1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
66 EXPORT_SYMBOL(mem_cgroup_subsys
);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly
;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata
= 1;
79 static int really_do_swap_account __initdata
= 0;
83 #define do_swap_account 0
88 * Statistics for memory cgroup.
90 enum mem_cgroup_stat_index
{
92 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
95 MEM_CGROUP_STAT_RSS
, /* # of pages charged as anon rss */
96 MEM_CGROUP_STAT_FILE_MAPPED
, /* # of pages charged as file rss */
97 MEM_CGROUP_STAT_SWAP
, /* # of pages, swapped out */
98 MEM_CGROUP_STAT_NSTATS
,
101 static const char * const mem_cgroup_stat_names
[] = {
108 enum mem_cgroup_events_index
{
109 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
110 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
111 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
112 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
113 MEM_CGROUP_EVENTS_NSTATS
,
116 static const char * const mem_cgroup_events_names
[] = {
123 static const char * const mem_cgroup_lru_names
[] = {
132 * Per memcg event counter is incremented at every pagein/pageout. With THP,
133 * it will be incremated by the number of pages. This counter is used for
134 * for trigger some periodic events. This is straightforward and better
135 * than using jiffies etc. to handle periodic memcg event.
137 enum mem_cgroup_events_target
{
138 MEM_CGROUP_TARGET_THRESH
,
139 MEM_CGROUP_TARGET_SOFTLIMIT
,
140 MEM_CGROUP_TARGET_NUMAINFO
,
143 #define THRESHOLDS_EVENTS_TARGET 128
144 #define SOFTLIMIT_EVENTS_TARGET 1024
145 #define NUMAINFO_EVENTS_TARGET 1024
147 struct mem_cgroup_stat_cpu
{
148 long count
[MEM_CGROUP_STAT_NSTATS
];
149 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
150 unsigned long nr_page_events
;
151 unsigned long targets
[MEM_CGROUP_NTARGETS
];
154 struct mem_cgroup_reclaim_iter
{
155 /* css_id of the last scanned hierarchy member */
157 /* scan generation, increased every round-trip */
158 unsigned int generation
;
162 * per-zone information in memory controller.
164 struct mem_cgroup_per_zone
{
165 struct lruvec lruvec
;
166 unsigned long lru_size
[NR_LRU_LISTS
];
168 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
170 struct rb_node tree_node
; /* RB tree node */
171 unsigned long long usage_in_excess
;/* Set to the value by which */
172 /* the soft limit is exceeded*/
174 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
175 /* use container_of */
178 struct mem_cgroup_per_node
{
179 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
182 struct mem_cgroup_lru_info
{
183 struct mem_cgroup_per_node
*nodeinfo
[0];
187 * Cgroups above their limits are maintained in a RB-Tree, independent of
188 * their hierarchy representation
191 struct mem_cgroup_tree_per_zone
{
192 struct rb_root rb_root
;
196 struct mem_cgroup_tree_per_node
{
197 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
200 struct mem_cgroup_tree
{
201 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
204 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
206 struct mem_cgroup_threshold
{
207 struct eventfd_ctx
*eventfd
;
212 struct mem_cgroup_threshold_ary
{
213 /* An array index points to threshold just below or equal to usage. */
214 int current_threshold
;
215 /* Size of entries[] */
217 /* Array of thresholds */
218 struct mem_cgroup_threshold entries
[0];
221 struct mem_cgroup_thresholds
{
222 /* Primary thresholds array */
223 struct mem_cgroup_threshold_ary
*primary
;
225 * Spare threshold array.
226 * This is needed to make mem_cgroup_unregister_event() "never fail".
227 * It must be able to store at least primary->size - 1 entries.
229 struct mem_cgroup_threshold_ary
*spare
;
233 struct mem_cgroup_eventfd_list
{
234 struct list_head list
;
235 struct eventfd_ctx
*eventfd
;
238 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
239 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
242 * The memory controller data structure. The memory controller controls both
243 * page cache and RSS per cgroup. We would eventually like to provide
244 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
245 * to help the administrator determine what knobs to tune.
247 * TODO: Add a water mark for the memory controller. Reclaim will begin when
248 * we hit the water mark. May be even add a low water mark, such that
249 * no reclaim occurs from a cgroup at it's low water mark, this is
250 * a feature that will be implemented much later in the future.
253 struct cgroup_subsys_state css
;
255 * the counter to account for memory usage
257 struct res_counter res
;
261 * the counter to account for mem+swap usage.
263 struct res_counter memsw
;
266 * rcu_freeing is used only when freeing struct mem_cgroup,
267 * so put it into a union to avoid wasting more memory.
268 * It must be disjoint from the css field. It could be
269 * in a union with the res field, but res plays a much
270 * larger part in mem_cgroup life than memsw, and might
271 * be of interest, even at time of free, when debugging.
272 * So share rcu_head with the less interesting memsw.
274 struct rcu_head rcu_freeing
;
276 * We also need some space for a worker in deferred freeing.
277 * By the time we call it, rcu_freeing is no longer in use.
279 struct work_struct work_freeing
;
283 * the counter to account for kernel memory usage.
285 struct res_counter kmem
;
287 * Should the accounting and control be hierarchical, per subtree?
290 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
298 /* OOM-Killer disable */
299 int oom_kill_disable
;
301 /* set when res.limit == memsw.limit */
302 bool memsw_is_minimum
;
304 /* protect arrays of thresholds */
305 struct mutex thresholds_lock
;
307 /* thresholds for memory usage. RCU-protected */
308 struct mem_cgroup_thresholds thresholds
;
310 /* thresholds for mem+swap usage. RCU-protected */
311 struct mem_cgroup_thresholds memsw_thresholds
;
313 /* For oom notifier event fd */
314 struct list_head oom_notify
;
317 * Should we move charges of a task when a task is moved into this
318 * mem_cgroup ? And what type of charges should we move ?
320 unsigned long move_charge_at_immigrate
;
322 * set > 0 if pages under this cgroup are moving to other cgroup.
324 atomic_t moving_account
;
325 /* taken only while moving_account > 0 */
326 spinlock_t move_lock
;
330 struct mem_cgroup_stat_cpu __percpu
*stat
;
332 * used when a cpu is offlined or other synchronizations
333 * See mem_cgroup_read_stat().
335 struct mem_cgroup_stat_cpu nocpu_base
;
336 spinlock_t pcp_counter_lock
;
338 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
339 struct tcp_memcontrol tcp_mem
;
341 #if defined(CONFIG_MEMCG_KMEM)
342 /* analogous to slab_common's slab_caches list. per-memcg */
343 struct list_head memcg_slab_caches
;
344 /* Not a spinlock, we can take a lot of time walking the list */
345 struct mutex slab_caches_mutex
;
346 /* Index in the kmem_cache->memcg_params->memcg_caches array */
350 int last_scanned_node
;
352 nodemask_t scan_nodes
;
353 atomic_t numainfo_events
;
354 atomic_t numainfo_updating
;
357 * Per cgroup active and inactive list, similar to the
358 * per zone LRU lists.
360 * WARNING: This has to be the last element of the struct. Don't
361 * add new fields after this point.
363 struct mem_cgroup_lru_info info
;
366 static size_t memcg_size(void)
368 return sizeof(struct mem_cgroup
) +
369 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
372 /* internal only representation about the status of kmem accounting. */
374 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
375 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
376 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
379 /* We account when limit is on, but only after call sites are patched */
380 #define KMEM_ACCOUNTED_MASK \
381 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
383 #ifdef CONFIG_MEMCG_KMEM
384 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
386 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
389 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
391 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
394 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
396 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
399 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
401 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
404 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
407 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
413 &memcg
->kmem_account_flags
);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct
{
430 spinlock_t lock
; /* for from, to */
431 struct mem_cgroup
*from
;
432 struct mem_cgroup
*to
;
433 unsigned long immigrate_flags
;
434 unsigned long precharge
;
435 unsigned long moved_charge
;
436 unsigned long moved_swap
;
437 struct task_struct
*moving_task
; /* a task moving charges */
438 wait_queue_head_t waitq
; /* a waitq for other context */
440 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
441 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON
,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
486 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
487 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
488 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
489 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
492 * The memcg_create_mutex will be held whenever a new cgroup is created.
493 * As a consequence, any change that needs to protect against new child cgroups
494 * appearing has to hold it as well.
496 static DEFINE_MUTEX(memcg_create_mutex
);
498 static void mem_cgroup_get(struct mem_cgroup
*memcg
);
499 static void mem_cgroup_put(struct mem_cgroup
*memcg
);
502 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
504 return container_of(s
, struct mem_cgroup
, css
);
507 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
509 return (memcg
== root_mem_cgroup
);
512 /* Writing them here to avoid exposing memcg's inner layout */
513 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
515 void sock_update_memcg(struct sock
*sk
)
517 if (mem_cgroup_sockets_enabled
) {
518 struct mem_cgroup
*memcg
;
519 struct cg_proto
*cg_proto
;
521 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
523 /* Socket cloning can throw us here with sk_cgrp already
524 * filled. It won't however, necessarily happen from
525 * process context. So the test for root memcg given
526 * the current task's memcg won't help us in this case.
528 * Respecting the original socket's memcg is a better
529 * decision in this case.
532 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
533 mem_cgroup_get(sk
->sk_cgrp
->memcg
);
538 memcg
= mem_cgroup_from_task(current
);
539 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
540 if (!mem_cgroup_is_root(memcg
) && memcg_proto_active(cg_proto
)) {
541 mem_cgroup_get(memcg
);
542 sk
->sk_cgrp
= cg_proto
;
547 EXPORT_SYMBOL(sock_update_memcg
);
549 void sock_release_memcg(struct sock
*sk
)
551 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
552 struct mem_cgroup
*memcg
;
553 WARN_ON(!sk
->sk_cgrp
->memcg
);
554 memcg
= sk
->sk_cgrp
->memcg
;
555 mem_cgroup_put(memcg
);
559 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
561 if (!memcg
|| mem_cgroup_is_root(memcg
))
564 return &memcg
->tcp_mem
.cg_proto
;
566 EXPORT_SYMBOL(tcp_proto_cgroup
);
568 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
570 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
572 static_key_slow_dec(&memcg_socket_limit_enabled
);
575 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
580 #ifdef CONFIG_MEMCG_KMEM
582 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
583 * There are two main reasons for not using the css_id for this:
584 * 1) this works better in sparse environments, where we have a lot of memcgs,
585 * but only a few kmem-limited. Or also, if we have, for instance, 200
586 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
587 * 200 entry array for that.
589 * 2) In order not to violate the cgroup API, we would like to do all memory
590 * allocation in ->create(). At that point, we haven't yet allocated the
591 * css_id. Having a separate index prevents us from messing with the cgroup
594 * The current size of the caches array is stored in
595 * memcg_limited_groups_array_size. It will double each time we have to
598 static DEFINE_IDA(kmem_limited_groups
);
599 int memcg_limited_groups_array_size
;
602 * MIN_SIZE is different than 1, because we would like to avoid going through
603 * the alloc/free process all the time. In a small machine, 4 kmem-limited
604 * cgroups is a reasonable guess. In the future, it could be a parameter or
605 * tunable, but that is strictly not necessary.
607 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
608 * this constant directly from cgroup, but it is understandable that this is
609 * better kept as an internal representation in cgroup.c. In any case, the
610 * css_id space is not getting any smaller, and we don't have to necessarily
611 * increase ours as well if it increases.
613 #define MEMCG_CACHES_MIN_SIZE 4
614 #define MEMCG_CACHES_MAX_SIZE 65535
617 * A lot of the calls to the cache allocation functions are expected to be
618 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
619 * conditional to this static branch, we'll have to allow modules that does
620 * kmem_cache_alloc and the such to see this symbol as well
622 struct static_key memcg_kmem_enabled_key
;
623 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
625 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
627 if (memcg_kmem_is_active(memcg
)) {
628 static_key_slow_dec(&memcg_kmem_enabled_key
);
629 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
632 * This check can't live in kmem destruction function,
633 * since the charges will outlive the cgroup
635 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
638 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
641 #endif /* CONFIG_MEMCG_KMEM */
643 static void disarm_static_keys(struct mem_cgroup
*memcg
)
645 disarm_sock_keys(memcg
);
646 disarm_kmem_keys(memcg
);
649 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
651 static struct mem_cgroup_per_zone
*
652 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
654 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
655 return &memcg
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
658 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
663 static struct mem_cgroup_per_zone
*
664 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
666 int nid
= page_to_nid(page
);
667 int zid
= page_zonenum(page
);
669 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
672 static struct mem_cgroup_tree_per_zone
*
673 soft_limit_tree_node_zone(int nid
, int zid
)
675 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
678 static struct mem_cgroup_tree_per_zone
*
679 soft_limit_tree_from_page(struct page
*page
)
681 int nid
= page_to_nid(page
);
682 int zid
= page_zonenum(page
);
684 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
688 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
689 struct mem_cgroup_per_zone
*mz
,
690 struct mem_cgroup_tree_per_zone
*mctz
,
691 unsigned long long new_usage_in_excess
)
693 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
694 struct rb_node
*parent
= NULL
;
695 struct mem_cgroup_per_zone
*mz_node
;
700 mz
->usage_in_excess
= new_usage_in_excess
;
701 if (!mz
->usage_in_excess
)
705 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
707 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
710 * We can't avoid mem cgroups that are over their soft
711 * limit by the same amount
713 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
716 rb_link_node(&mz
->tree_node
, parent
, p
);
717 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
722 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
723 struct mem_cgroup_per_zone
*mz
,
724 struct mem_cgroup_tree_per_zone
*mctz
)
728 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
733 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
734 struct mem_cgroup_per_zone
*mz
,
735 struct mem_cgroup_tree_per_zone
*mctz
)
737 spin_lock(&mctz
->lock
);
738 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
739 spin_unlock(&mctz
->lock
);
743 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
745 unsigned long long excess
;
746 struct mem_cgroup_per_zone
*mz
;
747 struct mem_cgroup_tree_per_zone
*mctz
;
748 int nid
= page_to_nid(page
);
749 int zid
= page_zonenum(page
);
750 mctz
= soft_limit_tree_from_page(page
);
753 * Necessary to update all ancestors when hierarchy is used.
754 * because their event counter is not touched.
756 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
757 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
758 excess
= res_counter_soft_limit_excess(&memcg
->res
);
760 * We have to update the tree if mz is on RB-tree or
761 * mem is over its softlimit.
763 if (excess
|| mz
->on_tree
) {
764 spin_lock(&mctz
->lock
);
765 /* if on-tree, remove it */
767 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
769 * Insert again. mz->usage_in_excess will be updated.
770 * If excess is 0, no tree ops.
772 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
773 spin_unlock(&mctz
->lock
);
778 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
781 struct mem_cgroup_per_zone
*mz
;
782 struct mem_cgroup_tree_per_zone
*mctz
;
784 for_each_node(node
) {
785 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
786 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
787 mctz
= soft_limit_tree_node_zone(node
, zone
);
788 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
793 static struct mem_cgroup_per_zone
*
794 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
796 struct rb_node
*rightmost
= NULL
;
797 struct mem_cgroup_per_zone
*mz
;
801 rightmost
= rb_last(&mctz
->rb_root
);
803 goto done
; /* Nothing to reclaim from */
805 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
807 * Remove the node now but someone else can add it back,
808 * we will to add it back at the end of reclaim to its correct
809 * position in the tree.
811 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
812 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
813 !css_tryget(&mz
->memcg
->css
))
819 static struct mem_cgroup_per_zone
*
820 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
822 struct mem_cgroup_per_zone
*mz
;
824 spin_lock(&mctz
->lock
);
825 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
826 spin_unlock(&mctz
->lock
);
831 * Implementation Note: reading percpu statistics for memcg.
833 * Both of vmstat[] and percpu_counter has threshold and do periodic
834 * synchronization to implement "quick" read. There are trade-off between
835 * reading cost and precision of value. Then, we may have a chance to implement
836 * a periodic synchronizion of counter in memcg's counter.
838 * But this _read() function is used for user interface now. The user accounts
839 * memory usage by memory cgroup and he _always_ requires exact value because
840 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
841 * have to visit all online cpus and make sum. So, for now, unnecessary
842 * synchronization is not implemented. (just implemented for cpu hotplug)
844 * If there are kernel internal actions which can make use of some not-exact
845 * value, and reading all cpu value can be performance bottleneck in some
846 * common workload, threashold and synchonization as vmstat[] should be
849 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
850 enum mem_cgroup_stat_index idx
)
856 for_each_online_cpu(cpu
)
857 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
858 #ifdef CONFIG_HOTPLUG_CPU
859 spin_lock(&memcg
->pcp_counter_lock
);
860 val
+= memcg
->nocpu_base
.count
[idx
];
861 spin_unlock(&memcg
->pcp_counter_lock
);
867 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
870 int val
= (charge
) ? 1 : -1;
871 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
874 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
875 enum mem_cgroup_events_index idx
)
877 unsigned long val
= 0;
880 for_each_online_cpu(cpu
)
881 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
882 #ifdef CONFIG_HOTPLUG_CPU
883 spin_lock(&memcg
->pcp_counter_lock
);
884 val
+= memcg
->nocpu_base
.events
[idx
];
885 spin_unlock(&memcg
->pcp_counter_lock
);
890 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
891 bool anon
, int nr_pages
)
896 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
897 * counted as CACHE even if it's on ANON LRU.
900 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
903 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
906 /* pagein of a big page is an event. So, ignore page size */
908 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
910 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
911 nr_pages
= -nr_pages
; /* for event */
914 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
920 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
922 struct mem_cgroup_per_zone
*mz
;
924 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
925 return mz
->lru_size
[lru
];
929 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
930 unsigned int lru_mask
)
932 struct mem_cgroup_per_zone
*mz
;
934 unsigned long ret
= 0;
936 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
939 if (BIT(lru
) & lru_mask
)
940 ret
+= mz
->lru_size
[lru
];
946 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
947 int nid
, unsigned int lru_mask
)
952 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
953 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
959 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
960 unsigned int lru_mask
)
965 for_each_node_state(nid
, N_MEMORY
)
966 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
970 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
971 enum mem_cgroup_events_target target
)
973 unsigned long val
, next
;
975 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
976 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
977 /* from time_after() in jiffies.h */
978 if ((long)next
- (long)val
< 0) {
980 case MEM_CGROUP_TARGET_THRESH
:
981 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
983 case MEM_CGROUP_TARGET_SOFTLIMIT
:
984 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
986 case MEM_CGROUP_TARGET_NUMAINFO
:
987 next
= val
+ NUMAINFO_EVENTS_TARGET
;
992 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
999 * Check events in order.
