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_subsys_state
*prev_css
, *next_css
;
1088 * Root is not visited by cgroup iterators so it needs an
1094 prev_css
= (last_visited
== root
) ? NULL
: &last_visited
->css
;
1096 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1099 * Even if we found a group we have to make sure it is
1100 * alive. css && !memcg means that the groups should be
1101 * skipped and we should continue the tree walk.
1102 * last_visited css is safe to use because it is
1103 * protected by css_get and the tree walk is rcu safe.
1106 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1108 if (css_tryget(&mem
->css
))
1111 prev_css
= next_css
;
1119 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1122 * When a group in the hierarchy below root is destroyed, the
1123 * hierarchy iterator can no longer be trusted since it might
1124 * have pointed to the destroyed group. Invalidate it.
1126 atomic_inc(&root
->dead_count
);
1129 static struct mem_cgroup
*
1130 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1131 struct mem_cgroup
*root
,
1134 struct mem_cgroup
*position
= NULL
;
1136 * A cgroup destruction happens in two stages: offlining and
1137 * release. They are separated by a RCU grace period.
1139 * If the iterator is valid, we may still race with an
1140 * offlining. The RCU lock ensures the object won't be
1141 * released, tryget will fail if we lost the race.
1143 *sequence
= atomic_read(&root
->dead_count
);
1144 if (iter
->last_dead_count
== *sequence
) {
1146 position
= iter
->last_visited
;
1147 if (position
&& !css_tryget(&position
->css
))
1153 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1154 struct mem_cgroup
*last_visited
,
1155 struct mem_cgroup
*new_position
,
1159 css_put(&last_visited
->css
);
1161 * We store the sequence count from the time @last_visited was
1162 * loaded successfully instead of rereading it here so that we
1163 * don't lose destruction events in between. We could have
1164 * raced with the destruction of @new_position after all.
1166 iter
->last_visited
= new_position
;
1168 iter
->last_dead_count
= sequence
;
1172 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1173 * @root: hierarchy root
1174 * @prev: previously returned memcg, NULL on first invocation
1175 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1177 * Returns references to children of the hierarchy below @root, or
1178 * @root itself, or %NULL after a full round-trip.
1180 * Caller must pass the return value in @prev on subsequent
1181 * invocations for reference counting, or use mem_cgroup_iter_break()
1182 * to cancel a hierarchy walk before the round-trip is complete.
1184 * Reclaimers can specify a zone and a priority level in @reclaim to
1185 * divide up the memcgs in the hierarchy among all concurrent
1186 * reclaimers operating on the same zone and priority.
1188 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1189 struct mem_cgroup
*prev
,
1190 struct mem_cgroup_reclaim_cookie
*reclaim
)
1192 struct mem_cgroup
*memcg
= NULL
;
1193 struct mem_cgroup
*last_visited
= NULL
;
1195 if (mem_cgroup_disabled())
1199 root
= root_mem_cgroup
;
1201 if (prev
&& !reclaim
)
1202 last_visited
= prev
;
1204 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1212 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1213 int uninitialized_var(seq
);
1216 int nid
= zone_to_nid(reclaim
->zone
);
1217 int zid
= zone_idx(reclaim
->zone
);
1218 struct mem_cgroup_per_zone
*mz
;
1220 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1221 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1222 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1223 iter
->last_visited
= NULL
;
1227 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1230 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1233 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1237 else if (!prev
&& memcg
)
1238 reclaim
->generation
= iter
->generation
;
1247 if (prev
&& prev
!= root
)
1248 css_put(&prev
->css
);
1254 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1255 * @root: hierarchy root
1256 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1258 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1259 struct mem_cgroup
*prev
)
1262 root
= root_mem_cgroup
;
1263 if (prev
&& prev
!= root
)
1264 css_put(&prev
->css
);
1268 * Iteration constructs for visiting all cgroups (under a tree). If
1269 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1270 * be used for reference counting.
1272 #define for_each_mem_cgroup_tree(iter, root) \
1273 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1275 iter = mem_cgroup_iter(root, iter, NULL))
1277 #define for_each_mem_cgroup(iter) \
1278 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1280 iter = mem_cgroup_iter(NULL, iter, NULL))
1282 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1284 struct mem_cgroup
*memcg
;
1287 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1288 if (unlikely(!memcg
))
1293 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1296 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1304 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1307 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1308 * @zone: zone of the wanted lruvec
1309 * @memcg: memcg of the wanted lruvec
1311 * Returns the lru list vector holding pages for the given @zone and
1312 * @mem. This can be the global zone lruvec, if the memory controller
1315 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1316 struct mem_cgroup
*memcg
)
1318 struct mem_cgroup_per_zone
*mz
;
1319 struct lruvec
*lruvec
;
1321 if (mem_cgroup_disabled()) {
1322 lruvec
= &zone
->lruvec
;
1326 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1327 lruvec
= &mz
->lruvec
;
1330 * Since a node can be onlined after the mem_cgroup was created,
1331 * we have to be prepared to initialize lruvec->zone here;
1332 * and if offlined then reonlined, we need to reinitialize it.
1334 if (unlikely(lruvec
->zone
!= zone
))
1335 lruvec
->zone
= zone
;
1340 * Following LRU functions are allowed to be used without PCG_LOCK.
1341 * Operations are called by routine of global LRU independently from memcg.
1342 * What we have to take care of here is validness of pc->mem_cgroup.
1344 * Changes to pc->mem_cgroup happens when
1347 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1348 * It is added to LRU before charge.
1349 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1350 * When moving account, the page is not on LRU. It's isolated.
1354 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1356 * @zone: zone of the page
1358 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1360 struct mem_cgroup_per_zone
*mz
;
1361 struct mem_cgroup
*memcg
;
1362 struct page_cgroup
*pc
;
1363 struct lruvec
*lruvec
;
1365 if (mem_cgroup_disabled()) {
1366 lruvec
= &zone
->lruvec
;
1370 pc
= lookup_page_cgroup(page
);
1371 memcg
= pc
->mem_cgroup
;
1374 * Surreptitiously switch any uncharged offlist page to root:
1375 * an uncharged page off lru does nothing to secure
1376 * its former mem_cgroup from sudden removal.
1378 * Our caller holds lru_lock, and PageCgroupUsed is updated
1379 * under page_cgroup lock: between them, they make all uses
1380 * of pc->mem_cgroup safe.
1382 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1383 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1385 mz
= page_cgroup_zoneinfo(memcg
, page
);
1386 lruvec
= &mz
->lruvec
;
1389 * Since a node can be onlined after the mem_cgroup was created,
1390 * we have to be prepared to initialize lruvec->zone here;
1391 * and if offlined then reonlined, we need to reinitialize it.
1393 if (unlikely(lruvec
->zone
!= zone
))
1394 lruvec
->zone
= zone
;
1399 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1400 * @lruvec: mem_cgroup per zone lru vector
1401 * @lru: index of lru list the page is sitting on
1402 * @nr_pages: positive when adding or negative when removing
1404 * This function must be called when a page is added to or removed from an
1407 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1410 struct mem_cgroup_per_zone
*mz
;
1411 unsigned long *lru_size
;
1413 if (mem_cgroup_disabled())
1416 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1417 lru_size
= mz
->lru_size
+ lru
;
1418 *lru_size
+= nr_pages
;
1419 VM_BUG_ON((long)(*lru_size
) < 0);
1423 * Checks whether given mem is same or in the root_mem_cgroup's
1426 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1427 struct mem_cgroup
*memcg
)
1429 if (root_memcg
== memcg
)
1431 if (!root_memcg
->use_hierarchy
|| !memcg
)
1433 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1436 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1437 struct mem_cgroup
*memcg
)
1442 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1447 bool task_in_mem_cgroup(struct task_struct
*task
,
1448 const struct mem_cgroup
*memcg
)
1450 struct mem_cgroup
*curr
= NULL
;
1451 struct task_struct
*p
;
1454 p
= find_lock_task_mm(task
);
1456 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1460 * All threads may have already detached their mm's, but the oom
1461 * killer still needs to detect if they have already been oom
1462 * killed to prevent needlessly killing additional tasks.
1465 curr
= mem_cgroup_from_task(task
);
1467 css_get(&curr
->css
);
1473 * We should check use_hierarchy of "memcg" not "curr". Because checking
1474 * use_hierarchy of "curr" here make this function true if hierarchy is
1475 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1476 * hierarchy(even if use_hierarchy is disabled in "memcg").
1478 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1479 css_put(&curr
->css
);
1483 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1485 unsigned long inactive_ratio
;
1486 unsigned long inactive
;
1487 unsigned long active
;
1490 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1491 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1493 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1495 inactive_ratio
= int_sqrt(10 * gb
);
1499 return inactive
* inactive_ratio
< active
;
1502 #define mem_cgroup_from_res_counter(counter, member) \
1503 container_of(counter, struct mem_cgroup, member)
1506 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1507 * @memcg: the memory cgroup
1509 * Returns the maximum amount of memory @mem can be charged with, in
1512 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1514 unsigned long long margin
;
1516 margin
= res_counter_margin(&memcg
->res
);
1517 if (do_swap_account
)
1518 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1519 return margin
>> PAGE_SHIFT
;
1522 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1525 if (!css_parent(&memcg
->css
))
1526 return vm_swappiness
;
1528 return memcg
->swappiness
;
1532 * memcg->moving_account is used for checking possibility that some thread is
1533 * calling move_account(). When a thread on CPU-A starts moving pages under
1534 * a memcg, other threads should check memcg->moving_account under
1535 * rcu_read_lock(), like this:
1539 * memcg->moving_account+1 if (memcg->mocing_account)
1541 * synchronize_rcu() update something.
1546 /* for quick checking without looking up memcg */
1547 atomic_t memcg_moving __read_mostly
;
1549 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1551 atomic_inc(&memcg_moving
);
1552 atomic_inc(&memcg
->moving_account
);
1556 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1559 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1560 * We check NULL in callee rather than caller.
1563 atomic_dec(&memcg_moving
);
1564 atomic_dec(&memcg
->moving_account
);
1569 * 2 routines for checking "mem" is under move_account() or not.
1571 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1572 * is used for avoiding races in accounting. If true,
1573 * pc->mem_cgroup may be overwritten.
1575 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1576 * under hierarchy of moving cgroups. This is for
1577 * waiting at hith-memory prressure caused by "move".
1580 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1582 VM_BUG_ON(!rcu_read_lock_held());
1583 return atomic_read(&memcg
->moving_account
) > 0;
1586 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1588 struct mem_cgroup
*from
;
1589 struct mem_cgroup
*to
;
1592 * Unlike task_move routines, we access mc.to, mc.from not under
1593 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1595 spin_lock(&mc
.lock
);
1601 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1602 || mem_cgroup_same_or_subtree(memcg
, to
);
1604 spin_unlock(&mc
.lock
);
1608 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1610 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1611 if (mem_cgroup_under_move(memcg
)) {
1613 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1614 /* moving charge context might have finished. */
1617 finish_wait(&mc
.waitq
, &wait
);
1625 * Take this lock when
1626 * - a code tries to modify page's memcg while it's USED.
1627 * - a code tries to modify page state accounting in a memcg.
1628 * see mem_cgroup_stolen(), too.
1630 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1631 unsigned long *flags
)
1633 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1636 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1637 unsigned long *flags
)
1639 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1642 #define K(x) ((x) << (PAGE_SHIFT-10))
1644 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1645 * @memcg: The memory cgroup that went over limit
1646 * @p: Task that is going to be killed
1648 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1651 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1653 struct cgroup
*task_cgrp
;
1654 struct cgroup
*mem_cgrp
;
1656 * Need a buffer in BSS, can't rely on allocations. The code relies
1657 * on the assumption that OOM is serialized for memory controller.
1658 * If this assumption is broken, revisit this code.
1660 static char memcg_name
[PATH_MAX
];
1662 struct mem_cgroup
*iter
;
1670 mem_cgrp
= memcg
->css
.cgroup
;
1671 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1673 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1676 * Unfortunately, we are unable to convert to a useful name
1677 * But we'll still print out the usage information
1684 pr_info("Task in %s killed", memcg_name
);
1687 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1695 * Continues from above, so we don't need an KERN_ level
1697 pr_cont(" as a result of limit of %s\n", memcg_name
);
1700 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1701 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1702 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1703 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1704 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1705 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1706 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1707 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1708 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1709 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1710 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1711 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1713 for_each_mem_cgroup_tree(iter
, memcg
) {
1714 pr_info("Memory cgroup stats");
1717 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1719 pr_cont(" for %s", memcg_name
);
1723 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1724 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1726 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1727 K(mem_cgroup_read_stat(iter
, i
)));
1730 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1731 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1732 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1739 * This function returns the number of memcg under hierarchy tree. Returns
1740 * 1(self count) if no children.
1742 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1745 struct mem_cgroup
*iter
;
1747 for_each_mem_cgroup_tree(iter
, memcg
)
1753 * Return the memory (and swap, if configured) limit for a memcg.
1755 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1759 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1762 * Do not consider swap space if we cannot swap due to swappiness
1764 if (mem_cgroup_swappiness(memcg
)) {
1767 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1768 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1771 * If memsw is finite and limits the amount of swap space
1772 * available to this memcg, return that limit.
1774 limit
= min(limit
, memsw
);
1780 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1783 struct mem_cgroup
*iter
;
1784 unsigned long chosen_points
= 0;
1785 unsigned long totalpages
;
1786 unsigned int points
= 0;
1787 struct task_struct
*chosen
= NULL
;
1790 * If current has a pending SIGKILL or is exiting, then automatically
1791 * select it. The goal is to allow it to allocate so that it may
1792 * quickly exit and free its memory.
1794 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1795 set_thread_flag(TIF_MEMDIE
);
1799 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1800 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1801 for_each_mem_cgroup_tree(iter
, memcg
) {
1802 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1803 struct cgroup_iter it
;
1804 struct task_struct
*task
;
1806 cgroup_iter_start(cgroup
, &it
);
1807 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1808 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1810 case OOM_SCAN_SELECT
:
1812 put_task_struct(chosen
);
1814 chosen_points
= ULONG_MAX
;
1815 get_task_struct(chosen
);
1817 case OOM_SCAN_CONTINUE
:
1819 case OOM_SCAN_ABORT
:
1820 cgroup_iter_end(cgroup
, &it
);
1821 mem_cgroup_iter_break(memcg
, iter
);
1823 put_task_struct(chosen
);
1828 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1829 if (points
> chosen_points
) {
1831 put_task_struct(chosen
);
1833 chosen_points
= points
;
1834 get_task_struct(chosen
);
1837 cgroup_iter_end(cgroup
, &it
);
1842 points
= chosen_points
* 1000 / totalpages
;
1843 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1844 NULL
, "Memory cgroup out of memory");
1847 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1849 unsigned long flags
)
1851 unsigned long total
= 0;
1852 bool noswap
= false;
1855 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1857 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1860 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1862 drain_all_stock_async(memcg
);
1863 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1865 * Allow limit shrinkers, which are triggered directly
1866 * by userspace, to catch signals and stop reclaim
1867 * after minimal progress, regardless of the margin.
1869 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1871 if (mem_cgroup_margin(memcg
))
1874 * If nothing was reclaimed after two attempts, there
1875 * may be no reclaimable pages in this hierarchy.
1884 * test_mem_cgroup_node_reclaimable
1885 * @memcg: the target memcg
1886 * @nid: the node ID to be checked.
1887 * @noswap : specify true here if the user wants flle only information.
1889 * This function returns whether the specified memcg contains any
1890 * reclaimable pages on a node. Returns true if there are any reclaimable
1891 * pages in the node.
1893 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1894 int nid
, bool noswap
)
1896 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1898 if (noswap
|| !total_swap_pages
)
1900 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1905 #if MAX_NUMNODES > 1
1908 * Always updating the nodemask is not very good - even if we have an empty
1909 * list or the wrong list here, we can start from some node and traverse all
1910 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1913 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1917 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1918 * pagein/pageout changes since the last update.
