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/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
70 EXPORT_SYMBOL(memory_cgrp_subsys
);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly
;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata
= 1;
83 static int really_do_swap_account __initdata
;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names
[] = {
100 enum mem_cgroup_events_index
{
101 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS
,
108 static const char * const mem_cgroup_events_names
[] = {
115 static const char * const mem_cgroup_lru_names
[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target
{
130 MEM_CGROUP_TARGET_THRESH
,
131 MEM_CGROUP_TARGET_SOFTLIMIT
,
132 MEM_CGROUP_TARGET_NUMAINFO
,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu
{
140 long count
[MEM_CGROUP_STAT_NSTATS
];
141 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
142 unsigned long nr_page_events
;
143 unsigned long targets
[MEM_CGROUP_NTARGETS
];
146 struct mem_cgroup_reclaim_iter
{
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup
*last_visited
;
154 /* scan generation, increased every round-trip */
155 unsigned int generation
;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone
{
162 struct lruvec lruvec
;
163 unsigned long lru_size
[NR_LRU_LISTS
];
165 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
167 struct rb_node tree_node
; /* RB tree node */
168 unsigned long long usage_in_excess
;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node
{
176 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone
{
185 struct rb_root rb_root
;
189 struct mem_cgroup_tree_per_node
{
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
193 struct mem_cgroup_tree
{
194 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
199 struct mem_cgroup_threshold
{
200 struct eventfd_ctx
*eventfd
;
205 struct mem_cgroup_threshold_ary
{
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold
;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries
[0];
214 struct mem_cgroup_thresholds
{
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary
*primary
;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary
*spare
;
226 struct mem_cgroup_eventfd_list
{
227 struct list_head list
;
228 struct eventfd_ctx
*eventfd
;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event
{
236 * memcg which the event belongs to.
238 struct mem_cgroup
*memcg
;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx
*eventfd
;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list
;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event
)(struct mem_cgroup
*memcg
,
253 struct eventfd_ctx
*eventfd
, const char *args
);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event
)(struct mem_cgroup
*memcg
,
260 struct eventfd_ctx
*eventfd
);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t
*wqh
;
268 struct work_struct remove
;
271 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
272 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css
;
288 * the counter to account for memory usage
290 struct res_counter res
;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure
;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw
;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem
;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups
;
315 /* OOM-Killer disable */
316 int oom_kill_disable
;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum
;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock
;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds
;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds
;
330 /* For oom notifier event fd */
331 struct list_head oom_notify
;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate
;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account
;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock
;
347 struct mem_cgroup_stat_cpu __percpu
*stat
;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base
;
353 spinlock_t pcp_counter_lock
;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem
;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list, but per-memcg;
361 * protected by memcg_slab_mutex */
362 struct list_head memcg_slab_caches
;
363 /* Index in the kmem_cache->memcg_params->memcg_caches array */
367 int last_scanned_node
;
369 nodemask_t scan_nodes
;
370 atomic_t numainfo_events
;
371 atomic_t numainfo_updating
;
374 /* List of events which userspace want to receive */
375 struct list_head event_list
;
376 spinlock_t event_list_lock
;
378 struct mem_cgroup_per_node
*nodeinfo
[0];
379 /* WARNING: nodeinfo must be the last member here */
382 /* internal only representation about the status of kmem accounting. */
384 KMEM_ACCOUNTED_ACTIVE
, /* accounted by this cgroup itself */
385 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
388 #ifdef CONFIG_MEMCG_KMEM
389 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
391 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
394 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
396 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
399 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
402 * Our caller must use css_get() first, because memcg_uncharge_kmem()
403 * will call css_put() if it sees the memcg is dead.
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
407 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
413 &memcg
->kmem_account_flags
);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct
{
430 spinlock_t lock
; /* for from, to */
431 struct mem_cgroup
*from
;
432 struct mem_cgroup
*to
;
433 unsigned long immigrate_flags
;
434 unsigned long precharge
;
435 unsigned long moved_charge
;
436 unsigned long moved_swap
;
437 struct task_struct
*moving_task
; /* a task moving charges */
438 wait_queue_head_t waitq
; /* a waitq for other context */
440 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
441 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON
,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
486 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
487 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
488 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
489 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
492 * The memcg_create_mutex will be held whenever a new cgroup is created.
493 * As a consequence, any change that needs to protect against new child cgroups
494 * appearing has to hold it as well.
496 static DEFINE_MUTEX(memcg_create_mutex
);
498 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
500 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
503 /* Some nice accessors for the vmpressure. */
504 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
507 memcg
= root_mem_cgroup
;
508 return &memcg
->vmpressure
;
511 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
513 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
516 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
518 return (memcg
== root_mem_cgroup
);
522 * We restrict the id in the range of [1, 65535], so it can fit into
525 #define MEM_CGROUP_ID_MAX USHRT_MAX
527 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
529 return memcg
->css
.id
;
532 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
534 struct cgroup_subsys_state
*css
;
536 css
= css_from_id(id
, &memory_cgrp_subsys
);
537 return mem_cgroup_from_css(css
);
540 /* Writing them here to avoid exposing memcg's inner layout */
541 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
543 void sock_update_memcg(struct sock
*sk
)
545 if (mem_cgroup_sockets_enabled
) {
546 struct mem_cgroup
*memcg
;
547 struct cg_proto
*cg_proto
;
549 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
551 /* Socket cloning can throw us here with sk_cgrp already
552 * filled. It won't however, necessarily happen from
553 * process context. So the test for root memcg given
554 * the current task's memcg won't help us in this case.
556 * Respecting the original socket's memcg is a better
557 * decision in this case.
560 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
561 css_get(&sk
->sk_cgrp
->memcg
->css
);
566 memcg
= mem_cgroup_from_task(current
);
567 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
568 if (!mem_cgroup_is_root(memcg
) &&
569 memcg_proto_active(cg_proto
) &&
570 css_tryget_online(&memcg
->css
)) {
571 sk
->sk_cgrp
= cg_proto
;
576 EXPORT_SYMBOL(sock_update_memcg
);
578 void sock_release_memcg(struct sock
*sk
)
580 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
581 struct mem_cgroup
*memcg
;
582 WARN_ON(!sk
->sk_cgrp
->memcg
);
583 memcg
= sk
->sk_cgrp
->memcg
;
584 css_put(&sk
->sk_cgrp
->memcg
->css
);
588 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
590 if (!memcg
|| mem_cgroup_is_root(memcg
))
593 return &memcg
->tcp_mem
;
595 EXPORT_SYMBOL(tcp_proto_cgroup
);
597 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
599 if (!memcg_proto_activated(&memcg
->tcp_mem
))
601 static_key_slow_dec(&memcg_socket_limit_enabled
);
604 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
609 #ifdef CONFIG_MEMCG_KMEM
611 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
612 * The main reason for not using cgroup id for this:
613 * this works better in sparse environments, where we have a lot of memcgs,
614 * but only a few kmem-limited. Or also, if we have, for instance, 200
615 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
616 * 200 entry array for that.
618 * The current size of the caches array is stored in
619 * memcg_limited_groups_array_size. It will double each time we have to
622 static DEFINE_IDA(kmem_limited_groups
);
623 int memcg_limited_groups_array_size
;
626 * MIN_SIZE is different than 1, because we would like to avoid going through
627 * the alloc/free process all the time. In a small machine, 4 kmem-limited
628 * cgroups is a reasonable guess. In the future, it could be a parameter or
629 * tunable, but that is strictly not necessary.
631 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
632 * this constant directly from cgroup, but it is understandable that this is
633 * better kept as an internal representation in cgroup.c. In any case, the
634 * cgrp_id space is not getting any smaller, and we don't have to necessarily
635 * increase ours as well if it increases.
637 #define MEMCG_CACHES_MIN_SIZE 4
638 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
641 * A lot of the calls to the cache allocation functions are expected to be
642 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
643 * conditional to this static branch, we'll have to allow modules that does
644 * kmem_cache_alloc and the such to see this symbol as well
646 struct static_key memcg_kmem_enabled_key
;
647 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
649 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
651 if (memcg_kmem_is_active(memcg
)) {
652 static_key_slow_dec(&memcg_kmem_enabled_key
);
653 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
656 * This check can't live in kmem destruction function,
657 * since the charges will outlive the cgroup
659 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
662 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
665 #endif /* CONFIG_MEMCG_KMEM */
667 static void disarm_static_keys(struct mem_cgroup
*memcg
)
669 disarm_sock_keys(memcg
);
670 disarm_kmem_keys(memcg
);
673 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
675 static struct mem_cgroup_per_zone
*
676 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
678 int nid
= zone_to_nid(zone
);
679 int zid
= zone_idx(zone
);
681 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
684 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
689 static struct mem_cgroup_per_zone
*
690 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
692 int nid
= page_to_nid(page
);
693 int zid
= page_zonenum(page
);
695 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
698 static struct mem_cgroup_tree_per_zone
*
699 soft_limit_tree_node_zone(int nid
, int zid
)
701 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
704 static struct mem_cgroup_tree_per_zone
*
705 soft_limit_tree_from_page(struct page
*page
)
707 int nid
= page_to_nid(page
);
708 int zid
= page_zonenum(page
);
710 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
713 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
714 struct mem_cgroup_tree_per_zone
*mctz
,
715 unsigned long long new_usage_in_excess
)
717 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
718 struct rb_node
*parent
= NULL
;
719 struct mem_cgroup_per_zone
*mz_node
;
724 mz
->usage_in_excess
= new_usage_in_excess
;
725 if (!mz
->usage_in_excess
)
729 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
731 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
734 * We can't avoid mem cgroups that are over their soft
735 * limit by the same amount
737 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
740 rb_link_node(&mz
->tree_node
, parent
, p
);
741 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
745 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
746 struct mem_cgroup_tree_per_zone
*mctz
)
750 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
754 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
755 struct mem_cgroup_tree_per_zone
*mctz
)
757 spin_lock(&mctz
->lock
);
758 __mem_cgroup_remove_exceeded(mz
, mctz
);
759 spin_unlock(&mctz
->lock
);
763 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
765 unsigned long long excess
;
766 struct mem_cgroup_per_zone
*mz
;
767 struct mem_cgroup_tree_per_zone
*mctz
;
769 mctz
= soft_limit_tree_from_page(page
);
771 * Necessary to update all ancestors when hierarchy is used.
772 * because their event counter is not touched.
774 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
775 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
776 excess
= res_counter_soft_limit_excess(&memcg
->res
);
778 * We have to update the tree if mz is on RB-tree or
779 * mem is over its softlimit.
781 if (excess
|| mz
->on_tree
) {
782 spin_lock(&mctz
->lock
);
783 /* if on-tree, remove it */
785 __mem_cgroup_remove_exceeded(mz
, mctz
);
787 * Insert again. mz->usage_in_excess will be updated.
788 * If excess is 0, no tree ops.
790 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
791 spin_unlock(&mctz
->lock
);
796 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
798 struct mem_cgroup_tree_per_zone
*mctz
;
799 struct mem_cgroup_per_zone
*mz
;
803 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
804 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
805 mctz
= soft_limit_tree_node_zone(nid
, zid
);
806 mem_cgroup_remove_exceeded(mz
, mctz
);
811 static struct mem_cgroup_per_zone
*
812 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
814 struct rb_node
*rightmost
= NULL
;
815 struct mem_cgroup_per_zone
*mz
;
819 rightmost
= rb_last(&mctz
->rb_root
);
821 goto done
; /* Nothing to reclaim from */
823 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
825 * Remove the node now but someone else can add it back,
826 * we will to add it back at the end of reclaim to its correct
827 * position in the tree.
829 __mem_cgroup_remove_exceeded(mz
, mctz
);
830 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
831 !css_tryget_online(&mz
->memcg
->css
))
837 static struct mem_cgroup_per_zone
*
838 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
840 struct mem_cgroup_per_zone
*mz
;
842 spin_lock(&mctz
->lock
);
843 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
844 spin_unlock(&mctz
->lock
);
849 * Implementation Note: reading percpu statistics for memcg.
851 * Both of vmstat[] and percpu_counter has threshold and do periodic
852 * synchronization to implement "quick" read. There are trade-off between
853 * reading cost and precision of value. Then, we may have a chance to implement
854 * a periodic synchronizion of counter in memcg's counter.
856 * But this _read() function is used for user interface now. The user accounts
857 * memory usage by memory cgroup and he _always_ requires exact value because
858 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
859 * have to visit all online cpus and make sum. So, for now, unnecessary
860 * synchronization is not implemented. (just implemented for cpu hotplug)
862 * If there are kernel internal actions which can make use of some not-exact
863 * value, and reading all cpu value can be performance bottleneck in some
864 * common workload, threashold and synchonization as vmstat[] should be
867 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
868 enum mem_cgroup_stat_index idx
)
874 for_each_online_cpu(cpu
)
875 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
876 #ifdef CONFIG_HOTPLUG_CPU
877 spin_lock(&memcg
->pcp_counter_lock
);
878 val
+= memcg
->nocpu_base
.count
[idx
];
879 spin_unlock(&memcg
->pcp_counter_lock
);
885 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
888 int val
= (charge
) ? 1 : -1;
889 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
892 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
893 enum mem_cgroup_events_index idx
)
895 unsigned long val
= 0;
899 for_each_online_cpu(cpu
)
900 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
901 #ifdef CONFIG_HOTPLUG_CPU
902 spin_lock(&memcg
->pcp_counter_lock
);
903 val
+= memcg
->nocpu_base
.events
[idx
];
904 spin_unlock(&memcg
->pcp_counter_lock
);
910 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
912 bool anon
, int nr_pages
)
915 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
916 * counted as CACHE even if it's on ANON LRU.
919 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
922 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
925 if (PageTransHuge(page
))
926 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
929 /* pagein of a big page is an event. So, ignore page size */
931 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
933 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
934 nr_pages
= -nr_pages
; /* for event */
937 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
940 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
942 struct mem_cgroup_per_zone
*mz
;
944 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
945 return mz
->lru_size
[lru
];
948 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
950 unsigned int lru_mask
)
952 unsigned long nr
= 0;
955 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
957 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
958 struct mem_cgroup_per_zone
*mz
;
962 if (!(BIT(lru
) & lru_mask
))
964 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
965 nr
+= mz
->lru_size
[lru
];
971 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
972 unsigned int lru_mask
)
974 unsigned long nr
= 0;
977 for_each_node_state(nid
, N_MEMORY
)
978 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
982 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
983 enum mem_cgroup_events_target target
)
985 unsigned long val
, next
;
987 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
988 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
989 /* from time_after() in jiffies.h */
990 if ((long)next
- (long)val
< 0) {
992 case MEM_CGROUP_TARGET_THRESH
:
993 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
995 case MEM_CGROUP_TARGET_SOFTLIMIT
:
996 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
998 case MEM_CGROUP_TARGET_NUMAINFO
:
999 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1004 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1011 * Check events in order.
1014 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1017 /* threshold event is triggered in finer grain than soft limit */
1018 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1019 MEM_CGROUP_TARGET_THRESH
))) {
1021 bool do_numainfo __maybe_unused
;
1023 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1024 MEM_CGROUP_TARGET_SOFTLIMIT
);
1025 #if MAX_NUMNODES > 1
1026 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1027 MEM_CGROUP_TARGET_NUMAINFO
);
1031 mem_cgroup_threshold(memcg
);
1032 if (unlikely(do_softlimit
))
1033 mem_cgroup_update_tree(memcg
, page
);
1034 #if MAX_NUMNODES > 1
1035 if (unlikely(do_numainfo
))
1036 atomic_inc(&memcg
->numainfo_events
);
1042 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1045 * mm_update_next_owner() may clear mm->owner to NULL
1046 * if it races with swapoff, page migration, etc.
1047 * So this can be called with p == NULL.
1052 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1055 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1057 struct mem_cgroup
*memcg
= NULL
;
1062 * Page cache insertions can happen withou an
1063 * actual mm context, e.g. during disk probing
1064 * on boot, loopback IO, acct() writes etc.
1067 memcg
= root_mem_cgroup
;
1069 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1070 if (unlikely(!memcg
))
1071 memcg
= root_mem_cgroup
;
1073 } while (!css_tryget_online(&memcg
->css
));
1079 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1080 * ref. count) or NULL if the whole root's subtree has been visited.
1082 * helper function to be used by mem_cgroup_iter
1084 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1085 struct mem_cgroup
*last_visited
)
1087 struct cgroup_subsys_state
*prev_css
, *next_css
;
1089 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1091 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1094 * Even if we found a group we have to make sure it is
1095 * alive. css && !memcg means that the groups should be
1096 * skipped and we should continue the tree walk.
1097 * last_visited css is safe to use because it is
1098 * protected by css_get and the tree walk is rcu safe.
1100 * We do not take a reference on the root of the tree walk
1101 * because we might race with the root removal when it would
1102 * be the only node in the iterated hierarchy and mem_cgroup_iter
1103 * would end up in an endless loop because it expects that at
1104 * least one valid node will be returned. Root cannot disappear
1105 * because caller of the iterator should hold it already so
1106 * skipping css reference should be safe.
1109 if ((next_css
== &root
->css
) ||
1110 ((next_css
->flags
& CSS_ONLINE
) &&
1111 css_tryget_online(next_css
)))
1112 return mem_cgroup_from_css(next_css
);
1114 prev_css
= next_css
;
1121 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1124 * When a group in the hierarchy below root is destroyed, the
1125 * hierarchy iterator can no longer be trusted since it might
1126 * have pointed to the destroyed group. Invalidate it.
1128 atomic_inc(&root
->dead_count
);
1131 static struct mem_cgroup
*
1132 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1133 struct mem_cgroup
*root
,
1136 struct mem_cgroup
*position
= NULL
;
1138 * A cgroup destruction happens in two stages: offlining and
1139 * release. They are separated by a RCU grace period.
1141 * If the iterator is valid, we may still race with an
1142 * offlining. The RCU lock ensures the object won't be
1143 * released, tryget will fail if we lost the race.
1145 *sequence
= atomic_read(&root
->dead_count
);
1146 if (iter
->last_dead_count
== *sequence
) {
1148 position
= iter
->last_visited
;
1151 * We cannot take a reference to root because we might race
1152 * with root removal and returning NULL would end up in
1153 * an endless loop on the iterator user level when root
1154 * would be returned all the time.
