroup
*prev
,
1064 struct mem_cgroup_reclaim_cookie
*reclaim
)
1066 struct mem_cgroup
*memcg
= NULL
;
1069 if (mem_cgroup_disabled())
1073 root
= root_mem_cgroup
;
1075 if (prev
&& !reclaim
)
1076 id
= css_id(&prev
->css
);
1078 if (prev
&& prev
!= root
)
1079 css_put(&prev
->css
);
1081 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1088 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1089 struct cgroup_subsys_state
*css
;
1092 int nid
= zone_to_nid(reclaim
->zone
);
1093 int zid
= zone_idx(reclaim
->zone
);
1094 struct mem_cgroup_per_zone
*mz
;
1096 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1097 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1098 if (prev
&& reclaim
->generation
!= iter
->generation
)
1100 id
= iter
->position
;
1104 css
= css_get_next(&mem_cgroup_subsys
, id
+ 1, &root
->css
, &id
);
1106 if (css
== &root
->css
|| css_tryget(css
))
1107 memcg
= mem_cgroup_from_css(css
);
1113 iter
->position
= id
;
1116 else if (!prev
&& memcg
)
1117 reclaim
->generation
= iter
->generation
;
1127 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1128 * @root: hierarchy root
1129 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1131 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1132 struct mem_cgroup
*prev
)
1135 root
= root_mem_cgroup
;
1136 if (prev
&& prev
!= root
)
1137 css_put(&prev
->css
);
1141 * Iteration constructs for visiting all cgroups (under a tree). If
1142 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1143 * be used for reference counting.
1145 #define for_each_mem_cgroup_tree(iter, root) \
1146 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1148 iter = mem_cgroup_iter(root, iter, NULL))
1150 #define for_each_mem_cgroup(iter) \
1151 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1153 iter = mem_cgroup_iter(NULL, iter, NULL))
1155 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1157 struct mem_cgroup
*memcg
;
1160 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1161 if (unlikely(!memcg
))
1166 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1169 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1177 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1180 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1181 * @zone: zone of the wanted lruvec
1182 * @memcg: memcg of the wanted lruvec
1184 * Returns the lru list vector holding pages for the given @zone and
1185 * @mem. This can be the global zone lruvec, if the memory controller
1188 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1189 struct mem_cgroup
*memcg
)
1191 struct mem_cgroup_per_zone
*mz
;
1192 struct lruvec
*lruvec
;
1194 if (mem_cgroup_disabled()) {
1195 lruvec
= &zone
->lruvec
;
1199 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1200 lruvec
= &mz
->lruvec
;
1203 * Since a node can be onlined after the mem_cgroup was created,
1204 * we have to be prepared to initialize lruvec->zone here;
1205 * and if offlined then reonlined, we need to reinitialize it.
1207 if (unlikely(lruvec
->zone
!= zone
))
1208 lruvec
->zone
= zone
;
1213 * Following LRU functions are allowed to be used without PCG_LOCK.
1214 * Operations are called by routine of global LRU independently from memcg.
1215 * What we have to take care of here is validness of pc->mem_cgroup.
1217 * Changes to pc->mem_cgroup happens when
1220 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1221 * It is added to LRU before charge.
1222 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1223 * When moving account, the page is not on LRU. It's isolated.
1227 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1229 * @zone: zone of the page
1231 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1233 struct mem_cgroup_per_zone
*mz
;
1234 struct mem_cgroup
*memcg
;
1235 struct page_cgroup
*pc
;
1236 struct lruvec
*lruvec
;
1238 if (mem_cgroup_disabled()) {
1239 lruvec
= &zone
->lruvec
;
1243 pc
= lookup_page_cgroup(page
);
1244 memcg
= pc
->mem_cgroup
;
1247 * Surreptitiously switch any uncharged offlist page to root:
1248 * an uncharged page off lru does nothing to secure
1249 * its former mem_cgroup from sudden removal.
1251 * Our caller holds lru_lock, and PageCgroupUsed is updated
1252 * under page_cgroup lock: between them, they make all uses
1253 * of pc->mem_cgroup safe.
1255 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1256 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1258 mz
= page_cgroup_zoneinfo(memcg
, page
);
1259 lruvec
= &mz
->lruvec
;
1262 * Since a node can be onlined after the mem_cgroup was created,
1263 * we have to be prepared to initialize lruvec->zone here;
1264 * and if offlined then reonlined, we need to reinitialize it.
1266 if (unlikely(lruvec
->zone
!= zone
))
1267 lruvec
->zone
= zone
;
1272 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1273 * @lruvec: mem_cgroup per zone lru vector
1274 * @lru: index of lru list the page is sitting on
1275 * @nr_pages: positive when adding or negative when removing
1277 * This function must be called when a page is added to or removed from an
1280 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1283 struct mem_cgroup_per_zone
*mz
;
1284 unsigned long *lru_size
;
1286 if (mem_cgroup_disabled())
1289 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1290 lru_size
= mz
->lru_size
+ lru
;
1291 *lru_size
+= nr_pages
;
1292 VM_BUG_ON((long)(*lru_size
) < 0);
1296 * Checks whether given mem is same or in the root_mem_cgroup's
1299 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1300 struct mem_cgroup
*memcg
)
1302 if (root_memcg
== memcg
)
1304 if (!root_memcg
->use_hierarchy
|| !memcg
)
1306 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1309 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1310 struct mem_cgroup
*memcg
)
1315 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1320 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*memcg
)
1323 struct mem_cgroup
*curr
= NULL
;
1324 struct task_struct
*p
;
1326 p
= find_lock_task_mm(task
);
1328 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1332 * All threads may have already detached their mm's, but the oom
1333 * killer still needs to detect if they have already been oom
1334 * killed to prevent needlessly killing additional tasks.
1337 curr
= mem_cgroup_from_task(task
);
1339 css_get(&curr
->css
);
1345 * We should check use_hierarchy of "memcg" not "curr". Because checking
1346 * use_hierarchy of "curr" here make this function true if hierarchy is
1347 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1348 * hierarchy(even if use_hierarchy is disabled in "memcg").
1350 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1351 css_put(&curr
->css
);
1355 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1357 unsigned long inactive_ratio
;
1358 unsigned long inactive
;
1359 unsigned long active
;
1362 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1363 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1365 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1367 inactive_ratio
= int_sqrt(10 * gb
);
1371 return inactive
* inactive_ratio
< active
;
1374 int mem_cgroup_inactive_file_is_low(struct lruvec
*lruvec
)
1376 unsigned long active
;
1377 unsigned long inactive
;
1379 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1380 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1382 return (active
> inactive
);
1385 #define mem_cgroup_from_res_counter(counter, member) \
1386 container_of(counter, struct mem_cgroup, member)
1389 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1390 * @memcg: the memory cgroup
1392 * Returns the maximum amount of memory @mem can be charged with, in
1395 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1397 unsigned long long margin
;
1399 margin
= res_counter_margin(&memcg
->res
);
1400 if (do_swap_account
)
1401 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1402 return margin
>> PAGE_SHIFT
;
1405 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1407 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1410 if (cgrp
->parent
== NULL
)
1411 return vm_swappiness
;
1413 return memcg
->swappiness
;
1417 * memcg->moving_account is used for checking possibility that some thread is
1418 * calling move_account(). When a thread on CPU-A starts moving pages under
1419 * a memcg, other threads should check memcg->moving_account under
1420 * rcu_read_lock(), like this:
1424 * memcg->moving_account+1 if (memcg->mocing_account)
1426 * synchronize_rcu() update something.
1431 /* for quick checking without looking up memcg */
1432 atomic_t memcg_moving __read_mostly
;
1434 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1436 atomic_inc(&memcg_moving
);
1437 atomic_inc(&memcg
->moving_account
);
1441 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1444 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1445 * We check NULL in callee rather than caller.
1448 atomic_dec(&memcg_moving
);
1449 atomic_dec(&memcg
->moving_account
);
1454 * 2 routines for checking "mem" is under move_account() or not.
1456 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1457 * is used for avoiding races in accounting. If true,
1458 * pc->mem_cgroup may be overwritten.
1460 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1461 * under hierarchy of moving cgroups. This is for
1462 * waiting at hith-memory prressure caused by "move".
1465 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1467 VM_BUG_ON(!rcu_read_lock_held());
1468 return atomic_read(&memcg
->moving_account
) > 0;
1471 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1473 struct mem_cgroup
*from
;
1474 struct mem_cgroup
*to
;
1477 * Unlike task_move routines, we access mc.to, mc.from not under
1478 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1480 spin_lock(&mc
.lock
);
1486 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1487 || mem_cgroup_same_or_subtree(memcg
, to
);
1489 spin_unlock(&mc
.lock
);
1493 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1495 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1496 if (mem_cgroup_under_move(memcg
)) {
1498 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1499 /* moving charge context might have finished. */
1502 finish_wait(&mc
.waitq
, &wait
);
1510 * Take this lock when
1511 * - a code tries to modify page's memcg while it's USED.
1512 * - a code tries to modify page state accounting in a memcg.
1513 * see mem_cgroup_stolen(), too.
1515 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1516 unsigned long *flags
)
1518 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1521 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1522 unsigned long *flags
)
1524 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1528 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1529 * @memcg: The memory cgroup that went over limit
1530 * @p: Task that is going to be killed
1532 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1535 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1537 struct cgroup
*task_cgrp
;
1538 struct cgroup
*mem_cgrp
;
1540 * Need a buffer in BSS, can't rely on allocations. The code relies
1541 * on the assumption that OOM is serialized for memory controller.
1542 * If this assumption is broken, revisit this code.
1544 static char memcg_name
[PATH_MAX
];
1552 mem_cgrp
= memcg
->css
.cgroup
;
1553 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1555 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1558 * Unfortunately, we are unable to convert to a useful name
1559 * But we'll still print out the usage information
1566 printk(KERN_INFO
"Task in %s killed", memcg_name
);
1569 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1577 * Continues from above, so we don't need an KERN_ level
1579 printk(KERN_CONT
" as a result of limit of %s\n", memcg_name
);
1582 printk(KERN_INFO
"memory: usage %llukB, limit %llukB, failcnt %llu\n",
1583 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1584 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1585 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1586 printk(KERN_INFO
"memory+swap: usage %llukB, limit %llukB, "
1588 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1589 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1590 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1591 printk(KERN_INFO
"kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1592 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1593 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1594 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1598 * This function returns the number of memcg under hierarchy tree. Returns
1599 * 1(self count) if no children.
1601 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1604 struct mem_cgroup
*iter
;
1606 for_each_mem_cgroup_tree(iter
, memcg
)
1612 * Return the memory (and swap, if configured) limit for a memcg.
1614 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1618 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1621 * Do not consider swap space if we cannot swap due to swappiness
1623 if (mem_cgroup_swappiness(memcg
)) {
1626 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1627 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1630 * If memsw is finite and limits the amount of swap space
1631 * available to this memcg, return that limit.
1633 limit
= min(limit
, memsw
);
1639 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1642 struct mem_cgroup
*iter
;
1643 unsigned long chosen_points
= 0;
1644 unsigned long totalpages
;
1645 unsigned int points
= 0;
1646 struct task_struct
*chosen
= NULL
;
1649 * If current has a pending SIGKILL, then automatically select it. The
1650 * goal is to allow it to allocate so that it may quickly exit and free
1653 if (fatal_signal_pending(current
)) {
1654 set_thread_flag(TIF_MEMDIE
);
1658 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1659 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1660 for_each_mem_cgroup_tree(iter
, memcg
) {
1661 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1662 struct cgroup_iter it
;
1663 struct task_struct
*task
;
1665 cgroup_iter_start(cgroup
, &it
);
1666 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1667 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1669 case OOM_SCAN_SELECT
:
1671 put_task_struct(chosen
);
1673 chosen_points
= ULONG_MAX
;
1674 get_task_struct(chosen
);
1676 case OOM_SCAN_CONTINUE
:
1678 case OOM_SCAN_ABORT
:
1679 cgroup_iter_end(cgroup
, &it
);
1680 mem_cgroup_iter_break(memcg
, iter
);
1682 put_task_struct(chosen
);
1687 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1688 if (points
> chosen_points
) {
1690 put_task_struct(chosen
);
1692 chosen_points
= points
;
1693 get_task_struct(chosen
);
1696 cgroup_iter_end(cgroup
, &it
);
1701 points
= chosen_points
* 1000 / totalpages
;
1702 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1703 NULL
, "Memory cgroup out of memory");
1706 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1708 unsigned long flags
)
1710 unsigned long total
= 0;
1711 bool noswap
= false;
1714 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1716 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1719 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1721 drain_all_stock_async(memcg
);
1722 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1724 * Allow limit shrinkers, which are triggered directly
1725 * by userspace, to catch signals and stop reclaim
1726 * after minimal progress, regardless of the margin.
1728 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1730 if (mem_cgroup_margin(memcg
))
1733 * If nothing was reclaimed after two attempts, there
1734 * may be no reclaimable pages in this hierarchy.
1743 * test_mem_cgroup_node_reclaimable
1744 * @memcg: the target memcg
1745 * @nid: the node ID to be checked.
1746 * @noswap : specify true here if the user wants flle only information.
1748 * This function returns whether the specified memcg contains any
1749 * reclaimable pages on a node. Returns true if there are any reclaimable
1750 * pages in the node.
1752 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1753 int nid
, bool noswap
)
1755 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1757 if (noswap
|| !total_swap_pages
)
1759 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1764 #if MAX_NUMNODES > 1
1767 * Always updating the nodemask is not very good - even if we have an empty
1768 * list or the wrong list here, we can start from some node and traverse all
1769 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1772 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1776 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1777 * pagein/pageout changes since the last update.
1779 if (!atomic_read(&memcg
->numainfo_events
))
1781 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1784 /* make a nodemask where this memcg uses memory from */
1785 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1787 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1789 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1790 node_clear(nid
, memcg
->scan_nodes
);
1793 atomic_set(&memcg
->numainfo_events
, 0);
1794 atomic_set(&memcg
->numainfo_updating
, 0);
1798 * Selecting a node where we start reclaim from. Because what we need is just
1799 * reducing usage counter, start from anywhere is O,K. Considering
1800 * memory reclaim from current node, there are pros. and cons.
1802 * Freeing memory from current node means freeing memory from a node which
1803 * we'll use or we've used. So, it may make LRU bad. And if several threads
1804 * hit limits, it will see a contention on a node. But freeing from remote
1805 * node means more costs for memory reclaim because of memory latency.
1807 * Now, we use round-robin. Better algorithm is welcomed.
1809 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1813 mem_cgroup_may_update_nodemask(memcg
);
1814 node
= memcg
->last_scanned_node
;
1816 node
= next_node(node
, memcg
->scan_nodes
);
1817 if (node
== MAX_NUMNODES
)
1818 node
= first_node(memcg
->scan_nodes
);
1820 * We call this when we hit limit, not when pages are added to LRU.
1821 * No LRU may hold pages because all pages are UNEVICTABLE or
1822 * memcg is too small and all pages are not on LRU. In that case,
1823 * we use curret node.
1825 if (unlikely(node
== MAX_NUMNODES
))
1826 node
= numa_node_id();
1828 memcg
->last_scanned_node
= node
;
1833 * Check all nodes whether it contains reclaimable pages or not.
1834 * For quick scan, we make use of scan_nodes. This will allow us to skip
1835 * unused nodes. But scan_nodes is lazily updated and may not cotain
1836 * enough new information. We need to do double check.
1838 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1843 * quick check...making use of scan_node.
1844 * We can skip unused nodes.