1920 if (!atomic_read(&memcg
->numainfo_events
))
1922 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1925 /* make a nodemask where this memcg uses memory from */
1926 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1928 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1930 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1931 node_clear(nid
, memcg
->scan_nodes
);
1934 atomic_set(&memcg
->numainfo_events
, 0);
1935 atomic_set(&memcg
->numainfo_updating
, 0);
1939 * Selecting a node where we start reclaim from. Because what we need is just
1940 * reducing usage counter, start from anywhere is O,K. Considering
1941 * memory reclaim from current node, there are pros. and cons.
1943 * Freeing memory from current node means freeing memory from a node which
1944 * we'll use or we've used. So, it may make LRU bad. And if several threads
1945 * hit limits, it will see a contention on a node. But freeing from remote
1946 * node means more costs for memory reclaim because of memory latency.
1948 * Now, we use round-robin. Better algorithm is welcomed.
1950 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1954 mem_cgroup_may_update_nodemask(memcg
);
1955 node
= memcg
->last_scanned_node
;
1957 node
= next_node(node
, memcg
->scan_nodes
);
1958 if (node
== MAX_NUMNODES
)
1959 node
= first_node(memcg
->scan_nodes
);
1961 * We call this when we hit limit, not when pages are added to LRU.
1962 * No LRU may hold pages because all pages are UNEVICTABLE or
1963 * memcg is too small and all pages are not on LRU. In that case,
1964 * we use curret node.
1966 if (unlikely(node
== MAX_NUMNODES
))
1967 node
= numa_node_id();
1969 memcg
->last_scanned_node
= node
;
1974 * Check all nodes whether it contains reclaimable pages or not.
1975 * For quick scan, we make use of scan_nodes. This will allow us to skip
1976 * unused nodes. But scan_nodes is lazily updated and may not cotain
1977 * enough new information. We need to do double check.
1979 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1984 * quick check...making use of scan_node.
1985 * We can skip unused nodes.
1987 if (!nodes_empty(memcg
->scan_nodes
)) {
1988 for (nid
= first_node(memcg
->scan_nodes
);
1990 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1992 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1997 * Check rest of nodes.
1999 for_each_node_state(nid
, N_MEMORY
) {
2000 if (node_isset(nid
, memcg
->scan_nodes
))
2002 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2009 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2014 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2016 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2020 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2023 unsigned long *total_scanned
)
2025 struct mem_cgroup
*victim
= NULL
;
2028 unsigned long excess
;
2029 unsigned long nr_scanned
;
2030 struct mem_cgroup_reclaim_cookie reclaim
= {
2035 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2038 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2043 * If we have not been able to reclaim
2044 * anything, it might because there are
2045 * no reclaimable pages under this hierarchy
2050 * We want to do more targeted reclaim.
2051 * excess >> 2 is not to excessive so as to
2052 * reclaim too much, nor too less that we keep
2053 * coming back to reclaim from this cgroup
2055 if (total
>= (excess
>> 2) ||
2056 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2061 if (!mem_cgroup_reclaimable(victim
, false))
2063 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2065 *total_scanned
+= nr_scanned
;
2066 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2069 mem_cgroup_iter_break(root_memcg
, victim
);
2074 * Check OOM-Killer is already running under our hierarchy.
2075 * If someone is running, return false.
2076 * Has to be called with memcg_oom_lock
2078 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
2080 struct mem_cgroup
*iter
, *failed
= NULL
;
2082 for_each_mem_cgroup_tree(iter
, memcg
) {
2083 if (iter
->oom_lock
) {
2085 * this subtree of our hierarchy is already locked
2086 * so we cannot give a lock.
2089 mem_cgroup_iter_break(memcg
, iter
);
2092 iter
->oom_lock
= true;
2099 * OK, we failed to lock the whole subtree so we have to clean up
2100 * what we set up to the failing subtree
2102 for_each_mem_cgroup_tree(iter
, memcg
) {
2103 if (iter
== failed
) {
2104 mem_cgroup_iter_break(memcg
, iter
);
2107 iter
->oom_lock
= false;
2113 * Has to be called with memcg_oom_lock
2115 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2117 struct mem_cgroup
*iter
;
2119 for_each_mem_cgroup_tree(iter
, memcg
)
2120 iter
->oom_lock
= false;
2124 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2126 struct mem_cgroup
*iter
;
2128 for_each_mem_cgroup_tree(iter
, memcg
)
2129 atomic_inc(&iter
->under_oom
);
2132 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2134 struct mem_cgroup
*iter
;
2137 * When a new child is created while the hierarchy is under oom,
2138 * mem_cgroup_oom_lock() may not be called. We have to use
2139 * atomic_add_unless() here.
2141 for_each_mem_cgroup_tree(iter
, memcg
)
2142 atomic_add_unless(&iter
->under_oom
, -1, 0);
2145 static DEFINE_SPINLOCK(memcg_oom_lock
);
2146 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2148 struct oom_wait_info
{
2149 struct mem_cgroup
*memcg
;
2153 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2154 unsigned mode
, int sync
, void *arg
)
2156 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2157 struct mem_cgroup
*oom_wait_memcg
;
2158 struct oom_wait_info
*oom_wait_info
;
2160 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2161 oom_wait_memcg
= oom_wait_info
->memcg
;
2164 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2165 * Then we can use css_is_ancestor without taking care of RCU.
2167 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2168 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2170 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2173 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2175 /* for filtering, pass "memcg" as argument. */
2176 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2179 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2181 if (memcg
&& atomic_read(&memcg
->under_oom
))
2182 memcg_wakeup_oom(memcg
);
2186 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2188 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2191 struct oom_wait_info owait
;
2192 bool locked
, need_to_kill
;
2194 owait
.memcg
= memcg
;
2195 owait
.wait
.flags
= 0;
2196 owait
.wait
.func
= memcg_oom_wake_function
;
2197 owait
.wait
.private = current
;
2198 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2199 need_to_kill
= true;
2200 mem_cgroup_mark_under_oom(memcg
);
2202 /* At first, try to OOM lock hierarchy under memcg.*/
2203 spin_lock(&memcg_oom_lock
);
2204 locked
= mem_cgroup_oom_lock(memcg
);
2206 * Even if signal_pending(), we can't quit charge() loop without
2207 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2208 * under OOM is always welcomed, use TASK_KILLABLE here.
2210 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2211 if (!locked
|| memcg
->oom_kill_disable
)
2212 need_to_kill
= false;
2214 mem_cgroup_oom_notify(memcg
);
2215 spin_unlock(&memcg_oom_lock
);
2218 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2219 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2222 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2224 spin_lock(&memcg_oom_lock
);
2226 mem_cgroup_oom_unlock(memcg
);
2227 memcg_wakeup_oom(memcg
);
2228 spin_unlock(&memcg_oom_lock
);
2230 mem_cgroup_unmark_under_oom(memcg
);
2232 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2234 /* Give chance to dying process */
2235 schedule_timeout_uninterruptible(1);
2240 * Currently used to update mapped file statistics, but the routine can be
2241 * generalized to update other statistics as well.
2243 * Notes: Race condition
2245 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2246 * it tends to be costly. But considering some conditions, we doesn't need
2247 * to do so _always_.
2249 * Considering "charge", lock_page_cgroup() is not required because all
2250 * file-stat operations happen after a page is attached to radix-tree. There
2251 * are no race with "charge".
2253 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2254 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2255 * if there are race with "uncharge". Statistics itself is properly handled
2258 * Considering "move", this is an only case we see a race. To make the race
2259 * small, we check mm->moving_account and detect there are possibility of race
2260 * If there is, we take a lock.
2263 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2264 bool *locked
, unsigned long *flags
)
2266 struct mem_cgroup
*memcg
;
2267 struct page_cgroup
*pc
;
2269 pc
= lookup_page_cgroup(page
);
2271 memcg
= pc
->mem_cgroup
;
2272 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2275 * If this memory cgroup is not under account moving, we don't
2276 * need to take move_lock_mem_cgroup(). Because we already hold
2277 * rcu_read_lock(), any calls to move_account will be delayed until
2278 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2280 if (!mem_cgroup_stolen(memcg
))
2283 move_lock_mem_cgroup(memcg
, flags
);
2284 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2285 move_unlock_mem_cgroup(memcg
, flags
);
2291 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2293 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2296 * It's guaranteed that pc->mem_cgroup never changes while
2297 * lock is held because a routine modifies pc->mem_cgroup
2298 * should take move_lock_mem_cgroup().
2300 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2303 void mem_cgroup_update_page_stat(struct page
*page
,
2304 enum mem_cgroup_page_stat_item idx
, int val
)
2306 struct mem_cgroup
*memcg
;
2307 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2308 unsigned long uninitialized_var(flags
);
2310 if (mem_cgroup_disabled())
2313 memcg
= pc
->mem_cgroup
;
2314 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2318 case MEMCG_NR_FILE_MAPPED
:
2319 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2325 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2329 * size of first charge trial. "32" comes from vmscan.c's magic value.
2330 * TODO: maybe necessary to use big numbers in big irons.
2332 #define CHARGE_BATCH 32U
2333 struct memcg_stock_pcp
{
2334 struct mem_cgroup
*cached
; /* this never be root cgroup */
2335 unsigned int nr_pages
;
2336 struct work_struct work
;
2337 unsigned long flags
;
2338 #define FLUSHING_CACHED_CHARGE 0
2340 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2341 static DEFINE_MUTEX(percpu_charge_mutex
);
2344 * consume_stock: Try to consume stocked charge on this cpu.
2345 * @memcg: memcg to consume from.
2346 * @nr_pages: how many pages to charge.
2348 * The charges will only happen if @memcg matches the current cpu's memcg
2349 * stock, and at least @nr_pages are available in that stock. Failure to
2350 * service an allocation will refill the stock.
2352 * returns true if successful, false otherwise.
2354 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2356 struct memcg_stock_pcp
*stock
;
2359 if (nr_pages
> CHARGE_BATCH
)
2362 stock
= &get_cpu_var(memcg_stock
);
2363 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2364 stock
->nr_pages
-= nr_pages
;
2365 else /* need to call res_counter_charge */
2367 put_cpu_var(memcg_stock
);
2372 * Returns stocks cached in percpu to res_counter and reset cached information.
2374 static void drain_stock(struct memcg_stock_pcp
*stock
)
2376 struct mem_cgroup
*old
= stock
->cached
;
2378 if (stock
->nr_pages
) {
2379 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2381 res_counter_uncharge(&old
->res
, bytes
);
2382 if (do_swap_account
)
2383 res_counter_uncharge(&old
->memsw
, bytes
);
2384 stock
->nr_pages
= 0;
2386 stock
->cached
= NULL
;
2390 * This must be called under preempt disabled or must be called by
2391 * a thread which is pinned to local cpu.
2393 static void drain_local_stock(struct work_struct
*dummy
)
2395 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2397 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2400 static void __init
memcg_stock_init(void)
2404 for_each_possible_cpu(cpu
) {
2405 struct memcg_stock_pcp
*stock
=
2406 &per_cpu(memcg_stock
, cpu
);
2407 INIT_WORK(&stock
->work
, drain_local_stock
);
2412 * Cache charges(val) which is from res_counter, to local per_cpu area.
2413 * This will be consumed by consume_stock() function, later.
2415 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2417 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2419 if (stock
->cached
!= memcg
) { /* reset if necessary */
2421 stock
->cached
= memcg
;
2423 stock
->nr_pages
+= nr_pages
;
2424 put_cpu_var(memcg_stock
);
2428 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2429 * of the hierarchy under it. sync flag says whether we should block
2430 * until the work is done.
2432 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2436 /* Notify other cpus that system-wide "drain" is running */
2439 for_each_online_cpu(cpu
) {
2440 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2441 struct mem_cgroup
*memcg
;
2443 memcg
= stock
->cached
;
2444 if (!memcg
|| !stock
->nr_pages
)
2446 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2448 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2450 drain_local_stock(&stock
->work
);
2452 schedule_work_on(cpu
, &stock
->work
);
2460 for_each_online_cpu(cpu
) {
2461 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2462 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2463 flush_work(&stock
->work
);
2470 * Tries to drain stocked charges in other cpus. This function is asynchronous
2471 * and just put a work per cpu for draining localy on each cpu. Caller can
2472 * expects some charges will be back to res_counter later but cannot wait for
2475 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2478 * If someone calls draining, avoid adding more kworker runs.
2480 if (!mutex_trylock(&percpu_charge_mutex
))
2482 drain_all_stock(root_memcg
, false);
2483 mutex_unlock(&percpu_charge_mutex
);
2486 /* This is a synchronous drain interface. */
2487 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2489 /* called when force_empty is called */
2490 mutex_lock(&percpu_charge_mutex
);
2491 drain_all_stock(root_memcg
, true);
2492 mutex_unlock(&percpu_charge_mutex
);
2496 * This function drains percpu counter value from DEAD cpu and
2497 * move it to local cpu. Note that this function can be preempted.
2499 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2503 spin_lock(&memcg
->pcp_counter_lock
);
2504 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2505 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2507 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2508 memcg
->nocpu_base
.count
[i
] += x
;
2510 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2511 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2513 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2514 memcg
->nocpu_base
.events
[i
] += x
;
2516 spin_unlock(&memcg
->pcp_counter_lock
);
2519 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2520 unsigned long action
,
2523 int cpu
= (unsigned long)hcpu
;
2524 struct memcg_stock_pcp
*stock
;
2525 struct mem_cgroup
*iter
;
2527 if (action
== CPU_ONLINE
)
2530 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2533 for_each_mem_cgroup(iter
)
2534 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2536 stock
= &per_cpu(memcg_stock
, cpu
);
2542 /* See __mem_cgroup_try_charge() for details */
2544 CHARGE_OK
, /* success */
2545 CHARGE_RETRY
, /* need to retry but retry is not bad */
2546 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2547 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2548 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2551 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2552 unsigned int nr_pages
, unsigned int min_pages
,
2555 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2556 struct mem_cgroup
*mem_over_limit
;
2557 struct res_counter
*fail_res
;
2558 unsigned long flags
= 0;
2561 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2564 if (!do_swap_account
)
2566 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2570 res_counter_uncharge(&memcg
->res
, csize
);
2571 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2572 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2574 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2576 * Never reclaim on behalf of optional batching, retry with a
2577 * single page instead.
2579 if (nr_pages
> min_pages
)
2580 return CHARGE_RETRY
;
2582 if (!(gfp_mask
& __GFP_WAIT
))
2583 return CHARGE_WOULDBLOCK
;
2585 if (gfp_mask
& __GFP_NORETRY
)
2586 return CHARGE_NOMEM
;
2588 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2589 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2590 return CHARGE_RETRY
;
2592 * Even though the limit is exceeded at this point, reclaim
2593 * may have been able to free some pages. Retry the charge
2594 * before killing the task.
2596 * Only for regular pages, though: huge pages are rather
2597 * unlikely to succeed so close to the limit, and we fall back
2598 * to regular pages anyway in case of failure.
2600 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2601 return CHARGE_RETRY
;
2604 * At task move, charge accounts can be doubly counted. So, it's
2605 * better to wait until the end of task_move if something is going on.
2607 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2608 return CHARGE_RETRY
;
2610 /* If we don't need to call oom-killer at el, return immediately */
2612 return CHARGE_NOMEM
;
2614 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2615 return CHARGE_OOM_DIE
;
2617 return CHARGE_RETRY
;
2621 * __mem_cgroup_try_charge() does
2622 * 1. detect memcg to be charged against from passed *mm and *ptr,
2623 * 2. update res_counter
2624 * 3. call memory reclaim if necessary.
2626 * In some special case, if the task is fatal, fatal_signal_pending() or
2627 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2628 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2629 * as possible without any hazards. 2: all pages should have a valid
2630 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2631 * pointer, that is treated as a charge to root_mem_cgroup.
2633 * So __mem_cgroup_try_charge() will return
2634 * 0 ... on success, filling *ptr with a valid memcg pointer.
