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
)
759 spin_lock_irqsave(&mctz
->lock
, flags
);
760 __mem_cgroup_remove_exceeded(mz
, mctz
);
761 spin_unlock_irqrestore(&mctz
->lock
, flags
);
765 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
767 unsigned long long excess
;
768 struct mem_cgroup_per_zone
*mz
;
769 struct mem_cgroup_tree_per_zone
*mctz
;
771 mctz
= soft_limit_tree_from_page(page
);
773 * Necessary to update all ancestors when hierarchy is used.
774 * because their event counter is not touched.
776 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
777 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
778 excess
= res_counter_soft_limit_excess(&memcg
->res
);
780 * We have to update the tree if mz is on RB-tree or
781 * mem is over its softlimit.
783 if (excess
|| mz
->on_tree
) {
786 spin_lock_irqsave(&mctz
->lock
, flags
);
787 /* if on-tree, remove it */
789 __mem_cgroup_remove_exceeded(mz
, mctz
);
791 * Insert again. mz->usage_in_excess will be updated.
792 * If excess is 0, no tree ops.
794 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
795 spin_unlock_irqrestore(&mctz
->lock
, flags
);
800 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
802 struct mem_cgroup_tree_per_zone
*mctz
;
803 struct mem_cgroup_per_zone
*mz
;
807 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
808 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
809 mctz
= soft_limit_tree_node_zone(nid
, zid
);
810 mem_cgroup_remove_exceeded(mz
, mctz
);
815 static struct mem_cgroup_per_zone
*
816 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
818 struct rb_node
*rightmost
= NULL
;
819 struct mem_cgroup_per_zone
*mz
;
823 rightmost
= rb_last(&mctz
->rb_root
);
825 goto done
; /* Nothing to reclaim from */
827 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
829 * Remove the node now but someone else can add it back,
830 * we will to add it back at the end of reclaim to its correct
831 * position in the tree.
833 __mem_cgroup_remove_exceeded(mz
, mctz
);
834 if (!res_counter_soft_limit_excess(&mz
->memcg
->res
) ||
835 !css_tryget_online(&mz
->memcg
->css
))
841 static struct mem_cgroup_per_zone
*
842 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
844 struct mem_cgroup_per_zone
*mz
;
846 spin_lock_irq(&mctz
->lock
);
847 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
848 spin_unlock_irq(&mctz
->lock
);
853 * Implementation Note: reading percpu statistics for memcg.
855 * Both of vmstat[] and percpu_counter has threshold and do periodic
856 * synchronization to implement "quick" read. There are trade-off between
857 * reading cost and precision of value. Then, we may have a chance to implement
858 * a periodic synchronizion of counter in memcg's counter.
860 * But this _read() function is used for user interface now. The user accounts
861 * memory usage by memory cgroup and he _always_ requires exact value because
862 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
863 * have to visit all online cpus and make sum. So, for now, unnecessary
864 * synchronization is not implemented. (just implemented for cpu hotplug)
866 * If there are kernel internal actions which can make use of some not-exact
867 * value, and reading all cpu value can be performance bottleneck in some
868 * common workload, threashold and synchonization as vmstat[] should be
871 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
872 enum mem_cgroup_stat_index idx
)
878 for_each_online_cpu(cpu
)
879 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
880 #ifdef CONFIG_HOTPLUG_CPU
881 spin_lock(&memcg
->pcp_counter_lock
);
882 val
+= memcg
->nocpu_base
.count
[idx
];
883 spin_unlock(&memcg
->pcp_counter_lock
);
889 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
890 enum mem_cgroup_events_index idx
)
892 unsigned long val
= 0;
896 for_each_online_cpu(cpu
)
897 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
898 #ifdef CONFIG_HOTPLUG_CPU
899 spin_lock(&memcg
->pcp_counter_lock
);
900 val
+= memcg
->nocpu_base
.events
[idx
];
901 spin_unlock(&memcg
->pcp_counter_lock
);
907 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
912 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
913 * counted as CACHE even if it's on ANON LRU.
916 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
919 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
922 if (PageTransHuge(page
))
923 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
926 /* pagein of a big page is an event. So, ignore page size */
928 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
930 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
931 nr_pages
= -nr_pages
; /* for event */
934 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
937 unsigned long mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
939 struct mem_cgroup_per_zone
*mz
;
941 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
942 return mz
->lru_size
[lru
];
945 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
947 unsigned int lru_mask
)
949 unsigned long nr
= 0;
952 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
954 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
955 struct mem_cgroup_per_zone
*mz
;
959 if (!(BIT(lru
) & lru_mask
))
961 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
962 nr
+= mz
->lru_size
[lru
];
968 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
969 unsigned int lru_mask
)
971 unsigned long nr
= 0;
974 for_each_node_state(nid
, N_MEMORY
)
975 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
979 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
980 enum mem_cgroup_events_target target
)
982 unsigned long val
, next
;
984 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
985 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
986 /* from time_after() in jiffies.h */
987 if ((long)next
- (long)val
< 0) {
989 case MEM_CGROUP_TARGET_THRESH
:
990 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
992 case MEM_CGROUP_TARGET_SOFTLIMIT
:
993 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
995 case MEM_CGROUP_TARGET_NUMAINFO
:
996 next
= val
+ NUMAINFO_EVENTS_TARGET
;
1001 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
1008 * Check events in order.
1011 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
1013 /* threshold event is triggered in finer grain than soft limit */
1014 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
1015 MEM_CGROUP_TARGET_THRESH
))) {
1017 bool do_numainfo __maybe_unused
;
1019 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
1020 MEM_CGROUP_TARGET_SOFTLIMIT
);
1021 #if MAX_NUMNODES > 1
1022 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
1023 MEM_CGROUP_TARGET_NUMAINFO
);
1025 mem_cgroup_threshold(memcg
);
1026 if (unlikely(do_softlimit
))
1027 mem_cgroup_update_tree(memcg
, page
);
1028 #if MAX_NUMNODES > 1
1029 if (unlikely(do_numainfo
))
1030 atomic_inc(&memcg
->numainfo_events
);
1035 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
1038 * mm_update_next_owner() may clear mm->owner to NULL
1039 * if it races with swapoff, page migration, etc.
1040 * So this can be called with p == NULL.
1045 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
1048 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
1050 struct mem_cgroup
*memcg
= NULL
;
1055 * Page cache insertions can happen withou an
1056 * actual mm context, e.g. during disk probing
1057 * on boot, loopback IO, acct() writes etc.
1060 memcg
= root_mem_cgroup
;
1062 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1063 if (unlikely(!memcg
))
1064 memcg
= root_mem_cgroup
;
1066 } while (!css_tryget_online(&memcg
->css
));
1072 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1073 * ref. count) or NULL if the whole root's subtree has been visited.
1075 * helper function to be used by mem_cgroup_iter
1077 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
1078 struct mem_cgroup
*last_visited
)
1080 struct cgroup_subsys_state
*prev_css
, *next_css
;
1082 prev_css
= last_visited
? &last_visited
->css
: NULL
;
1084 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
1087 * Even if we found a group we have to make sure it is
1088 * alive. css && !memcg means that the groups should be
1089 * skipped and we should continue the tree walk.
1090 * last_visited css is safe to use because it is
1091 * protected by css_get and the tree walk is rcu safe.
1093 * We do not take a reference on the root of the tree walk
1094 * because we might race with the root removal when it would
1095 * be the only node in the iterated hierarchy and mem_cgroup_iter
1096 * would end up in an endless loop because it expects that at
1097 * least one valid node will be returned. Root cannot disappear
1098 * because caller of the iterator should hold it already so
1099 * skipping css reference should be safe.
1102 if ((next_css
== &root
->css
) ||
1103 ((next_css
->flags
& CSS_ONLINE
) &&
1104 css_tryget_online(next_css
)))
1105 return mem_cgroup_from_css(next_css
);
1107 prev_css
= next_css
;
1114 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1117 * When a group in the hierarchy below root is destroyed, the
1118 * hierarchy iterator can no longer be trusted since it might
1119 * have pointed to the destroyed group. Invalidate it.
1121 atomic_inc(&root
->dead_count
);
1124 static struct mem_cgroup
*
1125 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1126 struct mem_cgroup
*root
,
1129 struct mem_cgroup
*position
= NULL
;
1131 * A cgroup destruction happens in two stages: offlining and
1132 * release. They are separated by a RCU grace period.
1134 * If the iterator is valid, we may still race with an
1135 * offlining. The RCU lock ensures the object won't be
1136 * released, tryget will fail if we lost the race.
1138 *sequence
= atomic_read(&root
->dead_count
);
1139 if (iter
->last_dead_count
== *sequence
) {
1141 position
= iter
->last_visited
;
1144 * We cannot take a reference to root because we might race
1145 * with root removal and returning NULL would end up in
1146 * an endless loop on the iterator user level when root
1147 * would be returned all the time.
1149 if (position
&& position
!= root
&&
1150 !css_tryget_online(&position
->css
))
1156 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1157 struct mem_cgroup
*last_visited
,
1158 struct mem_cgroup
*new_position
,
1159 struct mem_cgroup
*root
,
1162 /* root reference counting symmetric to mem_cgroup_iter_load */
1163 if (last_visited
&& last_visited
!= root
)
1164 css_put(&last_visited
->css
);
1166 * We store the sequence count from the time @last_visited was
1167 * loaded successfully instead of rereading it here so that we
1168 * don't lose destruction events in between. We could have
1169 * raced with the destruction of @new_position after all.
1171 iter
->last_visited
= new_position
;
1173 iter
->last_dead_count
= sequence
;
1177 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1178 * @root: hierarchy root
1179 * @prev: previously returned memcg, NULL on first invocation
1180 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1182 * Returns references to children of the hierarchy below @root, or
1183 * @root itself, or %NULL after a full round-trip.
1185 * Caller must pass the return value in @prev on subsequent
1186 * invocations for reference counting, or use mem_cgroup_iter_break()
1187 * to cancel a hierarchy walk before the round-trip is complete.
1189 * Reclaimers can specify a zone and a priority level in @reclaim to
1190 * divide up the memcgs in the hierarchy among all concurrent
1191 * reclaimers operating on the same zone and priority.
1193 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
1194 struct mem_cgroup
*prev
,
1195 struct mem_cgroup_reclaim_cookie
*reclaim
)
1197 struct mem_cgroup
*memcg
= NULL
;
1198 struct mem_cgroup
*last_visited
= NULL
;
1200 if (mem_cgroup_disabled())
1204 root
= root_mem_cgroup
;
1206 if (prev
&& !reclaim
)
1207 last_visited
= prev
;
1209 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1217 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1218 int uninitialized_var(seq
);
1221 struct mem_cgroup_per_zone
*mz
;
1223 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
1224 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1225 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1226 iter
->last_visited
= NULL
;
1230 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1233 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1236 mem_cgroup_iter_update(iter
, last_visited
, memcg
, root
,
1241 else if (!prev
&& memcg
)
1242 reclaim
->generation
= iter
->generation
;
1251 if (prev
&& prev
!= root
)
1252 css_put(&prev
->css
);
1258 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1259 * @root: hierarchy root
1260 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1262 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1263 struct mem_cgroup
*prev
)
1266 root
= root_mem_cgroup
;
1267 if (prev
&& prev
!= root
)
1268 css_put(&prev
->css
);
1272 * Iteration constructs for visiting all cgroups (under a tree). If
1273 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1274 * be used for reference counting.
1276 #define for_each_mem_cgroup_tree(iter, root) \
1277 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1279 iter = mem_cgroup_iter(root, iter, NULL))
1281 #define for_each_mem_cgroup(iter) \
1282 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1284 iter = mem_cgroup_iter(NULL, iter, NULL))
1286 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1288 struct mem_cgroup
*memcg
;
1291 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1292 if (unlikely(!memcg
))
1297 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1300 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1308 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1311 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1312 * @zone: zone of the wanted lruvec
1313 * @memcg: memcg of the wanted lruvec
1315 * Returns the lru list vector holding pages for the given @zone and
1316 * @mem. This can be the global zone lruvec, if the memory controller
1319 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1320 struct mem_cgroup
*memcg
)
1322 struct mem_cgroup_per_zone
*mz
;
1323 struct lruvec
*lruvec
;
1325 if (mem_cgroup_disabled()) {
1326 lruvec
= &zone
->lruvec
;
1330 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1331 lruvec
= &mz
->lruvec
;
1334 * Since a node can be onlined after the mem_cgroup was created,
1335 * we have to be prepared to initialize lruvec->zone here;
1336 * and if offlined then reonlined, we need to reinitialize it.
1338 if (unlikely(lruvec
->zone
!= zone
))
1339 lruvec
->zone
= zone
;
1344 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1346 * @zone: zone of the page
1348 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1350 struct mem_cgroup_per_zone
*mz
;
1351 struct mem_cgroup
*memcg
;
1352 struct page_cgroup
*pc
;
1353 struct lruvec
*lruvec
;
1355 if (mem_cgroup_disabled()) {
1356 lruvec
= &zone
->lruvec
;
1360 pc
= lookup_page_cgroup(page
);
1361 memcg
= pc
->mem_cgroup
;
1364 * Surreptitiously switch any uncharged offlist page to root:
1365 * an uncharged page off lru does nothing to secure
1366 * its former mem_cgroup from sudden removal.
1368 * Our caller holds lru_lock, and PageCgroupUsed is updated
1369 * under page_cgroup lock: between them, they make all uses
1370 * of pc->mem_cgroup safe.
1372 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1373 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1375 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1376 lruvec
= &mz
->lruvec
;
1379 * Since a node can be onlined after the mem_cgroup was created,
1380 * we have to be prepared to initialize lruvec->zone here;
1381 * and if offlined then reonlined, we need to reinitialize it.
1383 if (unlikely(lruvec
->zone
!= zone
))
1384 lruvec
->zone
= zone
;
1389 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1390 * @lruvec: mem_cgroup per zone lru vector
1391 * @lru: index of lru list the page is sitting on
1392 * @nr_pages: positive when adding or negative when removing
1394 * This function must be called when a page is added to or removed from an
1397 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1400 struct mem_cgroup_per_zone
*mz
;
1401 unsigned long *lru_size
;
1403 if (mem_cgroup_disabled())
1406 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1407 lru_size
= mz
->lru_size
+ lru
;
1408 *lru_size
+= nr_pages
;
1409 VM_BUG_ON((long)(*lru_size
) < 0);
1413 * Checks whether given mem is same or in the root_mem_cgroup's
1416 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1417 struct mem_cgroup
*memcg
)
1419 if (root_memcg
== memcg
)
1421 if (!root_memcg
->use_hierarchy
|| !memcg
)
1423 return cgroup_is_descendant(memcg
->css
.cgroup
, root_memcg
->css
.cgroup
);
1426 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1427 struct mem_cgroup
*memcg
)
1432 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1437 bool task_in_mem_cgroup(struct task_struct
*task
,
1438 const struct mem_cgroup
*memcg
)
1440 struct mem_cgroup
*curr
= NULL
;
1441 struct task_struct
*p
;
1444 p
= find_lock_task_mm(task
);
1446 curr
= get_mem_cgroup_from_mm(p
->mm
);
1450 * All threads may have already detached their mm's, but the oom
1451 * killer still needs to detect if they have already been oom
1452 * killed to prevent needlessly killing additional tasks.
1455 curr
= mem_cgroup_from_task(task
);
1457 css_get(&curr
->css
);
1461 * We should check use_hierarchy of "memcg" not "curr". Because checking
1462 * use_hierarchy of "curr" here make this function true if hierarchy is
1463 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1464 * hierarchy(even if use_hierarchy is disabled in "memcg").
1466 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1467 css_put(&curr
->css
);
1471 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1473 unsigned long inactive_ratio
;
1474 unsigned long inactive
;
1475 unsigned long active
;
1478 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1479 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1481 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1483 inactive_ratio
= int_sqrt(10 * gb
);
1487 return inactive
* inactive_ratio
< active
;
1490 #define mem_cgroup_from_res_counter(counter, member) \
1491 container_of(counter, struct mem_cgroup, member)
1494 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1495 * @memcg: the memory cgroup
1497 * Returns the maximum amount of memory @mem can be charged with, in
1500 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1502 unsigned long long margin
;
1504 margin
= res_counter_margin(&memcg
->res
);
1505 if (do_swap_account
)
1506 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1507 return margin
>> PAGE_SHIFT
;
1510 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1513 if (mem_cgroup_disabled() || !memcg
->css
.parent
)
1514 return vm_swappiness
;
1516 return memcg
->swappiness
;
1520 * memcg->moving_account is used for checking possibility that some thread is
1521 * calling move_account(). When a thread on CPU-A starts moving pages under
1522 * a memcg, other threads should check memcg->moving_account under
1523 * rcu_read_lock(), like this:
1527 * memcg->moving_account+1 if (memcg->mocing_account)
1529 * synchronize_rcu() update something.
