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>
64 #include <asm/uaccess.h>
66 #include <trace/events/vmscan.h>
68 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
69 EXPORT_SYMBOL(mem_cgroup_subsys
);
71 #define MEM_CGROUP_RECLAIM_RETRIES 5
72 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
74 #ifdef CONFIG_MEMCG_SWAP
75 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
76 int do_swap_account __read_mostly
;
78 /* for remember boot option*/
79 #ifdef CONFIG_MEMCG_SWAP_ENABLED
80 static int really_do_swap_account __initdata
= 1;
82 static int really_do_swap_account __initdata
= 0;
86 #define do_swap_account 0
90 static const char * const mem_cgroup_stat_names
[] = {
99 enum mem_cgroup_events_index
{
100 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
101 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
102 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
103 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
104 MEM_CGROUP_EVENTS_NSTATS
,
107 static const char * const mem_cgroup_events_names
[] = {
114 static const char * const mem_cgroup_lru_names
[] = {
123 * Per memcg event counter is incremented at every pagein/pageout. With THP,
124 * it will be incremated by the number of pages. This counter is used for
125 * for trigger some periodic events. This is straightforward and better
126 * than using jiffies etc. to handle periodic memcg event.
128 enum mem_cgroup_events_target
{
129 MEM_CGROUP_TARGET_THRESH
,
130 MEM_CGROUP_TARGET_SOFTLIMIT
,
131 MEM_CGROUP_TARGET_NUMAINFO
,
134 #define THRESHOLDS_EVENTS_TARGET 128
135 #define SOFTLIMIT_EVENTS_TARGET 1024
136 #define NUMAINFO_EVENTS_TARGET 1024
138 struct mem_cgroup_stat_cpu
{
139 long count
[MEM_CGROUP_STAT_NSTATS
];
140 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
141 unsigned long nr_page_events
;
142 unsigned long targets
[MEM_CGROUP_NTARGETS
];
145 struct mem_cgroup_reclaim_iter
{
147 * last scanned hierarchy member. Valid only if last_dead_count
148 * matches memcg->dead_count of the hierarchy root group.
150 struct mem_cgroup
*last_visited
;
151 unsigned long last_dead_count
;
153 /* scan generation, increased every round-trip */
154 unsigned int generation
;
158 * per-zone information in memory controller.
160 struct mem_cgroup_per_zone
{
161 struct lruvec lruvec
;
162 unsigned long lru_size
[NR_LRU_LISTS
];
164 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
166 struct rb_node tree_node
; /* RB tree node */
167 unsigned long long usage_in_excess
;/* Set to the value by which */
168 /* the soft limit is exceeded*/
170 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
171 /* use container_of */
174 struct mem_cgroup_per_node
{
175 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
179 * Cgroups above their limits are maintained in a RB-Tree, independent of
180 * their hierarchy representation
183 struct mem_cgroup_tree_per_zone
{
184 struct rb_root rb_root
;
188 struct mem_cgroup_tree_per_node
{
189 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
192 struct mem_cgroup_tree
{
193 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
196 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
198 struct mem_cgroup_threshold
{
199 struct eventfd_ctx
*eventfd
;
204 struct mem_cgroup_threshold_ary
{
205 /* An array index points to threshold just below or equal to usage. */
206 int current_threshold
;
207 /* Size of entries[] */
209 /* Array of thresholds */
210 struct mem_cgroup_threshold entries
[0];
213 struct mem_cgroup_thresholds
{
214 /* Primary thresholds array */
215 struct mem_cgroup_threshold_ary
*primary
;
217 * Spare threshold array.
218 * This is needed to make mem_cgroup_unregister_event() "never fail".
219 * It must be able to store at least primary->size - 1 entries.
221 struct mem_cgroup_threshold_ary
*spare
;
225 struct mem_cgroup_eventfd_list
{
226 struct list_head list
;
227 struct eventfd_ctx
*eventfd
;
230 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
231 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
234 * The memory controller data structure. The memory controller controls both
235 * page cache and RSS per cgroup. We would eventually like to provide
236 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
237 * to help the administrator determine what knobs to tune.
239 * TODO: Add a water mark for the memory controller. Reclaim will begin when
240 * we hit the water mark. May be even add a low water mark, such that
241 * no reclaim occurs from a cgroup at it's low water mark, this is
242 * a feature that will be implemented much later in the future.
245 struct cgroup_subsys_state css
;
247 * the counter to account for memory usage
249 struct res_counter res
;
251 /* vmpressure notifications */
252 struct vmpressure vmpressure
;
255 * the counter to account for mem+swap usage.
257 struct res_counter memsw
;
260 * the counter to account for kernel memory usage.
262 struct res_counter kmem
;
264 * Should the accounting and control be hierarchical, per subtree?
267 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
271 atomic_t oom_wakeups
;
274 /* OOM-Killer disable */
275 int oom_kill_disable
;
277 /* set when res.limit == memsw.limit */
278 bool memsw_is_minimum
;
280 /* protect arrays of thresholds */
281 struct mutex thresholds_lock
;
283 /* thresholds for memory usage. RCU-protected */
284 struct mem_cgroup_thresholds thresholds
;
286 /* thresholds for mem+swap usage. RCU-protected */
287 struct mem_cgroup_thresholds memsw_thresholds
;
289 /* For oom notifier event fd */
290 struct list_head oom_notify
;
293 * Should we move charges of a task when a task is moved into this
294 * mem_cgroup ? And what type of charges should we move ?
296 unsigned long move_charge_at_immigrate
;
298 * set > 0 if pages under this cgroup are moving to other cgroup.
300 atomic_t moving_account
;
301 /* taken only while moving_account > 0 */
302 spinlock_t move_lock
;
306 struct mem_cgroup_stat_cpu __percpu
*stat
;
308 * used when a cpu is offlined or other synchronizations
309 * See mem_cgroup_read_stat().
311 struct mem_cgroup_stat_cpu nocpu_base
;
312 spinlock_t pcp_counter_lock
;
315 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
316 struct tcp_memcontrol tcp_mem
;
318 #if defined(CONFIG_MEMCG_KMEM)
319 /* analogous to slab_common's slab_caches list. per-memcg */
320 struct list_head memcg_slab_caches
;
321 /* Not a spinlock, we can take a lot of time walking the list */
322 struct mutex slab_caches_mutex
;
323 /* Index in the kmem_cache->memcg_params->memcg_caches array */
327 int last_scanned_node
;
329 nodemask_t scan_nodes
;
330 atomic_t numainfo_events
;
331 atomic_t numainfo_updating
;
334 struct mem_cgroup_per_node
*nodeinfo
[0];
335 /* WARNING: nodeinfo must be the last member here */
338 static size_t memcg_size(void)
340 return sizeof(struct mem_cgroup
) +
341 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
344 /* internal only representation about the status of kmem accounting. */
346 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
347 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
348 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
351 /* We account when limit is on, but only after call sites are patched */
352 #define KMEM_ACCOUNTED_MASK \
353 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
355 #ifdef CONFIG_MEMCG_KMEM
356 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
358 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
361 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
363 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
366 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
368 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
371 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
373 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
376 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
379 * Our caller must use css_get() first, because memcg_uncharge_kmem()
380 * will call css_put() if it sees the memcg is dead.
383 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
384 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
387 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
389 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
390 &memcg
->kmem_account_flags
);
394 /* Stuffs for move charges at task migration. */
396 * Types of charges to be moved. "move_charge_at_immitgrate" and
397 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
400 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
401 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
405 /* "mc" and its members are protected by cgroup_mutex */
406 static struct move_charge_struct
{
407 spinlock_t lock
; /* for from, to */
408 struct mem_cgroup
*from
;
409 struct mem_cgroup
*to
;
410 unsigned long immigrate_flags
;
411 unsigned long precharge
;
412 unsigned long moved_charge
;
413 unsigned long moved_swap
;
414 struct task_struct
*moving_task
; /* a task moving charges */
415 wait_queue_head_t waitq
; /* a waitq for other context */
417 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
418 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
421 static bool move_anon(void)
423 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
426 static bool move_file(void)
428 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
432 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
433 * limit reclaim to prevent infinite loops, if they ever occur.
435 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
436 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
439 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
440 MEM_CGROUP_CHARGE_TYPE_ANON
,
441 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
442 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
446 /* for encoding cft->private value on file */
454 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
455 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
456 #define MEMFILE_ATTR(val) ((val) & 0xffff)
457 /* Used for OOM nofiier */
458 #define OOM_CONTROL (0)
461 * Reclaim flags for mem_cgroup_hierarchical_reclaim
463 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
464 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
465 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
466 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
469 * The memcg_create_mutex will be held whenever a new cgroup is created.
470 * As a consequence, any change that needs to protect against new child cgroups
471 * appearing has to hold it as well.
473 static DEFINE_MUTEX(memcg_create_mutex
);
475 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
477 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
480 /* Some nice accessors for the vmpressure. */
481 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
484 memcg
= root_mem_cgroup
;
485 return &memcg
->vmpressure
;
488 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
490 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
493 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
495 return &mem_cgroup_from_css(css
)->vmpressure
;
498 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
500 return (memcg
== root_mem_cgroup
);
503 /* Writing them here to avoid exposing memcg's inner layout */
504 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
506 void sock_update_memcg(struct sock
*sk
)
508 if (mem_cgroup_sockets_enabled
) {
509 struct mem_cgroup
*memcg
;
510 struct cg_proto
*cg_proto
;
512 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
514 /* Socket cloning can throw us here with sk_cgrp already
515 * filled. It won't however, necessarily happen from
516 * process context. So the test for root memcg given
517 * the current task's memcg won't help us in this case.
519 * Respecting the original socket's memcg is a better
520 * decision in this case.
523 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
524 css_get(&sk
->sk_cgrp
->memcg
->css
);
529 memcg
= mem_cgroup_from_task(current
);
530 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
531 if (!mem_cgroup_is_root(memcg
) &&
532 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
533 sk
->sk_cgrp
= cg_proto
;
538 EXPORT_SYMBOL(sock_update_memcg
);
540 void sock_release_memcg(struct sock
*sk
)
542 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
543 struct mem_cgroup
*memcg
;
544 WARN_ON(!sk
->sk_cgrp
->memcg
);
545 memcg
= sk
->sk_cgrp
->memcg
;
546 css_put(&sk
->sk_cgrp
->memcg
->css
);
550 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
552 if (!memcg
|| mem_cgroup_is_root(memcg
))
555 return &memcg
->tcp_mem
.cg_proto
;
557 EXPORT_SYMBOL(tcp_proto_cgroup
);
559 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
561 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
563 static_key_slow_dec(&memcg_socket_limit_enabled
);
566 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
571 #ifdef CONFIG_MEMCG_KMEM
573 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
574 * There are two main reasons for not using the css_id for this:
575 * 1) this works better in sparse environments, where we have a lot of memcgs,
576 * but only a few kmem-limited. Or also, if we have, for instance, 200
577 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
578 * 200 entry array for that.
580 * 2) In order not to violate the cgroup API, we would like to do all memory
581 * allocation in ->create(). At that point, we haven't yet allocated the
582 * css_id. Having a separate index prevents us from messing with the cgroup
585 * The current size of the caches array is stored in
586 * memcg_limited_groups_array_size. It will double each time we have to
589 static DEFINE_IDA(kmem_limited_groups
);
590 int memcg_limited_groups_array_size
;
593 * MIN_SIZE is different than 1, because we would like to avoid going through
594 * the alloc/free process all the time. In a small machine, 4 kmem-limited
595 * cgroups is a reasonable guess. In the future, it could be a parameter or
596 * tunable, but that is strictly not necessary.
598 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
599 * this constant directly from cgroup, but it is understandable that this is
600 * better kept as an internal representation in cgroup.c. In any case, the
601 * css_id space is not getting any smaller, and we don't have to necessarily
602 * increase ours as well if it increases.
604 #define MEMCG_CACHES_MIN_SIZE 4
605 #define MEMCG_CACHES_MAX_SIZE 65535
608 * A lot of the calls to the cache allocation functions are expected to be
609 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
610 * conditional to this static branch, we'll have to allow modules that does
611 * kmem_cache_alloc and the such to see this symbol as well
613 struct static_key memcg_kmem_enabled_key
;
614 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
616 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
618 if (memcg_kmem_is_active(memcg
)) {
619 static_key_slow_dec(&memcg_kmem_enabled_key
);
620 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
623 * This check can't live in kmem destruction function,
624 * since the charges will outlive the cgroup
626 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
629 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
632 #endif /* CONFIG_MEMCG_KMEM */
634 static void disarm_static_keys(struct mem_cgroup
*memcg
)
636 disarm_sock_keys(memcg
);
637 disarm_kmem_keys(memcg
);
640 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
642 static struct mem_cgroup_per_zone
*
643 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
645 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
646 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
649 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
654 static struct mem_cgroup_per_zone
*
655 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
657 int nid
= page_to_nid(page
);
658 int zid
= page_zonenum(page
);
660 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
663 static struct mem_cgroup_tree_per_zone
*
664 soft_limit_tree_node_zone(int nid
, int zid
)
666 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
669 static struct mem_cgroup_tree_per_zone
*
670 soft_limit_tree_from_page(struct page
*page
)
672 int nid
= page_to_nid(page
);
673 int zid
= page_zonenum(page
);
675 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
679 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
680 struct mem_cgroup_per_zone
*mz
,
681 struct mem_cgroup_tree_per_zone
*mctz
,
682 unsigned long long new_usage_in_excess
)
684 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
685 struct rb_node
*parent
= NULL
;
686 struct mem_cgroup_per_zone
*mz_node
;
691 mz
->usage_in_excess
= new_usage_in_excess
;
692 if (!mz
->usage_in_excess
)
696 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
698 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
701 * We can't avoid mem cgroups that are over their soft
702 * limit by the same amount
704 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
707 rb_link_node(&mz
->tree_node
, parent
, p
);
708 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
713 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
714 struct mem_cgroup_per_zone
*mz
,
715 struct mem_cgroup_tree_per_zone
*mctz
)
719 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
724 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
725 struct mem_cgroup_per_zone
*mz
,
726 struct mem_cgroup_tree_per_zone
*mctz
)
728 spin_lock(&mctz
->lock
);
729 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
730 spin_unlock(&mctz
->lock
);
734 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
736 unsigned long long excess
;
737 struct mem_cgroup_per_zone
*mz
;
738 struct mem_cgroup_tree_per_zone
*mctz
;
739 int nid
= page_to_nid(page
);
740 int zid
= page_zonenum(page
);
741 mctz
= soft_limit_tree_from_page(page
);
744 * Necessary to update all ancestors when hierarchy is used.
745 * because their event counter is not touched.
747 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
748 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
749 excess
= res_counter_soft_limit_excess(&memcg
->res
);
751 * We have to update the tree if mz is on RB-tree or
752 * mem is over its softlimit.
754 if (excess
|| mz
->on_tree
) {
755 spin_lock(&mctz
->lock
);
756 /* if on-tree, remove it */
758 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
760 * Insert again. mz->usage_in_excess will be updated.
761 * If excess is 0, no tree ops.
763 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
764 spin_unlock(&mctz
->lock
);
769 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
772 struct mem_cgroup_per_zone
*mz
;
773 struct mem_cgroup_tree_per_zone
*mctz
;
775 for_each_node(node
) {
776 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
777 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
778 mctz
= soft_limit_tree_node_zone(node
, zone
);
779 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
784 static struct mem_cgroup_per_zone
*
785 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
787 struct rb_node
*rightmost
= NULL
;
788 struct mem_cgroup_per_zone
*mz
;
792 rightmost
= rb_last(&mctz
->rb_root
);
794 goto done
; /* Nothing to reclaim from */
796 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
798 * Remove the node now but someone else can add it back,
799 * we will to add it back at the end of reclaim to its correct
800 * position in the tree.
802 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
803 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
804 !css_tryget(&mz
->memcg
->css
))
810 static struct mem_cgroup_per_zone
*
811 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
813 struct mem_cgroup_per_zone
*mz
;
815 spin_lock(&mctz
->lock
);
816 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
817 spin_unlock(&mctz
->lock
);
822 * Implementation Note: reading percpu statistics for memcg.
824 * Both of vmstat[] and percpu_counter has threshold and do periodic
825 * synchronization to implement "quick" read. There are trade-off between
826 * reading cost and precision of value. Then, we may have a chance to implement
827 * a periodic synchronizion of counter in memcg's counter.
829 * But this _read() function is used for user interface now. The user accounts
830 * memory usage by memory cgroup and he _always_ requires exact value because
831 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
832 * have to visit all online cpus and make sum. So, for now, unnecessary
833 * synchronization is not implemented. (just implemented for cpu hotplug)
835 * If there are kernel internal actions which can make use of some not-exact
836 * value, and reading all cpu value can be performance bottleneck in some
837 * common workload, threashold and synchonization as vmstat[] should be
840 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
841 enum mem_cgroup_stat_index idx
)
847 for_each_online_cpu(cpu
)
848 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
849 #ifdef CONFIG_HOTPLUG_CPU
850 spin_lock(&memcg
->pcp_counter_lock
);
851 val
+= memcg
->nocpu_base
.count
[idx
];
852 spin_unlock(&memcg
->pcp_counter_lock
);
858 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
861 int val
= (charge
) ? 1 : -1;
862 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
865 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
866 enum mem_cgroup_events_index idx
)
868 unsigned long val
= 0;
872 for_each_online_cpu(cpu
)
873 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
874 #ifdef CONFIG_HOTPLUG_CPU
875 spin_lock(&memcg
->pcp_counter_lock
);
876 val
+= memcg
->nocpu_base
.events
[idx
];
877 spin_unlock(&memcg
->pcp_counter_lock
);
883 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
885 bool anon
, int nr_pages
)
890 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
891 * counted as CACHE even if it's on ANON LRU.
894 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
897 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
900 if (PageTransHuge(page
))
901 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
904 /* pagein of a big page is an event. So, ignore page size */
906 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
908 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
909 nr_pages
= -nr_pages
; /* for event */
912 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
918 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
920 struct mem_cgroup_per_zone
*mz
;
922 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
923 return mz
->lru_size
[lru
];
927 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
928 unsigned int lru_mask
)
930 struct mem_cgroup_per_zone
*mz
;
932 unsigned long ret
= 0;
934 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
937 if (BIT(lru
) & lru_mask
)
938 ret
+= mz
->lru_size
[lru
];
944 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
945 int nid
, unsigned int lru_mask
)
950 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
951 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
957 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
958 unsigned int lru_mask
)
963 for_each_node_state(nid
, N_MEMORY
)
964 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
968 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
969 enum mem_cgroup_events_target target
)
971 unsigned long val
, next
;
973 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
974 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
975 /* from time_after() in jiffies.h */
976 if ((long)next
- (long)val
< 0) {
978 case MEM_CGROUP_TARGET_THRESH
:
979 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
981 case MEM_CGROUP_TARGET_SOFTLIMIT
:
982 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
984 case MEM_CGROUP_TARGET_NUMAINFO
:
985 next
= val
+ NUMAINFO_EVENTS_TARGET
;
990 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
997 * Check events in order.