1156 if (position
&& position
!= root
&&
1157 !css_tryget_online(&position
->css
))
1163 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1164 struct mem_cgroup
*last_visited
,
1165 struct mem_cgroup
*new_position
,
1166 struct mem_cgroup
*root
,
1169 /* root reference counting symmetric to mem_cgroup_iter_load */
1170 if (last_visited
&& last_visited
!= root
)
1171 css_put(&last_visited
->css
);
1173 * We store the sequence count from the time @last_visited was
1174 * loaded successfully instead of rereading it here so that we
1175 * don't lose destruction events in between. We could have
1176 * raced with the destruction of @new_position after all.
1178 iter
->last_visited
= new_position
;
1180 iter
->last_dead_count
= sequence
;
1184 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1185 * @root: hierarchy root
1186 * @prev: previously returned memcg, NULL on first invocation
1187 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1189 * Returns references to children of the hierarchy below @root, or
1190 * @root itself, or %NULL after a full round-trip.
1192 * Caller must pass the return value in @prev on subsequent
1193 * invocations for reference counting, or use mem_cgroup_iter_break()
1194 * to cancel a hierarchy walk before the round-trip is complete.
1196 * Reclaimers can specify a zone and a priority level in @reclaim to
1197 * divide up the memcgs in the hierarchy among all concurrent
1198 * reclaimers operating on the same zone and priority.
1200 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1201 struct mem_cgroup
*prev
,
1202 struct mem_cgroup_reclaim_cookie
*reclaim
)
1204 struct mem_cgroup
*memcg
= NULL
;
1205 struct mem_cgroup
*last_visited
= NULL
;
1207 if (mem_cgroup_disabled())
1211 root
= root_mem_cgroup
;
1213 if (prev
&& !reclaim
)
1214 last_visited
= prev
;
1216 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1224 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1225 int uninitialized_var(seq
);
1228 struct mem_cgroup_per_zone
*mz
;
1230 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1231 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1232 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1233 iter
->last_visited
= NULL
;
1237 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1240 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1243 mem_cgroup_iter_update(iter
, last_visited
, memcg
, root
,
1248 else if (!prev
&& memcg
)
1249 reclaim
->generation
= iter
->generation
;
1258 if (prev
&& prev
!= root
)
1259 css_put(&prev
->css
);
1265 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1266 * @root: hierarchy root
1267 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1269 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1270 struct mem_cgroup
*prev
)
1273 root
= root_mem_cgroup
;
1274 if (prev
&& prev
!= root
)
1275 css_put(&prev
->css
);
1279 * Iteration constructs for visiting all cgroups (under a tree). If
1280 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1281 * be used for reference counting.
1283 #define for_each_mem_cgroup_tree(iter, root) \
1284 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1286 iter = mem_cgroup_iter(root, iter, NULL))
1288 #define for_each_mem_cgroup(iter) \
1289 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1291 iter = mem_cgroup_iter(NULL, iter, NULL))
1293 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1295 struct mem_cgroup
*memcg
;
1298 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1299 if (unlikely(!memcg
))
1304 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1307 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1315 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1318 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1319 * @zone: zone of the wanted lruvec
1320 * @memcg: memcg of the wanted lruvec
1322 * Returns the lru list vector holding pages for the given @zone and
1323 * @mem. This can be the global zone lruvec, if the memory controller
1326 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1327 struct mem_cgroup
*memcg
)
1329 struct mem_cgroup_per_zone
*mz
;
1330 struct lruvec
*lruvec
;
1332 if (mem_cgroup_disabled()) {
1333 lruvec
= &zone
->lruvec
;
1337 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1338 lruvec
= &mz
->lruvec
;
1341 * Since a node can be onlined after the mem_cgroup was created,
1342 * we have to be prepared to initialize lruvec->zone here;
1343 * and if offlined then reonlined, we need to reinitialize it.
1345 if (unlikely(lruvec
->zone
!= zone
))
1346 lruvec
->zone
= zone
;
1351 * Following LRU functions are allowed to be used without PCG_LOCK.
1352 * Operations are called by routine of global LRU independently from memcg.
1353 * What we have to take care of here is validness of pc->mem_cgroup.
1355 * Changes to pc->mem_cgroup happens when
1358 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1359 * It is added to LRU before charge.
1360 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1361 * When moving account, the page is not on LRU. It's isolated.
1365 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1367 * @zone: zone of the page
1369 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1371 struct mem_cgroup_per_zone
*mz
;
1372 struct mem_cgroup
*memcg
;
1373 struct page_cgroup
*pc
;
1374 struct lruvec
*lruvec
;
1376 if (mem_cgroup_disabled()) {
1377 lruvec
= &zone
->lruvec
;
1381 pc
= lookup_page_cgroup(page
);
1382 memcg
= pc
->mem_cgroup
;
1385 * Surreptitiously switch any uncharged offlist page to root:
1386 * an uncharged page off lru does nothing to secure
1387 * its former mem_cgroup from sudden removal.
1389 * Our caller holds lru_lock, and PageCgroupUsed is updated
1390 * under page_cgroup lock: between them, they make all uses
1391 * of pc->mem_cgroup safe.
1393 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1394 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1396 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1397 lruvec
= &mz
->lruvec
;
1400 * Since a node can be onlined after the mem_cgroup was created,
1401 * we have to be prepared to initialize lruvec->zone here;
1402 * and if offlined then reonlined, we need to reinitialize it.
1404 if (unlikely(lruvec
->zone
!= zone
))
1405 lruvec
->zone
= zone
;
1410 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1411 * @lruvec: mem_cgroup per zone lru vector
1412 * @lru: index of lru list the page is sitting on
1413 * @nr_pages: positive when adding or negative when removing
1415 * This function must be called when a page is added to or removed from an
1418 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1421 struct mem_cgroup_per_zone
*mz
;
1422 unsigned long *lru_size
;
1424 if (mem_cgroup_disabled())
1427 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1428 lru_size
= mz
->lru_size
+ lru
;
1429 *lru_size
+= nr_pages
;
1430 VM_BUG_ON((long)(*lru_size
) < 0);
1434 * Checks whether given mem is same or in the root_mem_cgroup's
1437 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1438 struct mem_cgroup
*memcg
)
1440 if (root_memcg
== memcg
)
1442 if (!root_memcg
->use_hierarchy
|| !memcg
)
1444 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1447 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1448 struct mem_cgroup
*memcg
)
1453 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1458 bool task_in_mem_cgroup(struct task_struct
*task
,
1459 const struct mem_cgroup
*memcg
)
1461 struct mem_cgroup
*curr
= NULL
;
1462 struct task_struct
*p
;
1465 p
= find_lock_task_mm(task
);
1467 curr
= get_mem_cgroup_from_mm(p
->mm
);
1471 * All threads may have already detached their mm's, but the oom
1472 * killer still needs to detect if they have already been oom
1473 * killed to prevent needlessly killing additional tasks.
1476 curr
= mem_cgroup_from_task(task
);
1478 css_get(&curr
->css
);
1482 * We should check use_hierarchy of "memcg" not "curr". Because checking
1483 * use_hierarchy of "curr" here make this function true if hierarchy is
1484 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1485 * hierarchy(even if use_hierarchy is disabled in "memcg").
1487 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1488 css_put(&curr
->css
);
1492 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1494 unsigned long inactive_ratio
;
1495 unsigned long inactive
;
1496 unsigned long active
;
1499 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1500 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1502 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1504 inactive_ratio
= int_sqrt(10 * gb
);
1508 return inactive
* inactive_ratio
< active
;
1511 #define mem_cgroup_from_res_counter(counter, member) \
1512 container_of(counter, struct mem_cgroup, member)
1515 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1516 * @memcg: the memory cgroup
1518 * Returns the maximum amount of memory @mem can be charged with, in
1521 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1523 unsigned long long margin
;
1525 margin
= res_counter_margin(&memcg
->res
);
1526 if (do_swap_account
)
1527 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1528 return margin
>> PAGE_SHIFT
;
1531 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1534 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1535 return vm_swappiness
;
1537 return memcg
->swappiness
;
1541 * memcg->moving_account is used for checking possibility that some thread is
1542 * calling move_account(). When a thread on CPU-A starts moving pages under
1543 * a memcg, other threads should check memcg->moving_account under
1544 * rcu_read_lock(), like this:
1548 * memcg->moving_account+1 if (memcg->mocing_account)
1550 * synchronize_rcu() update something.
1555 /* for quick checking without looking up memcg */
1556 atomic_t memcg_moving __read_mostly
;
1558 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1560 atomic_inc(&memcg_moving
);
1561 atomic_inc(&memcg
->moving_account
);
1565 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1568 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1569 * We check NULL in callee rather than caller.
1572 atomic_dec(&memcg_moving
);
1573 atomic_dec(&memcg
->moving_account
);
1578 * A routine for checking "mem" is under move_account() or not.
1580 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1581 * moving cgroups. This is for waiting at high-memory pressure
1584 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1586 struct mem_cgroup
*from
;
1587 struct mem_cgroup
*to
;
1590 * Unlike task_move routines, we access mc.to, mc.from not under
1591 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1593 spin_lock(&mc
.lock
);
1599 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1600 || mem_cgroup_same_or_subtree(memcg
, to
);
1602 spin_unlock(&mc
.lock
);
1606 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1608 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1609 if (mem_cgroup_under_move(memcg
)) {
1611 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1612 /* moving charge context might have finished. */
1615 finish_wait(&mc
.waitq
, &wait
);
1623 * Take this lock when
1624 * - a code tries to modify page's memcg while it's USED.
1625 * - a code tries to modify page state accounting in a memcg.
1627 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1628 unsigned long *flags
)
1630 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1633 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1634 unsigned long *flags
)
1636 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1639 #define K(x) ((x) << (PAGE_SHIFT-10))
1641 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1642 * @memcg: The memory cgroup that went over limit
1643 * @p: Task that is going to be killed
1645 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1648 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1650 /* oom_info_lock ensures that parallel ooms do not interleave */
1651 static DEFINE_MUTEX(oom_info_lock
);
1652 struct mem_cgroup
*iter
;
1658 mutex_lock(&oom_info_lock
);
1661 pr_info("Task in ");
1662 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1663 pr_info(" killed as a result of limit of ");
1664 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1669 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1670 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1671 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1672 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1673 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1674 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1675 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1676 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1677 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1678 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1679 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1680 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1682 for_each_mem_cgroup_tree(iter
, memcg
) {
1683 pr_info("Memory cgroup stats for ");
1684 pr_cont_cgroup_path(iter
->css
.cgroup
);
1687 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1688 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1690 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1691 K(mem_cgroup_read_stat(iter
, i
)));
1694 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1695 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1696 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1700 mutex_unlock(&oom_info_lock
);
1704 * This function returns the number of memcg under hierarchy tree. Returns
1705 * 1(self count) if no children.
1707 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1710 struct mem_cgroup
*iter
;
1712 for_each_mem_cgroup_tree(iter
, memcg
)
1718 * Return the memory (and swap, if configured) limit for a memcg.
1720 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1724 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1727 * Do not consider swap space if we cannot swap due to swappiness
1729 if (mem_cgroup_swappiness(memcg
)) {
1732 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1733 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1736 * If memsw is finite and limits the amount of swap space
1737 * available to this memcg, return that limit.
1739 limit
= min(limit
, memsw
);
1745 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1748 struct mem_cgroup
*iter
;
1749 unsigned long chosen_points
= 0;
1750 unsigned long totalpages
;
1751 unsigned int points
= 0;
1752 struct task_struct
*chosen
= NULL
;
1755 * If current has a pending SIGKILL or is exiting, then automatically
1756 * select it. The goal is to allow it to allocate so that it may
1757 * quickly exit and free its memory.
1759 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1760 set_thread_flag(TIF_MEMDIE
);
1764 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1765 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1766 for_each_mem_cgroup_tree(iter
, memcg
) {
1767 struct css_task_iter it
;
1768 struct task_struct
*task
;
1770 css_task_iter_start(&iter
->css
, &it
);
1771 while ((task
= css_task_iter_next(&it
))) {
1772 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1774 case OOM_SCAN_SELECT
:
1776 put_task_struct(chosen
);
1778 chosen_points
= ULONG_MAX
;
1779 get_task_struct(chosen
);
1781 case OOM_SCAN_CONTINUE
:
1783 case OOM_SCAN_ABORT
:
1784 css_task_iter_end(&it
);
1785 mem_cgroup_iter_break(memcg
, iter
);
1787 put_task_struct(chosen
);
1792 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1793 if (!points
|| points
< chosen_points
)
1795 /* Prefer thread group leaders for display purposes */
1796 if (points
== chosen_points
&&
1797 thread_group_leader(chosen
))
1801 put_task_struct(chosen
);
1803 chosen_points
= points
;
1804 get_task_struct(chosen
);
1806 css_task_iter_end(&it
);
1811 points
= chosen_points
* 1000 / totalpages
;
1812 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1813 NULL
, "Memory cgroup out of memory");
1816 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1818 unsigned long flags
)
1820 unsigned long total
= 0;
1821 bool noswap
= false;
1824 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1826 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1829 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1831 drain_all_stock_async(memcg
);
1832 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1834 * Allow limit shrinkers, which are triggered directly
1835 * by userspace, to catch signals and stop reclaim
1836 * after minimal progress, regardless of the margin.
1838 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1840 if (mem_cgroup_margin(memcg
))
1843 * If nothing was reclaimed after two attempts, there
1844 * may be no reclaimable pages in this hierarchy.
1853 * test_mem_cgroup_node_reclaimable
1854 * @memcg: the target memcg
1855 * @nid: the node ID to be checked.
1856 * @noswap : specify true here if the user wants flle only information.
1858 * This function returns whether the specified memcg contains any
1859 * reclaimable pages on a node. Returns true if there are any reclaimable
1860 * pages in the node.
1862 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1863 int nid
, bool noswap
)
1865 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1867 if (noswap
|| !total_swap_pages
)
1869 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1874 #if MAX_NUMNODES > 1
1877 * Always updating the nodemask is not very good - even if we have an empty
1878 * list or the wrong list here, we can start from some node and traverse all
1879 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1882 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1886 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1887 * pagein/pageout changes since the last update.
1889 if (!atomic_read(&memcg
->numainfo_events
))
1891 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1894 /* make a nodemask where this memcg uses memory from */
1895 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1897 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1899 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1900 node_clear(nid
, memcg
->scan_nodes
);
1903 atomic_set(&memcg
->numainfo_events
, 0);
1904 atomic_set(&memcg
->numainfo_updating
, 0);
1908 * Selecting a node where we start reclaim from. Because what we need is just
1909 * reducing usage counter, start from anywhere is O,K. Considering
1910 * memory reclaim from current node, there are pros. and cons.
1912 * Freeing memory from current node means freeing memory from a node which
1913 * we'll use or we've used. So, it may make LRU bad. And if several threads
1914 * hit limits, it will see a contention on a node. But freeing from remote
1915 * node means more costs for memory reclaim because of memory latency.
1917 * Now, we use round-robin. Better algorithm is welcomed.
1919 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1923 mem_cgroup_may_update_nodemask(memcg
);
1924 node
= memcg
->last_scanned_node
;
1926 node
= next_node(node
, memcg
->scan_nodes
);
1927 if (node
== MAX_NUMNODES
)
1928 node
= first_node(memcg
->scan_nodes
);
1930 * We call this when we hit limit, not when pages are added to LRU.
1931 * No LRU may hold pages because all pages are UNEVICTABLE or
1932 * memcg is too small and all pages are not on LRU. In that case,
1933 * we use curret node.
1935 if (unlikely(node
== MAX_NUMNODES
))
1936 node
= numa_node_id();
1938 memcg
->last_scanned_node
= node
;
1943 * Check all nodes whether it contains reclaimable pages or not.
1944 * For quick scan, we make use of scan_nodes. This will allow us to skip
1945 * unused nodes. But scan_nodes is lazily updated and may not cotain
1946 * enough new information. We need to do double check.
1948 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1953 * quick check...making use of scan_node.
1954 * We can skip unused nodes.
1956 if (!nodes_empty(memcg
->scan_nodes
)) {
1957 for (nid
= first_node(memcg
->scan_nodes
);
1959 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1961 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1966 * Check rest of nodes.
1968 for_each_node_state(nid
, N_MEMORY
) {
1969 if (node_isset(nid
, memcg
->scan_nodes
))
1971 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1978 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1983 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1985 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1989 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1992 unsigned long *total_scanned
)
1994 struct mem_cgroup
*victim
= NULL
;
1997 unsigned long excess
;
1998 unsigned long nr_scanned
;
1999 struct mem_cgroup_reclaim_cookie reclaim
= {
2004 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
2007 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
2012 * If we have not been able to reclaim
2013 * anything, it might because there are
2014 * no reclaimable pages under this hierarchy
2019 * We want to do more targeted reclaim.
2020 * excess >> 2 is not to excessive so as to
2021 * reclaim too much, nor too less that we keep
2022 * coming back to reclaim from this cgroup
2024 if (total
>= (excess
>> 2) ||
2025 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2030 if (!mem_cgroup_reclaimable(victim
, false))
2032 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2034 *total_scanned
+= nr_scanned
;
2035 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2038 mem_cgroup_iter_break(root_memcg
, victim
);
2042 #ifdef CONFIG_LOCKDEP
2043 static struct lockdep_map memcg_oom_lock_dep_map
= {
2044 .name
= "memcg_oom_lock",
2048 static DEFINE_SPINLOCK(memcg_oom_lock
);
2051 * Check OOM-Killer is already running under our hierarchy.
2052 * If someone is running, return false.
2054 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2056 struct mem_cgroup
*iter
, *failed
= NULL
;
2058 spin_lock(&memcg_oom_lock
);
2060 for_each_mem_cgroup_tree(iter
, memcg
) {
2061 if (iter
->oom_lock
) {
2063 * this subtree of our hierarchy is already locked
2064 * so we cannot give a lock.
