1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
67 EXPORT_SYMBOL(mem_cgroup_subsys
);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly
;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata
= 1;
80 static int really_do_swap_account __initdata
= 0;
84 #define do_swap_account 0
88 static const char * const mem_cgroup_stat_names
[] = {
97 enum mem_cgroup_events_index
{
98 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
99 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
100 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS
,
105 static const char * const mem_cgroup_events_names
[] = {
112 static const char * const mem_cgroup_lru_names
[] = {
121 * Per memcg event counter is incremented at every pagein/pageout. With THP,
122 * it will be incremated by the number of pages. This counter is used for
123 * for trigger some periodic events. This is straightforward and better
124 * than using jiffies etc. to handle periodic memcg event.
126 enum mem_cgroup_events_target
{
127 MEM_CGROUP_TARGET_THRESH
,
128 MEM_CGROUP_TARGET_SOFTLIMIT
,
129 MEM_CGROUP_TARGET_NUMAINFO
,
132 #define THRESHOLDS_EVENTS_TARGET 128
133 #define SOFTLIMIT_EVENTS_TARGET 1024
134 #define NUMAINFO_EVENTS_TARGET 1024
136 struct mem_cgroup_stat_cpu
{
137 long count
[MEM_CGROUP_STAT_NSTATS
];
138 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
139 unsigned long nr_page_events
;
140 unsigned long targets
[MEM_CGROUP_NTARGETS
];
143 struct mem_cgroup_reclaim_iter
{
145 * last scanned hierarchy member. Valid only if last_dead_count
146 * matches memcg->dead_count of the hierarchy root group.
148 struct mem_cgroup
*last_visited
;
149 unsigned long last_dead_count
;
151 /* scan generation, increased every round-trip */
152 unsigned int generation
;
156 * per-zone information in memory controller.
158 struct mem_cgroup_per_zone
{
159 struct lruvec lruvec
;
160 unsigned long lru_size
[NR_LRU_LISTS
];
162 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
164 struct rb_node tree_node
; /* RB tree node */
165 unsigned long long usage_in_excess
;/* Set to the value by which */
166 /* the soft limit is exceeded*/
168 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
169 /* use container_of */
172 struct mem_cgroup_per_node
{
173 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
177 * Cgroups above their limits are maintained in a RB-Tree, independent of
178 * their hierarchy representation
181 struct mem_cgroup_tree_per_zone
{
182 struct rb_root rb_root
;
186 struct mem_cgroup_tree_per_node
{
187 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
190 struct mem_cgroup_tree
{
191 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
194 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
196 struct mem_cgroup_threshold
{
197 struct eventfd_ctx
*eventfd
;
202 struct mem_cgroup_threshold_ary
{
203 /* An array index points to threshold just below or equal to usage. */
204 int current_threshold
;
205 /* Size of entries[] */
207 /* Array of thresholds */
208 struct mem_cgroup_threshold entries
[0];
211 struct mem_cgroup_thresholds
{
212 /* Primary thresholds array */
213 struct mem_cgroup_threshold_ary
*primary
;
215 * Spare threshold array.
216 * This is needed to make mem_cgroup_unregister_event() "never fail".
217 * It must be able to store at least primary->size - 1 entries.
219 struct mem_cgroup_threshold_ary
*spare
;
223 struct mem_cgroup_eventfd_list
{
224 struct list_head list
;
225 struct eventfd_ctx
*eventfd
;
228 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
229 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
232 * The memory controller data structure. The memory controller controls both
233 * page cache and RSS per cgroup. We would eventually like to provide
234 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
235 * to help the administrator determine what knobs to tune.
237 * TODO: Add a water mark for the memory controller. Reclaim will begin when
238 * we hit the water mark. May be even add a low water mark, such that
239 * no reclaim occurs from a cgroup at it's low water mark, this is
240 * a feature that will be implemented much later in the future.
243 struct cgroup_subsys_state css
;
245 * the counter to account for memory usage
247 struct res_counter res
;
249 /* vmpressure notifications */
250 struct vmpressure vmpressure
;
253 * the counter to account for mem+swap usage.
255 struct res_counter memsw
;
258 * the counter to account for kernel memory usage.
260 struct res_counter kmem
;
262 * Should the accounting and control be hierarchical, per subtree?
265 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
269 atomic_t oom_wakeups
;
272 /* OOM-Killer disable */
273 int oom_kill_disable
;
275 /* set when res.limit == memsw.limit */
276 bool memsw_is_minimum
;
278 /* protect arrays of thresholds */
279 struct mutex thresholds_lock
;
281 /* thresholds for memory usage. RCU-protected */
282 struct mem_cgroup_thresholds thresholds
;
284 /* thresholds for mem+swap usage. RCU-protected */
285 struct mem_cgroup_thresholds memsw_thresholds
;
287 /* For oom notifier event fd */
288 struct list_head oom_notify
;
291 * Should we move charges of a task when a task is moved into this
292 * mem_cgroup ? And what type of charges should we move ?
294 unsigned long move_charge_at_immigrate
;
296 * set > 0 if pages under this cgroup are moving to other cgroup.
298 atomic_t moving_account
;
299 /* taken only while moving_account > 0 */
300 spinlock_t move_lock
;
304 struct mem_cgroup_stat_cpu __percpu
*stat
;
306 * used when a cpu is offlined or other synchronizations
307 * See mem_cgroup_read_stat().
309 struct mem_cgroup_stat_cpu nocpu_base
;
310 spinlock_t pcp_counter_lock
;
313 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
314 struct tcp_memcontrol tcp_mem
;
316 #if defined(CONFIG_MEMCG_KMEM)
317 /* analogous to slab_common's slab_caches list. per-memcg */
318 struct list_head memcg_slab_caches
;
319 /* Not a spinlock, we can take a lot of time walking the list */
320 struct mutex slab_caches_mutex
;
321 /* Index in the kmem_cache->memcg_params->memcg_caches array */
325 int last_scanned_node
;
327 nodemask_t scan_nodes
;
328 atomic_t numainfo_events
;
329 atomic_t numainfo_updating
;
332 struct mem_cgroup_per_node
*nodeinfo
[0];
333 /* WARNING: nodeinfo must be the last member here */
336 static size_t memcg_size(void)
338 return sizeof(struct mem_cgroup
) +
339 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
342 /* internal only representation about the status of kmem accounting. */
344 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
345 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
346 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
349 /* We account when limit is on, but only after call sites are patched */
350 #define KMEM_ACCOUNTED_MASK \
351 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
353 #ifdef CONFIG_MEMCG_KMEM
354 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
356 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
359 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
361 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
364 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
366 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
369 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
371 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
374 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
377 * Our caller must use css_get() first, because memcg_uncharge_kmem()
378 * will call css_put() if it sees the memcg is dead.
381 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
382 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
385 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
387 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
388 &memcg
->kmem_account_flags
);
392 /* Stuffs for move charges at task migration. */
394 * Types of charges to be moved. "move_charge_at_immitgrate" and
395 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
398 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
399 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
403 /* "mc" and its members are protected by cgroup_mutex */
404 static struct move_charge_struct
{
405 spinlock_t lock
; /* for from, to */
406 struct mem_cgroup
*from
;
407 struct mem_cgroup
*to
;
408 unsigned long immigrate_flags
;
409 unsigned long precharge
;
410 unsigned long moved_charge
;
411 unsigned long moved_swap
;
412 struct task_struct
*moving_task
; /* a task moving charges */
413 wait_queue_head_t waitq
; /* a waitq for other context */
415 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
416 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
419 static bool move_anon(void)
421 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
424 static bool move_file(void)
426 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
430 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
431 * limit reclaim to prevent infinite loops, if they ever occur.
433 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
434 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
437 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
438 MEM_CGROUP_CHARGE_TYPE_ANON
,
439 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
440 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
444 /* for encoding cft->private value on file */
452 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
453 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
454 #define MEMFILE_ATTR(val) ((val) & 0xffff)
455 /* Used for OOM nofiier */
456 #define OOM_CONTROL (0)
459 * Reclaim flags for mem_cgroup_hierarchical_reclaim
461 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
462 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
463 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
464 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
467 * The memcg_create_mutex will be held whenever a new cgroup is created.
468 * As a consequence, any change that needs to protect against new child cgroups
469 * appearing has to hold it as well.
471 static DEFINE_MUTEX(memcg_create_mutex
);
473 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
475 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
478 /* Some nice accessors for the vmpressure. */
479 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
482 memcg
= root_mem_cgroup
;
483 return &memcg
->vmpressure
;
486 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
488 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
491 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
493 return &mem_cgroup_from_css(css
)->vmpressure
;
496 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
498 return (memcg
== root_mem_cgroup
);
501 /* Writing them here to avoid exposing memcg's inner layout */
502 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
504 void sock_update_memcg(struct sock
*sk
)
506 if (mem_cgroup_sockets_enabled
) {
507 struct mem_cgroup
*memcg
;
508 struct cg_proto
*cg_proto
;
510 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
512 /* Socket cloning can throw us here with sk_cgrp already
513 * filled. It won't however, necessarily happen from
514 * process context. So the test for root memcg given
515 * the current task's memcg won't help us in this case.
517 * Respecting the original socket's memcg is a better
518 * decision in this case.
521 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
522 css_get(&sk
->sk_cgrp
->memcg
->css
);
527 memcg
= mem_cgroup_from_task(current
);
528 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
529 if (!mem_cgroup_is_root(memcg
) &&
530 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
531 sk
->sk_cgrp
= cg_proto
;
536 EXPORT_SYMBOL(sock_update_memcg
);
538 void sock_release_memcg(struct sock
*sk
)
540 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
541 struct mem_cgroup
*memcg
;
542 WARN_ON(!sk
->sk_cgrp
->memcg
);
543 memcg
= sk
->sk_cgrp
->memcg
;
544 css_put(&sk
->sk_cgrp
->memcg
->css
);
548 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
550 if (!memcg
|| mem_cgroup_is_root(memcg
))
553 return &memcg
->tcp_mem
.cg_proto
;
555 EXPORT_SYMBOL(tcp_proto_cgroup
);
557 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
559 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
561 static_key_slow_dec(&memcg_socket_limit_enabled
);
564 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
569 #ifdef CONFIG_MEMCG_KMEM
571 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
572 * There are two main reasons for not using the css_id for this:
573 * 1) this works better in sparse environments, where we have a lot of memcgs,
574 * but only a few kmem-limited. Or also, if we have, for instance, 200
575 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
576 * 200 entry array for that.
578 * 2) In order not to violate the cgroup API, we would like to do all memory
579 * allocation in ->create(). At that point, we haven't yet allocated the
580 * css_id. Having a separate index prevents us from messing with the cgroup
583 * The current size of the caches array is stored in
584 * memcg_limited_groups_array_size. It will double each time we have to
587 static DEFINE_IDA(kmem_limited_groups
);
588 int memcg_limited_groups_array_size
;
591 * MIN_SIZE is different than 1, because we would like to avoid going through
592 * the alloc/free process all the time. In a small machine, 4 kmem-limited
593 * cgroups is a reasonable guess. In the future, it could be a parameter or
594 * tunable, but that is strictly not necessary.
596 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
597 * this constant directly from cgroup, but it is understandable that this is
598 * better kept as an internal representation in cgroup.c. In any case, the
599 * css_id space is not getting any smaller, and we don't have to necessarily
600 * increase ours as well if it increases.
602 #define MEMCG_CACHES_MIN_SIZE 4
603 #define MEMCG_CACHES_MAX_SIZE 65535
606 * A lot of the calls to the cache allocation functions are expected to be
607 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
608 * conditional to this static branch, we'll have to allow modules that does
609 * kmem_cache_alloc and the such to see this symbol as well
611 struct static_key memcg_kmem_enabled_key
;
612 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
614 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
616 if (memcg_kmem_is_active(memcg
)) {
617 static_key_slow_dec(&memcg_kmem_enabled_key
);
618 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
621 * This check can't live in kmem destruction function,
622 * since the charges will outlive the cgroup
624 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
627 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
630 #endif /* CONFIG_MEMCG_KMEM */
632 static void disarm_static_keys(struct mem_cgroup
*memcg
)
634 disarm_sock_keys(memcg
);
635 disarm_kmem_keys(memcg
);
638 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
640 static struct mem_cgroup_per_zone
*
641 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
643 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
644 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
647 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
652 static struct mem_cgroup_per_zone
*
653 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
655 int nid
= page_to_nid(page
);
656 int zid
= page_zonenum(page
);
658 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
661 static struct mem_cgroup_tree_per_zone
*
662 soft_limit_tree_node_zone(int nid
, int zid
)
664 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
667 static struct mem_cgroup_tree_per_zone
*
668 soft_limit_tree_from_page(struct page
*page
)
670 int nid
= page_to_nid(page
);
671 int zid
= page_zonenum(page
);
673 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
677 __mem_cgroup_insert_exceeded(struct mem_cgroup
*memcg
,
678 struct mem_cgroup_per_zone
*mz
,
679 struct mem_cgroup_tree_per_zone
*mctz
,
680 unsigned long long new_usage_in_excess
)
682 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
683 struct rb_node
*parent
= NULL
;
684 struct mem_cgroup_per_zone
*mz_node
;
689 mz
->usage_in_excess
= new_usage_in_excess
;
690 if (!mz
->usage_in_excess
)
694 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
696 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
699 * We can't avoid mem cgroups that are over their soft
700 * limit by the same amount
702 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
705 rb_link_node(&mz
->tree_node
, parent
, p
);
706 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
711 __mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
712 struct mem_cgroup_per_zone
*mz
,
713 struct mem_cgroup_tree_per_zone
*mctz
)
717 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
722 mem_cgroup_remove_exceeded(struct mem_cgroup
*memcg
,
723 struct mem_cgroup_per_zone
*mz
,
724 struct mem_cgroup_tree_per_zone
*mctz
)
726 spin_lock(&mctz
->lock
);
727 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
728 spin_unlock(&mctz
->lock
);
732 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
734 unsigned long long excess
;
735 struct mem_cgroup_per_zone
*mz
;
736 struct mem_cgroup_tree_per_zone
*mctz
;
737 int nid
= page_to_nid(page
);
738 int zid
= page_zonenum(page
);
739 mctz
= soft_limit_tree_from_page(page
);
742 * Necessary to update all ancestors when hierarchy is used.
743 * because their event counter is not touched.
745 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
746 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
747 excess
= res_counter_soft_limit_excess(&memcg
->res
);
749 * We have to update the tree if mz is on RB-tree or
750 * mem is over its softlimit.
752 if (excess
|| mz
->on_tree
) {
753 spin_lock(&mctz
->lock
);
754 /* if on-tree, remove it */
756 __mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
758 * Insert again. mz->usage_in_excess will be updated.
759 * If excess is 0, no tree ops.
761 __mem_cgroup_insert_exceeded(memcg
, mz
, mctz
, excess
);
762 spin_unlock(&mctz
->lock
);
767 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
770 struct mem_cgroup_per_zone
*mz
;
771 struct mem_cgroup_tree_per_zone
*mctz
;
773 for_each_node(node
) {
774 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
775 mz
= mem_cgroup_zoneinfo(memcg
, node
, zone
);
776 mctz
= soft_limit_tree_node_zone(node
, zone
);
777 mem_cgroup_remove_exceeded(memcg
, mz
, mctz
);
782 static struct mem_cgroup_per_zone
*
783 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
785 struct rb_node
*rightmost
= NULL
;
786 struct mem_cgroup_per_zone
*mz
;
790 rightmost
= rb_last(&mctz
->rb_root
);
792 goto done
; /* Nothing to reclaim from */
794 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
796 * Remove the node now but someone else can add it back,
797 * we will to add it back at the end of reclaim to its correct
798 * position in the tree.
800 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
801 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
802 !css_tryget(&mz
->memcg
->css
))
808 static struct mem_cgroup_per_zone
*
809 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
811 struct mem_cgroup_per_zone
*mz
;
813 spin_lock(&mctz
->lock
);
814 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
815 spin_unlock(&mctz
->lock
);
820 * Implementation Note: reading percpu statistics for memcg.
822 * Both of vmstat[] and percpu_counter has threshold and do periodic
823 * synchronization to implement "quick" read. There are trade-off between
824 * reading cost and precision of value. Then, we may have a chance to implement
825 * a periodic synchronizion of counter in memcg's counter.
827 * But this _read() function is used for user interface now. The user accounts
828 * memory usage by memory cgroup and he _always_ requires exact value because
829 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
830 * have to visit all online cpus and make sum. So, for now, unnecessary
831 * synchronization is not implemented. (just implemented for cpu hotplug)
833 * If there are kernel internal actions which can make use of some not-exact
834 * value, and reading all cpu value can be performance bottleneck in some
835 * common workload, threashold and synchonization as vmstat[] should be
838 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
839 enum mem_cgroup_stat_index idx
)
845 for_each_online_cpu(cpu
)
846 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
847 #ifdef CONFIG_HOTPLUG_CPU
848 spin_lock(&memcg
->pcp_counter_lock
);
849 val
+= memcg
->nocpu_base
.count
[idx
];
850 spin_unlock(&memcg
->pcp_counter_lock
);
856 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
859 int val
= (charge
) ? 1 : -1;
860 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
863 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
864 enum mem_cgroup_events_index idx
)
866 unsigned long val
= 0;
870 for_each_online_cpu(cpu
)
871 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
872 #ifdef CONFIG_HOTPLUG_CPU
873 spin_lock(&memcg
->pcp_counter_lock
);
874 val
+= memcg
->nocpu_base
.events
[idx
];
875 spin_unlock(&memcg
->pcp_counter_lock
);
881 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
883 bool anon
, int nr_pages
)
888 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
889 * counted as CACHE even if it's on ANON LRU.
892 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
895 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
898 if (PageTransHuge(page
))
899 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
902 /* pagein of a big page is an event. So, ignore page size */
904 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
906 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
907 nr_pages
= -nr_pages
; /* for event */
910 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
916 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
918 struct mem_cgroup_per_zone
*mz
;
920 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
921 return mz
->lru_size
[lru
];
925 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
926 unsigned int lru_mask
)
928 struct mem_cgroup_per_zone
*mz
;
930 unsigned long ret
= 0;
932 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
935 if (BIT(lru
) & lru_mask
)
936 ret
+= mz
->lru_size
[lru
];
942 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
943 int nid
, unsigned int lru_mask
)
948 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
949 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
955 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
956 unsigned int lru_mask
)
961 for_each_node_state(nid
, N_MEMORY
)
962 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
966 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
967 enum mem_cgroup_events_target target
)
969 unsigned long val
, next
;
971 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
972 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
973 /* from time_after() in jiffies.h */
974 if ((long)next
- (long)val
< 0) {
976 case MEM_CGROUP_TARGET_THRESH
:
977 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
979 case MEM_CGROUP_TARGET_SOFTLIMIT
:
980 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
982 case MEM_CGROUP_TARGET_NUMAINFO
:
983 next
= val
+ NUMAINFO_EVENTS_TARGET
;
988 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
995 * Check events in order.
