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>
57 #include <linux/lockdep.h>
61 #include <net/tcp_memcontrol.h>
63 #include <asm/uaccess.h>
65 #include <trace/events/vmscan.h>
67 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
68 EXPORT_SYMBOL(mem_cgroup_subsys
);
70 #define MEM_CGROUP_RECLAIM_RETRIES 5
71 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
73 #ifdef CONFIG_MEMCG_SWAP
74 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
75 int do_swap_account __read_mostly
;
77 /* for remember boot option*/
78 #ifdef CONFIG_MEMCG_SWAP_ENABLED
79 static int really_do_swap_account __initdata
= 1;
81 static int really_do_swap_account __initdata
= 0;
85 #define do_swap_account 0
89 static const char * const mem_cgroup_stat_names
[] = {
98 enum mem_cgroup_events_index
{
99 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
100 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
101 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
102 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
103 MEM_CGROUP_EVENTS_NSTATS
,
106 static const char * const mem_cgroup_events_names
[] = {
113 static const char * const mem_cgroup_lru_names
[] = {
122 * Per memcg event counter is incremented at every pagein/pageout. With THP,
123 * it will be incremated by the number of pages. This counter is used for
124 * for trigger some periodic events. This is straightforward and better
125 * than using jiffies etc. to handle periodic memcg event.
127 enum mem_cgroup_events_target
{
128 MEM_CGROUP_TARGET_THRESH
,
129 MEM_CGROUP_TARGET_SOFTLIMIT
,
130 MEM_CGROUP_TARGET_NUMAINFO
,
133 #define THRESHOLDS_EVENTS_TARGET 128
134 #define SOFTLIMIT_EVENTS_TARGET 1024
135 #define NUMAINFO_EVENTS_TARGET 1024
137 struct mem_cgroup_stat_cpu
{
138 long count
[MEM_CGROUP_STAT_NSTATS
];
139 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
140 unsigned long nr_page_events
;
141 unsigned long targets
[MEM_CGROUP_NTARGETS
];
144 struct mem_cgroup_reclaim_iter
{
146 * last scanned hierarchy member. Valid only if last_dead_count
147 * matches memcg->dead_count of the hierarchy root group.
149 struct mem_cgroup
*last_visited
;
150 unsigned long last_dead_count
;
152 /* scan generation, increased every round-trip */
153 unsigned int generation
;
157 * per-zone information in memory controller.
159 struct mem_cgroup_per_zone
{
160 struct lruvec lruvec
;
161 unsigned long lru_size
[NR_LRU_LISTS
];
163 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
165 struct rb_node tree_node
; /* RB tree node */
166 unsigned long long usage_in_excess
;/* Set to the value by which */
167 /* the soft limit is exceeded*/
169 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
170 /* use container_of */
173 struct mem_cgroup_per_node
{
174 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
178 * Cgroups above their limits are maintained in a RB-Tree, independent of
179 * their hierarchy representation
182 struct mem_cgroup_tree_per_zone
{
183 struct rb_root rb_root
;
187 struct mem_cgroup_tree_per_node
{
188 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
191 struct mem_cgroup_tree
{
192 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
195 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
197 struct mem_cgroup_threshold
{
198 struct eventfd_ctx
*eventfd
;
203 struct mem_cgroup_threshold_ary
{
204 /* An array index points to threshold just below or equal to usage. */
205 int current_threshold
;
206 /* Size of entries[] */
208 /* Array of thresholds */
209 struct mem_cgroup_threshold entries
[0];
212 struct mem_cgroup_thresholds
{
213 /* Primary thresholds array */
214 struct mem_cgroup_threshold_ary
*primary
;
216 * Spare threshold array.
217 * This is needed to make mem_cgroup_unregister_event() "never fail".
218 * It must be able to store at least primary->size - 1 entries.
220 struct mem_cgroup_threshold_ary
*spare
;
224 struct mem_cgroup_eventfd_list
{
225 struct list_head list
;
226 struct eventfd_ctx
*eventfd
;
229 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
230 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
233 * The memory controller data structure. The memory controller controls both
234 * page cache and RSS per cgroup. We would eventually like to provide
235 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
236 * to help the administrator determine what knobs to tune.
238 * TODO: Add a water mark for the memory controller. Reclaim will begin when
239 * we hit the water mark. May be even add a low water mark, such that
240 * no reclaim occurs from a cgroup at it's low water mark, this is
241 * a feature that will be implemented much later in the future.
244 struct cgroup_subsys_state css
;
246 * the counter to account for memory usage
248 struct res_counter res
;
250 /* vmpressure notifications */
251 struct vmpressure vmpressure
;
254 * the counter to account for mem+swap usage.
256 struct res_counter memsw
;
259 * the counter to account for kernel memory usage.
261 struct res_counter kmem
;
263 * Should the accounting and control be hierarchical, per subtree?
266 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
270 atomic_t oom_wakeups
;
273 /* OOM-Killer disable */
274 int oom_kill_disable
;
276 /* set when res.limit == memsw.limit */
277 bool memsw_is_minimum
;
279 /* protect arrays of thresholds */
280 struct mutex thresholds_lock
;
282 /* thresholds for memory usage. RCU-protected */
283 struct mem_cgroup_thresholds thresholds
;
285 /* thresholds for mem+swap usage. RCU-protected */
286 struct mem_cgroup_thresholds memsw_thresholds
;
288 /* For oom notifier event fd */
289 struct list_head oom_notify
;
292 * Should we move charges of a task when a task is moved into this
293 * mem_cgroup ? And what type of charges should we move ?
295 unsigned long move_charge_at_immigrate
;
297 * set > 0 if pages under this cgroup are moving to other cgroup.
299 atomic_t moving_account
;
300 /* taken only while moving_account > 0 */
301 spinlock_t move_lock
;
305 struct mem_cgroup_stat_cpu __percpu
*stat
;
307 * used when a cpu is offlined or other synchronizations
308 * See mem_cgroup_read_stat().
310 struct mem_cgroup_stat_cpu nocpu_base
;
311 spinlock_t pcp_counter_lock
;
314 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
315 struct tcp_memcontrol tcp_mem
;
317 #if defined(CONFIG_MEMCG_KMEM)
318 /* analogous to slab_common's slab_caches list. per-memcg */
319 struct list_head memcg_slab_caches
;
320 /* Not a spinlock, we can take a lot of time walking the list */
321 struct mutex slab_caches_mutex
;
322 /* Index in the kmem_cache->memcg_params->memcg_caches array */
326 int last_scanned_node
;
328 nodemask_t scan_nodes
;
329 atomic_t numainfo_events
;
330 atomic_t numainfo_updating
;
333 struct mem_cgroup_per_node
*nodeinfo
[0];
334 /* WARNING: nodeinfo must be the last member here */
337 static size_t memcg_size(void)
339 return sizeof(struct mem_cgroup
) +
340 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
343 /* internal only representation about the status of kmem accounting. */
345 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
346 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
347 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
350 /* We account when limit is on, but only after call sites are patched */
351 #define KMEM_ACCOUNTED_MASK \
352 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
354 #ifdef CONFIG_MEMCG_KMEM
355 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
357 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
360 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
362 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
365 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
367 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
370 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
372 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
375 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
378 * Our caller must use css_get() first, because memcg_uncharge_kmem()
379 * will call css_put() if it sees the memcg is dead.
382 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
383 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
386 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
388 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
389 &memcg
->kmem_account_flags
);
393 /* Stuffs for move charges at task migration. */
395 * Types of charges to be moved. "move_charge_at_immitgrate" and
396 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
399 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
400 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
404 /* "mc" and its members are protected by cgroup_mutex */
405 static struct move_charge_struct
{
406 spinlock_t lock
; /* for from, to */
407 struct mem_cgroup
*from
;
408 struct mem_cgroup
*to
;
409 unsigned long immigrate_flags
;
410 unsigned long precharge
;
411 unsigned long moved_charge
;
412 unsigned long moved_swap
;
413 struct task_struct
*moving_task
; /* a task moving charges */
414 wait_queue_head_t waitq
; /* a waitq for other context */
416 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
417 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
420 static bool move_anon(void)
422 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
425 static bool move_file(void)
427 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
431 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
432 * limit reclaim to prevent infinite loops, if they ever occur.
434 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
435 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
438 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
439 MEM_CGROUP_CHARGE_TYPE_ANON
,
440 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
441 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
445 /* for encoding cft->private value on file */
453 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
454 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
455 #define MEMFILE_ATTR(val) ((val) & 0xffff)
456 /* Used for OOM nofiier */
457 #define OOM_CONTROL (0)
460 * Reclaim flags for mem_cgroup_hierarchical_reclaim
462 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
463 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
464 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
465 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
468 * The memcg_create_mutex will be held whenever a new cgroup is created.
469 * As a consequence, any change that needs to protect against new child cgroups
470 * appearing has to hold it as well.
472 static DEFINE_MUTEX(memcg_create_mutex
);
474 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
476 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
479 /* Some nice accessors for the vmpressure. */
480 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
483 memcg
= root_mem_cgroup
;
484 return &memcg
->vmpressure
;
487 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
489 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
492 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
494 return &mem_cgroup_from_css(css
)->vmpressure
;
497 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
499 return (memcg
== root_mem_cgroup
);
503 * We restrict the id in the range of [1, 65535], so it can fit into
506 #define MEM_CGROUP_ID_MAX USHRT_MAX
508 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
511 * The ID of the root cgroup is 0, but memcg treat 0 as an
512 * invalid ID, so we return (cgroup_id + 1).
514 return memcg
->css
.cgroup
->id
+ 1;
517 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
519 struct cgroup_subsys_state
*css
;
521 css
= css_from_id(id
- 1, &mem_cgroup_subsys
);
522 return mem_cgroup_from_css(css
);
525 /* Writing them here to avoid exposing memcg's inner layout */
526 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
528 void sock_update_memcg(struct sock
*sk
)
530 if (mem_cgroup_sockets_enabled
) {
531 struct mem_cgroup
*memcg
;
532 struct cg_proto
*cg_proto
;
534 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
536 /* Socket cloning can throw us here with sk_cgrp already
537 * filled. It won't however, necessarily happen from
538 * process context. So the test for root memcg given
539 * the current task's memcg won't help us in this case.
541 * Respecting the original socket's memcg is a better
542 * decision in this case.
545 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
546 css_get(&sk
->sk_cgrp
->memcg
->css
);
551 memcg
= mem_cgroup_from_task(current
);
552 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
553 if (!mem_cgroup_is_root(memcg
) &&
554 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
555 sk
->sk_cgrp
= cg_proto
;
560 EXPORT_SYMBOL(sock_update_memcg
);
562 void sock_release_memcg(struct sock
*sk
)
564 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
565 struct mem_cgroup
*memcg
;
566 WARN_ON(!sk
->sk_cgrp
->memcg
);
567 memcg
= sk
->sk_cgrp
->memcg
;
568 css_put(&sk
->sk_cgrp
->memcg
->css
);
572 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
574 if (!memcg
|| mem_cgroup_is_root(memcg
))
577 return &memcg
->tcp_mem
.cg_proto
;
579 EXPORT_SYMBOL(tcp_proto_cgroup
);
581 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
583 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
585 static_key_slow_dec(&memcg_socket_limit_enabled
);
588 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
593 #ifdef CONFIG_MEMCG_KMEM
595 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
596 * The main reason for not using cgroup id for this:
597 * this works better in sparse environments, where we have a lot of memcgs,
598 * but only a few kmem-limited. Or also, if we have, for instance, 200
599 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
600 * 200 entry array for that.
602 * The current size of the caches array is stored in
603 * memcg_limited_groups_array_size. It will double each time we have to
606 static DEFINE_IDA(kmem_limited_groups
);
607 int memcg_limited_groups_array_size
;
610 * MIN_SIZE is different than 1, because we would like to avoid going through
611 * the alloc/free process all the time. In a small machine, 4 kmem-limited
612 * cgroups is a reasonable guess. In the future, it could be a parameter or
613 * tunable, but that is strictly not necessary.
615 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
616 * this constant directly from cgroup, but it is understandable that this is
617 * better kept as an internal representation in cgroup.c. In any case, the
618 * cgrp_id space is not getting any smaller, and we don't have to necessarily
619 * increase ours as well if it increases.
621 #define MEMCG_CACHES_MIN_SIZE 4
622 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
625 * A lot of the calls to the cache allocation functions are expected to be
626 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
627 * conditional to this static branch, we'll have to allow modules that does
628 * kmem_cache_alloc and the such to see this symbol as well
630 struct static_key memcg_kmem_enabled_key
;
631 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
633 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
635 if (memcg_kmem_is_active(memcg
)) {
636 static_key_slow_dec(&memcg_kmem_enabled_key
);
637 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
640 * This check can't live in kmem destruction function,
641 * since the charges will outlive the cgroup
643 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
646 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
649 #endif /* CONFIG_MEMCG_KMEM */
651 static void disarm_static_keys(struct mem_cgroup
*memcg
)
653 disarm_sock_keys(memcg
);
654 disarm_kmem_keys(memcg
);
657 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
659 static struct mem_cgroup_per_zone
*
660 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
662 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
663 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
666 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
671 static struct mem_cgroup_per_zone
*
672 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
674 int nid
= page_to_nid(page
);
675 int zid
= page_zonenum(page
);
677 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
680 static struct mem_cgroup_tree_per_zone
*
681 soft_limit_tree_node_zone(int nid
, int zid
)
683 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
686 static struct mem_cgroup_tree_per_zone
*
687 soft_limit_tree_from_page(struct page
*page
)
689 int nid
= page_to_nid(page
);
690 int zid
= page_zonenum(page
);
692 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
696 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
697 struct mem_cgroup_per_zone
*mz
,
698 struct mem_cgroup_tree_per_zone
*mctz
,
699 unsigned long long new_usage_in_excess
)
701 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
702 struct rb_node
*parent
= NULL
;
703 struct mem_cgroup_per_zone
*mz_node
;
708 mz
->usage_in_excess
= new_usage_in_excess
;
709 if (!mz
->usage_in_excess
)
713 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
715 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
718 * We can't avoid mem cgroups that are over their soft
719 * limit by the same amount
721 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
724 rb_link_node(&mz
->tree_node
, parent
, p
);
725 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
730 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
731 struct mem_cgroup_per_zone
*mz
,
732 struct mem_cgroup_tree_per_zone
*mctz
)
736 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
741 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
742 struct mem_cgroup_per_zone
*mz
,
743 struct mem_cgroup_tree_per_zone
*mctz
)
745 spin_lock(&mctz
->lock
);
746 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
747 spin_unlock(&mctz
->lock
);
751 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
753 unsigned long long excess
;
754 struct mem_cgroup_per_zone
*mz
;
755 struct mem_cgroup_tree_per_zone
*mctz
;
756 int nid
= page_to_nid(page
);
757 int zid
= page_zonenum(page
);
758 mctz
= soft_limit_tree_from_page(page
);
761 * Necessary to update all ancestors when hierarchy is used.
762 * because their event counter is not touched.
764 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
765 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
766 excess
= res_counter_soft_limit_excess(&memcg
->res
);
768 * We have to update the tree if mz is on RB-tree or
769 * mem is over its softlimit.
771 if (excess
|| mz
->on_tree
) {
772 spin_lock(&mctz
->lock
);
773 /* if on-tree, remove it */
775 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
777 * Insert again. mz->usage_in_excess will be updated.
778 * If excess is 0, no tree ops.
780 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
781 spin_unlock(&mctz
->lock
);
786 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
789 struct mem_cgroup_per_zone
*mz
;
790 struct mem_cgroup_tree_per_zone
*mctz
;
792 for_each_node(node
) {
793 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
794 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
795 mctz
= soft_limit_tree_node_zone(node
, zone
);
796 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
801 static struct mem_cgroup_per_zone
*
802 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
804 struct rb_node
*rightmost
= NULL
;
805 struct mem_cgroup_per_zone
*mz
;
809 rightmost
= rb_last(&mctz
->rb_root
);
811 goto done
; /* Nothing to reclaim from */
813 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
815 * Remove the node now but someone else can add it back,
816 * we will to add it back at the end of reclaim to its correct
817 * position in the tree.
819 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
820 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
821 !css_tryget(&mz
->memcg
->css
))
827 static struct mem_cgroup_per_zone
*
828 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
830 struct mem_cgroup_per_zone
*mz
;
832 spin_lock(&mctz
->lock
);
833 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
834 spin_unlock(&mctz
->lock
);
839 * Implementation Note: reading percpu statistics for memcg.
841 * Both of vmstat[] and percpu_counter has threshold and do periodic
842 * synchronization to implement "quick" read. There are trade-off between
843 * reading cost and precision of value. Then, we may have a chance to implement
844 * a periodic synchronizion of counter in memcg's counter.
846 * But this _read() function is used for user interface now. The user accounts
847 * memory usage by memory cgroup and he _always_ requires exact value because
848 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
849 * have to visit all online cpus and make sum. So, for now, unnecessary
850 * synchronization is not implemented. (just implemented for cpu hotplug)
852 * If there are kernel internal actions which can make use of some not-exact
853 * value, and reading all cpu value can be performance bottleneck in some
854 * common workload, threashold and synchonization as vmstat[] should be
857 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
858 enum mem_cgroup_stat_index idx
)
864 for_each_online_cpu(cpu
)
865 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
866 #ifdef CONFIG_HOTPLUG_CPU
867 spin_lock(&memcg
->pcp_counter_lock
);
868 val
+= memcg
->nocpu_base
.count
[idx
];
869 spin_unlock(&memcg
->pcp_counter_lock
);
875 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
878 int val
= (charge
) ? 1 : -1;
879 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
882 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
883 enum mem_cgroup_events_index idx
)
885 unsigned long val
= 0;
889 for_each_online_cpu(cpu
)
890 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
891 #ifdef CONFIG_HOTPLUG_CPU
892 spin_lock(&memcg
->pcp_counter_lock
);
893 val
+= memcg
->nocpu_base
.events
[idx
];
894 spin_unlock(&memcg
->pcp_counter_lock
);
900 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
902 bool anon
, int nr_pages
)
907 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
908 * counted as CACHE even if it's on ANON LRU.
911 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
914 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
917 if (PageTransHuge(page
))
918 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
921 /* pagein of a big page is an event. So, ignore page size */
923 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
925 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
926 nr_pages
= -nr_pages
; /* for event */
929 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
935 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
937 struct mem_cgroup_per_zone
*mz
;
939 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
940 return mz
->lru_size
[lru
];
944 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
945 unsigned int lru_mask
)
947 struct mem_cgroup_per_zone
*mz
;
949 unsigned long ret
= 0;
951 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
954 if (BIT(lru
) & lru_mask
)
955 ret
+= mz
->lru_size
[lru
];
961 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
962 int nid
, unsigned int lru_mask
)
967 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
968 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
974 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
975 unsigned int lru_mask
)
980 for_each_node_state(nid
, N_MEMORY
)
981 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
985 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
986 enum mem_cgroup_events_target target
)
988 unsigned long val
, next
;
990 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
991 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
992 /* from time_after() in jiffies.h */
993 if ((long)next
- (long)val
< 0) {
995 case MEM_CGROUP_TARGET_THRESH
:
996 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
998 case MEM_CGROUP_TARGET_SOFTLIMIT
:
999 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
1001 case MEM_CGROUP_TARGET_NUMAINFO
:
1002 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1007 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1014 * Check events in order.