2067 mem_cgroup_iter_break(memcg
, iter
);
2070 iter
->oom_lock
= true;
2075 * OK, we failed to lock the whole subtree so we have
2076 * to clean up what we set up to the failing subtree
2078 for_each_mem_cgroup_tree(iter
, memcg
) {
2079 if (iter
== failed
) {
2080 mem_cgroup_iter_break(memcg
, iter
);
2083 iter
->oom_lock
= false;
2086 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
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 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2099 for_each_mem_cgroup_tree(iter
, memcg
)
2100 iter
->oom_lock
= false;
2101 spin_unlock(&memcg_oom_lock
);
2104 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2106 struct mem_cgroup
*iter
;
2108 for_each_mem_cgroup_tree(iter
, memcg
)
2109 atomic_inc(&iter
->under_oom
);
2112 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2114 struct mem_cgroup
*iter
;
2117 * When a new child is created while the hierarchy is under oom,
2118 * mem_cgroup_oom_lock() may not be called. We have to use
2119 * atomic_add_unless() here.
2121 for_each_mem_cgroup_tree(iter
, memcg
)
2122 atomic_add_unless(&iter
->under_oom
, -1, 0);
2125 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2127 struct oom_wait_info
{
2128 struct mem_cgroup
*memcg
;
2132 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2133 unsigned mode
, int sync
, void *arg
)
2135 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2136 struct mem_cgroup
*oom_wait_memcg
;
2137 struct oom_wait_info
*oom_wait_info
;
2139 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2140 oom_wait_memcg
= oom_wait_info
->memcg
;
2143 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2144 * Then we can use css_is_ancestor without taking care of RCU.
2146 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2147 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2149 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2152 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2154 atomic_inc(&memcg
->oom_wakeups
);
2155 /* for filtering, pass "memcg" as argument. */
2156 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2159 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2161 if (memcg
&& atomic_read(&memcg
->under_oom
))
2162 memcg_wakeup_oom(memcg
);
2165 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2167 if (!current
->memcg_oom
.may_oom
)
2170 * We are in the middle of the charge context here, so we
2171 * don't want to block when potentially sitting on a callstack
2172 * that holds all kinds of filesystem and mm locks.
2174 * Also, the caller may handle a failed allocation gracefully
2175 * (like optional page cache readahead) and so an OOM killer
2176 * invocation might not even be necessary.
2178 * That's why we don't do anything here except remember the
2179 * OOM context and then deal with it at the end of the page
2180 * fault when the stack is unwound, the locks are released,
2181 * and when we know whether the fault was overall successful.
2183 css_get(&memcg
->css
);
2184 current
->memcg_oom
.memcg
= memcg
;
2185 current
->memcg_oom
.gfp_mask
= mask
;
2186 current
->memcg_oom
.order
= order
;
2190 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2191 * @handle: actually kill/wait or just clean up the OOM state
2193 * This has to be called at the end of a page fault if the memcg OOM
2194 * handler was enabled.
2196 * Memcg supports userspace OOM handling where failed allocations must
2197 * sleep on a waitqueue until the userspace task resolves the
2198 * situation. Sleeping directly in the charge context with all kinds
2199 * of locks held is not a good idea, instead we remember an OOM state
2200 * in the task and mem_cgroup_oom_synchronize() has to be called at
2201 * the end of the page fault to complete the OOM handling.
2203 * Returns %true if an ongoing memcg OOM situation was detected and
2204 * completed, %false otherwise.
2206 bool mem_cgroup_oom_synchronize(bool handle
)
2208 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2209 struct oom_wait_info owait
;
2212 /* OOM is global, do not handle */
2219 owait
.memcg
= memcg
;
2220 owait
.wait
.flags
= 0;
2221 owait
.wait
.func
= memcg_oom_wake_function
;
2222 owait
.wait
.private = current
;
2223 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2225 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2226 mem_cgroup_mark_under_oom(memcg
);
2228 locked
= mem_cgroup_oom_trylock(memcg
);
2231 mem_cgroup_oom_notify(memcg
);
2233 if (locked
&& !memcg
->oom_kill_disable
) {
2234 mem_cgroup_unmark_under_oom(memcg
);
2235 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2236 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2237 current
->memcg_oom
.order
);
2240 mem_cgroup_unmark_under_oom(memcg
);
2241 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2245 mem_cgroup_oom_unlock(memcg
);
2247 * There is no guarantee that an OOM-lock contender
2248 * sees the wakeups triggered by the OOM kill
2249 * uncharges. Wake any sleepers explicitely.
2251 memcg_oom_recover(memcg
);
2254 current
->memcg_oom
.memcg
= NULL
;
2255 css_put(&memcg
->css
);
2260 * Used to update mapped file or writeback or other statistics.
2262 * Notes: Race condition
2264 * We usually use lock_page_cgroup() 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 memcg->moving_account and detect there are possibility
2279 * of race or not. 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().
2299 VM_BUG_ON(!rcu_read_lock_held());
2300 if (atomic_read(&memcg
->moving_account
) <= 0)
2303 move_lock_mem_cgroup(memcg
, flags
);
2304 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2305 move_unlock_mem_cgroup(memcg
, flags
);
2311 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2313 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2316 * It's guaranteed that pc->mem_cgroup never changes while
2317 * lock is held because a routine modifies pc->mem_cgroup
2318 * should take move_lock_mem_cgroup().
2320 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2323 void mem_cgroup_update_page_stat(struct page
*page
,
2324 enum mem_cgroup_stat_index idx
, int val
)
2326 struct mem_cgroup
*memcg
;
2327 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2328 unsigned long uninitialized_var(flags
);
2330 if (mem_cgroup_disabled())
2333 VM_BUG_ON(!rcu_read_lock_held());
2334 memcg
= pc
->mem_cgroup
;
2335 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2338 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2342 * size of first charge trial. "32" comes from vmscan.c's magic value.
2343 * TODO: maybe necessary to use big numbers in big irons.
2345 #define CHARGE_BATCH 32U
2346 struct memcg_stock_pcp
{
2347 struct mem_cgroup
*cached
; /* this never be root cgroup */
2348 unsigned int nr_pages
;
2349 struct work_struct work
;
2350 unsigned long flags
;
2351 #define FLUSHING_CACHED_CHARGE 0
2353 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2354 static DEFINE_MUTEX(percpu_charge_mutex
);
2357 * consume_stock: Try to consume stocked charge on this cpu.
2358 * @memcg: memcg to consume from.
2359 * @nr_pages: how many pages to charge.
2361 * The charges will only happen if @memcg matches the current cpu's memcg
2362 * stock, and at least @nr_pages are available in that stock. Failure to
2363 * service an allocation will refill the stock.
2365 * returns true if successful, false otherwise.
2367 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2369 struct memcg_stock_pcp
*stock
;
2372 if (nr_pages
> CHARGE_BATCH
)
2375 stock
= &get_cpu_var(memcg_stock
);
2376 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2377 stock
->nr_pages
-= nr_pages
;
2378 else /* need to call res_counter_charge */
2380 put_cpu_var(memcg_stock
);
2385 * Returns stocks cached in percpu to res_counter and reset cached information.
2387 static void drain_stock(struct memcg_stock_pcp
*stock
)
2389 struct mem_cgroup
*old
= stock
->cached
;
2391 if (stock
->nr_pages
) {
2392 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2394 res_counter_uncharge(&old
->res
, bytes
);
2395 if (do_swap_account
)
2396 res_counter_uncharge(&old
->memsw
, bytes
);
2397 stock
->nr_pages
= 0;
2399 stock
->cached
= NULL
;
2403 * This must be called under preempt disabled or must be called by
2404 * a thread which is pinned to local cpu.
2406 static void drain_local_stock(struct work_struct
*dummy
)
2408 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2410 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2413 static void __init
memcg_stock_init(void)
2417 for_each_possible_cpu(cpu
) {
2418 struct memcg_stock_pcp
*stock
=
2419 &per_cpu(memcg_stock
, cpu
);
2420 INIT_WORK(&stock
->work
, drain_local_stock
);
2425 * Cache charges(val) which is from res_counter, to local per_cpu area.
2426 * This will be consumed by consume_stock() function, later.
2428 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2430 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2432 if (stock
->cached
!= memcg
) { /* reset if necessary */
2434 stock
->cached
= memcg
;
2436 stock
->nr_pages
+= nr_pages
;
2437 put_cpu_var(memcg_stock
);
2441 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2442 * of the hierarchy under it. sync flag says whether we should block
2443 * until the work is done.
2445 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2449 /* Notify other cpus that system-wide "drain" is running */
2452 for_each_online_cpu(cpu
) {
2453 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2454 struct mem_cgroup
*memcg
;
2456 memcg
= stock
->cached
;
2457 if (!memcg
|| !stock
->nr_pages
)
2459 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2461 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2463 drain_local_stock(&stock
->work
);
2465 schedule_work_on(cpu
, &stock
->work
);
2473 for_each_online_cpu(cpu
) {
2474 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2475 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2476 flush_work(&stock
->work
);
2483 * Tries to drain stocked charges in other cpus. This function is asynchronous
2484 * and just put a work per cpu for draining localy on each cpu. Caller can
2485 * expects some charges will be back to res_counter later but cannot wait for
2488 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2491 * If someone calls draining, avoid adding more kworker runs.
2493 if (!mutex_trylock(&percpu_charge_mutex
))
2495 drain_all_stock(root_memcg
, false);
2496 mutex_unlock(&percpu_charge_mutex
);
2499 /* This is a synchronous drain interface. */
2500 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2502 /* called when force_empty is called */
2503 mutex_lock(&percpu_charge_mutex
);
2504 drain_all_stock(root_memcg
, true);
2505 mutex_unlock(&percpu_charge_mutex
);
2509 * This function drains percpu counter value from DEAD cpu and
2510 * move it to local cpu. Note that this function can be preempted.
2512 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2516 spin_lock(&memcg
->pcp_counter_lock
);
2517 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2518 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2520 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2521 memcg
->nocpu_base
.count
[i
] += x
;
2523 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2524 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2526 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2527 memcg
->nocpu_base
.events
[i
] += x
;
2529 spin_unlock(&memcg
->pcp_counter_lock
);
2532 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2533 unsigned long action
,
2536 int cpu
= (unsigned long)hcpu
;
2537 struct memcg_stock_pcp
*stock
;
2538 struct mem_cgroup
*iter
;
2540 if (action
== CPU_ONLINE
)
2543 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2546 for_each_mem_cgroup(iter
)
2547 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2549 stock
= &per_cpu(memcg_stock
, cpu
);
2555 * mem_cgroup_try_charge - try charging a memcg
2556 * @memcg: memcg to charge
2557 * @nr_pages: number of pages to charge
2559 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2560 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2562 static int mem_cgroup_try_charge(struct mem_cgroup
*memcg
,
2564 unsigned int nr_pages
)
2566 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2567 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2568 struct mem_cgroup
*mem_over_limit
;
2569 struct res_counter
*fail_res
;
2570 unsigned long nr_reclaimed
;
2571 unsigned long flags
= 0;
2572 unsigned long long size
;
2576 if (consume_stock(memcg
, nr_pages
))
2579 size
= batch
* PAGE_SIZE
;
2580 if (!res_counter_charge(&memcg
->res
, size
, &fail_res
)) {
2581 if (!do_swap_account
)
2583 if (!res_counter_charge(&memcg
->memsw
, size
, &fail_res
))
2585 res_counter_uncharge(&memcg
->res
, size
);
2586 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2587 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2589 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2591 if (batch
> nr_pages
) {
2597 * Unlike in global OOM situations, memcg is not in a physical
2598 * memory shortage. Allow dying and OOM-killed tasks to
2599 * bypass the last charges so that they can exit quickly and
2600 * free their memory.
2602 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2603 fatal_signal_pending(current
) ||
2604 current
->flags
& PF_EXITING
))
2607 if (unlikely(task_in_memcg_oom(current
)))
2610 if (!(gfp_mask
& __GFP_WAIT
))
2613 nr_reclaimed
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2615 if (mem_cgroup_margin(mem_over_limit
) >= batch
)
2618 if (gfp_mask
& __GFP_NORETRY
)
2621 * Even though the limit is exceeded at this point, reclaim
2622 * may have been able to free some pages. Retry the charge
2623 * before killing the task.
2625 * Only for regular pages, though: huge pages are rather
2626 * unlikely to succeed so close to the limit, and we fall back
2627 * to regular pages anyway in case of failure.
2629 if (nr_reclaimed
&& batch
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2632 * At task move, charge accounts can be doubly counted. So, it's
2633 * better to wait until the end of task_move if something is going on.
2635 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2641 if (gfp_mask
& __GFP_NOFAIL
)
2644 if (fatal_signal_pending(current
))
2647 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(batch
));
2649 if (!(gfp_mask
& __GFP_NOFAIL
))
2652 memcg
= root_mem_cgroup
;
2657 if (batch
> nr_pages
)
2658 refill_stock(memcg
, batch
- nr_pages
);
2664 * mem_cgroup_try_charge_mm - try charging a mm
2665 * @mm: mm_struct to charge
2666 * @nr_pages: number of pages to charge
2667 * @oom: trigger OOM if reclaim fails
2669 * Returns the charged mem_cgroup associated with the given mm_struct or
2670 * NULL the charge failed.
2672 static struct mem_cgroup
*mem_cgroup_try_charge_mm(struct mm_struct
*mm
,
2674 unsigned int nr_pages
)
2677 struct mem_cgroup
*memcg
;
2680 memcg
= get_mem_cgroup_from_mm(mm
);
2681 ret
= mem_cgroup_try_charge(memcg
, gfp_mask
, nr_pages
);
2682 css_put(&memcg
->css
);
2684 memcg
= root_mem_cgroup
;
2692 * Somemtimes we have to undo a charge we got by try_charge().
2693 * This function is for that and do uncharge, put css's refcnt.
2694 * gotten by try_charge().
2696 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2697 unsigned int nr_pages
)
2699 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2701 res_counter_uncharge(&memcg
->res
, bytes
);
2702 if (do_swap_account
)
2703 res_counter_uncharge(&memcg
->memsw
, bytes
);
2707 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2708 * This is useful when moving usage to parent cgroup.
2710 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2711 unsigned int nr_pages
)
2713 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2715 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2716 if (do_swap_account
)
2717 res_counter_uncharge_until(&memcg
->memsw
,
2718 memcg
->memsw
.parent
, bytes
);
2722 * A helper function to get mem_cgroup from ID. must be called under
2723 * rcu_read_lock(). The caller is responsible for calling
2724 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2725 * refcnt from swap can be called against removed memcg.)
2727 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2729 /* ID 0 is unused ID */
2732 return mem_cgroup_from_id(id
);
2735 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2737 struct mem_cgroup
*memcg
= NULL
;
2738 struct page_cgroup
*pc
;
2742 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2744 pc
= lookup_page_cgroup(page
);
2745 lock_page_cgroup(pc
);
2746 if (PageCgroupUsed(pc
)) {
2747 memcg
= pc
->mem_cgroup
;
2748 if (memcg
&& !css_tryget_online(&memcg
->css
))
2750 } else if (PageSwapCache(page
)) {
2751 ent
.val
= page_private(page
);
2752 id
= lookup_swap_cgroup_id(ent
);
2754 memcg
= mem_cgroup_lookup(id
);
2755 if (memcg
&& !css_tryget_online(&memcg
->css
))
2759 unlock_page_cgroup(pc
);
2763 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2765 unsigned int nr_pages
,
2766 enum charge_type ctype
,
2769 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2770 struct zone
*uninitialized_var(zone
);
2771 struct lruvec
*lruvec
;
2772 bool was_on_lru
= false;
2775 lock_page_cgroup(pc
);
2776 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2778 * we don't need page_cgroup_lock about tail pages, becase they are not
2779 * accessed by any other context at this point.
2783 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2784 * may already be on some other mem_cgroup's LRU. Take care of it.
2787 zone
= page_zone(page
);
2788 spin_lock_irq(&zone
->lru_lock
);
2789 if (PageLRU(page
)) {
2790 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2792 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2797 pc
->mem_cgroup
= memcg
;
2799 * We access a page_cgroup asynchronously without lock_page_cgroup().
2800 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2801 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2802 * before USED bit, we need memory barrier here.
2803 * See mem_cgroup_add_lru_list(), etc.
2806 SetPageCgroupUsed(pc
);
2810 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2811 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2813 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2815 spin_unlock_irq(&zone
->lru_lock
);
2818 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2823 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2824 unlock_page_cgroup(pc
);
2827 * "charge_statistics" updated event counter. Then, check it.
2828 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2829 * if they exceeds softlimit.
2831 memcg_check_events(memcg
, page
);
2834 static DEFINE_MUTEX(set_limit_mutex
);
2836 #ifdef CONFIG_MEMCG_KMEM
2838 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2839 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2841 static DEFINE_MUTEX(memcg_slab_mutex
);
2843 static DEFINE_MUTEX(activate_kmem_mutex
);
2845 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2847 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2848 memcg_kmem_is_active(memcg
);
2852 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2853 * in the memcg_cache_params struct.
2855 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2857 struct kmem_cache
*cachep
;
2859 VM_BUG_ON(p
->is_root_cache
);
2860 cachep
= p
->root_cache
;
2861 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2864 #ifdef CONFIG_SLABINFO
2865 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
2867 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
2868 struct memcg_cache_params
*params
;
2870 if (!memcg_can_account_kmem(memcg
))
2873 print_slabinfo_header(m
);
2875 mutex_lock(&memcg_slab_mutex
);
2876 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2877 cache_show(memcg_params_to_cache(params
), m
);
2878 mutex_unlock(&memcg_slab_mutex
);
2884 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2886 struct res_counter
*fail_res
;
2889 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2893 ret
= mem_cgroup_try_charge(memcg
, gfp
, size
>> PAGE_SHIFT
);
2894 if (ret
== -EINTR
) {
2896 * mem_cgroup_try_charge() chosed to bypass to root due to
2897 * OOM kill or fatal signal. Since our only options are to
2898 * either fail the allocation or charge it to this cgroup, do
2899 * it as a temporary condition. But we can't fail. From a
2900 * kmem/slab perspective, the cache has already been selected,
2901 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2904 * This condition will only trigger if the task entered
2905 * memcg_charge_kmem in a sane state, but was OOM-killed during
2906 * mem_cgroup_try_charge() above. Tasks that were already
2907 * dying when the allocation triggers should have been already
2908 * directed to the root cgroup in memcontrol.h
2910 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2911 if (do_swap_account
)
2912 res_counter_charge_nofail(&memcg
->memsw
, size
,
2916 res_counter_uncharge(&memcg
->kmem
, size
);
2921 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2923 res_counter_uncharge(&memcg
->res
, size
);
2924 if (do_swap_account
)
2925 res_counter_uncharge(&memcg
->memsw
, size
);
2928 if (res_counter_uncharge(&memcg
->kmem
, size
))
2932 * Releases a reference taken in kmem_cgroup_css_offline in case
2933 * this last uncharge is racing with the offlining code or it is
2934 * outliving the memcg existence.