1846 if (!nodes_empty(memcg
->scan_nodes
)) {
1847 for (nid
= first_node(memcg
->scan_nodes
);
1849 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1851 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1856 * Check rest of nodes.
1858 for_each_node_state(nid
, N_MEMORY
) {
1859 if (node_isset(nid
, memcg
->scan_nodes
))
1861 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1868 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1873 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1875 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1879 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1882 unsigned long *total_scanned
)
1884 struct mem_cgroup
*victim
= NULL
;
1887 unsigned long excess
;
1888 unsigned long nr_scanned
;
1889 struct mem_cgroup_reclaim_cookie reclaim
= {
1894 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
1897 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1902 * If we have not been able to reclaim
1903 * anything, it might because there are
1904 * no reclaimable pages under this hierarchy
1909 * We want to do more targeted reclaim.
1910 * excess >> 2 is not to excessive so as to
1911 * reclaim too much, nor too less that we keep
1912 * coming back to reclaim from this cgroup
1914 if (total
>= (excess
>> 2) ||
1915 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1920 if (!mem_cgroup_reclaimable(victim
, false))
1922 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1924 *total_scanned
+= nr_scanned
;
1925 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
1928 mem_cgroup_iter_break(root_memcg
, victim
);
1933 * Check OOM-Killer is already running under our hierarchy.
1934 * If someone is running, return false.
1935 * Has to be called with memcg_oom_lock
1937 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
1939 struct mem_cgroup
*iter
, *failed
= NULL
;
1941 for_each_mem_cgroup_tree(iter
, memcg
) {
1942 if (iter
->oom_lock
) {
1944 * this subtree of our hierarchy is already locked
1945 * so we cannot give a lock.
1948 mem_cgroup_iter_break(memcg
, iter
);
1951 iter
->oom_lock
= true;
1958 * OK, we failed to lock the whole subtree so we have to clean up
1959 * what we set up to the failing subtree
1961 for_each_mem_cgroup_tree(iter
, memcg
) {
1962 if (iter
== failed
) {
1963 mem_cgroup_iter_break(memcg
, iter
);
1966 iter
->oom_lock
= false;
1972 * Has to be called with memcg_oom_lock
1974 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1976 struct mem_cgroup
*iter
;
1978 for_each_mem_cgroup_tree(iter
, memcg
)
1979 iter
->oom_lock
= false;
1983 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1985 struct mem_cgroup
*iter
;
1987 for_each_mem_cgroup_tree(iter
, memcg
)
1988 atomic_inc(&iter
->under_oom
);
1991 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1993 struct mem_cgroup
*iter
;
1996 * When a new child is created while the hierarchy is under oom,
1997 * mem_cgroup_oom_lock() may not be called. We have to use
1998 * atomic_add_unless() here.
2000 for_each_mem_cgroup_tree(iter
, memcg
)
2001 atomic_add_unless(&iter
->under_oom
, -1, 0);
2004 static DEFINE_SPINLOCK(memcg_oom_lock
);
2005 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2007 struct oom_wait_info
{
2008 struct mem_cgroup
*memcg
;
2012 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2013 unsigned mode
, int sync
, void *arg
)
2015 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2016 struct mem_cgroup
*oom_wait_memcg
;
2017 struct oom_wait_info
*oom_wait_info
;
2019 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2020 oom_wait_memcg
= oom_wait_info
->memcg
;
2023 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2024 * Then we can use css_is_ancestor without taking care of RCU.
2026 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2027 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2029 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2032 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2034 /* for filtering, pass "memcg" as argument. */
2035 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2038 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2040 if (memcg
&& atomic_read(&memcg
->under_oom
))
2041 memcg_wakeup_oom(memcg
);
2045 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2047 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2050 struct oom_wait_info owait
;
2051 bool locked
, need_to_kill
;
2053 owait
.memcg
= memcg
;
2054 owait
.wait
.flags
= 0;
2055 owait
.wait
.func
= memcg_oom_wake_function
;
2056 owait
.wait
.private = current
;
2057 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2058 need_to_kill
= true;
2059 mem_cgroup_mark_under_oom(memcg
);
2061 /* At first, try to OOM lock hierarchy under memcg.*/
2062 spin_lock(&memcg_oom_lock
);
2063 locked
= mem_cgroup_oom_lock(memcg
);
2065 * Even if signal_pending(), we can't quit charge() loop without
2066 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2067 * under OOM is always welcomed, use TASK_KILLABLE here.
2069 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2070 if (!locked
|| memcg
->oom_kill_disable
)
2071 need_to_kill
= false;
2073 mem_cgroup_oom_notify(memcg
);
2074 spin_unlock(&memcg_oom_lock
);
2077 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2078 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2081 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2083 spin_lock(&memcg_oom_lock
);
2085 mem_cgroup_oom_unlock(memcg
);
2086 memcg_wakeup_oom(memcg
);
2087 spin_unlock(&memcg_oom_lock
);
2089 mem_cgroup_unmark_under_oom(memcg
);
2091 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2093 /* Give chance to dying process */
2094 schedule_timeout_uninterruptible(1);
2099 * Currently used to update mapped file statistics, but the routine can be
2100 * generalized to update other statistics as well.
2102 * Notes: Race condition
2104 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2105 * it tends to be costly. But considering some conditions, we doesn't need
2106 * to do so _always_.
2108 * Considering "charge", lock_page_cgroup() is not required because all
2109 * file-stat operations happen after a page is attached to radix-tree. There
2110 * are no race with "charge".
2112 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2113 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2114 * if there are race with "uncharge". Statistics itself is properly handled
2117 * Considering "move", this is an only case we see a race. To make the race
2118 * small, we check mm->moving_account and detect there are possibility of race
2119 * If there is, we take a lock.
2122 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2123 bool *locked
, unsigned long *flags
)
2125 struct mem_cgroup
*memcg
;
2126 struct page_cgroup
*pc
;
2128 pc
= lookup_page_cgroup(page
);
2130 memcg
= pc
->mem_cgroup
;
2131 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2134 * If this memory cgroup is not under account moving, we don't
2135 * need to take move_lock_mem_cgroup(). Because we already hold
2136 * rcu_read_lock(), any calls to move_account will be delayed until
2137 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2139 if (!mem_cgroup_stolen(memcg
))
2142 move_lock_mem_cgroup(memcg
, flags
);
2143 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2144 move_unlock_mem_cgroup(memcg
, flags
);
2150 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2152 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2155 * It's guaranteed that pc->mem_cgroup never changes while
2156 * lock is held because a routine modifies pc->mem_cgroup
2157 * should take move_lock_mem_cgroup().
2159 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2162 void mem_cgroup_update_page_stat(struct page
*page
,
2163 enum mem_cgroup_page_stat_item idx
, int val
)
2165 struct mem_cgroup
*memcg
;
2166 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2167 unsigned long uninitialized_var(flags
);
2169 if (mem_cgroup_disabled())
2172 memcg
= pc
->mem_cgroup
;
2173 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2177 case MEMCG_NR_FILE_MAPPED
:
2178 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2184 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2188 * size of first charge trial. "32" comes from vmscan.c's magic value.
2189 * TODO: maybe necessary to use big numbers in big irons.
2191 #define CHARGE_BATCH 32U
2192 struct memcg_stock_pcp
{
2193 struct mem_cgroup
*cached
; /* this never be root cgroup */
2194 unsigned int nr_pages
;
2195 struct work_struct work
;
2196 unsigned long flags
;
2197 #define FLUSHING_CACHED_CHARGE 0
2199 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2200 static DEFINE_MUTEX(percpu_charge_mutex
);
2203 * consume_stock: Try to consume stocked charge on this cpu.
2204 * @memcg: memcg to consume from.
2205 * @nr_pages: how many pages to charge.
2207 * The charges will only happen if @memcg matches the current cpu's memcg
2208 * stock, and at least @nr_pages are available in that stock. Failure to
2209 * service an allocation will refill the stock.
2211 * returns true if successful, false otherwise.
2213 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2215 struct memcg_stock_pcp
*stock
;
2218 if (nr_pages
> CHARGE_BATCH
)
2221 stock
= &get_cpu_var(memcg_stock
);
2222 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2223 stock
->nr_pages
-= nr_pages
;
2224 else /* need to call res_counter_charge */
2226 put_cpu_var(memcg_stock
);
2231 * Returns stocks cached in percpu to res_counter and reset cached information.
2233 static void drain_stock(struct memcg_stock_pcp
*stock
)
2235 struct mem_cgroup
*old
= stock
->cached
;
2237 if (stock
->nr_pages
) {
2238 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2240 res_counter_uncharge(&old
->res
, bytes
);
2241 if (do_swap_account
)
2242 res_counter_uncharge(&old
->memsw
, bytes
);
2243 stock
->nr_pages
= 0;
2245 stock
->cached
= NULL
;
2249 * This must be called under preempt disabled or must be called by
2250 * a thread which is pinned to local cpu.
2252 static void drain_local_stock(struct work_struct
*dummy
)
2254 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2256 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2260 * Cache charges(val) which is from res_counter, to local per_cpu area.
2261 * This will be consumed by consume_stock() function, later.
2263 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2265 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2267 if (stock
->cached
!= memcg
) { /* reset if necessary */
2269 stock
->cached
= memcg
;
2271 stock
->nr_pages
+= nr_pages
;
2272 put_cpu_var(memcg_stock
);
2276 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2277 * of the hierarchy under it. sync flag says whether we should block
2278 * until the work is done.
2280 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2284 /* Notify other cpus that system-wide "drain" is running */
2287 for_each_online_cpu(cpu
) {
2288 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2289 struct mem_cgroup
*memcg
;
2291 memcg
= stock
->cached
;
2292 if (!memcg
|| !stock
->nr_pages
)
2294 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2296 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2298 drain_local_stock(&stock
->work
);
2300 schedule_work_on(cpu
, &stock
->work
);
2308 for_each_online_cpu(cpu
) {
2309 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2310 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2311 flush_work(&stock
->work
);
2318 * Tries to drain stocked charges in other cpus. This function is asynchronous
2319 * and just put a work per cpu for draining localy on each cpu. Caller can
2320 * expects some charges will be back to res_counter later but cannot wait for
2323 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2326 * If someone calls draining, avoid adding more kworker runs.
2328 if (!mutex_trylock(&percpu_charge_mutex
))
2330 drain_all_stock(root_memcg
, false);
2331 mutex_unlock(&percpu_charge_mutex
);
2334 /* This is a synchronous drain interface. */
2335 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2337 /* called when force_empty is called */
2338 mutex_lock(&percpu_charge_mutex
);
2339 drain_all_stock(root_memcg
, true);
2340 mutex_unlock(&percpu_charge_mutex
);
2344 * This function drains percpu counter value from DEAD cpu and
2345 * move it to local cpu. Note that this function can be preempted.
2347 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2351 spin_lock(&memcg
->pcp_counter_lock
);
2352 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2353 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2355 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2356 memcg
->nocpu_base
.count
[i
] += x
;
2358 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2359 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2361 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2362 memcg
->nocpu_base
.events
[i
] += x
;
2364 spin_unlock(&memcg
->pcp_counter_lock
);
2367 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2368 unsigned long action
,
2371 int cpu
= (unsigned long)hcpu
;
2372 struct memcg_stock_pcp
*stock
;
2373 struct mem_cgroup
*iter
;
2375 if (action
== CPU_ONLINE
)
2378 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2381 for_each_mem_cgroup(iter
)
2382 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2384 stock
= &per_cpu(memcg_stock
, cpu
);
2390 /* See __mem_cgroup_try_charge() for details */
2392 CHARGE_OK
, /* success */
2393 CHARGE_RETRY
, /* need to retry but retry is not bad */
2394 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2395 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2396 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2399 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2400 unsigned int nr_pages
, unsigned int min_pages
,
2403 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2404 struct mem_cgroup
*mem_over_limit
;
2405 struct res_counter
*fail_res
;
2406 unsigned long flags
= 0;
2409 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2412 if (!do_swap_account
)
2414 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2418 res_counter_uncharge(&memcg
->res
, csize
);
2419 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2420 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2422 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2424 * Never reclaim on behalf of optional batching, retry with a
2425 * single page instead.
2427 if (nr_pages
> min_pages
)
2428 return CHARGE_RETRY
;
2430 if (!(gfp_mask
& __GFP_WAIT
))
2431 return CHARGE_WOULDBLOCK
;
2433 if (gfp_mask
& __GFP_NORETRY
)
2434 return CHARGE_NOMEM
;
2436 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2437 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2438 return CHARGE_RETRY
;
2440 * Even though the limit is exceeded at this point, reclaim
2441 * may have been able to free some pages. Retry the charge
2442 * before killing the task.
2444 * Only for regular pages, though: huge pages are rather
2445 * unlikely to succeed so close to the limit, and we fall back
2446 * to regular pages anyway in case of failure.
2448 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2449 return CHARGE_RETRY
;
2452 * At task move, charge accounts can be doubly counted. So, it's
2453 * better to wait until the end of task_move if something is going on.
2455 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2456 return CHARGE_RETRY
;
2458 /* If we don't need to call oom-killer at el, return immediately */
2460 return CHARGE_NOMEM
;
2462 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2463 return CHARGE_OOM_DIE
;
2465 return CHARGE_RETRY
;
2469 * __mem_cgroup_try_charge() does
2470 * 1. detect memcg to be charged against from passed *mm and *ptr,
2471 * 2. update res_counter
2472 * 3. call memory reclaim if necessary.
2474 * In some special case, if the task is fatal, fatal_signal_pending() or
2475 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2476 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2477 * as possible without any hazards. 2: all pages should have a valid
2478 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2479 * pointer, that is treated as a charge to root_mem_cgroup.
2481 * So __mem_cgroup_try_charge() will return
2482 * 0 ... on success, filling *ptr with a valid memcg pointer.
2483 * -ENOMEM ... charge failure because of resource limits.
2484 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2486 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2487 * the oom-killer can be invoked.
2489 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2491 unsigned int nr_pages
,
2492 struct mem_cgroup
**ptr
,
2495 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2496 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2497 struct mem_cgroup
*memcg
= NULL
;
2501 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2502 * in system level. So, allow to go ahead dying process in addition to
2505 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2506 || fatal_signal_pending(current
)))
2510 * We always charge the cgroup the mm_struct belongs to.
2511 * The mm_struct's mem_cgroup changes on task migration if the
2512 * thread group leader migrates. It's possible that mm is not
2513 * set, if so charge the root memcg (happens for pagecache usage).
2516 *ptr
= root_mem_cgroup
;
2518 if (*ptr
) { /* css should be a valid one */
2520 if (mem_cgroup_is_root(memcg
))
2522 if (consume_stock(memcg
, nr_pages
))
2524 css_get(&memcg
->css
);
2526 struct task_struct
*p
;
2529 p
= rcu_dereference(mm
->owner
);
2531 * Because we don't have task_lock(), "p" can exit.
2532 * In that case, "memcg" can point to root or p can be NULL with
2533 * race with swapoff. Then, we have small risk of mis-accouning.
2534 * But such kind of mis-account by race always happens because
2535 * we don't have cgroup_mutex(). It's overkill and we allo that
2537 * (*) swapoff at el will charge against mm-struct not against
2538 * task-struct. So, mm->owner can be NULL.
2540 memcg
= mem_cgroup_from_task(p
);
2542 memcg
= root_mem_cgroup
;
2543 if (mem_cgroup_is_root(memcg
)) {
2547 if (consume_stock(memcg
, nr_pages
)) {
2549 * It seems dagerous to access memcg without css_get().
2550 * But considering how consume_stok works, it's not
2551 * necessary. If consume_stock success, some charges
2552 * from this memcg are cached on this cpu. So, we
2553 * don't need to call css_get()/css_tryget() before
2554 * calling consume_stock().