998 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1001 /* threshold event is triggered in finer grain than soft limit */
1002 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1003 MEM_CGROUP_TARGET_THRESH
))) {
1005 bool do_numainfo __maybe_unused
;
1007 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1008 MEM_CGROUP_TARGET_SOFTLIMIT
);
1009 #if MAX_NUMNODES > 1
1010 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1011 MEM_CGROUP_TARGET_NUMAINFO
);
1015 mem_cgroup_threshold(memcg
);
1016 if (unlikely(do_softlimit
))
1017 mem_cgroup_update_tree(memcg
, page
);
1018 #if MAX_NUMNODES > 1
1019 if (unlikely(do_numainfo
))
1020 atomic_inc(&memcg
->numainfo_events
);
1026 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1029 * mm_update_next_owner() may clear mm->owner to NULL
1030 * if it races with swapoff, page migration, etc.
1031 * So this can be called with p == NULL.
1036 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
1039 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1041 struct mem_cgroup
*memcg
= NULL
;
1046 * Because we have no locks, mm->owner's may be being moved to other
1047 * cgroup. We use css_tryget() here even if this looks
1048 * pessimistic (rather than adding locks here).
1052 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1053 if (unlikely(!memcg
))
1055 } while (!css_tryget(&memcg
->css
));
1061 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1062 * ref. count) or NULL if the whole root's subtree has been visited.
1064 * helper function to be used by mem_cgroup_iter
1066 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1067 struct mem_cgroup
*last_visited
)
1069 struct cgroup_subsys_state
*prev_css
, *next_css
;
1071 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1073 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1076 * Even if we found a group we have to make sure it is
1077 * alive. css && !memcg means that the groups should be
1078 * skipped and we should continue the tree walk.
1079 * last_visited css is safe to use because it is
1080 * protected by css_get and the tree walk is rcu safe.
1083 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
1085 if (css_tryget(&mem
->css
))
1088 prev_css
= next_css
;
1096 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1099 * When a group in the hierarchy below root is destroyed, the
1100 * hierarchy iterator can no longer be trusted since it might
1101 * have pointed to the destroyed group. Invalidate it.
1103 atomic_inc(&root
->dead_count
);
1106 static struct mem_cgroup
*
1107 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1108 struct mem_cgroup
*root
,
1111 struct mem_cgroup
*position
= NULL
;
1113 * A cgroup destruction happens in two stages: offlining and
1114 * release. They are separated by a RCU grace period.
1116 * If the iterator is valid, we may still race with an
1117 * offlining. The RCU lock ensures the object won't be
1118 * released, tryget will fail if we lost the race.
1120 *sequence
= atomic_read(&root
->dead_count
);
1121 if (iter
->last_dead_count
== *sequence
) {
1123 position
= iter
->last_visited
;
1124 if (position
&& !css_tryget(&position
->css
))
1130 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1131 struct mem_cgroup
*last_visited
,
1132 struct mem_cgroup
*new_position
,
1136 css_put(&last_visited
->css
);
1138 * We store the sequence count from the time @last_visited was
1139 * loaded successfully instead of rereading it here so that we
1140 * don't lose destruction events in between. We could have
1141 * raced with the destruction of @new_position after all.
1143 iter
->last_visited
= new_position
;
1145 iter
->last_dead_count
= sequence
;
1149 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1150 * @root: hierarchy root
1151 * @prev: previously returned memcg, NULL on first invocation
1152 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1154 * Returns references to children of the hierarchy below @root, or
1155 * @root itself, or %NULL after a full round-trip.
1157 * Caller must pass the return value in @prev on subsequent
1158 * invocations for reference counting, or use mem_cgroup_iter_break()
1159 * to cancel a hierarchy walk before the round-trip is complete.
1161 * Reclaimers can specify a zone and a priority level in @reclaim to
1162 * divide up the memcgs in the hierarchy among all concurrent
1163 * reclaimers operating on the same zone and priority.
1165 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1166 struct mem_cgroup
*prev
,
1167 struct mem_cgroup_reclaim_cookie
*reclaim
)
1169 struct mem_cgroup
*memcg
= NULL
;
1170 struct mem_cgroup
*last_visited
= NULL
;
1172 if (mem_cgroup_disabled())
1176 root
= root_mem_cgroup
;
1178 if (prev
&& !reclaim
)
1179 last_visited
= prev
;
1181 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1189 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1190 int uninitialized_var(seq
);
1193 int nid
= zone_to_nid(reclaim
->zone
);
1194 int zid
= zone_idx(reclaim
->zone
);
1195 struct mem_cgroup_per_zone
*mz
;
1197 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1198 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1199 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1200 iter
->last_visited
= NULL
;
1204 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1207 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1210 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1214 else if (!prev
&& memcg
)
1215 reclaim
->generation
= iter
->generation
;
1224 if (prev
&& prev
!= root
)
1225 css_put(&prev
->css
);
1231 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1232 * @root: hierarchy root
1233 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1235 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1236 struct mem_cgroup
*prev
)
1239 root
= root_mem_cgroup
;
1240 if (prev
&& prev
!= root
)
1241 css_put(&prev
->css
);
1245 * Iteration constructs for visiting all cgroups (under a tree). If
1246 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1247 * be used for reference counting.
1249 #define for_each_mem_cgroup_tree(iter, root) \
1250 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1252 iter = mem_cgroup_iter(root, iter, NULL))
1254 #define for_each_mem_cgroup(iter) \
1255 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1257 iter = mem_cgroup_iter(NULL, iter, NULL))
1259 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1261 struct mem_cgroup
*memcg
;
1264 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1265 if (unlikely(!memcg
))
1270 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1273 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1281 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1284 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1285 * @zone: zone of the wanted lruvec
1286 * @memcg: memcg of the wanted lruvec
1288 * Returns the lru list vector holding pages for the given @zone and
1289 * @mem. This can be the global zone lruvec, if the memory controller
1292 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1293 struct mem_cgroup
*memcg
)
1295 struct mem_cgroup_per_zone
*mz
;
1296 struct lruvec
*lruvec
;
1298 if (mem_cgroup_disabled()) {
1299 lruvec
= &zone
->lruvec
;
1303 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1304 lruvec
= &mz
->lruvec
;
1307 * Since a node can be onlined after the mem_cgroup was created,
1308 * we have to be prepared to initialize lruvec->zone here;
1309 * and if offlined then reonlined, we need to reinitialize it.
1311 if (unlikely(lruvec
->zone
!= zone
))
1312 lruvec
->zone
= zone
;
1317 * Following LRU functions are allowed to be used without PCG_LOCK.
1318 * Operations are called by routine of global LRU independently from memcg.
1319 * What we have to take care of here is validness of pc->mem_cgroup.
1321 * Changes to pc->mem_cgroup happens when
1324 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1325 * It is added to LRU before charge.
1326 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1327 * When moving account, the page is not on LRU. It's isolated.
1331 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1333 * @zone: zone of the page
1335 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1337 struct mem_cgroup_per_zone
*mz
;
1338 struct mem_cgroup
*memcg
;
1339 struct page_cgroup
*pc
;
1340 struct lruvec
*lruvec
;
1342 if (mem_cgroup_disabled()) {
1343 lruvec
= &zone
->lruvec
;
1347 pc
= lookup_page_cgroup(page
);
1348 memcg
= pc
->mem_cgroup
;
1351 * Surreptitiously switch any uncharged offlist page to root:
1352 * an uncharged page off lru does nothing to secure
1353 * its former mem_cgroup from sudden removal.
1355 * Our caller holds lru_lock, and PageCgroupUsed is updated
1356 * under page_cgroup lock: between them, they make all uses
1357 * of pc->mem_cgroup safe.
1359 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1360 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1362 mz
= page_cgroup_zoneinfo(memcg
, page
);
1363 lruvec
= &mz
->lruvec
;
1366 * Since a node can be onlined after the mem_cgroup was created,
1367 * we have to be prepared to initialize lruvec->zone here;
1368 * and if offlined then reonlined, we need to reinitialize it.
1370 if (unlikely(lruvec
->zone
!= zone
))
1371 lruvec
->zone
= zone
;
1376 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1377 * @lruvec: mem_cgroup per zone lru vector
1378 * @lru: index of lru list the page is sitting on
1379 * @nr_pages: positive when adding or negative when removing
1381 * This function must be called when a page is added to or removed from an
1384 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1387 struct mem_cgroup_per_zone
*mz
;
1388 unsigned long *lru_size
;
1390 if (mem_cgroup_disabled())
1393 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1394 lru_size
= mz
->lru_size
+ lru
;
1395 *lru_size
+= nr_pages
;
1396 VM_BUG_ON((long)(*lru_size
) < 0);
1400 * Checks whether given mem is same or in the root_mem_cgroup's
1403 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1404 struct mem_cgroup
*memcg
)
1406 if (root_memcg
== memcg
)
1408 if (!root_memcg
->use_hierarchy
|| !memcg
)
1410 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1413 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1414 struct mem_cgroup
*memcg
)
1419 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1424 bool task_in_mem_cgroup(struct task_struct
*task
,
1425 const struct mem_cgroup
*memcg
)
1427 struct mem_cgroup
*curr
= NULL
;
1428 struct task_struct
*p
;
1431 p
= find_lock_task_mm(task
);
1433 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1437 * All threads may have already detached their mm's, but the oom
1438 * killer still needs to detect if they have already been oom
1439 * killed to prevent needlessly killing additional tasks.
1442 curr
= mem_cgroup_from_task(task
);
1444 css_get(&curr
->css
);
1450 * We should check use_hierarchy of "memcg" not "curr". Because checking
1451 * use_hierarchy of "curr" here make this function true if hierarchy is
1452 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1453 * hierarchy(even if use_hierarchy is disabled in "memcg").
1455 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1456 css_put(&curr
->css
);
1460 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1462 unsigned long inactive_ratio
;
1463 unsigned long inactive
;
1464 unsigned long active
;
1467 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1468 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1470 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1472 inactive_ratio
= int_sqrt(10 * gb
);
1476 return inactive
* inactive_ratio
< active
;
1479 #define mem_cgroup_from_res_counter(counter, member) \
1480 container_of(counter, struct mem_cgroup, member)
1483 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1484 * @memcg: the memory cgroup
1486 * Returns the maximum amount of memory @mem can be charged with, in
1489 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1491 unsigned long long margin
;
1493 margin
= res_counter_margin(&memcg
->res
);
1494 if (do_swap_account
)
1495 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1496 return margin
>> PAGE_SHIFT
;
1499 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1502 if (!css_parent(&memcg
->css
))
1503 return vm_swappiness
;
1505 return memcg
->swappiness
;
1509 * memcg->moving_account is used for checking possibility that some thread is
1510 * calling move_account(). When a thread on CPU-A starts moving pages under
1511 * a memcg, other threads should check memcg->moving_account under
1512 * rcu_read_lock(), like this:
1516 * memcg->moving_account+1 if (memcg->mocing_account)
1518 * synchronize_rcu() update something.
1523 /* for quick checking without looking up memcg */
1524 atomic_t memcg_moving __read_mostly
;
1526 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1528 atomic_inc(&memcg_moving
);
1529 atomic_inc(&memcg
->moving_account
);
1533 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1536 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1537 * We check NULL in callee rather than caller.
1540 atomic_dec(&memcg_moving
);
1541 atomic_dec(&memcg
->moving_account
);
1546 * 2 routines for checking "mem" is under move_account() or not.
1548 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1549 * is used for avoiding races in accounting. If true,
1550 * pc->mem_cgroup may be overwritten.
1552 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1553 * under hierarchy of moving cgroups. This is for
1554 * waiting at hith-memory prressure caused by "move".
1557 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1559 VM_BUG_ON(!rcu_read_lock_held());
1560 return atomic_read(&memcg
->moving_account
) > 0;
1563 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1565 struct mem_cgroup
*from
;
1566 struct mem_cgroup
*to
;
1569 * Unlike task_move routines, we access mc.to, mc.from not under
1570 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1572 spin_lock(&mc
.lock
);
1578 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1579 || mem_cgroup_same_or_subtree(memcg
, to
);
1581 spin_unlock(&mc
.lock
);
1585 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1587 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1588 if (mem_cgroup_under_move(memcg
)) {
1590 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1591 /* moving charge context might have finished. */
1594 finish_wait(&mc
.waitq
, &wait
);
1602 * Take this lock when
1603 * - a code tries to modify page's memcg while it's USED.
1604 * - a code tries to modify page state accounting in a memcg.
1605 * see mem_cgroup_stolen(), too.
1607 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1608 unsigned long *flags
)
1610 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1613 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1614 unsigned long *flags
)
1616 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1619 #define K(x) ((x) << (PAGE_SHIFT-10))
1621 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1622 * @memcg: The memory cgroup that went over limit
1623 * @p: Task that is going to be killed
1625 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1628 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1630 struct cgroup
*task_cgrp
;
1631 struct cgroup
*mem_cgrp
;
1633 * Need a buffer in BSS, can't rely on allocations. The code relies
1634 * on the assumption that OOM is serialized for memory controller.
1635 * If this assumption is broken, revisit this code.
1637 static char memcg_name
[PATH_MAX
];
1639 struct mem_cgroup
*iter
;
1647 mem_cgrp
= memcg
->css
.cgroup
;
1648 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1650 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1653 * Unfortunately, we are unable to convert to a useful name
1654 * But we'll still print out the usage information
1661 pr_info("Task in %s killed", memcg_name
);
1664 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1672 * Continues from above, so we don't need an KERN_ level
1674 pr_cont(" as a result of limit of %s\n", memcg_name
);
1677 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1678 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1679 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1680 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1681 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1682 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1683 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1684 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1685 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1686 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1687 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1688 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1690 for_each_mem_cgroup_tree(iter
, memcg
) {
1691 pr_info("Memory cgroup stats");
1694 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1696 pr_cont(" for %s", memcg_name
);
1700 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1701 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1703 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1704 K(mem_cgroup_read_stat(iter
, i
)));
1707 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1708 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1709 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1716 * This function returns the number of memcg under hierarchy tree. Returns
1717 * 1(self count) if no children.
1719 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1722 struct mem_cgroup
*iter
;
1724 for_each_mem_cgroup_tree(iter
, memcg
)
1730 * Return the memory (and swap, if configured) limit for a memcg.
1732 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1736 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1739 * Do not consider swap space if we cannot swap due to swappiness
1741 if (mem_cgroup_swappiness(memcg
)) {
1744 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1745 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1748 * If memsw is finite and limits the amount of swap space
1749 * available to this memcg, return that limit.
1751 limit
= min(limit
, memsw
);
1757 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1760 struct mem_cgroup
*iter
;
1761 unsigned long chosen_points
= 0;
1762 unsigned long totalpages
;
1763 unsigned int points
= 0;
1764 struct task_struct
*chosen
= NULL
;
1767 * If current has a pending SIGKILL or is exiting, then automatically
1768 * select it. The goal is to allow it to allocate so that it may
1769 * quickly exit and free its memory.
1771 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1772 set_thread_flag(TIF_MEMDIE
);
1776 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1777 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1778 for_each_mem_cgroup_tree(iter
, memcg
) {
1779 struct css_task_iter it
;
1780 struct task_struct
*task
;
1782 css_task_iter_start(&iter
->css
, &it
);
1783 while ((task
= css_task_iter_next(&it
))) {
1784 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1786 case OOM_SCAN_SELECT
:
1788 put_task_struct(chosen
);
1790 chosen_points
= ULONG_MAX
;
1791 get_task_struct(chosen
);
1793 case OOM_SCAN_CONTINUE
:
1795 case OOM_SCAN_ABORT
:
1796 css_task_iter_end(&it
);
1797 mem_cgroup_iter_break(memcg
, iter
);
1799 put_task_struct(chosen
);
1804 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1805 if (points
> chosen_points
) {
1807 put_task_struct(chosen
);
1809 chosen_points
= points
;
1810 get_task_struct(chosen
);
1813 css_task_iter_end(&it
);
1818 points
= chosen_points
* 1000 / totalpages
;
1819 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1820 NULL
, "Memory cgroup out of memory");
1823 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1825 unsigned long flags
)
1827 unsigned long total
= 0;
1828 bool noswap
= false;
1831 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1833 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1836 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1838 drain_all_stock_async(memcg
);
1839 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1841 * Allow limit shrinkers, which are triggered directly
1842 * by userspace, to catch signals and stop reclaim
1843 * after minimal progress, regardless of the margin.
1845 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1847 if (mem_cgroup_margin(memcg
))
1850 * If nothing was reclaimed after two attempts, there
1851 * may be no reclaimable pages in this hierarchy.
1860 * test_mem_cgroup_node_reclaimable
1861 * @memcg: the target memcg
1862 * @nid: the node ID to be checked.
1863 * @noswap : specify true here if the user wants flle only information.
1865 * This function returns whether the specified memcg contains any
1866 * reclaimable pages on a node. Returns true if there are any reclaimable
1867 * pages in the node.
1869 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1870 int nid
, bool noswap
)
1872 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1874 if (noswap
|| !total_swap_pages
)
1876 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1881 #if MAX_NUMNODES > 1
1884 * Always updating the nodemask is not very good - even if we have an empty
1885 * list or the wrong list here, we can start from some node and traverse all
1886 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1889 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1893 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1894 * pagein/pageout changes since the last update.
1896 if (!atomic_read(&memcg
->numainfo_events
))
1898 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1901 /* make a nodemask where this memcg uses memory from */
1902 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1904 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1906 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1907 node_clear(nid
, memcg
->scan_nodes
);
1910 atomic_set(&memcg
->numainfo_events
, 0);
1911 atomic_set(&memcg
->numainfo_updating
, 0);
1915 * Selecting a node where we start reclaim from. Because what we need is just
1916 * reducing usage counter, start from anywhere is O,K. Considering
1917 * memory reclaim from current node, there are pros. and cons.
1919 * Freeing memory from current node means freeing memory from a node which
1920 * we'll use or we've used. So, it may make LRU bad. And if several threads
1921 * hit limits, it will see a contention on a node. But freeing from remote
1922 * node means more costs for memory reclaim because of memory latency.
1924 * Now, we use round-robin. Better algorithm is welcomed.
1926 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1930 mem_cgroup_may_update_nodemask(memcg
);
1931 node
= memcg
->last_scanned_node
;
1933 node
= next_node(node
, memcg
->scan_nodes
);
1934 if (node
== MAX_NUMNODES
)
1935 node
= first_node(memcg
->scan_nodes
);
1937 * We call this when we hit limit, not when pages are added to LRU.
