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
[] = {
123 static const char * const mem_cgroup_lru_names
[] = {
132 * Per memcg event counter is incremented at every pagein/pageout. With THP,
133 * it will be incremated by the number of pages. This counter is used for
134 * for trigger some periodic events. This is straightforward and better
135 * than using jiffies etc. to handle periodic memcg event.
137 enum mem_cgroup_events_target
{
138 MEM_CGROUP_TARGET_THRESH
,
139 MEM_CGROUP_TARGET_SOFTLIMIT
,
140 MEM_CGROUP_TARGET_NUMAINFO
,
143 #define THRESHOLDS_EVENTS_TARGET 128
144 #define SOFTLIMIT_EVENTS_TARGET 1024
145 #define NUMAINFO_EVENTS_TARGET 1024
147 struct mem_cgroup_stat_cpu
{
148 long count
[MEM_CGROUP_STAT_NSTATS
];
149 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
150 unsigned long nr_page_events
;
151 unsigned long targets
[MEM_CGROUP_NTARGETS
];
154 struct mem_cgroup_reclaim_iter
{
156 * last scanned hierarchy member. Valid only if last_dead_count
157 * matches memcg->dead_count of the hierarchy root group.
159 struct mem_cgroup
*last_visited
;
160 unsigned long last_dead_count
;
162 /* scan generation, increased every round-trip */
163 unsigned int generation
;
167 * per-zone information in memory controller.
169 struct mem_cgroup_per_zone
{
170 struct lruvec lruvec
;
171 unsigned long lru_size
[NR_LRU_LISTS
];
173 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
175 struct rb_node tree_node
; /* RB tree node */
176 unsigned long long usage_in_excess
;/* Set to the value by which */
177 /* the soft limit is exceeded*/
179 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
180 /* use container_of */
183 struct mem_cgroup_per_node
{
184 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
187 struct mem_cgroup_lru_info
{
188 struct mem_cgroup_per_node
*nodeinfo
[0];
192 * Cgroups above their limits are maintained in a RB-Tree, independent of
193 * their hierarchy representation
196 struct mem_cgroup_tree_per_zone
{
197 struct rb_root rb_root
;
201 struct mem_cgroup_tree_per_node
{
202 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
205 struct mem_cgroup_tree
{
206 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
209 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
211 struct mem_cgroup_threshold
{
212 struct eventfd_ctx
*eventfd
;
217 struct mem_cgroup_threshold_ary
{
218 /* An array index points to threshold just below or equal to usage. */
219 int current_threshold
;
220 /* Size of entries[] */
222 /* Array of thresholds */
223 struct mem_cgroup_threshold entries
[0];
226 struct mem_cgroup_thresholds
{
227 /* Primary thresholds array */
228 struct mem_cgroup_threshold_ary
*primary
;
230 * Spare threshold array.
231 * This is needed to make mem_cgroup_unregister_event() "never fail".
232 * It must be able to store at least primary->size - 1 entries.
234 struct mem_cgroup_threshold_ary
*spare
;
238 struct mem_cgroup_eventfd_list
{
239 struct list_head list
;
240 struct eventfd_ctx
*eventfd
;
243 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
244 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
247 * The memory controller data structure. The memory controller controls both
248 * page cache and RSS per cgroup. We would eventually like to provide
249 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
250 * to help the administrator determine what knobs to tune.
252 * TODO: Add a water mark for the memory controller. Reclaim will begin when
253 * we hit the water mark. May be even add a low water mark, such that
254 * no reclaim occurs from a cgroup at it's low water mark, this is
255 * a feature that will be implemented much later in the future.
258 struct cgroup_subsys_state css
;
260 * the counter to account for memory usage
262 struct res_counter res
;
266 * the counter to account for mem+swap usage.
268 struct res_counter memsw
;
271 * rcu_freeing is used only when freeing struct mem_cgroup,
272 * so put it into a union to avoid wasting more memory.
273 * It must be disjoint from the css field. It could be
274 * in a union with the res field, but res plays a much
275 * larger part in mem_cgroup life than memsw, and might
276 * be of interest, even at time of free, when debugging.
277 * So share rcu_head with the less interesting memsw.
279 struct rcu_head rcu_freeing
;
281 * We also need some space for a worker in deferred freeing.
282 * By the time we call it, rcu_freeing is no longer in use.
284 struct work_struct work_freeing
;
288 * the counter to account for kernel memory usage.
290 struct res_counter kmem
;
292 * Should the accounting and control be hierarchical, per subtree?
295 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
303 /* OOM-Killer disable */
304 int oom_kill_disable
;
306 /* set when res.limit == memsw.limit */
307 bool memsw_is_minimum
;
309 /* protect arrays of thresholds */
310 struct mutex thresholds_lock
;
312 /* thresholds for memory usage. RCU-protected */
313 struct mem_cgroup_thresholds thresholds
;
315 /* thresholds for mem+swap usage. RCU-protected */
316 struct mem_cgroup_thresholds memsw_thresholds
;
318 /* For oom notifier event fd */
319 struct list_head oom_notify
;
322 * Should we move charges of a task when a task is moved into this
323 * mem_cgroup ? And what type of charges should we move ?
325 unsigned long move_charge_at_immigrate
;
327 * set > 0 if pages under this cgroup are moving to other cgroup.
329 atomic_t moving_account
;
330 /* taken only while moving_account > 0 */
331 spinlock_t move_lock
;
335 struct mem_cgroup_stat_cpu __percpu
*stat
;
337 * used when a cpu is offlined or other synchronizations
338 * See mem_cgroup_read_stat().
340 struct mem_cgroup_stat_cpu nocpu_base
;
341 spinlock_t pcp_counter_lock
;
344 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
345 struct tcp_memcontrol tcp_mem
;
347 #if defined(CONFIG_MEMCG_KMEM)
348 /* analogous to slab_common's slab_caches list. per-memcg */
349 struct list_head memcg_slab_caches
;
350 /* Not a spinlock, we can take a lot of time walking the list */
351 struct mutex slab_caches_mutex
;
352 /* Index in the kmem_cache->memcg_params->memcg_caches array */
356 int last_scanned_node
;
358 nodemask_t scan_nodes
;
359 atomic_t numainfo_events
;
360 atomic_t numainfo_updating
;
363 * Per cgroup active and inactive list, similar to the
364 * per zone LRU lists.
366 * WARNING: This has to be the last element of the struct. Don't
367 * add new fields after this point.
369 struct mem_cgroup_lru_info info
;
372 static size_t memcg_size(void)
374 return sizeof(struct mem_cgroup
) +
375 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
378 /* internal only representation about the status of kmem accounting. */
380 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
381 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
382 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
385 /* We account when limit is on, but only after call sites are patched */
386 #define KMEM_ACCOUNTED_MASK \
387 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
392 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
395 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
400 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
402 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
405 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
407 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
410 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
412 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
413 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
416 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
418 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
419 &memcg
->kmem_account_flags
);
423 /* Stuffs for move charges at task migration. */
425 * Types of charges to be moved. "move_charge_at_immitgrate" and
426 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
429 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
430 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
434 /* "mc" and its members are protected by cgroup_mutex */
435 static struct move_charge_struct
{
436 spinlock_t lock
; /* for from, to */
437 struct mem_cgroup
*from
;
438 struct mem_cgroup
*to
;
439 unsigned long immigrate_flags
;
440 unsigned long precharge
;
441 unsigned long moved_charge
;
442 unsigned long moved_swap
;
443 struct task_struct
*moving_task
; /* a task moving charges */
444 wait_queue_head_t waitq
; /* a waitq for other context */
446 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
447 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
450 static bool move_anon(void)
452 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
455 static bool move_file(void)
457 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
461 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
462 * limit reclaim to prevent infinite loops, if they ever occur.
464 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
465 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
468 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
469 MEM_CGROUP_CHARGE_TYPE_ANON
,
470 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
471 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
475 /* for encoding cft->private value on file */
483 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
484 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
485 #define MEMFILE_ATTR(val) ((val) & 0xffff)
486 /* Used for OOM nofiier */
487 #define OOM_CONTROL (0)
490 * Reclaim flags for mem_cgroup_hierarchical_reclaim
492 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
493 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
494 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
495 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
498 * The memcg_create_mutex will be held whenever a new cgroup is created.
499 * As a consequence, any change that needs to protect against new child cgroups
500 * appearing has to hold it as well.
502 static DEFINE_MUTEX(memcg_create_mutex
);
504 static void mem_cgroup_get(struct mem_cgroup
*memcg
);
505 static void mem_cgroup_put(struct mem_cgroup
*memcg
);
508 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
510 return container_of(s
, struct mem_cgroup
, css
);
513 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
515 return (memcg
== root_mem_cgroup
);
518 /* Writing them here to avoid exposing memcg's inner layout */
519 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
521 void sock_update_memcg(struct sock
*sk
)
523 if (mem_cgroup_sockets_enabled
) {
524 struct mem_cgroup
*memcg
;
525 struct cg_proto
*cg_proto
;
527 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
529 /* Socket cloning can throw us here with sk_cgrp already
530 * filled. It won't however, necessarily happen from
531 * process context. So the test for root memcg given
532 * the current task's memcg won't help us in this case.
534 * Respecting the original socket's memcg is a better
535 * decision in this case.
538 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
539 mem_cgroup_get(sk
->sk_cgrp
->memcg
);
544 memcg
= mem_cgroup_from_task(current
);
545 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
546 if (!mem_cgroup_is_root(memcg
) && memcg_proto_active(cg_proto
)) {
547 mem_cgroup_get(memcg
);
548 sk
->sk_cgrp
= cg_proto
;
553 EXPORT_SYMBOL(sock_update_memcg
);
555 void sock_release_memcg(struct sock
*sk
)
557 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
558 struct mem_cgroup
*memcg
;
559 WARN_ON(!sk
->sk_cgrp
->memcg
);
560 memcg
= sk
->sk_cgrp
->memcg
;
561 mem_cgroup_put(memcg
);
565 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
567 if (!memcg
|| mem_cgroup_is_root(memcg
))
570 return &memcg
->tcp_mem
.cg_proto
;
572 EXPORT_SYMBOL(tcp_proto_cgroup
);
574 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
576 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
578 static_key_slow_dec(&memcg_socket_limit_enabled
);
581 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
586 #ifdef CONFIG_MEMCG_KMEM
588 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
589 * There are two main reasons for not using the css_id for this:
590 * 1) this works better in sparse environments, where we have a lot of memcgs,
591 * but only a few kmem-limited. Or also, if we have, for instance, 200
592 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
593 * 200 entry array for that.
595 * 2) In order not to violate the cgroup API, we would like to do all memory
596 * allocation in ->create(). At that point, we haven't yet allocated the
597 * css_id. Having a separate index prevents us from messing with the cgroup
600 * The current size of the caches array is stored in
601 * memcg_limited_groups_array_size. It will double each time we have to
604 static DEFINE_IDA(kmem_limited_groups
);
605 int memcg_limited_groups_array_size
;
608 * MIN_SIZE is different than 1, because we would like to avoid going through
609 * the alloc/free process all the time. In a small machine, 4 kmem-limited
610 * cgroups is a reasonable guess. In the future, it could be a parameter or
611 * tunable, but that is strictly not necessary.
613 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
614 * this constant directly from cgroup, but it is understandable that this is
615 * better kept as an internal representation in cgroup.c. In any case, the
616 * css_id space is not getting any smaller, and we don't have to necessarily
617 * increase ours as well if it increases.
619 #define MEMCG_CACHES_MIN_SIZE 4
620 #define MEMCG_CACHES_MAX_SIZE 65535
623 * A lot of the calls to the cache allocation functions are expected to be
624 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
625 * conditional to this static branch, we'll have to allow modules that does
626 * kmem_cache_alloc and the such to see this symbol as well
628 struct static_key memcg_kmem_enabled_key
;
629 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
631 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
633 if (memcg_kmem_is_active(memcg
)) {
634 static_key_slow_dec(&memcg_kmem_enabled_key
);
635 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
638 * This check can't live in kmem destruction function,
639 * since the charges will outlive the cgroup
641 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
644 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
647 #endif /* CONFIG_MEMCG_KMEM */
649 static void disarm_static_keys(struct mem_cgroup
*memcg
)
651 disarm_sock_keys(memcg
);
652 disarm_kmem_keys(memcg
);
655 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
657 static struct mem_cgroup_per_zone
*
658 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
660 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
661 return &memcg
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
664 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
669 static struct mem_cgroup_per_zone
*
670 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
672 int nid
= page_to_nid(page
);
673 int zid
= page_zonenum(page
);
675 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
678 static struct mem_cgroup_tree_per_zone
*
679 soft_limit_tree_node_zone(int nid
, int zid
)
681 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
684 static struct mem_cgroup_tree_per_zone
*
685 soft_limit_tree_from_page(struct page
*page
)
687 int nid
= page_to_nid(page
);
688 int zid
= page_zonenum(page
);
690 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
694 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
695 struct mem_cgroup_per_zone
*mz
,
696 struct mem_cgroup_tree_per_zone
*mctz
,
697 unsigned long long new_usage_in_excess
)
699 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
700 struct rb_node
*parent
= NULL
;
701 struct mem_cgroup_per_zone
*mz_node
;
706 mz
->usage_in_excess
= new_usage_in_excess
;
707 if (!mz
->usage_in_excess
)
711 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
713 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
716 * We can't avoid mem cgroups that are over their soft
717 * limit by the same amount
719 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
722 rb_link_node(&mz
->tree_node
, parent
, p
);
723 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
728 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
729 struct mem_cgroup_per_zone
*mz
,
730 struct mem_cgroup_tree_per_zone
*mctz
)
734 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
739 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
740 struct mem_cgroup_per_zone
*mz
,
741 struct mem_cgroup_tree_per_zone
*mctz
)
743 spin_lock(&mctz
->lock
);
744 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
745 spin_unlock(&mctz
->lock
);
749 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
751 unsigned long long excess
;
752 struct mem_cgroup_per_zone
*mz
;
753 struct mem_cgroup_tree_per_zone
*mctz
;
754 int nid
= page_to_nid(page
);
755 int zid
= page_zonenum(page
);
756 mctz
= soft_limit_tree_from_page(page
);
759 * Necessary to update all ancestors when hierarchy is used.
760 * because their event counter is not touched.
762 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
763 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
764 excess
= res_counter_soft_limit_excess(&memcg
->res
);
766 * We have to update the tree if mz is on RB-tree or
767 * mem is over its softlimit.
769 if (excess
|| mz
->on_tree
) {
770 spin_lock(&mctz
->lock
);
771 /* if on-tree, remove it */
773 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
775 * Insert again. mz->usage_in_excess will be updated.
776 * If excess is 0, no tree ops.
778 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
779 spin_unlock(&mctz
->lock
);
784 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
787 struct mem_cgroup_per_zone
*mz
;
788 struct mem_cgroup_tree_per_zone
*mctz
;
790 for_each_node(node
) {
791 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
792 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
793 mctz
= soft_limit_tree_node_zone(node
, zone
);
794 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
799 static struct mem_cgroup_per_zone
*
800 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
802 struct rb_node
*rightmost
= NULL
;
803 struct mem_cgroup_per_zone
*mz
;
807 rightmost
= rb_last(&mctz
->rb_root
);
809 goto done
; /* Nothing to reclaim from */
811 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
813 * Remove the node now but someone else can add it back,
814 * we will to add it back at the end of reclaim to its correct
815 * position in the tree.
817 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
818 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
819 !css_tryget(&mz
->memcg
->css
))
825 static struct mem_cgroup_per_zone
*
826 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
828 struct mem_cgroup_per_zone
*mz
;
830 spin_lock(&mctz
->lock
);
831 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
832 spin_unlock(&mctz
->lock
);
837 * Implementation Note: reading percpu statistics for memcg.
839 * Both of vmstat[] and percpu_counter has threshold and do periodic
840 * synchronization to implement "quick" read. There are trade-off between
841 * reading cost and precision of value. Then, we may have a chance to implement
842 * a periodic synchronizion of counter in memcg's counter.
844 * But this _read() function is used for user interface now. The user accounts
845 * memory usage by memory cgroup and he _always_ requires exact value because
846 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
847 * have to visit all online cpus and make sum. So, for now, unnecessary
848 * synchronization is not implemented. (just implemented for cpu hotplug)
850 * If there are kernel internal actions which can make use of some not-exact
851 * value, and reading all cpu value can be performance bottleneck in some
852 * common workload, threashold and synchonization as vmstat[] should be
855 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
856 enum mem_cgroup_stat_index idx
)
862 for_each_online_cpu(cpu
)
863 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
864 #ifdef CONFIG_HOTPLUG_CPU
865 spin_lock(&memcg
->pcp_counter_lock
);
866 val
+= memcg
->nocpu_base
.count
[idx
];
867 spin_unlock(&memcg
->pcp_counter_lock
);
873 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
876 int val
= (charge
) ? 1 : -1;
877 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
880 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
881 enum mem_cgroup_events_index idx
)
883 unsigned long val
= 0;
886 for_each_online_cpu(cpu
)
887 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
888 #ifdef CONFIG_HOTPLUG_CPU
889 spin_lock(&memcg
->pcp_counter_lock
);
890 val
+= memcg
->nocpu_base
.events
[idx
];
891 spin_unlock(&memcg
->pcp_counter_lock
);
896 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
897 bool anon
, int nr_pages
)
902 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
903 * counted as CACHE even if it's on ANON LRU.
906 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
909 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
912 /* pagein of a big page is an event. So, ignore page size */
914 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
916 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
917 nr_pages
= -nr_pages
; /* for event */
920 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
926 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
928 struct mem_cgroup_per_zone
*mz
;
930 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
931 return mz
->lru_size
[lru
];
935 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
936 unsigned int lru_mask
)
938 struct mem_cgroup_per_zone
*mz
;
940 unsigned long ret
= 0;
942 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
945 if (BIT(lru
) & lru_mask
)
946 ret
+= mz
->lru_size
[lru
];
952 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
953 int nid
, unsigned int lru_mask
)
958 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
959 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
965 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
966 unsigned int lru_mask
)
971 for_each_node_state(nid
, N_MEMORY
)
972 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
976 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
977 enum mem_cgroup_events_target target
)
979 unsigned long val
, next
;
981 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
982 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
983 /* from time_after() in jiffies.h */
984 if ((long)next
- (long)val
< 0) {
986 case MEM_CGROUP_TARGET_THRESH
:
987 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
989 case MEM_CGROUP_TARGET_SOFTLIMIT
:
990 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
992 case MEM_CGROUP_TARGET_NUMAINFO
:
993 next
= val
+ NUMAINFO_EVENTS_TARGET
;
998 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1005 * Check events in order.