1017 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1020 /* threshold event is triggered in finer grain than soft limit */
1021 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1022 MEM_CGROUP_TARGET_THRESH
))) {
1024 bool do_numainfo __maybe_unused
;
1026 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1027 MEM_CGROUP_TARGET_SOFTLIMIT
);
1028 #if MAX_NUMNODES > 1
1029 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1030 MEM_CGROUP_TARGET_NUMAINFO
);
1034 mem_cgroup_threshold(memcg
);
1035 if (unlikely(do_softlimit
))
1036 mem_cgroup_update_tree(memcg
, page
);
1037 #if MAX_NUMNODES > 1
1038 if (unlikely(do_numainfo
))
1039 atomic_inc(&memcg
->numainfo_events
);
1045 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1048 * mm_update_next_owner() may clear mm->owner to NULL
1049 * if it races with swapoff, page migration, etc.
1050 * So this can be called with p == NULL.
1055 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
1058 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1060 struct mem_cgroup
*memcg
= NULL
;
1065 * Because we have no locks, mm->owner's may be being moved to other
1066 * cgroup. We use css_tryget() here even if this looks
1067 * pessimistic (rather than adding locks here).
1071 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1072 if (unlikely(!memcg
))
1074 } while (!css_tryget(&memcg
->css
));
1080 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1081 * ref. count) or NULL if the whole root's subtree has been visited.
1083 * helper function to be used by mem_cgroup_iter
1085 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1086 struct mem_cgroup
*last_visited
)
1088 struct cgroup_subsys_state
*prev_css
, *next_css
;
1090 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1092 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1095 * Even if we found a group we have to make sure it is
1096 * alive. css && !memcg means that the groups should be
1097 * skipped and we should continue the tree walk.
1098 * last_visited css is safe to use because it is
1099 * protected by css_get and the tree walk is rcu safe.
1102 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1104 if (css_tryget(&mem
->css
))
1107 prev_css
= next_css
;
1115 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1118 * When a group in the hierarchy below root is destroyed, the
1119 * hierarchy iterator can no longer be trusted since it might
1120 * have pointed to the destroyed group. Invalidate it.
1122 atomic_inc(&root
->dead_count
);
1125 static struct mem_cgroup
*
1126 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1127 struct mem_cgroup
*root
,
1130 struct mem_cgroup
*position
= NULL
;
1132 * A cgroup destruction happens in two stages: offlining and
1133 * release. They are separated by a RCU grace period.
1135 * If the iterator is valid, we may still race with an
1136 * offlining. The RCU lock ensures the object won't be
1137 * released, tryget will fail if we lost the race.
1139 *sequence
= atomic_read(&root
->dead_count
);
1140 if (iter
->last_dead_count
== *sequence
) {
1142 position
= iter
->last_visited
;
1143 if (position
&& !css_tryget(&position
->css
))
1149 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1150 struct mem_cgroup
*last_visited
,
1151 struct mem_cgroup
*new_position
,
1155 css_put(&last_visited
->css
);
1157 * We store the sequence count from the time @last_visited was
1158 * loaded successfully instead of rereading it here so that we
1159 * don't lose destruction events in between. We could have
1160 * raced with the destruction of @new_position after all.
1162 iter
->last_visited
= new_position
;
1164 iter
->last_dead_count
= sequence
;
1168 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1169 * @root: hierarchy root
1170 * @prev: previously returned memcg, NULL on first invocation
1171 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1173 * Returns references to children of the hierarchy below @root, or
1174 * @root itself, or %NULL after a full round-trip.
1176 * Caller must pass the return value in @prev on subsequent
1177 * invocations for reference counting, or use mem_cgroup_iter_break()
1178 * to cancel a hierarchy walk before the round-trip is complete.
1180 * Reclaimers can specify a zone and a priority level in @reclaim to
1181 * divide up the memcgs in the hierarchy among all concurrent
1182 * reclaimers operating on the same zone and priority.
1184 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1185 struct mem_cgroup
*prev
,
1186 struct mem_cgroup_reclaim_cookie
*reclaim
)
1188 struct mem_cgroup
*memcg
= NULL
;
1189 struct mem_cgroup
*last_visited
= NULL
;
1191 if (mem_cgroup_disabled())
1195 root
= root_mem_cgroup
;
1197 if (prev
&& !reclaim
)
1198 last_visited
= prev
;
1200 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1208 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1209 int uninitialized_var(seq
);
1212 int nid
= zone_to_nid(reclaim
->zone
);
1213 int zid
= zone_idx(reclaim
->zone
);
1214 struct mem_cgroup_per_zone
*mz
;
1216 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1217 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1218 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1219 iter
->last_visited
= NULL
;
1223 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1226 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1229 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1233 else if (!prev
&& memcg
)
1234 reclaim
->generation
= iter
->generation
;
1243 if (prev
&& prev
!= root
)
1244 css_put(&prev
->css
);
1250 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1251 * @root: hierarchy root
1252 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1254 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1255 struct mem_cgroup
*prev
)
1258 root
= root_mem_cgroup
;
1259 if (prev
&& prev
!= root
)
1260 css_put(&prev
->css
);
1264 * Iteration constructs for visiting all cgroups (under a tree). If
1265 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1266 * be used for reference counting.
1268 #define for_each_mem_cgroup_tree(iter, root) \
1269 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1271 iter = mem_cgroup_iter(root, iter, NULL))
1273 #define for_each_mem_cgroup(iter) \
1274 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1276 iter = mem_cgroup_iter(NULL, iter, NULL))
1278 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1280 struct mem_cgroup
*memcg
;
1283 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1284 if (unlikely(!memcg
))
1289 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1292 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1300 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1303 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1304 * @zone: zone of the wanted lruvec
1305 * @memcg: memcg of the wanted lruvec
1307 * Returns the lru list vector holding pages for the given @zone and
1308 * @mem. This can be the global zone lruvec, if the memory controller
1311 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1312 struct mem_cgroup
*memcg
)
1314 struct mem_cgroup_per_zone
*mz
;
1315 struct lruvec
*lruvec
;
1317 if (mem_cgroup_disabled()) {
1318 lruvec
= &zone
->lruvec
;
1322 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1323 lruvec
= &mz
->lruvec
;
1326 * Since a node can be onlined after the mem_cgroup was created,
1327 * we have to be prepared to initialize lruvec->zone here;
1328 * and if offlined then reonlined, we need to reinitialize it.
1330 if (unlikely(lruvec
->zone
!= zone
))
1331 lruvec
->zone
= zone
;
1336 * Following LRU functions are allowed to be used without PCG_LOCK.
1337 * Operations are called by routine of global LRU independently from memcg.
1338 * What we have to take care of here is validness of pc->mem_cgroup.
1340 * Changes to pc->mem_cgroup happens when
1343 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1344 * It is added to LRU before charge.
1345 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1346 * When moving account, the page is not on LRU. It's isolated.
1350 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1352 * @zone: zone of the page
1354 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1356 struct mem_cgroup_per_zone
*mz
;
1357 struct mem_cgroup
*memcg
;
1358 struct page_cgroup
*pc
;
1359 struct lruvec
*lruvec
;
1361 if (mem_cgroup_disabled()) {
1362 lruvec
= &zone
->lruvec
;
1366 pc
= lookup_page_cgroup(page
);
1367 memcg
= pc
->mem_cgroup
;
1370 * Surreptitiously switch any uncharged offlist page to root:
1371 * an uncharged page off lru does nothing to secure
1372 * its former mem_cgroup from sudden removal.
1374 * Our caller holds lru_lock, and PageCgroupUsed is updated
1375 * under page_cgroup lock: between them, they make all uses
1376 * of pc->mem_cgroup safe.
1378 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1379 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1381 mz
= page_cgroup_zoneinfo(memcg
, page
);
1382 lruvec
= &mz
->lruvec
;
1385 * Since a node can be onlined after the mem_cgroup was created,
1386 * we have to be prepared to initialize lruvec->zone here;
1387 * and if offlined then reonlined, we need to reinitialize it.
1389 if (unlikely(lruvec
->zone
!= zone
))
1390 lruvec
->zone
= zone
;
1395 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1396 * @lruvec: mem_cgroup per zone lru vector
1397 * @lru: index of lru list the page is sitting on
1398 * @nr_pages: positive when adding or negative when removing
1400 * This function must be called when a page is added to or removed from an
1403 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1406 struct mem_cgroup_per_zone
*mz
;
1407 unsigned long *lru_size
;
1409 if (mem_cgroup_disabled())
1412 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1413 lru_size
= mz
->lru_size
+ lru
;
1414 *lru_size
+= nr_pages
;
1415 VM_BUG_ON((long)(*lru_size
) < 0);
1419 * Checks whether given mem is same or in the root_mem_cgroup's
1422 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1423 struct mem_cgroup
*memcg
)
1425 if (root_memcg
== memcg
)
1427 if (!root_memcg
->use_hierarchy
|| !memcg
)
1429 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1432 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1433 struct mem_cgroup
*memcg
)
1438 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1443 bool task_in_mem_cgroup(struct task_struct
*task
,
1444 const struct mem_cgroup
*memcg
)
1446 struct mem_cgroup
*curr
= NULL
;
1447 struct task_struct
*p
;
1450 p
= find_lock_task_mm(task
);
1452 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1456 * All threads may have already detached their mm's, but the oom
1457 * killer still needs to detect if they have already been oom
1458 * killed to prevent needlessly killing additional tasks.
1461 curr
= mem_cgroup_from_task(task
);
1463 css_get(&curr
->css
);
1469 * We should check use_hierarchy of "memcg" not "curr". Because checking
1470 * use_hierarchy of "curr" here make this function true if hierarchy is
1471 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1472 * hierarchy(even if use_hierarchy is disabled in "memcg").
1474 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1475 css_put(&curr
->css
);
1479 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1481 unsigned long inactive_ratio
;
1482 unsigned long inactive
;
1483 unsigned long active
;
1486 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1487 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1489 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1491 inactive_ratio
= int_sqrt(10 * gb
);
1495 return inactive
* inactive_ratio
< active
;
1498 #define mem_cgroup_from_res_counter(counter, member) \
1499 container_of(counter, struct mem_cgroup, member)
1502 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1503 * @memcg: the memory cgroup
1505 * Returns the maximum amount of memory @mem can be charged with, in
1508 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1510 unsigned long long margin
;
1512 margin
= res_counter_margin(&memcg
->res
);
1513 if (do_swap_account
)
1514 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1515 return margin
>> PAGE_SHIFT
;
1518 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1521 if (!css_parent(&memcg
->css
))
1522 return vm_swappiness
;
1524 return memcg
->swappiness
;
1528 * memcg->moving_account is used for checking possibility that some thread is
1529 * calling move_account(). When a thread on CPU-A starts moving pages under
1530 * a memcg, other threads should check memcg->moving_account under
1531 * rcu_read_lock(), like this:
1535 * memcg->moving_account+1 if (memcg->mocing_account)
1537 * synchronize_rcu() update something.
1542 /* for quick checking without looking up memcg */
1543 atomic_t memcg_moving __read_mostly
;
1545 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1547 atomic_inc(&memcg_moving
);
1548 atomic_inc(&memcg
->moving_account
);
1552 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1555 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1556 * We check NULL in callee rather than caller.
1559 atomic_dec(&memcg_moving
);
1560 atomic_dec(&memcg
->moving_account
);
1565 * 2 routines for checking "mem" is under move_account() or not.
1567 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1568 * is used for avoiding races in accounting. If true,
1569 * pc->mem_cgroup may be overwritten.
1571 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1572 * under hierarchy of moving cgroups. This is for
1573 * waiting at hith-memory prressure caused by "move".
1576 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1578 VM_BUG_ON(!rcu_read_lock_held());
1579 return atomic_read(&memcg
->moving_account
) > 0;
1582 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1584 struct mem_cgroup
*from
;
1585 struct mem_cgroup
*to
;
1588 * Unlike task_move routines, we access mc.to, mc.from not under
1589 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1591 spin_lock(&mc
.lock
);
1597 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1598 || mem_cgroup_same_or_subtree(memcg
, to
);
1600 spin_unlock(&mc
.lock
);
1604 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1606 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1607 if (mem_cgroup_under_move(memcg
)) {
1609 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1610 /* moving charge context might have finished. */
1613 finish_wait(&mc
.waitq
, &wait
);
1621 * Take this lock when
1622 * - a code tries to modify page's memcg while it's USED.
1623 * - a code tries to modify page state accounting in a memcg.
1624 * see mem_cgroup_stolen(), too.
1626 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1627 unsigned long *flags
)
1629 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1632 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1633 unsigned long *flags
)
1635 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1638 #define K(x) ((x) << (PAGE_SHIFT-10))
1640 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1641 * @memcg: The memory cgroup that went over limit
1642 * @p: Task that is going to be killed
1644 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1647 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1649 struct cgroup
*task_cgrp
;
1650 struct cgroup
*mem_cgrp
;
1652 * Need a buffer in BSS, can't rely on allocations. The code relies
1653 * on the assumption that OOM is serialized for memory controller.
1654 * If this assumption is broken, revisit this code.
1656 static char memcg_name
[PATH_MAX
];
1658 struct mem_cgroup
*iter
;
1666 mem_cgrp
= memcg
->css
.cgroup
;
1667 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1669 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1672 * Unfortunately, we are unable to convert to a useful name
1673 * But we'll still print out the usage information
1680 pr_info("Task in %s killed", memcg_name
);
1683 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1691 * Continues from above, so we don't need an KERN_ level
1693 pr_cont(" as a result of limit of %s\n", memcg_name
);
1696 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1697 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1698 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1699 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1700 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1701 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1702 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1703 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1704 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1705 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1706 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1707 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1709 for_each_mem_cgroup_tree(iter
, memcg
) {
1710 pr_info("Memory cgroup stats");
1713 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1715 pr_cont(" for %s", memcg_name
);
1719 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1720 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1722 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1723 K(mem_cgroup_read_stat(iter
, i
)));
1726 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1727 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1728 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1735 * This function returns the number of memcg under hierarchy tree. Returns
1736 * 1(self count) if no children.
1738 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1741 struct mem_cgroup
*iter
;
1743 for_each_mem_cgroup_tree(iter
, memcg
)
1749 * Return the memory (and swap, if configured) limit for a memcg.
1751 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1755 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1758 * Do not consider swap space if we cannot swap due to swappiness
1760 if (mem_cgroup_swappiness(memcg
)) {
1763 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1764 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1767 * If memsw is finite and limits the amount of swap space
1768 * available to this memcg, return that limit.
1770 limit
= min(limit
, memsw
);
1776 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1779 struct mem_cgroup
*iter
;
1780 unsigned long chosen_points
= 0;
1781 unsigned long totalpages
;
1782 unsigned int points
= 0;
1783 struct task_struct
*chosen
= NULL
;
1786 * If current has a pending SIGKILL or is exiting, then automatically
1787 * select it. The goal is to allow it to allocate so that it may
1788 * quickly exit and free its memory.
1790 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1791 set_thread_flag(TIF_MEMDIE
);
1795 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1796 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1797 for_each_mem_cgroup_tree(iter
, memcg
) {
1798 struct css_task_iter it
;
1799 struct task_struct
*task
;
1801 css_task_iter_start(&iter
->css
, &it
);
1802 while ((task
= css_task_iter_next(&it
))) {
1803 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1805 case OOM_SCAN_SELECT
:
1807 put_task_struct(chosen
);
1809 chosen_points
= ULONG_MAX
;
1810 get_task_struct(chosen
);
1812 case OOM_SCAN_CONTINUE
:
1814 case OOM_SCAN_ABORT
:
1815 css_task_iter_end(&it
);
1816 mem_cgroup_iter_break(memcg
, iter
);
1818 put_task_struct(chosen
);
1823 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1824 if (points
> chosen_points
) {
1826 put_task_struct(chosen
);
1828 chosen_points
= points
;
1829 get_task_struct(chosen
);
1832 css_task_iter_end(&it
);
1837 points
= chosen_points
* 1000 / totalpages
;
1838 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1839 NULL
, "Memory cgroup out of memory");
1842 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1844 unsigned long flags
)
1846 unsigned long total
= 0;
1847 bool noswap
= false;
1850 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1852 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1855 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1857 drain_all_stock_async(memcg
);
1858 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1860 * Allow limit shrinkers, which are triggered directly
1861 * by userspace, to catch signals and stop reclaim
1862 * after minimal progress, regardless of the margin.
1864 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1866 if (mem_cgroup_margin(memcg
))
1869 * If nothing was reclaimed after two attempts, there
1870 * may be no reclaimable pages in this hierarchy.
1879 * test_mem_cgroup_node_reclaimable
1880 * @memcg: the target memcg
1881 * @nid: the node ID to be checked.
1882 * @noswap : specify true here if the user wants flle only information.
1884 * This function returns whether the specified memcg contains any
1885 * reclaimable pages on a node. Returns true if there are any reclaimable
1886 * pages in the node.
1888 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1889 int nid
, bool noswap
)
1891 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1893 if (noswap
|| !total_swap_pages
)
1895 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1900 #if MAX_NUMNODES > 1
1903 * Always updating the nodemask is not very good - even if we have an empty
1904 * list or the wrong list here, we can start from some node and traverse all
1905 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1908 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1912 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1913 * pagein/pageout changes since the last update.
1915 if (!atomic_read(&memcg
->numainfo_events
))
1917 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1920 /* make a nodemask where this memcg uses memory from */
1921 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1923 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1925 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1926 node_clear(nid
, memcg
->scan_nodes
);
1929 atomic_set(&memcg
->numainfo_events
, 0);
1930 atomic_set(&memcg
->numainfo_updating
, 0);
1934 * Selecting a node where we start reclaim from. Because what we need is just
1935 * reducing usage counter, start from anywhere is O,K. Considering
1936 * memory reclaim from current node, there are pros. and cons.
1938 * Freeing memory from current node means freeing memory from a node which
1939 * we'll use or we've used. So, it may make LRU bad. And if several threads
1940 * hit limits, it will see a contention on a node. But freeing from remote
1941 * node means more costs for memory reclaim because of memory latency.
1943 * Now, we use round-robin. Better algorithm is welcomed.
