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
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly
;
76 EXPORT_SYMBOL(memory_cgrp_subsys
);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
80 struct cgroup_subsys_state
*mem_cgroup_root_css __read_mostly
;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 int do_swap_account __read_mostly
;
86 #define do_swap_account 0
89 static const char * const mem_cgroup_stat_names
[] = {
99 static const char * const mem_cgroup_events_names
[] = {
106 static const char * const mem_cgroup_lru_names
[] = {
114 #define THRESHOLDS_EVENTS_TARGET 128
115 #define SOFTLIMIT_EVENTS_TARGET 1024
116 #define NUMAINFO_EVENTS_TARGET 1024
119 * Cgroups above their limits are maintained in a RB-Tree, independent of
120 * their hierarchy representation
123 struct mem_cgroup_tree_per_zone
{
124 struct rb_root rb_root
;
128 struct mem_cgroup_tree_per_node
{
129 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
132 struct mem_cgroup_tree
{
133 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
136 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
139 struct mem_cgroup_eventfd_list
{
140 struct list_head list
;
141 struct eventfd_ctx
*eventfd
;
145 * cgroup_event represents events which userspace want to receive.
147 struct mem_cgroup_event
{
149 * memcg which the event belongs to.
151 struct mem_cgroup
*memcg
;
153 * eventfd to signal userspace about the event.
155 struct eventfd_ctx
*eventfd
;
157 * Each of these stored in a list by the cgroup.
159 struct list_head list
;
161 * register_event() callback will be used to add new userspace
162 * waiter for changes related to this event. Use eventfd_signal()
163 * on eventfd to send notification to userspace.
165 int (*register_event
)(struct mem_cgroup
*memcg
,
166 struct eventfd_ctx
*eventfd
, const char *args
);
168 * unregister_event() callback will be called when userspace closes
169 * the eventfd or on cgroup removing. This callback must be set,
170 * if you want provide notification functionality.
172 void (*unregister_event
)(struct mem_cgroup
*memcg
,
173 struct eventfd_ctx
*eventfd
);
175 * All fields below needed to unregister event when
176 * userspace closes eventfd.
179 wait_queue_head_t
*wqh
;
181 struct work_struct remove
;
184 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
185 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
187 /* Stuffs for move charges at task migration. */
189 * Types of charges to be moved.
191 #define MOVE_ANON 0x1U
192 #define MOVE_FILE 0x2U
193 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
195 /* "mc" and its members are protected by cgroup_mutex */
196 static struct move_charge_struct
{
197 spinlock_t lock
; /* for from, to */
198 struct mem_cgroup
*from
;
199 struct mem_cgroup
*to
;
201 unsigned long precharge
;
202 unsigned long moved_charge
;
203 unsigned long moved_swap
;
204 struct task_struct
*moving_task
; /* a task moving charges */
205 wait_queue_head_t waitq
; /* a waitq for other context */
207 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
208 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
212 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
213 * limit reclaim to prevent infinite loops, if they ever occur.
215 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
216 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
219 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
220 MEM_CGROUP_CHARGE_TYPE_ANON
,
221 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
222 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
226 /* for encoding cft->private value on file */
234 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
235 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
236 #define MEMFILE_ATTR(val) ((val) & 0xffff)
237 /* Used for OOM nofiier */
238 #define OOM_CONTROL (0)
241 * The memcg_create_mutex will be held whenever a new cgroup is created.
242 * As a consequence, any change that needs to protect against new child cgroups
243 * appearing has to hold it as well.
245 static DEFINE_MUTEX(memcg_create_mutex
);
247 /* Some nice accessors for the vmpressure. */
248 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
251 memcg
= root_mem_cgroup
;
252 return &memcg
->vmpressure
;
255 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
257 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
260 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
262 return (memcg
== root_mem_cgroup
);
266 * We restrict the id in the range of [1, 65535], so it can fit into
269 #define MEM_CGROUP_ID_MAX USHRT_MAX
271 static inline unsigned short mem_cgroup_id(struct mem_cgroup
*memcg
)
273 return memcg
->css
.id
;
277 * A helper function to get mem_cgroup from ID. must be called under
278 * rcu_read_lock(). The caller is responsible for calling
279 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
280 * refcnt from swap can be called against removed memcg.)
282 static inline struct mem_cgroup
*mem_cgroup_from_id(unsigned short id
)
284 struct cgroup_subsys_state
*css
;
286 css
= css_from_id(id
, &memory_cgrp_subsys
);
287 return mem_cgroup_from_css(css
);
290 /* Writing them here to avoid exposing memcg's inner layout */
291 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
293 void sock_update_memcg(struct sock
*sk
)
295 if (mem_cgroup_sockets_enabled
) {
296 struct mem_cgroup
*memcg
;
297 struct cg_proto
*cg_proto
;
299 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
301 /* Socket cloning can throw us here with sk_cgrp already
302 * filled. It won't however, necessarily happen from
303 * process context. So the test for root memcg given
304 * the current task's memcg won't help us in this case.
306 * Respecting the original socket's memcg is a better
307 * decision in this case.
310 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
311 css_get(&sk
->sk_cgrp
->memcg
->css
);
316 memcg
= mem_cgroup_from_task(current
);
317 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
318 if (cg_proto
&& test_bit(MEMCG_SOCK_ACTIVE
, &cg_proto
->flags
) &&
319 css_tryget_online(&memcg
->css
)) {
320 sk
->sk_cgrp
= cg_proto
;
325 EXPORT_SYMBOL(sock_update_memcg
);
327 void sock_release_memcg(struct sock
*sk
)
329 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
330 struct mem_cgroup
*memcg
;
331 WARN_ON(!sk
->sk_cgrp
->memcg
);
332 memcg
= sk
->sk_cgrp
->memcg
;
333 css_put(&sk
->sk_cgrp
->memcg
->css
);
337 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
339 if (!memcg
|| mem_cgroup_is_root(memcg
))
342 return &memcg
->tcp_mem
;
344 EXPORT_SYMBOL(tcp_proto_cgroup
);
348 #ifdef CONFIG_MEMCG_KMEM
350 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
351 * The main reason for not using cgroup id for this:
352 * this works better in sparse environments, where we have a lot of memcgs,
353 * but only a few kmem-limited. Or also, if we have, for instance, 200
354 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
355 * 200 entry array for that.
357 * The current size of the caches array is stored in memcg_nr_cache_ids. It
358 * will double each time we have to increase it.
360 static DEFINE_IDA(memcg_cache_ida
);
361 int memcg_nr_cache_ids
;
363 /* Protects memcg_nr_cache_ids */
364 static DECLARE_RWSEM(memcg_cache_ids_sem
);
366 void memcg_get_cache_ids(void)
368 down_read(&memcg_cache_ids_sem
);
371 void memcg_put_cache_ids(void)
373 up_read(&memcg_cache_ids_sem
);
377 * MIN_SIZE is different than 1, because we would like to avoid going through
378 * the alloc/free process all the time. In a small machine, 4 kmem-limited
379 * cgroups is a reasonable guess. In the future, it could be a parameter or
380 * tunable, but that is strictly not necessary.
382 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
383 * this constant directly from cgroup, but it is understandable that this is
384 * better kept as an internal representation in cgroup.c. In any case, the
385 * cgrp_id space is not getting any smaller, and we don't have to necessarily
386 * increase ours as well if it increases.
388 #define MEMCG_CACHES_MIN_SIZE 4
389 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
392 * A lot of the calls to the cache allocation functions are expected to be
393 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
394 * conditional to this static branch, we'll have to allow modules that does
395 * kmem_cache_alloc and the such to see this symbol as well
397 struct static_key memcg_kmem_enabled_key
;
398 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
400 #endif /* CONFIG_MEMCG_KMEM */
402 static struct mem_cgroup_per_zone
*
403 mem_cgroup_zone_zoneinfo(struct mem_cgroup
*memcg
, struct zone
*zone
)
405 int nid
= zone_to_nid(zone
);
406 int zid
= zone_idx(zone
);
408 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
412 * mem_cgroup_css_from_page - css of the memcg associated with a page
413 * @page: page of interest
415 * If memcg is bound to the default hierarchy, css of the memcg associated
416 * with @page is returned. The returned css remains associated with @page
417 * until it is released.
419 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
422 * XXX: The above description of behavior on the default hierarchy isn't
423 * strictly true yet as replace_page_cache_page() can modify the
424 * association before @page is released even on the default hierarchy;
425 * however, the current and planned usages don't mix the the two functions
426 * and replace_page_cache_page() will soon be updated to make the invariant
429 struct cgroup_subsys_state
*mem_cgroup_css_from_page(struct page
*page
)
431 struct mem_cgroup
*memcg
;
435 memcg
= page
->mem_cgroup
;
437 if (!memcg
|| !cgroup_on_dfl(memcg
->css
.cgroup
))
438 memcg
= root_mem_cgroup
;
445 * page_cgroup_ino - return inode number of the memcg a page is charged to
448 * Look up the closest online ancestor of the memory cgroup @page is charged to
449 * and return its inode number or 0 if @page is not charged to any cgroup. It
450 * is safe to call this function without holding a reference to @page.
452 * Note, this function is inherently racy, because there is nothing to prevent
453 * the cgroup inode from getting torn down and potentially reallocated a moment
454 * after page_cgroup_ino() returns, so it only should be used by callers that
455 * do not care (such as procfs interfaces).
457 ino_t
page_cgroup_ino(struct page
*page
)
459 struct mem_cgroup
*memcg
;
460 unsigned long ino
= 0;
463 memcg
= READ_ONCE(page
->mem_cgroup
);
464 while (memcg
&& !(memcg
->css
.flags
& CSS_ONLINE
))
465 memcg
= parent_mem_cgroup(memcg
);
467 ino
= cgroup_ino(memcg
->css
.cgroup
);
472 static struct mem_cgroup_per_zone
*
473 mem_cgroup_page_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
475 int nid
= page_to_nid(page
);
476 int zid
= page_zonenum(page
);
478 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
481 static struct mem_cgroup_tree_per_zone
*
482 soft_limit_tree_node_zone(int nid
, int zid
)
484 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
487 static struct mem_cgroup_tree_per_zone
*
488 soft_limit_tree_from_page(struct page
*page
)
490 int nid
= page_to_nid(page
);
491 int zid
= page_zonenum(page
);
493 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
496 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone
*mz
,
497 struct mem_cgroup_tree_per_zone
*mctz
,
498 unsigned long new_usage_in_excess
)
500 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
501 struct rb_node
*parent
= NULL
;
502 struct mem_cgroup_per_zone
*mz_node
;
507 mz
->usage_in_excess
= new_usage_in_excess
;
508 if (!mz
->usage_in_excess
)
512 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
514 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
517 * We can't avoid mem cgroups that are over their soft
518 * limit by the same amount
520 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
523 rb_link_node(&mz
->tree_node
, parent
, p
);
524 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
528 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
529 struct mem_cgroup_tree_per_zone
*mctz
)
533 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
537 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone
*mz
,
538 struct mem_cgroup_tree_per_zone
*mctz
)
542 spin_lock_irqsave(&mctz
->lock
, flags
);
543 __mem_cgroup_remove_exceeded(mz
, mctz
);
544 spin_unlock_irqrestore(&mctz
->lock
, flags
);
547 static unsigned long soft_limit_excess(struct mem_cgroup
*memcg
)
549 unsigned long nr_pages
= page_counter_read(&memcg
->memory
);
550 unsigned long soft_limit
= READ_ONCE(memcg
->soft_limit
);
551 unsigned long excess
= 0;
553 if (nr_pages
> soft_limit
)
554 excess
= nr_pages
- soft_limit
;
559 static void mem_cgroup_update_tree(struct mem_cgroup
*memcg
, struct page
*page
)
561 unsigned long excess
;
562 struct mem_cgroup_per_zone
*mz
;
563 struct mem_cgroup_tree_per_zone
*mctz
;
565 mctz
= soft_limit_tree_from_page(page
);
567 * Necessary to update all ancestors when hierarchy is used.
568 * because their event counter is not touched.
570 for (; memcg
; memcg
= parent_mem_cgroup(memcg
)) {
571 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
572 excess
= soft_limit_excess(memcg
);
574 * We have to update the tree if mz is on RB-tree or
575 * mem is over its softlimit.
577 if (excess
|| mz
->on_tree
) {
580 spin_lock_irqsave(&mctz
->lock
, flags
);
581 /* if on-tree, remove it */
583 __mem_cgroup_remove_exceeded(mz
, mctz
);
585 * Insert again. mz->usage_in_excess will be updated.
586 * If excess is 0, no tree ops.
588 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
589 spin_unlock_irqrestore(&mctz
->lock
, flags
);
594 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*memcg
)
596 struct mem_cgroup_tree_per_zone
*mctz
;
597 struct mem_cgroup_per_zone
*mz
;
601 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
602 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
603 mctz
= soft_limit_tree_node_zone(nid
, zid
);
604 mem_cgroup_remove_exceeded(mz
, mctz
);
609 static struct mem_cgroup_per_zone
*
610 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
612 struct rb_node
*rightmost
= NULL
;
613 struct mem_cgroup_per_zone
*mz
;
617 rightmost
= rb_last(&mctz
->rb_root
);
619 goto done
; /* Nothing to reclaim from */
621 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
623 * Remove the node now but someone else can add it back,
624 * we will to add it back at the end of reclaim to its correct
625 * position in the tree.
627 __mem_cgroup_remove_exceeded(mz
, mctz
);
628 if (!soft_limit_excess(mz
->memcg
) ||
629 !css_tryget_online(&mz
->memcg
->css
))
635 static struct mem_cgroup_per_zone
*
636 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
638 struct mem_cgroup_per_zone
*mz
;
640 spin_lock_irq(&mctz
->lock
);
641 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
642 spin_unlock_irq(&mctz
->lock
);
647 * Implementation Note: reading percpu statistics for memcg.
649 * Both of vmstat[] and percpu_counter has threshold and do periodic
650 * synchronization to implement "quick" read. There are trade-off between
651 * reading cost and precision of value. Then, we may have a chance to implement
652 * a periodic synchronizion of counter in memcg's counter.
654 * But this _read() function is used for user interface now. The user accounts
655 * memory usage by memory cgroup and he _always_ requires exact value because
656 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
657 * have to visit all online cpus and make sum. So, for now, unnecessary
658 * synchronization is not implemented. (just implemented for cpu hotplug)
660 * If there are kernel internal actions which can make use of some not-exact
661 * value, and reading all cpu value can be performance bottleneck in some
662 * common workload, threashold and synchonization as vmstat[] should be
665 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
666 enum mem_cgroup_stat_index idx
)
671 for_each_possible_cpu(cpu
)
672 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
676 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
677 enum mem_cgroup_events_index idx
)
679 unsigned long val
= 0;
682 for_each_possible_cpu(cpu
)
683 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
687 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
692 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
693 * counted as CACHE even if it's on ANON LRU.
696 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
699 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
702 if (PageTransHuge(page
))
703 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
706 /* pagein of a big page is an event. So, ignore page size */
708 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
710 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
711 nr_pages
= -nr_pages
; /* for event */
714 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
717 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
719 unsigned int lru_mask
)
721 unsigned long nr
= 0;
724 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
726 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
727 struct mem_cgroup_per_zone
*mz
;
731 if (!(BIT(lru
) & lru_mask
))
733 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
734 nr
+= mz
->lru_size
[lru
];
740 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
741 unsigned int lru_mask
)
743 unsigned long nr
= 0;
746 for_each_node_state(nid
, N_MEMORY
)
747 nr
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
751 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
752 enum mem_cgroup_events_target target
)
754 unsigned long val
, next
;
756 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
757 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
758 /* from time_after() in jiffies.h */
759 if ((long)next
- (long)val
< 0) {
761 case MEM_CGROUP_TARGET_THRESH
:
762 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
764 case MEM_CGROUP_TARGET_SOFTLIMIT
:
765 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
767 case MEM_CGROUP_TARGET_NUMAINFO
:
768 next
= val
+ NUMAINFO_EVENTS_TARGET
;
773 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
780 * Check events in order.
783 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
785 /* threshold event is triggered in finer grain than soft limit */
786 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
787 MEM_CGROUP_TARGET_THRESH
))) {
789 bool do_numainfo __maybe_unused
;
791 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
792 MEM_CGROUP_TARGET_SOFTLIMIT
);
794 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
795 MEM_CGROUP_TARGET_NUMAINFO
);
797 mem_cgroup_threshold(memcg
);
798 if (unlikely(do_softlimit
))
799 mem_cgroup_update_tree(memcg
, page
);
801 if (unlikely(do_numainfo
))
802 atomic_inc(&memcg
->numainfo_events
);
807 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
810 * mm_update_next_owner() may clear mm->owner to NULL
811 * if it races with swapoff, page migration, etc.
812 * So this can be called with p == NULL.
817 return mem_cgroup_from_css(task_css(p
, memory_cgrp_id
));
819 EXPORT_SYMBOL(mem_cgroup_from_task
);
821 static struct mem_cgroup
*get_mem_cgroup_from_mm(struct mm_struct
*mm
)
823 struct mem_cgroup
*memcg
= NULL
;
828 * Page cache insertions can happen withou an
829 * actual mm context, e.g. during disk probing
830 * on boot, loopback IO, acct() writes etc.
833 memcg
= root_mem_cgroup
;
835 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
836 if (unlikely(!memcg
))
837 memcg
= root_mem_cgroup
;
839 } while (!css_tryget_online(&memcg
->css
));
845 * mem_cgroup_iter - iterate over memory cgroup hierarchy
846 * @root: hierarchy root
847 * @prev: previously returned memcg, NULL on first invocation
848 * @reclaim: cookie for shared reclaim walks, NULL for full walks
850 * Returns references to children of the hierarchy below @root, or
851 * @root itself, or %NULL after a full round-trip.
