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
[MAX_NUMNODES
];
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 * Per cgroup active and inactive list, similar to the
288 * per zone LRU lists.
290 struct mem_cgroup_lru_info info
;
291 int last_scanned_node
;
293 nodemask_t scan_nodes
;
294 atomic_t numainfo_events
;
295 atomic_t numainfo_updating
;
298 * Should the accounting and control be hierarchical, per subtree?
301 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
309 /* OOM-Killer disable */
310 int oom_kill_disable
;
312 /* set when res.limit == memsw.limit */
313 bool memsw_is_minimum
;
315 /* protect arrays of thresholds */
316 struct mutex thresholds_lock
;
318 /* thresholds for memory usage. RCU-protected */
319 struct mem_cgroup_thresholds thresholds
;
321 /* thresholds for mem+swap usage. RCU-protected */
322 struct mem_cgroup_thresholds memsw_thresholds
;
324 /* For oom notifier event fd */
325 struct list_head oom_notify
;
328 * Should we move charges of a task when a task is moved into this
329 * mem_cgroup ? And what type of charges should we move ?
331 unsigned long move_charge_at_immigrate
;
333 * set > 0 if pages under this cgroup are moving to other cgroup.
335 atomic_t moving_account
;
336 /* taken only while moving_account > 0 */
337 spinlock_t move_lock
;
341 struct mem_cgroup_stat_cpu __percpu
*stat
;
343 * used when a cpu is offlined or other synchronizations
344 * See mem_cgroup_read_stat().
346 struct mem_cgroup_stat_cpu nocpu_base
;
347 spinlock_t pcp_counter_lock
;
349 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
350 struct tcp_memcontrol tcp_mem
;
352 #if defined(CONFIG_MEMCG_KMEM)
353 /* analogous to slab_common's slab_caches list. per-memcg */
354 struct list_head memcg_slab_caches
;
355 /* Not a spinlock, we can take a lot of time walking the list */
356 struct mutex slab_caches_mutex
;
357 /* Index in the kmem_cache->memcg_params->memcg_caches array */
362 /* internal only representation about the status of kmem accounting. */
364 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
365 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
366 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
369 /* We account when limit is on, but only after call sites are patched */
370 #define KMEM_ACCOUNTED_MASK \
371 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
373 #ifdef CONFIG_MEMCG_KMEM
374 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
376 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
379 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
381 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
384 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
386 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
389 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
391 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
394 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
396 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
397 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
400 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
402 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
403 &memcg
->kmem_account_flags
);
407 /* Stuffs for move charges at task migration. */
409 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
410 * left-shifted bitmap of these types.
413 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
414 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
418 /* "mc" and its members are protected by cgroup_mutex */
419 static struct move_charge_struct
{
420 spinlock_t lock
; /* for from, to */
421 struct mem_cgroup
*from
;
422 struct mem_cgroup
*to
;
423 unsigned long precharge
;
424 unsigned long moved_charge
;
425 unsigned long moved_swap
;
426 struct task_struct
*moving_task
; /* a task moving charges */
427 wait_queue_head_t waitq
; /* a waitq for other context */
429 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
430 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
433 static bool move_anon(void)
435 return test_bit(MOVE_CHARGE_TYPE_ANON
,
436 &mc
.to
->move_charge_at_immigrate
);
439 static bool move_file(void)
441 return test_bit(MOVE_CHARGE_TYPE_FILE
,
442 &mc
.to
->move_charge_at_immigrate
);
446 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
447 * limit reclaim to prevent infinite loops, if they ever occur.
449 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
450 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
453 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
454 MEM_CGROUP_CHARGE_TYPE_ANON
,
455 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
456 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
460 /* for encoding cft->private value on file */
468 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
469 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
470 #define MEMFILE_ATTR(val) ((val) & 0xffff)
471 /* Used for OOM nofiier */
472 #define OOM_CONTROL (0)
475 * Reclaim flags for mem_cgroup_hierarchical_reclaim
477 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
478 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
479 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
480 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
482 static void mem_cgroup_get(struct mem_cgroup
*memcg
);
483 static void mem_cgroup_put(struct mem_cgroup
*memcg
);
486 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
488 return container_of(s
, struct mem_cgroup
, css
);
491 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
493 return (memcg
== root_mem_cgroup
);
496 /* Writing them here to avoid exposing memcg's inner layout */
497 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
499 void sock_update_memcg(struct sock
*sk
)
501 if (mem_cgroup_sockets_enabled
) {
502 struct mem_cgroup
*memcg
;
503 struct cg_proto
*cg_proto
;
505 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
507 /* Socket cloning can throw us here with sk_cgrp already
508 * filled. It won't however, necessarily happen from
509 * process context. So the test for root memcg given
510 * the current task's memcg won't help us in this case.
512 * Respecting the original socket's memcg is a better
513 * decision in this case.
516 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
517 mem_cgroup_get(sk
->sk_cgrp
->memcg
);
522 memcg
= mem_cgroup_from_task(current
);
523 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
524 if (!mem_cgroup_is_root(memcg
) && memcg_proto_active(cg_proto
)) {
525 mem_cgroup_get(memcg
);
526 sk
->sk_cgrp
= cg_proto
;
531 EXPORT_SYMBOL(sock_update_memcg
);
533 void sock_release_memcg(struct sock
*sk
)
535 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
536 struct mem_cgroup
*memcg
;
537 WARN_ON(!sk
->sk_cgrp
->memcg
);
538 memcg
= sk
->sk_cgrp
->memcg
;
539 mem_cgroup_put(memcg
);
543 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
545 if (!memcg
|| mem_cgroup_is_root(memcg
))
548 return &memcg
->tcp_mem
.cg_proto
;
550 EXPORT_SYMBOL(tcp_proto_cgroup
);
552 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
554 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
556 static_key_slow_dec(&memcg_socket_limit_enabled
);
559 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
564 #ifdef CONFIG_MEMCG_KMEM
566 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
567 * There are two main reasons for not using the css_id for this:
568 * 1) this works better in sparse environments, where we have a lot of memcgs,
569 * but only a few kmem-limited. Or also, if we have, for instance, 200
570 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
571 * 200 entry array for that.
573 * 2) In order not to violate the cgroup API, we would like to do all memory
574 * allocation in ->create(). At that point, we haven't yet allocated the
575 * css_id. Having a separate index prevents us from messing with the cgroup
578 * The current size of the caches array is stored in
579 * memcg_limited_groups_array_size. It will double each time we have to
582 static DEFINE_IDA(kmem_limited_groups
);
583 int memcg_limited_groups_array_size
;
586 * MIN_SIZE is different than 1, because we would like to avoid going through
587 * the alloc/free process all the time. In a small machine, 4 kmem-limited
588 * cgroups is a reasonable guess. In the future, it could be a parameter or
589 * tunable, but that is strictly not necessary.
591 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
592 * this constant directly from cgroup, but it is understandable that this is
593 * better kept as an internal representation in cgroup.c. In any case, the
594 * css_id space is not getting any smaller, and we don't have to necessarily
595 * increase ours as well if it increases.
597 #define MEMCG_CACHES_MIN_SIZE 4
598 #define MEMCG_CACHES_MAX_SIZE 65535
601 * A lot of the calls to the cache allocation functions are expected to be
602 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
603 * conditional to this static branch, we'll have to allow modules that does
604 * kmem_cache_alloc and the such to see this symbol as well
606 struct static_key memcg_kmem_enabled_key
;
607 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
609 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
611 if (memcg_kmem_is_active(memcg
)) {
612 static_key_slow_dec(&memcg_kmem_enabled_key
);
613 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
616 * This check can't live in kmem destruction function,
617 * since the charges will outlive the cgroup
619 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
622 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
625 #endif /* CONFIG_MEMCG_KMEM */
627 static void disarm_static_keys(struct mem_cgroup
*memcg
)
629 disarm_sock_keys(memcg
);
630 disarm_kmem_keys(memcg
);
633 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
635 static struct mem_cgroup_per_zone
*
636 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
638 return &memcg
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
641 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
646 static struct mem_cgroup_per_zone
*
647 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
649 int nid
= page_to_nid(page
);
650 int zid
= page_zonenum(page
);
652 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
655 static struct mem_cgroup_tree_per_zone
*
656 soft_limit_tree_node_zone(int nid
, int zid
)
658 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
661 static struct mem_cgroup_tree_per_zone
*
662 soft_limit_tree_from_page(struct page
*page
)
664 int nid
= page_to_nid(page
);
665 int zid
= page_zonenum(page
);
667 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
671 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
672 struct mem_cgroup_per_zone
*mz
,
673 struct mem_cgroup_tree_per_zone
*mctz
,
674 unsigned long long new_usage_in_excess
)
676 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
677 struct rb_node
*parent
= NULL
;
678 struct mem_cgroup_per_zone
*mz_node
;
683 mz
->usage_in_excess
= new_usage_in_excess
;
684 if (!mz
->usage_in_excess
)
688 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
690 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
693 * We can't avoid mem cgroups that are over their soft
694 * limit by the same amount
696 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
699 rb_link_node(&mz
->tree_node
, parent
, p
);
700 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
705 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
706 struct mem_cgroup_per_zone
*mz
,
707 struct mem_cgroup_tree_per_zone
*mctz
)
711 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
716 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
717 struct mem_cgroup_per_zone
*mz
,
718 struct mem_cgroup_tree_per_zone
*mctz
)
720 spin_lock(&mctz
->lock
);
721 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
722 spin_unlock(&mctz
->lock
);
726 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
728 unsigned long long excess
;
729 struct mem_cgroup_per_zone
*mz
;
730 struct mem_cgroup_tree_per_zone
*mctz
;
731 int nid
= page_to_nid(page
);
732 int zid
= page_zonenum(page
);
733 mctz
= soft_limit_tree_from_page(page
);
736 * Necessary to update all ancestors when hierarchy is used.
737 * because their event counter is not touched.
739 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
740 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
741 excess
= res_counter_soft_limit_excess(&memcg
->res
);
743 * We have to update the tree if mz is on RB-tree or
744 * mem is over its softlimit.
746 if (excess
|| mz
->on_tree
) {
747 spin_lock(&mctz
->lock
);
748 /* if on-tree, remove it */
750 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
752 * Insert again. mz->usage_in_excess will be updated.
753 * If excess is 0, no tree ops.
755 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
756 spin_unlock(&mctz
->lock
);
761 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
764 struct mem_cgroup_per_zone
*mz
;
765 struct mem_cgroup_tree_per_zone
*mctz
;
767 for_each_node(node
) {
768 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
769 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
770 mctz
= soft_limit_tree_node_zone(node
, zone
);
771 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
776 static struct mem_cgroup_per_zone
*
777 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
779 struct rb_node
*rightmost
= NULL
;
780 struct mem_cgroup_per_zone
*mz
;
784 rightmost
= rb_last(&mctz
->rb_root
);
786 goto done
; /* Nothing to reclaim from */
788 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
790 * Remove the node now but someone else can add it back,
791 * we will to add it back at the end of reclaim to its correct
792 * position in the tree.
794 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
795 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
796 !css_tryget(&mz
->memcg
->css
))
802 static struct mem_cgroup_per_zone
*
803 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
805 struct mem_cgroup_per_zone
*mz
;
807 spin_lock(&mctz
->lock
);
808 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
809 spin_unlock(&mctz
->lock
);
814 * Implementation Note: reading percpu statistics for memcg.
816 * Both of vmstat[] and percpu_counter has threshold and do periodic
817 * synchronization to implement "quick" read. There are trade-off between
818 * reading cost and precision of value. Then, we may have a chance to implement
819 * a periodic synchronizion of counter in memcg's counter.
821 * But this _read() function is used for user interface now. The user accounts
822 * memory usage by memory cgroup and he _always_ requires exact value because
823 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
824 * have to visit all online cpus and make sum. So, for now, unnecessary
825 * synchronization is not implemented. (just implemented for cpu hotplug)
827 * If there are kernel internal actions which can make use of some not-exact
828 * value, and reading all cpu value can be performance bottleneck in some
829 * common workload, threashold and synchonization as vmstat[] should be
832 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
833 enum mem_cgroup_stat_index idx
)
839 for_each_online_cpu(cpu
)
840 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
841 #ifdef CONFIG_HOTPLUG_CPU
842 spin_lock(&memcg
->pcp_counter_lock
);
843 val
+= memcg
->nocpu_base
.count
[idx
];
844 spin_unlock(&memcg
->pcp_counter_lock
);
850 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
853 int val
= (charge
) ? 1 : -1;
854 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
857 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
858 enum mem_cgroup_events_index idx
)
860 unsigned long val
= 0;
863 for_each_online_cpu(cpu
)
864 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
865 #ifdef CONFIG_HOTPLUG_CPU
866 spin_lock(&memcg
->pcp_counter_lock
);
867 val
+= memcg
->nocpu_base
.events
[idx
];
868 spin_unlock(&memcg
->pcp_counter_lock
);
873 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
874 bool anon
, int nr_pages
)
879 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
880 * counted as CACHE even if it's on ANON LRU.
883 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
886 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
889 /* pagein of a big page is an event. So, ignore page size */
891 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
893 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
894 nr_pages
= -nr_pages
; /* for event */
897 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
903 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
905 struct mem_cgroup_per_zone
*mz
;
907 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
908 return mz
->lru_size
[lru
];
912 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
913 unsigned int lru_mask
)
915 struct mem_cgroup_per_zone
*mz
;
917 unsigned long ret
= 0;
919 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
922 if (BIT(lru
) & lru_mask
)
923 ret
+= mz
->lru_size
[lru
];
929 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
930 int nid
, unsigned int lru_mask
)
935 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
936 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
942 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
943 unsigned int lru_mask
)
948 for_each_node_state(nid
, N_MEMORY
)
949 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
953 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
954 enum mem_cgroup_events_target target
)
956 unsigned long val
, next
;
958 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
959 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
960 /* from time_after() in jiffies.h */
961 if ((long)next
- (long)val
< 0) {
963 case MEM_CGROUP_TARGET_THRESH
:
964 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
966 case MEM_CGROUP_TARGET_SOFTLIMIT
:
967 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
969 case MEM_CGROUP_TARGET_NUMAINFO
:
970 next
= val
+ NUMAINFO_EVENTS_TARGET
;
975 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
982 * Check events in order.
985 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
988 /* threshold event is triggered in finer grain than soft limit */
989 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
990 MEM_CGROUP_TARGET_THRESH
))) {
992 bool do_numainfo __maybe_unused
;
994 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
995 MEM_CGROUP_TARGET_SOFTLIMIT
);
997 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
998 MEM_CGROUP_TARGET_NUMAINFO
);
1002 mem_cgroup_threshold(memcg
);
1003 if (unlikely(do_softlimit
))
1004 mem_cgroup_update_tree(memcg
, page
);
1005 #if MAX_NUMNODES > 1
1006 if (unlikely(do_numainfo
))
1007 atomic_inc(&memcg
->numainfo_events
);
1013 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
1015 return mem_cgroup_from_css(
1016 cgroup_subsys_state(cont
, mem_cgroup_subsys_id
));
1019 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1022 * mm_update_next_owner() may clear mm->owner to NULL
1023 * if it races with swapoff, page migration, etc.
1024 * So this can be called with p == NULL.
1029 return mem_cgroup_from_css(task_subsys_state(p
, mem_cgroup_subsys_id
));
1032 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1034 struct mem_cgroup
*memcg
= NULL
;
1039 * Because we have no locks, mm->owner's may be being moved to other
1040 * cgroup. We use css_tryget() here even if this looks
1041 * pessimistic (rather than adding locks here).
1045 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1046 if (unlikely(!memcg
))
1048 } while (!css_tryget(&memcg
->css
));
1054 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1055 * @root: hierarchy root
1056 * @prev: previously returned memcg, NULL on first invocation
1057 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1059 * Returns references to children of the hierarchy below @root, or
1060 * @root itself, or %NULL after a full round-trip.
1062 * Caller must pass the return value in @prev on subsequent
1063 * invocations for reference counting, or use mem_cgroup_iter_break()
1064 * to cancel a hierarchy walk before the round-trip is complete.
1066 * Reclaimers can specify a zone and a priority level in @reclaim to
1067 * divide up the memcgs in the hierarchy among all concurrent
1068 * reclaimers operating on the same zone and priority.
1070 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1071 struct mem_cgroup
*prev
,
1072 struct mem_cgroup_reclaim_cookie
*reclaim
)
1074 struct mem_cgroup
*memcg
= NULL
;
1077 if (mem_cgroup_disabled())
1081 root
= root_mem_cgroup
;
1083 if (prev
&& !reclaim
)
1084 id
= css_id(&prev
->css
);
1086 if (prev
&& prev
!= root
)
1087 css_put(&prev
->css
);
1089 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1096 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1097 struct cgroup_subsys_state
*css
;
1100 int nid
= zone_to_nid(reclaim
->zone
);
1101 int zid
= zone_idx(reclaim
->zone
);
1102 struct mem_cgroup_per_zone
*mz
;
1104 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1105 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1106 if (prev
&& reclaim
->generation
!= iter
->generation
)
1108 id
= iter
->position
;
1112 css
= css_get_next(&mem_cgroup_subsys
, id
+ 1, &root
->css
, &id
);
1114 if (css
== &root
->css
|| css_tryget(css
))
1115 memcg
= mem_cgroup_from_css(css
);
1121 iter
->position
= id
;
1124 else if (!prev
&& memcg
)
1125 reclaim
->generation
= iter
->generation
;
1135 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1136 * @root: hierarchy root
1137 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1139 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1140 struct mem_cgroup
*prev
)
1143 root
= root_mem_cgroup
;
1144 if (prev
&& prev
!= root
)
1145 css_put(&prev
->css
);
1149 * Iteration constructs for visiting all cgroups (under a tree). If
1150 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1151 * be used for reference counting.