1534 /* for quick checking without looking up memcg */
1535 atomic_t memcg_moving __read_mostly
;
1537 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1539 atomic_inc(&memcg_moving
);
1540 atomic_inc(&memcg
->moving_account
);
1544 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1547 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1548 * We check NULL in callee rather than caller.
1551 atomic_dec(&memcg_moving
);
1552 atomic_dec(&memcg
->moving_account
);
1557 * A routine for checking "mem" is under move_account() or not.
1559 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1560 * moving cgroups. This is for waiting at high-memory pressure
1563 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1565 struct mem_cgroup
*from
;
1566 struct mem_cgroup
*to
;
1569 * Unlike task_move routines, we access mc.to, mc.from not under
1570 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1572 spin_lock(&mc
.lock
);
1578 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1579 || mem_cgroup_same_or_subtree(memcg
, to
);
1581 spin_unlock(&mc
.lock
);
1585 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1587 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1588 if (mem_cgroup_under_move(memcg
)) {
1590 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1591 /* moving charge context might have finished. */
1594 finish_wait(&mc
.waitq
, &wait
);
1602 * Take this lock when
1603 * - a code tries to modify page's memcg while it's USED.
1604 * - a code tries to modify page state accounting in a memcg.
1606 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1607 unsigned long *flags
)
1609 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1612 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1613 unsigned long *flags
)
1615 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1618 #define K(x) ((x) << (PAGE_SHIFT-10))
1620 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1621 * @memcg: The memory cgroup that went over limit
1622 * @p: Task that is going to be killed
1624 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1627 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1629 /* oom_info_lock ensures that parallel ooms do not interleave */
1630 static DEFINE_MUTEX(oom_info_lock
);
1631 struct mem_cgroup
*iter
;
1637 mutex_lock(&oom_info_lock
);
1640 pr_info("Task in ");
1641 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1642 pr_info(" killed as a result of limit of ");
1643 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1648 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1649 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1650 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1651 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1652 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1653 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1654 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1655 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1656 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1657 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1658 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1659 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1661 for_each_mem_cgroup_tree(iter
, memcg
) {
1662 pr_info("Memory cgroup stats for ");
1663 pr_cont_cgroup_path(iter
->css
.cgroup
);
1666 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1667 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1669 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1670 K(mem_cgroup_read_stat(iter
, i
)));
1673 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1674 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1675 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1679 mutex_unlock(&oom_info_lock
);
1683 * This function returns the number of memcg under hierarchy tree. Returns
1684 * 1(self count) if no children.
1686 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1689 struct mem_cgroup
*iter
;
1691 for_each_mem_cgroup_tree(iter
, memcg
)
1697 * Return the memory (and swap, if configured) limit for a memcg.
1699 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1703 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1706 * Do not consider swap space if we cannot swap due to swappiness
1708 if (mem_cgroup_swappiness(memcg
)) {
1711 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1712 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1715 * If memsw is finite and limits the amount of swap space
1716 * available to this memcg, return that limit.
1718 limit
= min(limit
, memsw
);
1724 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1727 struct mem_cgroup
*iter
;
1728 unsigned long chosen_points
= 0;
1729 unsigned long totalpages
;
1730 unsigned int points
= 0;
1731 struct task_struct
*chosen
= NULL
;
1734 * If current has a pending SIGKILL or is exiting, then automatically
1735 * select it. The goal is to allow it to allocate so that it may
1736 * quickly exit and free its memory.
1738 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1739 set_thread_flag(TIF_MEMDIE
);
1743 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1744 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1745 for_each_mem_cgroup_tree(iter
, memcg
) {
1746 struct css_task_iter it
;
1747 struct task_struct
*task
;
1749 css_task_iter_start(&iter
->css
, &it
);
1750 while ((task
= css_task_iter_next(&it
))) {
1751 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1753 case OOM_SCAN_SELECT
:
1755 put_task_struct(chosen
);
1757 chosen_points
= ULONG_MAX
;
1758 get_task_struct(chosen
);
1760 case OOM_SCAN_CONTINUE
:
1762 case OOM_SCAN_ABORT
:
1763 css_task_iter_end(&it
);
1764 mem_cgroup_iter_break(memcg
, iter
);
1766 put_task_struct(chosen
);
1771 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1772 if (!points
|| points
< chosen_points
)
1774 /* Prefer thread group leaders for display purposes */
1775 if (points
== chosen_points
&&
1776 thread_group_leader(chosen
))
1780 put_task_struct(chosen
);
1782 chosen_points
= points
;
1783 get_task_struct(chosen
);
1785 css_task_iter_end(&it
);
1790 points
= chosen_points
* 1000 / totalpages
;
1791 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1792 NULL
, "Memory cgroup out of memory");
1795 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1797 unsigned long flags
)
1799 unsigned long total
= 0;
1800 bool noswap
= false;
1803 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1805 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1808 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1810 drain_all_stock_async(memcg
);
1811 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1813 * Allow limit shrinkers, which are triggered directly
1814 * by userspace, to catch signals and stop reclaim
1815 * after minimal progress, regardless of the margin.
1817 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1819 if (mem_cgroup_margin(memcg
))
1822 * If nothing was reclaimed after two attempts, there
1823 * may be no reclaimable pages in this hierarchy.
1832 * test_mem_cgroup_node_reclaimable
1833 * @memcg: the target memcg
1834 * @nid: the node ID to be checked.
1835 * @noswap : specify true here if the user wants flle only information.
1837 * This function returns whether the specified memcg contains any
1838 * reclaimable pages on a node. Returns true if there are any reclaimable
1839 * pages in the node.
1841 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1842 int nid
, bool noswap
)
1844 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1846 if (noswap
|| !total_swap_pages
)
1848 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1853 #if MAX_NUMNODES > 1
1856 * Always updating the nodemask is not very good - even if we have an empty
1857 * list or the wrong list here, we can start from some node and traverse all
1858 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1861 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1865 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1866 * pagein/pageout changes since the last update.
1868 if (!atomic_read(&memcg
->numainfo_events
))
1870 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1873 /* make a nodemask where this memcg uses memory from */
1874 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1876 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1878 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1879 node_clear(nid
, memcg
->scan_nodes
);
1882 atomic_set(&memcg
->numainfo_events
, 0);
1883 atomic_set(&memcg
->numainfo_updating
, 0);
1887 * Selecting a node where we start reclaim from. Because what we need is just
1888 * reducing usage counter, start from anywhere is O,K. Considering
1889 * memory reclaim from current node, there are pros. and cons.
1891 * Freeing memory from current node means freeing memory from a node which
1892 * we'll use or we've used. So, it may make LRU bad. And if several threads
1893 * hit limits, it will see a contention on a node. But freeing from remote
1894 * node means more costs for memory reclaim because of memory latency.
1896 * Now, we use round-robin. Better algorithm is welcomed.
1898 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1902 mem_cgroup_may_update_nodemask(memcg
);
1903 node
= memcg
->last_scanned_node
;
1905 node
= next_node(node
, memcg
->scan_nodes
);
1906 if (node
== MAX_NUMNODES
)
1907 node
= first_node(memcg
->scan_nodes
);
1909 * We call this when we hit limit, not when pages are added to LRU.
1910 * No LRU may hold pages because all pages are UNEVICTABLE or
1911 * memcg is too small and all pages are not on LRU. In that case,
1912 * we use curret node.
1914 if (unlikely(node
== MAX_NUMNODES
))
1915 node
= numa_node_id();
1917 memcg
->last_scanned_node
= node
;
1922 * Check all nodes whether it contains reclaimable pages or not.
1923 * For quick scan, we make use of scan_nodes. This will allow us to skip
1924 * unused nodes. But scan_nodes is lazily updated and may not cotain
1925 * enough new information. We need to do double check.
1927 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1932 * quick check...making use of scan_node.
1933 * We can skip unused nodes.
1935 if (!nodes_empty(memcg
->scan_nodes
)) {
1936 for (nid
= first_node(memcg
->scan_nodes
);
1938 nid
= next_node(nid
, memcg
->scan_nodes
)) {
1940 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1945 * Check rest of nodes.
1947 for_each_node_state(nid
, N_MEMORY
) {
1948 if (node_isset(nid
, memcg
->scan_nodes
))
1950 if (test_mem_cgroup_node_reclaimable(memcg
, nid
, noswap
))
1957 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1962 static bool mem_cgroup_reclaimable(struct mem_cgroup
*memcg
, bool noswap
)
1964 return test_mem_cgroup_node_reclaimable(memcg
, 0, noswap
);
1968 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1971 unsigned long *total_scanned
)
1973 struct mem_cgroup
*victim
= NULL
;
1976 unsigned long excess
;
1977 unsigned long nr_scanned
;
1978 struct mem_cgroup_reclaim_cookie reclaim
= {
1983 excess
= res_counter_soft_limit_excess(&root_memcg
->res
) >> PAGE_SHIFT
;
1986 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1991 * If we have not been able to reclaim
1992 * anything, it might because there are
1993 * no reclaimable pages under this hierarchy
1998 * We want to do more targeted reclaim.
1999 * excess >> 2 is not to excessive so as to
2000 * reclaim too much, nor too less that we keep
2001 * coming back to reclaim from this cgroup
2003 if (total
>= (excess
>> 2) ||
2004 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
2009 if (!mem_cgroup_reclaimable(victim
, false))
2011 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
2013 *total_scanned
+= nr_scanned
;
2014 if (!res_counter_soft_limit_excess(&root_memcg
->res
))
2017 mem_cgroup_iter_break(root_memcg
, victim
);
2021 #ifdef CONFIG_LOCKDEP
2022 static struct lockdep_map memcg_oom_lock_dep_map
= {
2023 .name
= "memcg_oom_lock",
2027 static DEFINE_SPINLOCK(memcg_oom_lock
);
2030 * Check OOM-Killer is already running under our hierarchy.
2031 * If someone is running, return false.
2033 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
2035 struct mem_cgroup
*iter
, *failed
= NULL
;
2037 spin_lock(&memcg_oom_lock
);
2039 for_each_mem_cgroup_tree(iter
, memcg
) {
2040 if (iter
->oom_lock
) {
2042 * this subtree of our hierarchy is already locked
2043 * so we cannot give a lock.
2046 mem_cgroup_iter_break(memcg
, iter
);
2049 iter
->oom_lock
= true;
2054 * OK, we failed to lock the whole subtree so we have
2055 * to clean up what we set up to the failing subtree
2057 for_each_mem_cgroup_tree(iter
, memcg
) {
2058 if (iter
== failed
) {
2059 mem_cgroup_iter_break(memcg
, iter
);
2062 iter
->oom_lock
= false;
2065 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
2067 spin_unlock(&memcg_oom_lock
);
2072 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
2074 struct mem_cgroup
*iter
;
2076 spin_lock(&memcg_oom_lock
);
2077 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
2078 for_each_mem_cgroup_tree(iter
, memcg
)
2079 iter
->oom_lock
= false;
2080 spin_unlock(&memcg_oom_lock
);
2083 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
2085 struct mem_cgroup
*iter
;
2087 for_each_mem_cgroup_tree(iter
, memcg
)
2088 atomic_inc(&iter
->under_oom
);
2091 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
2093 struct mem_cgroup
*iter
;
2096 * When a new child is created while the hierarchy is under oom,
2097 * mem_cgroup_oom_lock() may not be called. We have to use
2098 * atomic_add_unless() here.
2100 for_each_mem_cgroup_tree(iter
, memcg
)
2101 atomic_add_unless(&iter
->under_oom
, -1, 0);
2104 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
2106 struct oom_wait_info
{
2107 struct mem_cgroup
*memcg
;
2111 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2112 unsigned mode
, int sync
, void *arg
)
2114 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2115 struct mem_cgroup
*oom_wait_memcg
;
2116 struct oom_wait_info
*oom_wait_info
;
2118 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2119 oom_wait_memcg
= oom_wait_info
->memcg
;
2122 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2123 * Then we can use css_is_ancestor without taking care of RCU.
2125 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2126 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2128 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2131 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2133 atomic_inc(&memcg
->oom_wakeups
);
2134 /* for filtering, pass "memcg" as argument. */
2135 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2138 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2140 if (memcg
&& atomic_read(&memcg
->under_oom
))
2141 memcg_wakeup_oom(memcg
);
2144 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2146 if (!current
->memcg_oom
.may_oom
)
2149 * We are in the middle of the charge context here, so we
2150 * don't want to block when potentially sitting on a callstack
2151 * that holds all kinds of filesystem and mm locks.
2153 * Also, the caller may handle a failed allocation gracefully
2154 * (like optional page cache readahead) and so an OOM killer
2155 * invocation might not even be necessary.
2157 * That's why we don't do anything here except remember the
2158 * OOM context and then deal with it at the end of the page
2159 * fault when the stack is unwound, the locks are released,
2160 * and when we know whether the fault was overall successful.
2162 css_get(&memcg
->css
);
2163 current
->memcg_oom
.memcg
= memcg
;
2164 current
->memcg_oom
.gfp_mask
= mask
;
2165 current
->memcg_oom
.order
= order
;
2169 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2170 * @handle: actually kill/wait or just clean up the OOM state
2172 * This has to be called at the end of a page fault if the memcg OOM
2173 * handler was enabled.
2175 * Memcg supports userspace OOM handling where failed allocations must
2176 * sleep on a waitqueue until the userspace task resolves the
2177 * situation. Sleeping directly in the charge context with all kinds
2178 * of locks held is not a good idea, instead we remember an OOM state
2179 * in the task and mem_cgroup_oom_synchronize() has to be called at
2180 * the end of the page fault to complete the OOM handling.
2182 * Returns %true if an ongoing memcg OOM situation was detected and
2183 * completed, %false otherwise.
2185 bool mem_cgroup_oom_synchronize(bool handle
)
2187 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
2188 struct oom_wait_info owait
;
2191 /* OOM is global, do not handle */
2198 owait
.memcg
= memcg
;
2199 owait
.wait
.flags
= 0;
2200 owait
.wait
.func
= memcg_oom_wake_function
;
2201 owait
.wait
.private = current
;
2202 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2204 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2205 mem_cgroup_mark_under_oom(memcg
);
2207 locked
= mem_cgroup_oom_trylock(memcg
);
2210 mem_cgroup_oom_notify(memcg
);
2212 if (locked
&& !memcg
->oom_kill_disable
) {
2213 mem_cgroup_unmark_under_oom(memcg
);
2214 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2215 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
2216 current
->memcg_oom
.order
);
2219 mem_cgroup_unmark_under_oom(memcg
);
2220 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2224 mem_cgroup_oom_unlock(memcg
);
2226 * There is no guarantee that an OOM-lock contender
2227 * sees the wakeups triggered by the OOM kill
2228 * uncharges. Wake any sleepers explicitely.
2230 memcg_oom_recover(memcg
);
2233 current
->memcg_oom
.memcg
= NULL
;
2234 css_put(&memcg
->css
);
2239 * Used to update mapped file or writeback or other statistics.
2241 * Notes: Race condition
2243 * Charging occurs during page instantiation, while the page is
2244 * unmapped and locked in page migration, or while the page table is
2245 * locked in THP migration. No race is possible.
2247 * Uncharge happens to pages with zero references, no race possible.
2249 * Charge moving between groups is protected by checking mm->moving
2250 * account and taking the move_lock in the slowpath.
2253 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2254 bool *locked
, unsigned long *flags
)
2256 struct mem_cgroup
*memcg
;
2257 struct page_cgroup
*pc
;
2259 pc
= lookup_page_cgroup(page
);
2261 memcg
= pc
->mem_cgroup
;
2262 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2265 * If this memory cgroup is not under account moving, we don't
2266 * need to take move_lock_mem_cgroup(). Because we already hold
2267 * rcu_read_lock(), any calls to move_account will be delayed until
2268 * rcu_read_unlock().
2270 VM_BUG_ON(!rcu_read_lock_held());
2271 if (atomic_read(&memcg
->moving_account
) <= 0)
2274 move_lock_mem_cgroup(memcg
, flags
);
2275 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2276 move_unlock_mem_cgroup(memcg
, flags
);
2282 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2284 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2287 * It's guaranteed that pc->mem_cgroup never changes while
2288 * lock is held because a routine modifies pc->mem_cgroup
2289 * should take move_lock_mem_cgroup().
2291 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2294 void mem_cgroup_update_page_stat(struct page
*page
,
2295 enum mem_cgroup_stat_index idx
, int val
)
2297 struct mem_cgroup
*memcg
;
2298 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2299 unsigned long uninitialized_var(flags
);
2301 if (mem_cgroup_disabled())
2304 VM_BUG_ON(!rcu_read_lock_held());
2305 memcg
= pc
->mem_cgroup
;
2306 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2309 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2313 * size of first charge trial. "32" comes from vmscan.c's magic value.