2936 * The memory barrier imposed by test&clear is paired with the
2937 * explicit one in memcg_kmem_mark_dead().
2939 if (memcg_kmem_test_and_clear_dead(memcg
))
2940 css_put(&memcg
->css
);
2944 * helper for acessing a memcg's index. It will be used as an index in the
2945 * child cache array in kmem_cache, and also to derive its name. This function
2946 * will return -1 when this is not a kmem-limited memcg.
2948 int memcg_cache_id(struct mem_cgroup
*memcg
)
2950 return memcg
? memcg
->kmemcg_id
: -1;
2953 static size_t memcg_caches_array_size(int num_groups
)
2956 if (num_groups
<= 0)
2959 size
= 2 * num_groups
;
2960 if (size
< MEMCG_CACHES_MIN_SIZE
)
2961 size
= MEMCG_CACHES_MIN_SIZE
;
2962 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2963 size
= MEMCG_CACHES_MAX_SIZE
;
2969 * We should update the current array size iff all caches updates succeed. This
2970 * can only be done from the slab side. The slab mutex needs to be held when
2973 void memcg_update_array_size(int num
)
2975 if (num
> memcg_limited_groups_array_size
)
2976 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2979 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2981 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2983 VM_BUG_ON(!is_root_cache(s
));
2985 if (num_groups
> memcg_limited_groups_array_size
) {
2987 struct memcg_cache_params
*new_params
;
2988 ssize_t size
= memcg_caches_array_size(num_groups
);
2990 size
*= sizeof(void *);
2991 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
2993 new_params
= kzalloc(size
, GFP_KERNEL
);
2997 new_params
->is_root_cache
= true;
3000 * There is the chance it will be bigger than
3001 * memcg_limited_groups_array_size, if we failed an allocation
3002 * in a cache, in which case all caches updated before it, will
3003 * have a bigger array.
3005 * But if that is the case, the data after
3006 * memcg_limited_groups_array_size is certainly unused
3008 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3009 if (!cur_params
->memcg_caches
[i
])
3011 new_params
->memcg_caches
[i
] =
3012 cur_params
->memcg_caches
[i
];
3016 * Ideally, we would wait until all caches succeed, and only
3017 * then free the old one. But this is not worth the extra
3018 * pointer per-cache we'd have to have for this.
3020 * It is not a big deal if some caches are left with a size
3021 * bigger than the others. And all updates will reset this
3024 rcu_assign_pointer(s
->memcg_params
, new_params
);
3026 kfree_rcu(cur_params
, rcu_head
);
3031 int memcg_alloc_cache_params(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3032 struct kmem_cache
*root_cache
)
3036 if (!memcg_kmem_enabled())
3040 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3041 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3043 size
= sizeof(struct memcg_cache_params
);
3045 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3046 if (!s
->memcg_params
)
3050 s
->memcg_params
->memcg
= memcg
;
3051 s
->memcg_params
->root_cache
= root_cache
;
3052 css_get(&memcg
->css
);
3054 s
->memcg_params
->is_root_cache
= true;
3059 void memcg_free_cache_params(struct kmem_cache
*s
)
3061 if (!s
->memcg_params
)
3063 if (!s
->memcg_params
->is_root_cache
)
3064 css_put(&s
->memcg_params
->memcg
->css
);
3065 kfree(s
->memcg_params
);
3068 static void memcg_register_cache(struct mem_cgroup
*memcg
,
3069 struct kmem_cache
*root_cache
)
3071 static char memcg_name_buf
[NAME_MAX
+ 1]; /* protected by
3073 struct kmem_cache
*cachep
;
3076 lockdep_assert_held(&memcg_slab_mutex
);
3078 id
= memcg_cache_id(memcg
);
3081 * Since per-memcg caches are created asynchronously on first
3082 * allocation (see memcg_kmem_get_cache()), several threads can try to
3083 * create the same cache, but only one of them may succeed.
3085 if (cache_from_memcg_idx(root_cache
, id
))
3088 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
3089 cachep
= memcg_create_kmem_cache(memcg
, root_cache
, memcg_name_buf
);
3091 * If we could not create a memcg cache, do not complain, because
3092 * that's not critical at all as we can always proceed with the root
3098 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3101 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3102 * barrier here to ensure nobody will see the kmem_cache partially
3107 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
3108 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
3111 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
3113 struct kmem_cache
*root_cache
;
3114 struct mem_cgroup
*memcg
;
3117 lockdep_assert_held(&memcg_slab_mutex
);
3119 BUG_ON(is_root_cache(cachep
));
3121 root_cache
= cachep
->memcg_params
->root_cache
;
3122 memcg
= cachep
->memcg_params
->memcg
;
3123 id
= memcg_cache_id(memcg
);
3125 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
3126 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
3128 list_del(&cachep
->memcg_params
->list
);
3130 kmem_cache_destroy(cachep
);
3134 * During the creation a new cache, we need to disable our accounting mechanism
3135 * altogether. This is true even if we are not creating, but rather just
3136 * enqueing new caches to be created.
3138 * This is because that process will trigger allocations; some visible, like
3139 * explicit kmallocs to auxiliary data structures, name strings and internal
3140 * cache structures; some well concealed, like INIT_WORK() that can allocate
3141 * objects during debug.
3143 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3144 * to it. This may not be a bounded recursion: since the first cache creation
3145 * failed to complete (waiting on the allocation), we'll just try to create the
3146 * cache again, failing at the same point.
3148 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3149 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3150 * inside the following two functions.
3152 static inline void memcg_stop_kmem_account(void)
3154 VM_BUG_ON(!current
->mm
);
3155 current
->memcg_kmem_skip_account
++;
3158 static inline void memcg_resume_kmem_account(void)
3160 VM_BUG_ON(!current
->mm
);
3161 current
->memcg_kmem_skip_account
--;
3164 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
3166 struct kmem_cache
*c
;
3169 mutex_lock(&memcg_slab_mutex
);
3170 for_each_memcg_cache_index(i
) {
3171 c
= cache_from_memcg_idx(s
, i
);
3175 memcg_unregister_cache(c
);
3177 if (cache_from_memcg_idx(s
, i
))
3180 mutex_unlock(&memcg_slab_mutex
);
3184 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3186 struct kmem_cache
*cachep
;
3187 struct memcg_cache_params
*params
, *tmp
;
3189 if (!memcg_kmem_is_active(memcg
))
3192 mutex_lock(&memcg_slab_mutex
);
3193 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
3194 cachep
= memcg_params_to_cache(params
);
3195 kmem_cache_shrink(cachep
);
3196 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3197 memcg_unregister_cache(cachep
);
3199 mutex_unlock(&memcg_slab_mutex
);
3202 struct memcg_register_cache_work
{
3203 struct mem_cgroup
*memcg
;
3204 struct kmem_cache
*cachep
;
3205 struct work_struct work
;
3208 static void memcg_register_cache_func(struct work_struct
*w
)
3210 struct memcg_register_cache_work
*cw
=
3211 container_of(w
, struct memcg_register_cache_work
, work
);
3212 struct mem_cgroup
*memcg
= cw
->memcg
;
3213 struct kmem_cache
*cachep
= cw
->cachep
;
3215 mutex_lock(&memcg_slab_mutex
);
3216 memcg_register_cache(memcg
, cachep
);
3217 mutex_unlock(&memcg_slab_mutex
);
3219 css_put(&memcg
->css
);
3224 * Enqueue the creation of a per-memcg kmem_cache.
3226 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3227 struct kmem_cache
*cachep
)
3229 struct memcg_register_cache_work
*cw
;
3231 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
3233 css_put(&memcg
->css
);
3238 cw
->cachep
= cachep
;
3240 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
3241 schedule_work(&cw
->work
);
3244 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3245 struct kmem_cache
*cachep
)
3248 * We need to stop accounting when we kmalloc, because if the
3249 * corresponding kmalloc cache is not yet created, the first allocation
3250 * in __memcg_schedule_register_cache will recurse.
3252 * However, it is better to enclose the whole function. Depending on
3253 * the debugging options enabled, INIT_WORK(), for instance, can
3254 * trigger an allocation. This too, will make us recurse. Because at
3255 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3256 * the safest choice is to do it like this, wrapping the whole function.
3258 memcg_stop_kmem_account();
3259 __memcg_schedule_register_cache(memcg
, cachep
);
3260 memcg_resume_kmem_account();
3263 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
3267 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
,
3268 PAGE_SIZE
<< order
);
3270 atomic_add(1 << order
, &cachep
->memcg_params
->nr_pages
);
3274 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
3276 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, PAGE_SIZE
<< order
);
3277 atomic_sub(1 << order
, &cachep
->memcg_params
->nr_pages
);
3281 * Return the kmem_cache we're supposed to use for a slab allocation.
3282 * We try to use the current memcg's version of the cache.
3284 * If the cache does not exist yet, if we are the first user of it,
3285 * we either create it immediately, if possible, or create it asynchronously
3287 * In the latter case, we will let the current allocation go through with
3288 * the original cache.
3290 * Can't be called in interrupt context or from kernel threads.
3291 * This function needs to be called with rcu_read_lock() held.
3293 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3296 struct mem_cgroup
*memcg
;
3297 struct kmem_cache
*memcg_cachep
;
3299 VM_BUG_ON(!cachep
->memcg_params
);
3300 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3302 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3306 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3308 if (!memcg_can_account_kmem(memcg
))
3311 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3312 if (likely(memcg_cachep
)) {
3313 cachep
= memcg_cachep
;
3317 /* The corresponding put will be done in the workqueue. */
3318 if (!css_tryget_online(&memcg
->css
))
3323 * If we are in a safe context (can wait, and not in interrupt
3324 * context), we could be be predictable and return right away.
3325 * This would guarantee that the allocation being performed
3326 * already belongs in the new cache.
3328 * However, there are some clashes that can arrive from locking.
3329 * For instance, because we acquire the slab_mutex while doing
3330 * memcg_create_kmem_cache, this means no further allocation
3331 * could happen with the slab_mutex held. So it's better to
3334 memcg_schedule_register_cache(memcg
, cachep
);
3342 * We need to verify if the allocation against current->mm->owner's memcg is
3343 * possible for the given order. But the page is not allocated yet, so we'll
3344 * need a further commit step to do the final arrangements.
3346 * It is possible for the task to switch cgroups in this mean time, so at
3347 * commit time, we can't rely on task conversion any longer. We'll then use
3348 * the handle argument to return to the caller which cgroup we should commit
3349 * against. We could also return the memcg directly and avoid the pointer
3350 * passing, but a boolean return value gives better semantics considering
3351 * the compiled-out case as well.
3353 * Returning true means the allocation is possible.
3356 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3358 struct mem_cgroup
*memcg
;
3364 * Disabling accounting is only relevant for some specific memcg
3365 * internal allocations. Therefore we would initially not have such
3366 * check here, since direct calls to the page allocator that are
3367 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3368 * outside memcg core. We are mostly concerned with cache allocations,
3369 * and by having this test at memcg_kmem_get_cache, we are already able
3370 * to relay the allocation to the root cache and bypass the memcg cache
3373 * There is one exception, though: the SLUB allocator does not create
3374 * large order caches, but rather service large kmallocs directly from
3375 * the page allocator. Therefore, the following sequence when backed by
3376 * the SLUB allocator:
3378 * memcg_stop_kmem_account();
3379 * kmalloc(<large_number>)
3380 * memcg_resume_kmem_account();
3382 * would effectively ignore the fact that we should skip accounting,
3383 * since it will drive us directly to this function without passing
3384 * through the cache selector memcg_kmem_get_cache. Such large
3385 * allocations are extremely rare but can happen, for instance, for the
3386 * cache arrays. We bring this test here.
3388 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3391 memcg
= get_mem_cgroup_from_mm(current
->mm
);
3393 if (!memcg_can_account_kmem(memcg
)) {
3394 css_put(&memcg
->css
);
3398 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3402 css_put(&memcg
->css
);
3406 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3409 struct page_cgroup
*pc
;
3411 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3413 /* The page allocation failed. Revert */
3415 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3419 pc
= lookup_page_cgroup(page
);
3420 lock_page_cgroup(pc
);
3421 pc
->mem_cgroup
= memcg
;
3422 SetPageCgroupUsed(pc
);
3423 unlock_page_cgroup(pc
);
3426 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3428 struct mem_cgroup
*memcg
= NULL
;
3429 struct page_cgroup
*pc
;
3432 pc
= lookup_page_cgroup(page
);
3434 * Fast unlocked return. Theoretically might have changed, have to
3435 * check again after locking.
3437 if (!PageCgroupUsed(pc
))
3440 lock_page_cgroup(pc
);
3441 if (PageCgroupUsed(pc
)) {
3442 memcg
= pc
->mem_cgroup
;
3443 ClearPageCgroupUsed(pc
);
3445 unlock_page_cgroup(pc
);
3448 * We trust that only if there is a memcg associated with the page, it
3449 * is a valid allocation
3454 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3455 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3458 static inline void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3461 #endif /* CONFIG_MEMCG_KMEM */
3463 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3465 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3467 * Because tail pages are not marked as "used", set it. We're under
3468 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3469 * charge/uncharge will be never happen and move_account() is done under
3470 * compound_lock(), so we don't have to take care of races.
3472 void mem_cgroup_split_huge_fixup(struct page
*head
)
3474 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3475 struct page_cgroup
*pc
;
3476 struct mem_cgroup
*memcg
;
3479 if (mem_cgroup_disabled())
3482 memcg
= head_pc
->mem_cgroup
;
3483 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3485 pc
->mem_cgroup
= memcg
;
3486 smp_wmb();/* see __commit_charge() */
3487 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3489 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3492 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3495 * mem_cgroup_move_account - move account of the page
3497 * @nr_pages: number of regular pages (>1 for huge pages)
3498 * @pc: page_cgroup of the page.
3499 * @from: mem_cgroup which the page is moved from.
3500 * @to: mem_cgroup which the page is moved to. @from != @to.
3502 * The caller must confirm following.
3503 * - page is not on LRU (isolate_page() is useful.)
3504 * - compound_lock is held when nr_pages > 1
3506 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3509 static int mem_cgroup_move_account(struct page
*page
,
3510 unsigned int nr_pages
,
3511 struct page_cgroup
*pc
,
3512 struct mem_cgroup
*from
,
3513 struct mem_cgroup
*to
)
3515 unsigned long flags
;
3517 bool anon
= PageAnon(page
);
3519 VM_BUG_ON(from
== to
);
3520 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3522 * The page is isolated from LRU. So, collapse function
3523 * will not handle this page. But page splitting can happen.
3524 * Do this check under compound_page_lock(). The caller should
3528 if (nr_pages
> 1 && !PageTransHuge(page
))
3531 lock_page_cgroup(pc
);
3534 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3537 move_lock_mem_cgroup(from
, &flags
);
3539 if (!anon
&& page_mapped(page
)) {
3540 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3542 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3546 if (PageWriteback(page
)) {
3547 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3549 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3553 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3555 /* caller should have done css_get */
3556 pc
->mem_cgroup
= to
;
3557 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3558 move_unlock_mem_cgroup(from
, &flags
);
3561 unlock_page_cgroup(pc
);
3565 memcg_check_events(to
, page
);
3566 memcg_check_events(from
, page
);
3572 * mem_cgroup_move_parent - moves page to the parent group
3573 * @page: the page to move
3574 * @pc: page_cgroup of the page
3575 * @child: page's cgroup
3577 * move charges to its parent or the root cgroup if the group has no
3578 * parent (aka use_hierarchy==0).
3579 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3580 * mem_cgroup_move_account fails) the failure is always temporary and
3581 * it signals a race with a page removal/uncharge or migration. In the
3582 * first case the page is on the way out and it will vanish from the LRU
3583 * on the next attempt and the call should be retried later.
3584 * Isolation from the LRU fails only if page has been isolated from
3585 * the LRU since we looked at it and that usually means either global
3586 * reclaim or migration going on. The page will either get back to the
3588 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3589 * (!PageCgroupUsed) or moved to a different group. The page will
3590 * disappear in the next attempt.
3592 static int mem_cgroup_move_parent(struct page
*page
,
3593 struct page_cgroup
*pc
,
3594 struct mem_cgroup
*child
)
3596 struct mem_cgroup
*parent
;
3597 unsigned int nr_pages
;
3598 unsigned long uninitialized_var(flags
);
3601 VM_BUG_ON(mem_cgroup_is_root(child
));
3604 if (!get_page_unless_zero(page
))
3606 if (isolate_lru_page(page
))
3609 nr_pages
= hpage_nr_pages(page
);
3611 parent
= parent_mem_cgroup(child
);
3613 * If no parent, move charges to root cgroup.
3616 parent
= root_mem_cgroup
;
3619 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3620 flags
= compound_lock_irqsave(page
);
3623 ret
= mem_cgroup_move_account(page
, nr_pages
,
3626 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3629 compound_unlock_irqrestore(page
, flags
);
3630 putback_lru_page(page
);
3637 int mem_cgroup_charge_anon(struct page
*page
,
3638 struct mm_struct
*mm
, gfp_t gfp_mask
)
3640 unsigned int nr_pages
= 1;
3641 struct mem_cgroup
*memcg
;
3643 if (mem_cgroup_disabled())
3646 VM_BUG_ON_PAGE(page_mapped(page
), page
);
3647 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
3650 if (PageTransHuge(page
)) {
3651 nr_pages
<<= compound_order(page
);
3652 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3655 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, nr_pages
);
3658 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
,
3659 MEM_CGROUP_CHARGE_TYPE_ANON
, false);
3664 * While swap-in, try_charge -> commit or cancel, the page is locked.