1000 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1003 /* threshold event is triggered in finer grain than soft limit */
1004 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1005 MEM_CGROUP_TARGET_THRESH
))) {
1007 bool do_numainfo __maybe_unused
;
1009 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1010 MEM_CGROUP_TARGET_SOFTLIMIT
);
1011 #if MAX_NUMNODES > 1
1012 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1013 MEM_CGROUP_TARGET_NUMAINFO
);
1017 mem_cgroup_threshold(memcg
);
1018 if (unlikely(do_softlimit
))
1019 mem_cgroup_update_tree(memcg
, page
);
1020 #if MAX_NUMNODES > 1
1021 if (unlikely(do_numainfo
))
1022 atomic_inc(&memcg
->numainfo_events
);
1028 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1031 * mm_update_next_owner() may clear mm->owner to NULL
1032 * if it races with swapoff, page migration, etc.
1033 * So this can be called with p == NULL.
1038 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
1041 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1043 struct mem_cgroup
*memcg
= NULL
;
1048 * Because we have no locks, mm->owner's may be being moved to other
1049 * cgroup. We use css_tryget() here even if this looks
1050 * pessimistic (rather than adding locks here).
1054 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1055 if (unlikely(!memcg
))
1057 } while (!css_tryget(&memcg
->css
));
1063 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1064 * ref. count) or NULL if the whole root's subtree has been visited.
1066 * helper function to be used by mem_cgroup_iter
1068 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1069 struct mem_cgroup
*last_visited
)
1071 struct cgroup_subsys_state
*prev_css
, *next_css
;
1073 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1075 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1078 * Even if we found a group we have to make sure it is
1079 * alive. css && !memcg means that the groups should be
1080 * skipped and we should continue the tree walk.
1081 * last_visited css is safe to use because it is
1082 * protected by css_get and the tree walk is rcu safe.
1085 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1087 if (css_tryget(&mem
->css
))
1090 prev_css
= next_css
;
1098 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1101 * When a group in the hierarchy below root is destroyed, the
1102 * hierarchy iterator can no longer be trusted since it might
1103 * have pointed to the destroyed group. Invalidate it.
1105 atomic_inc(&root
->dead_count
);
1108 static struct mem_cgroup
*
1109 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1110 struct mem_cgroup
*root
,
1113 struct mem_cgroup
*position
= NULL
;
1115 * A cgroup destruction happens in two stages: offlining and
1116 * release. They are separated by a RCU grace period.
1118 * If the iterator is valid, we may still race with an
1119 * offlining. The RCU lock ensures the object won't be
1120 * released, tryget will fail if we lost the race.
1122 *sequence
= atomic_read(&root
->dead_count
);
1123 if (iter
->last_dead_count
== *sequence
) {
1125 position
= iter
->last_visited
;
1126 if (position
&& !css_tryget(&position
->css
))
1132 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1133 struct mem_cgroup
*last_visited
,
1134 struct mem_cgroup
*new_position
,
1138 css_put(&last_visited
->css
);
1140 * We store the sequence count from the time @last_visited was
1141 * loaded successfully instead of rereading it here so that we
1142 * don't lose destruction events in between. We could have
1143 * raced with the destruction of @new_position after all.
1145 iter
->last_visited
= new_position
;
1147 iter
->last_dead_count
= sequence
;
1151 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1152 * @root: hierarchy root
1153 * @prev: previously returned memcg, NULL on first invocation
1154 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1156 * Returns references to children of the hierarchy below @root, or
1157 * @root itself, or %NULL after a full round-trip.
1159 * Caller must pass the return value in @prev on subsequent
1160 * invocations for reference counting, or use mem_cgroup_iter_break()
1161 * to cancel a hierarchy walk before the round-trip is complete.
1163 * Reclaimers can specify a zone and a priority level in @reclaim to
1164 * divide up the memcgs in the hierarchy among all concurrent
1165 * reclaimers operating on the same zone and priority.
1167 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1168 struct mem_cgroup
*prev
,
1169 struct mem_cgroup_reclaim_cookie
*reclaim
)
1171 struct mem_cgroup
*memcg
= NULL
;
1172 struct mem_cgroup
*last_visited
= NULL
;
1174 if (mem_cgroup_disabled())
1178 root
= root_mem_cgroup
;
1180 if (prev
&& !reclaim
)
1181 last_visited
= prev
;
1183 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1191 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1192 int uninitialized_var(seq
);
1195 int nid
= zone_to_nid(reclaim
->zone
);
1196 int zid
= zone_idx(reclaim
->zone
);
1197 struct mem_cgroup_per_zone
*mz
;
1199 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1200 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1201 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1202 iter
->last_visited
= NULL
;
1206 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1209 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1212 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1216 else if (!prev
&& memcg
)
1217 reclaim
->generation
= iter
->generation
;
1226 if (prev
&& prev
!= root
)
1227 css_put(&prev
->css
);
1233 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1234 * @root: hierarchy root
1235 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1237 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1238 struct mem_cgroup
*prev
)
1241 root
= root_mem_cgroup
;
1242 if (prev
&& prev
!= root
)
1243 css_put(&prev
->css
);
1247 * Iteration constructs for visiting all cgroups (under a tree). If
1248 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1249 * be used for reference counting.
1251 #define for_each_mem_cgroup_tree(iter, root) \
1252 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1254 iter = mem_cgroup_iter(root, iter, NULL))
1256 #define for_each_mem_cgroup(iter) \
1257 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1259 iter = mem_cgroup_iter(NULL, iter, NULL))
1261 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1263 struct mem_cgroup
*memcg
;
1266 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1267 if (unlikely(!memcg
))
1272 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1275 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1283 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1286 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1287 * @zone: zone of the wanted lruvec
1288 * @memcg: memcg of the wanted lruvec
1290 * Returns the lru list vector holding pages for the given @zone and
1291 * @mem. This can be the global zone lruvec, if the memory controller
1294 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1295 struct mem_cgroup
*memcg
)
1297 struct mem_cgroup_per_zone
*mz
;
1298 struct lruvec
*lruvec
;
1300 if (mem_cgroup_disabled()) {
1301 lruvec
= &zone
->lruvec
;
1305 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1306 lruvec
= &mz
->lruvec
;
1309 * Since a node can be onlined after the mem_cgroup was created,
1310 * we have to be prepared to initialize lruvec->zone here;
1311 * and if offlined then reonlined, we need to reinitialize it.
1313 if (unlikely(lruvec
->zone
!= zone
))
1314 lruvec
->zone
= zone
;
1319 * Following LRU functions are allowed to be used without PCG_LOCK.
1320 * Operations are called by routine of global LRU independently from memcg.
1321 * What we have to take care of here is validness of pc->mem_cgroup.
1323 * Changes to pc->mem_cgroup happens when
1326 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1327 * It is added to LRU before charge.
1328 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1329 * When moving account, the page is not on LRU. It's isolated.
1333 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1335 * @zone: zone of the page
1337 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1339 struct mem_cgroup_per_zone
*mz
;
1340 struct mem_cgroup
*memcg
;
1341 struct page_cgroup
*pc
;
1342 struct lruvec
*lruvec
;
1344 if (mem_cgroup_disabled()) {
1345 lruvec
= &zone
->lruvec
;
1349 pc
= lookup_page_cgroup(page
);
1350 memcg
= pc
->mem_cgroup
;
1353 * Surreptitiously switch any uncharged offlist page to root:
1354 * an uncharged page off lru does nothing to secure
1355 * its former mem_cgroup from sudden removal.
1357 * Our caller holds lru_lock, and PageCgroupUsed is updated
1358 * under page_cgroup lock: between them, they make all uses
1359 * of pc->mem_cgroup safe.
1361 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1362 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1364 mz
= page_cgroup_zoneinfo(memcg
, page
);
1365 lruvec
= &mz
->lruvec
;
1368 * Since a node can be onlined after the mem_cgroup was created,
1369 * we have to be prepared to initialize lruvec->zone here;
1370 * and if offlined then reonlined, we need to reinitialize it.
1372 if (unlikely(lruvec
->zone
!= zone
))
1373 lruvec
->zone
= zone
;
1378 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1379 * @lruvec: mem_cgroup per zone lru vector
1380 * @lru: index of lru list the page is sitting on
1381 * @nr_pages: positive when adding or negative when removing
1383 * This function must be called when a page is added to or removed from an
1386 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1389 struct mem_cgroup_per_zone
*mz
;
1390 unsigned long *lru_size
;
1392 if (mem_cgroup_disabled())
1395 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1396 lru_size
= mz
->lru_size
+ lru
;
1397 *lru_size
+= nr_pages
;
1398 VM_BUG_ON((long)(*lru_size
) < 0);
1402 * Checks whether given mem is same or in the root_mem_cgroup's
1405 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1406 struct mem_cgroup
*memcg
)
1408 if (root_memcg
== memcg
)
1410 if (!root_memcg
->use_hierarchy
|| !memcg
)
1412 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1415 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1416 struct mem_cgroup
*memcg
)
1421 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1426 bool task_in_mem_cgroup(struct task_struct
*task
,
1427 const struct mem_cgroup
*memcg
)
1429 struct mem_cgroup
*curr
= NULL
;
1430 struct task_struct
*p
;
1433 p
= find_lock_task_mm(task
);
1435 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1439 * All threads may have already detached their mm's, but the oom
1440 * killer still needs to detect if they have already been oom
1441 * killed to prevent needlessly killing additional tasks.
1444 curr
= mem_cgroup_from_task(task
);
1446 css_get(&curr
->css
);
1452 * We should check use_hierarchy of "memcg" not "curr". Because checking
1453 * use_hierarchy of "curr" here make this function true if hierarchy is
1454 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1455 * hierarchy(even if use_hierarchy is disabled in "memcg").
1457 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1458 css_put(&curr
->css
);
1462 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1464 unsigned long inactive_ratio
;
1465 unsigned long inactive
;
1466 unsigned long active
;
1469 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1470 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1472 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1474 inactive_ratio
= int_sqrt(10 * gb
);
1478 return inactive
* inactive_ratio
< active
;
1481 #define mem_cgroup_from_res_counter(counter, member) \
1482 container_of(counter, struct mem_cgroup, member)
1485 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1486 * @memcg: the memory cgroup
1488 * Returns the maximum amount of memory @mem can be charged with, in
1491 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1493 unsigned long long margin
;
1495 margin
= res_counter_margin(&memcg
->res
);
1496 if (do_swap_account
)
1497 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1498 return margin
>> PAGE_SHIFT
;
1501 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1504 if (!css_parent(&memcg
->css
))
1505 return vm_swappiness
;
1507 return memcg
->swappiness
;
1511 * memcg->moving_account is used for checking possibility that some thread is
1512 * calling move_account(). When a thread on CPU-A starts moving pages under
1513 * a memcg, other threads should check memcg->moving_account under
1514 * rcu_read_lock(), like this:
1518 * memcg->moving_account+1 if (memcg->mocing_account)
1520 * synchronize_rcu() update something.
1525 /* for quick checking without looking up memcg */
1526 atomic_t memcg_moving __read_mostly
;
1528 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1530 atomic_inc(&memcg_moving
);
1531 atomic_inc(&memcg
->moving_account
);
1535 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1538 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1539 * We check NULL in callee rather than caller.
1542 atomic_dec(&memcg_moving
);
1543 atomic_dec(&memcg
->moving_account
);
1548 * 2 routines for checking "mem" is under move_account() or not.
1550 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1551 * is used for avoiding races in accounting. If true,
1552 * pc->mem_cgroup may be overwritten.
1554 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1555 * under hierarchy of moving cgroups. This is for
1556 * waiting at hith-memory prressure caused by "move".
1559 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1561 VM_BUG_ON(!rcu_read_lock_held());
1562 return atomic_read(&memcg
->moving_account
) > 0;
1565 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1567 struct mem_cgroup
*from
;
1568 struct mem_cgroup
*to
;
1571 * Unlike task_move routines, we access mc.to, mc.from not under
1572 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1574 spin_lock(&mc
.lock
);
1580 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1581 || mem_cgroup_same_or_subtree(memcg
, to
);
1583 spin_unlock(&mc
.lock
);
1587 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1589 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1590 if (mem_cgroup_under_move(memcg
)) {
1592 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1593 /* moving charge context might have finished. */
1596 finish_wait(&mc
.waitq
, &wait
);
1604 * Take this lock when
1605 * - a code tries to modify page's memcg while it's USED.
1606 * - a code tries to modify page state accounting in a memcg.
1607 * see mem_cgroup_stolen(), too.
1609 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1610 unsigned long *flags
)
1612 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1615 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1616 unsigned long *flags
)
1618 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1621 #define K(x) ((x) << (PAGE_SHIFT-10))
1623 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1624 * @memcg: The memory cgroup that went over limit
1625 * @p: Task that is going to be killed
1627 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1630 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1632 struct cgroup
*task_cgrp
;
1633 struct cgroup
*mem_cgrp
;
1635 * Need a buffer in BSS, can't rely on allocations. The code relies
1636 * on the assumption that OOM is serialized for memory controller.
1637 * If this assumption is broken, revisit this code.
1639 static char memcg_name
[PATH_MAX
];
1641 struct mem_cgroup
*iter
;
1649 mem_cgrp
= memcg
->css
.cgroup
;
1650 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1652 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1655 * Unfortunately, we are unable to convert to a useful name
1656 * But we'll still print out the usage information
1663 pr_info("Task in %s killed", memcg_name
);
1666 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1674 * Continues from above, so we don't need an KERN_ level
1676 pr_cont(" as a result of limit of %s\n", memcg_name
);
1679 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1680 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1681 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1682 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1683 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1684 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1685 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1686 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1687 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1688 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1689 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1690 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1692 for_each_mem_cgroup_tree(iter
, memcg
) {
1693 pr_info("Memory cgroup stats");
1696 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1698 pr_cont(" for %s", memcg_name
);
1702 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1703 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1705 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1706 K(mem_cgroup_read_stat(iter
, i
)));
1709 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1710 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1711 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1718 * This function returns the number of memcg under hierarchy tree. Returns
1719 * 1(self count) if no children.
1721 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1724 struct mem_cgroup
*iter
;
1726 for_each_mem_cgroup_tree(iter
, memcg
)
1732 * Return the memory (and swap, if configured) limit for a memcg.
1734 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1738 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1741 * Do not consider swap space if we cannot swap due to swappiness
1743 if (mem_cgroup_swappiness(memcg
)) {
1746 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1747 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1750 * If memsw is finite and limits the amount of swap space
1751 * available to this memcg, return that limit.
1753 limit
= min(limit
, memsw
);
1759 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1762 struct mem_cgroup
*iter
;
1763 unsigned long chosen_points
= 0;
1764 unsigned long totalpages
;
1765 unsigned int points
= 0;
1766 struct task_struct
*chosen
= NULL
;
1769 * If current has a pending SIGKILL or is exiting, then automatically
1770 * select it. The goal is to allow it to allocate so that it may
1771 * quickly exit and free its memory.
1773 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1774 set_thread_flag(TIF_MEMDIE
);
1778 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1779 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1780 for_each_mem_cgroup_tree(iter
, memcg
) {
1781 struct css_task_iter it
;
1782 struct task_struct
*task
;
1784 css_task_iter_start(&iter
->css
, &it
);
1785 while ((task
= css_task_iter_next(&it
))) {
1786 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1788 case OOM_SCAN_SELECT
:
1790 put_task_struct(chosen
);
1792 chosen_points
= ULONG_MAX
;
1793 get_task_struct(chosen
);
1795 case OOM_SCAN_CONTINUE
:
1797 case OOM_SCAN_ABORT
:
1798 css_task_iter_end(&it
);
1799 mem_cgroup_iter_break(memcg
, iter
);
1801 put_task_struct(chosen
);
1806 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1807 if (points
> chosen_points
) {
1809 put_task_struct(chosen
);
1811 chosen_points
= points
;
1812 get_task_struct(chosen
);
1815 css_task_iter_end(&it
);
1820 points
= chosen_points
* 1000 / totalpages
;
1821 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1822 NULL
, "Memory cgroup out of memory");
1825 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1827 unsigned long flags
)
1829 unsigned long total
= 0;
1830 bool noswap
= false;
1833 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1835 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1838 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1840 drain_all_stock_async(memcg
);
1841 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1843 * Allow limit shrinkers, which are triggered directly
1844 * by userspace, to catch signals and stop reclaim
1845 * after minimal progress, regardless of the margin.
1847 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1849 if (mem_cgroup_margin(memcg
))
1852 * If nothing was reclaimed after two attempts, there
1853 * may be no reclaimable pages in this hierarchy.
1862 * test_mem_cgroup_node_reclaimable
1863 * @memcg: the target memcg
1864 * @nid: the node ID to be checked.
1865 * @noswap : specify true here if the user wants flle only information.
1867 * This function returns whether the specified memcg contains any
1868 * reclaimable pages on a node. Returns true if there are any reclaimable
1869 * pages in the node.
1871 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1872 int nid
, bool noswap
)
1874 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1876 if (noswap
|| !total_swap_pages
)
1878 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1883 #if MAX_NUMNODES > 1
1886 * Always updating the nodemask is not very good - even if we have an empty
1887 * list or the wrong list here, we can start from some node and traverse all
1888 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1891 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1895 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1896 * pagein/pageout changes since the last update.
1898 if (!atomic_read(&memcg
->numainfo_events
))
1900 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1903 /* make a nodemask where this memcg uses memory from */
1904 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1906 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1908 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1909 node_clear(nid
, memcg
->scan_nodes
);
1912 atomic_set(&memcg
->numainfo_events
, 0);
1913 atomic_set(&memcg
->numainfo_updating
, 0);
1917 * Selecting a node where we start reclaim from. Because what we need is just
1918 * reducing usage counter, start from anywhere is O,K. Considering
1919 * memory reclaim from current node, there are pros. and cons.
1921 * Freeing memory from current node means freeing memory from a node which
1922 * we'll use or we've used. So, it may make LRU bad. And if several threads
1923 * hit limits, it will see a contention on a node. But freeing from remote
1924 * node means more costs for memory reclaim because of memory latency.
1926 * Now, we use round-robin. Better algorithm is welcomed.
1928 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1932 mem_cgroup_may_update_nodemask(memcg
);
1933 node
= memcg
->last_scanned_node
;
1935 node
= next_node(node
, memcg
->scan_nodes
);
1936 if (node
== MAX_NUMNODES
)
1937 node
= first_node(memcg
->scan_nodes
);
1939 * We call this when we hit limit, not when pages are added to LRU.