2314 * TODO: maybe necessary to use big numbers in big irons.
2316 #define CHARGE_BATCH 32U
2317 struct memcg_stock_pcp
{
2318 struct mem_cgroup
*cached
; /* this never be root cgroup */
2319 unsigned int nr_pages
;
2320 struct work_struct work
;
2321 unsigned long flags
;
2322 #define FLUSHING_CACHED_CHARGE 0
2324 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2325 static DEFINE_MUTEX(percpu_charge_mutex
);
2328 * consume_stock: Try to consume stocked charge on this cpu.
2329 * @memcg: memcg to consume from.
2330 * @nr_pages: how many pages to charge.
2332 * The charges will only happen if @memcg matches the current cpu's memcg
2333 * stock, and at least @nr_pages are available in that stock. Failure to
2334 * service an allocation will refill the stock.
2336 * returns true if successful, false otherwise.
2338 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2340 struct memcg_stock_pcp
*stock
;
2343 if (nr_pages
> CHARGE_BATCH
)
2346 stock
= &get_cpu_var(memcg_stock
);
2347 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2348 stock
->nr_pages
-= nr_pages
;
2349 else /* need to call res_counter_charge */
2351 put_cpu_var(memcg_stock
);
2356 * Returns stocks cached in percpu to res_counter and reset cached information.
2358 static void drain_stock(struct memcg_stock_pcp
*stock
)
2360 struct mem_cgroup
*old
= stock
->cached
;
2362 if (stock
->nr_pages
) {
2363 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2365 res_counter_uncharge(&old
->res
, bytes
);
2366 if (do_swap_account
)
2367 res_counter_uncharge(&old
->memsw
, bytes
);
2368 stock
->nr_pages
= 0;
2370 stock
->cached
= NULL
;
2374 * This must be called under preempt disabled or must be called by
2375 * a thread which is pinned to local cpu.
2377 static void drain_local_stock(struct work_struct
*dummy
)
2379 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
2381 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2384 static void __init
memcg_stock_init(void)
2388 for_each_possible_cpu(cpu
) {
2389 struct memcg_stock_pcp
*stock
=
2390 &per_cpu(memcg_stock
, cpu
);
2391 INIT_WORK(&stock
->work
, drain_local_stock
);
2396 * Cache charges(val) which is from res_counter, to local per_cpu area.
2397 * This will be consumed by consume_stock() function, later.
2399 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2401 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2403 if (stock
->cached
!= memcg
) { /* reset if necessary */
2405 stock
->cached
= memcg
;
2407 stock
->nr_pages
+= nr_pages
;
2408 put_cpu_var(memcg_stock
);
2412 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2413 * of the hierarchy under it. sync flag says whether we should block
2414 * until the work is done.
2416 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2420 /* Notify other cpus that system-wide "drain" is running */
2423 for_each_online_cpu(cpu
) {
2424 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2425 struct mem_cgroup
*memcg
;
2427 memcg
= stock
->cached
;
2428 if (!memcg
|| !stock
->nr_pages
)
2430 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2432 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2434 drain_local_stock(&stock
->work
);
2436 schedule_work_on(cpu
, &stock
->work
);
2444 for_each_online_cpu(cpu
) {
2445 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2446 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2447 flush_work(&stock
->work
);
2454 * Tries to drain stocked charges in other cpus. This function is asynchronous
2455 * and just put a work per cpu for draining localy on each cpu. Caller can
2456 * expects some charges will be back to res_counter later but cannot wait for
2459 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2462 * If someone calls draining, avoid adding more kworker runs.
2464 if (!mutex_trylock(&percpu_charge_mutex
))
2466 drain_all_stock(root_memcg
, false);
2467 mutex_unlock(&percpu_charge_mutex
);
2470 /* This is a synchronous drain interface. */
2471 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2473 /* called when force_empty is called */
2474 mutex_lock(&percpu_charge_mutex
);
2475 drain_all_stock(root_memcg
, true);
2476 mutex_unlock(&percpu_charge_mutex
);
2480 * This function drains percpu counter value from DEAD cpu and
2481 * move it to local cpu. Note that this function can be preempted.
2483 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2487 spin_lock(&memcg
->pcp_counter_lock
);
2488 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2489 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2491 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2492 memcg
->nocpu_base
.count
[i
] += x
;
2494 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2495 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2497 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2498 memcg
->nocpu_base
.events
[i
] += x
;
2500 spin_unlock(&memcg
->pcp_counter_lock
);
2503 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2504 unsigned long action
,
2507 int cpu
= (unsigned long)hcpu
;
2508 struct memcg_stock_pcp
*stock
;
2509 struct mem_cgroup
*iter
;
2511 if (action
== CPU_ONLINE
)
2514 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2517 for_each_mem_cgroup(iter
)
2518 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2520 stock
= &per_cpu(memcg_stock
, cpu
);
2525 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2526 unsigned int nr_pages
)
2528 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2529 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2530 struct mem_cgroup
*mem_over_limit
;
2531 struct res_counter
*fail_res
;
2532 unsigned long nr_reclaimed
;
2533 unsigned long flags
= 0;
2534 unsigned long long size
;
2538 if (consume_stock(memcg
, nr_pages
))
2541 size
= batch
* PAGE_SIZE
;
2542 if (!res_counter_charge(&memcg
->res
, size
, &fail_res
)) {
2543 if (!do_swap_account
)
2545 if (!res_counter_charge(&memcg
->memsw
, size
, &fail_res
))
2547 res_counter_uncharge(&memcg
->res
, size
);
2548 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2549 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2551 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2553 if (batch
> nr_pages
) {
2559 * Unlike in global OOM situations, memcg is not in a physical
2560 * memory shortage. Allow dying and OOM-killed tasks to
2561 * bypass the last charges so that they can exit quickly and
2562 * free their memory.
2564 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2565 fatal_signal_pending(current
) ||
2566 current
->flags
& PF_EXITING
))
2569 if (unlikely(task_in_memcg_oom(current
)))
2572 if (!(gfp_mask
& __GFP_WAIT
))
2575 nr_reclaimed
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2577 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2580 if (gfp_mask
& __GFP_NORETRY
)
2583 * Even though the limit is exceeded at this point, reclaim
2584 * may have been able to free some pages. Retry the charge
2585 * before killing the task.
2587 * Only for regular pages, though: huge pages are rather
2588 * unlikely to succeed so close to the limit, and we fall back
2589 * to regular pages anyway in case of failure.
2591 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2594 * At task move, charge accounts can be doubly counted. So, it's
2595 * better to wait until the end of task_move if something is going on.
2597 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2603 if (gfp_mask
& __GFP_NOFAIL
)
2606 if (fatal_signal_pending(current
))
2609 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2611 if (!(gfp_mask
& __GFP_NOFAIL
))
2614 memcg
= root_mem_cgroup
;
2619 if (batch
> nr_pages
)
2620 refill_stock(memcg
, batch
- nr_pages
);
2625 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2627 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2629 res_counter_uncharge(&memcg
->res
, bytes
);
2630 if (do_swap_account
)
2631 res_counter_uncharge(&memcg
->memsw
, bytes
);
2635 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2636 * This is useful when moving usage to parent cgroup.
2638 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2639 unsigned int nr_pages
)
2641 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2643 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2644 if (do_swap_account
)
2645 res_counter_uncharge_until(&memcg
->memsw
,
2646 memcg
->memsw
.parent
, bytes
);
2650 * A helper function to get mem_cgroup from ID. must be called under
2651 * rcu_read_lock(). The caller is responsible for calling
2652 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2653 * refcnt from swap can be called against removed memcg.)
2655 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2657 /* ID 0 is unused ID */
2660 return mem_cgroup_from_id(id
);
2664 * try_get_mem_cgroup_from_page - look up page's memcg association
2667 * Look up, get a css reference, and return the memcg that owns @page.
2669 * The page must be locked to prevent racing with swap-in and page
2670 * cache charges. If coming from an unlocked page table, the caller
2671 * must ensure the page is on the LRU or this can race with charging.
2673 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2675 struct mem_cgroup
*memcg
= NULL
;
2676 struct page_cgroup
*pc
;
2680 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
2682 pc
= lookup_page_cgroup(page
);
2683 if (PageCgroupUsed(pc
)) {
2684 memcg
= pc
->mem_cgroup
;
2685 if (memcg
&& !css_tryget_online(&memcg
->css
))
2687 } else if (PageSwapCache(page
)) {
2688 ent
.val
= page_private(page
);
2689 id
= lookup_swap_cgroup_id(ent
);
2691 memcg
= mem_cgroup_lookup(id
);
2692 if (memcg
&& !css_tryget_online(&memcg
->css
))
2699 static void lock_page_lru(struct page
*page
, int *isolated
)
2701 struct zone
*zone
= page_zone(page
);
2703 spin_lock_irq(&zone
->lru_lock
);
2704 if (PageLRU(page
)) {
2705 struct lruvec
*lruvec
;
2707 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2709 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2715 static void unlock_page_lru(struct page
*page
, int isolated
)
2717 struct zone
*zone
= page_zone(page
);
2720 struct lruvec
*lruvec
;
2722 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2723 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2725 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2727 spin_unlock_irq(&zone
->lru_lock
);
2730 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2733 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2736 VM_BUG_ON_PAGE(PageCgroupUsed(pc
), page
);
2738 * we don't need page_cgroup_lock about tail pages, becase they are not
2739 * accessed by any other context at this point.
2743 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2744 * may already be on some other mem_cgroup's LRU. Take care of it.
2747 lock_page_lru(page
, &isolated
);
2750 * Nobody should be changing or seriously looking at
2751 * pc->mem_cgroup and pc->flags at this point:
2753 * - the page is uncharged
2755 * - the page is off-LRU
2757 * - an anonymous fault has exclusive page access, except for
2758 * a locked page table
2760 * - a page cache insertion, a swapin fault, or a migration
2761 * have the page locked
2763 pc
->mem_cgroup
= memcg
;
2764 pc
->flags
= PCG_USED
| PCG_MEM
| (do_swap_account
? PCG_MEMSW
: 0);
2767 unlock_page_lru(page
, isolated
);
2770 static DEFINE_MUTEX(set_limit_mutex
);
2772 #ifdef CONFIG_MEMCG_KMEM
2774 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2775 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2777 static DEFINE_MUTEX(memcg_slab_mutex
);
2779 static DEFINE_MUTEX(activate_kmem_mutex
);
2781 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2783 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2784 memcg_kmem_is_active(memcg
);
2788 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2789 * in the memcg_cache_params struct.
2791 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2793 struct kmem_cache
*cachep
;
2795 VM_BUG_ON(p
->is_root_cache
);
2796 cachep
= p
->root_cache
;
2797 return cache_from_memcg_idx(cachep
, memcg_cache_id(p
->memcg
));
2800 #ifdef CONFIG_SLABINFO
2801 static int mem_cgroup_slabinfo_read(struct seq_file
*m
, void *v
)
2803 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
2804 struct memcg_cache_params
*params
;
2806 if (!memcg_can_account_kmem(memcg
))
2809 print_slabinfo_header(m
);
2811 mutex_lock(&memcg_slab_mutex
);
2812 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2813 cache_show(memcg_params_to_cache(params
), m
);
2814 mutex_unlock(&memcg_slab_mutex
);
2820 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2822 struct res_counter
*fail_res
;
2825 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2829 ret
= try_charge(memcg
, gfp
, size
>> PAGE_SHIFT
);
2830 if (ret
== -EINTR
) {
2832 * try_charge() chose to bypass to root due to OOM kill or
2833 * fatal signal. Since our only options are to either fail
2834 * the allocation or charge it to this cgroup, do it as a
2835 * temporary condition. But we can't fail. From a kmem/slab
2836 * perspective, the cache has already been selected, by
2837 * mem_cgroup_kmem_get_cache(), so it is too late to change
2840 * This condition will only trigger if the task entered
2841 * memcg_charge_kmem in a sane state, but was OOM-killed
2842 * during try_charge() above. Tasks that were already dying
2843 * when the allocation triggers should have been already
2844 * directed to the root cgroup in memcontrol.h
2846 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2847 if (do_swap_account
)
2848 res_counter_charge_nofail(&memcg
->memsw
, size
,
2852 res_counter_uncharge(&memcg
->kmem
, size
);
2857 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2859 res_counter_uncharge(&memcg
->res
, size
);
2860 if (do_swap_account
)
2861 res_counter_uncharge(&memcg
->memsw
, size
);
2864 if (res_counter_uncharge(&memcg
->kmem
, size
))
2868 * Releases a reference taken in kmem_cgroup_css_offline in case
2869 * this last uncharge is racing with the offlining code or it is
2870 * outliving the memcg existence.
2872 * The memory barrier imposed by test&clear is paired with the
2873 * explicit one in memcg_kmem_mark_dead().
2875 if (memcg_kmem_test_and_clear_dead(memcg
))
2876 css_put(&memcg
->css
);
2880 * helper for acessing a memcg's index. It will be used as an index in the
2881 * child cache array in kmem_cache, and also to derive its name. This function
2882 * will return -1 when this is not a kmem-limited memcg.
2884 int memcg_cache_id(struct mem_cgroup
*memcg
)
2886 return memcg
? memcg
->kmemcg_id
: -1;
2889 static size_t memcg_caches_array_size(int num_groups
)
2892 if (num_groups
<= 0)
2895 size
= 2 * num_groups
;
2896 if (size
< MEMCG_CACHES_MIN_SIZE
)
2897 size
= MEMCG_CACHES_MIN_SIZE
;
2898 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2899 size
= MEMCG_CACHES_MAX_SIZE
;
2905 * We should update the current array size iff all caches updates succeed. This
2906 * can only be done from the slab side. The slab mutex needs to be held when
2909 void memcg_update_array_size(int num
)
2911 if (num
> memcg_limited_groups_array_size
)
2912 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2915 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2917 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2919 VM_BUG_ON(!is_root_cache(s
));
2921 if (num_groups
> memcg_limited_groups_array_size
) {
2923 struct memcg_cache_params
*new_params
;
2924 ssize_t size
= memcg_caches_array_size(num_groups
);
2926 size
*= sizeof(void *);
2927 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
2929 new_params
= kzalloc(size
, GFP_KERNEL
);
2933 new_params
->is_root_cache
= true;
2936 * There is the chance it will be bigger than
2937 * memcg_limited_groups_array_size, if we failed an allocation
2938 * in a cache, in which case all caches updated before it, will
2939 * have a bigger array.
2941 * But if that is the case, the data after
2942 * memcg_limited_groups_array_size is certainly unused
2944 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
2945 if (!cur_params
->memcg_caches
[i
])
2947 new_params
->memcg_caches
[i
] =
2948 cur_params
->memcg_caches
[i
];
2952 * Ideally, we would wait until all caches succeed, and only
2953 * then free the old one. But this is not worth the extra
2954 * pointer per-cache we'd have to have for this.
2956 * It is not a big deal if some caches are left with a size
2957 * bigger than the others. And all updates will reset this
2960 rcu_assign_pointer(s
->memcg_params
, new_params
);
2962 kfree_rcu(cur_params
, rcu_head
);
2967 int memcg_alloc_cache_params(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
2968 struct kmem_cache
*root_cache
)
2972 if (!memcg_kmem_enabled())
2976 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
2977 size
+= memcg_limited_groups_array_size
* sizeof(void *);
2979 size
= sizeof(struct memcg_cache_params
);
2981 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2982 if (!s
->memcg_params
)
2986 s
->memcg_params
->memcg
= memcg
;
2987 s
->memcg_params
->root_cache
= root_cache
;
2988 css_get(&memcg
->css
);
2990 s
->memcg_params
->is_root_cache
= true;
2995 void memcg_free_cache_params(struct kmem_cache
*s
)
2997 if (!s
->memcg_params
)
2999 if (!s
->memcg_params
->is_root_cache
)
3000 css_put(&s
->memcg_params
->memcg
->css
);
3001 kfree(s
->memcg_params
);
3004 static void memcg_register_cache(struct mem_cgroup
*memcg
,
3005 struct kmem_cache
*root_cache
)
3007 static char memcg_name_buf
[NAME_MAX
+ 1]; /* protected by
3009 struct kmem_cache
*cachep
;
3012 lockdep_assert_held(&memcg_slab_mutex
);
3014 id
= memcg_cache_id(memcg
);
3017 * Since per-memcg caches are created asynchronously on first
3018 * allocation (see memcg_kmem_get_cache()), several threads can try to
3019 * create the same cache, but only one of them may succeed.