1002 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1005 /* threshold event is triggered in finer grain than soft limit */
1006 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1007 MEM_CGROUP_TARGET_THRESH
))) {
1009 bool do_numainfo __maybe_unused
;
1011 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1012 MEM_CGROUP_TARGET_SOFTLIMIT
);
1013 #if MAX_NUMNODES > 1
1014 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1015 MEM_CGROUP_TARGET_NUMAINFO
);
1019 mem_cgroup_threshold(memcg
);
1020 if (unlikely(do_softlimit
))
1021 mem_cgroup_update_tree(memcg
, page
);
1022 #if MAX_NUMNODES > 1
1023 if (unlikely(do_numainfo
))
1024 atomic_inc(&memcg
->numainfo_events
);
1030 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
1032 return mem_cgroup_from_css(
1033 cgroup_subsys_state(cont
, mem_cgroup_subsys_id
));
1036 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1039 * mm_update_next_owner() may clear mm->owner to NULL
1040 * if it races with swapoff, page migration, etc.
1041 * So this can be called with p == NULL.
1046 return mem_cgroup_from_css(task_subsys_state(p
, mem_cgroup_subsys_id
));
1049 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1051 struct mem_cgroup
*memcg
= NULL
;
1056 * Because we have no locks, mm->owner's may be being moved to other
1057 * cgroup. We use css_tryget() here even if this looks
1058 * pessimistic (rather than adding locks here).
1062 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1063 if (unlikely(!memcg
))
1065 } while (!css_tryget(&memcg
->css
));
1071 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1072 * @root: hierarchy root
1073 * @prev: previously returned memcg, NULL on first invocation
1074 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1076 * Returns references to children of the hierarchy below @root, or
1077 * @root itself, or %NULL after a full round-trip.
1079 * Caller must pass the return value in @prev on subsequent
1080 * invocations for reference counting, or use mem_cgroup_iter_break()
1081 * to cancel a hierarchy walk before the round-trip is complete.
1083 * Reclaimers can specify a zone and a priority level in @reclaim to
1084 * divide up the memcgs in the hierarchy among all concurrent
1085 * reclaimers operating on the same zone and priority.
1087 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1088 struct mem_cgroup
*prev
,
1089 struct mem_cgroup_reclaim_cookie
*reclaim
)
1091 struct mem_cgroup
*memcg
= NULL
;
1094 if (mem_cgroup_disabled())
1098 root
= root_mem_cgroup
;
1100 if (prev
&& !reclaim
)
1101 id
= css_id(&prev
->css
);
1103 if (prev
&& prev
!= root
)
1104 css_put(&prev
->css
);
1106 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1113 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1114 struct cgroup_subsys_state
*css
;
1117 int nid
= zone_to_nid(reclaim
->zone
);
1118 int zid
= zone_idx(reclaim
->zone
);
1119 struct mem_cgroup_per_zone
*mz
;
1121 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1122 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1123 if (prev
&& reclaim
->generation
!= iter
->generation
)
1125 id
= iter
->position
;
1129 css
= css_get_next(&mem_cgroup_subsys
, id
+ 1, &root
->css
, &id
);
1131 if (css
== &root
->css
|| css_tryget(css
))
1132 memcg
= mem_cgroup_from_css(css
);
1138 iter
->position
= id
;
1141 else if (!prev
&& memcg
)
1142 reclaim
->generation
= iter
->generation
;
1152 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1153 * @root: hierarchy root
1154 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1156 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1157 struct mem_cgroup
*prev
)
1160 root
= root_mem_cgroup
;
1161 if (prev
&& prev
!= root
)
1162 css_put(&prev
->css
);
1166 * Iteration constructs for visiting all cgroups (under a tree). If
1167 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1168 * be used for reference counting.
1170 #define for_each_mem_cgroup_tree(iter, root) \
1171 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1173 iter = mem_cgroup_iter(root, iter, NULL))
1175 #define for_each_mem_cgroup(iter) \
1176 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1178 iter = mem_cgroup_iter(NULL, iter, NULL))
1180 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1182 struct mem_cgroup
*memcg
;
1185 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1186 if (unlikely(!memcg
))
1191 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1194 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1202 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1205 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1206 * @zone: zone of the wanted lruvec
1207 * @memcg: memcg of the wanted lruvec
1209 * Returns the lru list vector holding pages for the given @zone and
1210 * @mem. This can be the global zone lruvec, if the memory controller
1213 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1214 struct mem_cgroup
*memcg
)
1216 struct mem_cgroup_per_zone
*mz
;
1217 struct lruvec
*lruvec
;
1219 if (mem_cgroup_disabled()) {
1220 lruvec
= &zone
->lruvec
;
1224 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1225 lruvec
= &mz
->lruvec
;
1228 * Since a node can be onlined after the mem_cgroup was created,
1229 * we have to be prepared to initialize lruvec->zone here;
1230 * and if offlined then reonlined, we need to reinitialize it.
1232 if (unlikely(lruvec
->zone
!= zone
))
1233 lruvec
->zone
= zone
;
1238 * Following LRU functions are allowed to be used without PCG_LOCK.
1239 * Operations are called by routine of global LRU independently from memcg.
1240 * What we have to take care of here is validness of pc->mem_cgroup.
1242 * Changes to pc->mem_cgroup happens when
1245 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1246 * It is added to LRU before charge.
1247 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1248 * When moving account, the page is not on LRU. It's isolated.
1252 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1254 * @zone: zone of the page
1256 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1258 struct mem_cgroup_per_zone
*mz
;
1259 struct mem_cgroup
*memcg
;
1260 struct page_cgroup
*pc
;
1261 struct lruvec
*lruvec
;
1263 if (mem_cgroup_disabled()) {
1264 lruvec
= &zone
->lruvec
;
1268 pc
= lookup_page_cgroup(page
);
1269 memcg
= pc
->mem_cgroup
;
1272 * Surreptitiously switch any uncharged offlist page to root:
1273 * an uncharged page off lru does nothing to secure
1274 * its former mem_cgroup from sudden removal.
1276 * Our caller holds lru_lock, and PageCgroupUsed is updated
1277 * under page_cgroup lock: between them, they make all uses
1278 * of pc->mem_cgroup safe.
1280 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1281 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1283 mz
= page_cgroup_zoneinfo(memcg
, page
);
1284 lruvec
= &mz
->lruvec
;
1287 * Since a node can be onlined after the mem_cgroup was created,
1288 * we have to be prepared to initialize lruvec->zone here;
1289 * and if offlined then reonlined, we need to reinitialize it.
1291 if (unlikely(lruvec
->zone
!= zone
))
1292 lruvec
->zone
= zone
;
1297 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1298 * @lruvec: mem_cgroup per zone lru vector
1299 * @lru: index of lru list the page is sitting on
1300 * @nr_pages: positive when adding or negative when removing
1302 * This function must be called when a page is added to or removed from an
1305 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1308 struct mem_cgroup_per_zone
*mz
;
1309 unsigned long *lru_size
;
1311 if (mem_cgroup_disabled())
1314 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1315 lru_size
= mz
->lru_size
+ lru
;
1316 *lru_size
+= nr_pages
;
1317 VM_BUG_ON((long)(*lru_size
) < 0);
1321 * Checks whether given mem is same or in the root_mem_cgroup's
1324 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1325 struct mem_cgroup
*memcg
)
1327 if (root_memcg
== memcg
)
1329 if (!root_memcg
->use_hierarchy
|| !memcg
)
1331 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1334 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1335 struct mem_cgroup
*memcg
)
1340 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1345 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*memcg
)
1348 struct mem_cgroup
*curr
= NULL
;
1349 struct task_struct
*p
;
1351 p
= find_lock_task_mm(task
);
1353 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1357 * All threads may have already detached their mm's, but the oom
1358 * killer still needs to detect if they have already been oom
1359 * killed to prevent needlessly killing additional tasks.
1362 curr
= mem_cgroup_from_task(task
);
1364 css_get(&curr
->css
);
1370 * We should check use_hierarchy of "memcg" not "curr". Because checking
1371 * use_hierarchy of "curr" here make this function true if hierarchy is
1372 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1373 * hierarchy(even if use_hierarchy is disabled in "memcg").
1375 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1376 css_put(&curr
->css
);
1380 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1382 unsigned long inactive_ratio
;
1383 unsigned long inactive
;
1384 unsigned long active
;
1387 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1388 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1390 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1392 inactive_ratio
= int_sqrt(10 * gb
);
1396 return inactive
* inactive_ratio
< active
;
1399 #define mem_cgroup_from_res_counter(counter, member) \
1400 container_of(counter, struct mem_cgroup, member)
1403 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1404 * @memcg: the memory cgroup
1406 * Returns the maximum amount of memory @mem can be charged with, in
1409 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1411 unsigned long long margin
;
1413 margin
= res_counter_margin(&memcg
->res
);
1414 if (do_swap_account
)
1415 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1416 return margin
>> PAGE_SHIFT
;
1419 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1421 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1424 if (cgrp
->parent
== NULL
)
1425 return vm_swappiness
;
1427 return memcg
->swappiness
;
1431 * memcg->moving_account is used for checking possibility that some thread is
1432 * calling move_account(). When a thread on CPU-A starts moving pages under
1433 * a memcg, other threads should check memcg->moving_account under
1434 * rcu_read_lock(), like this:
1438 * memcg->moving_account+1 if (memcg->mocing_account)
1440 * synchronize_rcu() update something.
1445 /* for quick checking without looking up memcg */
1446 atomic_t memcg_moving __read_mostly
;
1448 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1450 atomic_inc(&memcg_moving
);
1451 atomic_inc(&memcg
->moving_account
);
1455 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1458 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1459 * We check NULL in callee rather than caller.
1462 atomic_dec(&memcg_moving
);
1463 atomic_dec(&memcg
->moving_account
);
1468 * 2 routines for checking "mem" is under move_account() or not.
1470 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1471 * is used for avoiding races in accounting. If true,
1472 * pc->mem_cgroup may be overwritten.
1474 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1475 * under hierarchy of moving cgroups. This is for
1476 * waiting at hith-memory prressure caused by "move".
1479 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1481 VM_BUG_ON(!rcu_read_lock_held());
1482 return atomic_read(&memcg
->moving_account
) > 0;
1485 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1487 struct mem_cgroup
*from
;
1488 struct mem_cgroup
*to
;
1491 * Unlike task_move routines, we access mc.to, mc.from not under
1492 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1494 spin_lock(&mc
.lock
);
1500 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1501 || mem_cgroup_same_or_subtree(memcg
, to
);
1503 spin_unlock(&mc
.lock
);
1507 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1509 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1510 if (mem_cgroup_under_move(memcg
)) {
1512 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1513 /* moving charge context might have finished. */
1516 finish_wait(&mc
.waitq
, &wait
);
1524 * Take this lock when
1525 * - a code tries to modify page's memcg while it's USED.
1526 * - a code tries to modify page state accounting in a memcg.
1527 * see mem_cgroup_stolen(), too.
1529 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1530 unsigned long *flags
)
1532 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1535 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1536 unsigned long *flags
)
1538 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1541 #define K(x) ((x) << (PAGE_SHIFT-10))
1543 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1544 * @memcg: The memory cgroup that went over limit
1545 * @p: Task that is going to be killed
1547 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1550 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1552 struct cgroup
*task_cgrp
;
1553 struct cgroup
*mem_cgrp
;
1555 * Need a buffer in BSS, can't rely on allocations. The code relies
1556 * on the assumption that OOM is serialized for memory controller.
1557 * If this assumption is broken, revisit this code.
1559 static char memcg_name
[PATH_MAX
];
1561 struct mem_cgroup
*iter
;
1569 mem_cgrp
= memcg
->css
.cgroup
;
1570 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1572 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1575 * Unfortunately, we are unable to convert to a useful name
1576 * But we'll still print out the usage information
1583 pr_info("Task in %s killed", memcg_name
);
1586 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1594 * Continues from above, so we don't need an KERN_ level
1596 pr_cont(" as a result of limit of %s\n", memcg_name
);
1599 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1600 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1601 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1602 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1603 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1604 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1605 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1606 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1607 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1608 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1609 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1610 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1612 for_each_mem_cgroup_tree(iter
, memcg
) {
1613 pr_info("Memory cgroup stats");
1616 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1618 pr_cont(" for %s", memcg_name
);
1622 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1623 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1625 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1626 K(mem_cgroup_read_stat(iter
, i
)));
1629 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1630 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1631 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1638 * This function returns the number of memcg under hierarchy tree. Returns
1639 * 1(self count) if no children.
1641 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1644 struct mem_cgroup
*iter
;
1646 for_each_mem_cgroup_tree(iter
, memcg
)
1652 * Return the memory (and swap, if configured) limit for a memcg.
1654 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1658 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1661 * Do not consider swap space if we cannot swap due to swappiness
1663 if (mem_cgroup_swappiness(memcg
)) {
1666 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1667 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1670 * If memsw is finite and limits the amount of swap space
1671 * available to this memcg, return that limit.
1673 limit
= min(limit
, memsw
);
1679 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1682 struct mem_cgroup
*iter
;
1683 unsigned long chosen_points
= 0;
1684 unsigned long totalpages
;
1685 unsigned int points
= 0;
1686 struct task_struct
*chosen
= NULL
;
1689 * If current has a pending SIGKILL, then automatically select it. The
1690 * goal is to allow it to allocate so that it may quickly exit and free
1693 if (fatal_signal_pending(current
)) {
1694 set_thread_flag(TIF_MEMDIE
);
1698 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1699 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1700 for_each_mem_cgroup_tree(iter
, memcg
) {
1701 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1702 struct cgroup_iter it
;
1703 struct task_struct
*task
;
1705 cgroup_iter_start(cgroup
, &it
);
1706 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1707 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1709 case OOM_SCAN_SELECT
:
1711 put_task_struct(chosen
);
1713 chosen_points
= ULONG_MAX
;
1714 get_task_struct(chosen
);
1716 case OOM_SCAN_CONTINUE
:
1718 case OOM_SCAN_ABORT
:
1719 cgroup_iter_end(cgroup
, &it
);
1720 mem_cgroup_iter_break(memcg
, iter
);
1722 put_task_struct(chosen
);
1727 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1728 if (points
> chosen_points
) {
1730 put_task_struct(chosen
);
1732 chosen_points
= points
;
1733 get_task_struct(chosen
);
1736 cgroup_iter_end(cgroup
, &it
);
1741 points
= chosen_points
* 1000 / totalpages
;
1742 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1743 NULL
, "Memory cgroup out of memory");
1746 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1748 unsigned long flags
)
1750 unsigned long total
= 0;
1751 bool noswap
= false;
1754 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1756 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1759 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1761 drain_all_stock_async(memcg
);
1762 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1764 * Allow limit shrinkers, which are triggered directly
1765 * by userspace, to catch signals and stop reclaim
1766 * after minimal progress, regardless of the margin.
1768 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1770 if (mem_cgroup_margin(memcg
))
1773 * If nothing was reclaimed after two attempts, there
1774 * may be no reclaimable pages in this hierarchy.
1783 * test_mem_cgroup_node_reclaimable
1784 * @memcg: the target memcg
1785 * @nid: the node ID to be checked.
1786 * @noswap : specify true here if the user wants flle only information.
1788 * This function returns whether the specified memcg contains any
1789 * reclaimable pages on a node. Returns true if there are any reclaimable
1790 * pages in the node.
1792 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1793 int nid
, bool noswap
)
1795 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1797 if (noswap
|| !total_swap_pages
)
1799 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1804 #if MAX_NUMNODES > 1
1807 * Always updating the nodemask is not very good - even if we have an empty
1808 * list or the wrong list here, we can start from some node and traverse all
1809 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1812 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1816 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1817 * pagein/pageout changes since the last update.
1819 if (!atomic_read(&memcg
->numainfo_events
))
1821 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1824 /* make a nodemask where this memcg uses memory from */
1825 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1827 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1829 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1830 node_clear(nid
, memcg
->scan_nodes
);
1833 atomic_set(&memcg
->numainfo_events
, 0);
1834 atomic_set(&memcg
->numainfo_updating
, 0);
1838 * Selecting a node where we start reclaim from. Because what we need is just
1839 * reducing usage counter, start from anywhere is O,K. Considering
1840 * memory reclaim from current node, there are pros. and cons.
1842 * Freeing memory from current node means freeing memory from a node which
1843 * we'll use or we've used. So, it may make LRU bad. And if several threads
1844 * hit limits, it will see a contention on a node. But freeing from remote
1845 * node means more costs for memory reclaim because of memory latency.
1847 * Now, we use round-robin. Better algorithm is welcomed.
1849 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1853 mem_cgroup_may_update_nodemask(memcg
);
1854 node
= memcg
->last_scanned_node
;
1856 node
= next_node(node
, memcg
->scan_nodes
);
1857 if (node
== MAX_NUMNODES
)
1858 node
= first_node(memcg
->scan_nodes
);
1860 * We call this when we hit limit, not when pages are added to LRU.
1861 * No LRU may hold pages because all pages are UNEVICTABLE or
1862 * memcg is too small and all pages are not on LRU. In that case,
1863 * we use curret node.
1865 if (unlikely(node
== MAX_NUMNODES
))
1866 node
= numa_node_id();
1868 memcg
->last_scanned_node
= node
;
1873 * Check all nodes whether it contains reclaimable pages or not.
1874 * For quick scan, we make use of scan_nodes. This will allow us to skip
1875 * unused nodes. But scan_nodes is lazily updated and may not cotain
1876 * enough new information. We need to do double check.
1878 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1883 * quick check...making use of scan_node.
1884 * We can skip unused nodes.
1886 if (!nodes_empty(memcg
->scan_nodes
)) {
1887 for (nid
= first_node(memcg
->scan_nodes
);
1889 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1891 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1896 * Check rest of nodes.