1940 * No LRU may hold pages because all pages are UNEVICTABLE or
1941 * memcg is too small and all pages are not on LRU. In that case,
1942 * we use curret node.
1944 if (unlikely(node
== MAX_NUMNODES
))
1945 node
= numa_node_id();
1947 memcg
->last_scanned_node
= node
;
1952 * Check all nodes whether it contains reclaimable pages or not.
1953 * For quick scan, we make use of scan_nodes. This will allow us to skip
1954 * unused nodes. But scan_nodes is lazily updated and may not cotain
1955 * enough new information. We need to do double check.
1957 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1962 * quick check...making use of scan_node.
1963 * We can skip unused nodes.
1965 if (!nodes_empty(memcg
->scan_nodes
)) {
1966 for (nid
= first_node(memcg
->scan_nodes
);
1968 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1970 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1975 * Check rest of nodes.
1977 for_each_node_state(nid
, N_MEMORY
) {
1978 if (node_isset(nid
, memcg
->scan_nodes
))
1980 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1987 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1992 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1994 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1998 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
2001 unsigned long *total_scanned
)
2003 struct mem_cgroup
*victim
= NULL
;
2006 unsigned long excess
;
2007 unsigned long nr_scanned
;
2008 struct mem_cgroup_reclaim_cookie reclaim
= {
2013 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2016 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2021 * If we have not been able to reclaim
2022 * anything, it might because there are
2023 * no reclaimable pages under this hierarchy
2028 * We want to do more targeted reclaim.
2029 * excess >> 2 is not to excessive so as to
2030 * reclaim too much, nor too less that we keep
2031 * coming back to reclaim from this cgroup
2033 if (total
>= (excess
>> 2) ||
2034 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2039 if (!mem_cgroup_reclaimable(victim
, false))
2041 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2043 *total_scanned
+= nr_scanned
;
2044 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2047 mem_cgroup_iter_break(root_memcg
, victim
);
2051 #ifdef CONFIG_LOCKDEP
2052 static struct lockdep_map memcg_oom_lock_dep_map
= {
2053 .name
= "memcg_oom_lock",
2057 static DEFINE_SPINLOCK(memcg_oom_lock
);
2060 * Check OOM-Killer is already running under our hierarchy.
2061 * If someone is running, return false.
2063 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2065 struct mem_cgroup
*iter
, *failed
= NULL
;
2067 spin_lock(&memcg_oom_lock
);
2069 for_each_mem_cgroup_tree(iter
, memcg
) {
2070 if (iter
->oom_lock
) {
2072 * this subtree of our hierarchy is already locked
2073 * so we cannot give a lock.
2076 mem_cgroup_iter_break(memcg
, iter
);
2079 iter
->oom_lock
= true;
2084 * OK, we failed to lock the whole subtree so we have
2085 * to clean up what we set up to the failing subtree
2087 for_each_mem_cgroup_tree(iter
, memcg
) {
2088 if (iter
== failed
) {
2089 mem_cgroup_iter_break(memcg
, iter
);
2092 iter
->oom_lock
= false;
2095 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2097 spin_unlock(&memcg_oom_lock
);
2102 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2104 struct mem_cgroup
*iter
;
2106 spin_lock(&memcg_oom_lock
);
2107 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2108 for_each_mem_cgroup_tree(iter
, memcg
)
2109 iter
->oom_lock
= false;
2110 spin_unlock(&memcg_oom_lock
);
2113 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2115 struct mem_cgroup
*iter
;
2117 for_each_mem_cgroup_tree(iter
, memcg
)
2118 atomic_inc(&iter
->under_oom
);
2121 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2123 struct mem_cgroup
*iter
;
2126 * When a new child is created while the hierarchy is under oom,
2127 * mem_cgroup_oom_lock() may not be called. We have to use
2128 * atomic_add_unless() here.
2130 for_each_mem_cgroup_tree(iter
, memcg
)
2131 atomic_add_unless(&iter
->under_oom
, -1, 0);
2134 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2136 struct oom_wait_info
{
2137 struct mem_cgroup
*memcg
;
2141 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2142 unsigned mode
, int sync
, void *arg
)
2144 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2145 struct mem_cgroup
*oom_wait_memcg
;
2146 struct oom_wait_info
*oom_wait_info
;
2148 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2149 oom_wait_memcg
= oom_wait_info
->memcg
;
2152 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2153 * Then we can use css_is_ancestor without taking care of RCU.
2155 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2156 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2158 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2161 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2163 atomic_inc(&memcg
->oom_wakeups
);
2164 /* for filtering, pass "memcg" as argument. */
2165 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2168 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2170 if (memcg
&& atomic_read(&memcg
->under_oom
))
2171 memcg_wakeup_oom(memcg
);
2174 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2176 if (!current
->memcg_oom
.may_oom
)
2179 * We are in the middle of the charge context here, so we
2180 * don't want to block when potentially sitting on a callstack
2181 * that holds all kinds of filesystem and mm locks.
2183 * Also, the caller may handle a failed allocation gracefully
2184 * (like optional page cache readahead) and so an OOM killer
2185 * invocation might not even be necessary.
2187 * That's why we don't do anything here except remember the
2188 * OOM context and then deal with it at the end of the page
2189 * fault when the stack is unwound, the locks are released,
2190 * and when we know whether the fault was overall successful.
2192 css_get(&memcg
->css
);
2193 current
->memcg_oom
.memcg
= memcg
;
2194 current
->memcg_oom
.gfp_mask
= mask
;
2195 current
->memcg_oom
.order
= order
;
2199 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2200 * @handle: actually kill/wait or just clean up the OOM state
2202 * This has to be called at the end of a page fault if the memcg OOM
2203 * handler was enabled.
2205 * Memcg supports userspace OOM handling where failed allocations must
2206 * sleep on a waitqueue until the userspace task resolves the
2207 * situation. Sleeping directly in the charge context with all kinds
2208 * of locks held is not a good idea, instead we remember an OOM state
2209 * in the task and mem_cgroup_oom_synchronize() has to be called at
2210 * the end of the page fault to complete the OOM handling.
2212 * Returns %true if an ongoing memcg OOM situation was detected and
2213 * completed, %false otherwise.
2215 bool mem_cgroup_oom_synchronize(bool handle
)
2217 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2218 struct oom_wait_info owait
;
2221 /* OOM is global, do not handle */
2228 owait
.memcg
= memcg
;
2229 owait
.wait
.flags
= 0;
2230 owait
.wait
.func
= memcg_oom_wake_function
;
2231 owait
.wait
.private = current
;
2232 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2234 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2235 mem_cgroup_mark_under_oom(memcg
);
2237 locked
= mem_cgroup_oom_trylock(memcg
);
2240 mem_cgroup_oom_notify(memcg
);
2242 if (locked
&& !memcg
->oom_kill_disable
) {
2243 mem_cgroup_unmark_under_oom(memcg
);
2244 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2245 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2246 current
->memcg_oom
.order
);
2249 mem_cgroup_unmark_under_oom(memcg
);
2250 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2254 mem_cgroup_oom_unlock(memcg
);
2256 * There is no guarantee that an OOM-lock contender
2257 * sees the wakeups triggered by the OOM kill
2258 * uncharges. Wake any sleepers explicitely.
2260 memcg_oom_recover(memcg
);
2263 current
->memcg_oom
.memcg
= NULL
;
2264 css_put(&memcg
->css
);
2269 * Currently used to update mapped file statistics, but the routine can be
2270 * generalized to update other statistics as well.
2272 * Notes: Race condition
2274 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2275 * it tends to be costly. But considering some conditions, we doesn't need
2276 * to do so _always_.
2278 * Considering "charge", lock_page_cgroup() is not required because all
2279 * file-stat operations happen after a page is attached to radix-tree. There
2280 * are no race with "charge".
2282 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2283 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2284 * if there are race with "uncharge". Statistics itself is properly handled
2287 * Considering "move", this is an only case we see a race. To make the race
2288 * small, we check mm->moving_account and detect there are possibility of race
2289 * If there is, we take a lock.
2292 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2293 bool *locked
, unsigned long *flags
)
2295 struct mem_cgroup
*memcg
;
2296 struct page_cgroup
*pc
;
2298 pc
= lookup_page_cgroup(page
);
2300 memcg
= pc
->mem_cgroup
;
2301 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2304 * If this memory cgroup is not under account moving, we don't
2305 * need to take move_lock_mem_cgroup(). Because we already hold
2306 * rcu_read_lock(), any calls to move_account will be delayed until
2307 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2309 if (!mem_cgroup_stolen(memcg
))
2312 move_lock_mem_cgroup(memcg
, flags
);
2313 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2314 move_unlock_mem_cgroup(memcg
, flags
);
2320 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2322 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2325 * It's guaranteed that pc->mem_cgroup never changes while
2326 * lock is held because a routine modifies pc->mem_cgroup
2327 * should take move_lock_mem_cgroup().
2329 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2332 void mem_cgroup_update_page_stat(struct page
*page
,
2333 enum mem_cgroup_stat_index idx
, int val
)
2335 struct mem_cgroup
*memcg
;
2336 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2337 unsigned long uninitialized_var(flags
);
2339 if (mem_cgroup_disabled())
2342 VM_BUG_ON(!rcu_read_lock_held());
2343 memcg
= pc
->mem_cgroup
;
2344 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2347 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2351 * size of first charge trial. "32" comes from vmscan.c's magic value.
2352 * TODO: maybe necessary to use big numbers in big irons.
2354 #define CHARGE_BATCH 32U
2355 struct memcg_stock_pcp
{
2356 struct mem_cgroup
*cached
; /* this never be root cgroup */
2357 unsigned int nr_pages
;
2358 struct work_struct work
;
2359 unsigned long flags
;
2360 #define FLUSHING_CACHED_CHARGE 0
2362 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2363 static DEFINE_MUTEX(percpu_charge_mutex
);
2366 * consume_stock: Try to consume stocked charge on this cpu.
2367 * @memcg: memcg to consume from.
2368 * @nr_pages: how many pages to charge.
2370 * The charges will only happen if @memcg matches the current cpu's memcg
2371 * stock, and at least @nr_pages are available in that stock. Failure to
2372 * service an allocation will refill the stock.
2374 * returns true if successful, false otherwise.
2376 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2378 struct memcg_stock_pcp
*stock
;
2381 if (nr_pages
> CHARGE_BATCH
)
2384 stock
= &get_cpu_var(memcg_stock
);
2385 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2386 stock
->nr_pages
-= nr_pages
;
2387 else /* need to call res_counter_charge */
2389 put_cpu_var(memcg_stock
);
2394 * Returns stocks cached in percpu to res_counter and reset cached information.
2396 static void drain_stock(struct memcg_stock_pcp
*stock
)
2398 struct mem_cgroup
*old
= stock
->cached
;
2400 if (stock
->nr_pages
) {
2401 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2403 res_counter_uncharge(&old
->res
, bytes
);
2404 if (do_swap_account
)
2405 res_counter_uncharge(&old
->memsw
, bytes
);
2406 stock
->nr_pages
= 0;
2408 stock
->cached
= NULL
;
2412 * This must be called under preempt disabled or must be called by
2413 * a thread which is pinned to local cpu.
2415 static void drain_local_stock(struct work_struct
*dummy
)
2417 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2419 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2422 static void __init
memcg_stock_init(void)
2426 for_each_possible_cpu(cpu
) {
2427 struct memcg_stock_pcp
*stock
=
2428 &per_cpu(memcg_stock
, cpu
);
2429 INIT_WORK(&stock
->work
, drain_local_stock
);
2434 * Cache charges(val) which is from res_counter, to local per_cpu area.
2435 * This will be consumed by consume_stock() function, later.
2437 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2439 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2441 if (stock
->cached
!= memcg
) { /* reset if necessary */
2443 stock
->cached
= memcg
;
2445 stock
->nr_pages
+= nr_pages
;
2446 put_cpu_var(memcg_stock
);
2450 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2451 * of the hierarchy under it. sync flag says whether we should block
2452 * until the work is done.
2454 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2458 /* Notify other cpus that system-wide "drain" is running */
2461 for_each_online_cpu(cpu
) {
2462 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2463 struct mem_cgroup
*memcg
;
2465 memcg
= stock
->cached
;
2466 if (!memcg
|| !stock
->nr_pages
)
2468 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2470 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2472 drain_local_stock(&stock
->work
);
2474 schedule_work_on(cpu
, &stock
->work
);
2482 for_each_online_cpu(cpu
) {
2483 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2484 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2485 flush_work(&stock
->work
);
2492 * Tries to drain stocked charges in other cpus. This function is asynchronous
2493 * and just put a work per cpu for draining localy on each cpu. Caller can
2494 * expects some charges will be back to res_counter later but cannot wait for
2497 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2500 * If someone calls draining, avoid adding more kworker runs.
2502 if (!mutex_trylock(&percpu_charge_mutex
))
2504 drain_all_stock(root_memcg
, false);
2505 mutex_unlock(&percpu_charge_mutex
);
2508 /* This is a synchronous drain interface. */
2509 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2511 /* called when force_empty is called */
2512 mutex_lock(&percpu_charge_mutex
);
2513 drain_all_stock(root_memcg
, true);
2514 mutex_unlock(&percpu_charge_mutex
);
2518 * This function drains percpu counter value from DEAD cpu and
2519 * move it to local cpu. Note that this function can be preempted.
2521 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2525 spin_lock(&memcg
->pcp_counter_lock
);
2526 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2527 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2529 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2530 memcg
->nocpu_base
.count
[i
] += x
;
2532 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2533 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2535 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2536 memcg
->nocpu_base
.events
[i
] += x
;
2538 spin_unlock(&memcg
->pcp_counter_lock
);
2541 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2542 unsigned long action
,
2545 int cpu
= (unsigned long)hcpu
;
2546 struct memcg_stock_pcp
*stock
;
2547 struct mem_cgroup
*iter
;
2549 if (action
== CPU_ONLINE
)
2552 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2555 for_each_mem_cgroup(iter
)
2556 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2558 stock
= &per_cpu(memcg_stock
, cpu
);
2564 /* See __mem_cgroup_try_charge() for details */
2566 CHARGE_OK
, /* success */
2567 CHARGE_RETRY
, /* need to retry but retry is not bad */
2568 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2569 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2572 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2573 unsigned int nr_pages
, unsigned int min_pages
,
2576 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2577 struct mem_cgroup
*mem_over_limit
;
2578 struct res_counter
*fail_res
;
2579 unsigned long flags
= 0;
2582 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2585 if (!do_swap_account
)
2587 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2591 res_counter_uncharge(&memcg
->res
, csize
);
2592 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2593 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2595 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2597 * Never reclaim on behalf of optional batching, retry with a
2598 * single page instead.
2600 if (nr_pages
> min_pages
)
2601 return CHARGE_RETRY
;
2603 if (!(gfp_mask
& __GFP_WAIT
))
2604 return CHARGE_WOULDBLOCK
;
2606 if (gfp_mask
& __GFP_NORETRY
)
2607 return CHARGE_NOMEM
;
2609 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2610 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2611 return CHARGE_RETRY
;
2613 * Even though the limit is exceeded at this point, reclaim
2614 * may have been able to free some pages. Retry the charge
2615 * before killing the task.
2617 * Only for regular pages, though: huge pages are rather
2618 * unlikely to succeed so close to the limit, and we fall back
2619 * to regular pages anyway in case of failure.
2621 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2622 return CHARGE_RETRY
;
2625 * At task move, charge accounts can be doubly counted. So, it's
2626 * better to wait until the end of task_move if something is going on.
2628 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2629 return CHARGE_RETRY
;
2632 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2634 return CHARGE_NOMEM
;
2638 * __mem_cgroup_try_charge() does
2639 * 1. detect memcg to be charged against from passed *mm and *ptr,
2640 * 2. update res_counter
2641 * 3. call memory reclaim if necessary.
2643 * In some special case, if the task is fatal, fatal_signal_pending() or
2644 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2645 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2646 * as possible without any hazards. 2: all pages should have a valid
2647 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2648 * pointer, that is treated as a charge to root_mem_cgroup.
2650 * So __mem_cgroup_try_charge() will return
2651 * 0 ... on success, filling *ptr with a valid memcg pointer.
2652 * -ENOMEM ... charge failure because of resource limits.
2653 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2655 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2656 * the oom-killer can be invoked.
2658 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2660 unsigned int nr_pages
,
2661 struct mem_cgroup
**ptr
,
2664 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2665 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2666 struct mem_cgroup
*memcg
= NULL
;
2670 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2671 * in system level. So, allow to go ahead dying process in addition to
2674 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2675 || fatal_signal_pending(current
)))
2678 if (unlikely(task_in_memcg_oom(current
)))
2682 * We always charge the cgroup the mm_struct belongs to.
2683 * The mm_struct's mem_cgroup changes on task migration if the
2684 * thread group leader migrates. It's possible that mm is not
2685 * set, if so charge the root memcg (happens for pagecache usage).
2688 *ptr
= root_mem_cgroup
;
2690 if (*ptr
) { /* css should be a valid one */
2692 if (mem_cgroup_is_root(memcg
))
2694 if (consume_stock(memcg
, nr_pages
))
2696 css_get(&memcg
->css
);
2698 struct task_struct
*p
;
2701 p
= rcu_dereference(mm
->owner
);
2703 * Because we don't have task_lock(), "p" can exit.
2704 * In that case, "memcg" can point to root or p can be NULL with
2705 * race with swapoff. Then, we have small risk of mis-accouning.
2706 * But such kind of mis-account by race always happens because
2707 * we don't have cgroup_mutex(). It's overkill and we allo that
2709 * (*) swapoff at el will charge against mm-struct not against
2710 * task-struct. So, mm->owner can be NULL.
2712 memcg
= mem_cgroup_from_task(p
);
2714 memcg
= root_mem_cgroup
;
2715 if (mem_cgroup_is_root(memcg
)) {
2719 if (consume_stock(memcg
, nr_pages
)) {
2721 * It seems dagerous to access memcg without css_get().
2722 * But considering how consume_stok works, it's not
2723 * necessary. If consume_stock success, some charges
2724 * from this memcg are cached on this cpu. So, we
2725 * don't need to call css_get()/css_tryget() before
2726 * calling consume_stock().
2731 /* after here, we may be blocked. we need to get refcnt */
2732 if (!css_tryget(&memcg
->css
)) {
2740 bool invoke_oom
= oom
&& !nr_oom_retries
;
2742 /* If killed, bypass charge */
2743 if (fatal_signal_pending(current
)) {
2744 css_put(&memcg
->css
);
2748 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2749 nr_pages
, invoke_oom
);
2753 case CHARGE_RETRY
: /* not in OOM situation but retry */
2755 css_put(&memcg
->css
);
2758 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2759 css_put(&memcg
->css
);
2761 case CHARGE_NOMEM
: /* OOM routine works */
2762 if (!oom
|| invoke_oom
) {
2763 css_put(&memcg
->css
);
2769 } while (ret
!= CHARGE_OK
);
2771 if (batch
> nr_pages
)
2772 refill_stock(memcg
, batch
- nr_pages
);
2773 css_put(&memcg
->css
);
2778 if (!(gfp_mask
& __GFP_NOFAIL
)) {
2783 *ptr
= root_mem_cgroup
;
2788 * Somemtimes we have to undo a charge we got by try_charge().