2559 /* after here, we may be blocked. we need to get refcnt */
2560 if (!css_tryget(&memcg
->css
)) {
2570 /* If killed, bypass charge */
2571 if (fatal_signal_pending(current
)) {
2572 css_put(&memcg
->css
);
2577 if (oom
&& !nr_oom_retries
) {
2579 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2582 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2587 case CHARGE_RETRY
: /* not in OOM situation but retry */
2589 css_put(&memcg
->css
);
2592 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2593 css_put(&memcg
->css
);
2595 case CHARGE_NOMEM
: /* OOM routine works */
2597 css_put(&memcg
->css
);
2600 /* If oom, we never return -ENOMEM */
2603 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2604 css_put(&memcg
->css
);
2607 } while (ret
!= CHARGE_OK
);
2609 if (batch
> nr_pages
)
2610 refill_stock(memcg
, batch
- nr_pages
);
2611 css_put(&memcg
->css
);
2619 *ptr
= root_mem_cgroup
;
2624 * Somemtimes we have to undo a charge we got by try_charge().
2625 * This function is for that and do uncharge, put css's refcnt.
2626 * gotten by try_charge().
2628 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2629 unsigned int nr_pages
)
2631 if (!mem_cgroup_is_root(memcg
)) {
2632 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2634 res_counter_uncharge(&memcg
->res
, bytes
);
2635 if (do_swap_account
)
2636 res_counter_uncharge(&memcg
->memsw
, bytes
);
2641 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2642 * This is useful when moving usage to parent cgroup.
2644 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2645 unsigned int nr_pages
)
2647 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2649 if (mem_cgroup_is_root(memcg
))
2652 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2653 if (do_swap_account
)
2654 res_counter_uncharge_until(&memcg
->memsw
,
2655 memcg
->memsw
.parent
, bytes
);
2659 * A helper function to get mem_cgroup from ID. must be called under
2660 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2661 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2662 * called against removed memcg.)
2664 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2666 struct cgroup_subsys_state
*css
;
2668 /* ID 0 is unused ID */
2671 css
= css_lookup(&mem_cgroup_subsys
, id
);
2674 return mem_cgroup_from_css(css
);
2677 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2679 struct mem_cgroup
*memcg
= NULL
;
2680 struct page_cgroup
*pc
;
2684 VM_BUG_ON(!PageLocked(page
));
2686 pc
= lookup_page_cgroup(page
);
2687 lock_page_cgroup(pc
);
2688 if (PageCgroupUsed(pc
)) {
2689 memcg
= pc
->mem_cgroup
;
2690 if (memcg
&& !css_tryget(&memcg
->css
))
2692 } else if (PageSwapCache(page
)) {
2693 ent
.val
= page_private(page
);
2694 id
= lookup_swap_cgroup_id(ent
);
2696 memcg
= mem_cgroup_lookup(id
);
2697 if (memcg
&& !css_tryget(&memcg
->css
))
2701 unlock_page_cgroup(pc
);
2705 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2707 unsigned int nr_pages
,
2708 enum charge_type ctype
,
2711 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2712 struct zone
*uninitialized_var(zone
);
2713 struct lruvec
*lruvec
;
2714 bool was_on_lru
= false;
2717 lock_page_cgroup(pc
);
2718 VM_BUG_ON(PageCgroupUsed(pc
));
2720 * we don't need page_cgroup_lock about tail pages, becase they are not
2721 * accessed by any other context at this point.
2725 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2726 * may already be on some other mem_cgroup's LRU. Take care of it.
2729 zone
= page_zone(page
);
2730 spin_lock_irq(&zone
->lru_lock
);
2731 if (PageLRU(page
)) {
2732 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2734 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2739 pc
->mem_cgroup
= memcg
;
2741 * We access a page_cgroup asynchronously without lock_page_cgroup().
2742 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2743 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2744 * before USED bit, we need memory barrier here.
2745 * See mem_cgroup_add_lru_list(), etc.
2748 SetPageCgroupUsed(pc
);
2752 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2753 VM_BUG_ON(PageLRU(page
));
2755 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2757 spin_unlock_irq(&zone
->lru_lock
);
2760 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2765 mem_cgroup_charge_statistics(memcg
, anon
, nr_pages
);
2766 unlock_page_cgroup(pc
);
2769 * "charge_statistics" updated event counter. Then, check it.
2770 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2771 * if they exceeds softlimit.
2773 memcg_check_events(memcg
, page
);
2776 static DEFINE_MUTEX(set_limit_mutex
);
2778 #ifdef CONFIG_MEMCG_KMEM
2779 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2781 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2782 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2786 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2787 * in the memcg_cache_params struct.
2789 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2791 struct kmem_cache
*cachep
;
2793 VM_BUG_ON(p
->is_root_cache
);
2794 cachep
= p
->root_cache
;
2795 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2798 #ifdef CONFIG_SLABINFO
2799 static int mem_cgroup_slabinfo_read(struct cgroup
*cont
, struct cftype
*cft
,
2802 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
2803 struct memcg_cache_params
*params
;
2805 if (!memcg_can_account_kmem(memcg
))
2808 print_slabinfo_header(m
);
2810 mutex_lock(&memcg
->slab_caches_mutex
);
2811 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2812 cache_show(memcg_params_to_cache(params
), m
);
2813 mutex_unlock(&memcg
->slab_caches_mutex
);
2819 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2821 struct res_counter
*fail_res
;
2822 struct mem_cgroup
*_memcg
;
2826 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2831 * Conditions under which we can wait for the oom_killer. Those are
2832 * the same conditions tested by the core page allocator
2834 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2837 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2840 if (ret
== -EINTR
) {
2842 * __mem_cgroup_try_charge() chosed to bypass to root due to
2843 * OOM kill or fatal signal. Since our only options are to
2844 * either fail the allocation or charge it to this cgroup, do
2845 * it as a temporary condition. But we can't fail. From a
2846 * kmem/slab perspective, the cache has already been selected,
2847 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2850 * This condition will only trigger if the task entered
2851 * memcg_charge_kmem in a sane state, but was OOM-killed during
2852 * __mem_cgroup_try_charge() above. Tasks that were already
2853 * dying when the allocation triggers should have been already
2854 * directed to the root cgroup in memcontrol.h
2856 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2857 if (do_swap_account
)
2858 res_counter_charge_nofail(&memcg
->memsw
, size
,
2862 res_counter_uncharge(&memcg
->kmem
, size
);
2867 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2869 res_counter_uncharge(&memcg
->res
, size
);
2870 if (do_swap_account
)
2871 res_counter_uncharge(&memcg
->memsw
, size
);
2874 if (res_counter_uncharge(&memcg
->kmem
, size
))
2877 if (memcg_kmem_test_and_clear_dead(memcg
))
2878 mem_cgroup_put(memcg
);
2881 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2886 mutex_lock(&memcg
->slab_caches_mutex
);
2887 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2888 mutex_unlock(&memcg
->slab_caches_mutex
);
2892 * helper for acessing a memcg's index. It will be used as an index in the
2893 * child cache array in kmem_cache, and also to derive its name. This function
2894 * will return -1 when this is not a kmem-limited memcg.
2896 int memcg_cache_id(struct mem_cgroup
*memcg
)
2898 return memcg
? memcg
->kmemcg_id
: -1;
2902 * This ends up being protected by the set_limit mutex, during normal
2903 * operation, because that is its main call site.
2905 * But when we create a new cache, we can call this as well if its parent
2906 * is kmem-limited. That will have to hold set_limit_mutex as well.
2908 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2912 num
= ida_simple_get(&kmem_limited_groups
,
2913 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2917 * After this point, kmem_accounted (that we test atomically in
2918 * the beginning of this conditional), is no longer 0. This
2919 * guarantees only one process will set the following boolean
2920 * to true. We don't need test_and_set because we're protected
2921 * by the set_limit_mutex anyway.
2923 memcg_kmem_set_activated(memcg
);
2925 ret
= memcg_update_all_caches(num
+1);
2927 ida_simple_remove(&kmem_limited_groups
, num
);
2928 memcg_kmem_clear_activated(memcg
);
2932 memcg
->kmemcg_id
= num
;
2933 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
2934 mutex_init(&memcg
->slab_caches_mutex
);
2938 static size_t memcg_caches_array_size(int num_groups
)
2941 if (num_groups
<= 0)
2944 size
= 2 * num_groups
;
2945 if (size
< MEMCG_CACHES_MIN_SIZE
)
2946 size
= MEMCG_CACHES_MIN_SIZE
;
2947 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2948 size
= MEMCG_CACHES_MAX_SIZE
;
2954 * We should update the current array size iff all caches updates succeed. This
2955 * can only be done from the slab side. The slab mutex needs to be held when
2958 void memcg_update_array_size(int num
)
2960 if (num
> memcg_limited_groups_array_size
)
2961 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2964 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2966 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2968 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
2970 if (num_groups
> memcg_limited_groups_array_size
) {
2972 ssize_t size
= memcg_caches_array_size(num_groups
);
2974 size
*= sizeof(void *);
2975 size
+= sizeof(struct memcg_cache_params
);
2977 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2978 if (!s
->memcg_params
) {
2979 s
->memcg_params
= cur_params
;
2983 s
->memcg_params
->is_root_cache
= true;
2986 * There is the chance it will be bigger than
2987 * memcg_limited_groups_array_size, if we failed an allocation
2988 * in a cache, in which case all caches updated before it, will
2989 * have a bigger array.
2991 * But if that is the case, the data after
2992 * memcg_limited_groups_array_size is certainly unused
2994 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
2995 if (!cur_params
->memcg_caches
[i
])
2997 s
->memcg_params
->memcg_caches
[i
] =
2998 cur_params
->memcg_caches
[i
];
3002 * Ideally, we would wait until all caches succeed, and only
3003 * then free the old one. But this is not worth the extra
3004 * pointer per-cache we'd have to have for this.
3006 * It is not a big deal if some caches are left with a size
3007 * bigger than the others. And all updates will reset this
3015 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
)
3017 size_t size
= sizeof(struct memcg_cache_params
);
3019 if (!memcg_kmem_enabled())
3023 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3025 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3026 if (!s
->memcg_params
)
3030 s
->memcg_params
->memcg
= memcg
;
3034 void memcg_release_cache(struct kmem_cache
*s
)
3036 struct kmem_cache
*root
;
3037 struct mem_cgroup
*memcg
;
3041 * This happens, for instance, when a root cache goes away before we
3044 if (!s
->memcg_params
)
3047 if (s
->memcg_params
->is_root_cache
)
3050 memcg
= s
->memcg_params
->memcg
;
3051 id
= memcg_cache_id(memcg
);
3053 root
= s
->memcg_params
->root_cache
;
3054 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3055 mem_cgroup_put(memcg
);
3057 mutex_lock(&memcg
->slab_caches_mutex
);
3058 list_del(&s
->memcg_params
->list
);
3059 mutex_unlock(&memcg
->slab_caches_mutex
);
3062 kfree(s
->memcg_params
);
3066 * During the creation a new cache, we need to disable our accounting mechanism
3067 * altogether. This is true even if we are not creating, but rather just
3068 * enqueing new caches to be created.
3070 * This is because that process will trigger allocations; some visible, like
3071 * explicit kmallocs to auxiliary data structures, name strings and internal
3072 * cache structures; some well concealed, like INIT_WORK() that can allocate
3073 * objects during debug.
3075 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3076 * to it. This may not be a bounded recursion: since the first cache creation
3077 * failed to complete (waiting on the allocation), we'll just try to create the
3078 * cache again, failing at the same point.
3080 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3081 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3082 * inside the following two functions.
3084 static inline void memcg_stop_kmem_account(void)
3086 VM_BUG_ON(!current
->mm
);
3087 current
->memcg_kmem_skip_account
++;
3090 static inline void memcg_resume_kmem_account(void)
3092 VM_BUG_ON(!current
->mm
);
3093 current
->memcg_kmem_skip_account
--;
3096 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3098 struct kmem_cache
*cachep
;
3099 struct memcg_cache_params
*p
;
3101 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3103 cachep
= memcg_params_to_cache(p
);
3106 * If we get down to 0 after shrink, we could delete right away.
3107 * However, memcg_release_pages() already puts us back in the workqueue
3108 * in that case. If we proceed deleting, we'll get a dangling
3109 * reference, and removing the object from the workqueue in that case
3110 * is unnecessary complication. We are not a fast path.
3112 * Note that this case is fundamentally different from racing with
3113 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3114 * kmem_cache_shrink, not only we would be reinserting a dead cache
3115 * into the queue, but doing so from inside the worker racing to
3118 * So if we aren't down to zero, we'll just schedule a worker and try
3121 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3122 kmem_cache_shrink(cachep
);
3123 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3126 kmem_cache_destroy(cachep
);
3129 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3131 if (!cachep
->memcg_params
->dead
)
3135 * There are many ways in which we can get here.
3137 * We can get to a memory-pressure situation while the delayed work is
3138 * still pending to run. The vmscan shrinkers can then release all
3139 * cache memory and get us to destruction. If this is the case, we'll
3140 * be executed twice, which is a bug (the second time will execute over
3141 * bogus data). In this case, cancelling the work should be fine.
3143 * But we can also get here from the worker itself, if
3144 * kmem_cache_shrink is enough to shake all the remaining objects and
3145 * get the page count to 0. In this case, we'll deadlock if we try to
3146 * cancel the work (the worker runs with an internal lock held, which
3147 * is the same lock we would hold for cancel_work_sync().)
3149 * Since we can't possibly know who got us here, just refrain from
3150 * running if there is already work pending
3152 if (work_pending(&cachep
->memcg_params
->destroy
))
3155 * We have to defer the actual destroying to a workqueue, because
3156 * we might currently be in a context that cannot sleep.
3158 schedule_work(&cachep
->memcg_params
->destroy
);
3161 static char *memcg_cache_name(struct mem_cgroup
*memcg
, struct kmem_cache
*s
)
3164 struct dentry
*dentry
;
3167 dentry
= rcu_dereference(memcg
->css
.cgroup
->dentry
);
3170 BUG_ON(dentry
== NULL
);
3172 name
= kasprintf(GFP_KERNEL
, "%s(%d:%s)", s
->name
,
3173 memcg_cache_id(memcg
), dentry
->d_name
.name
);
3178 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3179 struct kmem_cache
*s
)
3182 struct kmem_cache
*new;
3184 name
= memcg_cache_name(memcg
, s
);
3188 new = kmem_cache_create_memcg(memcg
, name
, s
->object_size
, s
->align
,
3189 (s
->flags
& ~SLAB_PANIC
), s
->ctor
);
3192 new->allocflags
|= __GFP_KMEMCG
;
3199 * This lock protects updaters, not readers. We want readers to be as fast as
3200 * they can, and they will either see NULL or a valid cache value. Our model
3201 * allow them to see NULL, in which case the root memcg will be selected.
3203 * We need this lock because multiple allocations to the same cache from a non
3204 * will span more than one worker. Only one of them can create the cache.
3206 static DEFINE_MUTEX(memcg_cache_mutex
);
3207 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3208 struct kmem_cache
*cachep
)
3210 struct kmem_cache
*new_cachep
;
3213 BUG_ON(!memcg_can_account_kmem(memcg
));
3215 idx
= memcg_cache_id(memcg
);
3217 mutex_lock(&memcg_cache_mutex
);
3218 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3222 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3223 if (new_cachep
== NULL
) {
3224 new_cachep
= cachep
;
3228 mem_cgroup_get(memcg
);
3229 new_cachep
->memcg_params
->root_cache
= cachep
;
3230 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3232 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3234 * the readers won't lock, make sure everybody sees the updated value,
3235 * so they won't put stuff in the queue again for no reason
3239 mutex_unlock(&memcg_cache_mutex
);
3243 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3245 struct kmem_cache
*c
;
3248 if (!s
->memcg_params
)
3250 if (!s
->memcg_params
->is_root_cache
)
3254 * If the cache is being destroyed, we trust that there is no one else
3255 * requesting objects from it. Even if there are, the sanity checks in
3256 * kmem_cache_destroy should caught this ill-case.