1898 for_each_node_state(nid
, N_MEMORY
) {
1899 if (node_isset(nid
, memcg
->scan_nodes
))
1901 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1908 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1913 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1915 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1919 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1922 unsigned long *total_scanned
)
1924 struct mem_cgroup
*victim
= NULL
;
1927 unsigned long excess
;
1928 unsigned long nr_scanned
;
1929 struct mem_cgroup_reclaim_cookie reclaim
= {
1934 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
1937 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1942 * If we have not been able to reclaim
1943 * anything, it might because there are
1944 * no reclaimable pages under this hierarchy
1949 * We want to do more targeted reclaim.
1950 * excess >> 2 is not to excessive so as to
1951 * reclaim too much, nor too less that we keep
1952 * coming back to reclaim from this cgroup
1954 if (total
>= (excess
>> 2) ||
1955 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1960 if (!mem_cgroup_reclaimable(victim
, false))
1962 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1964 *total_scanned
+= nr_scanned
;
1965 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
1968 mem_cgroup_iter_break(root_memcg
, victim
);
1973 * Check OOM-Killer is already running under our hierarchy.
1974 * If someone is running, return false.
1975 * Has to be called with memcg_oom_lock
1977 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
1979 struct mem_cgroup
*iter
, *failed
= NULL
;
1981 for_each_mem_cgroup_tree(iter
, memcg
) {
1982 if (iter
->oom_lock
) {
1984 * this subtree of our hierarchy is already locked
1985 * so we cannot give a lock.
1988 mem_cgroup_iter_break(memcg
, iter
);
1991 iter
->oom_lock
= true;
1998 * OK, we failed to lock the whole subtree so we have to clean up
1999 * what we set up to the failing subtree
2001 for_each_mem_cgroup_tree(iter
, memcg
) {
2002 if (iter
== failed
) {
2003 mem_cgroup_iter_break(memcg
, iter
);
2006 iter
->oom_lock
= false;
2012 * Has to be called with memcg_oom_lock
2014 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2016 struct mem_cgroup
*iter
;
2018 for_each_mem_cgroup_tree(iter
, memcg
)
2019 iter
->oom_lock
= false;
2023 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2025 struct mem_cgroup
*iter
;
2027 for_each_mem_cgroup_tree(iter
, memcg
)
2028 atomic_inc(&iter
->under_oom
);
2031 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2033 struct mem_cgroup
*iter
;
2036 * When a new child is created while the hierarchy is under oom,
2037 * mem_cgroup_oom_lock() may not be called. We have to use
2038 * atomic_add_unless() here.
2040 for_each_mem_cgroup_tree(iter
, memcg
)
2041 atomic_add_unless(&iter
->under_oom
, -1, 0);
2044 static DEFINE_SPINLOCK(memcg_oom_lock
);
2045 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2047 struct oom_wait_info
{
2048 struct mem_cgroup
*memcg
;
2052 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2053 unsigned mode
, int sync
, void *arg
)
2055 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2056 struct mem_cgroup
*oom_wait_memcg
;
2057 struct oom_wait_info
*oom_wait_info
;
2059 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2060 oom_wait_memcg
= oom_wait_info
->memcg
;
2063 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2064 * Then we can use css_is_ancestor without taking care of RCU.
2066 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2067 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2069 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2072 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2074 /* for filtering, pass "memcg" as argument. */
2075 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2078 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2080 if (memcg
&& atomic_read(&memcg
->under_oom
))
2081 memcg_wakeup_oom(memcg
);
2085 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2087 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2090 struct oom_wait_info owait
;
2091 bool locked
, need_to_kill
;
2093 owait
.memcg
= memcg
;
2094 owait
.wait
.flags
= 0;
2095 owait
.wait
.func
= memcg_oom_wake_function
;
2096 owait
.wait
.private = current
;
2097 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2098 need_to_kill
= true;
2099 mem_cgroup_mark_under_oom(memcg
);
2101 /* At first, try to OOM lock hierarchy under memcg.*/
2102 spin_lock(&memcg_oom_lock
);
2103 locked
= mem_cgroup_oom_lock(memcg
);
2105 * Even if signal_pending(), we can't quit charge() loop without
2106 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2107 * under OOM is always welcomed, use TASK_KILLABLE here.
2109 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2110 if (!locked
|| memcg
->oom_kill_disable
)
2111 need_to_kill
= false;
2113 mem_cgroup_oom_notify(memcg
);
2114 spin_unlock(&memcg_oom_lock
);
2117 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2118 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2121 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2123 spin_lock(&memcg_oom_lock
);
2125 mem_cgroup_oom_unlock(memcg
);
2126 memcg_wakeup_oom(memcg
);
2127 spin_unlock(&memcg_oom_lock
);
2129 mem_cgroup_unmark_under_oom(memcg
);
2131 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2133 /* Give chance to dying process */
2134 schedule_timeout_uninterruptible(1);
2139 * Currently used to update mapped file statistics, but the routine can be
2140 * generalized to update other statistics as well.
2142 * Notes: Race condition
2144 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2145 * it tends to be costly. But considering some conditions, we doesn't need
2146 * to do so _always_.
2148 * Considering "charge", lock_page_cgroup() is not required because all
2149 * file-stat operations happen after a page is attached to radix-tree. There
2150 * are no race with "charge".
2152 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2153 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2154 * if there are race with "uncharge". Statistics itself is properly handled
2157 * Considering "move", this is an only case we see a race. To make the race
2158 * small, we check mm->moving_account and detect there are possibility of race
2159 * If there is, we take a lock.
2162 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2163 bool *locked
, unsigned long *flags
)
2165 struct mem_cgroup
*memcg
;
2166 struct page_cgroup
*pc
;
2168 pc
= lookup_page_cgroup(page
);
2170 memcg
= pc
->mem_cgroup
;
2171 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2174 * If this memory cgroup is not under account moving, we don't
2175 * need to take move_lock_mem_cgroup(). Because we already hold
2176 * rcu_read_lock(), any calls to move_account will be delayed until
2177 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2179 if (!mem_cgroup_stolen(memcg
))
2182 move_lock_mem_cgroup(memcg
, flags
);
2183 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2184 move_unlock_mem_cgroup(memcg
, flags
);
2190 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2192 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2195 * It's guaranteed that pc->mem_cgroup never changes while
2196 * lock is held because a routine modifies pc->mem_cgroup
2197 * should take move_lock_mem_cgroup().
2199 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2202 void mem_cgroup_update_page_stat(struct page
*page
,
2203 enum mem_cgroup_page_stat_item idx
, int val
)
2205 struct mem_cgroup
*memcg
;
2206 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2207 unsigned long uninitialized_var(flags
);
2209 if (mem_cgroup_disabled())
2212 memcg
= pc
->mem_cgroup
;
2213 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2217 case MEMCG_NR_FILE_MAPPED
:
2218 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2224 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2228 * size of first charge trial. "32" comes from vmscan.c's magic value.
2229 * TODO: maybe necessary to use big numbers in big irons.
2231 #define CHARGE_BATCH 32U
2232 struct memcg_stock_pcp
{
2233 struct mem_cgroup
*cached
; /* this never be root cgroup */
2234 unsigned int nr_pages
;
2235 struct work_struct work
;
2236 unsigned long flags
;
2237 #define FLUSHING_CACHED_CHARGE 0
2239 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2240 static DEFINE_MUTEX(percpu_charge_mutex
);
2243 * consume_stock: Try to consume stocked charge on this cpu.
2244 * @memcg: memcg to consume from.
2245 * @nr_pages: how many pages to charge.
2247 * The charges will only happen if @memcg matches the current cpu's memcg
2248 * stock, and at least @nr_pages are available in that stock. Failure to
2249 * service an allocation will refill the stock.
2251 * returns true if successful, false otherwise.
2253 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2255 struct memcg_stock_pcp
*stock
;
2258 if (nr_pages
> CHARGE_BATCH
)
2261 stock
= &get_cpu_var(memcg_stock
);
2262 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2263 stock
->nr_pages
-= nr_pages
;
2264 else /* need to call res_counter_charge */
2266 put_cpu_var(memcg_stock
);
2271 * Returns stocks cached in percpu to res_counter and reset cached information.
2273 static void drain_stock(struct memcg_stock_pcp
*stock
)
2275 struct mem_cgroup
*old
= stock
->cached
;
2277 if (stock
->nr_pages
) {
2278 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2280 res_counter_uncharge(&old
->res
, bytes
);
2281 if (do_swap_account
)
2282 res_counter_uncharge(&old
->memsw
, bytes
);
2283 stock
->nr_pages
= 0;
2285 stock
->cached
= NULL
;
2289 * This must be called under preempt disabled or must be called by
2290 * a thread which is pinned to local cpu.
2292 static void drain_local_stock(struct work_struct
*dummy
)
2294 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2296 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2300 * Cache charges(val) which is from res_counter, to local per_cpu area.
2301 * This will be consumed by consume_stock() function, later.
2303 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2305 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2307 if (stock
->cached
!= memcg
) { /* reset if necessary */
2309 stock
->cached
= memcg
;
2311 stock
->nr_pages
+= nr_pages
;
2312 put_cpu_var(memcg_stock
);
2316 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2317 * of the hierarchy under it. sync flag says whether we should block
2318 * until the work is done.
2320 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2324 /* Notify other cpus that system-wide "drain" is running */
2327 for_each_online_cpu(cpu
) {
2328 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2329 struct mem_cgroup
*memcg
;
2331 memcg
= stock
->cached
;
2332 if (!memcg
|| !stock
->nr_pages
)
2334 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2336 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2338 drain_local_stock(&stock
->work
);
2340 schedule_work_on(cpu
, &stock
->work
);
2348 for_each_online_cpu(cpu
) {
2349 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2350 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2351 flush_work(&stock
->work
);
2358 * Tries to drain stocked charges in other cpus. This function is asynchronous
2359 * and just put a work per cpu for draining localy on each cpu. Caller can
2360 * expects some charges will be back to res_counter later but cannot wait for
2363 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2366 * If someone calls draining, avoid adding more kworker runs.
2368 if (!mutex_trylock(&percpu_charge_mutex
))
2370 drain_all_stock(root_memcg
, false);
2371 mutex_unlock(&percpu_charge_mutex
);
2374 /* This is a synchronous drain interface. */
2375 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2377 /* called when force_empty is called */
2378 mutex_lock(&percpu_charge_mutex
);
2379 drain_all_stock(root_memcg
, true);
2380 mutex_unlock(&percpu_charge_mutex
);
2384 * This function drains percpu counter value from DEAD cpu and
2385 * move it to local cpu. Note that this function can be preempted.
2387 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2391 spin_lock(&memcg
->pcp_counter_lock
);
2392 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2393 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2395 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2396 memcg
->nocpu_base
.count
[i
] += x
;
2398 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2399 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2401 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2402 memcg
->nocpu_base
.events
[i
] += x
;
2404 spin_unlock(&memcg
->pcp_counter_lock
);
2407 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2408 unsigned long action
,
2411 int cpu
= (unsigned long)hcpu
;
2412 struct memcg_stock_pcp
*stock
;
2413 struct mem_cgroup
*iter
;
2415 if (action
== CPU_ONLINE
)
2418 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2421 for_each_mem_cgroup(iter
)
2422 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2424 stock
= &per_cpu(memcg_stock
, cpu
);
2430 /* See __mem_cgroup_try_charge() for details */
2432 CHARGE_OK
, /* success */
2433 CHARGE_RETRY
, /* need to retry but retry is not bad */
2434 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2435 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2436 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2439 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2440 unsigned int nr_pages
, unsigned int min_pages
,
2443 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2444 struct mem_cgroup
*mem_over_limit
;
2445 struct res_counter
*fail_res
;
2446 unsigned long flags
= 0;
2449 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2452 if (!do_swap_account
)
2454 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2458 res_counter_uncharge(&memcg
->res
, csize
);
2459 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2460 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2462 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2464 * Never reclaim on behalf of optional batching, retry with a
2465 * single page instead.
2467 if (nr_pages
> min_pages
)
2468 return CHARGE_RETRY
;
2470 if (!(gfp_mask
& __GFP_WAIT
))
2471 return CHARGE_WOULDBLOCK
;
2473 if (gfp_mask
& __GFP_NORETRY
)
2474 return CHARGE_NOMEM
;
2476 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2477 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2478 return CHARGE_RETRY
;
2480 * Even though the limit is exceeded at this point, reclaim
2481 * may have been able to free some pages. Retry the charge
2482 * before killing the task.
2484 * Only for regular pages, though: huge pages are rather
2485 * unlikely to succeed so close to the limit, and we fall back
2486 * to regular pages anyway in case of failure.
2488 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2489 return CHARGE_RETRY
;
2492 * At task move, charge accounts can be doubly counted. So, it's
2493 * better to wait until the end of task_move if something is going on.
2495 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2496 return CHARGE_RETRY
;
2498 /* If we don't need to call oom-killer at el, return immediately */
2500 return CHARGE_NOMEM
;
2502 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2503 return CHARGE_OOM_DIE
;
2505 return CHARGE_RETRY
;
2509 * __mem_cgroup_try_charge() does
2510 * 1. detect memcg to be charged against from passed *mm and *ptr,
2511 * 2. update res_counter
2512 * 3. call memory reclaim if necessary.
2514 * In some special case, if the task is fatal, fatal_signal_pending() or
2515 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2516 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2517 * as possible without any hazards. 2: all pages should have a valid
2518 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2519 * pointer, that is treated as a charge to root_mem_cgroup.
2521 * So __mem_cgroup_try_charge() will return
2522 * 0 ... on success, filling *ptr with a valid memcg pointer.
2523 * -ENOMEM ... charge failure because of resource limits.
2524 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2526 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2527 * the oom-killer can be invoked.
2529 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2531 unsigned int nr_pages
,
2532 struct mem_cgroup
**ptr
,
2535 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2536 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2537 struct mem_cgroup
*memcg
= NULL
;
2541 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2542 * in system level. So, allow to go ahead dying process in addition to
2545 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2546 || fatal_signal_pending(current
)))
2550 * We always charge the cgroup the mm_struct belongs to.
2551 * The mm_struct's mem_cgroup changes on task migration if the
2552 * thread group leader migrates. It's possible that mm is not
2553 * set, if so charge the root memcg (happens for pagecache usage).
2556 *ptr
= root_mem_cgroup
;
2558 if (*ptr
) { /* css should be a valid one */
2560 if (mem_cgroup_is_root(memcg
))
2562 if (consume_stock(memcg
, nr_pages
))
2564 css_get(&memcg
->css
);
2566 struct task_struct
*p
;
2569 p
= rcu_dereference(mm
->owner
);
2571 * Because we don't have task_lock(), "p" can exit.
2572 * In that case, "memcg" can point to root or p can be NULL with
2573 * race with swapoff. Then, we have small risk of mis-accouning.
2574 * But such kind of mis-account by race always happens because
2575 * we don't have cgroup_mutex(). It's overkill and we allo that
2577 * (*) swapoff at el will charge against mm-struct not against
2578 * task-struct. So, mm->owner can be NULL.
2580 memcg
= mem_cgroup_from_task(p
);
2582 memcg
= root_mem_cgroup
;
2583 if (mem_cgroup_is_root(memcg
)) {
2587 if (consume_stock(memcg
, nr_pages
)) {
2589 * It seems dagerous to access memcg without css_get().
2590 * But considering how consume_stok works, it's not
2591 * necessary. If consume_stock success, some charges
2592 * from this memcg are cached on this cpu. So, we
2593 * don't need to call css_get()/css_tryget() before
2594 * calling consume_stock().
2599 /* after here, we may be blocked. we need to get refcnt */
2600 if (!css_tryget(&memcg
->css
)) {
2610 /* If killed, bypass charge */
2611 if (fatal_signal_pending(current
)) {
2612 css_put(&memcg
->css
);
2617 if (oom
&& !nr_oom_retries
) {
2619 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2622 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2627 case CHARGE_RETRY
: /* not in OOM situation but retry */
2629 css_put(&memcg
->css
);
2632 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2633 css_put(&memcg
->css
);
2635 case CHARGE_NOMEM
: /* OOM routine works */
2637 css_put(&memcg
->css
);
2640 /* If oom, we never return -ENOMEM */
2643 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2644 css_put(&memcg
->css
);
2647 } while (ret
!= CHARGE_OK
);
2649 if (batch
> nr_pages
)
2650 refill_stock(memcg
, batch
- nr_pages
);
2651 css_put(&memcg
->css
);
2659 *ptr
= root_mem_cgroup
;
2664 * Somemtimes we have to undo a charge we got by try_charge().
2665 * This function is for that and do uncharge, put css's refcnt.
2666 * gotten by try_charge().
2668 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2669 unsigned int nr_pages
)
2671 if (!mem_cgroup_is_root(memcg
)) {
2672 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2674 res_counter_uncharge(&memcg
->res
, bytes
);
2675 if (do_swap_account
)
2676 res_counter_uncharge(&memcg
->memsw
, bytes
);
2681 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2682 * This is useful when moving usage to parent cgroup.
2684 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2685 unsigned int nr_pages
)
2687 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2689 if (mem_cgroup_is_root(memcg
))
2692 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2693 if (do_swap_account
)
2694 res_counter_uncharge_until(&memcg
->memsw
,
2695 memcg
->memsw
.parent
, bytes
);
2699 * A helper function to get mem_cgroup from ID. must be called under
2700 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2701 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2702 * called against removed memcg.)
2704 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2706 struct cgroup_subsys_state
*css
;
2708 /* ID 0 is unused ID */
2711 css
= css_lookup(&mem_cgroup_subsys
, id
);
2714 return mem_cgroup_from_css(css
);
2717 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2719 struct mem_cgroup
*memcg
= NULL
;
2720 struct page_cgroup
*pc
;
2724 VM_BUG_ON(!PageLocked(page
));
2726 pc
= lookup_page_cgroup(page
);
2727 lock_page_cgroup(pc
);
2728 if (PageCgroupUsed(pc
)) {
2729 memcg
= pc
->mem_cgroup
;
2730 if (memcg
&& !css_tryget(&memcg
->css
))
2732 } else if (PageSwapCache(page
)) {
2733 ent
.val
= page_private(page
);
2734 id
= lookup_swap_cgroup_id(ent
);
2736 memcg
= mem_cgroup_lookup(id
);
2737 if (memcg
&& !css_tryget(&memcg
->css
))
2741 unlock_page_cgroup(pc
);
2745 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2747 unsigned int nr_pages
,
2748 enum charge_type ctype
,
2751 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2752 struct zone
*uninitialized_var(zone
);
2753 struct lruvec
*lruvec
;
2754 bool was_on_lru
= false;
2757 lock_page_cgroup(pc
);
2758 VM_BUG_ON(PageCgroupUsed(pc
));
2760 * we don't need page_cgroup_lock about tail pages, becase they are not
2761 * accessed by any other context at this point.