853 * Caller must pass the return value in @prev on subsequent
854 * invocations for reference counting, or use mem_cgroup_iter_break()
855 * to cancel a hierarchy walk before the round-trip is complete.
857 * Reclaimers can specify a zone and a priority level in @reclaim to
858 * divide up the memcgs in the hierarchy among all concurrent
859 * reclaimers operating on the same zone and priority.
861 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
862 struct mem_cgroup
*prev
,
863 struct mem_cgroup_reclaim_cookie
*reclaim
)
865 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
866 struct cgroup_subsys_state
*css
= NULL
;
867 struct mem_cgroup
*memcg
= NULL
;
868 struct mem_cgroup
*pos
= NULL
;
870 if (mem_cgroup_disabled())
874 root
= root_mem_cgroup
;
876 if (prev
&& !reclaim
)
879 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
888 struct mem_cgroup_per_zone
*mz
;
890 mz
= mem_cgroup_zone_zoneinfo(root
, reclaim
->zone
);
891 iter
= &mz
->iter
[reclaim
->priority
];
893 if (prev
&& reclaim
->generation
!= iter
->generation
)
897 pos
= READ_ONCE(iter
->position
);
899 * A racing update may change the position and
900 * put the last reference, hence css_tryget(),
901 * or retry to see the updated position.
903 } while (pos
&& !css_tryget(&pos
->css
));
910 css
= css_next_descendant_pre(css
, &root
->css
);
913 * Reclaimers share the hierarchy walk, and a
914 * new one might jump in right at the end of
915 * the hierarchy - make sure they see at least
916 * one group and restart from the beginning.
924 * Verify the css and acquire a reference. The root
925 * is provided by the caller, so we know it's alive
926 * and kicking, and don't take an extra reference.
928 memcg
= mem_cgroup_from_css(css
);
930 if (css
== &root
->css
)
933 if (css_tryget(css
)) {
935 * Make sure the memcg is initialized:
936 * mem_cgroup_css_online() orders the the
937 * initialization against setting the flag.
939 if (smp_load_acquire(&memcg
->initialized
))
949 if (cmpxchg(&iter
->position
, pos
, memcg
) == pos
) {
951 css_get(&memcg
->css
);
957 * pairs with css_tryget when dereferencing iter->position
966 reclaim
->generation
= iter
->generation
;
972 if (prev
&& prev
!= root
)
979 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
980 * @root: hierarchy root
981 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
983 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
984 struct mem_cgroup
*prev
)
987 root
= root_mem_cgroup
;
988 if (prev
&& prev
!= root
)
993 * Iteration constructs for visiting all cgroups (under a tree). If
994 * loops are exited prematurely (break), mem_cgroup_iter_break() must
995 * be used for reference counting.
997 #define for_each_mem_cgroup_tree(iter, root) \
998 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1000 iter = mem_cgroup_iter(root, iter, NULL))
1002 #define for_each_mem_cgroup(iter) \
1003 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1005 iter = mem_cgroup_iter(NULL, iter, NULL))
1008 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1009 * @zone: zone of the wanted lruvec
1010 * @memcg: memcg of the wanted lruvec
1012 * Returns the lru list vector holding pages for the given @zone and
1013 * @mem. This can be the global zone lruvec, if the memory controller
1016 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1017 struct mem_cgroup
*memcg
)
1019 struct mem_cgroup_per_zone
*mz
;
1020 struct lruvec
*lruvec
;
1022 if (mem_cgroup_disabled()) {
1023 lruvec
= &zone
->lruvec
;
1027 mz
= mem_cgroup_zone_zoneinfo(memcg
, zone
);
1028 lruvec
= &mz
->lruvec
;
1031 * Since a node can be onlined after the mem_cgroup was created,
1032 * we have to be prepared to initialize lruvec->zone here;
1033 * and if offlined then reonlined, we need to reinitialize it.
1035 if (unlikely(lruvec
->zone
!= zone
))
1036 lruvec
->zone
= zone
;
1041 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1043 * @zone: zone of the page
1045 * This function is only safe when following the LRU page isolation
1046 * and putback protocol: the LRU lock must be held, and the page must
1047 * either be PageLRU() or the caller must have isolated/allocated it.
1049 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1051 struct mem_cgroup_per_zone
*mz
;
1052 struct mem_cgroup
*memcg
;
1053 struct lruvec
*lruvec
;
1055 if (mem_cgroup_disabled()) {
1056 lruvec
= &zone
->lruvec
;
1060 memcg
= page
->mem_cgroup
;
1062 * Swapcache readahead pages are added to the LRU - and
1063 * possibly migrated - before they are charged.
1066 memcg
= root_mem_cgroup
;
1068 mz
= mem_cgroup_page_zoneinfo(memcg
, page
);
1069 lruvec
= &mz
->lruvec
;
1072 * Since a node can be onlined after the mem_cgroup was created,
1073 * we have to be prepared to initialize lruvec->zone here;
1074 * and if offlined then reonlined, we need to reinitialize it.
1076 if (unlikely(lruvec
->zone
!= zone
))
1077 lruvec
->zone
= zone
;
1082 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1083 * @lruvec: mem_cgroup per zone lru vector
1084 * @lru: index of lru list the page is sitting on
1085 * @nr_pages: positive when adding or negative when removing
1087 * This function must be called when a page is added to or removed from an
1090 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1093 struct mem_cgroup_per_zone
*mz
;
1094 unsigned long *lru_size
;
1096 if (mem_cgroup_disabled())
1099 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1100 lru_size
= mz
->lru_size
+ lru
;
1101 *lru_size
+= nr_pages
;
1102 VM_BUG_ON((long)(*lru_size
) < 0);
1105 bool task_in_mem_cgroup(struct task_struct
*task
, struct mem_cgroup
*memcg
)
1107 struct mem_cgroup
*task_memcg
;
1108 struct task_struct
*p
;
1111 p
= find_lock_task_mm(task
);
1113 task_memcg
= get_mem_cgroup_from_mm(p
->mm
);
1117 * All threads may have already detached their mm's, but the oom
1118 * killer still needs to detect if they have already been oom
1119 * killed to prevent needlessly killing additional tasks.
1122 task_memcg
= mem_cgroup_from_task(task
);
1123 css_get(&task_memcg
->css
);
1126 ret
= mem_cgroup_is_descendant(task_memcg
, memcg
);
1127 css_put(&task_memcg
->css
);
1131 #define mem_cgroup_from_counter(counter, member) \
1132 container_of(counter, struct mem_cgroup, member)
1135 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1136 * @memcg: the memory cgroup
1138 * Returns the maximum amount of memory @mem can be charged with, in
1141 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1143 unsigned long margin
= 0;
1144 unsigned long count
;
1145 unsigned long limit
;
1147 count
= page_counter_read(&memcg
->memory
);
1148 limit
= READ_ONCE(memcg
->memory
.limit
);
1150 margin
= limit
- count
;
1152 if (do_swap_account
) {
1153 count
= page_counter_read(&memcg
->memsw
);
1154 limit
= READ_ONCE(memcg
->memsw
.limit
);
1156 margin
= min(margin
, limit
- count
);
1163 * A routine for checking "mem" is under move_account() or not.
1165 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1166 * moving cgroups. This is for waiting at high-memory pressure
1169 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1171 struct mem_cgroup
*from
;
1172 struct mem_cgroup
*to
;
1175 * Unlike task_move routines, we access mc.to, mc.from not under
1176 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1178 spin_lock(&mc
.lock
);
1184 ret
= mem_cgroup_is_descendant(from
, memcg
) ||
1185 mem_cgroup_is_descendant(to
, memcg
);
1187 spin_unlock(&mc
.lock
);
1191 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1193 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1194 if (mem_cgroup_under_move(memcg
)) {
1196 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1197 /* moving charge context might have finished. */
1200 finish_wait(&mc
.waitq
, &wait
);
1207 #define K(x) ((x) << (PAGE_SHIFT-10))
1209 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1210 * @memcg: The memory cgroup that went over limit
1211 * @p: Task that is going to be killed
1213 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1216 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1218 /* oom_info_lock ensures that parallel ooms do not interleave */
1219 static DEFINE_MUTEX(oom_info_lock
);
1220 struct mem_cgroup
*iter
;
1223 mutex_lock(&oom_info_lock
);
1227 pr_info("Task in ");
1228 pr_cont_cgroup_path(task_cgroup(p
, memory_cgrp_id
));
1229 pr_cont(" killed as a result of limit of ");
1231 pr_info("Memory limit reached of cgroup ");
1234 pr_cont_cgroup_path(memcg
->css
.cgroup
);
1239 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1240 K((u64
)page_counter_read(&memcg
->memory
)),
1241 K((u64
)memcg
->memory
.limit
), memcg
->memory
.failcnt
);
1242 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1243 K((u64
)page_counter_read(&memcg
->memsw
)),
1244 K((u64
)memcg
->memsw
.limit
), memcg
->memsw
.failcnt
);
1245 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1246 K((u64
)page_counter_read(&memcg
->kmem
)),
1247 K((u64
)memcg
->kmem
.limit
), memcg
->kmem
.failcnt
);
1249 for_each_mem_cgroup_tree(iter
, memcg
) {
1250 pr_info("Memory cgroup stats for ");
1251 pr_cont_cgroup_path(iter
->css
.cgroup
);
1254 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1255 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1257 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1258 K(mem_cgroup_read_stat(iter
, i
)));
1261 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1262 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1263 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1267 mutex_unlock(&oom_info_lock
);
1271 * This function returns the number of memcg under hierarchy tree. Returns
1272 * 1(self count) if no children.
1274 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1277 struct mem_cgroup
*iter
;
1279 for_each_mem_cgroup_tree(iter
, memcg
)
1285 * Return the memory (and swap, if configured) limit for a memcg.
1287 static unsigned long mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1289 unsigned long limit
;
1291 limit
= memcg
->memory
.limit
;
1292 if (mem_cgroup_swappiness(memcg
)) {
1293 unsigned long memsw_limit
;
1295 memsw_limit
= memcg
->memsw
.limit
;
1296 limit
= min(limit
+ total_swap_pages
, memsw_limit
);
1301 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1304 struct oom_control oc
= {
1307 .gfp_mask
= gfp_mask
,
1310 struct mem_cgroup
*iter
;
1311 unsigned long chosen_points
= 0;
1312 unsigned long totalpages
;
1313 unsigned int points
= 0;
1314 struct task_struct
*chosen
= NULL
;
1316 mutex_lock(&oom_lock
);
1319 * If current has a pending SIGKILL or is exiting, then automatically
1320 * select it. The goal is to allow it to allocate so that it may
1321 * quickly exit and free its memory.
1323 if (fatal_signal_pending(current
) || task_will_free_mem(current
)) {
1324 mark_oom_victim(current
);
1328 check_panic_on_oom(&oc
, CONSTRAINT_MEMCG
, memcg
);
1329 totalpages
= mem_cgroup_get_limit(memcg
) ? : 1;
1330 for_each_mem_cgroup_tree(iter
, memcg
) {
1331 struct css_task_iter it
;
1332 struct task_struct
*task
;
1334 css_task_iter_start(&iter
->css
, &it
);
1335 while ((task
= css_task_iter_next(&it
))) {
1336 switch (oom_scan_process_thread(&oc
, task
, totalpages
)) {
1337 case OOM_SCAN_SELECT
:
1339 put_task_struct(chosen
);
1341 chosen_points
= ULONG_MAX
;
1342 get_task_struct(chosen
);
1344 case OOM_SCAN_CONTINUE
:
1346 case OOM_SCAN_ABORT
:
1347 css_task_iter_end(&it
);
1348 mem_cgroup_iter_break(memcg
, iter
);
1350 put_task_struct(chosen
);
1355 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1356 if (!points
|| points
< chosen_points
)
1358 /* Prefer thread group leaders for display purposes */
1359 if (points
== chosen_points
&&
1360 thread_group_leader(chosen
))
1364 put_task_struct(chosen
);
1366 chosen_points
= points
;
1367 get_task_struct(chosen
);
1369 css_task_iter_end(&it
);
1373 points
= chosen_points
* 1000 / totalpages
;
1374 oom_kill_process(&oc
, chosen
, points
, totalpages
, memcg
,
1375 "Memory cgroup out of memory");
1378 mutex_unlock(&oom_lock
);
1381 #if MAX_NUMNODES > 1
1384 * test_mem_cgroup_node_reclaimable
1385 * @memcg: the target memcg
1386 * @nid: the node ID to be checked.
1387 * @noswap : specify true here if the user wants flle only information.
1389 * This function returns whether the specified memcg contains any
1390 * reclaimable pages on a node. Returns true if there are any reclaimable
1391 * pages in the node.
1393 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1394 int nid
, bool noswap
)
1396 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1398 if (noswap
|| !total_swap_pages
)
1400 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1407 * Always updating the nodemask is not very good - even if we have an empty
1408 * list or the wrong list here, we can start from some node and traverse all
1409 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1412 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1416 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1417 * pagein/pageout changes since the last update.
1419 if (!atomic_read(&memcg
->numainfo_events
))
1421 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1424 /* make a nodemask where this memcg uses memory from */
1425 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1427 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1429 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1430 node_clear(nid
, memcg
->scan_nodes
);
1433 atomic_set(&memcg
->numainfo_events
, 0);
1434 atomic_set(&memcg
->numainfo_updating
, 0);
1438 * Selecting a node where we start reclaim from. Because what we need is just
1439 * reducing usage counter, start from anywhere is O,K. Considering
1440 * memory reclaim from current node, there are pros. and cons.
1442 * Freeing memory from current node means freeing memory from a node which
1443 * we'll use or we've used. So, it may make LRU bad. And if several threads
1444 * hit limits, it will see a contention on a node. But freeing from remote
1445 * node means more costs for memory reclaim because of memory latency.
1447 * Now, we use round-robin. Better algorithm is welcomed.
1449 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1453 mem_cgroup_may_update_nodemask(memcg
);
1454 node
= memcg
->last_scanned_node
;
1456 node
= next_node(node
, memcg
->scan_nodes
);
1457 if (node
== MAX_NUMNODES
)
1458 node
= first_node(memcg
->scan_nodes
);
1460 * We call this when we hit limit, not when pages are added to LRU.
1461 * No LRU may hold pages because all pages are UNEVICTABLE or
1462 * memcg is too small and all pages are not on LRU. In that case,
1463 * we use curret node.
1465 if (unlikely(node
== MAX_NUMNODES
))
1466 node
= numa_node_id();
1468 memcg
->last_scanned_node
= node
;
1472 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1478 static int mem_cgroup_soft_reclaim(struct mem_cgroup
*root_memcg
,
1481 unsigned long *total_scanned
)
1483 struct mem_cgroup
*victim
= NULL
;
1486 unsigned long excess
;
1487 unsigned long nr_scanned
;
1488 struct mem_cgroup_reclaim_cookie reclaim
= {
1493 excess
= soft_limit_excess(root_memcg
);
1496 victim
= mem_cgroup_iter(root_memcg
, victim
, &reclaim
);
1501 * If we have not been able to reclaim
1502 * anything, it might because there are
1503 * no reclaimable pages under this hierarchy
1508 * We want to do more targeted reclaim.
1509 * excess >> 2 is not to excessive so as to
1510 * reclaim too much, nor too less that we keep
1511 * coming back to reclaim from this cgroup
1513 if (total
>= (excess
>> 2) ||
1514 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
))
1519 total
+= mem_cgroup_shrink_node_zone(victim
, gfp_mask
, false,
1521 *total_scanned
+= nr_scanned
;
1522 if (!soft_limit_excess(root_memcg
))
1525 mem_cgroup_iter_break(root_memcg
, victim
);
1529 #ifdef CONFIG_LOCKDEP
1530 static struct lockdep_map memcg_oom_lock_dep_map
= {
1531 .name
= "memcg_oom_lock",
1535 static DEFINE_SPINLOCK(memcg_oom_lock
);
1538 * Check OOM-Killer is already running under our hierarchy.
1539 * If someone is running, return false.
1541 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1543 struct mem_cgroup
*iter
, *failed
= NULL
;
1545 spin_lock(&memcg_oom_lock
);
1547 for_each_mem_cgroup_tree(iter
, memcg
) {
1548 if (iter
->oom_lock
) {
1550 * this subtree of our hierarchy is already locked
1551 * so we cannot give a lock.