1008 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1011 /* threshold event is triggered in finer grain than soft limit */
1012 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1013 MEM_CGROUP_TARGET_THRESH
))) {
1015 bool do_numainfo __maybe_unused
;
1017 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1018 MEM_CGROUP_TARGET_SOFTLIMIT
);
1019 #if MAX_NUMNODES > 1
1020 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1021 MEM_CGROUP_TARGET_NUMAINFO
);
1025 mem_cgroup_threshold(memcg
);
1026 if (unlikely(do_softlimit
))
1027 mem_cgroup_update_tree(memcg
, page
);
1028 #if MAX_NUMNODES > 1
1029 if (unlikely(do_numainfo
))
1030 atomic_inc(&memcg
->numainfo_events
);
1036 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
1038 return mem_cgroup_from_css(
1039 cgroup_subsys_state(cont
, mem_cgroup_subsys_id
));
1042 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1045 * mm_update_next_owner() may clear mm->owner to NULL
1046 * if it races with swapoff, page migration, etc.
1047 * So this can be called with p == NULL.
1052 return mem_cgroup_from_css(task_subsys_state(p
, mem_cgroup_subsys_id
));
1055 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1057 struct mem_cgroup
*memcg
= NULL
;
1062 * Because we have no locks, mm->owner's may be being moved to other
1063 * cgroup. We use css_tryget() here even if this looks
1064 * pessimistic (rather than adding locks here).
1068 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1069 if (unlikely(!memcg
))
1071 } while (!css_tryget(&memcg
->css
));
1077 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1078 * @root: hierarchy root
1079 * @prev: previously returned memcg, NULL on first invocation
1080 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1082 * Returns references to children of the hierarchy below @root, or
1083 * @root itself, or %NULL after a full round-trip.
1085 * Caller must pass the return value in @prev on subsequent
1086 * invocations for reference counting, or use mem_cgroup_iter_break()
1087 * to cancel a hierarchy walk before the round-trip is complete.
1089 * Reclaimers can specify a zone and a priority level in @reclaim to
1090 * divide up the memcgs in the hierarchy among all concurrent
1091 * reclaimers operating on the same zone and priority.
1093 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1094 struct mem_cgroup
*prev
,
1095 struct mem_cgroup_reclaim_cookie
*reclaim
)
1097 struct mem_cgroup
*memcg
= NULL
;
1098 struct mem_cgroup
*last_visited
= NULL
;
1099 unsigned long uninitialized_var(dead_count
);
1101 if (mem_cgroup_disabled())
1105 root
= root_mem_cgroup
;
1107 if (prev
&& !reclaim
)
1108 last_visited
= prev
;
1110 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1118 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1121 int nid
= zone_to_nid(reclaim
->zone
);
1122 int zid
= zone_idx(reclaim
->zone
);
1123 struct mem_cgroup_per_zone
*mz
;
1125 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1126 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1127 last_visited
= iter
->last_visited
;
1128 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1129 iter
->last_visited
= NULL
;
1134 * If the dead_count mismatches, a destruction
1135 * has happened or is happening concurrently.
1136 * If the dead_count matches, a destruction
1137 * might still happen concurrently, but since
1138 * we checked under RCU, that destruction
1139 * won't free the object until we release the
1140 * RCU reader lock. Thus, the dead_count
1141 * check verifies the pointer is still valid,
1142 * css_tryget() verifies the cgroup pointed to
1145 dead_count
= atomic_read(&root
->dead_count
);
1147 last_visited
= iter
->last_visited
;
1149 if ((dead_count
!= iter
->last_dead_count
) ||
1150 !css_tryget(&last_visited
->css
)) {
1151 last_visited
= NULL
;
1157 * Root is not visited by cgroup iterators so it needs an
1160 if (!last_visited
) {
1163 struct cgroup
*prev_cgroup
, *next_cgroup
;
1165 prev_cgroup
= (last_visited
== root
) ? NULL
1166 : last_visited
->css
.cgroup
;
1168 next_cgroup
= cgroup_next_descendant_pre(
1169 prev_cgroup
, root
->css
.cgroup
);
1172 * Even if we found a group we have to make sure it is
1173 * alive. css && !memcg means that the groups should be
1174 * skipped and we should continue the tree walk.
1175 * last_visited css is safe to use because it is
1176 * protected by css_get and the tree walk is rcu safe.
1179 struct mem_cgroup
*mem
= mem_cgroup_from_cont(
1181 if (css_tryget(&mem
->css
))
1184 prev_cgroup
= next_cgroup
;
1192 css_put(&last_visited
->css
);
1194 iter
->last_visited
= memcg
;
1196 iter
->last_dead_count
= dead_count
;
1200 else if (!prev
&& memcg
)
1201 reclaim
->generation
= iter
->generation
;
1210 if (prev
&& prev
!= root
)
1211 css_put(&prev
->css
);
1217 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1218 * @root: hierarchy root
1219 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1221 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1222 struct mem_cgroup
*prev
)
1225 root
= root_mem_cgroup
;
1226 if (prev
&& prev
!= root
)
1227 css_put(&prev
->css
);
1231 * Iteration constructs for visiting all cgroups (under a tree). If
1232 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1233 * be used for reference counting.
1235 #define for_each_mem_cgroup_tree(iter, root) \
1236 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1238 iter = mem_cgroup_iter(root, iter, NULL))
1240 #define for_each_mem_cgroup(iter) \
1241 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1243 iter = mem_cgroup_iter(NULL, iter, NULL))
1245 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1247 struct mem_cgroup
*memcg
;
1250 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1251 if (unlikely(!memcg
))
1256 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1259 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1267 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1270 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1271 * @zone: zone of the wanted lruvec
1272 * @memcg: memcg of the wanted lruvec
1274 * Returns the lru list vector holding pages for the given @zone and
1275 * @mem. This can be the global zone lruvec, if the memory controller
1278 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1279 struct mem_cgroup
*memcg
)
1281 struct mem_cgroup_per_zone
*mz
;
1282 struct lruvec
*lruvec
;
1284 if (mem_cgroup_disabled()) {
1285 lruvec
= &zone
->lruvec
;
1289 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1290 lruvec
= &mz
->lruvec
;
1293 * Since a node can be onlined after the mem_cgroup was created,
1294 * we have to be prepared to initialize lruvec->zone here;
1295 * and if offlined then reonlined, we need to reinitialize it.
1297 if (unlikely(lruvec
->zone
!= zone
))
1298 lruvec
->zone
= zone
;
1303 * Following LRU functions are allowed to be used without PCG_LOCK.
1304 * Operations are called by routine of global LRU independently from memcg.
1305 * What we have to take care of here is validness of pc->mem_cgroup.
1307 * Changes to pc->mem_cgroup happens when
1310 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1311 * It is added to LRU before charge.
1312 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1313 * When moving account, the page is not on LRU. It's isolated.
1317 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1319 * @zone: zone of the page
1321 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1323 struct mem_cgroup_per_zone
*mz
;
1324 struct mem_cgroup
*memcg
;
1325 struct page_cgroup
*pc
;
1326 struct lruvec
*lruvec
;
1328 if (mem_cgroup_disabled()) {
1329 lruvec
= &zone
->lruvec
;
1333 pc
= lookup_page_cgroup(page
);
1334 memcg
= pc
->mem_cgroup
;
1337 * Surreptitiously switch any uncharged offlist page to root:
1338 * an uncharged page off lru does nothing to secure
1339 * its former mem_cgroup from sudden removal.
1341 * Our caller holds lru_lock, and PageCgroupUsed is updated
1342 * under page_cgroup lock: between them, they make all uses
1343 * of pc->mem_cgroup safe.
1345 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1346 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1348 mz
= page_cgroup_zoneinfo(memcg
, page
);
1349 lruvec
= &mz
->lruvec
;
1352 * Since a node can be onlined after the mem_cgroup was created,
1353 * we have to be prepared to initialize lruvec->zone here;
1354 * and if offlined then reonlined, we need to reinitialize it.
1356 if (unlikely(lruvec
->zone
!= zone
))
1357 lruvec
->zone
= zone
;
1362 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1363 * @lruvec: mem_cgroup per zone lru vector
1364 * @lru: index of lru list the page is sitting on
1365 * @nr_pages: positive when adding or negative when removing
1367 * This function must be called when a page is added to or removed from an
1370 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1373 struct mem_cgroup_per_zone
*mz
;
1374 unsigned long *lru_size
;
1376 if (mem_cgroup_disabled())
1379 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1380 lru_size
= mz
->lru_size
+ lru
;
1381 *lru_size
+= nr_pages
;
1382 VM_BUG_ON((long)(*lru_size
) < 0);
1386 * Checks whether given mem is same or in the root_mem_cgroup's
1389 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1390 struct mem_cgroup
*memcg
)
1392 if (root_memcg
== memcg
)
1394 if (!root_memcg
->use_hierarchy
|| !memcg
)
1396 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1399 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1400 struct mem_cgroup
*memcg
)
1405 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1410 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*memcg
)
1413 struct mem_cgroup
*curr
= NULL
;
1414 struct task_struct
*p
;
1416 p
= find_lock_task_mm(task
);
1418 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1422 * All threads may have already detached their mm's, but the oom
1423 * killer still needs to detect if they have already been oom
1424 * killed to prevent needlessly killing additional tasks.
1427 curr
= mem_cgroup_from_task(task
);
1429 css_get(&curr
->css
);
1435 * We should check use_hierarchy of "memcg" not "curr". Because checking
1436 * use_hierarchy of "curr" here make this function true if hierarchy is
1437 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1438 * hierarchy(even if use_hierarchy is disabled in "memcg").
1440 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1441 css_put(&curr
->css
);
1445 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1447 unsigned long inactive_ratio
;
1448 unsigned long inactive
;
1449 unsigned long active
;
1452 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1453 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1455 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1457 inactive_ratio
= int_sqrt(10 * gb
);
1461 return inactive
* inactive_ratio
< active
;
1464 #define mem_cgroup_from_res_counter(counter, member) \
1465 container_of(counter, struct mem_cgroup, member)
1468 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1469 * @memcg: the memory cgroup
1471 * Returns the maximum amount of memory @mem can be charged with, in
1474 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1476 unsigned long long margin
;
1478 margin
= res_counter_margin(&memcg
->res
);
1479 if (do_swap_account
)
1480 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1481 return margin
>> PAGE_SHIFT
;
1484 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1486 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1489 if (cgrp
->parent
== NULL
)
1490 return vm_swappiness
;
1492 return memcg
->swappiness
;
1496 * memcg->moving_account is used for checking possibility that some thread is
1497 * calling move_account(). When a thread on CPU-A starts moving pages under
1498 * a memcg, other threads should check memcg->moving_account under
1499 * rcu_read_lock(), like this:
1503 * memcg->moving_account+1 if (memcg->mocing_account)
1505 * synchronize_rcu() update something.
1510 /* for quick checking without looking up memcg */
1511 atomic_t memcg_moving __read_mostly
;
1513 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1515 atomic_inc(&memcg_moving
);
1516 atomic_inc(&memcg
->moving_account
);
1520 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1523 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1524 * We check NULL in callee rather than caller.
1527 atomic_dec(&memcg_moving
);
1528 atomic_dec(&memcg
->moving_account
);
1533 * 2 routines for checking "mem" is under move_account() or not.
1535 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1536 * is used for avoiding races in accounting. If true,
1537 * pc->mem_cgroup may be overwritten.
1539 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1540 * under hierarchy of moving cgroups. This is for
1541 * waiting at hith-memory prressure caused by "move".
1544 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1546 VM_BUG_ON(!rcu_read_lock_held());
1547 return atomic_read(&memcg
->moving_account
) > 0;
1550 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1552 struct mem_cgroup
*from
;
1553 struct mem_cgroup
*to
;
1556 * Unlike task_move routines, we access mc.to, mc.from not under
1557 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1559 spin_lock(&mc
.lock
);
1565 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1566 || mem_cgroup_same_or_subtree(memcg
, to
);
1568 spin_unlock(&mc
.lock
);
1572 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1574 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1575 if (mem_cgroup_under_move(memcg
)) {
1577 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1578 /* moving charge context might have finished. */
1581 finish_wait(&mc
.waitq
, &wait
);
1589 * Take this lock when
1590 * - a code tries to modify page's memcg while it's USED.
1591 * - a code tries to modify page state accounting in a memcg.
1592 * see mem_cgroup_stolen(), too.
1594 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1595 unsigned long *flags
)
1597 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1600 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1601 unsigned long *flags
)
1603 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1606 #define K(x) ((x) << (PAGE_SHIFT-10))
1608 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1609 * @memcg: The memory cgroup that went over limit
1610 * @p: Task that is going to be killed
1612 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1615 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1617 struct cgroup
*task_cgrp
;
1618 struct cgroup
*mem_cgrp
;
1620 * Need a buffer in BSS, can't rely on allocations. The code relies
1621 * on the assumption that OOM is serialized for memory controller.
1622 * If this assumption is broken, revisit this code.
1624 static char memcg_name
[PATH_MAX
];
1626 struct mem_cgroup
*iter
;
1634 mem_cgrp
= memcg
->css
.cgroup
;
1635 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1637 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1640 * Unfortunately, we are unable to convert to a useful name
1641 * But we'll still print out the usage information
1648 pr_info("Task in %s killed", memcg_name
);
1651 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1659 * Continues from above, so we don't need an KERN_ level
1661 pr_cont(" as a result of limit of %s\n", memcg_name
);
1664 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1665 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1666 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1667 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1668 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1669 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1670 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1671 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1672 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1673 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1674 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1675 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1677 for_each_mem_cgroup_tree(iter
, memcg
) {
1678 pr_info("Memory cgroup stats");
1681 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1683 pr_cont(" for %s", memcg_name
);
1687 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1688 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1690 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1691 K(mem_cgroup_read_stat(iter
, i
)));
1694 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1695 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1696 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1703 * This function returns the number of memcg under hierarchy tree. Returns
1704 * 1(self count) if no children.
1706 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1709 struct mem_cgroup
*iter
;
1711 for_each_mem_cgroup_tree(iter
, memcg
)
1717 * Return the memory (and swap, if configured) limit for a memcg.
1719 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1723 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1726 * Do not consider swap space if we cannot swap due to swappiness
1728 if (mem_cgroup_swappiness(memcg
)) {
1731 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1732 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1735 * If memsw is finite and limits the amount of swap space
1736 * available to this memcg, return that limit.
1738 limit
= min(limit
, memsw
);
1744 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1747 struct mem_cgroup
*iter
;
1748 unsigned long chosen_points
= 0;
1749 unsigned long totalpages
;
1750 unsigned int points
= 0;
1751 struct task_struct
*chosen
= NULL
;
1754 * If current has a pending SIGKILL, then automatically select it. The
1755 * goal is to allow it to allocate so that it may quickly exit and free
1758 if (fatal_signal_pending(current
)) {
1759 set_thread_flag(TIF_MEMDIE
);
1763 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1764 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1765 for_each_mem_cgroup_tree(iter
, memcg
) {
1766 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1767 struct cgroup_iter it
;
1768 struct task_struct
*task
;
1770 cgroup_iter_start(cgroup
, &it
);
1771 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1772 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1774 case OOM_SCAN_SELECT
:
1776 put_task_struct(chosen
);
1778 chosen_points
= ULONG_MAX
;
1779 get_task_struct(chosen
);
1781 case OOM_SCAN_CONTINUE
:
1783 case OOM_SCAN_ABORT
:
1784 cgroup_iter_end(cgroup
, &it
);
1785 mem_cgroup_iter_break(memcg
, iter
);
1787 put_task_struct(chosen
);
1792 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1793 if (points
> chosen_points
) {
1795 put_task_struct(chosen
);
1797 chosen_points
= points
;
1798 get_task_struct(chosen
);
1801 cgroup_iter_end(cgroup
, &it
);
1806 points
= chosen_points
* 1000 / totalpages
;
1807 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1808 NULL
, "Memory cgroup out of memory");
1811 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1813 unsigned long flags
)
1815 unsigned long total
= 0;
1816 bool noswap
= false;
1819 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1821 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1824 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1826 drain_all_stock_async(memcg
);
1827 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1829 * Allow limit shrinkers, which are triggered directly
1830 * by userspace, to catch signals and stop reclaim
1831 * after minimal progress, regardless of the margin.
1833 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1835 if (mem_cgroup_margin(memcg
))
1838 * If nothing was reclaimed after two attempts, there
1839 * may be no reclaimable pages in this hierarchy.
1848 * test_mem_cgroup_node_reclaimable
1849 * @memcg: the target memcg
1850 * @nid: the node ID to be checked.
1851 * @noswap : specify true here if the user wants flle only information.
1853 * This function returns whether the specified memcg contains any
1854 * reclaimable pages on a node. Returns true if there are any reclaimable
1855 * pages in the node.
1857 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1858 int nid
, bool noswap
)
1860 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1862 if (noswap
|| !total_swap_pages
)
1864 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1869 #if MAX_NUMNODES > 1
1872 * Always updating the nodemask is not very good - even if we have an empty
1873 * list or the wrong list here, we can start from some node and traverse all
1874 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1877 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1881 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1882 * pagein/pageout changes since the last update.
1884 if (!atomic_read(&memcg
->numainfo_events
))
1886 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1889 /* make a nodemask where this memcg uses memory from */
1890 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1892 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1894 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1895 node_clear(nid
, memcg
->scan_nodes
);
1898 atomic_set(&memcg
->numainfo_events
, 0);
1899 atomic_set(&memcg
->numainfo_updating
, 0);
1903 * Selecting a node where we start reclaim from. Because what we need is just
1904 * reducing usage counter, start from anywhere is O,K. Considering
1905 * memory reclaim from current node, there are pros. and cons.
1907 * Freeing memory from current node means freeing memory from a node which
1908 * we'll use or we've used. So, it may make LRU bad. And if several threads
1909 * hit limits, it will see a contention on a node. But freeing from remote
1910 * node means more costs for memory reclaim because of memory latency.
1912 * Now, we use round-robin. Better algorithm is welcomed.
1914 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1918 mem_cgroup_may_update_nodemask(memcg
);
1919 node
= memcg
->last_scanned_node
;
1921 node
= next_node(node
, memcg
->scan_nodes
);
1922 if (node
== MAX_NUMNODES
)
1923 node
= first_node(memcg
->scan_nodes
);
1925 * We call this when we hit limit, not when pages are added to LRU.