1938 * No LRU may hold pages because all pages are UNEVICTABLE or
1939 * memcg is too small and all pages are not on LRU. In that case,
1940 * we use curret node.
1942 if (unlikely(node
== MAX_NUMNODES
))
1943 node
= numa_node_id();
1945 memcg
->last_scanned_node
= node
;
1950 * Check all nodes whether it contains reclaimable pages or not.
1951 * For quick scan, we make use of scan_nodes. This will allow us to skip
1952 * unused nodes. But scan_nodes is lazily updated and may not cotain
1953 * enough new information. We need to do double check.
1955 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1960 * quick check...making use of scan_node.
1961 * We can skip unused nodes.
1963 if (!nodes_empty(memcg
->scan_nodes
)) {
1964 for (nid
= first_node(memcg
->scan_nodes
);
1966 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1968 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1973 * Check rest of nodes.
1975 for_each_node_state(nid
, N_MEMORY
) {
1976 if (node_isset(nid
, memcg
->scan_nodes
))
1978 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1985 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1990 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1992 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1996 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1999 unsigned long *total_scanned
)
2001 struct mem_cgroup
*victim
= NULL
;
2004 unsigned long excess
;
2005 unsigned long nr_scanned
;
2006 struct mem_cgroup_reclaim_cookie reclaim
= {
2011 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2014 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2019 * If we have not been able to reclaim
2020 * anything, it might because there are
2021 * no reclaimable pages under this hierarchy
2026 * We want to do more targeted reclaim.
2027 * excess >> 2 is not to excessive so as to
2028 * reclaim too much, nor too less that we keep
2029 * coming back to reclaim from this cgroup
2031 if (total
>= (excess
>> 2) ||
2032 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2037 if (!mem_cgroup_reclaimable(victim
, false))
2039 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2041 *total_scanned
+= nr_scanned
;
2042 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2045 mem_cgroup_iter_break(root_memcg
, victim
);
2049 static DEFINE_SPINLOCK(memcg_oom_lock
);
2052 * Check OOM-Killer is already running under our hierarchy.
2053 * If someone is running, return false.
2055 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2057 struct mem_cgroup
*iter
, *failed
= NULL
;
2059 spin_lock(&memcg_oom_lock
);
2061 for_each_mem_cgroup_tree(iter
, memcg
) {
2062 if (iter
->oom_lock
) {
2064 * this subtree of our hierarchy is already locked
2065 * so we cannot give a lock.
2068 mem_cgroup_iter_break(memcg
, iter
);
2071 iter
->oom_lock
= true;
2076 * OK, we failed to lock the whole subtree so we have
2077 * to clean up what we set up to the failing subtree
2079 for_each_mem_cgroup_tree(iter
, memcg
) {
2080 if (iter
== failed
) {
2081 mem_cgroup_iter_break(memcg
, iter
);
2084 iter
->oom_lock
= false;
2088 spin_unlock(&memcg_oom_lock
);
2093 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2095 struct mem_cgroup
*iter
;
2097 spin_lock(&memcg_oom_lock
);
2098 for_each_mem_cgroup_tree(iter
, memcg
)
2099 iter
->oom_lock
= false;
2100 spin_unlock(&memcg_oom_lock
);
2103 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2105 struct mem_cgroup
*iter
;
2107 for_each_mem_cgroup_tree(iter
, memcg
)
2108 atomic_inc(&iter
->under_oom
);
2111 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2113 struct mem_cgroup
*iter
;
2116 * When a new child is created while the hierarchy is under oom,
2117 * mem_cgroup_oom_lock() may not be called. We have to use
2118 * atomic_add_unless() here.
2120 for_each_mem_cgroup_tree(iter
, memcg
)
2121 atomic_add_unless(&iter
->under_oom
, -1, 0);
2124 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2126 struct oom_wait_info
{
2127 struct mem_cgroup
*memcg
;
2131 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2132 unsigned mode
, int sync
, void *arg
)
2134 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2135 struct mem_cgroup
*oom_wait_memcg
;
2136 struct oom_wait_info
*oom_wait_info
;
2138 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2139 oom_wait_memcg
= oom_wait_info
->memcg
;
2142 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2143 * Then we can use css_is_ancestor without taking care of RCU.
2145 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2146 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2148 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2151 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2153 atomic_inc(&memcg
->oom_wakeups
);
2154 /* for filtering, pass "memcg" as argument. */
2155 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2158 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2160 if (memcg
&& atomic_read(&memcg
->under_oom
))
2161 memcg_wakeup_oom(memcg
);
2164 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2166 if (!current
->memcg_oom
.may_oom
)
2169 * We are in the middle of the charge context here, so we
2170 * don't want to block when potentially sitting on a callstack
2171 * that holds all kinds of filesystem and mm locks.
2173 * Also, the caller may handle a failed allocation gracefully
2174 * (like optional page cache readahead) and so an OOM killer
2175 * invocation might not even be necessary.
2177 * That's why we don't do anything here except remember the
2178 * OOM context and then deal with it at the end of the page
2179 * fault when the stack is unwound, the locks are released,
2180 * and when we know whether the fault was overall successful.
2182 css_get(&memcg
->css
);
2183 current
->memcg_oom
.memcg
= memcg
;
2184 current
->memcg_oom
.gfp_mask
= mask
;
2185 current
->memcg_oom
.order
= order
;
2189 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2190 * @handle: actually kill/wait or just clean up the OOM state
2192 * This has to be called at the end of a page fault if the memcg OOM
2193 * handler was enabled.
2195 * Memcg supports userspace OOM handling where failed allocations must
2196 * sleep on a waitqueue until the userspace task resolves the
2197 * situation. Sleeping directly in the charge context with all kinds
2198 * of locks held is not a good idea, instead we remember an OOM state
2199 * in the task and mem_cgroup_oom_synchronize() has to be called at
2200 * the end of the page fault to complete the OOM handling.
2202 * Returns %true if an ongoing memcg OOM situation was detected and
2203 * completed, %false otherwise.
2205 bool mem_cgroup_oom_synchronize(bool handle
)
2207 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2208 struct oom_wait_info owait
;
2211 /* OOM is global, do not handle */
2218 owait
.memcg
= memcg
;
2219 owait
.wait
.flags
= 0;
2220 owait
.wait
.func
= memcg_oom_wake_function
;
2221 owait
.wait
.private = current
;
2222 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2224 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2225 mem_cgroup_mark_under_oom(memcg
);
2227 locked
= mem_cgroup_oom_trylock(memcg
);
2230 mem_cgroup_oom_notify(memcg
);
2232 if (locked
&& !memcg
->oom_kill_disable
) {
2233 mem_cgroup_unmark_under_oom(memcg
);
2234 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2235 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2236 current
->memcg_oom
.order
);
2239 mem_cgroup_unmark_under_oom(memcg
);
2240 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2244 mem_cgroup_oom_unlock(memcg
);
2246 * There is no guarantee that an OOM-lock contender
2247 * sees the wakeups triggered by the OOM kill
2248 * uncharges. Wake any sleepers explicitely.
2250 memcg_oom_recover(memcg
);
2253 current
->memcg_oom
.memcg
= NULL
;
2254 css_put(&memcg
->css
);
2259 * Currently used to update mapped file statistics, but the routine can be
2260 * generalized to update other statistics as well.
2262 * Notes: Race condition
2264 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2265 * it tends to be costly. But considering some conditions, we doesn't need
2266 * to do so _always_.
2268 * Considering "charge", lock_page_cgroup() is not required because all
2269 * file-stat operations happen after a page is attached to radix-tree. There
2270 * are no race with "charge".
2272 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2273 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2274 * if there are race with "uncharge". Statistics itself is properly handled
2277 * Considering "move", this is an only case we see a race. To make the race
2278 * small, we check mm->moving_account and detect there are possibility of race
2279 * If there is, we take a lock.
2282 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2283 bool *locked
, unsigned long *flags
)
2285 struct mem_cgroup
*memcg
;
2286 struct page_cgroup
*pc
;
2288 pc
= lookup_page_cgroup(page
);
2290 memcg
= pc
->mem_cgroup
;
2291 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2294 * If this memory cgroup is not under account moving, we don't
2295 * need to take move_lock_mem_cgroup(). Because we already hold
2296 * rcu_read_lock(), any calls to move_account will be delayed until
2297 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2299 if (!mem_cgroup_stolen(memcg
))
2302 move_lock_mem_cgroup(memcg
, flags
);
2303 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2304 move_unlock_mem_cgroup(memcg
, flags
);
2310 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2312 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2315 * It's guaranteed that pc->mem_cgroup never changes while
2316 * lock is held because a routine modifies pc->mem_cgroup
2317 * should take move_lock_mem_cgroup().
2319 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2322 void mem_cgroup_update_page_stat(struct page
*page
,
2323 enum mem_cgroup_stat_index idx
, int val
)
2325 struct mem_cgroup
*memcg
;
2326 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2327 unsigned long uninitialized_var(flags
);
2329 if (mem_cgroup_disabled())
2332 VM_BUG_ON(!rcu_read_lock_held());
2333 memcg
= pc
->mem_cgroup
;
2334 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2337 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2341 * size of first charge trial. "32" comes from vmscan.c's magic value.
2342 * TODO: maybe necessary to use big numbers in big irons.
2344 #define CHARGE_BATCH 32U
2345 struct memcg_stock_pcp
{
2346 struct mem_cgroup
*cached
; /* this never be root cgroup */
2347 unsigned int nr_pages
;
2348 struct work_struct work
;
2349 unsigned long flags
;
2350 #define FLUSHING_CACHED_CHARGE 0
2352 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2353 static DEFINE_MUTEX(percpu_charge_mutex
);
2356 * consume_stock: Try to consume stocked charge on this cpu.
2357 * @memcg: memcg to consume from.
2358 * @nr_pages: how many pages to charge.
2360 * The charges will only happen if @memcg matches the current cpu's memcg
2361 * stock, and at least @nr_pages are available in that stock. Failure to
2362 * service an allocation will refill the stock.
2364 * returns true if successful, false otherwise.
2366 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2368 struct memcg_stock_pcp
*stock
;
2371 if (nr_pages
> CHARGE_BATCH
)
2374 stock
= &get_cpu_var(memcg_stock
);
2375 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2376 stock
->nr_pages
-= nr_pages
;
2377 else /* need to call res_counter_charge */
2379 put_cpu_var(memcg_stock
);
2384 * Returns stocks cached in percpu to res_counter and reset cached information.
2386 static void drain_stock(struct memcg_stock_pcp
*stock
)
2388 struct mem_cgroup
*old
= stock
->cached
;
2390 if (stock
->nr_pages
) {
2391 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2393 res_counter_uncharge(&old
->res
, bytes
);
2394 if (do_swap_account
)
2395 res_counter_uncharge(&old
->memsw
, bytes
);
2396 stock
->nr_pages
= 0;
2398 stock
->cached
= NULL
;
2402 * This must be called under preempt disabled or must be called by
2403 * a thread which is pinned to local cpu.
2405 static void drain_local_stock(struct work_struct
*dummy
)
2407 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2409 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2412 static void __init
memcg_stock_init(void)
2416 for_each_possible_cpu(cpu
) {
2417 struct memcg_stock_pcp
*stock
=
2418 &per_cpu(memcg_stock
, cpu
);
2419 INIT_WORK(&stock
->work
, drain_local_stock
);
2424 * Cache charges(val) which is from res_counter, to local per_cpu area.
2425 * This will be consumed by consume_stock() function, later.
2427 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2429 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2431 if (stock
->cached
!= memcg
) { /* reset if necessary */
2433 stock
->cached
= memcg
;
2435 stock
->nr_pages
+= nr_pages
;
2436 put_cpu_var(memcg_stock
);
2440 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2441 * of the hierarchy under it. sync flag says whether we should block
2442 * until the work is done.
2444 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2448 /* Notify other cpus that system-wide "drain" is running */
2451 for_each_online_cpu(cpu
) {
2452 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2453 struct mem_cgroup
*memcg
;
2455 memcg
= stock
->cached
;
2456 if (!memcg
|| !stock
->nr_pages
)
2458 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2460 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2462 drain_local_stock(&stock
->work
);
2464 schedule_work_on(cpu
, &stock
->work
);
2472 for_each_online_cpu(cpu
) {
2473 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2474 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2475 flush_work(&stock
->work
);
2482 * Tries to drain stocked charges in other cpus. This function is asynchronous
2483 * and just put a work per cpu for draining localy on each cpu. Caller can
2484 * expects some charges will be back to res_counter later but cannot wait for
2487 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2490 * If someone calls draining, avoid adding more kworker runs.
2492 if (!mutex_trylock(&percpu_charge_mutex
))
2494 drain_all_stock(root_memcg
, false);
2495 mutex_unlock(&percpu_charge_mutex
);
2498 /* This is a synchronous drain interface. */
2499 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2501 /* called when force_empty is called */
2502 mutex_lock(&percpu_charge_mutex
);
2503 drain_all_stock(root_memcg
, true);
2504 mutex_unlock(&percpu_charge_mutex
);
2508 * This function drains percpu counter value from DEAD cpu and
2509 * move it to local cpu. Note that this function can be preempted.
2511 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2515 spin_lock(&memcg
->pcp_counter_lock
);
2516 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2517 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2519 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2520 memcg
->nocpu_base
.count
[i
] += x
;
2522 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2523 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2525 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2526 memcg
->nocpu_base
.events
[i
] += x
;
2528 spin_unlock(&memcg
->pcp_counter_lock
);
2531 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2532 unsigned long action
,
2535 int cpu
= (unsigned long)hcpu
;
2536 struct memcg_stock_pcp
*stock
;
2537 struct mem_cgroup
*iter
;
2539 if (action
== CPU_ONLINE
)
2542 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2545 for_each_mem_cgroup(iter
)
2546 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2548 stock
= &per_cpu(memcg_stock
, cpu
);
2554 /* See __mem_cgroup_try_charge() for details */
2556 CHARGE_OK
, /* success */
2557 CHARGE_RETRY
, /* need to retry but retry is not bad */
2558 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2559 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2562 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2563 unsigned int nr_pages
, unsigned int min_pages
,
2566 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2567 struct mem_cgroup
*mem_over_limit
;
2568 struct res_counter
*fail_res
;
2569 unsigned long flags
= 0;
2572 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2575 if (!do_swap_account
)
2577 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2581 res_counter_uncharge(&memcg
->res
, csize
);
2582 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2583 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2585 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2587 * Never reclaim on behalf of optional batching, retry with a
2588 * single page instead.
2590 if (nr_pages
> min_pages
)
2591 return CHARGE_RETRY
;
2593 if (!(gfp_mask
& __GFP_WAIT
))
2594 return CHARGE_WOULDBLOCK
;
2596 if (gfp_mask
& __GFP_NORETRY
)
2597 return CHARGE_NOMEM
;
2599 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2600 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2601 return CHARGE_RETRY
;
2603 * Even though the limit is exceeded at this point, reclaim
2604 * may have been able to free some pages. Retry the charge
2605 * before killing the task.
2607 * Only for regular pages, though: huge pages are rather
2608 * unlikely to succeed so close to the limit, and we fall back
2609 * to regular pages anyway in case of failure.
2611 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2612 return CHARGE_RETRY
;
2615 * At task move, charge accounts can be doubly counted. So, it's
2616 * better to wait until the end of task_move if something is going on.
2618 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2619 return CHARGE_RETRY
;
2622 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2624 return CHARGE_NOMEM
;
2628 * __mem_cgroup_try_charge() does
2629 * 1. detect memcg to be charged against from passed *mm and *ptr,
2630 * 2. update res_counter
2631 * 3. call memory reclaim if necessary.
2633 * In some special case, if the task is fatal, fatal_signal_pending() or
2634 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2635 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2636 * as possible without any hazards. 2: all pages should have a valid
2637 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2638 * pointer, that is treated as a charge to root_mem_cgroup.
2640 * So __mem_cgroup_try_charge() will return
2641 * 0 ... on success, filling *ptr with a valid memcg pointer.
2642 * -ENOMEM ... charge failure because of resource limits.
2643 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2645 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2646 * the oom-killer can be invoked.
2648 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2650 unsigned int nr_pages
,
2651 struct mem_cgroup
**ptr
,
2654 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2655 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2656 struct mem_cgroup
*memcg
= NULL
;
2660 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2661 * in system level. So, allow to go ahead dying process in addition to
2664 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2665 || fatal_signal_pending(current
)))
2668 if (unlikely(task_in_memcg_oom(current
)))
2672 * We always charge the cgroup the mm_struct belongs to.
2673 * The mm_struct's mem_cgroup changes on task migration if the
2674 * thread group leader migrates. It's possible that mm is not
2675 * set, if so charge the root memcg (happens for pagecache usage).
2678 *ptr
= root_mem_cgroup
;
2680 if (*ptr
) { /* css should be a valid one */
2682 if (mem_cgroup_is_root(memcg
))
2684 if (consume_stock(memcg
, nr_pages
))
2686 css_get(&memcg
->css
);
2688 struct task_struct
*p
;
2691 p
= rcu_dereference(mm
->owner
);
2693 * Because we don't have task_lock(), "p" can exit.
2694 * In that case, "memcg" can point to root or p can be NULL with
2695 * race with swapoff. Then, we have small risk of mis-accouning.
2696 * But such kind of mis-account by race always happens because
2697 * we don't have cgroup_mutex(). It's overkill and we allo that
2699 * (*) swapoff at el will charge against mm-struct not against
2700 * task-struct. So, mm->owner can be NULL.
2702 memcg
= mem_cgroup_from_task(p
);
2704 memcg
= root_mem_cgroup
;
2705 if (mem_cgroup_is_root(memcg
)) {
2709 if (consume_stock(memcg
, nr_pages
)) {
2711 * It seems dagerous to access memcg without css_get().
2712 * But considering how consume_stok works, it's not
2713 * necessary. If consume_stock success, some charges
2714 * from this memcg are cached on this cpu. So, we
2715 * don't need to call css_get()/css_tryget() before
2716 * calling consume_stock().
2721 /* after here, we may be blocked. we need to get refcnt */
2722 if (!css_tryget(&memcg
->css
)) {
2730 bool invoke_oom
= oom
&& !nr_oom_retries
;
2732 /* If killed, bypass charge */
2733 if (fatal_signal_pending(current
)) {
2734 css_put(&memcg
->css
);
2738 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2739 nr_pages
, invoke_oom
);
2743 case CHARGE_RETRY
: /* not in OOM situation but retry */
2745 css_put(&memcg
->css
);
2748 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2749 css_put(&memcg
->css
);
2751 case CHARGE_NOMEM
: /* OOM routine works */
2752 if (!oom
|| invoke_oom
) {
2753 css_put(&memcg
->css
);
2759 } while (ret
!= CHARGE_OK
);
2761 if (batch
> nr_pages
)
2762 refill_stock(memcg
, batch
- nr_pages
);
2763 css_put(&memcg
->css
);
2771 *ptr
= root_mem_cgroup
;
2776 * Somemtimes we have to undo a charge we got by try_charge().
