1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/slab.h>
43 #include <linux/swap.h>
44 #include <linux/swapops.h>
45 #include <linux/spinlock.h>
46 #include <linux/eventfd.h>
47 #include <linux/sort.h>
49 #include <linux/seq_file.h>
50 #include <linux/vmalloc.h>
51 #include <linux/vmpressure.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
66 EXPORT_SYMBOL(mem_cgroup_subsys
);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup
*root_mem_cgroup __read_mostly
;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly
;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata
= 1;
79 static int really_do_swap_account __initdata
= 0;
83 #define do_swap_account 0
88 * Statistics for memory cgroup.
90 enum mem_cgroup_stat_index
{
92 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
95 MEM_CGROUP_STAT_RSS
, /* # of pages charged as anon rss */
96 MEM_CGROUP_STAT_RSS_HUGE
, /* # of pages charged as anon huge */
97 MEM_CGROUP_STAT_FILE_MAPPED
, /* # of pages charged as file rss */
98 MEM_CGROUP_STAT_SWAP
, /* # of pages, swapped out */
99 MEM_CGROUP_STAT_NSTATS
,
102 static const char * const mem_cgroup_stat_names
[] = {
110 enum mem_cgroup_events_index
{
111 MEM_CGROUP_EVENTS_PGPGIN
, /* # of pages paged in */
112 MEM_CGROUP_EVENTS_PGPGOUT
, /* # of pages paged out */
113 MEM_CGROUP_EVENTS_PGFAULT
, /* # of page-faults */
114 MEM_CGROUP_EVENTS_PGMAJFAULT
, /* # of major page-faults */
115 MEM_CGROUP_EVENTS_NSTATS
,
118 static const char * const mem_cgroup_events_names
[] = {
125 static const char * const mem_cgroup_lru_names
[] = {
134 * Per memcg event counter is incremented at every pagein/pageout. With THP,
135 * it will be incremated by the number of pages. This counter is used for
136 * for trigger some periodic events. This is straightforward and better
137 * than using jiffies etc. to handle periodic memcg event.
139 enum mem_cgroup_events_target
{
140 MEM_CGROUP_TARGET_THRESH
,
141 MEM_CGROUP_TARGET_NUMAINFO
,
144 #define THRESHOLDS_EVENTS_TARGET 128
145 #define SOFTLIMIT_EVENTS_TARGET 1024
146 #define NUMAINFO_EVENTS_TARGET 1024
148 struct mem_cgroup_stat_cpu
{
149 long count
[MEM_CGROUP_STAT_NSTATS
];
150 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
151 unsigned long nr_page_events
;
152 unsigned long targets
[MEM_CGROUP_NTARGETS
];
155 struct mem_cgroup_reclaim_iter
{
157 * last scanned hierarchy member. Valid only if last_dead_count
158 * matches memcg->dead_count of the hierarchy root group.
160 struct mem_cgroup
*last_visited
;
161 unsigned long last_dead_count
;
163 /* scan generation, increased every round-trip */
164 unsigned int generation
;
168 * per-zone information in memory controller.
170 struct mem_cgroup_per_zone
{
171 struct lruvec lruvec
;
172 unsigned long lru_size
[NR_LRU_LISTS
];
174 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
176 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
177 /* use container_of */
180 struct mem_cgroup_per_node
{
181 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
184 struct mem_cgroup_threshold
{
185 struct eventfd_ctx
*eventfd
;
190 struct mem_cgroup_threshold_ary
{
191 /* An array index points to threshold just below or equal to usage. */
192 int current_threshold
;
193 /* Size of entries[] */
195 /* Array of thresholds */
196 struct mem_cgroup_threshold entries
[0];
199 struct mem_cgroup_thresholds
{
200 /* Primary thresholds array */
201 struct mem_cgroup_threshold_ary
*primary
;
203 * Spare threshold array.
204 * This is needed to make mem_cgroup_unregister_event() "never fail".
205 * It must be able to store at least primary->size - 1 entries.
207 struct mem_cgroup_threshold_ary
*spare
;
211 struct mem_cgroup_eventfd_list
{
212 struct list_head list
;
213 struct eventfd_ctx
*eventfd
;
216 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
217 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
220 * The memory controller data structure. The memory controller controls both
221 * page cache and RSS per cgroup. We would eventually like to provide
222 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
223 * to help the administrator determine what knobs to tune.