1926 * No LRU may hold pages because all pages are UNEVICTABLE or
1927 * memcg is too small and all pages are not on LRU. In that case,
1928 * we use curret node.
1930 if (unlikely(node
== MAX_NUMNODES
))
1931 node
= numa_node_id();
1933 memcg
->last_scanned_node
= node
;
1938 * Check all nodes whether it contains reclaimable pages or not.
1939 * For quick scan, we make use of scan_nodes. This will allow us to skip
1940 * unused nodes. But scan_nodes is lazily updated and may not cotain
1941 * enough new information. We need to do double check.
1943 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1948 * quick check...making use of scan_node.
1949 * We can skip unused nodes.
1951 if (!nodes_empty(memcg
->scan_nodes
)) {
1952 for (nid
= first_node(memcg
->scan_nodes
);
1954 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1956 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1961 * Check rest of nodes.
1963 for_each_node_state(nid
, N_MEMORY
) {
1964 if (node_isset(nid
, memcg
->scan_nodes
))
1966 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1973 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1978 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1980 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1984 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1987 unsigned long *total_scanned
)
1989 struct mem_cgroup
*victim
= NULL
;
1992 unsigned long excess
;
1993 unsigned long nr_scanned
;
1994 struct mem_cgroup_reclaim_cookie reclaim
= {
1999 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2002 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2007 * If we have not been able to reclaim
2008 * anything, it might because there are
2009 * no reclaimable pages under this hierarchy
2014 * We want to do more targeted reclaim.
2015 * excess >> 2 is not to excessive so as to
2016 * reclaim too much, nor too less that we keep
2017 * coming back to reclaim from this cgroup
2019 if (total
>= (excess
>> 2) ||
2020 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2025 if (!mem_cgroup_reclaimable(victim
, false))
2027 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2029 *total_scanned
+= nr_scanned
;
2030 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2033 mem_cgroup_iter_break(root_memcg
, victim
);
2038 * Check OOM-Killer is already running under our hierarchy.
2039 * If someone is running, return false.
2040 * Has to be called with memcg_oom_lock
2042 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
2044 struct mem_cgroup
*iter
, *failed
= NULL
;
2046 for_each_mem_cgroup_tree(iter
, memcg
) {
2047 if (iter
->oom_lock
) {
2049 * this subtree of our hierarchy is already locked
2050 * so we cannot give a lock.
2053 mem_cgroup_iter_break(memcg
, iter
);
2056 iter
->oom_lock
= true;
2063 * OK, we failed to lock the whole subtree so we have to clean up
2064 * what we set up to the failing subtree
2066 for_each_mem_cgroup_tree(iter
, memcg
) {
2067 if (iter
== failed
) {
2068 mem_cgroup_iter_break(memcg
, iter
);
2071 iter
->oom_lock
= false;
2077 * Has to be called with memcg_oom_lock
2079 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2081 struct mem_cgroup
*iter
;
2083 for_each_mem_cgroup_tree(iter
, memcg
)
2084 iter
->oom_lock
= false;
2088 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2090 struct mem_cgroup
*iter
;
2092 for_each_mem_cgroup_tree(iter
, memcg
)
2093 atomic_inc(&iter
->under_oom
);
2096 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2098 struct mem_cgroup
*iter
;
2101 * When a new child is created while the hierarchy is under oom,
2102 * mem_cgroup_oom_lock() may not be called. We have to use
2103 * atomic_add_unless() here.
2105 for_each_mem_cgroup_tree(iter
, memcg
)
2106 atomic_add_unless(&iter
->under_oom
, -1, 0);
2109 static DEFINE_SPINLOCK(memcg_oom_lock
);
2110 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2112 struct oom_wait_info
{
2113 struct mem_cgroup
*memcg
;
2117 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2118 unsigned mode
, int sync
, void *arg
)
2120 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2121 struct mem_cgroup
*oom_wait_memcg
;
2122 struct oom_wait_info
*oom_wait_info
;
2124 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2125 oom_wait_memcg
= oom_wait_info
->memcg
;
2128 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2129 * Then we can use css_is_ancestor without taking care of RCU.
2131 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2132 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2134 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2137 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2139 /* for filtering, pass "memcg" as argument. */
2140 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2143 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2145 if (memcg
&& atomic_read(&memcg
->under_oom
))
2146 memcg_wakeup_oom(memcg
);
2150 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2152 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2155 struct oom_wait_info owait
;
2156 bool locked
, need_to_kill
;
2158 owait
.memcg
= memcg
;
2159 owait
.wait
.flags
= 0;
2160 owait
.wait
.func
= memcg_oom_wake_function
;
2161 owait
.wait
.private = current
;
2162 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2163 need_to_kill
= true;
2164 mem_cgroup_mark_under_oom(memcg
);
2166 /* At first, try to OOM lock hierarchy under memcg.*/
2167 spin_lock(&memcg_oom_lock
);
2168 locked
= mem_cgroup_oom_lock(memcg
);
2170 * Even if signal_pending(), we can't quit charge() loop without
2171 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2172 * under OOM is always welcomed, use TASK_KILLABLE here.
2174 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2175 if (!locked
|| memcg
->oom_kill_disable
)
2176 need_to_kill
= false;
2178 mem_cgroup_oom_notify(memcg
);
2179 spin_unlock(&memcg_oom_lock
);
2182 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2183 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2186 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2188 spin_lock(&memcg_oom_lock
);
2190 mem_cgroup_oom_unlock(memcg
);
2191 memcg_wakeup_oom(memcg
);
2192 spin_unlock(&memcg_oom_lock
);
2194 mem_cgroup_unmark_under_oom(memcg
);
2196 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2198 /* Give chance to dying process */
2199 schedule_timeout_uninterruptible(1);
2204 * Currently used to update mapped file statistics, but the routine can be
2205 * generalized to update other statistics as well.
2207 * Notes: Race condition
2209 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2210 * it tends to be costly. But considering some conditions, we doesn't need
2211 * to do so _always_.
2213 * Considering "charge", lock_page_cgroup() is not required because all
2214 * file-stat operations happen after a page is attached to radix-tree. There
2215 * are no race with "charge".
2217 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2218 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2219 * if there are race with "uncharge". Statistics itself is properly handled
2222 * Considering "move", this is an only case we see a race. To make the race
2223 * small, we check mm->moving_account and detect there are possibility of race
2224 * If there is, we take a lock.
2227 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2228 bool *locked
, unsigned long *flags
)
2230 struct mem_cgroup
*memcg
;
2231 struct page_cgroup
*pc
;
2233 pc
= lookup_page_cgroup(page
);
2235 memcg
= pc
->mem_cgroup
;
2236 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2239 * If this memory cgroup is not under account moving, we don't
2240 * need to take move_lock_mem_cgroup(). Because we already hold
2241 * rcu_read_lock(), any calls to move_account will be delayed until
2242 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2244 if (!mem_cgroup_stolen(memcg
))
2247 move_lock_mem_cgroup(memcg
, flags
);
2248 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2249 move_unlock_mem_cgroup(memcg
, flags
);
2255 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2257 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2260 * It's guaranteed that pc->mem_cgroup never changes while
2261 * lock is held because a routine modifies pc->mem_cgroup
2262 * should take move_lock_mem_cgroup().
2264 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2267 void mem_cgroup_update_page_stat(struct page
*page
,
2268 enum mem_cgroup_page_stat_item idx
, int val
)
2270 struct mem_cgroup
*memcg
;
2271 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2272 unsigned long uninitialized_var(flags
);
2274 if (mem_cgroup_disabled())
2277 memcg
= pc
->mem_cgroup
;
2278 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2282 case MEMCG_NR_FILE_MAPPED
:
2283 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2289 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2293 * size of first charge trial. "32" comes from vmscan.c's magic value.
2294 * TODO: maybe necessary to use big numbers in big irons.
2296 #define CHARGE_BATCH 32U
2297 struct memcg_stock_pcp
{
2298 struct mem_cgroup
*cached
; /* this never be root cgroup */
2299 unsigned int nr_pages
;
2300 struct work_struct work
;
2301 unsigned long flags
;
2302 #define FLUSHING_CACHED_CHARGE 0
2304 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2305 static DEFINE_MUTEX(percpu_charge_mutex
);
2308 * consume_stock: Try to consume stocked charge on this cpu.
2309 * @memcg: memcg to consume from.
2310 * @nr_pages: how many pages to charge.
2312 * The charges will only happen if @memcg matches the current cpu's memcg
2313 * stock, and at least @nr_pages are available in that stock. Failure to
2314 * service an allocation will refill the stock.
2316 * returns true if successful, false otherwise.
2318 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2320 struct memcg_stock_pcp
*stock
;
2323 if (nr_pages
> CHARGE_BATCH
)
2326 stock
= &get_cpu_var(memcg_stock
);
2327 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2328 stock
->nr_pages
-= nr_pages
;
2329 else /* need to call res_counter_charge */
2331 put_cpu_var(memcg_stock
);
2336 * Returns stocks cached in percpu to res_counter and reset cached information.
2338 static void drain_stock(struct memcg_stock_pcp
*stock
)
2340 struct mem_cgroup
*old
= stock
->cached
;
2342 if (stock
->nr_pages
) {
2343 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2345 res_counter_uncharge(&old
->res
, bytes
);
2346 if (do_swap_account
)
2347 res_counter_uncharge(&old
->memsw
, bytes
);
2348 stock
->nr_pages
= 0;
2350 stock
->cached
= NULL
;
2354 * This must be called under preempt disabled or must be called by
2355 * a thread which is pinned to local cpu.
2357 static void drain_local_stock(struct work_struct
*dummy
)
2359 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2361 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2364 static void __init
memcg_stock_init(void)
2368 for_each_possible_cpu(cpu
) {
2369 struct memcg_stock_pcp
*stock
=
2370 &per_cpu(memcg_stock
, cpu
);
2371 INIT_WORK(&stock
->work
, drain_local_stock
);
2376 * Cache charges(val) which is from res_counter, to local per_cpu area.
2377 * This will be consumed by consume_stock() function, later.
2379 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2381 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2383 if (stock
->cached
!= memcg
) { /* reset if necessary */
2385 stock
->cached
= memcg
;
2387 stock
->nr_pages
+= nr_pages
;
2388 put_cpu_var(memcg_stock
);
2392 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2393 * of the hierarchy under it. sync flag says whether we should block
2394 * until the work is done.
2396 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2400 /* Notify other cpus that system-wide "drain" is running */
2403 for_each_online_cpu(cpu
) {
2404 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2405 struct mem_cgroup
*memcg
;
2407 memcg
= stock
->cached
;
2408 if (!memcg
|| !stock
->nr_pages
)
2410 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2412 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2414 drain_local_stock(&stock
->work
);
2416 schedule_work_on(cpu
, &stock
->work
);
2424 for_each_online_cpu(cpu
) {
2425 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2426 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2427 flush_work(&stock
->work
);
2434 * Tries to drain stocked charges in other cpus. This function is asynchronous
2435 * and just put a work per cpu for draining localy on each cpu. Caller can
2436 * expects some charges will be back to res_counter later but cannot wait for
2439 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2442 * If someone calls draining, avoid adding more kworker runs.
2444 if (!mutex_trylock(&percpu_charge_mutex
))
2446 drain_all_stock(root_memcg
, false);
2447 mutex_unlock(&percpu_charge_mutex
);
2450 /* This is a synchronous drain interface. */
2451 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2453 /* called when force_empty is called */
2454 mutex_lock(&percpu_charge_mutex
);
2455 drain_all_stock(root_memcg
, true);
2456 mutex_unlock(&percpu_charge_mutex
);
2460 * This function drains percpu counter value from DEAD cpu and
2461 * move it to local cpu. Note that this function can be preempted.
2463 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2467 spin_lock(&memcg
->pcp_counter_lock
);
2468 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2469 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2471 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2472 memcg
->nocpu_base
.count
[i
] += x
;
2474 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2475 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2477 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2478 memcg
->nocpu_base
.events
[i
] += x
;
2480 spin_unlock(&memcg
->pcp_counter_lock
);
2483 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2484 unsigned long action
,
2487 int cpu
= (unsigned long)hcpu
;
2488 struct memcg_stock_pcp
*stock
;
2489 struct mem_cgroup
*iter
;
2491 if (action
== CPU_ONLINE
)
2494 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2497 for_each_mem_cgroup(iter
)
2498 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2500 stock
= &per_cpu(memcg_stock
, cpu
);
2506 /* See __mem_cgroup_try_charge() for details */
2508 CHARGE_OK
, /* success */
2509 CHARGE_RETRY
, /* need to retry but retry is not bad */
2510 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2511 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2512 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2515 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2516 unsigned int nr_pages
, unsigned int min_pages
,
2519 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2520 struct mem_cgroup
*mem_over_limit
;
2521 struct res_counter
*fail_res
;
2522 unsigned long flags
= 0;
2525 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2528 if (!do_swap_account
)
2530 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2534 res_counter_uncharge(&memcg
->res
, csize
);
2535 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2536 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2538 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2540 * Never reclaim on behalf of optional batching, retry with a
2541 * single page instead.
2543 if (nr_pages
> min_pages
)
2544 return CHARGE_RETRY
;
2546 if (!(gfp_mask
& __GFP_WAIT
))
2547 return CHARGE_WOULDBLOCK
;
2549 if (gfp_mask
& __GFP_NORETRY
)
2550 return CHARGE_NOMEM
;
2552 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2553 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2554 return CHARGE_RETRY
;
2556 * Even though the limit is exceeded at this point, reclaim
2557 * may have been able to free some pages. Retry the charge
2558 * before killing the task.
2560 * Only for regular pages, though: huge pages are rather
2561 * unlikely to succeed so close to the limit, and we fall back
2562 * to regular pages anyway in case of failure.
2564 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2565 return CHARGE_RETRY
;
2568 * At task move, charge accounts can be doubly counted. So, it's
2569 * better to wait until the end of task_move if something is going on.
2571 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2572 return CHARGE_RETRY
;
2574 /* If we don't need to call oom-killer at el, return immediately */
2576 return CHARGE_NOMEM
;
2578 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2579 return CHARGE_OOM_DIE
;
2581 return CHARGE_RETRY
;
2585 * __mem_cgroup_try_charge() does
2586 * 1. detect memcg to be charged against from passed *mm and *ptr,
2587 * 2. update res_counter
2588 * 3. call memory reclaim if necessary.
2590 * In some special case, if the task is fatal, fatal_signal_pending() or
2591 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2592 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2593 * as possible without any hazards. 2: all pages should have a valid
2594 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2595 * pointer, that is treated as a charge to root_mem_cgroup.
2597 * So __mem_cgroup_try_charge() will return
2598 * 0 ... on success, filling *ptr with a valid memcg pointer.
2599 * -ENOMEM ... charge failure because of resource limits.
2600 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2602 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2603 * the oom-killer can be invoked.
2605 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2607 unsigned int nr_pages
,
2608 struct mem_cgroup
**ptr
,
2611 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2612 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2613 struct mem_cgroup
*memcg
= NULL
;
2617 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2618 * in system level. So, allow to go ahead dying process in addition to
2621 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2622 || fatal_signal_pending(current
)))
2626 * We always charge the cgroup the mm_struct belongs to.
2627 * The mm_struct's mem_cgroup changes on task migration if the
2628 * thread group leader migrates. It's possible that mm is not
2629 * set, if so charge the root memcg (happens for pagecache usage).
2632 *ptr
= root_mem_cgroup
;
2634 if (*ptr
) { /* css should be a valid one */
2636 if (mem_cgroup_is_root(memcg
))
2638 if (consume_stock(memcg
, nr_pages
))
2640 css_get(&memcg
->css
);
2642 struct task_struct
*p
;
2645 p
= rcu_dereference(mm
->owner
);
2647 * Because we don't have task_lock(), "p" can exit.
2648 * In that case, "memcg" can point to root or p can be NULL with
2649 * race with swapoff. Then, we have small risk of mis-accouning.
2650 * But such kind of mis-account by race always happens because
2651 * we don't have cgroup_mutex(). It's overkill and we allo that
2653 * (*) swapoff at el will charge against mm-struct not against
2654 * task-struct. So, mm->owner can be NULL.
2656 memcg
= mem_cgroup_from_task(p
);
2658 memcg
= root_mem_cgroup
;
2659 if (mem_cgroup_is_root(memcg
)) {
2663 if (consume_stock(memcg
, nr_pages
)) {
2665 * It seems dagerous to access memcg without css_get().
2666 * But considering how consume_stok works, it's not
2667 * necessary. If consume_stock success, some charges
2668 * from this memcg are cached on this cpu. So, we
2669 * don't need to call css_get()/css_tryget() before
2670 * calling consume_stock().
2675 /* after here, we may be blocked. we need to get refcnt */
2676 if (!css_tryget(&memcg
->css
)) {
2686 /* If killed, bypass charge */
2687 if (fatal_signal_pending(current
)) {
2688 css_put(&memcg
->css
);
2693 if (oom
&& !nr_oom_retries
) {
2695 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2698 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2703 case CHARGE_RETRY
: /* not in OOM situation but retry */
2705 css_put(&memcg
->css
);
2708 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2709 css_put(&memcg
->css
);
2711 case CHARGE_NOMEM
: /* OOM routine works */
2713 css_put(&memcg
->css
);
2716 /* If oom, we never return -ENOMEM */
2719 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2720 css_put(&memcg
->css
);
2723 } while (ret
!= CHARGE_OK
);
2725 if (batch
> nr_pages
)
2726 refill_stock(memcg
, batch
- nr_pages
);
2727 css_put(&memcg
->css
);
2735 *ptr
= root_mem_cgroup
;
2740 * Somemtimes we have to undo a charge we got by try_charge().
2741 * This function is for that and do uncharge, put css's refcnt.
2742 * gotten by try_charge().
2744 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2745 unsigned int nr_pages
)
2747 if (!mem_cgroup_is_root(memcg
)) {
2748 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2750 res_counter_uncharge(&memcg
->res
, bytes
);
2751 if (do_swap_account
)
2752 res_counter_uncharge(&memcg
->memsw
, bytes
);
2757 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2758 * This is useful when moving usage to parent cgroup.
2760 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2761 unsigned int nr_pages
)
2763 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2765 if (mem_cgroup_is_root(memcg
))
2768 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2769 if (do_swap_account
)
2770 res_counter_uncharge_until(&memcg
->memsw
,
2771 memcg
->memsw
.parent
, bytes
);
2775 * A helper function to get mem_cgroup from ID. must be called under
2776 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2777 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2778 * called against removed memcg.)