3665 * And when try_charge() successfully returns, one refcnt to memcg without
3666 * struct page_cgroup is acquired. This refcnt will be consumed by
3667 * "commit()" or removed by "cancel()"
3669 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3672 struct mem_cgroup
**memcgp
)
3674 struct mem_cgroup
*memcg
= NULL
;
3675 struct page_cgroup
*pc
;
3678 pc
= lookup_page_cgroup(page
);
3680 * Every swap fault against a single page tries to charge the
3681 * page, bail as early as possible. shmem_unuse() encounters
3682 * already charged pages, too. The USED bit is protected by
3683 * the page lock, which serializes swap cache removal, which
3684 * in turn serializes uncharging.
3686 if (PageCgroupUsed(pc
))
3688 if (do_swap_account
)
3689 memcg
= try_get_mem_cgroup_from_page(page
);
3691 memcg
= get_mem_cgroup_from_mm(mm
);
3692 ret
= mem_cgroup_try_charge(memcg
, mask
, 1);
3693 css_put(&memcg
->css
);
3695 memcg
= root_mem_cgroup
;
3703 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3704 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3706 if (mem_cgroup_disabled()) {
3711 * A racing thread's fault, or swapoff, may have already
3712 * updated the pte, and even removed page from swap cache: in
3713 * those cases unuse_pte()'s pte_same() test will fail; but
3714 * there's also a KSM case which does need to charge the page.
3716 if (!PageSwapCache(page
)) {
3717 struct mem_cgroup
*memcg
;
3719 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, 1);
3725 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3728 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3730 if (mem_cgroup_disabled())
3734 __mem_cgroup_cancel_charge(memcg
, 1);
3738 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3739 enum charge_type ctype
)
3741 if (mem_cgroup_disabled())
3746 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3748 * Now swap is on-memory. This means this page may be
3749 * counted both as mem and swap....double count.
3750 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3751 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3752 * may call delete_from_swap_cache() before reach here.
3754 if (do_swap_account
&& PageSwapCache(page
)) {
3755 swp_entry_t ent
= {.val
= page_private(page
)};
3756 mem_cgroup_uncharge_swap(ent
);
3760 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3761 struct mem_cgroup
*memcg
)
3763 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3764 MEM_CGROUP_CHARGE_TYPE_ANON
);
3767 int mem_cgroup_charge_file(struct page
*page
, struct mm_struct
*mm
,
3770 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3771 struct mem_cgroup
*memcg
;
3774 if (mem_cgroup_disabled())
3776 if (PageCompound(page
))
3779 if (PageSwapCache(page
)) { /* shmem */
3780 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3784 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3788 memcg
= mem_cgroup_try_charge_mm(mm
, gfp_mask
, 1);
3791 __mem_cgroup_commit_charge(memcg
, page
, 1, type
, false);
3795 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3796 unsigned int nr_pages
,
3797 const enum charge_type ctype
)
3799 struct memcg_batch_info
*batch
= NULL
;
3800 bool uncharge_memsw
= true;
3802 /* If swapout, usage of swap doesn't decrease */
3803 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3804 uncharge_memsw
= false;
3806 batch
= ¤t
->memcg_batch
;
3808 * In usual, we do css_get() when we remember memcg pointer.
3809 * But in this case, we keep res->usage until end of a series of
3810 * uncharges. Then, it's ok to ignore memcg's refcnt.
3813 batch
->memcg
= memcg
;
3815 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3816 * In those cases, all pages freed continuously can be expected to be in
3817 * the same cgroup and we have chance to coalesce uncharges.
3818 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3819 * because we want to do uncharge as soon as possible.
3822 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3823 goto direct_uncharge
;
3826 goto direct_uncharge
;
3829 * In typical case, batch->memcg == mem. This means we can
3830 * merge a series of uncharges to an uncharge of res_counter.
3831 * If not, we uncharge res_counter ony by one.
3833 if (batch
->memcg
!= memcg
)
3834 goto direct_uncharge
;
3835 /* remember freed charge and uncharge it later */
3838 batch
->memsw_nr_pages
++;
3841 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3843 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3844 if (unlikely(batch
->memcg
!= memcg
))
3845 memcg_oom_recover(memcg
);
3849 * uncharge if !page_mapped(page)
3851 static struct mem_cgroup
*
3852 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3855 struct mem_cgroup
*memcg
= NULL
;
3856 unsigned int nr_pages
= 1;
3857 struct page_cgroup
*pc
;
3860 if (mem_cgroup_disabled())
3863 if (PageTransHuge(page
)) {
3864 nr_pages
<<= compound_order(page
);
3865 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3868 * Check if our page_cgroup is valid
3870 pc
= lookup_page_cgroup(page
);
3871 if (unlikely(!PageCgroupUsed(pc
)))
3874 lock_page_cgroup(pc
);
3876 memcg
= pc
->mem_cgroup
;
3878 if (!PageCgroupUsed(pc
))
3881 anon
= PageAnon(page
);
3884 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3886 * Generally PageAnon tells if it's the anon statistics to be
3887 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3888 * used before page reached the stage of being marked PageAnon.
3892 case MEM_CGROUP_CHARGE_TYPE_DROP
:
3893 /* See mem_cgroup_prepare_migration() */
3894 if (page_mapped(page
))
3897 * Pages under migration may not be uncharged. But
3898 * end_migration() /must/ be the one uncharging the
3899 * unused post-migration page and so it has to call
3900 * here with the migration bit still set. See the
3901 * res_counter handling below.
3903 if (!end_migration
&& PageCgroupMigration(pc
))
3906 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
3907 if (!PageAnon(page
)) { /* Shared memory */
3908 if (page
->mapping
&& !page_is_file_cache(page
))
3910 } else if (page_mapped(page
)) /* Anon */
3917 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
3919 ClearPageCgroupUsed(pc
);
3921 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3922 * freed from LRU. This is safe because uncharged page is expected not
3923 * to be reused (freed soon). Exception is SwapCache, it's handled by
3924 * special functions.
3927 unlock_page_cgroup(pc
);
3929 * even after unlock, we have memcg->res.usage here and this memcg
3930 * will never be freed, so it's safe to call css_get().
3932 memcg_check_events(memcg
, page
);
3933 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
3934 mem_cgroup_swap_statistics(memcg
, true);
3935 css_get(&memcg
->css
);
3938 * Migration does not charge the res_counter for the
3939 * replacement page, so leave it alone when phasing out the
3940 * page that is unused after the migration.
3943 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
3948 unlock_page_cgroup(pc
);
3952 void mem_cgroup_uncharge_page(struct page
*page
)
3955 if (page_mapped(page
))
3957 VM_BUG_ON_PAGE(page
->mapping
&& !PageAnon(page
), page
);
3959 * If the page is in swap cache, uncharge should be deferred
3960 * to the swap path, which also properly accounts swap usage
3961 * and handles memcg lifetime.
3963 * Note that this check is not stable and reclaim may add the
3964 * page to swap cache at any time after this. However, if the
3965 * page is not in swap cache by the time page->mapcount hits
3966 * 0, there won't be any page table references to the swap
3967 * slot, and reclaim will free it and not actually write the
3970 if (PageSwapCache(page
))
3972 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
3975 void mem_cgroup_uncharge_cache_page(struct page
*page
)
3977 VM_BUG_ON_PAGE(page_mapped(page
), page
);
3978 VM_BUG_ON_PAGE(page
->mapping
, page
);
3979 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
3983 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3984 * In that cases, pages are freed continuously and we can expect pages
3985 * are in the same memcg. All these calls itself limits the number of
3986 * pages freed at once, then uncharge_start/end() is called properly.
3987 * This may be called prural(2) times in a context,
3990 void mem_cgroup_uncharge_start(void)
3992 current
->memcg_batch
.do_batch
++;
3993 /* We can do nest. */
3994 if (current
->memcg_batch
.do_batch
== 1) {
3995 current
->memcg_batch
.memcg
= NULL
;
3996 current
->memcg_batch
.nr_pages
= 0;
3997 current
->memcg_batch
.memsw_nr_pages
= 0;
4001 void mem_cgroup_uncharge_end(void)
4003 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4005 if (!batch
->do_batch
)
4009 if (batch
->do_batch
) /* If stacked, do nothing. */
4015 * This "batch->memcg" is valid without any css_get/put etc...
4016 * bacause we hide charges behind us.
4018 if (batch
->nr_pages
)
4019 res_counter_uncharge(&batch
->memcg
->res
,
4020 batch
->nr_pages
* PAGE_SIZE
);
4021 if (batch
->memsw_nr_pages
)
4022 res_counter_uncharge(&batch
->memcg
->memsw
,
4023 batch
->memsw_nr_pages
* PAGE_SIZE
);
4024 memcg_oom_recover(batch
->memcg
);
4025 /* forget this pointer (for sanity check) */
4026 batch
->memcg
= NULL
;
4031 * called after __delete_from_swap_cache() and drop "page" account.
4032 * memcg information is recorded to swap_cgroup of "ent"
4035 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4037 struct mem_cgroup
*memcg
;
4038 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4040 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4041 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4043 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4046 * record memcg information, if swapout && memcg != NULL,
4047 * css_get() was called in uncharge().
4049 if (do_swap_account
&& swapout
&& memcg
)
4050 swap_cgroup_record(ent
, mem_cgroup_id(memcg
));
4054 #ifdef CONFIG_MEMCG_SWAP
4056 * called from swap_entry_free(). remove record in swap_cgroup and
4057 * uncharge "memsw" account.
4059 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4061 struct mem_cgroup
*memcg
;
4064 if (!do_swap_account
)
4067 id
= swap_cgroup_record(ent
, 0);
4069 memcg
= mem_cgroup_lookup(id
);
4072 * We uncharge this because swap is freed. This memcg can
4073 * be obsolete one. We avoid calling css_tryget_online().
4075 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4076 mem_cgroup_swap_statistics(memcg
, false);
4077 css_put(&memcg
->css
);
4083 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4084 * @entry: swap entry to be moved
4085 * @from: mem_cgroup which the entry is moved from
4086 * @to: mem_cgroup which the entry is moved to
4088 * It succeeds only when the swap_cgroup's record for this entry is the same
4089 * as the mem_cgroup's id of @from.
4091 * Returns 0 on success, -EINVAL on failure.
4093 * The caller must have charged to @to, IOW, called res_counter_charge() about
4094 * both res and memsw, and called css_get().
4096 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4097 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4099 unsigned short old_id
, new_id
;
4101 old_id
= mem_cgroup_id(from
);
4102 new_id
= mem_cgroup_id(to
);
4104 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4105 mem_cgroup_swap_statistics(from
, false);
4106 mem_cgroup_swap_statistics(to
, true);
4108 * This function is only called from task migration context now.
4109 * It postpones res_counter and refcount handling till the end
4110 * of task migration(mem_cgroup_clear_mc()) for performance
4111 * improvement. But we cannot postpone css_get(to) because if
4112 * the process that has been moved to @to does swap-in, the
4113 * refcount of @to might be decreased to 0.
4115 * We are in attach() phase, so the cgroup is guaranteed to be
4116 * alive, so we can just call css_get().
4124 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4125 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4132 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4135 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4136 struct mem_cgroup
**memcgp
)
4138 struct mem_cgroup
*memcg
= NULL
;
4139 unsigned int nr_pages
= 1;
4140 struct page_cgroup
*pc
;
4141 enum charge_type ctype
;
4145 if (mem_cgroup_disabled())
4148 if (PageTransHuge(page
))
4149 nr_pages
<<= compound_order(page
);
4151 pc
= lookup_page_cgroup(page
);
4152 lock_page_cgroup(pc
);
4153 if (PageCgroupUsed(pc
)) {
4154 memcg
= pc
->mem_cgroup
;
4155 css_get(&memcg
->css
);
4157 * At migrating an anonymous page, its mapcount goes down
4158 * to 0 and uncharge() will be called. But, even if it's fully
4159 * unmapped, migration may fail and this page has to be
4160 * charged again. We set MIGRATION flag here and delay uncharge
4161 * until end_migration() is called
4163 * Corner Case Thinking
4165 * When the old page was mapped as Anon and it's unmap-and-freed
4166 * while migration was ongoing.
4167 * If unmap finds the old page, uncharge() of it will be delayed
4168 * until end_migration(). If unmap finds a new page, it's
4169 * uncharged when it make mapcount to be 1->0. If unmap code
4170 * finds swap_migration_entry, the new page will not be mapped
4171 * and end_migration() will find it(mapcount==0).
4174 * When the old page was mapped but migraion fails, the kernel
4175 * remaps it. A charge for it is kept by MIGRATION flag even
4176 * if mapcount goes down to 0. We can do remap successfully
4177 * without charging it again.
4180 * The "old" page is under lock_page() until the end of
4181 * migration, so, the old page itself will not be swapped-out.
4182 * If the new page is swapped out before end_migraton, our
4183 * hook to usual swap-out path will catch the event.
4186 SetPageCgroupMigration(pc
);
4188 unlock_page_cgroup(pc
);
4190 * If the page is not charged at this point,
4198 * We charge new page before it's used/mapped. So, even if unlock_page()
4199 * is called before end_migration, we can catch all events on this new
4200 * page. In the case new page is migrated but not remapped, new page's
4201 * mapcount will be finally 0 and we call uncharge in end_migration().
4204 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4206 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4208 * The page is committed to the memcg, but it's not actually
4209 * charged to the res_counter since we plan on replacing the
4210 * old one and only one page is going to be left afterwards.
4212 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4215 /* remove redundant charge if migration failed*/
4216 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4217 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4219 struct page
*used
, *unused
;
4220 struct page_cgroup
*pc
;
4226 if (!migration_ok
) {
4233 anon
= PageAnon(used
);
4234 __mem_cgroup_uncharge_common(unused
,
4235 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4236 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4238 css_put(&memcg
->css
);
4240 * We disallowed uncharge of pages under migration because mapcount
4241 * of the page goes down to zero, temporarly.
4242 * Clear the flag and check the page should be charged.
4244 pc
= lookup_page_cgroup(oldpage
);
4245 lock_page_cgroup(pc
);
4246 ClearPageCgroupMigration(pc
);
4247 unlock_page_cgroup(pc
);
4250 * If a page is a file cache, radix-tree replacement is very atomic
4251 * and we can skip this check. When it was an Anon page, its mapcount
4252 * goes down to 0. But because we added MIGRATION flage, it's not
4253 * uncharged yet. There are several case but page->mapcount check
4254 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4255 * check. (see prepare_charge() also)
4258 mem_cgroup_uncharge_page(used
);
4262 * At replace page cache, newpage is not under any memcg but it's on
4263 * LRU. So, this function doesn't touch res_counter but handles LRU
4264 * in correct way. Both pages are locked so we cannot race with uncharge.
4266 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4267 struct page
*newpage
)
4269 struct mem_cgroup
*memcg
= NULL
;
4270 struct page_cgroup
*pc
;
4271 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4273 if (mem_cgroup_disabled())
4276 pc
= lookup_page_cgroup(oldpage
);
4277 /* fix accounting on old pages */
4278 lock_page_cgroup(pc
);
4279 if (PageCgroupUsed(pc
)) {
4280 memcg
= pc
->mem_cgroup
;
4281 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4282 ClearPageCgroupUsed(pc
);
4284 unlock_page_cgroup(pc
);
4287 * When called from shmem_replace_page(), in some cases the
4288 * oldpage has already been charged, and in some cases not.
4293 * Even if newpage->mapping was NULL before starting replacement,
4294 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4295 * LRU while we overwrite pc->mem_cgroup.
4297 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4300 #ifdef CONFIG_DEBUG_VM
4301 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4303 struct page_cgroup
*pc
;
4305 pc
= lookup_page_cgroup(page
);
4307 * Can be NULL while feeding pages into the page allocator for
4308 * the first time, i.e. during boot or memory hotplug;
4309 * or when mem_cgroup_disabled().
4311 if (likely(pc
) && PageCgroupUsed(pc
))
4316 bool mem_cgroup_bad_page_check(struct page
*page
)
4318 if (mem_cgroup_disabled())
4321 return lookup_page_cgroup_used(page
) != NULL
;
4324 void mem_cgroup_print_bad_page(struct page
*page
)
4326 struct page_cgroup
*pc
;
4328 pc
= lookup_page_cgroup_used(page
);
4330 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4331 pc
, pc
->flags
, pc
->mem_cgroup
);
4336 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4337 unsigned long long val
)
4340 u64 memswlimit
, memlimit
;
4342 int children
= mem_cgroup_count_children(memcg
);
4343 u64 curusage
, oldusage
;
4347 * For keeping hierarchical_reclaim simple, how long we should retry
4348 * is depends on callers. We set our retry-count to be function
4349 * of # of children which we should visit in this loop.
4351 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4353 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4356 while (retry_count
) {
4357 if (signal_pending(current
)) {
4362 * Rather than hide all in some function, I do this in
4363 * open coded manner. You see what this really does.
4364 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4366 mutex_lock(&set_limit_mutex
);
4367 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4368 if (memswlimit
< val
) {
4370 mutex_unlock(&set_limit_mutex
);
4374 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4378 ret
= res_counter_set_limit(&memcg
->res
, val
);
4380 if (memswlimit
== val
)
4381 memcg
->memsw_is_minimum
= true;
4383 memcg
->memsw_is_minimum
= false;
4385 mutex_unlock(&set_limit_mutex
);
4390 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4391 MEM_CGROUP_RECLAIM_SHRINK
);
4392 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4393 /* Usage is reduced ? */
4394 if (curusage
>= oldusage
)
4397 oldusage
= curusage
;
4399 if (!ret
&& enlarge
)
4400 memcg_oom_recover(memcg
);
4405 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4406 unsigned long long val
)
4409 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4410 int children
= mem_cgroup_count_children(memcg
);
4414 /* see mem_cgroup_resize_res_limit */
4415 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4416 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4417 while (retry_count
) {
4418 if (signal_pending(current
)) {
4423 * Rather than hide all in some function, I do this in
4424 * open coded manner. You see what this really does.