225 * TODO: Add a water mark for the memory controller. Reclaim will begin when
226 * we hit the water mark. May be even add a low water mark, such that
227 * no reclaim occurs from a cgroup at it's low water mark, this is
228 * a feature that will be implemented much later in the future.
231 struct cgroup_subsys_state css
;
233 * the counter to account for memory usage
235 struct res_counter res
;
237 /* vmpressure notifications */
238 struct vmpressure vmpressure
;
241 * the counter to account for mem+swap usage.
243 struct res_counter memsw
;
246 * the counter to account for kernel memory usage.
248 struct res_counter kmem
;
250 * Should the accounting and control be hierarchical, per subtree?
253 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
259 /* OOM-Killer disable */
260 int oom_kill_disable
;
262 /* set when res.limit == memsw.limit */
263 bool memsw_is_minimum
;
265 /* protect arrays of thresholds */
266 struct mutex thresholds_lock
;
268 /* thresholds for memory usage. RCU-protected */
269 struct mem_cgroup_thresholds thresholds
;
271 /* thresholds for mem+swap usage. RCU-protected */
272 struct mem_cgroup_thresholds memsw_thresholds
;
274 /* For oom notifier event fd */
275 struct list_head oom_notify
;
278 * Should we move charges of a task when a task is moved into this
279 * mem_cgroup ? And what type of charges should we move ?
281 unsigned long move_charge_at_immigrate
;
283 * set > 0 if pages under this cgroup are moving to other cgroup.
285 atomic_t moving_account
;
286 /* taken only while moving_account > 0 */
287 spinlock_t move_lock
;
291 struct mem_cgroup_stat_cpu __percpu
*stat
;
293 * used when a cpu is offlined or other synchronizations
294 * See mem_cgroup_read_stat().
296 struct mem_cgroup_stat_cpu nocpu_base
;
297 spinlock_t pcp_counter_lock
;
300 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
301 struct tcp_memcontrol tcp_mem
;
303 #if defined(CONFIG_MEMCG_KMEM)
304 /* analogous to slab_common's slab_caches list. per-memcg */
305 struct list_head memcg_slab_caches
;
306 /* Not a spinlock, we can take a lot of time walking the list */
307 struct mutex slab_caches_mutex
;
308 /* Index in the kmem_cache->memcg_params->memcg_caches array */
312 int last_scanned_node
;
314 nodemask_t scan_nodes
;
315 atomic_t numainfo_events
;
316 atomic_t numainfo_updating
;
319 struct mem_cgroup_per_node
*nodeinfo
[0];
320 /* WARNING: nodeinfo must be the last member here */
323 static size_t memcg_size(void)
325 return sizeof(struct mem_cgroup
) +
326 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
329 /* internal only representation about the status of kmem accounting. */
331 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
332 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
333 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
336 /* We account when limit is on, but only after call sites are patched */
337 #define KMEM_ACCOUNTED_MASK \
338 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
340 #ifdef CONFIG_MEMCG_KMEM
341 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
343 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
346 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
348 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
351 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
353 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
356 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
358 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
361 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
364 * Our caller must use css_get() first, because memcg_uncharge_kmem()
365 * will call css_put() if it sees the memcg is dead.
368 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
369 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
372 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
374 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
375 &memcg
->kmem_account_flags
);
379 /* Stuffs for move charges at task migration. */
381 * Types of charges to be moved. "move_charge_at_immitgrate" and
382 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
385 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
386 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
390 /* "mc" and its members are protected by cgroup_mutex */
391 static struct move_charge_struct
{
392 spinlock_t lock
; /* for from, to */
393 struct mem_cgroup
*from
;
394 struct mem_cgroup
*to
;
395 unsigned long immigrate_flags
;
396 unsigned long precharge
;
397 unsigned long moved_charge
;
398 unsigned long moved_swap
;
399 struct task_struct
*moving_task
; /* a task moving charges */
400 wait_queue_head_t waitq
; /* a waitq for other context */
402 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
403 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
406 static bool move_anon(void)
408 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
411 static bool move_file(void)
413 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
417 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
418 * limit reclaim to prevent infinite loops, if they ever occur.
420 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
423 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
424 MEM_CGROUP_CHARGE_TYPE_ANON
,
425 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
426 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
430 /* for encoding cft->private value on file */
438 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
439 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
440 #define MEMFILE_ATTR(val) ((val) & 0xffff)
441 /* Used for OOM nofiier */
442 #define OOM_CONTROL (0)
445 * Reclaim flags for mem_cgroup_hierarchical_reclaim
447 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
448 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
449 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
450 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
453 * The memcg_create_mutex will be held whenever a new cgroup is created.