2635 * -ENOMEM ... charge failure because of resource limits.
2636 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2638 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2639 * the oom-killer can be invoked.
2641 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2643 unsigned int nr_pages
,
2644 struct mem_cgroup
**ptr
,
2647 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2648 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2649 struct mem_cgroup
*memcg
= NULL
;
2653 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2654 * in system level. So, allow to go ahead dying process in addition to
2657 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2658 || fatal_signal_pending(current
)))
2662 * We always charge the cgroup the mm_struct belongs to.
2663 * The mm_struct's mem_cgroup changes on task migration if the
2664 * thread group leader migrates. It's possible that mm is not
2665 * set, if so charge the root memcg (happens for pagecache usage).
2668 *ptr
= root_mem_cgroup
;
2670 if (*ptr
) { /* css should be a valid one */
2672 if (mem_cgroup_is_root(memcg
))
2674 if (consume_stock(memcg
, nr_pages
))
2676 css_get(&memcg
->css
);
2678 struct task_struct
*p
;
2681 p
= rcu_dereference(mm
->owner
);
2683 * Because we don't have task_lock(), "p" can exit.
2684 * In that case, "memcg" can point to root or p can be NULL with
2685 * race with swapoff. Then, we have small risk of mis-accouning.
2686 * But such kind of mis-account by race always happens because
2687 * we don't have cgroup_mutex(). It's overkill and we allo that
2689 * (*) swapoff at el will charge against mm-struct not against
2690 * task-struct. So, mm->owner can be NULL.
2692 memcg
= mem_cgroup_from_task(p
);
2694 memcg
= root_mem_cgroup
;
2695 if (mem_cgroup_is_root(memcg
)) {
2699 if (consume_stock(memcg
, nr_pages
)) {
2701 * It seems dagerous to access memcg without css_get().
2702 * But considering how consume_stok works, it's not
2703 * necessary. If consume_stock success, some charges
2704 * from this memcg are cached on this cpu. So, we
2705 * don't need to call css_get()/css_tryget() before
2706 * calling consume_stock().
2711 /* after here, we may be blocked. we need to get refcnt */
2712 if (!css_tryget(&memcg
->css
)) {
2722 /* If killed, bypass charge */
2723 if (fatal_signal_pending(current
)) {
2724 css_put(&memcg
->css
);
2729 if (oom
&& !nr_oom_retries
) {
2731 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2734 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2739 case CHARGE_RETRY
: /* not in OOM situation but retry */
2741 css_put(&memcg
->css
);
2744 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2745 css_put(&memcg
->css
);
2747 case CHARGE_NOMEM
: /* OOM routine works */
2749 css_put(&memcg
->css
);
2752 /* If oom, we never return -ENOMEM */
2755 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2756 css_put(&memcg
->css
);
2759 } while (ret
!= CHARGE_OK
);
2761 if (batch
> nr_pages
)
2762 refill_stock(memcg
, batch
- nr_pages
);
2763 css_put(&memcg
->css
);
2771 *ptr
= root_mem_cgroup
;
2776 * Somemtimes we have to undo a charge we got by try_charge().
2777 * This function is for that and do uncharge, put css's refcnt.
2778 * gotten by try_charge().
2780 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2781 unsigned int nr_pages
)
2783 if (!mem_cgroup_is_root(memcg
)) {
2784 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2786 res_counter_uncharge(&memcg
->res
, bytes
);
2787 if (do_swap_account
)
2788 res_counter_uncharge(&memcg
->memsw
, bytes
);
2793 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2794 * This is useful when moving usage to parent cgroup.
2796 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2797 unsigned int nr_pages
)
2799 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2801 if (mem_cgroup_is_root(memcg
))
2804 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2805 if (do_swap_account
)
2806 res_counter_uncharge_until(&memcg
->memsw
,
2807 memcg
->memsw
.parent
, bytes
);
2811 * A helper function to get mem_cgroup from ID. must be called under
2812 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2813 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2814 * called against removed memcg.)
2816 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2818 struct cgroup_subsys_state
*css
;
2820 /* ID 0 is unused ID */
2823 css
= css_lookup(&mem_cgroup_subsys
, id
);
2826 return mem_cgroup_from_css(css
);
2829 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2831 struct mem_cgroup
*memcg
= NULL
;
2832 struct page_cgroup
*pc
;
2836 VM_BUG_ON(!PageLocked(page
));
2838 pc
= lookup_page_cgroup(page
);
2839 lock_page_cgroup(pc
);
2840 if (PageCgroupUsed(pc
)) {
2841 memcg
= pc
->mem_cgroup
;
2842 if (memcg
&& !css_tryget(&memcg
->css
))
2844 } else if (PageSwapCache(page
)) {
2845 ent
.val
= page_private(page
);
2846 id
= lookup_swap_cgroup_id(ent
);
2848 memcg
= mem_cgroup_lookup(id
);
2849 if (memcg
&& !css_tryget(&memcg
->css
))
2853 unlock_page_cgroup(pc
);
2857 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2859 unsigned int nr_pages
,
2860 enum charge_type ctype
,
2863 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2864 struct zone
*uninitialized_var(zone
);
2865 struct lruvec
*lruvec
;
2866 bool was_on_lru
= false;
2869 lock_page_cgroup(pc
);
2870 VM_BUG_ON(PageCgroupUsed(pc
));
2872 * we don't need page_cgroup_lock about tail pages, becase they are not
2873 * accessed by any other context at this point.
2877 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2878 * may already be on some other mem_cgroup's LRU. Take care of it.
2881 zone
= page_zone(page
);
2882 spin_lock_irq(&zone
->lru_lock
);
2883 if (PageLRU(page
)) {
2884 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2886 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2891 pc
->mem_cgroup
= memcg
;
2893 * We access a page_cgroup asynchronously without lock_page_cgroup().
2894 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2895 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2896 * before USED bit, we need memory barrier here.
2897 * See mem_cgroup_add_lru_list(), etc.
2900 SetPageCgroupUsed(pc
);
2904 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2905 VM_BUG_ON(PageLRU(page
));
2907 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2909 spin_unlock_irq(&zone
->lru_lock
);
2912 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2917 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2918 unlock_page_cgroup(pc
);
2921 * "charge_statistics" updated event counter. Then, check it.
2922 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2923 * if they exceeds softlimit.
2925 memcg_check_events(memcg
, page
);
2928 static DEFINE_MUTEX(set_limit_mutex
);
2930 #ifdef CONFIG_MEMCG_KMEM
2931 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2933 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2934 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2938 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2939 * in the memcg_cache_params struct.
2941 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2943 struct kmem_cache
*cachep
;
2945 VM_BUG_ON(p
->is_root_cache
);
2946 cachep
= p
->root_cache
;
2947 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2950 #ifdef CONFIG_SLABINFO
2951 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2952 struct cftype
*cft
, struct seq_file
*m
)
2954 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2955 struct memcg_cache_params
*params
;
2957 if (!memcg_can_account_kmem(memcg
))
2960 print_slabinfo_header(m
);
2962 mutex_lock(&memcg
->slab_caches_mutex
);
2963 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2964 cache_show(memcg_params_to_cache(params
), m
);
2965 mutex_unlock(&memcg
->slab_caches_mutex
);
2971 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2973 struct res_counter
*fail_res
;
2974 struct mem_cgroup
*_memcg
;
2978 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2983 * Conditions under which we can wait for the oom_killer. Those are
2984 * the same conditions tested by the core page allocator
2986 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2989 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2992 if (ret
== -EINTR
) {
2994 * __mem_cgroup_try_charge() chosed to bypass to root due to
2995 * OOM kill or fatal signal. Since our only options are to
2996 * either fail the allocation or charge it to this cgroup, do
2997 * it as a temporary condition. But we can't fail. From a
2998 * kmem/slab perspective, the cache has already been selected,
2999 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3002 * This condition will only trigger if the task entered
3003 * memcg_charge_kmem in a sane state, but was OOM-killed during
3004 * __mem_cgroup_try_charge() above. Tasks that were already
3005 * dying when the allocation triggers should have been already
3006 * directed to the root cgroup in memcontrol.h
3008 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3009 if (do_swap_account
)
3010 res_counter_charge_nofail(&memcg
->memsw
, size
,
3014 res_counter_uncharge(&memcg
->kmem
, size
);
3019 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3021 res_counter_uncharge(&memcg
->res
, size
);
3022 if (do_swap_account
)
3023 res_counter_uncharge(&memcg
->memsw
, size
);
3026 if (res_counter_uncharge(&memcg
->kmem
, size
))
3030 * Releases a reference taken in kmem_cgroup_css_offline in case
3031 * this last uncharge is racing with the offlining code or it is
3032 * outliving the memcg existence.
3034 * The memory barrier imposed by test&clear is paired with the
3035 * explicit one in memcg_kmem_mark_dead().
3037 if (memcg_kmem_test_and_clear_dead(memcg
))
3038 css_put(&memcg
->css
);
3041 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3046 mutex_lock(&memcg
->slab_caches_mutex
);
3047 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3048 mutex_unlock(&memcg
->slab_caches_mutex
);
3052 * helper for acessing a memcg's index. It will be used as an index in the
3053 * child cache array in kmem_cache, and also to derive its name. This function
3054 * will return -1 when this is not a kmem-limited memcg.
3056 int memcg_cache_id(struct mem_cgroup
*memcg
)
3058 return memcg
? memcg
->kmemcg_id
: -1;
3062 * This ends up being protected by the set_limit mutex, during normal
3063 * operation, because that is its main call site.
3065 * But when we create a new cache, we can call this as well if its parent
3066 * is kmem-limited. That will have to hold set_limit_mutex as well.
3068 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3072 num
= ida_simple_get(&kmem_limited_groups
,
3073 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3077 * After this point, kmem_accounted (that we test atomically in
3078 * the beginning of this conditional), is no longer 0. This
3079 * guarantees only one process will set the following boolean
3080 * to true. We don't need test_and_set because we're protected
3081 * by the set_limit_mutex anyway.
3083 memcg_kmem_set_activated(memcg
);
3085 ret
= memcg_update_all_caches(num
+1);
3087 ida_simple_remove(&kmem_limited_groups
, num
);
3088 memcg_kmem_clear_activated(memcg
);
3092 memcg
->kmemcg_id
= num
;
3093 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3094 mutex_init(&memcg
->slab_caches_mutex
);
3098 static size_t memcg_caches_array_size(int num_groups
)
3101 if (num_groups
<= 0)
3104 size
= 2 * num_groups
;
3105 if (size
< MEMCG_CACHES_MIN_SIZE
)
3106 size
= MEMCG_CACHES_MIN_SIZE
;
3107 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3108 size
= MEMCG_CACHES_MAX_SIZE
;
3114 * We should update the current array size iff all caches updates succeed. This
3115 * can only be done from the slab side. The slab mutex needs to be held when
3118 void memcg_update_array_size(int num
)
3120 if (num
> memcg_limited_groups_array_size
)
3121 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3124 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3126 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3128 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3130 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3132 if (num_groups
> memcg_limited_groups_array_size
) {
3134 ssize_t size
= memcg_caches_array_size(num_groups
);
3136 size
*= sizeof(void *);
3137 size
+= sizeof(struct memcg_cache_params
);
3139 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3140 if (!s
->memcg_params
) {
3141 s
->memcg_params
= cur_params
;
3145 s
->memcg_params
->is_root_cache
= true;
3148 * There is the chance it will be bigger than
3149 * memcg_limited_groups_array_size, if we failed an allocation
3150 * in a cache, in which case all caches updated before it, will
3151 * have a bigger array.
3153 * But if that is the case, the data after
3154 * memcg_limited_groups_array_size is certainly unused
3156 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3157 if (!cur_params
->memcg_caches
[i
])
3159 s
->memcg_params
->memcg_caches
[i
] =
3160 cur_params
->memcg_caches
[i
];
3164 * Ideally, we would wait until all caches succeed, and only
3165 * then free the old one. But this is not worth the extra
3166 * pointer per-cache we'd have to have for this.
3168 * It is not a big deal if some caches are left with a size
3169 * bigger than the others. And all updates will reset this
3177 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3178 struct kmem_cache
*root_cache
)
3180 size_t size
= sizeof(struct memcg_cache_params
);
3182 if (!memcg_kmem_enabled())
3186 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3188 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3189 if (!s
->memcg_params
)
3192 INIT_WORK(&s
->memcg_params
->destroy
,
3193 kmem_cache_destroy_work_func
);
3195 s
->memcg_params
->memcg
= memcg
;
3196 s
->memcg_params
->root_cache
= root_cache
;
3198 s
->memcg_params
->is_root_cache
= true;
3203 void memcg_release_cache(struct kmem_cache
*s
)
3205 struct kmem_cache
*root
;
3206 struct mem_cgroup
*memcg
;
3210 * This happens, for instance, when a root cache goes away before we
3213 if (!s
->memcg_params
)
3216 if (s
->memcg_params
->is_root_cache
)
3219 memcg
= s
->memcg_params
->memcg
;
3220 id
= memcg_cache_id(memcg
);
3222 root
= s
->memcg_params
->root_cache
;
3223 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3225 mutex_lock(&memcg
->slab_caches_mutex
);
3226 list_del(&s
->memcg_params
->list
);
3227 mutex_unlock(&memcg
->slab_caches_mutex
);
3229 css_put(&memcg
->css
);
3231 kfree(s
->memcg_params
);
3235 * During the creation a new cache, we need to disable our accounting mechanism
3236 * altogether. This is true even if we are not creating, but rather just
3237 * enqueing new caches to be created.
3239 * This is because that process will trigger allocations; some visible, like
3240 * explicit kmallocs to auxiliary data structures, name strings and internal
3241 * cache structures; some well concealed, like INIT_WORK() that can allocate
3242 * objects during debug.
3244 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3245 * to it. This may not be a bounded recursion: since the first cache creation
3246 * failed to complete (waiting on the allocation), we'll just try to create the
3247 * cache again, failing at the same point.
3249 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3250 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3251 * inside the following two functions.
3253 static inline void memcg_stop_kmem_account(void)
3255 VM_BUG_ON(!current
->mm
);
3256 current
->memcg_kmem_skip_account
++;
3259 static inline void memcg_resume_kmem_account(void)
3261 VM_BUG_ON(!current
->mm
);
3262 current
->memcg_kmem_skip_account
--;
3265 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3267 struct kmem_cache
*cachep
;
3268 struct memcg_cache_params
*p
;
3270 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3272 cachep
= memcg_params_to_cache(p
);
3275 * If we get down to 0 after shrink, we could delete right away.
3276 * However, memcg_release_pages() already puts us back in the workqueue
3277 * in that case. If we proceed deleting, we'll get a dangling
3278 * reference, and removing the object from the workqueue in that case
3279 * is unnecessary complication. We are not a fast path.
3281 * Note that this case is fundamentally different from racing with
3282 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3283 * kmem_cache_shrink, not only we would be reinserting a dead cache
3284 * into the queue, but doing so from inside the worker racing to
3287 * So if we aren't down to zero, we'll just schedule a worker and try
3290 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3291 kmem_cache_shrink(cachep
);
3292 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3295 kmem_cache_destroy(cachep
);
3298 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3300 if (!cachep
->memcg_params
->dead
)
3304 * There are many ways in which we can get here.
3306 * We can get to a memory-pressure situation while the delayed work is
3307 * still pending to run. The vmscan shrinkers can then release all
3308 * cache memory and get us to destruction. If this is the case, we'll
3309 * be executed twice, which is a bug (the second time will execute over
3310 * bogus data). In this case, cancelling the work should be fine.
3312 * But we can also get here from the worker itself, if
3313 * kmem_cache_shrink is enough to shake all the remaining objects and
3314 * get the page count to 0. In this case, we'll deadlock if we try to
3315 * cancel the work (the worker runs with an internal lock held, which
3316 * is the same lock we would hold for cancel_work_sync().)