1945 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1949 mem_cgroup_may_update_nodemask(memcg
);
1950 node
= memcg
->last_scanned_node
;
1952 node
= next_node(node
, memcg
->scan_nodes
);
1953 if (node
== MAX_NUMNODES
)
1954 node
= first_node(memcg
->scan_nodes
);
1956 * We call this when we hit limit, not when pages are added to LRU.
1957 * No LRU may hold pages because all pages are UNEVICTABLE or
1958 * memcg is too small and all pages are not on LRU. In that case,
1959 * we use curret node.
1961 if (unlikely(node
== MAX_NUMNODES
))
1962 node
= numa_node_id();
1964 memcg
->last_scanned_node
= node
;
1969 * Check all nodes whether it contains reclaimable pages or not.
1970 * For quick scan, we make use of scan_nodes. This will allow us to skip
1971 * unused nodes. But scan_nodes is lazily updated and may not cotain
1972 * enough new information. We need to do double check.
1974 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1979 * quick check...making use of scan_node.
1980 * We can skip unused nodes.
1982 if (!nodes_empty(memcg
->scan_nodes
)) {
1983 for (nid
= first_node(memcg
->scan_nodes
);
1985 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1987 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1992 * Check rest of nodes.
1994 for_each_node_state(nid
, N_MEMORY
) {
1995 if (node_isset(nid
, memcg
->scan_nodes
))
1997 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
2004 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
2009 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
2011 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
2015 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2018 unsigned long *total_scanned
)
2020 struct mem_cgroup
*victim
= NULL
;
2023 unsigned long excess
;
2024 unsigned long nr_scanned
;
2025 struct mem_cgroup_reclaim_cookie reclaim
= {
2030 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2033 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2038 * If we have not been able to reclaim
2039 * anything, it might because there are
2040 * no reclaimable pages under this hierarchy
2045 * We want to do more targeted reclaim.
2046 * excess >> 2 is not to excessive so as to
2047 * reclaim too much, nor too less that we keep
2048 * coming back to reclaim from this cgroup
2050 if (total
>= (excess
>> 2) ||
2051 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2056 if (!mem_cgroup_reclaimable(victim
, false))
2058 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2060 *total_scanned
+= nr_scanned
;
2061 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2064 mem_cgroup_iter_break(root_memcg
, victim
);
2068 #ifdef CONFIG_LOCKDEP
2069 static struct lockdep_map memcg_oom_lock_dep_map
= {
2070 .name
= "memcg_oom_lock",
2074 static DEFINE_SPINLOCK(memcg_oom_lock
);
2077 * Check OOM-Killer is already running under our hierarchy.
2078 * If someone is running, return false.
2080 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2082 struct mem_cgroup
*iter
, *failed
= NULL
;
2084 spin_lock(&memcg_oom_lock
);
2086 for_each_mem_cgroup_tree(iter
, memcg
) {
2087 if (iter
->oom_lock
) {
2089 * this subtree of our hierarchy is already locked
2090 * so we cannot give a lock.
2093 mem_cgroup_iter_break(memcg
, iter
);
2096 iter
->oom_lock
= true;
2101 * OK, we failed to lock the whole subtree so we have
2102 * to clean up what we set up to the failing subtree
2104 for_each_mem_cgroup_tree(iter
, memcg
) {
2105 if (iter
== failed
) {
2106 mem_cgroup_iter_break(memcg
, iter
);
2109 iter
->oom_lock
= false;
2112 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2114 spin_unlock(&memcg_oom_lock
);
2119 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2121 struct mem_cgroup
*iter
;
2123 spin_lock(&memcg_oom_lock
);
2124 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2125 for_each_mem_cgroup_tree(iter
, memcg
)
2126 iter
->oom_lock
= false;
2127 spin_unlock(&memcg_oom_lock
);
2130 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2132 struct mem_cgroup
*iter
;
2134 for_each_mem_cgroup_tree(iter
, memcg
)
2135 atomic_inc(&iter
->under_oom
);
2138 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2140 struct mem_cgroup
*iter
;
2143 * When a new child is created while the hierarchy is under oom,
2144 * mem_cgroup_oom_lock() may not be called. We have to use
2145 * atomic_add_unless() here.
2147 for_each_mem_cgroup_tree(iter
, memcg
)
2148 atomic_add_unless(&iter
->under_oom
, -1, 0);
2151 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2153 struct oom_wait_info
{
2154 struct mem_cgroup
*memcg
;
2158 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2159 unsigned mode
, int sync
, void *arg
)
2161 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2162 struct mem_cgroup
*oom_wait_memcg
;
2163 struct oom_wait_info
*oom_wait_info
;
2165 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2166 oom_wait_memcg
= oom_wait_info
->memcg
;
2169 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2170 * Then we can use css_is_ancestor without taking care of RCU.
2172 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2173 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2175 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2178 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2180 atomic_inc(&memcg
->oom_wakeups
);
2181 /* for filtering, pass "memcg" as argument. */
2182 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2185 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2187 if (memcg
&& atomic_read(&memcg
->under_oom
))
2188 memcg_wakeup_oom(memcg
);
2191 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2193 if (!current
->memcg_oom
.may_oom
)
2196 * We are in the middle of the charge context here, so we
2197 * don't want to block when potentially sitting on a callstack
2198 * that holds all kinds of filesystem and mm locks.
2200 * Also, the caller may handle a failed allocation gracefully
2201 * (like optional page cache readahead) and so an OOM killer
2202 * invocation might not even be necessary.
2204 * That's why we don't do anything here except remember the
2205 * OOM context and then deal with it at the end of the page
2206 * fault when the stack is unwound, the locks are released,
2207 * and when we know whether the fault was overall successful.
2209 css_get(&memcg
->css
);
2210 current
->memcg_oom
.memcg
= memcg
;
2211 current
->memcg_oom
.gfp_mask
= mask
;
2212 current
->memcg_oom
.order
= order
;
2216 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2217 * @handle: actually kill/wait or just clean up the OOM state
2219 * This has to be called at the end of a page fault if the memcg OOM
2220 * handler was enabled.
2222 * Memcg supports userspace OOM handling where failed allocations must
2223 * sleep on a waitqueue until the userspace task resolves the
2224 * situation. Sleeping directly in the charge context with all kinds
2225 * of locks held is not a good idea, instead we remember an OOM state
2226 * in the task and mem_cgroup_oom_synchronize() has to be called at
2227 * the end of the page fault to complete the OOM handling.
2229 * Returns %true if an ongoing memcg OOM situation was detected and
2230 * completed, %false otherwise.
2232 bool mem_cgroup_oom_synchronize(bool handle
)
2234 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2235 struct oom_wait_info owait
;
2238 /* OOM is global, do not handle */
2245 owait
.memcg
= memcg
;
2246 owait
.wait
.flags
= 0;
2247 owait
.wait
.func
= memcg_oom_wake_function
;
2248 owait
.wait
.private = current
;
2249 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2251 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2252 mem_cgroup_mark_under_oom(memcg
);
2254 locked
= mem_cgroup_oom_trylock(memcg
);
2257 mem_cgroup_oom_notify(memcg
);
2259 if (locked
&& !memcg
->oom_kill_disable
) {
2260 mem_cgroup_unmark_under_oom(memcg
);
2261 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2262 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2263 current
->memcg_oom
.order
);
2266 mem_cgroup_unmark_under_oom(memcg
);
2267 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2271 mem_cgroup_oom_unlock(memcg
);
2273 * There is no guarantee that an OOM-lock contender
2274 * sees the wakeups triggered by the OOM kill
2275 * uncharges. Wake any sleepers explicitely.
2277 memcg_oom_recover(memcg
);
2280 current
->memcg_oom
.memcg
= NULL
;
2281 css_put(&memcg
->css
);
2286 * Currently used to update mapped file statistics, but the routine can be
2287 * generalized to update other statistics as well.
2289 * Notes: Race condition
2291 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2292 * it tends to be costly. But considering some conditions, we doesn't need
2293 * to do so _always_.
2295 * Considering "charge", lock_page_cgroup() is not required because all
2296 * file-stat operations happen after a page is attached to radix-tree. There
2297 * are no race with "charge".
2299 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2300 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2301 * if there are race with "uncharge". Statistics itself is properly handled
2304 * Considering "move", this is an only case we see a race. To make the race
2305 * small, we check mm->moving_account and detect there are possibility of race
2306 * If there is, we take a lock.
2309 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2310 bool *locked
, unsigned long *flags
)
2312 struct mem_cgroup
*memcg
;
2313 struct page_cgroup
*pc
;
2315 pc
= lookup_page_cgroup(page
);
2317 memcg
= pc
->mem_cgroup
;
2318 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2321 * If this memory cgroup is not under account moving, we don't
2322 * need to take move_lock_mem_cgroup(). Because we already hold
2323 * rcu_read_lock(), any calls to move_account will be delayed until
2324 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2326 if (!mem_cgroup_stolen(memcg
))
2329 move_lock_mem_cgroup(memcg
, flags
);
2330 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2331 move_unlock_mem_cgroup(memcg
, flags
);
2337 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2339 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2342 * It's guaranteed that pc->mem_cgroup never changes while
2343 * lock is held because a routine modifies pc->mem_cgroup
2344 * should take move_lock_mem_cgroup().
2346 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2349 void mem_cgroup_update_page_stat(struct page
*page
,
2350 enum mem_cgroup_stat_index idx
, int val
)
2352 struct mem_cgroup
*memcg
;
2353 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2354 unsigned long uninitialized_var(flags
);
2356 if (mem_cgroup_disabled())
2359 VM_BUG_ON(!rcu_read_lock_held());
2360 memcg
= pc
->mem_cgroup
;
2361 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2364 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2368 * size of first charge trial. "32" comes from vmscan.c's magic value.
2369 * TODO: maybe necessary to use big numbers in big irons.
2371 #define CHARGE_BATCH 32U
2372 struct memcg_stock_pcp
{
2373 struct mem_cgroup
*cached
; /* this never be root cgroup */
2374 unsigned int nr_pages
;
2375 struct work_struct work
;
2376 unsigned long flags
;
2377 #define FLUSHING_CACHED_CHARGE 0
2379 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2380 static DEFINE_MUTEX(percpu_charge_mutex
);
2383 * consume_stock: Try to consume stocked charge on this cpu.
2384 * @memcg: memcg to consume from.
2385 * @nr_pages: how many pages to charge.
2387 * The charges will only happen if @memcg matches the current cpu's memcg
2388 * stock, and at least @nr_pages are available in that stock. Failure to
2389 * service an allocation will refill the stock.
2391 * returns true if successful, false otherwise.
2393 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2395 struct memcg_stock_pcp
*stock
;
2398 if (nr_pages
> CHARGE_BATCH
)
2401 stock
= &get_cpu_var(memcg_stock
);
2402 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2403 stock
->nr_pages
-= nr_pages
;
2404 else /* need to call res_counter_charge */
2406 put_cpu_var(memcg_stock
);
2411 * Returns stocks cached in percpu to res_counter and reset cached information.
2413 static void drain_stock(struct memcg_stock_pcp
*stock
)
2415 struct mem_cgroup
*old
= stock
->cached
;
2417 if (stock
->nr_pages
) {
2418 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2420 res_counter_uncharge(&old
->res
, bytes
);
2421 if (do_swap_account
)
2422 res_counter_uncharge(&old
->memsw
, bytes
);
2423 stock
->nr_pages
= 0;
2425 stock
->cached
= NULL
;
2429 * This must be called under preempt disabled or must be called by
2430 * a thread which is pinned to local cpu.
2432 static void drain_local_stock(struct work_struct
*dummy
)
2434 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2436 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2439 static void __init
memcg_stock_init(void)
2443 for_each_possible_cpu(cpu
) {
2444 struct memcg_stock_pcp
*stock
=
2445 &per_cpu(memcg_stock
, cpu
);
2446 INIT_WORK(&stock
->work
, drain_local_stock
);
2451 * Cache charges(val) which is from res_counter, to local per_cpu area.
2452 * This will be consumed by consume_stock() function, later.
2454 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2456 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2458 if (stock
->cached
!= memcg
) { /* reset if necessary */
2460 stock
->cached
= memcg
;
2462 stock
->nr_pages
+= nr_pages
;
2463 put_cpu_var(memcg_stock
);
2467 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2468 * of the hierarchy under it. sync flag says whether we should block
2469 * until the work is done.
2471 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2475 /* Notify other cpus that system-wide "drain" is running */
2478 for_each_online_cpu(cpu
) {
2479 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2480 struct mem_cgroup
*memcg
;
2482 memcg
= stock
->cached
;
2483 if (!memcg
|| !stock
->nr_pages
)
2485 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2487 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2489 drain_local_stock(&stock
->work
);
2491 schedule_work_on(cpu
, &stock
->work
);
2499 for_each_online_cpu(cpu
) {
2500 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2501 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2502 flush_work(&stock
->work
);
2509 * Tries to drain stocked charges in other cpus. This function is asynchronous
2510 * and just put a work per cpu for draining localy on each cpu. Caller can
2511 * expects some charges will be back to res_counter later but cannot wait for
2514 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2517 * If someone calls draining, avoid adding more kworker runs.
2519 if (!mutex_trylock(&percpu_charge_mutex
))
2521 drain_all_stock(root_memcg
, false);
2522 mutex_unlock(&percpu_charge_mutex
);
2525 /* This is a synchronous drain interface. */
2526 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2528 /* called when force_empty is called */
2529 mutex_lock(&percpu_charge_mutex
);
2530 drain_all_stock(root_memcg
, true);
2531 mutex_unlock(&percpu_charge_mutex
);
2535 * This function drains percpu counter value from DEAD cpu and
2536 * move it to local cpu. Note that this function can be preempted.
2538 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2542 spin_lock(&memcg
->pcp_counter_lock
);
2543 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2544 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2546 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2547 memcg
->nocpu_base
.count
[i
] += x
;
2549 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2550 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2552 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2553 memcg
->nocpu_base
.events
[i
] += x
;
2555 spin_unlock(&memcg
->pcp_counter_lock
);
2558 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2559 unsigned long action
,
2562 int cpu
= (unsigned long)hcpu
;
2563 struct memcg_stock_pcp
*stock
;
2564 struct mem_cgroup
*iter
;
2566 if (action
== CPU_ONLINE
)
2569 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2572 for_each_mem_cgroup(iter
)
2573 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2575 stock
= &per_cpu(memcg_stock
, cpu
);
2581 /* See __mem_cgroup_try_charge() for details */
2583 CHARGE_OK
, /* success */
2584 CHARGE_RETRY
, /* need to retry but retry is not bad */
2585 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2586 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2589 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2590 unsigned int nr_pages
, unsigned int min_pages
,
2593 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2594 struct mem_cgroup
*mem_over_limit
;
2595 struct res_counter
*fail_res
;
2596 unsigned long flags
= 0;
2599 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2602 if (!do_swap_account
)
2604 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2608 res_counter_uncharge(&memcg
->res
, csize
);
2609 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2610 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2612 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2614 * Never reclaim on behalf of optional batching, retry with a
2615 * single page instead.
2617 if (nr_pages
> min_pages
)
2618 return CHARGE_RETRY
;
2620 if (!(gfp_mask
& __GFP_WAIT
))
2621 return CHARGE_WOULDBLOCK
;
2623 if (gfp_mask
& __GFP_NORETRY
)
2624 return CHARGE_NOMEM
;
2626 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2627 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2628 return CHARGE_RETRY
;
2630 * Even though the limit is exceeded at this point, reclaim
2631 * may have been able to free some pages. Retry the charge
2632 * before killing the task.
2634 * Only for regular pages, though: huge pages are rather
2635 * unlikely to succeed so close to the limit, and we fall back
2636 * to regular pages anyway in case of failure.
2638 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2639 return CHARGE_RETRY
;
2642 * At task move, charge accounts can be doubly counted. So, it's
2643 * better to wait until the end of task_move if something is going on.
2645 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2646 return CHARGE_RETRY
;
2649 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2651 return CHARGE_NOMEM
;
2655 * __mem_cgroup_try_charge() does
2656 * 1. detect memcg to be charged against from passed *mm and *ptr,
2657 * 2. update res_counter
2658 * 3. call memory reclaim if necessary.
2660 * In some special case, if the task is fatal, fatal_signal_pending() or
2661 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2662 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2663 * as possible without any hazards. 2: all pages should have a valid
2664 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2665 * pointer, that is treated as a charge to root_mem_cgroup.
2667 * So __mem_cgroup_try_charge() will return
2668 * 0 ... on success, filling *ptr with a valid memcg pointer.
2669 * -ENOMEM ... charge failure because of resource limits.
2670 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2672 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2673 * the oom-killer can be invoked.
2675 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2677 unsigned int nr_pages
,
2678 struct mem_cgroup
**ptr
,
2681 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2682 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2683 struct mem_cgroup
*memcg
= NULL
;
2687 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2688 * in system level. So, allow to go ahead dying process in addition to
2691 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2692 || fatal_signal_pending(current
)))
2695 if (unlikely(task_in_memcg_oom(current
)))
2699 * We always charge the cgroup the mm_struct belongs to.
2700 * The mm_struct's mem_cgroup changes on task migration if the
2701 * thread group leader migrates. It's possible that mm is not
2702 * set, if so charge the root memcg (happens for pagecache usage).
2705 *ptr
= root_mem_cgroup
;
2707 if (*ptr
) { /* css should be a valid one */
2709 if (mem_cgroup_is_root(memcg
))
2711 if (consume_stock(memcg
, nr_pages
))
2713 css_get(&memcg
->css
);
2715 struct task_struct
*p
;
2718 p
= rcu_dereference(mm
->owner
);
2720 * Because we don't have task_lock(), "p" can exit.
2721 * In that case, "memcg" can point to root or p can be NULL with
2722 * race with swapoff. Then, we have small risk of mis-accouning.
2723 * But such kind of mis-account by race always happens because
2724 * we don't have cgroup_mutex(). It's overkill and we allo that
2726 * (*) swapoff at el will charge against mm-struct not against
2727 * task-struct. So, mm->owner can be NULL.
2729 memcg
= mem_cgroup_from_task(p
);
2731 memcg
= root_mem_cgroup
;
2732 if (mem_cgroup_is_root(memcg
)) {
2736 if (consume_stock(memcg
, nr_pages
)) {
2738 * It seems dagerous to access memcg without css_get().
2739 * But considering how consume_stok works, it's not
2740 * necessary. If consume_stock success, some charges
2741 * from this memcg are cached on this cpu. So, we
2742 * don't need to call css_get()/css_tryget() before
2743 * calling consume_stock().