454 * As a consequence, any change that needs to protect against new child cgroups
455 * appearing has to hold it as well.
457 static DEFINE_MUTEX(memcg_create_mutex
);
459 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
461 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
464 /* Some nice accessors for the vmpressure. */
465 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
468 memcg
= root_mem_cgroup
;
469 return &memcg
->vmpressure
;
472 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
474 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
477 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
479 return &mem_cgroup_from_css(css
)->vmpressure
;
482 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
484 return (memcg
== root_mem_cgroup
);
487 /* Writing them here to avoid exposing memcg's inner layout */
488 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
490 void sock_update_memcg(struct sock
*sk
)
492 if (mem_cgroup_sockets_enabled
) {
493 struct mem_cgroup
*memcg
;
494 struct cg_proto
*cg_proto
;
496 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
498 /* Socket cloning can throw us here with sk_cgrp already
499 * filled. It won't however, necessarily happen from
500 * process context. So the test for root memcg given
501 * the current task's memcg won't help us in this case.
503 * Respecting the original socket's memcg is a better
504 * decision in this case.
507 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
508 css_get(&sk
->sk_cgrp
->memcg
->css
);
513 memcg
= mem_cgroup_from_task(current
);
514 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
515 if (!mem_cgroup_is_root(memcg
) &&
516 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
517 sk
->sk_cgrp
= cg_proto
;
522 EXPORT_SYMBOL(sock_update_memcg
);
524 void sock_release_memcg(struct sock
*sk
)
526 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
527 struct mem_cgroup
*memcg
;
528 WARN_ON(!sk
->sk_cgrp
->memcg
);
529 memcg
= sk
->sk_cgrp
->memcg
;
530 css_put(&sk
->sk_cgrp
->memcg
->css
);
534 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
536 if (!memcg
|| mem_cgroup_is_root(memcg
))
539 return &memcg
->tcp_mem
.cg_proto
;
541 EXPORT_SYMBOL(tcp_proto_cgroup
);
543 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
545 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
547 static_key_slow_dec(&memcg_socket_limit_enabled
);
550 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
555 #ifdef CONFIG_MEMCG_KMEM
557 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
558 * There are two main reasons for not using the css_id for this:
559 * 1) this works better in sparse environments, where we have a lot of memcgs,
560 * but only a few kmem-limited. Or also, if we have, for instance, 200
561 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
562 * 200 entry array for that.
564 * 2) In order not to violate the cgroup API, we would like to do all memory
565 * allocation in ->create(). At that point, we haven't yet allocated the
566 * css_id. Having a separate index prevents us from messing with the cgroup
569 * The current size of the caches array is stored in
570 * memcg_limited_groups_array_size. It will double each time we have to
573 static DEFINE_IDA(kmem_limited_groups
);
574 int memcg_limited_groups_array_size
;
577 * MIN_SIZE is different than 1, because we would like to avoid going through
578 * the alloc/free process all the time. In a small machine, 4 kmem-limited
579 * cgroups is a reasonable guess. In the future, it could be a parameter or
580 * tunable, but that is strictly not necessary.
582 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
583 * this constant directly from cgroup, but it is understandable that this is
584 * better kept as an internal representation in cgroup.c. In any case, the
585 * css_id space is not getting any smaller, and we don't have to necessarily
586 * increase ours as well if it increases.
588 #define MEMCG_CACHES_MIN_SIZE 4
589 #define MEMCG_CACHES_MAX_SIZE 65535
592 * A lot of the calls to the cache allocation functions are expected to be
593 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
594 * conditional to this static branch, we'll have to allow modules that does
595 * kmem_cache_alloc and the such to see this symbol as well
597 struct static_key memcg_kmem_enabled_key
;
598 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
600 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
602 if (memcg_kmem_is_active(memcg
)) {
603 static_key_slow_dec(&memcg_kmem_enabled_key
);
604 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
607 * This check can't live in kmem destruction function,
608 * since the charges will outlive the cgroup
610 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
613 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
616 #endif /* CONFIG_MEMCG_KMEM */
618 static void disarm_static_keys(struct mem_cgroup
*memcg
)
620 disarm_sock_keys(memcg
);
621 disarm_kmem_keys(memcg
);
624 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
626 static struct mem_cgroup_per_zone
*
627 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
629 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
630 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
633 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
638 static struct mem_cgroup_per_zone
*
639 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
641 int nid
= page_to_nid(page
);
642 int zid
= page_zonenum(page
);
644 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
648 * Implementation Note: reading percpu statistics for memcg.