2780 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2782 struct cgroup_subsys_state
*css
;
2784 /* ID 0 is unused ID */
2787 css
= css_lookup(&mem_cgroup_subsys
, id
);
2790 return mem_cgroup_from_css(css
);
2793 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2795 struct mem_cgroup
*memcg
= NULL
;
2796 struct page_cgroup
*pc
;
2800 VM_BUG_ON(!PageLocked(page
));
2802 pc
= lookup_page_cgroup(page
);
2803 lock_page_cgroup(pc
);
2804 if (PageCgroupUsed(pc
)) {
2805 memcg
= pc
->mem_cgroup
;
2806 if (memcg
&& !css_tryget(&memcg
->css
))
2808 } else if (PageSwapCache(page
)) {
2809 ent
.val
= page_private(page
);
2810 id
= lookup_swap_cgroup_id(ent
);
2812 memcg
= mem_cgroup_lookup(id
);
2813 if (memcg
&& !css_tryget(&memcg
->css
))
2817 unlock_page_cgroup(pc
);
2821 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2823 unsigned int nr_pages
,
2824 enum charge_type ctype
,
2827 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2828 struct zone
*uninitialized_var(zone
);
2829 struct lruvec
*lruvec
;
2830 bool was_on_lru
= false;
2833 lock_page_cgroup(pc
);
2834 VM_BUG_ON(PageCgroupUsed(pc
));
2836 * we don't need page_cgroup_lock about tail pages, becase they are not
2837 * accessed by any other context at this point.
2841 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2842 * may already be on some other mem_cgroup's LRU. Take care of it.
2845 zone
= page_zone(page
);
2846 spin_lock_irq(&zone
->lru_lock
);
2847 if (PageLRU(page
)) {
2848 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2850 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2855 pc
->mem_cgroup
= memcg
;
2857 * We access a page_cgroup asynchronously without lock_page_cgroup().
2858 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2859 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2860 * before USED bit, we need memory barrier here.
2861 * See mem_cgroup_add_lru_list(), etc.
2864 SetPageCgroupUsed(pc
);
2868 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2869 VM_BUG_ON(PageLRU(page
));
2871 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2873 spin_unlock_irq(&zone
->lru_lock
);
2876 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2881 mem_cgroup_charge_statistics(memcg
, anon
, nr_pages
);
2882 unlock_page_cgroup(pc
);
2885 * "charge_statistics" updated event counter. Then, check it.
2886 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2887 * if they exceeds softlimit.
2889 memcg_check_events(memcg
, page
);
2892 static DEFINE_MUTEX(set_limit_mutex
);
2894 #ifdef CONFIG_MEMCG_KMEM
2895 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2897 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2898 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2902 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2903 * in the memcg_cache_params struct.
2905 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2907 struct kmem_cache
*cachep
;
2909 VM_BUG_ON(p
->is_root_cache
);
2910 cachep
= p
->root_cache
;
2911 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2914 #ifdef CONFIG_SLABINFO
2915 static int mem_cgroup_slabinfo_read(struct cgroup
*cont
, struct cftype
*cft
,
2918 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
2919 struct memcg_cache_params
*params
;
2921 if (!memcg_can_account_kmem(memcg
))
2924 print_slabinfo_header(m
);
2926 mutex_lock(&memcg
->slab_caches_mutex
);
2927 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2928 cache_show(memcg_params_to_cache(params
), m
);
2929 mutex_unlock(&memcg
->slab_caches_mutex
);
2935 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2937 struct res_counter
*fail_res
;
2938 struct mem_cgroup
*_memcg
;
2942 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2947 * Conditions under which we can wait for the oom_killer. Those are
2948 * the same conditions tested by the core page allocator
2950 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2953 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2956 if (ret
== -EINTR
) {
2958 * __mem_cgroup_try_charge() chosed to bypass to root due to
2959 * OOM kill or fatal signal. Since our only options are to
2960 * either fail the allocation or charge it to this cgroup, do
2961 * it as a temporary condition. But we can't fail. From a
2962 * kmem/slab perspective, the cache has already been selected,
2963 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2966 * This condition will only trigger if the task entered
2967 * memcg_charge_kmem in a sane state, but was OOM-killed during
2968 * __mem_cgroup_try_charge() above. Tasks that were already
2969 * dying when the allocation triggers should have been already
2970 * directed to the root cgroup in memcontrol.h
2972 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2973 if (do_swap_account
)
2974 res_counter_charge_nofail(&memcg
->memsw
, size
,
2978 res_counter_uncharge(&memcg
->kmem
, size
);
2983 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2985 res_counter_uncharge(&memcg
->res
, size
);
2986 if (do_swap_account
)
2987 res_counter_uncharge(&memcg
->memsw
, size
);
2990 if (res_counter_uncharge(&memcg
->kmem
, size
))
2993 if (memcg_kmem_test_and_clear_dead(memcg
))
2994 mem_cgroup_put(memcg
);
2997 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3002 mutex_lock(&memcg
->slab_caches_mutex
);
3003 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3004 mutex_unlock(&memcg
->slab_caches_mutex
);
3008 * helper for acessing a memcg's index. It will be used as an index in the
3009 * child cache array in kmem_cache, and also to derive its name. This function
3010 * will return -1 when this is not a kmem-limited memcg.
3012 int memcg_cache_id(struct mem_cgroup
*memcg
)
3014 return memcg
? memcg
->kmemcg_id
: -1;
3018 * This ends up being protected by the set_limit mutex, during normal
3019 * operation, because that is its main call site.
3021 * But when we create a new cache, we can call this as well if its parent
3022 * is kmem-limited. That will have to hold set_limit_mutex as well.
3024 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3028 num
= ida_simple_get(&kmem_limited_groups
,
3029 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3033 * After this point, kmem_accounted (that we test atomically in
3034 * the beginning of this conditional), is no longer 0. This
3035 * guarantees only one process will set the following boolean
3036 * to true. We don't need test_and_set because we're protected
3037 * by the set_limit_mutex anyway.
3039 memcg_kmem_set_activated(memcg
);
3041 ret
= memcg_update_all_caches(num
+1);
3043 ida_simple_remove(&kmem_limited_groups
, num
);
3044 memcg_kmem_clear_activated(memcg
);
3048 memcg
->kmemcg_id
= num
;
3049 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3050 mutex_init(&memcg
->slab_caches_mutex
);
3054 static size_t memcg_caches_array_size(int num_groups
)
3057 if (num_groups
<= 0)
3060 size
= 2 * num_groups
;
3061 if (size
< MEMCG_CACHES_MIN_SIZE
)
3062 size
= MEMCG_CACHES_MIN_SIZE
;
3063 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3064 size
= MEMCG_CACHES_MAX_SIZE
;
3070 * We should update the current array size iff all caches updates succeed. This
3071 * can only be done from the slab side. The slab mutex needs to be held when
3074 void memcg_update_array_size(int num
)
3076 if (num
> memcg_limited_groups_array_size
)
3077 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3080 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3082 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3084 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3086 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3088 if (num_groups
> memcg_limited_groups_array_size
) {
3090 ssize_t size
= memcg_caches_array_size(num_groups
);
3092 size
*= sizeof(void *);
3093 size
+= sizeof(struct memcg_cache_params
);
3095 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3096 if (!s
->memcg_params
) {
3097 s
->memcg_params
= cur_params
;
3101 INIT_WORK(&s
->memcg_params
->destroy
,
3102 kmem_cache_destroy_work_func
);
3103 s
->memcg_params
->is_root_cache
= true;
3106 * There is the chance it will be bigger than
3107 * memcg_limited_groups_array_size, if we failed an allocation
3108 * in a cache, in which case all caches updated before it, will
3109 * have a bigger array.
3111 * But if that is the case, the data after
3112 * memcg_limited_groups_array_size is certainly unused
3114 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3115 if (!cur_params
->memcg_caches
[i
])
3117 s
->memcg_params
->memcg_caches
[i
] =
3118 cur_params
->memcg_caches
[i
];
3122 * Ideally, we would wait until all caches succeed, and only
3123 * then free the old one. But this is not worth the extra
3124 * pointer per-cache we'd have to have for this.
3126 * It is not a big deal if some caches are left with a size
3127 * bigger than the others. And all updates will reset this
3135 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3136 struct kmem_cache
*root_cache
)
3138 size_t size
= sizeof(struct memcg_cache_params
);
3140 if (!memcg_kmem_enabled())
3144 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3146 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3147 if (!s
->memcg_params
)
3150 INIT_WORK(&s
->memcg_params
->destroy
,
3151 kmem_cache_destroy_work_func
);
3153 s
->memcg_params
->memcg
= memcg
;
3154 s
->memcg_params
->root_cache
= root_cache
;
3156 s
->memcg_params
->is_root_cache
= true;
3161 void memcg_release_cache(struct kmem_cache
*s
)
3163 struct kmem_cache
*root
;
3164 struct mem_cgroup
*memcg
;
3168 * This happens, for instance, when a root cache goes away before we
3171 if (!s
->memcg_params
)
3174 if (s
->memcg_params
->is_root_cache
)
3177 memcg
= s
->memcg_params
->memcg
;
3178 id
= memcg_cache_id(memcg
);
3180 root
= s
->memcg_params
->root_cache
;
3181 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3182 mem_cgroup_put(memcg
);
3184 mutex_lock(&memcg
->slab_caches_mutex
);
3185 list_del(&s
->memcg_params
->list
);
3186 mutex_unlock(&memcg
->slab_caches_mutex
);
3189 kfree(s
->memcg_params
);
3193 * During the creation a new cache, we need to disable our accounting mechanism
3194 * altogether. This is true even if we are not creating, but rather just
3195 * enqueing new caches to be created.
3197 * This is because that process will trigger allocations; some visible, like
3198 * explicit kmallocs to auxiliary data structures, name strings and internal
3199 * cache structures; some well concealed, like INIT_WORK() that can allocate
3200 * objects during debug.
3202 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3203 * to it. This may not be a bounded recursion: since the first cache creation
3204 * failed to complete (waiting on the allocation), we'll just try to create the
3205 * cache again, failing at the same point.
3207 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3208 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3209 * inside the following two functions.
3211 static inline void memcg_stop_kmem_account(void)
3213 VM_BUG_ON(!current
->mm
);
3214 current
->memcg_kmem_skip_account
++;
3217 static inline void memcg_resume_kmem_account(void)
3219 VM_BUG_ON(!current
->mm
);
3220 current
->memcg_kmem_skip_account
--;
3223 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3225 struct kmem_cache
*cachep
;
3226 struct memcg_cache_params
*p
;
3228 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3230 cachep
= memcg_params_to_cache(p
);
3233 * If we get down to 0 after shrink, we could delete right away.
3234 * However, memcg_release_pages() already puts us back in the workqueue
3235 * in that case. If we proceed deleting, we'll get a dangling
3236 * reference, and removing the object from the workqueue in that case
3237 * is unnecessary complication. We are not a fast path.
3239 * Note that this case is fundamentally different from racing with
3240 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3241 * kmem_cache_shrink, not only we would be reinserting a dead cache
3242 * into the queue, but doing so from inside the worker racing to
3245 * So if we aren't down to zero, we'll just schedule a worker and try
3248 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3249 kmem_cache_shrink(cachep
);
3250 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3253 kmem_cache_destroy(cachep
);
3256 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3258 if (!cachep
->memcg_params
->dead
)
3262 * There are many ways in which we can get here.
3264 * We can get to a memory-pressure situation while the delayed work is
3265 * still pending to run. The vmscan shrinkers can then release all
3266 * cache memory and get us to destruction. If this is the case, we'll
3267 * be executed twice, which is a bug (the second time will execute over
3268 * bogus data). In this case, cancelling the work should be fine.
3270 * But we can also get here from the worker itself, if
3271 * kmem_cache_shrink is enough to shake all the remaining objects and
3272 * get the page count to 0. In this case, we'll deadlock if we try to
3273 * cancel the work (the worker runs with an internal lock held, which
3274 * is the same lock we would hold for cancel_work_sync().)
3276 * Since we can't possibly know who got us here, just refrain from
3277 * running if there is already work pending
3279 if (work_pending(&cachep
->memcg_params
->destroy
))
3282 * We have to defer the actual destroying to a workqueue, because
3283 * we might currently be in a context that cannot sleep.
3285 schedule_work(&cachep
->memcg_params
->destroy
);
3288 static char *memcg_cache_name(struct mem_cgroup
*memcg
, struct kmem_cache
*s
)
3291 struct dentry
*dentry
;
3294 dentry
= rcu_dereference(memcg
->css
.cgroup
->dentry
);
3297 BUG_ON(dentry
== NULL
);
3299 name
= kasprintf(GFP_KERNEL
, "%s(%d:%s)", s
->name
,
3300 memcg_cache_id(memcg
), dentry
->d_name
.name
);
3305 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3306 struct kmem_cache
*s
)
3309 struct kmem_cache
*new;
3311 name
= memcg_cache_name(memcg
, s
);
3315 new = kmem_cache_create_memcg(memcg
, name
, s
->object_size
, s
->align
,
3316 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3319 new->allocflags
|= __GFP_KMEMCG
;
3326 * This lock protects updaters, not readers. We want readers to be as fast as
3327 * they can, and they will either see NULL or a valid cache value. Our model
3328 * allow them to see NULL, in which case the root memcg will be selected.
3330 * We need this lock because multiple allocations to the same cache from a non
3331 * will span more than one worker. Only one of them can create the cache.
3333 static DEFINE_MUTEX(memcg_cache_mutex
);
3334 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3335 struct kmem_cache
*cachep
)
3337 struct kmem_cache
*new_cachep
;
3340 BUG_ON(!memcg_can_account_kmem(memcg
));
3342 idx
= memcg_cache_id(memcg
);
3344 mutex_lock(&memcg_cache_mutex
);
3345 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3349 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3350 if (new_cachep
== NULL
) {
3351 new_cachep
= cachep
;
3355 mem_cgroup_get(memcg
);
3356 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3358 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3360 * the readers won't lock, make sure everybody sees the updated value,
3361 * so they won't put stuff in the queue again for no reason
3365 mutex_unlock(&memcg_cache_mutex
);
3369 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3371 struct kmem_cache
*c
;
3374 if (!s
->memcg_params
)
3376 if (!s
->memcg_params
->is_root_cache
)
3380 * If the cache is being destroyed, we trust that there is no one else
3381 * requesting objects from it. Even if there are, the sanity checks in
3382 * kmem_cache_destroy should caught this ill-case.
3384 * Still, we don't want anyone else freeing memcg_caches under our
3385 * noses, which can happen if a new memcg comes to life. As usual,
3386 * we'll take the set_limit_mutex to protect ourselves against this.
3388 mutex_lock(&set_limit_mutex
);
3389 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3390 c
= s
->memcg_params
->memcg_caches
[i
];
3395 * We will now manually delete the caches, so to avoid races
3396 * we need to cancel all pending destruction workers and
3397 * proceed with destruction ourselves.
3399 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3400 * and that could spawn the workers again: it is likely that
3401 * the cache still have active pages until this very moment.
3402 * This would lead us back to mem_cgroup_destroy_cache.
3404 * But that will not execute at all if the "dead" flag is not
3405 * set, so flip it down to guarantee we are in control.
3407 c
->memcg_params
->dead
= false;
3408 cancel_work_sync(&c
->memcg_params
->destroy
);
3409 kmem_cache_destroy(c
);
3411 mutex_unlock(&set_limit_mutex
);
3414 struct create_work
{
3415 struct mem_cgroup
*memcg
;
3416 struct kmem_cache
*cachep
;
3417 struct work_struct work
;
3420 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3422 struct kmem_cache
*cachep
;
3423 struct memcg_cache_params
*params
;
3425 if (!memcg_kmem_is_active(memcg
))
3428 mutex_lock(&memcg
->slab_caches_mutex
);
3429 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3430 cachep
= memcg_params_to_cache(params
);
3431 cachep
->memcg_params
->dead
= true;
3432 schedule_work(&cachep
->memcg_params
->destroy
);
3434 mutex_unlock(&memcg
->slab_caches_mutex
);
3437 static void memcg_create_cache_work_func(struct work_struct
*w
)
3439 struct create_work
*cw
;
3441 cw
= container_of(w
, struct create_work
, work
);
3442 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3443 /* Drop the reference gotten when we enqueued. */
3444 css_put(&cw
->memcg
->css
);
3449 * Enqueue the creation of a per-memcg kmem_cache.
3450 * Called with rcu_read_lock.
3452 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3453 struct kmem_cache
*cachep
)
3455 struct create_work
*cw
;
3457 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3461 /* The corresponding put will be done in the workqueue. */
3462 if (!css_tryget(&memcg
->css
)) {
3468 cw
->cachep
= cachep
;
3470 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3471 schedule_work(&cw
->work
);
3474 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3475 struct kmem_cache
*cachep
)
3478 * We need to stop accounting when we kmalloc, because if the
3479 * corresponding kmalloc cache is not yet created, the first allocation
3480 * in __memcg_create_cache_enqueue will recurse.
3482 * However, it is better to enclose the whole function. Depending on
3483 * the debugging options enabled, INIT_WORK(), for instance, can
3484 * trigger an allocation. This too, will make us recurse. Because at
3485 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3486 * the safest choice is to do it like this, wrapping the whole function.
3488 memcg_stop_kmem_account();
3489 __memcg_create_cache_enqueue(memcg
, cachep
);
3490 memcg_resume_kmem_account();
3493 * Return the kmem_cache we're supposed to use for a slab allocation.
3494 * We try to use the current memcg's version of the cache.
3496 * If the cache does not exist yet, if we are the first user of it,
3497 * we either create it immediately, if possible, or create it asynchronously
3499 * In the latter case, we will let the current allocation go through with
3500 * the original cache.
3502 * Can't be called in interrupt context or from kernel threads.
3503 * This function needs to be called with rcu_read_lock() held.
3505 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3508 struct mem_cgroup
*memcg
;
3511 VM_BUG_ON(!cachep
->memcg_params
);
3512 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3514 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3518 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3521 if (!memcg_can_account_kmem(memcg
))
3524 idx
= memcg_cache_id(memcg
);
3527 * barrier to mare sure we're always seeing the up to date value. The
3528 * code updating memcg_caches will issue a write barrier to match this.
3530 read_barrier_depends();
3531 if (unlikely(cachep
->memcg_params
->memcg_caches
[idx
] == NULL
)) {
3533 * If we are in a safe context (can wait, and not in interrupt
3534 * context), we could be be predictable and return right away.
3535 * This would guarantee that the allocation being performed
3536 * already belongs in the new cache.