3021 if (cache_from_memcg_idx(root_cache
, id
))
3024 cgroup_name(memcg
->css
.cgroup
, memcg_name_buf
, NAME_MAX
+ 1);
3025 cachep
= memcg_create_kmem_cache(memcg
, root_cache
, memcg_name_buf
);
3027 * If we could not create a memcg cache, do not complain, because
3028 * that's not critical at all as we can always proceed with the root
3034 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
3037 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3038 * barrier here to ensure nobody will see the kmem_cache partially
3043 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
]);
3044 root_cache
->memcg_params
->memcg_caches
[id
] = cachep
;
3047 static void memcg_unregister_cache(struct kmem_cache
*cachep
)
3049 struct kmem_cache
*root_cache
;
3050 struct mem_cgroup
*memcg
;
3053 lockdep_assert_held(&memcg_slab_mutex
);
3055 BUG_ON(is_root_cache(cachep
));
3057 root_cache
= cachep
->memcg_params
->root_cache
;
3058 memcg
= cachep
->memcg_params
->memcg
;
3059 id
= memcg_cache_id(memcg
);
3061 BUG_ON(root_cache
->memcg_params
->memcg_caches
[id
] != cachep
);
3062 root_cache
->memcg_params
->memcg_caches
[id
] = NULL
;
3064 list_del(&cachep
->memcg_params
->list
);
3066 kmem_cache_destroy(cachep
);
3070 * During the creation a new cache, we need to disable our accounting mechanism
3071 * altogether. This is true even if we are not creating, but rather just
3072 * enqueing new caches to be created.
3074 * This is because that process will trigger allocations; some visible, like
3075 * explicit kmallocs to auxiliary data structures, name strings and internal
3076 * cache structures; some well concealed, like INIT_WORK() that can allocate
3077 * objects during debug.
3079 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3080 * to it. This may not be a bounded recursion: since the first cache creation
3081 * failed to complete (waiting on the allocation), we'll just try to create the
3082 * cache again, failing at the same point.
3084 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3085 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3086 * inside the following two functions.
3088 static inline void memcg_stop_kmem_account(void)
3090 VM_BUG_ON(!current
->mm
);
3091 current
->memcg_kmem_skip_account
++;
3094 static inline void memcg_resume_kmem_account(void)
3096 VM_BUG_ON(!current
->mm
);
3097 current
->memcg_kmem_skip_account
--;
3100 int __memcg_cleanup_cache_params(struct kmem_cache
*s
)
3102 struct kmem_cache
*c
;
3105 mutex_lock(&memcg_slab_mutex
);
3106 for_each_memcg_cache_index(i
) {
3107 c
= cache_from_memcg_idx(s
, i
);
3111 memcg_unregister_cache(c
);
3113 if (cache_from_memcg_idx(s
, i
))
3116 mutex_unlock(&memcg_slab_mutex
);
3120 static void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3122 struct kmem_cache
*cachep
;
3123 struct memcg_cache_params
*params
, *tmp
;
3125 if (!memcg_kmem_is_active(memcg
))
3128 mutex_lock(&memcg_slab_mutex
);
3129 list_for_each_entry_safe(params
, tmp
, &memcg
->memcg_slab_caches
, list
) {
3130 cachep
= memcg_params_to_cache(params
);
3131 kmem_cache_shrink(cachep
);
3132 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3133 memcg_unregister_cache(cachep
);
3135 mutex_unlock(&memcg_slab_mutex
);
3138 struct memcg_register_cache_work
{
3139 struct mem_cgroup
*memcg
;
3140 struct kmem_cache
*cachep
;
3141 struct work_struct work
;
3144 static void memcg_register_cache_func(struct work_struct
*w
)
3146 struct memcg_register_cache_work
*cw
=
3147 container_of(w
, struct memcg_register_cache_work
, work
);
3148 struct mem_cgroup
*memcg
= cw
->memcg
;
3149 struct kmem_cache
*cachep
= cw
->cachep
;
3151 mutex_lock(&memcg_slab_mutex
);
3152 memcg_register_cache(memcg
, cachep
);
3153 mutex_unlock(&memcg_slab_mutex
);
3155 css_put(&memcg
->css
);
3160 * Enqueue the creation of a per-memcg kmem_cache.
3162 static void __memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3163 struct kmem_cache
*cachep
)
3165 struct memcg_register_cache_work
*cw
;
3167 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
3169 css_put(&memcg
->css
);
3174 cw
->cachep
= cachep
;
3176 INIT_WORK(&cw
->work
, memcg_register_cache_func
);
3177 schedule_work(&cw
->work
);
3180 static void memcg_schedule_register_cache(struct mem_cgroup
*memcg
,
3181 struct kmem_cache
*cachep
)
3184 * We need to stop accounting when we kmalloc, because if the
3185 * corresponding kmalloc cache is not yet created, the first allocation
3186 * in __memcg_schedule_register_cache will recurse.
3188 * However, it is better to enclose the whole function. Depending on
3189 * the debugging options enabled, INIT_WORK(), for instance, can
3190 * trigger an allocation. This too, will make us recurse. Because at
3191 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3192 * the safest choice is to do it like this, wrapping the whole function.
3194 memcg_stop_kmem_account();
3195 __memcg_schedule_register_cache(memcg
, cachep
);
3196 memcg_resume_kmem_account();
3199 int __memcg_charge_slab(struct kmem_cache
*cachep
, gfp_t gfp
, int order
)
3203 res
= memcg_charge_kmem(cachep
->memcg_params
->memcg
, gfp
,
3204 PAGE_SIZE
<< order
);
3206 atomic_add(1 << order
, &cachep
->memcg_params
->nr_pages
);
3210 void __memcg_uncharge_slab(struct kmem_cache
*cachep
, int order
)
3212 memcg_uncharge_kmem(cachep
->memcg_params
->memcg
, PAGE_SIZE
<< order
);
3213 atomic_sub(1 << order
, &cachep
->memcg_params
->nr_pages
);
3217 * Return the kmem_cache we're supposed to use for a slab allocation.
3218 * We try to use the current memcg's version of the cache.
3220 * If the cache does not exist yet, if we are the first user of it,
3221 * we either create it immediately, if possible, or create it asynchronously
3223 * In the latter case, we will let the current allocation go through with
3224 * the original cache.
3226 * Can't be called in interrupt context or from kernel threads.
3227 * This function needs to be called with rcu_read_lock() held.
3229 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3232 struct mem_cgroup
*memcg
;
3233 struct kmem_cache
*memcg_cachep
;
3235 VM_BUG_ON(!cachep
->memcg_params
);
3236 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3238 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3242 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3244 if (!memcg_can_account_kmem(memcg
))
3247 memcg_cachep
= cache_from_memcg_idx(cachep
, memcg_cache_id(memcg
));
3248 if (likely(memcg_cachep
)) {
3249 cachep
= memcg_cachep
;
3253 /* The corresponding put will be done in the workqueue. */
3254 if (!css_tryget_online(&memcg
->css
))
3259 * If we are in a safe context (can wait, and not in interrupt
3260 * context), we could be be predictable and return right away.
3261 * This would guarantee that the allocation being performed
3262 * already belongs in the new cache.
3264 * However, there are some clashes that can arrive from locking.
3265 * For instance, because we acquire the slab_mutex while doing
3266 * memcg_create_kmem_cache, this means no further allocation
3267 * could happen with the slab_mutex held. So it's better to
3270 memcg_schedule_register_cache(memcg
, cachep
);
3278 * We need to verify if the allocation against current->mm->owner's memcg is
3279 * possible for the given order. But the page is not allocated yet, so we'll
3280 * need a further commit step to do the final arrangements.
3282 * It is possible for the task to switch cgroups in this mean time, so at
3283 * commit time, we can't rely on task conversion any longer. We'll then use
3284 * the handle argument to return to the caller which cgroup we should commit
3285 * against. We could also return the memcg directly and avoid the pointer
3286 * passing, but a boolean return value gives better semantics considering
3287 * the compiled-out case as well.
3289 * Returning true means the allocation is possible.
3292 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3294 struct mem_cgroup
*memcg
;
3300 * Disabling accounting is only relevant for some specific memcg
3301 * internal allocations. Therefore we would initially not have such
3302 * check here, since direct calls to the page allocator that are
3303 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3304 * outside memcg core. We are mostly concerned with cache allocations,
3305 * and by having this test at memcg_kmem_get_cache, we are already able
3306 * to relay the allocation to the root cache and bypass the memcg cache
3309 * There is one exception, though: the SLUB allocator does not create
3310 * large order caches, but rather service large kmallocs directly from
3311 * the page allocator. Therefore, the following sequence when backed by
3312 * the SLUB allocator:
3314 * memcg_stop_kmem_account();
3315 * kmalloc(<large_number>)
3316 * memcg_resume_kmem_account();
3318 * would effectively ignore the fact that we should skip accounting,
3319 * since it will drive us directly to this function without passing
3320 * through the cache selector memcg_kmem_get_cache. Such large
3321 * allocations are extremely rare but can happen, for instance, for the
3322 * cache arrays. We bring this test here.
3324 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3327 memcg
= get_mem_cgroup_from_mm(current
->mm
);
3329 if (!memcg_can_account_kmem(memcg
)) {
3330 css_put(&memcg
->css
);
3334 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3338 css_put(&memcg
->css
);
3342 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3345 struct page_cgroup
*pc
;
3347 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3349 /* The page allocation failed. Revert */
3351 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3355 * The page is freshly allocated and not visible to any
3356 * outside callers yet. Set up pc non-atomically.
3358 pc
= lookup_page_cgroup(page
);
3359 pc
->mem_cgroup
= memcg
;
3360 pc
->flags
= PCG_USED
;
3363 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3365 struct mem_cgroup
*memcg
= NULL
;
3366 struct page_cgroup
*pc
;
3369 pc
= lookup_page_cgroup(page
);
3370 if (!PageCgroupUsed(pc
))
3373 memcg
= pc
->mem_cgroup
;
3377 * We trust that only if there is a memcg associated with the page, it
3378 * is a valid allocation
3383 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
3384 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3387 static inline void memcg_unregister_all_caches(struct mem_cgroup
*memcg
)
3390 #endif /* CONFIG_MEMCG_KMEM */
3392 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3395 * Because tail pages are not marked as "used", set it. We're under
3396 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3397 * charge/uncharge will be never happen and move_account() is done under
3398 * compound_lock(), so we don't have to take care of races.
3400 void mem_cgroup_split_huge_fixup(struct page
*head
)
3402 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3403 struct page_cgroup
*pc
;
3404 struct mem_cgroup
*memcg
;
3407 if (mem_cgroup_disabled())
3410 memcg
= head_pc
->mem_cgroup
;
3411 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3413 pc
->mem_cgroup
= memcg
;
3414 pc
->flags
= head_pc
->flags
;
3416 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3419 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3422 * mem_cgroup_move_account - move account of the page
3424 * @nr_pages: number of regular pages (>1 for huge pages)
3425 * @pc: page_cgroup of the page.
3426 * @from: mem_cgroup which the page is moved from.
3427 * @to: mem_cgroup which the page is moved to. @from != @to.
3429 * The caller must confirm following.
3430 * - page is not on LRU (isolate_page() is useful.)
3431 * - compound_lock is held when nr_pages > 1
3433 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3436 static int mem_cgroup_move_account(struct page
*page
,
3437 unsigned int nr_pages
,
3438 struct page_cgroup
*pc
,
3439 struct mem_cgroup
*from
,
3440 struct mem_cgroup
*to
)
3442 unsigned long flags
;
3445 VM_BUG_ON(from
== to
);
3446 VM_BUG_ON_PAGE(PageLRU(page
), page
);
3448 * The page is isolated from LRU. So, collapse function
3449 * will not handle this page. But page splitting can happen.
3450 * Do this check under compound_page_lock(). The caller should
3454 if (nr_pages
> 1 && !PageTransHuge(page
))
3458 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3459 * of its source page while we change it: page migration takes
3460 * both pages off the LRU, but page cache replacement doesn't.
3462 if (!trylock_page(page
))
3466 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3469 move_lock_mem_cgroup(from
, &flags
);
3471 if (!PageAnon(page
) && page_mapped(page
)) {
3472 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3474 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
3478 if (PageWriteback(page
)) {
3479 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3481 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
3486 * It is safe to change pc->mem_cgroup here because the page
3487 * is referenced, charged, and isolated - we can't race with
3488 * uncharging, charging, migration, or LRU putback.
3491 /* caller should have done css_get */
3492 pc
->mem_cgroup
= to
;
3493 move_unlock_mem_cgroup(from
, &flags
);
3496 local_irq_disable();
3497 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
3498 memcg_check_events(to
, page
);
3499 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
3500 memcg_check_events(from
, page
);
3509 * mem_cgroup_move_parent - moves page to the parent group
3510 * @page: the page to move
3511 * @pc: page_cgroup of the page
3512 * @child: page's cgroup
3514 * move charges to its parent or the root cgroup if the group has no
3515 * parent (aka use_hierarchy==0).
3516 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3517 * mem_cgroup_move_account fails) the failure is always temporary and
3518 * it signals a race with a page removal/uncharge or migration. In the
3519 * first case the page is on the way out and it will vanish from the LRU
3520 * on the next attempt and the call should be retried later.
3521 * Isolation from the LRU fails only if page has been isolated from
3522 * the LRU since we looked at it and that usually means either global
3523 * reclaim or migration going on. The page will either get back to the
3525 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3526 * (!PageCgroupUsed) or moved to a different group. The page will
3527 * disappear in the next attempt.
3529 static int mem_cgroup_move_parent(struct page
*page
,
3530 struct page_cgroup
*pc
,
3531 struct mem_cgroup
*child
)
3533 struct mem_cgroup
*parent
;
3534 unsigned int nr_pages
;
3535 unsigned long uninitialized_var(flags
);
3538 VM_BUG_ON(mem_cgroup_is_root(child
));
3541 if (!get_page_unless_zero(page
))
3543 if (isolate_lru_page(page
))
3546 nr_pages
= hpage_nr_pages(page
);
3548 parent
= parent_mem_cgroup(child
);
3550 * If no parent, move charges to root cgroup.
3553 parent
= root_mem_cgroup
;
3556 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
3557 flags
= compound_lock_irqsave(page
);
3560 ret
= mem_cgroup_move_account(page
, nr_pages
,
3563 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3566 compound_unlock_irqrestore(page
, flags
);
3567 putback_lru_page(page
);
3574 #ifdef CONFIG_MEMCG_SWAP
3575 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
3578 int val
= (charge
) ? 1 : -1;
3579 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
3583 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3584 * @entry: swap entry to be moved
3585 * @from: mem_cgroup which the entry is moved from
3586 * @to: mem_cgroup which the entry is moved to
3588 * It succeeds only when the swap_cgroup's record for this entry is the same
3589 * as the mem_cgroup's id of @from.
3591 * Returns 0 on success, -EINVAL on failure.
3593 * The caller must have charged to @to, IOW, called res_counter_charge() about
3594 * both res and memsw, and called css_get().
3596 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
3597 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3599 unsigned short old_id
, new_id
;
3601 old_id
= mem_cgroup_id(from
);
3602 new_id
= mem_cgroup_id(to
);
3604 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
3605 mem_cgroup_swap_statistics(from
, false);
3606 mem_cgroup_swap_statistics(to
, true);
3608 * This function is only called from task migration context now.
3609 * It postpones res_counter and refcount handling till the end
3610 * of task migration(mem_cgroup_clear_mc()) for performance
3611 * improvement. But we cannot postpone css_get(to) because if
3612 * the process that has been moved to @to does swap-in, the
3613 * refcount of @to might be decreased to 0.
3615 * We are in attach() phase, so the cgroup is guaranteed to be
3616 * alive, so we can just call css_get().
3624 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
3625 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
3631 #ifdef CONFIG_DEBUG_VM
3632 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
3634 struct page_cgroup
*pc
;
3636 pc
= lookup_page_cgroup(page
);
3638 * Can be NULL while feeding pages into the page allocator for
3639 * the first time, i.e. during boot or memory hotplug;
3640 * or when mem_cgroup_disabled().
3642 if (likely(pc
) && PageCgroupUsed(pc
))
3647 bool mem_cgroup_bad_page_check(struct page
*page
)
3649 if (mem_cgroup_disabled())
3652 return lookup_page_cgroup_used(page
) != NULL
;
3655 void mem_cgroup_print_bad_page(struct page
*page
)
3657 struct page_cgroup
*pc
;
3659 pc
= lookup_page_cgroup_used(page
);
3661 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3662 pc
, pc
->flags
, pc
->mem_cgroup
);
3667 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3668 unsigned long long val
)
3671 u64 memswlimit
, memlimit
;
3673 int children
= mem_cgroup_count_children(memcg
);
3674 u64 curusage
, oldusage
;
3678 * For keeping hierarchical_reclaim simple, how long we should retry
3679 * is depends on callers. We set our retry-count to be function
3680 * of # of children which we should visit in this loop.
3682 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
3684 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3687 while (retry_count
) {
3688 if (signal_pending(current
)) {
3693 * Rather than hide all in some function, I do this in
3694 * open coded manner. You see what this really does.