2777 * This function is for that and do uncharge, put css's refcnt.
2778 * gotten by try_charge().
2780 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2781 unsigned int nr_pages
)
2783 if (!mem_cgroup_is_root(memcg
)) {
2784 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2786 res_counter_uncharge(&memcg
->res
, bytes
);
2787 if (do_swap_account
)
2788 res_counter_uncharge(&memcg
->memsw
, bytes
);
2793 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2794 * This is useful when moving usage to parent cgroup.
2796 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2797 unsigned int nr_pages
)
2799 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2801 if (mem_cgroup_is_root(memcg
))
2804 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2805 if (do_swap_account
)
2806 res_counter_uncharge_until(&memcg
->memsw
,
2807 memcg
->memsw
.parent
, bytes
);
2811 * A helper function to get mem_cgroup from ID. must be called under
2812 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2813 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2814 * called against removed memcg.)
2816 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2818 struct cgroup_subsys_state
*css
;
2820 /* ID 0 is unused ID */
2823 css
= css_lookup(&mem_cgroup_subsys
, id
);
2826 return mem_cgroup_from_css(css
);
2829 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2831 struct mem_cgroup
*memcg
= NULL
;
2832 struct page_cgroup
*pc
;
2836 VM_BUG_ON(!PageLocked(page
));
2838 pc
= lookup_page_cgroup(page
);
2839 lock_page_cgroup(pc
);
2840 if (PageCgroupUsed(pc
)) {
2841 memcg
= pc
->mem_cgroup
;
2842 if (memcg
&& !css_tryget(&memcg
->css
))
2844 } else if (PageSwapCache(page
)) {
2845 ent
.val
= page_private(page
);
2846 id
= lookup_swap_cgroup_id(ent
);
2848 memcg
= mem_cgroup_lookup(id
);
2849 if (memcg
&& !css_tryget(&memcg
->css
))
2853 unlock_page_cgroup(pc
);
2857 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2859 unsigned int nr_pages
,
2860 enum charge_type ctype
,
2863 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2864 struct zone
*uninitialized_var(zone
);
2865 struct lruvec
*lruvec
;
2866 bool was_on_lru
= false;
2869 lock_page_cgroup(pc
);
2870 VM_BUG_ON(PageCgroupUsed(pc
));
2872 * we don't need page_cgroup_lock about tail pages, becase they are not
2873 * accessed by any other context at this point.
2877 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2878 * may already be on some other mem_cgroup's LRU. Take care of it.
2881 zone
= page_zone(page
);
2882 spin_lock_irq(&zone
->lru_lock
);
2883 if (PageLRU(page
)) {
2884 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2886 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2891 pc
->mem_cgroup
= memcg
;
2893 * We access a page_cgroup asynchronously without lock_page_cgroup().
2894 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2895 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2896 * before USED bit, we need memory barrier here.
2897 * See mem_cgroup_add_lru_list(), etc.
2900 SetPageCgroupUsed(pc
);
2904 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2905 VM_BUG_ON(PageLRU(page
));
2907 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2909 spin_unlock_irq(&zone
->lru_lock
);
2912 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2917 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2918 unlock_page_cgroup(pc
);
2921 * "charge_statistics" updated event counter. Then, check it.
2922 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2923 * if they exceeds softlimit.
2925 memcg_check_events(memcg
, page
);
2928 static DEFINE_MUTEX(set_limit_mutex
);
2930 #ifdef CONFIG_MEMCG_KMEM
2931 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2933 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2934 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2938 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2939 * in the memcg_cache_params struct.
2941 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2943 struct kmem_cache
*cachep
;
2945 VM_BUG_ON(p
->is_root_cache
);
2946 cachep
= p
->root_cache
;
2947 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2950 #ifdef CONFIG_SLABINFO
2951 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2952 struct cftype
*cft
, struct seq_file
*m
)
2954 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2955 struct memcg_cache_params
*params
;
2957 if (!memcg_can_account_kmem(memcg
))
2960 print_slabinfo_header(m
);
2962 mutex_lock(&memcg
->slab_caches_mutex
);
2963 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2964 cache_show(memcg_params_to_cache(params
), m
);
2965 mutex_unlock(&memcg
->slab_caches_mutex
);
2971 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2973 struct res_counter
*fail_res
;
2974 struct mem_cgroup
*_memcg
;
2978 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2983 * Conditions under which we can wait for the oom_killer. Those are
2984 * the same conditions tested by the core page allocator
2986 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2989 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2992 if (ret
== -EINTR
) {
2994 * __mem_cgroup_try_charge() chosed to bypass to root due to
2995 * OOM kill or fatal signal. Since our only options are to
2996 * either fail the allocation or charge it to this cgroup, do
2997 * it as a temporary condition. But we can't fail. From a
2998 * kmem/slab perspective, the cache has already been selected,
2999 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3002 * This condition will only trigger if the task entered
3003 * memcg_charge_kmem in a sane state, but was OOM-killed during
3004 * __mem_cgroup_try_charge() above. Tasks that were already
3005 * dying when the allocation triggers should have been already
3006 * directed to the root cgroup in memcontrol.h
3008 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
3009 if (do_swap_account
)
3010 res_counter_charge_nofail(&memcg
->memsw
, size
,
3014 res_counter_uncharge(&memcg
->kmem
, size
);
3019 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
3021 res_counter_uncharge(&memcg
->res
, size
);
3022 if (do_swap_account
)
3023 res_counter_uncharge(&memcg
->memsw
, size
);
3026 if (res_counter_uncharge(&memcg
->kmem
, size
))
3030 * Releases a reference taken in kmem_cgroup_css_offline in case
3031 * this last uncharge is racing with the offlining code or it is
3032 * outliving the memcg existence.
3034 * The memory barrier imposed by test&clear is paired with the
3035 * explicit one in memcg_kmem_mark_dead().
3037 if (memcg_kmem_test_and_clear_dead(memcg
))
3038 css_put(&memcg
->css
);
3041 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
3046 mutex_lock(&memcg
->slab_caches_mutex
);
3047 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3048 mutex_unlock(&memcg
->slab_caches_mutex
);
3052 * helper for acessing a memcg's index. It will be used as an index in the
3053 * child cache array in kmem_cache, and also to derive its name. This function
3054 * will return -1 when this is not a kmem-limited memcg.
3056 int memcg_cache_id(struct mem_cgroup
*memcg
)
3058 return memcg
? memcg
->kmemcg_id
: -1;
3062 * This ends up being protected by the set_limit mutex, during normal
3063 * operation, because that is its main call site.
3065 * But when we create a new cache, we can call this as well if its parent
3066 * is kmem-limited. That will have to hold set_limit_mutex as well.
3068 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
3072 num
= ida_simple_get(&kmem_limited_groups
,
3073 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
3077 * After this point, kmem_accounted (that we test atomically in
3078 * the beginning of this conditional), is no longer 0. This
3079 * guarantees only one process will set the following boolean
3080 * to true. We don't need test_and_set because we're protected
3081 * by the set_limit_mutex anyway.
3083 memcg_kmem_set_activated(memcg
);
3085 ret
= memcg_update_all_caches(num
+1);
3087 ida_simple_remove(&kmem_limited_groups
, num
);
3088 memcg_kmem_clear_activated(memcg
);
3092 memcg
->kmemcg_id
= num
;
3093 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3094 mutex_init(&memcg
->slab_caches_mutex
);
3098 static size_t memcg_caches_array_size(int num_groups
)
3101 if (num_groups
<= 0)
3104 size
= 2 * num_groups
;
3105 if (size
< MEMCG_CACHES_MIN_SIZE
)
3106 size
= MEMCG_CACHES_MIN_SIZE
;
3107 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3108 size
= MEMCG_CACHES_MAX_SIZE
;
3114 * We should update the current array size iff all caches updates succeed. This
3115 * can only be done from the slab side. The slab mutex needs to be held when
3118 void memcg_update_array_size(int num
)
3120 if (num
> memcg_limited_groups_array_size
)
3121 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3124 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3126 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3128 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3130 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3132 if (num_groups
> memcg_limited_groups_array_size
) {
3134 ssize_t size
= memcg_caches_array_size(num_groups
);
3136 size
*= sizeof(void *);
3137 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3139 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3140 if (!s
->memcg_params
) {
3141 s
->memcg_params
= cur_params
;
3145 s
->memcg_params
->is_root_cache
= true;
3148 * There is the chance it will be bigger than
3149 * memcg_limited_groups_array_size, if we failed an allocation
3150 * in a cache, in which case all caches updated before it, will
3151 * have a bigger array.
3153 * But if that is the case, the data after
3154 * memcg_limited_groups_array_size is certainly unused
3156 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3157 if (!cur_params
->memcg_caches
[i
])
3159 s
->memcg_params
->memcg_caches
[i
] =
3160 cur_params
->memcg_caches
[i
];
3164 * Ideally, we would wait until all caches succeed, and only
3165 * then free the old one. But this is not worth the extra
3166 * pointer per-cache we'd have to have for this.
3168 * It is not a big deal if some caches are left with a size
3169 * bigger than the others. And all updates will reset this
3177 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3178 struct kmem_cache
*root_cache
)
3182 if (!memcg_kmem_enabled())
3186 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3187 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3189 size
= sizeof(struct memcg_cache_params
);
3191 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3192 if (!s
->memcg_params
)
3196 s
->memcg_params
->memcg
= memcg
;
3197 s
->memcg_params
->root_cache
= root_cache
;
3198 INIT_WORK(&s
->memcg_params
->destroy
,
3199 kmem_cache_destroy_work_func
);
3201 s
->memcg_params
->is_root_cache
= true;
3206 void memcg_release_cache(struct kmem_cache
*s
)
3208 struct kmem_cache
*root
;
3209 struct mem_cgroup
*memcg
;
3213 * This happens, for instance, when a root cache goes away before we
3216 if (!s
->memcg_params
)
3219 if (s
->memcg_params
->is_root_cache
)
3222 memcg
= s
->memcg_params
->memcg
;
3223 id
= memcg_cache_id(memcg
);
3225 root
= s
->memcg_params
->root_cache
;
3226 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3228 mutex_lock(&memcg
->slab_caches_mutex
);
3229 list_del(&s
->memcg_params
->list
);
3230 mutex_unlock(&memcg
->slab_caches_mutex
);
3232 css_put(&memcg
->css
);
3234 kfree(s
->memcg_params
);
3238 * During the creation a new cache, we need to disable our accounting mechanism
3239 * altogether. This is true even if we are not creating, but rather just
3240 * enqueing new caches to be created.
3242 * This is because that process will trigger allocations; some visible, like
3243 * explicit kmallocs to auxiliary data structures, name strings and internal
3244 * cache structures; some well concealed, like INIT_WORK() that can allocate
3245 * objects during debug.
3247 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3248 * to it. This may not be a bounded recursion: since the first cache creation
3249 * failed to complete (waiting on the allocation), we'll just try to create the
3250 * cache again, failing at the same point.
3252 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3253 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3254 * inside the following two functions.
3256 static inline void memcg_stop_kmem_account(void)
3258 VM_BUG_ON(!current
->mm
);
3259 current
->memcg_kmem_skip_account
++;
3262 static inline void memcg_resume_kmem_account(void)
3264 VM_BUG_ON(!current
->mm
);
3265 current
->memcg_kmem_skip_account
--;
3268 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3270 struct kmem_cache
*cachep
;
3271 struct memcg_cache_params
*p
;
3273 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3275 cachep
= memcg_params_to_cache(p
);
3278 * If we get down to 0 after shrink, we could delete right away.
3279 * However, memcg_release_pages() already puts us back in the workqueue
3280 * in that case. If we proceed deleting, we'll get a dangling
3281 * reference, and removing the object from the workqueue in that case
3282 * is unnecessary complication. We are not a fast path.
3284 * Note that this case is fundamentally different from racing with
3285 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3286 * kmem_cache_shrink, not only we would be reinserting a dead cache
3287 * into the queue, but doing so from inside the worker racing to
3290 * So if we aren't down to zero, we'll just schedule a worker and try
3293 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3294 kmem_cache_shrink(cachep
);
3295 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3298 kmem_cache_destroy(cachep
);
3301 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3303 if (!cachep
->memcg_params
->dead
)
3307 * There are many ways in which we can get here.
3309 * We can get to a memory-pressure situation while the delayed work is
3310 * still pending to run. The vmscan shrinkers can then release all
3311 * cache memory and get us to destruction. If this is the case, we'll
3312 * be executed twice, which is a bug (the second time will execute over
3313 * bogus data). In this case, cancelling the work should be fine.
3315 * But we can also get here from the worker itself, if
3316 * kmem_cache_shrink is enough to shake all the remaining objects and
3317 * get the page count to 0. In this case, we'll deadlock if we try to
3318 * cancel the work (the worker runs with an internal lock held, which
3319 * is the same lock we would hold for cancel_work_sync().)
3321 * Since we can't possibly know who got us here, just refrain from
3322 * running if there is already work pending
3324 if (work_pending(&cachep
->memcg_params
->destroy
))
3327 * We have to defer the actual destroying to a workqueue, because
3328 * we might currently be in a context that cannot sleep.
3330 schedule_work(&cachep
->memcg_params
->destroy
);
3334 * This lock protects updaters, not readers. We want readers to be as fast as
3335 * they can, and they will either see NULL or a valid cache value. Our model
3336 * allow them to see NULL, in which case the root memcg will be selected.
3338 * We need this lock because multiple allocations to the same cache from a non
3339 * will span more than one worker. Only one of them can create the cache.
3341 static DEFINE_MUTEX(memcg_cache_mutex
);
3344 * Called with memcg_cache_mutex held
3346 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3347 struct kmem_cache
*s
)
3349 struct kmem_cache
*new;
3350 static char *tmp_name
= NULL
;
3352 lockdep_assert_held(&memcg_cache_mutex
);
3355 * kmem_cache_create_memcg duplicates the given name and
3356 * cgroup_name for this name requires RCU context.
3357 * This static temporary buffer is used to prevent from
3358 * pointless shortliving allocation.
3361 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3367 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3368 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3371 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3372 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3375 new->allocflags
|= __GFP_KMEMCG
;
3380 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3381 struct kmem_cache
*cachep
)
3383 struct kmem_cache
*new_cachep
;
3386 BUG_ON(!memcg_can_account_kmem(memcg
));
3388 idx
= memcg_cache_id(memcg
);
3390 mutex_lock(&memcg_cache_mutex
);
3391 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3393 css_put(&memcg
->css
);
3397 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3398 if (new_cachep
== NULL
) {
3399 new_cachep
= cachep
;
3400 css_put(&memcg
->css
);
3404 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3406 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3408 * the readers won't lock, make sure everybody sees the updated value,
3409 * so they won't put stuff in the queue again for no reason
3413 mutex_unlock(&memcg_cache_mutex
);
3417 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3419 struct kmem_cache
*c
;
3422 if (!s
->memcg_params
)
3424 if (!s
->memcg_params
->is_root_cache
)
3428 * If the cache is being destroyed, we trust that there is no one else
3429 * requesting objects from it. Even if there are, the sanity checks in
3430 * kmem_cache_destroy should caught this ill-case.
3432 * Still, we don't want anyone else freeing memcg_caches under our
3433 * noses, which can happen if a new memcg comes to life. As usual,
3434 * we'll take the set_limit_mutex to protect ourselves against this.
3436 mutex_lock(&set_limit_mutex
);
3437 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3438 c
= s
->memcg_params
->memcg_caches
[i
];
3443 * We will now manually delete the caches, so to avoid races
3444 * we need to cancel all pending destruction workers and
3445 * proceed with destruction ourselves.
3447 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3448 * and that could spawn the workers again: it is likely that
3449 * the cache still have active pages until this very moment.
3450 * This would lead us back to mem_cgroup_destroy_cache.
3452 * But that will not execute at all if the "dead" flag is not
3453 * set, so flip it down to guarantee we are in control.
3455 c
->memcg_params
->dead
= false;
3456 cancel_work_sync(&c
->memcg_params
->destroy
);
3457 kmem_cache_destroy(c
);
3459 mutex_unlock(&set_limit_mutex
);
3462 struct create_work
{
3463 struct mem_cgroup
*memcg
;
3464 struct kmem_cache
*cachep
;
3465 struct work_struct work
;
3468 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3470 struct kmem_cache
*cachep
;
3471 struct memcg_cache_params
*params
;
3473 if (!memcg_kmem_is_active(memcg
))
3476 mutex_lock(&memcg
->slab_caches_mutex
);
3477 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3478 cachep
= memcg_params_to_cache(params
);
3479 cachep
->memcg_params
->dead
= true;
3480 schedule_work(&cachep
->memcg_params
->destroy
);
3482 mutex_unlock(&memcg
->slab_caches_mutex
);
3485 static void memcg_create_cache_work_func(struct work_struct
*w
)
3487 struct create_work
*cw
;
3489 cw
= container_of(w
, struct create_work
, work
);
3490 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3495 * Enqueue the creation of a per-memcg kmem_cache.
3497 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3498 struct kmem_cache
*cachep
)
3500 struct create_work
*cw
;
3502 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3504 css_put(&memcg
->css
);
3509 cw
->cachep
= cachep
;
3511 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3512 schedule_work(&cw
->work
);
3515 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3516 struct kmem_cache
*cachep
)
3519 * We need to stop accounting when we kmalloc, because if the
3520 * corresponding kmalloc cache is not yet created, the first allocation
3521 * in __memcg_create_cache_enqueue will recurse.
3523 * However, it is better to enclose the whole function. Depending on
3524 * the debugging options enabled, INIT_WORK(), for instance, can
3525 * trigger an allocation. This too, will make us recurse. Because at
3526 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3527 * the safest choice is to do it like this, wrapping the whole function.
3529 memcg_stop_kmem_account();
3530 __memcg_create_cache_enqueue(memcg
, cachep
);
3531 memcg_resume_kmem_account();
3534 * Return the kmem_cache we're supposed to use for a slab allocation.
3535 * We try to use the current memcg's version of the cache.
3537 * If the cache does not exist yet, if we are the first user of it,
3538 * we either create it immediately, if possible, or create it asynchronously
3540 * In the latter case, we will let the current allocation go through with
3541 * the original cache.
3543 * Can't be called in interrupt context or from kernel threads.
3544 * This function needs to be called with rcu_read_lock() held.
3546 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3549 struct mem_cgroup
*memcg
;
3552 VM_BUG_ON(!cachep
->memcg_params
);
3553 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3555 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3559 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3561 if (!memcg_can_account_kmem(memcg
))
3564 idx
= memcg_cache_id(memcg
);
3567 * barrier to mare sure we're always seeing the up to date value. The
3568 * code updating memcg_caches will issue a write barrier to match this.