3538 * However, there are some clashes that can arrive from locking.
3539 * For instance, because we acquire the slab_mutex while doing
3540 * kmem_cache_dup, this means no further allocation could happen
3541 * with the slab_mutex held.
3543 * Also, because cache creation issue get_online_cpus(), this
3544 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3545 * that ends up reversed during cpu hotplug. (cpuset allocates
3546 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3547 * better to defer everything.
3549 memcg_create_cache_enqueue(memcg
, cachep
);
3553 return cachep
->memcg_params
->memcg_caches
[idx
];
3555 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3558 * We need to verify if the allocation against current->mm->owner's memcg is
3559 * possible for the given order. But the page is not allocated yet, so we'll
3560 * need a further commit step to do the final arrangements.
3562 * It is possible for the task to switch cgroups in this mean time, so at
3563 * commit time, we can't rely on task conversion any longer. We'll then use
3564 * the handle argument to return to the caller which cgroup we should commit
3565 * against. We could also return the memcg directly and avoid the pointer
3566 * passing, but a boolean return value gives better semantics considering
3567 * the compiled-out case as well.
3569 * Returning true means the allocation is possible.
3572 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3574 struct mem_cgroup
*memcg
;
3578 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3581 * very rare case described in mem_cgroup_from_task. Unfortunately there
3582 * isn't much we can do without complicating this too much, and it would
3583 * be gfp-dependent anyway. Just let it go
3585 if (unlikely(!memcg
))
3588 if (!memcg_can_account_kmem(memcg
)) {
3589 css_put(&memcg
->css
);
3593 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3597 css_put(&memcg
->css
);
3601 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3604 struct page_cgroup
*pc
;
3606 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3608 /* The page allocation failed. Revert */
3610 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3614 pc
= lookup_page_cgroup(page
);
3615 lock_page_cgroup(pc
);
3616 pc
->mem_cgroup
= memcg
;
3617 SetPageCgroupUsed(pc
);
3618 unlock_page_cgroup(pc
);
3621 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3623 struct mem_cgroup
*memcg
= NULL
;
3624 struct page_cgroup
*pc
;
3627 pc
= lookup_page_cgroup(page
);
3629 * Fast unlocked return. Theoretically might have changed, have to
3630 * check again after locking.
3632 if (!PageCgroupUsed(pc
))
3635 lock_page_cgroup(pc
);
3636 if (PageCgroupUsed(pc
)) {
3637 memcg
= pc
->mem_cgroup
;
3638 ClearPageCgroupUsed(pc
);
3640 unlock_page_cgroup(pc
);
3643 * We trust that only if there is a memcg associated with the page, it
3644 * is a valid allocation
3649 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3650 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3653 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3656 #endif /* CONFIG_MEMCG_KMEM */
3658 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3660 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3662 * Because tail pages are not marked as "used", set it. We're under
3663 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3664 * charge/uncharge will be never happen and move_account() is done under
3665 * compound_lock(), so we don't have to take care of races.
3667 void mem_cgroup_split_huge_fixup(struct page
*head
)
3669 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3670 struct page_cgroup
*pc
;
3673 if (mem_cgroup_disabled())
3675 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3677 pc
->mem_cgroup
= head_pc
->mem_cgroup
;
3678 smp_wmb();/* see __commit_charge() */
3679 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3682 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3685 * mem_cgroup_move_account - move account of the page
3687 * @nr_pages: number of regular pages (>1 for huge pages)
3688 * @pc: page_cgroup of the page.
3689 * @from: mem_cgroup which the page is moved from.
3690 * @to: mem_cgroup which the page is moved to. @from != @to.
3692 * The caller must confirm following.
3693 * - page is not on LRU (isolate_page() is useful.)
3694 * - compound_lock is held when nr_pages > 1
3696 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3699 static int mem_cgroup_move_account(struct page
*page
,
3700 unsigned int nr_pages
,
3701 struct page_cgroup
*pc
,
3702 struct mem_cgroup
*from
,
3703 struct mem_cgroup
*to
)
3705 unsigned long flags
;
3707 bool anon
= PageAnon(page
);
3709 VM_BUG_ON(from
== to
);
3710 VM_BUG_ON(PageLRU(page
));
3712 * The page is isolated from LRU. So, collapse function
3713 * will not handle this page. But page splitting can happen.
3714 * Do this check under compound_page_lock(). The caller should
3718 if (nr_pages
> 1 && !PageTransHuge(page
))
3721 lock_page_cgroup(pc
);
3724 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3727 move_lock_mem_cgroup(from
, &flags
);
3729 if (!anon
&& page_mapped(page
)) {
3730 /* Update mapped_file data for mem_cgroup */
3732 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3733 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3736 mem_cgroup_charge_statistics(from
, anon
, -nr_pages
);
3738 /* caller should have done css_get */
3739 pc
->mem_cgroup
= to
;
3740 mem_cgroup_charge_statistics(to
, anon
, nr_pages
);
3741 move_unlock_mem_cgroup(from
, &flags
);
3744 unlock_page_cgroup(pc
);
3748 memcg_check_events(to
, page
);
3749 memcg_check_events(from
, page
);
3755 * mem_cgroup_move_parent - moves page to the parent group
3756 * @page: the page to move
3757 * @pc: page_cgroup of the page
3758 * @child: page's cgroup
3760 * move charges to its parent or the root cgroup if the group has no
3761 * parent (aka use_hierarchy==0).
3762 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3763 * mem_cgroup_move_account fails) the failure is always temporary and
3764 * it signals a race with a page removal/uncharge or migration. In the
3765 * first case the page is on the way out and it will vanish from the LRU
3766 * on the next attempt and the call should be retried later.
3767 * Isolation from the LRU fails only if page has been isolated from
3768 * the LRU since we looked at it and that usually means either global
3769 * reclaim or migration going on. The page will either get back to the
3771 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3772 * (!PageCgroupUsed) or moved to a different group. The page will
3773 * disappear in the next attempt.
3775 static int mem_cgroup_move_parent(struct page
*page
,
3776 struct page_cgroup
*pc
,
3777 struct mem_cgroup
*child
)
3779 struct mem_cgroup
*parent
;
3780 unsigned int nr_pages
;
3781 unsigned long uninitialized_var(flags
);
3784 VM_BUG_ON(mem_cgroup_is_root(child
));
3787 if (!get_page_unless_zero(page
))
3789 if (isolate_lru_page(page
))
3792 nr_pages
= hpage_nr_pages(page
);
3794 parent
= parent_mem_cgroup(child
);
3796 * If no parent, move charges to root cgroup.
3799 parent
= root_mem_cgroup
;
3802 VM_BUG_ON(!PageTransHuge(page
));
3803 flags
= compound_lock_irqsave(page
);
3806 ret
= mem_cgroup_move_account(page
, nr_pages
,
3809 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3812 compound_unlock_irqrestore(page
, flags
);
3813 putback_lru_page(page
);
3821 * Charge the memory controller for page usage.
3823 * 0 if the charge was successful
3824 * < 0 if the cgroup is over its limit
3826 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3827 gfp_t gfp_mask
, enum charge_type ctype
)
3829 struct mem_cgroup
*memcg
= NULL
;
3830 unsigned int nr_pages
= 1;
3834 if (PageTransHuge(page
)) {
3835 nr_pages
<<= compound_order(page
);
3836 VM_BUG_ON(!PageTransHuge(page
));
3838 * Never OOM-kill a process for a huge page. The
3839 * fault handler will fall back to regular pages.
3844 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3847 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3851 int mem_cgroup_newpage_charge(struct page
*page
,
3852 struct mm_struct
*mm
, gfp_t gfp_mask
)
3854 if (mem_cgroup_disabled())
3856 VM_BUG_ON(page_mapped(page
));
3857 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3859 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3860 MEM_CGROUP_CHARGE_TYPE_ANON
);
3864 * While swap-in, try_charge -> commit or cancel, the page is locked.
3865 * And when try_charge() successfully returns, one refcnt to memcg without
3866 * struct page_cgroup is acquired. This refcnt will be consumed by
3867 * "commit()" or removed by "cancel()"
3869 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3872 struct mem_cgroup
**memcgp
)
3874 struct mem_cgroup
*memcg
;
3875 struct page_cgroup
*pc
;
3878 pc
= lookup_page_cgroup(page
);
3880 * Every swap fault against a single page tries to charge the
3881 * page, bail as early as possible. shmem_unuse() encounters
3882 * already charged pages, too. The USED bit is protected by
3883 * the page lock, which serializes swap cache removal, which
3884 * in turn serializes uncharging.
3886 if (PageCgroupUsed(pc
))
3888 if (!do_swap_account
)
3890 memcg
= try_get_mem_cgroup_from_page(page
);
3894 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3895 css_put(&memcg
->css
);
3900 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3906 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3907 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3910 if (mem_cgroup_disabled())
3913 * A racing thread's fault, or swapoff, may have already
3914 * updated the pte, and even removed page from swap cache: in
3915 * those cases unuse_pte()'s pte_same() test will fail; but
3916 * there's also a KSM case which does need to charge the page.
3918 if (!PageSwapCache(page
)) {
3921 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3926 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3929 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3931 if (mem_cgroup_disabled())
3935 __mem_cgroup_cancel_charge(memcg
, 1);
3939 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3940 enum charge_type ctype
)
3942 if (mem_cgroup_disabled())
3947 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3949 * Now swap is on-memory. This means this page may be
3950 * counted both as mem and swap....double count.
3951 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3952 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3953 * may call delete_from_swap_cache() before reach here.
3955 if (do_swap_account
&& PageSwapCache(page
)) {
3956 swp_entry_t ent
= {.val
= page_private(page
)};
3957 mem_cgroup_uncharge_swap(ent
);
3961 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3962 struct mem_cgroup
*memcg
)
3964 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3965 MEM_CGROUP_CHARGE_TYPE_ANON
);
3968 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3971 struct mem_cgroup
*memcg
= NULL
;
3972 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3975 if (mem_cgroup_disabled())
3977 if (PageCompound(page
))
3980 if (!PageSwapCache(page
))
3981 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3982 else { /* page is swapcache/shmem */
3983 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3986 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3991 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3992 unsigned int nr_pages
,
3993 const enum charge_type ctype
)
3995 struct memcg_batch_info
*batch
= NULL
;
3996 bool uncharge_memsw
= true;
3998 /* If swapout, usage of swap doesn't decrease */
3999 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4000 uncharge_memsw
= false;
4002 batch
= ¤t
->memcg_batch
;
4004 * In usual, we do css_get() when we remember memcg pointer.
4005 * But in this case, we keep res->usage until end of a series of
4006 * uncharges. Then, it's ok to ignore memcg's refcnt.
4009 batch
->memcg
= memcg
;
4011 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4012 * In those cases, all pages freed continuously can be expected to be in
4013 * the same cgroup and we have chance to coalesce uncharges.
4014 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4015 * because we want to do uncharge as soon as possible.
4018 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4019 goto direct_uncharge
;
4022 goto direct_uncharge
;
4025 * In typical case, batch->memcg == mem. This means we can
4026 * merge a series of uncharges to an uncharge of res_counter.
4027 * If not, we uncharge res_counter ony by one.
4029 if (batch
->memcg
!= memcg
)
4030 goto direct_uncharge
;
4031 /* remember freed charge and uncharge it later */
4034 batch
->memsw_nr_pages
++;
4037 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4039 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4040 if (unlikely(batch
->memcg
!= memcg
))
4041 memcg_oom_recover(memcg
);
4045 * uncharge if !page_mapped(page)
4047 static struct mem_cgroup
*
4048 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4051 struct mem_cgroup
*memcg
= NULL
;
4052 unsigned int nr_pages
= 1;
4053 struct page_cgroup
*pc
;
4056 if (mem_cgroup_disabled())
4059 VM_BUG_ON(PageSwapCache(page
));
4061 if (PageTransHuge(page
)) {
4062 nr_pages
<<= compound_order(page
);
4063 VM_BUG_ON(!PageTransHuge(page
));
4066 * Check if our page_cgroup is valid
4068 pc
= lookup_page_cgroup(page
);
4069 if (unlikely(!PageCgroupUsed(pc
)))
4072 lock_page_cgroup(pc
);
4074 memcg
= pc
->mem_cgroup
;
4076 if (!PageCgroupUsed(pc
))
4079 anon
= PageAnon(page
);
4082 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4084 * Generally PageAnon tells if it's the anon statistics to be
4085 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4086 * used before page reached the stage of being marked PageAnon.
4090 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4091 /* See mem_cgroup_prepare_migration() */
4092 if (page_mapped(page
))
4095 * Pages under migration may not be uncharged. But
4096 * end_migration() /must/ be the one uncharging the
4097 * unused post-migration page and so it has to call
4098 * here with the migration bit still set. See the
4099 * res_counter handling below.
4101 if (!end_migration
&& PageCgroupMigration(pc
))
4104 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4105 if (!PageAnon(page
)) { /* Shared memory */
4106 if (page
->mapping
&& !page_is_file_cache(page
))
4108 } else if (page_mapped(page
)) /* Anon */
4115 mem_cgroup_charge_statistics(memcg
, anon
, -nr_pages
);
4117 ClearPageCgroupUsed(pc
);
4119 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4120 * freed from LRU. This is safe because uncharged page is expected not
4121 * to be reused (freed soon). Exception is SwapCache, it's handled by
4122 * special functions.
4125 unlock_page_cgroup(pc
);
4127 * even after unlock, we have memcg->res.usage here and this memcg
4128 * will never be freed.
4130 memcg_check_events(memcg
, page
);
4131 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4132 mem_cgroup_swap_statistics(memcg
, true);
4133 mem_cgroup_get(memcg
);
4136 * Migration does not charge the res_counter for the
4137 * replacement page, so leave it alone when phasing out the
4138 * page that is unused after the migration.
4140 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4141 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4146 unlock_page_cgroup(pc
);
4150 void mem_cgroup_uncharge_page(struct page
*page
)
4153 if (page_mapped(page
))
4155 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4156 if (PageSwapCache(page
))
4158 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4161 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4163 VM_BUG_ON(page_mapped(page
));
4164 VM_BUG_ON(page
->mapping
);
4165 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4169 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4170 * In that cases, pages are freed continuously and we can expect pages
4171 * are in the same memcg. All these calls itself limits the number of
4172 * pages freed at once, then uncharge_start/end() is called properly.
4173 * This may be called prural(2) times in a context,
4176 void mem_cgroup_uncharge_start(void)
4178 current
->memcg_batch
.do_batch
++;
4179 /* We can do nest. */
4180 if (current
->memcg_batch
.do_batch
== 1) {
4181 current
->memcg_batch
.memcg
= NULL
;
4182 current
->memcg_batch
.nr_pages
= 0;
4183 current
->memcg_batch
.memsw_nr_pages
= 0;
4187 void mem_cgroup_uncharge_end(void)
4189 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4191 if (!batch
->do_batch
)
4195 if (batch
->do_batch
) /* If stacked, do nothing. */
4201 * This "batch->memcg" is valid without any css_get/put etc...
4202 * bacause we hide charges behind us.
4204 if (batch
->nr_pages
)
4205 res_counter_uncharge(&batch
->memcg
->res
,
4206 batch
->nr_pages
* PAGE_SIZE
);
4207 if (batch
->memsw_nr_pages
)
4208 res_counter_uncharge(&batch
->memcg
->memsw
,
4209 batch
->memsw_nr_pages
* PAGE_SIZE
);
4210 memcg_oom_recover(batch
->memcg
);
4211 /* forget this pointer (for sanity check) */
4212 batch
->memcg
= NULL
;
4217 * called after __delete_from_swap_cache() and drop "page" account.
4218 * memcg information is recorded to swap_cgroup of "ent"
4221 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4223 struct mem_cgroup
*memcg
;
4224 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4226 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4227 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4229 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4232 * record memcg information, if swapout && memcg != NULL,
4233 * mem_cgroup_get() was called in uncharge().
4235 if (do_swap_account
&& swapout
&& memcg
)
4236 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4240 #ifdef CONFIG_MEMCG_SWAP
4242 * called from swap_entry_free(). remove record in swap_cgroup and
4243 * uncharge "memsw" account.
4245 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4247 struct mem_cgroup
*memcg
;
4250 if (!do_swap_account
)
4253 id
= swap_cgroup_record(ent
, 0);
4255 memcg
= mem_cgroup_lookup(id
);
4258 * We uncharge this because swap is freed.
4259 * This memcg can be obsolete one. We avoid calling css_tryget
4261 if (!mem_cgroup_is_root(memcg
))
4262 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4263 mem_cgroup_swap_statistics(memcg
, false);
4264 mem_cgroup_put(memcg
);
4270 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4271 * @entry: swap entry to be moved
4272 * @from: mem_cgroup which the entry is moved from
4273 * @to: mem_cgroup which the entry is moved to
4275 * It succeeds only when the swap_cgroup's record for this entry is the same
4276 * as the mem_cgroup's id of @from.
4278 * Returns 0 on success, -EINVAL on failure.
4280 * The caller must have charged to @to, IOW, called res_counter_charge() about
4281 * both res and memsw, and called css_get().
4283 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4284 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4286 unsigned short old_id
, new_id
;
4288 old_id
= css_id(&from
->css
);
4289 new_id
= css_id(&to
->css
);
4291 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4292 mem_cgroup_swap_statistics(from
, false);
4293 mem_cgroup_swap_statistics(to
, true);
4295 * This function is only called from task migration context now.
4296 * It postpones res_counter and refcount handling till the end
4297 * of task migration(mem_cgroup_clear_mc()) for performance
4298 * improvement. But we cannot postpone mem_cgroup_get(to)
4299 * because if the process that has been moved to @to does
4300 * swap-in, the refcount of @to might be decreased to 0.
4308 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4309 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4316 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4319 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4320 struct mem_cgroup
**memcgp
)
4322 struct mem_cgroup
*memcg
= NULL
;
4323 unsigned int nr_pages
= 1;
4324 struct page_cgroup
*pc
;
4325 enum charge_type ctype
;
4329 if (mem_cgroup_disabled())
4332 if (PageTransHuge(page
))
4333 nr_pages
<<= compound_order(page
);
4335 pc
= lookup_page_cgroup(page
);
4336 lock_page_cgroup(pc
);
4337 if (PageCgroupUsed(pc
)) {
4338 memcg
= pc
->mem_cgroup
;
4339 css_get(&memcg
->css
);
4341 * At migrating an anonymous page, its mapcount goes down
4342 * to 0 and uncharge() will be called. But, even if it's fully
4343 * unmapped, migration may fail and this page has to be
4344 * charged again. We set MIGRATION flag here and delay uncharge
4345 * until end_migration() is called
4347 * Corner Case Thinking
4349 * When the old page was mapped as Anon and it's unmap-and-freed
4350 * while migration was ongoing.