2789 * This function is for that and do uncharge, put css's refcnt.
2790 * gotten by try_charge().
2792 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2793 unsigned int nr_pages
)
2795 if (!mem_cgroup_is_root(memcg
)) {
2796 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2798 res_counter_uncharge(&memcg
->res
, bytes
);
2799 if (do_swap_account
)
2800 res_counter_uncharge(&memcg
->memsw
, bytes
);
2805 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2806 * This is useful when moving usage to parent cgroup.
2808 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2809 unsigned int nr_pages
)
2811 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2813 if (mem_cgroup_is_root(memcg
))
2816 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2817 if (do_swap_account
)
2818 res_counter_uncharge_until(&memcg
->memsw
,
2819 memcg
->memsw
.parent
, bytes
);
2823 * A helper function to get mem_cgroup from ID. must be called under
2824 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2825 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2826 * called against removed memcg.)
2828 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2830 struct cgroup_subsys_state
*css
;
2832 /* ID 0 is unused ID */
2835 css
= css_lookup(&mem_cgroup_subsys
, id
);
2838 return mem_cgroup_from_css(css
);
2841 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2843 struct mem_cgroup
*memcg
= NULL
;
2844 struct page_cgroup
*pc
;
2848 VM_BUG_ON(!PageLocked(page
));
2850 pc
= lookup_page_cgroup(page
);
2851 lock_page_cgroup(pc
);
2852 if (PageCgroupUsed(pc
)) {
2853 memcg
= pc
->mem_cgroup
;
2854 if (memcg
&& !css_tryget(&memcg
->css
))
2856 } else if (PageSwapCache(page
)) {
2857 ent
.val
= page_private(page
);
2858 id
= lookup_swap_cgroup_id(ent
);
2860 memcg
= mem_cgroup_lookup(id
);
2861 if (memcg
&& !css_tryget(&memcg
->css
))
2865 unlock_page_cgroup(pc
);
2869 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2871 unsigned int nr_pages
,
2872 enum charge_type ctype
,
2875 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2876 struct zone
*uninitialized_var(zone
);
2877 struct lruvec
*lruvec
;
2878 bool was_on_lru
= false;
2881 lock_page_cgroup(pc
);
2882 VM_BUG_ON(PageCgroupUsed(pc
));
2884 * we don't need page_cgroup_lock about tail pages, becase they are not
2885 * accessed by any other context at this point.
2889 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2890 * may already be on some other mem_cgroup's LRU. Take care of it.
2893 zone
= page_zone(page
);
2894 spin_lock_irq(&zone
->lru_lock
);
2895 if (PageLRU(page
)) {
2896 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2898 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2903 pc
->mem_cgroup
= memcg
;
2905 * We access a page_cgroup asynchronously without lock_page_cgroup().
2906 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2907 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2908 * before USED bit, we need memory barrier here.
2909 * See mem_cgroup_add_lru_list(), etc.
2912 SetPageCgroupUsed(pc
);
2916 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2917 VM_BUG_ON(PageLRU(page
));
2919 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2921 spin_unlock_irq(&zone
->lru_lock
);
2924 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2929 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2930 unlock_page_cgroup(pc
);
2933 * "charge_statistics" updated event counter. Then, check it.
2934 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2935 * if they exceeds softlimit.
2937 memcg_check_events(memcg
, page
);
2940 static DEFINE_MUTEX(set_limit_mutex
);
2942 #ifdef CONFIG_MEMCG_KMEM
2943 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2945 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2946 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2950 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2951 * in the memcg_cache_params struct.
2953 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2955 struct kmem_cache
*cachep
;
2957 VM_BUG_ON(p
->is_root_cache
);
2958 cachep
= p
->root_cache
;
2959 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2962 #ifdef CONFIG_SLABINFO
2963 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2964 struct cftype
*cft
, struct seq_file
*m
)
2966 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2967 struct memcg_cache_params
*params
;
2969 if (!memcg_can_account_kmem(memcg
))
2972 print_slabinfo_header(m
);
2974 mutex_lock(&memcg
->slab_caches_mutex
);
2975 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2976 cache_show(memcg_params_to_cache(params
), m
);
2977 mutex_unlock(&memcg
->slab_caches_mutex
);
2983 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2985 struct res_counter
*fail_res
;
2986 struct mem_cgroup
*_memcg
;
2989 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2994 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2995 &_memcg
, oom_gfp_allowed(gfp
));
2997 if (ret
== -EINTR
) {
2999 * __mem_cgroup_try_charge() chosed to bypass to root due to
3000 * OOM kill or fatal signal. Since our only options are to
3001 * either fail the allocation or charge it to this cgroup, do
3002 * it as a temporary condition. But we can't fail. From a
3003 * kmem/slab perspective, the cache has already been selected,
3004 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3007 * This condition will only trigger if the task entered
3008 * memcg_charge_kmem in a sane state, but was OOM-killed during
3009 * __mem_cgroup_try_charge() above. Tasks that were already
3010 * dying when the allocation triggers should have been already
3011 * directed to the root cgroup in memcontrol.h
3013 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3014 if (do_swap_account
)
3015 res_counter_charge_nofail(&memcg
->memsw
, size
,
3019 res_counter_uncharge(&memcg
->kmem
, size
);
3024 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3026 res_counter_uncharge(&memcg
->res
, size
);
3027 if (do_swap_account
)
3028 res_counter_uncharge(&memcg
->memsw
, size
);
3031 if (res_counter_uncharge(&memcg
->kmem
, size
))
3035 * Releases a reference taken in kmem_cgroup_css_offline in case
3036 * this last uncharge is racing with the offlining code or it is
3037 * outliving the memcg existence.
3039 * The memory barrier imposed by test&clear is paired with the
3040 * explicit one in memcg_kmem_mark_dead().
3042 if (memcg_kmem_test_and_clear_dead(memcg
))
3043 css_put(&memcg
->css
);
3046 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3051 mutex_lock(&memcg
->slab_caches_mutex
);
3052 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3053 mutex_unlock(&memcg
->slab_caches_mutex
);
3057 * helper for acessing a memcg's index. It will be used as an index in the
3058 * child cache array in kmem_cache, and also to derive its name. This function
3059 * will return -1 when this is not a kmem-limited memcg.
3061 int memcg_cache_id(struct mem_cgroup
*memcg
)
3063 return memcg
? memcg
->kmemcg_id
: -1;
3067 * This ends up being protected by the set_limit mutex, during normal
3068 * operation, because that is its main call site.
3070 * But when we create a new cache, we can call this as well if its parent
3071 * is kmem-limited. That will have to hold set_limit_mutex as well.
3073 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3077 num
= ida_simple_get(&kmem_limited_groups
,
3078 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3082 * After this point, kmem_accounted (that we test atomically in
3083 * the beginning of this conditional), is no longer 0. This
3084 * guarantees only one process will set the following boolean
3085 * to true. We don't need test_and_set because we're protected
3086 * by the set_limit_mutex anyway.
3088 memcg_kmem_set_activated(memcg
);
3090 ret
= memcg_update_all_caches(num
+1);
3092 ida_simple_remove(&kmem_limited_groups
, num
);
3093 memcg_kmem_clear_activated(memcg
);
3097 memcg
->kmemcg_id
= num
;
3098 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3099 mutex_init(&memcg
->slab_caches_mutex
);
3103 static size_t memcg_caches_array_size(int num_groups
)
3106 if (num_groups
<= 0)
3109 size
= 2 * num_groups
;
3110 if (size
< MEMCG_CACHES_MIN_SIZE
)
3111 size
= MEMCG_CACHES_MIN_SIZE
;
3112 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3113 size
= MEMCG_CACHES_MAX_SIZE
;
3119 * We should update the current array size iff all caches updates succeed. This
3120 * can only be done from the slab side. The slab mutex needs to be held when
3123 void memcg_update_array_size(int num
)
3125 if (num
> memcg_limited_groups_array_size
)
3126 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3129 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3131 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3133 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3135 VM_BUG_ON(!is_root_cache(s
));
3137 if (num_groups
> memcg_limited_groups_array_size
) {
3139 ssize_t size
= memcg_caches_array_size(num_groups
);
3141 size
*= sizeof(void *);
3142 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3144 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3145 if (!s
->memcg_params
) {
3146 s
->memcg_params
= cur_params
;
3150 s
->memcg_params
->is_root_cache
= true;
3153 * There is the chance it will be bigger than
3154 * memcg_limited_groups_array_size, if we failed an allocation
3155 * in a cache, in which case all caches updated before it, will
3156 * have a bigger array.
3158 * But if that is the case, the data after
3159 * memcg_limited_groups_array_size is certainly unused
3161 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3162 if (!cur_params
->memcg_caches
[i
])
3164 s
->memcg_params
->memcg_caches
[i
] =
3165 cur_params
->memcg_caches
[i
];
3169 * Ideally, we would wait until all caches succeed, and only
3170 * then free the old one. But this is not worth the extra
3171 * pointer per-cache we'd have to have for this.
3173 * It is not a big deal if some caches are left with a size
3174 * bigger than the others. And all updates will reset this
3182 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3183 struct kmem_cache
*root_cache
)
3187 if (!memcg_kmem_enabled())
3191 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3192 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3194 size
= sizeof(struct memcg_cache_params
);
3196 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3197 if (!s
->memcg_params
)
3201 s
->memcg_params
->memcg
= memcg
;
3202 s
->memcg_params
->root_cache
= root_cache
;
3203 INIT_WORK(&s
->memcg_params
->destroy
,
3204 kmem_cache_destroy_work_func
);
3206 s
->memcg_params
->is_root_cache
= true;
3211 void memcg_release_cache(struct kmem_cache
*s
)
3213 struct kmem_cache
*root
;
3214 struct mem_cgroup
*memcg
;
3218 * This happens, for instance, when a root cache goes away before we
3221 if (!s
->memcg_params
)
3224 if (s
->memcg_params
->is_root_cache
)
3227 memcg
= s
->memcg_params
->memcg
;
3228 id
= memcg_cache_id(memcg
);
3230 root
= s
->memcg_params
->root_cache
;
3231 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3233 mutex_lock(&memcg
->slab_caches_mutex
);
3234 list_del(&s
->memcg_params
->list
);
3235 mutex_unlock(&memcg
->slab_caches_mutex
);
3237 css_put(&memcg
->css
);
3239 kfree(s
->memcg_params
);
3243 * During the creation a new cache, we need to disable our accounting mechanism
3244 * altogether. This is true even if we are not creating, but rather just
3245 * enqueing new caches to be created.
3247 * This is because that process will trigger allocations; some visible, like
3248 * explicit kmallocs to auxiliary data structures, name strings and internal
3249 * cache structures; some well concealed, like INIT_WORK() that can allocate
3250 * objects during debug.
3252 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3253 * to it. This may not be a bounded recursion: since the first cache creation
3254 * failed to complete (waiting on the allocation), we'll just try to create the
3255 * cache again, failing at the same point.
3257 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3258 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3259 * inside the following two functions.
3261 static inline void memcg_stop_kmem_account(void)
3263 VM_BUG_ON(!current
->mm
);
3264 current
->memcg_kmem_skip_account
++;
3267 static inline void memcg_resume_kmem_account(void)
3269 VM_BUG_ON(!current
->mm
);
3270 current
->memcg_kmem_skip_account
--;
3273 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3275 struct kmem_cache
*cachep
;
3276 struct memcg_cache_params
*p
;
3278 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3280 cachep
= memcg_params_to_cache(p
);
3283 * If we get down to 0 after shrink, we could delete right away.
3284 * However, memcg_release_pages() already puts us back in the workqueue
3285 * in that case. If we proceed deleting, we'll get a dangling
3286 * reference, and removing the object from the workqueue in that case
3287 * is unnecessary complication. We are not a fast path.
3289 * Note that this case is fundamentally different from racing with
3290 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3291 * kmem_cache_shrink, not only we would be reinserting a dead cache
3292 * into the queue, but doing so from inside the worker racing to
3295 * So if we aren't down to zero, we'll just schedule a worker and try
3298 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3299 kmem_cache_shrink(cachep
);
3300 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3303 kmem_cache_destroy(cachep
);
3306 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3308 if (!cachep
->memcg_params
->dead
)
3312 * There are many ways in which we can get here.
3314 * We can get to a memory-pressure situation while the delayed work is
3315 * still pending to run. The vmscan shrinkers can then release all
3316 * cache memory and get us to destruction. If this is the case, we'll
3317 * be executed twice, which is a bug (the second time will execute over
3318 * bogus data). In this case, cancelling the work should be fine.
3320 * But we can also get here from the worker itself, if
3321 * kmem_cache_shrink is enough to shake all the remaining objects and
3322 * get the page count to 0. In this case, we'll deadlock if we try to
3323 * cancel the work (the worker runs with an internal lock held, which
3324 * is the same lock we would hold for cancel_work_sync().)
3326 * Since we can't possibly know who got us here, just refrain from
3327 * running if there is already work pending
3329 if (work_pending(&cachep
->memcg_params
->destroy
))
3332 * We have to defer the actual destroying to a workqueue, because
3333 * we might currently be in a context that cannot sleep.
3335 schedule_work(&cachep
->memcg_params
->destroy
);
3339 * This lock protects updaters, not readers. We want readers to be as fast as
3340 * they can, and they will either see NULL or a valid cache value. Our model
3341 * allow them to see NULL, in which case the root memcg will be selected.
3343 * We need this lock because multiple allocations to the same cache from a non
3344 * will span more than one worker. Only one of them can create the cache.
3346 static DEFINE_MUTEX(memcg_cache_mutex
);
3349 * Called with memcg_cache_mutex held
3351 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3352 struct kmem_cache
*s
)
3354 struct kmem_cache
*new;
3355 static char *tmp_name
= NULL
;
3357 lockdep_assert_held(&memcg_cache_mutex
);
3360 * kmem_cache_create_memcg duplicates the given name and
3361 * cgroup_name for this name requires RCU context.
3362 * This static temporary buffer is used to prevent from
3363 * pointless shortliving allocation.
3366 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3372 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3373 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3376 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3377 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3380 new->allocflags
|= __GFP_KMEMCG
;
3385 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3386 struct kmem_cache
*cachep
)
3388 struct kmem_cache
*new_cachep
;
3391 BUG_ON(!memcg_can_account_kmem(memcg
));
3393 idx
= memcg_cache_id(memcg
);
3395 mutex_lock(&memcg_cache_mutex
);
3396 new_cachep
= cache_from_memcg_idx(cachep
, idx
);
3398 css_put(&memcg
->css
);
3402 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3403 if (new_cachep
== NULL
) {
3404 new_cachep
= cachep
;
3405 css_put(&memcg
->css
);
3409 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3411 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3413 * the readers won't lock, make sure everybody sees the updated value,
3414 * so they won't put stuff in the queue again for no reason
3418 mutex_unlock(&memcg_cache_mutex
);
3422 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3424 struct kmem_cache
*c
;
3427 if (!s
->memcg_params
)
3429 if (!s
->memcg_params
->is_root_cache
)
3433 * If the cache is being destroyed, we trust that there is no one else
3434 * requesting objects from it. Even if there are, the sanity checks in
3435 * kmem_cache_destroy should caught this ill-case.
3437 * Still, we don't want anyone else freeing memcg_caches under our
3438 * noses, which can happen if a new memcg comes to life. As usual,
3439 * we'll take the set_limit_mutex to protect ourselves against this.
3441 mutex_lock(&set_limit_mutex
);
3442 for_each_memcg_cache_index(i
) {
3443 c
= cache_from_memcg_idx(s
, i
);
3448 * We will now manually delete the caches, so to avoid races
3449 * we need to cancel all pending destruction workers and
3450 * proceed with destruction ourselves.
3452 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3453 * and that could spawn the workers again: it is likely that
3454 * the cache still have active pages until this very moment.
3455 * This would lead us back to mem_cgroup_destroy_cache.
3457 * But that will not execute at all if the "dead" flag is not
3458 * set, so flip it down to guarantee we are in control.
3460 c
->memcg_params
->dead
= false;
3461 cancel_work_sync(&c
->memcg_params
->destroy
);
3462 kmem_cache_destroy(c
);
3464 mutex_unlock(&set_limit_mutex
);
3467 struct create_work
{
3468 struct mem_cgroup
*memcg
;
3469 struct kmem_cache
*cachep
;
3470 struct work_struct work
;
3473 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3475 struct kmem_cache
*cachep
;
3476 struct memcg_cache_params
*params
;
3478 if (!memcg_kmem_is_active(memcg
))
3481 mutex_lock(&memcg
->slab_caches_mutex
);
3482 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3483 cachep
= memcg_params_to_cache(params
);
3484 cachep
->memcg_params
->dead
= true;
3485 schedule_work(&cachep
->memcg_params
->destroy
);
3487 mutex_unlock(&memcg
->slab_caches_mutex
);
3490 static void memcg_create_cache_work_func(struct work_struct
*w
)
3492 struct create_work
*cw
;
3494 cw
= container_of(w
, struct create_work
, work
);
3495 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3500 * Enqueue the creation of a per-memcg kmem_cache.
3502 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3503 struct kmem_cache
*cachep
)
3505 struct create_work
*cw
;
3507 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3509 css_put(&memcg
->css
);
3514 cw
->cachep
= cachep
;
3516 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3517 schedule_work(&cw
->work
);
3520 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3521 struct kmem_cache
*cachep
)
3524 * We need to stop accounting when we kmalloc, because if the
3525 * corresponding kmalloc cache is not yet created, the first allocation
3526 * in __memcg_create_cache_enqueue will recurse.
3528 * However, it is better to enclose the whole function. Depending on
3529 * the debugging options enabled, INIT_WORK(), for instance, can
3530 * trigger an allocation. This too, will make us recurse. Because at
3531 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3532 * the safest choice is to do it like this, wrapping the whole function.
3534 memcg_stop_kmem_account();
3535 __memcg_create_cache_enqueue(memcg
, cachep
);
3536 memcg_resume_kmem_account();
3539 * Return the kmem_cache we're supposed to use for a slab allocation.
3540 * We try to use the current memcg's version of the cache.
3542 * If the cache does not exist yet, if we are the first user of it,
3543 * we either create it immediately, if possible, or create it asynchronously
3545 * In the latter case, we will let the current allocation go through with
3546 * the original cache.
3548 * Can't be called in interrupt context or from kernel threads.
