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/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
67 EXPORT_SYMBOL(mem_cgroup_subsys
);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly
;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata
= 1;
80 static int really_do_swap_account __initdata
= 0;
84 #define do_swap_account 0
89 * Statistics for memory cgroup.
91 enum mem_cgroup_stat_index
{
93 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
95 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
96 MEM_CGROUP_STAT_RSS
, /* # of pages charged as anon rss */
97 MEM_CGROUP_STAT_RSS_HUGE
, /* # of pages charged as anon huge */
98 MEM_CGROUP_STAT_FILE_MAPPED
, /* # of pages charged as file rss */
99 MEM_CGROUP_STAT_SWAP
, /* # of pages, swapped out */
100 MEM_CGROUP_STAT_NSTATS
,
103 static const char * const mem_cgroup_stat_names
[] = {
111 enum mem_cgroup_events_index
{
112 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
113 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
114 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
115 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
116 MEM_CGROUP_EVENTS_NSTATS
,
119 static const char * const mem_cgroup_events_names
[] = {
126 static const char * const mem_cgroup_lru_names
[] = {
135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
136 * it will be incremated by the number of pages. This counter is used for
137 * for trigger some periodic events. This is straightforward and better
138 * than using jiffies etc. to handle periodic memcg event.
140 enum mem_cgroup_events_target
{
141 MEM_CGROUP_TARGET_THRESH
,
142 MEM_CGROUP_TARGET_SOFTLIMIT
,
143 MEM_CGROUP_TARGET_NUMAINFO
,
146 #define THRESHOLDS_EVENTS_TARGET 128
147 #define SOFTLIMIT_EVENTS_TARGET 1024
148 #define NUMAINFO_EVENTS_TARGET 1024
150 struct mem_cgroup_stat_cpu
{
151 long count
[MEM_CGROUP_STAT_NSTATS
];
152 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
153 unsigned long nr_page_events
;
154 unsigned long targets
[MEM_CGROUP_NTARGETS
];
157 struct mem_cgroup_reclaim_iter
{
159 * last scanned hierarchy member. Valid only if last_dead_count
160 * matches memcg->dead_count of the hierarchy root group.
162 struct mem_cgroup
*last_visited
;
163 unsigned long last_dead_count
;
165 /* scan generation, increased every round-trip */
166 unsigned int generation
;
170 * per-zone information in memory controller.
172 struct mem_cgroup_per_zone
{
173 struct lruvec lruvec
;
174 unsigned long lru_size
[NR_LRU_LISTS
];
176 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
178 struct rb_node tree_node
; /* RB tree node */
179 unsigned long long usage_in_excess
;/* Set to the value by which */
180 /* the soft limit is exceeded*/
182 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
183 /* use container_of */
186 struct mem_cgroup_per_node
{
187 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
191 * Cgroups above their limits are maintained in a RB-Tree, independent of
192 * their hierarchy representation
195 struct mem_cgroup_tree_per_zone
{
196 struct rb_root rb_root
;
200 struct mem_cgroup_tree_per_node
{
201 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
204 struct mem_cgroup_tree
{
205 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
208 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
210 struct mem_cgroup_threshold
{
211 struct eventfd_ctx
*eventfd
;
216 struct mem_cgroup_threshold_ary
{
217 /* An array index points to threshold just below or equal to usage. */
218 int current_threshold
;
219 /* Size of entries[] */
221 /* Array of thresholds */
222 struct mem_cgroup_threshold entries
[0];
225 struct mem_cgroup_thresholds
{
226 /* Primary thresholds array */
227 struct mem_cgroup_threshold_ary
*primary
;
229 * Spare threshold array.
230 * This is needed to make mem_cgroup_unregister_event() "never fail".
231 * It must be able to store at least primary->size - 1 entries.
233 struct mem_cgroup_threshold_ary
*spare
;
237 struct mem_cgroup_eventfd_list
{
238 struct list_head list
;
239 struct eventfd_ctx
*eventfd
;
242 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
243 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
246 * The memory controller data structure. The memory controller controls both
247 * page cache and RSS per cgroup. We would eventually like to provide
248 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
249 * to help the administrator determine what knobs to tune.
251 * TODO: Add a water mark for the memory controller. Reclaim will begin when
252 * we hit the water mark. May be even add a low water mark, such that
253 * no reclaim occurs from a cgroup at it's low water mark, this is
254 * a feature that will be implemented much later in the future.
257 struct cgroup_subsys_state css
;
259 * the counter to account for memory usage
261 struct res_counter res
;
263 /* vmpressure notifications */
264 struct vmpressure vmpressure
;
267 * the counter to account for mem+swap usage.
269 struct res_counter memsw
;
272 * the counter to account for kernel memory usage.
274 struct res_counter kmem
;
276 * Should the accounting and control be hierarchical, per subtree?
279 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
285 /* OOM-Killer disable */
286 int oom_kill_disable
;
288 /* set when res.limit == memsw.limit */
289 bool memsw_is_minimum
;
291 /* protect arrays of thresholds */
292 struct mutex thresholds_lock
;
294 /* thresholds for memory usage. RCU-protected */
295 struct mem_cgroup_thresholds thresholds
;
297 /* thresholds for mem+swap usage. RCU-protected */
298 struct mem_cgroup_thresholds memsw_thresholds
;
300 /* For oom notifier event fd */
301 struct list_head oom_notify
;
304 * Should we move charges of a task when a task is moved into this
305 * mem_cgroup ? And what type of charges should we move ?
307 unsigned long move_charge_at_immigrate
;
309 * set > 0 if pages under this cgroup are moving to other cgroup.
311 atomic_t moving_account
;
312 /* taken only while moving_account > 0 */
313 spinlock_t move_lock
;
317 struct mem_cgroup_stat_cpu __percpu
*stat
;
319 * used when a cpu is offlined or other synchronizations
320 * See mem_cgroup_read_stat().
322 struct mem_cgroup_stat_cpu nocpu_base
;
323 spinlock_t pcp_counter_lock
;
326 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
327 struct tcp_memcontrol tcp_mem
;
329 #if defined(CONFIG_MEMCG_KMEM)
330 /* analogous to slab_common's slab_caches list. per-memcg */
331 struct list_head memcg_slab_caches
;
332 /* Not a spinlock, we can take a lot of time walking the list */
333 struct mutex slab_caches_mutex
;
334 /* Index in the kmem_cache->memcg_params->memcg_caches array */
338 int last_scanned_node
;
340 nodemask_t scan_nodes
;
341 atomic_t numainfo_events
;
342 atomic_t numainfo_updating
;
345 struct mem_cgroup_per_node
*nodeinfo
[0];
346 /* WARNING: nodeinfo must be the last member here */
349 static size_t memcg_size(void)
351 return sizeof(struct mem_cgroup
) +
352 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
355 /* internal only representation about the status of kmem accounting. */
357 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
358 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
359 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
362 /* We account when limit is on, but only after call sites are patched */
363 #define KMEM_ACCOUNTED_MASK \
364 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
366 #ifdef CONFIG_MEMCG_KMEM
367 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
369 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
372 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
374 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
377 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
379 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
382 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
384 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
387 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
390 * Our caller must use css_get() first, because memcg_uncharge_kmem()
391 * will call css_put() if it sees the memcg is dead.
394 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
395 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
398 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
400 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
401 &memcg
->kmem_account_flags
);
405 /* Stuffs for move charges at task migration. */
407 * Types of charges to be moved. "move_charge_at_immitgrate" and
408 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
411 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
412 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
416 /* "mc" and its members are protected by cgroup_mutex */
417 static struct move_charge_struct
{
418 spinlock_t lock
; /* for from, to */
419 struct mem_cgroup
*from
;
420 struct mem_cgroup
*to
;
421 unsigned long immigrate_flags
;
422 unsigned long precharge
;
423 unsigned long moved_charge
;
424 unsigned long moved_swap
;
425 struct task_struct
*moving_task
; /* a task moving charges */
426 wait_queue_head_t waitq
; /* a waitq for other context */
428 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
429 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
432 static bool move_anon(void)
434 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
437 static bool move_file(void)
439 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
443 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
444 * limit reclaim to prevent infinite loops, if they ever occur.
446 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
447 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
450 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
451 MEM_CGROUP_CHARGE_TYPE_ANON
,
452 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
453 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
457 /* for encoding cft->private value on file */
465 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
466 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
467 #define MEMFILE_ATTR(val) ((val) & 0xffff)
468 /* Used for OOM nofiier */
469 #define OOM_CONTROL (0)
472 * Reclaim flags for mem_cgroup_hierarchical_reclaim
474 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
475 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
476 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
477 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
480 * The memcg_create_mutex will be held whenever a new cgroup is created.
481 * As a consequence, any change that needs to protect against new child cgroups
482 * appearing has to hold it as well.
484 static DEFINE_MUTEX(memcg_create_mutex
);
486 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
488 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
491 /* Some nice accessors for the vmpressure. */
492 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
495 memcg
= root_mem_cgroup
;
496 return &memcg
->vmpressure
;
499 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
501 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
504 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
506 return &mem_cgroup_from_css(css
)->vmpressure
;
509 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
511 return (memcg
== root_mem_cgroup
);
514 /* Writing them here to avoid exposing memcg's inner layout */
515 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
517 void sock_update_memcg(struct sock
*sk
)
519 if (mem_cgroup_sockets_enabled
) {
520 struct mem_cgroup
*memcg
;
521 struct cg_proto
*cg_proto
;
523 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
525 /* Socket cloning can throw us here with sk_cgrp already
526 * filled. It won't however, necessarily happen from
527 * process context. So the test for root memcg given
528 * the current task's memcg won't help us in this case.
530 * Respecting the original socket's memcg is a better
531 * decision in this case.
534 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
535 css_get(&sk
->sk_cgrp
->memcg
->css
);
540 memcg
= mem_cgroup_from_task(current
);
541 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
542 if (!mem_cgroup_is_root(memcg
) &&
543 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
544 sk
->sk_cgrp
= cg_proto
;
549 EXPORT_SYMBOL(sock_update_memcg
);
551 void sock_release_memcg(struct sock
*sk
)
553 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
554 struct mem_cgroup
*memcg
;
555 WARN_ON(!sk
->sk_cgrp
->memcg
);
556 memcg
= sk
->sk_cgrp
->memcg
;
557 css_put(&sk
->sk_cgrp
->memcg
->css
);
561 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
563 if (!memcg
|| mem_cgroup_is_root(memcg
))
566 return &memcg
->tcp_mem
.cg_proto
;
568 EXPORT_SYMBOL(tcp_proto_cgroup
);
570 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
572 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
574 static_key_slow_dec(&memcg_socket_limit_enabled
);
577 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
582 #ifdef CONFIG_MEMCG_KMEM
584 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
585 * There are two main reasons for not using the css_id for this:
586 * 1) this works better in sparse environments, where we have a lot of memcgs,
587 * but only a few kmem-limited. Or also, if we have, for instance, 200
588 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
589 * 200 entry array for that.
591 * 2) In order not to violate the cgroup API, we would like to do all memory
592 * allocation in ->create(). At that point, we haven't yet allocated the
593 * css_id. Having a separate index prevents us from messing with the cgroup
596 * The current size of the caches array is stored in
597 * memcg_limited_groups_array_size. It will double each time we have to
600 static DEFINE_IDA(kmem_limited_groups
);
601 int memcg_limited_groups_array_size
;
604 * MIN_SIZE is different than 1, because we would like to avoid going through
605 * the alloc/free process all the time. In a small machine, 4 kmem-limited
606 * cgroups is a reasonable guess. In the future, it could be a parameter or
607 * tunable, but that is strictly not necessary.
609 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
610 * this constant directly from cgroup, but it is understandable that this is
611 * better kept as an internal representation in cgroup.c. In any case, the
612 * css_id space is not getting any smaller, and we don't have to necessarily
613 * increase ours as well if it increases.
615 #define MEMCG_CACHES_MIN_SIZE 4
616 #define MEMCG_CACHES_MAX_SIZE 65535
619 * A lot of the calls to the cache allocation functions are expected to be
620 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
621 * conditional to this static branch, we'll have to allow modules that does
622 * kmem_cache_alloc and the such to see this symbol as well
624 struct static_key memcg_kmem_enabled_key
;
625 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
627 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
629 if (memcg_kmem_is_active(memcg
)) {
630 static_key_slow_dec(&memcg_kmem_enabled_key
);
631 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
634 * This check can't live in kmem destruction function,
635 * since the charges will outlive the cgroup
637 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
640 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
643 #endif /* CONFIG_MEMCG_KMEM */
645 static void disarm_static_keys(struct mem_cgroup
*memcg
)
647 disarm_sock_keys(memcg
);
648 disarm_kmem_keys(memcg
);
651 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
653 static struct mem_cgroup_per_zone
*
654 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
656 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
657 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
660 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
665 static struct mem_cgroup_per_zone
*
666 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
668 int nid
= page_to_nid(page
);
669 int zid
= page_zonenum(page
);
671 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
674 static struct mem_cgroup_tree_per_zone
*
675 soft_limit_tree_node_zone(int nid
, int zid
)
677 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
680 static struct mem_cgroup_tree_per_zone
*
681 soft_limit_tree_from_page(struct page
*page
)
683 int nid
= page_to_nid(page
);
684 int zid
= page_zonenum(page
);
686 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
690 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
691 struct mem_cgroup_per_zone
*mz
,
692 struct mem_cgroup_tree_per_zone
*mctz
,
693 unsigned long long new_usage_in_excess
)
695 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
696 struct rb_node
*parent
= NULL
;
697 struct mem_cgroup_per_zone
*mz_node
;
702 mz
->usage_in_excess
= new_usage_in_excess
;
703 if (!mz
->usage_in_excess
)
707 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
709 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
712 * We can't avoid mem cgroups that are over their soft
713 * limit by the same amount
715 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
718 rb_link_node(&mz
->tree_node
, parent
, p
);
719 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
724 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
725 struct mem_cgroup_per_zone
*mz
,
726 struct mem_cgroup_tree_per_zone
*mctz
)
730 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
735 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
736 struct mem_cgroup_per_zone
*mz
,
737 struct mem_cgroup_tree_per_zone
*mctz
)
739 spin_lock(&mctz
->lock
);
740 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
741 spin_unlock(&mctz
->lock
);
745 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
747 unsigned long long excess
;
748 struct mem_cgroup_per_zone
*mz
;
749 struct mem_cgroup_tree_per_zone
*mctz
;
750 int nid
= page_to_nid(page
);
751 int zid
= page_zonenum(page
);
752 mctz
= soft_limit_tree_from_page(page
);
755 * Necessary to update all ancestors when hierarchy is used.
756 * because their event counter is not touched.
758 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
759 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
760 excess
= res_counter_soft_limit_excess(&memcg
->res
);
762 * We have to update the tree if mz is on RB-tree or
763 * mem is over its softlimit.
765 if (excess
|| mz
->on_tree
) {
766 spin_lock(&mctz
->lock
);
767 /* if on-tree, remove it */
769 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
771 * Insert again. mz->usage_in_excess will be updated.
772 * If excess is 0, no tree ops.
774 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
775 spin_unlock(&mctz
->lock
);
780 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
783 struct mem_cgroup_per_zone
*mz
;
784 struct mem_cgroup_tree_per_zone
*mctz
;
786 for_each_node(node
) {
787 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
788 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
789 mctz
= soft_limit_tree_node_zone(node
, zone
);
790 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
795 static struct mem_cgroup_per_zone
*
796 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
798 struct rb_node
*rightmost
= NULL
;
799 struct mem_cgroup_per_zone
*mz
;
803 rightmost
= rb_last(&mctz
->rb_root
);
805 goto done
; /* Nothing to reclaim from */
807 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
809 * Remove the node now but someone else can add it back,
810 * we will to add it back at the end of reclaim to its correct
811 * position in the tree.
813 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
814 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
815 !css_tryget(&mz
->memcg
->css
))
821 static struct mem_cgroup_per_zone
*
822 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
824 struct mem_cgroup_per_zone
*mz
;
826 spin_lock(&mctz
->lock
);
827 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
828 spin_unlock(&mctz
->lock
);
833 * Implementation Note: reading percpu statistics for memcg.
835 * Both of vmstat[] and percpu_counter has threshold and do periodic
836 * synchronization to implement "quick" read. There are trade-off between
837 * reading cost and precision of value. Then, we may have a chance to implement
838 * a periodic synchronizion of counter in memcg's counter.
840 * But this _read() function is used for user interface now. The user accounts
841 * memory usage by memory cgroup and he _always_ requires exact value because
842 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
843 * have to visit all online cpus and make sum. So, for now, unnecessary
844 * synchronization is not implemented. (just implemented for cpu hotplug)
846 * If there are kernel internal actions which can make use of some not-exact
847 * value, and reading all cpu value can be performance bottleneck in some
848 * common workload, threashold and synchonization as vmstat[] should be
851 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
852 enum mem_cgroup_stat_index idx
)
858 for_each_online_cpu(cpu
)
859 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
860 #ifdef CONFIG_HOTPLUG_CPU
861 spin_lock(&memcg
->pcp_counter_lock
);
862 val
+= memcg
->nocpu_base
.count
[idx
];
863 spin_unlock(&memcg
->pcp_counter_lock
);
869 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
872 int val
= (charge
) ? 1 : -1;
873 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
876 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
877 enum mem_cgroup_events_index idx
)
879 unsigned long val
= 0;
882 for_each_online_cpu(cpu
)
883 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
884 #ifdef CONFIG_HOTPLUG_CPU
885 spin_lock(&memcg
->pcp_counter_lock
);
886 val
+= memcg
->nocpu_base
.events
[idx
];
887 spin_unlock(&memcg
->pcp_counter_lock
);
892 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
894 bool anon
, int nr_pages
)
899 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
900 * counted as CACHE even if it's on ANON LRU.