1554 mem_cgroup_iter_break(memcg
, iter
);
1557 iter
->oom_lock
= true;
1562 * OK, we failed to lock the whole subtree so we have
1563 * to clean up what we set up to the failing subtree
1565 for_each_mem_cgroup_tree(iter
, memcg
) {
1566 if (iter
== failed
) {
1567 mem_cgroup_iter_break(memcg
, iter
);
1570 iter
->oom_lock
= false;
1573 mutex_acquire(&memcg_oom_lock_dep_map
, 0, 1, _RET_IP_
);
1575 spin_unlock(&memcg_oom_lock
);
1580 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1582 struct mem_cgroup
*iter
;
1584 spin_lock(&memcg_oom_lock
);
1585 mutex_release(&memcg_oom_lock_dep_map
, 1, _RET_IP_
);
1586 for_each_mem_cgroup_tree(iter
, memcg
)
1587 iter
->oom_lock
= false;
1588 spin_unlock(&memcg_oom_lock
);
1591 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1593 struct mem_cgroup
*iter
;
1595 spin_lock(&memcg_oom_lock
);
1596 for_each_mem_cgroup_tree(iter
, memcg
)
1598 spin_unlock(&memcg_oom_lock
);
1601 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1603 struct mem_cgroup
*iter
;
1606 * When a new child is created while the hierarchy is under oom,
1607 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1609 spin_lock(&memcg_oom_lock
);
1610 for_each_mem_cgroup_tree(iter
, memcg
)
1611 if (iter
->under_oom
> 0)
1613 spin_unlock(&memcg_oom_lock
);
1616 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1618 struct oom_wait_info
{
1619 struct mem_cgroup
*memcg
;
1623 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1624 unsigned mode
, int sync
, void *arg
)
1626 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1627 struct mem_cgroup
*oom_wait_memcg
;
1628 struct oom_wait_info
*oom_wait_info
;
1630 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1631 oom_wait_memcg
= oom_wait_info
->memcg
;
1633 if (!mem_cgroup_is_descendant(wake_memcg
, oom_wait_memcg
) &&
1634 !mem_cgroup_is_descendant(oom_wait_memcg
, wake_memcg
))
1636 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1639 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1642 * For the following lockless ->under_oom test, the only required
1643 * guarantee is that it must see the state asserted by an OOM when
1644 * this function is called as a result of userland actions
1645 * triggered by the notification of the OOM. This is trivially
1646 * achieved by invoking mem_cgroup_mark_under_oom() before
1647 * triggering notification.
1649 if (memcg
&& memcg
->under_oom
)
1650 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1653 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
1655 if (!current
->memcg_oom
.may_oom
)
1658 * We are in the middle of the charge context here, so we
1659 * don't want to block when potentially sitting on a callstack
1660 * that holds all kinds of filesystem and mm locks.
1662 * Also, the caller may handle a failed allocation gracefully
1663 * (like optional page cache readahead) and so an OOM killer
1664 * invocation might not even be necessary.
1666 * That's why we don't do anything here except remember the
1667 * OOM context and then deal with it at the end of the page
1668 * fault when the stack is unwound, the locks are released,
1669 * and when we know whether the fault was overall successful.
1671 css_get(&memcg
->css
);
1672 current
->memcg_oom
.memcg
= memcg
;
1673 current
->memcg_oom
.gfp_mask
= mask
;
1674 current
->memcg_oom
.order
= order
;
1678 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1679 * @handle: actually kill/wait or just clean up the OOM state
1681 * This has to be called at the end of a page fault if the memcg OOM
1682 * handler was enabled.
1684 * Memcg supports userspace OOM handling where failed allocations must
1685 * sleep on a waitqueue until the userspace task resolves the
1686 * situation. Sleeping directly in the charge context with all kinds
1687 * of locks held is not a good idea, instead we remember an OOM state
1688 * in the task and mem_cgroup_oom_synchronize() has to be called at
1689 * the end of the page fault to complete the OOM handling.
1691 * Returns %true if an ongoing memcg OOM situation was detected and
1692 * completed, %false otherwise.
1694 bool mem_cgroup_oom_synchronize(bool handle
)
1696 struct mem_cgroup
*memcg
= current
->memcg_oom
.memcg
;
1697 struct oom_wait_info owait
;
1700 /* OOM is global, do not handle */
1704 if (!handle
|| oom_killer_disabled
)
1707 owait
.memcg
= memcg
;
1708 owait
.wait
.flags
= 0;
1709 owait
.wait
.func
= memcg_oom_wake_function
;
1710 owait
.wait
.private = current
;
1711 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1713 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1714 mem_cgroup_mark_under_oom(memcg
);
1716 locked
= mem_cgroup_oom_trylock(memcg
);
1719 mem_cgroup_oom_notify(memcg
);
1721 if (locked
&& !memcg
->oom_kill_disable
) {
1722 mem_cgroup_unmark_under_oom(memcg
);
1723 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1724 mem_cgroup_out_of_memory(memcg
, current
->memcg_oom
.gfp_mask
,
1725 current
->memcg_oom
.order
);
1728 mem_cgroup_unmark_under_oom(memcg
);
1729 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1733 mem_cgroup_oom_unlock(memcg
);
1735 * There is no guarantee that an OOM-lock contender
1736 * sees the wakeups triggered by the OOM kill
1737 * uncharges. Wake any sleepers explicitely.
1739 memcg_oom_recover(memcg
);
1742 current
->memcg_oom
.memcg
= NULL
;
1743 css_put(&memcg
->css
);
1748 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1749 * @page: page that is going to change accounted state
1751 * This function must mark the beginning of an accounted page state
1752 * change to prevent double accounting when the page is concurrently
1753 * being moved to another memcg:
1755 * memcg = mem_cgroup_begin_page_stat(page);
1756 * if (TestClearPageState(page))
1757 * mem_cgroup_update_page_stat(memcg, state, -1);
1758 * mem_cgroup_end_page_stat(memcg);
1760 struct mem_cgroup
*mem_cgroup_begin_page_stat(struct page
*page
)
1762 struct mem_cgroup
*memcg
;
1763 unsigned long flags
;
1766 * The RCU lock is held throughout the transaction. The fast
1767 * path can get away without acquiring the memcg->move_lock
1768 * because page moving starts with an RCU grace period.
1770 * The RCU lock also protects the memcg from being freed when
1771 * the page state that is going to change is the only thing
1772 * preventing the page from being uncharged.
1773 * E.g. end-writeback clearing PageWriteback(), which allows
1774 * migration to go ahead and uncharge the page before the
1775 * account transaction might be complete.
1779 if (mem_cgroup_disabled())
1782 memcg
= page
->mem_cgroup
;
1783 if (unlikely(!memcg
))
1786 if (atomic_read(&memcg
->moving_account
) <= 0)
1789 spin_lock_irqsave(&memcg
->move_lock
, flags
);
1790 if (memcg
!= page
->mem_cgroup
) {
1791 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1796 * When charge migration first begins, we can have locked and
1797 * unlocked page stat updates happening concurrently. Track
1798 * the task who has the lock for mem_cgroup_end_page_stat().
1800 memcg
->move_lock_task
= current
;
1801 memcg
->move_lock_flags
= flags
;
1805 EXPORT_SYMBOL(mem_cgroup_begin_page_stat
);
1808 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1809 * @memcg: the memcg that was accounted against
1811 void mem_cgroup_end_page_stat(struct mem_cgroup
*memcg
)
1813 if (memcg
&& memcg
->move_lock_task
== current
) {
1814 unsigned long flags
= memcg
->move_lock_flags
;
1816 memcg
->move_lock_task
= NULL
;
1817 memcg
->move_lock_flags
= 0;
1819 spin_unlock_irqrestore(&memcg
->move_lock
, flags
);
1824 EXPORT_SYMBOL(mem_cgroup_end_page_stat
);
1827 * size of first charge trial. "32" comes from vmscan.c's magic value.
1828 * TODO: maybe necessary to use big numbers in big irons.
1830 #define CHARGE_BATCH 32U
1831 struct memcg_stock_pcp
{
1832 struct mem_cgroup
*cached
; /* this never be root cgroup */
1833 unsigned int nr_pages
;
1834 struct work_struct work
;
1835 unsigned long flags
;
1836 #define FLUSHING_CACHED_CHARGE 0
1838 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1839 static DEFINE_MUTEX(percpu_charge_mutex
);
1842 * consume_stock: Try to consume stocked charge on this cpu.
1843 * @memcg: memcg to consume from.
1844 * @nr_pages: how many pages to charge.
1846 * The charges will only happen if @memcg matches the current cpu's memcg
1847 * stock, and at least @nr_pages are available in that stock. Failure to
1848 * service an allocation will refill the stock.
1850 * returns true if successful, false otherwise.
1852 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1854 struct memcg_stock_pcp
*stock
;
1857 if (nr_pages
> CHARGE_BATCH
)
1860 stock
= &get_cpu_var(memcg_stock
);
1861 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
) {
1862 stock
->nr_pages
-= nr_pages
;
1865 put_cpu_var(memcg_stock
);
1870 * Returns stocks cached in percpu and reset cached information.
1872 static void drain_stock(struct memcg_stock_pcp
*stock
)
1874 struct mem_cgroup
*old
= stock
->cached
;
1876 if (stock
->nr_pages
) {
1877 page_counter_uncharge(&old
->memory
, stock
->nr_pages
);
1878 if (do_swap_account
)
1879 page_counter_uncharge(&old
->memsw
, stock
->nr_pages
);
1880 css_put_many(&old
->css
, stock
->nr_pages
);
1881 stock
->nr_pages
= 0;
1883 stock
->cached
= NULL
;
1887 * This must be called under preempt disabled or must be called by
1888 * a thread which is pinned to local cpu.
1890 static void drain_local_stock(struct work_struct
*dummy
)
1892 struct memcg_stock_pcp
*stock
= this_cpu_ptr(&memcg_stock
);
1894 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
1898 * Cache charges(val) to local per_cpu area.
1899 * This will be consumed by consume_stock() function, later.
1901 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
1903 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1905 if (stock
->cached
!= memcg
) { /* reset if necessary */
1907 stock
->cached
= memcg
;
1909 stock
->nr_pages
+= nr_pages
;
1910 put_cpu_var(memcg_stock
);
1914 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1915 * of the hierarchy under it.
1917 static void drain_all_stock(struct mem_cgroup
*root_memcg
)
1921 /* If someone's already draining, avoid adding running more workers. */
1922 if (!mutex_trylock(&percpu_charge_mutex
))
1924 /* Notify other cpus that system-wide "drain" is running */
1927 for_each_online_cpu(cpu
) {
1928 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1929 struct mem_cgroup
*memcg
;
1931 memcg
= stock
->cached
;
1932 if (!memcg
|| !stock
->nr_pages
)
1934 if (!mem_cgroup_is_descendant(memcg
, root_memcg
))
1936 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
1938 drain_local_stock(&stock
->work
);
1940 schedule_work_on(cpu
, &stock
->work
);
1945 mutex_unlock(&percpu_charge_mutex
);
1948 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1949 unsigned long action
,
1952 int cpu
= (unsigned long)hcpu
;
1953 struct memcg_stock_pcp
*stock
;
1955 if (action
== CPU_ONLINE
)
1958 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
1961 stock
= &per_cpu(memcg_stock
, cpu
);
1966 static int try_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1967 unsigned int nr_pages
)
1969 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
1970 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1971 struct mem_cgroup
*mem_over_limit
;
1972 struct page_counter
*counter
;
1973 unsigned long nr_reclaimed
;
1974 bool may_swap
= true;
1975 bool drained
= false;
1978 if (mem_cgroup_is_root(memcg
))
1981 if (consume_stock(memcg
, nr_pages
))
1984 if (!do_swap_account
||
1985 !page_counter_try_charge(&memcg
->memsw
, batch
, &counter
)) {
1986 if (!page_counter_try_charge(&memcg
->memory
, batch
, &counter
))
1988 if (do_swap_account
)
1989 page_counter_uncharge(&memcg
->memsw
, batch
);
1990 mem_over_limit
= mem_cgroup_from_counter(counter
, memory
);
1992 mem_over_limit
= mem_cgroup_from_counter(counter
, memsw
);
1996 if (batch
> nr_pages
) {
2002 * Unlike in global OOM situations, memcg is not in a physical
2003 * memory shortage. Allow dying and OOM-killed tasks to
2004 * bypass the last charges so that they can exit quickly and
2005 * free their memory.
2007 if (unlikely(test_thread_flag(TIF_MEMDIE
) ||
2008 fatal_signal_pending(current
) ||
2009 current
->flags
& PF_EXITING
))
2012 if (unlikely(task_in_memcg_oom(current
)))
2015 if (!(gfp_mask
& __GFP_WAIT
))
2018 mem_cgroup_events(mem_over_limit
, MEMCG_MAX
, 1);
2020 nr_reclaimed
= try_to_free_mem_cgroup_pages(mem_over_limit
, nr_pages
,
2021 gfp_mask
, may_swap
);
2023 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2027 drain_all_stock(mem_over_limit
);
2032 if (gfp_mask
& __GFP_NORETRY
)
2035 * Even though the limit is exceeded at this point, reclaim
2036 * may have been able to free some pages. Retry the charge
2037 * before killing the task.
2039 * Only for regular pages, though: huge pages are rather
2040 * unlikely to succeed so close to the limit, and we fall back
2041 * to regular pages anyway in case of failure.
2043 if (nr_reclaimed
&& nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
))
2046 * At task move, charge accounts can be doubly counted. So, it's
2047 * better to wait until the end of task_move if something is going on.
2049 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2055 if (gfp_mask
& __GFP_NOFAIL
)
2058 if (fatal_signal_pending(current
))
2061 mem_cgroup_events(mem_over_limit
, MEMCG_OOM
, 1);
2063 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(nr_pages
));
2065 if (!(gfp_mask
& __GFP_NOFAIL
))
2071 css_get_many(&memcg
->css
, batch
);
2072 if (batch
> nr_pages
)
2073 refill_stock(memcg
, batch
- nr_pages
);
2074 if (!(gfp_mask
& __GFP_WAIT
))
2077 * If the hierarchy is above the normal consumption range,
2078 * make the charging task trim their excess contribution.
2081 if (page_counter_read(&memcg
->memory
) <= memcg
->high
)
2083 mem_cgroup_events(memcg
, MEMCG_HIGH
, 1);
2084 try_to_free_mem_cgroup_pages(memcg
, nr_pages
, gfp_mask
, true);
2085 } while ((memcg
= parent_mem_cgroup(memcg
)));
2090 static void cancel_charge(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2092 if (mem_cgroup_is_root(memcg
))
2095 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2096 if (do_swap_account
)
2097 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2099 css_put_many(&memcg
->css
, nr_pages
);
2102 static void lock_page_lru(struct page
*page
, int *isolated
)
2104 struct zone
*zone
= page_zone(page
);
2106 spin_lock_irq(&zone
->lru_lock
);
2107 if (PageLRU(page
)) {
2108 struct lruvec
*lruvec
;
2110 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2112 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2118 static void unlock_page_lru(struct page
*page
, int isolated
)
2120 struct zone
*zone
= page_zone(page
);
2123 struct lruvec
*lruvec
;
2125 lruvec
= mem_cgroup_page_lruvec(page
, zone
);
2126 VM_BUG_ON_PAGE(PageLRU(page
), page
);
2128 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2130 spin_unlock_irq(&zone
->lru_lock
);
2133 static void commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2138 VM_BUG_ON_PAGE(page
->mem_cgroup
, page
);
2141 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2142 * may already be on some other mem_cgroup's LRU. Take care of it.
2145 lock_page_lru(page
, &isolated
);
2148 * Nobody should be changing or seriously looking at
2149 * page->mem_cgroup at this point:
2151 * - the page is uncharged
2153 * - the page is off-LRU
2155 * - an anonymous fault has exclusive page access, except for
2156 * a locked page table
2158 * - a page cache insertion, a swapin fault, or a migration
2159 * have the page locked
2161 page
->mem_cgroup
= memcg
;
2164 unlock_page_lru(page
, isolated
);
2167 #ifdef CONFIG_MEMCG_KMEM
2168 int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
,
2169 unsigned long nr_pages
)
2171 struct page_counter
*counter
;
2174 ret
= page_counter_try_charge(&memcg
->kmem
, nr_pages
, &counter
);
2178 ret
= try_charge(memcg
, gfp
, nr_pages
);
2179 if (ret
== -EINTR
) {
2181 * try_charge() chose to bypass to root due to OOM kill or
2182 * fatal signal. Since our only options are to either fail
2183 * the allocation or charge it to this cgroup, do it as a
2184 * temporary condition. But we can't fail. From a kmem/slab
2185 * perspective, the cache has already been selected, by
2186 * mem_cgroup_kmem_get_cache(), so it is too late to change
2189 * This condition will only trigger if the task entered
2190 * memcg_charge_kmem in a sane state, but was OOM-killed
2191 * during try_charge() above. Tasks that were already dying
2192 * when the allocation triggers should have been already
2193 * directed to the root cgroup in memcontrol.h
2195 page_counter_charge(&memcg
->memory
, nr_pages
);
2196 if (do_swap_account
)
2197 page_counter_charge(&memcg
->memsw
, nr_pages
);
2198 css_get_many(&memcg
->css
, nr_pages
);
2201 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2206 void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, unsigned long nr_pages
)
2208 page_counter_uncharge(&memcg
->memory
, nr_pages
);
2209 if (do_swap_account
)
2210 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
2212 page_counter_uncharge(&memcg
->kmem
, nr_pages
);
2214 css_put_many(&memcg
->css
, nr_pages
);
2217 static int memcg_alloc_cache_id(void)
2222 id
= ida_simple_get(&memcg_cache_ida
,
2223 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2227 if (id
< memcg_nr_cache_ids
)
2231 * There's no space for the new id in memcg_caches arrays,
2232 * so we have to grow them.