2748 /* after here, we may be blocked. we need to get refcnt */
2749 if (!css_tryget(&memcg
->css
)) {
2757 bool invoke_oom
= oom
&& !nr_oom_retries
;
2759 /* If killed, bypass charge */
2760 if (fatal_signal_pending(current
)) {
2761 css_put(&memcg
->css
);
2765 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2766 nr_pages
, invoke_oom
);
2770 case CHARGE_RETRY
: /* not in OOM situation but retry */
2772 css_put(&memcg
->css
);
2775 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2776 css_put(&memcg
->css
);
2778 case CHARGE_NOMEM
: /* OOM routine works */
2779 if (!oom
|| invoke_oom
) {
2780 css_put(&memcg
->css
);
2786 } while (ret
!= CHARGE_OK
);
2788 if (batch
> nr_pages
)
2789 refill_stock(memcg
, batch
- nr_pages
);
2790 css_put(&memcg
->css
);
2795 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2800 *ptr
= root_mem_cgroup
;
2805 * Somemtimes we have to undo a charge we got by try_charge().
2806 * This function is for that and do uncharge, put css's refcnt.
2807 * gotten by try_charge().
2809 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2810 unsigned int nr_pages
)
2812 if (!mem_cgroup_is_root(memcg
)) {
2813 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2815 res_counter_uncharge(&memcg
->res
, bytes
);
2816 if (do_swap_account
)
2817 res_counter_uncharge(&memcg
->memsw
, bytes
);
2822 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2823 * This is useful when moving usage to parent cgroup.
2825 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2826 unsigned int nr_pages
)
2828 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2830 if (mem_cgroup_is_root(memcg
))
2833 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2834 if (do_swap_account
)
2835 res_counter_uncharge_until(&memcg
->memsw
,
2836 memcg
->memsw
.parent
, bytes
);
2840 * A helper function to get mem_cgroup from ID. must be called under
2841 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2842 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2843 * called against removed memcg.)
2845 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2847 /* ID 0 is unused ID */
2850 return mem_cgroup_from_id(id
);
2853 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2855 struct mem_cgroup
*memcg
= NULL
;
2856 struct page_cgroup
*pc
;
2860 VM_BUG_ON(!PageLocked(page
));
2862 pc
= lookup_page_cgroup(page
);
2863 lock_page_cgroup(pc
);
2864 if (PageCgroupUsed(pc
)) {
2865 memcg
= pc
->mem_cgroup
;
2866 if (memcg
&& !css_tryget(&memcg
->css
))
2868 } else if (PageSwapCache(page
)) {
2869 ent
.val
= page_private(page
);
2870 id
= lookup_swap_cgroup_id(ent
);
2872 memcg
= mem_cgroup_lookup(id
);
2873 if (memcg
&& !css_tryget(&memcg
->css
))
2877 unlock_page_cgroup(pc
);
2881 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2883 unsigned int nr_pages
,
2884 enum charge_type ctype
,
2887 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2888 struct zone
*uninitialized_var(zone
);
2889 struct lruvec
*lruvec
;
2890 bool was_on_lru
= false;
2893 lock_page_cgroup(pc
);
2894 VM_BUG_ON(PageCgroupUsed(pc
));
2896 * we don't need page_cgroup_lock about tail pages, becase they are not
2897 * accessed by any other context at this point.
2901 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2902 * may already be on some other mem_cgroup's LRU. Take care of it.
2905 zone
= page_zone(page
);
2906 spin_lock_irq(&zone
->lru_lock
);
2907 if (PageLRU(page
)) {
2908 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2910 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2915 pc
->mem_cgroup
= memcg
;
2917 * We access a page_cgroup asynchronously without lock_page_cgroup().
2918 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2919 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2920 * before USED bit, we need memory barrier here.
2921 * See mem_cgroup_add_lru_list(), etc.
2924 SetPageCgroupUsed(pc
);
2928 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2929 VM_BUG_ON(PageLRU(page
));
2931 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2933 spin_unlock_irq(&zone
->lru_lock
);
2936 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2941 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2942 unlock_page_cgroup(pc
);
2945 * "charge_statistics" updated event counter. Then, check it.
2946 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2947 * if they exceeds softlimit.
2949 memcg_check_events(memcg
, page
);
2952 static DEFINE_MUTEX(set_limit_mutex
);
2954 #ifdef CONFIG_MEMCG_KMEM
2955 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2957 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2958 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2962 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2963 * in the memcg_cache_params struct.
2965 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2967 struct kmem_cache
*cachep
;
2969 VM_BUG_ON(p
->is_root_cache
);
2970 cachep
= p
->root_cache
;
2971 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2974 #ifdef CONFIG_SLABINFO
2975 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2976 struct cftype
*cft
, struct seq_file
*m
)
2978 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2979 struct memcg_cache_params
*params
;
2981 if (!memcg_can_account_kmem(memcg
))
2984 print_slabinfo_header(m
);
2986 mutex_lock(&memcg
->slab_caches_mutex
);
2987 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2988 cache_show(memcg_params_to_cache(params
), m
);
2989 mutex_unlock(&memcg
->slab_caches_mutex
);
2995 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2997 struct res_counter
*fail_res
;
2998 struct mem_cgroup
*_memcg
;
3002 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
3007 * Conditions under which we can wait for the oom_killer. Those are
3008 * the same conditions tested by the core page allocator
3010 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
3013 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
3016 if (ret
== -EINTR
) {
3018 * __mem_cgroup_try_charge() chosed to bypass to root due to
3019 * OOM kill or fatal signal. Since our only options are to
3020 * either fail the allocation or charge it to this cgroup, do
3021 * it as a temporary condition. But we can't fail. From a
3022 * kmem/slab perspective, the cache has already been selected,
3023 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3026 * This condition will only trigger if the task entered
3027 * memcg_charge_kmem in a sane state, but was OOM-killed during
3028 * __mem_cgroup_try_charge() above. Tasks that were already
3029 * dying when the allocation triggers should have been already
3030 * directed to the root cgroup in memcontrol.h
3032 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3033 if (do_swap_account
)
3034 res_counter_charge_nofail(&memcg
->memsw
, size
,
3038 res_counter_uncharge(&memcg
->kmem
, size
);
3043 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3045 res_counter_uncharge(&memcg
->res
, size
);
3046 if (do_swap_account
)
3047 res_counter_uncharge(&memcg
->memsw
, size
);
3050 if (res_counter_uncharge(&memcg
->kmem
, size
))
3054 * Releases a reference taken in kmem_cgroup_css_offline in case
3055 * this last uncharge is racing with the offlining code or it is
3056 * outliving the memcg existence.
3058 * The memory barrier imposed by test&clear is paired with the
3059 * explicit one in memcg_kmem_mark_dead().
3061 if (memcg_kmem_test_and_clear_dead(memcg
))
3062 css_put(&memcg
->css
);
3065 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3070 mutex_lock(&memcg
->slab_caches_mutex
);
3071 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3072 mutex_unlock(&memcg
->slab_caches_mutex
);
3076 * helper for acessing a memcg's index. It will be used as an index in the
3077 * child cache array in kmem_cache, and also to derive its name. This function
3078 * will return -1 when this is not a kmem-limited memcg.
3080 int memcg_cache_id(struct mem_cgroup
*memcg
)
3082 return memcg
? memcg
->kmemcg_id
: -1;
3086 * This ends up being protected by the set_limit mutex, during normal
3087 * operation, because that is its main call site.
3089 * But when we create a new cache, we can call this as well if its parent
3090 * is kmem-limited. That will have to hold set_limit_mutex as well.
3092 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3096 num
= ida_simple_get(&kmem_limited_groups
,
3097 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3101 * After this point, kmem_accounted (that we test atomically in
3102 * the beginning of this conditional), is no longer 0. This
3103 * guarantees only one process will set the following boolean
3104 * to true. We don't need test_and_set because we're protected
3105 * by the set_limit_mutex anyway.
3107 memcg_kmem_set_activated(memcg
);
3109 ret
= memcg_update_all_caches(num
+1);
3111 ida_simple_remove(&kmem_limited_groups
, num
);
3112 memcg_kmem_clear_activated(memcg
);
3116 memcg
->kmemcg_id
= num
;
3117 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3118 mutex_init(&memcg
->slab_caches_mutex
);
3122 static size_t memcg_caches_array_size(int num_groups
)
3125 if (num_groups
<= 0)
3128 size
= 2 * num_groups
;
3129 if (size
< MEMCG_CACHES_MIN_SIZE
)
3130 size
= MEMCG_CACHES_MIN_SIZE
;
3131 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3132 size
= MEMCG_CACHES_MAX_SIZE
;
3138 * We should update the current array size iff all caches updates succeed. This
3139 * can only be done from the slab side. The slab mutex needs to be held when
3142 void memcg_update_array_size(int num
)
3144 if (num
> memcg_limited_groups_array_size
)
3145 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3148 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3150 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3152 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3154 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3156 if (num_groups
> memcg_limited_groups_array_size
) {
3158 ssize_t size
= memcg_caches_array_size(num_groups
);
3160 size
*= sizeof(void *);
3161 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3163 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3164 if (!s
->memcg_params
) {
3165 s
->memcg_params
= cur_params
;
3169 s
->memcg_params
->is_root_cache
= true;
3172 * There is the chance it will be bigger than
3173 * memcg_limited_groups_array_size, if we failed an allocation
3174 * in a cache, in which case all caches updated before it, will
3175 * have a bigger array.
3177 * But if that is the case, the data after
3178 * memcg_limited_groups_array_size is certainly unused
3180 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3181 if (!cur_params
->memcg_caches
[i
])
3183 s
->memcg_params
->memcg_caches
[i
] =
3184 cur_params
->memcg_caches
[i
];
3188 * Ideally, we would wait until all caches succeed, and only
3189 * then free the old one. But this is not worth the extra
3190 * pointer per-cache we'd have to have for this.
3192 * It is not a big deal if some caches are left with a size
3193 * bigger than the others. And all updates will reset this
3201 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3202 struct kmem_cache
*root_cache
)
3206 if (!memcg_kmem_enabled())
3210 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3211 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3213 size
= sizeof(struct memcg_cache_params
);
3215 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3216 if (!s
->memcg_params
)
3220 s
->memcg_params
->memcg
= memcg
;
3221 s
->memcg_params
->root_cache
= root_cache
;
3222 INIT_WORK(&s
->memcg_params
->destroy
,
3223 kmem_cache_destroy_work_func
);
3225 s
->memcg_params
->is_root_cache
= true;
3230 void memcg_release_cache(struct kmem_cache
*s
)
3232 struct kmem_cache
*root
;
3233 struct mem_cgroup
*memcg
;
3237 * This happens, for instance, when a root cache goes away before we
3240 if (!s
->memcg_params
)
3243 if (s
->memcg_params
->is_root_cache
)
3246 memcg
= s
->memcg_params
->memcg
;
3247 id
= memcg_cache_id(memcg
);
3249 root
= s
->memcg_params
->root_cache
;
3250 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3252 mutex_lock(&memcg
->slab_caches_mutex
);
3253 list_del(&s
->memcg_params
->list
);
3254 mutex_unlock(&memcg
->slab_caches_mutex
);
3256 css_put(&memcg
->css
);
3258 kfree(s
->memcg_params
);
3262 * During the creation a new cache, we need to disable our accounting mechanism
3263 * altogether. This is true even if we are not creating, but rather just
3264 * enqueing new caches to be created.
3266 * This is because that process will trigger allocations; some visible, like
3267 * explicit kmallocs to auxiliary data structures, name strings and internal
3268 * cache structures; some well concealed, like INIT_WORK() that can allocate
3269 * objects during debug.
3271 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3272 * to it. This may not be a bounded recursion: since the first cache creation
3273 * failed to complete (waiting on the allocation), we'll just try to create the
3274 * cache again, failing at the same point.
3276 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3277 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3278 * inside the following two functions.
3280 static inline void memcg_stop_kmem_account(void)
3282 VM_BUG_ON(!current
->mm
);
3283 current
->memcg_kmem_skip_account
++;
3286 static inline void memcg_resume_kmem_account(void)
3288 VM_BUG_ON(!current
->mm
);
3289 current
->memcg_kmem_skip_account
--;
3292 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3294 struct kmem_cache
*cachep
;
3295 struct memcg_cache_params
*p
;
3297 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3299 cachep
= memcg_params_to_cache(p
);
3302 * If we get down to 0 after shrink, we could delete right away.
3303 * However, memcg_release_pages() already puts us back in the workqueue
3304 * in that case. If we proceed deleting, we'll get a dangling
3305 * reference, and removing the object from the workqueue in that case
3306 * is unnecessary complication. We are not a fast path.
3308 * Note that this case is fundamentally different from racing with
3309 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3310 * kmem_cache_shrink, not only we would be reinserting a dead cache
3311 * into the queue, but doing so from inside the worker racing to
3314 * So if we aren't down to zero, we'll just schedule a worker and try
3317 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3318 kmem_cache_shrink(cachep
);
3319 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3322 kmem_cache_destroy(cachep
);
3325 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3327 if (!cachep
->memcg_params
->dead
)
3331 * There are many ways in which we can get here.
3333 * We can get to a memory-pressure situation while the delayed work is
3334 * still pending to run. The vmscan shrinkers can then release all
3335 * cache memory and get us to destruction. If this is the case, we'll
3336 * be executed twice, which is a bug (the second time will execute over
3337 * bogus data). In this case, cancelling the work should be fine.
3339 * But we can also get here from the worker itself, if
3340 * kmem_cache_shrink is enough to shake all the remaining objects and
3341 * get the page count to 0. In this case, we'll deadlock if we try to
3342 * cancel the work (the worker runs with an internal lock held, which
3343 * is the same lock we would hold for cancel_work_sync().)
3345 * Since we can't possibly know who got us here, just refrain from
3346 * running if there is already work pending
3348 if (work_pending(&cachep
->memcg_params
->destroy
))
3351 * We have to defer the actual destroying to a workqueue, because
3352 * we might currently be in a context that cannot sleep.
3354 schedule_work(&cachep
->memcg_params
->destroy
);
3358 * This lock protects updaters, not readers. We want readers to be as fast as
3359 * they can, and they will either see NULL or a valid cache value. Our model
3360 * allow them to see NULL, in which case the root memcg will be selected.
3362 * We need this lock because multiple allocations to the same cache from a non
3363 * will span more than one worker. Only one of them can create the cache.
3365 static DEFINE_MUTEX(memcg_cache_mutex
);
3368 * Called with memcg_cache_mutex held
3370 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3371 struct kmem_cache
*s
)
3373 struct kmem_cache
*new;
3374 static char *tmp_name
= NULL
;
3376 lockdep_assert_held(&memcg_cache_mutex
);
3379 * kmem_cache_create_memcg duplicates the given name and
3380 * cgroup_name for this name requires RCU context.
3381 * This static temporary buffer is used to prevent from
3382 * pointless shortliving allocation.
3385 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3391 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3392 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3395 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3396 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3399 new->allocflags
|= __GFP_KMEMCG
;
3404 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3405 struct kmem_cache
*cachep
)
3407 struct kmem_cache
*new_cachep
;
3410 BUG_ON(!memcg_can_account_kmem(memcg
));
3412 idx
= memcg_cache_id(memcg
);
3414 mutex_lock(&memcg_cache_mutex
);
3415 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3417 css_put(&memcg
->css
);
3421 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3422 if (new_cachep
== NULL
) {
3423 new_cachep
= cachep
;
3424 css_put(&memcg
->css
);
3428 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3430 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3432 * the readers won't lock, make sure everybody sees the updated value,
3433 * so they won't put stuff in the queue again for no reason
3437 mutex_unlock(&memcg_cache_mutex
);
3441 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3443 struct kmem_cache
*c
;
3446 if (!s
->memcg_params
)
3448 if (!s
->memcg_params
->is_root_cache
)
3452 * If the cache is being destroyed, we trust that there is no one else
3453 * requesting objects from it. Even if there are, the sanity checks in
3454 * kmem_cache_destroy should caught this ill-case.
3456 * Still, we don't want anyone else freeing memcg_caches under our
3457 * noses, which can happen if a new memcg comes to life. As usual,
3458 * we'll take the set_limit_mutex to protect ourselves against this.
3460 mutex_lock(&set_limit_mutex
);
3461 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3462 c
= s
->memcg_params
->memcg_caches
[i
];
3467 * We will now manually delete the caches, so to avoid races
3468 * we need to cancel all pending destruction workers and
3469 * proceed with destruction ourselves.
3471 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3472 * and that could spawn the workers again: it is likely that
3473 * the cache still have active pages until this very moment.
3474 * This would lead us back to mem_cgroup_destroy_cache.
3476 * But that will not execute at all if the "dead" flag is not
3477 * set, so flip it down to guarantee we are in control.
3479 c
->memcg_params
->dead
= false;
3480 cancel_work_sync(&c
->memcg_params
->destroy
);
3481 kmem_cache_destroy(c
);
3483 mutex_unlock(&set_limit_mutex
);
3486 struct create_work
{
3487 struct mem_cgroup
*memcg
;
3488 struct kmem_cache
*cachep
;
3489 struct work_struct work
;
3492 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3494 struct kmem_cache
*cachep
;
3495 struct memcg_cache_params
*params
;
3497 if (!memcg_kmem_is_active(memcg
))
3500 mutex_lock(&memcg
->slab_caches_mutex
);
3501 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3502 cachep
= memcg_params_to_cache(params
);
3503 cachep
->memcg_params
->dead
= true;
3504 schedule_work(&cachep
->memcg_params
->destroy
);
3506 mutex_unlock(&memcg
->slab_caches_mutex
);
3509 static void memcg_create_cache_work_func(struct work_struct
*w
)
3511 struct create_work
*cw
;
3513 cw
= container_of(w
, struct create_work
, work
);
3514 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3519 * Enqueue the creation of a per-memcg kmem_cache.
3521 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3522 struct kmem_cache
*cachep
)
3524 struct create_work
*cw
;
3526 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3528 css_put(&memcg
->css
);
3533 cw
->cachep
= cachep
;
3535 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3536 schedule_work(&cw
->work
);
3539 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3540 struct kmem_cache
*cachep
)
3543 * We need to stop accounting when we kmalloc, because if the
3544 * corresponding kmalloc cache is not yet created, the first allocation
3545 * in __memcg_create_cache_enqueue will recurse.
3547 * However, it is better to enclose the whole function. Depending on
3548 * the debugging options enabled, INIT_WORK(), for instance, can
3549 * trigger an allocation. This too, will make us recurse. Because at
3550 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3551 * the safest choice is to do it like this, wrapping the whole function.
3553 memcg_stop_kmem_account();
3554 __memcg_create_cache_enqueue(memcg
, cachep
);
3555 memcg_resume_kmem_account();
3558 * Return the kmem_cache we're supposed to use for a slab allocation.
3559 * We try to use the current memcg's version of the cache.
3561 * If the cache does not exist yet, if we are the first user of it,
3562 * we either create it immediately, if possible, or create it asynchronously
3564 * In the latter case, we will let the current allocation go through with
3565 * the original cache.