903 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
906 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
909 if (PageTransHuge(page
))
910 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
913 /* pagein of a big page is an event. So, ignore page size */
915 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
917 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
918 nr_pages
= -nr_pages
; /* for event */
921 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
927 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
929 struct mem_cgroup_per_zone
*mz
;
931 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
932 return mz
->lru_size
[lru
];
936 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
937 unsigned int lru_mask
)
939 struct mem_cgroup_per_zone
*mz
;
941 unsigned long ret
= 0;
943 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
946 if (BIT(lru
) & lru_mask
)
947 ret
+= mz
->lru_size
[lru
];
953 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
954 int nid
, unsigned int lru_mask
)
959 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
960 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
966 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
967 unsigned int lru_mask
)
972 for_each_node_state(nid
, N_MEMORY
)
973 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
977 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
978 enum mem_cgroup_events_target target
)
980 unsigned long val
, next
;
982 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
983 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
984 /* from time_after() in jiffies.h */
985 if ((long)next
- (long)val
< 0) {
987 case MEM_CGROUP_TARGET_THRESH
:
988 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
990 case MEM_CGROUP_TARGET_SOFTLIMIT
:
991 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
993 case MEM_CGROUP_TARGET_NUMAINFO
:
994 next
= val
+ NUMAINFO_EVENTS_TARGET
;
999 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1006 * Check events in order.
1009 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1012 /* threshold event is triggered in finer grain than soft limit */
1013 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1014 MEM_CGROUP_TARGET_THRESH
))) {
1016 bool do_numainfo __maybe_unused
;
1018 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1019 MEM_CGROUP_TARGET_SOFTLIMIT
);
1020 #if MAX_NUMNODES > 1
1021 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1022 MEM_CGROUP_TARGET_NUMAINFO
);
1026 mem_cgroup_threshold(memcg
);
1027 if (unlikely(do_softlimit
))
1028 mem_cgroup_update_tree(memcg
, page
);
1029 #if MAX_NUMNODES > 1
1030 if (unlikely(do_numainfo
))
1031 atomic_inc(&memcg
->numainfo_events
);
1037 static inline struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
1039 return mem_cgroup_from_css(cgroup_css(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_css(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 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1078 * ref. count) or NULL if the whole root's subtree has been visited.
1080 * helper function to be used by mem_cgroup_iter
1082 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1083 struct mem_cgroup
*last_visited
)
1085 struct cgroup
*prev_cgroup
, *next_cgroup
;
1088 * Root is not visited by cgroup iterators so it needs an
1094 prev_cgroup
= (last_visited
== root
) ? NULL
1095 : last_visited
->css
.cgroup
;
1097 next_cgroup
= cgroup_next_descendant_pre(
1098 prev_cgroup
, root
->css
.cgroup
);
1101 * Even if we found a group we have to make sure it is
1102 * alive. css && !memcg means that the groups should be
1103 * skipped and we should continue the tree walk.
1104 * last_visited css is safe to use because it is
1105 * protected by css_get and the tree walk is rcu safe.
1108 struct mem_cgroup
*mem
= mem_cgroup_from_cont(
1110 if (css_tryget(&mem
->css
))
1113 prev_cgroup
= next_cgroup
;
1121 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1124 * When a group in the hierarchy below root is destroyed, the
1125 * hierarchy iterator can no longer be trusted since it might
1126 * have pointed to the destroyed group. Invalidate it.
1128 atomic_inc(&root
->dead_count
);
1131 static struct mem_cgroup
*
1132 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1133 struct mem_cgroup
*root
,
1136 struct mem_cgroup
*position
= NULL
;
1138 * A cgroup destruction happens in two stages: offlining and
1139 * release. They are separated by a RCU grace period.
1141 * If the iterator is valid, we may still race with an
1142 * offlining. The RCU lock ensures the object won't be
1143 * released, tryget will fail if we lost the race.
1145 *sequence
= atomic_read(&root
->dead_count
);
1146 if (iter
->last_dead_count
== *sequence
) {
1148 position
= iter
->last_visited
;
1149 if (position
&& !css_tryget(&position
->css
))
1155 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1156 struct mem_cgroup
*last_visited
,
1157 struct mem_cgroup
*new_position
,
1161 css_put(&last_visited
->css
);
1163 * We store the sequence count from the time @last_visited was
1164 * loaded successfully instead of rereading it here so that we
1165 * don't lose destruction events in between. We could have
1166 * raced with the destruction of @new_position after all.
1168 iter
->last_visited
= new_position
;
1170 iter
->last_dead_count
= sequence
;
1174 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1175 * @root: hierarchy root
1176 * @prev: previously returned memcg, NULL on first invocation
1177 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1179 * Returns references to children of the hierarchy below @root, or
1180 * @root itself, or %NULL after a full round-trip.
1182 * Caller must pass the return value in @prev on subsequent
1183 * invocations for reference counting, or use mem_cgroup_iter_break()
1184 * to cancel a hierarchy walk before the round-trip is complete.
1186 * Reclaimers can specify a zone and a priority level in @reclaim to
1187 * divide up the memcgs in the hierarchy among all concurrent
1188 * reclaimers operating on the same zone and priority.
1190 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1191 struct mem_cgroup
*prev
,
1192 struct mem_cgroup_reclaim_cookie
*reclaim
)
1194 struct mem_cgroup
*memcg
= NULL
;
1195 struct mem_cgroup
*last_visited
= NULL
;
1197 if (mem_cgroup_disabled())
1201 root
= root_mem_cgroup
;
1203 if (prev
&& !reclaim
)
1204 last_visited
= prev
;
1206 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1214 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1215 int uninitialized_var(seq
);
1218 int nid
= zone_to_nid(reclaim
->zone
);
1219 int zid
= zone_idx(reclaim
->zone
);
1220 struct mem_cgroup_per_zone
*mz
;
1222 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1223 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1224 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1225 iter
->last_visited
= NULL
;
1229 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1232 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1235 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1239 else if (!prev
&& memcg
)
1240 reclaim
->generation
= iter
->generation
;
1249 if (prev
&& prev
!= root
)
1250 css_put(&prev
->css
);
1256 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1257 * @root: hierarchy root
1258 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1260 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1261 struct mem_cgroup
*prev
)
1264 root
= root_mem_cgroup
;
1265 if (prev
&& prev
!= root
)
1266 css_put(&prev
->css
);
1270 * Iteration constructs for visiting all cgroups (under a tree). If
1271 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1272 * be used for reference counting.
1274 #define for_each_mem_cgroup_tree(iter, root) \
1275 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1277 iter = mem_cgroup_iter(root, iter, NULL))
1279 #define for_each_mem_cgroup(iter) \
1280 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1282 iter = mem_cgroup_iter(NULL, iter, NULL))
1284 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1286 struct mem_cgroup
*memcg
;
1289 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1290 if (unlikely(!memcg
))
1295 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1298 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1306 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1309 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1310 * @zone: zone of the wanted lruvec
1311 * @memcg: memcg of the wanted lruvec
1313 * Returns the lru list vector holding pages for the given @zone and
1314 * @mem. This can be the global zone lruvec, if the memory controller
1317 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1318 struct mem_cgroup
*memcg
)
1320 struct mem_cgroup_per_zone
*mz
;
1321 struct lruvec
*lruvec
;
1323 if (mem_cgroup_disabled()) {
1324 lruvec
= &zone
->lruvec
;
1328 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1329 lruvec
= &mz
->lruvec
;
1332 * Since a node can be onlined after the mem_cgroup was created,
1333 * we have to be prepared to initialize lruvec->zone here;
1334 * and if offlined then reonlined, we need to reinitialize it.
1336 if (unlikely(lruvec
->zone
!= zone
))
1337 lruvec
->zone
= zone
;
1342 * Following LRU functions are allowed to be used without PCG_LOCK.
1343 * Operations are called by routine of global LRU independently from memcg.
1344 * What we have to take care of here is validness of pc->mem_cgroup.
1346 * Changes to pc->mem_cgroup happens when
1349 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1350 * It is added to LRU before charge.
1351 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1352 * When moving account, the page is not on LRU. It's isolated.
1356 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1358 * @zone: zone of the page
1360 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1362 struct mem_cgroup_per_zone
*mz
;
1363 struct mem_cgroup
*memcg
;
1364 struct page_cgroup
*pc
;
1365 struct lruvec
*lruvec
;
1367 if (mem_cgroup_disabled()) {
1368 lruvec
= &zone
->lruvec
;
1372 pc
= lookup_page_cgroup(page
);
1373 memcg
= pc
->mem_cgroup
;
1376 * Surreptitiously switch any uncharged offlist page to root:
1377 * an uncharged page off lru does nothing to secure
1378 * its former mem_cgroup from sudden removal.
1380 * Our caller holds lru_lock, and PageCgroupUsed is updated
1381 * under page_cgroup lock: between them, they make all uses
1382 * of pc->mem_cgroup safe.
1384 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1385 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1387 mz
= page_cgroup_zoneinfo(memcg
, page
);
1388 lruvec
= &mz
->lruvec
;
1391 * Since a node can be onlined after the mem_cgroup was created,
1392 * we have to be prepared to initialize lruvec->zone here;
1393 * and if offlined then reonlined, we need to reinitialize it.
1395 if (unlikely(lruvec
->zone
!= zone
))
1396 lruvec
->zone
= zone
;
1401 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1402 * @lruvec: mem_cgroup per zone lru vector
1403 * @lru: index of lru list the page is sitting on
1404 * @nr_pages: positive when adding or negative when removing
1406 * This function must be called when a page is added to or removed from an
1409 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1412 struct mem_cgroup_per_zone
*mz
;
1413 unsigned long *lru_size
;
1415 if (mem_cgroup_disabled())
1418 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1419 lru_size
= mz
->lru_size
+ lru
;
1420 *lru_size
+= nr_pages
;
1421 VM_BUG_ON((long)(*lru_size
) < 0);
1425 * Checks whether given mem is same or in the root_mem_cgroup's
1428 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1429 struct mem_cgroup
*memcg
)
1431 if (root_memcg
== memcg
)
1433 if (!root_memcg
->use_hierarchy
|| !memcg
)
1435 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1438 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1439 struct mem_cgroup
*memcg
)
1444 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1449 bool task_in_mem_cgroup(struct task_struct
*task
,
1450 const struct mem_cgroup
*memcg
)
1452 struct mem_cgroup
*curr
= NULL
;
1453 struct task_struct
*p
;
1456 p
= find_lock_task_mm(task
);
1458 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1462 * All threads may have already detached their mm's, but the oom
1463 * killer still needs to detect if they have already been oom
1464 * killed to prevent needlessly killing additional tasks.
1467 curr
= mem_cgroup_from_task(task
);
1469 css_get(&curr
->css
);
1475 * We should check use_hierarchy of "memcg" not "curr". Because checking
1476 * use_hierarchy of "curr" here make this function true if hierarchy is
1477 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1478 * hierarchy(even if use_hierarchy is disabled in "memcg").
1480 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1481 css_put(&curr
->css
);
1485 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1487 unsigned long inactive_ratio
;
1488 unsigned long inactive
;
1489 unsigned long active
;
1492 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1493 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1495 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1497 inactive_ratio
= int_sqrt(10 * gb
);
1501 return inactive
* inactive_ratio
< active
;
1504 #define mem_cgroup_from_res_counter(counter, member) \
1505 container_of(counter, struct mem_cgroup, member)
1508 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1509 * @memcg: the memory cgroup
1511 * Returns the maximum amount of memory @mem can be charged with, in
1514 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1516 unsigned long long margin
;
1518 margin
= res_counter_margin(&memcg
->res
);
1519 if (do_swap_account
)
1520 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1521 return margin
>> PAGE_SHIFT
;
1524 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1527 if (!css_parent(&memcg
->css
))
1528 return vm_swappiness
;
1530 return memcg
->swappiness
;
1534 * memcg->moving_account is used for checking possibility that some thread is
1535 * calling move_account(). When a thread on CPU-A starts moving pages under
1536 * a memcg, other threads should check memcg->moving_account under
1537 * rcu_read_lock(), like this:
1541 * memcg->moving_account+1 if (memcg->mocing_account)
1543 * synchronize_rcu() update something.
1548 /* for quick checking without looking up memcg */
1549 atomic_t memcg_moving __read_mostly
;
1551 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1553 atomic_inc(&memcg_moving
);
1554 atomic_inc(&memcg
->moving_account
);
1558 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1561 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1562 * We check NULL in callee rather than caller.
1565 atomic_dec(&memcg_moving
);
1566 atomic_dec(&memcg
->moving_account
);
1571 * 2 routines for checking "mem" is under move_account() or not.
1573 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1574 * is used for avoiding races in accounting. If true,
1575 * pc->mem_cgroup may be overwritten.
1577 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1578 * under hierarchy of moving cgroups. This is for
1579 * waiting at hith-memory prressure caused by "move".
1582 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1584 VM_BUG_ON(!rcu_read_lock_held());
1585 return atomic_read(&memcg
->moving_account
) > 0;
1588 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1590 struct mem_cgroup
*from
;
1591 struct mem_cgroup
*to
;
1594 * Unlike task_move routines, we access mc.to, mc.from not under
1595 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1597 spin_lock(&mc
.lock
);
1603 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1604 || mem_cgroup_same_or_subtree(memcg
, to
);
1606 spin_unlock(&mc
.lock
);
1610 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1612 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1613 if (mem_cgroup_under_move(memcg
)) {
1615 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1616 /* moving charge context might have finished. */
1619 finish_wait(&mc
.waitq
, &wait
);
1627 * Take this lock when
1628 * - a code tries to modify page's memcg while it's USED.
1629 * - a code tries to modify page state accounting in a memcg.
1630 * see mem_cgroup_stolen(), too.
1632 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1633 unsigned long *flags
)
1635 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1638 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1639 unsigned long *flags
)
1641 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1644 #define K(x) ((x) << (PAGE_SHIFT-10))
1646 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1647 * @memcg: The memory cgroup that went over limit
1648 * @p: Task that is going to be killed
1650 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1653 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1655 struct cgroup
*task_cgrp
;
1656 struct cgroup
*mem_cgrp
;
1658 * Need a buffer in BSS, can't rely on allocations. The code relies
1659 * on the assumption that OOM is serialized for memory controller.
1660 * If this assumption is broken, revisit this code.
1662 static char memcg_name
[PATH_MAX
];
1664 struct mem_cgroup
*iter
;
1672 mem_cgrp
= memcg
->css
.cgroup
;
1673 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1675 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1678 * Unfortunately, we are unable to convert to a useful name
1679 * But we'll still print out the usage information
1686 pr_info("Task in %s killed", memcg_name
);
1689 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1697 * Continues from above, so we don't need an KERN_ level
1699 pr_cont(" as a result of limit of %s\n", memcg_name
);
1702 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1703 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1704 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1705 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1706 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1707 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1708 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1709 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1710 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1711 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1712 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1713 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1715 for_each_mem_cgroup_tree(iter
, memcg
) {
1716 pr_info("Memory cgroup stats");
1719 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1721 pr_cont(" for %s", memcg_name
);
1725 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1726 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1728 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1729 K(mem_cgroup_read_stat(iter
, i
)));
1732 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1733 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1734 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1741 * This function returns the number of memcg under hierarchy tree. Returns
1742 * 1(self count) if no children.
1744 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1747 struct mem_cgroup
*iter
;
1749 for_each_mem_cgroup_tree(iter
, memcg
)
1755 * Return the memory (and swap, if configured) limit for a memcg.
1757 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1761 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1764 * Do not consider swap space if we cannot swap due to swappiness
1766 if (mem_cgroup_swappiness(memcg
)) {
1769 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1770 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1773 * If memsw is finite and limits the amount of swap space
1774 * available to this memcg, return that limit.
1776 limit
= min(limit
, memsw
);
1782 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1785 struct mem_cgroup
*iter
;
1786 unsigned long chosen_points
= 0;
1787 unsigned long totalpages
;
1788 unsigned int points
= 0;
1789 struct task_struct
*chosen
= NULL
;
1792 * If current has a pending SIGKILL or is exiting, then automatically
1793 * select it. The goal is to allow it to allocate so that it may
1794 * quickly exit and free its memory.
1796 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1797 set_thread_flag(TIF_MEMDIE
);
1801 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1802 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1803 for_each_mem_cgroup_tree(iter
, memcg
) {
1804 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1805 struct cgroup_iter it
;
1806 struct task_struct
*task
;
1808 cgroup_iter_start(cgroup
, &it
);
1809 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1810 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1812 case OOM_SCAN_SELECT
:
1814 put_task_struct(chosen
);
1816 chosen_points
= ULONG_MAX
;
1817 get_task_struct(chosen
);
1819 case OOM_SCAN_CONTINUE
:
1821 case OOM_SCAN_ABORT
:
1822 cgroup_iter_end(cgroup
, &it
);
1823 mem_cgroup_iter_break(memcg
, iter
);
1825 put_task_struct(chosen
);
1830 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1831 if (points
> chosen_points
) {
1833 put_task_struct(chosen
);
1835 chosen_points
= points
;
1836 get_task_struct(chosen
);
1839 cgroup_iter_end(cgroup
, &it
);
1844 points
= chosen_points
* 1000 / totalpages
;
1845 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1846 NULL
, "Memory cgroup out of memory");
1849 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1851 unsigned long flags
)
1853 unsigned long total
= 0;
1854 bool noswap
= false;
1857 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1859 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1862 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1864 drain_all_stock_async(memcg
);
1865 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1867 * Allow limit shrinkers, which are triggered directly
1868 * by userspace, to catch signals and stop reclaim
1869 * after minimal progress, regardless of the margin.
1871 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1873 if (mem_cgroup_margin(memcg
))
1876 * If nothing was reclaimed after two attempts, there
1877 * may be no reclaimable pages in this hierarchy.
1886 * test_mem_cgroup_node_reclaimable
1887 * @memcg: the target memcg
1888 * @nid: the node ID to be checked.
1889 * @noswap : specify true here if the user wants flle only information.
1891 * This function returns whether the specified memcg contains any
1892 * reclaimable pages on a node. Returns true if there are any reclaimable
1893 * pages in the node.
1895 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1896 int nid
, bool noswap
)
1898 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1900 if (noswap
|| !total_swap_pages
)
1902 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1907 #if MAX_NUMNODES > 1
1910 * Always updating the nodemask is not very good - even if we have an empty
1911 * list or the wrong list here, we can start from some node and traverse all
1912 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1915 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1919 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1920 * pagein/pageout changes since the last update.
1922 if (!atomic_read(&memcg
->numainfo_events
))
1924 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1927 /* make a nodemask where this memcg uses memory from */
1928 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1930 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1932 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1933 node_clear(nid
, memcg
->scan_nodes
);
1936 atomic_set(&memcg
->numainfo_events
, 0);
1937 atomic_set(&memcg
->numainfo_updating
, 0);
1941 * Selecting a node where we start reclaim from. Because what we need is just
1942 * reducing usage counter, start from anywhere is O,K. Considering
1943 * memory reclaim from current node, there are pros. and cons.