3318 * Since we can't possibly know who got us here, just refrain from
3319 * running if there is already work pending
3321 if (work_pending(&cachep
->memcg_params
->destroy
))
3324 * We have to defer the actual destroying to a workqueue, because
3325 * we might currently be in a context that cannot sleep.
3327 schedule_work(&cachep
->memcg_params
->destroy
);
3331 * This lock protects updaters, not readers. We want readers to be as fast as
3332 * they can, and they will either see NULL or a valid cache value. Our model
3333 * allow them to see NULL, in which case the root memcg will be selected.
3335 * We need this lock because multiple allocations to the same cache from a non
3336 * will span more than one worker. Only one of them can create the cache.
3338 static DEFINE_MUTEX(memcg_cache_mutex
);
3341 * Called with memcg_cache_mutex held
3343 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3344 struct kmem_cache
*s
)
3346 struct kmem_cache
*new;
3347 static char *tmp_name
= NULL
;
3349 lockdep_assert_held(&memcg_cache_mutex
);
3352 * kmem_cache_create_memcg duplicates the given name and
3353 * cgroup_name for this name requires RCU context.
3354 * This static temporary buffer is used to prevent from
3355 * pointless shortliving allocation.
3358 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3364 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3365 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3368 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3369 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3372 new->allocflags
|= __GFP_KMEMCG
;
3377 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3378 struct kmem_cache
*cachep
)
3380 struct kmem_cache
*new_cachep
;
3383 BUG_ON(!memcg_can_account_kmem(memcg
));
3385 idx
= memcg_cache_id(memcg
);
3387 mutex_lock(&memcg_cache_mutex
);
3388 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3390 css_put(&memcg
->css
);
3394 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3395 if (new_cachep
== NULL
) {
3396 new_cachep
= cachep
;
3397 css_put(&memcg
->css
);
3401 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3403 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3405 * the readers won't lock, make sure everybody sees the updated value,
3406 * so they won't put stuff in the queue again for no reason
3410 mutex_unlock(&memcg_cache_mutex
);
3414 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3416 struct kmem_cache
*c
;
3419 if (!s
->memcg_params
)
3421 if (!s
->memcg_params
->is_root_cache
)
3425 * If the cache is being destroyed, we trust that there is no one else
3426 * requesting objects from it. Even if there are, the sanity checks in
3427 * kmem_cache_destroy should caught this ill-case.
3429 * Still, we don't want anyone else freeing memcg_caches under our
3430 * noses, which can happen if a new memcg comes to life. As usual,
3431 * we'll take the set_limit_mutex to protect ourselves against this.
3433 mutex_lock(&set_limit_mutex
);
3434 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3435 c
= s
->memcg_params
->memcg_caches
[i
];
3440 * We will now manually delete the caches, so to avoid races
3441 * we need to cancel all pending destruction workers and
3442 * proceed with destruction ourselves.
3444 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3445 * and that could spawn the workers again: it is likely that
3446 * the cache still have active pages until this very moment.
3447 * This would lead us back to mem_cgroup_destroy_cache.
3449 * But that will not execute at all if the "dead" flag is not
3450 * set, so flip it down to guarantee we are in control.
3452 c
->memcg_params
->dead
= false;
3453 cancel_work_sync(&c
->memcg_params
->destroy
);
3454 kmem_cache_destroy(c
);
3456 mutex_unlock(&set_limit_mutex
);
3459 struct create_work
{
3460 struct mem_cgroup
*memcg
;
3461 struct kmem_cache
*cachep
;
3462 struct work_struct work
;
3465 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3467 struct kmem_cache
*cachep
;
3468 struct memcg_cache_params
*params
;
3470 if (!memcg_kmem_is_active(memcg
))
3473 mutex_lock(&memcg
->slab_caches_mutex
);
3474 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3475 cachep
= memcg_params_to_cache(params
);
3476 cachep
->memcg_params
->dead
= true;
3477 schedule_work(&cachep
->memcg_params
->destroy
);
3479 mutex_unlock(&memcg
->slab_caches_mutex
);
3482 static void memcg_create_cache_work_func(struct work_struct
*w
)
3484 struct create_work
*cw
;
3486 cw
= container_of(w
, struct create_work
, work
);
3487 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3492 * Enqueue the creation of a per-memcg kmem_cache.
3494 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3495 struct kmem_cache
*cachep
)
3497 struct create_work
*cw
;
3499 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3501 css_put(&memcg
->css
);
3506 cw
->cachep
= cachep
;
3508 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3509 schedule_work(&cw
->work
);
3512 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3513 struct kmem_cache
*cachep
)
3516 * We need to stop accounting when we kmalloc, because if the
3517 * corresponding kmalloc cache is not yet created, the first allocation
3518 * in __memcg_create_cache_enqueue will recurse.
3520 * However, it is better to enclose the whole function. Depending on
3521 * the debugging options enabled, INIT_WORK(), for instance, can
3522 * trigger an allocation. This too, will make us recurse. Because at
3523 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3524 * the safest choice is to do it like this, wrapping the whole function.
3526 memcg_stop_kmem_account();
3527 __memcg_create_cache_enqueue(memcg
, cachep
);
3528 memcg_resume_kmem_account();
3531 * Return the kmem_cache we're supposed to use for a slab allocation.
3532 * We try to use the current memcg's version of the cache.
3534 * If the cache does not exist yet, if we are the first user of it,
3535 * we either create it immediately, if possible, or create it asynchronously
3537 * In the latter case, we will let the current allocation go through with
3538 * the original cache.
3540 * Can't be called in interrupt context or from kernel threads.
3541 * This function needs to be called with rcu_read_lock() held.
3543 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3546 struct mem_cgroup
*memcg
;
3549 VM_BUG_ON(!cachep
->memcg_params
);
3550 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3552 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3556 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3558 if (!memcg_can_account_kmem(memcg
))
3561 idx
= memcg_cache_id(memcg
);
3564 * barrier to mare sure we're always seeing the up to date value. The
3565 * code updating memcg_caches will issue a write barrier to match this.
3567 read_barrier_depends();
3568 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3569 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3573 /* The corresponding put will be done in the workqueue. */
3574 if (!css_tryget(&memcg
->css
))
3579 * If we are in a safe context (can wait, and not in interrupt
3580 * context), we could be be predictable and return right away.
3581 * This would guarantee that the allocation being performed
3582 * already belongs in the new cache.
3584 * However, there are some clashes that can arrive from locking.
3585 * For instance, because we acquire the slab_mutex while doing
3586 * kmem_cache_dup, this means no further allocation could happen
3587 * with the slab_mutex held.
3589 * Also, because cache creation issue get_online_cpus(), this
3590 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3591 * that ends up reversed during cpu hotplug. (cpuset allocates
3592 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3593 * better to defer everything.
3595 memcg_create_cache_enqueue(memcg
, cachep
);
3601 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3604 * We need to verify if the allocation against current->mm->owner's memcg is
3605 * possible for the given order. But the page is not allocated yet, so we'll
3606 * need a further commit step to do the final arrangements.
3608 * It is possible for the task to switch cgroups in this mean time, so at
3609 * commit time, we can't rely on task conversion any longer. We'll then use
3610 * the handle argument to return to the caller which cgroup we should commit
3611 * against. We could also return the memcg directly and avoid the pointer
3612 * passing, but a boolean return value gives better semantics considering
3613 * the compiled-out case as well.
3615 * Returning true means the allocation is possible.
3618 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3620 struct mem_cgroup
*memcg
;
3626 * Disabling accounting is only relevant for some specific memcg
3627 * internal allocations. Therefore we would initially not have such
3628 * check here, since direct calls to the page allocator that are marked
3629 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3630 * concerned with cache allocations, and by having this test at
3631 * memcg_kmem_get_cache, we are already able to relay the allocation to
3632 * the root cache and bypass the memcg cache altogether.
3634 * There is one exception, though: the SLUB allocator does not create
3635 * large order caches, but rather service large kmallocs directly from
3636 * the page allocator. Therefore, the following sequence when backed by
3637 * the SLUB allocator:
3639 * memcg_stop_kmem_account();
3640 * kmalloc(<large_number>)
3641 * memcg_resume_kmem_account();
3643 * would effectively ignore the fact that we should skip accounting,
3644 * since it will drive us directly to this function without passing
3645 * through the cache selector memcg_kmem_get_cache. Such large
3646 * allocations are extremely rare but can happen, for instance, for the
3647 * cache arrays. We bring this test here.
3649 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3652 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3655 * very rare case described in mem_cgroup_from_task. Unfortunately there
3656 * isn't much we can do without complicating this too much, and it would
3657 * be gfp-dependent anyway. Just let it go
3659 if (unlikely(!memcg
))
3662 if (!memcg_can_account_kmem(memcg
)) {
3663 css_put(&memcg
->css
);
3667 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3671 css_put(&memcg
->css
);
3675 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3678 struct page_cgroup
*pc
;
3680 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3682 /* The page allocation failed. Revert */
3684 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3688 pc
= lookup_page_cgroup(page
);
3689 lock_page_cgroup(pc
);
3690 pc
->mem_cgroup
= memcg
;
3691 SetPageCgroupUsed(pc
);
3692 unlock_page_cgroup(pc
);
3695 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3697 struct mem_cgroup
*memcg
= NULL
;
3698 struct page_cgroup
*pc
;
3701 pc
= lookup_page_cgroup(page
);
3703 * Fast unlocked return. Theoretically might have changed, have to
3704 * check again after locking.
3706 if (!PageCgroupUsed(pc
))
3709 lock_page_cgroup(pc
);
3710 if (PageCgroupUsed(pc
)) {
3711 memcg
= pc
->mem_cgroup
;
3712 ClearPageCgroupUsed(pc
);
3714 unlock_page_cgroup(pc
);
3717 * We trust that only if there is a memcg associated with the page, it
3718 * is a valid allocation
3723 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3724 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3727 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3730 #endif /* CONFIG_MEMCG_KMEM */
3732 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3734 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3736 * Because tail pages are not marked as "used", set it. We're under
3737 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3738 * charge/uncharge will be never happen and move_account() is done under
3739 * compound_lock(), so we don't have to take care of races.
3741 void mem_cgroup_split_huge_fixup(struct page
*head
)
3743 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3744 struct page_cgroup
*pc
;
3745 struct mem_cgroup
*memcg
;
3748 if (mem_cgroup_disabled())
3751 memcg
= head_pc
->mem_cgroup
;
3752 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3754 pc
->mem_cgroup
= memcg
;
3755 smp_wmb();/* see __commit_charge() */
3756 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3758 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3761 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3764 * mem_cgroup_move_account - move account of the page
3766 * @nr_pages: number of regular pages (>1 for huge pages)
3767 * @pc: page_cgroup of the page.
3768 * @from: mem_cgroup which the page is moved from.
3769 * @to: mem_cgroup which the page is moved to. @from != @to.
3771 * The caller must confirm following.
3772 * - page is not on LRU (isolate_page() is useful.)
3773 * - compound_lock is held when nr_pages > 1
3775 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3778 static int mem_cgroup_move_account(struct page
*page
,
3779 unsigned int nr_pages
,
3780 struct page_cgroup
*pc
,
3781 struct mem_cgroup
*from
,
3782 struct mem_cgroup
*to
)
3784 unsigned long flags
;
3786 bool anon
= PageAnon(page
);
3788 VM_BUG_ON(from
== to
);
3789 VM_BUG_ON(PageLRU(page
));
3791 * The page is isolated from LRU. So, collapse function
3792 * will not handle this page. But page splitting can happen.
3793 * Do this check under compound_page_lock(). The caller should
3797 if (nr_pages
> 1 && !PageTransHuge(page
))
3800 lock_page_cgroup(pc
);
3803 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3806 move_lock_mem_cgroup(from
, &flags
);
3808 if (!anon
&& page_mapped(page
)) {
3809 /* Update mapped_file data for mem_cgroup */
3811 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3812 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3815 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3817 /* caller should have done css_get */
3818 pc
->mem_cgroup
= to
;
3819 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3820 move_unlock_mem_cgroup(from
, &flags
);
3823 unlock_page_cgroup(pc
);
3827 memcg_check_events(to
, page
);
3828 memcg_check_events(from
, page
);
3834 * mem_cgroup_move_parent - moves page to the parent group
3835 * @page: the page to move
3836 * @pc: page_cgroup of the page
3837 * @child: page's cgroup
3839 * move charges to its parent or the root cgroup if the group has no
3840 * parent (aka use_hierarchy==0).
3841 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3842 * mem_cgroup_move_account fails) the failure is always temporary and
3843 * it signals a race with a page removal/uncharge or migration. In the
3844 * first case the page is on the way out and it will vanish from the LRU
3845 * on the next attempt and the call should be retried later.
3846 * Isolation from the LRU fails only if page has been isolated from
3847 * the LRU since we looked at it and that usually means either global
3848 * reclaim or migration going on. The page will either get back to the
3850 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3851 * (!PageCgroupUsed) or moved to a different group. The page will
3852 * disappear in the next attempt.
3854 static int mem_cgroup_move_parent(struct page
*page
,
3855 struct page_cgroup
*pc
,
3856 struct mem_cgroup
*child
)
3858 struct mem_cgroup
*parent
;
3859 unsigned int nr_pages
;
3860 unsigned long uninitialized_var(flags
);
3863 VM_BUG_ON(mem_cgroup_is_root(child
));
3866 if (!get_page_unless_zero(page
))
3868 if (isolate_lru_page(page
))
3871 nr_pages
= hpage_nr_pages(page
);
3873 parent
= parent_mem_cgroup(child
);
3875 * If no parent, move charges to root cgroup.
3878 parent
= root_mem_cgroup
;
3881 VM_BUG_ON(!PageTransHuge(page
));
3882 flags
= compound_lock_irqsave(page
);
3885 ret
= mem_cgroup_move_account(page
, nr_pages
,
3888 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3891 compound_unlock_irqrestore(page
, flags
);
3892 putback_lru_page(page
);
3900 * Charge the memory controller for page usage.
3902 * 0 if the charge was successful
3903 * < 0 if the cgroup is over its limit
3905 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3906 gfp_t gfp_mask
, enum charge_type ctype
)
3908 struct mem_cgroup
*memcg
= NULL
;
3909 unsigned int nr_pages
= 1;
3913 if (PageTransHuge(page
)) {
3914 nr_pages
<<= compound_order(page
);
3915 VM_BUG_ON(!PageTransHuge(page
));
3917 * Never OOM-kill a process for a huge page. The
3918 * fault handler will fall back to regular pages.
3923 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3926 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3930 int mem_cgroup_newpage_charge(struct page
*page
,
3931 struct mm_struct
*mm
, gfp_t gfp_mask
)
3933 if (mem_cgroup_disabled())
3935 VM_BUG_ON(page_mapped(page
));
3936 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3938 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3939 MEM_CGROUP_CHARGE_TYPE_ANON
);
3943 * While swap-in, try_charge -> commit or cancel, the page is locked.
3944 * And when try_charge() successfully returns, one refcnt to memcg without
3945 * struct page_cgroup is acquired. This refcnt will be consumed by
3946 * "commit()" or removed by "cancel()"
3948 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3951 struct mem_cgroup
**memcgp
)
3953 struct mem_cgroup
*memcg
;
3954 struct page_cgroup
*pc
;
3957 pc
= lookup_page_cgroup(page
);
3959 * Every swap fault against a single page tries to charge the
3960 * page, bail as early as possible. shmem_unuse() encounters
3961 * already charged pages, too. The USED bit is protected by
3962 * the page lock, which serializes swap cache removal, which
3963 * in turn serializes uncharging.
3965 if (PageCgroupUsed(pc
))
3967 if (!do_swap_account
)
3969 memcg
= try_get_mem_cgroup_from_page(page
);
3973 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3974 css_put(&memcg
->css
);
3979 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3985 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3986 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3989 if (mem_cgroup_disabled())
3992 * A racing thread's fault, or swapoff, may have already
3993 * updated the pte, and even removed page from swap cache: in
3994 * those cases unuse_pte()'s pte_same() test will fail; but
3995 * there's also a KSM case which does need to charge the page.