3567 * Can't be called in interrupt context or from kernel threads.
3568 * This function needs to be called with rcu_read_lock() held.
3570 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3573 struct mem_cgroup
*memcg
;
3576 VM_BUG_ON(!cachep
->memcg_params
);
3577 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3579 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3583 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3585 if (!memcg_can_account_kmem(memcg
))
3588 idx
= memcg_cache_id(memcg
);
3591 * barrier to mare sure we're always seeing the up to date value. The
3592 * code updating memcg_caches will issue a write barrier to match this.
3594 read_barrier_depends();
3595 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3596 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3600 /* The corresponding put will be done in the workqueue. */
3601 if (!css_tryget(&memcg
->css
))
3606 * If we are in a safe context (can wait, and not in interrupt
3607 * context), we could be be predictable and return right away.
3608 * This would guarantee that the allocation being performed
3609 * already belongs in the new cache.
3611 * However, there are some clashes that can arrive from locking.
3612 * For instance, because we acquire the slab_mutex while doing
3613 * kmem_cache_dup, this means no further allocation could happen
3614 * with the slab_mutex held.
3616 * Also, because cache creation issue get_online_cpus(), this
3617 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3618 * that ends up reversed during cpu hotplug. (cpuset allocates
3619 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3620 * better to defer everything.
3622 memcg_create_cache_enqueue(memcg
, cachep
);
3628 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3631 * We need to verify if the allocation against current->mm->owner's memcg is
3632 * possible for the given order. But the page is not allocated yet, so we'll
3633 * need a further commit step to do the final arrangements.
3635 * It is possible for the task to switch cgroups in this mean time, so at
3636 * commit time, we can't rely on task conversion any longer. We'll then use
3637 * the handle argument to return to the caller which cgroup we should commit
3638 * against. We could also return the memcg directly and avoid the pointer
3639 * passing, but a boolean return value gives better semantics considering
3640 * the compiled-out case as well.
3642 * Returning true means the allocation is possible.
3645 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3647 struct mem_cgroup
*memcg
;
3653 * Disabling accounting is only relevant for some specific memcg
3654 * internal allocations. Therefore we would initially not have such
3655 * check here, since direct calls to the page allocator that are marked
3656 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3657 * concerned with cache allocations, and by having this test at
3658 * memcg_kmem_get_cache, we are already able to relay the allocation to
3659 * the root cache and bypass the memcg cache altogether.
3661 * There is one exception, though: the SLUB allocator does not create
3662 * large order caches, but rather service large kmallocs directly from
3663 * the page allocator. Therefore, the following sequence when backed by
3664 * the SLUB allocator:
3666 * memcg_stop_kmem_account();
3667 * kmalloc(<large_number>)
3668 * memcg_resume_kmem_account();
3670 * would effectively ignore the fact that we should skip accounting,
3671 * since it will drive us directly to this function without passing
3672 * through the cache selector memcg_kmem_get_cache. Such large
3673 * allocations are extremely rare but can happen, for instance, for the
3674 * cache arrays. We bring this test here.
3676 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3679 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3682 * very rare case described in mem_cgroup_from_task. Unfortunately there
3683 * isn't much we can do without complicating this too much, and it would
3684 * be gfp-dependent anyway. Just let it go
3686 if (unlikely(!memcg
))
3689 if (!memcg_can_account_kmem(memcg
)) {
3690 css_put(&memcg
->css
);
3694 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3698 css_put(&memcg
->css
);
3702 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3705 struct page_cgroup
*pc
;
3707 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3709 /* The page allocation failed. Revert */
3711 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3715 pc
= lookup_page_cgroup(page
);
3716 lock_page_cgroup(pc
);
3717 pc
->mem_cgroup
= memcg
;
3718 SetPageCgroupUsed(pc
);
3719 unlock_page_cgroup(pc
);
3722 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3724 struct mem_cgroup
*memcg
= NULL
;
3725 struct page_cgroup
*pc
;
3728 pc
= lookup_page_cgroup(page
);
3730 * Fast unlocked return. Theoretically might have changed, have to
3731 * check again after locking.
3733 if (!PageCgroupUsed(pc
))
3736 lock_page_cgroup(pc
);
3737 if (PageCgroupUsed(pc
)) {
3738 memcg
= pc
->mem_cgroup
;
3739 ClearPageCgroupUsed(pc
);
3741 unlock_page_cgroup(pc
);
3744 * We trust that only if there is a memcg associated with the page, it
3745 * is a valid allocation
3750 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3751 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3754 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3757 #endif /* CONFIG_MEMCG_KMEM */
3759 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3761 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3763 * Because tail pages are not marked as "used", set it. We're under
3764 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3765 * charge/uncharge will be never happen and move_account() is done under
3766 * compound_lock(), so we don't have to take care of races.
3768 void mem_cgroup_split_huge_fixup(struct page
*head
)
3770 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3771 struct page_cgroup
*pc
;
3772 struct mem_cgroup
*memcg
;
3775 if (mem_cgroup_disabled())
3778 memcg
= head_pc
->mem_cgroup
;
3779 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3781 pc
->mem_cgroup
= memcg
;
3782 smp_wmb();/* see __commit_charge() */
3783 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3785 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3788 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3791 void mem_cgroup_move_account_page_stat(struct mem_cgroup
*from
,
3792 struct mem_cgroup
*to
,
3793 unsigned int nr_pages
,
3794 enum mem_cgroup_stat_index idx
)
3796 /* Update stat data for mem_cgroup */
3798 __this_cpu_sub(from
->stat
->count
[idx
], nr_pages
);
3799 __this_cpu_add(to
->stat
->count
[idx
], nr_pages
);
3804 * mem_cgroup_move_account - move account of the page
3806 * @nr_pages: number of regular pages (>1 for huge pages)
3807 * @pc: page_cgroup of the page.
3808 * @from: mem_cgroup which the page is moved from.
3809 * @to: mem_cgroup which the page is moved to. @from != @to.
3811 * The caller must confirm following.
3812 * - page is not on LRU (isolate_page() is useful.)
3813 * - compound_lock is held when nr_pages > 1
3815 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3818 static int mem_cgroup_move_account(struct page
*page
,
3819 unsigned int nr_pages
,
3820 struct page_cgroup
*pc
,
3821 struct mem_cgroup
*from
,
3822 struct mem_cgroup
*to
)
3824 unsigned long flags
;
3826 bool anon
= PageAnon(page
);
3828 VM_BUG_ON(from
== to
);
3829 VM_BUG_ON(PageLRU(page
));
3831 * The page is isolated from LRU. So, collapse function
3832 * will not handle this page. But page splitting can happen.
3833 * Do this check under compound_page_lock(). The caller should
3837 if (nr_pages
> 1 && !PageTransHuge(page
))
3840 lock_page_cgroup(pc
);
3843 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3846 move_lock_mem_cgroup(from
, &flags
);
3848 if (!anon
&& page_mapped(page
))
3849 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3850 MEM_CGROUP_STAT_FILE_MAPPED
);
3852 if (PageWriteback(page
))
3853 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3854 MEM_CGROUP_STAT_WRITEBACK
);
3856 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3858 /* caller should have done css_get */
3859 pc
->mem_cgroup
= to
;
3860 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3861 move_unlock_mem_cgroup(from
, &flags
);
3864 unlock_page_cgroup(pc
);
3868 memcg_check_events(to
, page
);
3869 memcg_check_events(from
, page
);
3875 * mem_cgroup_move_parent - moves page to the parent group
3876 * @page: the page to move
3877 * @pc: page_cgroup of the page
3878 * @child: page's cgroup
3880 * move charges to its parent or the root cgroup if the group has no
3881 * parent (aka use_hierarchy==0).
3882 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3883 * mem_cgroup_move_account fails) the failure is always temporary and
3884 * it signals a race with a page removal/uncharge or migration. In the
3885 * first case the page is on the way out and it will vanish from the LRU
3886 * on the next attempt and the call should be retried later.
3887 * Isolation from the LRU fails only if page has been isolated from
3888 * the LRU since we looked at it and that usually means either global
3889 * reclaim or migration going on. The page will either get back to the
3891 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3892 * (!PageCgroupUsed) or moved to a different group. The page will
3893 * disappear in the next attempt.
3895 static int mem_cgroup_move_parent(struct page
*page
,
3896 struct page_cgroup
*pc
,
3897 struct mem_cgroup
*child
)
3899 struct mem_cgroup
*parent
;
3900 unsigned int nr_pages
;
3901 unsigned long uninitialized_var(flags
);
3904 VM_BUG_ON(mem_cgroup_is_root(child
));
3907 if (!get_page_unless_zero(page
))
3909 if (isolate_lru_page(page
))
3912 nr_pages
= hpage_nr_pages(page
);
3914 parent
= parent_mem_cgroup(child
);
3916 * If no parent, move charges to root cgroup.
3919 parent
= root_mem_cgroup
;
3922 VM_BUG_ON(!PageTransHuge(page
));
3923 flags
= compound_lock_irqsave(page
);
3926 ret
= mem_cgroup_move_account(page
, nr_pages
,
3929 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3932 compound_unlock_irqrestore(page
, flags
);
3933 putback_lru_page(page
);
3941 * Charge the memory controller for page usage.
3943 * 0 if the charge was successful
3944 * < 0 if the cgroup is over its limit
3946 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3947 gfp_t gfp_mask
, enum charge_type ctype
)
3949 struct mem_cgroup
*memcg
= NULL
;
3950 unsigned int nr_pages
= 1;
3954 if (PageTransHuge(page
)) {
3955 nr_pages
<<= compound_order(page
);
3956 VM_BUG_ON(!PageTransHuge(page
));
3958 * Never OOM-kill a process for a huge page. The
3959 * fault handler will fall back to regular pages.
3964 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3967 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3971 int mem_cgroup_newpage_charge(struct page
*page
,
3972 struct mm_struct
*mm
, gfp_t gfp_mask
)
3974 if (mem_cgroup_disabled())
3976 VM_BUG_ON(page_mapped(page
));
3977 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3979 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3980 MEM_CGROUP_CHARGE_TYPE_ANON
);
3984 * While swap-in, try_charge -> commit or cancel, the page is locked.
3985 * And when try_charge() successfully returns, one refcnt to memcg without
3986 * struct page_cgroup is acquired. This refcnt will be consumed by
3987 * "commit()" or removed by "cancel()"
3989 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3992 struct mem_cgroup
**memcgp
)
3994 struct mem_cgroup
*memcg
;
3995 struct page_cgroup
*pc
;
3998 pc
= lookup_page_cgroup(page
);
4000 * Every swap fault against a single page tries to charge the
4001 * page, bail as early as possible. shmem_unuse() encounters
4002 * already charged pages, too. The USED bit is protected by
4003 * the page lock, which serializes swap cache removal, which
4004 * in turn serializes uncharging.
4006 if (PageCgroupUsed(pc
))
4008 if (!do_swap_account
)
4010 memcg
= try_get_mem_cgroup_from_page(page
);
4014 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
4015 css_put(&memcg
->css
);
4020 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
4026 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
4027 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
4030 if (mem_cgroup_disabled())
4033 * A racing thread's fault, or swapoff, may have already
4034 * updated the pte, and even removed page from swap cache: in
4035 * those cases unuse_pte()'s pte_same() test will fail; but
4036 * there's also a KSM case which does need to charge the page.
4038 if (!PageSwapCache(page
)) {
4041 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4046 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4049 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4051 if (mem_cgroup_disabled())
4055 __mem_cgroup_cancel_charge(memcg
, 1);
4059 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4060 enum charge_type ctype
)
4062 if (mem_cgroup_disabled())
4067 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4069 * Now swap is on-memory. This means this page may be
4070 * counted both as mem and swap....double count.
4071 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4072 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4073 * may call delete_from_swap_cache() before reach here.
4075 if (do_swap_account
&& PageSwapCache(page
)) {
4076 swp_entry_t ent
= {.val
= page_private(page
)};
4077 mem_cgroup_uncharge_swap(ent
);
4081 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4082 struct mem_cgroup
*memcg
)
4084 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4085 MEM_CGROUP_CHARGE_TYPE_ANON
);
4088 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4091 struct mem_cgroup
*memcg
= NULL
;
4092 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4095 if (mem_cgroup_disabled())
4097 if (PageCompound(page
))
4100 if (!PageSwapCache(page
))
4101 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4102 else { /* page is swapcache/shmem */
4103 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4106 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4111 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4112 unsigned int nr_pages
,
4113 const enum charge_type ctype
)
4115 struct memcg_batch_info
*batch
= NULL
;
4116 bool uncharge_memsw
= true;
4118 /* If swapout, usage of swap doesn't decrease */
4119 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4120 uncharge_memsw
= false;
4122 batch
= ¤t
->memcg_batch
;
4124 * In usual, we do css_get() when we remember memcg pointer.
4125 * But in this case, we keep res->usage until end of a series of
4126 * uncharges. Then, it's ok to ignore memcg's refcnt.
4129 batch
->memcg
= memcg
;
4131 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4132 * In those cases, all pages freed continuously can be expected to be in
4133 * the same cgroup and we have chance to coalesce uncharges.
4134 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4135 * because we want to do uncharge as soon as possible.
4138 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4139 goto direct_uncharge
;
4142 goto direct_uncharge
;
4145 * In typical case, batch->memcg == mem. This means we can
4146 * merge a series of uncharges to an uncharge of res_counter.
4147 * If not, we uncharge res_counter ony by one.
4149 if (batch
->memcg
!= memcg
)
4150 goto direct_uncharge
;
4151 /* remember freed charge and uncharge it later */
4154 batch
->memsw_nr_pages
++;
4157 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4159 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4160 if (unlikely(batch
->memcg
!= memcg
))
4161 memcg_oom_recover(memcg
);
4165 * uncharge if !page_mapped(page)
4167 static struct mem_cgroup
*
4168 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4171 struct mem_cgroup
*memcg
= NULL
;
4172 unsigned int nr_pages
= 1;
4173 struct page_cgroup
*pc
;
4176 if (mem_cgroup_disabled())
4179 if (PageTransHuge(page
)) {
4180 nr_pages
<<= compound_order(page
);
4181 VM_BUG_ON(!PageTransHuge(page
));
4184 * Check if our page_cgroup is valid
4186 pc
= lookup_page_cgroup(page
);
4187 if (unlikely(!PageCgroupUsed(pc
)))
4190 lock_page_cgroup(pc
);
4192 memcg
= pc
->mem_cgroup
;
4194 if (!PageCgroupUsed(pc
))
4197 anon
= PageAnon(page
);
4200 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4202 * Generally PageAnon tells if it's the anon statistics to be
4203 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4204 * used before page reached the stage of being marked PageAnon.
4208 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4209 /* See mem_cgroup_prepare_migration() */
4210 if (page_mapped(page
))
4213 * Pages under migration may not be uncharged. But
4214 * end_migration() /must/ be the one uncharging the
4215 * unused post-migration page and so it has to call
4216 * here with the migration bit still set. See the
4217 * res_counter handling below.
4219 if (!end_migration
&& PageCgroupMigration(pc
))
4222 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4223 if (!PageAnon(page
)) { /* Shared memory */
4224 if (page
->mapping
&& !page_is_file_cache(page
))
4226 } else if (page_mapped(page
)) /* Anon */
4233 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4235 ClearPageCgroupUsed(pc
);
4237 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4238 * freed from LRU. This is safe because uncharged page is expected not
4239 * to be reused (freed soon). Exception is SwapCache, it's handled by
4240 * special functions.
4243 unlock_page_cgroup(pc
);
4245 * even after unlock, we have memcg->res.usage here and this memcg
4246 * will never be freed, so it's safe to call css_get().
4248 memcg_check_events(memcg
, page
);
4249 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4250 mem_cgroup_swap_statistics(memcg
, true);
4251 css_get(&memcg
->css
);
4254 * Migration does not charge the res_counter for the
4255 * replacement page, so leave it alone when phasing out the
4256 * page that is unused after the migration.
4258 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4259 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4264 unlock_page_cgroup(pc
);
4268 void mem_cgroup_uncharge_page(struct page
*page
)
4271 if (page_mapped(page
))
4273 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4275 * If the page is in swap cache, uncharge should be deferred
4276 * to the swap path, which also properly accounts swap usage
4277 * and handles memcg lifetime.
4279 * Note that this check is not stable and reclaim may add the
4280 * page to swap cache at any time after this. However, if the
4281 * page is not in swap cache by the time page->mapcount hits
4282 * 0, there won't be any page table references to the swap
4283 * slot, and reclaim will free it and not actually write the
4286 if (PageSwapCache(page
))
4288 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4291 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4293 VM_BUG_ON(page_mapped(page
));
4294 VM_BUG_ON(page
->mapping
);
4295 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4299 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4300 * In that cases, pages are freed continuously and we can expect pages
4301 * are in the same memcg. All these calls itself limits the number of
4302 * pages freed at once, then uncharge_start/end() is called properly.
4303 * This may be called prural(2) times in a context,
4306 void mem_cgroup_uncharge_start(void)
4308 current
->memcg_batch
.do_batch
++;
4309 /* We can do nest. */
4310 if (current
->memcg_batch
.do_batch
== 1) {
4311 current
->memcg_batch
.memcg
= NULL
;
4312 current
->memcg_batch
.nr_pages
= 0;
4313 current
->memcg_batch
.memsw_nr_pages
= 0;
4317 void mem_cgroup_uncharge_end(void)
4319 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4321 if (!batch
->do_batch
)
4325 if (batch
->do_batch
) /* If stacked, do nothing. */
4331 * This "batch->memcg" is valid without any css_get/put etc...
4332 * bacause we hide charges behind us.
4334 if (batch
->nr_pages
)
4335 res_counter_uncharge(&batch
->memcg
->res
,
4336 batch
->nr_pages
* PAGE_SIZE
);
4337 if (batch
->memsw_nr_pages
)
4338 res_counter_uncharge(&batch
->memcg
->memsw
,
4339 batch
->memsw_nr_pages
* PAGE_SIZE
);
4340 memcg_oom_recover(batch
->memcg
);
4341 /* forget this pointer (for sanity check) */
4342 batch
->memcg
= NULL
;
4347 * called after __delete_from_swap_cache() and drop "page" account.
4348 * memcg information is recorded to swap_cgroup of "ent"
4351 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4353 struct mem_cgroup
*memcg
;
4354 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4356 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4357 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4359 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4362 * record memcg information, if swapout && memcg != NULL,
4363 * css_get() was called in uncharge().
4365 if (do_swap_account
&& swapout
&& memcg
)
4366 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4370 #ifdef CONFIG_MEMCG_SWAP
4372 * called from swap_entry_free(). remove record in swap_cgroup and
4373 * uncharge "memsw" account.