3258 * Still, we don't want anyone else freeing memcg_caches under our
3259 * noses, which can happen if a new memcg comes to life. As usual,
3260 * we'll take the set_limit_mutex to protect ourselves against this.
3262 mutex_lock(&set_limit_mutex
);
3263 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3264 c
= s
->memcg_params
->memcg_caches
[i
];
3269 * We will now manually delete the caches, so to avoid races
3270 * we need to cancel all pending destruction workers and
3271 * proceed with destruction ourselves.
3273 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3274 * and that could spawn the workers again: it is likely that
3275 * the cache still have active pages until this very moment.
3276 * This would lead us back to mem_cgroup_destroy_cache.
3278 * But that will not execute at all if the "dead" flag is not
3279 * set, so flip it down to guarantee we are in control.
3281 c
->memcg_params
->dead
= false;
3282 cancel_work_sync(&c
->memcg_params
->destroy
);
3283 kmem_cache_destroy(c
);
3285 mutex_unlock(&set_limit_mutex
);
3288 struct create_work
{
3289 struct mem_cgroup
*memcg
;
3290 struct kmem_cache
*cachep
;
3291 struct work_struct work
;
3294 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3296 struct kmem_cache
*cachep
;
3297 struct memcg_cache_params
*params
;
3299 if (!memcg_kmem_is_active(memcg
))
3302 mutex_lock(&memcg
->slab_caches_mutex
);
3303 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3304 cachep
= memcg_params_to_cache(params
);
3305 cachep
->memcg_params
->dead
= true;
3306 INIT_WORK(&cachep
->memcg_params
->destroy
,
3307 kmem_cache_destroy_work_func
);
3308 schedule_work(&cachep
->memcg_params
->destroy
);
3310 mutex_unlock(&memcg
->slab_caches_mutex
);
3313 static void memcg_create_cache_work_func(struct work_struct
*w
)
3315 struct create_work
*cw
;
3317 cw
= container_of(w
, struct create_work
, work
);
3318 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3319 /* Drop the reference gotten when we enqueued. */
3320 css_put(&cw
->memcg
->css
);
3325 * Enqueue the creation of a per-memcg kmem_cache.
3326 * Called with rcu_read_lock.
3328 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3329 struct kmem_cache
*cachep
)
3331 struct create_work
*cw
;
3333 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3337 /* The corresponding put will be done in the workqueue. */
3338 if (!css_tryget(&memcg
->css
)) {
3344 cw
->cachep
= cachep
;
3346 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3347 schedule_work(&cw
->work
);
3350 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3351 struct kmem_cache
*cachep
)
3354 * We need to stop accounting when we kmalloc, because if the
3355 * corresponding kmalloc cache is not yet created, the first allocation
3356 * in __memcg_create_cache_enqueue will recurse.
3358 * However, it is better to enclose the whole function. Depending on
3359 * the debugging options enabled, INIT_WORK(), for instance, can
3360 * trigger an allocation. This too, will make us recurse. Because at
3361 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3362 * the safest choice is to do it like this, wrapping the whole function.
3364 memcg_stop_kmem_account();
3365 __memcg_create_cache_enqueue(memcg
, cachep
);
3366 memcg_resume_kmem_account();
3369 * Return the kmem_cache we're supposed to use for a slab allocation.
3370 * We try to use the current memcg's version of the cache.
3372 * If the cache does not exist yet, if we are the first user of it,
3373 * we either create it immediately, if possible, or create it asynchronously
3375 * In the latter case, we will let the current allocation go through with
3376 * the original cache.
3378 * Can't be called in interrupt context or from kernel threads.
3379 * This function needs to be called with rcu_read_lock() held.
3381 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3384 struct mem_cgroup
*memcg
;
3387 VM_BUG_ON(!cachep
->memcg_params
);
3388 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3390 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3394 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3397 if (!memcg_can_account_kmem(memcg
))
3400 idx
= memcg_cache_id(memcg
);
3403 * barrier to mare sure we're always seeing the up to date value. The
3404 * code updating memcg_caches will issue a write barrier to match this.
3406 read_barrier_depends();
3407 if (unlikely(cachep
->memcg_params
->memcg_caches
[idx
] == NULL
)) {
3409 * If we are in a safe context (can wait, and not in interrupt
3410 * context), we could be be predictable and return right away.
3411 * This would guarantee that the allocation being performed
3412 * already belongs in the new cache.
3414 * However, there are some clashes that can arrive from locking.
3415 * For instance, because we acquire the slab_mutex while doing
3416 * kmem_cache_dup, this means no further allocation could happen
3417 * with the slab_mutex held.
3419 * Also, because cache creation issue get_online_cpus(), this
3420 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3421 * that ends up reversed during cpu hotplug. (cpuset allocates
3422 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3423 * better to defer everything.
3425 memcg_create_cache_enqueue(memcg
, cachep
);
3429 return cachep
->memcg_params
->memcg_caches
[idx
];
3431 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3434 * We need to verify if the allocation against current->mm->owner's memcg is
3435 * possible for the given order. But the page is not allocated yet, so we'll
3436 * need a further commit step to do the final arrangements.
3438 * It is possible for the task to switch cgroups in this mean time, so at
3439 * commit time, we can't rely on task conversion any longer. We'll then use
3440 * the handle argument to return to the caller which cgroup we should commit
3441 * against. We could also return the memcg directly and avoid the pointer
3442 * passing, but a boolean return value gives better semantics considering
3443 * the compiled-out case as well.
3445 * Returning true means the allocation is possible.
3448 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3450 struct mem_cgroup
*memcg
;
3454 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3457 * very rare case described in mem_cgroup_from_task. Unfortunately there
3458 * isn't much we can do without complicating this too much, and it would
3459 * be gfp-dependent anyway. Just let it go
3461 if (unlikely(!memcg
))
3464 if (!memcg_can_account_kmem(memcg
)) {
3465 css_put(&memcg
->css
);
3469 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3473 css_put(&memcg
->css
);
3477 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3480 struct page_cgroup
*pc
;
3482 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3484 /* The page allocation failed. Revert */
3486 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3490 pc
= lookup_page_cgroup(page
);
3491 lock_page_cgroup(pc
);
3492 pc
->mem_cgroup
= memcg
;
3493 SetPageCgroupUsed(pc
);
3494 unlock_page_cgroup(pc
);
3497 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3499 struct mem_cgroup
*memcg
= NULL
;
3500 struct page_cgroup
*pc
;
3503 pc
= lookup_page_cgroup(page
);
3505 * Fast unlocked return. Theoretically might have changed, have to
3506 * check again after locking.
3508 if (!PageCgroupUsed(pc
))
3511 lock_page_cgroup(pc
);
3512 if (PageCgroupUsed(pc
)) {
3513 memcg
= pc
->mem_cgroup
;
3514 ClearPageCgroupUsed(pc
);
3516 unlock_page_cgroup(pc
);
3519 * We trust that only if there is a memcg associated with the page, it
3520 * is a valid allocation
3525 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3526 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3529 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3532 #endif /* CONFIG_MEMCG_KMEM */
3534 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3536 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3538 * Because tail pages are not marked as "used", set it. We're under
3539 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3540 * charge/uncharge will be never happen and move_account() is done under
3541 * compound_lock(), so we don't have to take care of races.
3543 void mem_cgroup_split_huge_fixup(struct page
*head
)
3545 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3546 struct page_cgroup
*pc
;
3549 if (mem_cgroup_disabled())
3551 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3553 pc
->mem_cgroup
= head_pc
->mem_cgroup
;
3554 smp_wmb();/* see __commit_charge() */
3555 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3558 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3561 * mem_cgroup_move_account - move account of the page
3563 * @nr_pages: number of regular pages (>1 for huge pages)
3564 * @pc: page_cgroup of the page.
3565 * @from: mem_cgroup which the page is moved from.
3566 * @to: mem_cgroup which the page is moved to. @from != @to.
3568 * The caller must confirm following.
3569 * - page is not on LRU (isolate_page() is useful.)
3570 * - compound_lock is held when nr_pages > 1
3572 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3575 static int mem_cgroup_move_account(struct page
*page
,
3576 unsigned int nr_pages
,
3577 struct page_cgroup
*pc
,
3578 struct mem_cgroup
*from
,
3579 struct mem_cgroup
*to
)
3581 unsigned long flags
;
3583 bool anon
= PageAnon(page
);
3585 VM_BUG_ON(from
== to
);
3586 VM_BUG_ON(PageLRU(page
));
3588 * The page is isolated from LRU. So, collapse function
3589 * will not handle this page. But page splitting can happen.
3590 * Do this check under compound_page_lock(). The caller should
3594 if (nr_pages
> 1 && !PageTransHuge(page
))
3597 lock_page_cgroup(pc
);
3600 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3603 move_lock_mem_cgroup(from
, &flags
);
3605 if (!anon
&& page_mapped(page
)) {
3606 /* Update mapped_file data for mem_cgroup */
3608 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3609 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3612 mem_cgroup_charge_statistics(from
, anon
, -nr_pages
);
3614 /* caller should have done css_get */
3615 pc
->mem_cgroup
= to
;
3616 mem_cgroup_charge_statistics(to
, anon
, nr_pages
);
3617 move_unlock_mem_cgroup(from
, &flags
);
3620 unlock_page_cgroup(pc
);
3624 memcg_check_events(to
, page
);
3625 memcg_check_events(from
, page
);
3631 * mem_cgroup_move_parent - moves page to the parent group
3632 * @page: the page to move
3633 * @pc: page_cgroup of the page
3634 * @child: page's cgroup
3636 * move charges to its parent or the root cgroup if the group has no
3637 * parent (aka use_hierarchy==0).
3638 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3639 * mem_cgroup_move_account fails) the failure is always temporary and
3640 * it signals a race with a page removal/uncharge or migration. In the
3641 * first case the page is on the way out and it will vanish from the LRU
3642 * on the next attempt and the call should be retried later.
3643 * Isolation from the LRU fails only if page has been isolated from
3644 * the LRU since we looked at it and that usually means either global
3645 * reclaim or migration going on. The page will either get back to the
3647 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3648 * (!PageCgroupUsed) or moved to a different group. The page will
3649 * disappear in the next attempt.
3651 static int mem_cgroup_move_parent(struct page
*page
,
3652 struct page_cgroup
*pc
,
3653 struct mem_cgroup
*child
)
3655 struct mem_cgroup
*parent
;
3656 unsigned int nr_pages
;
3657 unsigned long uninitialized_var(flags
);
3660 VM_BUG_ON(mem_cgroup_is_root(child
));
3663 if (!get_page_unless_zero(page
))
3665 if (isolate_lru_page(page
))
3668 nr_pages
= hpage_nr_pages(page
);
3670 parent
= parent_mem_cgroup(child
);
3672 * If no parent, move charges to root cgroup.
3675 parent
= root_mem_cgroup
;
3678 VM_BUG_ON(!PageTransHuge(page
));
3679 flags
= compound_lock_irqsave(page
);
3682 ret
= mem_cgroup_move_account(page
, nr_pages
,
3685 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3688 compound_unlock_irqrestore(page
, flags
);
3689 putback_lru_page(page
);
3697 * Charge the memory controller for page usage.
3699 * 0 if the charge was successful
3700 * < 0 if the cgroup is over its limit
3702 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3703 gfp_t gfp_mask
, enum charge_type ctype
)
3705 struct mem_cgroup
*memcg
= NULL
;
3706 unsigned int nr_pages
= 1;
3710 if (PageTransHuge(page
)) {
3711 nr_pages
<<= compound_order(page
);
3712 VM_BUG_ON(!PageTransHuge(page
));
3714 * Never OOM-kill a process for a huge page. The
3715 * fault handler will fall back to regular pages.
3720 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3723 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3727 int mem_cgroup_newpage_charge(struct page
*page
,
3728 struct mm_struct
*mm
, gfp_t gfp_mask
)
3730 if (mem_cgroup_disabled())
3732 VM_BUG_ON(page_mapped(page
));
3733 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3735 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3736 MEM_CGROUP_CHARGE_TYPE_ANON
);
3740 * While swap-in, try_charge -> commit or cancel, the page is locked.
3741 * And when try_charge() successfully returns, one refcnt to memcg without
3742 * struct page_cgroup is acquired. This refcnt will be consumed by
3743 * "commit()" or removed by "cancel()"
3745 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3748 struct mem_cgroup
**memcgp
)
3750 struct mem_cgroup
*memcg
;
3751 struct page_cgroup
*pc
;
3754 pc
= lookup_page_cgroup(page
);
3756 * Every swap fault against a single page tries to charge the
3757 * page, bail as early as possible. shmem_unuse() encounters
3758 * already charged pages, too. The USED bit is protected by
3759 * the page lock, which serializes swap cache removal, which
3760 * in turn serializes uncharging.
3762 if (PageCgroupUsed(pc
))
3764 if (!do_swap_account
)
3766 memcg
= try_get_mem_cgroup_from_page(page
);
3770 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3771 css_put(&memcg
->css
);
3776 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3782 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3783 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3786 if (mem_cgroup_disabled())
3789 * A racing thread's fault, or swapoff, may have already
3790 * updated the pte, and even removed page from swap cache: in
3791 * those cases unuse_pte()'s pte_same() test will fail; but
3792 * there's also a KSM case which does need to charge the page.
3794 if (!PageSwapCache(page
)) {
3797 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3802 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3805 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3807 if (mem_cgroup_disabled())
3811 __mem_cgroup_cancel_charge(memcg
, 1);
3815 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3816 enum charge_type ctype
)
3818 if (mem_cgroup_disabled())
3823 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3825 * Now swap is on-memory. This means this page may be
3826 * counted both as mem and swap....double count.
3827 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3828 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3829 * may call delete_from_swap_cache() before reach here.
3831 if (do_swap_account
&& PageSwapCache(page
)) {
3832 swp_entry_t ent
= {.val
= page_private(page
)};
3833 mem_cgroup_uncharge_swap(ent
);
3837 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3838 struct mem_cgroup
*memcg
)
3840 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3841 MEM_CGROUP_CHARGE_TYPE_ANON
);
3844 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3847 struct mem_cgroup
*memcg
= NULL
;
3848 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3851 if (mem_cgroup_disabled())
3853 if (PageCompound(page
))
3856 if (!PageSwapCache(page
))
3857 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3858 else { /* page is swapcache/shmem */
3859 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3862 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3867 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3868 unsigned int nr_pages
,
3869 const enum charge_type ctype
)
3871 struct memcg_batch_info
*batch
= NULL
;
3872 bool uncharge_memsw
= true;
3874 /* If swapout, usage of swap doesn't decrease */
3875 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3876 uncharge_memsw
= false;
3878 batch
= ¤t
->memcg_batch
;
3880 * In usual, we do css_get() when we remember memcg pointer.
3881 * But in this case, we keep res->usage until end of a series of
3882 * uncharges. Then, it's ok to ignore memcg's refcnt.
3885 batch
->memcg
= memcg
;
3887 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3888 * In those cases, all pages freed continuously can be expected to be in
3889 * the same cgroup and we have chance to coalesce uncharges.
3890 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3891 * because we want to do uncharge as soon as possible.