3997 if (!PageSwapCache(page
)) {
4000 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4005 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4008 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4010 if (mem_cgroup_disabled())
4014 __mem_cgroup_cancel_charge(memcg
, 1);
4018 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4019 enum charge_type ctype
)
4021 if (mem_cgroup_disabled())
4026 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4028 * Now swap is on-memory. This means this page may be
4029 * counted both as mem and swap....double count.
4030 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4031 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4032 * may call delete_from_swap_cache() before reach here.
4034 if (do_swap_account
&& PageSwapCache(page
)) {
4035 swp_entry_t ent
= {.val
= page_private(page
)};
4036 mem_cgroup_uncharge_swap(ent
);
4040 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4041 struct mem_cgroup
*memcg
)
4043 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4044 MEM_CGROUP_CHARGE_TYPE_ANON
);
4047 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4050 struct mem_cgroup
*memcg
= NULL
;
4051 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4054 if (mem_cgroup_disabled())
4056 if (PageCompound(page
))
4059 if (!PageSwapCache(page
))
4060 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4061 else { /* page is swapcache/shmem */
4062 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4065 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4070 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4071 unsigned int nr_pages
,
4072 const enum charge_type ctype
)
4074 struct memcg_batch_info
*batch
= NULL
;
4075 bool uncharge_memsw
= true;
4077 /* If swapout, usage of swap doesn't decrease */
4078 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4079 uncharge_memsw
= false;
4081 batch
= ¤t
->memcg_batch
;
4083 * In usual, we do css_get() when we remember memcg pointer.
4084 * But in this case, we keep res->usage until end of a series of
4085 * uncharges. Then, it's ok to ignore memcg's refcnt.
4088 batch
->memcg
= memcg
;
4090 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4091 * In those cases, all pages freed continuously can be expected to be in
4092 * the same cgroup and we have chance to coalesce uncharges.
4093 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4094 * because we want to do uncharge as soon as possible.
4097 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4098 goto direct_uncharge
;
4101 goto direct_uncharge
;
4104 * In typical case, batch->memcg == mem. This means we can
4105 * merge a series of uncharges to an uncharge of res_counter.
4106 * If not, we uncharge res_counter ony by one.
4108 if (batch
->memcg
!= memcg
)
4109 goto direct_uncharge
;
4110 /* remember freed charge and uncharge it later */
4113 batch
->memsw_nr_pages
++;
4116 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4118 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4119 if (unlikely(batch
->memcg
!= memcg
))
4120 memcg_oom_recover(memcg
);
4124 * uncharge if !page_mapped(page)
4126 static struct mem_cgroup
*
4127 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4130 struct mem_cgroup
*memcg
= NULL
;
4131 unsigned int nr_pages
= 1;
4132 struct page_cgroup
*pc
;
4135 if (mem_cgroup_disabled())
4138 if (PageTransHuge(page
)) {
4139 nr_pages
<<= compound_order(page
);
4140 VM_BUG_ON(!PageTransHuge(page
));
4143 * Check if our page_cgroup is valid
4145 pc
= lookup_page_cgroup(page
);
4146 if (unlikely(!PageCgroupUsed(pc
)))
4149 lock_page_cgroup(pc
);
4151 memcg
= pc
->mem_cgroup
;
4153 if (!PageCgroupUsed(pc
))
4156 anon
= PageAnon(page
);
4159 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4161 * Generally PageAnon tells if it's the anon statistics to be
4162 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4163 * used before page reached the stage of being marked PageAnon.
4167 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4168 /* See mem_cgroup_prepare_migration() */
4169 if (page_mapped(page
))
4172 * Pages under migration may not be uncharged. But
4173 * end_migration() /must/ be the one uncharging the
4174 * unused post-migration page and so it has to call
4175 * here with the migration bit still set. See the
4176 * res_counter handling below.
4178 if (!end_migration
&& PageCgroupMigration(pc
))
4181 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4182 if (!PageAnon(page
)) { /* Shared memory */
4183 if (page
->mapping
&& !page_is_file_cache(page
))
4185 } else if (page_mapped(page
)) /* Anon */
4192 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4194 ClearPageCgroupUsed(pc
);
4196 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4197 * freed from LRU. This is safe because uncharged page is expected not
4198 * to be reused (freed soon). Exception is SwapCache, it's handled by
4199 * special functions.
4202 unlock_page_cgroup(pc
);
4204 * even after unlock, we have memcg->res.usage here and this memcg
4205 * will never be freed, so it's safe to call css_get().
4207 memcg_check_events(memcg
, page
);
4208 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4209 mem_cgroup_swap_statistics(memcg
, true);
4210 css_get(&memcg
->css
);
4213 * Migration does not charge the res_counter for the
4214 * replacement page, so leave it alone when phasing out the
4215 * page that is unused after the migration.
4217 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4218 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4223 unlock_page_cgroup(pc
);
4227 void mem_cgroup_uncharge_page(struct page
*page
)
4230 if (page_mapped(page
))
4232 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4234 * If the page is in swap cache, uncharge should be deferred
4235 * to the swap path, which also properly accounts swap usage
4236 * and handles memcg lifetime.
4238 * Note that this check is not stable and reclaim may add the
4239 * page to swap cache at any time after this. However, if the
4240 * page is not in swap cache by the time page->mapcount hits
4241 * 0, there won't be any page table references to the swap
4242 * slot, and reclaim will free it and not actually write the
4245 if (PageSwapCache(page
))
4247 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4250 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4252 VM_BUG_ON(page_mapped(page
));
4253 VM_BUG_ON(page
->mapping
);
4254 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4258 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4259 * In that cases, pages are freed continuously and we can expect pages
4260 * are in the same memcg. All these calls itself limits the number of
4261 * pages freed at once, then uncharge_start/end() is called properly.
4262 * This may be called prural(2) times in a context,
4265 void mem_cgroup_uncharge_start(void)
4267 current
->memcg_batch
.do_batch
++;
4268 /* We can do nest. */
4269 if (current
->memcg_batch
.do_batch
== 1) {
4270 current
->memcg_batch
.memcg
= NULL
;
4271 current
->memcg_batch
.nr_pages
= 0;
4272 current
->memcg_batch
.memsw_nr_pages
= 0;
4276 void mem_cgroup_uncharge_end(void)
4278 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4280 if (!batch
->do_batch
)
4284 if (batch
->do_batch
) /* If stacked, do nothing. */
4290 * This "batch->memcg" is valid without any css_get/put etc...
4291 * bacause we hide charges behind us.
4293 if (batch
->nr_pages
)
4294 res_counter_uncharge(&batch
->memcg
->res
,
4295 batch
->nr_pages
* PAGE_SIZE
);
4296 if (batch
->memsw_nr_pages
)
4297 res_counter_uncharge(&batch
->memcg
->memsw
,
4298 batch
->memsw_nr_pages
* PAGE_SIZE
);
4299 memcg_oom_recover(batch
->memcg
);
4300 /* forget this pointer (for sanity check) */
4301 batch
->memcg
= NULL
;
4306 * called after __delete_from_swap_cache() and drop "page" account.
4307 * memcg information is recorded to swap_cgroup of "ent"
4310 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4312 struct mem_cgroup
*memcg
;
4313 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4315 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4316 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4318 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4321 * record memcg information, if swapout && memcg != NULL,
4322 * css_get() was called in uncharge().
4324 if (do_swap_account
&& swapout
&& memcg
)
4325 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4329 #ifdef CONFIG_MEMCG_SWAP
4331 * called from swap_entry_free(). remove record in swap_cgroup and
4332 * uncharge "memsw" account.
4334 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4336 struct mem_cgroup
*memcg
;
4339 if (!do_swap_account
)
4342 id
= swap_cgroup_record(ent
, 0);
4344 memcg
= mem_cgroup_lookup(id
);
4347 * We uncharge this because swap is freed.
4348 * This memcg can be obsolete one. We avoid calling css_tryget
4350 if (!mem_cgroup_is_root(memcg
))
4351 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4352 mem_cgroup_swap_statistics(memcg
, false);
4353 css_put(&memcg
->css
);
4359 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4360 * @entry: swap entry to be moved
4361 * @from: mem_cgroup which the entry is moved from
4362 * @to: mem_cgroup which the entry is moved to
4364 * It succeeds only when the swap_cgroup's record for this entry is the same
4365 * as the mem_cgroup's id of @from.
4367 * Returns 0 on success, -EINVAL on failure.
4369 * The caller must have charged to @to, IOW, called res_counter_charge() about
4370 * both res and memsw, and called css_get().
4372 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4373 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4375 unsigned short old_id
, new_id
;
4377 old_id
= css_id(&from
->css
);
4378 new_id
= css_id(&to
->css
);
4380 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4381 mem_cgroup_swap_statistics(from
, false);
4382 mem_cgroup_swap_statistics(to
, true);
4384 * This function is only called from task migration context now.
4385 * It postpones res_counter and refcount handling till the end
4386 * of task migration(mem_cgroup_clear_mc()) for performance
4387 * improvement. But we cannot postpone css_get(to) because if
4388 * the process that has been moved to @to does swap-in, the
4389 * refcount of @to might be decreased to 0.
4391 * We are in attach() phase, so the cgroup is guaranteed to be
4392 * alive, so we can just call css_get().
4400 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4401 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4408 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4411 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4412 struct mem_cgroup
**memcgp
)
4414 struct mem_cgroup
*memcg
= NULL
;
4415 unsigned int nr_pages
= 1;
4416 struct page_cgroup
*pc
;
4417 enum charge_type ctype
;
4421 if (mem_cgroup_disabled())
4424 if (PageTransHuge(page
))
4425 nr_pages
<<= compound_order(page
);
4427 pc
= lookup_page_cgroup(page
);
4428 lock_page_cgroup(pc
);
4429 if (PageCgroupUsed(pc
)) {
4430 memcg
= pc
->mem_cgroup
;
4431 css_get(&memcg
->css
);
4433 * At migrating an anonymous page, its mapcount goes down
4434 * to 0 and uncharge() will be called. But, even if it's fully
4435 * unmapped, migration may fail and this page has to be
4436 * charged again. We set MIGRATION flag here and delay uncharge
4437 * until end_migration() is called
4439 * Corner Case Thinking
4441 * When the old page was mapped as Anon and it's unmap-and-freed
4442 * while migration was ongoing.
4443 * If unmap finds the old page, uncharge() of it will be delayed
4444 * until end_migration(). If unmap finds a new page, it's
4445 * uncharged when it make mapcount to be 1->0. If unmap code
4446 * finds swap_migration_entry, the new page will not be mapped
4447 * and end_migration() will find it(mapcount==0).
4450 * When the old page was mapped but migraion fails, the kernel
4451 * remaps it. A charge for it is kept by MIGRATION flag even
4452 * if mapcount goes down to 0. We can do remap successfully
4453 * without charging it again.
4456 * The "old" page is under lock_page() until the end of
4457 * migration, so, the old page itself will not be swapped-out.
4458 * If the new page is swapped out before end_migraton, our
4459 * hook to usual swap-out path will catch the event.
4462 SetPageCgroupMigration(pc
);
4464 unlock_page_cgroup(pc
);
4466 * If the page is not charged at this point,
4474 * We charge new page before it's used/mapped. So, even if unlock_page()
4475 * is called before end_migration, we can catch all events on this new
4476 * page. In the case new page is migrated but not remapped, new page's
4477 * mapcount will be finally 0 and we call uncharge in end_migration().
4480 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4482 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4484 * The page is committed to the memcg, but it's not actually
4485 * charged to the res_counter since we plan on replacing the
4486 * old one and only one page is going to be left afterwards.
4488 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4491 /* remove redundant charge if migration failed*/
4492 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4493 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4495 struct page
*used
, *unused
;
4496 struct page_cgroup
*pc
;
4502 if (!migration_ok
) {
4509 anon
= PageAnon(used
);
4510 __mem_cgroup_uncharge_common(unused
,
4511 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4512 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4514 css_put(&memcg
->css
);
4516 * We disallowed uncharge of pages under migration because mapcount
4517 * of the page goes down to zero, temporarly.
4518 * Clear the flag and check the page should be charged.
4520 pc
= lookup_page_cgroup(oldpage
);
4521 lock_page_cgroup(pc
);
4522 ClearPageCgroupMigration(pc
);
4523 unlock_page_cgroup(pc
);
4526 * If a page is a file cache, radix-tree replacement is very atomic
4527 * and we can skip this check. When it was an Anon page, its mapcount
4528 * goes down to 0. But because we added MIGRATION flage, it's not
4529 * uncharged yet. There are several case but page->mapcount check
4530 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4531 * check. (see prepare_charge() also)
4534 mem_cgroup_uncharge_page(used
);
4538 * At replace page cache, newpage is not under any memcg but it's on
4539 * LRU. So, this function doesn't touch res_counter but handles LRU
4540 * in correct way. Both pages are locked so we cannot race with uncharge.
4542 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4543 struct page
*newpage
)
4545 struct mem_cgroup
*memcg
= NULL
;
4546 struct page_cgroup
*pc
;
4547 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4549 if (mem_cgroup_disabled())
4552 pc
= lookup_page_cgroup(oldpage
);
4553 /* fix accounting on old pages */
4554 lock_page_cgroup(pc
);
4555 if (PageCgroupUsed(pc
)) {
4556 memcg
= pc
->mem_cgroup
;
4557 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4558 ClearPageCgroupUsed(pc
);
4560 unlock_page_cgroup(pc
);
4563 * When called from shmem_replace_page(), in some cases the
4564 * oldpage has already been charged, and in some cases not.
4569 * Even if newpage->mapping was NULL before starting replacement,
4570 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4571 * LRU while we overwrite pc->mem_cgroup.
4573 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4576 #ifdef CONFIG_DEBUG_VM
4577 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4579 struct page_cgroup
*pc
;
4581 pc
= lookup_page_cgroup(page
);
4583 * Can be NULL while feeding pages into the page allocator for
4584 * the first time, i.e. during boot or memory hotplug;
4585 * or when mem_cgroup_disabled().
4587 if (likely(pc
) && PageCgroupUsed(pc
))
4592 bool mem_cgroup_bad_page_check(struct page
*page
)
4594 if (mem_cgroup_disabled())
4597 return lookup_page_cgroup_used(page
) != NULL
;
4600 void mem_cgroup_print_bad_page(struct page
*page
)
4602 struct page_cgroup
*pc
;
4604 pc
= lookup_page_cgroup_used(page
);
4606 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4607 pc
, pc
->flags
, pc
->mem_cgroup
);
4612 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4613 unsigned long long val
)
4616 u64 memswlimit
, memlimit
;
4618 int children
= mem_cgroup_count_children(memcg
);
4619 u64 curusage
, oldusage
;
4623 * For keeping hierarchical_reclaim simple, how long we should retry
4624 * is depends on callers. We set our retry-count to be function
4625 * of # of children which we should visit in this loop.
4627 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4629 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4632 while (retry_count
) {
4633 if (signal_pending(current
)) {
4638 * Rather than hide all in some function, I do this in
4639 * open coded manner. You see what this really does.
4640 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4642 mutex_lock(&set_limit_mutex
);
4643 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4644 if (memswlimit
< val
) {
4646 mutex_unlock(&set_limit_mutex
);
4650 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4654 ret
= res_counter_set_limit(&memcg
->res
, val
);
4656 if (memswlimit
== val
)
4657 memcg
->memsw_is_minimum
= true;
4659 memcg
->memsw_is_minimum
= false;
4661 mutex_unlock(&set_limit_mutex
);
4666 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4667 MEM_CGROUP_RECLAIM_SHRINK
);
4668 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4669 /* Usage is reduced ? */
4670 if (curusage
>= oldusage
)
4673 oldusage
= curusage
;
4675 if (!ret
&& enlarge
)
4676 memcg_oom_recover(memcg
);
4681 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4682 unsigned long long val
)
4685 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4686 int children
= mem_cgroup_count_children(memcg
);
4690 /* see mem_cgroup_resize_res_limit */
4691 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4692 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4693 while (retry_count
) {
4694 if (signal_pending(current
)) {
4699 * Rather than hide all in some function, I do this in
4700 * open coded manner. You see what this really does.