4351 * If unmap finds the old page, uncharge() of it will be delayed
4352 * until end_migration(). If unmap finds a new page, it's
4353 * uncharged when it make mapcount to be 1->0. If unmap code
4354 * finds swap_migration_entry, the new page will not be mapped
4355 * and end_migration() will find it(mapcount==0).
4358 * When the old page was mapped but migraion fails, the kernel
4359 * remaps it. A charge for it is kept by MIGRATION flag even
4360 * if mapcount goes down to 0. We can do remap successfully
4361 * without charging it again.
4364 * The "old" page is under lock_page() until the end of
4365 * migration, so, the old page itself will not be swapped-out.
4366 * If the new page is swapped out before end_migraton, our
4367 * hook to usual swap-out path will catch the event.
4370 SetPageCgroupMigration(pc
);
4372 unlock_page_cgroup(pc
);
4374 * If the page is not charged at this point,
4382 * We charge new page before it's used/mapped. So, even if unlock_page()
4383 * is called before end_migration, we can catch all events on this new
4384 * page. In the case new page is migrated but not remapped, new page's
4385 * mapcount will be finally 0 and we call uncharge in end_migration().
4388 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4390 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4392 * The page is committed to the memcg, but it's not actually
4393 * charged to the res_counter since we plan on replacing the
4394 * old one and only one page is going to be left afterwards.
4396 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4399 /* remove redundant charge if migration failed*/
4400 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4401 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4403 struct page
*used
, *unused
;
4404 struct page_cgroup
*pc
;
4410 if (!migration_ok
) {
4417 anon
= PageAnon(used
);
4418 __mem_cgroup_uncharge_common(unused
,
4419 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4420 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4422 css_put(&memcg
->css
);
4424 * We disallowed uncharge of pages under migration because mapcount
4425 * of the page goes down to zero, temporarly.
4426 * Clear the flag and check the page should be charged.
4428 pc
= lookup_page_cgroup(oldpage
);
4429 lock_page_cgroup(pc
);
4430 ClearPageCgroupMigration(pc
);
4431 unlock_page_cgroup(pc
);
4434 * If a page is a file cache, radix-tree replacement is very atomic
4435 * and we can skip this check. When it was an Anon page, its mapcount
4436 * goes down to 0. But because we added MIGRATION flage, it's not
4437 * uncharged yet. There are several case but page->mapcount check
4438 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4439 * check. (see prepare_charge() also)
4442 mem_cgroup_uncharge_page(used
);
4446 * At replace page cache, newpage is not under any memcg but it's on
4447 * LRU. So, this function doesn't touch res_counter but handles LRU
4448 * in correct way. Both pages are locked so we cannot race with uncharge.
4450 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4451 struct page
*newpage
)
4453 struct mem_cgroup
*memcg
= NULL
;
4454 struct page_cgroup
*pc
;
4455 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4457 if (mem_cgroup_disabled())
4460 pc
= lookup_page_cgroup(oldpage
);
4461 /* fix accounting on old pages */
4462 lock_page_cgroup(pc
);
4463 if (PageCgroupUsed(pc
)) {
4464 memcg
= pc
->mem_cgroup
;
4465 mem_cgroup_charge_statistics(memcg
, false, -1);
4466 ClearPageCgroupUsed(pc
);
4468 unlock_page_cgroup(pc
);
4471 * When called from shmem_replace_page(), in some cases the
4472 * oldpage has already been charged, and in some cases not.
4477 * Even if newpage->mapping was NULL before starting replacement,
4478 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4479 * LRU while we overwrite pc->mem_cgroup.
4481 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4484 #ifdef CONFIG_DEBUG_VM
4485 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4487 struct page_cgroup
*pc
;
4489 pc
= lookup_page_cgroup(page
);
4491 * Can be NULL while feeding pages into the page allocator for
4492 * the first time, i.e. during boot or memory hotplug;
4493 * or when mem_cgroup_disabled().
4495 if (likely(pc
) && PageCgroupUsed(pc
))
4500 bool mem_cgroup_bad_page_check(struct page
*page
)
4502 if (mem_cgroup_disabled())
4505 return lookup_page_cgroup_used(page
) != NULL
;
4508 void mem_cgroup_print_bad_page(struct page
*page
)
4510 struct page_cgroup
*pc
;
4512 pc
= lookup_page_cgroup_used(page
);
4514 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4515 pc
, pc
->flags
, pc
->mem_cgroup
);
4520 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4521 unsigned long long val
)
4524 u64 memswlimit
, memlimit
;
4526 int children
= mem_cgroup_count_children(memcg
);
4527 u64 curusage
, oldusage
;
4531 * For keeping hierarchical_reclaim simple, how long we should retry
4532 * is depends on callers. We set our retry-count to be function
4533 * of # of children which we should visit in this loop.
4535 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4537 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4540 while (retry_count
) {
4541 if (signal_pending(current
)) {
4546 * Rather than hide all in some function, I do this in
4547 * open coded manner. You see what this really does.
4548 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4550 mutex_lock(&set_limit_mutex
);
4551 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4552 if (memswlimit
< val
) {
4554 mutex_unlock(&set_limit_mutex
);
4558 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4562 ret
= res_counter_set_limit(&memcg
->res
, val
);
4564 if (memswlimit
== val
)
4565 memcg
->memsw_is_minimum
= true;
4567 memcg
->memsw_is_minimum
= false;
4569 mutex_unlock(&set_limit_mutex
);
4574 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4575 MEM_CGROUP_RECLAIM_SHRINK
);
4576 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4577 /* Usage is reduced ? */
4578 if (curusage
>= oldusage
)
4581 oldusage
= curusage
;
4583 if (!ret
&& enlarge
)
4584 memcg_oom_recover(memcg
);
4589 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4590 unsigned long long val
)
4593 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4594 int children
= mem_cgroup_count_children(memcg
);
4598 /* see mem_cgroup_resize_res_limit */
4599 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4600 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4601 while (retry_count
) {
4602 if (signal_pending(current
)) {
4607 * Rather than hide all in some function, I do this in
4608 * open coded manner. You see what this really does.
4609 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4611 mutex_lock(&set_limit_mutex
);
4612 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4613 if (memlimit
> val
) {
4615 mutex_unlock(&set_limit_mutex
);
4618 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4619 if (memswlimit
< val
)
4621 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4623 if (memlimit
== val
)
4624 memcg
->memsw_is_minimum
= true;
4626 memcg
->memsw_is_minimum
= false;
4628 mutex_unlock(&set_limit_mutex
);
4633 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4634 MEM_CGROUP_RECLAIM_NOSWAP
|
4635 MEM_CGROUP_RECLAIM_SHRINK
);
4636 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4637 /* Usage is reduced ? */
4638 if (curusage
>= oldusage
)
4641 oldusage
= curusage
;
4643 if (!ret
&& enlarge
)
4644 memcg_oom_recover(memcg
);
4648 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4650 unsigned long *total_scanned
)
4652 unsigned long nr_reclaimed
= 0;
4653 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4654 unsigned long reclaimed
;
4656 struct mem_cgroup_tree_per_zone
*mctz
;
4657 unsigned long long excess
;
4658 unsigned long nr_scanned
;
4663 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4665 * This loop can run a while, specially if mem_cgroup's continuously
4666 * keep exceeding their soft limit and putting the system under
4673 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4678 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4679 gfp_mask
, &nr_scanned
);
4680 nr_reclaimed
+= reclaimed
;
4681 *total_scanned
+= nr_scanned
;
4682 spin_lock(&mctz
->lock
);
4685 * If we failed to reclaim anything from this memory cgroup
4686 * it is time to move on to the next cgroup
4692 * Loop until we find yet another one.
4694 * By the time we get the soft_limit lock
4695 * again, someone might have aded the
4696 * group back on the RB tree. Iterate to
4697 * make sure we get a different mem.
4698 * mem_cgroup_largest_soft_limit_node returns
4699 * NULL if no other cgroup is present on
4703 __mem_cgroup_largest_soft_limit_node(mctz
);
4705 css_put(&next_mz
->memcg
->css
);
4706 else /* next_mz == NULL or other memcg */
4710 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4711 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4713 * One school of thought says that we should not add
4714 * back the node to the tree if reclaim returns 0.
4715 * But our reclaim could return 0, simply because due
4716 * to priority we are exposing a smaller subset of
4717 * memory to reclaim from. Consider this as a longer
4720 /* If excess == 0, no tree ops */
4721 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4722 spin_unlock(&mctz
->lock
);
4723 css_put(&mz
->memcg
->css
);
4726 * Could not reclaim anything and there are no more
4727 * mem cgroups to try or we seem to be looping without
4728 * reclaiming anything.
4730 if (!nr_reclaimed
&&
4732 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4734 } while (!nr_reclaimed
);
4736 css_put(&next_mz
->memcg
->css
);
4737 return nr_reclaimed
;
4741 * mem_cgroup_force_empty_list - clears LRU of a group
4742 * @memcg: group to clear
4745 * @lru: lru to to clear
4747 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4748 * reclaim the pages page themselves - pages are moved to the parent (or root)
4751 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4752 int node
, int zid
, enum lru_list lru
)
4754 struct lruvec
*lruvec
;
4755 unsigned long flags
;
4756 struct list_head
*list
;
4760 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4761 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4762 list
= &lruvec
->lists
[lru
];
4766 struct page_cgroup
*pc
;
4769 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4770 if (list_empty(list
)) {
4771 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4774 page
= list_entry(list
->prev
, struct page
, lru
);
4776 list_move(&page
->lru
, list
);
4778 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4781 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4783 pc
= lookup_page_cgroup(page
);
4785 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4786 /* found lock contention or "pc" is obsolete. */
4791 } while (!list_empty(list
));
4795 * make mem_cgroup's charge to be 0 if there is no task by moving
4796 * all the charges and pages to the parent.
4797 * This enables deleting this mem_cgroup.
4799 * Caller is responsible for holding css reference on the memcg.
4801 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4807 /* This is for making all *used* pages to be on LRU. */
4808 lru_add_drain_all();
4809 drain_all_stock_sync(memcg
);
4810 mem_cgroup_start_move(memcg
);
4811 for_each_node_state(node
, N_MEMORY
) {
4812 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4815 mem_cgroup_force_empty_list(memcg
,
4820 mem_cgroup_end_move(memcg
);
4821 memcg_oom_recover(memcg
);
4825 * Kernel memory may not necessarily be trackable to a specific
4826 * process. So they are not migrated, and therefore we can't
4827 * expect their value to drop to 0 here.
4828 * Having res filled up with kmem only is enough.
4830 * This is a safety check because mem_cgroup_force_empty_list
4831 * could have raced with mem_cgroup_replace_page_cache callers
4832 * so the lru seemed empty but the page could have been added
4833 * right after the check. RES_USAGE should be safe as we always
4834 * charge before adding to the LRU.
4836 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4837 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4838 } while (usage
> 0);
4842 * This mainly exists for tests during the setting of set of use_hierarchy.
4843 * Since this is the very setting we are changing, the current hierarchy value
4846 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4850 /* bounce at first found */
4851 cgroup_for_each_child(pos
, memcg
->css
.cgroup
)
4857 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4858 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4859 * from mem_cgroup_count_children(), in the sense that we don't really care how
4860 * many children we have; we only need to know if we have any. It also counts
4861 * any memcg without hierarchy as infertile.
4863 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4865 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4869 * Reclaims as many pages from the given memcg as possible and moves
4870 * the rest to the parent.
4872 * Caller is responsible for holding css reference for memcg.
4874 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4876 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4877 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4879 /* returns EBUSY if there is a task or if we come here twice. */
4880 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4883 /* we call try-to-free pages for make this cgroup empty */
4884 lru_add_drain_all();
4885 /* try to free all pages in this cgroup */
4886 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4889 if (signal_pending(current
))
4892 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4896 /* maybe some writeback is necessary */
4897 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4902 mem_cgroup_reparent_charges(memcg
);
4907 static int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
4909 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4912 if (mem_cgroup_is_root(memcg
))
4914 css_get(&memcg
->css
);
4915 ret
= mem_cgroup_force_empty(memcg
);
4916 css_put(&memcg
->css
);
4922 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
4924 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
4927 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
4931 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4932 struct cgroup
*parent
= cont
->parent
;
4933 struct mem_cgroup
*parent_memcg
= NULL
;
4936 parent_memcg
= mem_cgroup_from_cont(parent
);
4938 mutex_lock(&memcg_create_mutex
);
4940 if (memcg
->use_hierarchy
== val
)
4944 * If parent's use_hierarchy is set, we can't make any modifications
4945 * in the child subtrees. If it is unset, then the change can
4946 * occur, provided the current cgroup has no children.
4948 * For the root cgroup, parent_mem is NULL, we allow value to be
4949 * set if there are no children.
4951 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4952 (val
== 1 || val
== 0)) {
4953 if (!__memcg_has_children(memcg
))
4954 memcg
->use_hierarchy
= val
;
4961 mutex_unlock(&memcg_create_mutex
);
4967 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4968 enum mem_cgroup_stat_index idx
)
4970 struct mem_cgroup
*iter
;
4973 /* Per-cpu values can be negative, use a signed accumulator */
4974 for_each_mem_cgroup_tree(iter
, memcg
)
4975 val
+= mem_cgroup_read_stat(iter
, idx
);
4977 if (val
< 0) /* race ? */
4982 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4986 if (!mem_cgroup_is_root(memcg
)) {
4988 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4990 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4993 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4994 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4997 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4999 return val
<< PAGE_SHIFT
;
5002 static ssize_t
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
,
5003 struct file
*file
, char __user
*buf
,
5004 size_t nbytes
, loff_t
*ppos
)
5006 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5012 type
= MEMFILE_TYPE(cft
->private);
5013 name
= MEMFILE_ATTR(cft
->private);
5015 if (!do_swap_account
&& type
== _MEMSWAP
)
5020 if (name
== RES_USAGE
)
5021 val
= mem_cgroup_usage(memcg
, false);
5023 val
= res_counter_read_u64(&memcg
->res
, name
);
5026 if (name
== RES_USAGE
)
5027 val
= mem_cgroup_usage(memcg
, true);
5029 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5032 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5038 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5039 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5042 static int memcg_update_kmem_limit(struct cgroup
*cont
, u64 val
)
5045 #ifdef CONFIG_MEMCG_KMEM
5046 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5048 * For simplicity, we won't allow this to be disabled. It also can't
5049 * be changed if the cgroup has children already, or if tasks had
5052 * If tasks join before we set the limit, a person looking at
5053 * kmem.usage_in_bytes will have no way to determine when it took
5054 * place, which makes the value quite meaningless.
5056 * After it first became limited, changes in the value of the limit are
5057 * of course permitted.
5059 mutex_lock(&memcg_create_mutex
);
5060 mutex_lock(&set_limit_mutex
);
5061 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
5062 if (cgroup_task_count(cont
) || memcg_has_children(memcg
)) {
5066 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5069 ret
= memcg_update_cache_sizes(memcg
);
5071 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
5074 static_key_slow_inc(&memcg_kmem_enabled_key
);
5076 * setting the active bit after the inc will guarantee no one
5077 * starts accounting before all call sites are patched
5079 memcg_kmem_set_active(memcg
);
5082 * kmem charges can outlive the cgroup. In the case of slab
5083 * pages, for instance, a page contain objects from various
5084 * processes, so it is unfeasible to migrate them away. We
5085 * need to reference count the memcg because of that.
5087 mem_cgroup_get(memcg
);
5089 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5091 mutex_unlock(&set_limit_mutex
);
5092 mutex_unlock(&memcg_create_mutex
);
5097 #ifdef CONFIG_MEMCG_KMEM
5098 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5101 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5105 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5107 * When that happen, we need to disable the static branch only on those
5108 * memcgs that enabled it. To achieve this, we would be forced to
5109 * complicate the code by keeping track of which memcgs were the ones
5110 * that actually enabled limits, and which ones got it from its
5113 * It is a lot simpler just to do static_key_slow_inc() on every child
5114 * that is accounted.
5116 if (!memcg_kmem_is_active(memcg
))
5120 * destroy(), called if we fail, will issue static_key_slow_inc() and
5121 * mem_cgroup_put() if kmem is enabled. We have to either call them
5122 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5123 * this more consistent, since it always leads to the same destroy path
5125 mem_cgroup_get(memcg
);
5126 static_key_slow_inc(&memcg_kmem_enabled_key
);
5128 mutex_lock(&set_limit_mutex
);
5129 ret
= memcg_update_cache_sizes(memcg
);
5130 mutex_unlock(&set_limit_mutex
);
5134 #endif /* CONFIG_MEMCG_KMEM */
5137 * The user of this function is...
5140 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
5143 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5146 unsigned long long val
;
5149 type
= MEMFILE_TYPE(cft
->private);
5150 name
= MEMFILE_ATTR(cft
->private);
5152 if (!do_swap_account
&& type
== _MEMSWAP
)
5157 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5161 /* This function does all necessary parse...reuse it */
5162 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5166 ret
= mem_cgroup_resize_limit(memcg
, val
);
5167 else if (type
== _MEMSWAP
)
5168 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5169 else if (type
== _KMEM
)
5170 ret
= memcg_update_kmem_limit(cont
, val
);
5174 case RES_SOFT_LIMIT
:
5175 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5179 * For memsw, soft limits are hard to implement in terms
5180 * of semantics, for now, we support soft limits for
5181 * control without swap
5184 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5189 ret
= -EINVAL
; /* should be BUG() ? */
5195 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5196 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5198 struct cgroup
*cgroup
;
5199 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5201 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5202 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5203 cgroup
= memcg
->css
.cgroup
;
5204 if (!memcg
->use_hierarchy
)
5207 while (cgroup
->parent
) {
5208 cgroup
= cgroup
->parent
;
5209 memcg
= mem_cgroup_from_cont(cgroup
);
5210 if (!memcg
->use_hierarchy
)
5212 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5213 min_limit
= min(min_limit
, tmp
);
5214 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5215 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5218 *mem_limit
= min_limit
;
5219 *memsw_limit
= min_memsw_limit
;
5222 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
5224 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5228 type
= MEMFILE_TYPE(event
);
5229 name
= MEMFILE_ATTR(event
);
5231 if (!do_swap_account
&& type
== _MEMSWAP
)
5237 res_counter_reset_max(&memcg
->res
);
5238 else if (type
== _MEMSWAP
)
5239 res_counter_reset_max(&memcg
->memsw
);
5240 else if (type
== _KMEM
)
5241 res_counter_reset_max(&memcg
->kmem
);
5247 res_counter_reset_failcnt(&memcg
->res
);
5248 else if (type
== _MEMSWAP
)
5249 res_counter_reset_failcnt(&memcg
->memsw
);
5250 else if (type
== _KMEM
)
5251 res_counter_reset_failcnt(&memcg
->kmem
);
5260 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
5263 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
5267 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5268 struct cftype
*cft
, u64 val
)
5270 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5272 if (val
>= (1 << NR_MOVE_TYPE
))
5276 * No kind of locking is needed in here, because ->can_attach() will
5277 * check this value once in the beginning of the process, and then carry
5278 * on with stale data. This means that changes to this value will only
5279 * affect task migrations starting after the change.