4425 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4427 mutex_lock(&set_limit_mutex
);
4428 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4429 if (memlimit
> val
) {
4431 mutex_unlock(&set_limit_mutex
);
4434 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4435 if (memswlimit
< val
)
4437 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4439 if (memlimit
== val
)
4440 memcg
->memsw_is_minimum
= true;
4442 memcg
->memsw_is_minimum
= false;
4444 mutex_unlock(&set_limit_mutex
);
4449 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4450 MEM_CGROUP_RECLAIM_NOSWAP
|
4451 MEM_CGROUP_RECLAIM_SHRINK
);
4452 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4453 /* Usage is reduced ? */
4454 if (curusage
>= oldusage
)
4457 oldusage
= curusage
;
4459 if (!ret
&& enlarge
)
4460 memcg_oom_recover(memcg
);
4464 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
4466 unsigned long *total_scanned
)
4468 unsigned long nr_reclaimed
= 0;
4469 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
4470 unsigned long reclaimed
;
4472 struct mem_cgroup_tree_per_zone
*mctz
;
4473 unsigned long long excess
;
4474 unsigned long nr_scanned
;
4479 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
4481 * This loop can run a while, specially if mem_cgroup's continuously
4482 * keep exceeding their soft limit and putting the system under
4489 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
4494 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
4495 gfp_mask
, &nr_scanned
);
4496 nr_reclaimed
+= reclaimed
;
4497 *total_scanned
+= nr_scanned
;
4498 spin_lock(&mctz
->lock
);
4501 * If we failed to reclaim anything from this memory cgroup
4502 * it is time to move on to the next cgroup
4508 * Loop until we find yet another one.
4510 * By the time we get the soft_limit lock
4511 * again, someone might have aded the
4512 * group back on the RB tree. Iterate to
4513 * make sure we get a different mem.
4514 * mem_cgroup_largest_soft_limit_node returns
4515 * NULL if no other cgroup is present on
4519 __mem_cgroup_largest_soft_limit_node(mctz
);
4521 css_put(&next_mz
->memcg
->css
);
4522 else /* next_mz == NULL or other memcg */
4526 __mem_cgroup_remove_exceeded(mz
, mctz
);
4527 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
4529 * One school of thought says that we should not add
4530 * back the node to the tree if reclaim returns 0.
4531 * But our reclaim could return 0, simply because due
4532 * to priority we are exposing a smaller subset of
4533 * memory to reclaim from. Consider this as a longer
4536 /* If excess == 0, no tree ops */
4537 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
4538 spin_unlock(&mctz
->lock
);
4539 css_put(&mz
->memcg
->css
);
4542 * Could not reclaim anything and there are no more
4543 * mem cgroups to try or we seem to be looping without
4544 * reclaiming anything.
4546 if (!nr_reclaimed
&&
4548 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
4550 } while (!nr_reclaimed
);
4552 css_put(&next_mz
->memcg
->css
);
4553 return nr_reclaimed
;
4557 * mem_cgroup_force_empty_list - clears LRU of a group
4558 * @memcg: group to clear
4561 * @lru: lru to to clear
4563 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4564 * reclaim the pages page themselves - pages are moved to the parent (or root)
4567 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4568 int node
, int zid
, enum lru_list lru
)
4570 struct lruvec
*lruvec
;
4571 unsigned long flags
;
4572 struct list_head
*list
;
4576 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4577 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4578 list
= &lruvec
->lists
[lru
];
4582 struct page_cgroup
*pc
;
4585 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4586 if (list_empty(list
)) {
4587 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4590 page
= list_entry(list
->prev
, struct page
, lru
);
4592 list_move(&page
->lru
, list
);
4594 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4597 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4599 pc
= lookup_page_cgroup(page
);
4601 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4602 /* found lock contention or "pc" is obsolete. */
4607 } while (!list_empty(list
));
4611 * make mem_cgroup's charge to be 0 if there is no task by moving
4612 * all the charges and pages to the parent.
4613 * This enables deleting this mem_cgroup.
4615 * Caller is responsible for holding css reference on the memcg.
4617 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4623 /* This is for making all *used* pages to be on LRU. */
4624 lru_add_drain_all();
4625 drain_all_stock_sync(memcg
);
4626 mem_cgroup_start_move(memcg
);
4627 for_each_node_state(node
, N_MEMORY
) {
4628 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4631 mem_cgroup_force_empty_list(memcg
,
4636 mem_cgroup_end_move(memcg
);
4637 memcg_oom_recover(memcg
);
4641 * Kernel memory may not necessarily be trackable to a specific
4642 * process. So they are not migrated, and therefore we can't
4643 * expect their value to drop to 0 here.
4644 * Having res filled up with kmem only is enough.
4646 * This is a safety check because mem_cgroup_force_empty_list
4647 * could have raced with mem_cgroup_replace_page_cache callers
4648 * so the lru seemed empty but the page could have been added
4649 * right after the check. RES_USAGE should be safe as we always
4650 * charge before adding to the LRU.
4652 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4653 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4654 } while (usage
> 0);
4658 * Test whether @memcg has children, dead or alive. Note that this
4659 * function doesn't care whether @memcg has use_hierarchy enabled and
4660 * returns %true if there are child csses according to the cgroup
4661 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
4663 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4668 * The lock does not prevent addition or deletion of children, but
4669 * it prevents a new child from being initialized based on this
4670 * parent in css_online(), so it's enough to decide whether
4671 * hierarchically inherited attributes can still be changed or not.
4673 lockdep_assert_held(&memcg_create_mutex
);
4676 ret
= css_next_child(NULL
, &memcg
->css
);
4682 * Reclaims as many pages from the given memcg as possible and moves
4683 * the rest to the parent.
4685 * Caller is responsible for holding css reference for memcg.
4687 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4689 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4691 /* we call try-to-free pages for make this cgroup empty */
4692 lru_add_drain_all();
4693 /* try to free all pages in this cgroup */
4694 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4697 if (signal_pending(current
))
4700 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4704 /* maybe some writeback is necessary */
4705 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4713 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
4714 char *buf
, size_t nbytes
,
4717 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4719 if (mem_cgroup_is_root(memcg
))
4721 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
4724 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4727 return mem_cgroup_from_css(css
)->use_hierarchy
;
4730 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4731 struct cftype
*cft
, u64 val
)
4734 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4735 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
4737 mutex_lock(&memcg_create_mutex
);
4739 if (memcg
->use_hierarchy
== val
)
4743 * If parent's use_hierarchy is set, we can't make any modifications
4744 * in the child subtrees. If it is unset, then the change can
4745 * occur, provided the current cgroup has no children.
4747 * For the root cgroup, parent_mem is NULL, we allow value to be
4748 * set if there are no children.
4750 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4751 (val
== 1 || val
== 0)) {
4752 if (!memcg_has_children(memcg
))
4753 memcg
->use_hierarchy
= val
;
4760 mutex_unlock(&memcg_create_mutex
);
4765 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
4768 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4769 enum res_type type
= MEMFILE_TYPE(cft
->private);
4770 int name
= MEMFILE_ATTR(cft
->private);
4774 return res_counter_read_u64(&memcg
->res
, name
);
4776 return res_counter_read_u64(&memcg
->memsw
, name
);
4778 return res_counter_read_u64(&memcg
->kmem
, name
);
4785 #ifdef CONFIG_MEMCG_KMEM
4786 /* should be called with activate_kmem_mutex held */
4787 static int __memcg_activate_kmem(struct mem_cgroup
*memcg
,
4788 unsigned long long limit
)
4793 if (memcg_kmem_is_active(memcg
))
4797 * We are going to allocate memory for data shared by all memory
4798 * cgroups so let's stop accounting here.
4800 memcg_stop_kmem_account();
4803 * For simplicity, we won't allow this to be disabled. It also can't
4804 * be changed if the cgroup has children already, or if tasks had
4807 * If tasks join before we set the limit, a person looking at
4808 * kmem.usage_in_bytes will have no way to determine when it took
4809 * place, which makes the value quite meaningless.
4811 * After it first became limited, changes in the value of the limit are
4812 * of course permitted.
4814 mutex_lock(&memcg_create_mutex
);
4815 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
4816 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
4818 mutex_unlock(&memcg_create_mutex
);
4822 memcg_id
= ida_simple_get(&kmem_limited_groups
,
4823 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
4830 * Make sure we have enough space for this cgroup in each root cache's
4833 mutex_lock(&memcg_slab_mutex
);
4834 err
= memcg_update_all_caches(memcg_id
+ 1);
4835 mutex_unlock(&memcg_slab_mutex
);
4839 memcg
->kmemcg_id
= memcg_id
;
4840 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
4843 * We couldn't have accounted to this cgroup, because it hasn't got the
4844 * active bit set yet, so this should succeed.
4846 err
= res_counter_set_limit(&memcg
->kmem
, limit
);
4849 static_key_slow_inc(&memcg_kmem_enabled_key
);
4851 * Setting the active bit after enabling static branching will
4852 * guarantee no one starts accounting before all call sites are
4855 memcg_kmem_set_active(memcg
);
4857 memcg_resume_kmem_account();
4861 ida_simple_remove(&kmem_limited_groups
, memcg_id
);
4865 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
4866 unsigned long long limit
)
4870 mutex_lock(&activate_kmem_mutex
);
4871 ret
= __memcg_activate_kmem(memcg
, limit
);
4872 mutex_unlock(&activate_kmem_mutex
);
4876 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4877 unsigned long long val
)
4881 if (!memcg_kmem_is_active(memcg
))
4882 ret
= memcg_activate_kmem(memcg
, val
);
4884 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4888 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4891 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4896 mutex_lock(&activate_kmem_mutex
);
4898 * If the parent cgroup is not kmem-active now, it cannot be activated
4899 * after this point, because it has at least one child already.
4901 if (memcg_kmem_is_active(parent
))
4902 ret
= __memcg_activate_kmem(memcg
, RES_COUNTER_MAX
);
4903 mutex_unlock(&activate_kmem_mutex
);
4907 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4908 unsigned long long val
)
4912 #endif /* CONFIG_MEMCG_KMEM */
4915 * The user of this function is...
4918 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
4919 char *buf
, size_t nbytes
, loff_t off
)
4921 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4924 unsigned long long val
;
4927 buf
= strstrip(buf
);
4928 type
= MEMFILE_TYPE(of_cft(of
)->private);
4929 name
= MEMFILE_ATTR(of_cft(of
)->private);
4933 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4937 /* This function does all necessary parse...reuse it */
4938 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4942 ret
= mem_cgroup_resize_limit(memcg
, val
);
4943 else if (type
== _MEMSWAP
)
4944 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
4945 else if (type
== _KMEM
)
4946 ret
= memcg_update_kmem_limit(memcg
, val
);
4950 case RES_SOFT_LIMIT
:
4951 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4955 * For memsw, soft limits are hard to implement in terms
4956 * of semantics, for now, we support soft limits for
4957 * control without swap
4960 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
4965 ret
= -EINVAL
; /* should be BUG() ? */
4968 return ret
?: nbytes
;
4971 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
4972 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
4974 unsigned long long min_limit
, min_memsw_limit
, tmp
;
4976 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4977 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4978 if (!memcg
->use_hierarchy
)
4981 while (memcg
->css
.parent
) {
4982 memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
4983 if (!memcg
->use_hierarchy
)
4985 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4986 min_limit
= min(min_limit
, tmp
);
4987 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4988 min_memsw_limit
= min(min_memsw_limit
, tmp
);
4991 *mem_limit
= min_limit
;
4992 *memsw_limit
= min_memsw_limit
;
4995 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
4996 size_t nbytes
, loff_t off
)
4998 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5002 type
= MEMFILE_TYPE(of_cft(of
)->private);
5003 name
= MEMFILE_ATTR(of_cft(of
)->private);
5008 res_counter_reset_max(&memcg
->res
);
5009 else if (type
== _MEMSWAP
)
5010 res_counter_reset_max(&memcg
->memsw
);
5011 else if (type
== _KMEM
)
5012 res_counter_reset_max(&memcg
->kmem
);
5018 res_counter_reset_failcnt(&memcg
->res
);
5019 else if (type
== _MEMSWAP
)
5020 res_counter_reset_failcnt(&memcg
->memsw
);
5021 else if (type
== _KMEM
)
5022 res_counter_reset_failcnt(&memcg
->kmem
);
5031 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5034 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5038 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5039 struct cftype
*cft
, u64 val
)
5041 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5043 if (val
>= (1 << NR_MOVE_TYPE
))
5047 * No kind of locking is needed in here, because ->can_attach() will
5048 * check this value once in the beginning of the process, and then carry
5049 * on with stale data. This means that changes to this value will only
5050 * affect task migrations starting after the change.
5052 memcg
->move_charge_at_immigrate
= val
;
5056 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5057 struct cftype
*cft
, u64 val
)
5064 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
5068 unsigned int lru_mask
;
5071 static const struct numa_stat stats
[] = {
5072 { "total", LRU_ALL
},
5073 { "file", LRU_ALL_FILE
},
5074 { "anon", LRU_ALL_ANON
},
5075 { "unevictable", BIT(LRU_UNEVICTABLE
) },
5077 const struct numa_stat
*stat
;
5080 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5082 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5083 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
5084 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
5085 for_each_node_state(nid
, N_MEMORY
) {
5086 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5088 seq_printf(m
, " N%d=%lu", nid
, nr
);
5093 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
5094 struct mem_cgroup
*iter
;
5097 for_each_mem_cgroup_tree(iter
, memcg
)
5098 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
5099 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
5100 for_each_node_state(nid
, N_MEMORY
) {
5102 for_each_mem_cgroup_tree(iter
, memcg
)
5103 nr
+= mem_cgroup_node_nr_lru_pages(
5104 iter
, nid
, stat
->lru_mask
);
5105 seq_printf(m
, " N%d=%lu", nid
, nr
);
5112 #endif /* CONFIG_NUMA */
5114 static inline void mem_cgroup_lru_names_not_uptodate(void)
5116 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5119 static int memcg_stat_show(struct seq_file
*m
, void *v
)
5121 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5122 struct mem_cgroup
*mi
;
5125 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5126 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5128 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5129 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5132 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5133 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5134 mem_cgroup_read_events(memcg
, i
));
5136 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5137 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5138 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5140 /* Hierarchical information */
5142 unsigned long long limit
, memsw_limit
;
5143 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5144 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5145 if (do_swap_account
)
5146 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5150 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5153 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5155 for_each_mem_cgroup_tree(mi
, memcg
)
5156 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5157 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5160 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5161 unsigned long long val
= 0;
5163 for_each_mem_cgroup_tree(mi
, memcg
)
5164 val
+= mem_cgroup_read_events(mi
, i
);
5165 seq_printf(m
, "total_%s %llu\n",
5166 mem_cgroup_events_names
[i
], val
);
5169 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5170 unsigned long long val
= 0;
5172 for_each_mem_cgroup_tree(mi
, memcg
)
5173 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5174 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5177 #ifdef CONFIG_DEBUG_VM
5180 struct mem_cgroup_per_zone
*mz
;
5181 struct zone_reclaim_stat
*rstat
;
5182 unsigned long recent_rotated
[2] = {0, 0};
5183 unsigned long recent_scanned
[2] = {0, 0};
5185 for_each_online_node(nid
)
5186 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5187 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
5188 rstat
= &mz
->lruvec
.reclaim_stat
;
5190 recent_rotated
[0] += rstat
->recent_rotated
[0];
5191 recent_rotated
[1] += rstat
->recent_rotated
[1];
5192 recent_scanned
[0] += rstat
->recent_scanned
[0];
5193 recent_scanned
[1] += rstat
->recent_scanned
[1];
5195 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5196 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5197 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5198 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5205 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5208 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5210 return mem_cgroup_swappiness(memcg
);
5213 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5214 struct cftype
*cft
, u64 val
)
5216 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5222 memcg
->swappiness
= val
;
5224 vm_swappiness
= val
;
5229 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5231 struct mem_cgroup_threshold_ary
*t
;
5237 t
= rcu_dereference(memcg
->thresholds
.primary
);
5239 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5245 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5247 usage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5250 * current_threshold points to threshold just below or equal to usage.
5251 * If it's not true, a threshold was crossed after last
5252 * call of __mem_cgroup_threshold().
5254 i
= t
->current_threshold
;
5257 * Iterate backward over array of thresholds starting from
5258 * current_threshold and check if a threshold is crossed.
5259 * If none of thresholds below usage is crossed, we read
5260 * only one element of the array here.
5262 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5263 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5265 /* i = current_threshold + 1 */
5269 * Iterate forward over array of thresholds starting from
5270 * current_threshold+1 and check if a threshold is crossed.
5271 * If none of thresholds above usage is crossed, we read
5272 * only one element of the array here.