3570 read_barrier_depends();
3571 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3572 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3576 /* The corresponding put will be done in the workqueue. */
3577 if (!css_tryget(&memcg
->css
))
3582 * If we are in a safe context (can wait, and not in interrupt
3583 * context), we could be be predictable and return right away.
3584 * This would guarantee that the allocation being performed
3585 * already belongs in the new cache.
3587 * However, there are some clashes that can arrive from locking.
3588 * For instance, because we acquire the slab_mutex while doing
3589 * kmem_cache_dup, this means no further allocation could happen
3590 * with the slab_mutex held.
3592 * Also, because cache creation issue get_online_cpus(), this
3593 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3594 * that ends up reversed during cpu hotplug. (cpuset allocates
3595 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3596 * better to defer everything.
3598 memcg_create_cache_enqueue(memcg
, cachep
);
3604 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3607 * We need to verify if the allocation against current->mm->owner's memcg is
3608 * possible for the given order. But the page is not allocated yet, so we'll
3609 * need a further commit step to do the final arrangements.
3611 * It is possible for the task to switch cgroups in this mean time, so at
3612 * commit time, we can't rely on task conversion any longer. We'll then use
3613 * the handle argument to return to the caller which cgroup we should commit
3614 * against. We could also return the memcg directly and avoid the pointer
3615 * passing, but a boolean return value gives better semantics considering
3616 * the compiled-out case as well.
3618 * Returning true means the allocation is possible.
3621 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3623 struct mem_cgroup
*memcg
;
3629 * Disabling accounting is only relevant for some specific memcg
3630 * internal allocations. Therefore we would initially not have such
3631 * check here, since direct calls to the page allocator that are marked
3632 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3633 * concerned with cache allocations, and by having this test at
3634 * memcg_kmem_get_cache, we are already able to relay the allocation to
3635 * the root cache and bypass the memcg cache altogether.
3637 * There is one exception, though: the SLUB allocator does not create
3638 * large order caches, but rather service large kmallocs directly from
3639 * the page allocator. Therefore, the following sequence when backed by
3640 * the SLUB allocator:
3642 * memcg_stop_kmem_account();
3643 * kmalloc(<large_number>)
3644 * memcg_resume_kmem_account();
3646 * would effectively ignore the fact that we should skip accounting,
3647 * since it will drive us directly to this function without passing
3648 * through the cache selector memcg_kmem_get_cache. Such large
3649 * allocations are extremely rare but can happen, for instance, for the
3650 * cache arrays. We bring this test here.
3652 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3655 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3658 * very rare case described in mem_cgroup_from_task. Unfortunately there
3659 * isn't much we can do without complicating this too much, and it would
3660 * be gfp-dependent anyway. Just let it go
3662 if (unlikely(!memcg
))
3665 if (!memcg_can_account_kmem(memcg
)) {
3666 css_put(&memcg
->css
);
3670 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3674 css_put(&memcg
->css
);
3678 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3681 struct page_cgroup
*pc
;
3683 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3685 /* The page allocation failed. Revert */
3687 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3691 pc
= lookup_page_cgroup(page
);
3692 lock_page_cgroup(pc
);
3693 pc
->mem_cgroup
= memcg
;
3694 SetPageCgroupUsed(pc
);
3695 unlock_page_cgroup(pc
);
3698 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3700 struct mem_cgroup
*memcg
= NULL
;
3701 struct page_cgroup
*pc
;
3704 pc
= lookup_page_cgroup(page
);
3706 * Fast unlocked return. Theoretically might have changed, have to
3707 * check again after locking.
3709 if (!PageCgroupUsed(pc
))
3712 lock_page_cgroup(pc
);
3713 if (PageCgroupUsed(pc
)) {
3714 memcg
= pc
->mem_cgroup
;
3715 ClearPageCgroupUsed(pc
);
3717 unlock_page_cgroup(pc
);
3720 * We trust that only if there is a memcg associated with the page, it
3721 * is a valid allocation
3726 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3727 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3730 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3733 #endif /* CONFIG_MEMCG_KMEM */
3735 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3737 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3739 * Because tail pages are not marked as "used", set it. We're under
3740 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3741 * charge/uncharge will be never happen and move_account() is done under
3742 * compound_lock(), so we don't have to take care of races.
3744 void mem_cgroup_split_huge_fixup(struct page
*head
)
3746 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3747 struct page_cgroup
*pc
;
3748 struct mem_cgroup
*memcg
;
3751 if (mem_cgroup_disabled())
3754 memcg
= head_pc
->mem_cgroup
;
3755 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3757 pc
->mem_cgroup
= memcg
;
3758 smp_wmb();/* see __commit_charge() */
3759 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3761 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3764 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3767 void mem_cgroup_move_account_page_stat(struct mem_cgroup
*from
,
3768 struct mem_cgroup
*to
,
3769 unsigned int nr_pages
,
3770 enum mem_cgroup_stat_index idx
)
3772 /* Update stat data for mem_cgroup */
3774 WARN_ON_ONCE(from
->stat
->count
[idx
] < nr_pages
);
3775 __this_cpu_add(from
->stat
->count
[idx
], -nr_pages
);
3776 __this_cpu_add(to
->stat
->count
[idx
], nr_pages
);
3781 * mem_cgroup_move_account - move account of the page
3783 * @nr_pages: number of regular pages (>1 for huge pages)
3784 * @pc: page_cgroup of the page.
3785 * @from: mem_cgroup which the page is moved from.
3786 * @to: mem_cgroup which the page is moved to. @from != @to.
3788 * The caller must confirm following.
3789 * - page is not on LRU (isolate_page() is useful.)
3790 * - compound_lock is held when nr_pages > 1
3792 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3795 static int mem_cgroup_move_account(struct page
*page
,
3796 unsigned int nr_pages
,
3797 struct page_cgroup
*pc
,
3798 struct mem_cgroup
*from
,
3799 struct mem_cgroup
*to
)
3801 unsigned long flags
;
3803 bool anon
= PageAnon(page
);
3805 VM_BUG_ON(from
== to
);
3806 VM_BUG_ON(PageLRU(page
));
3808 * The page is isolated from LRU. So, collapse function
3809 * will not handle this page. But page splitting can happen.
3810 * Do this check under compound_page_lock(). The caller should
3814 if (nr_pages
> 1 && !PageTransHuge(page
))
3817 lock_page_cgroup(pc
);
3820 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3823 move_lock_mem_cgroup(from
, &flags
);
3825 if (!anon
&& page_mapped(page
))
3826 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3827 MEM_CGROUP_STAT_FILE_MAPPED
);
3829 if (PageWriteback(page
))
3830 mem_cgroup_move_account_page_stat(from
, to
, nr_pages
,
3831 MEM_CGROUP_STAT_WRITEBACK
);
3833 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3835 /* caller should have done css_get */
3836 pc
->mem_cgroup
= to
;
3837 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3838 move_unlock_mem_cgroup(from
, &flags
);
3841 unlock_page_cgroup(pc
);
3845 memcg_check_events(to
, page
);
3846 memcg_check_events(from
, page
);
3852 * mem_cgroup_move_parent - moves page to the parent group
3853 * @page: the page to move
3854 * @pc: page_cgroup of the page
3855 * @child: page's cgroup
3857 * move charges to its parent or the root cgroup if the group has no
3858 * parent (aka use_hierarchy==0).
3859 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3860 * mem_cgroup_move_account fails) the failure is always temporary and
3861 * it signals a race with a page removal/uncharge or migration. In the
3862 * first case the page is on the way out and it will vanish from the LRU
3863 * on the next attempt and the call should be retried later.
3864 * Isolation from the LRU fails only if page has been isolated from
3865 * the LRU since we looked at it and that usually means either global
3866 * reclaim or migration going on. The page will either get back to the
3868 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3869 * (!PageCgroupUsed) or moved to a different group. The page will
3870 * disappear in the next attempt.
3872 static int mem_cgroup_move_parent(struct page
*page
,
3873 struct page_cgroup
*pc
,
3874 struct mem_cgroup
*child
)
3876 struct mem_cgroup
*parent
;
3877 unsigned int nr_pages
;
3878 unsigned long uninitialized_var(flags
);
3881 VM_BUG_ON(mem_cgroup_is_root(child
));
3884 if (!get_page_unless_zero(page
))
3886 if (isolate_lru_page(page
))
3889 nr_pages
= hpage_nr_pages(page
);
3891 parent
= parent_mem_cgroup(child
);
3893 * If no parent, move charges to root cgroup.
3896 parent
= root_mem_cgroup
;
3899 VM_BUG_ON(!PageTransHuge(page
));
3900 flags
= compound_lock_irqsave(page
);
3903 ret
= mem_cgroup_move_account(page
, nr_pages
,
3906 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3909 compound_unlock_irqrestore(page
, flags
);
3910 putback_lru_page(page
);
3918 * Charge the memory controller for page usage.
3920 * 0 if the charge was successful
3921 * < 0 if the cgroup is over its limit
3923 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3924 gfp_t gfp_mask
, enum charge_type ctype
)
3926 struct mem_cgroup
*memcg
= NULL
;
3927 unsigned int nr_pages
= 1;
3931 if (PageTransHuge(page
)) {
3932 nr_pages
<<= compound_order(page
);
3933 VM_BUG_ON(!PageTransHuge(page
));
3935 * Never OOM-kill a process for a huge page. The
3936 * fault handler will fall back to regular pages.
3941 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3944 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3948 int mem_cgroup_newpage_charge(struct page
*page
,
3949 struct mm_struct
*mm
, gfp_t gfp_mask
)
3951 if (mem_cgroup_disabled())
3953 VM_BUG_ON(page_mapped(page
));
3954 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3956 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3957 MEM_CGROUP_CHARGE_TYPE_ANON
);
3961 * While swap-in, try_charge -> commit or cancel, the page is locked.
3962 * And when try_charge() successfully returns, one refcnt to memcg without
3963 * struct page_cgroup is acquired. This refcnt will be consumed by
3964 * "commit()" or removed by "cancel()"
3966 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3969 struct mem_cgroup
**memcgp
)
3971 struct mem_cgroup
*memcg
;
3972 struct page_cgroup
*pc
;
3975 pc
= lookup_page_cgroup(page
);
3977 * Every swap fault against a single page tries to charge the
3978 * page, bail as early as possible. shmem_unuse() encounters
3979 * already charged pages, too. The USED bit is protected by
3980 * the page lock, which serializes swap cache removal, which
3981 * in turn serializes uncharging.
3983 if (PageCgroupUsed(pc
))
3985 if (!do_swap_account
)
3987 memcg
= try_get_mem_cgroup_from_page(page
);
3991 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3992 css_put(&memcg
->css
);
3997 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
4003 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
4004 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
4007 if (mem_cgroup_disabled())
4010 * A racing thread's fault, or swapoff, may have already
4011 * updated the pte, and even removed page from swap cache: in
4012 * those cases unuse_pte()'s pte_same() test will fail; but
4013 * there's also a KSM case which does need to charge the page.
4015 if (!PageSwapCache(page
)) {
4018 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
4023 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
4026 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
4028 if (mem_cgroup_disabled())
4032 __mem_cgroup_cancel_charge(memcg
, 1);
4036 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
4037 enum charge_type ctype
)
4039 if (mem_cgroup_disabled())
4044 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
4046 * Now swap is on-memory. This means this page may be
4047 * counted both as mem and swap....double count.
4048 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4049 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4050 * may call delete_from_swap_cache() before reach here.
4052 if (do_swap_account
&& PageSwapCache(page
)) {
4053 swp_entry_t ent
= {.val
= page_private(page
)};
4054 mem_cgroup_uncharge_swap(ent
);
4058 void mem_cgroup_commit_charge_swapin(struct page
*page
,
4059 struct mem_cgroup
*memcg
)
4061 __mem_cgroup_commit_charge_swapin(page
, memcg
,
4062 MEM_CGROUP_CHARGE_TYPE_ANON
);
4065 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
4068 struct mem_cgroup
*memcg
= NULL
;
4069 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4072 if (mem_cgroup_disabled())
4074 if (PageCompound(page
))
4077 if (!PageSwapCache(page
))
4078 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
4079 else { /* page is swapcache/shmem */
4080 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
4083 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
4088 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
4089 unsigned int nr_pages
,
4090 const enum charge_type ctype
)
4092 struct memcg_batch_info
*batch
= NULL
;
4093 bool uncharge_memsw
= true;
4095 /* If swapout, usage of swap doesn't decrease */
4096 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
4097 uncharge_memsw
= false;
4099 batch
= ¤t
->memcg_batch
;
4101 * In usual, we do css_get() when we remember memcg pointer.
4102 * But in this case, we keep res->usage until end of a series of
4103 * uncharges. Then, it's ok to ignore memcg's refcnt.
4106 batch
->memcg
= memcg
;
4108 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4109 * In those cases, all pages freed continuously can be expected to be in
4110 * the same cgroup and we have chance to coalesce uncharges.
4111 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4112 * because we want to do uncharge as soon as possible.
4115 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4116 goto direct_uncharge
;
4119 goto direct_uncharge
;
4122 * In typical case, batch->memcg == mem. This means we can
4123 * merge a series of uncharges to an uncharge of res_counter.
4124 * If not, we uncharge res_counter ony by one.
4126 if (batch
->memcg
!= memcg
)
4127 goto direct_uncharge
;
4128 /* remember freed charge and uncharge it later */
4131 batch
->memsw_nr_pages
++;
4134 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4136 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4137 if (unlikely(batch
->memcg
!= memcg
))
4138 memcg_oom_recover(memcg
);
4142 * uncharge if !page_mapped(page)
4144 static struct mem_cgroup
*
4145 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4148 struct mem_cgroup
*memcg
= NULL
;
4149 unsigned int nr_pages
= 1;
4150 struct page_cgroup
*pc
;
4153 if (mem_cgroup_disabled())
4156 if (PageTransHuge(page
)) {
4157 nr_pages
<<= compound_order(page
);
4158 VM_BUG_ON(!PageTransHuge(page
));
4161 * Check if our page_cgroup is valid
4163 pc
= lookup_page_cgroup(page
);
4164 if (unlikely(!PageCgroupUsed(pc
)))
4167 lock_page_cgroup(pc
);
4169 memcg
= pc
->mem_cgroup
;
4171 if (!PageCgroupUsed(pc
))
4174 anon
= PageAnon(page
);
4177 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4179 * Generally PageAnon tells if it's the anon statistics to be
4180 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4181 * used before page reached the stage of being marked PageAnon.
4185 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4186 /* See mem_cgroup_prepare_migration() */
4187 if (page_mapped(page
))
4190 * Pages under migration may not be uncharged. But
4191 * end_migration() /must/ be the one uncharging the
4192 * unused post-migration page and so it has to call
4193 * here with the migration bit still set. See the
4194 * res_counter handling below.
4196 if (!end_migration
&& PageCgroupMigration(pc
))
4199 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4200 if (!PageAnon(page
)) { /* Shared memory */
4201 if (page
->mapping
&& !page_is_file_cache(page
))
4203 } else if (page_mapped(page
)) /* Anon */
4210 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4212 ClearPageCgroupUsed(pc
);
4214 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4215 * freed from LRU. This is safe because uncharged page is expected not
4216 * to be reused (freed soon). Exception is SwapCache, it's handled by
4217 * special functions.
4220 unlock_page_cgroup(pc
);
4222 * even after unlock, we have memcg->res.usage here and this memcg
4223 * will never be freed, so it's safe to call css_get().
4225 memcg_check_events(memcg
, page
);
4226 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4227 mem_cgroup_swap_statistics(memcg
, true);
4228 css_get(&memcg
->css
);
4231 * Migration does not charge the res_counter for the
4232 * replacement page, so leave it alone when phasing out the
4233 * page that is unused after the migration.
4235 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4236 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4241 unlock_page_cgroup(pc
);
4245 void mem_cgroup_uncharge_page(struct page
*page
)
4248 if (page_mapped(page
))
4250 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4252 * If the page is in swap cache, uncharge should be deferred
4253 * to the swap path, which also properly accounts swap usage
4254 * and handles memcg lifetime.
4256 * Note that this check is not stable and reclaim may add the
4257 * page to swap cache at any time after this. However, if the
4258 * page is not in swap cache by the time page->mapcount hits
4259 * 0, there won't be any page table references to the swap
4260 * slot, and reclaim will free it and not actually write the
4263 if (PageSwapCache(page
))
4265 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4268 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4270 VM_BUG_ON(page_mapped(page
));
4271 VM_BUG_ON(page
->mapping
);
4272 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4276 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4277 * In that cases, pages are freed continuously and we can expect pages
4278 * are in the same memcg. All these calls itself limits the number of
4279 * pages freed at once, then uncharge_start/end() is called properly.
4280 * This may be called prural(2) times in a context,
4283 void mem_cgroup_uncharge_start(void)
4285 current
->memcg_batch
.do_batch
++;
4286 /* We can do nest. */
4287 if (current
->memcg_batch
.do_batch
== 1) {
4288 current
->memcg_batch
.memcg
= NULL
;
4289 current
->memcg_batch
.nr_pages
= 0;
4290 current
->memcg_batch
.memsw_nr_pages
= 0;
4294 void mem_cgroup_uncharge_end(void)
4296 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4298 if (!batch
->do_batch
)
4302 if (batch
->do_batch
) /* If stacked, do nothing. */
4308 * This "batch->memcg" is valid without any css_get/put etc...
4309 * bacause we hide charges behind us.
4311 if (batch
->nr_pages
)
4312 res_counter_uncharge(&batch
->memcg
->res
,
4313 batch
->nr_pages
* PAGE_SIZE
);
4314 if (batch
->memsw_nr_pages
)
4315 res_counter_uncharge(&batch
->memcg
->memsw
,
4316 batch
->memsw_nr_pages
* PAGE_SIZE
);
4317 memcg_oom_recover(batch
->memcg
);
4318 /* forget this pointer (for sanity check) */
4319 batch
->memcg
= NULL
;
4324 * called after __delete_from_swap_cache() and drop "page" account.
4325 * memcg information is recorded to swap_cgroup of "ent"
4328 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4330 struct mem_cgroup
*memcg
;
4331 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4333 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4334 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4336 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4339 * record memcg information, if swapout && memcg != NULL,
4340 * css_get() was called in uncharge().
4342 if (do_swap_account
&& swapout
&& memcg
)
4343 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4347 #ifdef CONFIG_MEMCG_SWAP
4349 * called from swap_entry_free(). remove record in swap_cgroup and
4350 * uncharge "memsw" account.
4352 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4354 struct mem_cgroup
*memcg
;
4357 if (!do_swap_account
)
4360 id
= swap_cgroup_record(ent
, 0);
4362 memcg
= mem_cgroup_lookup(id
);
4365 * We uncharge this because swap is freed.