2765 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2766 * may already be on some other mem_cgroup's LRU. Take care of it.
2769 zone
= page_zone(page
);
2770 spin_lock_irq(&zone
->lru_lock
);
2771 if (PageLRU(page
)) {
2772 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2774 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2779 pc
->mem_cgroup
= memcg
;
2781 * We access a page_cgroup asynchronously without lock_page_cgroup().
2782 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2783 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2784 * before USED bit, we need memory barrier here.
2785 * See mem_cgroup_add_lru_list(), etc.
2788 SetPageCgroupUsed(pc
);
2792 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2793 VM_BUG_ON(PageLRU(page
));
2795 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2797 spin_unlock_irq(&zone
->lru_lock
);
2800 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2805 mem_cgroup_charge_statistics(memcg
, anon
, nr_pages
);
2806 unlock_page_cgroup(pc
);
2809 * "charge_statistics" updated event counter. Then, check it.
2810 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2811 * if they exceeds softlimit.
2813 memcg_check_events(memcg
, page
);
2816 static DEFINE_MUTEX(set_limit_mutex
);
2818 #ifdef CONFIG_MEMCG_KMEM
2819 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2821 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2822 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2826 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2827 * in the memcg_cache_params struct.
2829 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2831 struct kmem_cache
*cachep
;
2833 VM_BUG_ON(p
->is_root_cache
);
2834 cachep
= p
->root_cache
;
2835 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2838 #ifdef CONFIG_SLABINFO
2839 static int mem_cgroup_slabinfo_read(struct cgroup
*cont
, struct cftype
*cft
,
2842 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
2843 struct memcg_cache_params
*params
;
2845 if (!memcg_can_account_kmem(memcg
))
2848 print_slabinfo_header(m
);
2850 mutex_lock(&memcg
->slab_caches_mutex
);
2851 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2852 cache_show(memcg_params_to_cache(params
), m
);
2853 mutex_unlock(&memcg
->slab_caches_mutex
);
2859 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2861 struct res_counter
*fail_res
;
2862 struct mem_cgroup
*_memcg
;
2866 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2871 * Conditions under which we can wait for the oom_killer. Those are
2872 * the same conditions tested by the core page allocator
2874 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2877 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2880 if (ret
== -EINTR
) {
2882 * __mem_cgroup_try_charge() chosed to bypass to root due to
2883 * OOM kill or fatal signal. Since our only options are to
2884 * either fail the allocation or charge it to this cgroup, do
2885 * it as a temporary condition. But we can't fail. From a
2886 * kmem/slab perspective, the cache has already been selected,
2887 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2890 * This condition will only trigger if the task entered
2891 * memcg_charge_kmem in a sane state, but was OOM-killed during
2892 * __mem_cgroup_try_charge() above. Tasks that were already
2893 * dying when the allocation triggers should have been already
2894 * directed to the root cgroup in memcontrol.h
2896 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2897 if (do_swap_account
)
2898 res_counter_charge_nofail(&memcg
->memsw
, size
,
2902 res_counter_uncharge(&memcg
->kmem
, size
);
2907 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2909 res_counter_uncharge(&memcg
->res
, size
);
2910 if (do_swap_account
)
2911 res_counter_uncharge(&memcg
->memsw
, size
);
2914 if (res_counter_uncharge(&memcg
->kmem
, size
))
2917 if (memcg_kmem_test_and_clear_dead(memcg
))
2918 mem_cgroup_put(memcg
);
2921 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2926 mutex_lock(&memcg
->slab_caches_mutex
);
2927 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2928 mutex_unlock(&memcg
->slab_caches_mutex
);
2932 * helper for acessing a memcg's index. It will be used as an index in the
2933 * child cache array in kmem_cache, and also to derive its name. This function
2934 * will return -1 when this is not a kmem-limited memcg.
2936 int memcg_cache_id(struct mem_cgroup
*memcg
)
2938 return memcg
? memcg
->kmemcg_id
: -1;
2942 * This ends up being protected by the set_limit mutex, during normal
2943 * operation, because that is its main call site.
2945 * But when we create a new cache, we can call this as well if its parent
2946 * is kmem-limited. That will have to hold set_limit_mutex as well.
2948 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2952 num
= ida_simple_get(&kmem_limited_groups
,
2953 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2957 * After this point, kmem_accounted (that we test atomically in
2958 * the beginning of this conditional), is no longer 0. This
2959 * guarantees only one process will set the following boolean
2960 * to true. We don't need test_and_set because we're protected
2961 * by the set_limit_mutex anyway.
2963 memcg_kmem_set_activated(memcg
);
2965 ret
= memcg_update_all_caches(num
+1);
2967 ida_simple_remove(&kmem_limited_groups
, num
);
2968 memcg_kmem_clear_activated(memcg
);
2972 memcg
->kmemcg_id
= num
;
2973 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
2974 mutex_init(&memcg
->slab_caches_mutex
);
2978 static size_t memcg_caches_array_size(int num_groups
)
2981 if (num_groups
<= 0)
2984 size
= 2 * num_groups
;
2985 if (size
< MEMCG_CACHES_MIN_SIZE
)
2986 size
= MEMCG_CACHES_MIN_SIZE
;
2987 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2988 size
= MEMCG_CACHES_MAX_SIZE
;
2994 * We should update the current array size iff all caches updates succeed. This
2995 * can only be done from the slab side. The slab mutex needs to be held when
2998 void memcg_update_array_size(int num
)
3000 if (num
> memcg_limited_groups_array_size
)
3001 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3004 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3006 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3008 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3010 if (num_groups
> memcg_limited_groups_array_size
) {
3012 ssize_t size
= memcg_caches_array_size(num_groups
);
3014 size
*= sizeof(void *);
3015 size
+= sizeof(struct memcg_cache_params
);
3017 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3018 if (!s
->memcg_params
) {
3019 s
->memcg_params
= cur_params
;
3023 s
->memcg_params
->is_root_cache
= true;
3026 * There is the chance it will be bigger than
3027 * memcg_limited_groups_array_size, if we failed an allocation
3028 * in a cache, in which case all caches updated before it, will
3029 * have a bigger array.
3031 * But if that is the case, the data after
3032 * memcg_limited_groups_array_size is certainly unused
3034 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3035 if (!cur_params
->memcg_caches
[i
])
3037 s
->memcg_params
->memcg_caches
[i
] =
3038 cur_params
->memcg_caches
[i
];
3042 * Ideally, we would wait until all caches succeed, and only
3043 * then free the old one. But this is not worth the extra
3044 * pointer per-cache we'd have to have for this.
3046 * It is not a big deal if some caches are left with a size
3047 * bigger than the others. And all updates will reset this
3055 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3056 struct kmem_cache
*root_cache
)
3058 size_t size
= sizeof(struct memcg_cache_params
);
3060 if (!memcg_kmem_enabled())
3064 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3066 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3067 if (!s
->memcg_params
)
3071 s
->memcg_params
->memcg
= memcg
;
3072 s
->memcg_params
->root_cache
= root_cache
;
3074 s
->memcg_params
->is_root_cache
= true;
3079 void memcg_release_cache(struct kmem_cache
*s
)
3081 struct kmem_cache
*root
;
3082 struct mem_cgroup
*memcg
;
3086 * This happens, for instance, when a root cache goes away before we
3089 if (!s
->memcg_params
)
3092 if (s
->memcg_params
->is_root_cache
)
3095 memcg
= s
->memcg_params
->memcg
;
3096 id
= memcg_cache_id(memcg
);
3098 root
= s
->memcg_params
->root_cache
;
3099 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3100 mem_cgroup_put(memcg
);
3102 mutex_lock(&memcg
->slab_caches_mutex
);
3103 list_del(&s
->memcg_params
->list
);
3104 mutex_unlock(&memcg
->slab_caches_mutex
);
3107 kfree(s
->memcg_params
);
3111 * During the creation a new cache, we need to disable our accounting mechanism
3112 * altogether. This is true even if we are not creating, but rather just
3113 * enqueing new caches to be created.
3115 * This is because that process will trigger allocations; some visible, like
3116 * explicit kmallocs to auxiliary data structures, name strings and internal
3117 * cache structures; some well concealed, like INIT_WORK() that can allocate
3118 * objects during debug.
3120 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3121 * to it. This may not be a bounded recursion: since the first cache creation
3122 * failed to complete (waiting on the allocation), we'll just try to create the
3123 * cache again, failing at the same point.
3125 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3126 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3127 * inside the following two functions.
3129 static inline void memcg_stop_kmem_account(void)
3131 VM_BUG_ON(!current
->mm
);
3132 current
->memcg_kmem_skip_account
++;
3135 static inline void memcg_resume_kmem_account(void)
3137 VM_BUG_ON(!current
->mm
);
3138 current
->memcg_kmem_skip_account
--;
3141 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3143 struct kmem_cache
*cachep
;
3144 struct memcg_cache_params
*p
;
3146 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3148 cachep
= memcg_params_to_cache(p
);
3151 * If we get down to 0 after shrink, we could delete right away.
3152 * However, memcg_release_pages() already puts us back in the workqueue
3153 * in that case. If we proceed deleting, we'll get a dangling
3154 * reference, and removing the object from the workqueue in that case
3155 * is unnecessary complication. We are not a fast path.
3157 * Note that this case is fundamentally different from racing with
3158 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3159 * kmem_cache_shrink, not only we would be reinserting a dead cache
3160 * into the queue, but doing so from inside the worker racing to
3163 * So if we aren't down to zero, we'll just schedule a worker and try
3166 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3167 kmem_cache_shrink(cachep
);
3168 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3171 kmem_cache_destroy(cachep
);
3174 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3176 if (!cachep
->memcg_params
->dead
)
3180 * There are many ways in which we can get here.
3182 * We can get to a memory-pressure situation while the delayed work is
3183 * still pending to run. The vmscan shrinkers can then release all
3184 * cache memory and get us to destruction. If this is the case, we'll
3185 * be executed twice, which is a bug (the second time will execute over
3186 * bogus data). In this case, cancelling the work should be fine.
3188 * But we can also get here from the worker itself, if
3189 * kmem_cache_shrink is enough to shake all the remaining objects and
3190 * get the page count to 0. In this case, we'll deadlock if we try to
3191 * cancel the work (the worker runs with an internal lock held, which
3192 * is the same lock we would hold for cancel_work_sync().)
3194 * Since we can't possibly know who got us here, just refrain from
3195 * running if there is already work pending
3197 if (work_pending(&cachep
->memcg_params
->destroy
))
3200 * We have to defer the actual destroying to a workqueue, because
3201 * we might currently be in a context that cannot sleep.
3203 schedule_work(&cachep
->memcg_params
->destroy
);
3206 static char *memcg_cache_name(struct mem_cgroup
*memcg
, struct kmem_cache
*s
)
3209 struct dentry
*dentry
;
3212 dentry
= rcu_dereference(memcg
->css
.cgroup
->dentry
);
3215 BUG_ON(dentry
== NULL
);
3217 name
= kasprintf(GFP_KERNEL
, "%s(%d:%s)", s
->name
,
3218 memcg_cache_id(memcg
), dentry
->d_name
.name
);
3223 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3224 struct kmem_cache
*s
)
3227 struct kmem_cache
*new;
3229 name
= memcg_cache_name(memcg
, s
);
3233 new = kmem_cache_create_memcg(memcg
, name
, s
->object_size
, s
->align
,
3234 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3237 new->allocflags
|= __GFP_KMEMCG
;
3244 * This lock protects updaters, not readers. We want readers to be as fast as
3245 * they can, and they will either see NULL or a valid cache value. Our model
3246 * allow them to see NULL, in which case the root memcg will be selected.
3248 * We need this lock because multiple allocations to the same cache from a non
3249 * will span more than one worker. Only one of them can create the cache.
3251 static DEFINE_MUTEX(memcg_cache_mutex
);
3252 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3253 struct kmem_cache
*cachep
)
3255 struct kmem_cache
*new_cachep
;
3258 BUG_ON(!memcg_can_account_kmem(memcg
));
3260 idx
= memcg_cache_id(memcg
);
3262 mutex_lock(&memcg_cache_mutex
);
3263 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3267 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3268 if (new_cachep
== NULL
) {
3269 new_cachep
= cachep
;
3273 mem_cgroup_get(memcg
);
3274 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3276 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3278 * the readers won't lock, make sure everybody sees the updated value,
3279 * so they won't put stuff in the queue again for no reason
3283 mutex_unlock(&memcg_cache_mutex
);
3287 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3289 struct kmem_cache
*c
;
3292 if (!s
->memcg_params
)
3294 if (!s
->memcg_params
->is_root_cache
)
3298 * If the cache is being destroyed, we trust that there is no one else
3299 * requesting objects from it. Even if there are, the sanity checks in
3300 * kmem_cache_destroy should caught this ill-case.
3302 * Still, we don't want anyone else freeing memcg_caches under our
3303 * noses, which can happen if a new memcg comes to life. As usual,
3304 * we'll take the set_limit_mutex to protect ourselves against this.
3306 mutex_lock(&set_limit_mutex
);
3307 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3308 c
= s
->memcg_params
->memcg_caches
[i
];
3313 * We will now manually delete the caches, so to avoid races
3314 * we need to cancel all pending destruction workers and
3315 * proceed with destruction ourselves.
3317 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3318 * and that could spawn the workers again: it is likely that
3319 * the cache still have active pages until this very moment.
3320 * This would lead us back to mem_cgroup_destroy_cache.
3322 * But that will not execute at all if the "dead" flag is not
3323 * set, so flip it down to guarantee we are in control.
3325 c
->memcg_params
->dead
= false;
3326 cancel_work_sync(&c
->memcg_params
->destroy
);
3327 kmem_cache_destroy(c
);
3329 mutex_unlock(&set_limit_mutex
);
3332 struct create_work
{
3333 struct mem_cgroup
*memcg
;
3334 struct kmem_cache
*cachep
;
3335 struct work_struct work
;
3338 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3340 struct kmem_cache
*cachep
;
3341 struct memcg_cache_params
*params
;
3343 if (!memcg_kmem_is_active(memcg
))
3346 mutex_lock(&memcg
->slab_caches_mutex
);
3347 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3348 cachep
= memcg_params_to_cache(params
);
3349 cachep
->memcg_params
->dead
= true;
3350 INIT_WORK(&cachep
->memcg_params
->destroy
,
3351 kmem_cache_destroy_work_func
);
3352 schedule_work(&cachep
->memcg_params
->destroy
);
3354 mutex_unlock(&memcg
->slab_caches_mutex
);
3357 static void memcg_create_cache_work_func(struct work_struct
*w
)
3359 struct create_work
*cw
;
3361 cw
= container_of(w
, struct create_work
, work
);
3362 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3363 /* Drop the reference gotten when we enqueued. */
3364 css_put(&cw
->memcg
->css
);
3369 * Enqueue the creation of a per-memcg kmem_cache.
3370 * Called with rcu_read_lock.
3372 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3373 struct kmem_cache
*cachep
)
3375 struct create_work
*cw
;
3377 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3381 /* The corresponding put will be done in the workqueue. */
3382 if (!css_tryget(&memcg
->css
)) {
3388 cw
->cachep
= cachep
;
3390 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3391 schedule_work(&cw
->work
);
3394 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3395 struct kmem_cache
*cachep
)
3398 * We need to stop accounting when we kmalloc, because if the
3399 * corresponding kmalloc cache is not yet created, the first allocation
3400 * in __memcg_create_cache_enqueue will recurse.
3402 * However, it is better to enclose the whole function. Depending on
3403 * the debugging options enabled, INIT_WORK(), for instance, can
3404 * trigger an allocation. This too, will make us recurse. Because at
3405 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3406 * the safest choice is to do it like this, wrapping the whole function.
3408 memcg_stop_kmem_account();
3409 __memcg_create_cache_enqueue(memcg
, cachep
);
3410 memcg_resume_kmem_account();
3413 * Return the kmem_cache we're supposed to use for a slab allocation.
3414 * We try to use the current memcg's version of the cache.
3416 * If the cache does not exist yet, if we are the first user of it,
3417 * we either create it immediately, if possible, or create it asynchronously
3419 * In the latter case, we will let the current allocation go through with
3420 * the original cache.
3422 * Can't be called in interrupt context or from kernel threads.
3423 * This function needs to be called with rcu_read_lock() held.
3425 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3428 struct mem_cgroup
*memcg
;
3431 VM_BUG_ON(!cachep
->memcg_params
);
3432 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3434 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3438 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3441 if (!memcg_can_account_kmem(memcg
))
3444 idx
= memcg_cache_id(memcg
);
3447 * barrier to mare sure we're always seeing the up to date value. The
3448 * code updating memcg_caches will issue a write barrier to match this.
3450 read_barrier_depends();
3451 if (unlikely(cachep
->memcg_params
->memcg_caches
[idx
] == NULL
)) {
3453 * If we are in a safe context (can wait, and not in interrupt
3454 * context), we could be be predictable and return right away.
3455 * This would guarantee that the allocation being performed
3456 * already belongs in the new cache.
3458 * However, there are some clashes that can arrive from locking.
3459 * For instance, because we acquire the slab_mutex while doing
3460 * kmem_cache_dup, this means no further allocation could happen
3461 * with the slab_mutex held.
3463 * Also, because cache creation issue get_online_cpus(), this
3464 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3465 * that ends up reversed during cpu hotplug. (cpuset allocates
3466 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3467 * better to defer everything.
3469 memcg_create_cache_enqueue(memcg
, cachep
);
3473 return cachep
->memcg_params
->memcg_caches
[idx
];
3475 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3478 * We need to verify if the allocation against current->mm->owner's memcg is
3479 * possible for the given order. But the page is not allocated yet, so we'll
3480 * need a further commit step to do the final arrangements.
3482 * It is possible for the task to switch cgroups in this mean time, so at
3483 * commit time, we can't rely on task conversion any longer. We'll then use
3484 * the handle argument to return to the caller which cgroup we should commit
3485 * against. We could also return the memcg directly and avoid the pointer
3486 * passing, but a boolean return value gives better semantics considering
3487 * the compiled-out case as well.