3894 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3895 goto direct_uncharge
;
3898 goto direct_uncharge
;
3901 * In typical case, batch->memcg == mem. This means we can
3902 * merge a series of uncharges to an uncharge of res_counter.
3903 * If not, we uncharge res_counter ony by one.
3905 if (batch
->memcg
!= memcg
)
3906 goto direct_uncharge
;
3907 /* remember freed charge and uncharge it later */
3910 batch
->memsw_nr_pages
++;
3913 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3915 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3916 if (unlikely(batch
->memcg
!= memcg
))
3917 memcg_oom_recover(memcg
);
3921 * uncharge if !page_mapped(page)
3923 static struct mem_cgroup
*
3924 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3927 struct mem_cgroup
*memcg
= NULL
;
3928 unsigned int nr_pages
= 1;
3929 struct page_cgroup
*pc
;
3932 if (mem_cgroup_disabled())
3935 VM_BUG_ON(PageSwapCache(page
));
3937 if (PageTransHuge(page
)) {
3938 nr_pages
<<= compound_order(page
);
3939 VM_BUG_ON(!PageTransHuge(page
));
3942 * Check if our page_cgroup is valid
3944 pc
= lookup_page_cgroup(page
);
3945 if (unlikely(!PageCgroupUsed(pc
)))
3948 lock_page_cgroup(pc
);
3950 memcg
= pc
->mem_cgroup
;
3952 if (!PageCgroupUsed(pc
))
3955 anon
= PageAnon(page
);
3958 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3960 * Generally PageAnon tells if it's the anon statistics to be
3961 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3962 * used before page reached the stage of being marked PageAnon.
3966 case MEM_CGROUP_CHARGE_TYPE_DROP
:
3967 /* See mem_cgroup_prepare_migration() */
3968 if (page_mapped(page
))
3971 * Pages under migration may not be uncharged. But
3972 * end_migration() /must/ be the one uncharging the
3973 * unused post-migration page and so it has to call
3974 * here with the migration bit still set. See the
3975 * res_counter handling below.
3977 if (!end_migration
&& PageCgroupMigration(pc
))
3980 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
3981 if (!PageAnon(page
)) { /* Shared memory */
3982 if (page
->mapping
&& !page_is_file_cache(page
))
3984 } else if (page_mapped(page
)) /* Anon */
3991 mem_cgroup_charge_statistics(memcg
, anon
, -nr_pages
);
3993 ClearPageCgroupUsed(pc
);
3995 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3996 * freed from LRU. This is safe because uncharged page is expected not
3997 * to be reused (freed soon). Exception is SwapCache, it's handled by
3998 * special functions.
4001 unlock_page_cgroup(pc
);
4003 * even after unlock, we have memcg->res.usage here and this memcg
4004 * will never be freed.
4006 memcg_check_events(memcg
, page
);
4007 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4008 mem_cgroup_swap_statistics(memcg
, true);
4009 mem_cgroup_get(memcg
);
4012 * Migration does not charge the res_counter for the
4013 * replacement page, so leave it alone when phasing out the
4014 * page that is unused after the migration.
4016 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4017 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4022 unlock_page_cgroup(pc
);
4026 void mem_cgroup_uncharge_page(struct page
*page
)
4029 if (page_mapped(page
))
4031 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4032 if (PageSwapCache(page
))
4034 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4037 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4039 VM_BUG_ON(page_mapped(page
));
4040 VM_BUG_ON(page
->mapping
);
4041 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4045 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4046 * In that cases, pages are freed continuously and we can expect pages
4047 * are in the same memcg. All these calls itself limits the number of
4048 * pages freed at once, then uncharge_start/end() is called properly.
4049 * This may be called prural(2) times in a context,
4052 void mem_cgroup_uncharge_start(void)
4054 current
->memcg_batch
.do_batch
++;
4055 /* We can do nest. */
4056 if (current
->memcg_batch
.do_batch
== 1) {
4057 current
->memcg_batch
.memcg
= NULL
;
4058 current
->memcg_batch
.nr_pages
= 0;
4059 current
->memcg_batch
.memsw_nr_pages
= 0;
4063 void mem_cgroup_uncharge_end(void)
4065 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4067 if (!batch
->do_batch
)
4071 if (batch
->do_batch
) /* If stacked, do nothing. */
4077 * This "batch->memcg" is valid without any css_get/put etc...
4078 * bacause we hide charges behind us.
4080 if (batch
->nr_pages
)
4081 res_counter_uncharge(&batch
->memcg
->res
,
4082 batch
->nr_pages
* PAGE_SIZE
);
4083 if (batch
->memsw_nr_pages
)
4084 res_counter_uncharge(&batch
->memcg
->memsw
,
4085 batch
->memsw_nr_pages
* PAGE_SIZE
);
4086 memcg_oom_recover(batch
->memcg
);
4087 /* forget this pointer (for sanity check) */
4088 batch
->memcg
= NULL
;
4093 * called after __delete_from_swap_cache() and drop "page" account.
4094 * memcg information is recorded to swap_cgroup of "ent"
4097 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4099 struct mem_cgroup
*memcg
;
4100 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4102 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4103 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4105 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4108 * record memcg information, if swapout && memcg != NULL,
4109 * mem_cgroup_get() was called in uncharge().
4111 if (do_swap_account
&& swapout
&& memcg
)
4112 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4116 #ifdef CONFIG_MEMCG_SWAP
4118 * called from swap_entry_free(). remove record in swap_cgroup and
4119 * uncharge "memsw" account.
4121 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4123 struct mem_cgroup
*memcg
;
4126 if (!do_swap_account
)
4129 id
= swap_cgroup_record(ent
, 0);
4131 memcg
= mem_cgroup_lookup(id
);
4134 * We uncharge this because swap is freed.
4135 * This memcg can be obsolete one. We avoid calling css_tryget
4137 if (!mem_cgroup_is_root(memcg
))
4138 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4139 mem_cgroup_swap_statistics(memcg
, false);
4140 mem_cgroup_put(memcg
);
4146 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4147 * @entry: swap entry to be moved
4148 * @from: mem_cgroup which the entry is moved from
4149 * @to: mem_cgroup which the entry is moved to
4151 * It succeeds only when the swap_cgroup's record for this entry is the same
4152 * as the mem_cgroup's id of @from.
4154 * Returns 0 on success, -EINVAL on failure.
4156 * The caller must have charged to @to, IOW, called res_counter_charge() about
4157 * both res and memsw, and called css_get().
4159 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4160 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4162 unsigned short old_id
, new_id
;
4164 old_id
= css_id(&from
->css
);
4165 new_id
= css_id(&to
->css
);
4167 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4168 mem_cgroup_swap_statistics(from
, false);
4169 mem_cgroup_swap_statistics(to
, true);
4171 * This function is only called from task migration context now.
4172 * It postpones res_counter and refcount handling till the end
4173 * of task migration(mem_cgroup_clear_mc()) for performance
4174 * improvement. But we cannot postpone mem_cgroup_get(to)
4175 * because if the process that has been moved to @to does
4176 * swap-in, the refcount of @to might be decreased to 0.
4184 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4185 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4192 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4195 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4196 struct mem_cgroup
**memcgp
)
4198 struct mem_cgroup
*memcg
= NULL
;
4199 unsigned int nr_pages
= 1;
4200 struct page_cgroup
*pc
;
4201 enum charge_type ctype
;
4205 if (mem_cgroup_disabled())
4208 if (PageTransHuge(page
))
4209 nr_pages
<<= compound_order(page
);
4211 pc
= lookup_page_cgroup(page
);
4212 lock_page_cgroup(pc
);
4213 if (PageCgroupUsed(pc
)) {
4214 memcg
= pc
->mem_cgroup
;
4215 css_get(&memcg
->css
);
4217 * At migrating an anonymous page, its mapcount goes down
4218 * to 0 and uncharge() will be called. But, even if it's fully
4219 * unmapped, migration may fail and this page has to be
4220 * charged again. We set MIGRATION flag here and delay uncharge
4221 * until end_migration() is called
4223 * Corner Case Thinking
4225 * When the old page was mapped as Anon and it's unmap-and-freed
4226 * while migration was ongoing.
4227 * If unmap finds the old page, uncharge() of it will be delayed
4228 * until end_migration(). If unmap finds a new page, it's
4229 * uncharged when it make mapcount to be 1->0. If unmap code
4230 * finds swap_migration_entry, the new page will not be mapped
4231 * and end_migration() will find it(mapcount==0).
4234 * When the old page was mapped but migraion fails, the kernel
4235 * remaps it. A charge for it is kept by MIGRATION flag even
4236 * if mapcount goes down to 0. We can do remap successfully
4237 * without charging it again.
4240 * The "old" page is under lock_page() until the end of
4241 * migration, so, the old page itself will not be swapped-out.
4242 * If the new page is swapped out before end_migraton, our
4243 * hook to usual swap-out path will catch the event.
4246 SetPageCgroupMigration(pc
);
4248 unlock_page_cgroup(pc
);
4250 * If the page is not charged at this point,
4258 * We charge new page before it's used/mapped. So, even if unlock_page()
4259 * is called before end_migration, we can catch all events on this new
4260 * page. In the case new page is migrated but not remapped, new page's
4261 * mapcount will be finally 0 and we call uncharge in end_migration().
4264 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4266 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4268 * The page is committed to the memcg, but it's not actually
4269 * charged to the res_counter since we plan on replacing the
4270 * old one and only one page is going to be left afterwards.
4272 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4275 /* remove redundant charge if migration failed*/
4276 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4277 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4279 struct page
*used
, *unused
;
4280 struct page_cgroup
*pc
;
4286 if (!migration_ok
) {
4293 anon
= PageAnon(used
);
4294 __mem_cgroup_uncharge_common(unused
,
4295 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4296 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4298 css_put(&memcg
->css
);
4300 * We disallowed uncharge of pages under migration because mapcount
4301 * of the page goes down to zero, temporarly.
4302 * Clear the flag and check the page should be charged.
4304 pc
= lookup_page_cgroup(oldpage
);
4305 lock_page_cgroup(pc
);
4306 ClearPageCgroupMigration(pc
);
4307 unlock_page_cgroup(pc
);
4310 * If a page is a file cache, radix-tree replacement is very atomic
4311 * and we can skip this check. When it was an Anon page, its mapcount
4312 * goes down to 0. But because we added MIGRATION flage, it's not
4313 * uncharged yet. There are several case but page->mapcount check
4314 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4315 * check. (see prepare_charge() also)
4318 mem_cgroup_uncharge_page(used
);
4322 * At replace page cache, newpage is not under any memcg but it's on
4323 * LRU. So, this function doesn't touch res_counter but handles LRU
4324 * in correct way. Both pages are locked so we cannot race with uncharge.
4326 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4327 struct page
*newpage
)
4329 struct mem_cgroup
*memcg
= NULL
;
4330 struct page_cgroup
*pc
;
4331 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4333 if (mem_cgroup_disabled())
4336 pc
= lookup_page_cgroup(oldpage
);
4337 /* fix accounting on old pages */
4338 lock_page_cgroup(pc
);
4339 if (PageCgroupUsed(pc
)) {
4340 memcg
= pc
->mem_cgroup
;
4341 mem_cgroup_charge_statistics(memcg
, false, -1);
4342 ClearPageCgroupUsed(pc
);
4344 unlock_page_cgroup(pc
);
4347 * When called from shmem_replace_page(), in some cases the
4348 * oldpage has already been charged, and in some cases not.
4353 * Even if newpage->mapping was NULL before starting replacement,
4354 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4355 * LRU while we overwrite pc->mem_cgroup.
4357 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4360 #ifdef CONFIG_DEBUG_VM
4361 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4363 struct page_cgroup
*pc
;
4365 pc
= lookup_page_cgroup(page
);
4367 * Can be NULL while feeding pages into the page allocator for
4368 * the first time, i.e. during boot or memory hotplug;
4369 * or when mem_cgroup_disabled().
4371 if (likely(pc
) && PageCgroupUsed(pc
))
4376 bool mem_cgroup_bad_page_check(struct page
*page
)
4378 if (mem_cgroup_disabled())
4381 return lookup_page_cgroup_used(page
) != NULL
;
4384 void mem_cgroup_print_bad_page(struct page
*page
)
4386 struct page_cgroup
*pc
;
4388 pc
= lookup_page_cgroup_used(page
);
4390 printk(KERN_ALERT
"pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4391 pc
, pc
->flags
, pc
->mem_cgroup
);
4396 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4397 unsigned long long val
)
4400 u64 memswlimit
, memlimit
;
4402 int children
= mem_cgroup_count_children(memcg
);
4403 u64 curusage
, oldusage
;
4407 * For keeping hierarchical_reclaim simple, how long we should retry
4408 * is depends on callers. We set our retry-count to be function
4409 * of # of children which we should visit in this loop.
4411 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4413 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4416 while (retry_count
) {
4417 if (signal_pending(current
)) {
4422 * Rather than hide all in some function, I do this in
4423 * open coded manner. You see what this really does.
4424 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4426 mutex_lock(&set_limit_mutex
);
4427 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4428 if (memswlimit
< val
) {
4430 mutex_unlock(&set_limit_mutex
);
4434 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4438 ret
= res_counter_set_limit(&memcg
->res
, val
);
4440 if (memswlimit
== val
)
4441 memcg
->memsw_is_minimum
= true;
4443 memcg
->memsw_is_minimum
= false;
4445 mutex_unlock(&set_limit_mutex
);
4450 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4451 MEM_CGROUP_RECLAIM_SHRINK
);
4452 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4453 /* Usage is reduced ? */
4454 if (curusage
>= oldusage
)
4457 oldusage
= curusage
;
4459 if (!ret
&& enlarge
)
4460 memcg_oom_recover(memcg
);
4465 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4466 unsigned long long val
)
4469 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4470 int children
= mem_cgroup_count_children(memcg
);
4474 /* see mem_cgroup_resize_res_limit */
4475 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4476 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4477 while (retry_count
) {
4478 if (signal_pending(current
)) {
4483 * Rather than hide all in some function, I do this in
4484 * open coded manner. You see what this really does.
4485 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4487 mutex_lock(&set_limit_mutex
);
4488 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4489 if (memlimit
> val
) {
4491 mutex_unlock(&set_limit_mutex
);
4494 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4495 if (memswlimit
< val
)
4497 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4499 if (memlimit
== val
)
4500 memcg
->memsw_is_minimum
= true;
4502 memcg
->memsw_is_minimum
= false;
4504 mutex_unlock(&set_limit_mutex
);
4509 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4510 MEM_CGROUP_RECLAIM_NOSWAP
|
4511 MEM_CGROUP_RECLAIM_SHRINK
);
4512 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4513 /* Usage is reduced ? */
4514 if (curusage
>= oldusage
)
4517 oldusage
= curusage
;
4519 if (!ret
&& enlarge
)
4520 memcg_oom_recover(memcg
);
4524 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4526 unsigned long *total_scanned
)
4528 unsigned long nr_reclaimed
= 0;
4529 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4530 unsigned long reclaimed
;
4532 struct mem_cgroup_tree_per_zone
*mctz
;
4533 unsigned long long excess
;
4534 unsigned long nr_scanned
;
4539 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4541 * This loop can run a while, specially if mem_cgroup's continuously
4542 * keep exceeding their soft limit and putting the system under
4549 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4554 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4555 gfp_mask
, &nr_scanned
);
4556 nr_reclaimed
+= reclaimed
;
4557 *total_scanned
+= nr_scanned
;
4558 spin_lock(&mctz
->lock
);
4561 * If we failed to reclaim anything from this memory cgroup
4562 * it is time to move on to the next cgroup
4568 * Loop until we find yet another one.
