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 static int memcg_limited_groups_array_size
;
577 * MIN_SIZE is different than 1, because we would like to avoid going through
578 * the alloc/free process all the time. In a small machine, 4 kmem-limited
579 * cgroups is a reasonable guess. In the future, it could be a parameter or
580 * tunable, but that is strictly not necessary.
582 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
583 * this constant directly from cgroup, but it is understandable that this is
584 * better kept as an internal representation in cgroup.c. In any case, the
585 * css_id space is not getting any smaller, and we don't have to necessarily
586 * increase ours as well if it increases.
588 #define MEMCG_CACHES_MIN_SIZE 4
589 #define MEMCG_CACHES_MAX_SIZE 65535
592 * A lot of the calls to the cache allocation functions are expected to be
593 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
594 * conditional to this static branch, we'll have to allow modules that does
595 * kmem_cache_alloc and the such to see this symbol as well
597 struct static_key memcg_kmem_enabled_key
;
598 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
600 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
602 if (memcg_kmem_is_active(memcg
)) {
603 static_key_slow_dec(&memcg_kmem_enabled_key
);
604 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
607 * This check can't live in kmem destruction function,
608 * since the charges will outlive the cgroup
610 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
613 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
616 #endif /* CONFIG_MEMCG_KMEM */
618 static void disarm_static_keys(struct mem_cgroup
*memcg
)
620 disarm_sock_keys(memcg
);
621 disarm_kmem_keys(memcg
);
624 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
626 static struct mem_cgroup_per_zone
*
627 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
629 return &memcg
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
632 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
637 static struct mem_cgroup_per_zone
*
638 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
640 int nid
= page_to_nid(page
);
641 int zid
= page_zonenum(page
);
643 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
646 static struct mem_cgroup_tree_per_zone
*
647 soft_limit_tree_node_zone(int nid
, int zid
)
649 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
652 static struct mem_cgroup_tree_per_zone
*
653 soft_limit_tree_from_page(struct page
*page
)
655 int nid
= page_to_nid(page
);
656 int zid
= page_zonenum(page
);
658 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
662 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
663 struct mem_cgroup_per_zone
*mz
,
664 struct mem_cgroup_tree_per_zone
*mctz
,
665 unsigned long long new_usage_in_excess
)
667 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
668 struct rb_node
*parent
= NULL
;
669 struct mem_cgroup_per_zone
*mz_node
;
674 mz
->usage_in_excess
= new_usage_in_excess
;
675 if (!mz
->usage_in_excess
)
679 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
681 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
684 * We can't avoid mem cgroups that are over their soft
685 * limit by the same amount
687 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
690 rb_link_node(&mz
->tree_node
, parent
, p
);
691 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
696 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
697 struct mem_cgroup_per_zone
*mz
,
698 struct mem_cgroup_tree_per_zone
*mctz
)
702 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
707 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
708 struct mem_cgroup_per_zone
*mz
,
709 struct mem_cgroup_tree_per_zone
*mctz
)
711 spin_lock(&mctz
->lock
);
712 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
713 spin_unlock(&mctz
->lock
);
717 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
719 unsigned long long excess
;
720 struct mem_cgroup_per_zone
*mz
;
721 struct mem_cgroup_tree_per_zone
*mctz
;
722 int nid
= page_to_nid(page
);
723 int zid
= page_zonenum(page
);
724 mctz
= soft_limit_tree_from_page(page
);
727 * Necessary to update all ancestors when hierarchy is used.
728 * because their event counter is not touched.
730 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
731 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
732 excess
= res_counter_soft_limit_excess(&memcg
->res
);
734 * We have to update the tree if mz is on RB-tree or
735 * mem is over its softlimit.
737 if (excess
|| mz
->on_tree
) {
738 spin_lock(&mctz
->lock
);
739 /* if on-tree, remove it */
741 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
743 * Insert again. mz->usage_in_excess will be updated.
744 * If excess is 0, no tree ops.
746 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
747 spin_unlock(&mctz
->lock
);
752 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
755 struct mem_cgroup_per_zone
*mz
;
756 struct mem_cgroup_tree_per_zone
*mctz
;
758 for_each_node(node
) {
759 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
760 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
761 mctz
= soft_limit_tree_node_zone(node
, zone
);
762 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
767 static struct mem_cgroup_per_zone
*
768 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
770 struct rb_node
*rightmost
= NULL
;
771 struct mem_cgroup_per_zone
*mz
;
775 rightmost
= rb_last(&mctz
->rb_root
);
777 goto done
; /* Nothing to reclaim from */
779 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
781 * Remove the node now but someone else can add it back,
782 * we will to add it back at the end of reclaim to its correct
783 * position in the tree.
785 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
786 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
787 !css_tryget(&mz
->memcg
->css
))
793 static struct mem_cgroup_per_zone
*
794 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
796 struct mem_cgroup_per_zone
*mz
;
798 spin_lock(&mctz
->lock
);
799 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
800 spin_unlock(&mctz
->lock
);
805 * Implementation Note: reading percpu statistics for memcg.
807 * Both of vmstat[] and percpu_counter has threshold and do periodic
808 * synchronization to implement "quick" read. There are trade-off between
809 * reading cost and precision of value. Then, we may have a chance to implement
810 * a periodic synchronizion of counter in memcg's counter.
812 * But this _read() function is used for user interface now. The user accounts
813 * memory usage by memory cgroup and he _always_ requires exact value because
814 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
815 * have to visit all online cpus and make sum. So, for now, unnecessary
816 * synchronization is not implemented. (just implemented for cpu hotplug)
818 * If there are kernel internal actions which can make use of some not-exact
819 * value, and reading all cpu value can be performance bottleneck in some
820 * common workload, threashold and synchonization as vmstat[] should be
823 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
824 enum mem_cgroup_stat_index idx
)
830 for_each_online_cpu(cpu
)
831 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
832 #ifdef CONFIG_HOTPLUG_CPU
833 spin_lock(&memcg
->pcp_counter_lock
);
834 val
+= memcg
->nocpu_base
.count
[idx
];
835 spin_unlock(&memcg
->pcp_counter_lock
);
841 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
844 int val
= (charge
) ? 1 : -1;
845 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
848 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
849 enum mem_cgroup_events_index idx
)
851 unsigned long val
= 0;
854 for_each_online_cpu(cpu
)
855 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
856 #ifdef CONFIG_HOTPLUG_CPU
857 spin_lock(&memcg
->pcp_counter_lock
);
858 val
+= memcg
->nocpu_base
.events
[idx
];
859 spin_unlock(&memcg
->pcp_counter_lock
);
864 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
865 bool anon
, int nr_pages
)
870 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
871 * counted as CACHE even if it's on ANON LRU.
874 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
877 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
880 /* pagein of a big page is an event. So, ignore page size */
882 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
884 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
885 nr_pages
= -nr_pages
; /* for event */
888 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
894 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
896 struct mem_cgroup_per_zone
*mz
;
898 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
899 return mz
->lru_size
[lru
];
903 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
904 unsigned int lru_mask
)
906 struct mem_cgroup_per_zone
*mz
;
908 unsigned long ret
= 0;
910 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
913 if (BIT(lru
) & lru_mask
)
914 ret
+= mz
->lru_size
[lru
];
920 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
921 int nid
, unsigned int lru_mask
)
926 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
927 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
933 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
934 unsigned int lru_mask
)
939 for_each_node_state(nid
, N_MEMORY
)
940 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
944 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
945 enum mem_cgroup_events_target target
)
947 unsigned long val
, next
;
949 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
950 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
951 /* from time_after() in jiffies.h */
952 if ((long)next
- (long)val
< 0) {
954 case MEM_CGROUP_TARGET_THRESH
:
955 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
957 case MEM_CGROUP_TARGET_SOFTLIMIT
:
958 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
960 case MEM_CGROUP_TARGET_NUMAINFO
:
961 next
= val
+ NUMAINFO_EVENTS_TARGET
;
966 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
973 * Check events in order.
976 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
979 /* threshold event is triggered in finer grain than soft limit */
980 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
981 MEM_CGROUP_TARGET_THRESH
))) {
983 bool do_numainfo __maybe_unused
;
985 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
986 MEM_CGROUP_TARGET_SOFTLIMIT
);
988 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
989 MEM_CGROUP_TARGET_NUMAINFO
);
993 mem_cgroup_threshold(memcg
);
994 if (unlikely(do_softlimit
))
995 mem_cgroup_update_tree(memcg
, page
);
997 if (unlikely(do_numainfo
))
998 atomic_inc(&memcg
->numainfo_events
);
1004 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
1006 return mem_cgroup_from_css(
1007 cgroup_subsys_state(cont
, mem_cgroup_subsys_id
));
1010 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1013 * mm_update_next_owner() may clear mm->owner to NULL
1014 * if it races with swapoff, page migration, etc.
1015 * So this can be called with p == NULL.
1020 return mem_cgroup_from_css(task_subsys_state(p
, mem_cgroup_subsys_id
));
1023 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1025 struct mem_cgroup
*memcg
= NULL
;
1030 * Because we have no locks, mm->owner's may be being moved to other
1031 * cgroup. We use css_tryget() here even if this looks
1032 * pessimistic (rather than adding locks here).
1036 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1037 if (unlikely(!memcg
))
1039 } while (!css_tryget(&memcg
->css
));
1045 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1046 * @root: hierarchy root
1047 * @prev: previously returned memcg, NULL on first invocation
1048 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1050 * Returns references to children of the hierarchy below @root, or
1051 * @root itself, or %NULL after a full round-trip.
1053 * Caller must pass the return value in @prev on subsequent
1054 * invocations for reference counting, or use mem_cgroup_iter_break()
1055 * to cancel a hierarchy walk before the round-trip is complete.
1057 * Reclaimers can specify a zone and a priority level in @reclaim to
1058 * divide up the memcgs in the hierarchy among all concurrent
1059 * reclaimers operating on the same zone and priority.
1061 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1062 struct mem_cgroup
*prev
,
1063 struct mem_cgroup_reclaim_cookie
*reclaim
)
1065 struct mem_cgroup
*memcg
= NULL
;
1068 if (mem_cgroup_disabled())
1072 root
= root_mem_cgroup
;
1074 if (prev
&& !reclaim
)
1075 id
= css_id(&prev
->css
);
1077 if (prev
&& prev
!= root
)
1078 css_put(&prev
->css
);
1080 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1087 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1088 struct cgroup_subsys_state
*css
;
1091 int nid
= zone_to_nid(reclaim
->zone
);
1092 int zid
= zone_idx(reclaim
->zone
);
1093 struct mem_cgroup_per_zone
*mz
;
1095 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1096 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1097 if (prev
&& reclaim
->generation
!= iter
->generation
)
1099 id
= iter
->position
;
1103 css
= css_get_next(&mem_cgroup_subsys
, id
+ 1, &root
->css
, &id
);
1105 if (css
== &root
->css
|| css_tryget(css
))
1106 memcg
= mem_cgroup_from_css(css
);
1112 iter
->position
= id
;
1115 else if (!prev
&& memcg
)
1116 reclaim
->generation
= iter
->generation
;
1126 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1127 * @root: hierarchy root
1128 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1130 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1131 struct mem_cgroup
*prev
)
1134 root
= root_mem_cgroup
;
1135 if (prev
&& prev
!= root
)
1136 css_put(&prev
->css
);
1140 * Iteration constructs for visiting all cgroups (under a tree). If
1141 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1142 * be used for reference counting.
1144 #define for_each_mem_cgroup_tree(iter, root) \
1145 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1147 iter = mem_cgroup_iter(root, iter, NULL))
1149 #define for_each_mem_cgroup(iter) \
1150 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1152 iter = mem_cgroup_iter(NULL, iter, NULL))
1154 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1156 struct mem_cgroup
*memcg
;
1159 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1160 if (unlikely(!memcg
))
1165 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1168 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1176 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1179 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1180 * @zone: zone of the wanted lruvec
1181 * @memcg: memcg of the wanted lruvec
1183 * Returns the lru list vector holding pages for the given @zone and
1184 * @mem. This can be the global zone lruvec, if the memory controller
1187 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1188 struct mem_cgroup
*memcg
)
1190 struct mem_cgroup_per_zone
*mz
;
1191 struct lruvec
*lruvec
;
1193 if (mem_cgroup_disabled()) {
1194 lruvec
= &zone
->lruvec
;
1198 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1199 lruvec
= &mz
->lruvec
;
1202 * Since a node can be onlined after the mem_cgroup was created,
1203 * we have to be prepared to initialize lruvec->zone here;
1204 * and if offlined then reonlined, we need to reinitialize it.
1206 if (unlikely(lruvec
->zone
!= zone
))
1207 lruvec
->zone
= zone
;
1212 * Following LRU functions are allowed to be used without PCG_LOCK.
1213 * Operations are called by routine of global LRU independently from memcg.
1214 * What we have to take care of here is validness of pc->mem_cgroup.
1216 * Changes to pc->mem_cgroup happens when
1219 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1220 * It is added to LRU before charge.
1221 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1222 * When moving account, the page is not on LRU. It's isolated.
1226 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1228 * @zone: zone of the page
1230 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1232 struct mem_cgroup_per_zone
*mz
;
1233 struct mem_cgroup
*memcg
;
1234 struct page_cgroup
*pc
;
1235 struct lruvec
*lruvec
;
1237 if (mem_cgroup_disabled()) {
1238 lruvec
= &zone
->lruvec
;
1242 pc
= lookup_page_cgroup(page
);
1243 memcg
= pc
->mem_cgroup
;
1246 * Surreptitiously switch any uncharged offlist page to root:
1247 * an uncharged page off lru does nothing to secure
1248 * its former mem_cgroup from sudden removal.
1250 * Our caller holds lru_lock, and PageCgroupUsed is updated
1251 * under page_cgroup lock: between them, they make all uses
1252 * of pc->mem_cgroup safe.
1254 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1255 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1257 mz
= page_cgroup_zoneinfo(memcg
, page
);
1258 lruvec
= &mz
->lruvec
;
1261 * Since a node can be onlined after the mem_cgroup was created,
1262 * we have to be prepared to initialize lruvec->zone here;
1263 * and if offlined then reonlined, we need to reinitialize it.
1265 if (unlikely(lruvec
->zone
!= zone
))
1266 lruvec
->zone
= zone
;
1271 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1272 * @lruvec: mem_cgroup per zone lru vector
1273 * @lru: index of lru list the page is sitting on
1274 * @nr_pages: positive when adding or negative when removing
1276 * This function must be called when a page is added to or removed from an
1279 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1282 struct mem_cgroup_per_zone
*mz
;
1283 unsigned long *lru_size
;
1285 if (mem_cgroup_disabled())
1288 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1289 lru_size
= mz
->lru_size
+ lru
;
1290 *lru_size
+= nr_pages
;
1291 VM_BUG_ON((long)(*lru_size
) < 0);
1295 * Checks whether given mem is same or in the root_mem_cgroup's
1298 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1299 struct mem_cgroup
*memcg
)
1301 if (root_memcg
== memcg
)
1303 if (!root_memcg
->use_hierarchy
|| !memcg
)
1305 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1308 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1309 struct mem_cgroup
*memcg
)
1314 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1319 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*memcg
)
1322 struct mem_cgroup
*curr
= NULL
;
1323 struct task_struct
*p
;
1325 p
= find_lock_task_mm(task
);
1327 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1331 * All threads may have already detached their mm's, but the oom
1332 * killer still needs to detect if they have already been oom
1333 * killed to prevent needlessly killing additional tasks.
1336 curr
= mem_cgroup_from_task(task
);
1338 css_get(&curr
->css
);
1344 * We should check use_hierarchy of "memcg" not "curr". Because checking
1345 * use_hierarchy of "curr" here make this function true if hierarchy is
1346 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1347 * hierarchy(even if use_hierarchy is disabled in "memcg").
1349 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1350 css_put(&curr
->css
);
1354 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1356 unsigned long inactive_ratio
;
1357 unsigned long inactive
;
1358 unsigned long active
;
1361 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1362 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1364 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1366 inactive_ratio
= int_sqrt(10 * gb
);
1370 return inactive
* inactive_ratio
< active
;
1373 int mem_cgroup_inactive_file_is_low(struct lruvec
*lruvec
)
1375 unsigned long active
;
1376 unsigned long inactive
;
1378 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1379 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1381 return (active
> inactive
);
1384 #define mem_cgroup_from_res_counter(counter, member) \
1385 container_of(counter, struct mem_cgroup, member)
1388 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1389 * @memcg: the memory cgroup
1391 * Returns the maximum amount of memory @mem can be charged with, in
1394 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1396 unsigned long long margin
;
1398 margin
= res_counter_margin(&memcg
->res
);
1399 if (do_swap_account
)
1400 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1401 return margin
>> PAGE_SHIFT
;
1404 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1406 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1409 if (cgrp
->parent
== NULL
)
1410 return vm_swappiness
;
1412 return memcg
->swappiness
;
1416 * memcg->moving_account is used for checking possibility that some thread is
1417 * calling move_account(). When a thread on CPU-A starts moving pages under
1418 * a memcg, other threads should check memcg->moving_account under
1419 * rcu_read_lock(), like this:
1423 * memcg->moving_account+1 if (memcg->mocing_account)
1425 * synchronize_rcu() update something.