3489 * Returning true means the allocation is possible.
3492 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3494 struct mem_cgroup
*memcg
;
3498 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3501 * very rare case described in mem_cgroup_from_task. Unfortunately there
3502 * isn't much we can do without complicating this too much, and it would
3503 * be gfp-dependent anyway. Just let it go
3505 if (unlikely(!memcg
))
3508 if (!memcg_can_account_kmem(memcg
)) {
3509 css_put(&memcg
->css
);
3513 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3517 css_put(&memcg
->css
);
3521 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3524 struct page_cgroup
*pc
;
3526 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3528 /* The page allocation failed. Revert */
3530 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3534 pc
= lookup_page_cgroup(page
);
3535 lock_page_cgroup(pc
);
3536 pc
->mem_cgroup
= memcg
;
3537 SetPageCgroupUsed(pc
);
3538 unlock_page_cgroup(pc
);
3541 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3543 struct mem_cgroup
*memcg
= NULL
;
3544 struct page_cgroup
*pc
;
3547 pc
= lookup_page_cgroup(page
);
3549 * Fast unlocked return. Theoretically might have changed, have to
3550 * check again after locking.
3552 if (!PageCgroupUsed(pc
))
3555 lock_page_cgroup(pc
);
3556 if (PageCgroupUsed(pc
)) {
3557 memcg
= pc
->mem_cgroup
;
3558 ClearPageCgroupUsed(pc
);
3560 unlock_page_cgroup(pc
);
3563 * We trust that only if there is a memcg associated with the page, it
3564 * is a valid allocation
3569 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3570 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3573 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3576 #endif /* CONFIG_MEMCG_KMEM */
3578 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3580 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3582 * Because tail pages are not marked as "used", set it. We're under
3583 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3584 * charge/uncharge will be never happen and move_account() is done under
3585 * compound_lock(), so we don't have to take care of races.
3587 void mem_cgroup_split_huge_fixup(struct page
*head
)
3589 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3590 struct page_cgroup
*pc
;
3593 if (mem_cgroup_disabled())
3595 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3597 pc
->mem_cgroup
= head_pc
->mem_cgroup
;
3598 smp_wmb();/* see __commit_charge() */
3599 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3602 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3605 * mem_cgroup_move_account - move account of the page
3607 * @nr_pages: number of regular pages (>1 for huge pages)
3608 * @pc: page_cgroup of the page.
3609 * @from: mem_cgroup which the page is moved from.
3610 * @to: mem_cgroup which the page is moved to. @from != @to.
3612 * The caller must confirm following.
3613 * - page is not on LRU (isolate_page() is useful.)
3614 * - compound_lock is held when nr_pages > 1
3616 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3619 static int mem_cgroup_move_account(struct page
*page
,
3620 unsigned int nr_pages
,
3621 struct page_cgroup
*pc
,
3622 struct mem_cgroup
*from
,
3623 struct mem_cgroup
*to
)
3625 unsigned long flags
;
3627 bool anon
= PageAnon(page
);
3629 VM_BUG_ON(from
== to
);
3630 VM_BUG_ON(PageLRU(page
));
3632 * The page is isolated from LRU. So, collapse function
3633 * will not handle this page. But page splitting can happen.
3634 * Do this check under compound_page_lock(). The caller should
3638 if (nr_pages
> 1 && !PageTransHuge(page
))
3641 lock_page_cgroup(pc
);
3644 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3647 move_lock_mem_cgroup(from
, &flags
);
3649 if (!anon
&& page_mapped(page
)) {
3650 /* Update mapped_file data for mem_cgroup */
3652 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3653 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3656 mem_cgroup_charge_statistics(from
, anon
, -nr_pages
);
3658 /* caller should have done css_get */
3659 pc
->mem_cgroup
= to
;
3660 mem_cgroup_charge_statistics(to
, anon
, nr_pages
);
3661 move_unlock_mem_cgroup(from
, &flags
);
3664 unlock_page_cgroup(pc
);
3668 memcg_check_events(to
, page
);
3669 memcg_check_events(from
, page
);
3675 * mem_cgroup_move_parent - moves page to the parent group
3676 * @page: the page to move
3677 * @pc: page_cgroup of the page
3678 * @child: page's cgroup
3680 * move charges to its parent or the root cgroup if the group has no
3681 * parent (aka use_hierarchy==0).
3682 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3683 * mem_cgroup_move_account fails) the failure is always temporary and
3684 * it signals a race with a page removal/uncharge or migration. In the
3685 * first case the page is on the way out and it will vanish from the LRU
3686 * on the next attempt and the call should be retried later.
3687 * Isolation from the LRU fails only if page has been isolated from
3688 * the LRU since we looked at it and that usually means either global
3689 * reclaim or migration going on. The page will either get back to the
3691 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3692 * (!PageCgroupUsed) or moved to a different group. The page will
3693 * disappear in the next attempt.
3695 static int mem_cgroup_move_parent(struct page
*page
,
3696 struct page_cgroup
*pc
,
3697 struct mem_cgroup
*child
)
3699 struct mem_cgroup
*parent
;
3700 unsigned int nr_pages
;
3701 unsigned long uninitialized_var(flags
);
3704 VM_BUG_ON(mem_cgroup_is_root(child
));
3707 if (!get_page_unless_zero(page
))
3709 if (isolate_lru_page(page
))
3712 nr_pages
= hpage_nr_pages(page
);
3714 parent
= parent_mem_cgroup(child
);
3716 * If no parent, move charges to root cgroup.
3719 parent
= root_mem_cgroup
;
3722 VM_BUG_ON(!PageTransHuge(page
));
3723 flags
= compound_lock_irqsave(page
);
3726 ret
= mem_cgroup_move_account(page
, nr_pages
,
3729 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3732 compound_unlock_irqrestore(page
, flags
);
3733 putback_lru_page(page
);
3741 * Charge the memory controller for page usage.
3743 * 0 if the charge was successful
3744 * < 0 if the cgroup is over its limit
3746 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3747 gfp_t gfp_mask
, enum charge_type ctype
)
3749 struct mem_cgroup
*memcg
= NULL
;
3750 unsigned int nr_pages
= 1;
3754 if (PageTransHuge(page
)) {
3755 nr_pages
<<= compound_order(page
);
3756 VM_BUG_ON(!PageTransHuge(page
));
3758 * Never OOM-kill a process for a huge page. The
3759 * fault handler will fall back to regular pages.
3764 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3767 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3771 int mem_cgroup_newpage_charge(struct page
*page
,
3772 struct mm_struct
*mm
, gfp_t gfp_mask
)
3774 if (mem_cgroup_disabled())
3776 VM_BUG_ON(page_mapped(page
));
3777 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3779 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3780 MEM_CGROUP_CHARGE_TYPE_ANON
);
3784 * While swap-in, try_charge -> commit or cancel, the page is locked.
3785 * And when try_charge() successfully returns, one refcnt to memcg without
3786 * struct page_cgroup is acquired. This refcnt will be consumed by
3787 * "commit()" or removed by "cancel()"
3789 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3792 struct mem_cgroup
**memcgp
)
3794 struct mem_cgroup
*memcg
;
3795 struct page_cgroup
*pc
;
3798 pc
= lookup_page_cgroup(page
);
3800 * Every swap fault against a single page tries to charge the
3801 * page, bail as early as possible. shmem_unuse() encounters
3802 * already charged pages, too. The USED bit is protected by
3803 * the page lock, which serializes swap cache removal, which
3804 * in turn serializes uncharging.
3806 if (PageCgroupUsed(pc
))
3808 if (!do_swap_account
)
3810 memcg
= try_get_mem_cgroup_from_page(page
);
3814 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3815 css_put(&memcg
->css
);
3820 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3826 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3827 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3830 if (mem_cgroup_disabled())
3833 * A racing thread's fault, or swapoff, may have already
3834 * updated the pte, and even removed page from swap cache: in
3835 * those cases unuse_pte()'s pte_same() test will fail; but
3836 * there's also a KSM case which does need to charge the page.
3838 if (!PageSwapCache(page
)) {
3841 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3846 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3849 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3851 if (mem_cgroup_disabled())
3855 __mem_cgroup_cancel_charge(memcg
, 1);
3859 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3860 enum charge_type ctype
)
3862 if (mem_cgroup_disabled())
3867 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3869 * Now swap is on-memory. This means this page may be
3870 * counted both as mem and swap....double count.
3871 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3872 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3873 * may call delete_from_swap_cache() before reach here.
3875 if (do_swap_account
&& PageSwapCache(page
)) {
3876 swp_entry_t ent
= {.val
= page_private(page
)};
3877 mem_cgroup_uncharge_swap(ent
);
3881 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3882 struct mem_cgroup
*memcg
)
3884 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3885 MEM_CGROUP_CHARGE_TYPE_ANON
);
3888 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3891 struct mem_cgroup
*memcg
= NULL
;
3892 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3895 if (mem_cgroup_disabled())
3897 if (PageCompound(page
))
3900 if (!PageSwapCache(page
))
3901 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3902 else { /* page is swapcache/shmem */
3903 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3906 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3911 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3912 unsigned int nr_pages
,
3913 const enum charge_type ctype
)
3915 struct memcg_batch_info
*batch
= NULL
;
3916 bool uncharge_memsw
= true;
3918 /* If swapout, usage of swap doesn't decrease */
3919 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3920 uncharge_memsw
= false;
3922 batch
= ¤t
->memcg_batch
;
3924 * In usual, we do css_get() when we remember memcg pointer.
3925 * But in this case, we keep res->usage until end of a series of
3926 * uncharges. Then, it's ok to ignore memcg's refcnt.
3929 batch
->memcg
= memcg
;
3931 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3932 * In those cases, all pages freed continuously can be expected to be in
3933 * the same cgroup and we have chance to coalesce uncharges.
3934 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3935 * because we want to do uncharge as soon as possible.
3938 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3939 goto direct_uncharge
;
3942 goto direct_uncharge
;
3945 * In typical case, batch->memcg == mem. This means we can
3946 * merge a series of uncharges to an uncharge of res_counter.
3947 * If not, we uncharge res_counter ony by one.
3949 if (batch
->memcg
!= memcg
)
3950 goto direct_uncharge
;
3951 /* remember freed charge and uncharge it later */
3954 batch
->memsw_nr_pages
++;
3957 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3959 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3960 if (unlikely(batch
->memcg
!= memcg
))
3961 memcg_oom_recover(memcg
);
3965 * uncharge if !page_mapped(page)
3967 static struct mem_cgroup
*
3968 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3971 struct mem_cgroup
*memcg
= NULL
;
3972 unsigned int nr_pages
= 1;
3973 struct page_cgroup
*pc
;
3976 if (mem_cgroup_disabled())
3979 VM_BUG_ON(PageSwapCache(page
));
3981 if (PageTransHuge(page
)) {
3982 nr_pages
<<= compound_order(page
);
3983 VM_BUG_ON(!PageTransHuge(page
));
3986 * Check if our page_cgroup is valid
3988 pc
= lookup_page_cgroup(page
);
3989 if (unlikely(!PageCgroupUsed(pc
)))
3992 lock_page_cgroup(pc
);
3994 memcg
= pc
->mem_cgroup
;
3996 if (!PageCgroupUsed(pc
))
3999 anon
= PageAnon(page
);
4002 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4004 * Generally PageAnon tells if it's the anon statistics to be
4005 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4006 * used before page reached the stage of being marked PageAnon.
4010 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4011 /* See mem_cgroup_prepare_migration() */
4012 if (page_mapped(page
))
4015 * Pages under migration may not be uncharged. But
4016 * end_migration() /must/ be the one uncharging the
4017 * unused post-migration page and so it has to call
4018 * here with the migration bit still set. See the
4019 * res_counter handling below.
4021 if (!end_migration
&& PageCgroupMigration(pc
))
4024 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4025 if (!PageAnon(page
)) { /* Shared memory */
4026 if (page
->mapping
&& !page_is_file_cache(page
))
4028 } else if (page_mapped(page
)) /* Anon */
4035 mem_cgroup_charge_statistics(memcg
, anon
, -nr_pages
);
4037 ClearPageCgroupUsed(pc
);
4039 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4040 * freed from LRU. This is safe because uncharged page is expected not
4041 * to be reused (freed soon). Exception is SwapCache, it's handled by
4042 * special functions.
4045 unlock_page_cgroup(pc
);
4047 * even after unlock, we have memcg->res.usage here and this memcg
4048 * will never be freed.
4050 memcg_check_events(memcg
, page
);
4051 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4052 mem_cgroup_swap_statistics(memcg
, true);
4053 mem_cgroup_get(memcg
);
4056 * Migration does not charge the res_counter for the
4057 * replacement page, so leave it alone when phasing out the
4058 * page that is unused after the migration.
4060 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4061 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4066 unlock_page_cgroup(pc
);
4070 void mem_cgroup_uncharge_page(struct page
*page
)
4073 if (page_mapped(page
))
4075 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4076 if (PageSwapCache(page
))
4078 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4081 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4083 VM_BUG_ON(page_mapped(page
));
4084 VM_BUG_ON(page
->mapping
);
4085 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4089 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4090 * In that cases, pages are freed continuously and we can expect pages
4091 * are in the same memcg. All these calls itself limits the number of
4092 * pages freed at once, then uncharge_start/end() is called properly.
4093 * This may be called prural(2) times in a context,
4096 void mem_cgroup_uncharge_start(void)
4098 current
->memcg_batch
.do_batch
++;
4099 /* We can do nest. */
4100 if (current
->memcg_batch
.do_batch
== 1) {
4101 current
->memcg_batch
.memcg
= NULL
;
4102 current
->memcg_batch
.nr_pages
= 0;
4103 current
->memcg_batch
.memsw_nr_pages
= 0;
4107 void mem_cgroup_uncharge_end(void)
4109 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4111 if (!batch
->do_batch
)
4115 if (batch
->do_batch
) /* If stacked, do nothing. */
4121 * This "batch->memcg" is valid without any css_get/put etc...
4122 * bacause we hide charges behind us.
4124 if (batch
->nr_pages
)
4125 res_counter_uncharge(&batch
->memcg
->res
,
4126 batch
->nr_pages
* PAGE_SIZE
);
4127 if (batch
->memsw_nr_pages
)
4128 res_counter_uncharge(&batch
->memcg
->memsw
,
4129 batch
->memsw_nr_pages
* PAGE_SIZE
);
4130 memcg_oom_recover(batch
->memcg
);
4131 /* forget this pointer (for sanity check) */
4132 batch
->memcg
= NULL
;
4137 * called after __delete_from_swap_cache() and drop "page" account.
4138 * memcg information is recorded to swap_cgroup of "ent"
4141 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4143 struct mem_cgroup
*memcg
;
4144 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4146 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4147 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4149 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4152 * record memcg information, if swapout && memcg != NULL,
4153 * mem_cgroup_get() was called in uncharge().
4155 if (do_swap_account
&& swapout
&& memcg
)
4156 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4160 #ifdef CONFIG_MEMCG_SWAP
4162 * called from swap_entry_free(). remove record in swap_cgroup and
4163 * uncharge "memsw" account.
4165 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4167 struct mem_cgroup
*memcg
;
4170 if (!do_swap_account
)
4173 id
= swap_cgroup_record(ent
, 0);
4175 memcg
= mem_cgroup_lookup(id
);
4178 * We uncharge this because swap is freed.
4179 * This memcg can be obsolete one. We avoid calling css_tryget
4181 if (!mem_cgroup_is_root(memcg
))
4182 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4183 mem_cgroup_swap_statistics(memcg
, false);
4184 mem_cgroup_put(memcg
);
4190 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4191 * @entry: swap entry to be moved
4192 * @from: mem_cgroup which the entry is moved from
4193 * @to: mem_cgroup which the entry is moved to
4195 * It succeeds only when the swap_cgroup's record for this entry is the same
4196 * as the mem_cgroup's id of @from.
4198 * Returns 0 on success, -EINVAL on failure.
4200 * The caller must have charged to @to, IOW, called res_counter_charge() about
4201 * both res and memsw, and called css_get().
4203 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4204 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4206 unsigned short old_id
, new_id
;
4208 old_id
= css_id(&from
->css
);
4209 new_id
= css_id(&to
->css
);
4211 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4212 mem_cgroup_swap_statistics(from
, false);
4213 mem_cgroup_swap_statistics(to
, true);
4215 * This function is only called from task migration context now.
4216 * It postpones res_counter and refcount handling till the end
4217 * of task migration(mem_cgroup_clear_mc()) for performance
4218 * improvement. But we cannot postpone mem_cgroup_get(to)
4219 * because if the process that has been moved to @to does
4220 * swap-in, the refcount of @to might be decreased to 0.
4228 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4229 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4236 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4239 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4240 struct mem_cgroup
**memcgp
)
4242 struct mem_cgroup
*memcg
= NULL
;
4243 unsigned int nr_pages
= 1;
4244 struct page_cgroup
*pc
;
4245 enum charge_type ctype
;
4249 if (mem_cgroup_disabled())
4252 if (PageTransHuge(page
))
4253 nr_pages
<<= compound_order(page
);
4255 pc
= lookup_page_cgroup(page
);
4256 lock_page_cgroup(pc
);
4257 if (PageCgroupUsed(pc
)) {
4258 memcg
= pc
->mem_cgroup
;
4259 css_get(&memcg
->css
);
4261 * At migrating an anonymous page, its mapcount goes down
4262 * to 0 and uncharge() will be called. But, even if it's fully
4263 * unmapped, migration may fail and this page has to be
4264 * charged again. We set MIGRATION flag here and delay uncharge
4265 * until end_migration() is called
4267 * Corner Case Thinking
4269 * When the old page was mapped as Anon and it's unmap-and-freed
4270 * while migration was ongoing.
4271 * If unmap finds the old page, uncharge() of it will be delayed
4272 * until end_migration(). If unmap finds a new page, it's
4273 * uncharged when it make mapcount to be 1->0. If unmap code
4274 * finds swap_migration_entry, the new page will not be mapped
4275 * and end_migration() will find it(mapcount==0).
4278 * When the old page was mapped but migraion fails, the kernel
4279 * remaps it. A charge for it is kept by MIGRATION flag even
4280 * if mapcount goes down to 0. We can do remap successfully
4281 * without charging it again.