4366 * This memcg can be obsolete one. We avoid calling css_tryget
4368 if (!mem_cgroup_is_root(memcg
))
4369 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4370 mem_cgroup_swap_statistics(memcg
, false);
4371 css_put(&memcg
->css
);
4377 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4378 * @entry: swap entry to be moved
4379 * @from: mem_cgroup which the entry is moved from
4380 * @to: mem_cgroup which the entry is moved to
4382 * It succeeds only when the swap_cgroup's record for this entry is the same
4383 * as the mem_cgroup's id of @from.
4385 * Returns 0 on success, -EINVAL on failure.
4387 * The caller must have charged to @to, IOW, called res_counter_charge() about
4388 * both res and memsw, and called css_get().
4390 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4391 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4393 unsigned short old_id
, new_id
;
4395 old_id
= css_id(&from
->css
);
4396 new_id
= css_id(&to
->css
);
4398 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4399 mem_cgroup_swap_statistics(from
, false);
4400 mem_cgroup_swap_statistics(to
, true);
4402 * This function is only called from task migration context now.
4403 * It postpones res_counter and refcount handling till the end
4404 * of task migration(mem_cgroup_clear_mc()) for performance
4405 * improvement. But we cannot postpone css_get(to) because if
4406 * the process that has been moved to @to does swap-in, the
4407 * refcount of @to might be decreased to 0.
4409 * We are in attach() phase, so the cgroup is guaranteed to be
4410 * alive, so we can just call css_get().
4418 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4419 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4426 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4429 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4430 struct mem_cgroup
**memcgp
)
4432 struct mem_cgroup
*memcg
= NULL
;
4433 unsigned int nr_pages
= 1;
4434 struct page_cgroup
*pc
;
4435 enum charge_type ctype
;
4439 if (mem_cgroup_disabled())
4442 if (PageTransHuge(page
))
4443 nr_pages
<<= compound_order(page
);
4445 pc
= lookup_page_cgroup(page
);
4446 lock_page_cgroup(pc
);
4447 if (PageCgroupUsed(pc
)) {
4448 memcg
= pc
->mem_cgroup
;
4449 css_get(&memcg
->css
);
4451 * At migrating an anonymous page, its mapcount goes down
4452 * to 0 and uncharge() will be called. But, even if it's fully
4453 * unmapped, migration may fail and this page has to be
4454 * charged again. We set MIGRATION flag here and delay uncharge
4455 * until end_migration() is called
4457 * Corner Case Thinking
4459 * When the old page was mapped as Anon and it's unmap-and-freed
4460 * while migration was ongoing.
4461 * If unmap finds the old page, uncharge() of it will be delayed
4462 * until end_migration(). If unmap finds a new page, it's
4463 * uncharged when it make mapcount to be 1->0. If unmap code
4464 * finds swap_migration_entry, the new page will not be mapped
4465 * and end_migration() will find it(mapcount==0).
4468 * When the old page was mapped but migraion fails, the kernel
4469 * remaps it. A charge for it is kept by MIGRATION flag even
4470 * if mapcount goes down to 0. We can do remap successfully
4471 * without charging it again.
4474 * The "old" page is under lock_page() until the end of
4475 * migration, so, the old page itself will not be swapped-out.
4476 * If the new page is swapped out before end_migraton, our
4477 * hook to usual swap-out path will catch the event.
4480 SetPageCgroupMigration(pc
);
4482 unlock_page_cgroup(pc
);
4484 * If the page is not charged at this point,
4492 * We charge new page before it's used/mapped. So, even if unlock_page()
4493 * is called before end_migration, we can catch all events on this new
4494 * page. In the case new page is migrated but not remapped, new page's
4495 * mapcount will be finally 0 and we call uncharge in end_migration().
4498 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4500 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4502 * The page is committed to the memcg, but it's not actually
4503 * charged to the res_counter since we plan on replacing the
4504 * old one and only one page is going to be left afterwards.
4506 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4509 /* remove redundant charge if migration failed*/
4510 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4511 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4513 struct page
*used
, *unused
;
4514 struct page_cgroup
*pc
;
4520 if (!migration_ok
) {
4527 anon
= PageAnon(used
);
4528 __mem_cgroup_uncharge_common(unused
,
4529 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4530 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4532 css_put(&memcg
->css
);
4534 * We disallowed uncharge of pages under migration because mapcount
4535 * of the page goes down to zero, temporarly.
4536 * Clear the flag and check the page should be charged.
4538 pc
= lookup_page_cgroup(oldpage
);
4539 lock_page_cgroup(pc
);
4540 ClearPageCgroupMigration(pc
);
4541 unlock_page_cgroup(pc
);
4544 * If a page is a file cache, radix-tree replacement is very atomic
4545 * and we can skip this check. When it was an Anon page, its mapcount
4546 * goes down to 0. But because we added MIGRATION flage, it's not
4547 * uncharged yet. There are several case but page->mapcount check
4548 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4549 * check. (see prepare_charge() also)
4552 mem_cgroup_uncharge_page(used
);
4556 * At replace page cache, newpage is not under any memcg but it's on
4557 * LRU. So, this function doesn't touch res_counter but handles LRU
4558 * in correct way. Both pages are locked so we cannot race with uncharge.
4560 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4561 struct page
*newpage
)
4563 struct mem_cgroup
*memcg
= NULL
;
4564 struct page_cgroup
*pc
;
4565 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4567 if (mem_cgroup_disabled())
4570 pc
= lookup_page_cgroup(oldpage
);
4571 /* fix accounting on old pages */
4572 lock_page_cgroup(pc
);
4573 if (PageCgroupUsed(pc
)) {
4574 memcg
= pc
->mem_cgroup
;
4575 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4576 ClearPageCgroupUsed(pc
);
4578 unlock_page_cgroup(pc
);
4581 * When called from shmem_replace_page(), in some cases the
4582 * oldpage has already been charged, and in some cases not.
4587 * Even if newpage->mapping was NULL before starting replacement,
4588 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4589 * LRU while we overwrite pc->mem_cgroup.
4591 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4594 #ifdef CONFIG_DEBUG_VM
4595 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4597 struct page_cgroup
*pc
;
4599 pc
= lookup_page_cgroup(page
);
4601 * Can be NULL while feeding pages into the page allocator for
4602 * the first time, i.e. during boot or memory hotplug;
4603 * or when mem_cgroup_disabled().
4605 if (likely(pc
) && PageCgroupUsed(pc
))
4610 bool mem_cgroup_bad_page_check(struct page
*page
)
4612 if (mem_cgroup_disabled())
4615 return lookup_page_cgroup_used(page
) != NULL
;
4618 void mem_cgroup_print_bad_page(struct page
*page
)
4620 struct page_cgroup
*pc
;
4622 pc
= lookup_page_cgroup_used(page
);
4624 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4625 pc
, pc
->flags
, pc
->mem_cgroup
);
4630 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4631 unsigned long long val
)
4634 u64 memswlimit
, memlimit
;
4636 int children
= mem_cgroup_count_children(memcg
);
4637 u64 curusage
, oldusage
;
4641 * For keeping hierarchical_reclaim simple, how long we should retry
4642 * is depends on callers. We set our retry-count to be function
4643 * of # of children which we should visit in this loop.
4645 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4647 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4650 while (retry_count
) {
4651 if (signal_pending(current
)) {
4656 * Rather than hide all in some function, I do this in
4657 * open coded manner. You see what this really does.
4658 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4660 mutex_lock(&set_limit_mutex
);
4661 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4662 if (memswlimit
< val
) {
4664 mutex_unlock(&set_limit_mutex
);
4668 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4672 ret
= res_counter_set_limit(&memcg
->res
, val
);
4674 if (memswlimit
== val
)
4675 memcg
->memsw_is_minimum
= true;
4677 memcg
->memsw_is_minimum
= false;
4679 mutex_unlock(&set_limit_mutex
);
4684 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4685 MEM_CGROUP_RECLAIM_SHRINK
);
4686 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4687 /* Usage is reduced ? */
4688 if (curusage
>= oldusage
)
4691 oldusage
= curusage
;
4693 if (!ret
&& enlarge
)
4694 memcg_oom_recover(memcg
);
4699 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4700 unsigned long long val
)
4703 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4704 int children
= mem_cgroup_count_children(memcg
);
4708 /* see mem_cgroup_resize_res_limit */
4709 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4710 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4711 while (retry_count
) {
4712 if (signal_pending(current
)) {
4717 * Rather than hide all in some function, I do this in
4718 * open coded manner. You see what this really does.
4719 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4721 mutex_lock(&set_limit_mutex
);
4722 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4723 if (memlimit
> val
) {
4725 mutex_unlock(&set_limit_mutex
);
4728 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4729 if (memswlimit
< val
)
4731 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4733 if (memlimit
== val
)
4734 memcg
->memsw_is_minimum
= true;
4736 memcg
->memsw_is_minimum
= false;
4738 mutex_unlock(&set_limit_mutex
);
4743 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4744 MEM_CGROUP_RECLAIM_NOSWAP
|
4745 MEM_CGROUP_RECLAIM_SHRINK
);
4746 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4747 /* Usage is reduced ? */
4748 if (curusage
>= oldusage
)
4751 oldusage
= curusage
;
4753 if (!ret
&& enlarge
)
4754 memcg_oom_recover(memcg
);
4758 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4760 unsigned long *total_scanned
)
4762 unsigned long nr_reclaimed
= 0;
4763 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4764 unsigned long reclaimed
;
4766 struct mem_cgroup_tree_per_zone
*mctz
;
4767 unsigned long long excess
;
4768 unsigned long nr_scanned
;
4773 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4775 * This loop can run a while, specially if mem_cgroup's continuously
4776 * keep exceeding their soft limit and putting the system under
4783 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4788 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4789 gfp_mask
, &nr_scanned
);
4790 nr_reclaimed
+= reclaimed
;
4791 *total_scanned
+= nr_scanned
;
4792 spin_lock(&mctz
->lock
);
4795 * If we failed to reclaim anything from this memory cgroup
4796 * it is time to move on to the next cgroup
4802 * Loop until we find yet another one.
4804 * By the time we get the soft_limit lock
4805 * again, someone might have aded the
4806 * group back on the RB tree. Iterate to
4807 * make sure we get a different mem.
4808 * mem_cgroup_largest_soft_limit_node returns
4809 * NULL if no other cgroup is present on
4813 __mem_cgroup_largest_soft_limit_node(mctz
);
4815 css_put(&next_mz
->memcg
->css
);
4816 else /* next_mz == NULL or other memcg */
4820 __mem_cgroup_remove_exceeded(mz
->memcg
, mz
, mctz
);
4821 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4823 * One school of thought says that we should not add
4824 * back the node to the tree if reclaim returns 0.
4825 * But our reclaim could return 0, simply because due
4826 * to priority we are exposing a smaller subset of
4827 * memory to reclaim from. Consider this as a longer
4830 /* If excess == 0, no tree ops */
4831 __mem_cgroup_insert_exceeded(mz
->memcg
, mz
, mctz
, excess
);
4832 spin_unlock(&mctz
->lock
);
4833 css_put(&mz
->memcg
->css
);
4836 * Could not reclaim anything and there are no more
4837 * mem cgroups to try or we seem to be looping without
4838 * reclaiming anything.
4840 if (!nr_reclaimed
&&
4842 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4844 } while (!nr_reclaimed
);
4846 css_put(&next_mz
->memcg
->css
);
4847 return nr_reclaimed
;
4851 * mem_cgroup_force_empty_list - clears LRU of a group
4852 * @memcg: group to clear
4855 * @lru: lru to to clear
4857 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4858 * reclaim the pages page themselves - pages are moved to the parent (or root)
4861 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4862 int node
, int zid
, enum lru_list lru
)
4864 struct lruvec
*lruvec
;
4865 unsigned long flags
;
4866 struct list_head
*list
;
4870 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4871 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4872 list
= &lruvec
->lists
[lru
];
4876 struct page_cgroup
*pc
;
4879 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4880 if (list_empty(list
)) {
4881 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4884 page
= list_entry(list
->prev
, struct page
, lru
);
4886 list_move(&page
->lru
, list
);
4888 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4891 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4893 pc
= lookup_page_cgroup(page
);
4895 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4896 /* found lock contention or "pc" is obsolete. */
4901 } while (!list_empty(list
));
4905 * make mem_cgroup's charge to be 0 if there is no task by moving
4906 * all the charges and pages to the parent.
4907 * This enables deleting this mem_cgroup.
4909 * Caller is responsible for holding css reference on the memcg.
4911 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4917 /* This is for making all *used* pages to be on LRU. */
4918 lru_add_drain_all();
4919 drain_all_stock_sync(memcg
);
4920 mem_cgroup_start_move(memcg
);
4921 for_each_node_state(node
, N_MEMORY
) {
4922 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4925 mem_cgroup_force_empty_list(memcg
,
4930 mem_cgroup_end_move(memcg
);
4931 memcg_oom_recover(memcg
);
4935 * Kernel memory may not necessarily be trackable to a specific
4936 * process. So they are not migrated, and therefore we can't
4937 * expect their value to drop to 0 here.
4938 * Having res filled up with kmem only is enough.
4940 * This is a safety check because mem_cgroup_force_empty_list
4941 * could have raced with mem_cgroup_replace_page_cache callers
4942 * so the lru seemed empty but the page could have been added
4943 * right after the check. RES_USAGE should be safe as we always
4944 * charge before adding to the LRU.
4946 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4947 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4948 } while (usage
> 0);
4952 * This mainly exists for tests during the setting of set of use_hierarchy.
4953 * Since this is the very setting we are changing, the current hierarchy value
4956 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4958 struct cgroup_subsys_state
*pos
;
4960 /* bounce at first found */
4961 css_for_each_child(pos
, &memcg
->css
)
4967 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4968 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4969 * from mem_cgroup_count_children(), in the sense that we don't really care how
4970 * many children we have; we only need to know if we have any. It also counts
4971 * any memcg without hierarchy as infertile.
4973 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4975 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4979 * Reclaims as many pages from the given memcg as possible and moves
4980 * the rest to the parent.
4982 * Caller is responsible for holding css reference for memcg.
4984 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4986 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4987 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4989 /* returns EBUSY if there is a task or if we come here twice. */
4990 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4993 /* we call try-to-free pages for make this cgroup empty */
4994 lru_add_drain_all();
4995 /* try to free all pages in this cgroup */
4996 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4999 if (signal_pending(current
))
5002 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
5006 /* maybe some writeback is necessary */
5007 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
5012 mem_cgroup_reparent_charges(memcg
);
5017 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
5020 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5022 if (mem_cgroup_is_root(memcg
))
5024 return mem_cgroup_force_empty(memcg
);
5027 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
5030 return mem_cgroup_from_css(css
)->use_hierarchy
;
5033 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
5034 struct cftype
*cft
, u64 val
)
5037 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5038 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5040 mutex_lock(&memcg_create_mutex
);
5042 if (memcg
->use_hierarchy
== val
)
5046 * If parent's use_hierarchy is set, we can't make any modifications
5047 * in the child subtrees. If it is unset, then the change can
5048 * occur, provided the current cgroup has no children.
5050 * For the root cgroup, parent_mem is NULL, we allow value to be
5051 * set if there are no children.
5053 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
5054 (val
== 1 || val
== 0)) {
5055 if (!__memcg_has_children(memcg
))
5056 memcg
->use_hierarchy
= val
;
5063 mutex_unlock(&memcg_create_mutex
);
5069 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
5070 enum mem_cgroup_stat_index idx
)
5072 struct mem_cgroup
*iter
;
5075 /* Per-cpu values can be negative, use a signed accumulator */
5076 for_each_mem_cgroup_tree(iter
, memcg
)
5077 val
+= mem_cgroup_read_stat(iter
, idx
);
5079 if (val
< 0) /* race ? */
5084 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
5088 if (!mem_cgroup_is_root(memcg
)) {
5090 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5092 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5096 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5097 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5099 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
5100 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
5103 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
5105 return val
<< PAGE_SHIFT
;
5108 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
5109 struct cftype
*cft
, struct file
*file
,
5110 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
5112 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5118 type
= MEMFILE_TYPE(cft
->private);
5119 name
= MEMFILE_ATTR(cft
->private);
5123 if (name
== RES_USAGE
)
5124 val
= mem_cgroup_usage(memcg
, false);
5126 val
= res_counter_read_u64(&memcg
->res
, name
);
5129 if (name
== RES_USAGE
)
5130 val
= mem_cgroup_usage(memcg
, true);
5132 val
= res_counter_read_u64(&memcg
->memsw
, name
);
5135 val
= res_counter_read_u64(&memcg
->kmem
, name
);
5141 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
5142 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
5145 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
5148 #ifdef CONFIG_MEMCG_KMEM
5149 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5151 * For simplicity, we won't allow this to be disabled. It also can't
5152 * be changed if the cgroup has children already, or if tasks had
5155 * If tasks join before we set the limit, a person looking at
5156 * kmem.usage_in_bytes will have no way to determine when it took
5157 * place, which makes the value quite meaningless.
5159 * After it first became limited, changes in the value of the limit are
5160 * of course permitted.
5162 mutex_lock(&memcg_create_mutex
);
5163 mutex_lock(&set_limit_mutex
);
5164 if (!memcg
->kmem_account_flags
&& val
!= RES_COUNTER_MAX
) {
5165 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
5169 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5172 ret
= memcg_update_cache_sizes(memcg
);
5174 res_counter_set_limit(&memcg
->kmem
, RES_COUNTER_MAX
);
5177 static_key_slow_inc(&memcg_kmem_enabled_key
);
5179 * setting the active bit after the inc will guarantee no one
5180 * starts accounting before all call sites are patched
5182 memcg_kmem_set_active(memcg
);
5184 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
5186 mutex_unlock(&set_limit_mutex
);
5187 mutex_unlock(&memcg_create_mutex
);
5192 #ifdef CONFIG_MEMCG_KMEM
5193 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5196 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5200 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5202 * When that happen, we need to disable the static branch only on those
5203 * memcgs that enabled it. To achieve this, we would be forced to
5204 * complicate the code by keeping track of which memcgs were the ones
5205 * that actually enabled limits, and which ones got it from its
5208 * It is a lot simpler just to do static_key_slow_inc() on every child
5209 * that is accounted.
5211 if (!memcg_kmem_is_active(memcg
))
5215 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5216 * memcg is active already. If the later initialization fails then the
5217 * cgroup core triggers the cleanup so we do not have to do it here.
5219 static_key_slow_inc(&memcg_kmem_enabled_key
);
5221 mutex_lock(&set_limit_mutex
);
5222 memcg_stop_kmem_account();
5223 ret
= memcg_update_cache_sizes(memcg
);
5224 memcg_resume_kmem_account();
5225 mutex_unlock(&set_limit_mutex
);
5229 #endif /* CONFIG_MEMCG_KMEM */
5232 * The user of this function is...