2234 down_write(&memcg_cache_ids_sem
);
2236 size
= 2 * (id
+ 1);
2237 if (size
< MEMCG_CACHES_MIN_SIZE
)
2238 size
= MEMCG_CACHES_MIN_SIZE
;
2239 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2240 size
= MEMCG_CACHES_MAX_SIZE
;
2242 err
= memcg_update_all_caches(size
);
2244 err
= memcg_update_all_list_lrus(size
);
2246 memcg_nr_cache_ids
= size
;
2248 up_write(&memcg_cache_ids_sem
);
2251 ida_simple_remove(&memcg_cache_ida
, id
);
2257 static void memcg_free_cache_id(int id
)
2259 ida_simple_remove(&memcg_cache_ida
, id
);
2262 struct memcg_kmem_cache_create_work
{
2263 struct mem_cgroup
*memcg
;
2264 struct kmem_cache
*cachep
;
2265 struct work_struct work
;
2268 static void memcg_kmem_cache_create_func(struct work_struct
*w
)
2270 struct memcg_kmem_cache_create_work
*cw
=
2271 container_of(w
, struct memcg_kmem_cache_create_work
, work
);
2272 struct mem_cgroup
*memcg
= cw
->memcg
;
2273 struct kmem_cache
*cachep
= cw
->cachep
;
2275 memcg_create_kmem_cache(memcg
, cachep
);
2277 css_put(&memcg
->css
);
2282 * Enqueue the creation of a per-memcg kmem_cache.
2284 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2285 struct kmem_cache
*cachep
)
2287 struct memcg_kmem_cache_create_work
*cw
;
2289 cw
= kmalloc(sizeof(*cw
), GFP_NOWAIT
);
2293 css_get(&memcg
->css
);
2296 cw
->cachep
= cachep
;
2297 INIT_WORK(&cw
->work
, memcg_kmem_cache_create_func
);
2299 schedule_work(&cw
->work
);
2302 static void memcg_schedule_kmem_cache_create(struct mem_cgroup
*memcg
,
2303 struct kmem_cache
*cachep
)
2306 * We need to stop accounting when we kmalloc, because if the
2307 * corresponding kmalloc cache is not yet created, the first allocation
2308 * in __memcg_schedule_kmem_cache_create will recurse.
2310 * However, it is better to enclose the whole function. Depending on
2311 * the debugging options enabled, INIT_WORK(), for instance, can
2312 * trigger an allocation. This too, will make us recurse. Because at
2313 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2314 * the safest choice is to do it like this, wrapping the whole function.
2316 current
->memcg_kmem_skip_account
= 1;
2317 __memcg_schedule_kmem_cache_create(memcg
, cachep
);
2318 current
->memcg_kmem_skip_account
= 0;
2322 * Return the kmem_cache we're supposed to use for a slab allocation.
2323 * We try to use the current memcg's version of the cache.
2325 * If the cache does not exist yet, if we are the first user of it,
2326 * we either create it immediately, if possible, or create it asynchronously
2328 * In the latter case, we will let the current allocation go through with
2329 * the original cache.
2331 * Can't be called in interrupt context or from kernel threads.
2332 * This function needs to be called with rcu_read_lock() held.
2334 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
)
2336 struct mem_cgroup
*memcg
;
2337 struct kmem_cache
*memcg_cachep
;
2340 VM_BUG_ON(!is_root_cache(cachep
));
2342 if (current
->memcg_kmem_skip_account
)
2345 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2346 kmemcg_id
= READ_ONCE(memcg
->kmemcg_id
);
2350 memcg_cachep
= cache_from_memcg_idx(cachep
, kmemcg_id
);
2351 if (likely(memcg_cachep
))
2352 return memcg_cachep
;
2355 * If we are in a safe context (can wait, and not in interrupt
2356 * context), we could be be predictable and return right away.
2357 * This would guarantee that the allocation being performed
2358 * already belongs in the new cache.
2360 * However, there are some clashes that can arrive from locking.
2361 * For instance, because we acquire the slab_mutex while doing
2362 * memcg_create_kmem_cache, this means no further allocation
2363 * could happen with the slab_mutex held. So it's better to
2366 memcg_schedule_kmem_cache_create(memcg
, cachep
);
2368 css_put(&memcg
->css
);
2372 void __memcg_kmem_put_cache(struct kmem_cache
*cachep
)
2374 if (!is_root_cache(cachep
))
2375 css_put(&cachep
->memcg_params
.memcg
->css
);
2379 * We need to verify if the allocation against current->mm->owner's memcg is
2380 * possible for the given order. But the page is not allocated yet, so we'll
2381 * need a further commit step to do the final arrangements.
2383 * It is possible for the task to switch cgroups in this mean time, so at
2384 * commit time, we can't rely on task conversion any longer. We'll then use
2385 * the handle argument to return to the caller which cgroup we should commit
2386 * against. We could also return the memcg directly and avoid the pointer
2387 * passing, but a boolean return value gives better semantics considering
2388 * the compiled-out case as well.
2390 * Returning true means the allocation is possible.
2393 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
2395 struct mem_cgroup
*memcg
;
2400 memcg
= get_mem_cgroup_from_mm(current
->mm
);
2402 if (!memcg_kmem_is_active(memcg
)) {
2403 css_put(&memcg
->css
);
2407 ret
= memcg_charge_kmem(memcg
, gfp
, 1 << order
);
2411 css_put(&memcg
->css
);
2415 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
2418 VM_BUG_ON(mem_cgroup_is_root(memcg
));
2420 /* The page allocation failed. Revert */
2422 memcg_uncharge_kmem(memcg
, 1 << order
);
2425 page
->mem_cgroup
= memcg
;
2428 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
2430 struct mem_cgroup
*memcg
= page
->mem_cgroup
;
2435 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg
), page
);
2437 memcg_uncharge_kmem(memcg
, 1 << order
);
2438 page
->mem_cgroup
= NULL
;
2441 struct mem_cgroup
*__mem_cgroup_from_kmem(void *ptr
)
2443 struct mem_cgroup
*memcg
= NULL
;
2444 struct kmem_cache
*cachep
;
2447 page
= virt_to_head_page(ptr
);
2448 if (PageSlab(page
)) {
2449 cachep
= page
->slab_cache
;
2450 if (!is_root_cache(cachep
))
2451 memcg
= cachep
->memcg_params
.memcg
;
2453 /* page allocated by alloc_kmem_pages */
2454 memcg
= page
->mem_cgroup
;
2458 #endif /* CONFIG_MEMCG_KMEM */
2460 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2463 * Because tail pages are not marked as "used", set it. We're under
2464 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2465 * charge/uncharge will be never happen and move_account() is done under
2466 * compound_lock(), so we don't have to take care of races.
2468 void mem_cgroup_split_huge_fixup(struct page
*head
)
2472 if (mem_cgroup_disabled())
2475 for (i
= 1; i
< HPAGE_PMD_NR
; i
++)
2476 head
[i
].mem_cgroup
= head
->mem_cgroup
;
2478 __this_cpu_sub(head
->mem_cgroup
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
2481 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2483 #ifdef CONFIG_MEMCG_SWAP
2484 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
2487 int val
= (charge
) ? 1 : -1;
2488 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
2492 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2493 * @entry: swap entry to be moved
2494 * @from: mem_cgroup which the entry is moved from
2495 * @to: mem_cgroup which the entry is moved to
2497 * It succeeds only when the swap_cgroup's record for this entry is the same
2498 * as the mem_cgroup's id of @from.
2500 * Returns 0 on success, -EINVAL on failure.
2502 * The caller must have charged to @to, IOW, called page_counter_charge() about
2503 * both res and memsw, and called css_get().
2505 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2506 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2508 unsigned short old_id
, new_id
;
2510 old_id
= mem_cgroup_id(from
);
2511 new_id
= mem_cgroup_id(to
);
2513 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2514 mem_cgroup_swap_statistics(from
, false);
2515 mem_cgroup_swap_statistics(to
, true);
2521 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2522 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
2528 static DEFINE_MUTEX(memcg_limit_mutex
);
2530 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2531 unsigned long limit
)
2533 unsigned long curusage
;
2534 unsigned long oldusage
;
2535 bool enlarge
= false;
2540 * For keeping hierarchical_reclaim simple, how long we should retry
2541 * is depends on callers. We set our retry-count to be function
2542 * of # of children which we should visit in this loop.
2544 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2545 mem_cgroup_count_children(memcg
);
2547 oldusage
= page_counter_read(&memcg
->memory
);
2550 if (signal_pending(current
)) {
2555 mutex_lock(&memcg_limit_mutex
);
2556 if (limit
> memcg
->memsw
.limit
) {
2557 mutex_unlock(&memcg_limit_mutex
);
2561 if (limit
> memcg
->memory
.limit
)
2563 ret
= page_counter_limit(&memcg
->memory
, limit
);
2564 mutex_unlock(&memcg_limit_mutex
);
2569 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, true);
2571 curusage
= page_counter_read(&memcg
->memory
);
2572 /* Usage is reduced ? */
2573 if (curusage
>= oldusage
)
2576 oldusage
= curusage
;
2577 } while (retry_count
);
2579 if (!ret
&& enlarge
)
2580 memcg_oom_recover(memcg
);
2585 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
2586 unsigned long limit
)
2588 unsigned long curusage
;
2589 unsigned long oldusage
;
2590 bool enlarge
= false;
2594 /* see mem_cgroup_resize_res_limit */
2595 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
*
2596 mem_cgroup_count_children(memcg
);
2598 oldusage
= page_counter_read(&memcg
->memsw
);
2601 if (signal_pending(current
)) {
2606 mutex_lock(&memcg_limit_mutex
);
2607 if (limit
< memcg
->memory
.limit
) {
2608 mutex_unlock(&memcg_limit_mutex
);
2612 if (limit
> memcg
->memsw
.limit
)
2614 ret
= page_counter_limit(&memcg
->memsw
, limit
);
2615 mutex_unlock(&memcg_limit_mutex
);
2620 try_to_free_mem_cgroup_pages(memcg
, 1, GFP_KERNEL
, false);
2622 curusage
= page_counter_read(&memcg
->memsw
);
2623 /* Usage is reduced ? */
2624 if (curusage
>= oldusage
)
2627 oldusage
= curusage
;
2628 } while (retry_count
);
2630 if (!ret
&& enlarge
)
2631 memcg_oom_recover(memcg
);
2636 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
2638 unsigned long *total_scanned
)
2640 unsigned long nr_reclaimed
= 0;
2641 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
2642 unsigned long reclaimed
;
2644 struct mem_cgroup_tree_per_zone
*mctz
;
2645 unsigned long excess
;
2646 unsigned long nr_scanned
;
2651 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
2653 * This loop can run a while, specially if mem_cgroup's continuously
2654 * keep exceeding their soft limit and putting the system under
2661 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
2666 reclaimed
= mem_cgroup_soft_reclaim(mz
->memcg
, zone
,
2667 gfp_mask
, &nr_scanned
);
2668 nr_reclaimed
+= reclaimed
;
2669 *total_scanned
+= nr_scanned
;
2670 spin_lock_irq(&mctz
->lock
);
2671 __mem_cgroup_remove_exceeded(mz
, mctz
);
2674 * If we failed to reclaim anything from this memory cgroup
2675 * it is time to move on to the next cgroup
2679 next_mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
2681 excess
= soft_limit_excess(mz
->memcg
);
2683 * One school of thought says that we should not add
2684 * back the node to the tree if reclaim returns 0.
2685 * But our reclaim could return 0, simply because due
2686 * to priority we are exposing a smaller subset of
2687 * memory to reclaim from. Consider this as a longer
2690 /* If excess == 0, no tree ops */
2691 __mem_cgroup_insert_exceeded(mz
, mctz
, excess
);
2692 spin_unlock_irq(&mctz
->lock
);
2693 css_put(&mz
->memcg
->css
);
2696 * Could not reclaim anything and there are no more
2697 * mem cgroups to try or we seem to be looping without
2698 * reclaiming anything.
2700 if (!nr_reclaimed
&&
2702 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
2704 } while (!nr_reclaimed
);
2706 css_put(&next_mz
->memcg
->css
);
2707 return nr_reclaimed
;
2711 * Test whether @memcg has children, dead or alive. Note that this
2712 * function doesn't care whether @memcg has use_hierarchy enabled and
2713 * returns %true if there are child csses according to the cgroup
2714 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2716 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
2721 * The lock does not prevent addition or deletion of children, but
2722 * it prevents a new child from being initialized based on this
2723 * parent in css_online(), so it's enough to decide whether
2724 * hierarchically inherited attributes can still be changed or not.
2726 lockdep_assert_held(&memcg_create_mutex
);
2729 ret
= css_next_child(NULL
, &memcg
->css
);
2735 * Reclaims as many pages from the given memcg as possible and moves
2736 * the rest to the parent.
2738 * Caller is responsible for holding css reference for memcg.
2740 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
2742 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2744 /* we call try-to-free pages for make this cgroup empty */
2745 lru_add_drain_all();
2746 /* try to free all pages in this cgroup */
2747 while (nr_retries
&& page_counter_read(&memcg
->memory
)) {
2750 if (signal_pending(current
))
2753 progress
= try_to_free_mem_cgroup_pages(memcg
, 1,
2757 /* maybe some writeback is necessary */
2758 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
2766 static ssize_t
mem_cgroup_force_empty_write(struct kernfs_open_file
*of
,
2767 char *buf
, size_t nbytes
,
2770 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
2772 if (mem_cgroup_is_root(memcg
))
2774 return mem_cgroup_force_empty(memcg
) ?: nbytes
;
2777 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
2780 return mem_cgroup_from_css(css
)->use_hierarchy
;
2783 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
2784 struct cftype
*cft
, u64 val
)
2787 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2788 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(memcg
->css
.parent
);
2790 mutex_lock(&memcg_create_mutex
);
2792 if (memcg
->use_hierarchy
== val
)
2796 * If parent's use_hierarchy is set, we can't make any modifications
2797 * in the child subtrees. If it is unset, then the change can
2798 * occur, provided the current cgroup has no children.
2800 * For the root cgroup, parent_mem is NULL, we allow value to be
2801 * set if there are no children.
2803 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
2804 (val
== 1 || val
== 0)) {
2805 if (!memcg_has_children(memcg
))
2806 memcg
->use_hierarchy
= val
;
2813 mutex_unlock(&memcg_create_mutex
);
2818 static unsigned long tree_stat(struct mem_cgroup
*memcg
,
2819 enum mem_cgroup_stat_index idx
)
2821 struct mem_cgroup
*iter
;
2824 /* Per-cpu values can be negative, use a signed accumulator */
2825 for_each_mem_cgroup_tree(iter
, memcg
)
2826 val
+= mem_cgroup_read_stat(iter
, idx
);
2828 if (val
< 0) /* race ? */
2833 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
2837 if (mem_cgroup_is_root(memcg
)) {
2838 val
= tree_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
2839 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_RSS
);
2841 val
+= tree_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
2844 val
= page_counter_read(&memcg
->memory
);
2846 val
= page_counter_read(&memcg
->memsw
);
2848 return val
<< PAGE_SHIFT
;
2859 static u64
mem_cgroup_read_u64(struct cgroup_subsys_state
*css
,
2862 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2863 struct page_counter
*counter
;
2865 switch (MEMFILE_TYPE(cft
->private)) {
2867 counter
= &memcg
->memory
;
2870 counter
= &memcg
->memsw
;
2873 counter
= &memcg
->kmem
;
2879 switch (MEMFILE_ATTR(cft
->private)) {
2881 if (counter
== &memcg
->memory
)
2882 return mem_cgroup_usage(memcg
, false);
2883 if (counter
== &memcg
->memsw
)
2884 return mem_cgroup_usage(memcg
, true);
2885 return (u64
)page_counter_read(counter
) * PAGE_SIZE
;
2887 return (u64
)counter
->limit
* PAGE_SIZE
;
2889 return (u64
)counter
->watermark
* PAGE_SIZE
;
2891 return counter
->failcnt
;
2892 case RES_SOFT_LIMIT
:
2893 return (u64
)memcg
->soft_limit
* PAGE_SIZE
;
2899 #ifdef CONFIG_MEMCG_KMEM
2900 static int memcg_activate_kmem(struct mem_cgroup
*memcg
,
2901 unsigned long nr_pages
)
2906 BUG_ON(memcg
->kmemcg_id
>= 0);
2907 BUG_ON(memcg
->kmem_acct_activated
);
2908 BUG_ON(memcg
->kmem_acct_active
);
2911 * For simplicity, we won't allow this to be disabled. It also can't
2912 * be changed if the cgroup has children already, or if tasks had
2915 * If tasks join before we set the limit, a person looking at
2916 * kmem.usage_in_bytes will have no way to determine when it took
2917 * place, which makes the value quite meaningless.
2919 * After it first became limited, changes in the value of the limit are
2920 * of course permitted.
2922 mutex_lock(&memcg_create_mutex
);
2923 if (cgroup_has_tasks(memcg
->css
.cgroup
) ||
2924 (memcg
->use_hierarchy
&& memcg_has_children(memcg
)))
2926 mutex_unlock(&memcg_create_mutex
);
2930 memcg_id
= memcg_alloc_cache_id();
2937 * We couldn't have accounted to this cgroup, because it hasn't got
2938 * activated yet, so this should succeed.
2940 err
= page_counter_limit(&memcg
->kmem
, nr_pages
);
2943 static_key_slow_inc(&memcg_kmem_enabled_key
);
2945 * A memory cgroup is considered kmem-active as soon as it gets
2946 * kmemcg_id. Setting the id after enabling static branching will
2947 * guarantee no one starts accounting before all call sites are
2950 memcg
->kmemcg_id
= memcg_id
;
2951 memcg
->kmem_acct_activated
= true;
2952 memcg
->kmem_acct_active
= true;
2957 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2958 unsigned long limit
)
2962 mutex_lock(&memcg_limit_mutex
);
2963 if (!memcg_kmem_is_active(memcg
))
2964 ret
= memcg_activate_kmem(memcg
, limit
);
2966 ret
= page_counter_limit(&memcg
->kmem
, limit
);
2967 mutex_unlock(&memcg_limit_mutex
);
2971 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
2974 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
2979 mutex_lock(&memcg_limit_mutex
);
2981 * If the parent cgroup is not kmem-active now, it cannot be activated
2982 * after this point, because it has at least one child already.