3695 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3697 mutex_lock(&set_limit_mutex
);
3698 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3699 if (memswlimit
< val
) {
3701 mutex_unlock(&set_limit_mutex
);
3705 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3709 ret
= res_counter_set_limit(&memcg
->res
, val
);
3711 if (memswlimit
== val
)
3712 memcg
->memsw_is_minimum
= true;
3714 memcg
->memsw_is_minimum
= false;
3716 mutex_unlock(&set_limit_mutex
);
3721 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
3722 MEM_CGROUP_RECLAIM_SHRINK
);
3723 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3724 /* Usage is reduced ? */
3725 if (curusage
>= oldusage
)
3728 oldusage
= curusage
;
3730 if (!ret
&& enlarge
)
3731 memcg_oom_recover(memcg
);
3736 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3737 unsigned long long val
)
3740 u64 memlimit
, memswlimit
, oldusage
, curusage
;
3741 int children
= mem_cgroup_count_children(memcg
);
3745 /* see mem_cgroup_resize_res_limit */
3746 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
3747 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3748 while (retry_count
) {
3749 if (signal_pending(current
)) {
3754 * Rather than hide all in some function, I do this in
3755 * open coded manner. You see what this really does.
3756 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3758 mutex_lock(&set_limit_mutex
);
3759 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3760 if (memlimit
> val
) {
3762 mutex_unlock(&set_limit_mutex
);
3765 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3766 if (memswlimit
< val
)
3768 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
3770 if (memlimit
== val
)
3771 memcg
->memsw_is_minimum
= true;
3773 memcg
->memsw_is_minimum
= false;
3775 mutex_unlock(&set_limit_mutex
);
3780 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
3781 MEM_CGROUP_RECLAIM_NOSWAP
|
3782 MEM_CGROUP_RECLAIM_SHRINK
);
3783 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3784 /* Usage is reduced ? */
3785 if (curusage
>= oldusage
)
3788 oldusage
= curusage
;
3790 if (!ret
&& enlarge
)
3791 memcg_oom_recover(memcg
);
3795 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3797 unsigned long *total_scanned
)
3799 unsigned long nr_reclaimed
= 0;
3800 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3801 unsigned long reclaimed
;
3803 struct mem_cgroup_tree_per_zone
*mctz
;
3804 unsigned long long excess
;
3805 unsigned long nr_scanned
;
3810 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3812 * This loop can run a while, specially if mem_cgroup's continuously
3813 * keep exceeding their soft limit and putting the system under
3820 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3825 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
3826 gfp_mask
, &nr_scanned
);
3827 nr_reclaimed
+= reclaimed
;
3828 *total_scanned
+= nr_scanned
;
3829 spin_lock_irq(&mctz
->lock
);
3832 * If we failed to reclaim anything from this memory cgroup
3833 * it is time to move on to the next cgroup
3839 * Loop until we find yet another one.
3841 * By the time we get the soft_limit lock
3842 * again, someone might have aded the
3843 * group back on the RB tree. Iterate to
3844 * make sure we get a different mem.
3845 * mem_cgroup_largest_soft_limit_node returns
3846 * NULL if no other cgroup is present on
3850 __mem_cgroup_largest_soft_limit_node(mctz
);
3852 css_put(&next_mz
->memcg
->css
);
3853 else /* next_mz == NULL or other memcg */
3857 __mem_cgroup_remove_exceeded(mz
, mctz
);
3858 excess
= res_counter_soft_limit_excess(&mz
->memcg
->res
);
3860 * One school of thought says that we should not add
3861 * back the node to the tree if reclaim returns 0.
3862 * But our reclaim could return 0, simply because due
3863 * to priority we are exposing a smaller subset of
3864 * memory to reclaim from. Consider this as a longer
3867 /* If excess == 0, no tree ops */
3868 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
3869 spin_unlock_irq(&mctz
->lock
);
3870 css_put(&mz
->memcg
->css
);
3873 * Could not reclaim anything and there are no more
3874 * mem cgroups to try or we seem to be looping without
3875 * reclaiming anything.
3877 if (!nr_reclaimed
&&
3879 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3881 } while (!nr_reclaimed
);
3883 css_put(&next_mz
->memcg
->css
);
3884 return nr_reclaimed
;
3888 * mem_cgroup_force_empty_list - clears LRU of a group
3889 * @memcg: group to clear
3892 * @lru: lru to to clear
3894 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3895 * reclaim the pages page themselves - pages are moved to the parent (or root)
3898 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
3899 int node
, int zid
, enum lru_list lru
)
3901 struct lruvec
*lruvec
;
3902 unsigned long flags
;
3903 struct list_head
*list
;
3907 zone
= &NODE_DATA(node
)->node_zones
[zid
];
3908 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
3909 list
= &lruvec
->lists
[lru
];
3913 struct page_cgroup
*pc
;
3916 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3917 if (list_empty(list
)) {
3918 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3921 page
= list_entry(list
->prev
, struct page
, lru
);
3923 list_move(&page
->lru
, list
);
3925 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3928 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3930 pc
= lookup_page_cgroup(page
);
3932 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
3933 /* found lock contention or "pc" is obsolete. */
3938 } while (!list_empty(list
));
3942 * make mem_cgroup's charge to be 0 if there is no task by moving
3943 * all the charges and pages to the parent.
3944 * This enables deleting this mem_cgroup.
3946 * Caller is responsible for holding css reference on the memcg.
3948 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
3954 /* This is for making all *used* pages to be on LRU. */
3955 lru_add_drain_all();
3956 drain_all_stock_sync(memcg
);
3957 mem_cgroup_start_move(memcg
);
3958 for_each_node_state(node
, N_MEMORY
) {
3959 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3962 mem_cgroup_force_empty_list(memcg
,
3967 mem_cgroup_end_move(memcg
);
3968 memcg_oom_recover(memcg
);
3972 * Kernel memory may not necessarily be trackable to a specific
3973 * process. So they are not migrated, and therefore we can't
3974 * expect their value to drop to 0 here.
3975 * Having res filled up with kmem only is enough.
3977 * This is a safety check because mem_cgroup_force_empty_list
3978 * could have raced with mem_cgroup_replace_page_cache callers
3979 * so the lru seemed empty but the page could have been added
3980 * right after the check. RES_USAGE should be safe as we always
3981 * charge before adding to the LRU.
3983 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
3984 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
3985 } while (usage
> 0);
3989 * Test whether @memcg has children, dead or alive. Note that this
3990 * function doesn't care whether @memcg has use_hierarchy enabled and
3991 * returns %true if there are child csses according to the cgroup
3992 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3994 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
3999 * The lock does not prevent addition or deletion of children, but
4000 * it prevents a new child from being initialized based on this
4001 * parent in css_online(), so it's enough to decide whether
4002 * hierarchically inherited attributes can still be changed or not.
4004 lockdep_assert_held(&memcg_create_mutex
);
4007 ret
= css_next_child(NULL
, &memcg
->css
);
4013 * Reclaims as many pages from the given memcg as possible and moves
4014 * the rest to the parent.
4016 * Caller is responsible for holding css reference for memcg.
4018 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4020 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4022 /* we call try-to-free pages for make this cgroup empty */
4023 lru_add_drain_all();
4024 /* try to free all pages in this cgroup */
4025 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4028 if (signal_pending(current
))
4031 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4035 /* maybe some writeback is necessary */
4036 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4044 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
4045 char *buf
, size_t nbytes
,
4048 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4050 if (mem_cgroup_is_root(memcg
))
4052 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
4055 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4058 return mem_cgroup_from_css(css
)->use_hierarchy
;
4061 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4062 struct cftype
*cft
, u64 val
)
4065 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4066 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
4068 mutex_lock(&memcg_create_mutex
);
4070 if (memcg
->use_hierarchy
== val
)
4074 * If parent's use_hierarchy is set, we can't make any modifications
4075 * in the child subtrees. If it is unset, then the change can
4076 * occur, provided the current cgroup has no children.
4078 * For the root cgroup, parent_mem is NULL, we allow value to be
4079 * set if there are no children.
4081 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4082 (val
== 1 || val
== 0)) {
4083 if (!memcg_has_children(memcg
))
4084 memcg
->use_hierarchy
= val
;
4091 mutex_unlock(&memcg_create_mutex
);
4096 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
4099 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4100 enum res_type type
= MEMFILE_TYPE(cft
->private);
4101 int name
= MEMFILE_ATTR(cft
->private);
4105 return res_counter_read_u64(&memcg
->res
, name
);
4107 return res_counter_read_u64(&memcg
->memsw
, name
);
4109 return res_counter_read_u64(&memcg
->kmem
, name
);
4116 #ifdef CONFIG_MEMCG_KMEM
4117 /* should be called with activate_kmem_mutex held */
4118 static int __memcg_activate_kmem(struct mem_cgroup
*memcg
,
4119 unsigned long long limit
)
4124 if (memcg_kmem_is_active(memcg
))
4128 * We are going to allocate memory for data shared by all memory
4129 * cgroups so let's stop accounting here.
4131 memcg_stop_kmem_account();
4134 * For simplicity, we won't allow this to be disabled. It also can't
4135 * be changed if the cgroup has children already, or if tasks had
4138 * If tasks join before we set the limit, a person looking at
4139 * kmem.usage_in_bytes will have no way to determine when it took
4140 * place, which makes the value quite meaningless.
4142 * After it first became limited, changes in the value of the limit are
4143 * of course permitted.
4145 mutex_lock(&memcg_create_mutex
);
4146 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
4147 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
4149 mutex_unlock(&memcg_create_mutex
);
4153 memcg_id
= ida_simple_get(&kmem_limited_groups
,
4154 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
4161 * Make sure we have enough space for this cgroup in each root cache's
4164 mutex_lock(&memcg_slab_mutex
);
4165 err
= memcg_update_all_caches(memcg_id
+ 1);
4166 mutex_unlock(&memcg_slab_mutex
);
4170 memcg
->kmemcg_id
= memcg_id
;
4171 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
4174 * We couldn't have accounted to this cgroup, because it hasn't got the
4175 * active bit set yet, so this should succeed.
4177 err
= res_counter_set_limit(&memcg
->kmem
, limit
);
4180 static_key_slow_inc(&memcg_kmem_enabled_key
);
4182 * Setting the active bit after enabling static branching will
4183 * guarantee no one starts accounting before all call sites are
4186 memcg_kmem_set_active(memcg
);
4188 memcg_resume_kmem_account();
4192 ida_simple_remove(&kmem_limited_groups
, memcg_id
);
4196 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
4197 unsigned long long limit
)
4201 mutex_lock(&activate_kmem_mutex
);
4202 ret
= __memcg_activate_kmem(memcg
, limit
);
4203 mutex_unlock(&activate_kmem_mutex
);
4207 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4208 unsigned long long val
)
4212 if (!memcg_kmem_is_active(memcg
))
4213 ret
= memcg_activate_kmem(memcg
, val
);
4215 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4219 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4222 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4227 mutex_lock(&activate_kmem_mutex
);
4229 * If the parent cgroup is not kmem-active now, it cannot be activated
4230 * after this point, because it has at least one child already.
4232 if (memcg_kmem_is_active(parent
))
4233 ret
= __memcg_activate_kmem(memcg
, RES_COUNTER_MAX
);
4234 mutex_unlock(&activate_kmem_mutex
);
4238 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
4239 unsigned long long val
)
4243 #endif /* CONFIG_MEMCG_KMEM */
4246 * The user of this function is...
4249 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
4250 char *buf
, size_t nbytes
, loff_t off
)
4252 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4255 unsigned long long val
;
4258 buf
= strstrip(buf
);
4259 type
= MEMFILE_TYPE(of_cft(of
)->private);
4260 name
= MEMFILE_ATTR(of_cft(of
)->private);
4264 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4268 /* This function does all necessary parse...reuse it */
4269 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4273 ret
= mem_cgroup_resize_limit(memcg
, val
);
4274 else if (type
== _MEMSWAP
)
4275 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
4276 else if (type
== _KMEM
)
4277 ret
= memcg_update_kmem_limit(memcg
, val
);
4281 case RES_SOFT_LIMIT
:
4282 ret
= res_counter_memparse_write_strategy(buf
, &val
);
4286 * For memsw, soft limits are hard to implement in terms
4287 * of semantics, for now, we support soft limits for
4288 * control without swap
4291 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
4296 ret
= -EINVAL
; /* should be BUG() ? */
4299 return ret
?: nbytes
;
4302 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
4303 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
4305 unsigned long long min_limit
, min_memsw_limit
, tmp
;
4307 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4308 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4309 if (!memcg
->use_hierarchy
)
4312 while (memcg
->css
.parent
) {
4313 memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
4314 if (!memcg
->use_hierarchy
)
4316 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4317 min_limit
= min(min_limit
, tmp
);
4318 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4319 min_memsw_limit
= min(min_memsw_limit
, tmp
);
4322 *mem_limit
= min_limit
;
4323 *memsw_limit
= min_memsw_limit
;
4326 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
4327 size_t nbytes
, loff_t off
)
4329 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
4333 type
= MEMFILE_TYPE(of_cft(of
)->private);
4334 name
= MEMFILE_ATTR(of_cft(of
)->private);
4339 res_counter_reset_max(&memcg
->res
);
4340 else if (type
== _MEMSWAP
)
4341 res_counter_reset_max(&memcg
->memsw
);
4342 else if (type
== _KMEM
)
4343 res_counter_reset_max(&memcg
->kmem
);
4349 res_counter_reset_failcnt(&memcg
->res
);
4350 else if (type
== _MEMSWAP
)
4351 res_counter_reset_failcnt(&memcg
->memsw
);
4352 else if (type
== _KMEM
)
4353 res_counter_reset_failcnt(&memcg
->kmem
);
4362 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
4365 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
4369 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4370 struct cftype
*cft
, u64 val
)
4372 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4374 if (val
>= (1 << NR_MOVE_TYPE
))
4378 * No kind of locking is needed in here, because ->can_attach() will
4379 * check this value once in the beginning of the process, and then carry
4380 * on with stale data. This means that changes to this value will only
4381 * affect task migrations starting after the change.
4383 memcg
->move_charge_at_immigrate
= val
;
4387 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4388 struct cftype
*cft
, u64 val
)
4395 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
4399 unsigned int lru_mask
;
4402 static const struct numa_stat stats
[] = {
4403 { "total", LRU_ALL
},
4404 { "file", LRU_ALL_FILE
},
4405 { "anon", LRU_ALL_ANON
},
4406 { "unevictable", BIT(LRU_UNEVICTABLE
) },
4408 const struct numa_stat
*stat
;
4411 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4413 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4414 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
4415 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
4416 for_each_node_state(nid
, N_MEMORY
) {
4417 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
4419 seq_printf(m
, " N%d=%lu", nid
, nr
);
4424 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
4425 struct mem_cgroup
*iter
;
4428 for_each_mem_cgroup_tree(iter
, memcg
)
4429 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
4430 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
4431 for_each_node_state(nid
, N_MEMORY
) {
4433 for_each_mem_cgroup_tree(iter
, memcg
)
4434 nr
+= mem_cgroup_node_nr_lru_pages(
4435 iter
, nid
, stat
->lru_mask
);
4436 seq_printf(m
, " N%d=%lu", nid
, nr
);
4443 #endif /* CONFIG_NUMA */
4445 static inline void mem_cgroup_lru_names_not_uptodate(void)
4447 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
4450 static int memcg_stat_show(struct seq_file
*m
, void *v
)
4452 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
4453 struct mem_cgroup
*mi
;
4456 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
4457 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
4459 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
4460 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
4463 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
4464 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
4465 mem_cgroup_read_events(memcg
, i
));
4467 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
4468 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
4469 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
4471 /* Hierarchical information */
4473 unsigned long long limit
, memsw_limit
;
4474 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
4475 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
4476 if (do_swap_account
)
4477 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
4481 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
4484 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
4486 for_each_mem_cgroup_tree(mi
, memcg
)
4487 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
4488 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
4491 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
4492 unsigned long long val
= 0;
4494 for_each_mem_cgroup_tree(mi
, memcg
)
4495 val
+= mem_cgroup_read_events(mi
, i
);
4496 seq_printf(m
, "total_%s %llu\n",
4497 mem_cgroup_events_names
[i
], val
);
4500 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
4501 unsigned long long val
= 0;
4503 for_each_mem_cgroup_tree(mi
, memcg
)
4504 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
4505 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
4508 #ifdef CONFIG_DEBUG_VM
4511 struct mem_cgroup_per_zone
*mz
;
4512 struct zone_reclaim_stat
*rstat
;
4513 unsigned long recent_rotated
[2] = {0, 0};
4514 unsigned long recent_scanned
[2] = {0, 0};
4516 for_each_online_node(nid
)
4517 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4518 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
4519 rstat
= &mz
->lruvec
.reclaim_stat
;
4521 recent_rotated
[0] += rstat
->recent_rotated
[0];
4522 recent_rotated
[1] += rstat
->recent_rotated
[1];
4523 recent_scanned
[0] += rstat
->recent_scanned
[0];
4524 recent_scanned
[1] += rstat
->recent_scanned
[1];
4526 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
4527 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
4528 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
4529 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
4536 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
4539 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4541 return mem_cgroup_swappiness(memcg
);
4544 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
4545 struct cftype
*cft
, u64 val
)
4547 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4553 memcg
->swappiness
= val
;
4555 vm_swappiness
= val
;
4560 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
4562 struct mem_cgroup_threshold_ary
*t
;
4568 t
= rcu_dereference(memcg
->thresholds
.primary
);
4570 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
4576 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4578 usage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4581 * current_threshold points to threshold just below or equal to usage.