1430 /* for quick checking without looking up memcg */
1431 atomic_t memcg_moving __read_mostly
;
1433 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1435 atomic_inc(&memcg_moving
);
1436 atomic_inc(&memcg
->moving_account
);
1440 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1443 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1444 * We check NULL in callee rather than caller.
1447 atomic_dec(&memcg_moving
);
1448 atomic_dec(&memcg
->moving_account
);
1453 * 2 routines for checking "mem" is under move_account() or not.
1455 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1456 * is used for avoiding races in accounting. If true,
1457 * pc->mem_cgroup may be overwritten.
1459 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1460 * under hierarchy of moving cgroups. This is for
1461 * waiting at hith-memory prressure caused by "move".
1464 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1466 VM_BUG_ON(!rcu_read_lock_held());
1467 return atomic_read(&memcg
->moving_account
) > 0;
1470 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1472 struct mem_cgroup
*from
;
1473 struct mem_cgroup
*to
;
1476 * Unlike task_move routines, we access mc.to, mc.from not under
1477 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1479 spin_lock(&mc
.lock
);
1485 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1486 || mem_cgroup_same_or_subtree(memcg
, to
);
1488 spin_unlock(&mc
.lock
);
1492 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1494 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1495 if (mem_cgroup_under_move(memcg
)) {
1497 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1498 /* moving charge context might have finished. */
1501 finish_wait(&mc
.waitq
, &wait
);
1509 * Take this lock when
1510 * - a code tries to modify page's memcg while it's USED.
1511 * - a code tries to modify page state accounting in a memcg.
1512 * see mem_cgroup_stolen(), too.
1514 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1515 unsigned long *flags
)
1517 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1520 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1521 unsigned long *flags
)
1523 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1527 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1528 * @memcg: The memory cgroup that went over limit
1529 * @p: Task that is going to be killed
1531 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1534 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1536 struct cgroup
*task_cgrp
;
1537 struct cgroup
*mem_cgrp
;
1539 * Need a buffer in BSS, can't rely on allocations. The code relies
1540 * on the assumption that OOM is serialized for memory controller.
1541 * If this assumption is broken, revisit this code.
1543 static char memcg_name
[PATH_MAX
];
1551 mem_cgrp
= memcg
->css
.cgroup
;
1552 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1554 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1557 * Unfortunately, we are unable to convert to a useful name
1558 * But we'll still print out the usage information
1565 printk(KERN_INFO
"Task in %s killed", memcg_name
);
1568 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1576 * Continues from above, so we don't need an KERN_ level
1578 printk(KERN_CONT
" as a result of limit of %s\n", memcg_name
);
1581 printk(KERN_INFO
"memory: usage %llukB, limit %llukB, failcnt %llu\n",
1582 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1583 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1584 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1585 printk(KERN_INFO
"memory+swap: usage %llukB, limit %llukB, "
1587 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1588 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1589 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1590 printk(KERN_INFO
"kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1591 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1592 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1593 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1597 * This function returns the number of memcg under hierarchy tree. Returns
1598 * 1(self count) if no children.
1600 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1603 struct mem_cgroup
*iter
;
1605 for_each_mem_cgroup_tree(iter
, memcg
)
1611 * Return the memory (and swap, if configured) limit for a memcg.
1613 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1617 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1620 * Do not consider swap space if we cannot swap due to swappiness
1622 if (mem_cgroup_swappiness(memcg
)) {
1625 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1626 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1629 * If memsw is finite and limits the amount of swap space
1630 * available to this memcg, return that limit.
1632 limit
= min(limit
, memsw
);
1638 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1641 struct mem_cgroup
*iter
;
1642 unsigned long chosen_points
= 0;
1643 unsigned long totalpages
;
1644 unsigned int points
= 0;
1645 struct task_struct
*chosen
= NULL
;
1648 * If current has a pending SIGKILL, then automatically select it. The
1649 * goal is to allow it to allocate so that it may quickly exit and free
1652 if (fatal_signal_pending(current
)) {
1653 set_thread_flag(TIF_MEMDIE
);
1657 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1658 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1659 for_each_mem_cgroup_tree(iter
, memcg
) {
1660 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1661 struct cgroup_iter it
;
1662 struct task_struct
*task
;
1664 cgroup_iter_start(cgroup
, &it
);
1665 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1666 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1668 case OOM_SCAN_SELECT
:
1670 put_task_struct(chosen
);
1672 chosen_points
= ULONG_MAX
;
1673 get_task_struct(chosen
);
1675 case OOM_SCAN_CONTINUE
:
1677 case OOM_SCAN_ABORT
:
1678 cgroup_iter_end(cgroup
, &it
);
1679 mem_cgroup_iter_break(memcg
, iter
);
1681 put_task_struct(chosen
);
1686 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1687 if (points
> chosen_points
) {
1689 put_task_struct(chosen
);
1691 chosen_points
= points
;
1692 get_task_struct(chosen
);
1695 cgroup_iter_end(cgroup
, &it
);
1700 points
= chosen_points
* 1000 / totalpages
;
1701 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1702 NULL
, "Memory cgroup out of memory");
1705 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1707 unsigned long flags
)
1709 unsigned long total
= 0;
1710 bool noswap
= false;
1713 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1715 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1718 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1720 drain_all_stock_async(memcg
);
1721 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1723 * Allow limit shrinkers, which are triggered directly
1724 * by userspace, to catch signals and stop reclaim
1725 * after minimal progress, regardless of the margin.
1727 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1729 if (mem_cgroup_margin(memcg
))
1732 * If nothing was reclaimed after two attempts, there
1733 * may be no reclaimable pages in this hierarchy.
1742 * test_mem_cgroup_node_reclaimable
1743 * @memcg: the target memcg
1744 * @nid: the node ID to be checked.
1745 * @noswap : specify true here if the user wants flle only information.
1747 * This function returns whether the specified memcg contains any
1748 * reclaimable pages on a node. Returns true if there are any reclaimable
1749 * pages in the node.
1751 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1752 int nid
, bool noswap
)
1754 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1756 if (noswap
|| !total_swap_pages
)
1758 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1763 #if MAX_NUMNODES > 1
1766 * Always updating the nodemask is not very good - even if we have an empty
1767 * list or the wrong list here, we can start from some node and traverse all
1768 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1771 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1775 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1776 * pagein/pageout changes since the last update.
1778 if (!atomic_read(&memcg
->numainfo_events
))
1780 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1783 /* make a nodemask where this memcg uses memory from */
1784 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1786 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1788 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1789 node_clear(nid
, memcg
->scan_nodes
);
1792 atomic_set(&memcg
->numainfo_events
, 0);
1793 atomic_set(&memcg
->numainfo_updating
, 0);
1797 * Selecting a node where we start reclaim from. Because what we need is just
1798 * reducing usage counter, start from anywhere is O,K. Considering
1799 * memory reclaim from current node, there are pros. and cons.
1801 * Freeing memory from current node means freeing memory from a node which
1802 * we'll use or we've used. So, it may make LRU bad. And if several threads
1803 * hit limits, it will see a contention on a node. But freeing from remote
1804 * node means more costs for memory reclaim because of memory latency.
1806 * Now, we use round-robin. Better algorithm is welcomed.
1808 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1812 mem_cgroup_may_update_nodemask(memcg
);
1813 node
= memcg
->last_scanned_node
;
1815 node
= next_node(node
, memcg
->scan_nodes
);
1816 if (node
== MAX_NUMNODES
)
1817 node
= first_node(memcg
->scan_nodes
);
1819 * We call this when we hit limit, not when pages are added to LRU.
1820 * No LRU may hold pages because all pages are UNEVICTABLE or
1821 * memcg is too small and all pages are not on LRU. In that case,
1822 * we use curret node.
1824 if (unlikely(node
== MAX_NUMNODES
))
1825 node
= numa_node_id();
1827 memcg
->last_scanned_node
= node
;
1832 * Check all nodes whether it contains reclaimable pages or not.
1833 * For quick scan, we make use of scan_nodes. This will allow us to skip
1834 * unused nodes. But scan_nodes is lazily updated and may not cotain
1835 * enough new information. We need to do double check.
1837 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1842 * quick check...making use of scan_node.
1843 * We can skip unused nodes.
1845 if (!nodes_empty(memcg
->scan_nodes
)) {
1846 for (nid
= first_node(memcg
->scan_nodes
);
1848 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1850 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1855 * Check rest of nodes.
1857 for_each_node_state(nid
, N_MEMORY
) {
1858 if (node_isset(nid
, memcg
->scan_nodes
))
1860 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1867 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1872 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1874 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1878 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1881 unsigned long *total_scanned
)
1883 struct mem_cgroup
*victim
= NULL
;
1886 unsigned long excess
;
1887 unsigned long nr_scanned
;
1888 struct mem_cgroup_reclaim_cookie reclaim
= {
1893 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
1896 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1901 * If we have not been able to reclaim
1902 * anything, it might because there are
1903 * no reclaimable pages under this hierarchy
1908 * We want to do more targeted reclaim.
1909 * excess >> 2 is not to excessive so as to
1910 * reclaim too much, nor too less that we keep
1911 * coming back to reclaim from this cgroup
1913 if (total
>= (excess
>> 2) ||
1914 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1919 if (!mem_cgroup_reclaimable(victim
, false))
1921 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1923 *total_scanned
+= nr_scanned
;
1924 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
1927 mem_cgroup_iter_break(root_memcg
, victim
);
1932 * Check OOM-Killer is already running under our hierarchy.
1933 * If someone is running, return false.
1934 * Has to be called with memcg_oom_lock
1936 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
1938 struct mem_cgroup
*iter
, *failed
= NULL
;
1940 for_each_mem_cgroup_tree(iter
, memcg
) {
1941 if (iter
->oom_lock
) {
1943 * this subtree of our hierarchy is already locked
1944 * so we cannot give a lock.
1947 mem_cgroup_iter_break(memcg
, iter
);
1950 iter
->oom_lock
= true;
1957 * OK, we failed to lock the whole subtree so we have to clean up
1958 * what we set up to the failing subtree
1960 for_each_mem_cgroup_tree(iter
, memcg
) {
1961 if (iter
== failed
) {
1962 mem_cgroup_iter_break(memcg
, iter
);
1965 iter
->oom_lock
= false;
1971 * Has to be called with memcg_oom_lock
1973 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1975 struct mem_cgroup
*iter
;
1977 for_each_mem_cgroup_tree(iter
, memcg
)
1978 iter
->oom_lock
= false;
1982 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1984 struct mem_cgroup
*iter
;
1986 for_each_mem_cgroup_tree(iter
, memcg
)
1987 atomic_inc(&iter
->under_oom
);
1990 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1992 struct mem_cgroup
*iter
;
1995 * When a new child is created while the hierarchy is under oom,
1996 * mem_cgroup_oom_lock() may not be called. We have to use
1997 * atomic_add_unless() here.
1999 for_each_mem_cgroup_tree(iter
, memcg
)
2000 atomic_add_unless(&iter
->under_oom
, -1, 0);
2003 static DEFINE_SPINLOCK(memcg_oom_lock
);
2004 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2006 struct oom_wait_info
{
2007 struct mem_cgroup
*memcg
;
2011 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2012 unsigned mode
, int sync
, void *arg
)
2014 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2015 struct mem_cgroup
*oom_wait_memcg
;
2016 struct oom_wait_info
*oom_wait_info
;
2018 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2019 oom_wait_memcg
= oom_wait_info
->memcg
;
2022 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2023 * Then we can use css_is_ancestor without taking care of RCU.
2025 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2026 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2028 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2031 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2033 /* for filtering, pass "memcg" as argument. */
2034 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2037 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2039 if (memcg
&& atomic_read(&memcg
->under_oom
))
2040 memcg_wakeup_oom(memcg
);
2044 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2046 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2049 struct oom_wait_info owait
;
2050 bool locked
, need_to_kill
;
2052 owait
.memcg
= memcg
;
2053 owait
.wait
.flags
= 0;
2054 owait
.wait
.func
= memcg_oom_wake_function
;
2055 owait
.wait
.private = current
;
2056 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2057 need_to_kill
= true;
2058 mem_cgroup_mark_under_oom(memcg
);
2060 /* At first, try to OOM lock hierarchy under memcg.*/
2061 spin_lock(&memcg_oom_lock
);
2062 locked
= mem_cgroup_oom_lock(memcg
);
2064 * Even if signal_pending(), we can't quit charge() loop without
2065 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2066 * under OOM is always welcomed, use TASK_KILLABLE here.
2068 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2069 if (!locked
|| memcg
->oom_kill_disable
)
2070 need_to_kill
= false;
2072 mem_cgroup_oom_notify(memcg
);
2073 spin_unlock(&memcg_oom_lock
);
2076 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2077 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2080 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2082 spin_lock(&memcg_oom_lock
);
2084 mem_cgroup_oom_unlock(memcg
);
2085 memcg_wakeup_oom(memcg
);
2086 spin_unlock(&memcg_oom_lock
);
2088 mem_cgroup_unmark_under_oom(memcg
);
2090 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2092 /* Give chance to dying process */
2093 schedule_timeout_uninterruptible(1);
2098 * Currently used to update mapped file statistics, but the routine can be
2099 * generalized to update other statistics as well.
2101 * Notes: Race condition
2103 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2104 * it tends to be costly. But considering some conditions, we doesn't need
2105 * to do so _always_.
2107 * Considering "charge", lock_page_cgroup() is not required because all
2108 * file-stat operations happen after a page is attached to radix-tree. There
2109 * are no race with "charge".
2111 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2112 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2113 * if there are race with "uncharge". Statistics itself is properly handled
2116 * Considering "move", this is an only case we see a race. To make the race
2117 * small, we check mm->moving_account and detect there are possibility of race
2118 * If there is, we take a lock.
2121 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2122 bool *locked
, unsigned long *flags
)
2124 struct mem_cgroup
*memcg
;
2125 struct page_cgroup
*pc
;
2127 pc
= lookup_page_cgroup(page
);
2129 memcg
= pc
->mem_cgroup
;
2130 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2133 * If this memory cgroup is not under account moving, we don't
2134 * need to take move_lock_mem_cgroup(). Because we already hold
2135 * rcu_read_lock(), any calls to move_account will be delayed until
2136 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2138 if (!mem_cgroup_stolen(memcg
))
2141 move_lock_mem_cgroup(memcg
, flags
);
2142 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2143 move_unlock_mem_cgroup(memcg
, flags
);
2149 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2151 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2154 * It's guaranteed that pc->mem_cgroup never changes while
2155 * lock is held because a routine modifies pc->mem_cgroup
2156 * should take move_lock_mem_cgroup().
2158 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2161 void mem_cgroup_update_page_stat(struct page
*page
,
2162 enum mem_cgroup_page_stat_item idx
, int val
)
2164 struct mem_cgroup
*memcg
;
2165 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2166 unsigned long uninitialized_var(flags
);
2168 if (mem_cgroup_disabled())
2171 memcg
= pc
->mem_cgroup
;
2172 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2176 case MEMCG_NR_FILE_MAPPED
:
2177 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2183 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2187 * size of first charge trial. "32" comes from vmscan.c's magic value.
2188 * TODO: maybe necessary to use big numbers in big irons.
2190 #define CHARGE_BATCH 32U
2191 struct memcg_stock_pcp
{
2192 struct mem_cgroup
*cached
; /* this never be root cgroup */
2193 unsigned int nr_pages
;
2194 struct work_struct work
;
2195 unsigned long flags
;
2196 #define FLUSHING_CACHED_CHARGE 0
2198 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2199 static DEFINE_MUTEX(percpu_charge_mutex
);
2202 * consume_stock: Try to consume stocked charge on this cpu.
2203 * @memcg: memcg to consume from.
2204 * @nr_pages: how many pages to charge.
2206 * The charges will only happen if @memcg matches the current cpu's memcg
2207 * stock, and at least @nr_pages are available in that stock. Failure to
2208 * service an allocation will refill the stock.
2210 * returns true if successful, false otherwise.
2212 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2214 struct memcg_stock_pcp
*stock
;
2217 if (nr_pages
> CHARGE_BATCH
)
2220 stock
= &get_cpu_var(memcg_stock
);
2221 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2222 stock
->nr_pages
-= nr_pages
;
2223 else /* need to call res_counter_charge */
2225 put_cpu_var(memcg_stock
);
2230 * Returns stocks cached in percpu to res_counter and reset cached information.