5235 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5238 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5241 unsigned long long val
;
5244 type
= MEMFILE_TYPE(cft
->private);
5245 name
= MEMFILE_ATTR(cft
->private);
5249 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5253 /* This function does all necessary parse...reuse it */
5254 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5258 ret
= mem_cgroup_resize_limit(memcg
, val
);
5259 else if (type
== _MEMSWAP
)
5260 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5261 else if (type
== _KMEM
)
5262 ret
= memcg_update_kmem_limit(css
, val
);
5266 case RES_SOFT_LIMIT
:
5267 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5271 * For memsw, soft limits are hard to implement in terms
5272 * of semantics, for now, we support soft limits for
5273 * control without swap
5276 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5281 ret
= -EINVAL
; /* should be BUG() ? */
5287 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5288 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5290 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5292 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5293 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5294 if (!memcg
->use_hierarchy
)
5297 while (css_parent(&memcg
->css
)) {
5298 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5299 if (!memcg
->use_hierarchy
)
5301 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5302 min_limit
= min(min_limit
, tmp
);
5303 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5304 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5307 *mem_limit
= min_limit
;
5308 *memsw_limit
= min_memsw_limit
;
5311 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5313 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5317 type
= MEMFILE_TYPE(event
);
5318 name
= MEMFILE_ATTR(event
);
5323 res_counter_reset_max(&memcg
->res
);
5324 else if (type
== _MEMSWAP
)
5325 res_counter_reset_max(&memcg
->memsw
);
5326 else if (type
== _KMEM
)
5327 res_counter_reset_max(&memcg
->kmem
);
5333 res_counter_reset_failcnt(&memcg
->res
);
5334 else if (type
== _MEMSWAP
)
5335 res_counter_reset_failcnt(&memcg
->memsw
);
5336 else if (type
== _KMEM
)
5337 res_counter_reset_failcnt(&memcg
->kmem
);
5346 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5349 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5353 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5354 struct cftype
*cft
, u64 val
)
5356 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5358 if (val
>= (1 << NR_MOVE_TYPE
))
5362 * No kind of locking is needed in here, because ->can_attach() will
5363 * check this value once in the beginning of the process, and then carry
5364 * on with stale data. This means that changes to this value will only
5365 * affect task migrations starting after the change.
5367 memcg
->move_charge_at_immigrate
= val
;
5371 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5372 struct cftype
*cft
, u64 val
)
5379 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5380 struct cftype
*cft
, struct seq_file
*m
)
5383 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5384 unsigned long node_nr
;
5385 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5387 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5388 seq_printf(m
, "total=%lu", total_nr
);
5389 for_each_node_state(nid
, N_MEMORY
) {
5390 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5391 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5395 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5396 seq_printf(m
, "file=%lu", file_nr
);
5397 for_each_node_state(nid
, N_MEMORY
) {
5398 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5400 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5404 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5405 seq_printf(m
, "anon=%lu", anon_nr
);
5406 for_each_node_state(nid
, N_MEMORY
) {
5407 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5409 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5413 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5414 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5415 for_each_node_state(nid
, N_MEMORY
) {
5416 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5417 BIT(LRU_UNEVICTABLE
));
5418 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5423 #endif /* CONFIG_NUMA */
5425 static inline void mem_cgroup_lru_names_not_uptodate(void)
5427 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5430 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5433 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5434 struct mem_cgroup
*mi
;
5437 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5438 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5440 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5441 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5444 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5445 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5446 mem_cgroup_read_events(memcg
, i
));
5448 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5449 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5450 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5452 /* Hierarchical information */
5454 unsigned long long limit
, memsw_limit
;
5455 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5456 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5457 if (do_swap_account
)
5458 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5462 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5465 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5467 for_each_mem_cgroup_tree(mi
, memcg
)
5468 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5469 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5472 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5473 unsigned long long val
= 0;
5475 for_each_mem_cgroup_tree(mi
, memcg
)
5476 val
+= mem_cgroup_read_events(mi
, i
);
5477 seq_printf(m
, "total_%s %llu\n",
5478 mem_cgroup_events_names
[i
], val
);
5481 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5482 unsigned long long val
= 0;
5484 for_each_mem_cgroup_tree(mi
, memcg
)
5485 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5486 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5489 #ifdef CONFIG_DEBUG_VM
5492 struct mem_cgroup_per_zone
*mz
;
5493 struct zone_reclaim_stat
*rstat
;
5494 unsigned long recent_rotated
[2] = {0, 0};
5495 unsigned long recent_scanned
[2] = {0, 0};
5497 for_each_online_node(nid
)
5498 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5499 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5500 rstat
= &mz
->lruvec
.reclaim_stat
;
5502 recent_rotated
[0] += rstat
->recent_rotated
[0];
5503 recent_rotated
[1] += rstat
->recent_rotated
[1];
5504 recent_scanned
[0] += rstat
->recent_scanned
[0];
5505 recent_scanned
[1] += rstat
->recent_scanned
[1];
5507 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5508 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5509 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5510 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5517 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5520 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5522 return mem_cgroup_swappiness(memcg
);
5525 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5526 struct cftype
*cft
, u64 val
)
5528 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5529 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5531 if (val
> 100 || !parent
)
5534 mutex_lock(&memcg_create_mutex
);
5536 /* If under hierarchy, only empty-root can set this value */
5537 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5538 mutex_unlock(&memcg_create_mutex
);
5542 memcg
->swappiness
= val
;
5544 mutex_unlock(&memcg_create_mutex
);
5549 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5551 struct mem_cgroup_threshold_ary
*t
;
5557 t
= rcu_dereference(memcg
->thresholds
.primary
);
5559 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5564 usage
= mem_cgroup_usage(memcg
, swap
);
5567 * current_threshold points to threshold just below or equal to usage.
5568 * If it's not true, a threshold was crossed after last
5569 * call of __mem_cgroup_threshold().
5571 i
= t
->current_threshold
;
5574 * Iterate backward over array of thresholds starting from
5575 * current_threshold and check if a threshold is crossed.
5576 * If none of thresholds below usage is crossed, we read
5577 * only one element of the array here.
5579 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5580 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5582 /* i = current_threshold + 1 */
5586 * Iterate forward over array of thresholds starting from
5587 * current_threshold+1 and check if a threshold is crossed.
5588 * If none of thresholds above usage is crossed, we read
5589 * only one element of the array here.
5591 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5592 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5594 /* Update current_threshold */
5595 t
->current_threshold
= i
- 1;
5600 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5603 __mem_cgroup_threshold(memcg
, false);
5604 if (do_swap_account
)
5605 __mem_cgroup_threshold(memcg
, true);
5607 memcg
= parent_mem_cgroup(memcg
);
5611 static int compare_thresholds(const void *a
, const void *b
)
5613 const struct mem_cgroup_threshold
*_a
= a
;
5614 const struct mem_cgroup_threshold
*_b
= b
;
5616 if (_a
->threshold
> _b
->threshold
)
5619 if (_a
->threshold
< _b
->threshold
)
5625 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5627 struct mem_cgroup_eventfd_list
*ev
;
5629 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5630 eventfd_signal(ev
->eventfd
, 1);
5634 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5636 struct mem_cgroup
*iter
;
5638 for_each_mem_cgroup_tree(iter
, memcg
)
5639 mem_cgroup_oom_notify_cb(iter
);
5642 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5643 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5645 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5646 struct mem_cgroup_thresholds
*thresholds
;
5647 struct mem_cgroup_threshold_ary
*new;
5648 enum res_type type
= MEMFILE_TYPE(cft
->private);
5649 u64 threshold
, usage
;
5652 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5656 mutex_lock(&memcg
->thresholds_lock
);
5659 thresholds
= &memcg
->thresholds
;
5660 else if (type
== _MEMSWAP
)
5661 thresholds
= &memcg
->memsw_thresholds
;
5665 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5667 /* Check if a threshold crossed before adding a new one */
5668 if (thresholds
->primary
)
5669 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5671 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5673 /* Allocate memory for new array of thresholds */
5674 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5682 /* Copy thresholds (if any) to new array */
5683 if (thresholds
->primary
) {
5684 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5685 sizeof(struct mem_cgroup_threshold
));
5688 /* Add new threshold */
5689 new->entries
[size
- 1].eventfd
= eventfd
;
5690 new->entries
[size
- 1].threshold
= threshold
;
5692 /* Sort thresholds. Registering of new threshold isn't time-critical */
5693 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5694 compare_thresholds
, NULL
);
5696 /* Find current threshold */
5697 new->current_threshold
= -1;
5698 for (i
= 0; i
< size
; i
++) {
5699 if (new->entries
[i
].threshold
<= usage
) {
5701 * new->current_threshold will not be used until
5702 * rcu_assign_pointer(), so it's safe to increment
5705 ++new->current_threshold
;
5710 /* Free old spare buffer and save old primary buffer as spare */
5711 kfree(thresholds
->spare
);
5712 thresholds
->spare
= thresholds
->primary
;
5714 rcu_assign_pointer(thresholds
->primary
, new);
5716 /* To be sure that nobody uses thresholds */
5720 mutex_unlock(&memcg
->thresholds_lock
);
5725 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5726 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5728 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5729 struct mem_cgroup_thresholds
*thresholds
;
5730 struct mem_cgroup_threshold_ary
*new;
5731 enum res_type type
= MEMFILE_TYPE(cft
->private);
5735 mutex_lock(&memcg
->thresholds_lock
);
5737 thresholds
= &memcg
->thresholds
;
5738 else if (type
== _MEMSWAP
)
5739 thresholds
= &memcg
->memsw_thresholds
;
5743 if (!thresholds
->primary
)
5746 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5748 /* Check if a threshold crossed before removing */
5749 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5751 /* Calculate new number of threshold */
5753 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5754 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5758 new = thresholds
->spare
;
5760 /* Set thresholds array to NULL if we don't have thresholds */
5769 /* Copy thresholds and find current threshold */
5770 new->current_threshold
= -1;
5771 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5772 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5775 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5776 if (new->entries
[j
].threshold
<= usage
) {
5778 * new->current_threshold will not be used
5779 * until rcu_assign_pointer(), so it's safe to increment
5782 ++new->current_threshold
;
5788 /* Swap primary and spare array */
5789 thresholds
->spare
= thresholds
->primary
;
5790 /* If all events are unregistered, free the spare array */
5792 kfree(thresholds
->spare
);
5793 thresholds
->spare
= NULL
;
5796 rcu_assign_pointer(thresholds
->primary
, new);
5798 /* To be sure that nobody uses thresholds */
5801 mutex_unlock(&memcg
->thresholds_lock
);
5804 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5805 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5807 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5808 struct mem_cgroup_eventfd_list
*event
;
5809 enum res_type type
= MEMFILE_TYPE(cft
->private);
5811 BUG_ON(type
!= _OOM_TYPE
);
5812 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5816 spin_lock(&memcg_oom_lock
);
5818 event
->eventfd
= eventfd
;
5819 list_add(&event
->list
, &memcg
->oom_notify
);
5821 /* already in OOM ? */
5822 if (atomic_read(&memcg
->under_oom
))
5823 eventfd_signal(eventfd
, 1);
5824 spin_unlock(&memcg_oom_lock
);
5829 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5830 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5832 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5833 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5834 enum res_type type
= MEMFILE_TYPE(cft
->private);
5836 BUG_ON(type
!= _OOM_TYPE
);
5838 spin_lock(&memcg_oom_lock
);
5840 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5841 if (ev
->eventfd
== eventfd
) {
5842 list_del(&ev
->list
);
5847 spin_unlock(&memcg_oom_lock
);
5850 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5851 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5853 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5855 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5857 if (atomic_read(&memcg
->under_oom
))
5858 cb
->fill(cb
, "under_oom", 1);
5860 cb
->fill(cb
, "under_oom", 0);
5864 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5865 struct cftype
*cft
, u64 val
)
5867 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5868 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5870 /* cannot set to root cgroup and only 0 and 1 are allowed */
5871 if (!parent
|| !((val
== 0) || (val
== 1)))
5874 mutex_lock(&memcg_create_mutex
);
5875 /* oom-kill-disable is a flag for subhierarchy. */
5876 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5877 mutex_unlock(&memcg_create_mutex
);
5880 memcg
->oom_kill_disable
= val
;
5882 memcg_oom_recover(memcg
);
5883 mutex_unlock(&memcg_create_mutex
);
5887 #ifdef CONFIG_MEMCG_KMEM
5888 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5892 memcg
->kmemcg_id
= -1;
5893 ret
= memcg_propagate_kmem(memcg
);
5897 return mem_cgroup_sockets_init(memcg
, ss
);
5900 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5902 mem_cgroup_sockets_destroy(memcg
);
5905 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5907 if (!memcg_kmem_is_active(memcg
))
5911 * kmem charges can outlive the cgroup. In the case of slab
5912 * pages, for instance, a page contain objects from various
5913 * processes. As we prevent from taking a reference for every
5914 * such allocation we have to be careful when doing uncharge
5915 * (see memcg_uncharge_kmem) and here during offlining.
5917 * The idea is that that only the _last_ uncharge which sees
5918 * the dead memcg will drop the last reference. An additional
5919 * reference is taken here before the group is marked dead
5920 * which is then paired with css_put during uncharge resp. here.
5922 * Although this might sound strange as this path is called from
5923 * css_offline() when the referencemight have dropped down to 0
5924 * and shouldn't be incremented anymore (css_tryget would fail)
5925 * we do not have other options because of the kmem allocations
5928 css_get(&memcg
->css
);
5930 memcg_kmem_mark_dead(memcg
);
5932 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5935 if (memcg_kmem_test_and_clear_dead(memcg
))
5936 css_put(&memcg
->css
);
5939 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5944 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5948 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5953 static struct cftype mem_cgroup_files
[] = {
5955 .name
= "usage_in_bytes",
5956 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5957 .read
= mem_cgroup_read
,
5958 .register_event
= mem_cgroup_usage_register_event
,
5959 .unregister_event
= mem_cgroup_usage_unregister_event
,
5962 .name
= "max_usage_in_bytes",
5963 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5964 .trigger
= mem_cgroup_reset
,
5965 .read
= mem_cgroup_read
,
5968 .name
= "limit_in_bytes",
5969 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5970 .write_string
= mem_cgroup_write
,
5971 .read
= mem_cgroup_read
,
5974 .name
= "soft_limit_in_bytes",
5975 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5976 .write_string
= mem_cgroup_write
,
5977 .read
= mem_cgroup_read
,
5981 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5982 .trigger
= mem_cgroup_reset
,
5983 .read
= mem_cgroup_read
,
5987 .read_seq_string
= memcg_stat_show
,
5990 .name
= "force_empty",
5991 .trigger
= mem_cgroup_force_empty_write
,
5994 .name
= "use_hierarchy",
5995 .flags
= CFTYPE_INSANE
,
5996 .write_u64
= mem_cgroup_hierarchy_write
,
5997 .read_u64
= mem_cgroup_hierarchy_read
,
6000 .name
= "swappiness",
6001 .read_u64
= mem_cgroup_swappiness_read
,
6002 .write_u64
= mem_cgroup_swappiness_write
,
6005 .name
= "move_charge_at_immigrate",
6006 .read_u64
= mem_cgroup_move_charge_read
,
6007 .write_u64
= mem_cgroup_move_charge_write
,
6010 .name
= "oom_control",
6011 .read_map
= mem_cgroup_oom_control_read
,
6012 .write_u64
= mem_cgroup_oom_control_write
,
6013 .register_event
= mem_cgroup_oom_register_event
,
6014 .unregister_event
= mem_cgroup_oom_unregister_event
,
6015 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
6018 .name
= "pressure_level",
6019 .register_event
= vmpressure_register_event
,
6020 .unregister_event
= vmpressure_unregister_event
,
6024 .name
= "numa_stat",
6025 .read_seq_string
= memcg_numa_stat_show
,
6028 #ifdef CONFIG_MEMCG_KMEM
6030 .name
= "kmem.limit_in_bytes",
6031 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
6032 .write_string
= mem_cgroup_write
,
6033 .read
= mem_cgroup_read
,
6036 .name
= "kmem.usage_in_bytes",
6037 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
6038 .read
= mem_cgroup_read
,
6041 .name
= "kmem.failcnt",
6042 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
6043 .trigger
= mem_cgroup_reset
,
6044 .read
= mem_cgroup_read
,
6047 .name
= "kmem.max_usage_in_bytes",
6048 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
6049 .trigger
= mem_cgroup_reset
,
6050 .read
= mem_cgroup_read
,
6052 #ifdef CONFIG_SLABINFO
6054 .name
= "kmem.slabinfo",
6055 .read_seq_string
= mem_cgroup_slabinfo_read
,
6059 { }, /* terminate */
6062 #ifdef CONFIG_MEMCG_SWAP
6063 static struct cftype memsw_cgroup_files
[] = {
6065 .name
= "memsw.usage_in_bytes",
6066 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
6067 .read
= mem_cgroup_read
,
6068 .register_event
= mem_cgroup_usage_register_event
,
6069 .unregister_event
= mem_cgroup_usage_unregister_event
,
6072 .name
= "memsw.max_usage_in_bytes",
6073 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
6074 .trigger
= mem_cgroup_reset
,
6075 .read
= mem_cgroup_read
,
6078 .name
= "memsw.limit_in_bytes",
6079 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
6080 .write_string
= mem_cgroup_write
,
6081 .read
= mem_cgroup_read
,
6084 .name
= "memsw.failcnt",
6085 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
6086 .trigger
= mem_cgroup_reset
,
6087 .read
= mem_cgroup_read
,
6089 { }, /* terminate */
6092 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6094 struct mem_cgroup_per_node
*pn
;
6095 struct mem_cgroup_per_zone
*mz
;
6096 int zone
, tmp
= node
;
6098 * This routine is called against possible nodes.
6099 * But it's BUG to call kmalloc() against offline node.
6101 * TODO: this routine can waste much memory for nodes which will
6102 * never be onlined. It's better to use memory hotplug callback
6105 if (!node_state(node
, N_NORMAL_MEMORY
))
6107 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6111 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6112 mz
= &pn
->zoneinfo
[zone
];
6113 lruvec_init(&mz
->lruvec
);
6114 mz
->usage_in_excess
= 0;
6115 mz
->on_tree
= false;
6118 memcg
->nodeinfo
[node
] = pn
;
6122 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6124 kfree(memcg
->nodeinfo
[node
]);
6127 static struct mem_cgroup
*mem_cgroup_alloc(void)
6129 struct mem_cgroup
*memcg
;
6130 size_t size
= memcg_size();
6132 /* Can be very big if nr_node_ids is very big */
6133 if (size
< PAGE_SIZE
)
6134 memcg
= kzalloc(size
, GFP_KERNEL
);
6136 memcg
= vzalloc(size
);
6141 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6144 spin_lock_init(&memcg
->pcp_counter_lock
);
6148 if (size
< PAGE_SIZE
)
6156 * At destroying mem_cgroup, references from swap_cgroup can remain.