4284 * The "old" page is under lock_page() until the end of
4285 * migration, so, the old page itself will not be swapped-out.
4286 * If the new page is swapped out before end_migraton, our
4287 * hook to usual swap-out path will catch the event.
4290 SetPageCgroupMigration(pc
);
4292 unlock_page_cgroup(pc
);
4294 * If the page is not charged at this point,
4302 * We charge new page before it's used/mapped. So, even if unlock_page()
4303 * is called before end_migration, we can catch all events on this new
4304 * page. In the case new page is migrated but not remapped, new page's
4305 * mapcount will be finally 0 and we call uncharge in end_migration().
4308 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4310 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4312 * The page is committed to the memcg, but it's not actually
4313 * charged to the res_counter since we plan on replacing the
4314 * old one and only one page is going to be left afterwards.
4316 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4319 /* remove redundant charge if migration failed*/
4320 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4321 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4323 struct page
*used
, *unused
;
4324 struct page_cgroup
*pc
;
4330 if (!migration_ok
) {
4337 anon
= PageAnon(used
);
4338 __mem_cgroup_uncharge_common(unused
,
4339 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4340 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4342 css_put(&memcg
->css
);
4344 * We disallowed uncharge of pages under migration because mapcount
4345 * of the page goes down to zero, temporarly.
4346 * Clear the flag and check the page should be charged.
4348 pc
= lookup_page_cgroup(oldpage
);
4349 lock_page_cgroup(pc
);
4350 ClearPageCgroupMigration(pc
);
4351 unlock_page_cgroup(pc
);
4354 * If a page is a file cache, radix-tree replacement is very atomic
4355 * and we can skip this check. When it was an Anon page, its mapcount
4356 * goes down to 0. But because we added MIGRATION flage, it's not
4357 * uncharged yet. There are several case but page->mapcount check
4358 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4359 * check. (see prepare_charge() also)
4362 mem_cgroup_uncharge_page(used
);
4366 * At replace page cache, newpage is not under any memcg but it's on
4367 * LRU. So, this function doesn't touch res_counter but handles LRU
4368 * in correct way. Both pages are locked so we cannot race with uncharge.
4370 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4371 struct page
*newpage
)
4373 struct mem_cgroup
*memcg
= NULL
;
4374 struct page_cgroup
*pc
;
4375 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4377 if (mem_cgroup_disabled())
4380 pc
= lookup_page_cgroup(oldpage
);
4381 /* fix accounting on old pages */
4382 lock_page_cgroup(pc
);
4383 if (PageCgroupUsed(pc
)) {
4384 memcg
= pc
->mem_cgroup
;
4385 mem_cgroup_charge_statistics(memcg
, false, -1);
4386 ClearPageCgroupUsed(pc
);
4388 unlock_page_cgroup(pc
);
4391 * When called from shmem_replace_page(), in some cases the
4392 * oldpage has already been charged, and in some cases not.
4397 * Even if newpage->mapping was NULL before starting replacement,
4398 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4399 * LRU while we overwrite pc->mem_cgroup.
4401 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4404 #ifdef CONFIG_DEBUG_VM
4405 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4407 struct page_cgroup
*pc
;
4409 pc
= lookup_page_cgroup(page
);
4411 * Can be NULL while feeding pages into the page allocator for
4412 * the first time, i.e. during boot or memory hotplug;
4413 * or when mem_cgroup_disabled().
4415 if (likely(pc
) && PageCgroupUsed(pc
))
4420 bool mem_cgroup_bad_page_check(struct page
*page
)
4422 if (mem_cgroup_disabled())
4425 return lookup_page_cgroup_used(page
) != NULL
;
4428 void mem_cgroup_print_bad_page(struct page
*page
)
4430 struct page_cgroup
*pc
;
4432 pc
= lookup_page_cgroup_used(page
);
4434 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4435 pc
, pc
->flags
, pc
->mem_cgroup
);
4440 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4441 unsigned long long val
)
4444 u64 memswlimit
, memlimit
;
4446 int children
= mem_cgroup_count_children(memcg
);
4447 u64 curusage
, oldusage
;
4451 * For keeping hierarchical_reclaim simple, how long we should retry
4452 * is depends on callers. We set our retry-count to be function
4453 * of # of children which we should visit in this loop.
4455 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4457 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4460 while (retry_count
) {
4461 if (signal_pending(current
)) {
4466 * Rather than hide all in some function, I do this in
4467 * open coded manner. You see what this really does.
4468 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4470 mutex_lock(&set_limit_mutex
);
4471 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4472 if (memswlimit
< val
) {
4474 mutex_unlock(&set_limit_mutex
);
4478 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4482 ret
= res_counter_set_limit(&memcg
->res
, val
);
4484 if (memswlimit
== val
)
4485 memcg
->memsw_is_minimum
= true;
4487 memcg
->memsw_is_minimum
= false;
4489 mutex_unlock(&set_limit_mutex
);
4494 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4495 MEM_CGROUP_RECLAIM_SHRINK
);
4496 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4497 /* Usage is reduced ? */
4498 if (curusage
>= oldusage
)
4501 oldusage
= curusage
;
4503 if (!ret
&& enlarge
)
4504 memcg_oom_recover(memcg
);
4509 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4510 unsigned long long val
)
4513 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4514 int children
= mem_cgroup_count_children(memcg
);
4518 /* see mem_cgroup_resize_res_limit */
4519 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4520 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4521 while (retry_count
) {
4522 if (signal_pending(current
)) {
4527 * Rather than hide all in some function, I do this in
4528 * open coded manner. You see what this really does.
4529 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4531 mutex_lock(&set_limit_mutex
);
4532 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4533 if (memlimit
> val
) {
4535 mutex_unlock(&set_limit_mutex
);
4538 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4539 if (memswlimit
< val
)
4541 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4543 if (memlimit
== val
)
4544 memcg
->memsw_is_minimum
= true;
4546 memcg
->memsw_is_minimum
= false;
4548 mutex_unlock(&set_limit_mutex
);
4553 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4554 MEM_CGROUP_RECLAIM_NOSWAP
|
4555 MEM_CGROUP_RECLAIM_SHRINK
);
4556 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4557 /* Usage is reduced ? */
4558 if (curusage
>= oldusage
)
4561 oldusage
= curusage
;
4563 if (!ret
&& enlarge
)
4564 memcg_oom_recover(memcg
);
4568 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4570 unsigned long *total_scanned
)
4572 unsigned long nr_reclaimed
= 0;
4573 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4574 unsigned long reclaimed
;
4576 struct mem_cgroup_tree_per_zone
*mctz
;
4577 unsigned long long excess
;
4578 unsigned long nr_scanned
;
4583 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4585 * This loop can run a while, specially if mem_cgroup's continuously
4586 * keep exceeding their soft limit and putting the system under
4593 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4598 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4599 gfp_mask
, &nr_scanned
);
4600 nr_reclaimed
+= reclaimed
;
4601 *total_scanned
+= nr_scanned
;
4602 spin_lock(&mctz
->lock
);
4605 * If we failed to reclaim anything from this memory cgroup
4606 * it is time to move on to the next cgroup
4612 * Loop until we find yet another one.
4614 * By the time we get the soft_limit lock
4615 * again, someone might have aded the
4616 * group back on the RB tree. Iterate to
4617 * make sure we get a different mem.
4618 * mem_cgroup_largest_soft_limit_node returns
4619 * NULL if no other cgroup is present on
4623 __mem_cgroup_largest_soft_limit_node(mctz
);
4625 css_put(&next_mz
->memcg
->css
);
4626 else /* next_mz == NULL or other memcg */
4630 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4631 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4633 * One school of thought says that we should not add
4634 * back the node to the tree if reclaim returns 0.
4635 * But our reclaim could return 0, simply because due
4636 * to priority we are exposing a smaller subset of
4637 * memory to reclaim from. Consider this as a longer
4640 /* If excess == 0, no tree ops */
4641 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4642 spin_unlock(&mctz
->lock
);
4643 css_put(&mz
->memcg
->css
);
4646 * Could not reclaim anything and there are no more
4647 * mem cgroups to try or we seem to be looping without
4648 * reclaiming anything.
4650 if (!nr_reclaimed
&&
4652 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4654 } while (!nr_reclaimed
);
4656 css_put(&next_mz
->memcg
->css
);
4657 return nr_reclaimed
;
4661 * mem_cgroup_force_empty_list - clears LRU of a group
4662 * @memcg: group to clear
4665 * @lru: lru to to clear
4667 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4668 * reclaim the pages page themselves - pages are moved to the parent (or root)
4671 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4672 int node
, int zid
, enum lru_list lru
)
4674 struct lruvec
*lruvec
;
4675 unsigned long flags
;
4676 struct list_head
*list
;
4680 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4681 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4682 list
= &lruvec
->lists
[lru
];
4686 struct page_cgroup
*pc
;
4689 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4690 if (list_empty(list
)) {
4691 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4694 page
= list_entry(list
->prev
, struct page
, lru
);
4696 list_move(&page
->lru
, list
);
4698 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4701 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4703 pc
= lookup_page_cgroup(page
);
4705 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4706 /* found lock contention or "pc" is obsolete. */
4711 } while (!list_empty(list
));
4715 * make mem_cgroup's charge to be 0 if there is no task by moving
4716 * all the charges and pages to the parent.
4717 * This enables deleting this mem_cgroup.
4719 * Caller is responsible for holding css reference on the memcg.
4721 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4727 /* This is for making all *used* pages to be on LRU. */
4728 lru_add_drain_all();
4729 drain_all_stock_sync(memcg
);
4730 mem_cgroup_start_move(memcg
);
4731 for_each_node_state(node
, N_MEMORY
) {
4732 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4735 mem_cgroup_force_empty_list(memcg
,
4740 mem_cgroup_end_move(memcg
);
4741 memcg_oom_recover(memcg
);
4745 * Kernel memory may not necessarily be trackable to a specific
4746 * process. So they are not migrated, and therefore we can't
4747 * expect their value to drop to 0 here.
4748 * Having res filled up with kmem only is enough.
4750 * This is a safety check because mem_cgroup_force_empty_list
4751 * could have raced with mem_cgroup_replace_page_cache callers
4752 * so the lru seemed empty but the page could have been added
4753 * right after the check. RES_USAGE should be safe as we always
4754 * charge before adding to the LRU.
4756 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4757 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4758 } while (usage
> 0);
4762 * This mainly exists for tests during the setting of set of use_hierarchy.
4763 * Since this is the very setting we are changing, the current hierarchy value
4766 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4770 /* bounce at first found */
4771 cgroup_for_each_child(pos
, memcg
->css
.cgroup
)
4777 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4778 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4779 * from mem_cgroup_count_children(), in the sense that we don't really care how
4780 * many children we have; we only need to know if we have any. It also counts
4781 * any memcg without hierarchy as infertile.
4783 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4785 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4789 * Reclaims as many pages from the given memcg as possible and moves
4790 * the rest to the parent.
4792 * Caller is responsible for holding css reference for memcg.
4794 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4796 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4797 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4799 /* returns EBUSY if there is a task or if we come here twice. */
4800 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4803 /* we call try-to-free pages for make this cgroup empty */
4804 lru_add_drain_all();
4805 /* try to free all pages in this cgroup */
4806 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4809 if (signal_pending(current
))
4812 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4816 /* maybe some writeback is necessary */
4817 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4822 mem_cgroup_reparent_charges(memcg
);
4827 static int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
4829 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4832 if (mem_cgroup_is_root(memcg
))
4834 css_get(&memcg
->css
);
4835 ret
= mem_cgroup_force_empty(memcg
);
4836 css_put(&memcg
->css
);
4842 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
4844 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
4847 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
4851 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4852 struct cgroup
*parent
= cont
->parent
;
4853 struct mem_cgroup
*parent_memcg
= NULL
;
4856 parent_memcg
= mem_cgroup_from_cont(parent
);
4858 mutex_lock(&memcg_create_mutex
);
4860 if (memcg
->use_hierarchy
== val
)
4864 * If parent's use_hierarchy is set, we can't make any modifications
4865 * in the child subtrees. If it is unset, then the change can
4866 * occur, provided the current cgroup has no children.
4868 * For the root cgroup, parent_mem is NULL, we allow value to be
4869 * set if there are no children.
4871 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4872 (val
== 1 || val
== 0)) {
4873 if (!__memcg_has_children(memcg
))
4874 memcg
->use_hierarchy
= val
;
4881 mutex_unlock(&memcg_create_mutex
);
4887 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4888 enum mem_cgroup_stat_index idx
)
4890 struct mem_cgroup
*iter
;
4893 /* Per-cpu values can be negative, use a signed accumulator */
4894 for_each_mem_cgroup_tree(iter
, memcg
)
4895 val
+= mem_cgroup_read_stat(iter
, idx
);
4897 if (val
< 0) /* race ? */
4902 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4906 if (!mem_cgroup_is_root(memcg
)) {
4908 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4910 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4913 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4914 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4917 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4919 return val
<< PAGE_SHIFT
;
4922 static ssize_t
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
,
4923 struct file
*file
, char __user
*buf
,
4924 size_t nbytes
, loff_t
*ppos
)
4926 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4932 type
= MEMFILE_TYPE(cft
->private);
4933 name
= MEMFILE_ATTR(cft
->private);
4935 if (!do_swap_account
&& type
== _MEMSWAP
)
4940 if (name
== RES_USAGE
)
4941 val
= mem_cgroup_usage(memcg
, false);
4943 val
= res_counter_read_u64(&memcg
->res
, name
);
4946 if (name
== RES_USAGE
)
4947 val
= mem_cgroup_usage(memcg
, true);
4949 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4952 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4958 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4959 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4962 static int memcg_update_kmem_limit(struct cgroup
*cont
, u64 val
)
4965 #ifdef CONFIG_MEMCG_KMEM
4966 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4968 * For simplicity, we won't allow this to be disabled. It also can't
4969 * be changed if the cgroup has children already, or if tasks had
4972 * If tasks join before we set the limit, a person looking at
4973 * kmem.usage_in_bytes will have no way to determine when it took
4974 * place, which makes the value quite meaningless.
4976 * After it first became limited, changes in the value of the limit are
4977 * of course permitted.
4979 mutex_lock(&memcg_create_mutex
);
4980 mutex_lock(&set_limit_mutex
);
4981 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4982 if (cgroup_task_count(cont
) || memcg_has_children(memcg
)) {
4986 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4989 ret
= memcg_update_cache_sizes(memcg
);
4991 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4994 static_key_slow_inc(&memcg_kmem_enabled_key
);
4996 * setting the active bit after the inc will guarantee no one
4997 * starts accounting before all call sites are patched
4999 memcg_kmem_set_active(memcg
);
5002 * kmem charges can outlive the cgroup. In the case of slab
5003 * pages, for instance, a page contain objects from various
5004 * processes, so it is unfeasible to migrate them away. We
5005 * need to reference count the memcg because of that.
5007 mem_cgroup_get(memcg
);
5009 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5011 mutex_unlock(&set_limit_mutex
);
5012 mutex_unlock(&memcg_create_mutex
);
5017 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5020 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5024 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5025 #ifdef CONFIG_MEMCG_KMEM
5027 * When that happen, we need to disable the static branch only on those
5028 * memcgs that enabled it. To achieve this, we would be forced to
5029 * complicate the code by keeping track of which memcgs were the ones
5030 * that actually enabled limits, and which ones got it from its
5033 * It is a lot simpler just to do static_key_slow_inc() on every child
5034 * that is accounted.
5036 if (!memcg_kmem_is_active(memcg
))
5040 * destroy(), called if we fail, will issue static_key_slow_inc() and
5041 * mem_cgroup_put() if kmem is enabled. We have to either call them
5042 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5043 * this more consistent, since it always leads to the same destroy path
5045 mem_cgroup_get(memcg
);
5046 static_key_slow_inc(&memcg_kmem_enabled_key
);
5048 mutex_lock(&set_limit_mutex
);
5049 ret
= memcg_update_cache_sizes(memcg
);
5050 mutex_unlock(&set_limit_mutex
);
5057 * The user of this function is...
5060 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
5063 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5066 unsigned long long val
;
5069 type
= MEMFILE_TYPE(cft
->private);
5070 name
= MEMFILE_ATTR(cft
->private);
5072 if (!do_swap_account
&& type
== _MEMSWAP
)
5077 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5081 /* This function does all necessary parse...reuse it */
5082 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5086 ret
= mem_cgroup_resize_limit(memcg
, val
);
5087 else if (type
== _MEMSWAP
)
5088 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5089 else if (type
== _KMEM
)
5090 ret
= memcg_update_kmem_limit(cont
, val
);
5094 case RES_SOFT_LIMIT
:
5095 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5099 * For memsw, soft limits are hard to implement in terms
5100 * of semantics, for now, we support soft limits for
5101 * control without swap
5104 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5109 ret
= -EINVAL
; /* should be BUG() ? */
5115 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5116 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5118 struct cgroup
*cgroup
;
5119 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5121 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5122 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5123 cgroup
= memcg
->css
.cgroup
;
5124 if (!memcg
->use_hierarchy
)
5127 while (cgroup
->parent
) {
5128 cgroup
= cgroup
->parent
;
5129 memcg
= mem_cgroup_from_cont(cgroup
);
5130 if (!memcg
->use_hierarchy
)
5132 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5133 min_limit
= min(min_limit
, tmp
);
5134 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5135 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5138 *mem_limit
= min_limit
;
5139 *memsw_limit
= min_memsw_limit
;
5142 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
5144 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5148 type
= MEMFILE_TYPE(event
);
5149 name
= MEMFILE_ATTR(event
);
5151 if (!do_swap_account
&& type
== _MEMSWAP
)
5157 res_counter_reset_max(&memcg
->res
);
5158 else if (type
== _MEMSWAP
)
5159 res_counter_reset_max(&memcg
->memsw
);
5160 else if (type
== _KMEM
)
5161 res_counter_reset_max(&memcg
->kmem
);
5167 res_counter_reset_failcnt(&memcg
->res
);
5168 else if (type
== _MEMSWAP
)
5169 res_counter_reset_failcnt(&memcg
->memsw
);
5170 else if (type
== _KMEM
)
5171 res_counter_reset_failcnt(&memcg
->kmem
);
5180 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
5183 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
5187 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5188 struct cftype
*cft
, u64 val
)
5190 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5192 if (val
>= (1 << NR_MOVE_TYPE
))
5196 * No kind of locking is needed in here, because ->can_attach() will
5197 * check this value once in the beginning of the process, and then carry
5198 * on with stale data. This means that changes to this value will only
5199 * affect task migrations starting after the change.