2232 static void drain_stock(struct memcg_stock_pcp
*stock
)
2234 struct mem_cgroup
*old
= stock
->cached
;
2236 if (stock
->nr_pages
) {
2237 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2239 res_counter_uncharge(&old
->res
, bytes
);
2240 if (do_swap_account
)
2241 res_counter_uncharge(&old
->memsw
, bytes
);
2242 stock
->nr_pages
= 0;
2244 stock
->cached
= NULL
;
2248 * This must be called under preempt disabled or must be called by
2249 * a thread which is pinned to local cpu.
2251 static void drain_local_stock(struct work_struct
*dummy
)
2253 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2255 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2259 * Cache charges(val) which is from res_counter, to local per_cpu area.
2260 * This will be consumed by consume_stock() function, later.
2262 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2264 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2266 if (stock
->cached
!= memcg
) { /* reset if necessary */
2268 stock
->cached
= memcg
;
2270 stock
->nr_pages
+= nr_pages
;
2271 put_cpu_var(memcg_stock
);
2275 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2276 * of the hierarchy under it. sync flag says whether we should block
2277 * until the work is done.
2279 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2283 /* Notify other cpus that system-wide "drain" is running */
2286 for_each_online_cpu(cpu
) {
2287 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2288 struct mem_cgroup
*memcg
;
2290 memcg
= stock
->cached
;
2291 if (!memcg
|| !stock
->nr_pages
)
2293 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2295 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2297 drain_local_stock(&stock
->work
);
2299 schedule_work_on(cpu
, &stock
->work
);
2307 for_each_online_cpu(cpu
) {
2308 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2309 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2310 flush_work(&stock
->work
);
2317 * Tries to drain stocked charges in other cpus. This function is asynchronous
2318 * and just put a work per cpu for draining localy on each cpu. Caller can
2319 * expects some charges will be back to res_counter later but cannot wait for
2322 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2325 * If someone calls draining, avoid adding more kworker runs.
2327 if (!mutex_trylock(&percpu_charge_mutex
))
2329 drain_all_stock(root_memcg
, false);
2330 mutex_unlock(&percpu_charge_mutex
);
2333 /* This is a synchronous drain interface. */
2334 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2336 /* called when force_empty is called */
2337 mutex_lock(&percpu_charge_mutex
);
2338 drain_all_stock(root_memcg
, true);
2339 mutex_unlock(&percpu_charge_mutex
);
2343 * This function drains percpu counter value from DEAD cpu and
2344 * move it to local cpu. Note that this function can be preempted.
2346 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2350 spin_lock(&memcg
->pcp_counter_lock
);
2351 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2352 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2354 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2355 memcg
->nocpu_base
.count
[i
] += x
;
2357 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2358 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2360 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2361 memcg
->nocpu_base
.events
[i
] += x
;
2363 spin_unlock(&memcg
->pcp_counter_lock
);
2366 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2367 unsigned long action
,
2370 int cpu
= (unsigned long)hcpu
;
2371 struct memcg_stock_pcp
*stock
;
2372 struct mem_cgroup
*iter
;
2374 if (action
== CPU_ONLINE
)
2377 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2380 for_each_mem_cgroup(iter
)
2381 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2383 stock
= &per_cpu(memcg_stock
, cpu
);
2389 /* See __mem_cgroup_try_charge() for details */
2391 CHARGE_OK
, /* success */
2392 CHARGE_RETRY
, /* need to retry but retry is not bad */
2393 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2394 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2395 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2398 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2399 unsigned int nr_pages
, unsigned int min_pages
,
2402 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2403 struct mem_cgroup
*mem_over_limit
;
2404 struct res_counter
*fail_res
;
2405 unsigned long flags
= 0;
2408 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2411 if (!do_swap_account
)
2413 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2417 res_counter_uncharge(&memcg
->res
, csize
);
2418 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2419 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2421 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2423 * Never reclaim on behalf of optional batching, retry with a
2424 * single page instead.
2426 if (nr_pages
> min_pages
)
2427 return CHARGE_RETRY
;
2429 if (!(gfp_mask
& __GFP_WAIT
))
2430 return CHARGE_WOULDBLOCK
;
2432 if (gfp_mask
& __GFP_NORETRY
)
2433 return CHARGE_NOMEM
;
2435 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2436 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2437 return CHARGE_RETRY
;
2439 * Even though the limit is exceeded at this point, reclaim
2440 * may have been able to free some pages. Retry the charge
2441 * before killing the task.
2443 * Only for regular pages, though: huge pages are rather
2444 * unlikely to succeed so close to the limit, and we fall back
2445 * to regular pages anyway in case of failure.
2447 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2448 return CHARGE_RETRY
;
2451 * At task move, charge accounts can be doubly counted. So, it's
2452 * better to wait until the end of task_move if something is going on.
2454 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2455 return CHARGE_RETRY
;
2457 /* If we don't need to call oom-killer at el, return immediately */
2459 return CHARGE_NOMEM
;
2461 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2462 return CHARGE_OOM_DIE
;
2464 return CHARGE_RETRY
;
2468 * __mem_cgroup_try_charge() does
2469 * 1. detect memcg to be charged against from passed *mm and *ptr,
2470 * 2. update res_counter
2471 * 3. call memory reclaim if necessary.
2473 * In some special case, if the task is fatal, fatal_signal_pending() or
2474 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2475 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2476 * as possible without any hazards. 2: all pages should have a valid
2477 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2478 * pointer, that is treated as a charge to root_mem_cgroup.
2480 * So __mem_cgroup_try_charge() will return
2481 * 0 ... on success, filling *ptr with a valid memcg pointer.
2482 * -ENOMEM ... charge failure because of resource limits.
2483 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2485 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2486 * the oom-killer can be invoked.
2488 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2490 unsigned int nr_pages
,
2491 struct mem_cgroup
**ptr
,
2494 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2495 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2496 struct mem_cgroup
*memcg
= NULL
;
2500 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2501 * in system level. So, allow to go ahead dying process in addition to
2504 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2505 || fatal_signal_pending(current
)))
2509 * We always charge the cgroup the mm_struct belongs to.
2510 * The mm_struct's mem_cgroup changes on task migration if the
2511 * thread group leader migrates. It's possible that mm is not
2512 * set, if so charge the root memcg (happens for pagecache usage).
2515 *ptr
= root_mem_cgroup
;
2517 if (*ptr
) { /* css should be a valid one */
2519 if (mem_cgroup_is_root(memcg
))
2521 if (consume_stock(memcg
, nr_pages
))
2523 css_get(&memcg
->css
);
2525 struct task_struct
*p
;
2528 p
= rcu_dereference(mm
->owner
);
2530 * Because we don't have task_lock(), "p" can exit.
2531 * In that case, "memcg" can point to root or p can be NULL with
2532 * race with swapoff. Then, we have small risk of mis-accouning.
2533 * But such kind of mis-account by race always happens because
2534 * we don't have cgroup_mutex(). It's overkill and we allo that
2536 * (*) swapoff at el will charge against mm-struct not against
2537 * task-struct. So, mm->owner can be NULL.
2539 memcg
= mem_cgroup_from_task(p
);
2541 memcg
= root_mem_cgroup
;
2542 if (mem_cgroup_is_root(memcg
)) {
2546 if (consume_stock(memcg
, nr_pages
)) {
2548 * It seems dagerous to access memcg without css_get().
2549 * But considering how consume_stok works, it's not
2550 * necessary. If consume_stock success, some charges
2551 * from this memcg are cached on this cpu. So, we
2552 * don't need to call css_get()/css_tryget() before
2553 * calling consume_stock().
2558 /* after here, we may be blocked. we need to get refcnt */
2559 if (!css_tryget(&memcg
->css
)) {
2569 /* If killed, bypass charge */
2570 if (fatal_signal_pending(current
)) {
2571 css_put(&memcg
->css
);
2576 if (oom
&& !nr_oom_retries
) {
2578 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2581 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2586 case CHARGE_RETRY
: /* not in OOM situation but retry */
2588 css_put(&memcg
->css
);
2591 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2592 css_put(&memcg
->css
);
2594 case CHARGE_NOMEM
: /* OOM routine works */
2596 css_put(&memcg
->css
);
2599 /* If oom, we never return -ENOMEM */
2602 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2603 css_put(&memcg
->css
);
2606 } while (ret
!= CHARGE_OK
);
2608 if (batch
> nr_pages
)
2609 refill_stock(memcg
, batch
- nr_pages
);
2610 css_put(&memcg
->css
);
2618 *ptr
= root_mem_cgroup
;
2623 * Somemtimes we have to undo a charge we got by try_charge().
2624 * This function is for that and do uncharge, put css's refcnt.
2625 * gotten by try_charge().
2627 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2628 unsigned int nr_pages
)
2630 if (!mem_cgroup_is_root(memcg
)) {
2631 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2633 res_counter_uncharge(&memcg
->res
, bytes
);
2634 if (do_swap_account
)
2635 res_counter_uncharge(&memcg
->memsw
, bytes
);
2640 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2641 * This is useful when moving usage to parent cgroup.
2643 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2644 unsigned int nr_pages
)
2646 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2648 if (mem_cgroup_is_root(memcg
))
2651 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2652 if (do_swap_account
)
2653 res_counter_uncharge_until(&memcg
->memsw
,
2654 memcg
->memsw
.parent
, bytes
);
2658 * A helper function to get mem_cgroup from ID. must be called under
2659 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2660 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2661 * called against removed memcg.)
2663 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2665 struct cgroup_subsys_state
*css
;
2667 /* ID 0 is unused ID */
2670 css
= css_lookup(&mem_cgroup_subsys
, id
);
2673 return mem_cgroup_from_css(css
);
2676 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2678 struct mem_cgroup
*memcg
= NULL
;
2679 struct page_cgroup
*pc
;
2683 VM_BUG_ON(!PageLocked(page
));
2685 pc
= lookup_page_cgroup(page
);
2686 lock_page_cgroup(pc
);
2687 if (PageCgroupUsed(pc
)) {
2688 memcg
= pc
->mem_cgroup
;
2689 if (memcg
&& !css_tryget(&memcg
->css
))
2691 } else if (PageSwapCache(page
)) {
2692 ent
.val
= page_private(page
);
2693 id
= lookup_swap_cgroup_id(ent
);
2695 memcg
= mem_cgroup_lookup(id
);
2696 if (memcg
&& !css_tryget(&memcg
->css
))
2700 unlock_page_cgroup(pc
);
2704 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2706 unsigned int nr_pages
,
2707 enum charge_type ctype
,
2710 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2711 struct zone
*uninitialized_var(zone
);
2712 struct lruvec
*lruvec
;
2713 bool was_on_lru
= false;
2716 lock_page_cgroup(pc
);
2717 VM_BUG_ON(PageCgroupUsed(pc
));
2719 * we don't need page_cgroup_lock about tail pages, becase they are not
2720 * accessed by any other context at this point.
2724 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2725 * may already be on some other mem_cgroup's LRU. Take care of it.
2728 zone
= page_zone(page
);
2729 spin_lock_irq(&zone
->lru_lock
);
2730 if (PageLRU(page
)) {
2731 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2733 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2738 pc
->mem_cgroup
= memcg
;
2740 * We access a page_cgroup asynchronously without lock_page_cgroup().
2741 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2742 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2743 * before USED bit, we need memory barrier here.
2744 * See mem_cgroup_add_lru_list(), etc.
2747 SetPageCgroupUsed(pc
);
2751 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2752 VM_BUG_ON(PageLRU(page
));
2754 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2756 spin_unlock_irq(&zone
->lru_lock
);
2759 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2764 mem_cgroup_charge_statistics(memcg
, anon
, nr_pages
);
2765 unlock_page_cgroup(pc
);
2768 * "charge_statistics" updated event counter. Then, check it.
2769 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2770 * if they exceeds softlimit.
2772 memcg_check_events(memcg
, page
);
2775 static DEFINE_MUTEX(set_limit_mutex
);
2777 #ifdef CONFIG_MEMCG_KMEM
2778 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2780 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2781 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2785 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2786 * in the memcg_cache_params struct.
2788 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2790 struct kmem_cache
*cachep
;
2792 VM_BUG_ON(p
->is_root_cache
);
2793 cachep
= p
->root_cache
;
2794 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2797 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2799 struct res_counter
*fail_res
;
2800 struct mem_cgroup
*_memcg
;
2804 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2809 * Conditions under which we can wait for the oom_killer. Those are
2810 * the same conditions tested by the core page allocator
2812 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2815 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2818 if (ret
== -EINTR
) {
2820 * __mem_cgroup_try_charge() chosed to bypass to root due to
2821 * OOM kill or fatal signal. Since our only options are to
2822 * either fail the allocation or charge it to this cgroup, do
2823 * it as a temporary condition. But we can't fail. From a
2824 * kmem/slab perspective, the cache has already been selected,
2825 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2828 * This condition will only trigger if the task entered
2829 * memcg_charge_kmem in a sane state, but was OOM-killed during
2830 * __mem_cgroup_try_charge() above. Tasks that were already
2831 * dying when the allocation triggers should have been already
2832 * directed to the root cgroup in memcontrol.h
2834 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2835 if (do_swap_account
)
2836 res_counter_charge_nofail(&memcg
->memsw
, size
,
2840 res_counter_uncharge(&memcg
->kmem
, size
);
2845 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2847 res_counter_uncharge(&memcg
->res
, size
);
2848 if (do_swap_account
)
2849 res_counter_uncharge(&memcg
->memsw
, size
);
2852 if (res_counter_uncharge(&memcg
->kmem
, size
))
2855 if (memcg_kmem_test_and_clear_dead(memcg
))
2856 mem_cgroup_put(memcg
);
2859 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2864 mutex_lock(&memcg
->slab_caches_mutex
);
2865 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2866 mutex_unlock(&memcg
->slab_caches_mutex
);
2870 * helper for acessing a memcg's index. It will be used as an index in the
2871 * child cache array in kmem_cache, and also to derive its name. This function
2872 * will return -1 when this is not a kmem-limited memcg.
2874 int memcg_cache_id(struct mem_cgroup
*memcg
)
2876 return memcg
? memcg
->kmemcg_id
: -1;
2880 * This ends up being protected by the set_limit mutex, during normal
2881 * operation, because that is its main call site.
2883 * But when we create a new cache, we can call this as well if its parent
2884 * is kmem-limited. That will have to hold set_limit_mutex as well.
2886 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2890 num
= ida_simple_get(&kmem_limited_groups
,
2891 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2895 * After this point, kmem_accounted (that we test atomically in
2896 * the beginning of this conditional), is no longer 0. This
2897 * guarantees only one process will set the following boolean
2898 * to true. We don't need test_and_set because we're protected
2899 * by the set_limit_mutex anyway.
2901 memcg_kmem_set_activated(memcg
);
2903 ret
= memcg_update_all_caches(num
+1);
2905 ida_simple_remove(&kmem_limited_groups
, num
);
2906 memcg_kmem_clear_activated(memcg
);
2910 memcg
->kmemcg_id
= num
;
2911 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
2912 mutex_init(&memcg
->slab_caches_mutex
);
2916 static size_t memcg_caches_array_size(int num_groups
)
2919 if (num_groups
<= 0)
2922 size
= 2 * num_groups
;
2923 if (size
< MEMCG_CACHES_MIN_SIZE
)
2924 size
= MEMCG_CACHES_MIN_SIZE
;
2925 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2926 size
= MEMCG_CACHES_MAX_SIZE
;
2932 * We should update the current array size iff all caches updates succeed. This
2933 * can only be done from the slab side. The slab mutex needs to be held when
2936 void memcg_update_array_size(int num
)
2938 if (num
> memcg_limited_groups_array_size
)
2939 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2942 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2944 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2946 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
2948 if (num_groups
> memcg_limited_groups_array_size
) {
2950 ssize_t size
= memcg_caches_array_size(num_groups
);
2952 size
*= sizeof(void *);
2953 size
+= sizeof(struct memcg_cache_params
);
2955 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2956 if (!s
->memcg_params
) {
2957 s
->memcg_params
= cur_params
;
2961 s
->memcg_params
->is_root_cache
= true;
2964 * There is the chance it will be bigger than
2965 * memcg_limited_groups_array_size, if we failed an allocation
2966 * in a cache, in which case all caches updated before it, will
2967 * have a bigger array.
2969 * But if that is the case, the data after
2970 * memcg_limited_groups_array_size is certainly unused
2972 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
2973 if (!cur_params
->memcg_caches
[i
])
2975 s
->memcg_params
->memcg_caches
[i
] =
2976 cur_params
->memcg_caches
[i
];
2980 * Ideally, we would wait until all caches succeed, and only
2981 * then free the old one. But this is not worth the extra
2982 * pointer per-cache we'd have to have for this.