1153 #define for_each_mem_cgroup_tree(iter, root) \
1154 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1156 iter = mem_cgroup_iter(root, iter, NULL))
1158 #define for_each_mem_cgroup(iter) \
1159 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1161 iter = mem_cgroup_iter(NULL, iter, NULL))
1163 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1165 struct mem_cgroup
*memcg
;
1168 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1169 if (unlikely(!memcg
))
1174 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1177 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1185 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1188 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1189 * @zone: zone of the wanted lruvec
1190 * @memcg: memcg of the wanted lruvec
1192 * Returns the lru list vector holding pages for the given @zone and
1193 * @mem. This can be the global zone lruvec, if the memory controller
1196 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1197 struct mem_cgroup
*memcg
)
1199 struct mem_cgroup_per_zone
*mz
;
1200 struct lruvec
*lruvec
;
1202 if (mem_cgroup_disabled()) {
1203 lruvec
= &zone
->lruvec
;
1207 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1208 lruvec
= &mz
->lruvec
;
1211 * Since a node can be onlined after the mem_cgroup was created,
1212 * we have to be prepared to initialize lruvec->zone here;
1213 * and if offlined then reonlined, we need to reinitialize it.
1215 if (unlikely(lruvec
->zone
!= zone
))
1216 lruvec
->zone
= zone
;
1221 * Following LRU functions are allowed to be used without PCG_LOCK.
1222 * Operations are called by routine of global LRU independently from memcg.
1223 * What we have to take care of here is validness of pc->mem_cgroup.
1225 * Changes to pc->mem_cgroup happens when
1228 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1229 * It is added to LRU before charge.
1230 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1231 * When moving account, the page is not on LRU. It's isolated.
1235 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1237 * @zone: zone of the page
1239 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1241 struct mem_cgroup_per_zone
*mz
;
1242 struct mem_cgroup
*memcg
;
1243 struct page_cgroup
*pc
;
1244 struct lruvec
*lruvec
;
1246 if (mem_cgroup_disabled()) {
1247 lruvec
= &zone
->lruvec
;
1251 pc
= lookup_page_cgroup(page
);
1252 memcg
= pc
->mem_cgroup
;
1255 * Surreptitiously switch any uncharged offlist page to root:
1256 * an uncharged page off lru does nothing to secure
1257 * its former mem_cgroup from sudden removal.
1259 * Our caller holds lru_lock, and PageCgroupUsed is updated
1260 * under page_cgroup lock: between them, they make all uses
1261 * of pc->mem_cgroup safe.
1263 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1264 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1266 mz
= page_cgroup_zoneinfo(memcg
, page
);
1267 lruvec
= &mz
->lruvec
;
1270 * Since a node can be onlined after the mem_cgroup was created,
1271 * we have to be prepared to initialize lruvec->zone here;
1272 * and if offlined then reonlined, we need to reinitialize it.
1274 if (unlikely(lruvec
->zone
!= zone
))
1275 lruvec
->zone
= zone
;
1280 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1281 * @lruvec: mem_cgroup per zone lru vector
1282 * @lru: index of lru list the page is sitting on
1283 * @nr_pages: positive when adding or negative when removing
1285 * This function must be called when a page is added to or removed from an
1288 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1291 struct mem_cgroup_per_zone
*mz
;
1292 unsigned long *lru_size
;
1294 if (mem_cgroup_disabled())
1297 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1298 lru_size
= mz
->lru_size
+ lru
;
1299 *lru_size
+= nr_pages
;
1300 VM_BUG_ON((long)(*lru_size
) < 0);
1304 * Checks whether given mem is same or in the root_mem_cgroup's
1307 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1308 struct mem_cgroup
*memcg
)
1310 if (root_memcg
== memcg
)
1312 if (!root_memcg
->use_hierarchy
|| !memcg
)
1314 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1317 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1318 struct mem_cgroup
*memcg
)
1323 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1328 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*memcg
)
1331 struct mem_cgroup
*curr
= NULL
;
1332 struct task_struct
*p
;
1334 p
= find_lock_task_mm(task
);
1336 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1340 * All threads may have already detached their mm's, but the oom
1341 * killer still needs to detect if they have already been oom
1342 * killed to prevent needlessly killing additional tasks.
1345 curr
= mem_cgroup_from_task(task
);
1347 css_get(&curr
->css
);
1353 * We should check use_hierarchy of "memcg" not "curr". Because checking
1354 * use_hierarchy of "curr" here make this function true if hierarchy is
1355 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1356 * hierarchy(even if use_hierarchy is disabled in "memcg").
1358 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1359 css_put(&curr
->css
);
1363 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1365 unsigned long inactive_ratio
;
1366 unsigned long inactive
;
1367 unsigned long active
;
1370 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1371 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1373 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1375 inactive_ratio
= int_sqrt(10 * gb
);
1379 return inactive
* inactive_ratio
< active
;
1382 int mem_cgroup_inactive_file_is_low(struct lruvec
*lruvec
)
1384 unsigned long active
;
1385 unsigned long inactive
;
1387 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_FILE
);
1388 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_FILE
);
1390 return (active
> inactive
);
1393 #define mem_cgroup_from_res_counter(counter, member) \
1394 container_of(counter, struct mem_cgroup, member)
1397 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1398 * @memcg: the memory cgroup
1400 * Returns the maximum amount of memory @mem can be charged with, in
1403 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1405 unsigned long long margin
;
1407 margin
= res_counter_margin(&memcg
->res
);
1408 if (do_swap_account
)
1409 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1410 return margin
>> PAGE_SHIFT
;
1413 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1415 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1418 if (cgrp
->parent
== NULL
)
1419 return vm_swappiness
;
1421 return memcg
->swappiness
;
1425 * memcg->moving_account is used for checking possibility that some thread is
1426 * calling move_account(). When a thread on CPU-A starts moving pages under
1427 * a memcg, other threads should check memcg->moving_account under
1428 * rcu_read_lock(), like this:
1432 * memcg->moving_account+1 if (memcg->mocing_account)
1434 * synchronize_rcu() update something.
1439 /* for quick checking without looking up memcg */
1440 atomic_t memcg_moving __read_mostly
;
1442 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1444 atomic_inc(&memcg_moving
);
1445 atomic_inc(&memcg
->moving_account
);
1449 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1452 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1453 * We check NULL in callee rather than caller.
1456 atomic_dec(&memcg_moving
);
1457 atomic_dec(&memcg
->moving_account
);
1462 * 2 routines for checking "mem" is under move_account() or not.
1464 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1465 * is used for avoiding races in accounting. If true,
1466 * pc->mem_cgroup may be overwritten.
1468 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1469 * under hierarchy of moving cgroups. This is for
1470 * waiting at hith-memory prressure caused by "move".
1473 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1475 VM_BUG_ON(!rcu_read_lock_held());
1476 return atomic_read(&memcg
->moving_account
) > 0;
1479 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1481 struct mem_cgroup
*from
;
1482 struct mem_cgroup
*to
;
1485 * Unlike task_move routines, we access mc.to, mc.from not under
1486 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1488 spin_lock(&mc
.lock
);
1494 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1495 || mem_cgroup_same_or_subtree(memcg
, to
);
1497 spin_unlock(&mc
.lock
);
1501 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1503 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1504 if (mem_cgroup_under_move(memcg
)) {
1506 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1507 /* moving charge context might have finished. */
1510 finish_wait(&mc
.waitq
, &wait
);
1518 * Take this lock when
1519 * - a code tries to modify page's memcg while it's USED.
1520 * - a code tries to modify page state accounting in a memcg.
1521 * see mem_cgroup_stolen(), too.
1523 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1524 unsigned long *flags
)
1526 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1529 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1530 unsigned long *flags
)
1532 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1535 #define K(x) ((x) << (PAGE_SHIFT-10))
1537 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1538 * @memcg: The memory cgroup that went over limit
1539 * @p: Task that is going to be killed
1541 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1544 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1546 struct cgroup
*task_cgrp
;
1547 struct cgroup
*mem_cgrp
;
1549 * Need a buffer in BSS, can't rely on allocations. The code relies
1550 * on the assumption that OOM is serialized for memory controller.
1551 * If this assumption is broken, revisit this code.
1553 static char memcg_name
[PATH_MAX
];
1555 struct mem_cgroup
*iter
;
1563 mem_cgrp
= memcg
->css
.cgroup
;
1564 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1566 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1569 * Unfortunately, we are unable to convert to a useful name
1570 * But we'll still print out the usage information
1577 pr_info("Task in %s killed", memcg_name
);
1580 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1588 * Continues from above, so we don't need an KERN_ level
1590 pr_cont(" as a result of limit of %s\n", memcg_name
);
1593 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1594 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1595 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1596 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1597 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1598 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1599 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1600 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1601 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1602 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1603 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1604 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1606 for_each_mem_cgroup_tree(iter
, memcg
) {
1607 pr_info("Memory cgroup stats");
1610 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1612 pr_cont(" for %s", memcg_name
);
1616 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1617 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1619 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1620 K(mem_cgroup_read_stat(iter
, i
)));
1623 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1624 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1625 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1632 * This function returns the number of memcg under hierarchy tree. Returns
1633 * 1(self count) if no children.
1635 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1638 struct mem_cgroup
*iter
;
1640 for_each_mem_cgroup_tree(iter
, memcg
)
1646 * Return the memory (and swap, if configured) limit for a memcg.
1648 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1652 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1655 * Do not consider swap space if we cannot swap due to swappiness
1657 if (mem_cgroup_swappiness(memcg
)) {
1660 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1661 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1664 * If memsw is finite and limits the amount of swap space
1665 * available to this memcg, return that limit.
1667 limit
= min(limit
, memsw
);
1673 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1676 struct mem_cgroup
*iter
;
1677 unsigned long chosen_points
= 0;
1678 unsigned long totalpages
;
1679 unsigned int points
= 0;
1680 struct task_struct
*chosen
= NULL
;
1683 * If current has a pending SIGKILL, then automatically select it. The
1684 * goal is to allow it to allocate so that it may quickly exit and free
1687 if (fatal_signal_pending(current
)) {
1688 set_thread_flag(TIF_MEMDIE
);
1692 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1693 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1694 for_each_mem_cgroup_tree(iter
, memcg
) {
1695 struct cgroup
*cgroup
= iter
->css
.cgroup
;
1696 struct cgroup_iter it
;
1697 struct task_struct
*task
;
1699 cgroup_iter_start(cgroup
, &it
);
1700 while ((task
= cgroup_iter_next(cgroup
, &it
))) {
1701 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1703 case OOM_SCAN_SELECT
:
1705 put_task_struct(chosen
);
1707 chosen_points
= ULONG_MAX
;
1708 get_task_struct(chosen
);
1710 case OOM_SCAN_CONTINUE
:
1712 case OOM_SCAN_ABORT
:
1713 cgroup_iter_end(cgroup
, &it
);
1714 mem_cgroup_iter_break(memcg
, iter
);
1716 put_task_struct(chosen
);
1721 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1722 if (points
> chosen_points
) {
1724 put_task_struct(chosen
);
1726 chosen_points
= points
;
1727 get_task_struct(chosen
);
1730 cgroup_iter_end(cgroup
, &it
);
1735 points
= chosen_points
* 1000 / totalpages
;
1736 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1737 NULL
, "Memory cgroup out of memory");
1740 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1742 unsigned long flags
)
1744 unsigned long total
= 0;
1745 bool noswap
= false;
1748 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1750 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1753 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1755 drain_all_stock_async(memcg
);
1756 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1758 * Allow limit shrinkers, which are triggered directly
1759 * by userspace, to catch signals and stop reclaim
1760 * after minimal progress, regardless of the margin.
1762 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1764 if (mem_cgroup_margin(memcg
))
1767 * If nothing was reclaimed after two attempts, there
1768 * may be no reclaimable pages in this hierarchy.
1777 * test_mem_cgroup_node_reclaimable
1778 * @memcg: the target memcg
1779 * @nid: the node ID to be checked.
1780 * @noswap : specify true here if the user wants flle only information.
1782 * This function returns whether the specified memcg contains any
1783 * reclaimable pages on a node. Returns true if there are any reclaimable
1784 * pages in the node.
1786 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1787 int nid
, bool noswap
)
1789 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1791 if (noswap
|| !total_swap_pages
)
1793 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1798 #if MAX_NUMNODES > 1
1801 * Always updating the nodemask is not very good - even if we have an empty
1802 * list or the wrong list here, we can start from some node and traverse all
1803 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1806 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1810 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1811 * pagein/pageout changes since the last update.
1813 if (!atomic_read(&memcg
->numainfo_events
))
1815 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1818 /* make a nodemask where this memcg uses memory from */
1819 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1821 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1823 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1824 node_clear(nid
, memcg
->scan_nodes
);
1827 atomic_set(&memcg
->numainfo_events
, 0);
1828 atomic_set(&memcg
->numainfo_updating
, 0);
1832 * Selecting a node where we start reclaim from. Because what we need is just
1833 * reducing usage counter, start from anywhere is O,K. Considering
1834 * memory reclaim from current node, there are pros. and cons.
1836 * Freeing memory from current node means freeing memory from a node which
1837 * we'll use or we've used. So, it may make LRU bad. And if several threads
1838 * hit limits, it will see a contention on a node. But freeing from remote
1839 * node means more costs for memory reclaim because of memory latency.
1841 * Now, we use round-robin. Better algorithm is welcomed.
1843 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1847 mem_cgroup_may_update_nodemask(memcg
);
1848 node
= memcg
->last_scanned_node
;
1850 node
= next_node(node
, memcg
->scan_nodes
);
1851 if (node
== MAX_NUMNODES
)
1852 node
= first_node(memcg
->scan_nodes
);
1854 * We call this when we hit limit, not when pages are added to LRU.
1855 * No LRU may hold pages because all pages are UNEVICTABLE or
1856 * memcg is too small and all pages are not on LRU. In that case,
1857 * we use curret node.
1859 if (unlikely(node
== MAX_NUMNODES
))
1860 node
= numa_node_id();
1862 memcg
->last_scanned_node
= node
;
1867 * Check all nodes whether it contains reclaimable pages or not.
1868 * For quick scan, we make use of scan_nodes. This will allow us to skip
1869 * unused nodes. But scan_nodes is lazily updated and may not cotain
1870 * enough new information. We need to do double check.
1872 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1877 * quick check...making use of scan_node.
1878 * We can skip unused nodes.
1880 if (!nodes_empty(memcg
->scan_nodes
)) {
1881 for (nid
= first_node(memcg
->scan_nodes
);
1883 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1885 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1890 * Check rest of nodes.
1892 for_each_node_state(nid
, N_MEMORY
) {
1893 if (node_isset(nid
, memcg
->scan_nodes
))
1895 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1902 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1907 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1909 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1913 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1916 unsigned long *total_scanned
)
1918 struct mem_cgroup
*victim
= NULL
;
1921 unsigned long excess
;
1922 unsigned long nr_scanned
;
1923 struct mem_cgroup_reclaim_cookie reclaim
= {
1928 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
1931 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1936 * If we have not been able to reclaim
1937 * anything, it might because there are
1938 * no reclaimable pages under this hierarchy
1943 * We want to do more targeted reclaim.
1944 * excess >> 2 is not to excessive so as to
1945 * reclaim too much, nor too less that we keep
1946 * coming back to reclaim from this cgroup
1948 if (total
>= (excess
>> 2) ||
1949 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1954 if (!mem_cgroup_reclaimable(victim
, false))
1956 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1958 *total_scanned
+= nr_scanned
;
1959 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
1962 mem_cgroup_iter_break(root_memcg
, victim
);
1967 * Check OOM-Killer is already running under our hierarchy.
1968 * If someone is running, return false.
1969 * Has to be called with memcg_oom_lock
1971 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
1973 struct mem_cgroup
*iter
, *failed
= NULL
;
1975 for_each_mem_cgroup_tree(iter
, memcg
) {
1976 if (iter
->oom_lock
) {
1978 * this subtree of our hierarchy is already locked
1979 * so we cannot give a lock.
1982 mem_cgroup_iter_break(memcg
, iter
);
1985 iter
->oom_lock
= true;
1992 * OK, we failed to lock the whole subtree so we have to clean up
1993 * what we set up to the failing subtree
1995 for_each_mem_cgroup_tree(iter
, memcg
) {
1996 if (iter
== failed
) {
1997 mem_cgroup_iter_break(memcg
, iter
);
2000 iter
->oom_lock
= false;
2006 * Has to be called with memcg_oom_lock
2008 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2010 struct mem_cgroup
*iter
;
2012 for_each_mem_cgroup_tree(iter
, memcg
)
2013 iter
->oom_lock
= false;
2017 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2019 struct mem_cgroup
*iter
;
2021 for_each_mem_cgroup_tree(iter
, memcg
)
2022 atomic_inc(&iter
->under_oom
);
2025 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2027 struct mem_cgroup
*iter
;
2030 * When a new child is created while the hierarchy is under oom,
2031 * mem_cgroup_oom_lock() may not be called. We have to use
2032 * atomic_add_unless() here.
2034 for_each_mem_cgroup_tree(iter
, memcg
)
2035 atomic_add_unless(&iter
->under_oom
, -1, 0);
2038 static DEFINE_SPINLOCK(memcg_oom_lock
);
2039 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2041 struct oom_wait_info
{
2042 struct mem_cgroup
*memcg
;
2046 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2047 unsigned mode
, int sync
, void *arg
)
2049 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2050 struct mem_cgroup
*oom_wait_memcg
;
2051 struct oom_wait_info
*oom_wait_info
;
2053 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2054 oom_wait_memcg
= oom_wait_info
->memcg
;
2057 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2058 * Then we can use css_is_ancestor without taking care of RCU.