4570 * By the time we get the soft_limit lock
4571 * again, someone might have aded the
4572 * group back on the RB tree. Iterate to
4573 * make sure we get a different mem.
4574 * mem_cgroup_largest_soft_limit_node returns
4575 * NULL if no other cgroup is present on
4579 __mem_cgroup_largest_soft_limit_node(mctz
);
4581 css_put(&next_mz
->memcg
->css
);
4582 else /* next_mz == NULL or other memcg */
4586 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4587 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4589 * One school of thought says that we should not add
4590 * back the node to the tree if reclaim returns 0.
4591 * But our reclaim could return 0, simply because due
4592 * to priority we are exposing a smaller subset of
4593 * memory to reclaim from. Consider this as a longer
4596 /* If excess == 0, no tree ops */
4597 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4598 spin_unlock(&mctz
->lock
);
4599 css_put(&mz
->memcg
->css
);
4602 * Could not reclaim anything and there are no more
4603 * mem cgroups to try or we seem to be looping without
4604 * reclaiming anything.
4606 if (!nr_reclaimed
&&
4608 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4610 } while (!nr_reclaimed
);
4612 css_put(&next_mz
->memcg
->css
);
4613 return nr_reclaimed
;
4617 * mem_cgroup_force_empty_list - clears LRU of a group
4618 * @memcg: group to clear
4621 * @lru: lru to to clear
4623 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4624 * reclaim the pages page themselves - pages are moved to the parent (or root)
4627 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4628 int node
, int zid
, enum lru_list lru
)
4630 struct lruvec
*lruvec
;
4631 unsigned long flags
;
4632 struct list_head
*list
;
4636 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4637 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4638 list
= &lruvec
->lists
[lru
];
4642 struct page_cgroup
*pc
;
4645 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4646 if (list_empty(list
)) {
4647 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4650 page
= list_entry(list
->prev
, struct page
, lru
);
4652 list_move(&page
->lru
, list
);
4654 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4657 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4659 pc
= lookup_page_cgroup(page
);
4661 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4662 /* found lock contention or "pc" is obsolete. */
4667 } while (!list_empty(list
));
4671 * make mem_cgroup's charge to be 0 if there is no task by moving
4672 * all the charges and pages to the parent.
4673 * This enables deleting this mem_cgroup.
4675 * Caller is responsible for holding css reference on the memcg.
4677 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4683 /* This is for making all *used* pages to be on LRU. */
4684 lru_add_drain_all();
4685 drain_all_stock_sync(memcg
);
4686 mem_cgroup_start_move(memcg
);
4687 for_each_node_state(node
, N_MEMORY
) {
4688 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4691 mem_cgroup_force_empty_list(memcg
,
4696 mem_cgroup_end_move(memcg
);
4697 memcg_oom_recover(memcg
);
4701 * Kernel memory may not necessarily be trackable to a specific
4702 * process. So they are not migrated, and therefore we can't
4703 * expect their value to drop to 0 here.
4704 * Having res filled up with kmem only is enough.
4706 * This is a safety check because mem_cgroup_force_empty_list
4707 * could have raced with mem_cgroup_replace_page_cache callers
4708 * so the lru seemed empty but the page could have been added
4709 * right after the check. RES_USAGE should be safe as we always
4710 * charge before adding to the LRU.
4712 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4713 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4714 } while (usage
> 0);
4718 * Reclaims as many pages from the given memcg as possible and moves
4719 * the rest to the parent.
4721 * Caller is responsible for holding css reference for memcg.
4723 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4725 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4726 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4728 /* returns EBUSY if there is a task or if we come here twice. */
4729 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4732 /* we call try-to-free pages for make this cgroup empty */
4733 lru_add_drain_all();
4734 /* try to free all pages in this cgroup */
4735 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4738 if (signal_pending(current
))
4741 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4745 /* maybe some writeback is necessary */
4746 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4751 mem_cgroup_reparent_charges(memcg
);
4756 static int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
4758 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4761 if (mem_cgroup_is_root(memcg
))
4763 css_get(&memcg
->css
);
4764 ret
= mem_cgroup_force_empty(memcg
);
4765 css_put(&memcg
->css
);
4771 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
4773 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
4776 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
4780 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4781 struct cgroup
*parent
= cont
->parent
;
4782 struct mem_cgroup
*parent_memcg
= NULL
;
4785 parent_memcg
= mem_cgroup_from_cont(parent
);
4789 if (memcg
->use_hierarchy
== val
)
4793 * If parent's use_hierarchy is set, we can't make any modifications
4794 * in the child subtrees. If it is unset, then the change can
4795 * occur, provided the current cgroup has no children.
4797 * For the root cgroup, parent_mem is NULL, we allow value to be
4798 * set if there are no children.
4800 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4801 (val
== 1 || val
== 0)) {
4802 if (list_empty(&cont
->children
))
4803 memcg
->use_hierarchy
= val
;
4816 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4817 enum mem_cgroup_stat_index idx
)
4819 struct mem_cgroup
*iter
;
4822 /* Per-cpu values can be negative, use a signed accumulator */
4823 for_each_mem_cgroup_tree(iter
, memcg
)
4824 val
+= mem_cgroup_read_stat(iter
, idx
);
4826 if (val
< 0) /* race ? */
4831 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4835 if (!mem_cgroup_is_root(memcg
)) {
4837 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4839 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4842 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4843 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4846 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4848 return val
<< PAGE_SHIFT
;
4851 static ssize_t
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
,
4852 struct file
*file
, char __user
*buf
,
4853 size_t nbytes
, loff_t
*ppos
)
4855 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4861 type
= MEMFILE_TYPE(cft
->private);
4862 name
= MEMFILE_ATTR(cft
->private);
4864 if (!do_swap_account
&& type
== _MEMSWAP
)
4869 if (name
== RES_USAGE
)
4870 val
= mem_cgroup_usage(memcg
, false);
4872 val
= res_counter_read_u64(&memcg
->res
, name
);
4875 if (name
== RES_USAGE
)
4876 val
= mem_cgroup_usage(memcg
, true);
4878 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4881 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4887 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4888 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4891 static int memcg_update_kmem_limit(struct cgroup
*cont
, u64 val
)
4894 #ifdef CONFIG_MEMCG_KMEM
4895 bool must_inc_static_branch
= false;
4897 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4899 * For simplicity, we won't allow this to be disabled. It also can't
4900 * be changed if the cgroup has children already, or if tasks had
4903 * If tasks join before we set the limit, a person looking at
4904 * kmem.usage_in_bytes will have no way to determine when it took
4905 * place, which makes the value quite meaningless.
4907 * After it first became limited, changes in the value of the limit are
4908 * of course permitted.
4910 * Taking the cgroup_lock is really offensive, but it is so far the only
4911 * way to guarantee that no children will appear. There are plenty of
4912 * other offenders, and they should all go away. Fine grained locking
4913 * is probably the way to go here. When we are fully hierarchical, we
4914 * can also get rid of the use_hierarchy check.
4917 mutex_lock(&set_limit_mutex
);
4918 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4919 if (cgroup_task_count(cont
) || (memcg
->use_hierarchy
&&
4920 !list_empty(&cont
->children
))) {
4924 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4927 ret
= memcg_update_cache_sizes(memcg
);
4929 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4932 must_inc_static_branch
= true;
4934 * kmem charges can outlive the cgroup. In the case of slab
4935 * pages, for instance, a page contain objects from various
4936 * processes, so it is unfeasible to migrate them away. We
4937 * need to reference count the memcg because of that.
4939 mem_cgroup_get(memcg
);
4941 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4943 mutex_unlock(&set_limit_mutex
);
4947 * We are by now familiar with the fact that we can't inc the static
4948 * branch inside cgroup_lock. See disarm functions for details. A
4949 * worker here is overkill, but also wrong: After the limit is set, we
4950 * must start accounting right away. Since this operation can't fail,
4951 * we can safely defer it to here - no rollback will be needed.
4953 * The boolean used to control this is also safe, because
4954 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4955 * able to set it to true;
4957 if (must_inc_static_branch
) {
4958 static_key_slow_inc(&memcg_kmem_enabled_key
);
4960 * setting the active bit after the inc will guarantee no one
4961 * starts accounting before all call sites are patched
4963 memcg_kmem_set_active(memcg
);
4970 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4973 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4977 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
4978 #ifdef CONFIG_MEMCG_KMEM
4980 * When that happen, we need to disable the static branch only on those
4981 * memcgs that enabled it. To achieve this, we would be forced to
4982 * complicate the code by keeping track of which memcgs were the ones
4983 * that actually enabled limits, and which ones got it from its
4986 * It is a lot simpler just to do static_key_slow_inc() on every child
4987 * that is accounted.
4989 if (!memcg_kmem_is_active(memcg
))
4993 * destroy(), called if we fail, will issue static_key_slow_inc() and
4994 * mem_cgroup_put() if kmem is enabled. We have to either call them
4995 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
4996 * this more consistent, since it always leads to the same destroy path
4998 mem_cgroup_get(memcg
);
4999 static_key_slow_inc(&memcg_kmem_enabled_key
);
5001 mutex_lock(&set_limit_mutex
);
5002 ret
= memcg_update_cache_sizes(memcg
);
5003 mutex_unlock(&set_limit_mutex
);
5010 * The user of this function is...
5013 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
5016 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5019 unsigned long long val
;
5022 type
= MEMFILE_TYPE(cft
->private);
5023 name
= MEMFILE_ATTR(cft
->private);
5025 if (!do_swap_account
&& type
== _MEMSWAP
)
5030 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5034 /* This function does all necessary parse...reuse it */
5035 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5039 ret
= mem_cgroup_resize_limit(memcg
, val
);
5040 else if (type
== _MEMSWAP
)
5041 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5042 else if (type
== _KMEM
)
5043 ret
= memcg_update_kmem_limit(cont
, val
);
5047 case RES_SOFT_LIMIT
:
5048 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5052 * For memsw, soft limits are hard to implement in terms
5053 * of semantics, for now, we support soft limits for
5054 * control without swap
5057 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5062 ret
= -EINVAL
; /* should be BUG() ? */
5068 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5069 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5071 struct cgroup
*cgroup
;
5072 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5074 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5075 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5076 cgroup
= memcg
->css
.cgroup
;
5077 if (!memcg
->use_hierarchy
)
5080 while (cgroup
->parent
) {
5081 cgroup
= cgroup
->parent
;
5082 memcg
= mem_cgroup_from_cont(cgroup
);
5083 if (!memcg
->use_hierarchy
)
5085 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5086 min_limit
= min(min_limit
, tmp
);
5087 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5088 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5091 *mem_limit
= min_limit
;
5092 *memsw_limit
= min_memsw_limit
;
5095 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
5097 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5101 type
= MEMFILE_TYPE(event
);
5102 name
= MEMFILE_ATTR(event
);
5104 if (!do_swap_account
&& type
== _MEMSWAP
)
5110 res_counter_reset_max(&memcg
->res
);
5111 else if (type
== _MEMSWAP
)
5112 res_counter_reset_max(&memcg
->memsw
);
5113 else if (type
== _KMEM
)
5114 res_counter_reset_max(&memcg
->kmem
);
5120 res_counter_reset_failcnt(&memcg
->res
);
5121 else if (type
== _MEMSWAP
)
5122 res_counter_reset_failcnt(&memcg
->memsw
);
5123 else if (type
== _KMEM
)
5124 res_counter_reset_failcnt(&memcg
->kmem
);
5133 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
5136 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
5140 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5141 struct cftype
*cft
, u64 val
)
5143 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5145 if (val
>= (1 << NR_MOVE_TYPE
))
5148 * We check this value several times in both in can_attach() and
5149 * attach(), so we need cgroup lock to prevent this value from being
5153 memcg
->move_charge_at_immigrate
= val
;
5159 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5160 struct cftype
*cft
, u64 val
)
5167 static int memcg_numa_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5171 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5172 unsigned long node_nr
;
5173 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5175 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5176 seq_printf(m
, "total=%lu", total_nr
);
5177 for_each_node_state(nid
, N_MEMORY
) {
5178 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5179 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5183 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5184 seq_printf(m
, "file=%lu", file_nr
);
5185 for_each_node_state(nid
, N_MEMORY
) {
5186 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5188 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5192 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5193 seq_printf(m
, "anon=%lu", anon_nr
);
5194 for_each_node_state(nid
, N_MEMORY
) {
5195 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5197 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5201 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5202 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5203 for_each_node_state(nid
, N_MEMORY
) {
5204 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5205 BIT(LRU_UNEVICTABLE
));
5206 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5211 #endif /* CONFIG_NUMA */
5213 static const char * const mem_cgroup_lru_names
[] = {
5221 static inline void mem_cgroup_lru_names_not_uptodate(void)
5223 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5226 static int memcg_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5229 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5230 struct mem_cgroup
*mi
;
5233 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5234 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5236 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5237 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5240 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5241 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5242 mem_cgroup_read_events(memcg
, i
));
5244 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5245 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5246 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5248 /* Hierarchical information */
5250 unsigned long long limit
, memsw_limit
;
5251 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5252 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5253 if (do_swap_account
)
5254 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5258 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5261 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5263 for_each_mem_cgroup_tree(mi
, memcg
)
5264 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5265 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5268 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5269 unsigned long long val
= 0;
5271 for_each_mem_cgroup_tree(mi
, memcg
)
5272 val
+= mem_cgroup_read_events(mi
, i
);
5273 seq_printf(m
, "total_%s %llu\n",
5274 mem_cgroup_events_names
[i
], val
);
5277 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5278 unsigned long long val
= 0;
5280 for_each_mem_cgroup_tree(mi
, memcg
)
5281 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5282 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5285 #ifdef CONFIG_DEBUG_VM
5288 struct mem_cgroup_per_zone
*mz
;
5289 struct zone_reclaim_stat
*rstat
;
5290 unsigned long recent_rotated
[2] = {0, 0};
5291 unsigned long recent_scanned
[2] = {0, 0};
5293 for_each_online_node(nid
)
5294 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5295 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5296 rstat
= &mz
->lruvec
.reclaim_stat
;
5298 recent_rotated
[0] += rstat
->recent_rotated
[0];
5299 recent_rotated
[1] += rstat
->recent_rotated
[1];
5300 recent_scanned
[0] += rstat
->recent_scanned
[0];
5301 recent_scanned
[1] += rstat
->recent_scanned
[1];
5303 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5304 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5305 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5306 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5313 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
5315 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5317 return mem_cgroup_swappiness(memcg
);
5320 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
5323 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5324 struct mem_cgroup
*parent
;
5329 if (cgrp
->parent
== NULL
)
5332 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5336 /* If under hierarchy, only empty-root can set this value */
5337 if ((parent
->use_hierarchy
) ||
5338 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
5343 memcg
->swappiness
= val
;
5350 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5352 struct mem_cgroup_threshold_ary
*t
;
5358 t
= rcu_dereference(memcg
->thresholds
.primary
);
5360 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5365 usage
= mem_cgroup_usage(memcg
, swap
);
5368 * current_threshold points to threshold just below or equal to usage.
5369 * If it's not true, a threshold was crossed after last
5370 * call of __mem_cgroup_threshold().
5372 i
= t
->current_threshold
;
5375 * Iterate backward over array of thresholds starting from
5376 * current_threshold and check if a threshold is crossed.
5377 * If none of thresholds below usage is crossed, we read
5378 * only one element of the array here.
5380 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5381 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5383 /* i = current_threshold + 1 */
5387 * Iterate forward over array of thresholds starting from
5388 * current_threshold+1 and check if a threshold is crossed.