2984 * It is not a big deal if some caches are left with a size
2985 * bigger than the others. And all updates will reset this
2993 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
)
2995 size_t size
= sizeof(struct memcg_cache_params
);
2997 if (!memcg_kmem_enabled())
3001 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3003 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3004 if (!s
->memcg_params
)
3008 s
->memcg_params
->memcg
= memcg
;
3012 void memcg_release_cache(struct kmem_cache
*s
)
3014 struct kmem_cache
*root
;
3015 struct mem_cgroup
*memcg
;
3019 * This happens, for instance, when a root cache goes away before we
3022 if (!s
->memcg_params
)
3025 if (s
->memcg_params
->is_root_cache
)
3028 memcg
= s
->memcg_params
->memcg
;
3029 id
= memcg_cache_id(memcg
);
3031 root
= s
->memcg_params
->root_cache
;
3032 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3033 mem_cgroup_put(memcg
);
3035 mutex_lock(&memcg
->slab_caches_mutex
);
3036 list_del(&s
->memcg_params
->list
);
3037 mutex_unlock(&memcg
->slab_caches_mutex
);
3040 kfree(s
->memcg_params
);
3044 * During the creation a new cache, we need to disable our accounting mechanism
3045 * altogether. This is true even if we are not creating, but rather just
3046 * enqueing new caches to be created.
3048 * This is because that process will trigger allocations; some visible, like
3049 * explicit kmallocs to auxiliary data structures, name strings and internal
3050 * cache structures; some well concealed, like INIT_WORK() that can allocate
3051 * objects during debug.
3053 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3054 * to it. This may not be a bounded recursion: since the first cache creation
3055 * failed to complete (waiting on the allocation), we'll just try to create the
3056 * cache again, failing at the same point.
3058 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3059 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3060 * inside the following two functions.
3062 static inline void memcg_stop_kmem_account(void)
3064 VM_BUG_ON(!current
->mm
);
3065 current
->memcg_kmem_skip_account
++;
3068 static inline void memcg_resume_kmem_account(void)
3070 VM_BUG_ON(!current
->mm
);
3071 current
->memcg_kmem_skip_account
--;
3074 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3076 struct kmem_cache
*cachep
;
3077 struct memcg_cache_params
*p
;
3079 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3081 cachep
= memcg_params_to_cache(p
);
3084 * If we get down to 0 after shrink, we could delete right away.
3085 * However, memcg_release_pages() already puts us back in the workqueue
3086 * in that case. If we proceed deleting, we'll get a dangling
3087 * reference, and removing the object from the workqueue in that case
3088 * is unnecessary complication. We are not a fast path.
3090 * Note that this case is fundamentally different from racing with
3091 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3092 * kmem_cache_shrink, not only we would be reinserting a dead cache
3093 * into the queue, but doing so from inside the worker racing to
3096 * So if we aren't down to zero, we'll just schedule a worker and try
3099 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3100 kmem_cache_shrink(cachep
);
3101 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3104 kmem_cache_destroy(cachep
);
3107 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3109 if (!cachep
->memcg_params
->dead
)
3113 * There are many ways in which we can get here.
3115 * We can get to a memory-pressure situation while the delayed work is
3116 * still pending to run. The vmscan shrinkers can then release all
3117 * cache memory and get us to destruction. If this is the case, we'll
3118 * be executed twice, which is a bug (the second time will execute over
3119 * bogus data). In this case, cancelling the work should be fine.
3121 * But we can also get here from the worker itself, if
3122 * kmem_cache_shrink is enough to shake all the remaining objects and
3123 * get the page count to 0. In this case, we'll deadlock if we try to
3124 * cancel the work (the worker runs with an internal lock held, which
3125 * is the same lock we would hold for cancel_work_sync().)
3127 * Since we can't possibly know who got us here, just refrain from
3128 * running if there is already work pending
3130 if (work_pending(&cachep
->memcg_params
->destroy
))
3133 * We have to defer the actual destroying to a workqueue, because
3134 * we might currently be in a context that cannot sleep.
3136 schedule_work(&cachep
->memcg_params
->destroy
);
3139 static char *memcg_cache_name(struct mem_cgroup
*memcg
, struct kmem_cache
*s
)
3142 struct dentry
*dentry
;
3145 dentry
= rcu_dereference(memcg
->css
.cgroup
->dentry
);
3148 BUG_ON(dentry
== NULL
);
3150 name
= kasprintf(GFP_KERNEL
, "%s(%d:%s)", s
->name
,
3151 memcg_cache_id(memcg
), dentry
->d_name
.name
);
3156 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3157 struct kmem_cache
*s
)
3160 struct kmem_cache
*new;
3162 name
= memcg_cache_name(memcg
, s
);
3166 new = kmem_cache_create_memcg(memcg
, name
, s
->object_size
, s
->align
,
3167 (s
->flags
& ~SLAB_PANIC
), s
->ctor
);
3170 new->allocflags
|= __GFP_KMEMCG
;
3177 * This lock protects updaters, not readers. We want readers to be as fast as
3178 * they can, and they will either see NULL or a valid cache value. Our model
3179 * allow them to see NULL, in which case the root memcg will be selected.
3181 * We need this lock because multiple allocations to the same cache from a non
3182 * will span more than one worker. Only one of them can create the cache.
3184 static DEFINE_MUTEX(memcg_cache_mutex
);
3185 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3186 struct kmem_cache
*cachep
)
3188 struct kmem_cache
*new_cachep
;
3191 BUG_ON(!memcg_can_account_kmem(memcg
));
3193 idx
= memcg_cache_id(memcg
);
3195 mutex_lock(&memcg_cache_mutex
);
3196 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3200 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3201 if (new_cachep
== NULL
) {
3202 new_cachep
= cachep
;
3206 mem_cgroup_get(memcg
);
3207 new_cachep
->memcg_params
->root_cache
= cachep
;
3208 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3210 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3212 * the readers won't lock, make sure everybody sees the updated value,
3213 * so they won't put stuff in the queue again for no reason
3217 mutex_unlock(&memcg_cache_mutex
);
3221 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3223 struct kmem_cache
*c
;
3226 if (!s
->memcg_params
)
3228 if (!s
->memcg_params
->is_root_cache
)
3232 * If the cache is being destroyed, we trust that there is no one else
3233 * requesting objects from it. Even if there are, the sanity checks in
3234 * kmem_cache_destroy should caught this ill-case.
3236 * Still, we don't want anyone else freeing memcg_caches under our
3237 * noses, which can happen if a new memcg comes to life. As usual,
3238 * we'll take the set_limit_mutex to protect ourselves against this.
3240 mutex_lock(&set_limit_mutex
);
3241 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3242 c
= s
->memcg_params
->memcg_caches
[i
];
3247 * We will now manually delete the caches, so to avoid races
3248 * we need to cancel all pending destruction workers and
3249 * proceed with destruction ourselves.
3251 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3252 * and that could spawn the workers again: it is likely that
3253 * the cache still have active pages until this very moment.
3254 * This would lead us back to mem_cgroup_destroy_cache.
3256 * But that will not execute at all if the "dead" flag is not
3257 * set, so flip it down to guarantee we are in control.
3259 c
->memcg_params
->dead
= false;
3260 cancel_work_sync(&c
->memcg_params
->destroy
);
3261 kmem_cache_destroy(c
);
3263 mutex_unlock(&set_limit_mutex
);
3266 struct create_work
{
3267 struct mem_cgroup
*memcg
;
3268 struct kmem_cache
*cachep
;
3269 struct work_struct work
;
3272 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3274 struct kmem_cache
*cachep
;
3275 struct memcg_cache_params
*params
;
3277 if (!memcg_kmem_is_active(memcg
))
3280 mutex_lock(&memcg
->slab_caches_mutex
);
3281 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3282 cachep
= memcg_params_to_cache(params
);
3283 cachep
->memcg_params
->dead
= true;
3284 INIT_WORK(&cachep
->memcg_params
->destroy
,
3285 kmem_cache_destroy_work_func
);
3286 schedule_work(&cachep
->memcg_params
->destroy
);
3288 mutex_unlock(&memcg
->slab_caches_mutex
);
3291 static void memcg_create_cache_work_func(struct work_struct
*w
)
3293 struct create_work
*cw
;
3295 cw
= container_of(w
, struct create_work
, work
);
3296 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3297 /* Drop the reference gotten when we enqueued. */
3298 css_put(&cw
->memcg
->css
);
3303 * Enqueue the creation of a per-memcg kmem_cache.
3304 * Called with rcu_read_lock.
3306 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3307 struct kmem_cache
*cachep
)
3309 struct create_work
*cw
;
3311 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3315 /* The corresponding put will be done in the workqueue. */
3316 if (!css_tryget(&memcg
->css
)) {
3322 cw
->cachep
= cachep
;
3324 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3325 schedule_work(&cw
->work
);
3328 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3329 struct kmem_cache
*cachep
)
3332 * We need to stop accounting when we kmalloc, because if the
3333 * corresponding kmalloc cache is not yet created, the first allocation
3334 * in __memcg_create_cache_enqueue will recurse.
3336 * However, it is better to enclose the whole function. Depending on
3337 * the debugging options enabled, INIT_WORK(), for instance, can
3338 * trigger an allocation. This too, will make us recurse. Because at
3339 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3340 * the safest choice is to do it like this, wrapping the whole function.
3342 memcg_stop_kmem_account();
3343 __memcg_create_cache_enqueue(memcg
, cachep
);
3344 memcg_resume_kmem_account();
3347 * Return the kmem_cache we're supposed to use for a slab allocation.
3348 * We try to use the current memcg's version of the cache.
3350 * If the cache does not exist yet, if we are the first user of it,
3351 * we either create it immediately, if possible, or create it asynchronously
3353 * In the latter case, we will let the current allocation go through with
3354 * the original cache.
3356 * Can't be called in interrupt context or from kernel threads.
3357 * This function needs to be called with rcu_read_lock() held.
3359 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3362 struct mem_cgroup
*memcg
;
3365 VM_BUG_ON(!cachep
->memcg_params
);
3366 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3368 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3372 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3375 if (!memcg_can_account_kmem(memcg
))
3378 idx
= memcg_cache_id(memcg
);
3381 * barrier to mare sure we're always seeing the up to date value. The
3382 * code updating memcg_caches will issue a write barrier to match this.
3384 read_barrier_depends();
3385 if (unlikely(cachep
->memcg_params
->memcg_caches
[idx
] == NULL
)) {
3387 * If we are in a safe context (can wait, and not in interrupt
3388 * context), we could be be predictable and return right away.
3389 * This would guarantee that the allocation being performed
3390 * already belongs in the new cache.
3392 * However, there are some clashes that can arrive from locking.
3393 * For instance, because we acquire the slab_mutex while doing
3394 * kmem_cache_dup, this means no further allocation could happen
3395 * with the slab_mutex held.
3397 * Also, because cache creation issue get_online_cpus(), this
3398 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3399 * that ends up reversed during cpu hotplug. (cpuset allocates
3400 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3401 * better to defer everything.
3403 memcg_create_cache_enqueue(memcg
, cachep
);
3407 return cachep
->memcg_params
->memcg_caches
[idx
];
3409 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3412 * We need to verify if the allocation against current->mm->owner's memcg is
3413 * possible for the given order. But the page is not allocated yet, so we'll
3414 * need a further commit step to do the final arrangements.
3416 * It is possible for the task to switch cgroups in this mean time, so at
3417 * commit time, we can't rely on task conversion any longer. We'll then use
3418 * the handle argument to return to the caller which cgroup we should commit
3419 * against. We could also return the memcg directly and avoid the pointer
3420 * passing, but a boolean return value gives better semantics considering
3421 * the compiled-out case as well.
3423 * Returning true means the allocation is possible.
3426 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3428 struct mem_cgroup
*memcg
;
3432 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3435 * very rare case described in mem_cgroup_from_task. Unfortunately there
3436 * isn't much we can do without complicating this too much, and it would
3437 * be gfp-dependent anyway. Just let it go
3439 if (unlikely(!memcg
))
3442 if (!memcg_can_account_kmem(memcg
)) {
3443 css_put(&memcg
->css
);
3447 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3451 css_put(&memcg
->css
);
3455 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3458 struct page_cgroup
*pc
;
3460 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3462 /* The page allocation failed. Revert */
3464 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3468 pc
= lookup_page_cgroup(page
);
3469 lock_page_cgroup(pc
);
3470 pc
->mem_cgroup
= memcg
;
3471 SetPageCgroupUsed(pc
);
3472 unlock_page_cgroup(pc
);
3475 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3477 struct mem_cgroup
*memcg
= NULL
;
3478 struct page_cgroup
*pc
;
3481 pc
= lookup_page_cgroup(page
);
3483 * Fast unlocked return. Theoretically might have changed, have to
3484 * check again after locking.
3486 if (!PageCgroupUsed(pc
))
3489 lock_page_cgroup(pc
);
3490 if (PageCgroupUsed(pc
)) {
3491 memcg
= pc
->mem_cgroup
;
3492 ClearPageCgroupUsed(pc
);
3494 unlock_page_cgroup(pc
);
3497 * We trust that only if there is a memcg associated with the page, it
3498 * is a valid allocation
3503 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3504 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3507 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3510 #endif /* CONFIG_MEMCG_KMEM */
3512 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3514 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3516 * Because tail pages are not marked as "used", set it. We're under
3517 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3518 * charge/uncharge will be never happen and move_account() is done under
3519 * compound_lock(), so we don't have to take care of races.
3521 void mem_cgroup_split_huge_fixup(struct page
*head
)
3523 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3524 struct page_cgroup
*pc
;
3527 if (mem_cgroup_disabled())
3529 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3531 pc
->mem_cgroup
= head_pc
->mem_cgroup
;
3532 smp_wmb();/* see __commit_charge() */
3533 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3536 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3539 * mem_cgroup_move_account - move account of the page
3541 * @nr_pages: number of regular pages (>1 for huge pages)
3542 * @pc: page_cgroup of the page.
3543 * @from: mem_cgroup which the page is moved from.
3544 * @to: mem_cgroup which the page is moved to. @from != @to.
3546 * The caller must confirm following.
3547 * - page is not on LRU (isolate_page() is useful.)
3548 * - compound_lock is held when nr_pages > 1
3550 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3553 static int mem_cgroup_move_account(struct page
*page
,
3554 unsigned int nr_pages
,
3555 struct page_cgroup
*pc
,
3556 struct mem_cgroup
*from
,
3557 struct mem_cgroup
*to
)
3559 unsigned long flags
;
3561 bool anon
= PageAnon(page
);
3563 VM_BUG_ON(from
== to
);
3564 VM_BUG_ON(PageLRU(page
));
3566 * The page is isolated from LRU. So, collapse function
3567 * will not handle this page. But page splitting can happen.
3568 * Do this check under compound_page_lock(). The caller should
3572 if (nr_pages
> 1 && !PageTransHuge(page
))
3575 lock_page_cgroup(pc
);
3578 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3581 move_lock_mem_cgroup(from
, &flags
);
3583 if (!anon
&& page_mapped(page
)) {
3584 /* Update mapped_file data for mem_cgroup */
3586 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3587 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3590 mem_cgroup_charge_statistics(from
, anon
, -nr_pages
);
3592 /* caller should have done css_get */
3593 pc
->mem_cgroup
= to
;
3594 mem_cgroup_charge_statistics(to
, anon
, nr_pages
);
3595 move_unlock_mem_cgroup(from
, &flags
);
3598 unlock_page_cgroup(pc
);
3602 memcg_check_events(to
, page
);
3603 memcg_check_events(from
, page
);
3609 * mem_cgroup_move_parent - moves page to the parent group
3610 * @page: the page to move
3611 * @pc: page_cgroup of the page
3612 * @child: page's cgroup
3614 * move charges to its parent or the root cgroup if the group has no
3615 * parent (aka use_hierarchy==0).
3616 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3617 * mem_cgroup_move_account fails) the failure is always temporary and
3618 * it signals a race with a page removal/uncharge or migration. In the
3619 * first case the page is on the way out and it will vanish from the LRU
3620 * on the next attempt and the call should be retried later.
3621 * Isolation from the LRU fails only if page has been isolated from
3622 * the LRU since we looked at it and that usually means either global
3623 * reclaim or migration going on. The page will either get back to the
3625 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3626 * (!PageCgroupUsed) or moved to a different group. The page will
3627 * disappear in the next attempt.