4701 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4703 mutex_lock(&set_limit_mutex
);
4704 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4705 if (memlimit
> val
) {
4707 mutex_unlock(&set_limit_mutex
);
4710 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4711 if (memswlimit
< val
)
4713 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4715 if (memlimit
== val
)
4716 memcg
->memsw_is_minimum
= true;
4718 memcg
->memsw_is_minimum
= false;
4720 mutex_unlock(&set_limit_mutex
);
4725 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4726 MEM_CGROUP_RECLAIM_NOSWAP
|
4727 MEM_CGROUP_RECLAIM_SHRINK
);
4728 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4729 /* Usage is reduced ? */
4730 if (curusage
>= oldusage
)
4733 oldusage
= curusage
;
4735 if (!ret
&& enlarge
)
4736 memcg_oom_recover(memcg
);
4740 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4742 unsigned long *total_scanned
)
4744 unsigned long nr_reclaimed
= 0;
4745 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4746 unsigned long reclaimed
;
4748 struct mem_cgroup_tree_per_zone
*mctz
;
4749 unsigned long long excess
;
4750 unsigned long nr_scanned
;
4755 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4757 * This loop can run a while, specially if mem_cgroup's continuously
4758 * keep exceeding their soft limit and putting the system under
4765 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4770 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4771 gfp_mask
, &nr_scanned
);
4772 nr_reclaimed
+= reclaimed
;
4773 *total_scanned
+= nr_scanned
;
4774 spin_lock(&mctz
->lock
);
4777 * If we failed to reclaim anything from this memory cgroup
4778 * it is time to move on to the next cgroup
4784 * Loop until we find yet another one.
4786 * By the time we get the soft_limit lock
4787 * again, someone might have aded the
4788 * group back on the RB tree. Iterate to
4789 * make sure we get a different mem.
4790 * mem_cgroup_largest_soft_limit_node returns
4791 * NULL if no other cgroup is present on
4795 __mem_cgroup_largest_soft_limit_node(mctz
);
4797 css_put(&next_mz
->memcg
->css
);
4798 else /* next_mz == NULL or other memcg */
4802 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4803 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4805 * One school of thought says that we should not add
4806 * back the node to the tree if reclaim returns 0.
4807 * But our reclaim could return 0, simply because due
4808 * to priority we are exposing a smaller subset of
4809 * memory to reclaim from. Consider this as a longer
4812 /* If excess == 0, no tree ops */
4813 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4814 spin_unlock(&mctz
->lock
);
4815 css_put(&mz
->memcg
->css
);
4818 * Could not reclaim anything and there are no more
4819 * mem cgroups to try or we seem to be looping without
4820 * reclaiming anything.
4822 if (!nr_reclaimed
&&
4824 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4826 } while (!nr_reclaimed
);
4828 css_put(&next_mz
->memcg
->css
);
4829 return nr_reclaimed
;
4833 * mem_cgroup_force_empty_list - clears LRU of a group
4834 * @memcg: group to clear
4837 * @lru: lru to to clear
4839 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4840 * reclaim the pages page themselves - pages are moved to the parent (or root)
4843 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4844 int node
, int zid
, enum lru_list lru
)
4846 struct lruvec
*lruvec
;
4847 unsigned long flags
;
4848 struct list_head
*list
;
4852 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4853 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4854 list
= &lruvec
->lists
[lru
];
4858 struct page_cgroup
*pc
;
4861 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4862 if (list_empty(list
)) {
4863 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4866 page
= list_entry(list
->prev
, struct page
, lru
);
4868 list_move(&page
->lru
, list
);
4870 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4873 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4875 pc
= lookup_page_cgroup(page
);
4877 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4878 /* found lock contention or "pc" is obsolete. */
4883 } while (!list_empty(list
));
4887 * make mem_cgroup's charge to be 0 if there is no task by moving
4888 * all the charges and pages to the parent.
4889 * This enables deleting this mem_cgroup.
4891 * Caller is responsible for holding css reference on the memcg.
4893 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4899 /* This is for making all *used* pages to be on LRU. */
4900 lru_add_drain_all();
4901 drain_all_stock_sync(memcg
);
4902 mem_cgroup_start_move(memcg
);
4903 for_each_node_state(node
, N_MEMORY
) {
4904 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4907 mem_cgroup_force_empty_list(memcg
,
4912 mem_cgroup_end_move(memcg
);
4913 memcg_oom_recover(memcg
);
4917 * Kernel memory may not necessarily be trackable to a specific
4918 * process. So they are not migrated, and therefore we can't
4919 * expect their value to drop to 0 here.
4920 * Having res filled up with kmem only is enough.
4922 * This is a safety check because mem_cgroup_force_empty_list
4923 * could have raced with mem_cgroup_replace_page_cache callers
4924 * so the lru seemed empty but the page could have been added
4925 * right after the check. RES_USAGE should be safe as we always
4926 * charge before adding to the LRU.
4928 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4929 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4930 } while (usage
> 0);
4934 * This mainly exists for tests during the setting of set of use_hierarchy.
4935 * Since this is the very setting we are changing, the current hierarchy value
4938 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4940 struct cgroup_subsys_state
*pos
;
4942 /* bounce at first found */
4943 css_for_each_child(pos
, &memcg
->css
)
4949 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4950 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4951 * from mem_cgroup_count_children(), in the sense that we don't really care how
4952 * many children we have; we only need to know if we have any. It also counts
4953 * any memcg without hierarchy as infertile.
4955 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4957 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4961 * Reclaims as many pages from the given memcg as possible and moves
4962 * the rest to the parent.
4964 * Caller is responsible for holding css reference for memcg.
4966 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4968 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4969 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4971 /* returns EBUSY if there is a task or if we come here twice. */
4972 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4975 /* we call try-to-free pages for make this cgroup empty */
4976 lru_add_drain_all();
4977 /* try to free all pages in this cgroup */
4978 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4981 if (signal_pending(current
))
4984 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4988 /* maybe some writeback is necessary */
4989 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4994 mem_cgroup_reparent_charges(memcg
);
4999 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5002 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5005 if (mem_cgroup_is_root(memcg
))
5007 css_get(&memcg
->css
);
5008 ret
= mem_cgroup_force_empty(memcg
);
5009 css_put(&memcg
->css
);
5015 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5018 return mem_cgroup_from_css(css
)->use_hierarchy
;
5021 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5022 struct cftype
*cft
, u64 val
)
5025 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5026 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5028 mutex_lock(&memcg_create_mutex
);
5030 if (memcg
->use_hierarchy
== val
)
5034 * If parent's use_hierarchy is set, we can't make any modifications
5035 * in the child subtrees. If it is unset, then the change can
5036 * occur, provided the current cgroup has no children.
5038 * For the root cgroup, parent_mem is NULL, we allow value to be
5039 * set if there are no children.
5041 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5042 (val
== 1 || val
== 0)) {
5043 if (!__memcg_has_children(memcg
))
5044 memcg
->use_hierarchy
= val
;
5051 mutex_unlock(&memcg_create_mutex
);
5057 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5058 enum mem_cgroup_stat_index idx
)
5060 struct mem_cgroup
*iter
;
5063 /* Per-cpu values can be negative, use a signed accumulator */
5064 for_each_mem_cgroup_tree(iter
, memcg
)
5065 val
+= mem_cgroup_read_stat(iter
, idx
);
5067 if (val
< 0) /* race ? */
5072 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5076 if (!mem_cgroup_is_root(memcg
)) {
5078 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5080 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5084 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5085 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5087 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5088 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5091 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5093 return val
<< PAGE_SHIFT
;
5096 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
5097 struct cftype
*cft
, struct file
*file
,
5098 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
5100 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5106 type
= MEMFILE_TYPE(cft
->private);
5107 name
= MEMFILE_ATTR(cft
->private);
5111 if (name
== RES_USAGE
)
5112 val
= mem_cgroup_usage(memcg
, false);
5114 val
= res_counter_read_u64(&memcg
->res
, name
);
5117 if (name
== RES_USAGE
)
5118 val
= mem_cgroup_usage(memcg
, true);
5120 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5123 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5129 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5130 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5133 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5136 #ifdef CONFIG_MEMCG_KMEM
5137 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5139 * For simplicity, we won't allow this to be disabled. It also can't
5140 * be changed if the cgroup has children already, or if tasks had
5143 * If tasks join before we set the limit, a person looking at
5144 * kmem.usage_in_bytes will have no way to determine when it took
5145 * place, which makes the value quite meaningless.
5147 * After it first became limited, changes in the value of the limit are
5148 * of course permitted.
5150 mutex_lock(&memcg_create_mutex
);
5151 mutex_lock(&set_limit_mutex
);
5152 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
5153 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5157 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5160 ret
= memcg_update_cache_sizes(memcg
);
5162 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
5165 static_key_slow_inc(&memcg_kmem_enabled_key
);
5167 * setting the active bit after the inc will guarantee no one
5168 * starts accounting before all call sites are patched
5170 memcg_kmem_set_active(memcg
);
5172 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5174 mutex_unlock(&set_limit_mutex
);
5175 mutex_unlock(&memcg_create_mutex
);
5180 #ifdef CONFIG_MEMCG_KMEM
5181 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5184 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5188 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5190 * When that happen, we need to disable the static branch only on those
5191 * memcgs that enabled it. To achieve this, we would be forced to
5192 * complicate the code by keeping track of which memcgs were the ones
5193 * that actually enabled limits, and which ones got it from its
5196 * It is a lot simpler just to do static_key_slow_inc() on every child
5197 * that is accounted.
5199 if (!memcg_kmem_is_active(memcg
))
5203 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5204 * memcg is active already. If the later initialization fails then the
5205 * cgroup core triggers the cleanup so we do not have to do it here.
5207 static_key_slow_inc(&memcg_kmem_enabled_key
);
5209 mutex_lock(&set_limit_mutex
);
5210 memcg_stop_kmem_account();
5211 ret
= memcg_update_cache_sizes(memcg
);
5212 memcg_resume_kmem_account();
5213 mutex_unlock(&set_limit_mutex
);
5217 #endif /* CONFIG_MEMCG_KMEM */
5220 * The user of this function is...
5223 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5226 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5229 unsigned long long val
;
5232 type
= MEMFILE_TYPE(cft
->private);
5233 name
= MEMFILE_ATTR(cft
->private);
5237 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5241 /* This function does all necessary parse...reuse it */
5242 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5246 ret
= mem_cgroup_resize_limit(memcg
, val
);
5247 else if (type
== _MEMSWAP
)
5248 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5249 else if (type
== _KMEM
)
5250 ret
= memcg_update_kmem_limit(css
, val
);
5254 case RES_SOFT_LIMIT
:
5255 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5259 * For memsw, soft limits are hard to implement in terms
5260 * of semantics, for now, we support soft limits for
5261 * control without swap
5264 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5269 ret
= -EINVAL
; /* should be BUG() ? */
5275 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5276 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5278 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5280 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5281 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5282 if (!memcg
->use_hierarchy
)
5285 while (css_parent(&memcg
->css
)) {
5286 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5287 if (!memcg
->use_hierarchy
)
5289 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5290 min_limit
= min(min_limit
, tmp
);
5291 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5292 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5295 *mem_limit
= min_limit
;
5296 *memsw_limit
= min_memsw_limit
;
5299 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5301 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5305 type
= MEMFILE_TYPE(event
);
5306 name
= MEMFILE_ATTR(event
);
5311 res_counter_reset_max(&memcg
->res
);
5312 else if (type
== _MEMSWAP
)
5313 res_counter_reset_max(&memcg
->memsw
);
5314 else if (type
== _KMEM
)
5315 res_counter_reset_max(&memcg
->kmem
);
5321 res_counter_reset_failcnt(&memcg
->res
);
5322 else if (type
== _MEMSWAP
)
5323 res_counter_reset_failcnt(&memcg
->memsw
);
5324 else if (type
== _KMEM
)
5325 res_counter_reset_failcnt(&memcg
->kmem
);
5334 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5337 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5341 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5342 struct cftype
*cft
, u64 val
)
5344 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5346 if (val
>= (1 << NR_MOVE_TYPE
))
5350 * No kind of locking is needed in here, because ->can_attach() will
5351 * check this value once in the beginning of the process, and then carry
5352 * on with stale data. This means that changes to this value will only
5353 * affect task migrations starting after the change.
5355 memcg
->move_charge_at_immigrate
= val
;
5359 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5360 struct cftype
*cft
, u64 val
)
5367 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5368 struct cftype
*cft
, struct seq_file
*m
)
5371 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5372 unsigned long node_nr
;
5373 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5375 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5376 seq_printf(m
, "total=%lu", total_nr
);
5377 for_each_node_state(nid
, N_MEMORY
) {
5378 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5379 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5383 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5384 seq_printf(m
, "file=%lu", file_nr
);
5385 for_each_node_state(nid
, N_MEMORY
) {
5386 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5388 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5392 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5393 seq_printf(m
, "anon=%lu", anon_nr
);
5394 for_each_node_state(nid
, N_MEMORY
) {
5395 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5397 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5401 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5402 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5403 for_each_node_state(nid
, N_MEMORY
) {
5404 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5405 BIT(LRU_UNEVICTABLE
));
5406 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5411 #endif /* CONFIG_NUMA */
5413 static inline void mem_cgroup_lru_names_not_uptodate(void)
5415 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5418 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5421 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5422 struct mem_cgroup
*mi
;
5425 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5426 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5428 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5429 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5432 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5433 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5434 mem_cgroup_read_events(memcg
, i
));
5436 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5437 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5438 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5440 /* Hierarchical information */
5442 unsigned long long limit
, memsw_limit
;
5443 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5444 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5445 if (do_swap_account
)
5446 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5450 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5453 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5455 for_each_mem_cgroup_tree(mi
, memcg
)
5456 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5457 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5460 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5461 unsigned long long val
= 0;
5463 for_each_mem_cgroup_tree(mi
, memcg
)
5464 val
+= mem_cgroup_read_events(mi
, i
);
5465 seq_printf(m
, "total_%s %llu\n",
5466 mem_cgroup_events_names
[i
], val
);
5469 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5470 unsigned long long val
= 0;
5472 for_each_mem_cgroup_tree(mi
, memcg
)
5473 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5474 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5477 #ifdef CONFIG_DEBUG_VM
5480 struct mem_cgroup_per_zone
*mz
;
5481 struct zone_reclaim_stat
*rstat
;
5482 unsigned long recent_rotated
[2] = {0, 0};
5483 unsigned long recent_scanned
[2] = {0, 0};
5485 for_each_online_node(nid
)
5486 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5487 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5488 rstat
= &mz
->lruvec
.reclaim_stat
;
5490 recent_rotated
[0] += rstat
->recent_rotated
[0];
5491 recent_rotated
[1] += rstat
->recent_rotated
[1];
5492 recent_scanned
[0] += rstat
->recent_scanned
[0];
5493 recent_scanned
[1] += rstat
->recent_scanned
[1];
5495 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5496 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5497 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5498 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5505 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5508 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5510 return mem_cgroup_swappiness(memcg
);
5513 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5514 struct cftype
*cft
, u64 val
)
5516 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5517 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5519 if (val
> 100 || !parent
)
5522 mutex_lock(&memcg_create_mutex
);
5524 /* If under hierarchy, only empty-root can set this value */
5525 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5526 mutex_unlock(&memcg_create_mutex
);
5530 memcg
->swappiness
= val
;
5532 mutex_unlock(&memcg_create_mutex
);
5537 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5539 struct mem_cgroup_threshold_ary
*t
;
5545 t
= rcu_dereference(memcg
->thresholds
.primary
);
5547 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5552 usage
= mem_cgroup_usage(memcg
, swap
);
5555 * current_threshold points to threshold just below or equal to usage.