2060 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2061 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2063 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2066 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2068 /* for filtering, pass "memcg" as argument. */
2069 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2072 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2074 if (memcg
&& atomic_read(&memcg
->under_oom
))
2075 memcg_wakeup_oom(memcg
);
2079 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2081 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2084 struct oom_wait_info owait
;
2085 bool locked
, need_to_kill
;
2087 owait
.memcg
= memcg
;
2088 owait
.wait
.flags
= 0;
2089 owait
.wait
.func
= memcg_oom_wake_function
;
2090 owait
.wait
.private = current
;
2091 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2092 need_to_kill
= true;
2093 mem_cgroup_mark_under_oom(memcg
);
2095 /* At first, try to OOM lock hierarchy under memcg.*/
2096 spin_lock(&memcg_oom_lock
);
2097 locked
= mem_cgroup_oom_lock(memcg
);
2099 * Even if signal_pending(), we can't quit charge() loop without
2100 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2101 * under OOM is always welcomed, use TASK_KILLABLE here.
2103 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2104 if (!locked
|| memcg
->oom_kill_disable
)
2105 need_to_kill
= false;
2107 mem_cgroup_oom_notify(memcg
);
2108 spin_unlock(&memcg_oom_lock
);
2111 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2112 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2115 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2117 spin_lock(&memcg_oom_lock
);
2119 mem_cgroup_oom_unlock(memcg
);
2120 memcg_wakeup_oom(memcg
);
2121 spin_unlock(&memcg_oom_lock
);
2123 mem_cgroup_unmark_under_oom(memcg
);
2125 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2127 /* Give chance to dying process */
2128 schedule_timeout_uninterruptible(1);
2133 * Currently used to update mapped file statistics, but the routine can be
2134 * generalized to update other statistics as well.
2136 * Notes: Race condition
2138 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2139 * it tends to be costly. But considering some conditions, we doesn't need
2140 * to do so _always_.
2142 * Considering "charge", lock_page_cgroup() is not required because all
2143 * file-stat operations happen after a page is attached to radix-tree. There
2144 * are no race with "charge".
2146 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2147 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2148 * if there are race with "uncharge". Statistics itself is properly handled
2151 * Considering "move", this is an only case we see a race. To make the race
2152 * small, we check mm->moving_account and detect there are possibility of race
2153 * If there is, we take a lock.
2156 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2157 bool *locked
, unsigned long *flags
)
2159 struct mem_cgroup
*memcg
;
2160 struct page_cgroup
*pc
;
2162 pc
= lookup_page_cgroup(page
);
2164 memcg
= pc
->mem_cgroup
;
2165 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2168 * If this memory cgroup is not under account moving, we don't
2169 * need to take move_lock_mem_cgroup(). Because we already hold
2170 * rcu_read_lock(), any calls to move_account will be delayed until
2171 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2173 if (!mem_cgroup_stolen(memcg
))
2176 move_lock_mem_cgroup(memcg
, flags
);
2177 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2178 move_unlock_mem_cgroup(memcg
, flags
);
2184 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2186 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2189 * It's guaranteed that pc->mem_cgroup never changes while
2190 * lock is held because a routine modifies pc->mem_cgroup
2191 * should take move_lock_mem_cgroup().
2193 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2196 void mem_cgroup_update_page_stat(struct page
*page
,
2197 enum mem_cgroup_page_stat_item idx
, int val
)
2199 struct mem_cgroup
*memcg
;
2200 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2201 unsigned long uninitialized_var(flags
);
2203 if (mem_cgroup_disabled())
2206 memcg
= pc
->mem_cgroup
;
2207 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2211 case MEMCG_NR_FILE_MAPPED
:
2212 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2218 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2222 * size of first charge trial. "32" comes from vmscan.c's magic value.
2223 * TODO: maybe necessary to use big numbers in big irons.
2225 #define CHARGE_BATCH 32U
2226 struct memcg_stock_pcp
{
2227 struct mem_cgroup
*cached
; /* this never be root cgroup */
2228 unsigned int nr_pages
;
2229 struct work_struct work
;
2230 unsigned long flags
;
2231 #define FLUSHING_CACHED_CHARGE 0
2233 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2234 static DEFINE_MUTEX(percpu_charge_mutex
);
2237 * consume_stock: Try to consume stocked charge on this cpu.
2238 * @memcg: memcg to consume from.
2239 * @nr_pages: how many pages to charge.
2241 * The charges will only happen if @memcg matches the current cpu's memcg
2242 * stock, and at least @nr_pages are available in that stock. Failure to
2243 * service an allocation will refill the stock.
2245 * returns true if successful, false otherwise.
2247 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2249 struct memcg_stock_pcp
*stock
;
2252 if (nr_pages
> CHARGE_BATCH
)
2255 stock
= &get_cpu_var(memcg_stock
);
2256 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2257 stock
->nr_pages
-= nr_pages
;
2258 else /* need to call res_counter_charge */
2260 put_cpu_var(memcg_stock
);
2265 * Returns stocks cached in percpu to res_counter and reset cached information.
2267 static void drain_stock(struct memcg_stock_pcp
*stock
)
2269 struct mem_cgroup
*old
= stock
->cached
;
2271 if (stock
->nr_pages
) {
2272 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2274 res_counter_uncharge(&old
->res
, bytes
);
2275 if (do_swap_account
)
2276 res_counter_uncharge(&old
->memsw
, bytes
);
2277 stock
->nr_pages
= 0;
2279 stock
->cached
= NULL
;
2283 * This must be called under preempt disabled or must be called by
2284 * a thread which is pinned to local cpu.
2286 static void drain_local_stock(struct work_struct
*dummy
)
2288 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2290 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2294 * Cache charges(val) which is from res_counter, to local per_cpu area.
2295 * This will be consumed by consume_stock() function, later.
2297 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2299 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2301 if (stock
->cached
!= memcg
) { /* reset if necessary */
2303 stock
->cached
= memcg
;
2305 stock
->nr_pages
+= nr_pages
;
2306 put_cpu_var(memcg_stock
);
2310 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2311 * of the hierarchy under it. sync flag says whether we should block
2312 * until the work is done.
2314 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2318 /* Notify other cpus that system-wide "drain" is running */
2321 for_each_online_cpu(cpu
) {
2322 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2323 struct mem_cgroup
*memcg
;
2325 memcg
= stock
->cached
;
2326 if (!memcg
|| !stock
->nr_pages
)
2328 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2330 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2332 drain_local_stock(&stock
->work
);
2334 schedule_work_on(cpu
, &stock
->work
);
2342 for_each_online_cpu(cpu
) {
2343 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2344 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2345 flush_work(&stock
->work
);
2352 * Tries to drain stocked charges in other cpus. This function is asynchronous
2353 * and just put a work per cpu for draining localy on each cpu. Caller can
2354 * expects some charges will be back to res_counter later but cannot wait for
2357 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2360 * If someone calls draining, avoid adding more kworker runs.
2362 if (!mutex_trylock(&percpu_charge_mutex
))
2364 drain_all_stock(root_memcg
, false);
2365 mutex_unlock(&percpu_charge_mutex
);
2368 /* This is a synchronous drain interface. */
2369 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2371 /* called when force_empty is called */
2372 mutex_lock(&percpu_charge_mutex
);
2373 drain_all_stock(root_memcg
, true);
2374 mutex_unlock(&percpu_charge_mutex
);
2378 * This function drains percpu counter value from DEAD cpu and
2379 * move it to local cpu. Note that this function can be preempted.
2381 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2385 spin_lock(&memcg
->pcp_counter_lock
);
2386 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2387 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2389 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2390 memcg
->nocpu_base
.count
[i
] += x
;
2392 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2393 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2395 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2396 memcg
->nocpu_base
.events
[i
] += x
;
2398 spin_unlock(&memcg
->pcp_counter_lock
);
2401 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2402 unsigned long action
,
2405 int cpu
= (unsigned long)hcpu
;
2406 struct memcg_stock_pcp
*stock
;
2407 struct mem_cgroup
*iter
;
2409 if (action
== CPU_ONLINE
)
2412 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2415 for_each_mem_cgroup(iter
)
2416 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2418 stock
= &per_cpu(memcg_stock
, cpu
);
2424 /* See __mem_cgroup_try_charge() for details */
2426 CHARGE_OK
, /* success */
2427 CHARGE_RETRY
, /* need to retry but retry is not bad */
2428 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2429 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2430 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2433 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2434 unsigned int nr_pages
, unsigned int min_pages
,
2437 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2438 struct mem_cgroup
*mem_over_limit
;
2439 struct res_counter
*fail_res
;
2440 unsigned long flags
= 0;
2443 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2446 if (!do_swap_account
)
2448 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2452 res_counter_uncharge(&memcg
->res
, csize
);
2453 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2454 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2456 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2458 * Never reclaim on behalf of optional batching, retry with a
2459 * single page instead.
2461 if (nr_pages
> min_pages
)
2462 return CHARGE_RETRY
;
2464 if (!(gfp_mask
& __GFP_WAIT
))
2465 return CHARGE_WOULDBLOCK
;
2467 if (gfp_mask
& __GFP_NORETRY
)
2468 return CHARGE_NOMEM
;
2470 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2471 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2472 return CHARGE_RETRY
;
2474 * Even though the limit is exceeded at this point, reclaim
2475 * may have been able to free some pages. Retry the charge
2476 * before killing the task.
2478 * Only for regular pages, though: huge pages are rather
2479 * unlikely to succeed so close to the limit, and we fall back
2480 * to regular pages anyway in case of failure.
2482 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2483 return CHARGE_RETRY
;
2486 * At task move, charge accounts can be doubly counted. So, it's
2487 * better to wait until the end of task_move if something is going on.
2489 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2490 return CHARGE_RETRY
;
2492 /* If we don't need to call oom-killer at el, return immediately */
2494 return CHARGE_NOMEM
;
2496 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2497 return CHARGE_OOM_DIE
;
2499 return CHARGE_RETRY
;
2503 * __mem_cgroup_try_charge() does
2504 * 1. detect memcg to be charged against from passed *mm and *ptr,
2505 * 2. update res_counter
2506 * 3. call memory reclaim if necessary.
2508 * In some special case, if the task is fatal, fatal_signal_pending() or
2509 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2510 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2511 * as possible without any hazards. 2: all pages should have a valid
2512 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2513 * pointer, that is treated as a charge to root_mem_cgroup.
2515 * So __mem_cgroup_try_charge() will return
2516 * 0 ... on success, filling *ptr with a valid memcg pointer.
2517 * -ENOMEM ... charge failure because of resource limits.
2518 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2520 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2521 * the oom-killer can be invoked.
2523 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2525 unsigned int nr_pages
,
2526 struct mem_cgroup
**ptr
,
2529 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2530 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2531 struct mem_cgroup
*memcg
= NULL
;
2535 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2536 * in system level. So, allow to go ahead dying process in addition to
2539 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2540 || fatal_signal_pending(current
)))
2544 * We always charge the cgroup the mm_struct belongs to.
2545 * The mm_struct's mem_cgroup changes on task migration if the
2546 * thread group leader migrates. It's possible that mm is not
2547 * set, if so charge the root memcg (happens for pagecache usage).
2550 *ptr
= root_mem_cgroup
;
2552 if (*ptr
) { /* css should be a valid one */
2554 if (mem_cgroup_is_root(memcg
))
2556 if (consume_stock(memcg
, nr_pages
))
2558 css_get(&memcg
->css
);
2560 struct task_struct
*p
;
2563 p
= rcu_dereference(mm
->owner
);
2565 * Because we don't have task_lock(), "p" can exit.
2566 * In that case, "memcg" can point to root or p can be NULL with
2567 * race with swapoff. Then, we have small risk of mis-accouning.
2568 * But such kind of mis-account by race always happens because
2569 * we don't have cgroup_mutex(). It's overkill and we allo that
2571 * (*) swapoff at el will charge against mm-struct not against
2572 * task-struct. So, mm->owner can be NULL.
2574 memcg
= mem_cgroup_from_task(p
);
2576 memcg
= root_mem_cgroup
;
2577 if (mem_cgroup_is_root(memcg
)) {
2581 if (consume_stock(memcg
, nr_pages
)) {
2583 * It seems dagerous to access memcg without css_get().
2584 * But considering how consume_stok works, it's not
2585 * necessary. If consume_stock success, some charges
2586 * from this memcg are cached on this cpu. So, we
2587 * don't need to call css_get()/css_tryget() before
2588 * calling consume_stock().
2593 /* after here, we may be blocked. we need to get refcnt */
2594 if (!css_tryget(&memcg
->css
)) {
2604 /* If killed, bypass charge */
2605 if (fatal_signal_pending(current
)) {
2606 css_put(&memcg
->css
);
2611 if (oom
&& !nr_oom_retries
) {
2613 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2616 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2621 case CHARGE_RETRY
: /* not in OOM situation but retry */
2623 css_put(&memcg
->css
);
2626 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2627 css_put(&memcg
->css
);
2629 case CHARGE_NOMEM
: /* OOM routine works */
2631 css_put(&memcg
->css
);
2634 /* If oom, we never return -ENOMEM */
2637 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2638 css_put(&memcg
->css
);
2641 } while (ret
!= CHARGE_OK
);
2643 if (batch
> nr_pages
)
2644 refill_stock(memcg
, batch
- nr_pages
);
2645 css_put(&memcg
->css
);
2653 *ptr
= root_mem_cgroup
;
2658 * Somemtimes we have to undo a charge we got by try_charge().
2659 * This function is for that and do uncharge, put css's refcnt.
2660 * gotten by try_charge().
2662 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2663 unsigned int nr_pages
)
2665 if (!mem_cgroup_is_root(memcg
)) {
2666 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2668 res_counter_uncharge(&memcg
->res
, bytes
);
2669 if (do_swap_account
)
2670 res_counter_uncharge(&memcg
->memsw
, bytes
);
2675 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2676 * This is useful when moving usage to parent cgroup.
2678 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2679 unsigned int nr_pages
)
2681 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2683 if (mem_cgroup_is_root(memcg
))
2686 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2687 if (do_swap_account
)
2688 res_counter_uncharge_until(&memcg
->memsw
,
2689 memcg
->memsw
.parent
, bytes
);
2693 * A helper function to get mem_cgroup from ID. must be called under
2694 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2695 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2696 * called against removed memcg.)
2698 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2700 struct cgroup_subsys_state
*css
;
2702 /* ID 0 is unused ID */
2705 css
= css_lookup(&mem_cgroup_subsys
, id
);
2708 return mem_cgroup_from_css(css
);
2711 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2713 struct mem_cgroup
*memcg
= NULL
;
2714 struct page_cgroup
*pc
;
2718 VM_BUG_ON(!PageLocked(page
));
2720 pc
= lookup_page_cgroup(page
);
2721 lock_page_cgroup(pc
);
2722 if (PageCgroupUsed(pc
)) {
2723 memcg
= pc
->mem_cgroup
;
2724 if (memcg
&& !css_tryget(&memcg
->css
))
2726 } else if (PageSwapCache(page
)) {
2727 ent
.val
= page_private(page
);
2728 id
= lookup_swap_cgroup_id(ent
);
2730 memcg
= mem_cgroup_lookup(id
);
2731 if (memcg
&& !css_tryget(&memcg
->css
))
2735 unlock_page_cgroup(pc
);
2739 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2741 unsigned int nr_pages
,
2742 enum charge_type ctype
,
2745 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2746 struct zone
*uninitialized_var(zone
);
2747 struct lruvec
*lruvec
;
2748 bool was_on_lru
= false;
2751 lock_page_cgroup(pc
);
2752 VM_BUG_ON(PageCgroupUsed(pc
));
2754 * we don't need page_cgroup_lock about tail pages, becase they are not
2755 * accessed by any other context at this point.
2759 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2760 * may already be on some other mem_cgroup's LRU. Take care of it.
2763 zone
= page_zone(page
);
2764 spin_lock_irq(&zone
->lru_lock
);
2765 if (PageLRU(page
)) {
2766 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2768 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2773 pc
->mem_cgroup
= memcg
;
2775 * We access a page_cgroup asynchronously without lock_page_cgroup().
2776 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2777 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2778 * before USED bit, we need memory barrier here.
2779 * See mem_cgroup_add_lru_list(), etc.
2782 SetPageCgroupUsed(pc
);
2786 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2787 VM_BUG_ON(PageLRU(page
));
2789 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2791 spin_unlock_irq(&zone
->lru_lock
);
2794 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2799 mem_cgroup_charge_statistics(memcg
, anon
, nr_pages
);
2800 unlock_page_cgroup(pc
);
2803 * "charge_statistics" updated event counter. Then, check it.
2804 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2805 * if they exceeds softlimit.
2807 memcg_check_events(memcg
, page
);
2810 static DEFINE_MUTEX(set_limit_mutex
);
2812 #ifdef CONFIG_MEMCG_KMEM
2813 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2815 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2816 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2820 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2821 * in the memcg_cache_params struct.