4582 * If it's not true, a threshold was crossed after last
4583 * call of __mem_cgroup_threshold().
4585 i
= t
->current_threshold
;
4588 * Iterate backward over array of thresholds starting from
4589 * current_threshold and check if a threshold is crossed.
4590 * If none of thresholds below usage is crossed, we read
4591 * only one element of the array here.
4593 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
4594 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4596 /* i = current_threshold + 1 */
4600 * Iterate forward over array of thresholds starting from
4601 * current_threshold+1 and check if a threshold is crossed.
4602 * If none of thresholds above usage is crossed, we read
4603 * only one element of the array here.
4605 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
4606 eventfd_signal(t
->entries
[i
].eventfd
, 1);
4608 /* Update current_threshold */
4609 t
->current_threshold
= i
- 1;
4614 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
4617 __mem_cgroup_threshold(memcg
, false);
4618 if (do_swap_account
)
4619 __mem_cgroup_threshold(memcg
, true);
4621 memcg
= parent_mem_cgroup(memcg
);
4625 static int compare_thresholds(const void *a
, const void *b
)
4627 const struct mem_cgroup_threshold
*_a
= a
;
4628 const struct mem_cgroup_threshold
*_b
= b
;
4630 if (_a
->threshold
> _b
->threshold
)
4633 if (_a
->threshold
< _b
->threshold
)
4639 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
4641 struct mem_cgroup_eventfd_list
*ev
;
4643 spin_lock(&memcg_oom_lock
);
4645 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
4646 eventfd_signal(ev
->eventfd
, 1);
4648 spin_unlock(&memcg_oom_lock
);
4652 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
4654 struct mem_cgroup
*iter
;
4656 for_each_mem_cgroup_tree(iter
, memcg
)
4657 mem_cgroup_oom_notify_cb(iter
);
4660 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4661 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
4663 struct mem_cgroup_thresholds
*thresholds
;
4664 struct mem_cgroup_threshold_ary
*new;
4665 u64 threshold
, usage
;
4668 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
4672 mutex_lock(&memcg
->thresholds_lock
);
4675 thresholds
= &memcg
->thresholds
;
4676 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4677 } else if (type
== _MEMSWAP
) {
4678 thresholds
= &memcg
->memsw_thresholds
;
4679 usage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4683 /* Check if a threshold crossed before adding a new one */
4684 if (thresholds
->primary
)
4685 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4687 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
4689 /* Allocate memory for new array of thresholds */
4690 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
4698 /* Copy thresholds (if any) to new array */
4699 if (thresholds
->primary
) {
4700 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
4701 sizeof(struct mem_cgroup_threshold
));
4704 /* Add new threshold */
4705 new->entries
[size
- 1].eventfd
= eventfd
;
4706 new->entries
[size
- 1].threshold
= threshold
;
4708 /* Sort thresholds. Registering of new threshold isn't time-critical */
4709 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4710 compare_thresholds
, NULL
);
4712 /* Find current threshold */
4713 new->current_threshold
= -1;
4714 for (i
= 0; i
< size
; i
++) {
4715 if (new->entries
[i
].threshold
<= usage
) {
4717 * new->current_threshold will not be used until
4718 * rcu_assign_pointer(), so it's safe to increment
4721 ++new->current_threshold
;
4726 /* Free old spare buffer and save old primary buffer as spare */
4727 kfree(thresholds
->spare
);
4728 thresholds
->spare
= thresholds
->primary
;
4730 rcu_assign_pointer(thresholds
->primary
, new);
4732 /* To be sure that nobody uses thresholds */
4736 mutex_unlock(&memcg
->thresholds_lock
);
4741 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4742 struct eventfd_ctx
*eventfd
, const char *args
)
4744 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
4747 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
4748 struct eventfd_ctx
*eventfd
, const char *args
)
4750 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
4753 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4754 struct eventfd_ctx
*eventfd
, enum res_type type
)
4756 struct mem_cgroup_thresholds
*thresholds
;
4757 struct mem_cgroup_threshold_ary
*new;
4761 mutex_lock(&memcg
->thresholds_lock
);
4764 thresholds
= &memcg
->thresholds
;
4765 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4766 } else if (type
== _MEMSWAP
) {
4767 thresholds
= &memcg
->memsw_thresholds
;
4768 usage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4772 if (!thresholds
->primary
)
4775 /* Check if a threshold crossed before removing */
4776 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4778 /* Calculate new number of threshold */
4780 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4781 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4785 new = thresholds
->spare
;
4787 /* Set thresholds array to NULL if we don't have thresholds */
4796 /* Copy thresholds and find current threshold */
4797 new->current_threshold
= -1;
4798 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4799 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4802 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4803 if (new->entries
[j
].threshold
<= usage
) {
4805 * new->current_threshold will not be used
4806 * until rcu_assign_pointer(), so it's safe to increment
4809 ++new->current_threshold
;
4815 /* Swap primary and spare array */
4816 thresholds
->spare
= thresholds
->primary
;
4817 /* If all events are unregistered, free the spare array */
4819 kfree(thresholds
->spare
);
4820 thresholds
->spare
= NULL
;
4823 rcu_assign_pointer(thresholds
->primary
, new);
4825 /* To be sure that nobody uses thresholds */
4828 mutex_unlock(&memcg
->thresholds_lock
);
4831 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4832 struct eventfd_ctx
*eventfd
)
4834 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
4837 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
4838 struct eventfd_ctx
*eventfd
)
4840 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
4843 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
4844 struct eventfd_ctx
*eventfd
, const char *args
)
4846 struct mem_cgroup_eventfd_list
*event
;
4848 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4852 spin_lock(&memcg_oom_lock
);
4854 event
->eventfd
= eventfd
;
4855 list_add(&event
->list
, &memcg
->oom_notify
);
4857 /* already in OOM ? */
4858 if (atomic_read(&memcg
->under_oom
))
4859 eventfd_signal(eventfd
, 1);
4860 spin_unlock(&memcg_oom_lock
);
4865 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
4866 struct eventfd_ctx
*eventfd
)
4868 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4870 spin_lock(&memcg_oom_lock
);
4872 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
4873 if (ev
->eventfd
== eventfd
) {
4874 list_del(&ev
->list
);
4879 spin_unlock(&memcg_oom_lock
);
4882 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
4884 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
4886 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
4887 seq_printf(sf
, "under_oom %d\n", (bool)atomic_read(&memcg
->under_oom
));
4891 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
4892 struct cftype
*cft
, u64 val
)
4894 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4896 /* cannot set to root cgroup and only 0 and 1 are allowed */
4897 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
4900 memcg
->oom_kill_disable
= val
;
4902 memcg_oom_recover(memcg
);
4907 #ifdef CONFIG_MEMCG_KMEM
4908 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4912 memcg
->kmemcg_id
= -1;
4913 ret
= memcg_propagate_kmem(memcg
);
4917 return mem_cgroup_sockets_init(memcg
, ss
);
4920 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4922 mem_cgroup_sockets_destroy(memcg
);
4925 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
4927 if (!memcg_kmem_is_active(memcg
))
4931 * kmem charges can outlive the cgroup. In the case of slab
4932 * pages, for instance, a page contain objects from various
4933 * processes. As we prevent from taking a reference for every
4934 * such allocation we have to be careful when doing uncharge
4935 * (see memcg_uncharge_kmem) and here during offlining.
4937 * The idea is that that only the _last_ uncharge which sees
4938 * the dead memcg will drop the last reference. An additional
4939 * reference is taken here before the group is marked dead
4940 * which is then paired with css_put during uncharge resp. here.
4942 * Although this might sound strange as this path is called from
4943 * css_offline() when the referencemight have dropped down to 0 and
4944 * shouldn't be incremented anymore (css_tryget_online() would
4945 * fail) we do not have other options because of the kmem
4946 * allocations lifetime.
4948 css_get(&memcg
->css
);
4950 memcg_kmem_mark_dead(memcg
);
4952 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
4955 if (memcg_kmem_test_and_clear_dead(memcg
))
4956 css_put(&memcg
->css
);
4959 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
4964 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
4968 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
4974 * DO NOT USE IN NEW FILES.
4976 * "cgroup.event_control" implementation.
4978 * This is way over-engineered. It tries to support fully configurable
4979 * events for each user. Such level of flexibility is completely
4980 * unnecessary especially in the light of the planned unified hierarchy.
4982 * Please deprecate this and replace with something simpler if at all
4987 * Unregister event and free resources.
4989 * Gets called from workqueue.
4991 static void memcg_event_remove(struct work_struct
*work
)
4993 struct mem_cgroup_event
*event
=
4994 container_of(work
, struct mem_cgroup_event
, remove
);
4995 struct mem_cgroup
*memcg
= event
->memcg
;
4997 remove_wait_queue(event
->wqh
, &event
->wait
);
4999 event
->unregister_event(memcg
, event
->eventfd
);
5001 /* Notify userspace the event is going away. */
5002 eventfd_signal(event
->eventfd
, 1);
5004 eventfd_ctx_put(event
->eventfd
);
5006 css_put(&memcg
->css
);
5010 * Gets called on POLLHUP on eventfd when user closes it.
5012 * Called with wqh->lock held and interrupts disabled.
5014 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
5015 int sync
, void *key
)
5017 struct mem_cgroup_event
*event
=
5018 container_of(wait
, struct mem_cgroup_event
, wait
);
5019 struct mem_cgroup
*memcg
= event
->memcg
;
5020 unsigned long flags
= (unsigned long)key
;
5022 if (flags
& POLLHUP
) {
5024 * If the event has been detached at cgroup removal, we
5025 * can simply return knowing the other side will cleanup
5028 * We can't race against event freeing since the other
5029 * side will require wqh->lock via remove_wait_queue(),
5032 spin_lock(&memcg
->event_list_lock
);
5033 if (!list_empty(&event
->list
)) {
5034 list_del_init(&event
->list
);
5036 * We are in atomic context, but cgroup_event_remove()
5037 * may sleep, so we have to call it in workqueue.
5039 schedule_work(&event
->remove
);
5041 spin_unlock(&memcg
->event_list_lock
);
5047 static void memcg_event_ptable_queue_proc(struct file
*file
,
5048 wait_queue_head_t
*wqh
, poll_table
*pt
)
5050 struct mem_cgroup_event
*event
=
5051 container_of(pt
, struct mem_cgroup_event
, pt
);
5054 add_wait_queue(wqh
, &event
->wait
);
5058 * DO NOT USE IN NEW FILES.
5060 * Parse input and register new cgroup event handler.
5062 * Input must be in format '<event_fd> <control_fd> <args>'.
5063 * Interpretation of args is defined by control file implementation.
5065 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
5066 char *buf
, size_t nbytes
, loff_t off
)
5068 struct cgroup_subsys_state
*css
= of_css(of
);
5069 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5070 struct mem_cgroup_event
*event
;
5071 struct cgroup_subsys_state
*cfile_css
;
5072 unsigned int efd
, cfd
;
5079 buf
= strstrip(buf
);
5081 efd
= simple_strtoul(buf
, &endp
, 10);
5086 cfd
= simple_strtoul(buf
, &endp
, 10);
5087 if ((*endp
!= ' ') && (*endp
!= '\0'))
5091 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
5095 event
->memcg
= memcg
;
5096 INIT_LIST_HEAD(&event
->list
);
5097 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
5098 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
5099 INIT_WORK(&event
->remove
, memcg_event_remove
);
5107 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
5108 if (IS_ERR(event
->eventfd
)) {
5109 ret
= PTR_ERR(event
->eventfd
);
5116 goto out_put_eventfd
;
5119 /* the process need read permission on control file */
5120 /* AV: shouldn't we check that it's been opened for read instead? */
5121 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
5126 * Determine the event callbacks and set them in @event. This used
5127 * to be done via struct cftype but cgroup core no longer knows
5128 * about these events. The following is crude but the whole thing
5129 * is for compatibility anyway.
5131 * DO NOT ADD NEW FILES.
5133 name
= cfile
.file
->f_dentry
->d_name
.name
;
5135 if (!strcmp(name
, "memory.usage_in_bytes")) {
5136 event
->register_event
= mem_cgroup_usage_register_event
;
5137 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
5138 } else if (!strcmp(name
, "memory.oom_control")) {
5139 event
->register_event
= mem_cgroup_oom_register_event
;
5140 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
5141 } else if (!strcmp(name
, "memory.pressure_level")) {
5142 event
->register_event
= vmpressure_register_event
;
5143 event
->unregister_event
= vmpressure_unregister_event
;
5144 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
5145 event
->register_event
= memsw_cgroup_usage_register_event
;
5146 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
5153 * Verify @cfile should belong to @css. Also, remaining events are
5154 * automatically removed on cgroup destruction but the removal is
5155 * asynchronous, so take an extra ref on @css.
5157 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_dentry
->d_parent
,
5158 &memory_cgrp_subsys
);
5160 if (IS_ERR(cfile_css
))
5162 if (cfile_css
!= css
) {
5167 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
5171 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
5173 spin_lock(&memcg
->event_list_lock
);
5174 list_add(&event
->list
, &memcg
->event_list
);
5175 spin_unlock(&memcg
->event_list_lock
);
5187 eventfd_ctx_put(event
->eventfd
);
5196 static struct cftype mem_cgroup_files
[] = {
5198 .name
= "usage_in_bytes",
5199 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5200 .read_u64
= mem_cgroup_read_u64
,
5203 .name
= "max_usage_in_bytes",
5204 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5205 .write
= mem_cgroup_reset
,
5206 .read_u64
= mem_cgroup_read_u64
,
5209 .name
= "limit_in_bytes",
5210 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5211 .write
= mem_cgroup_write
,
5212 .read_u64
= mem_cgroup_read_u64
,
5215 .name
= "soft_limit_in_bytes",
5216 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5217 .write
= mem_cgroup_write
,
5218 .read_u64
= mem_cgroup_read_u64
,
5222 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5223 .write
= mem_cgroup_reset
,
5224 .read_u64
= mem_cgroup_read_u64
,
5228 .seq_show
= memcg_stat_show
,
5231 .name
= "force_empty",
5232 .write
= mem_cgroup_force_empty_write
,
5235 .name
= "use_hierarchy",
5236 .write_u64
= mem_cgroup_hierarchy_write
,
5237 .read_u64
= mem_cgroup_hierarchy_read
,
5240 .name
= "cgroup.event_control", /* XXX: for compat */
5241 .write
= memcg_write_event_control
,
5242 .flags
= CFTYPE_NO_PREFIX
,
5246 .name
= "swappiness",
5247 .read_u64
= mem_cgroup_swappiness_read
,
5248 .write_u64
= mem_cgroup_swappiness_write
,
5251 .name
= "move_charge_at_immigrate",
5252 .read_u64
= mem_cgroup_move_charge_read
,
5253 .write_u64
= mem_cgroup_move_charge_write
,
5256 .name
= "oom_control",
5257 .seq_show
= mem_cgroup_oom_control_read
,
5258 .write_u64
= mem_cgroup_oom_control_write
,
5259 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5262 .name
= "pressure_level",
5266 .name
= "numa_stat",
5267 .seq_show
= memcg_numa_stat_show
,
5270 #ifdef CONFIG_MEMCG_KMEM
5272 .name
= "kmem.limit_in_bytes",
5273 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5274 .write
= mem_cgroup_write
,
5275 .read_u64
= mem_cgroup_read_u64
,
5278 .name
= "kmem.usage_in_bytes",
5279 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5280 .read_u64
= mem_cgroup_read_u64
,
5283 .name
= "kmem.failcnt",
5284 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5285 .write
= mem_cgroup_reset
,
5286 .read_u64
= mem_cgroup_read_u64
,
5289 .name
= "kmem.max_usage_in_bytes",
5290 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5291 .write
= mem_cgroup_reset
,
5292 .read_u64
= mem_cgroup_read_u64
,
5294 #ifdef CONFIG_SLABINFO
5296 .name
= "kmem.slabinfo",
5297 .seq_show
= mem_cgroup_slabinfo_read
,
5301 { }, /* terminate */
5304 #ifdef CONFIG_MEMCG_SWAP
5305 static struct cftype memsw_cgroup_files
[] = {
5307 .name
= "memsw.usage_in_bytes",
5308 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5309 .read_u64
= mem_cgroup_read_u64
,
5312 .name
= "memsw.max_usage_in_bytes",
5313 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5314 .write
= mem_cgroup_reset
,
5315 .read_u64
= mem_cgroup_read_u64
,
5318 .name
= "memsw.limit_in_bytes",
5319 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5320 .write
= mem_cgroup_write
,
5321 .read_u64
= mem_cgroup_read_u64
,
5324 .name
= "memsw.failcnt",
5325 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5326 .write
= mem_cgroup_reset
,
5327 .read_u64
= mem_cgroup_read_u64
,
5329 { }, /* terminate */
5332 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5334 struct mem_cgroup_per_node
*pn
;
5335 struct mem_cgroup_per_zone
*mz
;
5336 int zone
, tmp
= node
;
5338 * This routine is called against possible nodes.