2984 if (memcg_kmem_is_active(parent
))
2985 ret
= memcg_activate_kmem(memcg
, PAGE_COUNTER_MAX
);
2986 mutex_unlock(&memcg_limit_mutex
);
2990 static int memcg_update_kmem_limit(struct mem_cgroup
*memcg
,
2991 unsigned long limit
)
2995 #endif /* CONFIG_MEMCG_KMEM */
2998 * The user of this function is...
3001 static ssize_t
mem_cgroup_write(struct kernfs_open_file
*of
,
3002 char *buf
, size_t nbytes
, loff_t off
)
3004 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3005 unsigned long nr_pages
;
3008 buf
= strstrip(buf
);
3009 ret
= page_counter_memparse(buf
, "-1", &nr_pages
);
3013 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3015 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3019 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3021 ret
= mem_cgroup_resize_limit(memcg
, nr_pages
);
3024 ret
= mem_cgroup_resize_memsw_limit(memcg
, nr_pages
);
3027 ret
= memcg_update_kmem_limit(memcg
, nr_pages
);
3031 case RES_SOFT_LIMIT
:
3032 memcg
->soft_limit
= nr_pages
;
3036 return ret
?: nbytes
;
3039 static ssize_t
mem_cgroup_reset(struct kernfs_open_file
*of
, char *buf
,
3040 size_t nbytes
, loff_t off
)
3042 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
3043 struct page_counter
*counter
;
3045 switch (MEMFILE_TYPE(of_cft(of
)->private)) {
3047 counter
= &memcg
->memory
;
3050 counter
= &memcg
->memsw
;
3053 counter
= &memcg
->kmem
;
3059 switch (MEMFILE_ATTR(of_cft(of
)->private)) {
3061 page_counter_reset_watermark(counter
);
3064 counter
->failcnt
= 0;
3073 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
3076 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
3080 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3081 struct cftype
*cft
, u64 val
)
3083 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3085 if (val
& ~MOVE_MASK
)
3089 * No kind of locking is needed in here, because ->can_attach() will
3090 * check this value once in the beginning of the process, and then carry
3091 * on with stale data. This means that changes to this value will only
3092 * affect task migrations starting after the change.
3094 memcg
->move_charge_at_immigrate
= val
;
3098 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
3099 struct cftype
*cft
, u64 val
)
3106 static int memcg_numa_stat_show(struct seq_file
*m
, void *v
)
3110 unsigned int lru_mask
;
3113 static const struct numa_stat stats
[] = {
3114 { "total", LRU_ALL
},
3115 { "file", LRU_ALL_FILE
},
3116 { "anon", LRU_ALL_ANON
},
3117 { "unevictable", BIT(LRU_UNEVICTABLE
) },
3119 const struct numa_stat
*stat
;
3122 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3124 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3125 nr
= mem_cgroup_nr_lru_pages(memcg
, stat
->lru_mask
);
3126 seq_printf(m
, "%s=%lu", stat
->name
, nr
);
3127 for_each_node_state(nid
, N_MEMORY
) {
3128 nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
3130 seq_printf(m
, " N%d=%lu", nid
, nr
);
3135 for (stat
= stats
; stat
< stats
+ ARRAY_SIZE(stats
); stat
++) {
3136 struct mem_cgroup
*iter
;
3139 for_each_mem_cgroup_tree(iter
, memcg
)
3140 nr
+= mem_cgroup_nr_lru_pages(iter
, stat
->lru_mask
);
3141 seq_printf(m
, "hierarchical_%s=%lu", stat
->name
, nr
);
3142 for_each_node_state(nid
, N_MEMORY
) {
3144 for_each_mem_cgroup_tree(iter
, memcg
)
3145 nr
+= mem_cgroup_node_nr_lru_pages(
3146 iter
, nid
, stat
->lru_mask
);
3147 seq_printf(m
, " N%d=%lu", nid
, nr
);
3154 #endif /* CONFIG_NUMA */
3156 static int memcg_stat_show(struct seq_file
*m
, void *v
)
3158 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
3159 unsigned long memory
, memsw
;
3160 struct mem_cgroup
*mi
;
3163 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names
) !=
3164 MEM_CGROUP_STAT_NSTATS
);
3165 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names
) !=
3166 MEM_CGROUP_EVENTS_NSTATS
);
3167 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
3169 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3170 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3172 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
3173 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
3176 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
3177 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
3178 mem_cgroup_read_events(memcg
, i
));
3180 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
3181 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
3182 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
3184 /* Hierarchical information */
3185 memory
= memsw
= PAGE_COUNTER_MAX
;
3186 for (mi
= memcg
; mi
; mi
= parent_mem_cgroup(mi
)) {
3187 memory
= min(memory
, mi
->memory
.limit
);
3188 memsw
= min(memsw
, mi
->memsw
.limit
);
3190 seq_printf(m
, "hierarchical_memory_limit %llu\n",
3191 (u64
)memory
* PAGE_SIZE
);
3192 if (do_swap_account
)
3193 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
3194 (u64
)memsw
* PAGE_SIZE
);
3196 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
3199 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
3201 for_each_mem_cgroup_tree(mi
, memcg
)
3202 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
3203 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
3206 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
3207 unsigned long long val
= 0;
3209 for_each_mem_cgroup_tree(mi
, memcg
)
3210 val
+= mem_cgroup_read_events(mi
, i
);
3211 seq_printf(m
, "total_%s %llu\n",
3212 mem_cgroup_events_names
[i
], val
);
3215 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
3216 unsigned long long val
= 0;
3218 for_each_mem_cgroup_tree(mi
, memcg
)
3219 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
3220 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
3223 #ifdef CONFIG_DEBUG_VM
3226 struct mem_cgroup_per_zone
*mz
;
3227 struct zone_reclaim_stat
*rstat
;
3228 unsigned long recent_rotated
[2] = {0, 0};
3229 unsigned long recent_scanned
[2] = {0, 0};
3231 for_each_online_node(nid
)
3232 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3233 mz
= &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
3234 rstat
= &mz
->lruvec
.reclaim_stat
;
3236 recent_rotated
[0] += rstat
->recent_rotated
[0];
3237 recent_rotated
[1] += rstat
->recent_rotated
[1];
3238 recent_scanned
[0] += rstat
->recent_scanned
[0];
3239 recent_scanned
[1] += rstat
->recent_scanned
[1];
3241 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
3242 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
3243 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
3244 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
3251 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
3254 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3256 return mem_cgroup_swappiness(memcg
);
3259 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
3260 struct cftype
*cft
, u64 val
)
3262 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3268 memcg
->swappiness
= val
;
3270 vm_swappiness
= val
;
3275 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3277 struct mem_cgroup_threshold_ary
*t
;
3278 unsigned long usage
;
3283 t
= rcu_dereference(memcg
->thresholds
.primary
);
3285 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3290 usage
= mem_cgroup_usage(memcg
, swap
);
3293 * current_threshold points to threshold just below or equal to usage.
3294 * If it's not true, a threshold was crossed after last
3295 * call of __mem_cgroup_threshold().
3297 i
= t
->current_threshold
;
3300 * Iterate backward over array of thresholds starting from
3301 * current_threshold and check if a threshold is crossed.
3302 * If none of thresholds below usage is crossed, we read
3303 * only one element of the array here.
3305 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3306 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3308 /* i = current_threshold + 1 */
3312 * Iterate forward over array of thresholds starting from
3313 * current_threshold+1 and check if a threshold is crossed.
3314 * If none of thresholds above usage is crossed, we read
3315 * only one element of the array here.
3317 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3318 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3320 /* Update current_threshold */
3321 t
->current_threshold
= i
- 1;
3326 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3329 __mem_cgroup_threshold(memcg
, false);
3330 if (do_swap_account
)
3331 __mem_cgroup_threshold(memcg
, true);
3333 memcg
= parent_mem_cgroup(memcg
);
3337 static int compare_thresholds(const void *a
, const void *b
)
3339 const struct mem_cgroup_threshold
*_a
= a
;
3340 const struct mem_cgroup_threshold
*_b
= b
;
3342 if (_a
->threshold
> _b
->threshold
)
3345 if (_a
->threshold
< _b
->threshold
)
3351 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
3353 struct mem_cgroup_eventfd_list
*ev
;
3355 spin_lock(&memcg_oom_lock
);
3357 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
3358 eventfd_signal(ev
->eventfd
, 1);
3360 spin_unlock(&memcg_oom_lock
);
3364 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
3366 struct mem_cgroup
*iter
;
3368 for_each_mem_cgroup_tree(iter
, memcg
)
3369 mem_cgroup_oom_notify_cb(iter
);
3372 static int __mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3373 struct eventfd_ctx
*eventfd
, const char *args
, enum res_type type
)
3375 struct mem_cgroup_thresholds
*thresholds
;
3376 struct mem_cgroup_threshold_ary
*new;
3377 unsigned long threshold
;
3378 unsigned long usage
;
3381 ret
= page_counter_memparse(args
, "-1", &threshold
);
3385 mutex_lock(&memcg
->thresholds_lock
);
3388 thresholds
= &memcg
->thresholds
;
3389 usage
= mem_cgroup_usage(memcg
, false);
3390 } else if (type
== _MEMSWAP
) {
3391 thresholds
= &memcg
->memsw_thresholds
;
3392 usage
= mem_cgroup_usage(memcg
, true);
3396 /* Check if a threshold crossed before adding a new one */
3397 if (thresholds
->primary
)
3398 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3400 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3402 /* Allocate memory for new array of thresholds */
3403 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3411 /* Copy thresholds (if any) to new array */
3412 if (thresholds
->primary
) {
3413 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3414 sizeof(struct mem_cgroup_threshold
));
3417 /* Add new threshold */
3418 new->entries
[size
- 1].eventfd
= eventfd
;
3419 new->entries
[size
- 1].threshold
= threshold
;
3421 /* Sort thresholds. Registering of new threshold isn't time-critical */
3422 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3423 compare_thresholds
, NULL
);
3425 /* Find current threshold */
3426 new->current_threshold
= -1;
3427 for (i
= 0; i
< size
; i
++) {
3428 if (new->entries
[i
].threshold
<= usage
) {
3430 * new->current_threshold will not be used until
3431 * rcu_assign_pointer(), so it's safe to increment
3434 ++new->current_threshold
;
3439 /* Free old spare buffer and save old primary buffer as spare */
3440 kfree(thresholds
->spare
);
3441 thresholds
->spare
= thresholds
->primary
;
3443 rcu_assign_pointer(thresholds
->primary
, new);
3445 /* To be sure that nobody uses thresholds */
3449 mutex_unlock(&memcg
->thresholds_lock
);
3454 static int mem_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3455 struct eventfd_ctx
*eventfd
, const char *args
)
3457 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEM
);
3460 static int memsw_cgroup_usage_register_event(struct mem_cgroup
*memcg
,
3461 struct eventfd_ctx
*eventfd
, const char *args
)
3463 return __mem_cgroup_usage_register_event(memcg
, eventfd
, args
, _MEMSWAP
);
3466 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3467 struct eventfd_ctx
*eventfd
, enum res_type type
)
3469 struct mem_cgroup_thresholds
*thresholds
;
3470 struct mem_cgroup_threshold_ary
*new;
3471 unsigned long usage
;
3474 mutex_lock(&memcg
->thresholds_lock
);
3477 thresholds
= &memcg
->thresholds
;
3478 usage
= mem_cgroup_usage(memcg
, false);
3479 } else if (type
== _MEMSWAP
) {
3480 thresholds
= &memcg
->memsw_thresholds
;
3481 usage
= mem_cgroup_usage(memcg
, true);
3485 if (!thresholds
->primary
)
3488 /* Check if a threshold crossed before removing */
3489 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3491 /* Calculate new number of threshold */
3493 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3494 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3498 new = thresholds
->spare
;
3500 /* Set thresholds array to NULL if we don't have thresholds */
3509 /* Copy thresholds and find current threshold */
3510 new->current_threshold
= -1;
3511 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3512 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3515 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3516 if (new->entries
[j
].threshold
<= usage
) {
3518 * new->current_threshold will not be used
3519 * until rcu_assign_pointer(), so it's safe to increment
3522 ++new->current_threshold
;
3528 /* Swap primary and spare array */
3529 thresholds
->spare
= thresholds
->primary
;
3530 /* If all events are unregistered, free the spare array */
3532 kfree(thresholds
->spare
);
3533 thresholds
->spare
= NULL
;
3536 rcu_assign_pointer(thresholds
->primary
, new);
3538 /* To be sure that nobody uses thresholds */
3541 mutex_unlock(&memcg
->thresholds_lock
);
3544 static void mem_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3545 struct eventfd_ctx
*eventfd
)
3547 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEM
);
3550 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup
*memcg
,
3551 struct eventfd_ctx
*eventfd
)
3553 return __mem_cgroup_usage_unregister_event(memcg
, eventfd
, _MEMSWAP
);
3556 static int mem_cgroup_oom_register_event(struct mem_cgroup
*memcg
,
3557 struct eventfd_ctx
*eventfd
, const char *args
)
3559 struct mem_cgroup_eventfd_list
*event
;
3561 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
3565 spin_lock(&memcg_oom_lock
);
3567 event
->eventfd
= eventfd
;
3568 list_add(&event
->list
, &memcg
->oom_notify
);
3570 /* already in OOM ? */
3571 if (memcg
->under_oom
)
3572 eventfd_signal(eventfd
, 1);
3573 spin_unlock(&memcg_oom_lock
);
3578 static void mem_cgroup_oom_unregister_event(struct mem_cgroup
*memcg
,
3579 struct eventfd_ctx
*eventfd
)
3581 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
3583 spin_lock(&memcg_oom_lock
);
3585 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
3586 if (ev
->eventfd
== eventfd
) {
3587 list_del(&ev
->list
);
3592 spin_unlock(&memcg_oom_lock
);
3595 static int mem_cgroup_oom_control_read(struct seq_file
*sf
, void *v
)
3597 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(sf
));
3599 seq_printf(sf
, "oom_kill_disable %d\n", memcg
->oom_kill_disable
);
3600 seq_printf(sf
, "under_oom %d\n", (bool)memcg
->under_oom
);
3604 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
3605 struct cftype
*cft
, u64 val
)
3607 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3609 /* cannot set to root cgroup and only 0 and 1 are allowed */
3610 if (!css
->parent
|| !((val
== 0) || (val
== 1)))
3613 memcg
->oom_kill_disable
= val
;
3615 memcg_oom_recover(memcg
);
3620 #ifdef CONFIG_MEMCG_KMEM
3621 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3625 ret
= memcg_propagate_kmem(memcg
);
3629 return mem_cgroup_sockets_init(memcg
, ss
);
3632 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3634 struct cgroup_subsys_state
*css
;
3635 struct mem_cgroup
*parent
, *child
;
3638 if (!memcg
->kmem_acct_active
)
3642 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3643 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3644 * guarantees no cache will be created for this cgroup after we are
3645 * done (see memcg_create_kmem_cache()).
3647 memcg
->kmem_acct_active
= false;
3649 memcg_deactivate_kmem_caches(memcg
);
3651 kmemcg_id
= memcg
->kmemcg_id
;
3652 BUG_ON(kmemcg_id
< 0);
3654 parent
= parent_mem_cgroup(memcg
);
3656 parent
= root_mem_cgroup
;
3659 * Change kmemcg_id of this cgroup and all its descendants to the
3660 * parent's id, and then move all entries from this cgroup's list_lrus
3661 * to ones of the parent. After we have finished, all list_lrus
3662 * corresponding to this cgroup are guaranteed to remain empty. The
3663 * ordering is imposed by list_lru_node->lock taken by
3664 * memcg_drain_all_list_lrus().
3666 css_for_each_descendant_pre(css
, &memcg
->css
) {
3667 child
= mem_cgroup_from_css(css
);
3668 BUG_ON(child
->kmemcg_id
!= kmemcg_id
);
3669 child
->kmemcg_id
= parent
->kmemcg_id
;
3670 if (!memcg
->use_hierarchy
)
3673 memcg_drain_all_list_lrus(kmemcg_id
, parent
->kmemcg_id
);
3675 memcg_free_cache_id(kmemcg_id
);
3678 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3680 if (memcg
->kmem_acct_activated
) {
3681 memcg_destroy_kmem_caches(memcg
);
3682 static_key_slow_dec(&memcg_kmem_enabled_key
);
3683 WARN_ON(page_counter_read(&memcg
->kmem
));
3685 mem_cgroup_sockets_destroy(memcg
);
3688 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
3693 static void memcg_deactivate_kmem(struct mem_cgroup
*memcg
)
3697 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
3702 #ifdef CONFIG_CGROUP_WRITEBACK
3704 struct list_head
*mem_cgroup_cgwb_list(struct mem_cgroup
*memcg
)
3706 return &memcg
->cgwb_list
;
3709 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3711 return wb_domain_init(&memcg
->cgwb_domain
, gfp
);
3714 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3716 wb_domain_exit(&memcg
->cgwb_domain
);
3719 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3721 wb_domain_size_changed(&memcg
->cgwb_domain
);
3724 struct wb_domain
*mem_cgroup_wb_domain(struct bdi_writeback
*wb
)
3726 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3728 if (!memcg
->css
.parent
)
3731 return &memcg
->cgwb_domain
;
3735 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3736 * @wb: bdi_writeback in question
3737 * @pavail: out parameter for number of available pages
3738 * @pdirty: out parameter for number of dirty pages
3739 * @pwriteback: out parameter for number of pages under writeback
3741 * Determine the numbers of available, dirty, and writeback pages in @wb's
3742 * memcg. Dirty and writeback are self-explanatory. Available is a bit
3745 * A memcg's headroom is "min(max, high) - used". The available memory is
3746 * calculated as the lowest headroom of itself and the ancestors plus the
3747 * number of pages already being used for file pages. Note that this
3748 * doesn't consider the actual amount of available memory in the system.