5389 * If none of thresholds above usage is crossed, we read
5390 * only one element of the array here.
5392 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5393 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5395 /* Update current_threshold */
5396 t
->current_threshold
= i
- 1;
5401 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5404 __mem_cgroup_threshold(memcg
, false);
5405 if (do_swap_account
)
5406 __mem_cgroup_threshold(memcg
, true);
5408 memcg
= parent_mem_cgroup(memcg
);
5412 static int compare_thresholds(const void *a
, const void *b
)
5414 const struct mem_cgroup_threshold
*_a
= a
;
5415 const struct mem_cgroup_threshold
*_b
= b
;
5417 return _a
->threshold
- _b
->threshold
;
5420 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5422 struct mem_cgroup_eventfd_list
*ev
;
5424 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5425 eventfd_signal(ev
->eventfd
, 1);
5429 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5431 struct mem_cgroup
*iter
;
5433 for_each_mem_cgroup_tree(iter
, memcg
)
5434 mem_cgroup_oom_notify_cb(iter
);
5437 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5438 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5440 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5441 struct mem_cgroup_thresholds
*thresholds
;
5442 struct mem_cgroup_threshold_ary
*new;
5443 enum res_type type
= MEMFILE_TYPE(cft
->private);
5444 u64 threshold
, usage
;
5447 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5451 mutex_lock(&memcg
->thresholds_lock
);
5454 thresholds
= &memcg
->thresholds
;
5455 else if (type
== _MEMSWAP
)
5456 thresholds
= &memcg
->memsw_thresholds
;
5460 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5462 /* Check if a threshold crossed before adding a new one */
5463 if (thresholds
->primary
)
5464 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5466 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5468 /* Allocate memory for new array of thresholds */
5469 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5477 /* Copy thresholds (if any) to new array */
5478 if (thresholds
->primary
) {
5479 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5480 sizeof(struct mem_cgroup_threshold
));
5483 /* Add new threshold */
5484 new->entries
[size
- 1].eventfd
= eventfd
;
5485 new->entries
[size
- 1].threshold
= threshold
;
5487 /* Sort thresholds. Registering of new threshold isn't time-critical */
5488 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5489 compare_thresholds
, NULL
);
5491 /* Find current threshold */
5492 new->current_threshold
= -1;
5493 for (i
= 0; i
< size
; i
++) {
5494 if (new->entries
[i
].threshold
<= usage
) {
5496 * new->current_threshold will not be used until
5497 * rcu_assign_pointer(), so it's safe to increment
5500 ++new->current_threshold
;
5505 /* Free old spare buffer and save old primary buffer as spare */
5506 kfree(thresholds
->spare
);
5507 thresholds
->spare
= thresholds
->primary
;
5509 rcu_assign_pointer(thresholds
->primary
, new);
5511 /* To be sure that nobody uses thresholds */
5515 mutex_unlock(&memcg
->thresholds_lock
);
5520 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5521 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5523 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5524 struct mem_cgroup_thresholds
*thresholds
;
5525 struct mem_cgroup_threshold_ary
*new;
5526 enum res_type type
= MEMFILE_TYPE(cft
->private);
5530 mutex_lock(&memcg
->thresholds_lock
);
5532 thresholds
= &memcg
->thresholds
;
5533 else if (type
== _MEMSWAP
)
5534 thresholds
= &memcg
->memsw_thresholds
;
5538 if (!thresholds
->primary
)
5541 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5543 /* Check if a threshold crossed before removing */
5544 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5546 /* Calculate new number of threshold */
5548 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5549 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5553 new = thresholds
->spare
;
5555 /* Set thresholds array to NULL if we don't have thresholds */
5564 /* Copy thresholds and find current threshold */
5565 new->current_threshold
= -1;
5566 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5567 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5570 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5571 if (new->entries
[j
].threshold
<= usage
) {
5573 * new->current_threshold will not be used
5574 * until rcu_assign_pointer(), so it's safe to increment
5577 ++new->current_threshold
;
5583 /* Swap primary and spare array */
5584 thresholds
->spare
= thresholds
->primary
;
5585 /* If all events are unregistered, free the spare array */
5587 kfree(thresholds
->spare
);
5588 thresholds
->spare
= NULL
;
5591 rcu_assign_pointer(thresholds
->primary
, new);
5593 /* To be sure that nobody uses thresholds */
5596 mutex_unlock(&memcg
->thresholds_lock
);
5599 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5600 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5602 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5603 struct mem_cgroup_eventfd_list
*event
;
5604 enum res_type type
= MEMFILE_TYPE(cft
->private);
5606 BUG_ON(type
!= _OOM_TYPE
);
5607 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5611 spin_lock(&memcg_oom_lock
);
5613 event
->eventfd
= eventfd
;
5614 list_add(&event
->list
, &memcg
->oom_notify
);
5616 /* already in OOM ? */
5617 if (atomic_read(&memcg
->under_oom
))
5618 eventfd_signal(eventfd
, 1);
5619 spin_unlock(&memcg_oom_lock
);
5624 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5625 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5627 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5628 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5629 enum res_type type
= MEMFILE_TYPE(cft
->private);
5631 BUG_ON(type
!= _OOM_TYPE
);
5633 spin_lock(&memcg_oom_lock
);
5635 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5636 if (ev
->eventfd
== eventfd
) {
5637 list_del(&ev
->list
);
5642 spin_unlock(&memcg_oom_lock
);
5645 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
5646 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5648 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5650 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5652 if (atomic_read(&memcg
->under_oom
))
5653 cb
->fill(cb
, "under_oom", 1);
5655 cb
->fill(cb
, "under_oom", 0);
5659 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
5660 struct cftype
*cft
, u64 val
)
5662 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5663 struct mem_cgroup
*parent
;
5665 /* cannot set to root cgroup and only 0 and 1 are allowed */
5666 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
5669 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5672 /* oom-kill-disable is a flag for subhierarchy. */
5673 if ((parent
->use_hierarchy
) ||
5674 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
5678 memcg
->oom_kill_disable
= val
;
5680 memcg_oom_recover(memcg
);
5685 #ifdef CONFIG_MEMCG_KMEM
5686 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5690 memcg
->kmemcg_id
= -1;
5691 ret
= memcg_propagate_kmem(memcg
);
5695 return mem_cgroup_sockets_init(memcg
, ss
);
5698 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5700 mem_cgroup_sockets_destroy(memcg
);
5702 memcg_kmem_mark_dead(memcg
);
5704 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5708 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5709 * path here, being careful not to race with memcg_uncharge_kmem: it is
5710 * possible that the charges went down to 0 between mark_dead and the
5711 * res_counter read, so in that case, we don't need the put
5713 if (memcg_kmem_test_and_clear_dead(memcg
))
5714 mem_cgroup_put(memcg
);
5717 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5722 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5727 static struct cftype mem_cgroup_files
[] = {
5729 .name
= "usage_in_bytes",
5730 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5731 .read
= mem_cgroup_read
,
5732 .register_event
= mem_cgroup_usage_register_event
,
5733 .unregister_event
= mem_cgroup_usage_unregister_event
,
5736 .name
= "max_usage_in_bytes",
5737 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5738 .trigger
= mem_cgroup_reset
,
5739 .read
= mem_cgroup_read
,
5742 .name
= "limit_in_bytes",
5743 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5744 .write_string
= mem_cgroup_write
,
5745 .read
= mem_cgroup_read
,
5748 .name
= "soft_limit_in_bytes",
5749 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5750 .write_string
= mem_cgroup_write
,
5751 .read
= mem_cgroup_read
,
5755 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5756 .trigger
= mem_cgroup_reset
,
5757 .read
= mem_cgroup_read
,
5761 .read_seq_string
= memcg_stat_show
,
5764 .name
= "force_empty",
5765 .trigger
= mem_cgroup_force_empty_write
,
5768 .name
= "use_hierarchy",
5769 .write_u64
= mem_cgroup_hierarchy_write
,
5770 .read_u64
= mem_cgroup_hierarchy_read
,
5773 .name
= "swappiness",
5774 .read_u64
= mem_cgroup_swappiness_read
,
5775 .write_u64
= mem_cgroup_swappiness_write
,
5778 .name
= "move_charge_at_immigrate",
5779 .read_u64
= mem_cgroup_move_charge_read
,
5780 .write_u64
= mem_cgroup_move_charge_write
,
5783 .name
= "oom_control",
5784 .read_map
= mem_cgroup_oom_control_read
,
5785 .write_u64
= mem_cgroup_oom_control_write
,
5786 .register_event
= mem_cgroup_oom_register_event
,
5787 .unregister_event
= mem_cgroup_oom_unregister_event
,
5788 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5792 .name
= "numa_stat",
5793 .read_seq_string
= memcg_numa_stat_show
,
5796 #ifdef CONFIG_MEMCG_SWAP
5798 .name
= "memsw.usage_in_bytes",
5799 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5800 .read
= mem_cgroup_read
,
5801 .register_event
= mem_cgroup_usage_register_event
,
5802 .unregister_event
= mem_cgroup_usage_unregister_event
,
5805 .name
= "memsw.max_usage_in_bytes",
5806 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5807 .trigger
= mem_cgroup_reset
,
5808 .read
= mem_cgroup_read
,
5811 .name
= "memsw.limit_in_bytes",
5812 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5813 .write_string
= mem_cgroup_write
,
5814 .read
= mem_cgroup_read
,
5817 .name
= "memsw.failcnt",
5818 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5819 .trigger
= mem_cgroup_reset
,
5820 .read
= mem_cgroup_read
,
5823 #ifdef CONFIG_MEMCG_KMEM
5825 .name
= "kmem.limit_in_bytes",
5826 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5827 .write_string
= mem_cgroup_write
,
5828 .read
= mem_cgroup_read
,
5831 .name
= "kmem.usage_in_bytes",
5832 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5833 .read
= mem_cgroup_read
,
5836 .name
= "kmem.failcnt",
5837 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5838 .trigger
= mem_cgroup_reset
,
5839 .read
= mem_cgroup_read
,
5842 .name
= "kmem.max_usage_in_bytes",
5843 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5844 .trigger
= mem_cgroup_reset
,
5845 .read
= mem_cgroup_read
,
5847 #ifdef CONFIG_SLABINFO
5849 .name
= "kmem.slabinfo",
5850 .read_seq_string
= mem_cgroup_slabinfo_read
,
5854 { }, /* terminate */
5857 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5859 struct mem_cgroup_per_node
*pn
;
5860 struct mem_cgroup_per_zone
*mz
;
5861 int zone
, tmp
= node
;
5863 * This routine is called against possible nodes.
5864 * But it's BUG to call kmalloc() against offline node.
5866 * TODO: this routine can waste much memory for nodes which will
5867 * never be onlined. It's better to use memory hotplug callback
5870 if (!node_state(node
, N_NORMAL_MEMORY
))
5872 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5876 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5877 mz
= &pn
->zoneinfo
[zone
];
5878 lruvec_init(&mz
->lruvec
);
5879 mz
->usage_in_excess
= 0;
5880 mz
->on_tree
= false;
5883 memcg
->info
.nodeinfo
[node
] = pn
;
5887 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5889 kfree(memcg
->info
.nodeinfo
[node
]);
5892 static struct mem_cgroup
*mem_cgroup_alloc(void)
5894 struct mem_cgroup
*memcg
;
5895 int size
= sizeof(struct mem_cgroup
);
5897 /* Can be very big if MAX_NUMNODES is very big */
5898 if (size
< PAGE_SIZE
)
5899 memcg
= kzalloc(size
, GFP_KERNEL
);
5901 memcg
= vzalloc(size
);
5906 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5909 spin_lock_init(&memcg
->pcp_counter_lock
);
5913 if (size
< PAGE_SIZE
)
5921 * At destroying mem_cgroup, references from swap_cgroup can remain.
5922 * (scanning all at force_empty is too costly...)
5924 * Instead of clearing all references at force_empty, we remember
5925 * the number of reference from swap_cgroup and free mem_cgroup when
5926 * it goes down to 0.
5928 * Removal of cgroup itself succeeds regardless of refs from swap.
5931 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5934 int size
= sizeof(struct mem_cgroup
);
5936 mem_cgroup_remove_from_trees(memcg
);
5937 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5940 free_mem_cgroup_per_zone_info(memcg
, node
);
5942 free_percpu(memcg
->stat
);
5945 * We need to make sure that (at least for now), the jump label
5946 * destruction code runs outside of the cgroup lock. This is because
5947 * get_online_cpus(), which is called from the static_branch update,
5948 * can't be called inside the cgroup_lock. cpusets are the ones
5949 * enforcing this dependency, so if they ever change, we might as well.
5951 * schedule_work() will guarantee this happens. Be careful if you need
5952 * to move this code around, and make sure it is outside
5955 disarm_static_keys(memcg
);
5956 if (size
< PAGE_SIZE
)
5964 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5965 * but in process context. The work_freeing structure is overlaid
5966 * on the rcu_freeing structure, which itself is overlaid on memsw.
5968 static void free_work(struct work_struct
*work
)
5970 struct mem_cgroup
*memcg
;
5972 memcg
= container_of(work
, struct mem_cgroup
, work_freeing
);
5973 __mem_cgroup_free(memcg
);
5976 static void free_rcu(struct rcu_head
*rcu_head
)
5978 struct mem_cgroup
*memcg
;
5980 memcg
= container_of(rcu_head
, struct mem_cgroup
, rcu_freeing
);
5981 INIT_WORK(&memcg
->work_freeing
, free_work
);
5982 schedule_work(&memcg
->work_freeing
);
5985 static void mem_cgroup_get(struct mem_cgroup
*memcg
)
5987 atomic_inc(&memcg
->refcnt
);
5990 static void __mem_cgroup_put(struct mem_cgroup
*memcg
, int count
)
5992 if (atomic_sub_and_test(count
, &memcg
->refcnt
)) {
5993 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5994 call_rcu(&memcg
->rcu_freeing
, free_rcu
);
5996 mem_cgroup_put(parent
);
6000 static void mem_cgroup_put(struct mem_cgroup
*memcg
)
6002 __mem_cgroup_put(memcg
, 1);
6006 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6008 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6010 if (!memcg
->res
.parent
)
6012 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6014 EXPORT_SYMBOL(parent_mem_cgroup
);
6016 #ifdef CONFIG_MEMCG_SWAP
6017 static void __init
enable_swap_cgroup(void)
6019 if (!mem_cgroup_disabled() && really_do_swap_account
)
6020 do_swap_account
= 1;
6023 static void __init
enable_swap_cgroup(void)
6028 static int mem_cgroup_soft_limit_tree_init(void)
6030 struct mem_cgroup_tree_per_node
*rtpn
;
6031 struct mem_cgroup_tree_per_zone
*rtpz
;
6032 int tmp
, node
, zone
;
6034 for_each_node(node
) {
6036 if (!node_state(node
, N_NORMAL_MEMORY
))
6038 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6042 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6044 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6045 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6046 rtpz
->rb_root
= RB_ROOT
;
6047 spin_lock_init(&rtpz
->lock
);
6053 for_each_node(node
) {
6054 if (!soft_limit_tree
.rb_tree_per_node
[node
])
6056 kfree(soft_limit_tree
.rb_tree_per_node
[node
]);
6057 soft_limit_tree
.rb_tree_per_node
[node
] = NULL
;
6063 static struct cgroup_subsys_state
* __ref
6064 mem_cgroup_css_alloc(struct cgroup
*cont
)
6066 struct mem_cgroup
*memcg
, *parent
;
6067 long error
= -ENOMEM
;
6070 memcg
= mem_cgroup_alloc();
6072 return ERR_PTR(error
);
6075 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6079 if (cont
->parent
== NULL
) {
6081 enable_swap_cgroup();
6083 if (mem_cgroup_soft_limit_tree_init())
6085 root_mem_cgroup
= memcg
;
6086 for_each_possible_cpu(cpu
) {
6087 struct memcg_stock_pcp
*stock
=
6088 &per_cpu(memcg_stock
, cpu
);
6089 INIT_WORK(&stock
->work
, drain_local_stock
);
6091 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6093 parent
= mem_cgroup_from_cont(cont
->parent
);
6094 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6095 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6098 if (parent
&& parent
->use_hierarchy
) {
6099 res_counter_init(&memcg
->res
, &parent
->res
);
6100 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6101 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6104 * We increment refcnt of the parent to ensure that we can
6105 * safely access it on res_counter_charge/uncharge.