3549 * This function needs to be called with rcu_read_lock() held.
3551 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3554 struct mem_cgroup
*memcg
;
3557 VM_BUG_ON(!cachep
->memcg_params
);
3558 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3560 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3564 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3566 if (!memcg_can_account_kmem(memcg
))
3569 idx
= memcg_cache_id(memcg
);
3572 * barrier to mare sure we're always seeing the up to date value. The
3573 * code updating memcg_caches will issue a write barrier to match this.
3575 read_barrier_depends();
3576 if (likely(cache_from_memcg_idx(cachep
, idx
))) {
3577 cachep
= cache_from_memcg_idx(cachep
, idx
);
3581 /* The corresponding put will be done in the workqueue. */
3582 if (!css_tryget(&memcg
->css
))
3587 * If we are in a safe context (can wait, and not in interrupt
3588 * context), we could be be predictable and return right away.
3589 * This would guarantee that the allocation being performed
3590 * already belongs in the new cache.
3592 * However, there are some clashes that can arrive from locking.
3593 * For instance, because we acquire the slab_mutex while doing
3594 * kmem_cache_dup, this means no further allocation could happen
3595 * with the slab_mutex held.
3597 * Also, because cache creation issue get_online_cpus(), this
3598 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3599 * that ends up reversed during cpu hotplug. (cpuset allocates
3600 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3601 * better to defer everything.
3603 memcg_create_cache_enqueue(memcg
, cachep
);
3609 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3612 * We need to verify if the allocation against current->mm->owner's memcg is
3613 * possible for the given order. But the page is not allocated yet, so we'll
3614 * need a further commit step to do the final arrangements.
3616 * It is possible for the task to switch cgroups in this mean time, so at
3617 * commit time, we can't rely on task conversion any longer. We'll then use
3618 * the handle argument to return to the caller which cgroup we should commit
3619 * against. We could also return the memcg directly and avoid the pointer
3620 * passing, but a boolean return value gives better semantics considering
3621 * the compiled-out case as well.
3623 * Returning true means the allocation is possible.
3626 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3628 struct mem_cgroup
*memcg
;
3634 * Disabling accounting is only relevant for some specific memcg
3635 * internal allocations. Therefore we would initially not have such
3636 * check here, since direct calls to the page allocator that are marked
3637 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3638 * concerned with cache allocations, and by having this test at
3639 * memcg_kmem_get_cache, we are already able to relay the allocation to
3640 * the root cache and bypass the memcg cache altogether.
3642 * There is one exception, though: the SLUB allocator does not create
3643 * large order caches, but rather service large kmallocs directly from
3644 * the page allocator. Therefore, the following sequence when backed by
3645 * the SLUB allocator:
3647 * memcg_stop_kmem_account();
3648 * kmalloc(<large_number>)
3649 * memcg_resume_kmem_account();
3651 * would effectively ignore the fact that we should skip accounting,
3652 * since it will drive us directly to this function without passing
3653 * through the cache selector memcg_kmem_get_cache. Such large
3654 * allocations are extremely rare but can happen, for instance, for the
3655 * cache arrays. We bring this test here.
3657 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3660 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3663 * very rare case described in mem_cgroup_from_task. Unfortunately there
3664 * isn't much we can do without complicating this too much, and it would
3665 * be gfp-dependent anyway. Just let it go
3667 if (unlikely(!memcg
))
3670 if (!memcg_can_account_kmem(memcg
)) {
3671 css_put(&memcg
->css
);
3675 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3679 css_put(&memcg
->css
);
3683 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3686 struct page_cgroup
*pc
;
3688 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3690 /* The page allocation failed. Revert */
3692 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3696 pc
= lookup_page_cgroup(page
);
3697 lock_page_cgroup(pc
);
3698 pc
->mem_cgroup
= memcg
;
3699 SetPageCgroupUsed(pc
);
3700 unlock_page_cgroup(pc
);
3703 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3705 struct mem_cgroup
*memcg
= NULL
;
3706 struct page_cgroup
*pc
;
3709 pc
= lookup_page_cgroup(page
);
3711 * Fast unlocked return. Theoretically might have changed, have to
3712 * check again after locking.
3714 if (!PageCgroupUsed(pc
))
3717 lock_page_cgroup(pc
);
3718 if (PageCgroupUsed(pc
)) {
3719 memcg
= pc
->mem_cgroup
;
3720 ClearPageCgroupUsed(pc
);
3722 unlock_page_cgroup(pc
);
3725 * We trust that only if there is a memcg associated with the page, it
3726 * is a valid allocation
3731 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3732 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3735 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3738 #endif /* CONFIG_MEMCG_KMEM */
3740 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3742 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3744 * Because tail pages are not marked as "used", set it. We're under
3745 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3746 * charge/uncharge will be never happen and move_account() is done under
3747 * compound_lock(), so we don't have to take care of races.
3749 void mem_cgroup_split_huge_fixup(struct page
*head
)
3751 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3752 struct page_cgroup
*pc
;
3753 struct mem_cgroup
*memcg
;
3756 if (mem_cgroup_disabled())
3759 memcg
= head_pc
->mem_cgroup
;
3760 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3762 pc
->mem_cgroup
= memcg
;
3763 smp_wmb();/* see __commit_charge() */
3764 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3766 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3769 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3772 void mem_cgroup_move_account_page_stat(struct mem_cgroup
*from
,
3773 struct mem_cgroup
*to
,
3774 unsigned int nr_pages
,
3775 enum mem_cgroup_stat_index idx
)
3777 /* Update stat data for mem_cgroup */
3779 __this_cpu_sub(from
->stat
->count
[idx
], nr_pages
);
3780 __this_cpu_add(to
->stat
->count
[idx
], nr_pages
);
3785 * mem_cgroup_move_account - move account of the page
3787 * @nr_pages: number of regular pages (>1 for huge pages)
3788 * @pc: page_cgroup of the page.
3789 * @from: mem_cgroup which the page is moved from.
3790 * @to: mem_cgroup which the page is moved to. @from != @to.
3792 * The caller must confirm following.
3793 * - page is not on LRU (isolate_page() is useful.)
3794 * - compound_lock is held when nr_pages > 1
3796 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3799 static int mem_cgroup_move_account(struct page
*page
,
3800 unsigned int nr_pages
,
3801 struct page_cgroup
*pc
,
3802 struct mem_cgroup
*from
,
3803 struct mem_cgroup
*to
)
3805 unsigned long flags
;
3807 bool anon
= PageAnon(page
);
3809 VM_BUG_ON(from
== to
);
3810 VM_BUG_ON(PageLRU(page
));
3812 * The page is isolated from LRU. So, collapse function
3813 * will not handle this page. But page splitting can happen.
3814 * Do this check under compound_page_lock(). The caller should
3818 if (nr_pages
> 1 && !PageTransHuge(page
))
3821 lock_page_cgroup(pc
);
3824 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3827 move_lock_mem_cgroup(from
, &flags
);
3829 if (!anon
&& page_mapped(page
))
3830 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3831 MEM_CGROUP_STAT_FILE_MAPPED
);
3833 if (PageWriteback(page
))
3834 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3835 MEM_CGROUP_STAT_WRITEBACK
);
3837 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3839 /* caller should have done css_get */
3840 pc
->mem_cgroup
= to
;
3841 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3842 move_unlock_mem_cgroup(from
, &flags
);
3845 unlock_page_cgroup(pc
);
3849 memcg_check_events(to
, page
);
3850 memcg_check_events(from
, page
);
3856 * mem_cgroup_move_parent - moves page to the parent group
3857 * @page: the page to move
3858 * @pc: page_cgroup of the page
3859 * @child: page's cgroup
3861 * move charges to its parent or the root cgroup if the group has no
3862 * parent (aka use_hierarchy==0).
3863 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3864 * mem_cgroup_move_account fails) the failure is always temporary and
3865 * it signals a race with a page removal/uncharge or migration. In the
3866 * first case the page is on the way out and it will vanish from the LRU
3867 * on the next attempt and the call should be retried later.
3868 * Isolation from the LRU fails only if page has been isolated from
3869 * the LRU since we looked at it and that usually means either global
3870 * reclaim or migration going on. The page will either get back to the
3872 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3873 * (!PageCgroupUsed) or moved to a different group. The page will
3874 * disappear in the next attempt.
3876 static int mem_cgroup_move_parent(struct page
*page
,
3877 struct page_cgroup
*pc
,
3878 struct mem_cgroup
*child
)
3880 struct mem_cgroup
*parent
;
3881 unsigned int nr_pages
;
3882 unsigned long uninitialized_var(flags
);
3885 VM_BUG_ON(mem_cgroup_is_root(child
));
3888 if (!get_page_unless_zero(page
))
3890 if (isolate_lru_page(page
))
3893 nr_pages
= hpage_nr_pages(page
);
3895 parent
= parent_mem_cgroup(child
);
3897 * If no parent, move charges to root cgroup.
3900 parent
= root_mem_cgroup
;
3903 VM_BUG_ON(!PageTransHuge(page
));
3904 flags
= compound_lock_irqsave(page
);
3907 ret
= mem_cgroup_move_account(page
, nr_pages
,
3910 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3913 compound_unlock_irqrestore(page
, flags
);
3914 putback_lru_page(page
);
3922 * Charge the memory controller for page usage.
3924 * 0 if the charge was successful
3925 * < 0 if the cgroup is over its limit
3927 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3928 gfp_t gfp_mask
, enum charge_type ctype
)
3930 struct mem_cgroup
*memcg
= NULL
;
3931 unsigned int nr_pages
= 1;
3935 if (PageTransHuge(page
)) {
3936 nr_pages
<<= compound_order(page
);
3937 VM_BUG_ON(!PageTransHuge(page
));
3939 * Never OOM-kill a process for a huge page. The
3940 * fault handler will fall back to regular pages.
3945 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3948 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3952 int mem_cgroup_newpage_charge(struct page
*page
,
3953 struct mm_struct
*mm
, gfp_t gfp_mask
)
3955 if (mem_cgroup_disabled())
3957 VM_BUG_ON(page_mapped(page
));
3958 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3960 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3961 MEM_CGROUP_CHARGE_TYPE_ANON
);
3965 * While swap-in, try_charge -> commit or cancel, the page is locked.
3966 * And when try_charge() successfully returns, one refcnt to memcg without
3967 * struct page_cgroup is acquired. This refcnt will be consumed by
3968 * "commit()" or removed by "cancel()"
3970 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3973 struct mem_cgroup
**memcgp
)
3975 struct mem_cgroup
*memcg
;
3976 struct page_cgroup
*pc
;
3979 pc
= lookup_page_cgroup(page
);
3981 * Every swap fault against a single page tries to charge the
3982 * page, bail as early as possible. shmem_unuse() encounters
3983 * already charged pages, too. The USED bit is protected by
3984 * the page lock, which serializes swap cache removal, which
3985 * in turn serializes uncharging.
3987 if (PageCgroupUsed(pc
))
3989 if (!do_swap_account
)
3991 memcg
= try_get_mem_cgroup_from_page(page
);
3995 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3996 css_put(&memcg
->css
);
4001 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
4007 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
4008 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
4011 if (mem_cgroup_disabled())
4014 * A racing thread's fault, or swapoff, may have already
4015 * updated the pte, and even removed page from swap cache: in
4016 * those cases unuse_pte()'s pte_same() test will fail; but
4017 * there's also a KSM case which does need to charge the page.
4019 if (!PageSwapCache(page
)) {
4022 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4027 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4030 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4032 if (mem_cgroup_disabled())
4036 __mem_cgroup_cancel_charge(memcg
, 1);
4040 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4041 enum charge_type ctype
)
4043 if (mem_cgroup_disabled())
4048 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4050 * Now swap is on-memory. This means this page may be
4051 * counted both as mem and swap....double count.
4052 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4053 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4054 * may call delete_from_swap_cache() before reach here.
4056 if (do_swap_account
&& PageSwapCache(page
)) {
4057 swp_entry_t ent
= {.val
= page_private(page
)};
4058 mem_cgroup_uncharge_swap(ent
);
4062 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4063 struct mem_cgroup
*memcg
)
4065 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4066 MEM_CGROUP_CHARGE_TYPE_ANON
);
4069 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4072 struct mem_cgroup
*memcg
= NULL
;
4073 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4076 if (mem_cgroup_disabled())
4078 if (PageCompound(page
))
4081 if (!PageSwapCache(page
))
4082 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4083 else { /* page is swapcache/shmem */
4084 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4087 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4092 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4093 unsigned int nr_pages
,
4094 const enum charge_type ctype
)
4096 struct memcg_batch_info
*batch
= NULL
;
4097 bool uncharge_memsw
= true;
4099 /* If swapout, usage of swap doesn't decrease */
4100 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4101 uncharge_memsw
= false;
4103 batch
= ¤t
->memcg_batch
;
4105 * In usual, we do css_get() when we remember memcg pointer.
4106 * But in this case, we keep res->usage until end of a series of
4107 * uncharges. Then, it's ok to ignore memcg's refcnt.
4110 batch
->memcg
= memcg
;
4112 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4113 * In those cases, all pages freed continuously can be expected to be in
4114 * the same cgroup and we have chance to coalesce uncharges.
4115 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4116 * because we want to do uncharge as soon as possible.
4119 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4120 goto direct_uncharge
;
4123 goto direct_uncharge
;
4126 * In typical case, batch->memcg == mem. This means we can
4127 * merge a series of uncharges to an uncharge of res_counter.
4128 * If not, we uncharge res_counter ony by one.
4130 if (batch
->memcg
!= memcg
)
4131 goto direct_uncharge
;
4132 /* remember freed charge and uncharge it later */
4135 batch
->memsw_nr_pages
++;
4138 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4140 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4141 if (unlikely(batch
->memcg
!= memcg
))
4142 memcg_oom_recover(memcg
);
4146 * uncharge if !page_mapped(page)
4148 static struct mem_cgroup
*
4149 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4152 struct mem_cgroup
*memcg
= NULL
;
4153 unsigned int nr_pages
= 1;
4154 struct page_cgroup
*pc
;
4157 if (mem_cgroup_disabled())
4160 if (PageTransHuge(page
)) {
4161 nr_pages
<<= compound_order(page
);
4162 VM_BUG_ON(!PageTransHuge(page
));
4165 * Check if our page_cgroup is valid
4167 pc
= lookup_page_cgroup(page
);
4168 if (unlikely(!PageCgroupUsed(pc
)))
4171 lock_page_cgroup(pc
);
4173 memcg
= pc
->mem_cgroup
;
4175 if (!PageCgroupUsed(pc
))
4178 anon
= PageAnon(page
);
4181 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4183 * Generally PageAnon tells if it's the anon statistics to be
4184 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4185 * used before page reached the stage of being marked PageAnon.
4189 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4190 /* See mem_cgroup_prepare_migration() */
4191 if (page_mapped(page
))
4194 * Pages under migration may not be uncharged. But
4195 * end_migration() /must/ be the one uncharging the
4196 * unused post-migration page and so it has to call
4197 * here with the migration bit still set. See the
4198 * res_counter handling below.
4200 if (!end_migration
&& PageCgroupMigration(pc
))
4203 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4204 if (!PageAnon(page
)) { /* Shared memory */
4205 if (page
->mapping
&& !page_is_file_cache(page
))
4207 } else if (page_mapped(page
)) /* Anon */
4214 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4216 ClearPageCgroupUsed(pc
);
4218 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4219 * freed from LRU. This is safe because uncharged page is expected not
4220 * to be reused (freed soon). Exception is SwapCache, it's handled by
4221 * special functions.
4224 unlock_page_cgroup(pc
);
4226 * even after unlock, we have memcg->res.usage here and this memcg
4227 * will never be freed, so it's safe to call css_get().
4229 memcg_check_events(memcg
, page
);
4230 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4231 mem_cgroup_swap_statistics(memcg
, true);
4232 css_get(&memcg
->css
);
4235 * Migration does not charge the res_counter for the
4236 * replacement page, so leave it alone when phasing out the
4237 * page that is unused after the migration.
4239 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4240 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4245 unlock_page_cgroup(pc
);
4249 void mem_cgroup_uncharge_page(struct page
*page
)
4252 if (page_mapped(page
))
4254 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4256 * If the page is in swap cache, uncharge should be deferred
4257 * to the swap path, which also properly accounts swap usage
4258 * and handles memcg lifetime.
4260 * Note that this check is not stable and reclaim may add the
4261 * page to swap cache at any time after this. However, if the
4262 * page is not in swap cache by the time page->mapcount hits
4263 * 0, there won't be any page table references to the swap
4264 * slot, and reclaim will free it and not actually write the
4267 if (PageSwapCache(page
))
4269 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4272 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4274 VM_BUG_ON(page_mapped(page
));
4275 VM_BUG_ON(page
->mapping
);
4276 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4280 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4281 * In that cases, pages are freed continuously and we can expect pages
4282 * are in the same memcg. All these calls itself limits the number of
4283 * pages freed at once, then uncharge_start/end() is called properly.
4284 * This may be called prural(2) times in a context,
4287 void mem_cgroup_uncharge_start(void)
4289 current
->memcg_batch
.do_batch
++;
4290 /* We can do nest. */
4291 if (current
->memcg_batch
.do_batch
== 1) {
4292 current
->memcg_batch
.memcg
= NULL
;
4293 current
->memcg_batch
.nr_pages
= 0;
4294 current
->memcg_batch
.memsw_nr_pages
= 0;
4298 void mem_cgroup_uncharge_end(void)
4300 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4302 if (!batch
->do_batch
)
4306 if (batch
->do_batch
) /* If stacked, do nothing. */
4312 * This "batch->memcg" is valid without any css_get/put etc...
4313 * bacause we hide charges behind us.
4315 if (batch
->nr_pages
)
4316 res_counter_uncharge(&batch
->memcg
->res
,
4317 batch
->nr_pages
* PAGE_SIZE
);
4318 if (batch
->memsw_nr_pages
)
4319 res_counter_uncharge(&batch
->memcg
->memsw
,
4320 batch
->memsw_nr_pages
* PAGE_SIZE
);
4321 memcg_oom_recover(batch
->memcg
);
4322 /* forget this pointer (for sanity check) */
4323 batch
->memcg
= NULL
;
4328 * called after __delete_from_swap_cache() and drop "page" account.
4329 * memcg information is recorded to swap_cgroup of "ent"
4332 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4334 struct mem_cgroup
*memcg
;
4335 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4337 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4338 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4340 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4343 * record memcg information, if swapout && memcg != NULL,
4344 * css_get() was called in uncharge().
4346 if (do_swap_account
&& swapout
&& memcg
)
4347 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4351 #ifdef CONFIG_MEMCG_SWAP
4353 * called from swap_entry_free(). remove record in swap_cgroup and
4354 * uncharge "memsw" account.