3749 * The caller should further cap *@pavail accordingly.
3751 void mem_cgroup_wb_stats(struct bdi_writeback
*wb
, unsigned long *pavail
,
3752 unsigned long *pdirty
, unsigned long *pwriteback
)
3754 struct mem_cgroup
*memcg
= mem_cgroup_from_css(wb
->memcg_css
);
3755 struct mem_cgroup
*parent
;
3756 unsigned long head_room
= PAGE_COUNTER_MAX
;
3757 unsigned long file_pages
;
3759 *pdirty
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_DIRTY
);
3761 /* this should eventually include NR_UNSTABLE_NFS */
3762 *pwriteback
= mem_cgroup_read_stat(memcg
, MEM_CGROUP_STAT_WRITEBACK
);
3764 file_pages
= mem_cgroup_nr_lru_pages(memcg
, (1 << LRU_INACTIVE_FILE
) |
3765 (1 << LRU_ACTIVE_FILE
));
3766 while ((parent
= parent_mem_cgroup(memcg
))) {
3767 unsigned long ceiling
= min(memcg
->memory
.limit
, memcg
->high
);
3768 unsigned long used
= page_counter_read(&memcg
->memory
);
3770 head_room
= min(head_room
, ceiling
- min(ceiling
, used
));
3774 *pavail
= file_pages
+ head_room
;
3777 #else /* CONFIG_CGROUP_WRITEBACK */
3779 static int memcg_wb_domain_init(struct mem_cgroup
*memcg
, gfp_t gfp
)
3784 static void memcg_wb_domain_exit(struct mem_cgroup
*memcg
)
3788 static void memcg_wb_domain_size_changed(struct mem_cgroup
*memcg
)
3792 #endif /* CONFIG_CGROUP_WRITEBACK */
3795 * DO NOT USE IN NEW FILES.
3797 * "cgroup.event_control" implementation.
3799 * This is way over-engineered. It tries to support fully configurable
3800 * events for each user. Such level of flexibility is completely
3801 * unnecessary especially in the light of the planned unified hierarchy.
3803 * Please deprecate this and replace with something simpler if at all
3808 * Unregister event and free resources.
3810 * Gets called from workqueue.
3812 static void memcg_event_remove(struct work_struct
*work
)
3814 struct mem_cgroup_event
*event
=
3815 container_of(work
, struct mem_cgroup_event
, remove
);
3816 struct mem_cgroup
*memcg
= event
->memcg
;
3818 remove_wait_queue(event
->wqh
, &event
->wait
);
3820 event
->unregister_event(memcg
, event
->eventfd
);
3822 /* Notify userspace the event is going away. */
3823 eventfd_signal(event
->eventfd
, 1);
3825 eventfd_ctx_put(event
->eventfd
);
3827 css_put(&memcg
->css
);
3831 * Gets called on POLLHUP on eventfd when user closes it.
3833 * Called with wqh->lock held and interrupts disabled.
3835 static int memcg_event_wake(wait_queue_t
*wait
, unsigned mode
,
3836 int sync
, void *key
)
3838 struct mem_cgroup_event
*event
=
3839 container_of(wait
, struct mem_cgroup_event
, wait
);
3840 struct mem_cgroup
*memcg
= event
->memcg
;
3841 unsigned long flags
= (unsigned long)key
;
3843 if (flags
& POLLHUP
) {
3845 * If the event has been detached at cgroup removal, we
3846 * can simply return knowing the other side will cleanup
3849 * We can't race against event freeing since the other
3850 * side will require wqh->lock via remove_wait_queue(),
3853 spin_lock(&memcg
->event_list_lock
);
3854 if (!list_empty(&event
->list
)) {
3855 list_del_init(&event
->list
);
3857 * We are in atomic context, but cgroup_event_remove()
3858 * may sleep, so we have to call it in workqueue.
3860 schedule_work(&event
->remove
);
3862 spin_unlock(&memcg
->event_list_lock
);
3868 static void memcg_event_ptable_queue_proc(struct file
*file
,
3869 wait_queue_head_t
*wqh
, poll_table
*pt
)
3871 struct mem_cgroup_event
*event
=
3872 container_of(pt
, struct mem_cgroup_event
, pt
);
3875 add_wait_queue(wqh
, &event
->wait
);
3879 * DO NOT USE IN NEW FILES.
3881 * Parse input and register new cgroup event handler.
3883 * Input must be in format '<event_fd> <control_fd> <args>'.
3884 * Interpretation of args is defined by control file implementation.
3886 static ssize_t
memcg_write_event_control(struct kernfs_open_file
*of
,
3887 char *buf
, size_t nbytes
, loff_t off
)
3889 struct cgroup_subsys_state
*css
= of_css(of
);
3890 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
3891 struct mem_cgroup_event
*event
;
3892 struct cgroup_subsys_state
*cfile_css
;
3893 unsigned int efd
, cfd
;
3900 buf
= strstrip(buf
);
3902 efd
= simple_strtoul(buf
, &endp
, 10);
3907 cfd
= simple_strtoul(buf
, &endp
, 10);
3908 if ((*endp
!= ' ') && (*endp
!= '\0'))
3912 event
= kzalloc(sizeof(*event
), GFP_KERNEL
);
3916 event
->memcg
= memcg
;
3917 INIT_LIST_HEAD(&event
->list
);
3918 init_poll_funcptr(&event
->pt
, memcg_event_ptable_queue_proc
);
3919 init_waitqueue_func_entry(&event
->wait
, memcg_event_wake
);
3920 INIT_WORK(&event
->remove
, memcg_event_remove
);
3928 event
->eventfd
= eventfd_ctx_fileget(efile
.file
);
3929 if (IS_ERR(event
->eventfd
)) {
3930 ret
= PTR_ERR(event
->eventfd
);
3937 goto out_put_eventfd
;
3940 /* the process need read permission on control file */
3941 /* AV: shouldn't we check that it's been opened for read instead? */
3942 ret
= inode_permission(file_inode(cfile
.file
), MAY_READ
);
3947 * Determine the event callbacks and set them in @event. This used
3948 * to be done via struct cftype but cgroup core no longer knows
3949 * about these events. The following is crude but the whole thing
3950 * is for compatibility anyway.
3952 * DO NOT ADD NEW FILES.
3954 name
= cfile
.file
->f_path
.dentry
->d_name
.name
;
3956 if (!strcmp(name
, "memory.usage_in_bytes")) {
3957 event
->register_event
= mem_cgroup_usage_register_event
;
3958 event
->unregister_event
= mem_cgroup_usage_unregister_event
;
3959 } else if (!strcmp(name
, "memory.oom_control")) {
3960 event
->register_event
= mem_cgroup_oom_register_event
;
3961 event
->unregister_event
= mem_cgroup_oom_unregister_event
;
3962 } else if (!strcmp(name
, "memory.pressure_level")) {
3963 event
->register_event
= vmpressure_register_event
;
3964 event
->unregister_event
= vmpressure_unregister_event
;
3965 } else if (!strcmp(name
, "memory.memsw.usage_in_bytes")) {
3966 event
->register_event
= memsw_cgroup_usage_register_event
;
3967 event
->unregister_event
= memsw_cgroup_usage_unregister_event
;
3974 * Verify @cfile should belong to @css. Also, remaining events are
3975 * automatically removed on cgroup destruction but the removal is
3976 * asynchronous, so take an extra ref on @css.
3978 cfile_css
= css_tryget_online_from_dir(cfile
.file
->f_path
.dentry
->d_parent
,
3979 &memory_cgrp_subsys
);
3981 if (IS_ERR(cfile_css
))
3983 if (cfile_css
!= css
) {
3988 ret
= event
->register_event(memcg
, event
->eventfd
, buf
);
3992 efile
.file
->f_op
->poll(efile
.file
, &event
->pt
);
3994 spin_lock(&memcg
->event_list_lock
);
3995 list_add(&event
->list
, &memcg
->event_list
);
3996 spin_unlock(&memcg
->event_list_lock
);
4008 eventfd_ctx_put(event
->eventfd
);
4017 static struct cftype mem_cgroup_legacy_files
[] = {
4019 .name
= "usage_in_bytes",
4020 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4021 .read_u64
= mem_cgroup_read_u64
,
4024 .name
= "max_usage_in_bytes",
4025 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4026 .write
= mem_cgroup_reset
,
4027 .read_u64
= mem_cgroup_read_u64
,
4030 .name
= "limit_in_bytes",
4031 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4032 .write
= mem_cgroup_write
,
4033 .read_u64
= mem_cgroup_read_u64
,
4036 .name
= "soft_limit_in_bytes",
4037 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4038 .write
= mem_cgroup_write
,
4039 .read_u64
= mem_cgroup_read_u64
,
4043 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4044 .write
= mem_cgroup_reset
,
4045 .read_u64
= mem_cgroup_read_u64
,
4049 .seq_show
= memcg_stat_show
,
4052 .name
= "force_empty",
4053 .write
= mem_cgroup_force_empty_write
,
4056 .name
= "use_hierarchy",
4057 .write_u64
= mem_cgroup_hierarchy_write
,
4058 .read_u64
= mem_cgroup_hierarchy_read
,
4061 .name
= "cgroup.event_control", /* XXX: for compat */
4062 .write
= memcg_write_event_control
,
4063 .flags
= CFTYPE_NO_PREFIX
,
4067 .name
= "swappiness",
4068 .read_u64
= mem_cgroup_swappiness_read
,
4069 .write_u64
= mem_cgroup_swappiness_write
,
4072 .name
= "move_charge_at_immigrate",
4073 .read_u64
= mem_cgroup_move_charge_read
,
4074 .write_u64
= mem_cgroup_move_charge_write
,
4077 .name
= "oom_control",
4078 .seq_show
= mem_cgroup_oom_control_read
,
4079 .write_u64
= mem_cgroup_oom_control_write
,
4080 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4083 .name
= "pressure_level",
4087 .name
= "numa_stat",
4088 .seq_show
= memcg_numa_stat_show
,
4091 #ifdef CONFIG_MEMCG_KMEM
4093 .name
= "kmem.limit_in_bytes",
4094 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
4095 .write
= mem_cgroup_write
,
4096 .read_u64
= mem_cgroup_read_u64
,
4099 .name
= "kmem.usage_in_bytes",
4100 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
4101 .read_u64
= mem_cgroup_read_u64
,
4104 .name
= "kmem.failcnt",
4105 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
4106 .write
= mem_cgroup_reset
,
4107 .read_u64
= mem_cgroup_read_u64
,
4110 .name
= "kmem.max_usage_in_bytes",
4111 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
4112 .write
= mem_cgroup_reset
,
4113 .read_u64
= mem_cgroup_read_u64
,
4115 #ifdef CONFIG_SLABINFO
4117 .name
= "kmem.slabinfo",
4118 .seq_start
= slab_start
,
4119 .seq_next
= slab_next
,
4120 .seq_stop
= slab_stop
,
4121 .seq_show
= memcg_slab_show
,
4125 { }, /* terminate */
4128 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4130 struct mem_cgroup_per_node
*pn
;
4131 struct mem_cgroup_per_zone
*mz
;
4132 int zone
, tmp
= node
;
4134 * This routine is called against possible nodes.
4135 * But it's BUG to call kmalloc() against offline node.
4137 * TODO: this routine can waste much memory for nodes which will
4138 * never be onlined. It's better to use memory hotplug callback
4141 if (!node_state(node
, N_NORMAL_MEMORY
))
4143 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4147 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4148 mz
= &pn
->zoneinfo
[zone
];
4149 lruvec_init(&mz
->lruvec
);
4150 mz
->usage_in_excess
= 0;
4151 mz
->on_tree
= false;
4154 memcg
->nodeinfo
[node
] = pn
;
4158 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
4160 kfree(memcg
->nodeinfo
[node
]);
4163 static struct mem_cgroup
*mem_cgroup_alloc(void)
4165 struct mem_cgroup
*memcg
;
4168 size
= sizeof(struct mem_cgroup
);
4169 size
+= nr_node_ids
* sizeof(struct mem_cgroup_per_node
*);
4171 memcg
= kzalloc(size
, GFP_KERNEL
);
4175 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4179 if (memcg_wb_domain_init(memcg
, GFP_KERNEL
))
4182 spin_lock_init(&memcg
->pcp_counter_lock
);
4186 free_percpu(memcg
->stat
);
4193 * At destroying mem_cgroup, references from swap_cgroup can remain.
4194 * (scanning all at force_empty is too costly...)
4196 * Instead of clearing all references at force_empty, we remember
4197 * the number of reference from swap_cgroup and free mem_cgroup when
4198 * it goes down to 0.
4200 * Removal of cgroup itself succeeds regardless of refs from swap.
4203 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
4207 mem_cgroup_remove_from_trees(memcg
);
4210 free_mem_cgroup_per_zone_info(memcg
, node
);
4212 free_percpu(memcg
->stat
);
4213 memcg_wb_domain_exit(memcg
);
4218 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4220 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
4222 if (!memcg
->memory
.parent
)
4224 return mem_cgroup_from_counter(memcg
->memory
.parent
, memory
);
4226 EXPORT_SYMBOL(parent_mem_cgroup
);
4228 static struct cgroup_subsys_state
* __ref
4229 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
4231 struct mem_cgroup
*memcg
;
4232 long error
= -ENOMEM
;
4235 memcg
= mem_cgroup_alloc();
4237 return ERR_PTR(error
);
4240 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
4244 if (parent_css
== NULL
) {
4245 root_mem_cgroup
= memcg
;
4246 mem_cgroup_root_css
= &memcg
->css
;
4247 page_counter_init(&memcg
->memory
, NULL
);
4248 memcg
->high
= PAGE_COUNTER_MAX
;
4249 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4250 page_counter_init(&memcg
->memsw
, NULL
);
4251 page_counter_init(&memcg
->kmem
, NULL
);
4254 memcg
->last_scanned_node
= MAX_NUMNODES
;
4255 INIT_LIST_HEAD(&memcg
->oom_notify
);
4256 memcg
->move_charge_at_immigrate
= 0;
4257 mutex_init(&memcg
->thresholds_lock
);
4258 spin_lock_init(&memcg
->move_lock
);
4259 vmpressure_init(&memcg
->vmpressure
);
4260 INIT_LIST_HEAD(&memcg
->event_list
);
4261 spin_lock_init(&memcg
->event_list_lock
);
4262 #ifdef CONFIG_MEMCG_KMEM
4263 memcg
->kmemcg_id
= -1;
4265 #ifdef CONFIG_CGROUP_WRITEBACK
4266 INIT_LIST_HEAD(&memcg
->cgwb_list
);
4271 __mem_cgroup_free(memcg
);
4272 return ERR_PTR(error
);
4276 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
4278 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4279 struct mem_cgroup
*parent
= mem_cgroup_from_css(css
->parent
);
4282 if (css
->id
> MEM_CGROUP_ID_MAX
)
4288 mutex_lock(&memcg_create_mutex
);
4290 memcg
->use_hierarchy
= parent
->use_hierarchy
;
4291 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
4292 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
4294 if (parent
->use_hierarchy
) {
4295 page_counter_init(&memcg
->memory
, &parent
->memory
);
4296 memcg
->high
= PAGE_COUNTER_MAX
;
4297 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4298 page_counter_init(&memcg
->memsw
, &parent
->memsw
);
4299 page_counter_init(&memcg
->kmem
, &parent
->kmem
);
4302 * No need to take a reference to the parent because cgroup
4303 * core guarantees its existence.
4306 page_counter_init(&memcg
->memory
, NULL
);
4307 memcg
->high
= PAGE_COUNTER_MAX
;
4308 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4309 page_counter_init(&memcg
->memsw
, NULL
);
4310 page_counter_init(&memcg
->kmem
, NULL
);
4312 * Deeper hierachy with use_hierarchy == false doesn't make
4313 * much sense so let cgroup subsystem know about this
4314 * unfortunate state in our controller.
4316 if (parent
!= root_mem_cgroup
)
4317 memory_cgrp_subsys
.broken_hierarchy
= true;
4319 mutex_unlock(&memcg_create_mutex
);
4321 ret
= memcg_init_kmem(memcg
, &memory_cgrp_subsys
);
4326 * Make sure the memcg is initialized: mem_cgroup_iter()
4327 * orders reading memcg->initialized against its callers
4328 * reading the memcg members.
4330 smp_store_release(&memcg
->initialized
, 1);
4335 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
4337 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4338 struct mem_cgroup_event
*event
, *tmp
;
4341 * Unregister events and notify userspace.
4342 * Notify userspace about cgroup removing only after rmdir of cgroup
4343 * directory to avoid race between userspace and kernelspace.
4345 spin_lock(&memcg
->event_list_lock
);
4346 list_for_each_entry_safe(event
, tmp
, &memcg
->event_list
, list
) {
4347 list_del_init(&event
->list
);
4348 schedule_work(&event
->remove
);
4350 spin_unlock(&memcg
->event_list_lock
);
4352 vmpressure_cleanup(&memcg
->vmpressure
);
4354 memcg_deactivate_kmem(memcg
);
4356 wb_memcg_offline(memcg
);
4359 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
4361 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4363 memcg_destroy_kmem(memcg
);
4364 __mem_cgroup_free(memcg
);
4368 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4369 * @css: the target css
4371 * Reset the states of the mem_cgroup associated with @css. This is
4372 * invoked when the userland requests disabling on the default hierarchy
4373 * but the memcg is pinned through dependency. The memcg should stop
4374 * applying policies and should revert to the vanilla state as it may be
4375 * made visible again.