3629 static int mem_cgroup_move_parent(struct page
*page
,
3630 struct page_cgroup
*pc
,
3631 struct mem_cgroup
*child
)
3633 struct mem_cgroup
*parent
;
3634 unsigned int nr_pages
;
3635 unsigned long uninitialized_var(flags
);
3638 VM_BUG_ON(mem_cgroup_is_root(child
));
3641 if (!get_page_unless_zero(page
))
3643 if (isolate_lru_page(page
))
3646 nr_pages
= hpage_nr_pages(page
);
3648 parent
= parent_mem_cgroup(child
);
3650 * If no parent, move charges to root cgroup.
3653 parent
= root_mem_cgroup
;
3656 VM_BUG_ON(!PageTransHuge(page
));
3657 flags
= compound_lock_irqsave(page
);
3660 ret
= mem_cgroup_move_account(page
, nr_pages
,
3663 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3666 compound_unlock_irqrestore(page
, flags
);
3667 putback_lru_page(page
);
3675 * Charge the memory controller for page usage.
3677 * 0 if the charge was successful
3678 * < 0 if the cgroup is over its limit
3680 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3681 gfp_t gfp_mask
, enum charge_type ctype
)
3683 struct mem_cgroup
*memcg
= NULL
;
3684 unsigned int nr_pages
= 1;
3688 if (PageTransHuge(page
)) {
3689 nr_pages
<<= compound_order(page
);
3690 VM_BUG_ON(!PageTransHuge(page
));
3692 * Never OOM-kill a process for a huge page. The
3693 * fault handler will fall back to regular pages.
3698 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3701 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3705 int mem_cgroup_newpage_charge(struct page
*page
,
3706 struct mm_struct
*mm
, gfp_t gfp_mask
)
3708 if (mem_cgroup_disabled())
3710 VM_BUG_ON(page_mapped(page
));
3711 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3713 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3714 MEM_CGROUP_CHARGE_TYPE_ANON
);
3718 * While swap-in, try_charge -> commit or cancel, the page is locked.
3719 * And when try_charge() successfully returns, one refcnt to memcg without
3720 * struct page_cgroup is acquired. This refcnt will be consumed by
3721 * "commit()" or removed by "cancel()"
3723 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3726 struct mem_cgroup
**memcgp
)
3728 struct mem_cgroup
*memcg
;
3729 struct page_cgroup
*pc
;
3732 pc
= lookup_page_cgroup(page
);
3734 * Every swap fault against a single page tries to charge the
3735 * page, bail as early as possible. shmem_unuse() encounters
3736 * already charged pages, too. The USED bit is protected by
3737 * the page lock, which serializes swap cache removal, which
3738 * in turn serializes uncharging.
3740 if (PageCgroupUsed(pc
))
3742 if (!do_swap_account
)
3744 memcg
= try_get_mem_cgroup_from_page(page
);
3748 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3749 css_put(&memcg
->css
);
3754 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3760 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3761 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3764 if (mem_cgroup_disabled())
3767 * A racing thread's fault, or swapoff, may have already
3768 * updated the pte, and even removed page from swap cache: in
3769 * those cases unuse_pte()'s pte_same() test will fail; but
3770 * there's also a KSM case which does need to charge the page.
3772 if (!PageSwapCache(page
)) {
3775 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3780 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3783 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3785 if (mem_cgroup_disabled())
3789 __mem_cgroup_cancel_charge(memcg
, 1);
3793 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3794 enum charge_type ctype
)
3796 if (mem_cgroup_disabled())
3801 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3803 * Now swap is on-memory. This means this page may be
3804 * counted both as mem and swap....double count.
3805 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3806 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3807 * may call delete_from_swap_cache() before reach here.
3809 if (do_swap_account
&& PageSwapCache(page
)) {
3810 swp_entry_t ent
= {.val
= page_private(page
)};
3811 mem_cgroup_uncharge_swap(ent
);
3815 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3816 struct mem_cgroup
*memcg
)
3818 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3819 MEM_CGROUP_CHARGE_TYPE_ANON
);
3822 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3825 struct mem_cgroup
*memcg
= NULL
;
3826 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3829 if (mem_cgroup_disabled())
3831 if (PageCompound(page
))
3834 if (!PageSwapCache(page
))
3835 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3836 else { /* page is swapcache/shmem */
3837 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3840 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3845 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3846 unsigned int nr_pages
,
3847 const enum charge_type ctype
)
3849 struct memcg_batch_info
*batch
= NULL
;
3850 bool uncharge_memsw
= true;
3852 /* If swapout, usage of swap doesn't decrease */
3853 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3854 uncharge_memsw
= false;
3856 batch
= ¤t
->memcg_batch
;
3858 * In usual, we do css_get() when we remember memcg pointer.
3859 * But in this case, we keep res->usage until end of a series of
3860 * uncharges. Then, it's ok to ignore memcg's refcnt.
3863 batch
->memcg
= memcg
;
3865 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3866 * In those cases, all pages freed continuously can be expected to be in
3867 * the same cgroup and we have chance to coalesce uncharges.
3868 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3869 * because we want to do uncharge as soon as possible.
3872 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3873 goto direct_uncharge
;
3876 goto direct_uncharge
;
3879 * In typical case, batch->memcg == mem. This means we can
3880 * merge a series of uncharges to an uncharge of res_counter.
3881 * If not, we uncharge res_counter ony by one.
3883 if (batch
->memcg
!= memcg
)
3884 goto direct_uncharge
;
3885 /* remember freed charge and uncharge it later */
3888 batch
->memsw_nr_pages
++;
3891 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3893 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3894 if (unlikely(batch
->memcg
!= memcg
))
3895 memcg_oom_recover(memcg
);
3899 * uncharge if !page_mapped(page)
3901 static struct mem_cgroup
*
3902 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3905 struct mem_cgroup
*memcg
= NULL
;
3906 unsigned int nr_pages
= 1;
3907 struct page_cgroup
*pc
;
3910 if (mem_cgroup_disabled())
3913 VM_BUG_ON(PageSwapCache(page
));
3915 if (PageTransHuge(page
)) {
3916 nr_pages
<<= compound_order(page
);
3917 VM_BUG_ON(!PageTransHuge(page
));
3920 * Check if our page_cgroup is valid
3922 pc
= lookup_page_cgroup(page
);
3923 if (unlikely(!PageCgroupUsed(pc
)))
3926 lock_page_cgroup(pc
);
3928 memcg
= pc
->mem_cgroup
;
3930 if (!PageCgroupUsed(pc
))
3933 anon
= PageAnon(page
);
3936 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3938 * Generally PageAnon tells if it's the anon statistics to be
3939 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3940 * used before page reached the stage of being marked PageAnon.
3944 case MEM_CGROUP_CHARGE_TYPE_DROP
:
3945 /* See mem_cgroup_prepare_migration() */
3946 if (page_mapped(page
))
3949 * Pages under migration may not be uncharged. But
3950 * end_migration() /must/ be the one uncharging the
3951 * unused post-migration page and so it has to call
3952 * here with the migration bit still set. See the
3953 * res_counter handling below.
3955 if (!end_migration
&& PageCgroupMigration(pc
))
3958 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
3959 if (!PageAnon(page
)) { /* Shared memory */
3960 if (page
->mapping
&& !page_is_file_cache(page
))
3962 } else if (page_mapped(page
)) /* Anon */
3969 mem_cgroup_charge_statistics(memcg
, anon
, -nr_pages
);
3971 ClearPageCgroupUsed(pc
);
3973 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3974 * freed from LRU. This is safe because uncharged page is expected not
3975 * to be reused (freed soon). Exception is SwapCache, it's handled by
3976 * special functions.
3979 unlock_page_cgroup(pc
);
3981 * even after unlock, we have memcg->res.usage here and this memcg
3982 * will never be freed.
3984 memcg_check_events(memcg
, page
);
3985 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
3986 mem_cgroup_swap_statistics(memcg
, true);
3987 mem_cgroup_get(memcg
);
3990 * Migration does not charge the res_counter for the
3991 * replacement page, so leave it alone when phasing out the
3992 * page that is unused after the migration.
3994 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
3995 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4000 unlock_page_cgroup(pc
);
4004 void mem_cgroup_uncharge_page(struct page
*page
)
4007 if (page_mapped(page
))
4009 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4010 if (PageSwapCache(page
))
4012 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4015 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4017 VM_BUG_ON(page_mapped(page
));
4018 VM_BUG_ON(page
->mapping
);
4019 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4023 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4024 * In that cases, pages are freed continuously and we can expect pages
4025 * are in the same memcg. All these calls itself limits the number of
4026 * pages freed at once, then uncharge_start/end() is called properly.
4027 * This may be called prural(2) times in a context,
4030 void mem_cgroup_uncharge_start(void)
4032 current
->memcg_batch
.do_batch
++;
4033 /* We can do nest. */
4034 if (current
->memcg_batch
.do_batch
== 1) {
4035 current
->memcg_batch
.memcg
= NULL
;
4036 current
->memcg_batch
.nr_pages
= 0;
4037 current
->memcg_batch
.memsw_nr_pages
= 0;
4041 void mem_cgroup_uncharge_end(void)
4043 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4045 if (!batch
->do_batch
)
4049 if (batch
->do_batch
) /* If stacked, do nothing. */
4055 * This "batch->memcg" is valid without any css_get/put etc...
4056 * bacause we hide charges behind us.
4058 if (batch
->nr_pages
)
4059 res_counter_uncharge(&batch
->memcg
->res
,
4060 batch
->nr_pages
* PAGE_SIZE
);
4061 if (batch
->memsw_nr_pages
)
4062 res_counter_uncharge(&batch
->memcg
->memsw
,
4063 batch
->memsw_nr_pages
* PAGE_SIZE
);
4064 memcg_oom_recover(batch
->memcg
);
4065 /* forget this pointer (for sanity check) */
4066 batch
->memcg
= NULL
;
4071 * called after __delete_from_swap_cache() and drop "page" account.
4072 * memcg information is recorded to swap_cgroup of "ent"
4075 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4077 struct mem_cgroup
*memcg
;
4078 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4080 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4081 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4083 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4086 * record memcg information, if swapout && memcg != NULL,
4087 * mem_cgroup_get() was called in uncharge().
4089 if (do_swap_account
&& swapout
&& memcg
)
4090 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4094 #ifdef CONFIG_MEMCG_SWAP
4096 * called from swap_entry_free(). remove record in swap_cgroup and
4097 * uncharge "memsw" account.
4099 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4101 struct mem_cgroup
*memcg
;
4104 if (!do_swap_account
)
4107 id
= swap_cgroup_record(ent
, 0);
4109 memcg
= mem_cgroup_lookup(id
);
4112 * We uncharge this because swap is freed.
4113 * This memcg can be obsolete one. We avoid calling css_tryget
4115 if (!mem_cgroup_is_root(memcg
))
4116 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4117 mem_cgroup_swap_statistics(memcg
, false);
4118 mem_cgroup_put(memcg
);
4124 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4125 * @entry: swap entry to be moved
4126 * @from: mem_cgroup which the entry is moved from
4127 * @to: mem_cgroup which the entry is moved to
4129 * It succeeds only when the swap_cgroup's record for this entry is the same
4130 * as the mem_cgroup's id of @from.
4132 * Returns 0 on success, -EINVAL on failure.
4134 * The caller must have charged to @to, IOW, called res_counter_charge() about
4135 * both res and memsw, and called css_get().
4137 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4138 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4140 unsigned short old_id
, new_id
;
4142 old_id
= css_id(&from
->css
);
4143 new_id
= css_id(&to
->css
);
4145 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4146 mem_cgroup_swap_statistics(from
, false);
4147 mem_cgroup_swap_statistics(to
, true);
4149 * This function is only called from task migration context now.
4150 * It postpones res_counter and refcount handling till the end
4151 * of task migration(mem_cgroup_clear_mc()) for performance
4152 * improvement. But we cannot postpone mem_cgroup_get(to)
4153 * because if the process that has been moved to @to does
4154 * swap-in, the refcount of @to might be decreased to 0.
4162 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4163 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4170 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4173 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4174 struct mem_cgroup
**memcgp
)
4176 struct mem_cgroup
*memcg
= NULL
;
4177 unsigned int nr_pages
= 1;
4178 struct page_cgroup
*pc
;
4179 enum charge_type ctype
;
4183 if (mem_cgroup_disabled())
4186 if (PageTransHuge(page
))
4187 nr_pages
<<= compound_order(page
);
4189 pc
= lookup_page_cgroup(page
);
4190 lock_page_cgroup(pc
);
4191 if (PageCgroupUsed(pc
)) {
4192 memcg
= pc
->mem_cgroup
;
4193 css_get(&memcg
->css
);
4195 * At migrating an anonymous page, its mapcount goes down
4196 * to 0 and uncharge() will be called. But, even if it's fully
4197 * unmapped, migration may fail and this page has to be
4198 * charged again. We set MIGRATION flag here and delay uncharge
4199 * until end_migration() is called
4201 * Corner Case Thinking
4203 * When the old page was mapped as Anon and it's unmap-and-freed
4204 * while migration was ongoing.
4205 * If unmap finds the old page, uncharge() of it will be delayed
4206 * until end_migration(). If unmap finds a new page, it's
4207 * uncharged when it make mapcount to be 1->0. If unmap code
4208 * finds swap_migration_entry, the new page will not be mapped
4209 * and end_migration() will find it(mapcount==0).
4212 * When the old page was mapped but migraion fails, the kernel
4213 * remaps it. A charge for it is kept by MIGRATION flag even
4214 * if mapcount goes down to 0. We can do remap successfully
4215 * without charging it again.
4218 * The "old" page is under lock_page() until the end of
4219 * migration, so, the old page itself will not be swapped-out.
4220 * If the new page is swapped out before end_migraton, our
4221 * hook to usual swap-out path will catch the event.
4224 SetPageCgroupMigration(pc
);
4226 unlock_page_cgroup(pc
);
4228 * If the page is not charged at this point,
4236 * We charge new page before it's used/mapped. So, even if unlock_page()
4237 * is called before end_migration, we can catch all events on this new
4238 * page. In the case new page is migrated but not remapped, new page's
4239 * mapcount will be finally 0 and we call uncharge in end_migration().
4242 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4244 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4246 * The page is committed to the memcg, but it's not actually
4247 * charged to the res_counter since we plan on replacing the
4248 * old one and only one page is going to be left afterwards.
4250 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4253 /* remove redundant charge if migration failed*/
4254 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4255 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4257 struct page
*used
, *unused
;
4258 struct page_cgroup
*pc
;
4264 if (!migration_ok
) {
4271 anon
= PageAnon(used
);
4272 __mem_cgroup_uncharge_common(unused
,
4273 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4274 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4276 css_put(&memcg
->css
);
4278 * We disallowed uncharge of pages under migration because mapcount
4279 * of the page goes down to zero, temporarly.
4280 * Clear the flag and check the page should be charged.
4282 pc
= lookup_page_cgroup(oldpage
);
4283 lock_page_cgroup(pc
);
4284 ClearPageCgroupMigration(pc
);
4285 unlock_page_cgroup(pc
);
4288 * If a page is a file cache, radix-tree replacement is very atomic
4289 * and we can skip this check. When it was an Anon page, its mapcount
4290 * goes down to 0. But because we added MIGRATION flage, it's not
4291 * uncharged yet. There are several case but page->mapcount check
4292 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4293 * check. (see prepare_charge() also)
4296 mem_cgroup_uncharge_page(used
);
4300 * At replace page cache, newpage is not under any memcg but it's on
4301 * LRU. So, this function doesn't touch res_counter but handles LRU
4302 * in correct way. Both pages are locked so we cannot race with uncharge.
4304 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4305 struct page
*newpage
)
4307 struct mem_cgroup
*memcg
= NULL
;
4308 struct page_cgroup
*pc
;
4309 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4311 if (mem_cgroup_disabled())
4314 pc
= lookup_page_cgroup(oldpage
);
4315 /* fix accounting on old pages */
4316 lock_page_cgroup(pc
);
4317 if (PageCgroupUsed(pc
)) {
4318 memcg
= pc
->mem_cgroup
;
4319 mem_cgroup_charge_statistics(memcg
, false, -1);
4320 ClearPageCgroupUsed(pc
);
4322 unlock_page_cgroup(pc
);
4325 * When called from shmem_replace_page(), in some cases the
4326 * oldpage has already been charged, and in some cases not.
4331 * Even if newpage->mapping was NULL before starting replacement,
4332 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4333 * LRU while we overwrite pc->mem_cgroup.
4335 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4338 #ifdef CONFIG_DEBUG_VM
4339 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4341 struct page_cgroup
*pc
;
4343 pc
= lookup_page_cgroup(page
);
4345 * Can be NULL while feeding pages into the page allocator for
4346 * the first time, i.e. during boot or memory hotplug;
4347 * or when mem_cgroup_disabled().