5201 memcg
->move_charge_at_immigrate
= val
;
5205 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5206 struct cftype
*cft
, u64 val
)
5213 static int memcg_numa_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5217 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5218 unsigned long node_nr
;
5219 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5221 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5222 seq_printf(m
, "total=%lu", total_nr
);
5223 for_each_node_state(nid
, N_MEMORY
) {
5224 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5225 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5229 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5230 seq_printf(m
, "file=%lu", file_nr
);
5231 for_each_node_state(nid
, N_MEMORY
) {
5232 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5234 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5238 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5239 seq_printf(m
, "anon=%lu", anon_nr
);
5240 for_each_node_state(nid
, N_MEMORY
) {
5241 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5243 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5247 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5248 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5249 for_each_node_state(nid
, N_MEMORY
) {
5250 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5251 BIT(LRU_UNEVICTABLE
));
5252 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5257 #endif /* CONFIG_NUMA */
5259 static inline void mem_cgroup_lru_names_not_uptodate(void)
5261 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5264 static int memcg_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5267 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5268 struct mem_cgroup
*mi
;
5271 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5272 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5274 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5275 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5278 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5279 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5280 mem_cgroup_read_events(memcg
, i
));
5282 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5283 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5284 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5286 /* Hierarchical information */
5288 unsigned long long limit
, memsw_limit
;
5289 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5290 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5291 if (do_swap_account
)
5292 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5296 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5299 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5301 for_each_mem_cgroup_tree(mi
, memcg
)
5302 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5303 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5306 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5307 unsigned long long val
= 0;
5309 for_each_mem_cgroup_tree(mi
, memcg
)
5310 val
+= mem_cgroup_read_events(mi
, i
);
5311 seq_printf(m
, "total_%s %llu\n",
5312 mem_cgroup_events_names
[i
], val
);
5315 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5316 unsigned long long val
= 0;
5318 for_each_mem_cgroup_tree(mi
, memcg
)
5319 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5320 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5323 #ifdef CONFIG_DEBUG_VM
5326 struct mem_cgroup_per_zone
*mz
;
5327 struct zone_reclaim_stat
*rstat
;
5328 unsigned long recent_rotated
[2] = {0, 0};
5329 unsigned long recent_scanned
[2] = {0, 0};
5331 for_each_online_node(nid
)
5332 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5333 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5334 rstat
= &mz
->lruvec
.reclaim_stat
;
5336 recent_rotated
[0] += rstat
->recent_rotated
[0];
5337 recent_rotated
[1] += rstat
->recent_rotated
[1];
5338 recent_scanned
[0] += rstat
->recent_scanned
[0];
5339 recent_scanned
[1] += rstat
->recent_scanned
[1];
5341 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5342 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5343 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5344 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5351 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
5353 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5355 return mem_cgroup_swappiness(memcg
);
5358 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
5361 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5362 struct mem_cgroup
*parent
;
5367 if (cgrp
->parent
== NULL
)
5370 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5372 mutex_lock(&memcg_create_mutex
);
5374 /* If under hierarchy, only empty-root can set this value */
5375 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5376 mutex_unlock(&memcg_create_mutex
);
5380 memcg
->swappiness
= val
;
5382 mutex_unlock(&memcg_create_mutex
);
5387 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5389 struct mem_cgroup_threshold_ary
*t
;
5395 t
= rcu_dereference(memcg
->thresholds
.primary
);
5397 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5402 usage
= mem_cgroup_usage(memcg
, swap
);
5405 * current_threshold points to threshold just below or equal to usage.
5406 * If it's not true, a threshold was crossed after last
5407 * call of __mem_cgroup_threshold().
5409 i
= t
->current_threshold
;
5412 * Iterate backward over array of thresholds starting from
5413 * current_threshold and check if a threshold is crossed.
5414 * If none of thresholds below usage is crossed, we read
5415 * only one element of the array here.
5417 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5418 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5420 /* i = current_threshold + 1 */
5424 * Iterate forward over array of thresholds starting from
5425 * current_threshold+1 and check if a threshold is crossed.
5426 * If none of thresholds above usage is crossed, we read
5427 * only one element of the array here.
5429 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5430 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5432 /* Update current_threshold */
5433 t
->current_threshold
= i
- 1;
5438 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5441 __mem_cgroup_threshold(memcg
, false);
5442 if (do_swap_account
)
5443 __mem_cgroup_threshold(memcg
, true);
5445 memcg
= parent_mem_cgroup(memcg
);
5449 static int compare_thresholds(const void *a
, const void *b
)
5451 const struct mem_cgroup_threshold
*_a
= a
;
5452 const struct mem_cgroup_threshold
*_b
= b
;
5454 return _a
->threshold
- _b
->threshold
;
5457 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5459 struct mem_cgroup_eventfd_list
*ev
;
5461 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5462 eventfd_signal(ev
->eventfd
, 1);
5466 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5468 struct mem_cgroup
*iter
;
5470 for_each_mem_cgroup_tree(iter
, memcg
)
5471 mem_cgroup_oom_notify_cb(iter
);
5474 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5475 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5477 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5478 struct mem_cgroup_thresholds
*thresholds
;
5479 struct mem_cgroup_threshold_ary
*new;
5480 enum res_type type
= MEMFILE_TYPE(cft
->private);
5481 u64 threshold
, usage
;
5484 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5488 mutex_lock(&memcg
->thresholds_lock
);
5491 thresholds
= &memcg
->thresholds
;
5492 else if (type
== _MEMSWAP
)
5493 thresholds
= &memcg
->memsw_thresholds
;
5497 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5499 /* Check if a threshold crossed before adding a new one */
5500 if (thresholds
->primary
)
5501 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5503 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5505 /* Allocate memory for new array of thresholds */
5506 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5514 /* Copy thresholds (if any) to new array */
5515 if (thresholds
->primary
) {
5516 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5517 sizeof(struct mem_cgroup_threshold
));
5520 /* Add new threshold */
5521 new->entries
[size
- 1].eventfd
= eventfd
;
5522 new->entries
[size
- 1].threshold
= threshold
;
5524 /* Sort thresholds. Registering of new threshold isn't time-critical */
5525 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5526 compare_thresholds
, NULL
);
5528 /* Find current threshold */
5529 new->current_threshold
= -1;
5530 for (i
= 0; i
< size
; i
++) {
5531 if (new->entries
[i
].threshold
<= usage
) {
5533 * new->current_threshold will not be used until
5534 * rcu_assign_pointer(), so it's safe to increment
5537 ++new->current_threshold
;
5542 /* Free old spare buffer and save old primary buffer as spare */
5543 kfree(thresholds
->spare
);
5544 thresholds
->spare
= thresholds
->primary
;
5546 rcu_assign_pointer(thresholds
->primary
, new);
5548 /* To be sure that nobody uses thresholds */
5552 mutex_unlock(&memcg
->thresholds_lock
);
5557 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5558 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5560 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5561 struct mem_cgroup_thresholds
*thresholds
;
5562 struct mem_cgroup_threshold_ary
*new;
5563 enum res_type type
= MEMFILE_TYPE(cft
->private);
5567 mutex_lock(&memcg
->thresholds_lock
);
5569 thresholds
= &memcg
->thresholds
;
5570 else if (type
== _MEMSWAP
)
5571 thresholds
= &memcg
->memsw_thresholds
;
5575 if (!thresholds
->primary
)
5578 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5580 /* Check if a threshold crossed before removing */
5581 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5583 /* Calculate new number of threshold */
5585 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5586 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5590 new = thresholds
->spare
;
5592 /* Set thresholds array to NULL if we don't have thresholds */
5601 /* Copy thresholds and find current threshold */
5602 new->current_threshold
= -1;
5603 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5604 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5607 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5608 if (new->entries
[j
].threshold
<= usage
) {
5610 * new->current_threshold will not be used
5611 * until rcu_assign_pointer(), so it's safe to increment
5614 ++new->current_threshold
;
5620 /* Swap primary and spare array */
5621 thresholds
->spare
= thresholds
->primary
;
5622 /* If all events are unregistered, free the spare array */
5624 kfree(thresholds
->spare
);
5625 thresholds
->spare
= NULL
;
5628 rcu_assign_pointer(thresholds
->primary
, new);
5630 /* To be sure that nobody uses thresholds */
5633 mutex_unlock(&memcg
->thresholds_lock
);
5636 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5637 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5639 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5640 struct mem_cgroup_eventfd_list
*event
;
5641 enum res_type type
= MEMFILE_TYPE(cft
->private);
5643 BUG_ON(type
!= _OOM_TYPE
);
5644 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5648 spin_lock(&memcg_oom_lock
);
5650 event
->eventfd
= eventfd
;
5651 list_add(&event
->list
, &memcg
->oom_notify
);
5653 /* already in OOM ? */
5654 if (atomic_read(&memcg
->under_oom
))
5655 eventfd_signal(eventfd
, 1);
5656 spin_unlock(&memcg_oom_lock
);
5661 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5662 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5664 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5665 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5666 enum res_type type
= MEMFILE_TYPE(cft
->private);
5668 BUG_ON(type
!= _OOM_TYPE
);
5670 spin_lock(&memcg_oom_lock
);
5672 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5673 if (ev
->eventfd
== eventfd
) {
5674 list_del(&ev
->list
);
5679 spin_unlock(&memcg_oom_lock
);
5682 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
5683 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5685 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5687 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5689 if (atomic_read(&memcg
->under_oom
))
5690 cb
->fill(cb
, "under_oom", 1);
5692 cb
->fill(cb
, "under_oom", 0);
5696 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
5697 struct cftype
*cft
, u64 val
)
5699 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5700 struct mem_cgroup
*parent
;
5702 /* cannot set to root cgroup and only 0 and 1 are allowed */
5703 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
5706 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5708 mutex_lock(&memcg_create_mutex
);
5709 /* oom-kill-disable is a flag for subhierarchy. */
5710 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5711 mutex_unlock(&memcg_create_mutex
);
5714 memcg
->oom_kill_disable
= val
;
5716 memcg_oom_recover(memcg
);
5717 mutex_unlock(&memcg_create_mutex
);
5721 #ifdef CONFIG_MEMCG_KMEM
5722 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5726 memcg
->kmemcg_id
= -1;
5727 ret
= memcg_propagate_kmem(memcg
);
5731 return mem_cgroup_sockets_init(memcg
, ss
);
5734 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5736 mem_cgroup_sockets_destroy(memcg
);
5738 memcg_kmem_mark_dead(memcg
);
5740 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5744 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5745 * path here, being careful not to race with memcg_uncharge_kmem: it is
5746 * possible that the charges went down to 0 between mark_dead and the
5747 * res_counter read, so in that case, we don't need the put
5749 if (memcg_kmem_test_and_clear_dead(memcg
))
5750 mem_cgroup_put(memcg
);
5753 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5758 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5763 static struct cftype mem_cgroup_files
[] = {
5765 .name
= "usage_in_bytes",
5766 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5767 .read
= mem_cgroup_read
,
5768 .register_event
= mem_cgroup_usage_register_event
,
5769 .unregister_event
= mem_cgroup_usage_unregister_event
,
5772 .name
= "max_usage_in_bytes",
5773 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5774 .trigger
= mem_cgroup_reset
,
5775 .read
= mem_cgroup_read
,
5778 .name
= "limit_in_bytes",
5779 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5780 .write_string
= mem_cgroup_write
,
5781 .read
= mem_cgroup_read
,
5784 .name
= "soft_limit_in_bytes",
5785 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5786 .write_string
= mem_cgroup_write
,
5787 .read
= mem_cgroup_read
,
5791 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5792 .trigger
= mem_cgroup_reset
,
5793 .read
= mem_cgroup_read
,
5797 .read_seq_string
= memcg_stat_show
,
5800 .name
= "force_empty",
5801 .trigger
= mem_cgroup_force_empty_write
,
5804 .name
= "use_hierarchy",
5805 .write_u64
= mem_cgroup_hierarchy_write
,
5806 .read_u64
= mem_cgroup_hierarchy_read
,
5809 .name
= "swappiness",
5810 .read_u64
= mem_cgroup_swappiness_read
,
5811 .write_u64
= mem_cgroup_swappiness_write
,
5814 .name
= "move_charge_at_immigrate",
5815 .read_u64
= mem_cgroup_move_charge_read
,
5816 .write_u64
= mem_cgroup_move_charge_write
,
5819 .name
= "oom_control",
5820 .read_map
= mem_cgroup_oom_control_read
,
5821 .write_u64
= mem_cgroup_oom_control_write
,
5822 .register_event
= mem_cgroup_oom_register_event
,
5823 .unregister_event
= mem_cgroup_oom_unregister_event
,
5824 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5828 .name
= "numa_stat",
5829 .read_seq_string
= memcg_numa_stat_show
,
5832 #ifdef CONFIG_MEMCG_KMEM
5834 .name
= "kmem.limit_in_bytes",
5835 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5836 .write_string
= mem_cgroup_write
,
5837 .read
= mem_cgroup_read
,
5840 .name
= "kmem.usage_in_bytes",
5841 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5842 .read
= mem_cgroup_read
,
5845 .name
= "kmem.failcnt",
5846 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5847 .trigger
= mem_cgroup_reset
,
5848 .read
= mem_cgroup_read
,
5851 .name
= "kmem.max_usage_in_bytes",
5852 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5853 .trigger
= mem_cgroup_reset
,
5854 .read
= mem_cgroup_read
,
5856 #ifdef CONFIG_SLABINFO
5858 .name
= "kmem.slabinfo",
5859 .read_seq_string
= mem_cgroup_slabinfo_read
,
5863 { }, /* terminate */
5866 #ifdef CONFIG_MEMCG_SWAP
5867 static struct cftype memsw_cgroup_files
[] = {
5869 .name
= "memsw.usage_in_bytes",
5870 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5871 .read
= mem_cgroup_read
,
5872 .register_event
= mem_cgroup_usage_register_event
,
5873 .unregister_event
= mem_cgroup_usage_unregister_event
,
5876 .name
= "memsw.max_usage_in_bytes",
5877 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5878 .trigger
= mem_cgroup_reset
,
5879 .read
= mem_cgroup_read
,
5882 .name
= "memsw.limit_in_bytes",
5883 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5884 .write_string
= mem_cgroup_write
,
5885 .read
= mem_cgroup_read
,
5888 .name
= "memsw.failcnt",
5889 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5890 .trigger
= mem_cgroup_reset
,
5891 .read
= mem_cgroup_read
,
5893 { }, /* terminate */
5896 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5898 struct mem_cgroup_per_node
*pn
;
5899 struct mem_cgroup_per_zone
*mz
;
5900 int zone
, tmp
= node
;
5902 * This routine is called against possible nodes.
5903 * But it's BUG to call kmalloc() against offline node.
5905 * TODO: this routine can waste much memory for nodes which will
5906 * never be onlined. It's better to use memory hotplug callback
5909 if (!node_state(node
, N_NORMAL_MEMORY
))
5911 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5915 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5916 mz
= &pn
->zoneinfo
[zone
];
5917 lruvec_init(&mz
->lruvec
);
5918 mz
->usage_in_excess
= 0;
5919 mz
->on_tree
= false;
5922 memcg
->info
.nodeinfo
[node
] = pn
;
5926 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5928 kfree(memcg
->info
.nodeinfo
[node
]);
5931 static struct mem_cgroup
*mem_cgroup_alloc(void)
5933 struct mem_cgroup
*memcg
;
5934 size_t size
= memcg_size();
5936 /* Can be very big if nr_node_ids is very big */
5937 if (size
< PAGE_SIZE
)
5938 memcg
= kzalloc(size
, GFP_KERNEL
);
5940 memcg
= vzalloc(size
);
5945 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5948 spin_lock_init(&memcg
->pcp_counter_lock
);
5952 if (size
< PAGE_SIZE
)
5960 * At destroying mem_cgroup, references from swap_cgroup can remain.
5961 * (scanning all at force_empty is too costly...)
5963 * Instead of clearing all references at force_empty, we remember
5964 * the number of reference from swap_cgroup and free mem_cgroup when
5965 * it goes down to 0.
5967 * Removal of cgroup itself succeeds regardless of refs from swap.
5970 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5973 size_t size
= memcg_size();
5975 mem_cgroup_remove_from_trees(memcg
);
5976 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5979 free_mem_cgroup_per_zone_info(memcg
, node
);
5981 free_percpu(memcg
->stat
);
5984 * We need to make sure that (at least for now), the jump label
5985 * destruction code runs outside of the cgroup lock. This is because
5986 * get_online_cpus(), which is called from the static_branch update,
5987 * can't be called inside the cgroup_lock. cpusets are the ones
5988 * enforcing this dependency, so if they ever change, we might as well.
5990 * schedule_work() will guarantee this happens. Be careful if you need
5991 * to move this code around, and make sure it is outside
5994 disarm_static_keys(memcg
);
5995 if (size
< PAGE_SIZE
)
6003 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6004 * but in process context. The work_freeing structure is overlaid
6005 * on the rcu_freeing structure, which itself is overlaid on memsw.