2823 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2825 struct kmem_cache
*cachep
;
2827 VM_BUG_ON(p
->is_root_cache
);
2828 cachep
= p
->root_cache
;
2829 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2832 #ifdef CONFIG_SLABINFO
2833 static int mem_cgroup_slabinfo_read(struct cgroup
*cont
, struct cftype
*cft
,
2836 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
2837 struct memcg_cache_params
*params
;
2839 if (!memcg_can_account_kmem(memcg
))
2842 print_slabinfo_header(m
);
2844 mutex_lock(&memcg
->slab_caches_mutex
);
2845 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2846 cache_show(memcg_params_to_cache(params
), m
);
2847 mutex_unlock(&memcg
->slab_caches_mutex
);
2853 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2855 struct res_counter
*fail_res
;
2856 struct mem_cgroup
*_memcg
;
2860 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2865 * Conditions under which we can wait for the oom_killer. Those are
2866 * the same conditions tested by the core page allocator
2868 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2871 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2874 if (ret
== -EINTR
) {
2876 * __mem_cgroup_try_charge() chosed to bypass to root due to
2877 * OOM kill or fatal signal. Since our only options are to
2878 * either fail the allocation or charge it to this cgroup, do
2879 * it as a temporary condition. But we can't fail. From a
2880 * kmem/slab perspective, the cache has already been selected,
2881 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2884 * This condition will only trigger if the task entered
2885 * memcg_charge_kmem in a sane state, but was OOM-killed during
2886 * __mem_cgroup_try_charge() above. Tasks that were already
2887 * dying when the allocation triggers should have been already
2888 * directed to the root cgroup in memcontrol.h
2890 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2891 if (do_swap_account
)
2892 res_counter_charge_nofail(&memcg
->memsw
, size
,
2896 res_counter_uncharge(&memcg
->kmem
, size
);
2901 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2903 res_counter_uncharge(&memcg
->res
, size
);
2904 if (do_swap_account
)
2905 res_counter_uncharge(&memcg
->memsw
, size
);
2908 if (res_counter_uncharge(&memcg
->kmem
, size
))
2911 if (memcg_kmem_test_and_clear_dead(memcg
))
2912 mem_cgroup_put(memcg
);
2915 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2920 mutex_lock(&memcg
->slab_caches_mutex
);
2921 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2922 mutex_unlock(&memcg
->slab_caches_mutex
);
2926 * helper for acessing a memcg's index. It will be used as an index in the
2927 * child cache array in kmem_cache, and also to derive its name. This function
2928 * will return -1 when this is not a kmem-limited memcg.
2930 int memcg_cache_id(struct mem_cgroup
*memcg
)
2932 return memcg
? memcg
->kmemcg_id
: -1;
2936 * This ends up being protected by the set_limit mutex, during normal
2937 * operation, because that is its main call site.
2939 * But when we create a new cache, we can call this as well if its parent
2940 * is kmem-limited. That will have to hold set_limit_mutex as well.
2942 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2946 num
= ida_simple_get(&kmem_limited_groups
,
2947 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2951 * After this point, kmem_accounted (that we test atomically in
2952 * the beginning of this conditional), is no longer 0. This
2953 * guarantees only one process will set the following boolean
2954 * to true. We don't need test_and_set because we're protected
2955 * by the set_limit_mutex anyway.
2957 memcg_kmem_set_activated(memcg
);
2959 ret
= memcg_update_all_caches(num
+1);
2961 ida_simple_remove(&kmem_limited_groups
, num
);
2962 memcg_kmem_clear_activated(memcg
);
2966 memcg
->kmemcg_id
= num
;
2967 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
2968 mutex_init(&memcg
->slab_caches_mutex
);
2972 static size_t memcg_caches_array_size(int num_groups
)
2975 if (num_groups
<= 0)
2978 size
= 2 * num_groups
;
2979 if (size
< MEMCG_CACHES_MIN_SIZE
)
2980 size
= MEMCG_CACHES_MIN_SIZE
;
2981 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2982 size
= MEMCG_CACHES_MAX_SIZE
;
2988 * We should update the current array size iff all caches updates succeed. This
2989 * can only be done from the slab side. The slab mutex needs to be held when
2992 void memcg_update_array_size(int num
)
2994 if (num
> memcg_limited_groups_array_size
)
2995 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2998 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3000 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3002 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3004 if (num_groups
> memcg_limited_groups_array_size
) {
3006 ssize_t size
= memcg_caches_array_size(num_groups
);
3008 size
*= sizeof(void *);
3009 size
+= sizeof(struct memcg_cache_params
);
3011 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3012 if (!s
->memcg_params
) {
3013 s
->memcg_params
= cur_params
;
3017 s
->memcg_params
->is_root_cache
= true;
3020 * There is the chance it will be bigger than
3021 * memcg_limited_groups_array_size, if we failed an allocation
3022 * in a cache, in which case all caches updated before it, will
3023 * have a bigger array.
3025 * But if that is the case, the data after
3026 * memcg_limited_groups_array_size is certainly unused
3028 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3029 if (!cur_params
->memcg_caches
[i
])
3031 s
->memcg_params
->memcg_caches
[i
] =
3032 cur_params
->memcg_caches
[i
];
3036 * Ideally, we would wait until all caches succeed, and only
3037 * then free the old one. But this is not worth the extra
3038 * pointer per-cache we'd have to have for this.
3040 * It is not a big deal if some caches are left with a size
3041 * bigger than the others. And all updates will reset this
3049 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3050 struct kmem_cache
*root_cache
)
3052 size_t size
= sizeof(struct memcg_cache_params
);
3054 if (!memcg_kmem_enabled())
3058 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3060 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3061 if (!s
->memcg_params
)
3065 s
->memcg_params
->memcg
= memcg
;
3066 s
->memcg_params
->root_cache
= root_cache
;
3068 s
->memcg_params
->is_root_cache
= true;
3073 void memcg_release_cache(struct kmem_cache
*s
)
3075 struct kmem_cache
*root
;
3076 struct mem_cgroup
*memcg
;
3080 * This happens, for instance, when a root cache goes away before we
3083 if (!s
->memcg_params
)
3086 if (s
->memcg_params
->is_root_cache
)
3089 memcg
= s
->memcg_params
->memcg
;
3090 id
= memcg_cache_id(memcg
);
3092 root
= s
->memcg_params
->root_cache
;
3093 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3094 mem_cgroup_put(memcg
);
3096 mutex_lock(&memcg
->slab_caches_mutex
);
3097 list_del(&s
->memcg_params
->list
);
3098 mutex_unlock(&memcg
->slab_caches_mutex
);
3101 kfree(s
->memcg_params
);
3105 * During the creation a new cache, we need to disable our accounting mechanism
3106 * altogether. This is true even if we are not creating, but rather just
3107 * enqueing new caches to be created.
3109 * This is because that process will trigger allocations; some visible, like
3110 * explicit kmallocs to auxiliary data structures, name strings and internal
3111 * cache structures; some well concealed, like INIT_WORK() that can allocate
3112 * objects during debug.
3114 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3115 * to it. This may not be a bounded recursion: since the first cache creation
3116 * failed to complete (waiting on the allocation), we'll just try to create the
3117 * cache again, failing at the same point.
3119 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3120 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3121 * inside the following two functions.
3123 static inline void memcg_stop_kmem_account(void)
3125 VM_BUG_ON(!current
->mm
);
3126 current
->memcg_kmem_skip_account
++;
3129 static inline void memcg_resume_kmem_account(void)
3131 VM_BUG_ON(!current
->mm
);
3132 current
->memcg_kmem_skip_account
--;
3135 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3137 struct kmem_cache
*cachep
;
3138 struct memcg_cache_params
*p
;
3140 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3142 cachep
= memcg_params_to_cache(p
);
3145 * If we get down to 0 after shrink, we could delete right away.
3146 * However, memcg_release_pages() already puts us back in the workqueue
3147 * in that case. If we proceed deleting, we'll get a dangling
3148 * reference, and removing the object from the workqueue in that case
3149 * is unnecessary complication. We are not a fast path.
3151 * Note that this case is fundamentally different from racing with
3152 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3153 * kmem_cache_shrink, not only we would be reinserting a dead cache
3154 * into the queue, but doing so from inside the worker racing to
3157 * So if we aren't down to zero, we'll just schedule a worker and try
3160 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3161 kmem_cache_shrink(cachep
);
3162 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3165 kmem_cache_destroy(cachep
);
3168 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3170 if (!cachep
->memcg_params
->dead
)
3174 * There are many ways in which we can get here.
3176 * We can get to a memory-pressure situation while the delayed work is
3177 * still pending to run. The vmscan shrinkers can then release all
3178 * cache memory and get us to destruction. If this is the case, we'll
3179 * be executed twice, which is a bug (the second time will execute over
3180 * bogus data). In this case, cancelling the work should be fine.
3182 * But we can also get here from the worker itself, if
3183 * kmem_cache_shrink is enough to shake all the remaining objects and
3184 * get the page count to 0. In this case, we'll deadlock if we try to
3185 * cancel the work (the worker runs with an internal lock held, which
3186 * is the same lock we would hold for cancel_work_sync().)
3188 * Since we can't possibly know who got us here, just refrain from
3189 * running if there is already work pending
3191 if (work_pending(&cachep
->memcg_params
->destroy
))
3194 * We have to defer the actual destroying to a workqueue, because
3195 * we might currently be in a context that cannot sleep.
3197 schedule_work(&cachep
->memcg_params
->destroy
);
3200 static char *memcg_cache_name(struct mem_cgroup
*memcg
, struct kmem_cache
*s
)
3203 struct dentry
*dentry
;
3206 dentry
= rcu_dereference(memcg
->css
.cgroup
->dentry
);
3209 BUG_ON(dentry
== NULL
);
3211 name
= kasprintf(GFP_KERNEL
, "%s(%d:%s)", s
->name
,
3212 memcg_cache_id(memcg
), dentry
->d_name
.name
);
3217 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3218 struct kmem_cache
*s
)
3221 struct kmem_cache
*new;
3223 name
= memcg_cache_name(memcg
, s
);
3227 new = kmem_cache_create_memcg(memcg
, name
, s
->object_size
, s
->align
,
3228 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3231 new->allocflags
|= __GFP_KMEMCG
;
3238 * This lock protects updaters, not readers. We want readers to be as fast as
3239 * they can, and they will either see NULL or a valid cache value. Our model
3240 * allow them to see NULL, in which case the root memcg will be selected.
3242 * We need this lock because multiple allocations to the same cache from a non
3243 * will span more than one worker. Only one of them can create the cache.
3245 static DEFINE_MUTEX(memcg_cache_mutex
);
3246 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3247 struct kmem_cache
*cachep
)
3249 struct kmem_cache
*new_cachep
;
3252 BUG_ON(!memcg_can_account_kmem(memcg
));
3254 idx
= memcg_cache_id(memcg
);
3256 mutex_lock(&memcg_cache_mutex
);
3257 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3261 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3262 if (new_cachep
== NULL
) {
3263 new_cachep
= cachep
;
3267 mem_cgroup_get(memcg
);
3268 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3270 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3272 * the readers won't lock, make sure everybody sees the updated value,
3273 * so they won't put stuff in the queue again for no reason
3277 mutex_unlock(&memcg_cache_mutex
);
3281 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3283 struct kmem_cache
*c
;
3286 if (!s
->memcg_params
)
3288 if (!s
->memcg_params
->is_root_cache
)
3292 * If the cache is being destroyed, we trust that there is no one else
3293 * requesting objects from it. Even if there are, the sanity checks in
3294 * kmem_cache_destroy should caught this ill-case.
3296 * Still, we don't want anyone else freeing memcg_caches under our
3297 * noses, which can happen if a new memcg comes to life. As usual,
3298 * we'll take the set_limit_mutex to protect ourselves against this.
3300 mutex_lock(&set_limit_mutex
);
3301 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3302 c
= s
->memcg_params
->memcg_caches
[i
];
3307 * We will now manually delete the caches, so to avoid races
3308 * we need to cancel all pending destruction workers and
3309 * proceed with destruction ourselves.
3311 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3312 * and that could spawn the workers again: it is likely that
3313 * the cache still have active pages until this very moment.
3314 * This would lead us back to mem_cgroup_destroy_cache.
3316 * But that will not execute at all if the "dead" flag is not
3317 * set, so flip it down to guarantee we are in control.
3319 c
->memcg_params
->dead
= false;
3320 cancel_work_sync(&c
->memcg_params
->destroy
);
3321 kmem_cache_destroy(c
);
3323 mutex_unlock(&set_limit_mutex
);
3326 struct create_work
{
3327 struct mem_cgroup
*memcg
;
3328 struct kmem_cache
*cachep
;
3329 struct work_struct work
;
3332 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3334 struct kmem_cache
*cachep
;
3335 struct memcg_cache_params
*params
;
3337 if (!memcg_kmem_is_active(memcg
))
3340 mutex_lock(&memcg
->slab_caches_mutex
);
3341 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3342 cachep
= memcg_params_to_cache(params
);
3343 cachep
->memcg_params
->dead
= true;
3344 INIT_WORK(&cachep
->memcg_params
->destroy
,
3345 kmem_cache_destroy_work_func
);
3346 schedule_work(&cachep
->memcg_params
->destroy
);
3348 mutex_unlock(&memcg
->slab_caches_mutex
);
3351 static void memcg_create_cache_work_func(struct work_struct
*w
)
3353 struct create_work
*cw
;
3355 cw
= container_of(w
, struct create_work
, work
);
3356 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3357 /* Drop the reference gotten when we enqueued. */
3358 css_put(&cw
->memcg
->css
);
3363 * Enqueue the creation of a per-memcg kmem_cache.
3364 * Called with rcu_read_lock.
3366 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3367 struct kmem_cache
*cachep
)
3369 struct create_work
*cw
;
3371 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3375 /* The corresponding put will be done in the workqueue. */
3376 if (!css_tryget(&memcg
->css
)) {
3382 cw
->cachep
= cachep
;
3384 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3385 schedule_work(&cw
->work
);
3388 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3389 struct kmem_cache
*cachep
)
3392 * We need to stop accounting when we kmalloc, because if the
3393 * corresponding kmalloc cache is not yet created, the first allocation
3394 * in __memcg_create_cache_enqueue will recurse.
3396 * However, it is better to enclose the whole function. Depending on
3397 * the debugging options enabled, INIT_WORK(), for instance, can
3398 * trigger an allocation. This too, will make us recurse. Because at
3399 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3400 * the safest choice is to do it like this, wrapping the whole function.
3402 memcg_stop_kmem_account();
3403 __memcg_create_cache_enqueue(memcg
, cachep
);
3404 memcg_resume_kmem_account();
3407 * Return the kmem_cache we're supposed to use for a slab allocation.
3408 * We try to use the current memcg's version of the cache.
3410 * If the cache does not exist yet, if we are the first user of it,
3411 * we either create it immediately, if possible, or create it asynchronously
3413 * In the latter case, we will let the current allocation go through with
3414 * the original cache.
3416 * Can't be called in interrupt context or from kernel threads.
3417 * This function needs to be called with rcu_read_lock() held.
3419 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3422 struct mem_cgroup
*memcg
;
3425 VM_BUG_ON(!cachep
->memcg_params
);
3426 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3428 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3432 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3435 if (!memcg_can_account_kmem(memcg
))
3438 idx
= memcg_cache_id(memcg
);
3441 * barrier to mare sure we're always seeing the up to date value. The
3442 * code updating memcg_caches will issue a write barrier to match this.
3444 read_barrier_depends();
3445 if (unlikely(cachep
->memcg_params
->memcg_caches
[idx
] == NULL
)) {
3447 * If we are in a safe context (can wait, and not in interrupt
3448 * context), we could be be predictable and return right away.
3449 * This would guarantee that the allocation being performed
3450 * already belongs in the new cache.
3452 * However, there are some clashes that can arrive from locking.
3453 * For instance, because we acquire the slab_mutex while doing
3454 * kmem_cache_dup, this means no further allocation could happen
3455 * with the slab_mutex held.
3457 * Also, because cache creation issue get_online_cpus(), this
3458 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3459 * that ends up reversed during cpu hotplug. (cpuset allocates
3460 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3461 * better to defer everything.
3463 memcg_create_cache_enqueue(memcg
, cachep
);
3467 return cachep
->memcg_params
->memcg_caches
[idx
];
3469 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3472 * We need to verify if the allocation against current->mm->owner's memcg is
3473 * possible for the given order. But the page is not allocated yet, so we'll
3474 * need a further commit step to do the final arrangements.
3476 * It is possible for the task to switch cgroups in this mean time, so at
3477 * commit time, we can't rely on task conversion any longer. We'll then use
3478 * the handle argument to return to the caller which cgroup we should commit
3479 * against. We could also return the memcg directly and avoid the pointer
3480 * passing, but a boolean return value gives better semantics considering
3481 * the compiled-out case as well.
3483 * Returning true means the allocation is possible.
3486 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3488 struct mem_cgroup
*memcg
;
3492 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3495 * very rare case described in mem_cgroup_from_task. Unfortunately there
3496 * isn't much we can do without complicating this too much, and it would
3497 * be gfp-dependent anyway. Just let it go
3499 if (unlikely(!memcg
))
3502 if (!memcg_can_account_kmem(memcg
)) {
3503 css_put(&memcg
->css
);
3507 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3511 css_put(&memcg
->css
);
3515 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3518 struct page_cgroup
*pc
;
3520 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3522 /* The page allocation failed. Revert */
3524 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3528 pc
= lookup_page_cgroup(page
);
3529 lock_page_cgroup(pc
);
3530 pc
->mem_cgroup
= memcg
;
3531 SetPageCgroupUsed(pc
);
3532 unlock_page_cgroup(pc
);
3535 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3537 struct mem_cgroup
*memcg
= NULL
;
3538 struct page_cgroup
*pc
;
3541 pc
= lookup_page_cgroup(page
);
3543 * Fast unlocked return. Theoretically might have changed, have to
3544 * check again after locking.
3546 if (!PageCgroupUsed(pc
))
3549 lock_page_cgroup(pc
);
3550 if (PageCgroupUsed(pc
)) {
3551 memcg
= pc
->mem_cgroup
;
3552 ClearPageCgroupUsed(pc
);
3554 unlock_page_cgroup(pc
);
3557 * We trust that only if there is a memcg associated with the page, it
3558 * is a valid allocation
3563 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3564 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3567 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3570 #endif /* CONFIG_MEMCG_KMEM */
3572 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3574 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3576 * Because tail pages are not marked as "used", set it. We're under
3577 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3578 * charge/uncharge will be never happen and move_account() is done under
3579 * compound_lock(), so we don't have to take care of races.