4356 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4358 struct mem_cgroup
*memcg
;
4361 if (!do_swap_account
)
4364 id
= swap_cgroup_record(ent
, 0);
4366 memcg
= mem_cgroup_lookup(id
);
4369 * We uncharge this because swap is freed.
4370 * This memcg can be obsolete one. We avoid calling css_tryget
4372 if (!mem_cgroup_is_root(memcg
))
4373 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4374 mem_cgroup_swap_statistics(memcg
, false);
4375 css_put(&memcg
->css
);
4381 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4382 * @entry: swap entry to be moved
4383 * @from: mem_cgroup which the entry is moved from
4384 * @to: mem_cgroup which the entry is moved to
4386 * It succeeds only when the swap_cgroup's record for this entry is the same
4387 * as the mem_cgroup's id of @from.
4389 * Returns 0 on success, -EINVAL on failure.
4391 * The caller must have charged to @to, IOW, called res_counter_charge() about
4392 * both res and memsw, and called css_get().
4394 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4395 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4397 unsigned short old_id
, new_id
;
4399 old_id
= css_id(&from
->css
);
4400 new_id
= css_id(&to
->css
);
4402 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4403 mem_cgroup_swap_statistics(from
, false);
4404 mem_cgroup_swap_statistics(to
, true);
4406 * This function is only called from task migration context now.
4407 * It postpones res_counter and refcount handling till the end
4408 * of task migration(mem_cgroup_clear_mc()) for performance
4409 * improvement. But we cannot postpone css_get(to) because if
4410 * the process that has been moved to @to does swap-in, the
4411 * refcount of @to might be decreased to 0.
4413 * We are in attach() phase, so the cgroup is guaranteed to be
4414 * alive, so we can just call css_get().
4422 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4423 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4430 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4433 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4434 struct mem_cgroup
**memcgp
)
4436 struct mem_cgroup
*memcg
= NULL
;
4437 unsigned int nr_pages
= 1;
4438 struct page_cgroup
*pc
;
4439 enum charge_type ctype
;
4443 if (mem_cgroup_disabled())
4446 if (PageTransHuge(page
))
4447 nr_pages
<<= compound_order(page
);
4449 pc
= lookup_page_cgroup(page
);
4450 lock_page_cgroup(pc
);
4451 if (PageCgroupUsed(pc
)) {
4452 memcg
= pc
->mem_cgroup
;
4453 css_get(&memcg
->css
);
4455 * At migrating an anonymous page, its mapcount goes down
4456 * to 0 and uncharge() will be called. But, even if it's fully
4457 * unmapped, migration may fail and this page has to be
4458 * charged again. We set MIGRATION flag here and delay uncharge
4459 * until end_migration() is called
4461 * Corner Case Thinking
4463 * When the old page was mapped as Anon and it's unmap-and-freed
4464 * while migration was ongoing.
4465 * If unmap finds the old page, uncharge() of it will be delayed
4466 * until end_migration(). If unmap finds a new page, it's
4467 * uncharged when it make mapcount to be 1->0. If unmap code
4468 * finds swap_migration_entry, the new page will not be mapped
4469 * and end_migration() will find it(mapcount==0).
4472 * When the old page was mapped but migraion fails, the kernel
4473 * remaps it. A charge for it is kept by MIGRATION flag even
4474 * if mapcount goes down to 0. We can do remap successfully
4475 * without charging it again.
4478 * The "old" page is under lock_page() until the end of
4479 * migration, so, the old page itself will not be swapped-out.
4480 * If the new page is swapped out before end_migraton, our
4481 * hook to usual swap-out path will catch the event.
4484 SetPageCgroupMigration(pc
);
4486 unlock_page_cgroup(pc
);
4488 * If the page is not charged at this point,
4496 * We charge new page before it's used/mapped. So, even if unlock_page()
4497 * is called before end_migration, we can catch all events on this new
4498 * page. In the case new page is migrated but not remapped, new page's
4499 * mapcount will be finally 0 and we call uncharge in end_migration().
4502 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4504 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4506 * The page is committed to the memcg, but it's not actually
4507 * charged to the res_counter since we plan on replacing the
4508 * old one and only one page is going to be left afterwards.
4510 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4513 /* remove redundant charge if migration failed*/
4514 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4515 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4517 struct page
*used
, *unused
;
4518 struct page_cgroup
*pc
;
4524 if (!migration_ok
) {
4531 anon
= PageAnon(used
);
4532 __mem_cgroup_uncharge_common(unused
,
4533 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4534 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4536 css_put(&memcg
->css
);
4538 * We disallowed uncharge of pages under migration because mapcount
4539 * of the page goes down to zero, temporarly.
4540 * Clear the flag and check the page should be charged.
4542 pc
= lookup_page_cgroup(oldpage
);
4543 lock_page_cgroup(pc
);
4544 ClearPageCgroupMigration(pc
);
4545 unlock_page_cgroup(pc
);
4548 * If a page is a file cache, radix-tree replacement is very atomic
4549 * and we can skip this check. When it was an Anon page, its mapcount
4550 * goes down to 0. But because we added MIGRATION flage, it's not
4551 * uncharged yet. There are several case but page->mapcount check
4552 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4553 * check. (see prepare_charge() also)
4556 mem_cgroup_uncharge_page(used
);
4560 * At replace page cache, newpage is not under any memcg but it's on
4561 * LRU. So, this function doesn't touch res_counter but handles LRU
4562 * in correct way. Both pages are locked so we cannot race with uncharge.
4564 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4565 struct page
*newpage
)
4567 struct mem_cgroup
*memcg
= NULL
;
4568 struct page_cgroup
*pc
;
4569 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4571 if (mem_cgroup_disabled())
4574 pc
= lookup_page_cgroup(oldpage
);
4575 /* fix accounting on old pages */
4576 lock_page_cgroup(pc
);
4577 if (PageCgroupUsed(pc
)) {
4578 memcg
= pc
->mem_cgroup
;
4579 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4580 ClearPageCgroupUsed(pc
);
4582 unlock_page_cgroup(pc
);
4585 * When called from shmem_replace_page(), in some cases the
4586 * oldpage has already been charged, and in some cases not.
4591 * Even if newpage->mapping was NULL before starting replacement,
4592 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4593 * LRU while we overwrite pc->mem_cgroup.
4595 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4598 #ifdef CONFIG_DEBUG_VM
4599 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4601 struct page_cgroup
*pc
;
4603 pc
= lookup_page_cgroup(page
);
4605 * Can be NULL while feeding pages into the page allocator for
4606 * the first time, i.e. during boot or memory hotplug;
4607 * or when mem_cgroup_disabled().
4609 if (likely(pc
) && PageCgroupUsed(pc
))
4614 bool mem_cgroup_bad_page_check(struct page
*page
)
4616 if (mem_cgroup_disabled())
4619 return lookup_page_cgroup_used(page
) != NULL
;
4622 void mem_cgroup_print_bad_page(struct page
*page
)
4624 struct page_cgroup
*pc
;
4626 pc
= lookup_page_cgroup_used(page
);
4628 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4629 pc
, pc
->flags
, pc
->mem_cgroup
);
4634 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4635 unsigned long long val
)
4638 u64 memswlimit
, memlimit
;
4640 int children
= mem_cgroup_count_children(memcg
);
4641 u64 curusage
, oldusage
;
4645 * For keeping hierarchical_reclaim simple, how long we should retry
4646 * is depends on callers. We set our retry-count to be function
4647 * of # of children which we should visit in this loop.
4649 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4651 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4654 while (retry_count
) {
4655 if (signal_pending(current
)) {
4660 * Rather than hide all in some function, I do this in
4661 * open coded manner. You see what this really does.
4662 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4664 mutex_lock(&set_limit_mutex
);
4665 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4666 if (memswlimit
< val
) {
4668 mutex_unlock(&set_limit_mutex
);
4672 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4676 ret
= res_counter_set_limit(&memcg
->res
, val
);
4678 if (memswlimit
== val
)
4679 memcg
->memsw_is_minimum
= true;
4681 memcg
->memsw_is_minimum
= false;
4683 mutex_unlock(&set_limit_mutex
);
4688 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4689 MEM_CGROUP_RECLAIM_SHRINK
);
4690 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4691 /* Usage is reduced ? */
4692 if (curusage
>= oldusage
)
4695 oldusage
= curusage
;
4697 if (!ret
&& enlarge
)
4698 memcg_oom_recover(memcg
);
4703 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4704 unsigned long long val
)
4707 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4708 int children
= mem_cgroup_count_children(memcg
);
4712 /* see mem_cgroup_resize_res_limit */
4713 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4714 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4715 while (retry_count
) {
4716 if (signal_pending(current
)) {
4721 * Rather than hide all in some function, I do this in
4722 * open coded manner. You see what this really does.
4723 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4725 mutex_lock(&set_limit_mutex
);
4726 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4727 if (memlimit
> val
) {
4729 mutex_unlock(&set_limit_mutex
);
4732 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4733 if (memswlimit
< val
)
4735 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4737 if (memlimit
== val
)
4738 memcg
->memsw_is_minimum
= true;
4740 memcg
->memsw_is_minimum
= false;
4742 mutex_unlock(&set_limit_mutex
);
4747 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4748 MEM_CGROUP_RECLAIM_NOSWAP
|
4749 MEM_CGROUP_RECLAIM_SHRINK
);
4750 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4751 /* Usage is reduced ? */
4752 if (curusage
>= oldusage
)
4755 oldusage
= curusage
;
4757 if (!ret
&& enlarge
)
4758 memcg_oom_recover(memcg
);
4762 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4764 unsigned long *total_scanned
)
4766 unsigned long nr_reclaimed
= 0;
4767 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4768 unsigned long reclaimed
;
4770 struct mem_cgroup_tree_per_zone
*mctz
;
4771 unsigned long long excess
;
4772 unsigned long nr_scanned
;
4777 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4779 * This loop can run a while, specially if mem_cgroup's continuously
4780 * keep exceeding their soft limit and putting the system under
4787 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4792 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4793 gfp_mask
, &nr_scanned
);
4794 nr_reclaimed
+= reclaimed
;
4795 *total_scanned
+= nr_scanned
;
4796 spin_lock(&mctz
->lock
);
4799 * If we failed to reclaim anything from this memory cgroup
4800 * it is time to move on to the next cgroup
4806 * Loop until we find yet another one.
4808 * By the time we get the soft_limit lock
4809 * again, someone might have aded the
4810 * group back on the RB tree. Iterate to
4811 * make sure we get a different mem.
4812 * mem_cgroup_largest_soft_limit_node returns
4813 * NULL if no other cgroup is present on
4817 __mem_cgroup_largest_soft_limit_node(mctz
);
4819 css_put(&next_mz
->memcg
->css
);
4820 else /* next_mz == NULL or other memcg */
4824 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4825 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4827 * One school of thought says that we should not add
4828 * back the node to the tree if reclaim returns 0.
4829 * But our reclaim could return 0, simply because due
4830 * to priority we are exposing a smaller subset of
4831 * memory to reclaim from. Consider this as a longer
4834 /* If excess == 0, no tree ops */
4835 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4836 spin_unlock(&mctz
->lock
);
4837 css_put(&mz
->memcg
->css
);
4840 * Could not reclaim anything and there are no more
4841 * mem cgroups to try or we seem to be looping without
4842 * reclaiming anything.
4844 if (!nr_reclaimed
&&
4846 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4848 } while (!nr_reclaimed
);
4850 css_put(&next_mz
->memcg
->css
);
4851 return nr_reclaimed
;
4855 * mem_cgroup_force_empty_list - clears LRU of a group
4856 * @memcg: group to clear
4859 * @lru: lru to to clear
4861 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4862 * reclaim the pages page themselves - pages are moved to the parent (or root)
4865 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4866 int node
, int zid
, enum lru_list lru
)
4868 struct lruvec
*lruvec
;
4869 unsigned long flags
;
4870 struct list_head
*list
;
4874 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4875 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4876 list
= &lruvec
->lists
[lru
];
4880 struct page_cgroup
*pc
;
4883 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4884 if (list_empty(list
)) {
4885 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4888 page
= list_entry(list
->prev
, struct page
, lru
);
4890 list_move(&page
->lru
, list
);
4892 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4895 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4897 pc
= lookup_page_cgroup(page
);
4899 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4900 /* found lock contention or "pc" is obsolete. */
4905 } while (!list_empty(list
));
4909 * make mem_cgroup's charge to be 0 if there is no task by moving
4910 * all the charges and pages to the parent.
4911 * This enables deleting this mem_cgroup.
4913 * Caller is responsible for holding css reference on the memcg.
4915 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4921 /* This is for making all *used* pages to be on LRU. */
4922 lru_add_drain_all();
4923 drain_all_stock_sync(memcg
);
4924 mem_cgroup_start_move(memcg
);
4925 for_each_node_state(node
, N_MEMORY
) {
4926 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4929 mem_cgroup_force_empty_list(memcg
,
4934 mem_cgroup_end_move(memcg
);
4935 memcg_oom_recover(memcg
);
4939 * Kernel memory may not necessarily be trackable to a specific
4940 * process. So they are not migrated, and therefore we can't
4941 * expect their value to drop to 0 here.
4942 * Having res filled up with kmem only is enough.
4944 * This is a safety check because mem_cgroup_force_empty_list
4945 * could have raced with mem_cgroup_replace_page_cache callers
4946 * so the lru seemed empty but the page could have been added
4947 * right after the check. RES_USAGE should be safe as we always
4948 * charge before adding to the LRU.
4950 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4951 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4952 } while (usage
> 0);
4955 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4957 lockdep_assert_held(&memcg_create_mutex
);
4959 * The lock does not prevent addition or deletion to the list
4960 * of children, but it prevents a new child from being
4961 * initialized based on this parent in css_online(), so it's
4962 * enough to decide whether hierarchically inherited
4963 * attributes can still be changed or not.
4965 return memcg
->use_hierarchy
&&
4966 !list_empty(&memcg
->css
.cgroup
->children
);
4970 * Reclaims as many pages from the given memcg as possible and moves
4971 * the rest to the parent.
4973 * Caller is responsible for holding css reference for memcg.
4975 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4977 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4978 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4980 /* returns EBUSY if there is a task or if we come here twice. */
4981 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4984 /* we call try-to-free pages for make this cgroup empty */
4985 lru_add_drain_all();
4986 /* try to free all pages in this cgroup */
4987 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4990 if (signal_pending(current
))
4993 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4997 /* maybe some writeback is necessary */
4998 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
5003 mem_cgroup_reparent_charges(memcg
);
5008 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5011 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5013 if (mem_cgroup_is_root(memcg
))
5015 return mem_cgroup_force_empty(memcg
);
5018 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5021 return mem_cgroup_from_css(css
)->use_hierarchy
;
5024 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5025 struct cftype
*cft
, u64 val
)
5028 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5029 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5031 mutex_lock(&memcg_create_mutex
);
5033 if (memcg
->use_hierarchy
== val
)
5037 * If parent's use_hierarchy is set, we can't make any modifications
5038 * in the child subtrees. If it is unset, then the change can
5039 * occur, provided the current cgroup has no children.
5041 * For the root cgroup, parent_mem is NULL, we allow value to be
5042 * set if there are no children.
5044 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5045 (val
== 1 || val
== 0)) {
5046 if (list_empty(&memcg
->css
.cgroup
->children
))
5047 memcg
->use_hierarchy
= val
;
5054 mutex_unlock(&memcg_create_mutex
);
5060 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5061 enum mem_cgroup_stat_index idx
)
5063 struct mem_cgroup
*iter
;
5066 /* Per-cpu values can be negative, use a signed accumulator */
5067 for_each_mem_cgroup_tree(iter
, memcg
)
5068 val
+= mem_cgroup_read_stat(iter
, idx
);
5070 if (val
< 0) /* race ? */
5075 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5079 if (!mem_cgroup_is_root(memcg
)) {
5081 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5083 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5087 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5088 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5090 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5091 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5094 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5096 return val
<< PAGE_SHIFT
;
5099 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
5100 struct cftype
*cft
, struct file
*file
,
5101 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
5103 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5109 type
= MEMFILE_TYPE(cft
->private);
5110 name
= MEMFILE_ATTR(cft
->private);
5114 if (name
== RES_USAGE
)
5115 val
= mem_cgroup_usage(memcg
, false);
5117 val
= res_counter_read_u64(&memcg
->res
, name
);
5120 if (name
== RES_USAGE
)
5121 val
= mem_cgroup_usage(memcg
, true);
5123 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5126 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5132 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5133 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5136 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5139 #ifdef CONFIG_MEMCG_KMEM
5140 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5142 * For simplicity, we won't allow this to be disabled. It also can't
5143 * be changed if the cgroup has children already, or if tasks had
5146 * If tasks join before we set the limit, a person looking at
5147 * kmem.usage_in_bytes will have no way to determine when it took
5148 * place, which makes the value quite meaningless.
5150 * After it first became limited, changes in the value of the limit are
5151 * of course permitted.
5153 mutex_lock(&memcg_create_mutex
);
5154 mutex_lock(&set_limit_mutex
);
5155 if (!memcg
->kmem_account_flags
&& val
!= RES_COUNTER_MAX
) {
5156 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5160 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5163 ret
= memcg_update_cache_sizes(memcg
);
5165 res_counter_set_limit(&memcg
->kmem
, RES_COUNTER_MAX
);
5168 static_key_slow_inc(&memcg_kmem_enabled_key
);
5170 * setting the active bit after the inc will guarantee no one
5171 * starts accounting before all call sites are patched
5173 memcg_kmem_set_active(memcg
);
5175 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5177 mutex_unlock(&set_limit_mutex
);
5178 mutex_unlock(&memcg_create_mutex
);
5183 #ifdef CONFIG_MEMCG_KMEM
5184 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5187 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5191 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5193 * When that happen, we need to disable the static branch only on those
5194 * memcgs that enabled it. To achieve this, we would be forced to
5195 * complicate the code by keeping track of which memcgs were the ones
5196 * that actually enabled limits, and which ones got it from its
5199 * It is a lot simpler just to do static_key_slow_inc() on every child
5200 * that is accounted.
5202 if (!memcg_kmem_is_active(memcg
))
5206 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5207 * memcg is active already. If the later initialization fails then the
5208 * cgroup core triggers the cleanup so we do not have to do it here.
5210 static_key_slow_inc(&memcg_kmem_enabled_key
);
5212 mutex_lock(&set_limit_mutex
);
5213 memcg_stop_kmem_account();
5214 ret
= memcg_update_cache_sizes(memcg
);
5215 memcg_resume_kmem_account();
5216 mutex_unlock(&set_limit_mutex
);
5220 #endif /* CONFIG_MEMCG_KMEM */
5223 * The user of this function is...