4375 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4377 struct mem_cgroup
*memcg
;
4380 if (!do_swap_account
)
4383 id
= swap_cgroup_record(ent
, 0);
4385 memcg
= mem_cgroup_lookup(id
);
4388 * We uncharge this because swap is freed.
4389 * This memcg can be obsolete one. We avoid calling css_tryget
4391 if (!mem_cgroup_is_root(memcg
))
4392 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4393 mem_cgroup_swap_statistics(memcg
, false);
4394 css_put(&memcg
->css
);
4400 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4401 * @entry: swap entry to be moved
4402 * @from: mem_cgroup which the entry is moved from
4403 * @to: mem_cgroup which the entry is moved to
4405 * It succeeds only when the swap_cgroup's record for this entry is the same
4406 * as the mem_cgroup's id of @from.
4408 * Returns 0 on success, -EINVAL on failure.
4410 * The caller must have charged to @to, IOW, called res_counter_charge() about
4411 * both res and memsw, and called css_get().
4413 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4414 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4416 unsigned short old_id
, new_id
;
4418 old_id
= mem_cgroup_id(from
);
4419 new_id
= mem_cgroup_id(to
);
4421 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4422 mem_cgroup_swap_statistics(from
, false);
4423 mem_cgroup_swap_statistics(to
, true);
4425 * This function is only called from task migration context now.
4426 * It postpones res_counter and refcount handling till the end
4427 * of task migration(mem_cgroup_clear_mc()) for performance
4428 * improvement. But we cannot postpone css_get(to) because if
4429 * the process that has been moved to @to does swap-in, the
4430 * refcount of @to might be decreased to 0.
4432 * We are in attach() phase, so the cgroup is guaranteed to be
4433 * alive, so we can just call css_get().
4441 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4442 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4449 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4452 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4453 struct mem_cgroup
**memcgp
)
4455 struct mem_cgroup
*memcg
= NULL
;
4456 unsigned int nr_pages
= 1;
4457 struct page_cgroup
*pc
;
4458 enum charge_type ctype
;
4462 if (mem_cgroup_disabled())
4465 if (PageTransHuge(page
))
4466 nr_pages
<<= compound_order(page
);
4468 pc
= lookup_page_cgroup(page
);
4469 lock_page_cgroup(pc
);
4470 if (PageCgroupUsed(pc
)) {
4471 memcg
= pc
->mem_cgroup
;
4472 css_get(&memcg
->css
);
4474 * At migrating an anonymous page, its mapcount goes down
4475 * to 0 and uncharge() will be called. But, even if it's fully
4476 * unmapped, migration may fail and this page has to be
4477 * charged again. We set MIGRATION flag here and delay uncharge
4478 * until end_migration() is called
4480 * Corner Case Thinking
4482 * When the old page was mapped as Anon and it's unmap-and-freed
4483 * while migration was ongoing.
4484 * If unmap finds the old page, uncharge() of it will be delayed
4485 * until end_migration(). If unmap finds a new page, it's
4486 * uncharged when it make mapcount to be 1->0. If unmap code
4487 * finds swap_migration_entry, the new page will not be mapped
4488 * and end_migration() will find it(mapcount==0).
4491 * When the old page was mapped but migraion fails, the kernel
4492 * remaps it. A charge for it is kept by MIGRATION flag even
4493 * if mapcount goes down to 0. We can do remap successfully
4494 * without charging it again.
4497 * The "old" page is under lock_page() until the end of
4498 * migration, so, the old page itself will not be swapped-out.
4499 * If the new page is swapped out before end_migraton, our
4500 * hook to usual swap-out path will catch the event.
4503 SetPageCgroupMigration(pc
);
4505 unlock_page_cgroup(pc
);
4507 * If the page is not charged at this point,
4515 * We charge new page before it's used/mapped. So, even if unlock_page()
4516 * is called before end_migration, we can catch all events on this new
4517 * page. In the case new page is migrated but not remapped, new page's
4518 * mapcount will be finally 0 and we call uncharge in end_migration().
4521 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4523 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4525 * The page is committed to the memcg, but it's not actually
4526 * charged to the res_counter since we plan on replacing the
4527 * old one and only one page is going to be left afterwards.
4529 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4532 /* remove redundant charge if migration failed*/
4533 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4534 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4536 struct page
*used
, *unused
;
4537 struct page_cgroup
*pc
;
4543 if (!migration_ok
) {
4550 anon
= PageAnon(used
);
4551 __mem_cgroup_uncharge_common(unused
,
4552 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4553 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4555 css_put(&memcg
->css
);
4557 * We disallowed uncharge of pages under migration because mapcount
4558 * of the page goes down to zero, temporarly.
4559 * Clear the flag and check the page should be charged.
4561 pc
= lookup_page_cgroup(oldpage
);
4562 lock_page_cgroup(pc
);
4563 ClearPageCgroupMigration(pc
);
4564 unlock_page_cgroup(pc
);
4567 * If a page is a file cache, radix-tree replacement is very atomic
4568 * and we can skip this check. When it was an Anon page, its mapcount
4569 * goes down to 0. But because we added MIGRATION flage, it's not
4570 * uncharged yet. There are several case but page->mapcount check
4571 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4572 * check. (see prepare_charge() also)
4575 mem_cgroup_uncharge_page(used
);
4579 * At replace page cache, newpage is not under any memcg but it's on
4580 * LRU. So, this function doesn't touch res_counter but handles LRU
4581 * in correct way. Both pages are locked so we cannot race with uncharge.
4583 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4584 struct page
*newpage
)
4586 struct mem_cgroup
*memcg
= NULL
;
4587 struct page_cgroup
*pc
;
4588 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4590 if (mem_cgroup_disabled())
4593 pc
= lookup_page_cgroup(oldpage
);
4594 /* fix accounting on old pages */
4595 lock_page_cgroup(pc
);
4596 if (PageCgroupUsed(pc
)) {
4597 memcg
= pc
->mem_cgroup
;
4598 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4599 ClearPageCgroupUsed(pc
);
4601 unlock_page_cgroup(pc
);
4604 * When called from shmem_replace_page(), in some cases the
4605 * oldpage has already been charged, and in some cases not.
4610 * Even if newpage->mapping was NULL before starting replacement,
4611 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4612 * LRU while we overwrite pc->mem_cgroup.
4614 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4617 #ifdef CONFIG_DEBUG_VM
4618 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4620 struct page_cgroup
*pc
;
4622 pc
= lookup_page_cgroup(page
);
4624 * Can be NULL while feeding pages into the page allocator for
4625 * the first time, i.e. during boot or memory hotplug;
4626 * or when mem_cgroup_disabled().
4628 if (likely(pc
) && PageCgroupUsed(pc
))
4633 bool mem_cgroup_bad_page_check(struct page
*page
)
4635 if (mem_cgroup_disabled())
4638 return lookup_page_cgroup_used(page
) != NULL
;
4641 void mem_cgroup_print_bad_page(struct page
*page
)
4643 struct page_cgroup
*pc
;
4645 pc
= lookup_page_cgroup_used(page
);
4647 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4648 pc
, pc
->flags
, pc
->mem_cgroup
);
4653 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4654 unsigned long long val
)
4657 u64 memswlimit
, memlimit
;
4659 int children
= mem_cgroup_count_children(memcg
);
4660 u64 curusage
, oldusage
;
4664 * For keeping hierarchical_reclaim simple, how long we should retry
4665 * is depends on callers. We set our retry-count to be function
4666 * of # of children which we should visit in this loop.
4668 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4670 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4673 while (retry_count
) {
4674 if (signal_pending(current
)) {
4679 * Rather than hide all in some function, I do this in
4680 * open coded manner. You see what this really does.
4681 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4683 mutex_lock(&set_limit_mutex
);
4684 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4685 if (memswlimit
< val
) {
4687 mutex_unlock(&set_limit_mutex
);
4691 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4695 ret
= res_counter_set_limit(&memcg
->res
, val
);
4697 if (memswlimit
== val
)
4698 memcg
->memsw_is_minimum
= true;
4700 memcg
->memsw_is_minimum
= false;
4702 mutex_unlock(&set_limit_mutex
);
4707 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4708 MEM_CGROUP_RECLAIM_SHRINK
);
4709 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4710 /* Usage is reduced ? */
4711 if (curusage
>= oldusage
)
4714 oldusage
= curusage
;
4716 if (!ret
&& enlarge
)
4717 memcg_oom_recover(memcg
);
4722 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4723 unsigned long long val
)
4726 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4727 int children
= mem_cgroup_count_children(memcg
);
4731 /* see mem_cgroup_resize_res_limit */
4732 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4733 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4734 while (retry_count
) {
4735 if (signal_pending(current
)) {
4740 * Rather than hide all in some function, I do this in
4741 * open coded manner. You see what this really does.
4742 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4744 mutex_lock(&set_limit_mutex
);
4745 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4746 if (memlimit
> val
) {
4748 mutex_unlock(&set_limit_mutex
);
4751 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4752 if (memswlimit
< val
)
4754 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4756 if (memlimit
== val
)
4757 memcg
->memsw_is_minimum
= true;
4759 memcg
->memsw_is_minimum
= false;
4761 mutex_unlock(&set_limit_mutex
);
4766 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4767 MEM_CGROUP_RECLAIM_NOSWAP
|
4768 MEM_CGROUP_RECLAIM_SHRINK
);
4769 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4770 /* Usage is reduced ? */
4771 if (curusage
>= oldusage
)
4774 oldusage
= curusage
;
4776 if (!ret
&& enlarge
)
4777 memcg_oom_recover(memcg
);
4781 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4783 unsigned long *total_scanned
)
4785 unsigned long nr_reclaimed
= 0;
4786 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4787 unsigned long reclaimed
;
4789 struct mem_cgroup_tree_per_zone
*mctz
;
4790 unsigned long long excess
;
4791 unsigned long nr_scanned
;
4796 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4798 * This loop can run a while, specially if mem_cgroup's continuously
4799 * keep exceeding their soft limit and putting the system under
4806 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4811 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4812 gfp_mask
, &nr_scanned
);
4813 nr_reclaimed
+= reclaimed
;
4814 *total_scanned
+= nr_scanned
;
4815 spin_lock(&mctz
->lock
);
4818 * If we failed to reclaim anything from this memory cgroup
4819 * it is time to move on to the next cgroup
4825 * Loop until we find yet another one.
4827 * By the time we get the soft_limit lock
4828 * again, someone might have aded the
4829 * group back on the RB tree. Iterate to
4830 * make sure we get a different mem.
4831 * mem_cgroup_largest_soft_limit_node returns
4832 * NULL if no other cgroup is present on
4836 __mem_cgroup_largest_soft_limit_node(mctz
);
4838 css_put(&next_mz
->memcg
->css
);
4839 else /* next_mz == NULL or other memcg */
4843 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4844 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4846 * One school of thought says that we should not add
4847 * back the node to the tree if reclaim returns 0.
4848 * But our reclaim could return 0, simply because due
4849 * to priority we are exposing a smaller subset of
4850 * memory to reclaim from. Consider this as a longer
4853 /* If excess == 0, no tree ops */
4854 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4855 spin_unlock(&mctz
->lock
);
4856 css_put(&mz
->memcg
->css
);
4859 * Could not reclaim anything and there are no more
4860 * mem cgroups to try or we seem to be looping without
4861 * reclaiming anything.
4863 if (!nr_reclaimed
&&
4865 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4867 } while (!nr_reclaimed
);
4869 css_put(&next_mz
->memcg
->css
);
4870 return nr_reclaimed
;
4874 * mem_cgroup_force_empty_list - clears LRU of a group
4875 * @memcg: group to clear
4878 * @lru: lru to to clear
4880 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4881 * reclaim the pages page themselves - pages are moved to the parent (or root)
4884 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4885 int node
, int zid
, enum lru_list lru
)
4887 struct lruvec
*lruvec
;
4888 unsigned long flags
;
4889 struct list_head
*list
;
4893 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4894 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4895 list
= &lruvec
->lists
[lru
];
4899 struct page_cgroup
*pc
;
4902 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4903 if (list_empty(list
)) {
4904 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4907 page
= list_entry(list
->prev
, struct page
, lru
);
4909 list_move(&page
->lru
, list
);
4911 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4914 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4916 pc
= lookup_page_cgroup(page
);
4918 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4919 /* found lock contention or "pc" is obsolete. */
4924 } while (!list_empty(list
));
4928 * make mem_cgroup's charge to be 0 if there is no task by moving
4929 * all the charges and pages to the parent.
4930 * This enables deleting this mem_cgroup.
4932 * Caller is responsible for holding css reference on the memcg.
4934 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4940 /* This is for making all *used* pages to be on LRU. */
4941 lru_add_drain_all();
4942 drain_all_stock_sync(memcg
);
4943 mem_cgroup_start_move(memcg
);
4944 for_each_node_state(node
, N_MEMORY
) {
4945 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4948 mem_cgroup_force_empty_list(memcg
,
4953 mem_cgroup_end_move(memcg
);
4954 memcg_oom_recover(memcg
);
4958 * Kernel memory may not necessarily be trackable to a specific
4959 * process. So they are not migrated, and therefore we can't
4960 * expect their value to drop to 0 here.
4961 * Having res filled up with kmem only is enough.
4963 * This is a safety check because mem_cgroup_force_empty_list
4964 * could have raced with mem_cgroup_replace_page_cache callers
4965 * so the lru seemed empty but the page could have been added
4966 * right after the check. RES_USAGE should be safe as we always
4967 * charge before adding to the LRU.
4969 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4970 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4971 } while (usage
> 0);
4974 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4976 lockdep_assert_held(&memcg_create_mutex
);
4978 * The lock does not prevent addition or deletion to the list
4979 * of children, but it prevents a new child from being
4980 * initialized based on this parent in css_online(), so it's
4981 * enough to decide whether hierarchically inherited
4982 * attributes can still be changed or not.
4984 return memcg
->use_hierarchy
&&
4985 !list_empty(&memcg
->css
.cgroup
->children
);
4989 * Reclaims as many pages from the given memcg as possible and moves
4990 * the rest to the parent.
4992 * Caller is responsible for holding css reference for memcg.
4994 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4996 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4997 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4999 /* returns EBUSY if there is a task or if we come here twice. */
5000 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
5003 /* we call try-to-free pages for make this cgroup empty */
5004 lru_add_drain_all();
5005 /* try to free all pages in this cgroup */
5006 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
5009 if (signal_pending(current
))
5012 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
5016 /* maybe some writeback is necessary */
5017 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
5022 mem_cgroup_reparent_charges(memcg
);
5027 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5030 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5032 if (mem_cgroup_is_root(memcg
))
5034 return mem_cgroup_force_empty(memcg
);
5037 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5040 return mem_cgroup_from_css(css
)->use_hierarchy
;
5043 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5044 struct cftype
*cft
, u64 val
)
5047 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5048 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5050 mutex_lock(&memcg_create_mutex
);
5052 if (memcg
->use_hierarchy
== val
)
5056 * If parent's use_hierarchy is set, we can't make any modifications
5057 * in the child subtrees. If it is unset, then the change can
5058 * occur, provided the current cgroup has no children.
5060 * For the root cgroup, parent_mem is NULL, we allow value to be
5061 * set if there are no children.
5063 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5064 (val
== 1 || val
== 0)) {
5065 if (list_empty(&memcg
->css
.cgroup
->children
))
5066 memcg
->use_hierarchy
= val
;
5073 mutex_unlock(&memcg_create_mutex
);
5079 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5080 enum mem_cgroup_stat_index idx
)
5082 struct mem_cgroup
*iter
;
5085 /* Per-cpu values can be negative, use a signed accumulator */
5086 for_each_mem_cgroup_tree(iter
, memcg
)
5087 val
+= mem_cgroup_read_stat(iter
, idx
);
5089 if (val
< 0) /* race ? */
5094 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5098 if (!mem_cgroup_is_root(memcg
)) {
5100 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5102 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5106 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5107 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5109 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5110 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5113 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5115 return val
<< PAGE_SHIFT
;
5118 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
5119 struct cftype
*cft
, struct file
*file
,
5120 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
5122 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5128 type
= MEMFILE_TYPE(cft
->private);
5129 name
= MEMFILE_ATTR(cft
->private);
5133 if (name
== RES_USAGE
)
5134 val
= mem_cgroup_usage(memcg
, false);
5136 val
= res_counter_read_u64(&memcg
->res
, name
);
5139 if (name
== RES_USAGE
)
5140 val
= mem_cgroup_usage(memcg
, true);
5142 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5145 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5151 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5152 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5155 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5158 #ifdef CONFIG_MEMCG_KMEM
5159 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5161 * For simplicity, we won't allow this to be disabled. It also can't
5162 * be changed if the cgroup has children already, or if tasks had
5165 * If tasks join before we set the limit, a person looking at
5166 * kmem.usage_in_bytes will have no way to determine when it took
5167 * place, which makes the value quite meaningless.
5169 * After it first became limited, changes in the value of the limit are
5170 * of course permitted.
5172 mutex_lock(&memcg_create_mutex
);
5173 mutex_lock(&set_limit_mutex
);
5174 if (!memcg
->kmem_account_flags
&& val
!= RES_COUNTER_MAX
) {
5175 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5179 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5182 ret
= memcg_update_cache_sizes(memcg
);
5184 res_counter_set_limit(&memcg
->kmem
, RES_COUNTER_MAX
);
5187 static_key_slow_inc(&memcg_kmem_enabled_key
);
5189 * setting the active bit after the inc will guarantee no one
5190 * starts accounting before all call sites are patched
5192 memcg_kmem_set_active(memcg
);
5194 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5196 mutex_unlock(&set_limit_mutex
);
5197 mutex_unlock(&memcg_create_mutex
);
5202 #ifdef CONFIG_MEMCG_KMEM
5203 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5206 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5210 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5212 * When that happen, we need to disable the static branch only on those
5213 * memcgs that enabled it. To achieve this, we would be forced to
5214 * complicate the code by keeping track of which memcgs were the ones
5215 * that actually enabled limits, and which ones got it from its
5218 * It is a lot simpler just to do static_key_slow_inc() on every child
5219 * that is accounted.
5221 if (!memcg_kmem_is_active(memcg
))
5225 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5226 * memcg is active already. If the later initialization fails then the
5227 * cgroup core triggers the cleanup so we do not have to do it here.
5229 static_key_slow_inc(&memcg_kmem_enabled_key
);
5231 mutex_lock(&set_limit_mutex
);
5232 memcg_stop_kmem_account();
5233 ret
= memcg_update_cache_sizes(memcg
);
5234 memcg_resume_kmem_account();
5235 mutex_unlock(&set_limit_mutex
);
5239 #endif /* CONFIG_MEMCG_KMEM */
5242 * The user of this function is...