4349 if (likely(pc
) && PageCgroupUsed(pc
))
4354 bool mem_cgroup_bad_page_check(struct page
*page
)
4356 if (mem_cgroup_disabled())
4359 return lookup_page_cgroup_used(page
) != NULL
;
4362 void mem_cgroup_print_bad_page(struct page
*page
)
4364 struct page_cgroup
*pc
;
4366 pc
= lookup_page_cgroup_used(page
);
4368 printk(KERN_ALERT
"pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4369 pc
, pc
->flags
, pc
->mem_cgroup
);
4374 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4375 unsigned long long val
)
4378 u64 memswlimit
, memlimit
;
4380 int children
= mem_cgroup_count_children(memcg
);
4381 u64 curusage
, oldusage
;
4385 * For keeping hierarchical_reclaim simple, how long we should retry
4386 * is depends on callers. We set our retry-count to be function
4387 * of # of children which we should visit in this loop.
4389 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4391 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4394 while (retry_count
) {
4395 if (signal_pending(current
)) {
4400 * Rather than hide all in some function, I do this in
4401 * open coded manner. You see what this really does.
4402 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4404 mutex_lock(&set_limit_mutex
);
4405 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4406 if (memswlimit
< val
) {
4408 mutex_unlock(&set_limit_mutex
);
4412 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4416 ret
= res_counter_set_limit(&memcg
->res
, val
);
4418 if (memswlimit
== val
)
4419 memcg
->memsw_is_minimum
= true;
4421 memcg
->memsw_is_minimum
= false;
4423 mutex_unlock(&set_limit_mutex
);
4428 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4429 MEM_CGROUP_RECLAIM_SHRINK
);
4430 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4431 /* Usage is reduced ? */
4432 if (curusage
>= oldusage
)
4435 oldusage
= curusage
;
4437 if (!ret
&& enlarge
)
4438 memcg_oom_recover(memcg
);
4443 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4444 unsigned long long val
)
4447 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4448 int children
= mem_cgroup_count_children(memcg
);
4452 /* see mem_cgroup_resize_res_limit */
4453 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4454 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4455 while (retry_count
) {
4456 if (signal_pending(current
)) {
4461 * Rather than hide all in some function, I do this in
4462 * open coded manner. You see what this really does.
4463 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4465 mutex_lock(&set_limit_mutex
);
4466 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4467 if (memlimit
> val
) {
4469 mutex_unlock(&set_limit_mutex
);
4472 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4473 if (memswlimit
< val
)
4475 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4477 if (memlimit
== val
)
4478 memcg
->memsw_is_minimum
= true;
4480 memcg
->memsw_is_minimum
= false;
4482 mutex_unlock(&set_limit_mutex
);
4487 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4488 MEM_CGROUP_RECLAIM_NOSWAP
|
4489 MEM_CGROUP_RECLAIM_SHRINK
);
4490 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4491 /* Usage is reduced ? */
4492 if (curusage
>= oldusage
)
4495 oldusage
= curusage
;
4497 if (!ret
&& enlarge
)
4498 memcg_oom_recover(memcg
);
4502 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4504 unsigned long *total_scanned
)
4506 unsigned long nr_reclaimed
= 0;
4507 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4508 unsigned long reclaimed
;
4510 struct mem_cgroup_tree_per_zone
*mctz
;
4511 unsigned long long excess
;
4512 unsigned long nr_scanned
;
4517 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4519 * This loop can run a while, specially if mem_cgroup's continuously
4520 * keep exceeding their soft limit and putting the system under
4527 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4532 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4533 gfp_mask
, &nr_scanned
);
4534 nr_reclaimed
+= reclaimed
;
4535 *total_scanned
+= nr_scanned
;
4536 spin_lock(&mctz
->lock
);
4539 * If we failed to reclaim anything from this memory cgroup
4540 * it is time to move on to the next cgroup
4546 * Loop until we find yet another one.
4548 * By the time we get the soft_limit lock
4549 * again, someone might have aded the
4550 * group back on the RB tree. Iterate to
4551 * make sure we get a different mem.
4552 * mem_cgroup_largest_soft_limit_node returns
4553 * NULL if no other cgroup is present on
4557 __mem_cgroup_largest_soft_limit_node(mctz
);
4559 css_put(&next_mz
->memcg
->css
);
4560 else /* next_mz == NULL or other memcg */
4564 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4565 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4567 * One school of thought says that we should not add
4568 * back the node to the tree if reclaim returns 0.
4569 * But our reclaim could return 0, simply because due
4570 * to priority we are exposing a smaller subset of
4571 * memory to reclaim from. Consider this as a longer
4574 /* If excess == 0, no tree ops */
4575 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4576 spin_unlock(&mctz
->lock
);
4577 css_put(&mz
->memcg
->css
);
4580 * Could not reclaim anything and there are no more
4581 * mem cgroups to try or we seem to be looping without
4582 * reclaiming anything.
4584 if (!nr_reclaimed
&&
4586 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4588 } while (!nr_reclaimed
);
4590 css_put(&next_mz
->memcg
->css
);
4591 return nr_reclaimed
;
4595 * mem_cgroup_force_empty_list - clears LRU of a group
4596 * @memcg: group to clear
4599 * @lru: lru to to clear
4601 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4602 * reclaim the pages page themselves - pages are moved to the parent (or root)
4605 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4606 int node
, int zid
, enum lru_list lru
)
4608 struct lruvec
*lruvec
;
4609 unsigned long flags
;
4610 struct list_head
*list
;
4614 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4615 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4616 list
= &lruvec
->lists
[lru
];
4620 struct page_cgroup
*pc
;
4623 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4624 if (list_empty(list
)) {
4625 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4628 page
= list_entry(list
->prev
, struct page
, lru
);
4630 list_move(&page
->lru
, list
);
4632 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4635 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4637 pc
= lookup_page_cgroup(page
);
4639 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4640 /* found lock contention or "pc" is obsolete. */
4645 } while (!list_empty(list
));
4649 * make mem_cgroup's charge to be 0 if there is no task by moving
4650 * all the charges and pages to the parent.
4651 * This enables deleting this mem_cgroup.
4653 * Caller is responsible for holding css reference on the memcg.
4655 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4661 /* This is for making all *used* pages to be on LRU. */
4662 lru_add_drain_all();
4663 drain_all_stock_sync(memcg
);
4664 mem_cgroup_start_move(memcg
);
4665 for_each_node_state(node
, N_MEMORY
) {
4666 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4669 mem_cgroup_force_empty_list(memcg
,
4674 mem_cgroup_end_move(memcg
);
4675 memcg_oom_recover(memcg
);
4679 * Kernel memory may not necessarily be trackable to a specific
4680 * process. So they are not migrated, and therefore we can't
4681 * expect their value to drop to 0 here.
4682 * Having res filled up with kmem only is enough.
4684 * This is a safety check because mem_cgroup_force_empty_list
4685 * could have raced with mem_cgroup_replace_page_cache callers
4686 * so the lru seemed empty but the page could have been added
4687 * right after the check. RES_USAGE should be safe as we always
4688 * charge before adding to the LRU.
4690 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4691 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4692 } while (usage
> 0);
4696 * Reclaims as many pages from the given memcg as possible and moves
4697 * the rest to the parent.
4699 * Caller is responsible for holding css reference for memcg.
4701 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4703 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4704 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4706 /* returns EBUSY if there is a task or if we come here twice. */
4707 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4710 /* we call try-to-free pages for make this cgroup empty */
4711 lru_add_drain_all();
4712 /* try to free all pages in this cgroup */
4713 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4716 if (signal_pending(current
))
4719 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4723 /* maybe some writeback is necessary */
4724 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4729 mem_cgroup_reparent_charges(memcg
);
4734 static int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
4736 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4739 if (mem_cgroup_is_root(memcg
))
4741 css_get(&memcg
->css
);
4742 ret
= mem_cgroup_force_empty(memcg
);
4743 css_put(&memcg
->css
);
4749 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
4751 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
4754 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
4758 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4759 struct cgroup
*parent
= cont
->parent
;
4760 struct mem_cgroup
*parent_memcg
= NULL
;
4763 parent_memcg
= mem_cgroup_from_cont(parent
);
4767 if (memcg
->use_hierarchy
== val
)
4771 * If parent's use_hierarchy is set, we can't make any modifications
4772 * in the child subtrees. If it is unset, then the change can
4773 * occur, provided the current cgroup has no children.
4775 * For the root cgroup, parent_mem is NULL, we allow value to be
4776 * set if there are no children.
4778 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4779 (val
== 1 || val
== 0)) {
4780 if (list_empty(&cont
->children
))
4781 memcg
->use_hierarchy
= val
;
4794 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4795 enum mem_cgroup_stat_index idx
)
4797 struct mem_cgroup
*iter
;
4800 /* Per-cpu values can be negative, use a signed accumulator */
4801 for_each_mem_cgroup_tree(iter
, memcg
)
4802 val
+= mem_cgroup_read_stat(iter
, idx
);
4804 if (val
< 0) /* race ? */
4809 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4813 if (!mem_cgroup_is_root(memcg
)) {
4815 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4817 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4820 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4821 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4824 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4826 return val
<< PAGE_SHIFT
;
4829 static ssize_t
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
,
4830 struct file
*file
, char __user
*buf
,
4831 size_t nbytes
, loff_t
*ppos
)
4833 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4839 type
= MEMFILE_TYPE(cft
->private);
4840 name
= MEMFILE_ATTR(cft
->private);
4842 if (!do_swap_account
&& type
== _MEMSWAP
)
4847 if (name
== RES_USAGE
)
4848 val
= mem_cgroup_usage(memcg
, false);
4850 val
= res_counter_read_u64(&memcg
->res
, name
);
4853 if (name
== RES_USAGE
)
4854 val
= mem_cgroup_usage(memcg
, true);
4856 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4859 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4865 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4866 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4869 static int memcg_update_kmem_limit(struct cgroup
*cont
, u64 val
)
4872 #ifdef CONFIG_MEMCG_KMEM
4873 bool must_inc_static_branch
= false;
4875 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4877 * For simplicity, we won't allow this to be disabled. It also can't
4878 * be changed if the cgroup has children already, or if tasks had
4881 * If tasks join before we set the limit, a person looking at
4882 * kmem.usage_in_bytes will have no way to determine when it took
4883 * place, which makes the value quite meaningless.
4885 * After it first became limited, changes in the value of the limit are
4886 * of course permitted.
4888 * Taking the cgroup_lock is really offensive, but it is so far the only
4889 * way to guarantee that no children will appear. There are plenty of
4890 * other offenders, and they should all go away. Fine grained locking
4891 * is probably the way to go here. When we are fully hierarchical, we
4892 * can also get rid of the use_hierarchy check.
4895 mutex_lock(&set_limit_mutex
);
4896 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4897 if (cgroup_task_count(cont
) || (memcg
->use_hierarchy
&&
4898 !list_empty(&cont
->children
))) {
4902 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4905 ret
= memcg_update_cache_sizes(memcg
);
4907 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4910 must_inc_static_branch
= true;
4912 * kmem charges can outlive the cgroup. In the case of slab
4913 * pages, for instance, a page contain objects from various
4914 * processes, so it is unfeasible to migrate them away. We
4915 * need to reference count the memcg because of that.
4917 mem_cgroup_get(memcg
);
4919 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4921 mutex_unlock(&set_limit_mutex
);
4925 * We are by now familiar with the fact that we can't inc the static
4926 * branch inside cgroup_lock. See disarm functions for details. A
4927 * worker here is overkill, but also wrong: After the limit is set, we
4928 * must start accounting right away. Since this operation can't fail,
4929 * we can safely defer it to here - no rollback will be needed.
4931 * The boolean used to control this is also safe, because
4932 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4933 * able to set it to true;
4935 if (must_inc_static_branch
) {
4936 static_key_slow_inc(&memcg_kmem_enabled_key
);
4938 * setting the active bit after the inc will guarantee no one
4939 * starts accounting before all call sites are patched
4941 memcg_kmem_set_active(memcg
);
4948 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4951 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4955 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
4956 #ifdef CONFIG_MEMCG_KMEM
4958 * When that happen, we need to disable the static branch only on those
4959 * memcgs that enabled it. To achieve this, we would be forced to
4960 * complicate the code by keeping track of which memcgs were the ones
4961 * that actually enabled limits, and which ones got it from its
4964 * It is a lot simpler just to do static_key_slow_inc() on every child
4965 * that is accounted.
4967 if (!memcg_kmem_is_active(memcg
))
4971 * destroy(), called if we fail, will issue static_key_slow_inc() and
4972 * mem_cgroup_put() if kmem is enabled. We have to either call them
4973 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
4974 * this more consistent, since it always leads to the same destroy path
4976 mem_cgroup_get(memcg
);
4977 static_key_slow_inc(&memcg_kmem_enabled_key
);
4979 mutex_lock(&set_limit_mutex
);
4980 ret
= memcg_update_cache_sizes(memcg
);
4981 mutex_unlock(&set_limit_mutex
);
4988 * The user of this function is...
4991 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
4994 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4997 unsigned long long val
;
5000 type
= MEMFILE_TYPE(cft
->private);
5001 name
= MEMFILE_ATTR(cft
->private);
5003 if (!do_swap_account
&& type
== _MEMSWAP
)
5008 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5012 /* This function does all necessary parse...reuse it */
5013 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5017 ret
= mem_cgroup_resize_limit(memcg
, val
);
5018 else if (type
== _MEMSWAP
)
5019 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5020 else if (type
== _KMEM
)
5021 ret
= memcg_update_kmem_limit(cont
, val
);
5025 case RES_SOFT_LIMIT
:
5026 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5030 * For memsw, soft limits are hard to implement in terms
5031 * of semantics, for now, we support soft limits for
5032 * control without swap
5035 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5040 ret
= -EINVAL
; /* should be BUG() ? */
5046 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5047 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5049 struct cgroup
*cgroup
;
5050 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5052 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5053 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5054 cgroup
= memcg
->css
.cgroup
;
5055 if (!memcg
->use_hierarchy
)
5058 while (cgroup
->parent
) {
5059 cgroup
= cgroup
->parent
;
5060 memcg
= mem_cgroup_from_cont(cgroup
);
5061 if (!memcg
->use_hierarchy
)
5063 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5064 min_limit
= min(min_limit
, tmp
);
5065 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5066 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5069 *mem_limit
= min_limit
;
5070 *memsw_limit
= min_memsw_limit
;
5073 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
5075 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5079 type
= MEMFILE_TYPE(event
);
5080 name
= MEMFILE_ATTR(event
);
5082 if (!do_swap_account
&& type
== _MEMSWAP
)
5088 res_counter_reset_max(&memcg
->res
);
5089 else if (type
== _MEMSWAP
)
5090 res_counter_reset_max(&memcg
->memsw
);
5091 else if (type
== _KMEM
)
5092 res_counter_reset_max(&memcg
->kmem
);
5098 res_counter_reset_failcnt(&memcg
->res
);
5099 else if (type
== _MEMSWAP
)
5100 res_counter_reset_failcnt(&memcg
->memsw
);
5101 else if (type
== _KMEM
)
5102 res_counter_reset_failcnt(&memcg
->kmem
);
5111 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
5114 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
5118 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5119 struct cftype
*cft
, u64 val
)
5121 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5123 if (val
>= (1 << NR_MOVE_TYPE
))
5126 * We check this value several times in both in can_attach() and
5127 * attach(), so we need cgroup lock to prevent this value from being
5131 memcg
->move_charge_at_immigrate
= val
;
5137 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5138 struct cftype
*cft
, u64 val
)
5145 static int memcg_numa_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5149 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5150 unsigned long node_nr
;
5151 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5153 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5154 seq_printf(m
, "total=%lu", total_nr
);
5155 for_each_node_state(nid
, N_MEMORY
) {
5156 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5157 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5161 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5162 seq_printf(m
, "file=%lu", file_nr
);
5163 for_each_node_state(nid
, N_MEMORY
) {
5164 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5166 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5170 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5171 seq_printf(m
, "anon=%lu", anon_nr
);
5172 for_each_node_state(nid
, N_MEMORY
) {
5173 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5175 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5179 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5180 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5181 for_each_node_state(nid
, N_MEMORY
) {
5182 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5183 BIT(LRU_UNEVICTABLE
));
5184 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5189 #endif /* CONFIG_NUMA */
5191 static const char * const mem_cgroup_lru_names
[] = {
5199 static inline void mem_cgroup_lru_names_not_uptodate(void)
5201 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5204 static int memcg_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5207 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5208 struct mem_cgroup
*mi
;
5211 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5212 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5214 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5215 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5218 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5219 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5220 mem_cgroup_read_events(memcg
, i
));
5222 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5223 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5224 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5226 /* Hierarchical information */
5228 unsigned long long limit
, memsw_limit
;
5229 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5230 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5231 if (do_swap_account
)
5232 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5236 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5239 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5241 for_each_mem_cgroup_tree(mi
, memcg
)
5242 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5243 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5246 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5247 unsigned long long val
= 0;
5249 for_each_mem_cgroup_tree(mi
, memcg
)
5250 val
+= mem_cgroup_read_events(mi
, i
);
5251 seq_printf(m
, "total_%s %llu\n",
5252 mem_cgroup_events_names
[i
], val
);
5255 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5256 unsigned long long val
= 0;
5258 for_each_mem_cgroup_tree(mi
, memcg
)
5259 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5260 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5263 #ifdef CONFIG_DEBUG_VM
5266 struct mem_cgroup_per_zone
*mz
;
5267 struct zone_reclaim_stat
*rstat
;
5268 unsigned long recent_rotated
[2] = {0, 0};
5269 unsigned long recent_scanned
[2] = {0, 0};
5271 for_each_online_node(nid
)
5272 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5273 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5274 rstat
= &mz
->lruvec
.reclaim_stat
;
5276 recent_rotated
[0] += rstat
->recent_rotated
[0];
5277 recent_rotated
[1] += rstat
->recent_rotated
[1];
5278 recent_scanned
[0] += rstat
->recent_scanned
[0];
5279 recent_scanned
[1] += rstat
->recent_scanned
[1];
5281 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5282 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5283 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5284 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5291 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
5293 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5295 return mem_cgroup_swappiness(memcg
);
5298 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
5301 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5302 struct mem_cgroup
*parent
;
5307 if (cgrp
->parent
== NULL
)
5310 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5314 /* If under hierarchy, only empty-root can set this value */
5315 if ((parent
->use_hierarchy
) ||
5316 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
5321 memcg
->swappiness
= val
;
5328 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5330 struct mem_cgroup_threshold_ary
*t
;
5336 t
= rcu_dereference(memcg
->thresholds
.primary
);
5338 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5343 usage
= mem_cgroup_usage(memcg
, swap
);
5346 * current_threshold points to threshold just below or equal to usage.