1945 * Freeing memory from current node means freeing memory from a node which
1946 * we'll use or we've used. So, it may make LRU bad. And if several threads
1947 * hit limits, it will see a contention on a node. But freeing from remote
1948 * node means more costs for memory reclaim because of memory latency.
1950 * Now, we use round-robin. Better algorithm is welcomed.
1952 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1956 mem_cgroup_may_update_nodemask(memcg
);
1957 node
= memcg
->last_scanned_node
;
1959 node
= next_node(node
, memcg
->scan_nodes
);
1960 if (node
== MAX_NUMNODES
)
1961 node
= first_node(memcg
->scan_nodes
);
1963 * We call this when we hit limit, not when pages are added to LRU.
1964 * No LRU may hold pages because all pages are UNEVICTABLE or
1965 * memcg is too small and all pages are not on LRU. In that case,
1966 * we use curret node.
1968 if (unlikely(node
== MAX_NUMNODES
))
1969 node
= numa_node_id();
1971 memcg
->last_scanned_node
= node
;
1976 * Check all nodes whether it contains reclaimable pages or not.
1977 * For quick scan, we make use of scan_nodes. This will allow us to skip
1978 * unused nodes. But scan_nodes is lazily updated and may not cotain
1979 * enough new information. We need to do double check.
1981 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1986 * quick check...making use of scan_node.
1987 * We can skip unused nodes.
1989 if (!nodes_empty(memcg
->scan_nodes
)) {
1990 for (nid
= first_node(memcg
->scan_nodes
);
1992 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1994 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1999 * Check rest of nodes.
2001 for_each_node_state(nid
, N_MEMORY
) {
2002 if (node_isset(nid
, memcg
->scan_nodes
))
2004 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2011 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2016 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2018 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2022 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2025 unsigned long *total_scanned
)
2027 struct mem_cgroup
*victim
= NULL
;
2030 unsigned long excess
;
2031 unsigned long nr_scanned
;
2032 struct mem_cgroup_reclaim_cookie reclaim
= {
2037 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2040 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2045 * If we have not been able to reclaim
2046 * anything, it might because there are
2047 * no reclaimable pages under this hierarchy
2052 * We want to do more targeted reclaim.
2053 * excess >> 2 is not to excessive so as to
2054 * reclaim too much, nor too less that we keep
2055 * coming back to reclaim from this cgroup
2057 if (total
>= (excess
>> 2) ||
2058 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2063 if (!mem_cgroup_reclaimable(victim
, false))
2065 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2067 *total_scanned
+= nr_scanned
;
2068 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2071 mem_cgroup_iter_break(root_memcg
, victim
);
2076 * Check OOM-Killer is already running under our hierarchy.
2077 * If someone is running, return false.
2078 * Has to be called with memcg_oom_lock
2080 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
2082 struct mem_cgroup
*iter
, *failed
= NULL
;
2084 for_each_mem_cgroup_tree(iter
, memcg
) {
2085 if (iter
->oom_lock
) {
2087 * this subtree of our hierarchy is already locked
2088 * so we cannot give a lock.
2091 mem_cgroup_iter_break(memcg
, iter
);
2094 iter
->oom_lock
= true;
2101 * OK, we failed to lock the whole subtree so we have to clean up
2102 * what we set up to the failing subtree
2104 for_each_mem_cgroup_tree(iter
, memcg
) {
2105 if (iter
== failed
) {
2106 mem_cgroup_iter_break(memcg
, iter
);
2109 iter
->oom_lock
= false;
2115 * Has to be called with memcg_oom_lock
2117 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2119 struct mem_cgroup
*iter
;
2121 for_each_mem_cgroup_tree(iter
, memcg
)
2122 iter
->oom_lock
= false;
2126 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2128 struct mem_cgroup
*iter
;
2130 for_each_mem_cgroup_tree(iter
, memcg
)
2131 atomic_inc(&iter
->under_oom
);
2134 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2136 struct mem_cgroup
*iter
;
2139 * When a new child is created while the hierarchy is under oom,
2140 * mem_cgroup_oom_lock() may not be called. We have to use
2141 * atomic_add_unless() here.
2143 for_each_mem_cgroup_tree(iter
, memcg
)
2144 atomic_add_unless(&iter
->under_oom
, -1, 0);
2147 static DEFINE_SPINLOCK(memcg_oom_lock
);
2148 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2150 struct oom_wait_info
{
2151 struct mem_cgroup
*memcg
;
2155 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2156 unsigned mode
, int sync
, void *arg
)
2158 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2159 struct mem_cgroup
*oom_wait_memcg
;
2160 struct oom_wait_info
*oom_wait_info
;
2162 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2163 oom_wait_memcg
= oom_wait_info
->memcg
;
2166 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2167 * Then we can use css_is_ancestor without taking care of RCU.
2169 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2170 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2172 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2175 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2177 /* for filtering, pass "memcg" as argument. */
2178 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2181 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2183 if (memcg
&& atomic_read(&memcg
->under_oom
))
2184 memcg_wakeup_oom(memcg
);
2188 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2190 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2193 struct oom_wait_info owait
;
2194 bool locked
, need_to_kill
;
2196 owait
.memcg
= memcg
;
2197 owait
.wait
.flags
= 0;
2198 owait
.wait
.func
= memcg_oom_wake_function
;
2199 owait
.wait
.private = current
;
2200 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2201 need_to_kill
= true;
2202 mem_cgroup_mark_under_oom(memcg
);
2204 /* At first, try to OOM lock hierarchy under memcg.*/
2205 spin_lock(&memcg_oom_lock
);
2206 locked
= mem_cgroup_oom_lock(memcg
);
2208 * Even if signal_pending(), we can't quit charge() loop without
2209 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2210 * under OOM is always welcomed, use TASK_KILLABLE here.
2212 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2213 if (!locked
|| memcg
->oom_kill_disable
)
2214 need_to_kill
= false;
2216 mem_cgroup_oom_notify(memcg
);
2217 spin_unlock(&memcg_oom_lock
);
2220 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2221 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2224 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2226 spin_lock(&memcg_oom_lock
);
2228 mem_cgroup_oom_unlock(memcg
);
2229 memcg_wakeup_oom(memcg
);
2230 spin_unlock(&memcg_oom_lock
);
2232 mem_cgroup_unmark_under_oom(memcg
);
2234 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2236 /* Give chance to dying process */
2237 schedule_timeout_uninterruptible(1);
2242 * Currently used to update mapped file statistics, but the routine can be
2243 * generalized to update other statistics as well.
2245 * Notes: Race condition
2247 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2248 * it tends to be costly. But considering some conditions, we doesn't need
2249 * to do so _always_.
2251 * Considering "charge", lock_page_cgroup() is not required because all
2252 * file-stat operations happen after a page is attached to radix-tree. There
2253 * are no race with "charge".
2255 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2256 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2257 * if there are race with "uncharge". Statistics itself is properly handled
2260 * Considering "move", this is an only case we see a race. To make the race
2261 * small, we check mm->moving_account and detect there are possibility of race
2262 * If there is, we take a lock.
2265 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2266 bool *locked
, unsigned long *flags
)
2268 struct mem_cgroup
*memcg
;
2269 struct page_cgroup
*pc
;
2271 pc
= lookup_page_cgroup(page
);
2273 memcg
= pc
->mem_cgroup
;
2274 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2277 * If this memory cgroup is not under account moving, we don't
2278 * need to take move_lock_mem_cgroup(). Because we already hold
2279 * rcu_read_lock(), any calls to move_account will be delayed until
2280 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2282 if (!mem_cgroup_stolen(memcg
))
2285 move_lock_mem_cgroup(memcg
, flags
);
2286 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2287 move_unlock_mem_cgroup(memcg
, flags
);
2293 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2295 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2298 * It's guaranteed that pc->mem_cgroup never changes while
2299 * lock is held because a routine modifies pc->mem_cgroup
2300 * should take move_lock_mem_cgroup().
2302 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2305 void mem_cgroup_update_page_stat(struct page
*page
,
2306 enum mem_cgroup_page_stat_item idx
, int val
)
2308 struct mem_cgroup
*memcg
;
2309 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2310 unsigned long uninitialized_var(flags
);
2312 if (mem_cgroup_disabled())
2315 memcg
= pc
->mem_cgroup
;
2316 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2320 case MEMCG_NR_FILE_MAPPED
:
2321 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2327 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2331 * size of first charge trial. "32" comes from vmscan.c's magic value.
2332 * TODO: maybe necessary to use big numbers in big irons.
2334 #define CHARGE_BATCH 32U
2335 struct memcg_stock_pcp
{
2336 struct mem_cgroup
*cached
; /* this never be root cgroup */
2337 unsigned int nr_pages
;
2338 struct work_struct work
;
2339 unsigned long flags
;
2340 #define FLUSHING_CACHED_CHARGE 0
2342 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2343 static DEFINE_MUTEX(percpu_charge_mutex
);
2346 * consume_stock: Try to consume stocked charge on this cpu.
2347 * @memcg: memcg to consume from.
2348 * @nr_pages: how many pages to charge.
2350 * The charges will only happen if @memcg matches the current cpu's memcg
2351 * stock, and at least @nr_pages are available in that stock. Failure to
2352 * service an allocation will refill the stock.
2354 * returns true if successful, false otherwise.
2356 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2358 struct memcg_stock_pcp
*stock
;
2361 if (nr_pages
> CHARGE_BATCH
)
2364 stock
= &get_cpu_var(memcg_stock
);
2365 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2366 stock
->nr_pages
-= nr_pages
;
2367 else /* need to call res_counter_charge */
2369 put_cpu_var(memcg_stock
);
2374 * Returns stocks cached in percpu to res_counter and reset cached information.
2376 static void drain_stock(struct memcg_stock_pcp
*stock
)
2378 struct mem_cgroup
*old
= stock
->cached
;
2380 if (stock
->nr_pages
) {
2381 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2383 res_counter_uncharge(&old
->res
, bytes
);
2384 if (do_swap_account
)
2385 res_counter_uncharge(&old
->memsw
, bytes
);
2386 stock
->nr_pages
= 0;
2388 stock
->cached
= NULL
;
2392 * This must be called under preempt disabled or must be called by
2393 * a thread which is pinned to local cpu.
2395 static void drain_local_stock(struct work_struct
*dummy
)
2397 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2399 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2402 static void __init
memcg_stock_init(void)
2406 for_each_possible_cpu(cpu
) {
2407 struct memcg_stock_pcp
*stock
=
2408 &per_cpu(memcg_stock
, cpu
);
2409 INIT_WORK(&stock
->work
, drain_local_stock
);
2414 * Cache charges(val) which is from res_counter, to local per_cpu area.
2415 * This will be consumed by consume_stock() function, later.
2417 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2419 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2421 if (stock
->cached
!= memcg
) { /* reset if necessary */
2423 stock
->cached
= memcg
;
2425 stock
->nr_pages
+= nr_pages
;
2426 put_cpu_var(memcg_stock
);
2430 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2431 * of the hierarchy under it. sync flag says whether we should block
2432 * until the work is done.
2434 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2438 /* Notify other cpus that system-wide "drain" is running */
2441 for_each_online_cpu(cpu
) {
2442 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2443 struct mem_cgroup
*memcg
;
2445 memcg
= stock
->cached
;
2446 if (!memcg
|| !stock
->nr_pages
)
2448 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2450 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2452 drain_local_stock(&stock
->work
);
2454 schedule_work_on(cpu
, &stock
->work
);
2462 for_each_online_cpu(cpu
) {
2463 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2464 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2465 flush_work(&stock
->work
);
2472 * Tries to drain stocked charges in other cpus. This function is asynchronous
2473 * and just put a work per cpu for draining localy on each cpu. Caller can
2474 * expects some charges will be back to res_counter later but cannot wait for
2477 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2480 * If someone calls draining, avoid adding more kworker runs.
2482 if (!mutex_trylock(&percpu_charge_mutex
))
2484 drain_all_stock(root_memcg
, false);
2485 mutex_unlock(&percpu_charge_mutex
);
2488 /* This is a synchronous drain interface. */
2489 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2491 /* called when force_empty is called */
2492 mutex_lock(&percpu_charge_mutex
);
2493 drain_all_stock(root_memcg
, true);
2494 mutex_unlock(&percpu_charge_mutex
);
2498 * This function drains percpu counter value from DEAD cpu and
2499 * move it to local cpu. Note that this function can be preempted.
2501 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2505 spin_lock(&memcg
->pcp_counter_lock
);
2506 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2507 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2509 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2510 memcg
->nocpu_base
.count
[i
] += x
;
2512 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2513 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2515 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2516 memcg
->nocpu_base
.events
[i
] += x
;
2518 spin_unlock(&memcg
->pcp_counter_lock
);
2521 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2522 unsigned long action
,
2525 int cpu
= (unsigned long)hcpu
;
2526 struct memcg_stock_pcp
*stock
;
2527 struct mem_cgroup
*iter
;
2529 if (action
== CPU_ONLINE
)
2532 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2535 for_each_mem_cgroup(iter
)
2536 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2538 stock
= &per_cpu(memcg_stock
, cpu
);
2544 /* See __mem_cgroup_try_charge() for details */
2546 CHARGE_OK
, /* success */
2547 CHARGE_RETRY
, /* need to retry but retry is not bad */
2548 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2549 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2550 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2553 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2554 unsigned int nr_pages
, unsigned int min_pages
,
2557 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2558 struct mem_cgroup
*mem_over_limit
;
2559 struct res_counter
*fail_res
;
2560 unsigned long flags
= 0;
2563 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2566 if (!do_swap_account
)
2568 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2572 res_counter_uncharge(&memcg
->res
, csize
);
2573 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2574 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2576 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2578 * Never reclaim on behalf of optional batching, retry with a
2579 * single page instead.
2581 if (nr_pages
> min_pages
)
2582 return CHARGE_RETRY
;
2584 if (!(gfp_mask
& __GFP_WAIT
))
2585 return CHARGE_WOULDBLOCK
;
2587 if (gfp_mask
& __GFP_NORETRY
)
2588 return CHARGE_NOMEM
;
2590 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2591 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2592 return CHARGE_RETRY
;
2594 * Even though the limit is exceeded at this point, reclaim
2595 * may have been able to free some pages. Retry the charge
2596 * before killing the task.
2598 * Only for regular pages, though: huge pages are rather
2599 * unlikely to succeed so close to the limit, and we fall back
2600 * to regular pages anyway in case of failure.
2602 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2603 return CHARGE_RETRY
;
2606 * At task move, charge accounts can be doubly counted. So, it's
2607 * better to wait until the end of task_move if something is going on.
2609 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2610 return CHARGE_RETRY
;
2612 /* If we don't need to call oom-killer at el, return immediately */
2614 return CHARGE_NOMEM
;
2616 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2617 return CHARGE_OOM_DIE
;
2619 return CHARGE_RETRY
;
2623 * __mem_cgroup_try_charge() does
2624 * 1. detect memcg to be charged against from passed *mm and *ptr,
2625 * 2. update res_counter
2626 * 3. call memory reclaim if necessary.
2628 * In some special case, if the task is fatal, fatal_signal_pending() or
2629 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2630 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2631 * as possible without any hazards. 2: all pages should have a valid
2632 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2633 * pointer, that is treated as a charge to root_mem_cgroup.
2635 * So __mem_cgroup_try_charge() will return
2636 * 0 ... on success, filling *ptr with a valid memcg pointer.
2637 * -ENOMEM ... charge failure because of resource limits.
2638 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2640 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2641 * the oom-killer can be invoked.
2643 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2645 unsigned int nr_pages
,
2646 struct mem_cgroup
**ptr
,
2649 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2650 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2651 struct mem_cgroup
*memcg
= NULL
;
2655 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2656 * in system level. So, allow to go ahead dying process in addition to
2659 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2660 || fatal_signal_pending(current
)))
2664 * We always charge the cgroup the mm_struct belongs to.
2665 * The mm_struct's mem_cgroup changes on task migration if the
2666 * thread group leader migrates. It's possible that mm is not
2667 * set, if so charge the root memcg (happens for pagecache usage).
2670 *ptr
= root_mem_cgroup
;
2672 if (*ptr
) { /* css should be a valid one */
2674 if (mem_cgroup_is_root(memcg
))
2676 if (consume_stock(memcg
, nr_pages
))
2678 css_get(&memcg
->css
);
2680 struct task_struct
*p
;
2683 p
= rcu_dereference(mm
->owner
);
2685 * Because we don't have task_lock(), "p" can exit.
2686 * In that case, "memcg" can point to root or p can be NULL with
2687 * race with swapoff. Then, we have small risk of mis-accouning.
2688 * But such kind of mis-account by race always happens because
2689 * we don't have cgroup_mutex(). It's overkill and we allo that
2691 * (*) swapoff at el will charge against mm-struct not against
2692 * task-struct. So, mm->owner can be NULL.
2694 memcg
= mem_cgroup_from_task(p
);
2696 memcg
= root_mem_cgroup
;
2697 if (mem_cgroup_is_root(memcg
)) {
2701 if (consume_stock(memcg
, nr_pages
)) {
2703 * It seems dagerous to access memcg without css_get().
2704 * But considering how consume_stok works, it's not
2705 * necessary. If consume_stock success, some charges
2706 * from this memcg are cached on this cpu. So, we
2707 * don't need to call css_get()/css_tryget() before
2708 * calling consume_stock().
2713 /* after here, we may be blocked. we need to get refcnt */
2714 if (!css_tryget(&memcg
->css
)) {
2724 /* If killed, bypass charge */
2725 if (fatal_signal_pending(current
)) {
2726 css_put(&memcg
->css
);
2731 if (oom
&& !nr_oom_retries
) {
2733 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2736 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2741 case CHARGE_RETRY
: /* not in OOM situation but retry */
2743 css_put(&memcg
->css
);
2746 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2747 css_put(&memcg
->css
);
2749 case CHARGE_NOMEM
: /* OOM routine works */
2751 css_put(&memcg
->css
);
2754 /* If oom, we never return -ENOMEM */
2757 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2758 css_put(&memcg
->css
);
2761 } while (ret
!= CHARGE_OK
);
2763 if (batch
> nr_pages
)
2764 refill_stock(memcg
, batch
- nr_pages
);
2765 css_put(&memcg
->css
);
2773 *ptr
= root_mem_cgroup
;
2778 * Somemtimes we have to undo a charge we got by try_charge().
2779 * This function is for that and do uncharge, put css's refcnt.
2780 * gotten by try_charge().