5245 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5248 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5251 unsigned long long val
;
5254 type
= MEMFILE_TYPE(cft
->private);
5255 name
= MEMFILE_ATTR(cft
->private);
5259 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5263 /* This function does all necessary parse...reuse it */
5264 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5268 ret
= mem_cgroup_resize_limit(memcg
, val
);
5269 else if (type
== _MEMSWAP
)
5270 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5271 else if (type
== _KMEM
)
5272 ret
= memcg_update_kmem_limit(css
, val
);
5276 case RES_SOFT_LIMIT
:
5277 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5281 * For memsw, soft limits are hard to implement in terms
5282 * of semantics, for now, we support soft limits for
5283 * control without swap
5286 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5291 ret
= -EINVAL
; /* should be BUG() ? */
5297 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5298 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5300 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5302 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5303 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5304 if (!memcg
->use_hierarchy
)
5307 while (css_parent(&memcg
->css
)) {
5308 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5309 if (!memcg
->use_hierarchy
)
5311 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5312 min_limit
= min(min_limit
, tmp
);
5313 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5314 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5317 *mem_limit
= min_limit
;
5318 *memsw_limit
= min_memsw_limit
;
5321 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5323 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5327 type
= MEMFILE_TYPE(event
);
5328 name
= MEMFILE_ATTR(event
);
5333 res_counter_reset_max(&memcg
->res
);
5334 else if (type
== _MEMSWAP
)
5335 res_counter_reset_max(&memcg
->memsw
);
5336 else if (type
== _KMEM
)
5337 res_counter_reset_max(&memcg
->kmem
);
5343 res_counter_reset_failcnt(&memcg
->res
);
5344 else if (type
== _MEMSWAP
)
5345 res_counter_reset_failcnt(&memcg
->memsw
);
5346 else if (type
== _KMEM
)
5347 res_counter_reset_failcnt(&memcg
->kmem
);
5356 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5359 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5363 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5364 struct cftype
*cft
, u64 val
)
5366 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5368 if (val
>= (1 << NR_MOVE_TYPE
))
5372 * No kind of locking is needed in here, because ->can_attach() will
5373 * check this value once in the beginning of the process, and then carry
5374 * on with stale data. This means that changes to this value will only
5375 * affect task migrations starting after the change.
5377 memcg
->move_charge_at_immigrate
= val
;
5381 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5382 struct cftype
*cft
, u64 val
)
5389 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5390 struct cftype
*cft
, struct seq_file
*m
)
5393 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5394 unsigned long node_nr
;
5395 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5397 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5398 seq_printf(m
, "total=%lu", total_nr
);
5399 for_each_node_state(nid
, N_MEMORY
) {
5400 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5401 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5405 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5406 seq_printf(m
, "file=%lu", file_nr
);
5407 for_each_node_state(nid
, N_MEMORY
) {
5408 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5410 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5414 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5415 seq_printf(m
, "anon=%lu", anon_nr
);
5416 for_each_node_state(nid
, N_MEMORY
) {
5417 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5419 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5423 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5424 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5425 for_each_node_state(nid
, N_MEMORY
) {
5426 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5427 BIT(LRU_UNEVICTABLE
));
5428 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5433 #endif /* CONFIG_NUMA */
5435 static inline void mem_cgroup_lru_names_not_uptodate(void)
5437 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5440 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5443 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5444 struct mem_cgroup
*mi
;
5447 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5448 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5450 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5451 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5454 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5455 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5456 mem_cgroup_read_events(memcg
, i
));
5458 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5459 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5460 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5462 /* Hierarchical information */
5464 unsigned long long limit
, memsw_limit
;
5465 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5466 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5467 if (do_swap_account
)
5468 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5472 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5475 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5477 for_each_mem_cgroup_tree(mi
, memcg
)
5478 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5479 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5482 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5483 unsigned long long val
= 0;
5485 for_each_mem_cgroup_tree(mi
, memcg
)
5486 val
+= mem_cgroup_read_events(mi
, i
);
5487 seq_printf(m
, "total_%s %llu\n",
5488 mem_cgroup_events_names
[i
], val
);
5491 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5492 unsigned long long val
= 0;
5494 for_each_mem_cgroup_tree(mi
, memcg
)
5495 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5496 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5499 #ifdef CONFIG_DEBUG_VM
5502 struct mem_cgroup_per_zone
*mz
;
5503 struct zone_reclaim_stat
*rstat
;
5504 unsigned long recent_rotated
[2] = {0, 0};
5505 unsigned long recent_scanned
[2] = {0, 0};
5507 for_each_online_node(nid
)
5508 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5509 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5510 rstat
= &mz
->lruvec
.reclaim_stat
;
5512 recent_rotated
[0] += rstat
->recent_rotated
[0];
5513 recent_rotated
[1] += rstat
->recent_rotated
[1];
5514 recent_scanned
[0] += rstat
->recent_scanned
[0];
5515 recent_scanned
[1] += rstat
->recent_scanned
[1];
5517 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5518 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5519 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5520 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5527 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5530 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5532 return mem_cgroup_swappiness(memcg
);
5535 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5536 struct cftype
*cft
, u64 val
)
5538 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5539 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5541 if (val
> 100 || !parent
)
5544 mutex_lock(&memcg_create_mutex
);
5546 /* If under hierarchy, only empty-root can set this value */
5547 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5548 mutex_unlock(&memcg_create_mutex
);
5552 memcg
->swappiness
= val
;
5554 mutex_unlock(&memcg_create_mutex
);
5559 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5561 struct mem_cgroup_threshold_ary
*t
;
5567 t
= rcu_dereference(memcg
->thresholds
.primary
);
5569 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5574 usage
= mem_cgroup_usage(memcg
, swap
);
5577 * current_threshold points to threshold just below or equal to usage.
5578 * If it's not true, a threshold was crossed after last
5579 * call of __mem_cgroup_threshold().
5581 i
= t
->current_threshold
;
5584 * Iterate backward over array of thresholds starting from
5585 * current_threshold and check if a threshold is crossed.
5586 * If none of thresholds below usage is crossed, we read
5587 * only one element of the array here.
5589 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5590 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5592 /* i = current_threshold + 1 */
5596 * Iterate forward over array of thresholds starting from
5597 * current_threshold+1 and check if a threshold is crossed.
5598 * If none of thresholds above usage is crossed, we read
5599 * only one element of the array here.
5601 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5602 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5604 /* Update current_threshold */
5605 t
->current_threshold
= i
- 1;
5610 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5613 __mem_cgroup_threshold(memcg
, false);
5614 if (do_swap_account
)
5615 __mem_cgroup_threshold(memcg
, true);
5617 memcg
= parent_mem_cgroup(memcg
);
5621 static int compare_thresholds(const void *a
, const void *b
)
5623 const struct mem_cgroup_threshold
*_a
= a
;
5624 const struct mem_cgroup_threshold
*_b
= b
;
5626 if (_a
->threshold
> _b
->threshold
)
5629 if (_a
->threshold
< _b
->threshold
)
5635 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5637 struct mem_cgroup_eventfd_list
*ev
;
5639 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5640 eventfd_signal(ev
->eventfd
, 1);
5644 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5646 struct mem_cgroup
*iter
;
5648 for_each_mem_cgroup_tree(iter
, memcg
)
5649 mem_cgroup_oom_notify_cb(iter
);
5652 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5653 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5655 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5656 struct mem_cgroup_thresholds
*thresholds
;
5657 struct mem_cgroup_threshold_ary
*new;
5658 enum res_type type
= MEMFILE_TYPE(cft
->private);
5659 u64 threshold
, usage
;
5662 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5666 mutex_lock(&memcg
->thresholds_lock
);
5669 thresholds
= &memcg
->thresholds
;
5670 else if (type
== _MEMSWAP
)
5671 thresholds
= &memcg
->memsw_thresholds
;
5675 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5677 /* Check if a threshold crossed before adding a new one */
5678 if (thresholds
->primary
)
5679 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5681 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5683 /* Allocate memory for new array of thresholds */
5684 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5692 /* Copy thresholds (if any) to new array */
5693 if (thresholds
->primary
) {
5694 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5695 sizeof(struct mem_cgroup_threshold
));
5698 /* Add new threshold */
5699 new->entries
[size
- 1].eventfd
= eventfd
;
5700 new->entries
[size
- 1].threshold
= threshold
;
5702 /* Sort thresholds. Registering of new threshold isn't time-critical */
5703 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5704 compare_thresholds
, NULL
);
5706 /* Find current threshold */
5707 new->current_threshold
= -1;
5708 for (i
= 0; i
< size
; i
++) {
5709 if (new->entries
[i
].threshold
<= usage
) {
5711 * new->current_threshold will not be used until
5712 * rcu_assign_pointer(), so it's safe to increment
5715 ++new->current_threshold
;
5720 /* Free old spare buffer and save old primary buffer as spare */
5721 kfree(thresholds
->spare
);
5722 thresholds
->spare
= thresholds
->primary
;
5724 rcu_assign_pointer(thresholds
->primary
, new);
5726 /* To be sure that nobody uses thresholds */
5730 mutex_unlock(&memcg
->thresholds_lock
);
5735 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5736 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5738 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5739 struct mem_cgroup_thresholds
*thresholds
;
5740 struct mem_cgroup_threshold_ary
*new;
5741 enum res_type type
= MEMFILE_TYPE(cft
->private);
5745 mutex_lock(&memcg
->thresholds_lock
);
5747 thresholds
= &memcg
->thresholds
;
5748 else if (type
== _MEMSWAP
)
5749 thresholds
= &memcg
->memsw_thresholds
;
5753 if (!thresholds
->primary
)
5756 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5758 /* Check if a threshold crossed before removing */
5759 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5761 /* Calculate new number of threshold */
5763 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5764 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5768 new = thresholds
->spare
;
5770 /* Set thresholds array to NULL if we don't have thresholds */
5779 /* Copy thresholds and find current threshold */
5780 new->current_threshold
= -1;
5781 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5782 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5785 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5786 if (new->entries
[j
].threshold
<= usage
) {
5788 * new->current_threshold will not be used
5789 * until rcu_assign_pointer(), so it's safe to increment
5792 ++new->current_threshold
;
5798 /* Swap primary and spare array */
5799 thresholds
->spare
= thresholds
->primary
;
5800 /* If all events are unregistered, free the spare array */
5802 kfree(thresholds
->spare
);
5803 thresholds
->spare
= NULL
;
5806 rcu_assign_pointer(thresholds
->primary
, new);
5808 /* To be sure that nobody uses thresholds */
5811 mutex_unlock(&memcg
->thresholds_lock
);
5814 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5815 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5817 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5818 struct mem_cgroup_eventfd_list
*event
;
5819 enum res_type type
= MEMFILE_TYPE(cft
->private);
5821 BUG_ON(type
!= _OOM_TYPE
);
5822 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5826 spin_lock(&memcg_oom_lock
);
5828 event
->eventfd
= eventfd
;
5829 list_add(&event
->list
, &memcg
->oom_notify
);
5831 /* already in OOM ? */
5832 if (atomic_read(&memcg
->under_oom
))
5833 eventfd_signal(eventfd
, 1);
5834 spin_unlock(&memcg_oom_lock
);
5839 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5840 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5842 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5843 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5844 enum res_type type
= MEMFILE_TYPE(cft
->private);
5846 BUG_ON(type
!= _OOM_TYPE
);
5848 spin_lock(&memcg_oom_lock
);
5850 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5851 if (ev
->eventfd
== eventfd
) {
5852 list_del(&ev
->list
);
5857 spin_unlock(&memcg_oom_lock
);
5860 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5861 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5863 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5865 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5867 if (atomic_read(&memcg
->under_oom
))
5868 cb
->fill(cb
, "under_oom", 1);
5870 cb
->fill(cb
, "under_oom", 0);
5874 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5875 struct cftype
*cft
, u64 val
)
5877 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5878 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5880 /* cannot set to root cgroup and only 0 and 1 are allowed */
5881 if (!parent
|| !((val
== 0) || (val
== 1)))
5884 mutex_lock(&memcg_create_mutex
);
5885 /* oom-kill-disable is a flag for subhierarchy. */
5886 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5887 mutex_unlock(&memcg_create_mutex
);
5890 memcg
->oom_kill_disable
= val
;
5892 memcg_oom_recover(memcg
);
5893 mutex_unlock(&memcg_create_mutex
);
5897 #ifdef CONFIG_MEMCG_KMEM
5898 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5902 memcg
->kmemcg_id
= -1;
5903 ret
= memcg_propagate_kmem(memcg
);
5907 return mem_cgroup_sockets_init(memcg
, ss
);
5910 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5912 mem_cgroup_sockets_destroy(memcg
);
5915 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5917 if (!memcg_kmem_is_active(memcg
))
5921 * kmem charges can outlive the cgroup. In the case of slab
5922 * pages, for instance, a page contain objects from various
5923 * processes. As we prevent from taking a reference for every
5924 * such allocation we have to be careful when doing uncharge
5925 * (see memcg_uncharge_kmem) and here during offlining.
5927 * The idea is that that only the _last_ uncharge which sees
5928 * the dead memcg will drop the last reference. An additional
5929 * reference is taken here before the group is marked dead
5930 * which is then paired with css_put during uncharge resp. here.
5932 * Although this might sound strange as this path is called from
5933 * css_offline() when the referencemight have dropped down to 0
5934 * and shouldn't be incremented anymore (css_tryget would fail)
5935 * we do not have other options because of the kmem allocations
5938 css_get(&memcg
->css
);
5940 memcg_kmem_mark_dead(memcg
);
5942 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5945 if (memcg_kmem_test_and_clear_dead(memcg
))
5946 css_put(&memcg
->css
);
5949 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5954 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5958 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5963 static struct cftype mem_cgroup_files
[] = {
5965 .name
= "usage_in_bytes",
5966 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5967 .read
= mem_cgroup_read
,
5968 .register_event
= mem_cgroup_usage_register_event
,
5969 .unregister_event
= mem_cgroup_usage_unregister_event
,
5972 .name
= "max_usage_in_bytes",
5973 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5974 .trigger
= mem_cgroup_reset
,
5975 .read
= mem_cgroup_read
,
5978 .name
= "limit_in_bytes",
5979 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5980 .write_string
= mem_cgroup_write
,
5981 .read
= mem_cgroup_read
,
5984 .name
= "soft_limit_in_bytes",
5985 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5986 .write_string
= mem_cgroup_write
,
5987 .read
= mem_cgroup_read
,
5991 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5992 .trigger
= mem_cgroup_reset
,
5993 .read
= mem_cgroup_read
,
5997 .read_seq_string
= memcg_stat_show
,
6000 .name
= "force_empty",
6001 .trigger
= mem_cgroup_force_empty_write
,
6004 .name
= "use_hierarchy",
6005 .flags
= CFTYPE_INSANE
,
6006 .write_u64
= mem_cgroup_hierarchy_write
,
6007 .read_u64
= mem_cgroup_hierarchy_read
,
6010 .name
= "swappiness",
6011 .read_u64
= mem_cgroup_swappiness_read
,
6012 .write_u64
= mem_cgroup_swappiness_write
,
6015 .name
= "move_charge_at_immigrate",
6016 .read_u64
= mem_cgroup_move_charge_read
,
6017 .write_u64
= mem_cgroup_move_charge_write
,
6020 .name
= "oom_control",
6021 .read_map
= mem_cgroup_oom_control_read
,
6022 .write_u64
= mem_cgroup_oom_control_write
,
6023 .register_event
= mem_cgroup_oom_register_event
,
6024 .unregister_event
= mem_cgroup_oom_unregister_event
,
6025 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6028 .name
= "pressure_level",
6029 .register_event
= vmpressure_register_event
,
6030 .unregister_event
= vmpressure_unregister_event
,
6034 .name
= "numa_stat",
6035 .read_seq_string
= memcg_numa_stat_show
,
6038 #ifdef CONFIG_MEMCG_KMEM
6040 .name
= "kmem.limit_in_bytes",
6041 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6042 .write_string
= mem_cgroup_write
,
6043 .read
= mem_cgroup_read
,
6046 .name
= "kmem.usage_in_bytes",
6047 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6048 .read
= mem_cgroup_read
,
6051 .name
= "kmem.failcnt",
6052 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6053 .trigger
= mem_cgroup_reset
,
6054 .read
= mem_cgroup_read
,
6057 .name
= "kmem.max_usage_in_bytes",
6058 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6059 .trigger
= mem_cgroup_reset
,
6060 .read
= mem_cgroup_read
,
6062 #ifdef CONFIG_SLABINFO
6064 .name
= "kmem.slabinfo",
6065 .read_seq_string
= mem_cgroup_slabinfo_read
,
6069 { }, /* terminate */
6072 #ifdef CONFIG_MEMCG_SWAP
6073 static struct cftype memsw_cgroup_files
[] = {
6075 .name
= "memsw.usage_in_bytes",
6076 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6077 .read
= mem_cgroup_read
,
6078 .register_event
= mem_cgroup_usage_register_event
,
6079 .unregister_event
= mem_cgroup_usage_unregister_event
,
6082 .name
= "memsw.max_usage_in_bytes",
6083 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6084 .trigger
= mem_cgroup_reset
,
6085 .read
= mem_cgroup_read
,
6088 .name
= "memsw.limit_in_bytes",
6089 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6090 .write_string
= mem_cgroup_write
,
6091 .read
= mem_cgroup_read
,
6094 .name
= "memsw.failcnt",
6095 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6096 .trigger
= mem_cgroup_reset
,
6097 .read
= mem_cgroup_read
,
6099 { }, /* terminate */
6102 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6104 struct mem_cgroup_per_node
*pn
;
6105 struct mem_cgroup_per_zone
*mz
;
6106 int zone
, tmp
= node
;
6108 * This routine is called against possible nodes.
6109 * But it's BUG to call kmalloc() against offline node.
6111 * TODO: this routine can waste much memory for nodes which will
6112 * never be onlined. It's better to use memory hotplug callback
6115 if (!node_state(node
, N_NORMAL_MEMORY
))
6117 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6121 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6122 mz
= &pn
->zoneinfo
[zone
];
6123 lruvec_init(&mz
->lruvec
);
6124 mz
->usage_in_excess
= 0;
6125 mz
->on_tree
= false;
6128 memcg
->nodeinfo
[node
] = pn
;
6132 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6134 kfree(memcg
->nodeinfo
[node
]);
6137 static struct mem_cgroup
*mem_cgroup_alloc(void)
6139 struct mem_cgroup
*memcg
;
6140 size_t size
= memcg_size();
6142 /* Can be very big if nr_node_ids is very big */
6143 if (size
< PAGE_SIZE
)
6144 memcg
= kzalloc(size
, GFP_KERNEL
);
6146 memcg
= vzalloc(size
);
6151 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6154 spin_lock_init(&memcg
->pcp_counter_lock
);
6158 if (size
< PAGE_SIZE
)
6166 * At destroying mem_cgroup, references from swap_cgroup can remain.