5347 * If it's not true, a threshold was crossed after last
5348 * call of __mem_cgroup_threshold().
5350 i
= t
->current_threshold
;
5353 * Iterate backward over array of thresholds starting from
5354 * current_threshold and check if a threshold is crossed.
5355 * If none of thresholds below usage is crossed, we read
5356 * only one element of the array here.
5358 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5359 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5361 /* i = current_threshold + 1 */
5365 * Iterate forward over array of thresholds starting from
5366 * current_threshold+1 and check if a threshold is crossed.
5367 * If none of thresholds above usage is crossed, we read
5368 * only one element of the array here.
5370 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5371 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5373 /* Update current_threshold */
5374 t
->current_threshold
= i
- 1;
5379 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5382 __mem_cgroup_threshold(memcg
, false);
5383 if (do_swap_account
)
5384 __mem_cgroup_threshold(memcg
, true);
5386 memcg
= parent_mem_cgroup(memcg
);
5390 static int compare_thresholds(const void *a
, const void *b
)
5392 const struct mem_cgroup_threshold
*_a
= a
;
5393 const struct mem_cgroup_threshold
*_b
= b
;
5395 return _a
->threshold
- _b
->threshold
;
5398 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5400 struct mem_cgroup_eventfd_list
*ev
;
5402 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5403 eventfd_signal(ev
->eventfd
, 1);
5407 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5409 struct mem_cgroup
*iter
;
5411 for_each_mem_cgroup_tree(iter
, memcg
)
5412 mem_cgroup_oom_notify_cb(iter
);
5415 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5416 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5418 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5419 struct mem_cgroup_thresholds
*thresholds
;
5420 struct mem_cgroup_threshold_ary
*new;
5421 enum res_type type
= MEMFILE_TYPE(cft
->private);
5422 u64 threshold
, usage
;
5425 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5429 mutex_lock(&memcg
->thresholds_lock
);
5432 thresholds
= &memcg
->thresholds
;
5433 else if (type
== _MEMSWAP
)
5434 thresholds
= &memcg
->memsw_thresholds
;
5438 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5440 /* Check if a threshold crossed before adding a new one */
5441 if (thresholds
->primary
)
5442 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5444 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5446 /* Allocate memory for new array of thresholds */
5447 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5455 /* Copy thresholds (if any) to new array */
5456 if (thresholds
->primary
) {
5457 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5458 sizeof(struct mem_cgroup_threshold
));
5461 /* Add new threshold */
5462 new->entries
[size
- 1].eventfd
= eventfd
;
5463 new->entries
[size
- 1].threshold
= threshold
;
5465 /* Sort thresholds. Registering of new threshold isn't time-critical */
5466 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5467 compare_thresholds
, NULL
);
5469 /* Find current threshold */
5470 new->current_threshold
= -1;
5471 for (i
= 0; i
< size
; i
++) {
5472 if (new->entries
[i
].threshold
<= usage
) {
5474 * new->current_threshold will not be used until
5475 * rcu_assign_pointer(), so it's safe to increment
5478 ++new->current_threshold
;
5483 /* Free old spare buffer and save old primary buffer as spare */
5484 kfree(thresholds
->spare
);
5485 thresholds
->spare
= thresholds
->primary
;
5487 rcu_assign_pointer(thresholds
->primary
, new);
5489 /* To be sure that nobody uses thresholds */
5493 mutex_unlock(&memcg
->thresholds_lock
);
5498 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5499 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5501 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5502 struct mem_cgroup_thresholds
*thresholds
;
5503 struct mem_cgroup_threshold_ary
*new;
5504 enum res_type type
= MEMFILE_TYPE(cft
->private);
5508 mutex_lock(&memcg
->thresholds_lock
);
5510 thresholds
= &memcg
->thresholds
;
5511 else if (type
== _MEMSWAP
)
5512 thresholds
= &memcg
->memsw_thresholds
;
5516 if (!thresholds
->primary
)
5519 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5521 /* Check if a threshold crossed before removing */
5522 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5524 /* Calculate new number of threshold */
5526 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5527 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5531 new = thresholds
->spare
;
5533 /* Set thresholds array to NULL if we don't have thresholds */
5542 /* Copy thresholds and find current threshold */
5543 new->current_threshold
= -1;
5544 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5545 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5548 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5549 if (new->entries
[j
].threshold
<= usage
) {
5551 * new->current_threshold will not be used
5552 * until rcu_assign_pointer(), so it's safe to increment
5555 ++new->current_threshold
;
5561 /* Swap primary and spare array */
5562 thresholds
->spare
= thresholds
->primary
;
5563 /* If all events are unregistered, free the spare array */
5565 kfree(thresholds
->spare
);
5566 thresholds
->spare
= NULL
;
5569 rcu_assign_pointer(thresholds
->primary
, new);
5571 /* To be sure that nobody uses thresholds */
5574 mutex_unlock(&memcg
->thresholds_lock
);
5577 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5578 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5580 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5581 struct mem_cgroup_eventfd_list
*event
;
5582 enum res_type type
= MEMFILE_TYPE(cft
->private);
5584 BUG_ON(type
!= _OOM_TYPE
);
5585 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5589 spin_lock(&memcg_oom_lock
);
5591 event
->eventfd
= eventfd
;
5592 list_add(&event
->list
, &memcg
->oom_notify
);
5594 /* already in OOM ? */
5595 if (atomic_read(&memcg
->under_oom
))
5596 eventfd_signal(eventfd
, 1);
5597 spin_unlock(&memcg_oom_lock
);
5602 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5603 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5605 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5606 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5607 enum res_type type
= MEMFILE_TYPE(cft
->private);
5609 BUG_ON(type
!= _OOM_TYPE
);
5611 spin_lock(&memcg_oom_lock
);
5613 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5614 if (ev
->eventfd
== eventfd
) {
5615 list_del(&ev
->list
);
5620 spin_unlock(&memcg_oom_lock
);
5623 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
5624 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5626 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5628 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5630 if (atomic_read(&memcg
->under_oom
))
5631 cb
->fill(cb
, "under_oom", 1);
5633 cb
->fill(cb
, "under_oom", 0);
5637 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
5638 struct cftype
*cft
, u64 val
)
5640 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5641 struct mem_cgroup
*parent
;
5643 /* cannot set to root cgroup and only 0 and 1 are allowed */
5644 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
5647 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5650 /* oom-kill-disable is a flag for subhierarchy. */
5651 if ((parent
->use_hierarchy
) ||
5652 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
5656 memcg
->oom_kill_disable
= val
;
5658 memcg_oom_recover(memcg
);
5663 #ifdef CONFIG_MEMCG_KMEM
5664 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5668 memcg
->kmemcg_id
= -1;
5669 ret
= memcg_propagate_kmem(memcg
);
5673 return mem_cgroup_sockets_init(memcg
, ss
);
5676 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5678 mem_cgroup_sockets_destroy(memcg
);
5680 memcg_kmem_mark_dead(memcg
);
5682 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5686 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5687 * path here, being careful not to race with memcg_uncharge_kmem: it is
5688 * possible that the charges went down to 0 between mark_dead and the
5689 * res_counter read, so in that case, we don't need the put
5691 if (memcg_kmem_test_and_clear_dead(memcg
))
5692 mem_cgroup_put(memcg
);
5695 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5700 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5705 static struct cftype mem_cgroup_files
[] = {
5707 .name
= "usage_in_bytes",
5708 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5709 .read
= mem_cgroup_read
,
5710 .register_event
= mem_cgroup_usage_register_event
,
5711 .unregister_event
= mem_cgroup_usage_unregister_event
,
5714 .name
= "max_usage_in_bytes",
5715 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5716 .trigger
= mem_cgroup_reset
,
5717 .read
= mem_cgroup_read
,
5720 .name
= "limit_in_bytes",
5721 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5722 .write_string
= mem_cgroup_write
,
5723 .read
= mem_cgroup_read
,
5726 .name
= "soft_limit_in_bytes",
5727 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5728 .write_string
= mem_cgroup_write
,
5729 .read
= mem_cgroup_read
,
5733 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5734 .trigger
= mem_cgroup_reset
,
5735 .read
= mem_cgroup_read
,
5739 .read_seq_string
= memcg_stat_show
,
5742 .name
= "force_empty",
5743 .trigger
= mem_cgroup_force_empty_write
,
5746 .name
= "use_hierarchy",
5747 .write_u64
= mem_cgroup_hierarchy_write
,
5748 .read_u64
= mem_cgroup_hierarchy_read
,
5751 .name
= "swappiness",
5752 .read_u64
= mem_cgroup_swappiness_read
,
5753 .write_u64
= mem_cgroup_swappiness_write
,
5756 .name
= "move_charge_at_immigrate",
5757 .read_u64
= mem_cgroup_move_charge_read
,
5758 .write_u64
= mem_cgroup_move_charge_write
,
5761 .name
= "oom_control",
5762 .read_map
= mem_cgroup_oom_control_read
,
5763 .write_u64
= mem_cgroup_oom_control_write
,
5764 .register_event
= mem_cgroup_oom_register_event
,
5765 .unregister_event
= mem_cgroup_oom_unregister_event
,
5766 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5770 .name
= "numa_stat",
5771 .read_seq_string
= memcg_numa_stat_show
,
5774 #ifdef CONFIG_MEMCG_SWAP
5776 .name
= "memsw.usage_in_bytes",
5777 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5778 .read
= mem_cgroup_read
,
5779 .register_event
= mem_cgroup_usage_register_event
,
5780 .unregister_event
= mem_cgroup_usage_unregister_event
,
5783 .name
= "memsw.max_usage_in_bytes",
5784 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5785 .trigger
= mem_cgroup_reset
,
5786 .read
= mem_cgroup_read
,
5789 .name
= "memsw.limit_in_bytes",
5790 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5791 .write_string
= mem_cgroup_write
,
5792 .read
= mem_cgroup_read
,
5795 .name
= "memsw.failcnt",
5796 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5797 .trigger
= mem_cgroup_reset
,
5798 .read
= mem_cgroup_read
,
5801 #ifdef CONFIG_MEMCG_KMEM
5803 .name
= "kmem.limit_in_bytes",
5804 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5805 .write_string
= mem_cgroup_write
,
5806 .read
= mem_cgroup_read
,
5809 .name
= "kmem.usage_in_bytes",
5810 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5811 .read
= mem_cgroup_read
,
5814 .name
= "kmem.failcnt",
5815 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5816 .trigger
= mem_cgroup_reset
,
5817 .read
= mem_cgroup_read
,
5820 .name
= "kmem.max_usage_in_bytes",
5821 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5822 .trigger
= mem_cgroup_reset
,
5823 .read
= mem_cgroup_read
,
5826 { }, /* terminate */
5829 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5831 struct mem_cgroup_per_node
*pn
;
5832 struct mem_cgroup_per_zone
*mz
;
5833 int zone
, tmp
= node
;
5835 * This routine is called against possible nodes.
5836 * But it's BUG to call kmalloc() against offline node.
5838 * TODO: this routine can waste much memory for nodes which will
5839 * never be onlined. It's better to use memory hotplug callback
5842 if (!node_state(node
, N_NORMAL_MEMORY
))
5844 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5848 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5849 mz
= &pn
->zoneinfo
[zone
];
5850 lruvec_init(&mz
->lruvec
);
5851 mz
->usage_in_excess
= 0;
5852 mz
->on_tree
= false;
5855 memcg
->info
.nodeinfo
[node
] = pn
;
5859 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5861 kfree(memcg
->info
.nodeinfo
[node
]);
5864 static struct mem_cgroup
*mem_cgroup_alloc(void)
5866 struct mem_cgroup
*memcg
;
5867 int size
= sizeof(struct mem_cgroup
);
5869 /* Can be very big if MAX_NUMNODES is very big */
5870 if (size
< PAGE_SIZE
)
5871 memcg
= kzalloc(size
, GFP_KERNEL
);
5873 memcg
= vzalloc(size
);
5878 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5881 spin_lock_init(&memcg
->pcp_counter_lock
);
5885 if (size
< PAGE_SIZE
)
5893 * At destroying mem_cgroup, references from swap_cgroup can remain.
5894 * (scanning all at force_empty is too costly...)
5896 * Instead of clearing all references at force_empty, we remember
5897 * the number of reference from swap_cgroup and free mem_cgroup when
5898 * it goes down to 0.
5900 * Removal of cgroup itself succeeds regardless of refs from swap.
5903 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5906 int size
= sizeof(struct mem_cgroup
);
5908 mem_cgroup_remove_from_trees(memcg
);
5909 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5912 free_mem_cgroup_per_zone_info(memcg
, node
);
5914 free_percpu(memcg
->stat
);
5917 * We need to make sure that (at least for now), the jump label
5918 * destruction code runs outside of the cgroup lock. This is because
5919 * get_online_cpus(), which is called from the static_branch update,
5920 * can't be called inside the cgroup_lock. cpusets are the ones
5921 * enforcing this dependency, so if they ever change, we might as well.
5923 * schedule_work() will guarantee this happens. Be careful if you need
5924 * to move this code around, and make sure it is outside
5927 disarm_static_keys(memcg
);
5928 if (size
< PAGE_SIZE
)
5936 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5937 * but in process context. The work_freeing structure is overlaid
5938 * on the rcu_freeing structure, which itself is overlaid on memsw.