5556 * If it's not true, a threshold was crossed after last
5557 * call of __mem_cgroup_threshold().
5559 i
= t
->current_threshold
;
5562 * Iterate backward over array of thresholds starting from
5563 * current_threshold and check if a threshold is crossed.
5564 * If none of thresholds below usage is crossed, we read
5565 * only one element of the array here.
5567 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5568 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5570 /* i = current_threshold + 1 */
5574 * Iterate forward over array of thresholds starting from
5575 * current_threshold+1 and check if a threshold is crossed.
5576 * If none of thresholds above usage is crossed, we read
5577 * only one element of the array here.
5579 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5580 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5582 /* Update current_threshold */
5583 t
->current_threshold
= i
- 1;
5588 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5591 __mem_cgroup_threshold(memcg
, false);
5592 if (do_swap_account
)
5593 __mem_cgroup_threshold(memcg
, true);
5595 memcg
= parent_mem_cgroup(memcg
);
5599 static int compare_thresholds(const void *a
, const void *b
)
5601 const struct mem_cgroup_threshold
*_a
= a
;
5602 const struct mem_cgroup_threshold
*_b
= b
;
5604 return _a
->threshold
- _b
->threshold
;
5607 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5609 struct mem_cgroup_eventfd_list
*ev
;
5611 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5612 eventfd_signal(ev
->eventfd
, 1);
5616 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5618 struct mem_cgroup
*iter
;
5620 for_each_mem_cgroup_tree(iter
, memcg
)
5621 mem_cgroup_oom_notify_cb(iter
);
5624 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5625 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5627 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5628 struct mem_cgroup_thresholds
*thresholds
;
5629 struct mem_cgroup_threshold_ary
*new;
5630 enum res_type type
= MEMFILE_TYPE(cft
->private);
5631 u64 threshold
, usage
;
5634 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5638 mutex_lock(&memcg
->thresholds_lock
);
5641 thresholds
= &memcg
->thresholds
;
5642 else if (type
== _MEMSWAP
)
5643 thresholds
= &memcg
->memsw_thresholds
;
5647 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5649 /* Check if a threshold crossed before adding a new one */
5650 if (thresholds
->primary
)
5651 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5653 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5655 /* Allocate memory for new array of thresholds */
5656 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5664 /* Copy thresholds (if any) to new array */
5665 if (thresholds
->primary
) {
5666 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5667 sizeof(struct mem_cgroup_threshold
));
5670 /* Add new threshold */
5671 new->entries
[size
- 1].eventfd
= eventfd
;
5672 new->entries
[size
- 1].threshold
= threshold
;
5674 /* Sort thresholds. Registering of new threshold isn't time-critical */
5675 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5676 compare_thresholds
, NULL
);
5678 /* Find current threshold */
5679 new->current_threshold
= -1;
5680 for (i
= 0; i
< size
; i
++) {
5681 if (new->entries
[i
].threshold
<= usage
) {
5683 * new->current_threshold will not be used until
5684 * rcu_assign_pointer(), so it's safe to increment
5687 ++new->current_threshold
;
5692 /* Free old spare buffer and save old primary buffer as spare */
5693 kfree(thresholds
->spare
);
5694 thresholds
->spare
= thresholds
->primary
;
5696 rcu_assign_pointer(thresholds
->primary
, new);
5698 /* To be sure that nobody uses thresholds */
5702 mutex_unlock(&memcg
->thresholds_lock
);
5707 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5708 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5710 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5711 struct mem_cgroup_thresholds
*thresholds
;
5712 struct mem_cgroup_threshold_ary
*new;
5713 enum res_type type
= MEMFILE_TYPE(cft
->private);
5717 mutex_lock(&memcg
->thresholds_lock
);
5719 thresholds
= &memcg
->thresholds
;
5720 else if (type
== _MEMSWAP
)
5721 thresholds
= &memcg
->memsw_thresholds
;
5725 if (!thresholds
->primary
)
5728 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5730 /* Check if a threshold crossed before removing */
5731 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5733 /* Calculate new number of threshold */
5735 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5736 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5740 new = thresholds
->spare
;
5742 /* Set thresholds array to NULL if we don't have thresholds */
5751 /* Copy thresholds and find current threshold */
5752 new->current_threshold
= -1;
5753 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5754 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5757 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5758 if (new->entries
[j
].threshold
<= usage
) {
5760 * new->current_threshold will not be used
5761 * until rcu_assign_pointer(), so it's safe to increment
5764 ++new->current_threshold
;
5770 /* Swap primary and spare array */
5771 thresholds
->spare
= thresholds
->primary
;
5772 /* If all events are unregistered, free the spare array */
5774 kfree(thresholds
->spare
);
5775 thresholds
->spare
= NULL
;
5778 rcu_assign_pointer(thresholds
->primary
, new);
5780 /* To be sure that nobody uses thresholds */
5783 mutex_unlock(&memcg
->thresholds_lock
);
5786 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5787 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5789 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5790 struct mem_cgroup_eventfd_list
*event
;
5791 enum res_type type
= MEMFILE_TYPE(cft
->private);
5793 BUG_ON(type
!= _OOM_TYPE
);
5794 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5798 spin_lock(&memcg_oom_lock
);
5800 event
->eventfd
= eventfd
;
5801 list_add(&event
->list
, &memcg
->oom_notify
);
5803 /* already in OOM ? */
5804 if (atomic_read(&memcg
->under_oom
))
5805 eventfd_signal(eventfd
, 1);
5806 spin_unlock(&memcg_oom_lock
);
5811 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5812 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5814 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5815 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5816 enum res_type type
= MEMFILE_TYPE(cft
->private);
5818 BUG_ON(type
!= _OOM_TYPE
);
5820 spin_lock(&memcg_oom_lock
);
5822 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5823 if (ev
->eventfd
== eventfd
) {
5824 list_del(&ev
->list
);
5829 spin_unlock(&memcg_oom_lock
);
5832 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5833 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5835 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5837 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5839 if (atomic_read(&memcg
->under_oom
))
5840 cb
->fill(cb
, "under_oom", 1);
5842 cb
->fill(cb
, "under_oom", 0);
5846 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5847 struct cftype
*cft
, u64 val
)
5849 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5850 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5852 /* cannot set to root cgroup and only 0 and 1 are allowed */
5853 if (!parent
|| !((val
== 0) || (val
== 1)))
5856 mutex_lock(&memcg_create_mutex
);
5857 /* oom-kill-disable is a flag for subhierarchy. */
5858 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5859 mutex_unlock(&memcg_create_mutex
);
5862 memcg
->oom_kill_disable
= val
;
5864 memcg_oom_recover(memcg
);
5865 mutex_unlock(&memcg_create_mutex
);
5869 #ifdef CONFIG_MEMCG_KMEM
5870 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5874 memcg
->kmemcg_id
= -1;
5875 ret
= memcg_propagate_kmem(memcg
);
5879 return mem_cgroup_sockets_init(memcg
, ss
);
5882 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5884 mem_cgroup_sockets_destroy(memcg
);
5887 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5889 if (!memcg_kmem_is_active(memcg
))
5893 * kmem charges can outlive the cgroup. In the case of slab
5894 * pages, for instance, a page contain objects from various
5895 * processes. As we prevent from taking a reference for every
5896 * such allocation we have to be careful when doing uncharge
5897 * (see memcg_uncharge_kmem) and here during offlining.
5899 * The idea is that that only the _last_ uncharge which sees
5900 * the dead memcg will drop the last reference. An additional
5901 * reference is taken here before the group is marked dead
5902 * which is then paired with css_put during uncharge resp. here.
5904 * Although this might sound strange as this path is called from
5905 * css_offline() when the referencemight have dropped down to 0
5906 * and shouldn't be incremented anymore (css_tryget would fail)
5907 * we do not have other options because of the kmem allocations
5910 css_get(&memcg
->css
);
5912 memcg_kmem_mark_dead(memcg
);
5914 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5917 if (memcg_kmem_test_and_clear_dead(memcg
))
5918 css_put(&memcg
->css
);
5921 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5926 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5930 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5935 static struct cftype mem_cgroup_files
[] = {
5937 .name
= "usage_in_bytes",
5938 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5939 .read
= mem_cgroup_read
,
5940 .register_event
= mem_cgroup_usage_register_event
,
5941 .unregister_event
= mem_cgroup_usage_unregister_event
,
5944 .name
= "max_usage_in_bytes",
5945 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5946 .trigger
= mem_cgroup_reset
,
5947 .read
= mem_cgroup_read
,
5950 .name
= "limit_in_bytes",
5951 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5952 .write_string
= mem_cgroup_write
,
5953 .read
= mem_cgroup_read
,
5956 .name
= "soft_limit_in_bytes",
5957 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5958 .write_string
= mem_cgroup_write
,
5959 .read
= mem_cgroup_read
,
5963 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5964 .trigger
= mem_cgroup_reset
,
5965 .read
= mem_cgroup_read
,
5969 .read_seq_string
= memcg_stat_show
,
5972 .name
= "force_empty",
5973 .trigger
= mem_cgroup_force_empty_write
,
5976 .name
= "use_hierarchy",
5977 .flags
= CFTYPE_INSANE
,
5978 .write_u64
= mem_cgroup_hierarchy_write
,
5979 .read_u64
= mem_cgroup_hierarchy_read
,
5982 .name
= "swappiness",
5983 .read_u64
= mem_cgroup_swappiness_read
,
5984 .write_u64
= mem_cgroup_swappiness_write
,
5987 .name
= "move_charge_at_immigrate",
5988 .read_u64
= mem_cgroup_move_charge_read
,
5989 .write_u64
= mem_cgroup_move_charge_write
,
5992 .name
= "oom_control",
5993 .read_map
= mem_cgroup_oom_control_read
,
5994 .write_u64
= mem_cgroup_oom_control_write
,
5995 .register_event
= mem_cgroup_oom_register_event
,
5996 .unregister_event
= mem_cgroup_oom_unregister_event
,
5997 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6000 .name
= "pressure_level",
6001 .register_event
= vmpressure_register_event
,
6002 .unregister_event
= vmpressure_unregister_event
,
6006 .name
= "numa_stat",
6007 .read_seq_string
= memcg_numa_stat_show
,
6010 #ifdef CONFIG_MEMCG_KMEM
6012 .name
= "kmem.limit_in_bytes",
6013 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6014 .write_string
= mem_cgroup_write
,
6015 .read
= mem_cgroup_read
,
6018 .name
= "kmem.usage_in_bytes",
6019 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6020 .read
= mem_cgroup_read
,
6023 .name
= "kmem.failcnt",
6024 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6025 .trigger
= mem_cgroup_reset
,
6026 .read
= mem_cgroup_read
,
6029 .name
= "kmem.max_usage_in_bytes",
6030 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6031 .trigger
= mem_cgroup_reset
,
6032 .read
= mem_cgroup_read
,
6034 #ifdef CONFIG_SLABINFO
6036 .name
= "kmem.slabinfo",
6037 .read_seq_string
= mem_cgroup_slabinfo_read
,
6041 { }, /* terminate */
6044 #ifdef CONFIG_MEMCG_SWAP
6045 static struct cftype memsw_cgroup_files
[] = {
6047 .name
= "memsw.usage_in_bytes",
6048 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6049 .read
= mem_cgroup_read
,
6050 .register_event
= mem_cgroup_usage_register_event
,
6051 .unregister_event
= mem_cgroup_usage_unregister_event
,
6054 .name
= "memsw.max_usage_in_bytes",
6055 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6056 .trigger
= mem_cgroup_reset
,
6057 .read
= mem_cgroup_read
,
6060 .name
= "memsw.limit_in_bytes",
6061 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6062 .write_string
= mem_cgroup_write
,
6063 .read
= mem_cgroup_read
,
6066 .name
= "memsw.failcnt",
6067 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6068 .trigger
= mem_cgroup_reset
,
6069 .read
= mem_cgroup_read
,
6071 { }, /* terminate */
6074 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6076 struct mem_cgroup_per_node
*pn
;
6077 struct mem_cgroup_per_zone
*mz
;
6078 int zone
, tmp
= node
;
6080 * This routine is called against possible nodes.
6081 * But it's BUG to call kmalloc() against offline node.
6083 * TODO: this routine can waste much memory for nodes which will
6084 * never be onlined. It's better to use memory hotplug callback
6087 if (!node_state(node
, N_NORMAL_MEMORY
))
6089 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6093 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6094 mz
= &pn
->zoneinfo
[zone
];
6095 lruvec_init(&mz
->lruvec
);
6096 mz
->usage_in_excess
= 0;
6097 mz
->on_tree
= false;
6100 memcg
->nodeinfo
[node
] = pn
;
6104 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6106 kfree(memcg
->nodeinfo
[node
]);
6109 static struct mem_cgroup
*mem_cgroup_alloc(void)
6111 struct mem_cgroup
*memcg
;
6112 size_t size
= memcg_size();
6114 /* Can be very big if nr_node_ids is very big */
6115 if (size
< PAGE_SIZE
)
6116 memcg
= kzalloc(size
, GFP_KERNEL
);
6118 memcg
= vzalloc(size
);
6123 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6126 spin_lock_init(&memcg
->pcp_counter_lock
);
6130 if (size
< PAGE_SIZE
)
6138 * At destroying mem_cgroup, references from swap_cgroup can remain.
6139 * (scanning all at force_empty is too costly...)
6141 * Instead of clearing all references at force_empty, we remember
6142 * the number of reference from swap_cgroup and free mem_cgroup when
6143 * it goes down to 0.
6145 * Removal of cgroup itself succeeds regardless of refs from swap.
6148 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6151 size_t size
= memcg_size();
6153 mem_cgroup_remove_from_trees(memcg
);
6154 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
6157 free_mem_cgroup_per_zone_info(memcg
, node
);
6159 free_percpu(memcg
->stat
);
6162 * We need to make sure that (at least for now), the jump label
6163 * destruction code runs outside of the cgroup lock. This is because
6164 * get_online_cpus(), which is called from the static_branch update,
6165 * can't be called inside the cgroup_lock. cpusets are the ones
6166 * enforcing this dependency, so if they ever change, we might as well.
6168 * schedule_work() will guarantee this happens. Be careful if you need
6169 * to move this code around, and make sure it is outside
6172 disarm_static_keys(memcg
);
6173 if (size
< PAGE_SIZE
)
6180 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6182 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6184 if (!memcg
->res
.parent
)
6186 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6188 EXPORT_SYMBOL(parent_mem_cgroup
);
6190 static void __init
mem_cgroup_soft_limit_tree_init(void)
6192 struct mem_cgroup_tree_per_node
*rtpn
;
6193 struct mem_cgroup_tree_per_zone
*rtpz
;
6194 int tmp
, node
, zone
;
6196 for_each_node(node
) {
6198 if (!node_state(node
, N_NORMAL_MEMORY
))
6200 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6203 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6205 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6206 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6207 rtpz
->rb_root
= RB_ROOT
;
6208 spin_lock_init(&rtpz
->lock
);
6213 static struct cgroup_subsys_state
* __ref
6214 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6216 struct mem_cgroup
*memcg
;
6217 long error
= -ENOMEM
;
6220 memcg
= mem_cgroup_alloc();
6222 return ERR_PTR(error
);
6225 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6229 if (parent_css
== NULL
) {
6230 root_mem_cgroup
= memcg
;
6231 res_counter_init(&memcg
->res
, NULL
);
6232 res_counter_init(&memcg
->memsw
, NULL
);
6233 res_counter_init(&memcg
->kmem
, NULL
);
6236 memcg
->last_scanned_node
= MAX_NUMNODES
;
6237 INIT_LIST_HEAD(&memcg
->oom_notify
);
6238 memcg
->move_charge_at_immigrate
= 0;
6239 mutex_init(&memcg
->thresholds_lock
);
6240 spin_lock_init(&memcg
->move_lock
);
6241 vmpressure_init(&memcg
->vmpressure
);
6246 __mem_cgroup_free(memcg
);
6247 return ERR_PTR(error
);
6251 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6253 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6254 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6260 mutex_lock(&memcg_create_mutex
);
6262 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6263 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6264 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6266 if (parent
->use_hierarchy
) {
6267 res_counter_init(&memcg
->res
, &parent
->res
);
6268 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6269 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6272 * No need to take a reference to the parent because cgroup
6273 * core guarantees its existence.