5339 * But it's BUG to call kmalloc() against offline node.
5341 * TODO: this routine can waste much memory for nodes which will
5342 * never be onlined. It's better to use memory hotplug callback
5345 if (!node_state(node
, N_NORMAL_MEMORY
))
5347 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5351 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5352 mz
= &pn
->zoneinfo
[zone
];
5353 lruvec_init(&mz
->lruvec
);
5354 mz
->usage_in_excess
= 0;
5355 mz
->on_tree
= false;
5358 memcg
->nodeinfo
[node
] = pn
;
5362 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5364 kfree(memcg
->nodeinfo
[node
]);
5367 static struct mem_cgroup
*mem_cgroup_alloc(void)
5369 struct mem_cgroup
*memcg
;
5372 size
= sizeof(struct mem_cgroup
);
5373 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
5375 memcg
= kzalloc(size
, GFP_KERNEL
);
5379 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5382 spin_lock_init(&memcg
->pcp_counter_lock
);
5391 * At destroying mem_cgroup, references from swap_cgroup can remain.
5392 * (scanning all at force_empty is too costly...)
5394 * Instead of clearing all references at force_empty, we remember
5395 * the number of reference from swap_cgroup and free mem_cgroup when
5396 * it goes down to 0.
5398 * Removal of cgroup itself succeeds regardless of refs from swap.
5401 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5405 mem_cgroup_remove_from_trees(memcg
);
5408 free_mem_cgroup_per_zone_info(memcg
, node
);
5410 free_percpu(memcg
->stat
);
5413 * We need to make sure that (at least for now), the jump label
5414 * destruction code runs outside of the cgroup lock. This is because
5415 * get_online_cpus(), which is called from the static_branch update,
5416 * can't be called inside the cgroup_lock. cpusets are the ones
5417 * enforcing this dependency, so if they ever change, we might as well.
5419 * schedule_work() will guarantee this happens. Be careful if you need
5420 * to move this code around, and make sure it is outside
5423 disarm_static_keys(memcg
);
5428 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5430 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
5432 if (!memcg
->res
.parent
)
5434 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
5436 EXPORT_SYMBOL(parent_mem_cgroup
);
5438 static void __init
mem_cgroup_soft_limit_tree_init(void)
5440 struct mem_cgroup_tree_per_node
*rtpn
;
5441 struct mem_cgroup_tree_per_zone
*rtpz
;
5442 int tmp
, node
, zone
;
5444 for_each_node(node
) {
5446 if (!node_state(node
, N_NORMAL_MEMORY
))
5448 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
5451 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5453 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5454 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5455 rtpz
->rb_root
= RB_ROOT
;
5456 spin_lock_init(&rtpz
->lock
);
5461 static struct cgroup_subsys_state
* __ref
5462 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5464 struct mem_cgroup
*memcg
;
5465 long error
= -ENOMEM
;
5468 memcg
= mem_cgroup_alloc();
5470 return ERR_PTR(error
);
5473 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
5477 if (parent_css
== NULL
) {
5478 root_mem_cgroup
= memcg
;
5479 res_counter_init(&memcg
->res
, NULL
);
5480 res_counter_init(&memcg
->memsw
, NULL
);
5481 res_counter_init(&memcg
->kmem
, NULL
);
5484 memcg
->last_scanned_node
= MAX_NUMNODES
;
5485 INIT_LIST_HEAD(&memcg
->oom_notify
);
5486 memcg
->move_charge_at_immigrate
= 0;
5487 mutex_init(&memcg
->thresholds_lock
);
5488 spin_lock_init(&memcg
->move_lock
);
5489 vmpressure_init(&memcg
->vmpressure
);
5490 INIT_LIST_HEAD(&memcg
->event_list
);
5491 spin_lock_init(&memcg
->event_list_lock
);
5496 __mem_cgroup_free(memcg
);
5497 return ERR_PTR(error
);
5501 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5503 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5504 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
5506 if (css
->id
> MEM_CGROUP_ID_MAX
)
5512 mutex_lock(&memcg_create_mutex
);
5514 memcg
->use_hierarchy
= parent
->use_hierarchy
;
5515 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5516 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5518 if (parent
->use_hierarchy
) {
5519 res_counter_init(&memcg
->res
, &parent
->res
);
5520 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
5521 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
5524 * No need to take a reference to the parent because cgroup
5525 * core guarantees its existence.
5528 res_counter_init(&memcg
->res
, &root_mem_cgroup
->res
);
5529 res_counter_init(&memcg
->memsw
, &root_mem_cgroup
->memsw
);
5530 res_counter_init(&memcg
->kmem
, &root_mem_cgroup
->kmem
);
5532 * Deeper hierachy with use_hierarchy == false doesn't make
5533 * much sense so let cgroup subsystem know about this
5534 * unfortunate state in our controller.
5536 if (parent
!= root_mem_cgroup
)
5537 memory_cgrp_subsys
.broken_hierarchy
= true;
5539 mutex_unlock(&memcg_create_mutex
);
5541 return memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
5545 * Announce all parents that a group from their hierarchy is gone.
5547 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
5549 struct mem_cgroup
*parent
= memcg
;
5551 while ((parent
= parent_mem_cgroup(parent
)))
5552 mem_cgroup_iter_invalidate(parent
);
5555 * if the root memcg is not hierarchical we have to check it
5558 if (!root_mem_cgroup
->use_hierarchy
)
5559 mem_cgroup_iter_invalidate(root_mem_cgroup
);
5562 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5564 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5565 struct mem_cgroup_event
*event
, *tmp
;
5566 struct cgroup_subsys_state
*iter
;
5569 * Unregister events and notify userspace.
5570 * Notify userspace about cgroup removing only after rmdir of cgroup
5571 * directory to avoid race between userspace and kernelspace.
5573 spin_lock(&memcg
->event_list_lock
);
5574 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
5575 list_del_init(&event
->list
);
5576 schedule_work(&event
->remove
);
5578 spin_unlock(&memcg
->event_list_lock
);
5580 kmem_cgroup_css_offline(memcg
);
5582 mem_cgroup_invalidate_reclaim_iterators(memcg
);
5585 * This requires that offlining is serialized. Right now that is
5586 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5588 css_for_each_descendant_post(iter
, css
)
5589 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter
));
5591 memcg_unregister_all_caches(memcg
);
5592 vmpressure_cleanup(&memcg
->vmpressure
);
5595 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5597 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5599 * XXX: css_offline() would be where we should reparent all
5600 * memory to prepare the cgroup for destruction. However,
5601 * memcg does not do css_tryget_online() and res_counter charging
5602 * under the same RCU lock region, which means that charging
5603 * could race with offlining. Offlining only happens to
5604 * cgroups with no tasks in them but charges can show up
5605 * without any tasks from the swapin path when the target
5606 * memcg is looked up from the swapout record and not from the
5607 * current task as it usually is. A race like this can leak
5608 * charges and put pages with stale cgroup pointers into
5612 * lookup_swap_cgroup_id()
5614 * mem_cgroup_lookup()
5615 * css_tryget_online()
5617 * disable css_tryget_online()
5620 * reparent_charges()
5621 * res_counter_charge()
5624 * pc->mem_cgroup = dead memcg
5627 * The bulk of the charges are still moved in offline_css() to
5628 * avoid pinning a lot of pages in case a long-term reference
5629 * like a swapout record is deferring the css_free() to long
5630 * after offlining. But this makes sure we catch any charges
5631 * made after offlining:
5633 mem_cgroup_reparent_charges(memcg
);
5635 memcg_destroy_kmem(memcg
);
5636 __mem_cgroup_free(memcg
);
5640 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5641 * @css: the target css
5643 * Reset the states of the mem_cgroup associated with @css. This is
5644 * invoked when the userland requests disabling on the default hierarchy
5645 * but the memcg is pinned through dependency. The memcg should stop
5646 * applying policies and should revert to the vanilla state as it may be
5647 * made visible again.
5649 * The current implementation only resets the essential configurations.
5650 * This needs to be expanded to cover all the visible parts.
5652 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
5654 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5656 mem_cgroup_resize_limit(memcg
, ULLONG_MAX
);
5657 mem_cgroup_resize_memsw_limit(memcg
, ULLONG_MAX
);
5658 memcg_update_kmem_limit(memcg
, ULLONG_MAX
);
5659 res_counter_set_soft_limit(&memcg
->res
, ULLONG_MAX
);
5663 /* Handlers for move charge at task migration. */
5664 static int mem_cgroup_do_precharge(unsigned long count
)
5668 /* Try a single bulk charge without reclaim first */
5669 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
5671 mc
.precharge
+= count
;
5674 if (ret
== -EINTR
) {
5675 cancel_charge(root_mem_cgroup
, count
);
5679 /* Try charges one by one with reclaim */
5681 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
5683 * In case of failure, any residual charges against
5684 * mc.to will be dropped by mem_cgroup_clear_mc()
5685 * later on. However, cancel any charges that are
5686 * bypassed to root right away or they'll be lost.
5689 cancel_charge(root_mem_cgroup
, 1);
5699 * get_mctgt_type - get target type of moving charge
5700 * @vma: the vma the pte to be checked belongs
5701 * @addr: the address corresponding to the pte to be checked
5702 * @ptent: the pte to be checked
5703 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5706 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5707 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5708 * move charge. if @target is not NULL, the page is stored in target->page
5709 * with extra refcnt got(Callers should handle it).
5710 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5711 * target for charge migration. if @target is not NULL, the entry is stored
5714 * Called with pte lock held.
5721 enum mc_target_type
{
5727 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
5728 unsigned long addr
, pte_t ptent
)
5730 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
5732 if (!page
|| !page_mapped(page
))
5734 if (PageAnon(page
)) {
5735 /* we don't move shared anon */
5738 } else if (!move_file())
5739 /* we ignore mapcount for file pages */
5741 if (!get_page_unless_zero(page
))
5748 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5749 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5751 struct page
*page
= NULL
;
5752 swp_entry_t ent
= pte_to_swp_entry(ptent
);
5754 if (!move_anon() || non_swap_entry(ent
))
5757 * Because lookup_swap_cache() updates some statistics counter,
5758 * we call find_get_page() with swapper_space directly.
5760 page
= find_get_page(swap_address_space(ent
), ent
.val
);
5761 if (do_swap_account
)
5762 entry
->val
= ent
.val
;
5767 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
5768 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5774 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
5775 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
5777 struct page
*page
= NULL
;
5778 struct address_space
*mapping
;
5781 if (!vma
->vm_file
) /* anonymous vma */
5786 mapping
= vma
->vm_file
->f_mapping
;
5787 if (pte_none(ptent
))
5788 pgoff
= linear_page_index(vma
, addr
);
5789 else /* pte_file(ptent) is true */
5790 pgoff
= pte_to_pgoff(ptent
);
5792 /* page is moved even if it's not RSS of this task(page-faulted). */
5794 /* shmem/tmpfs may report page out on swap: account for that too. */
5795 if (shmem_mapping(mapping
)) {
5796 page
= find_get_entry(mapping
, pgoff
);
5797 if (radix_tree_exceptional_entry(page
)) {
5798 swp_entry_t swp
= radix_to_swp_entry(page
);
5799 if (do_swap_account
)
5801 page
= find_get_page(swap_address_space(swp
), swp
.val
);
5804 page
= find_get_page(mapping
, pgoff
);
5806 page
= find_get_page(mapping
, pgoff
);
5811 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
5812 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
5814 struct page
*page
= NULL
;
5815 struct page_cgroup
*pc
;
5816 enum mc_target_type ret
= MC_TARGET_NONE
;
5817 swp_entry_t ent
= { .val
= 0 };
5819 if (pte_present(ptent
))
5820 page
= mc_handle_present_pte(vma
, addr
, ptent
);
5821 else if (is_swap_pte(ptent
))
5822 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
5823 else if (pte_none(ptent
) || pte_file(ptent
))
5824 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
5826 if (!page
&& !ent
.val
)
5829 pc
= lookup_page_cgroup(page
);
5831 * Do only loose check w/o serialization.
5832 * mem_cgroup_move_account() checks the pc is valid or
5833 * not under LRU exclusion.
5835 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
5836 ret
= MC_TARGET_PAGE
;
5838 target
->page
= page
;
5840 if (!ret
|| !target
)
5843 /* There is a swap entry and a page doesn't exist or isn't charged */
5844 if (ent
.val
&& !ret
&&
5845 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
5846 ret
= MC_TARGET_SWAP
;
5853 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5855 * We don't consider swapping or file mapped pages because THP does not
5856 * support them for now.
5857 * Caller should make sure that pmd_trans_huge(pmd) is true.
5859 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5860 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5862 struct page
*page
= NULL
;
5863 struct page_cgroup
*pc
;
5864 enum mc_target_type ret
= MC_TARGET_NONE
;
5866 page
= pmd_page(pmd
);
5867 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
5870 pc
= lookup_page_cgroup(page
);
5871 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
5872 ret
= MC_TARGET_PAGE
;
5875 target
->page
= page
;
5881 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
5882 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
5884 return MC_TARGET_NONE
;
5888 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
5889 unsigned long addr
, unsigned long end
,
5890 struct mm_walk
*walk
)
5892 struct vm_area_struct
*vma
= walk
->private;
5896 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
5897 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
5898 mc
.precharge
+= HPAGE_PMD_NR
;
5903 if (pmd_trans_unstable(pmd
))
5905 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
5906 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
5907 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
5908 mc
.precharge
++; /* increment precharge temporarily */
5909 pte_unmap_unlock(pte
- 1, ptl
);
5915 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
5917 unsigned long precharge
;
5918 struct vm_area_struct
*vma
;
5920 down_read(&mm
->mmap_sem
);
5921 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5922 struct mm_walk mem_cgroup_count_precharge_walk
= {
5923 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
5927 if (is_vm_hugetlb_page(vma
))
5929 walk_page_range(vma
->vm_start
, vma
->vm_end
,
5930 &mem_cgroup_count_precharge_walk
);
5932 up_read(&mm
->mmap_sem
);
5934 precharge
= mc
.precharge
;
5940 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
5942 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
5944 VM_BUG_ON(mc
.moving_task
);
5945 mc
.moving_task
= current
;
5946 return mem_cgroup_do_precharge(precharge
);
5949 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5950 static void __mem_cgroup_clear_mc(void)
5952 struct mem_cgroup
*from
= mc
.from
;
5953 struct mem_cgroup
*to
= mc
.to
;
5956 /* we must uncharge all the leftover precharges from mc.to */
5958 cancel_charge(mc
.to
, mc
.precharge
);
5962 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5963 * we must uncharge here.
5965 if (mc
.moved_charge
) {
5966 cancel_charge(mc
.from
, mc
.moved_charge
);
5967 mc
.moved_charge
= 0;
5969 /* we must fixup refcnts and charges */
5970 if (mc
.moved_swap
) {
5971 /* uncharge swap account from the old cgroup */
5972 res_counter_uncharge(&mc
.from
->memsw
,
5973 PAGE_SIZE
* mc
.moved_swap
);
5975 for (i
= 0; i
< mc
.moved_swap
; i
++)
5976 css_put(&mc
.from
->css
);
5979 * we charged both to->res and to->memsw, so we should
5982 res_counter_uncharge(&mc
.to
->res
,
5983 PAGE_SIZE
* mc
.moved_swap
);
5984 /* we've already done css_get(mc.to) */
5987 memcg_oom_recover(from
);
5988 memcg_oom_recover(to
);
5989 wake_up_all(&mc
.waitq
);
5992 static void mem_cgroup_clear_mc(void)
5994 struct mem_cgroup
*from
= mc
.from
;
5997 * we must clear moving_task before waking up waiters at the end of
6000 mc
.moving_task
= NULL
;
6001 __mem_cgroup_clear_mc();
6002 spin_lock(&mc
.lock
);
6005 spin_unlock(&mc
.lock
);
6006 mem_cgroup_end_move(from
);
6009 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6010 struct cgroup_taskset
*tset
)
6012 struct task_struct
*p
= cgroup_taskset_first(tset
);
6014 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6015 unsigned long move_charge_at_immigrate
;
6018 * We are now commited to this value whatever it is. Changes in this
6019 * tunable will only affect upcoming migrations, not the current one.