5940 static void free_work(struct work_struct
*work
)
5942 struct mem_cgroup
*memcg
;
5944 memcg
= container_of(work
, struct mem_cgroup
, work_freeing
);
5945 __mem_cgroup_free(memcg
);
5948 static void free_rcu(struct rcu_head
*rcu_head
)
5950 struct mem_cgroup
*memcg
;
5952 memcg
= container_of(rcu_head
, struct mem_cgroup
, rcu_freeing
);
5953 INIT_WORK(&memcg
->work_freeing
, free_work
);
5954 schedule_work(&memcg
->work_freeing
);
5957 static void mem_cgroup_get(struct mem_cgroup
*memcg
)
5959 atomic_inc(&memcg
->refcnt
);
5962 static void __mem_cgroup_put(struct mem_cgroup
*memcg
, int count
)
5964 if (atomic_sub_and_test(count
, &memcg
->refcnt
)) {
5965 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5966 call_rcu(&memcg
->rcu_freeing
, free_rcu
);
5968 mem_cgroup_put(parent
);
5972 static void mem_cgroup_put(struct mem_cgroup
*memcg
)
5974 __mem_cgroup_put(memcg
, 1);
5978 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5980 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
5982 if (!memcg
->res
.parent
)
5984 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
5986 EXPORT_SYMBOL(parent_mem_cgroup
);
5988 #ifdef CONFIG_MEMCG_SWAP
5989 static void __init
enable_swap_cgroup(void)
5991 if (!mem_cgroup_disabled() && really_do_swap_account
)
5992 do_swap_account
= 1;
5995 static void __init
enable_swap_cgroup(void)
6000 static int mem_cgroup_soft_limit_tree_init(void)
6002 struct mem_cgroup_tree_per_node
*rtpn
;
6003 struct mem_cgroup_tree_per_zone
*rtpz
;
6004 int tmp
, node
, zone
;
6006 for_each_node(node
) {
6008 if (!node_state(node
, N_NORMAL_MEMORY
))
6010 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6014 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6016 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6017 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6018 rtpz
->rb_root
= RB_ROOT
;
6019 spin_lock_init(&rtpz
->lock
);
6025 for_each_node(node
) {
6026 if (!soft_limit_tree
.rb_tree_per_node
[node
])
6028 kfree(soft_limit_tree
.rb_tree_per_node
[node
]);
6029 soft_limit_tree
.rb_tree_per_node
[node
] = NULL
;
6035 static struct cgroup_subsys_state
* __ref
6036 mem_cgroup_css_alloc(struct cgroup
*cont
)
6038 struct mem_cgroup
*memcg
, *parent
;
6039 long error
= -ENOMEM
;
6042 memcg
= mem_cgroup_alloc();
6044 return ERR_PTR(error
);
6047 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6051 if (cont
->parent
== NULL
) {
6053 enable_swap_cgroup();
6055 if (mem_cgroup_soft_limit_tree_init())
6057 root_mem_cgroup
= memcg
;
6058 for_each_possible_cpu(cpu
) {
6059 struct memcg_stock_pcp
*stock
=
6060 &per_cpu(memcg_stock
, cpu
);
6061 INIT_WORK(&stock
->work
, drain_local_stock
);
6063 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6065 parent
= mem_cgroup_from_cont(cont
->parent
);
6066 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6067 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6070 if (parent
&& parent
->use_hierarchy
) {
6071 res_counter_init(&memcg
->res
, &parent
->res
);
6072 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6073 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6076 * We increment refcnt of the parent to ensure that we can
6077 * safely access it on res_counter_charge/uncharge.
6078 * This refcnt will be decremented when freeing this
6079 * mem_cgroup(see mem_cgroup_put).
6081 mem_cgroup_get(parent
);
6083 res_counter_init(&memcg
->res
, NULL
);
6084 res_counter_init(&memcg
->memsw
, NULL
);
6085 res_counter_init(&memcg
->kmem
, NULL
);
6087 * Deeper hierachy with use_hierarchy == false doesn't make
6088 * much sense so let cgroup subsystem know about this
6089 * unfortunate state in our controller.
6091 if (parent
&& parent
!= root_mem_cgroup
)
6092 mem_cgroup_subsys
.broken_hierarchy
= true;
6094 memcg
->last_scanned_node
= MAX_NUMNODES
;
6095 INIT_LIST_HEAD(&memcg
->oom_notify
);
6098 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6099 atomic_set(&memcg
->refcnt
, 1);
6100 memcg
->move_charge_at_immigrate
= 0;
6101 mutex_init(&memcg
->thresholds_lock
);
6102 spin_lock_init(&memcg
->move_lock
);
6104 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6107 * We call put now because our (and parent's) refcnts
6108 * are already in place. mem_cgroup_put() will internally
6109 * call __mem_cgroup_free, so return directly
6111 mem_cgroup_put(memcg
);
6112 return ERR_PTR(error
);
6116 __mem_cgroup_free(memcg
);
6117 return ERR_PTR(error
);
6120 static void mem_cgroup_css_offline(struct cgroup
*cont
)
6122 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6124 mem_cgroup_reparent_charges(memcg
);
6125 mem_cgroup_destroy_all_caches(memcg
);
6128 static void mem_cgroup_css_free(struct cgroup
*cont
)
6130 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6132 kmem_cgroup_destroy(memcg
);
6134 mem_cgroup_put(memcg
);
6138 /* Handlers for move charge at task migration. */
6139 #define PRECHARGE_COUNT_AT_ONCE 256
6140 static int mem_cgroup_do_precharge(unsigned long count
)
6143 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6144 struct mem_cgroup
*memcg
= mc
.to
;
6146 if (mem_cgroup_is_root(memcg
)) {
6147 mc
.precharge
+= count
;
6148 /* we don't need css_get for root */
6151 /* try to charge at once */
6153 struct res_counter
*dummy
;
6155 * "memcg" cannot be under rmdir() because we've already checked
6156 * by cgroup_lock_live_cgroup() that it is not removed and we
6157 * are still under the same cgroup_mutex. So we can postpone
6160 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6162 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6163 PAGE_SIZE
* count
, &dummy
)) {
6164 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6167 mc
.precharge
+= count
;
6171 /* fall back to one by one charge */
6173 if (signal_pending(current
)) {
6177 if (!batch_count
--) {
6178 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6181 ret
= __mem_cgroup_try_charge(NULL
,
6182 GFP_KERNEL
, 1, &memcg
, false);
6184 /* mem_cgroup_clear_mc() will do uncharge later */
6192 * get_mctgt_type - get target type of moving charge
6193 * @vma: the vma the pte to be checked belongs
6194 * @addr: the address corresponding to the pte to be checked
6195 * @ptent: the pte to be checked
6196 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6199 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6200 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6201 * move charge. if @target is not NULL, the page is stored in target->page
6202 * with extra refcnt got(Callers should handle it).
6203 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6204 * target for charge migration. if @target is not NULL, the entry is stored
6207 * Called with pte lock held.
6214 enum mc_target_type
{
6220 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6221 unsigned long addr
, pte_t ptent
)
6223 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6225 if (!page
|| !page_mapped(page
))
6227 if (PageAnon(page
)) {
6228 /* we don't move shared anon */
6231 } else if (!move_file())
6232 /* we ignore mapcount for file pages */
6234 if (!get_page_unless_zero(page
))
6241 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6242 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6244 struct page
*page
= NULL
;
6245 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6247 if (!move_anon() || non_swap_entry(ent
))
6250 * Because lookup_swap_cache() updates some statistics counter,
6251 * we call find_get_page() with swapper_space directly.
6253 page
= find_get_page(&swapper_space
, ent
.val
);
6254 if (do_swap_account
)
6255 entry
->val
= ent
.val
;
6260 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6261 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6267 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6268 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6270 struct page
*page
= NULL
;
6271 struct address_space
*mapping
;
6274 if (!vma
->vm_file
) /* anonymous vma */
6279 mapping
= vma
->vm_file
->f_mapping
;
6280 if (pte_none(ptent
))
6281 pgoff
= linear_page_index(vma
, addr
);
6282 else /* pte_file(ptent) is true */
6283 pgoff
= pte_to_pgoff(ptent
);
6285 /* page is moved even if it's not RSS of this task(page-faulted). */
6286 page
= find_get_page(mapping
, pgoff
);
6289 /* shmem/tmpfs may report page out on swap: account for that too. */
6290 if (radix_tree_exceptional_entry(page
)) {
6291 swp_entry_t swap
= radix_to_swp_entry(page
);
6292 if (do_swap_account
)
6294 page
= find_get_page(&swapper_space
, swap
.val
);
6300 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6301 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6303 struct page
*page
= NULL
;
6304 struct page_cgroup
*pc
;
6305 enum mc_target_type ret
= MC_TARGET_NONE
;
6306 swp_entry_t ent
= { .val
= 0 };
6308 if (pte_present(ptent
))
6309 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6310 else if (is_swap_pte(ptent
))
6311 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6312 else if (pte_none(ptent
) || pte_file(ptent
))
6313 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6315 if (!page
&& !ent
.val
)
6318 pc
= lookup_page_cgroup(page
);
6320 * Do only loose check w/o page_cgroup lock.
6321 * mem_cgroup_move_account() checks the pc is valid or not under
6324 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6325 ret
= MC_TARGET_PAGE
;
6327 target
->page
= page
;
6329 if (!ret
|| !target
)
6332 /* There is a swap entry and a page doesn't exist or isn't charged */
6333 if (ent
.val
&& !ret
&&
6334 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6335 ret
= MC_TARGET_SWAP
;
6342 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6344 * We don't consider swapping or file mapped pages because THP does not
6345 * support them for now.
6346 * Caller should make sure that pmd_trans_huge(pmd) is true.
6348 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6349 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6351 struct page
*page
= NULL
;
6352 struct page_cgroup
*pc
;
6353 enum mc_target_type ret
= MC_TARGET_NONE
;
6355 page
= pmd_page(pmd
);
6356 VM_BUG_ON(!page
|| !PageHead(page
));
6359 pc
= lookup_page_cgroup(page
);
6360 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6361 ret
= MC_TARGET_PAGE
;
6364 target
->page
= page
;
6370 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6371 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6373 return MC_TARGET_NONE
;
6377 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6378 unsigned long addr
, unsigned long end
,
6379 struct mm_walk
*walk
)
6381 struct vm_area_struct
*vma
= walk
->private;
6385 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6386 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6387 mc
.precharge
+= HPAGE_PMD_NR
;
6388 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6392 if (pmd_trans_unstable(pmd
))
6394 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6395 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6396 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6397 mc
.precharge
++; /* increment precharge temporarily */
6398 pte_unmap_unlock(pte
- 1, ptl
);
6404 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6406 unsigned long precharge
;
6407 struct vm_area_struct
*vma
;
6409 down_read(&mm
->mmap_sem
);
6410 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6411 struct mm_walk mem_cgroup_count_precharge_walk
= {
6412 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6416 if (is_vm_hugetlb_page(vma
))
6418 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6419 &mem_cgroup_count_precharge_walk
);
6421 up_read(&mm
->mmap_sem
);
6423 precharge
= mc
.precharge
;
6429 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6431 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6433 VM_BUG_ON(mc
.moving_task
);
6434 mc
.moving_task
= current
;
6435 return mem_cgroup_do_precharge(precharge
);
6438 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6439 static void __mem_cgroup_clear_mc(void)
6441 struct mem_cgroup
*from
= mc
.from
;
6442 struct mem_cgroup
*to
= mc
.to
;
6444 /* we must uncharge all the leftover precharges from mc.to */
6446 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6450 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6451 * we must uncharge here.
6453 if (mc
.moved_charge
) {
6454 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6455 mc
.moved_charge
= 0;
6457 /* we must fixup refcnts and charges */
6458 if (mc
.moved_swap
) {
6459 /* uncharge swap account from the old cgroup */
6460 if (!mem_cgroup_is_root(mc
.from
))
6461 res_counter_uncharge(&mc
.from
->memsw
,
6462 PAGE_SIZE
* mc
.moved_swap
);
6463 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
6465 if (!mem_cgroup_is_root(mc
.to
)) {
6467 * we charged both to->res and to->memsw, so we should
6470 res_counter_uncharge(&mc
.to
->res
,
6471 PAGE_SIZE
* mc
.moved_swap
);
6473 /* we've already done mem_cgroup_get(mc.to) */
6476 memcg_oom_recover(from
);
6477 memcg_oom_recover(to
);
6478 wake_up_all(&mc
.waitq
);
6481 static void mem_cgroup_clear_mc(void)
6483 struct mem_cgroup
*from
= mc
.from
;
6486 * we must clear moving_task before waking up waiters at the end of
6489 mc
.moving_task
= NULL
;
6490 __mem_cgroup_clear_mc();
6491 spin_lock(&mc
.lock
);
6494 spin_unlock(&mc
.lock
);
6495 mem_cgroup_end_move(from
);
6498 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6499 struct cgroup_taskset
*tset
)
6501 struct task_struct
*p
= cgroup_taskset_first(tset
);
6503 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgroup
);
6505 if (memcg
->move_charge_at_immigrate
) {
6506 struct mm_struct
*mm
;
6507 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6509 VM_BUG_ON(from
== memcg
);
6511 mm
= get_task_mm(p
);
6514 /* We move charges only when we move a owner of the mm */
6515 if (mm
->owner
== p
) {
6518 VM_BUG_ON(mc
.precharge
);
6519 VM_BUG_ON(mc
.moved_charge
);
6520 VM_BUG_ON(mc
.moved_swap
);
6521 mem_cgroup_start_move(from
);
6522 spin_lock(&mc
.lock
);
6525 spin_unlock(&mc
.lock
);
6526 /* We set mc.moving_task later */
6528 ret
= mem_cgroup_precharge_mc(mm
);
6530 mem_cgroup_clear_mc();
6537 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6538 struct cgroup_taskset
*tset
)
6540 mem_cgroup_clear_mc();
6543 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6544 unsigned long addr
, unsigned long end
,
6545 struct mm_walk
*walk
)
6548 struct vm_area_struct
*vma
= walk
->private;
6551 enum mc_target_type target_type
;
6552 union mc_target target
;
6554 struct page_cgroup
*pc
;
6557 * We don't take compound_lock() here but no race with splitting thp
6559 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6560 * under splitting, which means there's no concurrent thp split,
6561 * - if another thread runs into split_huge_page() just after we
6562 * entered this if-block, the thread must wait for page table lock
6563 * to be unlocked in __split_huge_page_splitting(), where the main
6564 * part of thp split is not executed yet.
6566 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6567 if (mc
.precharge
< HPAGE_PMD_NR
) {
6568 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6571 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6572 if (target_type
== MC_TARGET_PAGE
) {
6574 if (!isolate_lru_page(page
)) {
6575 pc
= lookup_page_cgroup(page
);
6576 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6577 pc
, mc
.from
, mc
.to
)) {
6578 mc
.precharge
-= HPAGE_PMD_NR
;
6579 mc
.moved_charge
+= HPAGE_PMD_NR
;
6581 putback_lru_page(page
);
6585 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6589 if (pmd_trans_unstable(pmd
))
6592 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6593 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6594 pte_t ptent
= *(pte
++);
6600 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6601 case MC_TARGET_PAGE
:
6603 if (isolate_lru_page(page
))
6605 pc
= lookup_page_cgroup(page
);
6606 if (!mem_cgroup_move_account(page
, 1, pc
,
6609 /* we uncharge from mc.from later. */
6612 putback_lru_page(page
);
6613 put
: /* get_mctgt_type() gets the page */
6616 case MC_TARGET_SWAP
:
6618 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6620 /* we fixup refcnts and charges later. */
6628 pte_unmap_unlock(pte
- 1, ptl
);
6633 * We have consumed all precharges we got in can_attach().
6634 * We try charge one by one, but don't do any additional
6635 * charges to mc.to if we have failed in charge once in attach()
6638 ret
= mem_cgroup_do_precharge(1);
6646 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6648 struct vm_area_struct
*vma
;
6650 lru_add_drain_all();
6652 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6654 * Someone who are holding the mmap_sem might be waiting in
6655 * waitq. So we cancel all extra charges, wake up all waiters,
6656 * and retry. Because we cancel precharges, we might not be able
6657 * to move enough charges, but moving charge is a best-effort
6658 * feature anyway, so it wouldn't be a big problem.
6660 __mem_cgroup_clear_mc();
6664 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6666 struct mm_walk mem_cgroup_move_charge_walk
= {
6667 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6671 if (is_vm_hugetlb_page(vma
))
6673 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6674 &mem_cgroup_move_charge_walk
);
6677 * means we have consumed all precharges and failed in
6678 * doing additional charge. Just abandon here.
6682 up_read(&mm
->mmap_sem
);
6685 static void mem_cgroup_move_task(struct cgroup
*cont
,
6686 struct cgroup_taskset
*tset
)
6688 struct task_struct
*p
= cgroup_taskset_first(tset
);
6689 struct mm_struct
*mm
= get_task_mm(p
);
6693 mem_cgroup_move_charge(mm
);
6697 mem_cgroup_clear_mc();
6699 #else /* !CONFIG_MMU */
6700 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6701 struct cgroup_taskset
*tset
)
6705 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6706 struct cgroup_taskset
*tset
)
6709 static void mem_cgroup_move_task(struct cgroup
*cont
,
6710 struct cgroup_taskset
*tset
)
6715 struct cgroup_subsys mem_cgroup_subsys
= {
6717 .subsys_id
= mem_cgroup_subsys_id
,
6718 .css_alloc
= mem_cgroup_css_alloc
,
6719 .css_offline
= mem_cgroup_css_offline
,
6720 .css_free
= mem_cgroup_css_free
,
6721 .can_attach
= mem_cgroup_can_attach
,
6722 .cancel_attach
= mem_cgroup_cancel_attach
,
6723 .attach
= mem_cgroup_move_task
,
6724 .base_cftypes
= mem_cgroup_files
,
6729 #ifdef CONFIG_MEMCG_SWAP
6730 static int __init
enable_swap_account(char *s
)
6732 /* consider enabled if no parameter or 1 is given */
6733 if (!strcmp(s
, "1"))
6734 really_do_swap_account
= 1;
6735 else if (!strcmp(s
, "0"))
6736 really_do_swap_account
= 0;
6739 __setup("swapaccount=", enable_swap_account
);