2782 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2783 unsigned int nr_pages
)
2785 if (!mem_cgroup_is_root(memcg
)) {
2786 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2788 res_counter_uncharge(&memcg
->res
, bytes
);
2789 if (do_swap_account
)
2790 res_counter_uncharge(&memcg
->memsw
, bytes
);
2795 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2796 * This is useful when moving usage to parent cgroup.
2798 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2799 unsigned int nr_pages
)
2801 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2803 if (mem_cgroup_is_root(memcg
))
2806 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2807 if (do_swap_account
)
2808 res_counter_uncharge_until(&memcg
->memsw
,
2809 memcg
->memsw
.parent
, bytes
);
2813 * A helper function to get mem_cgroup from ID. must be called under
2814 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2815 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2816 * called against removed memcg.)
2818 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2820 struct cgroup_subsys_state
*css
;
2822 /* ID 0 is unused ID */
2825 css
= css_lookup(&mem_cgroup_subsys
, id
);
2828 return mem_cgroup_from_css(css
);
2831 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2833 struct mem_cgroup
*memcg
= NULL
;
2834 struct page_cgroup
*pc
;
2838 VM_BUG_ON(!PageLocked(page
));
2840 pc
= lookup_page_cgroup(page
);
2841 lock_page_cgroup(pc
);
2842 if (PageCgroupUsed(pc
)) {
2843 memcg
= pc
->mem_cgroup
;
2844 if (memcg
&& !css_tryget(&memcg
->css
))
2846 } else if (PageSwapCache(page
)) {
2847 ent
.val
= page_private(page
);
2848 id
= lookup_swap_cgroup_id(ent
);
2850 memcg
= mem_cgroup_lookup(id
);
2851 if (memcg
&& !css_tryget(&memcg
->css
))
2855 unlock_page_cgroup(pc
);
2859 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2861 unsigned int nr_pages
,
2862 enum charge_type ctype
,
2865 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2866 struct zone
*uninitialized_var(zone
);
2867 struct lruvec
*lruvec
;
2868 bool was_on_lru
= false;
2871 lock_page_cgroup(pc
);
2872 VM_BUG_ON(PageCgroupUsed(pc
));
2874 * we don't need page_cgroup_lock about tail pages, becase they are not
2875 * accessed by any other context at this point.
2879 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2880 * may already be on some other mem_cgroup's LRU. Take care of it.
2883 zone
= page_zone(page
);
2884 spin_lock_irq(&zone
->lru_lock
);
2885 if (PageLRU(page
)) {
2886 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2888 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2893 pc
->mem_cgroup
= memcg
;
2895 * We access a page_cgroup asynchronously without lock_page_cgroup().
2896 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2897 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2898 * before USED bit, we need memory barrier here.
2899 * See mem_cgroup_add_lru_list(), etc.
2902 SetPageCgroupUsed(pc
);
2906 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2907 VM_BUG_ON(PageLRU(page
));
2909 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2911 spin_unlock_irq(&zone
->lru_lock
);
2914 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2919 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2920 unlock_page_cgroup(pc
);
2923 * "charge_statistics" updated event counter. Then, check it.
2924 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2925 * if they exceeds softlimit.
2927 memcg_check_events(memcg
, page
);
2930 static DEFINE_MUTEX(set_limit_mutex
);
2932 #ifdef CONFIG_MEMCG_KMEM
2933 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2935 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2936 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2940 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2941 * in the memcg_cache_params struct.
2943 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2945 struct kmem_cache
*cachep
;
2947 VM_BUG_ON(p
->is_root_cache
);
2948 cachep
= p
->root_cache
;
2949 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2952 #ifdef CONFIG_SLABINFO
2953 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2954 struct cftype
*cft
, struct seq_file
*m
)
2956 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2957 struct memcg_cache_params
*params
;
2959 if (!memcg_can_account_kmem(memcg
))
2962 print_slabinfo_header(m
);
2964 mutex_lock(&memcg
->slab_caches_mutex
);
2965 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2966 cache_show(memcg_params_to_cache(params
), m
);
2967 mutex_unlock(&memcg
->slab_caches_mutex
);
2973 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2975 struct res_counter
*fail_res
;
2976 struct mem_cgroup
*_memcg
;
2980 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2985 * Conditions under which we can wait for the oom_killer. Those are
2986 * the same conditions tested by the core page allocator
2988 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2991 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2994 if (ret
== -EINTR
) {
2996 * __mem_cgroup_try_charge() chosed to bypass to root due to
2997 * OOM kill or fatal signal. Since our only options are to
2998 * either fail the allocation or charge it to this cgroup, do
2999 * it as a temporary condition. But we can't fail. From a
3000 * kmem/slab perspective, the cache has already been selected,
3001 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3004 * This condition will only trigger if the task entered
3005 * memcg_charge_kmem in a sane state, but was OOM-killed during
3006 * __mem_cgroup_try_charge() above. Tasks that were already
3007 * dying when the allocation triggers should have been already
3008 * directed to the root cgroup in memcontrol.h
3010 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3011 if (do_swap_account
)
3012 res_counter_charge_nofail(&memcg
->memsw
, size
,
3016 res_counter_uncharge(&memcg
->kmem
, size
);
3021 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3023 res_counter_uncharge(&memcg
->res
, size
);
3024 if (do_swap_account
)
3025 res_counter_uncharge(&memcg
->memsw
, size
);
3028 if (res_counter_uncharge(&memcg
->kmem
, size
))
3032 * Releases a reference taken in kmem_cgroup_css_offline in case
3033 * this last uncharge is racing with the offlining code or it is
3034 * outliving the memcg existence.
3036 * The memory barrier imposed by test&clear is paired with the
3037 * explicit one in memcg_kmem_mark_dead().
3039 if (memcg_kmem_test_and_clear_dead(memcg
))
3040 css_put(&memcg
->css
);
3043 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3048 mutex_lock(&memcg
->slab_caches_mutex
);
3049 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3050 mutex_unlock(&memcg
->slab_caches_mutex
);
3054 * helper for acessing a memcg's index. It will be used as an index in the
3055 * child cache array in kmem_cache, and also to derive its name. This function
3056 * will return -1 when this is not a kmem-limited memcg.
3058 int memcg_cache_id(struct mem_cgroup
*memcg
)
3060 return memcg
? memcg
->kmemcg_id
: -1;
3064 * This ends up being protected by the set_limit mutex, during normal
3065 * operation, because that is its main call site.
3067 * But when we create a new cache, we can call this as well if its parent
3068 * is kmem-limited. That will have to hold set_limit_mutex as well.
3070 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3074 num
= ida_simple_get(&kmem_limited_groups
,
3075 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3079 * After this point, kmem_accounted (that we test atomically in
3080 * the beginning of this conditional), is no longer 0. This
3081 * guarantees only one process will set the following boolean
3082 * to true. We don't need test_and_set because we're protected
3083 * by the set_limit_mutex anyway.
3085 memcg_kmem_set_activated(memcg
);
3087 ret
= memcg_update_all_caches(num
+1);
3089 ida_simple_remove(&kmem_limited_groups
, num
);
3090 memcg_kmem_clear_activated(memcg
);
3094 memcg
->kmemcg_id
= num
;
3095 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3096 mutex_init(&memcg
->slab_caches_mutex
);
3100 static size_t memcg_caches_array_size(int num_groups
)
3103 if (num_groups
<= 0)
3106 size
= 2 * num_groups
;
3107 if (size
< MEMCG_CACHES_MIN_SIZE
)
3108 size
= MEMCG_CACHES_MIN_SIZE
;
3109 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3110 size
= MEMCG_CACHES_MAX_SIZE
;
3116 * We should update the current array size iff all caches updates succeed. This
3117 * can only be done from the slab side. The slab mutex needs to be held when
3120 void memcg_update_array_size(int num
)
3122 if (num
> memcg_limited_groups_array_size
)
3123 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3126 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3128 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3130 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3132 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3134 if (num_groups
> memcg_limited_groups_array_size
) {
3136 ssize_t size
= memcg_caches_array_size(num_groups
);
3138 size
*= sizeof(void *);
3139 size
+= sizeof(struct memcg_cache_params
);
3141 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3142 if (!s
->memcg_params
) {
3143 s
->memcg_params
= cur_params
;
3147 s
->memcg_params
->is_root_cache
= true;
3150 * There is the chance it will be bigger than
3151 * memcg_limited_groups_array_size, if we failed an allocation
3152 * in a cache, in which case all caches updated before it, will
3153 * have a bigger array.
3155 * But if that is the case, the data after
3156 * memcg_limited_groups_array_size is certainly unused
3158 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3159 if (!cur_params
->memcg_caches
[i
])
3161 s
->memcg_params
->memcg_caches
[i
] =
3162 cur_params
->memcg_caches
[i
];
3166 * Ideally, we would wait until all caches succeed, and only
3167 * then free the old one. But this is not worth the extra
3168 * pointer per-cache we'd have to have for this.
3170 * It is not a big deal if some caches are left with a size
3171 * bigger than the others. And all updates will reset this
3179 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3180 struct kmem_cache
*root_cache
)
3182 size_t size
= sizeof(struct memcg_cache_params
);
3184 if (!memcg_kmem_enabled())
3188 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3190 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3191 if (!s
->memcg_params
)
3194 INIT_WORK(&s
->memcg_params
->destroy
,
3195 kmem_cache_destroy_work_func
);
3197 s
->memcg_params
->memcg
= memcg
;
3198 s
->memcg_params
->root_cache
= root_cache
;
3200 s
->memcg_params
->is_root_cache
= true;
3205 void memcg_release_cache(struct kmem_cache
*s
)
3207 struct kmem_cache
*root
;
3208 struct mem_cgroup
*memcg
;
3212 * This happens, for instance, when a root cache goes away before we
3215 if (!s
->memcg_params
)
3218 if (s
->memcg_params
->is_root_cache
)
3221 memcg
= s
->memcg_params
->memcg
;
3222 id
= memcg_cache_id(memcg
);
3224 root
= s
->memcg_params
->root_cache
;
3225 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3227 mutex_lock(&memcg
->slab_caches_mutex
);
3228 list_del(&s
->memcg_params
->list
);
3229 mutex_unlock(&memcg
->slab_caches_mutex
);
3231 css_put(&memcg
->css
);
3233 kfree(s
->memcg_params
);
3237 * During the creation a new cache, we need to disable our accounting mechanism
3238 * altogether. This is true even if we are not creating, but rather just
3239 * enqueing new caches to be created.
3241 * This is because that process will trigger allocations; some visible, like
3242 * explicit kmallocs to auxiliary data structures, name strings and internal
3243 * cache structures; some well concealed, like INIT_WORK() that can allocate
3244 * objects during debug.
3246 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3247 * to it. This may not be a bounded recursion: since the first cache creation
3248 * failed to complete (waiting on the allocation), we'll just try to create the
3249 * cache again, failing at the same point.
3251 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3252 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3253 * inside the following two functions.
3255 static inline void memcg_stop_kmem_account(void)
3257 VM_BUG_ON(!current
->mm
);
3258 current
->memcg_kmem_skip_account
++;
3261 static inline void memcg_resume_kmem_account(void)
3263 VM_BUG_ON(!current
->mm
);
3264 current
->memcg_kmem_skip_account
--;
3267 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3269 struct kmem_cache
*cachep
;
3270 struct memcg_cache_params
*p
;
3272 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3274 cachep
= memcg_params_to_cache(p
);
3277 * If we get down to 0 after shrink, we could delete right away.
3278 * However, memcg_release_pages() already puts us back in the workqueue
3279 * in that case. If we proceed deleting, we'll get a dangling
3280 * reference, and removing the object from the workqueue in that case
3281 * is unnecessary complication. We are not a fast path.
3283 * Note that this case is fundamentally different from racing with
3284 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3285 * kmem_cache_shrink, not only we would be reinserting a dead cache
3286 * into the queue, but doing so from inside the worker racing to
3289 * So if we aren't down to zero, we'll just schedule a worker and try
3292 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3293 kmem_cache_shrink(cachep
);
3294 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3297 kmem_cache_destroy(cachep
);
3300 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3302 if (!cachep
->memcg_params
->dead
)
3306 * There are many ways in which we can get here.
3308 * We can get to a memory-pressure situation while the delayed work is
3309 * still pending to run. The vmscan shrinkers can then release all
3310 * cache memory and get us to destruction. If this is the case, we'll
3311 * be executed twice, which is a bug (the second time will execute over
3312 * bogus data). In this case, cancelling the work should be fine.
3314 * But we can also get here from the worker itself, if
3315 * kmem_cache_shrink is enough to shake all the remaining objects and
3316 * get the page count to 0. In this case, we'll deadlock if we try to
3317 * cancel the work (the worker runs with an internal lock held, which
3318 * is the same lock we would hold for cancel_work_sync().)
3320 * Since we can't possibly know who got us here, just refrain from
3321 * running if there is already work pending
3323 if (work_pending(&cachep
->memcg_params
->destroy
))
3326 * We have to defer the actual destroying to a workqueue, because
3327 * we might currently be in a context that cannot sleep.
3329 schedule_work(&cachep
->memcg_params
->destroy
);
3333 * This lock protects updaters, not readers. We want readers to be as fast as
3334 * they can, and they will either see NULL or a valid cache value. Our model
3335 * allow them to see NULL, in which case the root memcg will be selected.
3337 * We need this lock because multiple allocations to the same cache from a non
3338 * will span more than one worker. Only one of them can create the cache.
3340 static DEFINE_MUTEX(memcg_cache_mutex
);
3343 * Called with memcg_cache_mutex held
3345 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3346 struct kmem_cache
*s
)
3348 struct kmem_cache
*new;
3349 static char *tmp_name
= NULL
;
3351 lockdep_assert_held(&memcg_cache_mutex
);
3354 * kmem_cache_create_memcg duplicates the given name and
3355 * cgroup_name for this name requires RCU context.
3356 * This static temporary buffer is used to prevent from
3357 * pointless shortliving allocation.
3360 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3366 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3367 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3370 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3371 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3374 new->allocflags
|= __GFP_KMEMCG
;
3379 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3380 struct kmem_cache
*cachep
)
3382 struct kmem_cache
*new_cachep
;
3385 BUG_ON(!memcg_can_account_kmem(memcg
));
3387 idx
= memcg_cache_id(memcg
);
3389 mutex_lock(&memcg_cache_mutex
);
3390 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3392 css_put(&memcg
->css
);
3396 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3397 if (new_cachep
== NULL
) {
3398 new_cachep
= cachep
;
3399 css_put(&memcg
->css
);
3403 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3405 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3407 * the readers won't lock, make sure everybody sees the updated value,
3408 * so they won't put stuff in the queue again for no reason
3412 mutex_unlock(&memcg_cache_mutex
);
3416 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3418 struct kmem_cache
*c
;
3421 if (!s
->memcg_params
)
3423 if (!s
->memcg_params
->is_root_cache
)
3427 * If the cache is being destroyed, we trust that there is no one else
3428 * requesting objects from it. Even if there are, the sanity checks in
3429 * kmem_cache_destroy should caught this ill-case.
3431 * Still, we don't want anyone else freeing memcg_caches under our
3432 * noses, which can happen if a new memcg comes to life. As usual,
3433 * we'll take the set_limit_mutex to protect ourselves against this.
3435 mutex_lock(&set_limit_mutex
);
3436 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3437 c
= s
->memcg_params
->memcg_caches
[i
];
3442 * We will now manually delete the caches, so to avoid races
3443 * we need to cancel all pending destruction workers and
3444 * proceed with destruction ourselves.
3446 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3447 * and that could spawn the workers again: it is likely that
3448 * the cache still have active pages until this very moment.
3449 * This would lead us back to mem_cgroup_destroy_cache.
3451 * But that will not execute at all if the "dead" flag is not
3452 * set, so flip it down to guarantee we are in control.
3454 c
->memcg_params
->dead
= false;
3455 cancel_work_sync(&c
->memcg_params
->destroy
);
3456 kmem_cache_destroy(c
);
3458 mutex_unlock(&set_limit_mutex
);
3461 struct create_work
{
3462 struct mem_cgroup
*memcg
;
3463 struct kmem_cache
*cachep
;
3464 struct work_struct work
;
3467 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3469 struct kmem_cache
*cachep
;
3470 struct memcg_cache_params
*params
;
3472 if (!memcg_kmem_is_active(memcg
))
3475 mutex_lock(&memcg
->slab_caches_mutex
);
3476 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3477 cachep
= memcg_params_to_cache(params
);
3478 cachep
->memcg_params
->dead
= true;
3479 schedule_work(&cachep
->memcg_params
->destroy
);
3481 mutex_unlock(&memcg
->slab_caches_mutex
);
3484 static void memcg_create_cache_work_func(struct work_struct
*w
)
3486 struct create_work
*cw
;
3488 cw
= container_of(w
, struct create_work
, work
);
3489 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3494 * Enqueue the creation of a per-memcg kmem_cache.
3496 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3497 struct kmem_cache
*cachep
)
3499 struct create_work
*cw
;
3501 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3503 css_put(&memcg
->css
);
3508 cw
->cachep
= cachep
;
3510 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3511 schedule_work(&cw
->work
);
3514 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3515 struct kmem_cache
*cachep
)
3518 * We need to stop accounting when we kmalloc, because if the
3519 * corresponding kmalloc cache is not yet created, the first allocation
3520 * in __memcg_create_cache_enqueue will recurse.
3522 * However, it is better to enclose the whole function. Depending on
3523 * the debugging options enabled, INIT_WORK(), for instance, can
3524 * trigger an allocation. This too, will make us recurse. Because at
3525 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3526 * the safest choice is to do it like this, wrapping the whole function.
3528 memcg_stop_kmem_account();
3529 __memcg_create_cache_enqueue(memcg
, cachep
);
3530 memcg_resume_kmem_account();
3533 * Return the kmem_cache we're supposed to use for a slab allocation.
3534 * We try to use the current memcg's version of the cache.
3536 * If the cache does not exist yet, if we are the first user of it,
3537 * we either create it immediately, if possible, or create it asynchronously
3539 * In the latter case, we will let the current allocation go through with
3540 * the original cache.
3542 * Can't be called in interrupt context or from kernel threads.
3543 * This function needs to be called with rcu_read_lock() held.
3545 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3548 struct mem_cgroup
*memcg
;
3551 VM_BUG_ON(!cachep
->memcg_params
);
3552 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3554 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3558 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3560 if (!memcg_can_account_kmem(memcg
))
3563 idx
= memcg_cache_id(memcg
);
3566 * barrier to mare sure we're always seeing the up to date value. The
3567 * code updating memcg_caches will issue a write barrier to match this.