6020 * So we need to save it, and keep it going.
6022 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6023 if (move_charge_at_immigrate
) {
6024 struct mm_struct
*mm
;
6025 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6027 VM_BUG_ON(from
== memcg
);
6029 mm
= get_task_mm(p
);
6032 /* We move charges only when we move a owner of the mm */
6033 if (mm
->owner
== p
) {
6036 VM_BUG_ON(mc
.precharge
);
6037 VM_BUG_ON(mc
.moved_charge
);
6038 VM_BUG_ON(mc
.moved_swap
);
6039 mem_cgroup_start_move(from
);
6040 spin_lock(&mc
.lock
);
6043 mc
.immigrate_flags
= move_charge_at_immigrate
;
6044 spin_unlock(&mc
.lock
);
6045 /* We set mc.moving_task later */
6047 ret
= mem_cgroup_precharge_mc(mm
);
6049 mem_cgroup_clear_mc();
6056 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6057 struct cgroup_taskset
*tset
)
6059 mem_cgroup_clear_mc();
6062 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6063 unsigned long addr
, unsigned long end
,
6064 struct mm_walk
*walk
)
6067 struct vm_area_struct
*vma
= walk
->private;
6070 enum mc_target_type target_type
;
6071 union mc_target target
;
6073 struct page_cgroup
*pc
;
6076 * We don't take compound_lock() here but no race with splitting thp
6078 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6079 * under splitting, which means there's no concurrent thp split,
6080 * - if another thread runs into split_huge_page() just after we
6081 * entered this if-block, the thread must wait for page table lock
6082 * to be unlocked in __split_huge_page_splitting(), where the main
6083 * part of thp split is not executed yet.
6085 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
6086 if (mc
.precharge
< HPAGE_PMD_NR
) {
6090 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6091 if (target_type
== MC_TARGET_PAGE
) {
6093 if (!isolate_lru_page(page
)) {
6094 pc
= lookup_page_cgroup(page
);
6095 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6096 pc
, mc
.from
, mc
.to
)) {
6097 mc
.precharge
-= HPAGE_PMD_NR
;
6098 mc
.moved_charge
+= HPAGE_PMD_NR
;
6100 putback_lru_page(page
);
6108 if (pmd_trans_unstable(pmd
))
6111 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6112 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6113 pte_t ptent
= *(pte
++);
6119 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6120 case MC_TARGET_PAGE
:
6122 if (isolate_lru_page(page
))
6124 pc
= lookup_page_cgroup(page
);
6125 if (!mem_cgroup_move_account(page
, 1, pc
,
6128 /* we uncharge from mc.from later. */
6131 putback_lru_page(page
);
6132 put
: /* get_mctgt_type() gets the page */
6135 case MC_TARGET_SWAP
:
6137 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6139 /* we fixup refcnts and charges later. */
6147 pte_unmap_unlock(pte
- 1, ptl
);
6152 * We have consumed all precharges we got in can_attach().
6153 * We try charge one by one, but don't do any additional
6154 * charges to mc.to if we have failed in charge once in attach()
6157 ret
= mem_cgroup_do_precharge(1);
6165 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6167 struct vm_area_struct
*vma
;
6169 lru_add_drain_all();
6171 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6173 * Someone who are holding the mmap_sem might be waiting in
6174 * waitq. So we cancel all extra charges, wake up all waiters,
6175 * and retry. Because we cancel precharges, we might not be able
6176 * to move enough charges, but moving charge is a best-effort
6177 * feature anyway, so it wouldn't be a big problem.
6179 __mem_cgroup_clear_mc();
6183 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6185 struct mm_walk mem_cgroup_move_charge_walk
= {
6186 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6190 if (is_vm_hugetlb_page(vma
))
6192 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6193 &mem_cgroup_move_charge_walk
);
6196 * means we have consumed all precharges and failed in
6197 * doing additional charge. Just abandon here.
6201 up_read(&mm
->mmap_sem
);
6204 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6205 struct cgroup_taskset
*tset
)
6207 struct task_struct
*p
= cgroup_taskset_first(tset
);
6208 struct mm_struct
*mm
= get_task_mm(p
);
6212 mem_cgroup_move_charge(mm
);
6216 mem_cgroup_clear_mc();
6218 #else /* !CONFIG_MMU */
6219 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6220 struct cgroup_taskset
*tset
)
6224 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6225 struct cgroup_taskset
*tset
)
6228 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6229 struct cgroup_taskset
*tset
)
6235 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6236 * to verify whether we're attached to the default hierarchy on each mount
6239 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6242 * use_hierarchy is forced on the default hierarchy. cgroup core
6243 * guarantees that @root doesn't have any children, so turning it
6244 * on for the root memcg is enough.
6246 if (cgroup_on_dfl(root_css
->cgroup
))
6247 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6250 struct cgroup_subsys memory_cgrp_subsys
= {
6251 .css_alloc
= mem_cgroup_css_alloc
,
6252 .css_online
= mem_cgroup_css_online
,
6253 .css_offline
= mem_cgroup_css_offline
,
6254 .css_free
= mem_cgroup_css_free
,
6255 .css_reset
= mem_cgroup_css_reset
,
6256 .can_attach
= mem_cgroup_can_attach
,
6257 .cancel_attach
= mem_cgroup_cancel_attach
,
6258 .attach
= mem_cgroup_move_task
,
6259 .bind
= mem_cgroup_bind
,
6260 .legacy_cftypes
= mem_cgroup_files
,
6264 #ifdef CONFIG_MEMCG_SWAP
6265 static int __init
enable_swap_account(char *s
)
6267 if (!strcmp(s
, "1"))
6268 really_do_swap_account
= 1;
6269 else if (!strcmp(s
, "0"))
6270 really_do_swap_account
= 0;
6273 __setup("swapaccount=", enable_swap_account
);
6275 static void __init
memsw_file_init(void)
6277 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
6278 memsw_cgroup_files
));
6281 static void __init
enable_swap_cgroup(void)
6283 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6284 do_swap_account
= 1;
6290 static void __init
enable_swap_cgroup(void)
6295 #ifdef CONFIG_MEMCG_SWAP
6297 * mem_cgroup_swapout - transfer a memsw charge to swap
6298 * @page: page whose memsw charge to transfer
6299 * @entry: swap entry to move the charge to
6301 * Transfer the memsw charge of @page to @entry.
6303 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
6305 struct page_cgroup
*pc
;
6306 unsigned short oldid
;
6308 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6309 VM_BUG_ON_PAGE(page_count(page
), page
);
6311 if (!do_swap_account
)
6314 pc
= lookup_page_cgroup(page
);
6316 /* Readahead page, never charged */
6317 if (!PageCgroupUsed(pc
))
6320 VM_BUG_ON_PAGE(!(pc
->flags
& PCG_MEMSW
), page
);
6322 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(pc
->mem_cgroup
));
6323 VM_BUG_ON_PAGE(oldid
, page
);
6325 pc
->flags
&= ~PCG_MEMSW
;
6326 css_get(&pc
->mem_cgroup
->css
);
6327 mem_cgroup_swap_statistics(pc
->mem_cgroup
, true);
6331 * mem_cgroup_uncharge_swap - uncharge a swap entry
6332 * @entry: swap entry to uncharge
6334 * Drop the memsw charge associated with @entry.
6336 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
6338 struct mem_cgroup
*memcg
;
6341 if (!do_swap_account
)
6344 id
= swap_cgroup_record(entry
, 0);
6346 memcg
= mem_cgroup_lookup(id
);
6348 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
6349 mem_cgroup_swap_statistics(memcg
, false);
6350 css_put(&memcg
->css
);
6357 * mem_cgroup_try_charge - try charging a page
6358 * @page: page to charge
6359 * @mm: mm context of the victim
6360 * @gfp_mask: reclaim mode
6361 * @memcgp: charged memcg return
6363 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6364 * pages according to @gfp_mask if necessary.
6366 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6367 * Otherwise, an error code is returned.
6369 * After page->mapping has been set up, the caller must finalize the
6370 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6371 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6373 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
6374 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
6376 struct mem_cgroup
*memcg
= NULL
;
6377 unsigned int nr_pages
= 1;
6380 if (mem_cgroup_disabled())
6383 if (PageSwapCache(page
)) {
6384 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
6386 * Every swap fault against a single page tries to charge the
6387 * page, bail as early as possible. shmem_unuse() encounters
6388 * already charged pages, too. The USED bit is protected by
6389 * the page lock, which serializes swap cache removal, which
6390 * in turn serializes uncharging.
6392 if (PageCgroupUsed(pc
))
6396 if (PageTransHuge(page
)) {
6397 nr_pages
<<= compound_order(page
);
6398 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6401 if (do_swap_account
&& PageSwapCache(page
))
6402 memcg
= try_get_mem_cgroup_from_page(page
);
6404 memcg
= get_mem_cgroup_from_mm(mm
);
6406 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
6408 css_put(&memcg
->css
);
6410 if (ret
== -EINTR
) {
6411 memcg
= root_mem_cgroup
;
6420 * mem_cgroup_commit_charge - commit a page charge
6421 * @page: page to charge
6422 * @memcg: memcg to charge the page to
6423 * @lrucare: page might be on LRU already
6425 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6426 * after page->mapping has been set up. This must happen atomically
6427 * as part of the page instantiation, i.e. under the page table lock
6428 * for anonymous pages, under the page lock for page and swap cache.
6430 * In addition, the page must not be on the LRU during the commit, to
6431 * prevent racing with task migration. If it might be, use @lrucare.
6433 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6435 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
6438 unsigned int nr_pages
= 1;
6440 VM_BUG_ON_PAGE(!page
->mapping
, page
);
6441 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
6443 if (mem_cgroup_disabled())
6446 * Swap faults will attempt to charge the same page multiple
6447 * times. But reuse_swap_page() might have removed the page
6448 * from swapcache already, so we can't check PageSwapCache().
6453 commit_charge(page
, memcg
, lrucare
);
6455 if (PageTransHuge(page
)) {
6456 nr_pages
<<= compound_order(page
);
6457 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6460 local_irq_disable();
6461 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
6462 memcg_check_events(memcg
, page
);
6465 if (do_swap_account
&& PageSwapCache(page
)) {
6466 swp_entry_t entry
= { .val
= page_private(page
) };
6468 * The swap entry might not get freed for a long time,
6469 * let's not wait for it. The page already received a
6470 * memory+swap charge, drop the swap entry duplicate.
6472 mem_cgroup_uncharge_swap(entry
);
6477 * mem_cgroup_cancel_charge - cancel a page charge
6478 * @page: page to charge
6479 * @memcg: memcg to charge the page to
6481 * Cancel a charge transaction started by mem_cgroup_try_charge().
6483 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
6485 unsigned int nr_pages
= 1;
6487 if (mem_cgroup_disabled())
6490 * Swap faults will attempt to charge the same page multiple
6491 * times. But reuse_swap_page() might have removed the page
6492 * from swapcache already, so we can't check PageSwapCache().
6497 if (PageTransHuge(page
)) {
6498 nr_pages
<<= compound_order(page
);
6499 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6502 cancel_charge(memcg
, nr_pages
);
6505 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
6506 unsigned long nr_mem
, unsigned long nr_memsw
,
6507 unsigned long nr_anon
, unsigned long nr_file
,
6508 unsigned long nr_huge
, struct page
*dummy_page
)
6510 unsigned long flags
;
6513 res_counter_uncharge(&memcg
->res
, nr_mem
* PAGE_SIZE
);
6515 res_counter_uncharge(&memcg
->memsw
, nr_memsw
* PAGE_SIZE
);
6517 memcg_oom_recover(memcg
);
6519 local_irq_save(flags
);
6520 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
6521 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
6522 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
6523 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
6524 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_anon
+ nr_file
);
6525 memcg_check_events(memcg
, dummy_page
);
6526 local_irq_restore(flags
);
6529 static void uncharge_list(struct list_head
*page_list
)
6531 struct mem_cgroup
*memcg
= NULL
;
6532 unsigned long nr_memsw
= 0;
6533 unsigned long nr_anon
= 0;
6534 unsigned long nr_file
= 0;
6535 unsigned long nr_huge
= 0;
6536 unsigned long pgpgout
= 0;
6537 unsigned long nr_mem
= 0;
6538 struct list_head
*next
;
6541 next
= page_list
->next
;
6543 unsigned int nr_pages
= 1;
6544 struct page_cgroup
*pc
;
6546 page
= list_entry(next
, struct page
, lru
);
6547 next
= page
->lru
.next
;
6549 VM_BUG_ON_PAGE(PageLRU(page
), page
);
6550 VM_BUG_ON_PAGE(page_count(page
), page
);
6552 pc
= lookup_page_cgroup(page
);
6553 if (!PageCgroupUsed(pc
))
6557 * Nobody should be changing or seriously looking at
6558 * pc->mem_cgroup and pc->flags at this point, we have
6559 * fully exclusive access to the page.
6562 if (memcg
!= pc
->mem_cgroup
) {
6564 uncharge_batch(memcg
, pgpgout
, nr_mem
, nr_memsw
,
6565 nr_anon
, nr_file
, nr_huge
, page
);
6566 pgpgout
= nr_mem
= nr_memsw
= 0;
6567 nr_anon
= nr_file
= nr_huge
= 0;
6569 memcg
= pc
->mem_cgroup
;
6572 if (PageTransHuge(page
)) {
6573 nr_pages
<<= compound_order(page
);
6574 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
6575 nr_huge
+= nr_pages
;
6579 nr_anon
+= nr_pages
;
6581 nr_file
+= nr_pages
;
6583 if (pc
->flags
& PCG_MEM
)
6585 if (pc
->flags
& PCG_MEMSW
)
6586 nr_memsw
+= nr_pages
;
6590 } while (next
!= page_list
);
6593 uncharge_batch(memcg
, pgpgout
, nr_mem
, nr_memsw
,
6594 nr_anon
, nr_file
, nr_huge
, page
);
6598 * mem_cgroup_uncharge - uncharge a page
6599 * @page: page to uncharge
6601 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6602 * mem_cgroup_commit_charge().
6604 void mem_cgroup_uncharge(struct page
*page
)
6606 struct page_cgroup
*pc
;
6608 if (mem_cgroup_disabled())
6611 /* Don't touch page->lru of any random page, pre-check: */
6612 pc
= lookup_page_cgroup(page
);
6613 if (!PageCgroupUsed(pc
))
6616 INIT_LIST_HEAD(&page
->lru
);
6617 uncharge_list(&page
->lru
);
6621 * mem_cgroup_uncharge_list - uncharge a list of page
6622 * @page_list: list of pages to uncharge
6624 * Uncharge a list of pages previously charged with
6625 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6627 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
6629 if (mem_cgroup_disabled())
6632 if (!list_empty(page_list
))
6633 uncharge_list(page_list
);
6637 * mem_cgroup_migrate - migrate a charge to another page
6638 * @oldpage: currently charged page
6639 * @newpage: page to transfer the charge to
6640 * @lrucare: both pages might be on the LRU already
6642 * Migrate the charge from @oldpage to @newpage.
6644 * Both pages must be locked, @newpage->mapping must be set up.
6646 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
6649 struct page_cgroup
*pc
;
6652 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
6653 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
6654 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
6655 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
6656 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
6657 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
6660 if (mem_cgroup_disabled())
6663 /* Page cache replacement: new page already charged? */
6664 pc
= lookup_page_cgroup(newpage
);
6665 if (PageCgroupUsed(pc
))
6668 /* Re-entrant migration: old page already uncharged? */
6669 pc
= lookup_page_cgroup(oldpage
);
6670 if (!PageCgroupUsed(pc
))
6673 VM_BUG_ON_PAGE(!(pc
->flags
& PCG_MEM
), oldpage
);
6674 VM_BUG_ON_PAGE(do_swap_account
&& !(pc
->flags
& PCG_MEMSW
), oldpage
);
6677 lock_page_lru(oldpage
, &isolated
);
6682 unlock_page_lru(oldpage
, isolated
);
6684 commit_charge(newpage
, pc
->mem_cgroup
, lrucare
);
6688 * subsys_initcall() for memory controller.
6690 * Some parts like hotcpu_notifier() have to be initialized from this context
6691 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6692 * everything that doesn't depend on a specific mem_cgroup structure should
6693 * be initialized from here.
6695 static int __init
mem_cgroup_init(void)
6697 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6698 enable_swap_cgroup();
6699 mem_cgroup_soft_limit_tree_init();
6703 subsys_initcall(mem_cgroup_init
);