3581 void mem_cgroup_split_huge_fixup(struct page
*head
)
3583 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3584 struct page_cgroup
*pc
;
3587 if (mem_cgroup_disabled())
3589 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3591 pc
->mem_cgroup
= head_pc
->mem_cgroup
;
3592 smp_wmb();/* see __commit_charge() */
3593 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3596 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3599 * mem_cgroup_move_account - move account of the page
3601 * @nr_pages: number of regular pages (>1 for huge pages)
3602 * @pc: page_cgroup of the page.
3603 * @from: mem_cgroup which the page is moved from.
3604 * @to: mem_cgroup which the page is moved to. @from != @to.
3606 * The caller must confirm following.
3607 * - page is not on LRU (isolate_page() is useful.)
3608 * - compound_lock is held when nr_pages > 1
3610 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3613 static int mem_cgroup_move_account(struct page
*page
,
3614 unsigned int nr_pages
,
3615 struct page_cgroup
*pc
,
3616 struct mem_cgroup
*from
,
3617 struct mem_cgroup
*to
)
3619 unsigned long flags
;
3621 bool anon
= PageAnon(page
);
3623 VM_BUG_ON(from
== to
);
3624 VM_BUG_ON(PageLRU(page
));
3626 * The page is isolated from LRU. So, collapse function
3627 * will not handle this page. But page splitting can happen.
3628 * Do this check under compound_page_lock(). The caller should
3632 if (nr_pages
> 1 && !PageTransHuge(page
))
3635 lock_page_cgroup(pc
);
3638 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3641 move_lock_mem_cgroup(from
, &flags
);
3643 if (!anon
&& page_mapped(page
)) {
3644 /* Update mapped_file data for mem_cgroup */
3646 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3647 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3650 mem_cgroup_charge_statistics(from
, anon
, -nr_pages
);
3652 /* caller should have done css_get */
3653 pc
->mem_cgroup
= to
;
3654 mem_cgroup_charge_statistics(to
, anon
, nr_pages
);
3655 move_unlock_mem_cgroup(from
, &flags
);
3658 unlock_page_cgroup(pc
);
3662 memcg_check_events(to
, page
);
3663 memcg_check_events(from
, page
);
3669 * mem_cgroup_move_parent - moves page to the parent group
3670 * @page: the page to move
3671 * @pc: page_cgroup of the page
3672 * @child: page's cgroup
3674 * move charges to its parent or the root cgroup if the group has no
3675 * parent (aka use_hierarchy==0).
3676 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3677 * mem_cgroup_move_account fails) the failure is always temporary and
3678 * it signals a race with a page removal/uncharge or migration. In the
3679 * first case the page is on the way out and it will vanish from the LRU
3680 * on the next attempt and the call should be retried later.
3681 * Isolation from the LRU fails only if page has been isolated from
3682 * the LRU since we looked at it and that usually means either global
3683 * reclaim or migration going on. The page will either get back to the
3685 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3686 * (!PageCgroupUsed) or moved to a different group. The page will
3687 * disappear in the next attempt.
3689 static int mem_cgroup_move_parent(struct page
*page
,
3690 struct page_cgroup
*pc
,
3691 struct mem_cgroup
*child
)
3693 struct mem_cgroup
*parent
;
3694 unsigned int nr_pages
;
3695 unsigned long uninitialized_var(flags
);
3698 VM_BUG_ON(mem_cgroup_is_root(child
));
3701 if (!get_page_unless_zero(page
))
3703 if (isolate_lru_page(page
))
3706 nr_pages
= hpage_nr_pages(page
);
3708 parent
= parent_mem_cgroup(child
);
3710 * If no parent, move charges to root cgroup.
3713 parent
= root_mem_cgroup
;
3716 VM_BUG_ON(!PageTransHuge(page
));
3717 flags
= compound_lock_irqsave(page
);
3720 ret
= mem_cgroup_move_account(page
, nr_pages
,
3723 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3726 compound_unlock_irqrestore(page
, flags
);
3727 putback_lru_page(page
);
3735 * Charge the memory controller for page usage.
3737 * 0 if the charge was successful
3738 * < 0 if the cgroup is over its limit
3740 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3741 gfp_t gfp_mask
, enum charge_type ctype
)
3743 struct mem_cgroup
*memcg
= NULL
;
3744 unsigned int nr_pages
= 1;
3748 if (PageTransHuge(page
)) {
3749 nr_pages
<<= compound_order(page
);
3750 VM_BUG_ON(!PageTransHuge(page
));
3752 * Never OOM-kill a process for a huge page. The
3753 * fault handler will fall back to regular pages.
3758 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3761 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3765 int mem_cgroup_newpage_charge(struct page
*page
,
3766 struct mm_struct
*mm
, gfp_t gfp_mask
)
3768 if (mem_cgroup_disabled())
3770 VM_BUG_ON(page_mapped(page
));
3771 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3773 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3774 MEM_CGROUP_CHARGE_TYPE_ANON
);
3778 * While swap-in, try_charge -> commit or cancel, the page is locked.
3779 * And when try_charge() successfully returns, one refcnt to memcg without
3780 * struct page_cgroup is acquired. This refcnt will be consumed by
3781 * "commit()" or removed by "cancel()"
3783 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3786 struct mem_cgroup
**memcgp
)
3788 struct mem_cgroup
*memcg
;
3789 struct page_cgroup
*pc
;
3792 pc
= lookup_page_cgroup(page
);
3794 * Every swap fault against a single page tries to charge the
3795 * page, bail as early as possible. shmem_unuse() encounters
3796 * already charged pages, too. The USED bit is protected by
3797 * the page lock, which serializes swap cache removal, which
3798 * in turn serializes uncharging.
3800 if (PageCgroupUsed(pc
))
3802 if (!do_swap_account
)
3804 memcg
= try_get_mem_cgroup_from_page(page
);
3808 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3809 css_put(&memcg
->css
);
3814 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3820 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3821 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3824 if (mem_cgroup_disabled())
3827 * A racing thread's fault, or swapoff, may have already
3828 * updated the pte, and even removed page from swap cache: in
3829 * those cases unuse_pte()'s pte_same() test will fail; but
3830 * there's also a KSM case which does need to charge the page.
3832 if (!PageSwapCache(page
)) {
3835 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3840 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3843 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3845 if (mem_cgroup_disabled())
3849 __mem_cgroup_cancel_charge(memcg
, 1);
3853 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3854 enum charge_type ctype
)
3856 if (mem_cgroup_disabled())
3861 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3863 * Now swap is on-memory. This means this page may be
3864 * counted both as mem and swap....double count.
3865 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3866 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3867 * may call delete_from_swap_cache() before reach here.
3869 if (do_swap_account
&& PageSwapCache(page
)) {
3870 swp_entry_t ent
= {.val
= page_private(page
)};
3871 mem_cgroup_uncharge_swap(ent
);
3875 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3876 struct mem_cgroup
*memcg
)
3878 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3879 MEM_CGROUP_CHARGE_TYPE_ANON
);
3882 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3885 struct mem_cgroup
*memcg
= NULL
;
3886 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3889 if (mem_cgroup_disabled())
3891 if (PageCompound(page
))
3894 if (!PageSwapCache(page
))
3895 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3896 else { /* page is swapcache/shmem */
3897 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3900 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3905 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3906 unsigned int nr_pages
,
3907 const enum charge_type ctype
)
3909 struct memcg_batch_info
*batch
= NULL
;
3910 bool uncharge_memsw
= true;
3912 /* If swapout, usage of swap doesn't decrease */
3913 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3914 uncharge_memsw
= false;
3916 batch
= ¤t
->memcg_batch
;
3918 * In usual, we do css_get() when we remember memcg pointer.
3919 * But in this case, we keep res->usage until end of a series of
3920 * uncharges. Then, it's ok to ignore memcg's refcnt.
3923 batch
->memcg
= memcg
;
3925 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3926 * In those cases, all pages freed continuously can be expected to be in
3927 * the same cgroup and we have chance to coalesce uncharges.
3928 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3929 * because we want to do uncharge as soon as possible.
3932 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3933 goto direct_uncharge
;
3936 goto direct_uncharge
;
3939 * In typical case, batch->memcg == mem. This means we can
3940 * merge a series of uncharges to an uncharge of res_counter.
3941 * If not, we uncharge res_counter ony by one.
3943 if (batch
->memcg
!= memcg
)
3944 goto direct_uncharge
;
3945 /* remember freed charge and uncharge it later */
3948 batch
->memsw_nr_pages
++;
3951 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3953 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3954 if (unlikely(batch
->memcg
!= memcg
))
3955 memcg_oom_recover(memcg
);
3959 * uncharge if !page_mapped(page)
3961 static struct mem_cgroup
*
3962 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3965 struct mem_cgroup
*memcg
= NULL
;
3966 unsigned int nr_pages
= 1;
3967 struct page_cgroup
*pc
;
3970 if (mem_cgroup_disabled())
3973 VM_BUG_ON(PageSwapCache(page
));
3975 if (PageTransHuge(page
)) {
3976 nr_pages
<<= compound_order(page
);
3977 VM_BUG_ON(!PageTransHuge(page
));
3980 * Check if our page_cgroup is valid
3982 pc
= lookup_page_cgroup(page
);
3983 if (unlikely(!PageCgroupUsed(pc
)))
3986 lock_page_cgroup(pc
);
3988 memcg
= pc
->mem_cgroup
;
3990 if (!PageCgroupUsed(pc
))
3993 anon
= PageAnon(page
);
3996 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3998 * Generally PageAnon tells if it's the anon statistics to be
3999 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4000 * used before page reached the stage of being marked PageAnon.
4004 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4005 /* See mem_cgroup_prepare_migration() */
4006 if (page_mapped(page
))
4009 * Pages under migration may not be uncharged. But
4010 * end_migration() /must/ be the one uncharging the
4011 * unused post-migration page and so it has to call
4012 * here with the migration bit still set. See the
4013 * res_counter handling below.
4015 if (!end_migration
&& PageCgroupMigration(pc
))
4018 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4019 if (!PageAnon(page
)) { /* Shared memory */
4020 if (page
->mapping
&& !page_is_file_cache(page
))
4022 } else if (page_mapped(page
)) /* Anon */
4029 mem_cgroup_charge_statistics(memcg
, anon
, -nr_pages
);
4031 ClearPageCgroupUsed(pc
);
4033 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4034 * freed from LRU. This is safe because uncharged page is expected not
4035 * to be reused (freed soon). Exception is SwapCache, it's handled by
4036 * special functions.
4039 unlock_page_cgroup(pc
);
4041 * even after unlock, we have memcg->res.usage here and this memcg
4042 * will never be freed.
4044 memcg_check_events(memcg
, page
);
4045 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4046 mem_cgroup_swap_statistics(memcg
, true);
4047 mem_cgroup_get(memcg
);
4050 * Migration does not charge the res_counter for the
4051 * replacement page, so leave it alone when phasing out the
4052 * page that is unused after the migration.
4054 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4055 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4060 unlock_page_cgroup(pc
);
4064 void mem_cgroup_uncharge_page(struct page
*page
)
4067 if (page_mapped(page
))
4069 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4070 if (PageSwapCache(page
))
4072 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4075 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4077 VM_BUG_ON(page_mapped(page
));
4078 VM_BUG_ON(page
->mapping
);
4079 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4083 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4084 * In that cases, pages are freed continuously and we can expect pages
4085 * are in the same memcg. All these calls itself limits the number of
4086 * pages freed at once, then uncharge_start/end() is called properly.
4087 * This may be called prural(2) times in a context,
4090 void mem_cgroup_uncharge_start(void)
4092 current
->memcg_batch
.do_batch
++;
4093 /* We can do nest. */
4094 if (current
->memcg_batch
.do_batch
== 1) {
4095 current
->memcg_batch
.memcg
= NULL
;
4096 current
->memcg_batch
.nr_pages
= 0;
4097 current
->memcg_batch
.memsw_nr_pages
= 0;
4101 void mem_cgroup_uncharge_end(void)
4103 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4105 if (!batch
->do_batch
)
4109 if (batch
->do_batch
) /* If stacked, do nothing. */
4115 * This "batch->memcg" is valid without any css_get/put etc...
4116 * bacause we hide charges behind us.
4118 if (batch
->nr_pages
)
4119 res_counter_uncharge(&batch
->memcg
->res
,
4120 batch
->nr_pages
* PAGE_SIZE
);
4121 if (batch
->memsw_nr_pages
)
4122 res_counter_uncharge(&batch
->memcg
->memsw
,
4123 batch
->memsw_nr_pages
* PAGE_SIZE
);
4124 memcg_oom_recover(batch
->memcg
);
4125 /* forget this pointer (for sanity check) */
4126 batch
->memcg
= NULL
;
4131 * called after __delete_from_swap_cache() and drop "page" account.
4132 * memcg information is recorded to swap_cgroup of "ent"
4135 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4137 struct mem_cgroup
*memcg
;
4138 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4140 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4141 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4143 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4146 * record memcg information, if swapout && memcg != NULL,
4147 * mem_cgroup_get() was called in uncharge().
4149 if (do_swap_account
&& swapout
&& memcg
)
4150 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4154 #ifdef CONFIG_MEMCG_SWAP
4156 * called from swap_entry_free(). remove record in swap_cgroup and
4157 * uncharge "memsw" account.
4159 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4161 struct mem_cgroup
*memcg
;
4164 if (!do_swap_account
)
4167 id
= swap_cgroup_record(ent
, 0);
4169 memcg
= mem_cgroup_lookup(id
);
4172 * We uncharge this because swap is freed.
4173 * This memcg can be obsolete one. We avoid calling css_tryget
4175 if (!mem_cgroup_is_root(memcg
))
4176 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4177 mem_cgroup_swap_statistics(memcg
, false);
4178 mem_cgroup_put(memcg
);
4184 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4185 * @entry: swap entry to be moved
4186 * @from: mem_cgroup which the entry is moved from
4187 * @to: mem_cgroup which the entry is moved to
4189 * It succeeds only when the swap_cgroup's record for this entry is the same
4190 * as the mem_cgroup's id of @from.
4192 * Returns 0 on success, -EINVAL on failure.
4194 * The caller must have charged to @to, IOW, called res_counter_charge() about
4195 * both res and memsw, and called css_get().
4197 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4198 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4200 unsigned short old_id
, new_id
;
4202 old_id
= css_id(&from
->css
);
4203 new_id
= css_id(&to
->css
);
4205 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4206 mem_cgroup_swap_statistics(from
, false);
4207 mem_cgroup_swap_statistics(to
, true);
4209 * This function is only called from task migration context now.
4210 * It postpones res_counter and refcount handling till the end
4211 * of task migration(mem_cgroup_clear_mc()) for performance
4212 * improvement. But we cannot postpone mem_cgroup_get(to)
4213 * because if the process that has been moved to @to does
4214 * swap-in, the refcount of @to might be decreased to 0.
4222 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4223 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4230 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4233 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4234 struct mem_cgroup
**memcgp
)
4236 struct mem_cgroup
*memcg
= NULL
;
4237 unsigned int nr_pages
= 1;
4238 struct page_cgroup
*pc
;
4239 enum charge_type ctype
;
4243 if (mem_cgroup_disabled())
4246 if (PageTransHuge(page
))
4247 nr_pages
<<= compound_order(page
);
4249 pc
= lookup_page_cgroup(page
);
4250 lock_page_cgroup(pc
);
4251 if (PageCgroupUsed(pc
)) {
4252 memcg
= pc
->mem_cgroup
;
4253 css_get(&memcg
->css
);
4255 * At migrating an anonymous page, its mapcount goes down
4256 * to 0 and uncharge() will be called. But, even if it's fully
4257 * unmapped, migration may fail and this page has to be
4258 * charged again. We set MIGRATION flag here and delay uncharge
4259 * until end_migration() is called
4261 * Corner Case Thinking
4263 * When the old page was mapped as Anon and it's unmap-and-freed
4264 * while migration was ongoing.
4265 * If unmap finds the old page, uncharge() of it will be delayed
4266 * until end_migration(). If unmap finds a new page, it's
4267 * uncharged when it make mapcount to be 1->0. If unmap code
4268 * finds swap_migration_entry, the new page will not be mapped
4269 * and end_migration() will find it(mapcount==0).
4272 * When the old page was mapped but migraion fails, the kernel
4273 * remaps it. A charge for it is kept by MIGRATION flag even
4274 * if mapcount goes down to 0. We can do remap successfully
4275 * without charging it again.
4278 * The "old" page is under lock_page() until the end of
4279 * migration, so, the old page itself will not be swapped-out.
4280 * If the new page is swapped out before end_migraton, our
4281 * hook to usual swap-out path will catch the event.
4284 SetPageCgroupMigration(pc
);
4286 unlock_page_cgroup(pc
);
4288 * If the page is not charged at this point,
4296 * We charge new page before it's used/mapped. So, even if unlock_page()
4297 * is called before end_migration, we can catch all events on this new
4298 * page. In the case new page is migrated but not remapped, new page's
4299 * mapcount will be finally 0 and we call uncharge in end_migration().
4302 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4304 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4306 * The page is committed to the memcg, but it's not actually
4307 * charged to the res_counter since we plan on replacing the
4308 * old one and only one page is going to be left afterwards.
4310 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4313 /* remove redundant charge if migration failed*/
4314 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4315 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4317 struct page
*used
, *unused
;
4318 struct page_cgroup
*pc
;
4324 if (!migration_ok
) {
4331 anon
= PageAnon(used
);
4332 __mem_cgroup_uncharge_common(unused
,
4333 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4334 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4336 css_put(&memcg
->css
);
4338 * We disallowed uncharge of pages under migration because mapcount
4339 * of the page goes down to zero, temporarly.