3569 read_barrier_depends();
3570 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3571 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3575 /* The corresponding put will be done in the workqueue. */
3576 if (!css_tryget(&memcg
->css
))
3581 * If we are in a safe context (can wait, and not in interrupt
3582 * context), we could be be predictable and return right away.
3583 * This would guarantee that the allocation being performed
3584 * already belongs in the new cache.
3586 * However, there are some clashes that can arrive from locking.
3587 * For instance, because we acquire the slab_mutex while doing
3588 * kmem_cache_dup, this means no further allocation could happen
3589 * with the slab_mutex held.
3591 * Also, because cache creation issue get_online_cpus(), this
3592 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3593 * that ends up reversed during cpu hotplug. (cpuset allocates
3594 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3595 * better to defer everything.
3597 memcg_create_cache_enqueue(memcg
, cachep
);
3603 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3606 * We need to verify if the allocation against current->mm->owner's memcg is
3607 * possible for the given order. But the page is not allocated yet, so we'll
3608 * need a further commit step to do the final arrangements.
3610 * It is possible for the task to switch cgroups in this mean time, so at
3611 * commit time, we can't rely on task conversion any longer. We'll then use
3612 * the handle argument to return to the caller which cgroup we should commit
3613 * against. We could also return the memcg directly and avoid the pointer
3614 * passing, but a boolean return value gives better semantics considering
3615 * the compiled-out case as well.
3617 * Returning true means the allocation is possible.
3620 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3622 struct mem_cgroup
*memcg
;
3628 * Disabling accounting is only relevant for some specific memcg
3629 * internal allocations. Therefore we would initially not have such
3630 * check here, since direct calls to the page allocator that are marked
3631 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3632 * concerned with cache allocations, and by having this test at
3633 * memcg_kmem_get_cache, we are already able to relay the allocation to
3634 * the root cache and bypass the memcg cache altogether.
3636 * There is one exception, though: the SLUB allocator does not create
3637 * large order caches, but rather service large kmallocs directly from
3638 * the page allocator. Therefore, the following sequence when backed by
3639 * the SLUB allocator:
3641 * memcg_stop_kmem_account();
3642 * kmalloc(<large_number>)
3643 * memcg_resume_kmem_account();
3645 * would effectively ignore the fact that we should skip accounting,
3646 * since it will drive us directly to this function without passing
3647 * through the cache selector memcg_kmem_get_cache. Such large
3648 * allocations are extremely rare but can happen, for instance, for the
3649 * cache arrays. We bring this test here.
3651 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3654 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3657 * very rare case described in mem_cgroup_from_task. Unfortunately there
3658 * isn't much we can do without complicating this too much, and it would
3659 * be gfp-dependent anyway. Just let it go
3661 if (unlikely(!memcg
))
3664 if (!memcg_can_account_kmem(memcg
)) {
3665 css_put(&memcg
->css
);
3669 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3673 css_put(&memcg
->css
);
3677 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3680 struct page_cgroup
*pc
;
3682 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3684 /* The page allocation failed. Revert */
3686 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3690 pc
= lookup_page_cgroup(page
);
3691 lock_page_cgroup(pc
);
3692 pc
->mem_cgroup
= memcg
;
3693 SetPageCgroupUsed(pc
);
3694 unlock_page_cgroup(pc
);
3697 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3699 struct mem_cgroup
*memcg
= NULL
;
3700 struct page_cgroup
*pc
;
3703 pc
= lookup_page_cgroup(page
);
3705 * Fast unlocked return. Theoretically might have changed, have to
3706 * check again after locking.
3708 if (!PageCgroupUsed(pc
))
3711 lock_page_cgroup(pc
);
3712 if (PageCgroupUsed(pc
)) {
3713 memcg
= pc
->mem_cgroup
;
3714 ClearPageCgroupUsed(pc
);
3716 unlock_page_cgroup(pc
);
3719 * We trust that only if there is a memcg associated with the page, it
3720 * is a valid allocation
3725 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3726 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3729 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3732 #endif /* CONFIG_MEMCG_KMEM */
3734 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3736 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3738 * Because tail pages are not marked as "used", set it. We're under
3739 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3740 * charge/uncharge will be never happen and move_account() is done under
3741 * compound_lock(), so we don't have to take care of races.
3743 void mem_cgroup_split_huge_fixup(struct page
*head
)
3745 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3746 struct page_cgroup
*pc
;
3747 struct mem_cgroup
*memcg
;
3750 if (mem_cgroup_disabled())
3753 memcg
= head_pc
->mem_cgroup
;
3754 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3756 pc
->mem_cgroup
= memcg
;
3757 smp_wmb();/* see __commit_charge() */
3758 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3760 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3763 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3766 * mem_cgroup_move_account - move account of the page
3768 * @nr_pages: number of regular pages (>1 for huge pages)
3769 * @pc: page_cgroup of the page.
3770 * @from: mem_cgroup which the page is moved from.
3771 * @to: mem_cgroup which the page is moved to. @from != @to.
3773 * The caller must confirm following.
3774 * - page is not on LRU (isolate_page() is useful.)
3775 * - compound_lock is held when nr_pages > 1
3777 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3780 static int mem_cgroup_move_account(struct page
*page
,
3781 unsigned int nr_pages
,
3782 struct page_cgroup
*pc
,
3783 struct mem_cgroup
*from
,
3784 struct mem_cgroup
*to
)
3786 unsigned long flags
;
3788 bool anon
= PageAnon(page
);
3790 VM_BUG_ON(from
== to
);
3791 VM_BUG_ON(PageLRU(page
));
3793 * The page is isolated from LRU. So, collapse function
3794 * will not handle this page. But page splitting can happen.
3795 * Do this check under compound_page_lock(). The caller should
3799 if (nr_pages
> 1 && !PageTransHuge(page
))
3802 lock_page_cgroup(pc
);
3805 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3808 move_lock_mem_cgroup(from
, &flags
);
3810 if (!anon
&& page_mapped(page
)) {
3811 /* Update mapped_file data for mem_cgroup */
3813 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3814 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3817 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3819 /* caller should have done css_get */
3820 pc
->mem_cgroup
= to
;
3821 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3822 move_unlock_mem_cgroup(from
, &flags
);
3825 unlock_page_cgroup(pc
);
3829 memcg_check_events(to
, page
);
3830 memcg_check_events(from
, page
);
3836 * mem_cgroup_move_parent - moves page to the parent group
3837 * @page: the page to move
3838 * @pc: page_cgroup of the page
3839 * @child: page's cgroup
3841 * move charges to its parent or the root cgroup if the group has no
3842 * parent (aka use_hierarchy==0).
3843 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3844 * mem_cgroup_move_account fails) the failure is always temporary and
3845 * it signals a race with a page removal/uncharge or migration. In the
3846 * first case the page is on the way out and it will vanish from the LRU
3847 * on the next attempt and the call should be retried later.
3848 * Isolation from the LRU fails only if page has been isolated from
3849 * the LRU since we looked at it and that usually means either global
3850 * reclaim or migration going on. The page will either get back to the
3852 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3853 * (!PageCgroupUsed) or moved to a different group. The page will
3854 * disappear in the next attempt.
3856 static int mem_cgroup_move_parent(struct page
*page
,
3857 struct page_cgroup
*pc
,
3858 struct mem_cgroup
*child
)
3860 struct mem_cgroup
*parent
;
3861 unsigned int nr_pages
;
3862 unsigned long uninitialized_var(flags
);
3865 VM_BUG_ON(mem_cgroup_is_root(child
));
3868 if (!get_page_unless_zero(page
))
3870 if (isolate_lru_page(page
))
3873 nr_pages
= hpage_nr_pages(page
);
3875 parent
= parent_mem_cgroup(child
);
3877 * If no parent, move charges to root cgroup.
3880 parent
= root_mem_cgroup
;
3883 VM_BUG_ON(!PageTransHuge(page
));
3884 flags
= compound_lock_irqsave(page
);
3887 ret
= mem_cgroup_move_account(page
, nr_pages
,
3890 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3893 compound_unlock_irqrestore(page
, flags
);
3894 putback_lru_page(page
);
3902 * Charge the memory controller for page usage.
3904 * 0 if the charge was successful
3905 * < 0 if the cgroup is over its limit
3907 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3908 gfp_t gfp_mask
, enum charge_type ctype
)
3910 struct mem_cgroup
*memcg
= NULL
;
3911 unsigned int nr_pages
= 1;
3915 if (PageTransHuge(page
)) {
3916 nr_pages
<<= compound_order(page
);
3917 VM_BUG_ON(!PageTransHuge(page
));
3919 * Never OOM-kill a process for a huge page. The
3920 * fault handler will fall back to regular pages.
3925 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3928 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3932 int mem_cgroup_newpage_charge(struct page
*page
,
3933 struct mm_struct
*mm
, gfp_t gfp_mask
)
3935 if (mem_cgroup_disabled())
3937 VM_BUG_ON(page_mapped(page
));
3938 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3940 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3941 MEM_CGROUP_CHARGE_TYPE_ANON
);
3945 * While swap-in, try_charge -> commit or cancel, the page is locked.
3946 * And when try_charge() successfully returns, one refcnt to memcg without
3947 * struct page_cgroup is acquired. This refcnt will be consumed by
3948 * "commit()" or removed by "cancel()"
3950 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3953 struct mem_cgroup
**memcgp
)
3955 struct mem_cgroup
*memcg
;
3956 struct page_cgroup
*pc
;
3959 pc
= lookup_page_cgroup(page
);
3961 * Every swap fault against a single page tries to charge the
3962 * page, bail as early as possible. shmem_unuse() encounters
3963 * already charged pages, too. The USED bit is protected by
3964 * the page lock, which serializes swap cache removal, which
3965 * in turn serializes uncharging.
3967 if (PageCgroupUsed(pc
))
3969 if (!do_swap_account
)
3971 memcg
= try_get_mem_cgroup_from_page(page
);
3975 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3976 css_put(&memcg
->css
);
3981 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3987 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3988 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3991 if (mem_cgroup_disabled())
3994 * A racing thread's fault, or swapoff, may have already
3995 * updated the pte, and even removed page from swap cache: in
3996 * those cases unuse_pte()'s pte_same() test will fail; but
3997 * there's also a KSM case which does need to charge the page.
3999 if (!PageSwapCache(page
)) {
4002 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4007 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4010 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4012 if (mem_cgroup_disabled())
4016 __mem_cgroup_cancel_charge(memcg
, 1);
4020 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4021 enum charge_type ctype
)
4023 if (mem_cgroup_disabled())
4028 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4030 * Now swap is on-memory. This means this page may be
4031 * counted both as mem and swap....double count.
4032 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4033 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4034 * may call delete_from_swap_cache() before reach here.
4036 if (do_swap_account
&& PageSwapCache(page
)) {
4037 swp_entry_t ent
= {.val
= page_private(page
)};
4038 mem_cgroup_uncharge_swap(ent
);
4042 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4043 struct mem_cgroup
*memcg
)
4045 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4046 MEM_CGROUP_CHARGE_TYPE_ANON
);
4049 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4052 struct mem_cgroup
*memcg
= NULL
;
4053 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4056 if (mem_cgroup_disabled())
4058 if (PageCompound(page
))
4061 if (!PageSwapCache(page
))
4062 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4063 else { /* page is swapcache/shmem */
4064 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4067 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4072 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4073 unsigned int nr_pages
,
4074 const enum charge_type ctype
)
4076 struct memcg_batch_info
*batch
= NULL
;
4077 bool uncharge_memsw
= true;
4079 /* If swapout, usage of swap doesn't decrease */
4080 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4081 uncharge_memsw
= false;
4083 batch
= ¤t
->memcg_batch
;
4085 * In usual, we do css_get() when we remember memcg pointer.
4086 * But in this case, we keep res->usage until end of a series of
4087 * uncharges. Then, it's ok to ignore memcg's refcnt.
4090 batch
->memcg
= memcg
;
4092 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4093 * In those cases, all pages freed continuously can be expected to be in
4094 * the same cgroup and we have chance to coalesce uncharges.
4095 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4096 * because we want to do uncharge as soon as possible.
4099 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4100 goto direct_uncharge
;
4103 goto direct_uncharge
;
4106 * In typical case, batch->memcg == mem. This means we can
4107 * merge a series of uncharges to an uncharge of res_counter.
4108 * If not, we uncharge res_counter ony by one.
4110 if (batch
->memcg
!= memcg
)
4111 goto direct_uncharge
;
4112 /* remember freed charge and uncharge it later */
4115 batch
->memsw_nr_pages
++;
4118 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4120 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4121 if (unlikely(batch
->memcg
!= memcg
))
4122 memcg_oom_recover(memcg
);
4126 * uncharge if !page_mapped(page)
4128 static struct mem_cgroup
*
4129 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4132 struct mem_cgroup
*memcg
= NULL
;
4133 unsigned int nr_pages
= 1;
4134 struct page_cgroup
*pc
;
4137 if (mem_cgroup_disabled())
4140 if (PageTransHuge(page
)) {
4141 nr_pages
<<= compound_order(page
);
4142 VM_BUG_ON(!PageTransHuge(page
));
4145 * Check if our page_cgroup is valid
4147 pc
= lookup_page_cgroup(page
);
4148 if (unlikely(!PageCgroupUsed(pc
)))
4151 lock_page_cgroup(pc
);
4153 memcg
= pc
->mem_cgroup
;
4155 if (!PageCgroupUsed(pc
))
4158 anon
= PageAnon(page
);
4161 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4163 * Generally PageAnon tells if it's the anon statistics to be
4164 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4165 * used before page reached the stage of being marked PageAnon.
4169 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4170 /* See mem_cgroup_prepare_migration() */
4171 if (page_mapped(page
))
4174 * Pages under migration may not be uncharged. But
4175 * end_migration() /must/ be the one uncharging the
4176 * unused post-migration page and so it has to call
4177 * here with the migration bit still set. See the
4178 * res_counter handling below.
4180 if (!end_migration
&& PageCgroupMigration(pc
))
4183 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4184 if (!PageAnon(page
)) { /* Shared memory */
4185 if (page
->mapping
&& !page_is_file_cache(page
))
4187 } else if (page_mapped(page
)) /* Anon */
4194 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4196 ClearPageCgroupUsed(pc
);
4198 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4199 * freed from LRU. This is safe because uncharged page is expected not
4200 * to be reused (freed soon). Exception is SwapCache, it's handled by
4201 * special functions.
4204 unlock_page_cgroup(pc
);
4206 * even after unlock, we have memcg->res.usage here and this memcg
4207 * will never be freed, so it's safe to call css_get().
4209 memcg_check_events(memcg
, page
);
4210 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4211 mem_cgroup_swap_statistics(memcg
, true);
4212 css_get(&memcg
->css
);
4215 * Migration does not charge the res_counter for the
4216 * replacement page, so leave it alone when phasing out the
4217 * page that is unused after the migration.
4219 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4220 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4225 unlock_page_cgroup(pc
);
4229 void mem_cgroup_uncharge_page(struct page
*page
)
4232 if (page_mapped(page
))
4234 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4236 * If the page is in swap cache, uncharge should be deferred
4237 * to the swap path, which also properly accounts swap usage
4238 * and handles memcg lifetime.
4240 * Note that this check is not stable and reclaim may add the
4241 * page to swap cache at any time after this. However, if the
4242 * page is not in swap cache by the time page->mapcount hits
4243 * 0, there won't be any page table references to the swap
4244 * slot, and reclaim will free it and not actually write the
4247 if (PageSwapCache(page
))
4249 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4252 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4254 VM_BUG_ON(page_mapped(page
));
4255 VM_BUG_ON(page
->mapping
);
4256 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4260 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4261 * In that cases, pages are freed continuously and we can expect pages
4262 * are in the same memcg. All these calls itself limits the number of
4263 * pages freed at once, then uncharge_start/end() is called properly.
4264 * This may be called prural(2) times in a context,
4267 void mem_cgroup_uncharge_start(void)
4269 current
->memcg_batch
.do_batch
++;
4270 /* We can do nest. */
4271 if (current
->memcg_batch
.do_batch
== 1) {
4272 current
->memcg_batch
.memcg
= NULL
;
4273 current
->memcg_batch
.nr_pages
= 0;
4274 current
->memcg_batch
.memsw_nr_pages
= 0;
4278 void mem_cgroup_uncharge_end(void)
4280 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4282 if (!batch
->do_batch
)
4286 if (batch
->do_batch
) /* If stacked, do nothing. */
4292 * This "batch->memcg" is valid without any css_get/put etc...
4293 * bacause we hide charges behind us.
4295 if (batch
->nr_pages
)
4296 res_counter_uncharge(&batch
->memcg
->res
,
4297 batch
->nr_pages
* PAGE_SIZE
);
4298 if (batch
->memsw_nr_pages
)
4299 res_counter_uncharge(&batch
->memcg
->memsw
,
4300 batch
->memsw_nr_pages
* PAGE_SIZE
);
4301 memcg_oom_recover(batch
->memcg
);
4302 /* forget this pointer (for sanity check) */
4303 batch
->memcg
= NULL
;
4308 * called after __delete_from_swap_cache() and drop "page" account.
4309 * memcg information is recorded to swap_cgroup of "ent"
4312 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4314 struct mem_cgroup
*memcg
;
4315 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4317 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4318 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4320 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4323 * record memcg information, if swapout && memcg != NULL,
4324 * css_get() was called in uncharge().
4326 if (do_swap_account
&& swapout
&& memcg
)
4327 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4331 #ifdef CONFIG_MEMCG_SWAP
4333 * called from swap_entry_free(). remove record in swap_cgroup and
4334 * uncharge "memsw" account.
4336 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4338 struct mem_cgroup
*memcg
;
4341 if (!do_swap_account
)
4344 id
= swap_cgroup_record(ent
, 0);
4346 memcg
= mem_cgroup_lookup(id
);
4349 * We uncharge this because swap is freed.
4350 * This memcg can be obsolete one. We avoid calling css_tryget
4352 if (!mem_cgroup_is_root(memcg
))
4353 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4354 mem_cgroup_swap_statistics(memcg
, false);
4355 css_put(&memcg
->css
);
4361 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4362 * @entry: swap entry to be moved
4363 * @from: mem_cgroup which the entry is moved from
4364 * @to: mem_cgroup which the entry is moved to
4366 * It succeeds only when the swap_cgroup's record for this entry is the same
4367 * as the mem_cgroup's id of @from.