6167 * (scanning all at force_empty is too costly...)
6169 * Instead of clearing all references at force_empty, we remember
6170 * the number of reference from swap_cgroup and free mem_cgroup when
6171 * it goes down to 0.
6173 * Removal of cgroup itself succeeds regardless of refs from swap.
6176 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6179 size_t size
= memcg_size();
6181 mem_cgroup_remove_from_trees(memcg
);
6184 free_mem_cgroup_per_zone_info(memcg
, node
);
6186 free_percpu(memcg
->stat
);
6189 * We need to make sure that (at least for now), the jump label
6190 * destruction code runs outside of the cgroup lock. This is because
6191 * get_online_cpus(), which is called from the static_branch update,
6192 * can't be called inside the cgroup_lock. cpusets are the ones
6193 * enforcing this dependency, so if they ever change, we might as well.
6195 * schedule_work() will guarantee this happens. Be careful if you need
6196 * to move this code around, and make sure it is outside
6199 disarm_static_keys(memcg
);
6200 if (size
< PAGE_SIZE
)
6207 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6209 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6211 if (!memcg
->res
.parent
)
6213 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6215 EXPORT_SYMBOL(parent_mem_cgroup
);
6217 static void __init
mem_cgroup_soft_limit_tree_init(void)
6219 struct mem_cgroup_tree_per_node
*rtpn
;
6220 struct mem_cgroup_tree_per_zone
*rtpz
;
6221 int tmp
, node
, zone
;
6223 for_each_node(node
) {
6225 if (!node_state(node
, N_NORMAL_MEMORY
))
6227 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6230 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6232 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6233 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6234 rtpz
->rb_root
= RB_ROOT
;
6235 spin_lock_init(&rtpz
->lock
);
6240 static struct cgroup_subsys_state
* __ref
6241 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6243 struct mem_cgroup
*memcg
;
6244 long error
= -ENOMEM
;
6247 memcg
= mem_cgroup_alloc();
6249 return ERR_PTR(error
);
6252 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6256 if (parent_css
== NULL
) {
6257 root_mem_cgroup
= memcg
;
6258 res_counter_init(&memcg
->res
, NULL
);
6259 res_counter_init(&memcg
->memsw
, NULL
);
6260 res_counter_init(&memcg
->kmem
, NULL
);
6263 memcg
->last_scanned_node
= MAX_NUMNODES
;
6264 INIT_LIST_HEAD(&memcg
->oom_notify
);
6265 memcg
->move_charge_at_immigrate
= 0;
6266 mutex_init(&memcg
->thresholds_lock
);
6267 spin_lock_init(&memcg
->move_lock
);
6268 vmpressure_init(&memcg
->vmpressure
);
6273 __mem_cgroup_free(memcg
);
6274 return ERR_PTR(error
);
6278 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6280 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6281 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6284 if (css
->cgroup
->id
> MEM_CGROUP_ID_MAX
)
6290 mutex_lock(&memcg_create_mutex
);
6292 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6293 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6294 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6296 if (parent
->use_hierarchy
) {
6297 res_counter_init(&memcg
->res
, &parent
->res
);
6298 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6299 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6302 * No need to take a reference to the parent because cgroup
6303 * core guarantees its existence.
6306 res_counter_init(&memcg
->res
, NULL
);
6307 res_counter_init(&memcg
->memsw
, NULL
);
6308 res_counter_init(&memcg
->kmem
, NULL
);
6310 * Deeper hierachy with use_hierarchy == false doesn't make
6311 * much sense so let cgroup subsystem know about this
6312 * unfortunate state in our controller.
6314 if (parent
!= root_mem_cgroup
)
6315 mem_cgroup_subsys
.broken_hierarchy
= true;
6318 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6319 mutex_unlock(&memcg_create_mutex
);
6324 * Announce all parents that a group from their hierarchy is gone.
6326 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6328 struct mem_cgroup
*parent
= memcg
;
6330 while ((parent
= parent_mem_cgroup(parent
)))
6331 mem_cgroup_iter_invalidate(parent
);
6334 * if the root memcg is not hierarchical we have to check it
6337 if (!root_mem_cgroup
->use_hierarchy
)
6338 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6341 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6343 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6345 kmem_cgroup_css_offline(memcg
);
6347 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6348 mem_cgroup_reparent_charges(memcg
);
6349 mem_cgroup_destroy_all_caches(memcg
);
6350 vmpressure_cleanup(&memcg
->vmpressure
);
6353 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6355 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6357 memcg_destroy_kmem(memcg
);
6358 __mem_cgroup_free(memcg
);
6362 /* Handlers for move charge at task migration. */
6363 #define PRECHARGE_COUNT_AT_ONCE 256
6364 static int mem_cgroup_do_precharge(unsigned long count
)
6367 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6368 struct mem_cgroup
*memcg
= mc
.to
;
6370 if (mem_cgroup_is_root(memcg
)) {
6371 mc
.precharge
+= count
;
6372 /* we don't need css_get for root */
6375 /* try to charge at once */
6377 struct res_counter
*dummy
;
6379 * "memcg" cannot be under rmdir() because we've already checked
6380 * by cgroup_lock_live_cgroup() that it is not removed and we
6381 * are still under the same cgroup_mutex. So we can postpone
6384 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6386 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6387 PAGE_SIZE
* count
, &dummy
)) {
6388 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6391 mc
.precharge
+= count
;
6395 /* fall back to one by one charge */
6397 if (signal_pending(current
)) {
6401 if (!batch_count
--) {
6402 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6405 ret
= __mem_cgroup_try_charge(NULL
,
6406 GFP_KERNEL
, 1, &memcg
, false);
6408 /* mem_cgroup_clear_mc() will do uncharge later */
6416 * get_mctgt_type - get target type of moving charge
6417 * @vma: the vma the pte to be checked belongs
6418 * @addr: the address corresponding to the pte to be checked
6419 * @ptent: the pte to be checked
6420 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6423 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6424 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6425 * move charge. if @target is not NULL, the page is stored in target->page
6426 * with extra refcnt got(Callers should handle it).
6427 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6428 * target for charge migration. if @target is not NULL, the entry is stored
6431 * Called with pte lock held.
6438 enum mc_target_type
{
6444 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6445 unsigned long addr
, pte_t ptent
)
6447 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6449 if (!page
|| !page_mapped(page
))
6451 if (PageAnon(page
)) {
6452 /* we don't move shared anon */
6455 } else if (!move_file())
6456 /* we ignore mapcount for file pages */
6458 if (!get_page_unless_zero(page
))
6465 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6466 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6468 struct page
*page
= NULL
;
6469 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6471 if (!move_anon() || non_swap_entry(ent
))
6474 * Because lookup_swap_cache() updates some statistics counter,
6475 * we call find_get_page() with swapper_space directly.
6477 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6478 if (do_swap_account
)
6479 entry
->val
= ent
.val
;
6484 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6485 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6491 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6492 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6494 struct page
*page
= NULL
;
6495 struct address_space
*mapping
;
6498 if (!vma
->vm_file
) /* anonymous vma */
6503 mapping
= vma
->vm_file
->f_mapping
;
6504 if (pte_none(ptent
))
6505 pgoff
= linear_page_index(vma
, addr
);
6506 else /* pte_file(ptent) is true */
6507 pgoff
= pte_to_pgoff(ptent
);
6509 /* page is moved even if it's not RSS of this task(page-faulted). */
6510 page
= find_get_page(mapping
, pgoff
);
6513 /* shmem/tmpfs may report page out on swap: account for that too. */
6514 if (radix_tree_exceptional_entry(page
)) {
6515 swp_entry_t swap
= radix_to_swp_entry(page
);
6516 if (do_swap_account
)
6518 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6524 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6525 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6527 struct page
*page
= NULL
;
6528 struct page_cgroup
*pc
;
6529 enum mc_target_type ret
= MC_TARGET_NONE
;
6530 swp_entry_t ent
= { .val
= 0 };
6532 if (pte_present(ptent
))
6533 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6534 else if (is_swap_pte(ptent
))
6535 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6536 else if (pte_none(ptent
) || pte_file(ptent
))
6537 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6539 if (!page
&& !ent
.val
)
6542 pc
= lookup_page_cgroup(page
);
6544 * Do only loose check w/o page_cgroup lock.
6545 * mem_cgroup_move_account() checks the pc is valid or not under
6548 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6549 ret
= MC_TARGET_PAGE
;
6551 target
->page
= page
;
6553 if (!ret
|| !target
)
6556 /* There is a swap entry and a page doesn't exist or isn't charged */
6557 if (ent
.val
&& !ret
&&
6558 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6559 ret
= MC_TARGET_SWAP
;
6566 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6568 * We don't consider swapping or file mapped pages because THP does not
6569 * support them for now.
6570 * Caller should make sure that pmd_trans_huge(pmd) is true.
6572 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6573 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6575 struct page
*page
= NULL
;
6576 struct page_cgroup
*pc
;
6577 enum mc_target_type ret
= MC_TARGET_NONE
;
6579 page
= pmd_page(pmd
);
6580 VM_BUG_ON(!page
|| !PageHead(page
));
6583 pc
= lookup_page_cgroup(page
);
6584 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6585 ret
= MC_TARGET_PAGE
;
6588 target
->page
= page
;
6594 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6595 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6597 return MC_TARGET_NONE
;
6601 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6602 unsigned long addr
, unsigned long end
,
6603 struct mm_walk
*walk
)
6605 struct vm_area_struct
*vma
= walk
->private;
6609 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6610 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6611 mc
.precharge
+= HPAGE_PMD_NR
;
6612 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6616 if (pmd_trans_unstable(pmd
))
6618 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6619 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6620 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6621 mc
.precharge
++; /* increment precharge temporarily */
6622 pte_unmap_unlock(pte
- 1, ptl
);
6628 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6630 unsigned long precharge
;
6631 struct vm_area_struct
*vma
;
6633 down_read(&mm
->mmap_sem
);
6634 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6635 struct mm_walk mem_cgroup_count_precharge_walk
= {
6636 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6640 if (is_vm_hugetlb_page(vma
))
6642 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6643 &mem_cgroup_count_precharge_walk
);
6645 up_read(&mm
->mmap_sem
);
6647 precharge
= mc
.precharge
;
6653 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6655 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6657 VM_BUG_ON(mc
.moving_task
);
6658 mc
.moving_task
= current
;
6659 return mem_cgroup_do_precharge(precharge
);
6662 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6663 static void __mem_cgroup_clear_mc(void)
6665 struct mem_cgroup
*from
= mc
.from
;
6666 struct mem_cgroup
*to
= mc
.to
;
6669 /* we must uncharge all the leftover precharges from mc.to */
6671 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6675 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6676 * we must uncharge here.
6678 if (mc
.moved_charge
) {
6679 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6680 mc
.moved_charge
= 0;
6682 /* we must fixup refcnts and charges */
6683 if (mc
.moved_swap
) {
6684 /* uncharge swap account from the old cgroup */
6685 if (!mem_cgroup_is_root(mc
.from
))
6686 res_counter_uncharge(&mc
.from
->memsw
,
6687 PAGE_SIZE
* mc
.moved_swap
);
6689 for (i
= 0; i
< mc
.moved_swap
; i
++)
6690 css_put(&mc
.from
->css
);
6692 if (!mem_cgroup_is_root(mc
.to
)) {
6694 * we charged both to->res and to->memsw, so we should
6697 res_counter_uncharge(&mc
.to
->res
,
6698 PAGE_SIZE
* mc
.moved_swap
);
6700 /* we've already done css_get(mc.to) */
6703 memcg_oom_recover(from
);
6704 memcg_oom_recover(to
);
6705 wake_up_all(&mc
.waitq
);
6708 static void mem_cgroup_clear_mc(void)
6710 struct mem_cgroup
*from
= mc
.from
;
6713 * we must clear moving_task before waking up waiters at the end of
6716 mc
.moving_task
= NULL
;
6717 __mem_cgroup_clear_mc();
6718 spin_lock(&mc
.lock
);
6721 spin_unlock(&mc
.lock
);
6722 mem_cgroup_end_move(from
);
6725 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6726 struct cgroup_taskset
*tset
)
6728 struct task_struct
*p
= cgroup_taskset_first(tset
);
6730 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6731 unsigned long move_charge_at_immigrate
;
6734 * We are now commited to this value whatever it is. Changes in this
6735 * tunable will only affect upcoming migrations, not the current one.
6736 * So we need to save it, and keep it going.
6738 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6739 if (move_charge_at_immigrate
) {
6740 struct mm_struct
*mm
;
6741 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6743 VM_BUG_ON(from
== memcg
);
6745 mm
= get_task_mm(p
);
6748 /* We move charges only when we move a owner of the mm */
6749 if (mm
->owner
== p
) {
6752 VM_BUG_ON(mc
.precharge
);
6753 VM_BUG_ON(mc
.moved_charge
);
6754 VM_BUG_ON(mc
.moved_swap
);
6755 mem_cgroup_start_move(from
);
6756 spin_lock(&mc
.lock
);
6759 mc
.immigrate_flags
= move_charge_at_immigrate
;
6760 spin_unlock(&mc
.lock
);
6761 /* We set mc.moving_task later */
6763 ret
= mem_cgroup_precharge_mc(mm
);
6765 mem_cgroup_clear_mc();
6772 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6773 struct cgroup_taskset
*tset
)
6775 mem_cgroup_clear_mc();
6778 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6779 unsigned long addr
, unsigned long end
,
6780 struct mm_walk
*walk
)
6783 struct vm_area_struct
*vma
= walk
->private;
6786 enum mc_target_type target_type
;
6787 union mc_target target
;
6789 struct page_cgroup
*pc
;
6792 * We don't take compound_lock() here but no race with splitting thp
6794 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6795 * under splitting, which means there's no concurrent thp split,
6796 * - if another thread runs into split_huge_page() just after we
6797 * entered this if-block, the thread must wait for page table lock
6798 * to be unlocked in __split_huge_page_splitting(), where the main
6799 * part of thp split is not executed yet.
6801 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6802 if (mc
.precharge
< HPAGE_PMD_NR
) {
6803 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6806 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6807 if (target_type
== MC_TARGET_PAGE
) {
6809 if (!isolate_lru_page(page
)) {
6810 pc
= lookup_page_cgroup(page
);
6811 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6812 pc
, mc
.from
, mc
.to
)) {
6813 mc
.precharge
-= HPAGE_PMD_NR
;
6814 mc
.moved_charge
+= HPAGE_PMD_NR
;
6816 putback_lru_page(page
);
6820 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6824 if (pmd_trans_unstable(pmd
))
6827 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6828 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6829 pte_t ptent
= *(pte
++);
6835 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6836 case MC_TARGET_PAGE
:
6838 if (isolate_lru_page(page
))
6840 pc
= lookup_page_cgroup(page
);
6841 if (!mem_cgroup_move_account(page
, 1, pc
,
6844 /* we uncharge from mc.from later. */
6847 putback_lru_page(page
);
6848 put
: /* get_mctgt_type() gets the page */
6851 case MC_TARGET_SWAP
:
6853 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6855 /* we fixup refcnts and charges later. */
6863 pte_unmap_unlock(pte
- 1, ptl
);
6868 * We have consumed all precharges we got in can_attach().
6869 * We try charge one by one, but don't do any additional
6870 * charges to mc.to if we have failed in charge once in attach()
6873 ret
= mem_cgroup_do_precharge(1);
6881 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6883 struct vm_area_struct
*vma
;
6885 lru_add_drain_all();
6887 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6889 * Someone who are holding the mmap_sem might be waiting in
6890 * waitq. So we cancel all extra charges, wake up all waiters,
6891 * and retry. Because we cancel precharges, we might not be able
6892 * to move enough charges, but moving charge is a best-effort
6893 * feature anyway, so it wouldn't be a big problem.
6895 __mem_cgroup_clear_mc();
6899 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6901 struct mm_walk mem_cgroup_move_charge_walk
= {
6902 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6906 if (is_vm_hugetlb_page(vma
))
6908 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6909 &mem_cgroup_move_charge_walk
);
6912 * means we have consumed all precharges and failed in
6913 * doing additional charge. Just abandon here.
6917 up_read(&mm
->mmap_sem
);
6920 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6921 struct cgroup_taskset
*tset
)
6923 struct task_struct
*p
= cgroup_taskset_first(tset
);
6924 struct mm_struct
*mm
= get_task_mm(p
);
6928 mem_cgroup_move_charge(mm
);
6932 mem_cgroup_clear_mc();
6934 #else /* !CONFIG_MMU */
6935 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6936 struct cgroup_taskset
*tset
)
6940 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6941 struct cgroup_taskset
*tset
)
6944 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6945 struct cgroup_taskset
*tset
)
6951 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6952 * to verify sane_behavior flag on each mount attempt.
6954 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6957 * use_hierarchy is forced with sane_behavior. cgroup core
6958 * guarantees that @root doesn't have any children, so turning it
6959 * on for the root memcg is enough.
6961 if (cgroup_sane_behavior(root_css
->cgroup
))
6962 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6965 struct cgroup_subsys mem_cgroup_subsys
= {
6967 .subsys_id
= mem_cgroup_subsys_id
,
6968 .css_alloc
= mem_cgroup_css_alloc
,
6969 .css_online
= mem_cgroup_css_online
,
6970 .css_offline
= mem_cgroup_css_offline
,
6971 .css_free
= mem_cgroup_css_free
,
6972 .can_attach
= mem_cgroup_can_attach
,
6973 .cancel_attach
= mem_cgroup_cancel_attach
,
6974 .attach
= mem_cgroup_move_task
,
6975 .bind
= mem_cgroup_bind
,
6976 .base_cftypes
= mem_cgroup_files
,
6980 #ifdef CONFIG_MEMCG_SWAP
6981 static int __init
enable_swap_account(char *s
)
6983 if (!strcmp(s
, "1"))
6984 really_do_swap_account
= 1;
6985 else if (!strcmp(s
, "0"))
6986 really_do_swap_account
= 0;
6989 __setup("swapaccount=", enable_swap_account
);
6991 static void __init
memsw_file_init(void)
6993 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6996 static void __init
enable_swap_cgroup(void)
6998 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6999 do_swap_account
= 1;
7005 static void __init
enable_swap_cgroup(void)
7011 * subsys_initcall() for memory controller.
7013 * Some parts like hotcpu_notifier() have to be initialized from this context
7014 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7015 * everything that doesn't depend on a specific mem_cgroup structure should
7016 * be initialized from here.
7018 static int __init
mem_cgroup_init(void)
7020 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7021 enable_swap_cgroup();
7022 mem_cgroup_soft_limit_tree_init();
7026 subsys_initcall(mem_cgroup_init
);