4377 * The current implementation only resets the essential configurations.
4378 * This needs to be expanded to cover all the visible parts.
4380 static void mem_cgroup_css_reset(struct cgroup_subsys_state
*css
)
4382 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4384 mem_cgroup_resize_limit(memcg
, PAGE_COUNTER_MAX
);
4385 mem_cgroup_resize_memsw_limit(memcg
, PAGE_COUNTER_MAX
);
4386 memcg_update_kmem_limit(memcg
, PAGE_COUNTER_MAX
);
4388 memcg
->high
= PAGE_COUNTER_MAX
;
4389 memcg
->soft_limit
= PAGE_COUNTER_MAX
;
4390 memcg_wb_domain_size_changed(memcg
);
4394 /* Handlers for move charge at task migration. */
4395 static int mem_cgroup_do_precharge(unsigned long count
)
4399 /* Try a single bulk charge without reclaim first */
4400 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_WAIT
, count
);
4402 mc
.precharge
+= count
;
4405 if (ret
== -EINTR
) {
4406 cancel_charge(root_mem_cgroup
, count
);
4410 /* Try charges one by one with reclaim */
4412 ret
= try_charge(mc
.to
, GFP_KERNEL
& ~__GFP_NORETRY
, 1);
4414 * In case of failure, any residual charges against
4415 * mc.to will be dropped by mem_cgroup_clear_mc()
4416 * later on. However, cancel any charges that are
4417 * bypassed to root right away or they'll be lost.
4420 cancel_charge(root_mem_cgroup
, 1);
4430 * get_mctgt_type - get target type of moving charge
4431 * @vma: the vma the pte to be checked belongs
4432 * @addr: the address corresponding to the pte to be checked
4433 * @ptent: the pte to be checked
4434 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4437 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4438 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4439 * move charge. if @target is not NULL, the page is stored in target->page
4440 * with extra refcnt got(Callers should handle it).
4441 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4442 * target for charge migration. if @target is not NULL, the entry is stored
4445 * Called with pte lock held.
4452 enum mc_target_type
{
4458 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4459 unsigned long addr
, pte_t ptent
)
4461 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4463 if (!page
|| !page_mapped(page
))
4465 if (PageAnon(page
)) {
4466 if (!(mc
.flags
& MOVE_ANON
))
4469 if (!(mc
.flags
& MOVE_FILE
))
4472 if (!get_page_unless_zero(page
))
4479 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4480 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4482 struct page
*page
= NULL
;
4483 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4485 if (!(mc
.flags
& MOVE_ANON
) || non_swap_entry(ent
))
4488 * Because lookup_swap_cache() updates some statistics counter,
4489 * we call find_get_page() with swapper_space directly.
4491 page
= find_get_page(swap_address_space(ent
), ent
.val
);
4492 if (do_swap_account
)
4493 entry
->val
= ent
.val
;
4498 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4499 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4505 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4506 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4508 struct page
*page
= NULL
;
4509 struct address_space
*mapping
;
4512 if (!vma
->vm_file
) /* anonymous vma */
4514 if (!(mc
.flags
& MOVE_FILE
))
4517 mapping
= vma
->vm_file
->f_mapping
;
4518 pgoff
= linear_page_index(vma
, addr
);
4520 /* page is moved even if it's not RSS of this task(page-faulted). */
4522 /* shmem/tmpfs may report page out on swap: account for that too. */
4523 if (shmem_mapping(mapping
)) {
4524 page
= find_get_entry(mapping
, pgoff
);
4525 if (radix_tree_exceptional_entry(page
)) {
4526 swp_entry_t swp
= radix_to_swp_entry(page
);
4527 if (do_swap_account
)
4529 page
= find_get_page(swap_address_space(swp
), swp
.val
);
4532 page
= find_get_page(mapping
, pgoff
);
4534 page
= find_get_page(mapping
, pgoff
);
4540 * mem_cgroup_move_account - move account of the page
4542 * @nr_pages: number of regular pages (>1 for huge pages)
4543 * @from: mem_cgroup which the page is moved from.
4544 * @to: mem_cgroup which the page is moved to. @from != @to.
4546 * The caller must confirm following.
4547 * - page is not on LRU (isolate_page() is useful.)
4548 * - compound_lock is held when nr_pages > 1
4550 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4553 static int mem_cgroup_move_account(struct page
*page
,
4554 unsigned int nr_pages
,
4555 struct mem_cgroup
*from
,
4556 struct mem_cgroup
*to
)
4558 unsigned long flags
;
4562 VM_BUG_ON(from
== to
);
4563 VM_BUG_ON_PAGE(PageLRU(page
), page
);
4565 * The page is isolated from LRU. So, collapse function
4566 * will not handle this page. But page splitting can happen.
4567 * Do this check under compound_page_lock(). The caller should
4571 if (nr_pages
> 1 && !PageTransHuge(page
))
4575 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4576 * of its source page while we change it: page migration takes
4577 * both pages off the LRU, but page cache replacement doesn't.
4579 if (!trylock_page(page
))
4583 if (page
->mem_cgroup
!= from
)
4586 anon
= PageAnon(page
);
4588 spin_lock_irqsave(&from
->move_lock
, flags
);
4590 if (!anon
&& page_mapped(page
)) {
4591 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4593 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
],
4598 * move_lock grabbed above and caller set from->moving_account, so
4599 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4600 * So mapping should be stable for dirty pages.
4602 if (!anon
&& PageDirty(page
)) {
4603 struct address_space
*mapping
= page_mapping(page
);
4605 if (mapping_cap_account_dirty(mapping
)) {
4606 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4608 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_DIRTY
],
4613 if (PageWriteback(page
)) {
4614 __this_cpu_sub(from
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4616 __this_cpu_add(to
->stat
->count
[MEM_CGROUP_STAT_WRITEBACK
],
4621 * It is safe to change page->mem_cgroup here because the page
4622 * is referenced, charged, and isolated - we can't race with
4623 * uncharging, charging, migration, or LRU putback.
4626 /* caller should have done css_get */
4627 page
->mem_cgroup
= to
;
4628 spin_unlock_irqrestore(&from
->move_lock
, flags
);
4632 local_irq_disable();
4633 mem_cgroup_charge_statistics(to
, page
, nr_pages
);
4634 memcg_check_events(to
, page
);
4635 mem_cgroup_charge_statistics(from
, page
, -nr_pages
);
4636 memcg_check_events(from
, page
);
4644 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
4645 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4647 struct page
*page
= NULL
;
4648 enum mc_target_type ret
= MC_TARGET_NONE
;
4649 swp_entry_t ent
= { .val
= 0 };
4651 if (pte_present(ptent
))
4652 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4653 else if (is_swap_pte(ptent
))
4654 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4655 else if (pte_none(ptent
))
4656 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4658 if (!page
&& !ent
.val
)
4662 * Do only loose check w/o serialization.
4663 * mem_cgroup_move_account() checks the page is valid or
4664 * not under LRU exclusion.
4666 if (page
->mem_cgroup
== mc
.from
) {
4667 ret
= MC_TARGET_PAGE
;
4669 target
->page
= page
;
4671 if (!ret
|| !target
)
4674 /* There is a swap entry and a page doesn't exist or isn't charged */
4675 if (ent
.val
&& !ret
&&
4676 mem_cgroup_id(mc
.from
) == lookup_swap_cgroup_id(ent
)) {
4677 ret
= MC_TARGET_SWAP
;
4684 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4686 * We don't consider swapping or file mapped pages because THP does not
4687 * support them for now.
4688 * Caller should make sure that pmd_trans_huge(pmd) is true.
4690 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4691 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4693 struct page
*page
= NULL
;
4694 enum mc_target_type ret
= MC_TARGET_NONE
;
4696 page
= pmd_page(pmd
);
4697 VM_BUG_ON_PAGE(!page
|| !PageHead(page
), page
);
4698 if (!(mc
.flags
& MOVE_ANON
))
4700 if (page
->mem_cgroup
== mc
.from
) {
4701 ret
= MC_TARGET_PAGE
;
4704 target
->page
= page
;
4710 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
4711 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
4713 return MC_TARGET_NONE
;
4717 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4718 unsigned long addr
, unsigned long end
,
4719 struct mm_walk
*walk
)
4721 struct vm_area_struct
*vma
= walk
->vma
;
4725 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4726 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
4727 mc
.precharge
+= HPAGE_PMD_NR
;
4732 if (pmd_trans_unstable(pmd
))
4734 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4735 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4736 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
4737 mc
.precharge
++; /* increment precharge temporarily */
4738 pte_unmap_unlock(pte
- 1, ptl
);
4744 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4746 unsigned long precharge
;
4748 struct mm_walk mem_cgroup_count_precharge_walk
= {
4749 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4752 down_read(&mm
->mmap_sem
);
4753 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk
);
4754 up_read(&mm
->mmap_sem
);
4756 precharge
= mc
.precharge
;
4762 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4764 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4766 VM_BUG_ON(mc
.moving_task
);
4767 mc
.moving_task
= current
;
4768 return mem_cgroup_do_precharge(precharge
);
4771 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4772 static void __mem_cgroup_clear_mc(void)
4774 struct mem_cgroup
*from
= mc
.from
;
4775 struct mem_cgroup
*to
= mc
.to
;
4777 /* we must uncharge all the leftover precharges from mc.to */
4779 cancel_charge(mc
.to
, mc
.precharge
);
4783 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4784 * we must uncharge here.
4786 if (mc
.moved_charge
) {
4787 cancel_charge(mc
.from
, mc
.moved_charge
);
4788 mc
.moved_charge
= 0;
4790 /* we must fixup refcnts and charges */
4791 if (mc
.moved_swap
) {
4792 /* uncharge swap account from the old cgroup */
4793 if (!mem_cgroup_is_root(mc
.from
))
4794 page_counter_uncharge(&mc
.from
->memsw
, mc
.moved_swap
);
4797 * we charged both to->memory and to->memsw, so we
4798 * should uncharge to->memory.
4800 if (!mem_cgroup_is_root(mc
.to
))
4801 page_counter_uncharge(&mc
.to
->memory
, mc
.moved_swap
);
4803 css_put_many(&mc
.from
->css
, mc
.moved_swap
);
4805 /* we've already done css_get(mc.to) */
4808 memcg_oom_recover(from
);
4809 memcg_oom_recover(to
);
4810 wake_up_all(&mc
.waitq
);
4813 static void mem_cgroup_clear_mc(void)
4816 * we must clear moving_task before waking up waiters at the end of
4819 mc
.moving_task
= NULL
;
4820 __mem_cgroup_clear_mc();
4821 spin_lock(&mc
.lock
);
4824 spin_unlock(&mc
.lock
);
4827 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
4828 struct cgroup_taskset
*tset
)
4830 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4831 struct mem_cgroup
*from
;
4832 struct task_struct
*p
;
4833 struct mm_struct
*mm
;
4834 unsigned long move_flags
;
4838 * We are now commited to this value whatever it is. Changes in this
4839 * tunable will only affect upcoming migrations, not the current one.
4840 * So we need to save it, and keep it going.
4842 move_flags
= READ_ONCE(memcg
->move_charge_at_immigrate
);
4846 p
= cgroup_taskset_first(tset
);
4847 from
= mem_cgroup_from_task(p
);
4849 VM_BUG_ON(from
== memcg
);
4851 mm
= get_task_mm(p
);
4854 /* We move charges only when we move a owner of the mm */
4855 if (mm
->owner
== p
) {
4858 VM_BUG_ON(mc
.precharge
);
4859 VM_BUG_ON(mc
.moved_charge
);
4860 VM_BUG_ON(mc
.moved_swap
);
4862 spin_lock(&mc
.lock
);
4865 mc
.flags
= move_flags
;
4866 spin_unlock(&mc
.lock
);
4867 /* We set mc.moving_task later */
4869 ret
= mem_cgroup_precharge_mc(mm
);
4871 mem_cgroup_clear_mc();
4877 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
4878 struct cgroup_taskset
*tset
)
4881 mem_cgroup_clear_mc();
4884 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4885 unsigned long addr
, unsigned long end
,
4886 struct mm_walk
*walk
)
4889 struct vm_area_struct
*vma
= walk
->vma
;
4892 enum mc_target_type target_type
;
4893 union mc_target target
;
4897 * We don't take compound_lock() here but no race with splitting thp
4899 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4900 * under splitting, which means there's no concurrent thp split,
4901 * - if another thread runs into split_huge_page() just after we
4902 * entered this if-block, the thread must wait for page table lock
4903 * to be unlocked in __split_huge_page_splitting(), where the main
4904 * part of thp split is not executed yet.
4906 if (pmd_trans_huge_lock(pmd
, vma
, &ptl
) == 1) {
4907 if (mc
.precharge
< HPAGE_PMD_NR
) {
4911 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
4912 if (target_type
== MC_TARGET_PAGE
) {
4914 if (!isolate_lru_page(page
)) {
4915 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
4917 mc
.precharge
-= HPAGE_PMD_NR
;
4918 mc
.moved_charge
+= HPAGE_PMD_NR
;
4920 putback_lru_page(page
);
4928 if (pmd_trans_unstable(pmd
))
4931 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4932 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4933 pte_t ptent
= *(pte
++);
4939 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
4940 case MC_TARGET_PAGE
:
4942 if (isolate_lru_page(page
))
4944 if (!mem_cgroup_move_account(page
, 1, mc
.from
, mc
.to
)) {
4946 /* we uncharge from mc.from later. */
4949 putback_lru_page(page
);
4950 put
: /* get_mctgt_type() gets the page */
4953 case MC_TARGET_SWAP
:
4955 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
4957 /* we fixup refcnts and charges later. */
4965 pte_unmap_unlock(pte
- 1, ptl
);
4970 * We have consumed all precharges we got in can_attach().
4971 * We try charge one by one, but don't do any additional
4972 * charges to mc.to if we have failed in charge once in attach()
4975 ret
= mem_cgroup_do_precharge(1);
4983 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4985 struct mm_walk mem_cgroup_move_charge_walk
= {
4986 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4990 lru_add_drain_all();
4992 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4993 * move_lock while we're moving its pages to another memcg.
4994 * Then wait for already started RCU-only updates to finish.
4996 atomic_inc(&mc
.from
->moving_account
);
4999 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
5001 * Someone who are holding the mmap_sem might be waiting in
5002 * waitq. So we cancel all extra charges, wake up all waiters,
5003 * and retry. Because we cancel precharges, we might not be able
5004 * to move enough charges, but moving charge is a best-effort
5005 * feature anyway, so it wouldn't be a big problem.
5007 __mem_cgroup_clear_mc();
5012 * When we have consumed all precharges and failed in doing
5013 * additional charge, the page walk just aborts.
5015 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk
);
5016 up_read(&mm
->mmap_sem
);
5017 atomic_dec(&mc
.from
->moving_account
);
5020 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5021 struct cgroup_taskset
*tset
)
5023 struct task_struct
*p
= cgroup_taskset_first(tset
);
5024 struct mm_struct
*mm
= get_task_mm(p
);
5028 mem_cgroup_move_charge(mm
);
5032 mem_cgroup_clear_mc();
5034 #else /* !CONFIG_MMU */
5035 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
5036 struct cgroup_taskset
*tset
)
5040 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
5041 struct cgroup_taskset
*tset
)
5044 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
5045 struct cgroup_taskset
*tset
)
5051 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5052 * to verify whether we're attached to the default hierarchy on each mount
5055 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
5058 * use_hierarchy is forced on the default hierarchy. cgroup core
5059 * guarantees that @root doesn't have any children, so turning it
5060 * on for the root memcg is enough.