4369 * Returns 0 on success, -EINVAL on failure.
4371 * The caller must have charged to @to, IOW, called res_counter_charge() about
4372 * both res and memsw, and called css_get().
4374 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4375 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4377 unsigned short old_id
, new_id
;
4379 old_id
= css_id(&from
->css
);
4380 new_id
= css_id(&to
->css
);
4382 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4383 mem_cgroup_swap_statistics(from
, false);
4384 mem_cgroup_swap_statistics(to
, true);
4386 * This function is only called from task migration context now.
4387 * It postpones res_counter and refcount handling till the end
4388 * of task migration(mem_cgroup_clear_mc()) for performance
4389 * improvement. But we cannot postpone css_get(to) because if
4390 * the process that has been moved to @to does swap-in, the
4391 * refcount of @to might be decreased to 0.
4393 * We are in attach() phase, so the cgroup is guaranteed to be
4394 * alive, so we can just call css_get().
4402 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4403 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4410 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4413 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4414 struct mem_cgroup
**memcgp
)
4416 struct mem_cgroup
*memcg
= NULL
;
4417 unsigned int nr_pages
= 1;
4418 struct page_cgroup
*pc
;
4419 enum charge_type ctype
;
4423 if (mem_cgroup_disabled())
4426 if (PageTransHuge(page
))
4427 nr_pages
<<= compound_order(page
);
4429 pc
= lookup_page_cgroup(page
);
4430 lock_page_cgroup(pc
);
4431 if (PageCgroupUsed(pc
)) {
4432 memcg
= pc
->mem_cgroup
;
4433 css_get(&memcg
->css
);
4435 * At migrating an anonymous page, its mapcount goes down
4436 * to 0 and uncharge() will be called. But, even if it's fully
4437 * unmapped, migration may fail and this page has to be
4438 * charged again. We set MIGRATION flag here and delay uncharge
4439 * until end_migration() is called
4441 * Corner Case Thinking
4443 * When the old page was mapped as Anon and it's unmap-and-freed
4444 * while migration was ongoing.
4445 * If unmap finds the old page, uncharge() of it will be delayed
4446 * until end_migration(). If unmap finds a new page, it's
4447 * uncharged when it make mapcount to be 1->0. If unmap code
4448 * finds swap_migration_entry, the new page will not be mapped
4449 * and end_migration() will find it(mapcount==0).
4452 * When the old page was mapped but migraion fails, the kernel
4453 * remaps it. A charge for it is kept by MIGRATION flag even
4454 * if mapcount goes down to 0. We can do remap successfully
4455 * without charging it again.
4458 * The "old" page is under lock_page() until the end of
4459 * migration, so, the old page itself will not be swapped-out.
4460 * If the new page is swapped out before end_migraton, our
4461 * hook to usual swap-out path will catch the event.
4464 SetPageCgroupMigration(pc
);
4466 unlock_page_cgroup(pc
);
4468 * If the page is not charged at this point,
4476 * We charge new page before it's used/mapped. So, even if unlock_page()
4477 * is called before end_migration, we can catch all events on this new
4478 * page. In the case new page is migrated but not remapped, new page's
4479 * mapcount will be finally 0 and we call uncharge in end_migration().
4482 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4484 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4486 * The page is committed to the memcg, but it's not actually
4487 * charged to the res_counter since we plan on replacing the
4488 * old one and only one page is going to be left afterwards.
4490 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4493 /* remove redundant charge if migration failed*/
4494 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4495 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4497 struct page
*used
, *unused
;
4498 struct page_cgroup
*pc
;
4504 if (!migration_ok
) {
4511 anon
= PageAnon(used
);
4512 __mem_cgroup_uncharge_common(unused
,
4513 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4514 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4516 css_put(&memcg
->css
);
4518 * We disallowed uncharge of pages under migration because mapcount
4519 * of the page goes down to zero, temporarly.
4520 * Clear the flag and check the page should be charged.
4522 pc
= lookup_page_cgroup(oldpage
);
4523 lock_page_cgroup(pc
);
4524 ClearPageCgroupMigration(pc
);
4525 unlock_page_cgroup(pc
);
4528 * If a page is a file cache, radix-tree replacement is very atomic
4529 * and we can skip this check. When it was an Anon page, its mapcount
4530 * goes down to 0. But because we added MIGRATION flage, it's not
4531 * uncharged yet. There are several case but page->mapcount check
4532 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4533 * check. (see prepare_charge() also)
4536 mem_cgroup_uncharge_page(used
);
4540 * At replace page cache, newpage is not under any memcg but it's on
4541 * LRU. So, this function doesn't touch res_counter but handles LRU
4542 * in correct way. Both pages are locked so we cannot race with uncharge.
4544 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4545 struct page
*newpage
)
4547 struct mem_cgroup
*memcg
= NULL
;
4548 struct page_cgroup
*pc
;
4549 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4551 if (mem_cgroup_disabled())
4554 pc
= lookup_page_cgroup(oldpage
);
4555 /* fix accounting on old pages */
4556 lock_page_cgroup(pc
);
4557 if (PageCgroupUsed(pc
)) {
4558 memcg
= pc
->mem_cgroup
;
4559 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4560 ClearPageCgroupUsed(pc
);
4562 unlock_page_cgroup(pc
);
4565 * When called from shmem_replace_page(), in some cases the
4566 * oldpage has already been charged, and in some cases not.
4571 * Even if newpage->mapping was NULL before starting replacement,
4572 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4573 * LRU while we overwrite pc->mem_cgroup.
4575 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4578 #ifdef CONFIG_DEBUG_VM
4579 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4581 struct page_cgroup
*pc
;
4583 pc
= lookup_page_cgroup(page
);
4585 * Can be NULL while feeding pages into the page allocator for
4586 * the first time, i.e. during boot or memory hotplug;
4587 * or when mem_cgroup_disabled().
4589 if (likely(pc
) && PageCgroupUsed(pc
))
4594 bool mem_cgroup_bad_page_check(struct page
*page
)
4596 if (mem_cgroup_disabled())
4599 return lookup_page_cgroup_used(page
) != NULL
;
4602 void mem_cgroup_print_bad_page(struct page
*page
)
4604 struct page_cgroup
*pc
;
4606 pc
= lookup_page_cgroup_used(page
);
4608 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4609 pc
, pc
->flags
, pc
->mem_cgroup
);
4614 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4615 unsigned long long val
)
4618 u64 memswlimit
, memlimit
;
4620 int children
= mem_cgroup_count_children(memcg
);
4621 u64 curusage
, oldusage
;
4625 * For keeping hierarchical_reclaim simple, how long we should retry
4626 * is depends on callers. We set our retry-count to be function
4627 * of # of children which we should visit in this loop.
4629 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4631 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4634 while (retry_count
) {
4635 if (signal_pending(current
)) {
4640 * Rather than hide all in some function, I do this in
4641 * open coded manner. You see what this really does.
4642 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4644 mutex_lock(&set_limit_mutex
);
4645 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4646 if (memswlimit
< val
) {
4648 mutex_unlock(&set_limit_mutex
);
4652 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4656 ret
= res_counter_set_limit(&memcg
->res
, val
);
4658 if (memswlimit
== val
)
4659 memcg
->memsw_is_minimum
= true;
4661 memcg
->memsw_is_minimum
= false;
4663 mutex_unlock(&set_limit_mutex
);
4668 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4669 MEM_CGROUP_RECLAIM_SHRINK
);
4670 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4671 /* Usage is reduced ? */
4672 if (curusage
>= oldusage
)
4675 oldusage
= curusage
;
4677 if (!ret
&& enlarge
)
4678 memcg_oom_recover(memcg
);
4683 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4684 unsigned long long val
)
4687 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4688 int children
= mem_cgroup_count_children(memcg
);
4692 /* see mem_cgroup_resize_res_limit */
4693 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4694 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4695 while (retry_count
) {
4696 if (signal_pending(current
)) {
4701 * Rather than hide all in some function, I do this in
4702 * open coded manner. You see what this really does.
4703 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4705 mutex_lock(&set_limit_mutex
);
4706 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4707 if (memlimit
> val
) {
4709 mutex_unlock(&set_limit_mutex
);
4712 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4713 if (memswlimit
< val
)
4715 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4717 if (memlimit
== val
)
4718 memcg
->memsw_is_minimum
= true;
4720 memcg
->memsw_is_minimum
= false;
4722 mutex_unlock(&set_limit_mutex
);
4727 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4728 MEM_CGROUP_RECLAIM_NOSWAP
|
4729 MEM_CGROUP_RECLAIM_SHRINK
);
4730 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4731 /* Usage is reduced ? */
4732 if (curusage
>= oldusage
)
4735 oldusage
= curusage
;
4737 if (!ret
&& enlarge
)
4738 memcg_oom_recover(memcg
);
4742 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4744 unsigned long *total_scanned
)
4746 unsigned long nr_reclaimed
= 0;
4747 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4748 unsigned long reclaimed
;
4750 struct mem_cgroup_tree_per_zone
*mctz
;
4751 unsigned long long excess
;
4752 unsigned long nr_scanned
;
4757 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4759 * This loop can run a while, specially if mem_cgroup's continuously
4760 * keep exceeding their soft limit and putting the system under
4767 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4772 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4773 gfp_mask
, &nr_scanned
);
4774 nr_reclaimed
+= reclaimed
;
4775 *total_scanned
+= nr_scanned
;
4776 spin_lock(&mctz
->lock
);
4779 * If we failed to reclaim anything from this memory cgroup
4780 * it is time to move on to the next cgroup
4786 * Loop until we find yet another one.
4788 * By the time we get the soft_limit lock
4789 * again, someone might have aded the
4790 * group back on the RB tree. Iterate to
4791 * make sure we get a different mem.
4792 * mem_cgroup_largest_soft_limit_node returns
4793 * NULL if no other cgroup is present on
4797 __mem_cgroup_largest_soft_limit_node(mctz
);
4799 css_put(&next_mz
->memcg
->css
);
4800 else /* next_mz == NULL or other memcg */
4804 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4805 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4807 * One school of thought says that we should not add
4808 * back the node to the tree if reclaim returns 0.
4809 * But our reclaim could return 0, simply because due
4810 * to priority we are exposing a smaller subset of
4811 * memory to reclaim from. Consider this as a longer
4814 /* If excess == 0, no tree ops */
4815 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4816 spin_unlock(&mctz
->lock
);
4817 css_put(&mz
->memcg
->css
);
4820 * Could not reclaim anything and there are no more
4821 * mem cgroups to try or we seem to be looping without
4822 * reclaiming anything.
4824 if (!nr_reclaimed
&&
4826 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4828 } while (!nr_reclaimed
);
4830 css_put(&next_mz
->memcg
->css
);
4831 return nr_reclaimed
;
4835 * mem_cgroup_force_empty_list - clears LRU of a group
4836 * @memcg: group to clear
4839 * @lru: lru to to clear
4841 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4842 * reclaim the pages page themselves - pages are moved to the parent (or root)
4845 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4846 int node
, int zid
, enum lru_list lru
)
4848 struct lruvec
*lruvec
;
4849 unsigned long flags
;
4850 struct list_head
*list
;
4854 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4855 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4856 list
= &lruvec
->lists
[lru
];
4860 struct page_cgroup
*pc
;
4863 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4864 if (list_empty(list
)) {
4865 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4868 page
= list_entry(list
->prev
, struct page
, lru
);
4870 list_move(&page
->lru
, list
);
4872 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4875 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4877 pc
= lookup_page_cgroup(page
);
4879 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4880 /* found lock contention or "pc" is obsolete. */
4885 } while (!list_empty(list
));
4889 * make mem_cgroup's charge to be 0 if there is no task by moving
4890 * all the charges and pages to the parent.
4891 * This enables deleting this mem_cgroup.
4893 * Caller is responsible for holding css reference on the memcg.
4895 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4901 /* This is for making all *used* pages to be on LRU. */
4902 lru_add_drain_all();
4903 drain_all_stock_sync(memcg
);
4904 mem_cgroup_start_move(memcg
);
4905 for_each_node_state(node
, N_MEMORY
) {
4906 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4909 mem_cgroup_force_empty_list(memcg
,
4914 mem_cgroup_end_move(memcg
);
4915 memcg_oom_recover(memcg
);
4919 * Kernel memory may not necessarily be trackable to a specific
4920 * process. So they are not migrated, and therefore we can't
4921 * expect their value to drop to 0 here.
4922 * Having res filled up with kmem only is enough.
4924 * This is a safety check because mem_cgroup_force_empty_list
4925 * could have raced with mem_cgroup_replace_page_cache callers
4926 * so the lru seemed empty but the page could have been added
4927 * right after the check. RES_USAGE should be safe as we always
4928 * charge before adding to the LRU.
4930 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4931 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4932 } while (usage
> 0);
4936 * This mainly exists for tests during the setting of set of use_hierarchy.
4937 * Since this is the very setting we are changing, the current hierarchy value
4940 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4944 /* bounce at first found */
4945 cgroup_for_each_child(pos
, memcg
->css
.cgroup
)
4951 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4952 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4953 * from mem_cgroup_count_children(), in the sense that we don't really care how
4954 * many children we have; we only need to know if we have any. It also counts
4955 * any memcg without hierarchy as infertile.
4957 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4959 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4963 * Reclaims as many pages from the given memcg as possible and moves
4964 * the rest to the parent.
4966 * Caller is responsible for holding css reference for memcg.
4968 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4970 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4971 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4973 /* returns EBUSY if there is a task or if we come here twice. */
4974 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4977 /* we call try-to-free pages for make this cgroup empty */
4978 lru_add_drain_all();
4979 /* try to free all pages in this cgroup */
4980 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4983 if (signal_pending(current
))
4986 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4990 /* maybe some writeback is necessary */
4991 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4996 mem_cgroup_reparent_charges(memcg
);
5001 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5004 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5007 if (mem_cgroup_is_root(memcg
))
5009 css_get(&memcg
->css
);
5010 ret
= mem_cgroup_force_empty(memcg
);
5011 css_put(&memcg
->css
);
5017 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5020 return mem_cgroup_from_css(css
)->use_hierarchy
;
5023 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5024 struct cftype
*cft
, u64 val
)
5027 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5028 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5030 mutex_lock(&memcg_create_mutex
);
5032 if (memcg
->use_hierarchy
== val
)
5036 * If parent's use_hierarchy is set, we can't make any modifications
5037 * in the child subtrees. If it is unset, then the change can
5038 * occur, provided the current cgroup has no children.
5040 * For the root cgroup, parent_mem is NULL, we allow value to be
5041 * set if there are no children.
5043 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5044 (val
== 1 || val
== 0)) {
5045 if (!__memcg_has_children(memcg
))
5046 memcg
->use_hierarchy
= val
;
5053 mutex_unlock(&memcg_create_mutex
);
5059 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5060 enum mem_cgroup_stat_index idx
)
5062 struct mem_cgroup
*iter
;
5065 /* Per-cpu values can be negative, use a signed accumulator */
5066 for_each_mem_cgroup_tree(iter
, memcg
)
5067 val
+= mem_cgroup_read_stat(iter
, idx
);
5069 if (val
< 0) /* race ? */
5074 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5078 if (!mem_cgroup_is_root(memcg
)) {
5080 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5082 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5086 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5087 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5089 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5090 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5093 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5095 return val
<< PAGE_SHIFT
;
5098 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
5099 struct cftype
*cft
, struct file
*file
,
5100 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
5102 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5108 type
= MEMFILE_TYPE(cft
->private);
5109 name
= MEMFILE_ATTR(cft
->private);
5113 if (name
== RES_USAGE
)
5114 val
= mem_cgroup_usage(memcg
, false);
5116 val
= res_counter_read_u64(&memcg
->res
, name
);
5119 if (name
== RES_USAGE
)
5120 val
= mem_cgroup_usage(memcg
, true);
5122 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5125 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5131 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5132 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5135 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5138 #ifdef CONFIG_MEMCG_KMEM
5139 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5141 * For simplicity, we won't allow this to be disabled. It also can't
5142 * be changed if the cgroup has children already, or if tasks had
5145 * If tasks join before we set the limit, a person looking at
5146 * kmem.usage_in_bytes will have no way to determine when it took
5147 * place, which makes the value quite meaningless.
5149 * After it first became limited, changes in the value of the limit are
5150 * of course permitted.
5152 mutex_lock(&memcg_create_mutex
);
5153 mutex_lock(&set_limit_mutex
);
5154 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
5155 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5159 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5162 ret
= memcg_update_cache_sizes(memcg
);
5164 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
5167 static_key_slow_inc(&memcg_kmem_enabled_key
);
5169 * setting the active bit after the inc will guarantee no one
5170 * starts accounting before all call sites are patched
5172 memcg_kmem_set_active(memcg
);
5174 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5176 mutex_unlock(&set_limit_mutex
);
5177 mutex_unlock(&memcg_create_mutex
);
5182 #ifdef CONFIG_MEMCG_KMEM
5183 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5186 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5190 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5192 * When that happen, we need to disable the static branch only on those
5193 * memcgs that enabled it. To achieve this, we would be forced to
5194 * complicate the code by keeping track of which memcgs were the ones
5195 * that actually enabled limits, and which ones got it from its
5198 * It is a lot simpler just to do static_key_slow_inc() on every child
5199 * that is accounted.
5201 if (!memcg_kmem_is_active(memcg
))
5205 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5206 * memcg is active already. If the later initialization fails then the
5207 * cgroup core triggers the cleanup so we do not have to do it here.
5209 static_key_slow_inc(&memcg_kmem_enabled_key
);
5211 mutex_lock(&set_limit_mutex
);
5212 memcg_stop_kmem_account();
5213 ret
= memcg_update_cache_sizes(memcg
);
5214 memcg_resume_kmem_account();
5215 mutex_unlock(&set_limit_mutex
);
5219 #endif /* CONFIG_MEMCG_KMEM */
5222 * The user of this function is...