5281 memcg
->move_charge_at_immigrate
= val
;
5285 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5286 struct cftype
*cft
, u64 val
)
5293 static int memcg_numa_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5297 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5298 unsigned long node_nr
;
5299 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5301 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5302 seq_printf(m
, "total=%lu", total_nr
);
5303 for_each_node_state(nid
, N_MEMORY
) {
5304 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5305 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5309 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5310 seq_printf(m
, "file=%lu", file_nr
);
5311 for_each_node_state(nid
, N_MEMORY
) {
5312 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5314 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5318 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5319 seq_printf(m
, "anon=%lu", anon_nr
);
5320 for_each_node_state(nid
, N_MEMORY
) {
5321 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5323 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5327 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5328 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5329 for_each_node_state(nid
, N_MEMORY
) {
5330 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5331 BIT(LRU_UNEVICTABLE
));
5332 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5337 #endif /* CONFIG_NUMA */
5339 static inline void mem_cgroup_lru_names_not_uptodate(void)
5341 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5344 static int memcg_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5347 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5348 struct mem_cgroup
*mi
;
5351 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5352 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5354 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5355 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5358 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5359 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5360 mem_cgroup_read_events(memcg
, i
));
5362 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5363 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5364 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5366 /* Hierarchical information */
5368 unsigned long long limit
, memsw_limit
;
5369 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5370 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5371 if (do_swap_account
)
5372 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5376 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5379 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5381 for_each_mem_cgroup_tree(mi
, memcg
)
5382 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5383 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5386 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5387 unsigned long long val
= 0;
5389 for_each_mem_cgroup_tree(mi
, memcg
)
5390 val
+= mem_cgroup_read_events(mi
, i
);
5391 seq_printf(m
, "total_%s %llu\n",
5392 mem_cgroup_events_names
[i
], val
);
5395 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5396 unsigned long long val
= 0;
5398 for_each_mem_cgroup_tree(mi
, memcg
)
5399 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5400 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5403 #ifdef CONFIG_DEBUG_VM
5406 struct mem_cgroup_per_zone
*mz
;
5407 struct zone_reclaim_stat
*rstat
;
5408 unsigned long recent_rotated
[2] = {0, 0};
5409 unsigned long recent_scanned
[2] = {0, 0};
5411 for_each_online_node(nid
)
5412 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5413 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5414 rstat
= &mz
->lruvec
.reclaim_stat
;
5416 recent_rotated
[0] += rstat
->recent_rotated
[0];
5417 recent_rotated
[1] += rstat
->recent_rotated
[1];
5418 recent_scanned
[0] += rstat
->recent_scanned
[0];
5419 recent_scanned
[1] += rstat
->recent_scanned
[1];
5421 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5422 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5423 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5424 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5431 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
5433 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5435 return mem_cgroup_swappiness(memcg
);
5438 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
5441 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5442 struct mem_cgroup
*parent
;
5447 if (cgrp
->parent
== NULL
)
5450 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5452 mutex_lock(&memcg_create_mutex
);
5454 /* If under hierarchy, only empty-root can set this value */
5455 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5456 mutex_unlock(&memcg_create_mutex
);
5460 memcg
->swappiness
= val
;
5462 mutex_unlock(&memcg_create_mutex
);
5467 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5469 struct mem_cgroup_threshold_ary
*t
;
5475 t
= rcu_dereference(memcg
->thresholds
.primary
);
5477 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5482 usage
= mem_cgroup_usage(memcg
, swap
);
5485 * current_threshold points to threshold just below or equal to usage.
5486 * If it's not true, a threshold was crossed after last
5487 * call of __mem_cgroup_threshold().
5489 i
= t
->current_threshold
;
5492 * Iterate backward over array of thresholds starting from
5493 * current_threshold and check if a threshold is crossed.
5494 * If none of thresholds below usage is crossed, we read
5495 * only one element of the array here.
5497 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5498 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5500 /* i = current_threshold + 1 */
5504 * Iterate forward over array of thresholds starting from
5505 * current_threshold+1 and check if a threshold is crossed.
5506 * If none of thresholds above usage is crossed, we read
5507 * only one element of the array here.
5509 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5510 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5512 /* Update current_threshold */
5513 t
->current_threshold
= i
- 1;
5518 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5521 __mem_cgroup_threshold(memcg
, false);
5522 if (do_swap_account
)
5523 __mem_cgroup_threshold(memcg
, true);
5525 memcg
= parent_mem_cgroup(memcg
);
5529 static int compare_thresholds(const void *a
, const void *b
)
5531 const struct mem_cgroup_threshold
*_a
= a
;
5532 const struct mem_cgroup_threshold
*_b
= b
;
5534 return _a
->threshold
- _b
->threshold
;
5537 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5539 struct mem_cgroup_eventfd_list
*ev
;
5541 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5542 eventfd_signal(ev
->eventfd
, 1);
5546 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5548 struct mem_cgroup
*iter
;
5550 for_each_mem_cgroup_tree(iter
, memcg
)
5551 mem_cgroup_oom_notify_cb(iter
);
5554 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5555 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5557 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5558 struct mem_cgroup_thresholds
*thresholds
;
5559 struct mem_cgroup_threshold_ary
*new;
5560 enum res_type type
= MEMFILE_TYPE(cft
->private);
5561 u64 threshold
, usage
;
5564 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5568 mutex_lock(&memcg
->thresholds_lock
);
5571 thresholds
= &memcg
->thresholds
;
5572 else if (type
== _MEMSWAP
)
5573 thresholds
= &memcg
->memsw_thresholds
;
5577 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5579 /* Check if a threshold crossed before adding a new one */
5580 if (thresholds
->primary
)
5581 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5583 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5585 /* Allocate memory for new array of thresholds */
5586 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5594 /* Copy thresholds (if any) to new array */
5595 if (thresholds
->primary
) {
5596 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5597 sizeof(struct mem_cgroup_threshold
));
5600 /* Add new threshold */
5601 new->entries
[size
- 1].eventfd
= eventfd
;
5602 new->entries
[size
- 1].threshold
= threshold
;
5604 /* Sort thresholds. Registering of new threshold isn't time-critical */
5605 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5606 compare_thresholds
, NULL
);
5608 /* Find current threshold */
5609 new->current_threshold
= -1;
5610 for (i
= 0; i
< size
; i
++) {
5611 if (new->entries
[i
].threshold
<= usage
) {
5613 * new->current_threshold will not be used until
5614 * rcu_assign_pointer(), so it's safe to increment
5617 ++new->current_threshold
;
5622 /* Free old spare buffer and save old primary buffer as spare */
5623 kfree(thresholds
->spare
);
5624 thresholds
->spare
= thresholds
->primary
;
5626 rcu_assign_pointer(thresholds
->primary
, new);
5628 /* To be sure that nobody uses thresholds */
5632 mutex_unlock(&memcg
->thresholds_lock
);
5637 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5638 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5640 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5641 struct mem_cgroup_thresholds
*thresholds
;
5642 struct mem_cgroup_threshold_ary
*new;
5643 enum res_type type
= MEMFILE_TYPE(cft
->private);
5647 mutex_lock(&memcg
->thresholds_lock
);
5649 thresholds
= &memcg
->thresholds
;
5650 else if (type
== _MEMSWAP
)
5651 thresholds
= &memcg
->memsw_thresholds
;
5655 if (!thresholds
->primary
)
5658 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5660 /* Check if a threshold crossed before removing */
5661 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5663 /* Calculate new number of threshold */
5665 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5666 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5670 new = thresholds
->spare
;
5672 /* Set thresholds array to NULL if we don't have thresholds */
5681 /* Copy thresholds and find current threshold */
5682 new->current_threshold
= -1;
5683 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5684 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5687 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5688 if (new->entries
[j
].threshold
<= usage
) {
5690 * new->current_threshold will not be used
5691 * until rcu_assign_pointer(), so it's safe to increment
5694 ++new->current_threshold
;
5700 /* Swap primary and spare array */
5701 thresholds
->spare
= thresholds
->primary
;
5702 /* If all events are unregistered, free the spare array */
5704 kfree(thresholds
->spare
);
5705 thresholds
->spare
= NULL
;
5708 rcu_assign_pointer(thresholds
->primary
, new);
5710 /* To be sure that nobody uses thresholds */
5713 mutex_unlock(&memcg
->thresholds_lock
);
5716 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5717 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5719 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5720 struct mem_cgroup_eventfd_list
*event
;
5721 enum res_type type
= MEMFILE_TYPE(cft
->private);
5723 BUG_ON(type
!= _OOM_TYPE
);
5724 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5728 spin_lock(&memcg_oom_lock
);
5730 event
->eventfd
= eventfd
;
5731 list_add(&event
->list
, &memcg
->oom_notify
);
5733 /* already in OOM ? */
5734 if (atomic_read(&memcg
->under_oom
))
5735 eventfd_signal(eventfd
, 1);
5736 spin_unlock(&memcg_oom_lock
);
5741 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5742 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5744 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5745 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5746 enum res_type type
= MEMFILE_TYPE(cft
->private);
5748 BUG_ON(type
!= _OOM_TYPE
);
5750 spin_lock(&memcg_oom_lock
);
5752 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5753 if (ev
->eventfd
== eventfd
) {
5754 list_del(&ev
->list
);
5759 spin_unlock(&memcg_oom_lock
);
5762 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
5763 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5765 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5767 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5769 if (atomic_read(&memcg
->under_oom
))
5770 cb
->fill(cb
, "under_oom", 1);
5772 cb
->fill(cb
, "under_oom", 0);
5776 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
5777 struct cftype
*cft
, u64 val
)
5779 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5780 struct mem_cgroup
*parent
;
5782 /* cannot set to root cgroup and only 0 and 1 are allowed */
5783 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
5786 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5788 mutex_lock(&memcg_create_mutex
);
5789 /* oom-kill-disable is a flag for subhierarchy. */
5790 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5791 mutex_unlock(&memcg_create_mutex
);
5794 memcg
->oom_kill_disable
= val
;
5796 memcg_oom_recover(memcg
);
5797 mutex_unlock(&memcg_create_mutex
);
5801 #ifdef CONFIG_MEMCG_KMEM
5802 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5806 memcg
->kmemcg_id
= -1;
5807 ret
= memcg_propagate_kmem(memcg
);
5811 return mem_cgroup_sockets_init(memcg
, ss
);
5814 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5816 mem_cgroup_sockets_destroy(memcg
);
5818 memcg_kmem_mark_dead(memcg
);
5820 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5824 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5825 * path here, being careful not to race with memcg_uncharge_kmem: it is
5826 * possible that the charges went down to 0 between mark_dead and the
5827 * res_counter read, so in that case, we don't need the put
5829 if (memcg_kmem_test_and_clear_dead(memcg
))
5830 mem_cgroup_put(memcg
);
5833 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5838 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5843 static struct cftype mem_cgroup_files
[] = {
5845 .name
= "usage_in_bytes",
5846 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5847 .read
= mem_cgroup_read
,
5848 .register_event
= mem_cgroup_usage_register_event
,
5849 .unregister_event
= mem_cgroup_usage_unregister_event
,
5852 .name
= "max_usage_in_bytes",
5853 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5854 .trigger
= mem_cgroup_reset
,
5855 .read
= mem_cgroup_read
,
5858 .name
= "limit_in_bytes",
5859 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5860 .write_string
= mem_cgroup_write
,
5861 .read
= mem_cgroup_read
,
5864 .name
= "soft_limit_in_bytes",
5865 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5866 .write_string
= mem_cgroup_write
,
5867 .read
= mem_cgroup_read
,
5871 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5872 .trigger
= mem_cgroup_reset
,
5873 .read
= mem_cgroup_read
,
5877 .read_seq_string
= memcg_stat_show
,
5880 .name
= "force_empty",
5881 .trigger
= mem_cgroup_force_empty_write
,
5884 .name
= "use_hierarchy",
5885 .write_u64
= mem_cgroup_hierarchy_write
,
5886 .read_u64
= mem_cgroup_hierarchy_read
,
5889 .name
= "swappiness",
5890 .read_u64
= mem_cgroup_swappiness_read
,
5891 .write_u64
= mem_cgroup_swappiness_write
,
5894 .name
= "move_charge_at_immigrate",
5895 .read_u64
= mem_cgroup_move_charge_read
,
5896 .write_u64
= mem_cgroup_move_charge_write
,
5899 .name
= "oom_control",
5900 .read_map
= mem_cgroup_oom_control_read
,
5901 .write_u64
= mem_cgroup_oom_control_write
,
5902 .register_event
= mem_cgroup_oom_register_event
,
5903 .unregister_event
= mem_cgroup_oom_unregister_event
,
5904 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5908 .name
= "numa_stat",
5909 .read_seq_string
= memcg_numa_stat_show
,
5912 #ifdef CONFIG_MEMCG_KMEM
5914 .name
= "kmem.limit_in_bytes",
5915 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5916 .write_string
= mem_cgroup_write
,
5917 .read
= mem_cgroup_read
,
5920 .name
= "kmem.usage_in_bytes",
5921 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5922 .read
= mem_cgroup_read
,
5925 .name
= "kmem.failcnt",
5926 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5927 .trigger
= mem_cgroup_reset
,
5928 .read
= mem_cgroup_read
,
5931 .name
= "kmem.max_usage_in_bytes",
5932 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5933 .trigger
= mem_cgroup_reset
,
5934 .read
= mem_cgroup_read
,
5936 #ifdef CONFIG_SLABINFO
5938 .name
= "kmem.slabinfo",
5939 .read_seq_string
= mem_cgroup_slabinfo_read
,
5943 { }, /* terminate */
5946 #ifdef CONFIG_MEMCG_SWAP
5947 static struct cftype memsw_cgroup_files
[] = {
5949 .name
= "memsw.usage_in_bytes",
5950 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5951 .read
= mem_cgroup_read
,
5952 .register_event
= mem_cgroup_usage_register_event
,
5953 .unregister_event
= mem_cgroup_usage_unregister_event
,
5956 .name
= "memsw.max_usage_in_bytes",
5957 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5958 .trigger
= mem_cgroup_reset
,
5959 .read
= mem_cgroup_read
,
5962 .name
= "memsw.limit_in_bytes",
5963 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5964 .write_string
= mem_cgroup_write
,
5965 .read
= mem_cgroup_read
,
5968 .name
= "memsw.failcnt",
5969 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5970 .trigger
= mem_cgroup_reset
,
5971 .read
= mem_cgroup_read
,
5973 { }, /* terminate */
5976 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5978 struct mem_cgroup_per_node
*pn
;
5979 struct mem_cgroup_per_zone
*mz
;
5980 int zone
, tmp
= node
;
5982 * This routine is called against possible nodes.
5983 * But it's BUG to call kmalloc() against offline node.
5985 * TODO: this routine can waste much memory for nodes which will
5986 * never be onlined. It's better to use memory hotplug callback
5989 if (!node_state(node
, N_NORMAL_MEMORY
))
5991 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5995 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5996 mz
= &pn
->zoneinfo
[zone
];
5997 lruvec_init(&mz
->lruvec
);
5998 mz
->usage_in_excess
= 0;
5999 mz
->on_tree
= false;
6002 memcg
->info
.nodeinfo
[node
] = pn
;
6006 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6008 kfree(memcg
->info
.nodeinfo
[node
]);
6011 static struct mem_cgroup
*mem_cgroup_alloc(void)
6013 struct mem_cgroup
*memcg
;
6014 size_t size
= memcg_size();
6016 /* Can be very big if nr_node_ids is very big */
6017 if (size
< PAGE_SIZE
)
6018 memcg
= kzalloc(size
, GFP_KERNEL
);
6020 memcg
= vzalloc(size
);
6025 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6028 spin_lock_init(&memcg
->pcp_counter_lock
);
6032 if (size
< PAGE_SIZE
)
6040 * At destroying mem_cgroup, references from swap_cgroup can remain.
6041 * (scanning all at force_empty is too costly...)
6043 * Instead of clearing all references at force_empty, we remember
6044 * the number of reference from swap_cgroup and free mem_cgroup when
6045 * it goes down to 0.
6047 * Removal of cgroup itself succeeds regardless of refs from swap.
6050 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6053 size_t size
= memcg_size();
6055 mem_cgroup_remove_from_trees(memcg
);
6056 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
6059 free_mem_cgroup_per_zone_info(memcg
, node
);
6061 free_percpu(memcg
->stat
);
6064 * We need to make sure that (at least for now), the jump label
6065 * destruction code runs outside of the cgroup lock. This is because
6066 * get_online_cpus(), which is called from the static_branch update,
6067 * can't be called inside the cgroup_lock. cpusets are the ones
6068 * enforcing this dependency, so if they ever change, we might as well.
6070 * schedule_work() will guarantee this happens. Be careful if you need
6071 * to move this code around, and make sure it is outside
6074 disarm_static_keys(memcg
);
6075 if (size
< PAGE_SIZE
)
6083 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6084 * but in process context. The work_freeing structure is overlaid
6085 * on the rcu_freeing structure, which itself is overlaid on memsw.