4340 * Clear the flag and check the page should be charged.
4342 pc
= lookup_page_cgroup(oldpage
);
4343 lock_page_cgroup(pc
);
4344 ClearPageCgroupMigration(pc
);
4345 unlock_page_cgroup(pc
);
4348 * If a page is a file cache, radix-tree replacement is very atomic
4349 * and we can skip this check. When it was an Anon page, its mapcount
4350 * goes down to 0. But because we added MIGRATION flage, it's not
4351 * uncharged yet. There are several case but page->mapcount check
4352 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4353 * check. (see prepare_charge() also)
4356 mem_cgroup_uncharge_page(used
);
4360 * At replace page cache, newpage is not under any memcg but it's on
4361 * LRU. So, this function doesn't touch res_counter but handles LRU
4362 * in correct way. Both pages are locked so we cannot race with uncharge.
4364 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4365 struct page
*newpage
)
4367 struct mem_cgroup
*memcg
= NULL
;
4368 struct page_cgroup
*pc
;
4369 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4371 if (mem_cgroup_disabled())
4374 pc
= lookup_page_cgroup(oldpage
);
4375 /* fix accounting on old pages */
4376 lock_page_cgroup(pc
);
4377 if (PageCgroupUsed(pc
)) {
4378 memcg
= pc
->mem_cgroup
;
4379 mem_cgroup_charge_statistics(memcg
, false, -1);
4380 ClearPageCgroupUsed(pc
);
4382 unlock_page_cgroup(pc
);
4385 * When called from shmem_replace_page(), in some cases the
4386 * oldpage has already been charged, and in some cases not.
4391 * Even if newpage->mapping was NULL before starting replacement,
4392 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4393 * LRU while we overwrite pc->mem_cgroup.
4395 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4398 #ifdef CONFIG_DEBUG_VM
4399 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4401 struct page_cgroup
*pc
;
4403 pc
= lookup_page_cgroup(page
);
4405 * Can be NULL while feeding pages into the page allocator for
4406 * the first time, i.e. during boot or memory hotplug;
4407 * or when mem_cgroup_disabled().
4409 if (likely(pc
) && PageCgroupUsed(pc
))
4414 bool mem_cgroup_bad_page_check(struct page
*page
)
4416 if (mem_cgroup_disabled())
4419 return lookup_page_cgroup_used(page
) != NULL
;
4422 void mem_cgroup_print_bad_page(struct page
*page
)
4424 struct page_cgroup
*pc
;
4426 pc
= lookup_page_cgroup_used(page
);
4428 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4429 pc
, pc
->flags
, pc
->mem_cgroup
);
4434 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4435 unsigned long long val
)
4438 u64 memswlimit
, memlimit
;
4440 int children
= mem_cgroup_count_children(memcg
);
4441 u64 curusage
, oldusage
;
4445 * For keeping hierarchical_reclaim simple, how long we should retry
4446 * is depends on callers. We set our retry-count to be function
4447 * of # of children which we should visit in this loop.
4449 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4451 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4454 while (retry_count
) {
4455 if (signal_pending(current
)) {
4460 * Rather than hide all in some function, I do this in
4461 * open coded manner. You see what this really does.
4462 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4464 mutex_lock(&set_limit_mutex
);
4465 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4466 if (memswlimit
< val
) {
4468 mutex_unlock(&set_limit_mutex
);
4472 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4476 ret
= res_counter_set_limit(&memcg
->res
, val
);
4478 if (memswlimit
== val
)
4479 memcg
->memsw_is_minimum
= true;
4481 memcg
->memsw_is_minimum
= false;
4483 mutex_unlock(&set_limit_mutex
);
4488 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4489 MEM_CGROUP_RECLAIM_SHRINK
);
4490 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4491 /* Usage is reduced ? */
4492 if (curusage
>= oldusage
)
4495 oldusage
= curusage
;
4497 if (!ret
&& enlarge
)
4498 memcg_oom_recover(memcg
);
4503 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4504 unsigned long long val
)
4507 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4508 int children
= mem_cgroup_count_children(memcg
);
4512 /* see mem_cgroup_resize_res_limit */
4513 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4514 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4515 while (retry_count
) {
4516 if (signal_pending(current
)) {
4521 * Rather than hide all in some function, I do this in
4522 * open coded manner. You see what this really does.
4523 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4525 mutex_lock(&set_limit_mutex
);
4526 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4527 if (memlimit
> val
) {
4529 mutex_unlock(&set_limit_mutex
);
4532 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4533 if (memswlimit
< val
)
4535 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4537 if (memlimit
== val
)
4538 memcg
->memsw_is_minimum
= true;
4540 memcg
->memsw_is_minimum
= false;
4542 mutex_unlock(&set_limit_mutex
);
4547 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4548 MEM_CGROUP_RECLAIM_NOSWAP
|
4549 MEM_CGROUP_RECLAIM_SHRINK
);
4550 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4551 /* Usage is reduced ? */
4552 if (curusage
>= oldusage
)
4555 oldusage
= curusage
;
4557 if (!ret
&& enlarge
)
4558 memcg_oom_recover(memcg
);
4562 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4564 unsigned long *total_scanned
)
4566 unsigned long nr_reclaimed
= 0;
4567 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4568 unsigned long reclaimed
;
4570 struct mem_cgroup_tree_per_zone
*mctz
;
4571 unsigned long long excess
;
4572 unsigned long nr_scanned
;
4577 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4579 * This loop can run a while, specially if mem_cgroup's continuously
4580 * keep exceeding their soft limit and putting the system under
4587 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4592 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4593 gfp_mask
, &nr_scanned
);
4594 nr_reclaimed
+= reclaimed
;
4595 *total_scanned
+= nr_scanned
;
4596 spin_lock(&mctz
->lock
);
4599 * If we failed to reclaim anything from this memory cgroup
4600 * it is time to move on to the next cgroup
4606 * Loop until we find yet another one.
4608 * By the time we get the soft_limit lock
4609 * again, someone might have aded the
4610 * group back on the RB tree. Iterate to
4611 * make sure we get a different mem.
4612 * mem_cgroup_largest_soft_limit_node returns
4613 * NULL if no other cgroup is present on
4617 __mem_cgroup_largest_soft_limit_node(mctz
);
4619 css_put(&next_mz
->memcg
->css
);
4620 else /* next_mz == NULL or other memcg */
4624 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4625 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4627 * One school of thought says that we should not add
4628 * back the node to the tree if reclaim returns 0.
4629 * But our reclaim could return 0, simply because due
4630 * to priority we are exposing a smaller subset of
4631 * memory to reclaim from. Consider this as a longer
4634 /* If excess == 0, no tree ops */
4635 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4636 spin_unlock(&mctz
->lock
);
4637 css_put(&mz
->memcg
->css
);
4640 * Could not reclaim anything and there are no more
4641 * mem cgroups to try or we seem to be looping without
4642 * reclaiming anything.
4644 if (!nr_reclaimed
&&
4646 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4648 } while (!nr_reclaimed
);
4650 css_put(&next_mz
->memcg
->css
);
4651 return nr_reclaimed
;
4655 * mem_cgroup_force_empty_list - clears LRU of a group
4656 * @memcg: group to clear
4659 * @lru: lru to to clear
4661 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4662 * reclaim the pages page themselves - pages are moved to the parent (or root)
4665 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4666 int node
, int zid
, enum lru_list lru
)
4668 struct lruvec
*lruvec
;
4669 unsigned long flags
;
4670 struct list_head
*list
;
4674 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4675 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4676 list
= &lruvec
->lists
[lru
];
4680 struct page_cgroup
*pc
;
4683 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4684 if (list_empty(list
)) {
4685 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4688 page
= list_entry(list
->prev
, struct page
, lru
);
4690 list_move(&page
->lru
, list
);
4692 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4695 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4697 pc
= lookup_page_cgroup(page
);
4699 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4700 /* found lock contention or "pc" is obsolete. */
4705 } while (!list_empty(list
));
4709 * make mem_cgroup's charge to be 0 if there is no task by moving
4710 * all the charges and pages to the parent.
4711 * This enables deleting this mem_cgroup.
4713 * Caller is responsible for holding css reference on the memcg.
4715 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4721 /* This is for making all *used* pages to be on LRU. */
4722 lru_add_drain_all();
4723 drain_all_stock_sync(memcg
);
4724 mem_cgroup_start_move(memcg
);
4725 for_each_node_state(node
, N_MEMORY
) {
4726 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4729 mem_cgroup_force_empty_list(memcg
,
4734 mem_cgroup_end_move(memcg
);
4735 memcg_oom_recover(memcg
);
4739 * Kernel memory may not necessarily be trackable to a specific
4740 * process. So they are not migrated, and therefore we can't
4741 * expect their value to drop to 0 here.
4742 * Having res filled up with kmem only is enough.
4744 * This is a safety check because mem_cgroup_force_empty_list
4745 * could have raced with mem_cgroup_replace_page_cache callers
4746 * so the lru seemed empty but the page could have been added
4747 * right after the check. RES_USAGE should be safe as we always
4748 * charge before adding to the LRU.
4750 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4751 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4752 } while (usage
> 0);
4756 * Reclaims as many pages from the given memcg as possible and moves
4757 * the rest to the parent.
4759 * Caller is responsible for holding css reference for memcg.
4761 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4763 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4764 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4766 /* returns EBUSY if there is a task or if we come here twice. */
4767 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4770 /* we call try-to-free pages for make this cgroup empty */
4771 lru_add_drain_all();
4772 /* try to free all pages in this cgroup */
4773 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4776 if (signal_pending(current
))
4779 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4783 /* maybe some writeback is necessary */
4784 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4789 mem_cgroup_reparent_charges(memcg
);
4794 static int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
4796 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4799 if (mem_cgroup_is_root(memcg
))
4801 css_get(&memcg
->css
);
4802 ret
= mem_cgroup_force_empty(memcg
);
4803 css_put(&memcg
->css
);
4809 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
4811 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
4814 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
4818 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4819 struct cgroup
*parent
= cont
->parent
;
4820 struct mem_cgroup
*parent_memcg
= NULL
;
4823 parent_memcg
= mem_cgroup_from_cont(parent
);
4827 if (memcg
->use_hierarchy
== val
)
4831 * If parent's use_hierarchy is set, we can't make any modifications
4832 * in the child subtrees. If it is unset, then the change can
4833 * occur, provided the current cgroup has no children.
4835 * For the root cgroup, parent_mem is NULL, we allow value to be
4836 * set if there are no children.
4838 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4839 (val
== 1 || val
== 0)) {
4840 if (list_empty(&cont
->children
))
4841 memcg
->use_hierarchy
= val
;
4854 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4855 enum mem_cgroup_stat_index idx
)
4857 struct mem_cgroup
*iter
;
4860 /* Per-cpu values can be negative, use a signed accumulator */
4861 for_each_mem_cgroup_tree(iter
, memcg
)
4862 val
+= mem_cgroup_read_stat(iter
, idx
);
4864 if (val
< 0) /* race ? */
4869 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4873 if (!mem_cgroup_is_root(memcg
)) {
4875 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4877 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4880 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4881 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4884 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4886 return val
<< PAGE_SHIFT
;
4889 static ssize_t
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
,
4890 struct file
*file
, char __user
*buf
,
4891 size_t nbytes
, loff_t
*ppos
)
4893 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4899 type
= MEMFILE_TYPE(cft
->private);
4900 name
= MEMFILE_ATTR(cft
->private);
4902 if (!do_swap_account
&& type
== _MEMSWAP
)
4907 if (name
== RES_USAGE
)
4908 val
= mem_cgroup_usage(memcg
, false);
4910 val
= res_counter_read_u64(&memcg
->res
, name
);
4913 if (name
== RES_USAGE
)
4914 val
= mem_cgroup_usage(memcg
, true);
4916 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4919 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4925 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4926 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4929 static int memcg_update_kmem_limit(struct cgroup
*cont
, u64 val
)
4932 #ifdef CONFIG_MEMCG_KMEM
4933 bool must_inc_static_branch
= false;
4935 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
4937 * For simplicity, we won't allow this to be disabled. It also can't
4938 * be changed if the cgroup has children already, or if tasks had
4941 * If tasks join before we set the limit, a person looking at
4942 * kmem.usage_in_bytes will have no way to determine when it took
4943 * place, which makes the value quite meaningless.
4945 * After it first became limited, changes in the value of the limit are
4946 * of course permitted.
4948 * Taking the cgroup_lock is really offensive, but it is so far the only
4949 * way to guarantee that no children will appear. There are plenty of
4950 * other offenders, and they should all go away. Fine grained locking
4951 * is probably the way to go here. When we are fully hierarchical, we
4952 * can also get rid of the use_hierarchy check.
4955 mutex_lock(&set_limit_mutex
);
4956 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4957 if (cgroup_task_count(cont
) || (memcg
->use_hierarchy
&&
4958 !list_empty(&cont
->children
))) {
4962 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4965 ret
= memcg_update_cache_sizes(memcg
);
4967 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4970 must_inc_static_branch
= true;
4972 * kmem charges can outlive the cgroup. In the case of slab
4973 * pages, for instance, a page contain objects from various
4974 * processes, so it is unfeasible to migrate them away. We
4975 * need to reference count the memcg because of that.
4977 mem_cgroup_get(memcg
);
4979 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4981 mutex_unlock(&set_limit_mutex
);
4985 * We are by now familiar with the fact that we can't inc the static
4986 * branch inside cgroup_lock. See disarm functions for details. A
4987 * worker here is overkill, but also wrong: After the limit is set, we
4988 * must start accounting right away. Since this operation can't fail,
4989 * we can safely defer it to here - no rollback will be needed.
4991 * The boolean used to control this is also safe, because
4992 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4993 * able to set it to true;
4995 if (must_inc_static_branch
) {
4996 static_key_slow_inc(&memcg_kmem_enabled_key
);
4998 * setting the active bit after the inc will guarantee no one
4999 * starts accounting before all call sites are patched
5001 memcg_kmem_set_active(memcg
);
5008 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5011 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5015 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5016 #ifdef CONFIG_MEMCG_KMEM
5018 * When that happen, we need to disable the static branch only on those
5019 * memcgs that enabled it. To achieve this, we would be forced to
5020 * complicate the code by keeping track of which memcgs were the ones
5021 * that actually enabled limits, and which ones got it from its
5024 * It is a lot simpler just to do static_key_slow_inc() on every child
5025 * that is accounted.
5027 if (!memcg_kmem_is_active(memcg
))
5031 * destroy(), called if we fail, will issue static_key_slow_inc() and
5032 * mem_cgroup_put() if kmem is enabled. We have to either call them
5033 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5034 * this more consistent, since it always leads to the same destroy path
5036 mem_cgroup_get(memcg
);
5037 static_key_slow_inc(&memcg_kmem_enabled_key
);
5039 mutex_lock(&set_limit_mutex
);
5040 ret
= memcg_update_cache_sizes(memcg
);
5041 mutex_unlock(&set_limit_mutex
);
5048 * The user of this function is...