6007 static void free_work(struct work_struct
*work
)
6009 struct mem_cgroup
*memcg
;
6011 memcg
= container_of(work
, struct mem_cgroup
, work_freeing
);
6012 __mem_cgroup_free(memcg
);
6015 static void free_rcu(struct rcu_head
*rcu_head
)
6017 struct mem_cgroup
*memcg
;
6019 memcg
= container_of(rcu_head
, struct mem_cgroup
, rcu_freeing
);
6020 INIT_WORK(&memcg
->work_freeing
, free_work
);
6021 schedule_work(&memcg
->work_freeing
);
6024 static void mem_cgroup_get(struct mem_cgroup
*memcg
)
6026 atomic_inc(&memcg
->refcnt
);
6029 static void __mem_cgroup_put(struct mem_cgroup
*memcg
, int count
)
6031 if (atomic_sub_and_test(count
, &memcg
->refcnt
)) {
6032 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
6033 call_rcu(&memcg
->rcu_freeing
, free_rcu
);
6035 mem_cgroup_put(parent
);
6039 static void mem_cgroup_put(struct mem_cgroup
*memcg
)
6041 __mem_cgroup_put(memcg
, 1);
6045 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6047 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6049 if (!memcg
->res
.parent
)
6051 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6053 EXPORT_SYMBOL(parent_mem_cgroup
);
6055 static int mem_cgroup_soft_limit_tree_init(void)
6057 struct mem_cgroup_tree_per_node
*rtpn
;
6058 struct mem_cgroup_tree_per_zone
*rtpz
;
6059 int tmp
, node
, zone
;
6061 for_each_node(node
) {
6063 if (!node_state(node
, N_NORMAL_MEMORY
))
6065 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6069 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6071 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6072 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6073 rtpz
->rb_root
= RB_ROOT
;
6074 spin_lock_init(&rtpz
->lock
);
6080 for_each_node(node
) {
6081 if (!soft_limit_tree
.rb_tree_per_node
[node
])
6083 kfree(soft_limit_tree
.rb_tree_per_node
[node
]);
6084 soft_limit_tree
.rb_tree_per_node
[node
] = NULL
;
6090 static struct cgroup_subsys_state
* __ref
6091 mem_cgroup_css_alloc(struct cgroup
*cont
)
6093 struct mem_cgroup
*memcg
;
6094 long error
= -ENOMEM
;
6097 memcg
= mem_cgroup_alloc();
6099 return ERR_PTR(error
);
6102 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6106 if (cont
->parent
== NULL
) {
6109 if (mem_cgroup_soft_limit_tree_init())
6111 root_mem_cgroup
= memcg
;
6112 for_each_possible_cpu(cpu
) {
6113 struct memcg_stock_pcp
*stock
=
6114 &per_cpu(memcg_stock
, cpu
);
6115 INIT_WORK(&stock
->work
, drain_local_stock
);
6118 res_counter_init(&memcg
->res
, NULL
);
6119 res_counter_init(&memcg
->memsw
, NULL
);
6120 res_counter_init(&memcg
->kmem
, NULL
);
6123 memcg
->last_scanned_node
= MAX_NUMNODES
;
6124 INIT_LIST_HEAD(&memcg
->oom_notify
);
6125 atomic_set(&memcg
->refcnt
, 1);
6126 memcg
->move_charge_at_immigrate
= 0;
6127 mutex_init(&memcg
->thresholds_lock
);
6128 spin_lock_init(&memcg
->move_lock
);
6133 __mem_cgroup_free(memcg
);
6134 return ERR_PTR(error
);
6138 mem_cgroup_css_online(struct cgroup
*cont
)
6140 struct mem_cgroup
*memcg
, *parent
;
6146 mutex_lock(&memcg_create_mutex
);
6147 memcg
= mem_cgroup_from_cont(cont
);
6148 parent
= mem_cgroup_from_cont(cont
->parent
);
6150 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6151 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6152 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6154 if (parent
->use_hierarchy
) {
6155 res_counter_init(&memcg
->res
, &parent
->res
);
6156 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6157 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6160 * We increment refcnt of the parent to ensure that we can
6161 * safely access it on res_counter_charge/uncharge.
6162 * This refcnt will be decremented when freeing this
6163 * mem_cgroup(see mem_cgroup_put).
6165 mem_cgroup_get(parent
);
6167 res_counter_init(&memcg
->res
, NULL
);
6168 res_counter_init(&memcg
->memsw
, NULL
);
6169 res_counter_init(&memcg
->kmem
, NULL
);
6171 * Deeper hierachy with use_hierarchy == false doesn't make
6172 * much sense so let cgroup subsystem know about this
6173 * unfortunate state in our controller.
6175 if (parent
!= root_mem_cgroup
)
6176 mem_cgroup_subsys
.broken_hierarchy
= true;
6179 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6180 mutex_unlock(&memcg_create_mutex
);
6183 * We call put now because our (and parent's) refcnts
6184 * are already in place. mem_cgroup_put() will internally
6185 * call __mem_cgroup_free, so return directly
6187 mem_cgroup_put(memcg
);
6188 if (parent
->use_hierarchy
)
6189 mem_cgroup_put(parent
);
6194 static void mem_cgroup_css_offline(struct cgroup
*cont
)
6196 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6198 mem_cgroup_reparent_charges(memcg
);
6199 mem_cgroup_destroy_all_caches(memcg
);
6202 static void mem_cgroup_css_free(struct cgroup
*cont
)
6204 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6206 kmem_cgroup_destroy(memcg
);
6208 mem_cgroup_put(memcg
);
6212 /* Handlers for move charge at task migration. */
6213 #define PRECHARGE_COUNT_AT_ONCE 256
6214 static int mem_cgroup_do_precharge(unsigned long count
)
6217 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6218 struct mem_cgroup
*memcg
= mc
.to
;
6220 if (mem_cgroup_is_root(memcg
)) {
6221 mc
.precharge
+= count
;
6222 /* we don't need css_get for root */
6225 /* try to charge at once */
6227 struct res_counter
*dummy
;
6229 * "memcg" cannot be under rmdir() because we've already checked
6230 * by cgroup_lock_live_cgroup() that it is not removed and we
6231 * are still under the same cgroup_mutex. So we can postpone
6234 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6236 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6237 PAGE_SIZE
* count
, &dummy
)) {
6238 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6241 mc
.precharge
+= count
;
6245 /* fall back to one by one charge */
6247 if (signal_pending(current
)) {
6251 if (!batch_count
--) {
6252 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6255 ret
= __mem_cgroup_try_charge(NULL
,
6256 GFP_KERNEL
, 1, &memcg
, false);
6258 /* mem_cgroup_clear_mc() will do uncharge later */
6266 * get_mctgt_type - get target type of moving charge
6267 * @vma: the vma the pte to be checked belongs
6268 * @addr: the address corresponding to the pte to be checked
6269 * @ptent: the pte to be checked
6270 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6273 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6274 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6275 * move charge. if @target is not NULL, the page is stored in target->page
6276 * with extra refcnt got(Callers should handle it).
6277 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6278 * target for charge migration. if @target is not NULL, the entry is stored
6281 * Called with pte lock held.
6288 enum mc_target_type
{
6294 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6295 unsigned long addr
, pte_t ptent
)
6297 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6299 if (!page
|| !page_mapped(page
))
6301 if (PageAnon(page
)) {
6302 /* we don't move shared anon */
6305 } else if (!move_file())
6306 /* we ignore mapcount for file pages */
6308 if (!get_page_unless_zero(page
))
6315 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6316 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6318 struct page
*page
= NULL
;
6319 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6321 if (!move_anon() || non_swap_entry(ent
))
6324 * Because lookup_swap_cache() updates some statistics counter,
6325 * we call find_get_page() with swapper_space directly.
6327 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6328 if (do_swap_account
)
6329 entry
->val
= ent
.val
;
6334 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6335 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6341 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6342 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6344 struct page
*page
= NULL
;
6345 struct address_space
*mapping
;
6348 if (!vma
->vm_file
) /* anonymous vma */
6353 mapping
= vma
->vm_file
->f_mapping
;
6354 if (pte_none(ptent
))
6355 pgoff
= linear_page_index(vma
, addr
);
6356 else /* pte_file(ptent) is true */
6357 pgoff
= pte_to_pgoff(ptent
);
6359 /* page is moved even if it's not RSS of this task(page-faulted). */
6360 page
= find_get_page(mapping
, pgoff
);
6363 /* shmem/tmpfs may report page out on swap: account for that too. */
6364 if (radix_tree_exceptional_entry(page
)) {
6365 swp_entry_t swap
= radix_to_swp_entry(page
);
6366 if (do_swap_account
)
6368 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6374 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6375 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6377 struct page
*page
= NULL
;
6378 struct page_cgroup
*pc
;
6379 enum mc_target_type ret
= MC_TARGET_NONE
;
6380 swp_entry_t ent
= { .val
= 0 };
6382 if (pte_present(ptent
))
6383 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6384 else if (is_swap_pte(ptent
))
6385 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6386 else if (pte_none(ptent
) || pte_file(ptent
))
6387 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6389 if (!page
&& !ent
.val
)
6392 pc
= lookup_page_cgroup(page
);
6394 * Do only loose check w/o page_cgroup lock.
6395 * mem_cgroup_move_account() checks the pc is valid or not under
6398 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6399 ret
= MC_TARGET_PAGE
;
6401 target
->page
= page
;
6403 if (!ret
|| !target
)
6406 /* There is a swap entry and a page doesn't exist or isn't charged */
6407 if (ent
.val
&& !ret
&&
6408 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6409 ret
= MC_TARGET_SWAP
;
6416 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6418 * We don't consider swapping or file mapped pages because THP does not
6419 * support them for now.
6420 * Caller should make sure that pmd_trans_huge(pmd) is true.
6422 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6423 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6425 struct page
*page
= NULL
;
6426 struct page_cgroup
*pc
;
6427 enum mc_target_type ret
= MC_TARGET_NONE
;
6429 page
= pmd_page(pmd
);
6430 VM_BUG_ON(!page
|| !PageHead(page
));
6433 pc
= lookup_page_cgroup(page
);
6434 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6435 ret
= MC_TARGET_PAGE
;
6438 target
->page
= page
;
6444 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6445 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6447 return MC_TARGET_NONE
;
6451 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6452 unsigned long addr
, unsigned long end
,
6453 struct mm_walk
*walk
)
6455 struct vm_area_struct
*vma
= walk
->private;
6459 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6460 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6461 mc
.precharge
+= HPAGE_PMD_NR
;
6462 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6466 if (pmd_trans_unstable(pmd
))
6468 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6469 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6470 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6471 mc
.precharge
++; /* increment precharge temporarily */
6472 pte_unmap_unlock(pte
- 1, ptl
);
6478 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6480 unsigned long precharge
;
6481 struct vm_area_struct
*vma
;
6483 down_read(&mm
->mmap_sem
);
6484 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6485 struct mm_walk mem_cgroup_count_precharge_walk
= {
6486 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6490 if (is_vm_hugetlb_page(vma
))
6492 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6493 &mem_cgroup_count_precharge_walk
);
6495 up_read(&mm
->mmap_sem
);
6497 precharge
= mc
.precharge
;
6503 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6505 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6507 VM_BUG_ON(mc
.moving_task
);
6508 mc
.moving_task
= current
;
6509 return mem_cgroup_do_precharge(precharge
);
6512 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6513 static void __mem_cgroup_clear_mc(void)
6515 struct mem_cgroup
*from
= mc
.from
;
6516 struct mem_cgroup
*to
= mc
.to
;
6518 /* we must uncharge all the leftover precharges from mc.to */
6520 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6524 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6525 * we must uncharge here.
6527 if (mc
.moved_charge
) {
6528 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6529 mc
.moved_charge
= 0;
6531 /* we must fixup refcnts and charges */
6532 if (mc
.moved_swap
) {
6533 /* uncharge swap account from the old cgroup */
6534 if (!mem_cgroup_is_root(mc
.from
))
6535 res_counter_uncharge(&mc
.from
->memsw
,
6536 PAGE_SIZE
* mc
.moved_swap
);
6537 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
6539 if (!mem_cgroup_is_root(mc
.to
)) {
6541 * we charged both to->res and to->memsw, so we should
6544 res_counter_uncharge(&mc
.to
->res
,
6545 PAGE_SIZE
* mc
.moved_swap
);
6547 /* we've already done mem_cgroup_get(mc.to) */
6550 memcg_oom_recover(from
);
6551 memcg_oom_recover(to
);
6552 wake_up_all(&mc
.waitq
);
6555 static void mem_cgroup_clear_mc(void)
6557 struct mem_cgroup
*from
= mc
.from
;
6560 * we must clear moving_task before waking up waiters at the end of
6563 mc
.moving_task
= NULL
;
6564 __mem_cgroup_clear_mc();
6565 spin_lock(&mc
.lock
);
6568 spin_unlock(&mc
.lock
);
6569 mem_cgroup_end_move(from
);
6572 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6573 struct cgroup_taskset
*tset
)
6575 struct task_struct
*p
= cgroup_taskset_first(tset
);
6577 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgroup
);
6578 unsigned long move_charge_at_immigrate
;
6581 * We are now commited to this value whatever it is. Changes in this
6582 * tunable will only affect upcoming migrations, not the current one.
6583 * So we need to save it, and keep it going.
6585 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6586 if (move_charge_at_immigrate
) {
6587 struct mm_struct
*mm
;
6588 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6590 VM_BUG_ON(from
== memcg
);
6592 mm
= get_task_mm(p
);
6595 /* We move charges only when we move a owner of the mm */
6596 if (mm
->owner
== p
) {
6599 VM_BUG_ON(mc
.precharge
);
6600 VM_BUG_ON(mc
.moved_charge
);
6601 VM_BUG_ON(mc
.moved_swap
);
6602 mem_cgroup_start_move(from
);
6603 spin_lock(&mc
.lock
);
6606 mc
.immigrate_flags
= move_charge_at_immigrate
;
6607 spin_unlock(&mc
.lock
);
6608 /* We set mc.moving_task later */
6610 ret
= mem_cgroup_precharge_mc(mm
);
6612 mem_cgroup_clear_mc();
6619 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6620 struct cgroup_taskset
*tset
)
6622 mem_cgroup_clear_mc();
6625 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6626 unsigned long addr
, unsigned long end
,
6627 struct mm_walk
*walk
)
6630 struct vm_area_struct
*vma
= walk
->private;
6633 enum mc_target_type target_type
;
6634 union mc_target target
;
6636 struct page_cgroup
*pc
;
6639 * We don't take compound_lock() here but no race with splitting thp
6641 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6642 * under splitting, which means there's no concurrent thp split,
6643 * - if another thread runs into split_huge_page() just after we
6644 * entered this if-block, the thread must wait for page table lock
6645 * to be unlocked in __split_huge_page_splitting(), where the main
6646 * part of thp split is not executed yet.
6648 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6649 if (mc
.precharge
< HPAGE_PMD_NR
) {
6650 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6653 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6654 if (target_type
== MC_TARGET_PAGE
) {
6656 if (!isolate_lru_page(page
)) {
6657 pc
= lookup_page_cgroup(page
);
6658 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6659 pc
, mc
.from
, mc
.to
)) {
6660 mc
.precharge
-= HPAGE_PMD_NR
;
6661 mc
.moved_charge
+= HPAGE_PMD_NR
;
6663 putback_lru_page(page
);
6667 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6671 if (pmd_trans_unstable(pmd
))
6674 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6675 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6676 pte_t ptent
= *(pte
++);
6682 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6683 case MC_TARGET_PAGE
:
6685 if (isolate_lru_page(page
))
6687 pc
= lookup_page_cgroup(page
);
6688 if (!mem_cgroup_move_account(page
, 1, pc
,
6691 /* we uncharge from mc.from later. */
6694 putback_lru_page(page
);
6695 put
: /* get_mctgt_type() gets the page */
6698 case MC_TARGET_SWAP
:
6700 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6702 /* we fixup refcnts and charges later. */
6710 pte_unmap_unlock(pte
- 1, ptl
);
6715 * We have consumed all precharges we got in can_attach().
6716 * We try charge one by one, but don't do any additional
6717 * charges to mc.to if we have failed in charge once in attach()
6720 ret
= mem_cgroup_do_precharge(1);
6728 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6730 struct vm_area_struct
*vma
;
6732 lru_add_drain_all();
6734 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6736 * Someone who are holding the mmap_sem might be waiting in
6737 * waitq. So we cancel all extra charges, wake up all waiters,
6738 * and retry. Because we cancel precharges, we might not be able
6739 * to move enough charges, but moving charge is a best-effort
6740 * feature anyway, so it wouldn't be a big problem.
6742 __mem_cgroup_clear_mc();
6746 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6748 struct mm_walk mem_cgroup_move_charge_walk
= {
6749 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6753 if (is_vm_hugetlb_page(vma
))
6755 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6756 &mem_cgroup_move_charge_walk
);
6759 * means we have consumed all precharges and failed in
6760 * doing additional charge. Just abandon here.
6764 up_read(&mm
->mmap_sem
);
6767 static void mem_cgroup_move_task(struct cgroup
*cont
,
6768 struct cgroup_taskset
*tset
)
6770 struct task_struct
*p
= cgroup_taskset_first(tset
);
6771 struct mm_struct
*mm
= get_task_mm(p
);
6775 mem_cgroup_move_charge(mm
);
6779 mem_cgroup_clear_mc();
6781 #else /* !CONFIG_MMU */
6782 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6783 struct cgroup_taskset
*tset
)
6787 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6788 struct cgroup_taskset
*tset
)
6791 static void mem_cgroup_move_task(struct cgroup
*cont
,
6792 struct cgroup_taskset
*tset
)
6797 struct cgroup_subsys mem_cgroup_subsys
= {
6799 .subsys_id
= mem_cgroup_subsys_id
,
6800 .css_alloc
= mem_cgroup_css_alloc
,
6801 .css_online
= mem_cgroup_css_online
,
6802 .css_offline
= mem_cgroup_css_offline
,
6803 .css_free
= mem_cgroup_css_free
,
6804 .can_attach
= mem_cgroup_can_attach
,
6805 .cancel_attach
= mem_cgroup_cancel_attach
,
6806 .attach
= mem_cgroup_move_task
,
6807 .base_cftypes
= mem_cgroup_files
,
6812 #ifdef CONFIG_MEMCG_SWAP
6813 static int __init
enable_swap_account(char *s
)
6815 /* consider enabled if no parameter or 1 is given */
6816 if (!strcmp(s
, "1"))
6817 really_do_swap_account
= 1;
6818 else if (!strcmp(s
, "0"))
6819 really_do_swap_account
= 0;
6822 __setup("swapaccount=", enable_swap_account
);
6824 static void __init
memsw_file_init(void)
6826 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6829 static void __init
enable_swap_cgroup(void)
6831 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6832 do_swap_account
= 1;
6838 static void __init
enable_swap_cgroup(void)
6844 * The rest of init is performed during ->css_alloc() for root css which
6845 * happens before initcalls. hotcpu_notifier() can't be done together as
6846 * it would introduce circular locking by adding cgroup_lock -> cpu hotplug
6847 * dependency. Do it from a subsys_initcall().
6849 static int __init
mem_cgroup_init(void)
6851 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6852 enable_swap_cgroup();
6855 subsys_initcall(mem_cgroup_init
);