5226 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5229 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5232 unsigned long long val
;
5235 type
= MEMFILE_TYPE(cft
->private);
5236 name
= MEMFILE_ATTR(cft
->private);
5240 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5244 /* This function does all necessary parse...reuse it */
5245 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5249 ret
= mem_cgroup_resize_limit(memcg
, val
);
5250 else if (type
== _MEMSWAP
)
5251 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5252 else if (type
== _KMEM
)
5253 ret
= memcg_update_kmem_limit(css
, val
);
5257 case RES_SOFT_LIMIT
:
5258 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5262 * For memsw, soft limits are hard to implement in terms
5263 * of semantics, for now, we support soft limits for
5264 * control without swap
5267 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5272 ret
= -EINVAL
; /* should be BUG() ? */
5278 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5279 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5281 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5283 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5284 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5285 if (!memcg
->use_hierarchy
)
5288 while (css_parent(&memcg
->css
)) {
5289 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5290 if (!memcg
->use_hierarchy
)
5292 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5293 min_limit
= min(min_limit
, tmp
);
5294 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5295 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5298 *mem_limit
= min_limit
;
5299 *memsw_limit
= min_memsw_limit
;
5302 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5304 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5308 type
= MEMFILE_TYPE(event
);
5309 name
= MEMFILE_ATTR(event
);
5314 res_counter_reset_max(&memcg
->res
);
5315 else if (type
== _MEMSWAP
)
5316 res_counter_reset_max(&memcg
->memsw
);
5317 else if (type
== _KMEM
)
5318 res_counter_reset_max(&memcg
->kmem
);
5324 res_counter_reset_failcnt(&memcg
->res
);
5325 else if (type
== _MEMSWAP
)
5326 res_counter_reset_failcnt(&memcg
->memsw
);
5327 else if (type
== _KMEM
)
5328 res_counter_reset_failcnt(&memcg
->kmem
);
5337 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5340 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5344 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5345 struct cftype
*cft
, u64 val
)
5347 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5349 if (val
>= (1 << NR_MOVE_TYPE
))
5353 * No kind of locking is needed in here, because ->can_attach() will
5354 * check this value once in the beginning of the process, and then carry
5355 * on with stale data. This means that changes to this value will only
5356 * affect task migrations starting after the change.
5358 memcg
->move_charge_at_immigrate
= val
;
5362 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5363 struct cftype
*cft
, u64 val
)
5370 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5371 struct cftype
*cft
, struct seq_file
*m
)
5375 unsigned int lru_mask
;
5378 static const struct numa_stat stats
[] = {
5379 { "total", LRU_ALL
},
5380 { "file", LRU_ALL_FILE
},
5381 { "anon", LRU_ALL_ANON
},
5382 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5384 const struct numa_stat
*stat
;
5387 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5389 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5390 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5391 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5392 for_each_node_state(nid
, N_MEMORY
) {
5393 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5395 seq_printf(m
, " N%d=%lu", nid
, nr
);
5400 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5401 struct mem_cgroup
*iter
;
5404 for_each_mem_cgroup_tree(iter
, memcg
)
5405 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5406 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5407 for_each_node_state(nid
, N_MEMORY
) {
5409 for_each_mem_cgroup_tree(iter
, memcg
)
5410 nr
+= mem_cgroup_node_nr_lru_pages(
5411 iter
, nid
, stat
->lru_mask
);
5412 seq_printf(m
, " N%d=%lu", nid
, nr
);
5419 #endif /* CONFIG_NUMA */
5421 static inline void mem_cgroup_lru_names_not_uptodate(void)
5423 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5426 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5429 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5430 struct mem_cgroup
*mi
;
5433 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5434 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5436 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5437 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5440 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5441 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5442 mem_cgroup_read_events(memcg
, i
));
5444 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5445 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5446 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5448 /* Hierarchical information */
5450 unsigned long long limit
, memsw_limit
;
5451 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5452 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5453 if (do_swap_account
)
5454 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5458 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5461 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5463 for_each_mem_cgroup_tree(mi
, memcg
)
5464 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5465 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5468 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5469 unsigned long long val
= 0;
5471 for_each_mem_cgroup_tree(mi
, memcg
)
5472 val
+= mem_cgroup_read_events(mi
, i
);
5473 seq_printf(m
, "total_%s %llu\n",
5474 mem_cgroup_events_names
[i
], val
);
5477 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5478 unsigned long long val
= 0;
5480 for_each_mem_cgroup_tree(mi
, memcg
)
5481 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5482 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5485 #ifdef CONFIG_DEBUG_VM
5488 struct mem_cgroup_per_zone
*mz
;
5489 struct zone_reclaim_stat
*rstat
;
5490 unsigned long recent_rotated
[2] = {0, 0};
5491 unsigned long recent_scanned
[2] = {0, 0};
5493 for_each_online_node(nid
)
5494 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5495 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5496 rstat
= &mz
->lruvec
.reclaim_stat
;
5498 recent_rotated
[0] += rstat
->recent_rotated
[0];
5499 recent_rotated
[1] += rstat
->recent_rotated
[1];
5500 recent_scanned
[0] += rstat
->recent_scanned
[0];
5501 recent_scanned
[1] += rstat
->recent_scanned
[1];
5503 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5504 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5505 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5506 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5513 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5516 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5518 return mem_cgroup_swappiness(memcg
);
5521 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5522 struct cftype
*cft
, u64 val
)
5524 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5525 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5527 if (val
> 100 || !parent
)
5530 mutex_lock(&memcg_create_mutex
);
5532 /* If under hierarchy, only empty-root can set this value */
5533 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5534 mutex_unlock(&memcg_create_mutex
);
5538 memcg
->swappiness
= val
;
5540 mutex_unlock(&memcg_create_mutex
);
5545 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5547 struct mem_cgroup_threshold_ary
*t
;
5553 t
= rcu_dereference(memcg
->thresholds
.primary
);
5555 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5560 usage
= mem_cgroup_usage(memcg
, swap
);
5563 * current_threshold points to threshold just below or equal to usage.
5564 * If it's not true, a threshold was crossed after last
5565 * call of __mem_cgroup_threshold().
5567 i
= t
->current_threshold
;
5570 * Iterate backward over array of thresholds starting from
5571 * current_threshold and check if a threshold is crossed.
5572 * If none of thresholds below usage is crossed, we read
5573 * only one element of the array here.
5575 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5576 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5578 /* i = current_threshold + 1 */
5582 * Iterate forward over array of thresholds starting from
5583 * current_threshold+1 and check if a threshold is crossed.
5584 * If none of thresholds above usage is crossed, we read
5585 * only one element of the array here.
5587 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5588 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5590 /* Update current_threshold */
5591 t
->current_threshold
= i
- 1;
5596 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5599 __mem_cgroup_threshold(memcg
, false);
5600 if (do_swap_account
)
5601 __mem_cgroup_threshold(memcg
, true);
5603 memcg
= parent_mem_cgroup(memcg
);
5607 static int compare_thresholds(const void *a
, const void *b
)
5609 const struct mem_cgroup_threshold
*_a
= a
;
5610 const struct mem_cgroup_threshold
*_b
= b
;
5612 if (_a
->threshold
> _b
->threshold
)
5615 if (_a
->threshold
< _b
->threshold
)
5621 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5623 struct mem_cgroup_eventfd_list
*ev
;
5625 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5626 eventfd_signal(ev
->eventfd
, 1);
5630 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5632 struct mem_cgroup
*iter
;
5634 for_each_mem_cgroup_tree(iter
, memcg
)
5635 mem_cgroup_oom_notify_cb(iter
);
5638 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5639 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5641 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5642 struct mem_cgroup_thresholds
*thresholds
;
5643 struct mem_cgroup_threshold_ary
*new;
5644 enum res_type type
= MEMFILE_TYPE(cft
->private);
5645 u64 threshold
, usage
;
5648 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5652 mutex_lock(&memcg
->thresholds_lock
);
5655 thresholds
= &memcg
->thresholds
;
5656 else if (type
== _MEMSWAP
)
5657 thresholds
= &memcg
->memsw_thresholds
;
5661 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5663 /* Check if a threshold crossed before adding a new one */
5664 if (thresholds
->primary
)
5665 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5667 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5669 /* Allocate memory for new array of thresholds */
5670 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5678 /* Copy thresholds (if any) to new array */
5679 if (thresholds
->primary
) {
5680 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5681 sizeof(struct mem_cgroup_threshold
));
5684 /* Add new threshold */
5685 new->entries
[size
- 1].eventfd
= eventfd
;
5686 new->entries
[size
- 1].threshold
= threshold
;
5688 /* Sort thresholds. Registering of new threshold isn't time-critical */
5689 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5690 compare_thresholds
, NULL
);
5692 /* Find current threshold */
5693 new->current_threshold
= -1;
5694 for (i
= 0; i
< size
; i
++) {
5695 if (new->entries
[i
].threshold
<= usage
) {
5697 * new->current_threshold will not be used until
5698 * rcu_assign_pointer(), so it's safe to increment
5701 ++new->current_threshold
;
5706 /* Free old spare buffer and save old primary buffer as spare */
5707 kfree(thresholds
->spare
);
5708 thresholds
->spare
= thresholds
->primary
;
5710 rcu_assign_pointer(thresholds
->primary
, new);
5712 /* To be sure that nobody uses thresholds */
5716 mutex_unlock(&memcg
->thresholds_lock
);
5721 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5722 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5724 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5725 struct mem_cgroup_thresholds
*thresholds
;
5726 struct mem_cgroup_threshold_ary
*new;
5727 enum res_type type
= MEMFILE_TYPE(cft
->private);
5731 mutex_lock(&memcg
->thresholds_lock
);
5733 thresholds
= &memcg
->thresholds
;
5734 else if (type
== _MEMSWAP
)
5735 thresholds
= &memcg
->memsw_thresholds
;
5739 if (!thresholds
->primary
)
5742 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5744 /* Check if a threshold crossed before removing */
5745 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5747 /* Calculate new number of threshold */
5749 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5750 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5754 new = thresholds
->spare
;
5756 /* Set thresholds array to NULL if we don't have thresholds */
5765 /* Copy thresholds and find current threshold */
5766 new->current_threshold
= -1;
5767 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5768 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5771 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5772 if (new->entries
[j
].threshold
<= usage
) {
5774 * new->current_threshold will not be used
5775 * until rcu_assign_pointer(), so it's safe to increment
5778 ++new->current_threshold
;
5784 /* Swap primary and spare array */
5785 thresholds
->spare
= thresholds
->primary
;
5786 /* If all events are unregistered, free the spare array */
5788 kfree(thresholds
->spare
);
5789 thresholds
->spare
= NULL
;
5792 rcu_assign_pointer(thresholds
->primary
, new);
5794 /* To be sure that nobody uses thresholds */
5797 mutex_unlock(&memcg
->thresholds_lock
);
5800 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5801 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5803 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5804 struct mem_cgroup_eventfd_list
*event
;
5805 enum res_type type
= MEMFILE_TYPE(cft
->private);
5807 BUG_ON(type
!= _OOM_TYPE
);
5808 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5812 spin_lock(&memcg_oom_lock
);
5814 event
->eventfd
= eventfd
;
5815 list_add(&event
->list
, &memcg
->oom_notify
);
5817 /* already in OOM ? */
5818 if (atomic_read(&memcg
->under_oom
))
5819 eventfd_signal(eventfd
, 1);
5820 spin_unlock(&memcg_oom_lock
);
5825 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5826 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5828 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5829 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5830 enum res_type type
= MEMFILE_TYPE(cft
->private);
5832 BUG_ON(type
!= _OOM_TYPE
);
5834 spin_lock(&memcg_oom_lock
);
5836 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5837 if (ev
->eventfd
== eventfd
) {
5838 list_del(&ev
->list
);
5843 spin_unlock(&memcg_oom_lock
);
5846 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5847 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5849 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5851 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5853 if (atomic_read(&memcg
->under_oom
))
5854 cb
->fill(cb
, "under_oom", 1);
5856 cb
->fill(cb
, "under_oom", 0);
5860 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5861 struct cftype
*cft
, u64 val
)
5863 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5864 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5866 /* cannot set to root cgroup and only 0 and 1 are allowed */
5867 if (!parent
|| !((val
== 0) || (val
== 1)))
5870 mutex_lock(&memcg_create_mutex
);
5871 /* oom-kill-disable is a flag for subhierarchy. */
5872 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5873 mutex_unlock(&memcg_create_mutex
);
5876 memcg
->oom_kill_disable
= val
;
5878 memcg_oom_recover(memcg
);
5879 mutex_unlock(&memcg_create_mutex
);
5883 #ifdef CONFIG_MEMCG_KMEM
5884 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5888 memcg
->kmemcg_id
= -1;
5889 ret
= memcg_propagate_kmem(memcg
);
5893 return mem_cgroup_sockets_init(memcg
, ss
);
5896 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5898 mem_cgroup_sockets_destroy(memcg
);
5901 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5903 if (!memcg_kmem_is_active(memcg
))
5907 * kmem charges can outlive the cgroup. In the case of slab
5908 * pages, for instance, a page contain objects from various
5909 * processes. As we prevent from taking a reference for every
5910 * such allocation we have to be careful when doing uncharge
5911 * (see memcg_uncharge_kmem) and here during offlining.
5913 * The idea is that that only the _last_ uncharge which sees
5914 * the dead memcg will drop the last reference. An additional
5915 * reference is taken here before the group is marked dead
5916 * which is then paired with css_put during uncharge resp. here.
5918 * Although this might sound strange as this path is called from
5919 * css_offline() when the referencemight have dropped down to 0
5920 * and shouldn't be incremented anymore (css_tryget would fail)
5921 * we do not have other options because of the kmem allocations
5924 css_get(&memcg
->css
);
5926 memcg_kmem_mark_dead(memcg
);
5928 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5931 if (memcg_kmem_test_and_clear_dead(memcg
))
5932 css_put(&memcg
->css
);
5935 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5940 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5944 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5949 static struct cftype mem_cgroup_files
[] = {
5951 .name
= "usage_in_bytes",
5952 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5953 .read
= mem_cgroup_read
,
5954 .register_event
= mem_cgroup_usage_register_event
,
5955 .unregister_event
= mem_cgroup_usage_unregister_event
,
5958 .name
= "max_usage_in_bytes",
5959 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5960 .trigger
= mem_cgroup_reset
,
5961 .read
= mem_cgroup_read
,
5964 .name
= "limit_in_bytes",
5965 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5966 .write_string
= mem_cgroup_write
,
5967 .read
= mem_cgroup_read
,
5970 .name
= "soft_limit_in_bytes",
5971 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5972 .write_string
= mem_cgroup_write
,
5973 .read
= mem_cgroup_read
,
5977 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5978 .trigger
= mem_cgroup_reset
,
5979 .read
= mem_cgroup_read
,
5983 .read_seq_string
= memcg_stat_show
,
5986 .name
= "force_empty",
5987 .trigger
= mem_cgroup_force_empty_write
,
5990 .name
= "use_hierarchy",
5991 .flags
= CFTYPE_INSANE
,
5992 .write_u64
= mem_cgroup_hierarchy_write
,
5993 .read_u64
= mem_cgroup_hierarchy_read
,
5996 .name
= "swappiness",
5997 .read_u64
= mem_cgroup_swappiness_read
,
5998 .write_u64
= mem_cgroup_swappiness_write
,
6001 .name
= "move_charge_at_immigrate",
6002 .read_u64
= mem_cgroup_move_charge_read
,
6003 .write_u64
= mem_cgroup_move_charge_write
,
6006 .name
= "oom_control",
6007 .read_map
= mem_cgroup_oom_control_read
,
6008 .write_u64
= mem_cgroup_oom_control_write
,
6009 .register_event
= mem_cgroup_oom_register_event
,
6010 .unregister_event
= mem_cgroup_oom_unregister_event
,
6011 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6014 .name
= "pressure_level",
6015 .register_event
= vmpressure_register_event
,
6016 .unregister_event
= vmpressure_unregister_event
,
6020 .name
= "numa_stat",
6021 .read_seq_string
= memcg_numa_stat_show
,
6024 #ifdef CONFIG_MEMCG_KMEM
6026 .name
= "kmem.limit_in_bytes",
6027 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6028 .write_string
= mem_cgroup_write
,
6029 .read
= mem_cgroup_read
,
6032 .name
= "kmem.usage_in_bytes",
6033 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6034 .read
= mem_cgroup_read
,
6037 .name
= "kmem.failcnt",
6038 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6039 .trigger
= mem_cgroup_reset
,
6040 .read
= mem_cgroup_read
,
6043 .name
= "kmem.max_usage_in_bytes",
6044 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6045 .trigger
= mem_cgroup_reset
,
6046 .read
= mem_cgroup_read
,
6048 #ifdef CONFIG_SLABINFO
6050 .name
= "kmem.slabinfo",
6051 .read_seq_string
= mem_cgroup_slabinfo_read
,
6055 { }, /* terminate */
6058 #ifdef CONFIG_MEMCG_SWAP
6059 static struct cftype memsw_cgroup_files
[] = {
6061 .name
= "memsw.usage_in_bytes",
6062 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6063 .read
= mem_cgroup_read
,
6064 .register_event
= mem_cgroup_usage_register_event
,
6065 .unregister_event
= mem_cgroup_usage_unregister_event
,
6068 .name
= "memsw.max_usage_in_bytes",
6069 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6070 .trigger
= mem_cgroup_reset
,
6071 .read
= mem_cgroup_read
,
6074 .name
= "memsw.limit_in_bytes",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6076 .write_string
= mem_cgroup_write
,
6077 .read
= mem_cgroup_read
,
6080 .name
= "memsw.failcnt",
6081 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6082 .trigger
= mem_cgroup_reset
,
6083 .read
= mem_cgroup_read
,
6085 { }, /* terminate */
6088 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6090 struct mem_cgroup_per_node
*pn
;
6091 struct mem_cgroup_per_zone
*mz
;
6092 int zone
, tmp
= node
;
6094 * This routine is called against possible nodes.
6095 * But it's BUG to call kmalloc() against offline node.
6097 * TODO: this routine can waste much memory for nodes which will
6098 * never be onlined. It's better to use memory hotplug callback
6101 if (!node_state(node
, N_NORMAL_MEMORY
))
6103 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6107 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6108 mz
= &pn
->zoneinfo
[zone
];
6109 lruvec_init(&mz
->lruvec
);
6110 mz
->usage_in_excess
= 0;
6111 mz
->on_tree
= false;
6114 memcg
->nodeinfo
[node
] = pn
;
6118 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6120 kfree(memcg
->nodeinfo
[node
]);
6123 static struct mem_cgroup
*mem_cgroup_alloc(void)
6125 struct mem_cgroup
*memcg
;
6126 size_t size
= memcg_size();
6128 /* Can be very big if nr_node_ids is very big */
6129 if (size
< PAGE_SIZE
)
6130 memcg
= kzalloc(size
, GFP_KERNEL
);
6132 memcg
= vzalloc(size
);
6137 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6140 spin_lock_init(&memcg
->pcp_counter_lock
);
6144 if (size
< PAGE_SIZE
)
6152 * At destroying mem_cgroup, references from swap_cgroup can remain.