5051 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
5054 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5057 unsigned long long val
;
5060 type
= MEMFILE_TYPE(cft
->private);
5061 name
= MEMFILE_ATTR(cft
->private);
5063 if (!do_swap_account
&& type
== _MEMSWAP
)
5068 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5072 /* This function does all necessary parse...reuse it */
5073 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5077 ret
= mem_cgroup_resize_limit(memcg
, val
);
5078 else if (type
== _MEMSWAP
)
5079 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5080 else if (type
== _KMEM
)
5081 ret
= memcg_update_kmem_limit(cont
, val
);
5085 case RES_SOFT_LIMIT
:
5086 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5090 * For memsw, soft limits are hard to implement in terms
5091 * of semantics, for now, we support soft limits for
5092 * control without swap
5095 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5100 ret
= -EINVAL
; /* should be BUG() ? */
5106 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5107 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5109 struct cgroup
*cgroup
;
5110 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5112 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5113 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5114 cgroup
= memcg
->css
.cgroup
;
5115 if (!memcg
->use_hierarchy
)
5118 while (cgroup
->parent
) {
5119 cgroup
= cgroup
->parent
;
5120 memcg
= mem_cgroup_from_cont(cgroup
);
5121 if (!memcg
->use_hierarchy
)
5123 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5124 min_limit
= min(min_limit
, tmp
);
5125 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5126 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5129 *mem_limit
= min_limit
;
5130 *memsw_limit
= min_memsw_limit
;
5133 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
5135 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5139 type
= MEMFILE_TYPE(event
);
5140 name
= MEMFILE_ATTR(event
);
5142 if (!do_swap_account
&& type
== _MEMSWAP
)
5148 res_counter_reset_max(&memcg
->res
);
5149 else if (type
== _MEMSWAP
)
5150 res_counter_reset_max(&memcg
->memsw
);
5151 else if (type
== _KMEM
)
5152 res_counter_reset_max(&memcg
->kmem
);
5158 res_counter_reset_failcnt(&memcg
->res
);
5159 else if (type
== _MEMSWAP
)
5160 res_counter_reset_failcnt(&memcg
->memsw
);
5161 else if (type
== _KMEM
)
5162 res_counter_reset_failcnt(&memcg
->kmem
);
5171 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
5174 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
5178 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5179 struct cftype
*cft
, u64 val
)
5181 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5183 if (val
>= (1 << NR_MOVE_TYPE
))
5186 * We check this value several times in both in can_attach() and
5187 * attach(), so we need cgroup lock to prevent this value from being
5191 memcg
->move_charge_at_immigrate
= val
;
5197 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
5198 struct cftype
*cft
, u64 val
)
5205 static int memcg_numa_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5209 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5210 unsigned long node_nr
;
5211 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5213 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5214 seq_printf(m
, "total=%lu", total_nr
);
5215 for_each_node_state(nid
, N_MEMORY
) {
5216 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5217 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5221 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5222 seq_printf(m
, "file=%lu", file_nr
);
5223 for_each_node_state(nid
, N_MEMORY
) {
5224 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5226 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5230 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5231 seq_printf(m
, "anon=%lu", anon_nr
);
5232 for_each_node_state(nid
, N_MEMORY
) {
5233 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5235 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5239 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5240 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5241 for_each_node_state(nid
, N_MEMORY
) {
5242 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5243 BIT(LRU_UNEVICTABLE
));
5244 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5249 #endif /* CONFIG_NUMA */
5251 static inline void mem_cgroup_lru_names_not_uptodate(void)
5253 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5256 static int memcg_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
5259 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
5260 struct mem_cgroup
*mi
;
5263 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5264 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5266 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5267 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5270 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5271 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5272 mem_cgroup_read_events(memcg
, i
));
5274 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5275 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5276 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5278 /* Hierarchical information */
5280 unsigned long long limit
, memsw_limit
;
5281 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5282 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5283 if (do_swap_account
)
5284 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5288 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5291 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5293 for_each_mem_cgroup_tree(mi
, memcg
)
5294 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5295 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5298 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5299 unsigned long long val
= 0;
5301 for_each_mem_cgroup_tree(mi
, memcg
)
5302 val
+= mem_cgroup_read_events(mi
, i
);
5303 seq_printf(m
, "total_%s %llu\n",
5304 mem_cgroup_events_names
[i
], val
);
5307 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5308 unsigned long long val
= 0;
5310 for_each_mem_cgroup_tree(mi
, memcg
)
5311 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5312 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5315 #ifdef CONFIG_DEBUG_VM
5318 struct mem_cgroup_per_zone
*mz
;
5319 struct zone_reclaim_stat
*rstat
;
5320 unsigned long recent_rotated
[2] = {0, 0};
5321 unsigned long recent_scanned
[2] = {0, 0};
5323 for_each_online_node(nid
)
5324 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5325 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5326 rstat
= &mz
->lruvec
.reclaim_stat
;
5328 recent_rotated
[0] += rstat
->recent_rotated
[0];
5329 recent_rotated
[1] += rstat
->recent_rotated
[1];
5330 recent_scanned
[0] += rstat
->recent_scanned
[0];
5331 recent_scanned
[1] += rstat
->recent_scanned
[1];
5333 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5334 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5335 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5336 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5343 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
5345 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5347 return mem_cgroup_swappiness(memcg
);
5350 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
5353 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5354 struct mem_cgroup
*parent
;
5359 if (cgrp
->parent
== NULL
)
5362 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5366 /* If under hierarchy, only empty-root can set this value */
5367 if ((parent
->use_hierarchy
) ||
5368 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
5373 memcg
->swappiness
= val
;
5380 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5382 struct mem_cgroup_threshold_ary
*t
;
5388 t
= rcu_dereference(memcg
->thresholds
.primary
);
5390 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5395 usage
= mem_cgroup_usage(memcg
, swap
);
5398 * current_threshold points to threshold just below or equal to usage.
5399 * If it's not true, a threshold was crossed after last
5400 * call of __mem_cgroup_threshold().
5402 i
= t
->current_threshold
;
5405 * Iterate backward over array of thresholds starting from
5406 * current_threshold and check if a threshold is crossed.
5407 * If none of thresholds below usage is crossed, we read
5408 * only one element of the array here.
5410 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5411 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5413 /* i = current_threshold + 1 */
5417 * Iterate forward over array of thresholds starting from
5418 * current_threshold+1 and check if a threshold is crossed.
5419 * If none of thresholds above usage is crossed, we read
5420 * only one element of the array here.
5422 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5423 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5425 /* Update current_threshold */
5426 t
->current_threshold
= i
- 1;
5431 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5434 __mem_cgroup_threshold(memcg
, false);
5435 if (do_swap_account
)
5436 __mem_cgroup_threshold(memcg
, true);
5438 memcg
= parent_mem_cgroup(memcg
);
5442 static int compare_thresholds(const void *a
, const void *b
)
5444 const struct mem_cgroup_threshold
*_a
= a
;
5445 const struct mem_cgroup_threshold
*_b
= b
;
5447 return _a
->threshold
- _b
->threshold
;
5450 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5452 struct mem_cgroup_eventfd_list
*ev
;
5454 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5455 eventfd_signal(ev
->eventfd
, 1);
5459 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5461 struct mem_cgroup
*iter
;
5463 for_each_mem_cgroup_tree(iter
, memcg
)
5464 mem_cgroup_oom_notify_cb(iter
);
5467 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
5468 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5470 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5471 struct mem_cgroup_thresholds
*thresholds
;
5472 struct mem_cgroup_threshold_ary
*new;
5473 enum res_type type
= MEMFILE_TYPE(cft
->private);
5474 u64 threshold
, usage
;
5477 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5481 mutex_lock(&memcg
->thresholds_lock
);
5484 thresholds
= &memcg
->thresholds
;
5485 else if (type
== _MEMSWAP
)
5486 thresholds
= &memcg
->memsw_thresholds
;
5490 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5492 /* Check if a threshold crossed before adding a new one */
5493 if (thresholds
->primary
)
5494 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5496 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5498 /* Allocate memory for new array of thresholds */
5499 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5507 /* Copy thresholds (if any) to new array */
5508 if (thresholds
->primary
) {
5509 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5510 sizeof(struct mem_cgroup_threshold
));
5513 /* Add new threshold */
5514 new->entries
[size
- 1].eventfd
= eventfd
;
5515 new->entries
[size
- 1].threshold
= threshold
;
5517 /* Sort thresholds. Registering of new threshold isn't time-critical */
5518 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5519 compare_thresholds
, NULL
);
5521 /* Find current threshold */
5522 new->current_threshold
= -1;
5523 for (i
= 0; i
< size
; i
++) {
5524 if (new->entries
[i
].threshold
<= usage
) {
5526 * new->current_threshold will not be used until
5527 * rcu_assign_pointer(), so it's safe to increment
5530 ++new->current_threshold
;
5535 /* Free old spare buffer and save old primary buffer as spare */
5536 kfree(thresholds
->spare
);
5537 thresholds
->spare
= thresholds
->primary
;
5539 rcu_assign_pointer(thresholds
->primary
, new);
5541 /* To be sure that nobody uses thresholds */
5545 mutex_unlock(&memcg
->thresholds_lock
);
5550 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
5551 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5553 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5554 struct mem_cgroup_thresholds
*thresholds
;
5555 struct mem_cgroup_threshold_ary
*new;
5556 enum res_type type
= MEMFILE_TYPE(cft
->private);
5560 mutex_lock(&memcg
->thresholds_lock
);
5562 thresholds
= &memcg
->thresholds
;
5563 else if (type
== _MEMSWAP
)
5564 thresholds
= &memcg
->memsw_thresholds
;
5568 if (!thresholds
->primary
)
5571 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5573 /* Check if a threshold crossed before removing */
5574 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5576 /* Calculate new number of threshold */
5578 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5579 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5583 new = thresholds
->spare
;
5585 /* Set thresholds array to NULL if we don't have thresholds */
5594 /* Copy thresholds and find current threshold */
5595 new->current_threshold
= -1;
5596 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5597 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5600 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5601 if (new->entries
[j
].threshold
<= usage
) {
5603 * new->current_threshold will not be used
5604 * until rcu_assign_pointer(), so it's safe to increment
5607 ++new->current_threshold
;
5613 /* Swap primary and spare array */
5614 thresholds
->spare
= thresholds
->primary
;
5615 /* If all events are unregistered, free the spare array */
5617 kfree(thresholds
->spare
);
5618 thresholds
->spare
= NULL
;
5621 rcu_assign_pointer(thresholds
->primary
, new);
5623 /* To be sure that nobody uses thresholds */
5626 mutex_unlock(&memcg
->thresholds_lock
);
5629 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
5630 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5632 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5633 struct mem_cgroup_eventfd_list
*event
;
5634 enum res_type type
= MEMFILE_TYPE(cft
->private);
5636 BUG_ON(type
!= _OOM_TYPE
);
5637 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5641 spin_lock(&memcg_oom_lock
);
5643 event
->eventfd
= eventfd
;
5644 list_add(&event
->list
, &memcg
->oom_notify
);
5646 /* already in OOM ? */
5647 if (atomic_read(&memcg
->under_oom
))
5648 eventfd_signal(eventfd
, 1);
5649 spin_unlock(&memcg_oom_lock
);
5654 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
5655 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5657 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5658 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5659 enum res_type type
= MEMFILE_TYPE(cft
->private);
5661 BUG_ON(type
!= _OOM_TYPE
);
5663 spin_lock(&memcg_oom_lock
);
5665 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5666 if (ev
->eventfd
== eventfd
) {
5667 list_del(&ev
->list
);
5672 spin_unlock(&memcg_oom_lock
);
5675 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
5676 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5678 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5680 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5682 if (atomic_read(&memcg
->under_oom
))
5683 cb
->fill(cb
, "under_oom", 1);
5685 cb
->fill(cb
, "under_oom", 0);
5689 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
5690 struct cftype
*cft
, u64 val
)
5692 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
5693 struct mem_cgroup
*parent
;
5695 /* cannot set to root cgroup and only 0 and 1 are allowed */
5696 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
5699 parent
= mem_cgroup_from_cont(cgrp
->parent
);
5702 /* oom-kill-disable is a flag for subhierarchy. */
5703 if ((parent
->use_hierarchy
) ||
5704 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
5708 memcg
->oom_kill_disable
= val
;
5710 memcg_oom_recover(memcg
);
5715 #ifdef CONFIG_MEMCG_KMEM
5716 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5720 memcg
->kmemcg_id
= -1;
5721 ret
= memcg_propagate_kmem(memcg
);
5725 return mem_cgroup_sockets_init(memcg
, ss
);
5728 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5730 mem_cgroup_sockets_destroy(memcg
);
5732 memcg_kmem_mark_dead(memcg
);
5734 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5738 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5739 * path here, being careful not to race with memcg_uncharge_kmem: it is
5740 * possible that the charges went down to 0 between mark_dead and the
5741 * res_counter read, so in that case, we don't need the put
5743 if (memcg_kmem_test_and_clear_dead(memcg
))
5744 mem_cgroup_put(memcg
);
5747 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5752 static void kmem_cgroup_destroy(struct mem_cgroup
*memcg
)
5757 static struct cftype mem_cgroup_files
[] = {
5759 .name
= "usage_in_bytes",
5760 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5761 .read
= mem_cgroup_read
,
5762 .register_event
= mem_cgroup_usage_register_event
,
5763 .unregister_event
= mem_cgroup_usage_unregister_event
,
5766 .name
= "max_usage_in_bytes",
5767 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5768 .trigger
= mem_cgroup_reset
,
5769 .read
= mem_cgroup_read
,
5772 .name
= "limit_in_bytes",
5773 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5774 .write_string
= mem_cgroup_write
,
5775 .read
= mem_cgroup_read
,
5778 .name
= "soft_limit_in_bytes",
5779 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5780 .write_string
= mem_cgroup_write
,
5781 .read
= mem_cgroup_read
,
5785 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5786 .trigger
= mem_cgroup_reset
,
5787 .read
= mem_cgroup_read
,
5791 .read_seq_string
= memcg_stat_show
,
5794 .name
= "force_empty",
5795 .trigger
= mem_cgroup_force_empty_write
,
5798 .name
= "use_hierarchy",
5799 .write_u64
= mem_cgroup_hierarchy_write
,
5800 .read_u64
= mem_cgroup_hierarchy_read
,
5803 .name
= "swappiness",
5804 .read_u64
= mem_cgroup_swappiness_read
,
5805 .write_u64
= mem_cgroup_swappiness_write
,
5808 .name
= "move_charge_at_immigrate",
5809 .read_u64
= mem_cgroup_move_charge_read
,
5810 .write_u64
= mem_cgroup_move_charge_write
,
5813 .name
= "oom_control",
5814 .read_map
= mem_cgroup_oom_control_read
,
5815 .write_u64
= mem_cgroup_oom_control_write
,
5816 .register_event
= mem_cgroup_oom_register_event
,
5817 .unregister_event
= mem_cgroup_oom_unregister_event
,
5818 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5822 .name
= "numa_stat",
5823 .read_seq_string
= memcg_numa_stat_show
,
5826 #ifdef CONFIG_MEMCG_SWAP
5828 .name
= "memsw.usage_in_bytes",
5829 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5830 .read
= mem_cgroup_read
,
5831 .register_event
= mem_cgroup_usage_register_event
,
5832 .unregister_event
= mem_cgroup_usage_unregister_event
,
5835 .name
= "memsw.max_usage_in_bytes",
5836 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5837 .trigger
= mem_cgroup_reset
,
5838 .read
= mem_cgroup_read
,
5841 .name
= "memsw.limit_in_bytes",
5842 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5843 .write_string
= mem_cgroup_write
,
5844 .read
= mem_cgroup_read
,
5847 .name
= "memsw.failcnt",
5848 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5849 .trigger
= mem_cgroup_reset
,
5850 .read
= mem_cgroup_read
,
5853 #ifdef CONFIG_MEMCG_KMEM
5855 .name
= "kmem.limit_in_bytes",
5856 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5857 .write_string
= mem_cgroup_write
,
5858 .read
= mem_cgroup_read
,
5861 .name
= "kmem.usage_in_bytes",
5862 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5863 .read
= mem_cgroup_read
,
5866 .name
= "kmem.failcnt",
5867 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5868 .trigger
= mem_cgroup_reset
,
5869 .read
= mem_cgroup_read
,
5872 .name
= "kmem.max_usage_in_bytes",
5873 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5874 .trigger
= mem_cgroup_reset
,
5875 .read
= mem_cgroup_read
,
5877 #ifdef CONFIG_SLABINFO
5879 .name
= "kmem.slabinfo",
5880 .read_seq_string
= mem_cgroup_slabinfo_read
,
5884 { }, /* terminate */
5887 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5889 struct mem_cgroup_per_node
*pn
;
5890 struct mem_cgroup_per_zone
*mz
;
5891 int zone
, tmp
= node
;
5893 * This routine is called against possible nodes.
5894 * But it's BUG to call kmalloc() against offline node.
5896 * TODO: this routine can waste much memory for nodes which will
5897 * never be onlined. It's better to use memory hotplug callback
5900 if (!node_state(node
, N_NORMAL_MEMORY
))
5902 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5906 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5907 mz
= &pn
->zoneinfo
[zone
];
5908 lruvec_init(&mz
->lruvec
);
5909 mz
->usage_in_excess
= 0;
5910 mz
->on_tree
= false;
5913 memcg
->info
.nodeinfo
[node
] = pn
;
5917 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5919 kfree(memcg
->info
.nodeinfo
[node
]);
5922 static struct mem_cgroup
*mem_cgroup_alloc(void)
5924 struct mem_cgroup
*memcg
;
5925 int size
= sizeof(struct mem_cgroup
);
5927 /* Can be very big if MAX_NUMNODES is very big */
5928 if (size
< PAGE_SIZE
)
5929 memcg
= kzalloc(size
, GFP_KERNEL
);
5931 memcg
= vzalloc(size
);
5936 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5939 spin_lock_init(&memcg
->pcp_counter_lock
);
5943 if (size
< PAGE_SIZE
)
5951 * At destroying mem_cgroup, references from swap_cgroup can remain.
5952 * (scanning all at force_empty is too costly...)
5954 * Instead of clearing all references at force_empty, we remember
5955 * the number of reference from swap_cgroup and free mem_cgroup when
5956 * it goes down to 0.
5958 * Removal of cgroup itself succeeds regardless of refs from swap.
5961 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5964 int size
= sizeof(struct mem_cgroup
);
5966 mem_cgroup_remove_from_trees(memcg
);
5967 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5970 free_mem_cgroup_per_zone_info(memcg
, node
);
5972 free_percpu(memcg
->stat
);
5975 * We need to make sure that (at least for now), the jump label
5976 * destruction code runs outside of the cgroup lock. This is because
5977 * get_online_cpus(), which is called from the static_branch update,
5978 * can't be called inside the cgroup_lock. cpusets are the ones
5979 * enforcing this dependency, so if they ever change, we might as well.
5981 * schedule_work() will guarantee this happens. Be careful if you need
5982 * to move this code around, and make sure it is outside
5985 disarm_static_keys(memcg
);
5986 if (size
< PAGE_SIZE
)
5994 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5995 * but in process context. The work_freeing structure is overlaid
5996 * on the rcu_freeing structure, which itself is overlaid on memsw.