6087 static void free_work(struct work_struct
*work
)
6089 struct mem_cgroup
*memcg
;
6091 memcg
= container_of(work
, struct mem_cgroup
, work_freeing
);
6092 __mem_cgroup_free(memcg
);
6095 static void free_rcu(struct rcu_head
*rcu_head
)
6097 struct mem_cgroup
*memcg
;
6099 memcg
= container_of(rcu_head
, struct mem_cgroup
, rcu_freeing
);
6100 INIT_WORK(&memcg
->work_freeing
, free_work
);
6101 schedule_work(&memcg
->work_freeing
);
6104 static void mem_cgroup_get(struct mem_cgroup
*memcg
)
6106 atomic_inc(&memcg
->refcnt
);
6109 static void __mem_cgroup_put(struct mem_cgroup
*memcg
, int count
)
6111 if (atomic_sub_and_test(count
, &memcg
->refcnt
)) {
6112 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
6113 call_rcu(&memcg
->rcu_freeing
, free_rcu
);
6115 mem_cgroup_put(parent
);
6119 static void mem_cgroup_put(struct mem_cgroup
*memcg
)
6121 __mem_cgroup_put(memcg
, 1);
6125 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6127 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6129 if (!memcg
->res
.parent
)
6131 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6133 EXPORT_SYMBOL(parent_mem_cgroup
);
6135 static void __init
mem_cgroup_soft_limit_tree_init(void)
6137 struct mem_cgroup_tree_per_node
*rtpn
;
6138 struct mem_cgroup_tree_per_zone
*rtpz
;
6139 int tmp
, node
, zone
;
6141 for_each_node(node
) {
6143 if (!node_state(node
, N_NORMAL_MEMORY
))
6145 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6148 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6150 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6151 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6152 rtpz
->rb_root
= RB_ROOT
;
6153 spin_lock_init(&rtpz
->lock
);
6158 static struct cgroup_subsys_state
* __ref
6159 mem_cgroup_css_alloc(struct cgroup
*cont
)
6161 struct mem_cgroup
*memcg
;
6162 long error
= -ENOMEM
;
6165 memcg
= mem_cgroup_alloc();
6167 return ERR_PTR(error
);
6170 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6174 if (cont
->parent
== NULL
) {
6175 root_mem_cgroup
= memcg
;
6176 res_counter_init(&memcg
->res
, NULL
);
6177 res_counter_init(&memcg
->memsw
, NULL
);
6178 res_counter_init(&memcg
->kmem
, NULL
);
6181 memcg
->last_scanned_node
= MAX_NUMNODES
;
6182 INIT_LIST_HEAD(&memcg
->oom_notify
);
6183 atomic_set(&memcg
->refcnt
, 1);
6184 memcg
->move_charge_at_immigrate
= 0;
6185 mutex_init(&memcg
->thresholds_lock
);
6186 spin_lock_init(&memcg
->move_lock
);
6191 __mem_cgroup_free(memcg
);
6192 return ERR_PTR(error
);
6196 mem_cgroup_css_online(struct cgroup
*cont
)
6198 struct mem_cgroup
*memcg
, *parent
;
6204 mutex_lock(&memcg_create_mutex
);
6205 memcg
= mem_cgroup_from_cont(cont
);
6206 parent
= mem_cgroup_from_cont(cont
->parent
);
6208 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6209 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6210 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6212 if (parent
->use_hierarchy
) {
6213 res_counter_init(&memcg
->res
, &parent
->res
);
6214 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6215 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6218 * We increment refcnt of the parent to ensure that we can
6219 * safely access it on res_counter_charge/uncharge.
6220 * This refcnt will be decremented when freeing this
6221 * mem_cgroup(see mem_cgroup_put).
6223 mem_cgroup_get(parent
);
6225 res_counter_init(&memcg
->res
, NULL
);
6226 res_counter_init(&memcg
->memsw
, NULL
);
6227 res_counter_init(&memcg
->kmem
, NULL
);
6229 * Deeper hierachy with use_hierarchy == false doesn't make
6230 * much sense so let cgroup subsystem know about this
6231 * unfortunate state in our controller.
6233 if (parent
!= root_mem_cgroup
)
6234 mem_cgroup_subsys
.broken_hierarchy
= true;
6237 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6238 mutex_unlock(&memcg_create_mutex
);
6241 * We call put now because our (and parent's) refcnts
6242 * are already in place. mem_cgroup_put() will internally
6243 * call __mem_cgroup_free, so return directly
6245 mem_cgroup_put(memcg
);
6246 if (parent
->use_hierarchy
)
6247 mem_cgroup_put(parent
);
6253 * Announce all parents that a group from their hierarchy is gone.
6255 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6257 struct mem_cgroup
*parent
= memcg
;
6259 while ((parent
= parent_mem_cgroup(parent
)))
6260 atomic_inc(&parent
->dead_count
);
6263 * if the root memcg is not hierarchical we have to check it
6266 if (!root_mem_cgroup
->use_hierarchy
)
6267 atomic_inc(&root_mem_cgroup
->dead_count
);
6270 static void mem_cgroup_css_offline(struct cgroup
*cont
)
6272 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6274 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6275 mem_cgroup_reparent_charges(memcg
);
6276 mem_cgroup_destroy_all_caches(memcg
);
6279 static void mem_cgroup_css_free(struct cgroup
*cont
)
6281 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6283 kmem_cgroup_destroy(memcg
);
6285 mem_cgroup_put(memcg
);
6289 /* Handlers for move charge at task migration. */
6290 #define PRECHARGE_COUNT_AT_ONCE 256
6291 static int mem_cgroup_do_precharge(unsigned long count
)
6294 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6295 struct mem_cgroup
*memcg
= mc
.to
;
6297 if (mem_cgroup_is_root(memcg
)) {
6298 mc
.precharge
+= count
;
6299 /* we don't need css_get for root */
6302 /* try to charge at once */
6304 struct res_counter
*dummy
;
6306 * "memcg" cannot be under rmdir() because we've already checked
6307 * by cgroup_lock_live_cgroup() that it is not removed and we
6308 * are still under the same cgroup_mutex. So we can postpone
6311 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6313 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6314 PAGE_SIZE
* count
, &dummy
)) {
6315 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6318 mc
.precharge
+= count
;
6322 /* fall back to one by one charge */
6324 if (signal_pending(current
)) {
6328 if (!batch_count
--) {
6329 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6332 ret
= __mem_cgroup_try_charge(NULL
,
6333 GFP_KERNEL
, 1, &memcg
, false);
6335 /* mem_cgroup_clear_mc() will do uncharge later */
6343 * get_mctgt_type - get target type of moving charge
6344 * @vma: the vma the pte to be checked belongs
6345 * @addr: the address corresponding to the pte to be checked
6346 * @ptent: the pte to be checked
6347 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6350 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6351 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6352 * move charge. if @target is not NULL, the page is stored in target->page
6353 * with extra refcnt got(Callers should handle it).
6354 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6355 * target for charge migration. if @target is not NULL, the entry is stored
6358 * Called with pte lock held.
6365 enum mc_target_type
{
6371 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6372 unsigned long addr
, pte_t ptent
)
6374 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6376 if (!page
|| !page_mapped(page
))
6378 if (PageAnon(page
)) {
6379 /* we don't move shared anon */
6382 } else if (!move_file())
6383 /* we ignore mapcount for file pages */
6385 if (!get_page_unless_zero(page
))
6392 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6393 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6395 struct page
*page
= NULL
;
6396 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6398 if (!move_anon() || non_swap_entry(ent
))
6401 * Because lookup_swap_cache() updates some statistics counter,
6402 * we call find_get_page() with swapper_space directly.
6404 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6405 if (do_swap_account
)
6406 entry
->val
= ent
.val
;
6411 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6412 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6418 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6419 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6421 struct page
*page
= NULL
;
6422 struct address_space
*mapping
;
6425 if (!vma
->vm_file
) /* anonymous vma */
6430 mapping
= vma
->vm_file
->f_mapping
;
6431 if (pte_none(ptent
))
6432 pgoff
= linear_page_index(vma
, addr
);
6433 else /* pte_file(ptent) is true */
6434 pgoff
= pte_to_pgoff(ptent
);
6436 /* page is moved even if it's not RSS of this task(page-faulted). */
6437 page
= find_get_page(mapping
, pgoff
);
6440 /* shmem/tmpfs may report page out on swap: account for that too. */
6441 if (radix_tree_exceptional_entry(page
)) {
6442 swp_entry_t swap
= radix_to_swp_entry(page
);
6443 if (do_swap_account
)
6445 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6451 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6452 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6454 struct page
*page
= NULL
;
6455 struct page_cgroup
*pc
;
6456 enum mc_target_type ret
= MC_TARGET_NONE
;
6457 swp_entry_t ent
= { .val
= 0 };
6459 if (pte_present(ptent
))
6460 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6461 else if (is_swap_pte(ptent
))
6462 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6463 else if (pte_none(ptent
) || pte_file(ptent
))
6464 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6466 if (!page
&& !ent
.val
)
6469 pc
= lookup_page_cgroup(page
);
6471 * Do only loose check w/o page_cgroup lock.
6472 * mem_cgroup_move_account() checks the pc is valid or not under
6475 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6476 ret
= MC_TARGET_PAGE
;
6478 target
->page
= page
;
6480 if (!ret
|| !target
)
6483 /* There is a swap entry and a page doesn't exist or isn't charged */
6484 if (ent
.val
&& !ret
&&
6485 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6486 ret
= MC_TARGET_SWAP
;
6493 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6495 * We don't consider swapping or file mapped pages because THP does not
6496 * support them for now.
6497 * Caller should make sure that pmd_trans_huge(pmd) is true.
6499 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6500 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6502 struct page
*page
= NULL
;
6503 struct page_cgroup
*pc
;
6504 enum mc_target_type ret
= MC_TARGET_NONE
;
6506 page
= pmd_page(pmd
);
6507 VM_BUG_ON(!page
|| !PageHead(page
));
6510 pc
= lookup_page_cgroup(page
);
6511 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6512 ret
= MC_TARGET_PAGE
;
6515 target
->page
= page
;
6521 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6522 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6524 return MC_TARGET_NONE
;
6528 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6529 unsigned long addr
, unsigned long end
,
6530 struct mm_walk
*walk
)
6532 struct vm_area_struct
*vma
= walk
->private;
6536 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6537 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6538 mc
.precharge
+= HPAGE_PMD_NR
;
6539 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6543 if (pmd_trans_unstable(pmd
))
6545 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6546 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6547 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6548 mc
.precharge
++; /* increment precharge temporarily */
6549 pte_unmap_unlock(pte
- 1, ptl
);
6555 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6557 unsigned long precharge
;
6558 struct vm_area_struct
*vma
;
6560 down_read(&mm
->mmap_sem
);
6561 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6562 struct mm_walk mem_cgroup_count_precharge_walk
= {
6563 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6567 if (is_vm_hugetlb_page(vma
))
6569 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6570 &mem_cgroup_count_precharge_walk
);
6572 up_read(&mm
->mmap_sem
);
6574 precharge
= mc
.precharge
;
6580 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6582 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6584 VM_BUG_ON(mc
.moving_task
);
6585 mc
.moving_task
= current
;
6586 return mem_cgroup_do_precharge(precharge
);
6589 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6590 static void __mem_cgroup_clear_mc(void)
6592 struct mem_cgroup
*from
= mc
.from
;
6593 struct mem_cgroup
*to
= mc
.to
;
6595 /* we must uncharge all the leftover precharges from mc.to */
6597 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6601 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6602 * we must uncharge here.
6604 if (mc
.moved_charge
) {
6605 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6606 mc
.moved_charge
= 0;
6608 /* we must fixup refcnts and charges */
6609 if (mc
.moved_swap
) {
6610 /* uncharge swap account from the old cgroup */
6611 if (!mem_cgroup_is_root(mc
.from
))
6612 res_counter_uncharge(&mc
.from
->memsw
,
6613 PAGE_SIZE
* mc
.moved_swap
);
6614 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
6616 if (!mem_cgroup_is_root(mc
.to
)) {
6618 * we charged both to->res and to->memsw, so we should
6621 res_counter_uncharge(&mc
.to
->res
,
6622 PAGE_SIZE
* mc
.moved_swap
);
6624 /* we've already done mem_cgroup_get(mc.to) */
6627 memcg_oom_recover(from
);
6628 memcg_oom_recover(to
);
6629 wake_up_all(&mc
.waitq
);
6632 static void mem_cgroup_clear_mc(void)
6634 struct mem_cgroup
*from
= mc
.from
;
6637 * we must clear moving_task before waking up waiters at the end of
6640 mc
.moving_task
= NULL
;
6641 __mem_cgroup_clear_mc();
6642 spin_lock(&mc
.lock
);
6645 spin_unlock(&mc
.lock
);
6646 mem_cgroup_end_move(from
);
6649 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6650 struct cgroup_taskset
*tset
)
6652 struct task_struct
*p
= cgroup_taskset_first(tset
);
6654 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgroup
);
6655 unsigned long move_charge_at_immigrate
;
6658 * We are now commited to this value whatever it is. Changes in this
6659 * tunable will only affect upcoming migrations, not the current one.
6660 * So we need to save it, and keep it going.
6662 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6663 if (move_charge_at_immigrate
) {
6664 struct mm_struct
*mm
;
6665 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6667 VM_BUG_ON(from
== memcg
);
6669 mm
= get_task_mm(p
);
6672 /* We move charges only when we move a owner of the mm */
6673 if (mm
->owner
== p
) {
6676 VM_BUG_ON(mc
.precharge
);
6677 VM_BUG_ON(mc
.moved_charge
);
6678 VM_BUG_ON(mc
.moved_swap
);
6679 mem_cgroup_start_move(from
);
6680 spin_lock(&mc
.lock
);
6683 mc
.immigrate_flags
= move_charge_at_immigrate
;
6684 spin_unlock(&mc
.lock
);
6685 /* We set mc.moving_task later */
6687 ret
= mem_cgroup_precharge_mc(mm
);
6689 mem_cgroup_clear_mc();
6696 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6697 struct cgroup_taskset
*tset
)
6699 mem_cgroup_clear_mc();
6702 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6703 unsigned long addr
, unsigned long end
,
6704 struct mm_walk
*walk
)
6707 struct vm_area_struct
*vma
= walk
->private;
6710 enum mc_target_type target_type
;
6711 union mc_target target
;
6713 struct page_cgroup
*pc
;
6716 * We don't take compound_lock() here but no race with splitting thp
6718 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6719 * under splitting, which means there's no concurrent thp split,
6720 * - if another thread runs into split_huge_page() just after we
6721 * entered this if-block, the thread must wait for page table lock
6722 * to be unlocked in __split_huge_page_splitting(), where the main
6723 * part of thp split is not executed yet.
6725 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6726 if (mc
.precharge
< HPAGE_PMD_NR
) {
6727 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6730 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6731 if (target_type
== MC_TARGET_PAGE
) {
6733 if (!isolate_lru_page(page
)) {
6734 pc
= lookup_page_cgroup(page
);
6735 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6736 pc
, mc
.from
, mc
.to
)) {
6737 mc
.precharge
-= HPAGE_PMD_NR
;
6738 mc
.moved_charge
+= HPAGE_PMD_NR
;
6740 putback_lru_page(page
);
6744 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6748 if (pmd_trans_unstable(pmd
))
6751 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6752 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6753 pte_t ptent
= *(pte
++);
6759 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6760 case MC_TARGET_PAGE
:
6762 if (isolate_lru_page(page
))
6764 pc
= lookup_page_cgroup(page
);
6765 if (!mem_cgroup_move_account(page
, 1, pc
,
6768 /* we uncharge from mc.from later. */
6771 putback_lru_page(page
);
6772 put
: /* get_mctgt_type() gets the page */
6775 case MC_TARGET_SWAP
:
6777 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6779 /* we fixup refcnts and charges later. */
6787 pte_unmap_unlock(pte
- 1, ptl
);
6792 * We have consumed all precharges we got in can_attach().
6793 * We try charge one by one, but don't do any additional
6794 * charges to mc.to if we have failed in charge once in attach()
6797 ret
= mem_cgroup_do_precharge(1);
6805 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6807 struct vm_area_struct
*vma
;
6809 lru_add_drain_all();
6811 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6813 * Someone who are holding the mmap_sem might be waiting in
6814 * waitq. So we cancel all extra charges, wake up all waiters,
6815 * and retry. Because we cancel precharges, we might not be able
6816 * to move enough charges, but moving charge is a best-effort
6817 * feature anyway, so it wouldn't be a big problem.
6819 __mem_cgroup_clear_mc();
6823 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6825 struct mm_walk mem_cgroup_move_charge_walk
= {
6826 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6830 if (is_vm_hugetlb_page(vma
))
6832 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6833 &mem_cgroup_move_charge_walk
);
6836 * means we have consumed all precharges and failed in
6837 * doing additional charge. Just abandon here.
6841 up_read(&mm
->mmap_sem
);
6844 static void mem_cgroup_move_task(struct cgroup
*cont
,
6845 struct cgroup_taskset
*tset
)
6847 struct task_struct
*p
= cgroup_taskset_first(tset
);
6848 struct mm_struct
*mm
= get_task_mm(p
);
6852 mem_cgroup_move_charge(mm
);
6856 mem_cgroup_clear_mc();
6858 #else /* !CONFIG_MMU */
6859 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6860 struct cgroup_taskset
*tset
)
6864 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6865 struct cgroup_taskset
*tset
)
6868 static void mem_cgroup_move_task(struct cgroup
*cont
,
6869 struct cgroup_taskset
*tset
)
6874 struct cgroup_subsys mem_cgroup_subsys
= {
6876 .subsys_id
= mem_cgroup_subsys_id
,
6877 .css_alloc
= mem_cgroup_css_alloc
,
6878 .css_online
= mem_cgroup_css_online
,
6879 .css_offline
= mem_cgroup_css_offline
,
6880 .css_free
= mem_cgroup_css_free
,
6881 .can_attach
= mem_cgroup_can_attach
,
6882 .cancel_attach
= mem_cgroup_cancel_attach
,
6883 .attach
= mem_cgroup_move_task
,
6884 .base_cftypes
= mem_cgroup_files
,
6889 #ifdef CONFIG_MEMCG_SWAP
6890 static int __init
enable_swap_account(char *s
)
6892 /* consider enabled if no parameter or 1 is given */
6893 if (!strcmp(s
, "1"))
6894 really_do_swap_account
= 1;
6895 else if (!strcmp(s
, "0"))
6896 really_do_swap_account
= 0;
6899 __setup("swapaccount=", enable_swap_account
);
6901 static void __init
memsw_file_init(void)
6903 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6906 static void __init
enable_swap_cgroup(void)
6908 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6909 do_swap_account
= 1;
6915 static void __init
enable_swap_cgroup(void)
6921 * subsys_initcall() for memory controller.
6923 * Some parts like hotcpu_notifier() have to be initialized from this context
6924 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6925 * everything that doesn't depend on a specific mem_cgroup structure should
6926 * be initialized from here.
6928 static int __init
mem_cgroup_init(void)
6930 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6931 enable_swap_cgroup();
6932 mem_cgroup_soft_limit_tree_init();
6936 subsys_initcall(mem_cgroup_init
);