5274 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5275 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5277 /* Update current_threshold */
5278 t
->current_threshold
= i
- 1;
5283 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5286 __mem_cgroup_threshold(memcg
, false);
5287 if (do_swap_account
)
5288 __mem_cgroup_threshold(memcg
, true);
5290 memcg
= parent_mem_cgroup(memcg
);
5294 static int compare_thresholds(const void *a
, const void *b
)
5296 const struct mem_cgroup_threshold
*_a
= a
;
5297 const struct mem_cgroup_threshold
*_b
= b
;
5299 if (_a
->threshold
> _b
->threshold
)
5302 if (_a
->threshold
< _b
->threshold
)
5308 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5310 struct mem_cgroup_eventfd_list
*ev
;
5312 spin_lock(&memcg_oom_lock
);
5314 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5315 eventfd_signal(ev
->eventfd
, 1);
5317 spin_unlock(&memcg_oom_lock
);
5321 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5323 struct mem_cgroup
*iter
;
5325 for_each_mem_cgroup_tree(iter
, memcg
)
5326 mem_cgroup_oom_notify_cb(iter
);
5329 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5330 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
5332 struct mem_cgroup_thresholds
*thresholds
;
5333 struct mem_cgroup_threshold_ary
*new;
5334 u64 threshold
, usage
;
5337 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5341 mutex_lock(&memcg
->thresholds_lock
);
5344 thresholds
= &memcg
->thresholds
;
5345 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5346 } else if (type
== _MEMSWAP
) {
5347 thresholds
= &memcg
->memsw_thresholds
;
5348 usage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5352 /* Check if a threshold crossed before adding a new one */
5353 if (thresholds
->primary
)
5354 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5356 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5358 /* Allocate memory for new array of thresholds */
5359 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5367 /* Copy thresholds (if any) to new array */
5368 if (thresholds
->primary
) {
5369 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5370 sizeof(struct mem_cgroup_threshold
));
5373 /* Add new threshold */
5374 new->entries
[size
- 1].eventfd
= eventfd
;
5375 new->entries
[size
- 1].threshold
= threshold
;
5377 /* Sort thresholds. Registering of new threshold isn't time-critical */
5378 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5379 compare_thresholds
, NULL
);
5381 /* Find current threshold */
5382 new->current_threshold
= -1;
5383 for (i
= 0; i
< size
; i
++) {
5384 if (new->entries
[i
].threshold
<= usage
) {
5386 * new->current_threshold will not be used until
5387 * rcu_assign_pointer(), so it's safe to increment
5390 ++new->current_threshold
;
5395 /* Free old spare buffer and save old primary buffer as spare */
5396 kfree(thresholds
->spare
);
5397 thresholds
->spare
= thresholds
->primary
;
5399 rcu_assign_pointer(thresholds
->primary
, new);
5401 /* To be sure that nobody uses thresholds */
5405 mutex_unlock(&memcg
->thresholds_lock
);
5410 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5411 struct eventfd_ctx
*eventfd
, const char *args
)
5413 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
5416 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
5417 struct eventfd_ctx
*eventfd
, const char *args
)
5419 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
5422 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5423 struct eventfd_ctx
*eventfd
, enum res_type type
)
5425 struct mem_cgroup_thresholds
*thresholds
;
5426 struct mem_cgroup_threshold_ary
*new;
5430 mutex_lock(&memcg
->thresholds_lock
);
5433 thresholds
= &memcg
->thresholds
;
5434 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
5435 } else if (type
== _MEMSWAP
) {
5436 thresholds
= &memcg
->memsw_thresholds
;
5437 usage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
5441 if (!thresholds
->primary
)
5444 /* Check if a threshold crossed before removing */
5445 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5447 /* Calculate new number of threshold */
5449 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5450 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5454 new = thresholds
->spare
;
5456 /* Set thresholds array to NULL if we don't have thresholds */
5465 /* Copy thresholds and find current threshold */
5466 new->current_threshold
= -1;
5467 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5468 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5471 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5472 if (new->entries
[j
].threshold
<= usage
) {
5474 * new->current_threshold will not be used
5475 * until rcu_assign_pointer(), so it's safe to increment
5478 ++new->current_threshold
;
5484 /* Swap primary and spare array */
5485 thresholds
->spare
= thresholds
->primary
;
5486 /* If all events are unregistered, free the spare array */
5488 kfree(thresholds
->spare
);
5489 thresholds
->spare
= NULL
;
5492 rcu_assign_pointer(thresholds
->primary
, new);
5494 /* To be sure that nobody uses thresholds */
5497 mutex_unlock(&memcg
->thresholds_lock
);
5500 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5501 struct eventfd_ctx
*eventfd
)
5503 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
5506 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
5507 struct eventfd_ctx
*eventfd
)
5509 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
5512 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
5513 struct eventfd_ctx
*eventfd
, const char *args
)
5515 struct mem_cgroup_eventfd_list
*event
;
5517 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5521 spin_lock(&memcg_oom_lock
);
5523 event
->eventfd
= eventfd
;
5524 list_add(&event
->list
, &memcg
->oom_notify
);
5526 /* already in OOM ? */
5527 if (atomic_read(&memcg
->under_oom
))
5528 eventfd_signal(eventfd
, 1);
5529 spin_unlock(&memcg_oom_lock
);
5534 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
5535 struct eventfd_ctx
*eventfd
)
5537 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5539 spin_lock(&memcg_oom_lock
);
5541 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5542 if (ev
->eventfd
== eventfd
) {
5543 list_del(&ev
->list
);
5548 spin_unlock(&memcg_oom_lock
);
5551 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
5553 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
5555 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
5556 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
5560 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5561 struct cftype
*cft
, u64 val
)
5563 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5565 /* cannot set to root cgroup and only 0 and 1 are allowed */
5566 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
5569 memcg
->oom_kill_disable
= val
;
5571 memcg_oom_recover(memcg
);
5576 #ifdef CONFIG_MEMCG_KMEM
5577 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5581 memcg
->kmemcg_id
= -1;
5582 ret
= memcg_propagate_kmem(memcg
);
5586 return mem_cgroup_sockets_init(memcg
, ss
);
5589 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5591 mem_cgroup_sockets_destroy(memcg
);
5594 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5596 if (!memcg_kmem_is_active(memcg
))
5600 * kmem charges can outlive the cgroup. In the case of slab
5601 * pages, for instance, a page contain objects from various
5602 * processes. As we prevent from taking a reference for every
5603 * such allocation we have to be careful when doing uncharge
5604 * (see memcg_uncharge_kmem) and here during offlining.
5606 * The idea is that that only the _last_ uncharge which sees
5607 * the dead memcg will drop the last reference. An additional
5608 * reference is taken here before the group is marked dead
5609 * which is then paired with css_put during uncharge resp. here.
5611 * Although this might sound strange as this path is called from
5612 * css_offline() when the referencemight have dropped down to 0 and
5613 * shouldn't be incremented anymore (css_tryget_online() would
5614 * fail) we do not have other options because of the kmem
5615 * allocations lifetime.
5617 css_get(&memcg
->css
);
5619 memcg_kmem_mark_dead(memcg
);
5621 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5624 if (memcg_kmem_test_and_clear_dead(memcg
))
5625 css_put(&memcg
->css
);
5628 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5633 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5637 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5643 * DO NOT USE IN NEW FILES.
5645 * "cgroup.event_control" implementation.
5647 * This is way over-engineered. It tries to support fully configurable
5648 * events for each user. Such level of flexibility is completely
5649 * unnecessary especially in the light of the planned unified hierarchy.
5651 * Please deprecate this and replace with something simpler if at all
5656 * Unregister event and free resources.
5658 * Gets called from workqueue.
5660 static void memcg_event_remove(struct work_struct
*work
)
5662 struct mem_cgroup_event
*event
=
5663 container_of(work
, struct mem_cgroup_event
, remove
);
5664 struct mem_cgroup
*memcg
= event
->memcg
;
5666 remove_wait_queue(event
->wqh
, &event
->wait
);
5668 event
->unregister_event(memcg
, event
->eventfd
);
5670 /* Notify userspace the event is going away. */
5671 eventfd_signal(event
->eventfd
, 1);
5673 eventfd_ctx_put(event
->eventfd
);
5675 css_put(&memcg
->css
);
5679 * Gets called on POLLHUP on eventfd when user closes it.
5681 * Called with wqh->lock held and interrupts disabled.
5683 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
5684 int sync
, void *key
)
5686 struct mem_cgroup_event
*event
=
5687 container_of(wait
, struct mem_cgroup_event
, wait
);
5688 struct mem_cgroup
*memcg
= event
->memcg
;
5689 unsigned long flags
= (unsigned long)key
;
5691 if (flags
& POLLHUP
) {
5693 * If the event has been detached at cgroup removal, we
5694 * can simply return knowing the other side will cleanup
5697 * We can't race against event freeing since the other
5698 * side will require wqh->lock via remove_wait_queue(),
5701 spin_lock(&memcg
->event_list_lock
);
5702 if (!list_empty(&event
->list
)) {
5703 list_del_init(&event
->list
);
5705 * We are in atomic context, but cgroup_event_remove()
5706 * may sleep, so we have to call it in workqueue.
5708 schedule_work(&event
->remove
);
5710 spin_unlock(&memcg
->event_list_lock
);
5716 static void memcg_event_ptable_queue_proc(struct file
*file
,
5717 wait_queue_head_t
*wqh
, poll_table
*pt
)
5719 struct mem_cgroup_event
*event
=
5720 container_of(pt
, struct mem_cgroup_event
, pt
);
5723 add_wait_queue(wqh
, &event
->wait
);
5727 * DO NOT USE IN NEW FILES.
5729 * Parse input and register new cgroup event handler.
5731 * Input must be in format '<event_fd> <control_fd> <args>'.
5732 * Interpretation of args is defined by control file implementation.
5734 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
5735 char *buf
, size_t nbytes
, loff_t off
)
5737 struct cgroup_subsys_state
*css
= of_css(of
);
5738 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5739 struct mem_cgroup_event
*event
;
5740 struct cgroup_subsys_state
*cfile_css
;
5741 unsigned int efd
, cfd
;
5748 buf
= strstrip(buf
);
5750 efd
= simple_strtoul(buf
, &endp
, 10);
5755 cfd
= simple_strtoul(buf
, &endp
, 10);
5756 if ((*endp
!= ' ') && (*endp
!= '\0'))
5760 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5764 event
->memcg
= memcg
;
5765 INIT_LIST_HEAD(&event
->list
);
5766 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
5767 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
5768 INIT_WORK(&event
->remove
, memcg_event_remove
);
5776 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
5777 if (IS_ERR(event
->eventfd
)) {
5778 ret
= PTR_ERR(event
->eventfd
);
5785 goto out_put_eventfd
;
5788 /* the process need read permission on control file */
5789 /* AV: shouldn't we check that it's been opened for read instead? */
5790 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
5795 * Determine the event callbacks and set them in @event. This used
5796 * to be done via struct cftype but cgroup core no longer knows
5797 * about these events. The following is crude but the whole thing
5798 * is for compatibility anyway.
5800 * DO NOT ADD NEW FILES.
5802 name
= cfile
.file
->f_dentry
->d_name
.name
;
5804 if (!strcmp(name
, "memory.usage_in_bytes")) {
5805 event
->register_event
= mem_cgroup_usage_register_event
;
5806 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
5807 } else if (!strcmp(name
, "memory.oom_control")) {
5808 event
->register_event
= mem_cgroup_oom_register_event
;
5809 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
5810 } else if (!strcmp(name
, "memory.pressure_level")) {
5811 event
->register_event
= vmpressure_register_event
;
5812 event
->unregister_event
= vmpressure_unregister_event
;
5813 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
5814 event
->register_event
= memsw_cgroup_usage_register_event
;
5815 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
5822 * Verify @cfile should belong to @css. Also, remaining events are
5823 * automatically removed on cgroup destruction but the removal is
5824 * asynchronous, so take an extra ref on @css.
5826 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_dentry
->d_parent
,
5827 &memory_cgrp_subsys
);
5829 if (IS_ERR(cfile_css
))
5831 if (cfile_css
!= css
) {
5836 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
5840 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
5842 spin_lock(&memcg
->event_list_lock
);
5843 list_add(&event
->list
, &memcg
->event_list
);
5844 spin_unlock(&memcg
->event_list_lock
);
5856 eventfd_ctx_put(event
->eventfd
);
5865 static struct cftype mem_cgroup_files
[] = {
5867 .name
= "usage_in_bytes",
5868 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5869 .read_u64
= mem_cgroup_read_u64
,
5872 .name
= "max_usage_in_bytes",
5873 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5874 .write
= mem_cgroup_reset
,
5875 .read_u64
= mem_cgroup_read_u64
,
5878 .name
= "limit_in_bytes",
5879 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5880 .write
= mem_cgroup_write
,
5881 .read_u64
= mem_cgroup_read_u64
,
5884 .name
= "soft_limit_in_bytes",
5885 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5886 .write
= mem_cgroup_write
,
5887 .read_u64
= mem_cgroup_read_u64
,
5891 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5892 .write
= mem_cgroup_reset
,
5893 .read_u64
= mem_cgroup_read_u64
,
5897 .seq_show
= memcg_stat_show
,
5900 .name
= "force_empty",
5901 .write
= mem_cgroup_force_empty_write
,
5904 .name
= "use_hierarchy",
5905 .write_u64
= mem_cgroup_hierarchy_write
,
5906 .read_u64
= mem_cgroup_hierarchy_read
,
5909 .name
= "cgroup.event_control", /* XXX: for compat */
5910 .write
= memcg_write_event_control
,
5911 .flags
= CFTYPE_NO_PREFIX
,
5915 .name
= "swappiness",
5916 .read_u64
= mem_cgroup_swappiness_read
,
5917 .write_u64
= mem_cgroup_swappiness_write
,
5920 .name
= "move_charge_at_immigrate",
5921 .read_u64
= mem_cgroup_move_charge_read
,
5922 .write_u64
= mem_cgroup_move_charge_write
,
5925 .name
= "oom_control",
5926 .seq_show
= mem_cgroup_oom_control_read
,
5927 .write_u64
= mem_cgroup_oom_control_write
,
5928 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5931 .name
= "pressure_level",
5935 .name
= "numa_stat",
5936 .seq_show
= memcg_numa_stat_show
,
5939 #ifdef CONFIG_MEMCG_KMEM
5941 .name
= "kmem.limit_in_bytes",
5942 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5943 .write
= mem_cgroup_write
,
5944 .read_u64
= mem_cgroup_read_u64
,
5947 .name
= "kmem.usage_in_bytes",
5948 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5949 .read_u64
= mem_cgroup_read_u64
,
5952 .name
= "kmem.failcnt",
5953 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5954 .write
= mem_cgroup_reset
,
5955 .read_u64
= mem_cgroup_read_u64
,
5958 .name
= "kmem.max_usage_in_bytes",
5959 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5960 .write
= mem_cgroup_reset
,
5961 .read_u64
= mem_cgroup_read_u64
,
5963 #ifdef CONFIG_SLABINFO
5965 .name
= "kmem.slabinfo",
5966 .seq_show
= mem_cgroup_slabinfo_read
,
5970 { }, /* terminate */
5973 #ifdef CONFIG_MEMCG_SWAP
5974 static struct cftype memsw_cgroup_files
[] = {
5976 .name
= "memsw.usage_in_bytes",
5977 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5978 .read_u64
= mem_cgroup_read_u64
,
5981 .name
= "memsw.max_usage_in_bytes",
5982 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5983 .write
= mem_cgroup_reset
,
5984 .read_u64
= mem_cgroup_read_u64
,
5987 .name
= "memsw.limit_in_bytes",
5988 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5989 .write
= mem_cgroup_write
,
5990 .read_u64
= mem_cgroup_read_u64
,
5993 .name
= "memsw.failcnt",
5994 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5995 .write
= mem_cgroup_reset
,
5996 .read_u64
= mem_cgroup_read_u64
,
5998 { }, /* terminate */
6001 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6003 struct mem_cgroup_per_node
*pn
;
6004 struct mem_cgroup_per_zone
*mz
;
6005 int zone
, tmp
= node
;
6007 * This routine is called against possible nodes.
6008 * But it's BUG to call kmalloc() against offline node.
6010 * TODO: this routine can waste much memory for nodes which will
6011 * never be onlined. It's better to use memory hotplug callback
6014 if (!node_state(node
, N_NORMAL_MEMORY
))
6016 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
6020 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6021 mz
= &pn
->zoneinfo
[zone
];
6022 lruvec_init(&mz
->lruvec
);
6023 mz
->usage_in_excess
= 0;
6024 mz
->on_tree
= false;
6027 memcg
->nodeinfo
[node
] = pn
;
6031 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
6033 kfree(memcg
->nodeinfo
[node
]);
6036 static struct mem_cgroup
*mem_cgroup_alloc(void)
6038 struct mem_cgroup
*memcg
;
6041 size
= sizeof(struct mem_cgroup
);
6042 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
6044 memcg
= kzalloc(size
, GFP_KERNEL
);
6048 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
6051 spin_lock_init(&memcg
->pcp_counter_lock
);
6060 * At destroying mem_cgroup, references from swap_cgroup can remain.
6061 * (scanning all at force_empty is too costly...)
6063 * Instead of clearing all references at force_empty, we remember
6064 * the number of reference from swap_cgroup and free mem_cgroup when
6065 * it goes down to 0.
6067 * Removal of cgroup itself succeeds regardless of refs from swap.
6070 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
6074 mem_cgroup_remove_from_trees(memcg
);
6077 free_mem_cgroup_per_zone_info(memcg
, node
);
6079 free_percpu(memcg
->stat
);
6082 * We need to make sure that (at least for now), the jump label
6083 * destruction code runs outside of the cgroup lock. This is because
6084 * get_online_cpus(), which is called from the static_branch update,
6085 * can't be called inside the cgroup_lock. cpusets are the ones
6086 * enforcing this dependency, so if they ever change, we might as well.
6088 * schedule_work() will guarantee this happens. Be careful if you need
6089 * to move this code around, and make sure it is outside
6092 disarm_static_keys(memcg
);
6097 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6099 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6101 if (!memcg
->res
.parent
)
6103 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6105 EXPORT_SYMBOL(parent_mem_cgroup
);
6107 static void __init
mem_cgroup_soft_limit_tree_init(void)
6109 struct mem_cgroup_tree_per_node
*rtpn
;
6110 struct mem_cgroup_tree_per_zone
*rtpz
;
6111 int tmp
, node
, zone
;
6113 for_each_node(node
) {
6115 if (!node_state(node
, N_NORMAL_MEMORY
))
6117 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
6120 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
6122 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
6123 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
6124 rtpz
->rb_root
= RB_ROOT
;
6125 spin_lock_init(&rtpz
->lock
);
6130 static struct cgroup_subsys_state
* __ref
6131 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6133 struct mem_cgroup
*memcg
;
6134 long error
= -ENOMEM
;
6137 memcg
= mem_cgroup_alloc();
6139 return ERR_PTR(error
);
6142 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6146 if (parent_css
== NULL
) {
6147 root_mem_cgroup
= memcg
;
6148 res_counter_init(&memcg
->res
, NULL
);
6149 res_counter_init(&memcg
->memsw
, NULL
);
6150 res_counter_init(&memcg
->kmem
, NULL
);
6153 memcg
->last_scanned_node
= MAX_NUMNODES
;
6154 INIT_LIST_HEAD(&memcg
->oom_notify
);
6155 memcg
->move_charge_at_immigrate
= 0;
6156 mutex_init(&memcg
->thresholds_lock
);
6157 spin_lock_init(&memcg
->move_lock
);
6158 vmpressure_init(&memcg
->vmpressure
);
6159 INIT_LIST_HEAD(&memcg
->event_list
);
6160 spin_lock_init(&memcg
->event_list_lock
);
6165 __mem_cgroup_free(memcg
);
6166 return ERR_PTR(error
);
6170 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6172 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6173 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
6175 if (css
->id
> MEM_CGROUP_ID_MAX
)
6181 mutex_lock(&memcg_create_mutex
);
6183 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6184 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6185 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6187 if (parent
->use_hierarchy
) {
6188 res_counter_init(&memcg
->res
, &parent
->res
);
6189 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6190 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6193 * No need to take a reference to the parent because cgroup
6194 * core guarantees its existence.