5225 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5228 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5231 unsigned long long val
;
5234 type
= MEMFILE_TYPE(cft
->private);
5235 name
= MEMFILE_ATTR(cft
->private);
5239 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5243 /* This function does all necessary parse...reuse it */
5244 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5248 ret
= mem_cgroup_resize_limit(memcg
, val
);
5249 else if (type
== _MEMSWAP
)
5250 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5251 else if (type
== _KMEM
)
5252 ret
= memcg_update_kmem_limit(css
, val
);
5256 case RES_SOFT_LIMIT
:
5257 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5261 * For memsw, soft limits are hard to implement in terms
5262 * of semantics, for now, we support soft limits for
5263 * control without swap
5266 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5271 ret
= -EINVAL
; /* should be BUG() ? */
5277 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5278 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5280 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5282 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5283 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5284 if (!memcg
->use_hierarchy
)
5287 while (css_parent(&memcg
->css
)) {
5288 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5289 if (!memcg
->use_hierarchy
)
5291 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5292 min_limit
= min(min_limit
, tmp
);
5293 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5294 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5297 *mem_limit
= min_limit
;
5298 *memsw_limit
= min_memsw_limit
;
5301 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5303 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5307 type
= MEMFILE_TYPE(event
);
5308 name
= MEMFILE_ATTR(event
);
5313 res_counter_reset_max(&memcg
->res
);
5314 else if (type
== _MEMSWAP
)
5315 res_counter_reset_max(&memcg
->memsw
);
5316 else if (type
== _KMEM
)
5317 res_counter_reset_max(&memcg
->kmem
);
5323 res_counter_reset_failcnt(&memcg
->res
);
5324 else if (type
== _MEMSWAP
)
5325 res_counter_reset_failcnt(&memcg
->memsw
);
5326 else if (type
== _KMEM
)
5327 res_counter_reset_failcnt(&memcg
->kmem
);
5336 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5339 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5343 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5344 struct cftype
*cft
, u64 val
)
5346 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5348 if (val
>= (1 << NR_MOVE_TYPE
))
5352 * No kind of locking is needed in here, because ->can_attach() will
5353 * check this value once in the beginning of the process, and then carry
5354 * on with stale data. This means that changes to this value will only
5355 * affect task migrations starting after the change.
5357 memcg
->move_charge_at_immigrate
= val
;
5361 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5362 struct cftype
*cft
, u64 val
)
5369 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5370 struct cftype
*cft
, struct seq_file
*m
)
5373 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5374 unsigned long node_nr
;
5375 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5377 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5378 seq_printf(m
, "total=%lu", total_nr
);
5379 for_each_node_state(nid
, N_MEMORY
) {
5380 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5381 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5385 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5386 seq_printf(m
, "file=%lu", file_nr
);
5387 for_each_node_state(nid
, N_MEMORY
) {
5388 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5390 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5394 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5395 seq_printf(m
, "anon=%lu", anon_nr
);
5396 for_each_node_state(nid
, N_MEMORY
) {
5397 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5399 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5403 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5404 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5405 for_each_node_state(nid
, N_MEMORY
) {
5406 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5407 BIT(LRU_UNEVICTABLE
));
5408 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5413 #endif /* CONFIG_NUMA */
5415 static inline void mem_cgroup_lru_names_not_uptodate(void)
5417 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5420 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5423 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5424 struct mem_cgroup
*mi
;
5427 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5428 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5430 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5431 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5434 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5435 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5436 mem_cgroup_read_events(memcg
, i
));
5438 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5439 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5440 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5442 /* Hierarchical information */
5444 unsigned long long limit
, memsw_limit
;
5445 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5446 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5447 if (do_swap_account
)
5448 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5452 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5455 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5457 for_each_mem_cgroup_tree(mi
, memcg
)
5458 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5459 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5462 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5463 unsigned long long val
= 0;
5465 for_each_mem_cgroup_tree(mi
, memcg
)
5466 val
+= mem_cgroup_read_events(mi
, i
);
5467 seq_printf(m
, "total_%s %llu\n",
5468 mem_cgroup_events_names
[i
], val
);
5471 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5472 unsigned long long val
= 0;
5474 for_each_mem_cgroup_tree(mi
, memcg
)
5475 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5476 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5479 #ifdef CONFIG_DEBUG_VM
5482 struct mem_cgroup_per_zone
*mz
;
5483 struct zone_reclaim_stat
*rstat
;
5484 unsigned long recent_rotated
[2] = {0, 0};
5485 unsigned long recent_scanned
[2] = {0, 0};
5487 for_each_online_node(nid
)
5488 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5489 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5490 rstat
= &mz
->lruvec
.reclaim_stat
;
5492 recent_rotated
[0] += rstat
->recent_rotated
[0];
5493 recent_rotated
[1] += rstat
->recent_rotated
[1];
5494 recent_scanned
[0] += rstat
->recent_scanned
[0];
5495 recent_scanned
[1] += rstat
->recent_scanned
[1];
5497 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5498 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5499 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5500 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5507 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5510 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5512 return mem_cgroup_swappiness(memcg
);
5515 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5516 struct cftype
*cft
, u64 val
)
5518 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5519 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5521 if (val
> 100 || !parent
)
5524 mutex_lock(&memcg_create_mutex
);
5526 /* If under hierarchy, only empty-root can set this value */
5527 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5528 mutex_unlock(&memcg_create_mutex
);
5532 memcg
->swappiness
= val
;
5534 mutex_unlock(&memcg_create_mutex
);
5539 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5541 struct mem_cgroup_threshold_ary
*t
;
5547 t
= rcu_dereference(memcg
->thresholds
.primary
);
5549 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5554 usage
= mem_cgroup_usage(memcg
, swap
);
5557 * current_threshold points to threshold just below or equal to usage.
5558 * If it's not true, a threshold was crossed after last
5559 * call of __mem_cgroup_threshold().
5561 i
= t
->current_threshold
;
5564 * Iterate backward over array of thresholds starting from
5565 * current_threshold and check if a threshold is crossed.
5566 * If none of thresholds below usage is crossed, we read
5567 * only one element of the array here.
5569 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5570 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5572 /* i = current_threshold + 1 */
5576 * Iterate forward over array of thresholds starting from
5577 * current_threshold+1 and check if a threshold is crossed.
5578 * If none of thresholds above usage is crossed, we read
5579 * only one element of the array here.
5581 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5582 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5584 /* Update current_threshold */
5585 t
->current_threshold
= i
- 1;
5590 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5593 __mem_cgroup_threshold(memcg
, false);
5594 if (do_swap_account
)
5595 __mem_cgroup_threshold(memcg
, true);
5597 memcg
= parent_mem_cgroup(memcg
);
5601 static int compare_thresholds(const void *a
, const void *b
)
5603 const struct mem_cgroup_threshold
*_a
= a
;
5604 const struct mem_cgroup_threshold
*_b
= b
;
5606 return _a
->threshold
- _b
->threshold
;
5609 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5611 struct mem_cgroup_eventfd_list
*ev
;
5613 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5614 eventfd_signal(ev
->eventfd
, 1);
5618 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5620 struct mem_cgroup
*iter
;
5622 for_each_mem_cgroup_tree(iter
, memcg
)
5623 mem_cgroup_oom_notify_cb(iter
);
5626 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5627 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5629 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5630 struct mem_cgroup_thresholds
*thresholds
;
5631 struct mem_cgroup_threshold_ary
*new;
5632 enum res_type type
= MEMFILE_TYPE(cft
->private);
5633 u64 threshold
, usage
;
5636 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5640 mutex_lock(&memcg
->thresholds_lock
);
5643 thresholds
= &memcg
->thresholds
;
5644 else if (type
== _MEMSWAP
)
5645 thresholds
= &memcg
->memsw_thresholds
;
5649 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5651 /* Check if a threshold crossed before adding a new one */
5652 if (thresholds
->primary
)
5653 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5655 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5657 /* Allocate memory for new array of thresholds */
5658 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5666 /* Copy thresholds (if any) to new array */
5667 if (thresholds
->primary
) {
5668 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5669 sizeof(struct mem_cgroup_threshold
));
5672 /* Add new threshold */
5673 new->entries
[size
- 1].eventfd
= eventfd
;
5674 new->entries
[size
- 1].threshold
= threshold
;
5676 /* Sort thresholds. Registering of new threshold isn't time-critical */
5677 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5678 compare_thresholds
, NULL
);
5680 /* Find current threshold */
5681 new->current_threshold
= -1;
5682 for (i
= 0; i
< size
; i
++) {
5683 if (new->entries
[i
].threshold
<= usage
) {
5685 * new->current_threshold will not be used until
5686 * rcu_assign_pointer(), so it's safe to increment
5689 ++new->current_threshold
;
5694 /* Free old spare buffer and save old primary buffer as spare */
5695 kfree(thresholds
->spare
);
5696 thresholds
->spare
= thresholds
->primary
;
5698 rcu_assign_pointer(thresholds
->primary
, new);
5700 /* To be sure that nobody uses thresholds */
5704 mutex_unlock(&memcg
->thresholds_lock
);
5709 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5710 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5712 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5713 struct mem_cgroup_thresholds
*thresholds
;
5714 struct mem_cgroup_threshold_ary
*new;
5715 enum res_type type
= MEMFILE_TYPE(cft
->private);
5719 mutex_lock(&memcg
->thresholds_lock
);
5721 thresholds
= &memcg
->thresholds
;
5722 else if (type
== _MEMSWAP
)
5723 thresholds
= &memcg
->memsw_thresholds
;
5727 if (!thresholds
->primary
)
5730 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5732 /* Check if a threshold crossed before removing */
5733 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5735 /* Calculate new number of threshold */
5737 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5738 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5742 new = thresholds
->spare
;
5744 /* Set thresholds array to NULL if we don't have thresholds */
5753 /* Copy thresholds and find current threshold */
5754 new->current_threshold
= -1;
5755 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5756 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5759 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5760 if (new->entries
[j
].threshold
<= usage
) {
5762 * new->current_threshold will not be used
5763 * until rcu_assign_pointer(), so it's safe to increment
5766 ++new->current_threshold
;
5772 /* Swap primary and spare array */
5773 thresholds
->spare
= thresholds
->primary
;
5774 /* If all events are unregistered, free the spare array */
5776 kfree(thresholds
->spare
);
5777 thresholds
->spare
= NULL
;
5780 rcu_assign_pointer(thresholds
->primary
, new);
5782 /* To be sure that nobody uses thresholds */
5785 mutex_unlock(&memcg
->thresholds_lock
);
5788 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5789 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5791 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5792 struct mem_cgroup_eventfd_list
*event
;
5793 enum res_type type
= MEMFILE_TYPE(cft
->private);
5795 BUG_ON(type
!= _OOM_TYPE
);
5796 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5800 spin_lock(&memcg_oom_lock
);
5802 event
->eventfd
= eventfd
;
5803 list_add(&event
->list
, &memcg
->oom_notify
);
5805 /* already in OOM ? */
5806 if (atomic_read(&memcg
->under_oom
))
5807 eventfd_signal(eventfd
, 1);
5808 spin_unlock(&memcg_oom_lock
);
5813 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5814 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5816 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5817 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5818 enum res_type type
= MEMFILE_TYPE(cft
->private);
5820 BUG_ON(type
!= _OOM_TYPE
);
5822 spin_lock(&memcg_oom_lock
);
5824 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5825 if (ev
->eventfd
== eventfd
) {
5826 list_del(&ev
->list
);
5831 spin_unlock(&memcg_oom_lock
);
5834 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5835 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5837 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5839 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5841 if (atomic_read(&memcg
->under_oom
))
5842 cb
->fill(cb
, "under_oom", 1);
5844 cb
->fill(cb
, "under_oom", 0);
5848 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5849 struct cftype
*cft
, u64 val
)
5851 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5852 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5854 /* cannot set to root cgroup and only 0 and 1 are allowed */
5855 if (!parent
|| !((val
== 0) || (val
== 1)))
5858 mutex_lock(&memcg_create_mutex
);
5859 /* oom-kill-disable is a flag for subhierarchy. */
5860 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5861 mutex_unlock(&memcg_create_mutex
);
5864 memcg
->oom_kill_disable
= val
;
5866 memcg_oom_recover(memcg
);
5867 mutex_unlock(&memcg_create_mutex
);
5871 #ifdef CONFIG_MEMCG_KMEM
5872 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5876 memcg
->kmemcg_id
= -1;
5877 ret
= memcg_propagate_kmem(memcg
);
5881 return mem_cgroup_sockets_init(memcg
, ss
);
5884 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5886 mem_cgroup_sockets_destroy(memcg
);
5889 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5891 if (!memcg_kmem_is_active(memcg
))
5895 * kmem charges can outlive the cgroup. In the case of slab
5896 * pages, for instance, a page contain objects from various
5897 * processes. As we prevent from taking a reference for every
5898 * such allocation we have to be careful when doing uncharge
5899 * (see memcg_uncharge_kmem) and here during offlining.
5901 * The idea is that that only the _last_ uncharge which sees
5902 * the dead memcg will drop the last reference. An additional
5903 * reference is taken here before the group is marked dead
5904 * which is then paired with css_put during uncharge resp. here.
5906 * Although this might sound strange as this path is called from
5907 * css_offline() when the referencemight have dropped down to 0
5908 * and shouldn't be incremented anymore (css_tryget would fail)
5909 * we do not have other options because of the kmem allocations
5912 css_get(&memcg
->css
);
5914 memcg_kmem_mark_dead(memcg
);
5916 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5919 if (memcg_kmem_test_and_clear_dead(memcg
))
5920 css_put(&memcg
->css
);
5923 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5928 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5932 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5937 static struct cftype mem_cgroup_files
[] = {
5939 .name
= "usage_in_bytes",
5940 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5941 .read
= mem_cgroup_read
,
5942 .register_event
= mem_cgroup_usage_register_event
,
5943 .unregister_event
= mem_cgroup_usage_unregister_event
,
5946 .name
= "max_usage_in_bytes",
5947 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5948 .trigger
= mem_cgroup_reset
,
5949 .read
= mem_cgroup_read
,
5952 .name
= "limit_in_bytes",
5953 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5954 .write_string
= mem_cgroup_write
,
5955 .read
= mem_cgroup_read
,
5958 .name
= "soft_limit_in_bytes",
5959 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5960 .write_string
= mem_cgroup_write
,
5961 .read
= mem_cgroup_read
,
5965 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5966 .trigger
= mem_cgroup_reset
,
5967 .read
= mem_cgroup_read
,
5971 .read_seq_string
= memcg_stat_show
,
5974 .name
= "force_empty",
5975 .trigger
= mem_cgroup_force_empty_write
,
5978 .name
= "use_hierarchy",
5979 .flags
= CFTYPE_INSANE
,
5980 .write_u64
= mem_cgroup_hierarchy_write
,
5981 .read_u64
= mem_cgroup_hierarchy_read
,
5984 .name
= "swappiness",
5985 .read_u64
= mem_cgroup_swappiness_read
,
5986 .write_u64
= mem_cgroup_swappiness_write
,
5989 .name
= "move_charge_at_immigrate",
5990 .read_u64
= mem_cgroup_move_charge_read
,
5991 .write_u64
= mem_cgroup_move_charge_write
,
5994 .name
= "oom_control",
5995 .read_map
= mem_cgroup_oom_control_read
,
5996 .write_u64
= mem_cgroup_oom_control_write
,
5997 .register_event
= mem_cgroup_oom_register_event
,
5998 .unregister_event
= mem_cgroup_oom_unregister_event
,
5999 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6002 .name
= "pressure_level",
6003 .register_event
= vmpressure_register_event
,
6004 .unregister_event
= vmpressure_unregister_event
,
6008 .name
= "numa_stat",
6009 .read_seq_string
= memcg_numa_stat_show
,
6012 #ifdef CONFIG_MEMCG_KMEM
6014 .name
= "kmem.limit_in_bytes",
6015 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6016 .write_string
= mem_cgroup_write
,
6017 .read
= mem_cgroup_read
,
6020 .name
= "kmem.usage_in_bytes",
6021 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6022 .read
= mem_cgroup_read
,
6025 .name
= "kmem.failcnt",
6026 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6027 .trigger
= mem_cgroup_reset
,
6028 .read
= mem_cgroup_read
,
6031 .name
= "kmem.max_usage_in_bytes",
6032 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6033 .trigger
= mem_cgroup_reset
,
6034 .read
= mem_cgroup_read
,
6036 #ifdef CONFIG_SLABINFO
6038 .name
= "kmem.slabinfo",
6039 .read_seq_string
= mem_cgroup_slabinfo_read
,
6043 { }, /* terminate */
6046 #ifdef CONFIG_MEMCG_SWAP
6047 static struct cftype memsw_cgroup_files
[] = {
6049 .name
= "memsw.usage_in_bytes",
6050 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6051 .read
= mem_cgroup_read
,
6052 .register_event
= mem_cgroup_usage_register_event
,
6053 .unregister_event
= mem_cgroup_usage_unregister_event
,
6056 .name
= "memsw.max_usage_in_bytes",
6057 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6058 .trigger
= mem_cgroup_reset
,
6059 .read
= mem_cgroup_read
,
6062 .name
= "memsw.limit_in_bytes",
6063 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6064 .write_string
= mem_cgroup_write
,
6065 .read
= mem_cgroup_read
,
6068 .name
= "memsw.failcnt",
6069 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6070 .trigger
= mem_cgroup_reset
,
6071 .read
= mem_cgroup_read
,
6073 { }, /* terminate */
6076 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6078 struct mem_cgroup_per_node
*pn
;
6079 struct mem_cgroup_per_zone
*mz
;
6080 int zone
, tmp
= node
;
6082 * This routine is called against possible nodes.
6083 * But it's BUG to call kmalloc() against offline node.
6085 * TODO: this routine can waste much memory for nodes which will
6086 * never be onlined. It's better to use memory hotplug callback
6089 if (!node_state(node
, N_NORMAL_MEMORY
))
6091 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6095 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6096 mz
= &pn
->zoneinfo
[zone
];
6097 lruvec_init(&mz
->lruvec
);
6098 mz
->usage_in_excess
= 0;
6099 mz
->on_tree
= false;
6102 memcg
->nodeinfo
[node
] = pn
;
6106 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6108 kfree(memcg
->nodeinfo
[node
]);
6111 static struct mem_cgroup
*mem_cgroup_alloc(void)
6113 struct mem_cgroup
*memcg
;
6114 size_t size
= memcg_size();
6116 /* Can be very big if nr_node_ids is very big */
6117 if (size
< PAGE_SIZE
)
6118 memcg
= kzalloc(size
, GFP_KERNEL
);
6120 memcg
= vzalloc(size
);
6125 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6128 spin_lock_init(&memcg
->pcp_counter_lock
);
6132 if (size
< PAGE_SIZE
)
6140 * At destroying mem_cgroup, references from swap_cgroup can remain.
6141 * (scanning all at force_empty is too costly...)