6276 res_counter_init(&memcg
->res
, NULL
);
6277 res_counter_init(&memcg
->memsw
, NULL
);
6278 res_counter_init(&memcg
->kmem
, NULL
);
6280 * Deeper hierachy with use_hierarchy == false doesn't make
6281 * much sense so let cgroup subsystem know about this
6282 * unfortunate state in our controller.
6284 if (parent
!= root_mem_cgroup
)
6285 mem_cgroup_subsys
.broken_hierarchy
= true;
6288 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6289 mutex_unlock(&memcg_create_mutex
);
6294 * Announce all parents that a group from their hierarchy is gone.
6296 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6298 struct mem_cgroup
*parent
= memcg
;
6300 while ((parent
= parent_mem_cgroup(parent
)))
6301 mem_cgroup_iter_invalidate(parent
);
6304 * if the root memcg is not hierarchical we have to check it
6307 if (!root_mem_cgroup
->use_hierarchy
)
6308 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6311 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6313 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6315 kmem_cgroup_css_offline(memcg
);
6317 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6318 mem_cgroup_reparent_charges(memcg
);
6319 mem_cgroup_destroy_all_caches(memcg
);
6322 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6324 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6326 memcg_destroy_kmem(memcg
);
6327 __mem_cgroup_free(memcg
);
6331 /* Handlers for move charge at task migration. */
6332 #define PRECHARGE_COUNT_AT_ONCE 256
6333 static int mem_cgroup_do_precharge(unsigned long count
)
6336 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6337 struct mem_cgroup
*memcg
= mc
.to
;
6339 if (mem_cgroup_is_root(memcg
)) {
6340 mc
.precharge
+= count
;
6341 /* we don't need css_get for root */
6344 /* try to charge at once */
6346 struct res_counter
*dummy
;
6348 * "memcg" cannot be under rmdir() because we've already checked
6349 * by cgroup_lock_live_cgroup() that it is not removed and we
6350 * are still under the same cgroup_mutex. So we can postpone
6353 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6355 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6356 PAGE_SIZE
* count
, &dummy
)) {
6357 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6360 mc
.precharge
+= count
;
6364 /* fall back to one by one charge */
6366 if (signal_pending(current
)) {
6370 if (!batch_count
--) {
6371 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6374 ret
= __mem_cgroup_try_charge(NULL
,
6375 GFP_KERNEL
, 1, &memcg
, false);
6377 /* mem_cgroup_clear_mc() will do uncharge later */
6385 * get_mctgt_type - get target type of moving charge
6386 * @vma: the vma the pte to be checked belongs
6387 * @addr: the address corresponding to the pte to be checked
6388 * @ptent: the pte to be checked
6389 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6392 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6393 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6394 * move charge. if @target is not NULL, the page is stored in target->page
6395 * with extra refcnt got(Callers should handle it).
6396 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6397 * target for charge migration. if @target is not NULL, the entry is stored
6400 * Called with pte lock held.
6407 enum mc_target_type
{
6413 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6414 unsigned long addr
, pte_t ptent
)
6416 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6418 if (!page
|| !page_mapped(page
))
6420 if (PageAnon(page
)) {
6421 /* we don't move shared anon */
6424 } else if (!move_file())
6425 /* we ignore mapcount for file pages */
6427 if (!get_page_unless_zero(page
))
6434 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6435 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6437 struct page
*page
= NULL
;
6438 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6440 if (!move_anon() || non_swap_entry(ent
))
6443 * Because lookup_swap_cache() updates some statistics counter,
6444 * we call find_get_page() with swapper_space directly.
6446 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6447 if (do_swap_account
)
6448 entry
->val
= ent
.val
;
6453 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6454 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6460 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6461 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6463 struct page
*page
= NULL
;
6464 struct address_space
*mapping
;
6467 if (!vma
->vm_file
) /* anonymous vma */
6472 mapping
= vma
->vm_file
->f_mapping
;
6473 if (pte_none(ptent
))
6474 pgoff
= linear_page_index(vma
, addr
);
6475 else /* pte_file(ptent) is true */
6476 pgoff
= pte_to_pgoff(ptent
);
6478 /* page is moved even if it's not RSS of this task(page-faulted). */
6479 page
= find_get_page(mapping
, pgoff
);
6482 /* shmem/tmpfs may report page out on swap: account for that too. */
6483 if (radix_tree_exceptional_entry(page
)) {
6484 swp_entry_t swap
= radix_to_swp_entry(page
);
6485 if (do_swap_account
)
6487 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6493 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6494 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6496 struct page
*page
= NULL
;
6497 struct page_cgroup
*pc
;
6498 enum mc_target_type ret
= MC_TARGET_NONE
;
6499 swp_entry_t ent
= { .val
= 0 };
6501 if (pte_present(ptent
))
6502 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6503 else if (is_swap_pte(ptent
))
6504 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6505 else if (pte_none(ptent
) || pte_file(ptent
))
6506 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6508 if (!page
&& !ent
.val
)
6511 pc
= lookup_page_cgroup(page
);
6513 * Do only loose check w/o page_cgroup lock.
6514 * mem_cgroup_move_account() checks the pc is valid or not under
6517 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6518 ret
= MC_TARGET_PAGE
;
6520 target
->page
= page
;
6522 if (!ret
|| !target
)
6525 /* There is a swap entry and a page doesn't exist or isn't charged */
6526 if (ent
.val
&& !ret
&&
6527 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6528 ret
= MC_TARGET_SWAP
;
6535 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6537 * We don't consider swapping or file mapped pages because THP does not
6538 * support them for now.
6539 * Caller should make sure that pmd_trans_huge(pmd) is true.
6541 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6542 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6544 struct page
*page
= NULL
;
6545 struct page_cgroup
*pc
;
6546 enum mc_target_type ret
= MC_TARGET_NONE
;
6548 page
= pmd_page(pmd
);
6549 VM_BUG_ON(!page
|| !PageHead(page
));
6552 pc
= lookup_page_cgroup(page
);
6553 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6554 ret
= MC_TARGET_PAGE
;
6557 target
->page
= page
;
6563 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6564 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6566 return MC_TARGET_NONE
;
6570 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6571 unsigned long addr
, unsigned long end
,
6572 struct mm_walk
*walk
)
6574 struct vm_area_struct
*vma
= walk
->private;
6578 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6579 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6580 mc
.precharge
+= HPAGE_PMD_NR
;
6581 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6585 if (pmd_trans_unstable(pmd
))
6587 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6588 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6589 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6590 mc
.precharge
++; /* increment precharge temporarily */
6591 pte_unmap_unlock(pte
- 1, ptl
);
6597 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6599 unsigned long precharge
;
6600 struct vm_area_struct
*vma
;
6602 down_read(&mm
->mmap_sem
);
6603 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6604 struct mm_walk mem_cgroup_count_precharge_walk
= {
6605 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6609 if (is_vm_hugetlb_page(vma
))
6611 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6612 &mem_cgroup_count_precharge_walk
);
6614 up_read(&mm
->mmap_sem
);
6616 precharge
= mc
.precharge
;
6622 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6624 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6626 VM_BUG_ON(mc
.moving_task
);
6627 mc
.moving_task
= current
;
6628 return mem_cgroup_do_precharge(precharge
);
6631 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6632 static void __mem_cgroup_clear_mc(void)
6634 struct mem_cgroup
*from
= mc
.from
;
6635 struct mem_cgroup
*to
= mc
.to
;
6638 /* we must uncharge all the leftover precharges from mc.to */
6640 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6644 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6645 * we must uncharge here.
6647 if (mc
.moved_charge
) {
6648 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6649 mc
.moved_charge
= 0;
6651 /* we must fixup refcnts and charges */
6652 if (mc
.moved_swap
) {
6653 /* uncharge swap account from the old cgroup */
6654 if (!mem_cgroup_is_root(mc
.from
))
6655 res_counter_uncharge(&mc
.from
->memsw
,
6656 PAGE_SIZE
* mc
.moved_swap
);
6658 for (i
= 0; i
< mc
.moved_swap
; i
++)
6659 css_put(&mc
.from
->css
);
6661 if (!mem_cgroup_is_root(mc
.to
)) {
6663 * we charged both to->res and to->memsw, so we should
6666 res_counter_uncharge(&mc
.to
->res
,
6667 PAGE_SIZE
* mc
.moved_swap
);
6669 /* we've already done css_get(mc.to) */
6672 memcg_oom_recover(from
);
6673 memcg_oom_recover(to
);
6674 wake_up_all(&mc
.waitq
);
6677 static void mem_cgroup_clear_mc(void)
6679 struct mem_cgroup
*from
= mc
.from
;
6682 * we must clear moving_task before waking up waiters at the end of
6685 mc
.moving_task
= NULL
;
6686 __mem_cgroup_clear_mc();
6687 spin_lock(&mc
.lock
);
6690 spin_unlock(&mc
.lock
);
6691 mem_cgroup_end_move(from
);
6694 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6695 struct cgroup_taskset
*tset
)
6697 struct task_struct
*p
= cgroup_taskset_first(tset
);
6699 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6700 unsigned long move_charge_at_immigrate
;
6703 * We are now commited to this value whatever it is. Changes in this
6704 * tunable will only affect upcoming migrations, not the current one.
6705 * So we need to save it, and keep it going.
6707 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6708 if (move_charge_at_immigrate
) {
6709 struct mm_struct
*mm
;
6710 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6712 VM_BUG_ON(from
== memcg
);
6714 mm
= get_task_mm(p
);
6717 /* We move charges only when we move a owner of the mm */
6718 if (mm
->owner
== p
) {
6721 VM_BUG_ON(mc
.precharge
);
6722 VM_BUG_ON(mc
.moved_charge
);
6723 VM_BUG_ON(mc
.moved_swap
);
6724 mem_cgroup_start_move(from
);
6725 spin_lock(&mc
.lock
);
6728 mc
.immigrate_flags
= move_charge_at_immigrate
;
6729 spin_unlock(&mc
.lock
);
6730 /* We set mc.moving_task later */
6732 ret
= mem_cgroup_precharge_mc(mm
);
6734 mem_cgroup_clear_mc();
6741 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6742 struct cgroup_taskset
*tset
)
6744 mem_cgroup_clear_mc();
6747 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6748 unsigned long addr
, unsigned long end
,
6749 struct mm_walk
*walk
)
6752 struct vm_area_struct
*vma
= walk
->private;
6755 enum mc_target_type target_type
;
6756 union mc_target target
;
6758 struct page_cgroup
*pc
;
6761 * We don't take compound_lock() here but no race with splitting thp
6763 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6764 * under splitting, which means there's no concurrent thp split,
6765 * - if another thread runs into split_huge_page() just after we
6766 * entered this if-block, the thread must wait for page table lock
6767 * to be unlocked in __split_huge_page_splitting(), where the main
6768 * part of thp split is not executed yet.
6770 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6771 if (mc
.precharge
< HPAGE_PMD_NR
) {
6772 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6775 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6776 if (target_type
== MC_TARGET_PAGE
) {
6778 if (!isolate_lru_page(page
)) {
6779 pc
= lookup_page_cgroup(page
);
6780 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6781 pc
, mc
.from
, mc
.to
)) {
6782 mc
.precharge
-= HPAGE_PMD_NR
;
6783 mc
.moved_charge
+= HPAGE_PMD_NR
;
6785 putback_lru_page(page
);
6789 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6793 if (pmd_trans_unstable(pmd
))
6796 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6797 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6798 pte_t ptent
= *(pte
++);
6804 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6805 case MC_TARGET_PAGE
:
6807 if (isolate_lru_page(page
))
6809 pc
= lookup_page_cgroup(page
);
6810 if (!mem_cgroup_move_account(page
, 1, pc
,
6813 /* we uncharge from mc.from later. */
6816 putback_lru_page(page
);
6817 put
: /* get_mctgt_type() gets the page */
6820 case MC_TARGET_SWAP
:
6822 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6824 /* we fixup refcnts and charges later. */
6832 pte_unmap_unlock(pte
- 1, ptl
);
6837 * We have consumed all precharges we got in can_attach().
6838 * We try charge one by one, but don't do any additional
6839 * charges to mc.to if we have failed in charge once in attach()
6842 ret
= mem_cgroup_do_precharge(1);
6850 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6852 struct vm_area_struct
*vma
;
6854 lru_add_drain_all();
6856 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6858 * Someone who are holding the mmap_sem might be waiting in
6859 * waitq. So we cancel all extra charges, wake up all waiters,
6860 * and retry. Because we cancel precharges, we might not be able
6861 * to move enough charges, but moving charge is a best-effort
6862 * feature anyway, so it wouldn't be a big problem.
6864 __mem_cgroup_clear_mc();
6868 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6870 struct mm_walk mem_cgroup_move_charge_walk
= {
6871 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6875 if (is_vm_hugetlb_page(vma
))
6877 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6878 &mem_cgroup_move_charge_walk
);
6881 * means we have consumed all precharges and failed in
6882 * doing additional charge. Just abandon here.
6886 up_read(&mm
->mmap_sem
);
6889 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6890 struct cgroup_taskset
*tset
)
6892 struct task_struct
*p
= cgroup_taskset_first(tset
);
6893 struct mm_struct
*mm
= get_task_mm(p
);
6897 mem_cgroup_move_charge(mm
);
6901 mem_cgroup_clear_mc();
6903 #else /* !CONFIG_MMU */
6904 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6905 struct cgroup_taskset
*tset
)
6909 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6910 struct cgroup_taskset
*tset
)
6913 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6914 struct cgroup_taskset
*tset
)
6920 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6921 * to verify sane_behavior flag on each mount attempt.
6923 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6926 * use_hierarchy is forced with sane_behavior. cgroup core
6927 * guarantees that @root doesn't have any children, so turning it
6928 * on for the root memcg is enough.
6930 if (cgroup_sane_behavior(root_css
->cgroup
))
6931 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6934 struct cgroup_subsys mem_cgroup_subsys
= {
6936 .subsys_id
= mem_cgroup_subsys_id
,
6937 .css_alloc
= mem_cgroup_css_alloc
,
6938 .css_online
= mem_cgroup_css_online
,
6939 .css_offline
= mem_cgroup_css_offline
,
6940 .css_free
= mem_cgroup_css_free
,
6941 .can_attach
= mem_cgroup_can_attach
,
6942 .cancel_attach
= mem_cgroup_cancel_attach
,
6943 .attach
= mem_cgroup_move_task
,
6944 .bind
= mem_cgroup_bind
,
6945 .base_cftypes
= mem_cgroup_files
,
6950 #ifdef CONFIG_MEMCG_SWAP
6951 static int __init
enable_swap_account(char *s
)
6953 /* consider enabled if no parameter or 1 is given */
6954 if (!strcmp(s
, "1"))
6955 really_do_swap_account
= 1;
6956 else if (!strcmp(s
, "0"))
6957 really_do_swap_account
= 0;
6960 __setup("swapaccount=", enable_swap_account
);
6962 static void __init
memsw_file_init(void)
6964 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6967 static void __init
enable_swap_cgroup(void)
6969 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6970 do_swap_account
= 1;
6976 static void __init
enable_swap_cgroup(void)
6982 * subsys_initcall() for memory controller.
6984 * Some parts like hotcpu_notifier() have to be initialized from this context
6985 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6986 * everything that doesn't depend on a specific mem_cgroup structure should
6987 * be initialized from here.
6989 static int __init
mem_cgroup_init(void)
6991 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6992 enable_swap_cgroup();
6993 mem_cgroup_soft_limit_tree_init();
6997 subsys_initcall(mem_cgroup_init
);