6157 * (scanning all at force_empty is too costly...)
6159 * Instead of clearing all references at force_empty, we remember
6160 * the number of reference from swap_cgroup and free mem_cgroup when
6161 * it goes down to 0.
6163 * Removal of cgroup itself succeeds regardless of refs from swap.
6166 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6169 size_t size
= memcg_size();
6171 mem_cgroup_remove_from_trees(memcg
);
6172 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
6175 free_mem_cgroup_per_zone_info(memcg
, node
);
6177 free_percpu(memcg
->stat
);
6180 * We need to make sure that (at least for now), the jump label
6181 * destruction code runs outside of the cgroup lock. This is because
6182 * get_online_cpus(), which is called from the static_branch update,
6183 * can't be called inside the cgroup_lock. cpusets are the ones
6184 * enforcing this dependency, so if they ever change, we might as well.
6186 * schedule_work() will guarantee this happens. Be careful if you need
6187 * to move this code around, and make sure it is outside
6190 disarm_static_keys(memcg
);
6191 if (size
< PAGE_SIZE
)
6198 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6200 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6202 if (!memcg
->res
.parent
)
6204 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6206 EXPORT_SYMBOL(parent_mem_cgroup
);
6208 static void __init
mem_cgroup_soft_limit_tree_init(void)
6210 struct mem_cgroup_tree_per_node
*rtpn
;
6211 struct mem_cgroup_tree_per_zone
*rtpz
;
6212 int tmp
, node
, zone
;
6214 for_each_node(node
) {
6216 if (!node_state(node
, N_NORMAL_MEMORY
))
6218 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6221 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6223 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6224 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6225 rtpz
->rb_root
= RB_ROOT
;
6226 spin_lock_init(&rtpz
->lock
);
6231 static struct cgroup_subsys_state
* __ref
6232 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6234 struct mem_cgroup
*memcg
;
6235 long error
= -ENOMEM
;
6238 memcg
= mem_cgroup_alloc();
6240 return ERR_PTR(error
);
6243 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6247 if (parent_css
== NULL
) {
6248 root_mem_cgroup
= memcg
;
6249 res_counter_init(&memcg
->res
, NULL
);
6250 res_counter_init(&memcg
->memsw
, NULL
);
6251 res_counter_init(&memcg
->kmem
, NULL
);
6254 memcg
->last_scanned_node
= MAX_NUMNODES
;
6255 INIT_LIST_HEAD(&memcg
->oom_notify
);
6256 memcg
->move_charge_at_immigrate
= 0;
6257 mutex_init(&memcg
->thresholds_lock
);
6258 spin_lock_init(&memcg
->move_lock
);
6259 vmpressure_init(&memcg
->vmpressure
);
6264 __mem_cgroup_free(memcg
);
6265 return ERR_PTR(error
);
6269 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6271 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6272 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6278 mutex_lock(&memcg_create_mutex
);
6280 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6281 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6282 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6284 if (parent
->use_hierarchy
) {
6285 res_counter_init(&memcg
->res
, &parent
->res
);
6286 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6287 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6290 * No need to take a reference to the parent because cgroup
6291 * core guarantees its existence.
6294 res_counter_init(&memcg
->res
, NULL
);
6295 res_counter_init(&memcg
->memsw
, NULL
);
6296 res_counter_init(&memcg
->kmem
, NULL
);
6298 * Deeper hierachy with use_hierarchy == false doesn't make
6299 * much sense so let cgroup subsystem know about this
6300 * unfortunate state in our controller.
6302 if (parent
!= root_mem_cgroup
)
6303 mem_cgroup_subsys
.broken_hierarchy
= true;
6306 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6307 mutex_unlock(&memcg_create_mutex
);
6312 * Announce all parents that a group from their hierarchy is gone.
6314 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6316 struct mem_cgroup
*parent
= memcg
;
6318 while ((parent
= parent_mem_cgroup(parent
)))
6319 mem_cgroup_iter_invalidate(parent
);
6322 * if the root memcg is not hierarchical we have to check it
6325 if (!root_mem_cgroup
->use_hierarchy
)
6326 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6329 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6331 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6333 kmem_cgroup_css_offline(memcg
);
6335 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6336 mem_cgroup_reparent_charges(memcg
);
6337 mem_cgroup_destroy_all_caches(memcg
);
6338 vmpressure_cleanup(&memcg
->vmpressure
);
6341 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6343 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6345 memcg_destroy_kmem(memcg
);
6346 __mem_cgroup_free(memcg
);
6350 /* Handlers for move charge at task migration. */
6351 #define PRECHARGE_COUNT_AT_ONCE 256
6352 static int mem_cgroup_do_precharge(unsigned long count
)
6355 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6356 struct mem_cgroup
*memcg
= mc
.to
;
6358 if (mem_cgroup_is_root(memcg
)) {
6359 mc
.precharge
+= count
;
6360 /* we don't need css_get for root */
6363 /* try to charge at once */
6365 struct res_counter
*dummy
;
6367 * "memcg" cannot be under rmdir() because we've already checked
6368 * by cgroup_lock_live_cgroup() that it is not removed and we
6369 * are still under the same cgroup_mutex. So we can postpone
6372 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6374 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6375 PAGE_SIZE
* count
, &dummy
)) {
6376 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6379 mc
.precharge
+= count
;
6383 /* fall back to one by one charge */
6385 if (signal_pending(current
)) {
6389 if (!batch_count
--) {
6390 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6393 ret
= __mem_cgroup_try_charge(NULL
,
6394 GFP_KERNEL
, 1, &memcg
, false);
6396 /* mem_cgroup_clear_mc() will do uncharge later */
6404 * get_mctgt_type - get target type of moving charge
6405 * @vma: the vma the pte to be checked belongs
6406 * @addr: the address corresponding to the pte to be checked
6407 * @ptent: the pte to be checked
6408 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6411 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6412 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6413 * move charge. if @target is not NULL, the page is stored in target->page
6414 * with extra refcnt got(Callers should handle it).
6415 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6416 * target for charge migration. if @target is not NULL, the entry is stored
6419 * Called with pte lock held.
6426 enum mc_target_type
{
6432 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6433 unsigned long addr
, pte_t ptent
)
6435 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6437 if (!page
|| !page_mapped(page
))
6439 if (PageAnon(page
)) {
6440 /* we don't move shared anon */
6443 } else if (!move_file())
6444 /* we ignore mapcount for file pages */
6446 if (!get_page_unless_zero(page
))
6453 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6454 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6456 struct page
*page
= NULL
;
6457 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6459 if (!move_anon() || non_swap_entry(ent
))
6462 * Because lookup_swap_cache() updates some statistics counter,
6463 * we call find_get_page() with swapper_space directly.
6465 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6466 if (do_swap_account
)
6467 entry
->val
= ent
.val
;
6472 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6473 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6479 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6480 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6482 struct page
*page
= NULL
;
6483 struct address_space
*mapping
;
6486 if (!vma
->vm_file
) /* anonymous vma */
6491 mapping
= vma
->vm_file
->f_mapping
;
6492 if (pte_none(ptent
))
6493 pgoff
= linear_page_index(vma
, addr
);
6494 else /* pte_file(ptent) is true */
6495 pgoff
= pte_to_pgoff(ptent
);
6497 /* page is moved even if it's not RSS of this task(page-faulted). */
6498 page
= find_get_page(mapping
, pgoff
);
6501 /* shmem/tmpfs may report page out on swap: account for that too. */
6502 if (radix_tree_exceptional_entry(page
)) {
6503 swp_entry_t swap
= radix_to_swp_entry(page
);
6504 if (do_swap_account
)
6506 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6512 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6513 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6515 struct page
*page
= NULL
;
6516 struct page_cgroup
*pc
;
6517 enum mc_target_type ret
= MC_TARGET_NONE
;
6518 swp_entry_t ent
= { .val
= 0 };
6520 if (pte_present(ptent
))
6521 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6522 else if (is_swap_pte(ptent
))
6523 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6524 else if (pte_none(ptent
) || pte_file(ptent
))
6525 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6527 if (!page
&& !ent
.val
)
6530 pc
= lookup_page_cgroup(page
);
6532 * Do only loose check w/o page_cgroup lock.
6533 * mem_cgroup_move_account() checks the pc is valid or not under
6536 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6537 ret
= MC_TARGET_PAGE
;
6539 target
->page
= page
;
6541 if (!ret
|| !target
)
6544 /* There is a swap entry and a page doesn't exist or isn't charged */
6545 if (ent
.val
&& !ret
&&
6546 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6547 ret
= MC_TARGET_SWAP
;
6554 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6556 * We don't consider swapping or file mapped pages because THP does not
6557 * support them for now.
6558 * Caller should make sure that pmd_trans_huge(pmd) is true.
6560 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6561 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6563 struct page
*page
= NULL
;
6564 struct page_cgroup
*pc
;
6565 enum mc_target_type ret
= MC_TARGET_NONE
;
6567 page
= pmd_page(pmd
);
6568 VM_BUG_ON(!page
|| !PageHead(page
));
6571 pc
= lookup_page_cgroup(page
);
6572 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6573 ret
= MC_TARGET_PAGE
;
6576 target
->page
= page
;
6582 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6583 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6585 return MC_TARGET_NONE
;
6589 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6590 unsigned long addr
, unsigned long end
,
6591 struct mm_walk
*walk
)
6593 struct vm_area_struct
*vma
= walk
->private;
6597 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6598 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6599 mc
.precharge
+= HPAGE_PMD_NR
;
6600 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6604 if (pmd_trans_unstable(pmd
))
6606 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6607 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6608 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6609 mc
.precharge
++; /* increment precharge temporarily */
6610 pte_unmap_unlock(pte
- 1, ptl
);
6616 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6618 unsigned long precharge
;
6619 struct vm_area_struct
*vma
;
6621 down_read(&mm
->mmap_sem
);
6622 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6623 struct mm_walk mem_cgroup_count_precharge_walk
= {
6624 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6628 if (is_vm_hugetlb_page(vma
))
6630 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6631 &mem_cgroup_count_precharge_walk
);
6633 up_read(&mm
->mmap_sem
);
6635 precharge
= mc
.precharge
;
6641 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6643 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6645 VM_BUG_ON(mc
.moving_task
);
6646 mc
.moving_task
= current
;
6647 return mem_cgroup_do_precharge(precharge
);
6650 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6651 static void __mem_cgroup_clear_mc(void)
6653 struct mem_cgroup
*from
= mc
.from
;
6654 struct mem_cgroup
*to
= mc
.to
;
6657 /* we must uncharge all the leftover precharges from mc.to */
6659 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6663 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6664 * we must uncharge here.
6666 if (mc
.moved_charge
) {
6667 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6668 mc
.moved_charge
= 0;
6670 /* we must fixup refcnts and charges */
6671 if (mc
.moved_swap
) {
6672 /* uncharge swap account from the old cgroup */
6673 if (!mem_cgroup_is_root(mc
.from
))
6674 res_counter_uncharge(&mc
.from
->memsw
,
6675 PAGE_SIZE
* mc
.moved_swap
);
6677 for (i
= 0; i
< mc
.moved_swap
; i
++)
6678 css_put(&mc
.from
->css
);
6680 if (!mem_cgroup_is_root(mc
.to
)) {
6682 * we charged both to->res and to->memsw, so we should
6685 res_counter_uncharge(&mc
.to
->res
,
6686 PAGE_SIZE
* mc
.moved_swap
);
6688 /* we've already done css_get(mc.to) */
6691 memcg_oom_recover(from
);
6692 memcg_oom_recover(to
);
6693 wake_up_all(&mc
.waitq
);
6696 static void mem_cgroup_clear_mc(void)
6698 struct mem_cgroup
*from
= mc
.from
;
6701 * we must clear moving_task before waking up waiters at the end of
6704 mc
.moving_task
= NULL
;
6705 __mem_cgroup_clear_mc();
6706 spin_lock(&mc
.lock
);
6709 spin_unlock(&mc
.lock
);
6710 mem_cgroup_end_move(from
);
6713 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6714 struct cgroup_taskset
*tset
)
6716 struct task_struct
*p
= cgroup_taskset_first(tset
);
6718 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6719 unsigned long move_charge_at_immigrate
;
6722 * We are now commited to this value whatever it is. Changes in this
6723 * tunable will only affect upcoming migrations, not the current one.
6724 * So we need to save it, and keep it going.
6726 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6727 if (move_charge_at_immigrate
) {
6728 struct mm_struct
*mm
;
6729 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6731 VM_BUG_ON(from
== memcg
);
6733 mm
= get_task_mm(p
);
6736 /* We move charges only when we move a owner of the mm */
6737 if (mm
->owner
== p
) {
6740 VM_BUG_ON(mc
.precharge
);
6741 VM_BUG_ON(mc
.moved_charge
);
6742 VM_BUG_ON(mc
.moved_swap
);
6743 mem_cgroup_start_move(from
);
6744 spin_lock(&mc
.lock
);
6747 mc
.immigrate_flags
= move_charge_at_immigrate
;
6748 spin_unlock(&mc
.lock
);
6749 /* We set mc.moving_task later */
6751 ret
= mem_cgroup_precharge_mc(mm
);
6753 mem_cgroup_clear_mc();
6760 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6761 struct cgroup_taskset
*tset
)
6763 mem_cgroup_clear_mc();
6766 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6767 unsigned long addr
, unsigned long end
,
6768 struct mm_walk
*walk
)
6771 struct vm_area_struct
*vma
= walk
->private;
6774 enum mc_target_type target_type
;
6775 union mc_target target
;
6777 struct page_cgroup
*pc
;
6780 * We don't take compound_lock() here but no race with splitting thp
6782 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6783 * under splitting, which means there's no concurrent thp split,
6784 * - if another thread runs into split_huge_page() just after we
6785 * entered this if-block, the thread must wait for page table lock
6786 * to be unlocked in __split_huge_page_splitting(), where the main
6787 * part of thp split is not executed yet.
6789 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6790 if (mc
.precharge
< HPAGE_PMD_NR
) {
6791 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6794 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6795 if (target_type
== MC_TARGET_PAGE
) {
6797 if (!isolate_lru_page(page
)) {
6798 pc
= lookup_page_cgroup(page
);
6799 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6800 pc
, mc
.from
, mc
.to
)) {
6801 mc
.precharge
-= HPAGE_PMD_NR
;
6802 mc
.moved_charge
+= HPAGE_PMD_NR
;
6804 putback_lru_page(page
);
6808 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6812 if (pmd_trans_unstable(pmd
))
6815 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6816 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6817 pte_t ptent
= *(pte
++);
6823 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6824 case MC_TARGET_PAGE
:
6826 if (isolate_lru_page(page
))
6828 pc
= lookup_page_cgroup(page
);
6829 if (!mem_cgroup_move_account(page
, 1, pc
,
6832 /* we uncharge from mc.from later. */
6835 putback_lru_page(page
);
6836 put
: /* get_mctgt_type() gets the page */
6839 case MC_TARGET_SWAP
:
6841 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6843 /* we fixup refcnts and charges later. */
6851 pte_unmap_unlock(pte
- 1, ptl
);
6856 * We have consumed all precharges we got in can_attach().
6857 * We try charge one by one, but don't do any additional
6858 * charges to mc.to if we have failed in charge once in attach()
6861 ret
= mem_cgroup_do_precharge(1);
6869 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6871 struct vm_area_struct
*vma
;
6873 lru_add_drain_all();
6875 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6877 * Someone who are holding the mmap_sem might be waiting in
6878 * waitq. So we cancel all extra charges, wake up all waiters,
6879 * and retry. Because we cancel precharges, we might not be able
6880 * to move enough charges, but moving charge is a best-effort
6881 * feature anyway, so it wouldn't be a big problem.
6883 __mem_cgroup_clear_mc();
6887 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6889 struct mm_walk mem_cgroup_move_charge_walk
= {
6890 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6894 if (is_vm_hugetlb_page(vma
))
6896 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6897 &mem_cgroup_move_charge_walk
);
6900 * means we have consumed all precharges and failed in
6901 * doing additional charge. Just abandon here.
6905 up_read(&mm
->mmap_sem
);
6908 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6909 struct cgroup_taskset
*tset
)
6911 struct task_struct
*p
= cgroup_taskset_first(tset
);
6912 struct mm_struct
*mm
= get_task_mm(p
);
6916 mem_cgroup_move_charge(mm
);
6920 mem_cgroup_clear_mc();
6922 #else /* !CONFIG_MMU */
6923 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6924 struct cgroup_taskset
*tset
)
6928 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6929 struct cgroup_taskset
*tset
)
6932 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6933 struct cgroup_taskset
*tset
)
6939 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6940 * to verify sane_behavior flag on each mount attempt.
6942 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6945 * use_hierarchy is forced with sane_behavior. cgroup core
6946 * guarantees that @root doesn't have any children, so turning it
6947 * on for the root memcg is enough.
6949 if (cgroup_sane_behavior(root_css
->cgroup
))
6950 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6953 struct cgroup_subsys mem_cgroup_subsys
= {
6955 .subsys_id
= mem_cgroup_subsys_id
,
6956 .css_alloc
= mem_cgroup_css_alloc
,
6957 .css_online
= mem_cgroup_css_online
,
6958 .css_offline
= mem_cgroup_css_offline
,
6959 .css_free
= mem_cgroup_css_free
,
6960 .can_attach
= mem_cgroup_can_attach
,
6961 .cancel_attach
= mem_cgroup_cancel_attach
,
6962 .attach
= mem_cgroup_move_task
,
6963 .bind
= mem_cgroup_bind
,
6964 .base_cftypes
= mem_cgroup_files
,
6969 #ifdef CONFIG_MEMCG_SWAP
6970 static int __init
enable_swap_account(char *s
)
6972 if (!strcmp(s
, "1"))
6973 really_do_swap_account
= 1;
6974 else if (!strcmp(s
, "0"))
6975 really_do_swap_account
= 0;
6978 __setup("swapaccount=", enable_swap_account
);
6980 static void __init
memsw_file_init(void)
6982 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6985 static void __init
enable_swap_cgroup(void)
6987 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6988 do_swap_account
= 1;
6994 static void __init
enable_swap_cgroup(void)
7000 * subsys_initcall() for memory controller.
7002 * Some parts like hotcpu_notifier() have to be initialized from this context
7003 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7004 * everything that doesn't depend on a specific mem_cgroup structure should
7005 * be initialized from here.
7007 static int __init
mem_cgroup_init(void)
7009 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
7010 enable_swap_cgroup();
7011 mem_cgroup_soft_limit_tree_init();
7015 subsys_initcall(mem_cgroup_init
);