6197 res_counter_init(&memcg
->res
, &root_mem_cgroup
->res
);
6198 res_counter_init(&memcg
->memsw
, &root_mem_cgroup
->memsw
);
6199 res_counter_init(&memcg
->kmem
, &root_mem_cgroup
->kmem
);
6201 * Deeper hierachy with use_hierarchy == false doesn't make
6202 * much sense so let cgroup subsystem know about this
6203 * unfortunate state in our controller.
6205 if (parent
!= root_mem_cgroup
)
6206 memory_cgrp_subsys
.broken_hierarchy
= true;
6208 mutex_unlock(&memcg_create_mutex
);
6210 return memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
6214 * Announce all parents that a group from their hierarchy is gone.
6216 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6218 struct mem_cgroup
*parent
= memcg
;
6220 while ((parent
= parent_mem_cgroup(parent
)))
6221 mem_cgroup_iter_invalidate(parent
);
6224 * if the root memcg is not hierarchical we have to check it
6227 if (!root_mem_cgroup
->use_hierarchy
)
6228 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6231 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6233 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6234 struct mem_cgroup_event
*event
, *tmp
;
6235 struct cgroup_subsys_state
*iter
;
6238 * Unregister events and notify userspace.
6239 * Notify userspace about cgroup removing only after rmdir of cgroup
6240 * directory to avoid race between userspace and kernelspace.
6242 spin_lock(&memcg
->event_list_lock
);
6243 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
6244 list_del_init(&event
->list
);
6245 schedule_work(&event
->remove
);
6247 spin_unlock(&memcg
->event_list_lock
);
6249 kmem_cgroup_css_offline(memcg
);
6251 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6254 * This requires that offlining is serialized. Right now that is
6255 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6257 css_for_each_descendant_post(iter
, css
)
6258 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter
));
6260 memcg_unregister_all_caches(memcg
);
6261 vmpressure_cleanup(&memcg
->vmpressure
);
6264 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6266 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6268 * XXX: css_offline() would be where we should reparent all
6269 * memory to prepare the cgroup for destruction. However,
6270 * memcg does not do css_tryget_online() and res_counter charging
6271 * under the same RCU lock region, which means that charging
6272 * could race with offlining. Offlining only happens to
6273 * cgroups with no tasks in them but charges can show up
6274 * without any tasks from the swapin path when the target
6275 * memcg is looked up from the swapout record and not from the
6276 * current task as it usually is. A race like this can leak
6277 * charges and put pages with stale cgroup pointers into
6281 * lookup_swap_cgroup_id()
6283 * mem_cgroup_lookup()
6284 * css_tryget_online()
6286 * disable css_tryget_online()
6289 * reparent_charges()
6290 * res_counter_charge()
6293 * pc->mem_cgroup = dead memcg
6296 * The bulk of the charges are still moved in offline_css() to
6297 * avoid pinning a lot of pages in case a long-term reference
6298 * like a swapout record is deferring the css_free() to long
6299 * after offlining. But this makes sure we catch any charges
6300 * made after offlining:
6302 mem_cgroup_reparent_charges(memcg
);
6304 memcg_destroy_kmem(memcg
);
6305 __mem_cgroup_free(memcg
);
6309 * mem_cgroup_css_reset - reset the states of a mem_cgroup
6310 * @css: the target css
6312 * Reset the states of the mem_cgroup associated with @css. This is
6313 * invoked when the userland requests disabling on the default hierarchy
6314 * but the memcg is pinned through dependency. The memcg should stop
6315 * applying policies and should revert to the vanilla state as it may be
6316 * made visible again.
6318 * The current implementation only resets the essential configurations.
6319 * This needs to be expanded to cover all the visible parts.
6321 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
6323 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6325 mem_cgroup_resize_limit(memcg
, ULLONG_MAX
);
6326 mem_cgroup_resize_memsw_limit(memcg
, ULLONG_MAX
);
6327 memcg_update_kmem_limit(memcg
, ULLONG_MAX
);
6328 res_counter_set_soft_limit(&memcg
->res
, ULLONG_MAX
);
6332 /* Handlers for move charge at task migration. */
6333 static int mem_cgroup_do_precharge(unsigned long count
)
6337 /* Try a single bulk charge without reclaim first */
6338 ret
= mem_cgroup_try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
6340 mc
.precharge
+= count
;
6343 if (ret
== -EINTR
) {
6344 __mem_cgroup_cancel_charge(root_mem_cgroup
, count
);
6348 /* Try charges one by one with reclaim */
6350 ret
= mem_cgroup_try_charge(mc
.to
,
6351 GFP_KERNEL
& ~__GFP_NORETRY
, 1);
6353 * In case of failure, any residual charges against
6354 * mc.to will be dropped by mem_cgroup_clear_mc()
6355 * later on. However, cancel any charges that are
6356 * bypassed to root right away or they'll be lost.
6359 __mem_cgroup_cancel_charge(root_mem_cgroup
, 1);
6369 * get_mctgt_type - get target type of moving charge
6370 * @vma: the vma the pte to be checked belongs
6371 * @addr: the address corresponding to the pte to be checked
6372 * @ptent: the pte to be checked
6373 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6376 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6377 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6378 * move charge. if @target is not NULL, the page is stored in target->page
6379 * with extra refcnt got(Callers should handle it).
6380 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6381 * target for charge migration. if @target is not NULL, the entry is stored
6384 * Called with pte lock held.
6391 enum mc_target_type
{
6397 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6398 unsigned long addr
, pte_t ptent
)
6400 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6402 if (!page
|| !page_mapped(page
))
6404 if (PageAnon(page
)) {
6405 /* we don't move shared anon */
6408 } else if (!move_file())
6409 /* we ignore mapcount for file pages */
6411 if (!get_page_unless_zero(page
))
6418 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6419 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6421 struct page
*page
= NULL
;
6422 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6424 if (!move_anon() || non_swap_entry(ent
))
6427 * Because lookup_swap_cache() updates some statistics counter,
6428 * we call find_get_page() with swapper_space directly.
6430 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6431 if (do_swap_account
)
6432 entry
->val
= ent
.val
;
6437 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6438 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6444 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6445 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6447 struct page
*page
= NULL
;
6448 struct address_space
*mapping
;
6451 if (!vma
->vm_file
) /* anonymous vma */
6456 mapping
= vma
->vm_file
->f_mapping
;
6457 if (pte_none(ptent
))
6458 pgoff
= linear_page_index(vma
, addr
);
6459 else /* pte_file(ptent) is true */
6460 pgoff
= pte_to_pgoff(ptent
);
6462 /* page is moved even if it's not RSS of this task(page-faulted). */
6464 /* shmem/tmpfs may report page out on swap: account for that too. */
6465 if (shmem_mapping(mapping
)) {
6466 page
= find_get_entry(mapping
, pgoff
);
6467 if (radix_tree_exceptional_entry(page
)) {
6468 swp_entry_t swp
= radix_to_swp_entry(page
);
6469 if (do_swap_account
)
6471 page
= find_get_page(swap_address_space(swp
), swp
.val
);
6474 page
= find_get_page(mapping
, pgoff
);
6476 page
= find_get_page(mapping
, pgoff
);
6481 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6482 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6484 struct page
*page
= NULL
;
6485 struct page_cgroup
*pc
;
6486 enum mc_target_type ret
= MC_TARGET_NONE
;
6487 swp_entry_t ent
= { .val
= 0 };
6489 if (pte_present(ptent
))
6490 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6491 else if (is_swap_pte(ptent
))
6492 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6493 else if (pte_none(ptent
) || pte_file(ptent
))
6494 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6496 if (!page
&& !ent
.val
)
6499 pc
= lookup_page_cgroup(page
);
6501 * Do only loose check w/o page_cgroup lock.
6502 * mem_cgroup_move_account() checks the pc is valid or not under
6505 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6506 ret
= MC_TARGET_PAGE
;
6508 target
->page
= page
;
6510 if (!ret
|| !target
)
6513 /* There is a swap entry and a page doesn't exist or isn't charged */
6514 if (ent
.val
&& !ret
&&
6515 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
6516 ret
= MC_TARGET_SWAP
;
6523 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6525 * We don't consider swapping or file mapped pages because THP does not
6526 * support them for now.
6527 * Caller should make sure that pmd_trans_huge(pmd) is true.
6529 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6530 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6532 struct page
*page
= NULL
;
6533 struct page_cgroup
*pc
;
6534 enum mc_target_type ret
= MC_TARGET_NONE
;
6536 page
= pmd_page(pmd
);
6537 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
6540 pc
= lookup_page_cgroup(page
);
6541 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6542 ret
= MC_TARGET_PAGE
;
6545 target
->page
= page
;
6551 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6552 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6554 return MC_TARGET_NONE
;
6558 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6559 unsigned long addr
, unsigned long end
,
6560 struct mm_walk
*walk
)
6562 struct vm_area_struct
*vma
= walk
->private;
6566 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6567 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6568 mc
.precharge
+= HPAGE_PMD_NR
;
6573 if (pmd_trans_unstable(pmd
))
6575 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6576 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6577 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6578 mc
.precharge
++; /* increment precharge temporarily */
6579 pte_unmap_unlock(pte
- 1, ptl
);
6585 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6587 unsigned long precharge
;
6588 struct vm_area_struct
*vma
;
6590 down_read(&mm
->mmap_sem
);
6591 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6592 struct mm_walk mem_cgroup_count_precharge_walk
= {
6593 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6597 if (is_vm_hugetlb_page(vma
))
6599 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6600 &mem_cgroup_count_precharge_walk
);
6602 up_read(&mm
->mmap_sem
);
6604 precharge
= mc
.precharge
;
6610 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6612 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6614 VM_BUG_ON(mc
.moving_task
);
6615 mc
.moving_task
= current
;
6616 return mem_cgroup_do_precharge(precharge
);
6619 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6620 static void __mem_cgroup_clear_mc(void)
6622 struct mem_cgroup
*from
= mc
.from
;
6623 struct mem_cgroup
*to
= mc
.to
;
6626 /* we must uncharge all the leftover precharges from mc.to */
6628 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6632 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6633 * we must uncharge here.
6635 if (mc
.moved_charge
) {
6636 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6637 mc
.moved_charge
= 0;
6639 /* we must fixup refcnts and charges */
6640 if (mc
.moved_swap
) {
6641 /* uncharge swap account from the old cgroup */
6642 res_counter_uncharge(&mc
.from
->memsw
,
6643 PAGE_SIZE
* mc
.moved_swap
);
6645 for (i
= 0; i
< mc
.moved_swap
; i
++)
6646 css_put(&mc
.from
->css
);
6649 * we charged both to->res and to->memsw, so we should
6652 res_counter_uncharge(&mc
.to
->res
,
6653 PAGE_SIZE
* mc
.moved_swap
);
6654 /* we've already done css_get(mc.to) */
6657 memcg_oom_recover(from
);
6658 memcg_oom_recover(to
);
6659 wake_up_all(&mc
.waitq
);
6662 static void mem_cgroup_clear_mc(void)
6664 struct mem_cgroup
*from
= mc
.from
;
6667 * we must clear moving_task before waking up waiters at the end of
6670 mc
.moving_task
= NULL
;
6671 __mem_cgroup_clear_mc();
6672 spin_lock(&mc
.lock
);
6675 spin_unlock(&mc
.lock
);
6676 mem_cgroup_end_move(from
);
6679 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6680 struct cgroup_taskset
*tset
)
6682 struct task_struct
*p
= cgroup_taskset_first(tset
);
6684 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6685 unsigned long move_charge_at_immigrate
;
6688 * We are now commited to this value whatever it is. Changes in this
6689 * tunable will only affect upcoming migrations, not the current one.
6690 * So we need to save it, and keep it going.
6692 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6693 if (move_charge_at_immigrate
) {
6694 struct mm_struct
*mm
;
6695 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6697 VM_BUG_ON(from
== memcg
);
6699 mm
= get_task_mm(p
);
6702 /* We move charges only when we move a owner of the mm */
6703 if (mm
->owner
== p
) {
6706 VM_BUG_ON(mc
.precharge
);
6707 VM_BUG_ON(mc
.moved_charge
);
6708 VM_BUG_ON(mc
.moved_swap
);
6709 mem_cgroup_start_move(from
);
6710 spin_lock(&mc
.lock
);
6713 mc
.immigrate_flags
= move_charge_at_immigrate
;
6714 spin_unlock(&mc
.lock
);
6715 /* We set mc.moving_task later */
6717 ret
= mem_cgroup_precharge_mc(mm
);
6719 mem_cgroup_clear_mc();
6726 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6727 struct cgroup_taskset
*tset
)
6729 mem_cgroup_clear_mc();
6732 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6733 unsigned long addr
, unsigned long end
,
6734 struct mm_walk
*walk
)
6737 struct vm_area_struct
*vma
= walk
->private;
6740 enum mc_target_type target_type
;
6741 union mc_target target
;
6743 struct page_cgroup
*pc
;
6746 * We don't take compound_lock() here but no race with splitting thp
6748 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6749 * under splitting, which means there's no concurrent thp split,
6750 * - if another thread runs into split_huge_page() just after we
6751 * entered this if-block, the thread must wait for page table lock
6752 * to be unlocked in __split_huge_page_splitting(), where the main
6753 * part of thp split is not executed yet.
6755 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6756 if (mc
.precharge
< HPAGE_PMD_NR
) {
6760 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6761 if (target_type
== MC_TARGET_PAGE
) {
6763 if (!isolate_lru_page(page
)) {
6764 pc
= lookup_page_cgroup(page
);
6765 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6766 pc
, mc
.from
, mc
.to
)) {
6767 mc
.precharge
-= HPAGE_PMD_NR
;
6768 mc
.moved_charge
+= HPAGE_PMD_NR
;
6770 putback_lru_page(page
);
6778 if (pmd_trans_unstable(pmd
))
6781 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6782 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6783 pte_t ptent
= *(pte
++);
6789 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6790 case MC_TARGET_PAGE
:
6792 if (isolate_lru_page(page
))
6794 pc
= lookup_page_cgroup(page
);
6795 if (!mem_cgroup_move_account(page
, 1, pc
,
6798 /* we uncharge from mc.from later. */
6801 putback_lru_page(page
);
6802 put
: /* get_mctgt_type() gets the page */
6805 case MC_TARGET_SWAP
:
6807 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6809 /* we fixup refcnts and charges later. */
6817 pte_unmap_unlock(pte
- 1, ptl
);
6822 * We have consumed all precharges we got in can_attach().
6823 * We try charge one by one, but don't do any additional
6824 * charges to mc.to if we have failed in charge once in attach()
6827 ret
= mem_cgroup_do_precharge(1);
6835 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6837 struct vm_area_struct
*vma
;
6839 lru_add_drain_all();
6841 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6843 * Someone who are holding the mmap_sem might be waiting in
6844 * waitq. So we cancel all extra charges, wake up all waiters,
6845 * and retry. Because we cancel precharges, we might not be able
6846 * to move enough charges, but moving charge is a best-effort
6847 * feature anyway, so it wouldn't be a big problem.
6849 __mem_cgroup_clear_mc();
6853 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6855 struct mm_walk mem_cgroup_move_charge_walk
= {
6856 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6860 if (is_vm_hugetlb_page(vma
))
6862 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6863 &mem_cgroup_move_charge_walk
);
6866 * means we have consumed all precharges and failed in
6867 * doing additional charge. Just abandon here.
6871 up_read(&mm
->mmap_sem
);
6874 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6875 struct cgroup_taskset
*tset
)
6877 struct task_struct
*p
= cgroup_taskset_first(tset
);
6878 struct mm_struct
*mm
= get_task_mm(p
);
6882 mem_cgroup_move_charge(mm
);
6886 mem_cgroup_clear_mc();
6888 #else /* !CONFIG_MMU */
6889 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6890 struct cgroup_taskset
*tset
)
6894 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6895 struct cgroup_taskset
*tset
)
6898 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6899 struct cgroup_taskset
*tset
)
6905 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6906 * to verify whether we're attached to the default hierarchy on each mount
6909 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6912 * use_hierarchy is forced on the default hierarchy. cgroup core
6913 * guarantees that @root doesn't have any children, so turning it
6914 * on for the root memcg is enough.
6916 if (cgroup_on_dfl(root_css
->cgroup
))
6917 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6920 struct cgroup_subsys memory_cgrp_subsys
= {
6921 .css_alloc
= mem_cgroup_css_alloc
,
6922 .css_online
= mem_cgroup_css_online
,
6923 .css_offline
= mem_cgroup_css_offline
,
6924 .css_free
= mem_cgroup_css_free
,
6925 .css_reset
= mem_cgroup_css_reset
,
6926 .can_attach
= mem_cgroup_can_attach
,
6927 .cancel_attach
= mem_cgroup_cancel_attach
,
6928 .attach
= mem_cgroup_move_task
,
6929 .bind
= mem_cgroup_bind
,
6930 .legacy_cftypes
= mem_cgroup_files
,
6934 #ifdef CONFIG_MEMCG_SWAP
6935 static int __init
enable_swap_account(char *s
)
6937 if (!strcmp(s
, "1"))
6938 really_do_swap_account
= 1;
6939 else if (!strcmp(s
, "0"))
6940 really_do_swap_account
= 0;
6943 __setup("swapaccount=", enable_swap_account
);
6945 static void __init
memsw_file_init(void)
6947 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6948 memsw_cgroup_files
));
6951 static void __init
enable_swap_cgroup(void)
6953 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6954 do_swap_account
= 1;
6960 static void __init
enable_swap_cgroup(void)
6966 * subsys_initcall() for memory controller.
6968 * Some parts like hotcpu_notifier() have to be initialized from this context
6969 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6970 * everything that doesn't depend on a specific mem_cgroup structure should
6971 * be initialized from here.
6973 static int __init
mem_cgroup_init(void)
6975 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6976 enable_swap_cgroup();
6977 mem_cgroup_soft_limit_tree_init();
6981 subsys_initcall(mem_cgroup_init
);