6153 * (scanning all at force_empty is too costly...)
6155 * Instead of clearing all references at force_empty, we remember
6156 * the number of reference from swap_cgroup and free mem_cgroup when
6157 * it goes down to 0.
6159 * Removal of cgroup itself succeeds regardless of refs from swap.
6162 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6165 size_t size
= memcg_size();
6167 mem_cgroup_remove_from_trees(memcg
);
6168 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
6171 free_mem_cgroup_per_zone_info(memcg
, node
);
6173 free_percpu(memcg
->stat
);
6176 * We need to make sure that (at least for now), the jump label
6177 * destruction code runs outside of the cgroup lock. This is because
6178 * get_online_cpus(), which is called from the static_branch update,
6179 * can't be called inside the cgroup_lock. cpusets are the ones
6180 * enforcing this dependency, so if they ever change, we might as well.
6182 * schedule_work() will guarantee this happens. Be careful if you need
6183 * to move this code around, and make sure it is outside
6186 disarm_static_keys(memcg
);
6187 if (size
< PAGE_SIZE
)
6194 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6196 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6198 if (!memcg
->res
.parent
)
6200 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6202 EXPORT_SYMBOL(parent_mem_cgroup
);
6204 static void __init
mem_cgroup_soft_limit_tree_init(void)
6206 struct mem_cgroup_tree_per_node
*rtpn
;
6207 struct mem_cgroup_tree_per_zone
*rtpz
;
6208 int tmp
, node
, zone
;
6210 for_each_node(node
) {
6212 if (!node_state(node
, N_NORMAL_MEMORY
))
6214 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6217 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6219 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6220 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6221 rtpz
->rb_root
= RB_ROOT
;
6222 spin_lock_init(&rtpz
->lock
);
6227 static struct cgroup_subsys_state
* __ref
6228 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6230 struct mem_cgroup
*memcg
;
6231 long error
= -ENOMEM
;
6234 memcg
= mem_cgroup_alloc();
6236 return ERR_PTR(error
);
6239 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6243 if (parent_css
== NULL
) {
6244 root_mem_cgroup
= memcg
;
6245 res_counter_init(&memcg
->res
, NULL
);
6246 res_counter_init(&memcg
->memsw
, NULL
);
6247 res_counter_init(&memcg
->kmem
, NULL
);
6250 memcg
->last_scanned_node
= MAX_NUMNODES
;
6251 INIT_LIST_HEAD(&memcg
->oom_notify
);
6252 memcg
->move_charge_at_immigrate
= 0;
6253 mutex_init(&memcg
->thresholds_lock
);
6254 spin_lock_init(&memcg
->move_lock
);
6255 vmpressure_init(&memcg
->vmpressure
);
6260 __mem_cgroup_free(memcg
);
6261 return ERR_PTR(error
);
6265 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6267 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6268 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6274 mutex_lock(&memcg_create_mutex
);
6276 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6277 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6278 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6280 if (parent
->use_hierarchy
) {
6281 res_counter_init(&memcg
->res
, &parent
->res
);
6282 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6283 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6286 * No need to take a reference to the parent because cgroup
6287 * core guarantees its existence.
6290 res_counter_init(&memcg
->res
, NULL
);
6291 res_counter_init(&memcg
->memsw
, NULL
);
6292 res_counter_init(&memcg
->kmem
, NULL
);
6294 * Deeper hierachy with use_hierarchy == false doesn't make
6295 * much sense so let cgroup subsystem know about this
6296 * unfortunate state in our controller.
6298 if (parent
!= root_mem_cgroup
)
6299 mem_cgroup_subsys
.broken_hierarchy
= true;
6302 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6303 mutex_unlock(&memcg_create_mutex
);
6308 * Announce all parents that a group from their hierarchy is gone.
6310 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6312 struct mem_cgroup
*parent
= memcg
;
6314 while ((parent
= parent_mem_cgroup(parent
)))
6315 mem_cgroup_iter_invalidate(parent
);
6318 * if the root memcg is not hierarchical we have to check it
6321 if (!root_mem_cgroup
->use_hierarchy
)
6322 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6325 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6327 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6329 kmem_cgroup_css_offline(memcg
);
6331 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6332 mem_cgroup_reparent_charges(memcg
);
6333 mem_cgroup_destroy_all_caches(memcg
);
6334 vmpressure_cleanup(&memcg
->vmpressure
);
6337 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6339 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6341 memcg_destroy_kmem(memcg
);
6342 __mem_cgroup_free(memcg
);
6346 /* Handlers for move charge at task migration. */
6347 #define PRECHARGE_COUNT_AT_ONCE 256
6348 static int mem_cgroup_do_precharge(unsigned long count
)
6351 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6352 struct mem_cgroup
*memcg
= mc
.to
;
6354 if (mem_cgroup_is_root(memcg
)) {
6355 mc
.precharge
+= count
;
6356 /* we don't need css_get for root */
6359 /* try to charge at once */
6361 struct res_counter
*dummy
;
6363 * "memcg" cannot be under rmdir() because we've already checked
6364 * by cgroup_lock_live_cgroup() that it is not removed and we
6365 * are still under the same cgroup_mutex. So we can postpone
6368 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6370 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6371 PAGE_SIZE
* count
, &dummy
)) {
6372 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6375 mc
.precharge
+= count
;
6379 /* fall back to one by one charge */
6381 if (signal_pending(current
)) {
6385 if (!batch_count
--) {
6386 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6389 ret
= __mem_cgroup_try_charge(NULL
,
6390 GFP_KERNEL
, 1, &memcg
, false);
6392 /* mem_cgroup_clear_mc() will do uncharge later */
6400 * get_mctgt_type - get target type of moving charge
6401 * @vma: the vma the pte to be checked belongs
6402 * @addr: the address corresponding to the pte to be checked
6403 * @ptent: the pte to be checked
6404 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6407 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6408 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6409 * move charge. if @target is not NULL, the page is stored in target->page
6410 * with extra refcnt got(Callers should handle it).
6411 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6412 * target for charge migration. if @target is not NULL, the entry is stored
6415 * Called with pte lock held.
6422 enum mc_target_type
{
6428 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6429 unsigned long addr
, pte_t ptent
)
6431 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6433 if (!page
|| !page_mapped(page
))
6435 if (PageAnon(page
)) {
6436 /* we don't move shared anon */
6439 } else if (!move_file())
6440 /* we ignore mapcount for file pages */
6442 if (!get_page_unless_zero(page
))
6449 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6450 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6452 struct page
*page
= NULL
;
6453 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6455 if (!move_anon() || non_swap_entry(ent
))
6458 * Because lookup_swap_cache() updates some statistics counter,
6459 * we call find_get_page() with swapper_space directly.
6461 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6462 if (do_swap_account
)
6463 entry
->val
= ent
.val
;
6468 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6469 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6475 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6476 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6478 struct page
*page
= NULL
;
6479 struct address_space
*mapping
;
6482 if (!vma
->vm_file
) /* anonymous vma */
6487 mapping
= vma
->vm_file
->f_mapping
;
6488 if (pte_none(ptent
))
6489 pgoff
= linear_page_index(vma
, addr
);
6490 else /* pte_file(ptent) is true */
6491 pgoff
= pte_to_pgoff(ptent
);
6493 /* page is moved even if it's not RSS of this task(page-faulted). */
6494 page
= find_get_page(mapping
, pgoff
);
6497 /* shmem/tmpfs may report page out on swap: account for that too. */
6498 if (radix_tree_exceptional_entry(page
)) {
6499 swp_entry_t swap
= radix_to_swp_entry(page
);
6500 if (do_swap_account
)
6502 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6508 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6509 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6511 struct page
*page
= NULL
;
6512 struct page_cgroup
*pc
;
6513 enum mc_target_type ret
= MC_TARGET_NONE
;
6514 swp_entry_t ent
= { .val
= 0 };
6516 if (pte_present(ptent
))
6517 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6518 else if (is_swap_pte(ptent
))
6519 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6520 else if (pte_none(ptent
) || pte_file(ptent
))
6521 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6523 if (!page
&& !ent
.val
)
6526 pc
= lookup_page_cgroup(page
);
6528 * Do only loose check w/o page_cgroup lock.
6529 * mem_cgroup_move_account() checks the pc is valid or not under
6532 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6533 ret
= MC_TARGET_PAGE
;
6535 target
->page
= page
;
6537 if (!ret
|| !target
)
6540 /* There is a swap entry and a page doesn't exist or isn't charged */
6541 if (ent
.val
&& !ret
&&
6542 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6543 ret
= MC_TARGET_SWAP
;
6550 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6552 * We don't consider swapping or file mapped pages because THP does not
6553 * support them for now.
6554 * Caller should make sure that pmd_trans_huge(pmd) is true.
6556 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6557 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6559 struct page
*page
= NULL
;
6560 struct page_cgroup
*pc
;
6561 enum mc_target_type ret
= MC_TARGET_NONE
;
6563 page
= pmd_page(pmd
);
6564 VM_BUG_ON(!page
|| !PageHead(page
));
6567 pc
= lookup_page_cgroup(page
);
6568 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6569 ret
= MC_TARGET_PAGE
;
6572 target
->page
= page
;
6578 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6579 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6581 return MC_TARGET_NONE
;
6585 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6586 unsigned long addr
, unsigned long end
,
6587 struct mm_walk
*walk
)
6589 struct vm_area_struct
*vma
= walk
->private;
6593 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6594 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6595 mc
.precharge
+= HPAGE_PMD_NR
;
6596 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6600 if (pmd_trans_unstable(pmd
))
6602 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6603 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6604 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6605 mc
.precharge
++; /* increment precharge temporarily */
6606 pte_unmap_unlock(pte
- 1, ptl
);
6612 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6614 unsigned long precharge
;
6615 struct vm_area_struct
*vma
;
6617 down_read(&mm
->mmap_sem
);
6618 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6619 struct mm_walk mem_cgroup_count_precharge_walk
= {
6620 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6624 if (is_vm_hugetlb_page(vma
))
6626 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6627 &mem_cgroup_count_precharge_walk
);
6629 up_read(&mm
->mmap_sem
);
6631 precharge
= mc
.precharge
;
6637 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6639 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6641 VM_BUG_ON(mc
.moving_task
);
6642 mc
.moving_task
= current
;
6643 return mem_cgroup_do_precharge(precharge
);
6646 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6647 static void __mem_cgroup_clear_mc(void)
6649 struct mem_cgroup
*from
= mc
.from
;
6650 struct mem_cgroup
*to
= mc
.to
;
6653 /* we must uncharge all the leftover precharges from mc.to */
6655 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6659 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6660 * we must uncharge here.
6662 if (mc
.moved_charge
) {
6663 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6664 mc
.moved_charge
= 0;
6666 /* we must fixup refcnts and charges */
6667 if (mc
.moved_swap
) {
6668 /* uncharge swap account from the old cgroup */
6669 if (!mem_cgroup_is_root(mc
.from
))
6670 res_counter_uncharge(&mc
.from
->memsw
,
6671 PAGE_SIZE
* mc
.moved_swap
);
6673 for (i
= 0; i
< mc
.moved_swap
; i
++)
6674 css_put(&mc
.from
->css
);
6676 if (!mem_cgroup_is_root(mc
.to
)) {
6678 * we charged both to->res and to->memsw, so we should
6681 res_counter_uncharge(&mc
.to
->res
,
6682 PAGE_SIZE
* mc
.moved_swap
);
6684 /* we've already done css_get(mc.to) */
6687 memcg_oom_recover(from
);
6688 memcg_oom_recover(to
);
6689 wake_up_all(&mc
.waitq
);
6692 static void mem_cgroup_clear_mc(void)
6694 struct mem_cgroup
*from
= mc
.from
;
6697 * we must clear moving_task before waking up waiters at the end of
6700 mc
.moving_task
= NULL
;
6701 __mem_cgroup_clear_mc();
6702 spin_lock(&mc
.lock
);
6705 spin_unlock(&mc
.lock
);
6706 mem_cgroup_end_move(from
);
6709 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6710 struct cgroup_taskset
*tset
)
6712 struct task_struct
*p
= cgroup_taskset_first(tset
);
6714 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6715 unsigned long move_charge_at_immigrate
;
6718 * We are now commited to this value whatever it is. Changes in this
6719 * tunable will only affect upcoming migrations, not the current one.
6720 * So we need to save it, and keep it going.
6722 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6723 if (move_charge_at_immigrate
) {
6724 struct mm_struct
*mm
;
6725 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6727 VM_BUG_ON(from
== memcg
);
6729 mm
= get_task_mm(p
);
6732 /* We move charges only when we move a owner of the mm */
6733 if (mm
->owner
== p
) {
6736 VM_BUG_ON(mc
.precharge
);
6737 VM_BUG_ON(mc
.moved_charge
);
6738 VM_BUG_ON(mc
.moved_swap
);
6739 mem_cgroup_start_move(from
);
6740 spin_lock(&mc
.lock
);
6743 mc
.immigrate_flags
= move_charge_at_immigrate
;
6744 spin_unlock(&mc
.lock
);
6745 /* We set mc.moving_task later */
6747 ret
= mem_cgroup_precharge_mc(mm
);
6749 mem_cgroup_clear_mc();
6756 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6757 struct cgroup_taskset
*tset
)
6759 mem_cgroup_clear_mc();
6762 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6763 unsigned long addr
, unsigned long end
,
6764 struct mm_walk
*walk
)
6767 struct vm_area_struct
*vma
= walk
->private;
6770 enum mc_target_type target_type
;
6771 union mc_target target
;
6773 struct page_cgroup
*pc
;
6776 * We don't take compound_lock() here but no race with splitting thp
6778 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6779 * under splitting, which means there's no concurrent thp split,
6780 * - if another thread runs into split_huge_page() just after we
6781 * entered this if-block, the thread must wait for page table lock
6782 * to be unlocked in __split_huge_page_splitting(), where the main
6783 * part of thp split is not executed yet.
6785 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6786 if (mc
.precharge
< HPAGE_PMD_NR
) {
6787 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6790 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6791 if (target_type
== MC_TARGET_PAGE
) {
6793 if (!isolate_lru_page(page
)) {
6794 pc
= lookup_page_cgroup(page
);
6795 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6796 pc
, mc
.from
, mc
.to
)) {
6797 mc
.precharge
-= HPAGE_PMD_NR
;
6798 mc
.moved_charge
+= HPAGE_PMD_NR
;
6800 putback_lru_page(page
);
6804 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6808 if (pmd_trans_unstable(pmd
))
6811 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6812 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6813 pte_t ptent
= *(pte
++);
6819 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6820 case MC_TARGET_PAGE
:
6822 if (isolate_lru_page(page
))
6824 pc
= lookup_page_cgroup(page
);
6825 if (!mem_cgroup_move_account(page
, 1, pc
,
6828 /* we uncharge from mc.from later. */
6831 putback_lru_page(page
);
6832 put
: /* get_mctgt_type() gets the page */
6835 case MC_TARGET_SWAP
:
6837 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6839 /* we fixup refcnts and charges later. */
6847 pte_unmap_unlock(pte
- 1, ptl
);
6852 * We have consumed all precharges we got in can_attach().
6853 * We try charge one by one, but don't do any additional
6854 * charges to mc.to if we have failed in charge once in attach()
6857 ret
= mem_cgroup_do_precharge(1);
6865 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6867 struct vm_area_struct
*vma
;
6869 lru_add_drain_all();
6871 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6873 * Someone who are holding the mmap_sem might be waiting in
6874 * waitq. So we cancel all extra charges, wake up all waiters,
6875 * and retry. Because we cancel precharges, we might not be able
6876 * to move enough charges, but moving charge is a best-effort
6877 * feature anyway, so it wouldn't be a big problem.
6879 __mem_cgroup_clear_mc();
6883 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6885 struct mm_walk mem_cgroup_move_charge_walk
= {
6886 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6890 if (is_vm_hugetlb_page(vma
))
6892 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6893 &mem_cgroup_move_charge_walk
);
6896 * means we have consumed all precharges and failed in
6897 * doing additional charge. Just abandon here.
6901 up_read(&mm
->mmap_sem
);
6904 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6905 struct cgroup_taskset
*tset
)
6907 struct task_struct
*p
= cgroup_taskset_first(tset
);
6908 struct mm_struct
*mm
= get_task_mm(p
);
6912 mem_cgroup_move_charge(mm
);
6916 mem_cgroup_clear_mc();
6918 #else /* !CONFIG_MMU */
6919 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6920 struct cgroup_taskset
*tset
)
6924 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6925 struct cgroup_taskset
*tset
)
6928 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6929 struct cgroup_taskset
*tset
)
6935 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6936 * to verify sane_behavior flag on each mount attempt.
6938 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6941 * use_hierarchy is forced with sane_behavior. cgroup core
6942 * guarantees that @root doesn't have any children, so turning it
6943 * on for the root memcg is enough.
6945 if (cgroup_sane_behavior(root_css
->cgroup
))
6946 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6949 struct cgroup_subsys mem_cgroup_subsys
= {
6951 .subsys_id
= mem_cgroup_subsys_id
,
6952 .css_alloc
= mem_cgroup_css_alloc
,
6953 .css_online
= mem_cgroup_css_online
,
6954 .css_offline
= mem_cgroup_css_offline
,
6955 .css_free
= mem_cgroup_css_free
,
6956 .can_attach
= mem_cgroup_can_attach
,
6957 .cancel_attach
= mem_cgroup_cancel_attach
,
6958 .attach
= mem_cgroup_move_task
,
6959 .bind
= mem_cgroup_bind
,
6960 .base_cftypes
= mem_cgroup_files
,
6965 #ifdef CONFIG_MEMCG_SWAP
6966 static int __init
enable_swap_account(char *s
)
6968 if (!strcmp(s
, "1"))
6969 really_do_swap_account
= 1;
6970 else if (!strcmp(s
, "0"))
6971 really_do_swap_account
= 0;
6974 __setup("swapaccount=", enable_swap_account
);
6976 static void __init
memsw_file_init(void)
6978 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6981 static void __init
enable_swap_cgroup(void)
6983 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6984 do_swap_account
= 1;
6990 static void __init
enable_swap_cgroup(void)
6996 * subsys_initcall() for memory controller.
6998 * Some parts like hotcpu_notifier() have to be initialized from this context
6999 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7000 * everything that doesn't depend on a specific mem_cgroup structure should
7001 * be initialized from here.
7003 static int __init
mem_cgroup_init(void)
7005 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7006 enable_swap_cgroup();
7007 mem_cgroup_soft_limit_tree_init();
7011 subsys_initcall(mem_cgroup_init
);