6106 * This refcnt will be decremented when freeing this
6107 * mem_cgroup(see mem_cgroup_put).
6109 mem_cgroup_get(parent
);
6111 res_counter_init(&memcg
->res
, NULL
);
6112 res_counter_init(&memcg
->memsw
, NULL
);
6113 res_counter_init(&memcg
->kmem
, NULL
);
6115 * Deeper hierachy with use_hierarchy == false doesn't make
6116 * much sense so let cgroup subsystem know about this
6117 * unfortunate state in our controller.
6119 if (parent
&& parent
!= root_mem_cgroup
)
6120 mem_cgroup_subsys
.broken_hierarchy
= true;
6122 memcg
->last_scanned_node
= MAX_NUMNODES
;
6123 INIT_LIST_HEAD(&memcg
->oom_notify
);
6126 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6127 atomic_set(&memcg
->refcnt
, 1);
6128 memcg
->move_charge_at_immigrate
= 0;
6129 mutex_init(&memcg
->thresholds_lock
);
6130 spin_lock_init(&memcg
->move_lock
);
6132 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6135 * We call put now because our (and parent's) refcnts
6136 * are already in place. mem_cgroup_put() will internally
6137 * call __mem_cgroup_free, so return directly
6139 mem_cgroup_put(memcg
);
6140 return ERR_PTR(error
);
6144 __mem_cgroup_free(memcg
);
6145 return ERR_PTR(error
);
6148 static void mem_cgroup_css_offline(struct cgroup
*cont
)
6150 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6152 mem_cgroup_reparent_charges(memcg
);
6153 mem_cgroup_destroy_all_caches(memcg
);
6156 static void mem_cgroup_css_free(struct cgroup
*cont
)
6158 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6160 kmem_cgroup_destroy(memcg
);
6162 mem_cgroup_put(memcg
);
6166 /* Handlers for move charge at task migration. */
6167 #define PRECHARGE_COUNT_AT_ONCE 256
6168 static int mem_cgroup_do_precharge(unsigned long count
)
6171 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6172 struct mem_cgroup
*memcg
= mc
.to
;
6174 if (mem_cgroup_is_root(memcg
)) {
6175 mc
.precharge
+= count
;
6176 /* we don't need css_get for root */
6179 /* try to charge at once */
6181 struct res_counter
*dummy
;
6183 * "memcg" cannot be under rmdir() because we've already checked
6184 * by cgroup_lock_live_cgroup() that it is not removed and we
6185 * are still under the same cgroup_mutex. So we can postpone
6188 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6190 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6191 PAGE_SIZE
* count
, &dummy
)) {
6192 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6195 mc
.precharge
+= count
;
6199 /* fall back to one by one charge */
6201 if (signal_pending(current
)) {
6205 if (!batch_count
--) {
6206 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6209 ret
= __mem_cgroup_try_charge(NULL
,
6210 GFP_KERNEL
, 1, &memcg
, false);
6212 /* mem_cgroup_clear_mc() will do uncharge later */
6220 * get_mctgt_type - get target type of moving charge
6221 * @vma: the vma the pte to be checked belongs
6222 * @addr: the address corresponding to the pte to be checked
6223 * @ptent: the pte to be checked
6224 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6227 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6228 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6229 * move charge. if @target is not NULL, the page is stored in target->page
6230 * with extra refcnt got(Callers should handle it).
6231 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6232 * target for charge migration. if @target is not NULL, the entry is stored
6235 * Called with pte lock held.
6242 enum mc_target_type
{
6248 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6249 unsigned long addr
, pte_t ptent
)
6251 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6253 if (!page
|| !page_mapped(page
))
6255 if (PageAnon(page
)) {
6256 /* we don't move shared anon */
6259 } else if (!move_file())
6260 /* we ignore mapcount for file pages */
6262 if (!get_page_unless_zero(page
))
6269 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6270 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6272 struct page
*page
= NULL
;
6273 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6275 if (!move_anon() || non_swap_entry(ent
))
6278 * Because lookup_swap_cache() updates some statistics counter,
6279 * we call find_get_page() with swapper_space directly.
6281 page
= find_get_page(&swapper_space
, ent
.val
);
6282 if (do_swap_account
)
6283 entry
->val
= ent
.val
;
6288 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6289 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6295 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6296 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6298 struct page
*page
= NULL
;
6299 struct address_space
*mapping
;
6302 if (!vma
->vm_file
) /* anonymous vma */
6307 mapping
= vma
->vm_file
->f_mapping
;
6308 if (pte_none(ptent
))
6309 pgoff
= linear_page_index(vma
, addr
);
6310 else /* pte_file(ptent) is true */
6311 pgoff
= pte_to_pgoff(ptent
);
6313 /* page is moved even if it's not RSS of this task(page-faulted). */
6314 page
= find_get_page(mapping
, pgoff
);
6317 /* shmem/tmpfs may report page out on swap: account for that too. */
6318 if (radix_tree_exceptional_entry(page
)) {
6319 swp_entry_t swap
= radix_to_swp_entry(page
);
6320 if (do_swap_account
)
6322 page
= find_get_page(&swapper_space
, swap
.val
);
6328 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6329 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6331 struct page
*page
= NULL
;
6332 struct page_cgroup
*pc
;
6333 enum mc_target_type ret
= MC_TARGET_NONE
;
6334 swp_entry_t ent
= { .val
= 0 };
6336 if (pte_present(ptent
))
6337 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6338 else if (is_swap_pte(ptent
))
6339 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6340 else if (pte_none(ptent
) || pte_file(ptent
))
6341 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6343 if (!page
&& !ent
.val
)
6346 pc
= lookup_page_cgroup(page
);
6348 * Do only loose check w/o page_cgroup lock.
6349 * mem_cgroup_move_account() checks the pc is valid or not under
6352 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6353 ret
= MC_TARGET_PAGE
;
6355 target
->page
= page
;
6357 if (!ret
|| !target
)
6360 /* There is a swap entry and a page doesn't exist or isn't charged */
6361 if (ent
.val
&& !ret
&&
6362 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6363 ret
= MC_TARGET_SWAP
;
6370 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6372 * We don't consider swapping or file mapped pages because THP does not
6373 * support them for now.
6374 * Caller should make sure that pmd_trans_huge(pmd) is true.
6376 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6377 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6379 struct page
*page
= NULL
;
6380 struct page_cgroup
*pc
;
6381 enum mc_target_type ret
= MC_TARGET_NONE
;
6383 page
= pmd_page(pmd
);
6384 VM_BUG_ON(!page
|| !PageHead(page
));
6387 pc
= lookup_page_cgroup(page
);
6388 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6389 ret
= MC_TARGET_PAGE
;
6392 target
->page
= page
;
6398 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6399 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6401 return MC_TARGET_NONE
;
6405 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6406 unsigned long addr
, unsigned long end
,
6407 struct mm_walk
*walk
)
6409 struct vm_area_struct
*vma
= walk
->private;
6413 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6414 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6415 mc
.precharge
+= HPAGE_PMD_NR
;
6416 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6420 if (pmd_trans_unstable(pmd
))
6422 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6423 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6424 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6425 mc
.precharge
++; /* increment precharge temporarily */
6426 pte_unmap_unlock(pte
- 1, ptl
);
6432 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6434 unsigned long precharge
;
6435 struct vm_area_struct
*vma
;
6437 down_read(&mm
->mmap_sem
);
6438 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6439 struct mm_walk mem_cgroup_count_precharge_walk
= {
6440 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6444 if (is_vm_hugetlb_page(vma
))
6446 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6447 &mem_cgroup_count_precharge_walk
);
6449 up_read(&mm
->mmap_sem
);
6451 precharge
= mc
.precharge
;
6457 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6459 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6461 VM_BUG_ON(mc
.moving_task
);
6462 mc
.moving_task
= current
;
6463 return mem_cgroup_do_precharge(precharge
);
6466 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6467 static void __mem_cgroup_clear_mc(void)
6469 struct mem_cgroup
*from
= mc
.from
;
6470 struct mem_cgroup
*to
= mc
.to
;
6472 /* we must uncharge all the leftover precharges from mc.to */
6474 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6478 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6479 * we must uncharge here.
6481 if (mc
.moved_charge
) {
6482 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6483 mc
.moved_charge
= 0;
6485 /* we must fixup refcnts and charges */
6486 if (mc
.moved_swap
) {
6487 /* uncharge swap account from the old cgroup */
6488 if (!mem_cgroup_is_root(mc
.from
))
6489 res_counter_uncharge(&mc
.from
->memsw
,
6490 PAGE_SIZE
* mc
.moved_swap
);
6491 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
6493 if (!mem_cgroup_is_root(mc
.to
)) {
6495 * we charged both to->res and to->memsw, so we should
6498 res_counter_uncharge(&mc
.to
->res
,
6499 PAGE_SIZE
* mc
.moved_swap
);
6501 /* we've already done mem_cgroup_get(mc.to) */
6504 memcg_oom_recover(from
);
6505 memcg_oom_recover(to
);
6506 wake_up_all(&mc
.waitq
);
6509 static void mem_cgroup_clear_mc(void)
6511 struct mem_cgroup
*from
= mc
.from
;
6514 * we must clear moving_task before waking up waiters at the end of
6517 mc
.moving_task
= NULL
;
6518 __mem_cgroup_clear_mc();
6519 spin_lock(&mc
.lock
);
6522 spin_unlock(&mc
.lock
);
6523 mem_cgroup_end_move(from
);
6526 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6527 struct cgroup_taskset
*tset
)
6529 struct task_struct
*p
= cgroup_taskset_first(tset
);
6531 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgroup
);
6533 if (memcg
->move_charge_at_immigrate
) {
6534 struct mm_struct
*mm
;
6535 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6537 VM_BUG_ON(from
== memcg
);
6539 mm
= get_task_mm(p
);
6542 /* We move charges only when we move a owner of the mm */
6543 if (mm
->owner
== p
) {
6546 VM_BUG_ON(mc
.precharge
);
6547 VM_BUG_ON(mc
.moved_charge
);
6548 VM_BUG_ON(mc
.moved_swap
);
6549 mem_cgroup_start_move(from
);
6550 spin_lock(&mc
.lock
);
6553 spin_unlock(&mc
.lock
);
6554 /* We set mc.moving_task later */
6556 ret
= mem_cgroup_precharge_mc(mm
);
6558 mem_cgroup_clear_mc();
6565 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6566 struct cgroup_taskset
*tset
)
6568 mem_cgroup_clear_mc();
6571 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6572 unsigned long addr
, unsigned long end
,
6573 struct mm_walk
*walk
)
6576 struct vm_area_struct
*vma
= walk
->private;
6579 enum mc_target_type target_type
;
6580 union mc_target target
;
6582 struct page_cgroup
*pc
;
6585 * We don't take compound_lock() here but no race with splitting thp
6587 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6588 * under splitting, which means there's no concurrent thp split,
6589 * - if another thread runs into split_huge_page() just after we
6590 * entered this if-block, the thread must wait for page table lock
6591 * to be unlocked in __split_huge_page_splitting(), where the main
6592 * part of thp split is not executed yet.
6594 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6595 if (mc
.precharge
< HPAGE_PMD_NR
) {
6596 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6599 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6600 if (target_type
== MC_TARGET_PAGE
) {
6602 if (!isolate_lru_page(page
)) {
6603 pc
= lookup_page_cgroup(page
);
6604 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6605 pc
, mc
.from
, mc
.to
)) {
6606 mc
.precharge
-= HPAGE_PMD_NR
;
6607 mc
.moved_charge
+= HPAGE_PMD_NR
;
6609 putback_lru_page(page
);
6613 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6617 if (pmd_trans_unstable(pmd
))
6620 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6621 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6622 pte_t ptent
= *(pte
++);
6628 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6629 case MC_TARGET_PAGE
:
6631 if (isolate_lru_page(page
))
6633 pc
= lookup_page_cgroup(page
);
6634 if (!mem_cgroup_move_account(page
, 1, pc
,
6637 /* we uncharge from mc.from later. */
6640 putback_lru_page(page
);
6641 put
: /* get_mctgt_type() gets the page */
6644 case MC_TARGET_SWAP
:
6646 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6648 /* we fixup refcnts and charges later. */
6656 pte_unmap_unlock(pte
- 1, ptl
);
6661 * We have consumed all precharges we got in can_attach().
6662 * We try charge one by one, but don't do any additional
6663 * charges to mc.to if we have failed in charge once in attach()
6666 ret
= mem_cgroup_do_precharge(1);
6674 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6676 struct vm_area_struct
*vma
;
6678 lru_add_drain_all();
6680 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6682 * Someone who are holding the mmap_sem might be waiting in
6683 * waitq. So we cancel all extra charges, wake up all waiters,
6684 * and retry. Because we cancel precharges, we might not be able
6685 * to move enough charges, but moving charge is a best-effort
6686 * feature anyway, so it wouldn't be a big problem.
6688 __mem_cgroup_clear_mc();
6692 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6694 struct mm_walk mem_cgroup_move_charge_walk
= {
6695 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6699 if (is_vm_hugetlb_page(vma
))
6701 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6702 &mem_cgroup_move_charge_walk
);
6705 * means we have consumed all precharges and failed in
6706 * doing additional charge. Just abandon here.
6710 up_read(&mm
->mmap_sem
);
6713 static void mem_cgroup_move_task(struct cgroup
*cont
,
6714 struct cgroup_taskset
*tset
)
6716 struct task_struct
*p
= cgroup_taskset_first(tset
);
6717 struct mm_struct
*mm
= get_task_mm(p
);
6721 mem_cgroup_move_charge(mm
);
6725 mem_cgroup_clear_mc();
6727 #else /* !CONFIG_MMU */
6728 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6729 struct cgroup_taskset
*tset
)
6733 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6734 struct cgroup_taskset
*tset
)
6737 static void mem_cgroup_move_task(struct cgroup
*cont
,
6738 struct cgroup_taskset
*tset
)
6743 struct cgroup_subsys mem_cgroup_subsys
= {
6745 .subsys_id
= mem_cgroup_subsys_id
,
6746 .css_alloc
= mem_cgroup_css_alloc
,
6747 .css_offline
= mem_cgroup_css_offline
,
6748 .css_free
= mem_cgroup_css_free
,
6749 .can_attach
= mem_cgroup_can_attach
,
6750 .cancel_attach
= mem_cgroup_cancel_attach
,
6751 .attach
= mem_cgroup_move_task
,
6752 .base_cftypes
= mem_cgroup_files
,
6757 #ifdef CONFIG_MEMCG_SWAP
6758 static int __init
enable_swap_account(char *s
)
6760 /* consider enabled if no parameter or 1 is given */
6761 if (!strcmp(s
, "1"))
6762 really_do_swap_account
= 1;
6763 else if (!strcmp(s
, "0"))
6764 really_do_swap_account
= 0;
6767 __setup("swapaccount=", enable_swap_account
);