5062 if (cgroup_on_dfl(root_css
->cgroup
))
5063 root_mem_cgroup
->use_hierarchy
= true;
5065 root_mem_cgroup
->use_hierarchy
= false;
5068 static u64
memory_current_read(struct cgroup_subsys_state
*css
,
5071 return mem_cgroup_usage(mem_cgroup_from_css(css
), false);
5074 static int memory_low_show(struct seq_file
*m
, void *v
)
5076 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5077 unsigned long low
= READ_ONCE(memcg
->low
);
5079 if (low
== PAGE_COUNTER_MAX
)
5080 seq_puts(m
, "max\n");
5082 seq_printf(m
, "%llu\n", (u64
)low
* PAGE_SIZE
);
5087 static ssize_t
memory_low_write(struct kernfs_open_file
*of
,
5088 char *buf
, size_t nbytes
, loff_t off
)
5090 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5094 buf
= strstrip(buf
);
5095 err
= page_counter_memparse(buf
, "max", &low
);
5104 static int memory_high_show(struct seq_file
*m
, void *v
)
5106 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5107 unsigned long high
= READ_ONCE(memcg
->high
);
5109 if (high
== PAGE_COUNTER_MAX
)
5110 seq_puts(m
, "max\n");
5112 seq_printf(m
, "%llu\n", (u64
)high
* PAGE_SIZE
);
5117 static ssize_t
memory_high_write(struct kernfs_open_file
*of
,
5118 char *buf
, size_t nbytes
, loff_t off
)
5120 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5124 buf
= strstrip(buf
);
5125 err
= page_counter_memparse(buf
, "max", &high
);
5131 memcg_wb_domain_size_changed(memcg
);
5135 static int memory_max_show(struct seq_file
*m
, void *v
)
5137 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5138 unsigned long max
= READ_ONCE(memcg
->memory
.limit
);
5140 if (max
== PAGE_COUNTER_MAX
)
5141 seq_puts(m
, "max\n");
5143 seq_printf(m
, "%llu\n", (u64
)max
* PAGE_SIZE
);
5148 static ssize_t
memory_max_write(struct kernfs_open_file
*of
,
5149 char *buf
, size_t nbytes
, loff_t off
)
5151 struct mem_cgroup
*memcg
= mem_cgroup_from_css(of_css(of
));
5155 buf
= strstrip(buf
);
5156 err
= page_counter_memparse(buf
, "max", &max
);
5160 err
= mem_cgroup_resize_limit(memcg
, max
);
5164 memcg_wb_domain_size_changed(memcg
);
5168 static int memory_events_show(struct seq_file
*m
, void *v
)
5170 struct mem_cgroup
*memcg
= mem_cgroup_from_css(seq_css(m
));
5172 seq_printf(m
, "low %lu\n", mem_cgroup_read_events(memcg
, MEMCG_LOW
));
5173 seq_printf(m
, "high %lu\n", mem_cgroup_read_events(memcg
, MEMCG_HIGH
));
5174 seq_printf(m
, "max %lu\n", mem_cgroup_read_events(memcg
, MEMCG_MAX
));
5175 seq_printf(m
, "oom %lu\n", mem_cgroup_read_events(memcg
, MEMCG_OOM
));
5180 static struct cftype memory_files
[] = {
5183 .read_u64
= memory_current_read
,
5187 .flags
= CFTYPE_NOT_ON_ROOT
,
5188 .seq_show
= memory_low_show
,
5189 .write
= memory_low_write
,
5193 .flags
= CFTYPE_NOT_ON_ROOT
,
5194 .seq_show
= memory_high_show
,
5195 .write
= memory_high_write
,
5199 .flags
= CFTYPE_NOT_ON_ROOT
,
5200 .seq_show
= memory_max_show
,
5201 .write
= memory_max_write
,
5205 .flags
= CFTYPE_NOT_ON_ROOT
,
5206 .seq_show
= memory_events_show
,
5211 struct cgroup_subsys memory_cgrp_subsys
= {
5212 .css_alloc
= mem_cgroup_css_alloc
,
5213 .css_online
= mem_cgroup_css_online
,
5214 .css_offline
= mem_cgroup_css_offline
,
5215 .css_free
= mem_cgroup_css_free
,
5216 .css_reset
= mem_cgroup_css_reset
,
5217 .can_attach
= mem_cgroup_can_attach
,
5218 .cancel_attach
= mem_cgroup_cancel_attach
,
5219 .attach
= mem_cgroup_move_task
,
5220 .bind
= mem_cgroup_bind
,
5221 .dfl_cftypes
= memory_files
,
5222 .legacy_cftypes
= mem_cgroup_legacy_files
,
5227 * mem_cgroup_low - check if memory consumption is below the normal range
5228 * @root: the highest ancestor to consider
5229 * @memcg: the memory cgroup to check
5231 * Returns %true if memory consumption of @memcg, and that of all
5232 * configurable ancestors up to @root, is below the normal range.
5234 bool mem_cgroup_low(struct mem_cgroup
*root
, struct mem_cgroup
*memcg
)
5236 if (mem_cgroup_disabled())
5240 * The toplevel group doesn't have a configurable range, so
5241 * it's never low when looked at directly, and it is not
5242 * considered an ancestor when assessing the hierarchy.
5245 if (memcg
== root_mem_cgroup
)
5248 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5251 while (memcg
!= root
) {
5252 memcg
= parent_mem_cgroup(memcg
);
5254 if (memcg
== root_mem_cgroup
)
5257 if (page_counter_read(&memcg
->memory
) >= memcg
->low
)
5264 * mem_cgroup_try_charge - try charging a page
5265 * @page: page to charge
5266 * @mm: mm context of the victim
5267 * @gfp_mask: reclaim mode
5268 * @memcgp: charged memcg return
5270 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5271 * pages according to @gfp_mask if necessary.
5273 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5274 * Otherwise, an error code is returned.
5276 * After page->mapping has been set up, the caller must finalize the
5277 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5278 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5280 int mem_cgroup_try_charge(struct page
*page
, struct mm_struct
*mm
,
5281 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
5283 struct mem_cgroup
*memcg
= NULL
;
5284 unsigned int nr_pages
= 1;
5287 if (mem_cgroup_disabled())
5290 if (PageSwapCache(page
)) {
5292 * Every swap fault against a single page tries to charge the
5293 * page, bail as early as possible. shmem_unuse() encounters
5294 * already charged pages, too. The USED bit is protected by
5295 * the page lock, which serializes swap cache removal, which
5296 * in turn serializes uncharging.
5298 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
5299 if (page
->mem_cgroup
)
5302 if (do_swap_account
) {
5303 swp_entry_t ent
= { .val
= page_private(page
), };
5304 unsigned short id
= lookup_swap_cgroup_id(ent
);
5307 memcg
= mem_cgroup_from_id(id
);
5308 if (memcg
&& !css_tryget_online(&memcg
->css
))
5314 if (PageTransHuge(page
)) {
5315 nr_pages
<<= compound_order(page
);
5316 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5320 memcg
= get_mem_cgroup_from_mm(mm
);
5322 ret
= try_charge(memcg
, gfp_mask
, nr_pages
);
5324 css_put(&memcg
->css
);
5326 if (ret
== -EINTR
) {
5327 memcg
= root_mem_cgroup
;
5336 * mem_cgroup_commit_charge - commit a page charge
5337 * @page: page to charge
5338 * @memcg: memcg to charge the page to
5339 * @lrucare: page might be on LRU already
5341 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5342 * after page->mapping has been set up. This must happen atomically
5343 * as part of the page instantiation, i.e. under the page table lock
5344 * for anonymous pages, under the page lock for page and swap cache.
5346 * In addition, the page must not be on the LRU during the commit, to
5347 * prevent racing with task migration. If it might be, use @lrucare.
5349 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5351 void mem_cgroup_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
5354 unsigned int nr_pages
= 1;
5356 VM_BUG_ON_PAGE(!page
->mapping
, page
);
5357 VM_BUG_ON_PAGE(PageLRU(page
) && !lrucare
, page
);
5359 if (mem_cgroup_disabled())
5362 * Swap faults will attempt to charge the same page multiple
5363 * times. But reuse_swap_page() might have removed the page
5364 * from swapcache already, so we can't check PageSwapCache().
5369 commit_charge(page
, memcg
, lrucare
);
5371 if (PageTransHuge(page
)) {
5372 nr_pages
<<= compound_order(page
);
5373 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5376 local_irq_disable();
5377 mem_cgroup_charge_statistics(memcg
, page
, nr_pages
);
5378 memcg_check_events(memcg
, page
);
5381 if (do_swap_account
&& PageSwapCache(page
)) {
5382 swp_entry_t entry
= { .val
= page_private(page
) };
5384 * The swap entry might not get freed for a long time,
5385 * let's not wait for it. The page already received a
5386 * memory+swap charge, drop the swap entry duplicate.
5388 mem_cgroup_uncharge_swap(entry
);
5393 * mem_cgroup_cancel_charge - cancel a page charge
5394 * @page: page to charge
5395 * @memcg: memcg to charge the page to
5397 * Cancel a charge transaction started by mem_cgroup_try_charge().
5399 void mem_cgroup_cancel_charge(struct page
*page
, struct mem_cgroup
*memcg
)
5401 unsigned int nr_pages
= 1;
5403 if (mem_cgroup_disabled())
5406 * Swap faults will attempt to charge the same page multiple
5407 * times. But reuse_swap_page() might have removed the page
5408 * from swapcache already, so we can't check PageSwapCache().
5413 if (PageTransHuge(page
)) {
5414 nr_pages
<<= compound_order(page
);
5415 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5418 cancel_charge(memcg
, nr_pages
);
5421 static void uncharge_batch(struct mem_cgroup
*memcg
, unsigned long pgpgout
,
5422 unsigned long nr_anon
, unsigned long nr_file
,
5423 unsigned long nr_huge
, struct page
*dummy_page
)
5425 unsigned long nr_pages
= nr_anon
+ nr_file
;
5426 unsigned long flags
;
5428 if (!mem_cgroup_is_root(memcg
)) {
5429 page_counter_uncharge(&memcg
->memory
, nr_pages
);
5430 if (do_swap_account
)
5431 page_counter_uncharge(&memcg
->memsw
, nr_pages
);
5432 memcg_oom_recover(memcg
);
5435 local_irq_save(flags
);
5436 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_anon
);
5437 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_file
);
5438 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
], nr_huge
);
5439 __this_cpu_add(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
], pgpgout
);
5440 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
5441 memcg_check_events(memcg
, dummy_page
);
5442 local_irq_restore(flags
);
5444 if (!mem_cgroup_is_root(memcg
))
5445 css_put_many(&memcg
->css
, nr_pages
);
5448 static void uncharge_list(struct list_head
*page_list
)
5450 struct mem_cgroup
*memcg
= NULL
;
5451 unsigned long nr_anon
= 0;
5452 unsigned long nr_file
= 0;
5453 unsigned long nr_huge
= 0;
5454 unsigned long pgpgout
= 0;
5455 struct list_head
*next
;
5458 next
= page_list
->next
;
5460 unsigned int nr_pages
= 1;
5462 page
= list_entry(next
, struct page
, lru
);
5463 next
= page
->lru
.next
;
5465 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5466 VM_BUG_ON_PAGE(page_count(page
), page
);
5468 if (!page
->mem_cgroup
)
5472 * Nobody should be changing or seriously looking at
5473 * page->mem_cgroup at this point, we have fully
5474 * exclusive access to the page.
5477 if (memcg
!= page
->mem_cgroup
) {
5479 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5481 pgpgout
= nr_anon
= nr_file
= nr_huge
= 0;
5483 memcg
= page
->mem_cgroup
;
5486 if (PageTransHuge(page
)) {
5487 nr_pages
<<= compound_order(page
);
5488 VM_BUG_ON_PAGE(!PageTransHuge(page
), page
);
5489 nr_huge
+= nr_pages
;
5493 nr_anon
+= nr_pages
;
5495 nr_file
+= nr_pages
;
5497 page
->mem_cgroup
= NULL
;
5500 } while (next
!= page_list
);
5503 uncharge_batch(memcg
, pgpgout
, nr_anon
, nr_file
,
5508 * mem_cgroup_uncharge - uncharge a page
5509 * @page: page to uncharge
5511 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5512 * mem_cgroup_commit_charge().
5514 void mem_cgroup_uncharge(struct page
*page
)
5516 if (mem_cgroup_disabled())
5519 /* Don't touch page->lru of any random page, pre-check: */
5520 if (!page
->mem_cgroup
)
5523 INIT_LIST_HEAD(&page
->lru
);
5524 uncharge_list(&page
->lru
);
5528 * mem_cgroup_uncharge_list - uncharge a list of page
5529 * @page_list: list of pages to uncharge
5531 * Uncharge a list of pages previously charged with
5532 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5534 void mem_cgroup_uncharge_list(struct list_head
*page_list
)
5536 if (mem_cgroup_disabled())
5539 if (!list_empty(page_list
))
5540 uncharge_list(page_list
);
5544 * mem_cgroup_migrate - migrate a charge to another page
5545 * @oldpage: currently charged page
5546 * @newpage: page to transfer the charge to
5547 * @lrucare: either or both pages might be on the LRU already
5549 * Migrate the charge from @oldpage to @newpage.
5551 * Both pages must be locked, @newpage->mapping must be set up.
5553 void mem_cgroup_migrate(struct page
*oldpage
, struct page
*newpage
,
5556 struct mem_cgroup
*memcg
;
5559 VM_BUG_ON_PAGE(!PageLocked(oldpage
), oldpage
);
5560 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
5561 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(oldpage
), oldpage
);
5562 VM_BUG_ON_PAGE(!lrucare
&& PageLRU(newpage
), newpage
);
5563 VM_BUG_ON_PAGE(PageAnon(oldpage
) != PageAnon(newpage
), newpage
);
5564 VM_BUG_ON_PAGE(PageTransHuge(oldpage
) != PageTransHuge(newpage
),
5567 if (mem_cgroup_disabled())
5570 /* Page cache replacement: new page already charged? */
5571 if (newpage
->mem_cgroup
)
5575 * Swapcache readahead pages can get migrated before being
5576 * charged, and migration from compaction can happen to an
5577 * uncharged page when the PFN walker finds a page that
5578 * reclaim just put back on the LRU but has not released yet.
5580 memcg
= oldpage
->mem_cgroup
;
5585 lock_page_lru(oldpage
, &isolated
);
5587 oldpage
->mem_cgroup
= NULL
;
5590 unlock_page_lru(oldpage
, isolated
);
5592 commit_charge(newpage
, memcg
, lrucare
);
5596 * subsys_initcall() for memory controller.
5598 * Some parts like hotcpu_notifier() have to be initialized from this context
5599 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5600 * everything that doesn't depend on a specific mem_cgroup structure should
5601 * be initialized from here.
5603 static int __init
mem_cgroup_init(void)
5607 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
5609 for_each_possible_cpu(cpu
)
5610 INIT_WORK(&per_cpu_ptr(&memcg_stock
, cpu
)->work
,
5613 for_each_node(node
) {
5614 struct mem_cgroup_tree_per_node
*rtpn
;
5617 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
,
5618 node_online(node
) ? node
: NUMA_NO_NODE
);
5620 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5621 struct mem_cgroup_tree_per_zone
*rtpz
;
5623 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
5624 rtpz
->rb_root
= RB_ROOT
;
5625 spin_lock_init(&rtpz
->lock
);
5627 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
5632 subsys_initcall(mem_cgroup_init
);
5634 #ifdef CONFIG_MEMCG_SWAP
5636 * mem_cgroup_swapout - transfer a memsw charge to swap
5637 * @page: page whose memsw charge to transfer
5638 * @entry: swap entry to move the charge to
5640 * Transfer the memsw charge of @page to @entry.
5642 void mem_cgroup_swapout(struct page
*page
, swp_entry_t entry
)
5644 struct mem_cgroup
*memcg
;
5645 unsigned short oldid
;
5647 VM_BUG_ON_PAGE(PageLRU(page
), page
);
5648 VM_BUG_ON_PAGE(page_count(page
), page
);
5650 if (!do_swap_account
)
5653 memcg
= page
->mem_cgroup
;
5655 /* Readahead page, never charged */
5659 oldid
= swap_cgroup_record(entry
, mem_cgroup_id(memcg
));
5660 VM_BUG_ON_PAGE(oldid
, page
);
5661 mem_cgroup_swap_statistics(memcg
, true);
5663 page
->mem_cgroup
= NULL
;
5665 if (!mem_cgroup_is_root(memcg
))
5666 page_counter_uncharge(&memcg
->memory
, 1);
5669 * Interrupts should be disabled here because the caller holds the
5670 * mapping->tree_lock lock which is taken with interrupts-off. It is
5671 * important here to have the interrupts disabled because it is the
5672 * only synchronisation we have for udpating the per-CPU variables.
5674 VM_BUG_ON(!irqs_disabled());
5675 mem_cgroup_charge_statistics(memcg
, page
, -1);
5676 memcg_check_events(memcg
, page
);
5680 * mem_cgroup_uncharge_swap - uncharge a swap entry
5681 * @entry: swap entry to uncharge
5683 * Drop the memsw charge associated with @entry.
5685 void mem_cgroup_uncharge_swap(swp_entry_t entry
)
5687 struct mem_cgroup
*memcg
;
5690 if (!do_swap_account
)
5693 id
= swap_cgroup_record(entry
, 0);
5695 memcg
= mem_cgroup_from_id(id
);
5697 if (!mem_cgroup_is_root(memcg
))
5698 page_counter_uncharge(&memcg
->memsw
, 1);
5699 mem_cgroup_swap_statistics(memcg
, false);
5700 css_put(&memcg
->css
);
5705 /* for remember boot option*/
5706 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5707 static int really_do_swap_account __initdata
= 1;
5709 static int really_do_swap_account __initdata
;
5712 static int __init
enable_swap_account(char *s
)
5714 if (!strcmp(s
, "1"))
5715 really_do_swap_account
= 1;
5716 else if (!strcmp(s
, "0"))
5717 really_do_swap_account
= 0;
5720 __setup("swapaccount=", enable_swap_account
);
5722 static struct cftype memsw_cgroup_files
[] = {
5724 .name
= "memsw.usage_in_bytes",
5725 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5726 .read_u64
= mem_cgroup_read_u64
,
5729 .name
= "memsw.max_usage_in_bytes",
5730 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5731 .write
= mem_cgroup_reset
,
5732 .read_u64
= mem_cgroup_read_u64
,
5735 .name
= "memsw.limit_in_bytes",
5736 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5737 .write
= mem_cgroup_write
,
5738 .read_u64
= mem_cgroup_read_u64
,
5741 .name
= "memsw.failcnt",
5742 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5743 .write
= mem_cgroup_reset
,
5744 .read_u64
= mem_cgroup_read_u64
,
5746 { }, /* terminate */
5749 static int __init
mem_cgroup_swap_init(void)
5751 if (!mem_cgroup_disabled() && really_do_swap_account
) {
5752 do_swap_account
= 1;
5753 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys
,
5754 memsw_cgroup_files
));
5758 subsys_initcall(mem_cgroup_swap_init
);
5760 #endif /* CONFIG_MEMCG_SWAP */