5998 static void free_work(struct work_struct
*work
)
6000 struct mem_cgroup
*memcg
;
6002 memcg
= container_of(work
, struct mem_cgroup
, work_freeing
);
6003 __mem_cgroup_free(memcg
);
6006 static void free_rcu(struct rcu_head
*rcu_head
)
6008 struct mem_cgroup
*memcg
;
6010 memcg
= container_of(rcu_head
, struct mem_cgroup
, rcu_freeing
);
6011 INIT_WORK(&memcg
->work_freeing
, free_work
);
6012 schedule_work(&memcg
->work_freeing
);
6015 static void mem_cgroup_get(struct mem_cgroup
*memcg
)
6017 atomic_inc(&memcg
->refcnt
);
6020 static void __mem_cgroup_put(struct mem_cgroup
*memcg
, int count
)
6022 if (atomic_sub_and_test(count
, &memcg
->refcnt
)) {
6023 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
6024 call_rcu(&memcg
->rcu_freeing
, free_rcu
);
6026 mem_cgroup_put(parent
);
6030 static void mem_cgroup_put(struct mem_cgroup
*memcg
)
6032 __mem_cgroup_put(memcg
, 1);
6036 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6038 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6040 if (!memcg
->res
.parent
)
6042 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6044 EXPORT_SYMBOL(parent_mem_cgroup
);
6046 #ifdef CONFIG_MEMCG_SWAP
6047 static void __init
enable_swap_cgroup(void)
6049 if (!mem_cgroup_disabled() && really_do_swap_account
)
6050 do_swap_account
= 1;
6053 static void __init
enable_swap_cgroup(void)
6058 static int mem_cgroup_soft_limit_tree_init(void)
6060 struct mem_cgroup_tree_per_node
*rtpn
;
6061 struct mem_cgroup_tree_per_zone
*rtpz
;
6062 int tmp
, node
, zone
;
6064 for_each_node(node
) {
6066 if (!node_state(node
, N_NORMAL_MEMORY
))
6068 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6072 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6074 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6075 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6076 rtpz
->rb_root
= RB_ROOT
;
6077 spin_lock_init(&rtpz
->lock
);
6083 for_each_node(node
) {
6084 if (!soft_limit_tree
.rb_tree_per_node
[node
])
6086 kfree(soft_limit_tree
.rb_tree_per_node
[node
]);
6087 soft_limit_tree
.rb_tree_per_node
[node
] = NULL
;
6093 static struct cgroup_subsys_state
* __ref
6094 mem_cgroup_css_alloc(struct cgroup
*cont
)
6096 struct mem_cgroup
*memcg
, *parent
;
6097 long error
= -ENOMEM
;
6100 memcg
= mem_cgroup_alloc();
6102 return ERR_PTR(error
);
6105 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6109 if (cont
->parent
== NULL
) {
6111 enable_swap_cgroup();
6113 if (mem_cgroup_soft_limit_tree_init())
6115 root_mem_cgroup
= memcg
;
6116 for_each_possible_cpu(cpu
) {
6117 struct memcg_stock_pcp
*stock
=
6118 &per_cpu(memcg_stock
, cpu
);
6119 INIT_WORK(&stock
->work
, drain_local_stock
);
6122 parent
= mem_cgroup_from_cont(cont
->parent
);
6123 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6124 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6127 if (parent
&& parent
->use_hierarchy
) {
6128 res_counter_init(&memcg
->res
, &parent
->res
);
6129 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6130 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6133 * We increment refcnt of the parent to ensure that we can
6134 * safely access it on res_counter_charge/uncharge.
6135 * This refcnt will be decremented when freeing this
6136 * mem_cgroup(see mem_cgroup_put).
6138 mem_cgroup_get(parent
);
6140 res_counter_init(&memcg
->res
, NULL
);
6141 res_counter_init(&memcg
->memsw
, NULL
);
6142 res_counter_init(&memcg
->kmem
, NULL
);
6144 * Deeper hierachy with use_hierarchy == false doesn't make
6145 * much sense so let cgroup subsystem know about this
6146 * unfortunate state in our controller.
6148 if (parent
&& parent
!= root_mem_cgroup
)
6149 mem_cgroup_subsys
.broken_hierarchy
= true;
6151 memcg
->last_scanned_node
= MAX_NUMNODES
;
6152 INIT_LIST_HEAD(&memcg
->oom_notify
);
6155 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6156 atomic_set(&memcg
->refcnt
, 1);
6157 memcg
->move_charge_at_immigrate
= 0;
6158 mutex_init(&memcg
->thresholds_lock
);
6159 spin_lock_init(&memcg
->move_lock
);
6161 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6164 * We call put now because our (and parent's) refcnts
6165 * are already in place. mem_cgroup_put() will internally
6166 * call __mem_cgroup_free, so return directly
6168 mem_cgroup_put(memcg
);
6169 return ERR_PTR(error
);
6173 __mem_cgroup_free(memcg
);
6174 return ERR_PTR(error
);
6177 static void mem_cgroup_css_offline(struct cgroup
*cont
)
6179 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6181 mem_cgroup_reparent_charges(memcg
);
6182 mem_cgroup_destroy_all_caches(memcg
);
6185 static void mem_cgroup_css_free(struct cgroup
*cont
)
6187 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
6189 kmem_cgroup_destroy(memcg
);
6191 mem_cgroup_put(memcg
);
6195 /* Handlers for move charge at task migration. */
6196 #define PRECHARGE_COUNT_AT_ONCE 256
6197 static int mem_cgroup_do_precharge(unsigned long count
)
6200 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6201 struct mem_cgroup
*memcg
= mc
.to
;
6203 if (mem_cgroup_is_root(memcg
)) {
6204 mc
.precharge
+= count
;
6205 /* we don't need css_get for root */
6208 /* try to charge at once */
6210 struct res_counter
*dummy
;
6212 * "memcg" cannot be under rmdir() because we've already checked
6213 * by cgroup_lock_live_cgroup() that it is not removed and we
6214 * are still under the same cgroup_mutex. So we can postpone
6217 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6219 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6220 PAGE_SIZE
* count
, &dummy
)) {
6221 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6224 mc
.precharge
+= count
;
6228 /* fall back to one by one charge */
6230 if (signal_pending(current
)) {
6234 if (!batch_count
--) {
6235 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6238 ret
= __mem_cgroup_try_charge(NULL
,
6239 GFP_KERNEL
, 1, &memcg
, false);
6241 /* mem_cgroup_clear_mc() will do uncharge later */
6249 * get_mctgt_type - get target type of moving charge
6250 * @vma: the vma the pte to be checked belongs
6251 * @addr: the address corresponding to the pte to be checked
6252 * @ptent: the pte to be checked
6253 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6256 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6257 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6258 * move charge. if @target is not NULL, the page is stored in target->page
6259 * with extra refcnt got(Callers should handle it).
6260 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6261 * target for charge migration. if @target is not NULL, the entry is stored
6264 * Called with pte lock held.
6271 enum mc_target_type
{
6277 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6278 unsigned long addr
, pte_t ptent
)
6280 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6282 if (!page
|| !page_mapped(page
))
6284 if (PageAnon(page
)) {
6285 /* we don't move shared anon */
6288 } else if (!move_file())
6289 /* we ignore mapcount for file pages */
6291 if (!get_page_unless_zero(page
))
6298 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6299 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6301 struct page
*page
= NULL
;
6302 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6304 if (!move_anon() || non_swap_entry(ent
))
6307 * Because lookup_swap_cache() updates some statistics counter,
6308 * we call find_get_page() with swapper_space directly.
6310 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6311 if (do_swap_account
)
6312 entry
->val
= ent
.val
;
6317 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6318 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6324 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6325 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6327 struct page
*page
= NULL
;
6328 struct address_space
*mapping
;
6331 if (!vma
->vm_file
) /* anonymous vma */
6336 mapping
= vma
->vm_file
->f_mapping
;
6337 if (pte_none(ptent
))
6338 pgoff
= linear_page_index(vma
, addr
);
6339 else /* pte_file(ptent) is true */
6340 pgoff
= pte_to_pgoff(ptent
);
6342 /* page is moved even if it's not RSS of this task(page-faulted). */
6343 page
= find_get_page(mapping
, pgoff
);
6346 /* shmem/tmpfs may report page out on swap: account for that too. */
6347 if (radix_tree_exceptional_entry(page
)) {
6348 swp_entry_t swap
= radix_to_swp_entry(page
);
6349 if (do_swap_account
)
6351 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6357 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6358 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6360 struct page
*page
= NULL
;
6361 struct page_cgroup
*pc
;
6362 enum mc_target_type ret
= MC_TARGET_NONE
;
6363 swp_entry_t ent
= { .val
= 0 };
6365 if (pte_present(ptent
))
6366 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6367 else if (is_swap_pte(ptent
))
6368 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6369 else if (pte_none(ptent
) || pte_file(ptent
))
6370 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6372 if (!page
&& !ent
.val
)
6375 pc
= lookup_page_cgroup(page
);
6377 * Do only loose check w/o page_cgroup lock.
6378 * mem_cgroup_move_account() checks the pc is valid or not under
6381 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6382 ret
= MC_TARGET_PAGE
;
6384 target
->page
= page
;
6386 if (!ret
|| !target
)
6389 /* There is a swap entry and a page doesn't exist or isn't charged */
6390 if (ent
.val
&& !ret
&&
6391 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6392 ret
= MC_TARGET_SWAP
;
6399 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6401 * We don't consider swapping or file mapped pages because THP does not
6402 * support them for now.
6403 * Caller should make sure that pmd_trans_huge(pmd) is true.
6405 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6406 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6408 struct page
*page
= NULL
;
6409 struct page_cgroup
*pc
;
6410 enum mc_target_type ret
= MC_TARGET_NONE
;
6412 page
= pmd_page(pmd
);
6413 VM_BUG_ON(!page
|| !PageHead(page
));
6416 pc
= lookup_page_cgroup(page
);
6417 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6418 ret
= MC_TARGET_PAGE
;
6421 target
->page
= page
;
6427 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6428 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6430 return MC_TARGET_NONE
;
6434 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6435 unsigned long addr
, unsigned long end
,
6436 struct mm_walk
*walk
)
6438 struct vm_area_struct
*vma
= walk
->private;
6442 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6443 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6444 mc
.precharge
+= HPAGE_PMD_NR
;
6445 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6449 if (pmd_trans_unstable(pmd
))
6451 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6452 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6453 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6454 mc
.precharge
++; /* increment precharge temporarily */
6455 pte_unmap_unlock(pte
- 1, ptl
);
6461 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6463 unsigned long precharge
;
6464 struct vm_area_struct
*vma
;
6466 down_read(&mm
->mmap_sem
);
6467 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6468 struct mm_walk mem_cgroup_count_precharge_walk
= {
6469 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6473 if (is_vm_hugetlb_page(vma
))
6475 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6476 &mem_cgroup_count_precharge_walk
);
6478 up_read(&mm
->mmap_sem
);
6480 precharge
= mc
.precharge
;
6486 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6488 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6490 VM_BUG_ON(mc
.moving_task
);
6491 mc
.moving_task
= current
;
6492 return mem_cgroup_do_precharge(precharge
);
6495 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6496 static void __mem_cgroup_clear_mc(void)
6498 struct mem_cgroup
*from
= mc
.from
;
6499 struct mem_cgroup
*to
= mc
.to
;
6501 /* we must uncharge all the leftover precharges from mc.to */
6503 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6507 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6508 * we must uncharge here.
6510 if (mc
.moved_charge
) {
6511 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6512 mc
.moved_charge
= 0;
6514 /* we must fixup refcnts and charges */
6515 if (mc
.moved_swap
) {
6516 /* uncharge swap account from the old cgroup */
6517 if (!mem_cgroup_is_root(mc
.from
))
6518 res_counter_uncharge(&mc
.from
->memsw
,
6519 PAGE_SIZE
* mc
.moved_swap
);
6520 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
6522 if (!mem_cgroup_is_root(mc
.to
)) {
6524 * we charged both to->res and to->memsw, so we should
6527 res_counter_uncharge(&mc
.to
->res
,
6528 PAGE_SIZE
* mc
.moved_swap
);
6530 /* we've already done mem_cgroup_get(mc.to) */
6533 memcg_oom_recover(from
);
6534 memcg_oom_recover(to
);
6535 wake_up_all(&mc
.waitq
);
6538 static void mem_cgroup_clear_mc(void)
6540 struct mem_cgroup
*from
= mc
.from
;
6543 * we must clear moving_task before waking up waiters at the end of
6546 mc
.moving_task
= NULL
;
6547 __mem_cgroup_clear_mc();
6548 spin_lock(&mc
.lock
);
6551 spin_unlock(&mc
.lock
);
6552 mem_cgroup_end_move(from
);
6555 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6556 struct cgroup_taskset
*tset
)
6558 struct task_struct
*p
= cgroup_taskset_first(tset
);
6560 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgroup
);
6562 if (memcg
->move_charge_at_immigrate
) {
6563 struct mm_struct
*mm
;
6564 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6566 VM_BUG_ON(from
== memcg
);
6568 mm
= get_task_mm(p
);
6571 /* We move charges only when we move a owner of the mm */
6572 if (mm
->owner
== p
) {
6575 VM_BUG_ON(mc
.precharge
);
6576 VM_BUG_ON(mc
.moved_charge
);
6577 VM_BUG_ON(mc
.moved_swap
);
6578 mem_cgroup_start_move(from
);
6579 spin_lock(&mc
.lock
);
6582 spin_unlock(&mc
.lock
);
6583 /* We set mc.moving_task later */
6585 ret
= mem_cgroup_precharge_mc(mm
);
6587 mem_cgroup_clear_mc();
6594 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6595 struct cgroup_taskset
*tset
)
6597 mem_cgroup_clear_mc();
6600 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6601 unsigned long addr
, unsigned long end
,
6602 struct mm_walk
*walk
)
6605 struct vm_area_struct
*vma
= walk
->private;
6608 enum mc_target_type target_type
;
6609 union mc_target target
;
6611 struct page_cgroup
*pc
;
6614 * We don't take compound_lock() here but no race with splitting thp
6616 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6617 * under splitting, which means there's no concurrent thp split,
6618 * - if another thread runs into split_huge_page() just after we
6619 * entered this if-block, the thread must wait for page table lock
6620 * to be unlocked in __split_huge_page_splitting(), where the main
6621 * part of thp split is not executed yet.
6623 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6624 if (mc
.precharge
< HPAGE_PMD_NR
) {
6625 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6628 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6629 if (target_type
== MC_TARGET_PAGE
) {
6631 if (!isolate_lru_page(page
)) {
6632 pc
= lookup_page_cgroup(page
);
6633 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6634 pc
, mc
.from
, mc
.to
)) {
6635 mc
.precharge
-= HPAGE_PMD_NR
;
6636 mc
.moved_charge
+= HPAGE_PMD_NR
;
6638 putback_lru_page(page
);
6642 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6646 if (pmd_trans_unstable(pmd
))
6649 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6650 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6651 pte_t ptent
= *(pte
++);
6657 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6658 case MC_TARGET_PAGE
:
6660 if (isolate_lru_page(page
))
6662 pc
= lookup_page_cgroup(page
);
6663 if (!mem_cgroup_move_account(page
, 1, pc
,
6666 /* we uncharge from mc.from later. */
6669 putback_lru_page(page
);
6670 put
: /* get_mctgt_type() gets the page */
6673 case MC_TARGET_SWAP
:
6675 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6677 /* we fixup refcnts and charges later. */
6685 pte_unmap_unlock(pte
- 1, ptl
);
6690 * We have consumed all precharges we got in can_attach().
6691 * We try charge one by one, but don't do any additional
6692 * charges to mc.to if we have failed in charge once in attach()
6695 ret
= mem_cgroup_do_precharge(1);
6703 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6705 struct vm_area_struct
*vma
;
6707 lru_add_drain_all();
6709 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6711 * Someone who are holding the mmap_sem might be waiting in
6712 * waitq. So we cancel all extra charges, wake up all waiters,
6713 * and retry. Because we cancel precharges, we might not be able
6714 * to move enough charges, but moving charge is a best-effort
6715 * feature anyway, so it wouldn't be a big problem.
6717 __mem_cgroup_clear_mc();
6721 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6723 struct mm_walk mem_cgroup_move_charge_walk
= {
6724 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6728 if (is_vm_hugetlb_page(vma
))
6730 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6731 &mem_cgroup_move_charge_walk
);
6734 * means we have consumed all precharges and failed in
6735 * doing additional charge. Just abandon here.
6739 up_read(&mm
->mmap_sem
);
6742 static void mem_cgroup_move_task(struct cgroup
*cont
,
6743 struct cgroup_taskset
*tset
)
6745 struct task_struct
*p
= cgroup_taskset_first(tset
);
6746 struct mm_struct
*mm
= get_task_mm(p
);
6750 mem_cgroup_move_charge(mm
);
6754 mem_cgroup_clear_mc();
6756 #else /* !CONFIG_MMU */
6757 static int mem_cgroup_can_attach(struct cgroup
*cgroup
,
6758 struct cgroup_taskset
*tset
)
6762 static void mem_cgroup_cancel_attach(struct cgroup
*cgroup
,
6763 struct cgroup_taskset
*tset
)
6766 static void mem_cgroup_move_task(struct cgroup
*cont
,
6767 struct cgroup_taskset
*tset
)
6772 struct cgroup_subsys mem_cgroup_subsys
= {
6774 .subsys_id
= mem_cgroup_subsys_id
,
6775 .css_alloc
= mem_cgroup_css_alloc
,
6776 .css_offline
= mem_cgroup_css_offline
,
6777 .css_free
= mem_cgroup_css_free
,
6778 .can_attach
= mem_cgroup_can_attach
,
6779 .cancel_attach
= mem_cgroup_cancel_attach
,
6780 .attach
= mem_cgroup_move_task
,
6781 .base_cftypes
= mem_cgroup_files
,
6787 * The rest of init is performed during ->css_alloc() for root css which
6788 * happens before initcalls. hotcpu_notifier() can't be done together as
6789 * it would introduce circular locking by adding cgroup_lock -> cpu hotplug
6790 * dependency. Do it from a subsys_initcall().
6792 static int __init
mem_cgroup_init(void)
6794 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6797 subsys_initcall(mem_cgroup_init
);
6799 #ifdef CONFIG_MEMCG_SWAP
6800 static int __init
enable_swap_account(char *s
)
6802 /* consider enabled if no parameter or 1 is given */
6803 if (!strcmp(s
, "1"))
6804 really_do_swap_account
= 1;
6805 else if (!strcmp(s
, "0"))
6806 really_do_swap_account
= 0;
6809 __setup("swapaccount=", enable_swap_account
);