6143 * Instead of clearing all references at force_empty, we remember
6144 * the number of reference from swap_cgroup and free mem_cgroup when
6145 * it goes down to 0.
6147 * Removal of cgroup itself succeeds regardless of refs from swap.
6150 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6153 size_t size
= memcg_size();
6155 mem_cgroup_remove_from_trees(memcg
);
6156 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
6159 free_mem_cgroup_per_zone_info(memcg
, node
);
6161 free_percpu(memcg
->stat
);
6164 * We need to make sure that (at least for now), the jump label
6165 * destruction code runs outside of the cgroup lock. This is because
6166 * get_online_cpus(), which is called from the static_branch update,
6167 * can't be called inside the cgroup_lock. cpusets are the ones
6168 * enforcing this dependency, so if they ever change, we might as well.
6170 * schedule_work() will guarantee this happens. Be careful if you need
6171 * to move this code around, and make sure it is outside
6174 disarm_static_keys(memcg
);
6175 if (size
< PAGE_SIZE
)
6182 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6184 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6186 if (!memcg
->res
.parent
)
6188 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6190 EXPORT_SYMBOL(parent_mem_cgroup
);
6192 static void __init
mem_cgroup_soft_limit_tree_init(void)
6194 struct mem_cgroup_tree_per_node
*rtpn
;
6195 struct mem_cgroup_tree_per_zone
*rtpz
;
6196 int tmp
, node
, zone
;
6198 for_each_node(node
) {
6200 if (!node_state(node
, N_NORMAL_MEMORY
))
6202 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6205 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6207 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6208 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6209 rtpz
->rb_root
= RB_ROOT
;
6210 spin_lock_init(&rtpz
->lock
);
6215 static struct cgroup_subsys_state
* __ref
6216 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6218 struct mem_cgroup
*memcg
;
6219 long error
= -ENOMEM
;
6222 memcg
= mem_cgroup_alloc();
6224 return ERR_PTR(error
);
6227 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6231 if (parent_css
== NULL
) {
6232 root_mem_cgroup
= memcg
;
6233 res_counter_init(&memcg
->res
, NULL
);
6234 res_counter_init(&memcg
->memsw
, NULL
);
6235 res_counter_init(&memcg
->kmem
, NULL
);
6238 memcg
->last_scanned_node
= MAX_NUMNODES
;
6239 INIT_LIST_HEAD(&memcg
->oom_notify
);
6240 memcg
->move_charge_at_immigrate
= 0;
6241 mutex_init(&memcg
->thresholds_lock
);
6242 spin_lock_init(&memcg
->move_lock
);
6243 vmpressure_init(&memcg
->vmpressure
);
6248 __mem_cgroup_free(memcg
);
6249 return ERR_PTR(error
);
6253 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6255 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6256 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6262 mutex_lock(&memcg_create_mutex
);
6264 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6265 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6266 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6268 if (parent
->use_hierarchy
) {
6269 res_counter_init(&memcg
->res
, &parent
->res
);
6270 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6271 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6274 * No need to take a reference to the parent because cgroup
6275 * core guarantees its existence.
6278 res_counter_init(&memcg
->res
, NULL
);
6279 res_counter_init(&memcg
->memsw
, NULL
);
6280 res_counter_init(&memcg
->kmem
, NULL
);
6282 * Deeper hierachy with use_hierarchy == false doesn't make
6283 * much sense so let cgroup subsystem know about this
6284 * unfortunate state in our controller.
6286 if (parent
!= root_mem_cgroup
)
6287 mem_cgroup_subsys
.broken_hierarchy
= true;
6290 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6291 mutex_unlock(&memcg_create_mutex
);
6296 * Announce all parents that a group from their hierarchy is gone.
6298 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6300 struct mem_cgroup
*parent
= memcg
;
6302 while ((parent
= parent_mem_cgroup(parent
)))
6303 mem_cgroup_iter_invalidate(parent
);
6306 * if the root memcg is not hierarchical we have to check it
6309 if (!root_mem_cgroup
->use_hierarchy
)
6310 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6313 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6315 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6317 kmem_cgroup_css_offline(memcg
);
6319 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6320 mem_cgroup_reparent_charges(memcg
);
6321 mem_cgroup_destroy_all_caches(memcg
);
6324 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6326 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6328 memcg_destroy_kmem(memcg
);
6329 __mem_cgroup_free(memcg
);
6333 /* Handlers for move charge at task migration. */
6334 #define PRECHARGE_COUNT_AT_ONCE 256
6335 static int mem_cgroup_do_precharge(unsigned long count
)
6338 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6339 struct mem_cgroup
*memcg
= mc
.to
;
6341 if (mem_cgroup_is_root(memcg
)) {
6342 mc
.precharge
+= count
;
6343 /* we don't need css_get for root */
6346 /* try to charge at once */
6348 struct res_counter
*dummy
;
6350 * "memcg" cannot be under rmdir() because we've already checked
6351 * by cgroup_lock_live_cgroup() that it is not removed and we
6352 * are still under the same cgroup_mutex. So we can postpone
6355 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6357 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6358 PAGE_SIZE
* count
, &dummy
)) {
6359 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6362 mc
.precharge
+= count
;
6366 /* fall back to one by one charge */
6368 if (signal_pending(current
)) {
6372 if (!batch_count
--) {
6373 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6376 ret
= __mem_cgroup_try_charge(NULL
,
6377 GFP_KERNEL
, 1, &memcg
, false);
6379 /* mem_cgroup_clear_mc() will do uncharge later */
6387 * get_mctgt_type - get target type of moving charge
6388 * @vma: the vma the pte to be checked belongs
6389 * @addr: the address corresponding to the pte to be checked
6390 * @ptent: the pte to be checked
6391 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6394 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6395 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6396 * move charge. if @target is not NULL, the page is stored in target->page
6397 * with extra refcnt got(Callers should handle it).
6398 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6399 * target for charge migration. if @target is not NULL, the entry is stored
6402 * Called with pte lock held.
6409 enum mc_target_type
{
6415 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6416 unsigned long addr
, pte_t ptent
)
6418 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6420 if (!page
|| !page_mapped(page
))
6422 if (PageAnon(page
)) {
6423 /* we don't move shared anon */
6426 } else if (!move_file())
6427 /* we ignore mapcount for file pages */
6429 if (!get_page_unless_zero(page
))
6436 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6437 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6439 struct page
*page
= NULL
;
6440 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6442 if (!move_anon() || non_swap_entry(ent
))
6445 * Because lookup_swap_cache() updates some statistics counter,
6446 * we call find_get_page() with swapper_space directly.
6448 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6449 if (do_swap_account
)
6450 entry
->val
= ent
.val
;
6455 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6456 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6462 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6463 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6465 struct page
*page
= NULL
;
6466 struct address_space
*mapping
;
6469 if (!vma
->vm_file
) /* anonymous vma */
6474 mapping
= vma
->vm_file
->f_mapping
;
6475 if (pte_none(ptent
))
6476 pgoff
= linear_page_index(vma
, addr
);
6477 else /* pte_file(ptent) is true */
6478 pgoff
= pte_to_pgoff(ptent
);
6480 /* page is moved even if it's not RSS of this task(page-faulted). */
6481 page
= find_get_page(mapping
, pgoff
);
6484 /* shmem/tmpfs may report page out on swap: account for that too. */
6485 if (radix_tree_exceptional_entry(page
)) {
6486 swp_entry_t swap
= radix_to_swp_entry(page
);
6487 if (do_swap_account
)
6489 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6495 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6496 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6498 struct page
*page
= NULL
;
6499 struct page_cgroup
*pc
;
6500 enum mc_target_type ret
= MC_TARGET_NONE
;
6501 swp_entry_t ent
= { .val
= 0 };
6503 if (pte_present(ptent
))
6504 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6505 else if (is_swap_pte(ptent
))
6506 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6507 else if (pte_none(ptent
) || pte_file(ptent
))
6508 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6510 if (!page
&& !ent
.val
)
6513 pc
= lookup_page_cgroup(page
);
6515 * Do only loose check w/o page_cgroup lock.
6516 * mem_cgroup_move_account() checks the pc is valid or not under
6519 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6520 ret
= MC_TARGET_PAGE
;
6522 target
->page
= page
;
6524 if (!ret
|| !target
)
6527 /* There is a swap entry and a page doesn't exist or isn't charged */
6528 if (ent
.val
&& !ret
&&
6529 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6530 ret
= MC_TARGET_SWAP
;
6537 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6539 * We don't consider swapping or file mapped pages because THP does not
6540 * support them for now.
6541 * Caller should make sure that pmd_trans_huge(pmd) is true.
6543 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6544 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6546 struct page
*page
= NULL
;
6547 struct page_cgroup
*pc
;
6548 enum mc_target_type ret
= MC_TARGET_NONE
;
6550 page
= pmd_page(pmd
);
6551 VM_BUG_ON(!page
|| !PageHead(page
));
6554 pc
= lookup_page_cgroup(page
);
6555 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6556 ret
= MC_TARGET_PAGE
;
6559 target
->page
= page
;
6565 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6566 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6568 return MC_TARGET_NONE
;
6572 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6573 unsigned long addr
, unsigned long end
,
6574 struct mm_walk
*walk
)
6576 struct vm_area_struct
*vma
= walk
->private;
6580 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6581 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6582 mc
.precharge
+= HPAGE_PMD_NR
;
6583 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6587 if (pmd_trans_unstable(pmd
))
6589 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6590 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6591 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6592 mc
.precharge
++; /* increment precharge temporarily */
6593 pte_unmap_unlock(pte
- 1, ptl
);
6599 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6601 unsigned long precharge
;
6602 struct vm_area_struct
*vma
;
6604 down_read(&mm
->mmap_sem
);
6605 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6606 struct mm_walk mem_cgroup_count_precharge_walk
= {
6607 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6611 if (is_vm_hugetlb_page(vma
))
6613 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6614 &mem_cgroup_count_precharge_walk
);
6616 up_read(&mm
->mmap_sem
);
6618 precharge
= mc
.precharge
;
6624 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6626 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6628 VM_BUG_ON(mc
.moving_task
);
6629 mc
.moving_task
= current
;
6630 return mem_cgroup_do_precharge(precharge
);
6633 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6634 static void __mem_cgroup_clear_mc(void)
6636 struct mem_cgroup
*from
= mc
.from
;
6637 struct mem_cgroup
*to
= mc
.to
;
6640 /* we must uncharge all the leftover precharges from mc.to */
6642 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6646 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6647 * we must uncharge here.
6649 if (mc
.moved_charge
) {
6650 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6651 mc
.moved_charge
= 0;
6653 /* we must fixup refcnts and charges */
6654 if (mc
.moved_swap
) {
6655 /* uncharge swap account from the old cgroup */
6656 if (!mem_cgroup_is_root(mc
.from
))
6657 res_counter_uncharge(&mc
.from
->memsw
,
6658 PAGE_SIZE
* mc
.moved_swap
);
6660 for (i
= 0; i
< mc
.moved_swap
; i
++)
6661 css_put(&mc
.from
->css
);
6663 if (!mem_cgroup_is_root(mc
.to
)) {
6665 * we charged both to->res and to->memsw, so we should
6668 res_counter_uncharge(&mc
.to
->res
,
6669 PAGE_SIZE
* mc
.moved_swap
);
6671 /* we've already done css_get(mc.to) */
6674 memcg_oom_recover(from
);
6675 memcg_oom_recover(to
);
6676 wake_up_all(&mc
.waitq
);
6679 static void mem_cgroup_clear_mc(void)
6681 struct mem_cgroup
*from
= mc
.from
;
6684 * we must clear moving_task before waking up waiters at the end of
6687 mc
.moving_task
= NULL
;
6688 __mem_cgroup_clear_mc();
6689 spin_lock(&mc
.lock
);
6692 spin_unlock(&mc
.lock
);
6693 mem_cgroup_end_move(from
);
6696 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6697 struct cgroup_taskset
*tset
)
6699 struct task_struct
*p
= cgroup_taskset_first(tset
);
6701 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6702 unsigned long move_charge_at_immigrate
;
6705 * We are now commited to this value whatever it is. Changes in this
6706 * tunable will only affect upcoming migrations, not the current one.
6707 * So we need to save it, and keep it going.
6709 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6710 if (move_charge_at_immigrate
) {
6711 struct mm_struct
*mm
;
6712 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6714 VM_BUG_ON(from
== memcg
);
6716 mm
= get_task_mm(p
);
6719 /* We move charges only when we move a owner of the mm */
6720 if (mm
->owner
== p
) {
6723 VM_BUG_ON(mc
.precharge
);
6724 VM_BUG_ON(mc
.moved_charge
);
6725 VM_BUG_ON(mc
.moved_swap
);
6726 mem_cgroup_start_move(from
);
6727 spin_lock(&mc
.lock
);
6730 mc
.immigrate_flags
= move_charge_at_immigrate
;
6731 spin_unlock(&mc
.lock
);
6732 /* We set mc.moving_task later */
6734 ret
= mem_cgroup_precharge_mc(mm
);
6736 mem_cgroup_clear_mc();
6743 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6744 struct cgroup_taskset
*tset
)
6746 mem_cgroup_clear_mc();
6749 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6750 unsigned long addr
, unsigned long end
,
6751 struct mm_walk
*walk
)
6754 struct vm_area_struct
*vma
= walk
->private;
6757 enum mc_target_type target_type
;
6758 union mc_target target
;
6760 struct page_cgroup
*pc
;
6763 * We don't take compound_lock() here but no race with splitting thp
6765 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6766 * under splitting, which means there's no concurrent thp split,
6767 * - if another thread runs into split_huge_page() just after we
6768 * entered this if-block, the thread must wait for page table lock
6769 * to be unlocked in __split_huge_page_splitting(), where the main
6770 * part of thp split is not executed yet.
6772 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6773 if (mc
.precharge
< HPAGE_PMD_NR
) {
6774 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6777 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6778 if (target_type
== MC_TARGET_PAGE
) {
6780 if (!isolate_lru_page(page
)) {
6781 pc
= lookup_page_cgroup(page
);
6782 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6783 pc
, mc
.from
, mc
.to
)) {
6784 mc
.precharge
-= HPAGE_PMD_NR
;
6785 mc
.moved_charge
+= HPAGE_PMD_NR
;
6787 putback_lru_page(page
);
6791 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6795 if (pmd_trans_unstable(pmd
))
6798 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6799 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6800 pte_t ptent
= *(pte
++);
6806 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6807 case MC_TARGET_PAGE
:
6809 if (isolate_lru_page(page
))
6811 pc
= lookup_page_cgroup(page
);
6812 if (!mem_cgroup_move_account(page
, 1, pc
,
6815 /* we uncharge from mc.from later. */
6818 putback_lru_page(page
);
6819 put
: /* get_mctgt_type() gets the page */
6822 case MC_TARGET_SWAP
:
6824 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6826 /* we fixup refcnts and charges later. */
6834 pte_unmap_unlock(pte
- 1, ptl
);
6839 * We have consumed all precharges we got in can_attach().
6840 * We try charge one by one, but don't do any additional
6841 * charges to mc.to if we have failed in charge once in attach()
6844 ret
= mem_cgroup_do_precharge(1);
6852 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6854 struct vm_area_struct
*vma
;
6856 lru_add_drain_all();
6858 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6860 * Someone who are holding the mmap_sem might be waiting in
6861 * waitq. So we cancel all extra charges, wake up all waiters,
6862 * and retry. Because we cancel precharges, we might not be able
6863 * to move enough charges, but moving charge is a best-effort
6864 * feature anyway, so it wouldn't be a big problem.
6866 __mem_cgroup_clear_mc();
6870 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6872 struct mm_walk mem_cgroup_move_charge_walk
= {
6873 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6877 if (is_vm_hugetlb_page(vma
))
6879 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6880 &mem_cgroup_move_charge_walk
);
6883 * means we have consumed all precharges and failed in
6884 * doing additional charge. Just abandon here.
6888 up_read(&mm
->mmap_sem
);
6891 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6892 struct cgroup_taskset
*tset
)
6894 struct task_struct
*p
= cgroup_taskset_first(tset
);
6895 struct mm_struct
*mm
= get_task_mm(p
);
6899 mem_cgroup_move_charge(mm
);
6903 mem_cgroup_clear_mc();
6905 #else /* !CONFIG_MMU */
6906 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6907 struct cgroup_taskset
*tset
)
6911 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6912 struct cgroup_taskset
*tset
)
6915 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6916 struct cgroup_taskset
*tset
)
6922 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6923 * to verify sane_behavior flag on each mount attempt.
6925 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6928 * use_hierarchy is forced with sane_behavior. cgroup core
6929 * guarantees that @root doesn't have any children, so turning it
6930 * on for the root memcg is enough.
6932 if (cgroup_sane_behavior(root_css
->cgroup
))
6933 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6936 struct cgroup_subsys mem_cgroup_subsys
= {
6938 .subsys_id
= mem_cgroup_subsys_id
,
6939 .css_alloc
= mem_cgroup_css_alloc
,
6940 .css_online
= mem_cgroup_css_online
,
6941 .css_offline
= mem_cgroup_css_offline
,
6942 .css_free
= mem_cgroup_css_free
,
6943 .can_attach
= mem_cgroup_can_attach
,
6944 .cancel_attach
= mem_cgroup_cancel_attach
,
6945 .attach
= mem_cgroup_move_task
,
6946 .bind
= mem_cgroup_bind
,
6947 .base_cftypes
= mem_cgroup_files
,
6952 #ifdef CONFIG_MEMCG_SWAP
6953 static int __init
enable_swap_account(char *s
)
6955 /* consider enabled if no parameter or 1 is given */
6956 if (!strcmp(s
, "1"))
6957 really_do_swap_account
= 1;
6958 else if (!strcmp(s
, "0"))
6959 really_do_swap_account
= 0;
6962 __setup("swapaccount=", enable_swap_account
);
6964 static void __init
memsw_file_init(void)
6966 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6969 static void __init
enable_swap_cgroup(void)
6971 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6972 do_swap_account
= 1;
6978 static void __init
enable_swap_cgroup(void)
6984 * subsys_initcall() for memory controller.
6986 * Some parts like hotcpu_notifier() have to be initialized from this context
6987 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6988 * everything that doesn't depend on a specific mem_cgroup structure should
6989 * be initialized from here.
6991 static int __init
mem_cgroup_init(void)
6993 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6994 enable_swap_cgroup();
6995 mem_cgroup_soft_limit_tree_init();
6999 subsys_initcall(mem_cgroup_init
);