650 * Both of vmstat[] and percpu_counter has threshold and do periodic
651 * synchronization to implement "quick" read. There are trade-off between
652 * reading cost and precision of value. Then, we may have a chance to implement
653 * a periodic synchronizion of counter in memcg's counter.
655 * But this _read() function is used for user interface now. The user accounts
656 * memory usage by memory cgroup and he _always_ requires exact value because
657 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
658 * have to visit all online cpus and make sum. So, for now, unnecessary
659 * synchronization is not implemented. (just implemented for cpu hotplug)
661 * If there are kernel internal actions which can make use of some not-exact
662 * value, and reading all cpu value can be performance bottleneck in some
663 * common workload, threashold and synchonization as vmstat[] should be
666 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
667 enum mem_cgroup_stat_index idx
)
673 for_each_online_cpu(cpu
)
674 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
675 #ifdef CONFIG_HOTPLUG_CPU
676 spin_lock(&memcg
->pcp_counter_lock
);
677 val
+= memcg
->nocpu_base
.count
[idx
];
678 spin_unlock(&memcg
->pcp_counter_lock
);
684 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
687 int val
= (charge
) ? 1 : -1;
688 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
691 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
692 enum mem_cgroup_events_index idx
)
694 unsigned long val
= 0;
697 for_each_online_cpu(cpu
)
698 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
699 #ifdef CONFIG_HOTPLUG_CPU
700 spin_lock(&memcg
->pcp_counter_lock
);
701 val
+= memcg
->nocpu_base
.events
[idx
];
702 spin_unlock(&memcg
->pcp_counter_lock
);
707 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
709 bool anon
, int nr_pages
)
714 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
715 * counted as CACHE even if it's on ANON LRU.
718 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
721 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
724 if (PageTransHuge(page
))
725 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
728 /* pagein of a big page is an event. So, ignore page size */
730 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
732 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
733 nr_pages
= -nr_pages
; /* for event */
736 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
742 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
744 struct mem_cgroup_per_zone
*mz
;
746 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
747 return mz
->lru_size
[lru
];
751 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
752 unsigned int lru_mask
)
754 struct mem_cgroup_per_zone
*mz
;
756 unsigned long ret
= 0;
758 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
761 if (BIT(lru
) & lru_mask
)
762 ret
+= mz
->lru_size
[lru
];
768 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
769 int nid
, unsigned int lru_mask
)
774 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
775 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
781 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
782 unsigned int lru_mask
)
787 for_each_node_state(nid
, N_MEMORY
)
788 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
792 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
793 enum mem_cgroup_events_target target
)
795 unsigned long val
, next
;
797 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
798 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
799 /* from time_after() in jiffies.h */
800 if ((long)next
- (long)val
< 0) {
802 case MEM_CGROUP_TARGET_THRESH
:
803 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
805 case MEM_CGROUP_TARGET_NUMAINFO
:
806 next
= val
+ NUMAINFO_EVENTS_TARGET
;
811 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
818 * Check events in order.
821 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
824 /* threshold event is triggered in finer grain than soft limit */
825 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
826 MEM_CGROUP_TARGET_THRESH
))) {
827 bool do_numainfo __maybe_unused
;
830 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
831 MEM_CGROUP_TARGET_NUMAINFO
);
835 mem_cgroup_threshold(memcg
);
837 if (unlikely(do_numainfo
))
838 atomic_inc(&memcg
->numainfo_events
);
844 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
847 * mm_update_next_owner() may clear mm->owner to NULL
848 * if it races with swapoff, page migration, etc.
849 * So this can be called with p == NULL.
854 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
857 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
859 struct mem_cgroup
*memcg
= NULL
;
864 * Because we have no locks, mm->owner's may be being moved to other
865 * cgroup. We use css_tryget() here even if this looks
866 * pessimistic (rather than adding locks here).
870 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
871 if (unlikely(!memcg
))
873 } while (!css_tryget(&memcg
->css
));
879 * Returns a next (in a pre-order walk) alive memcg (with elevated css
880 * ref. count) or NULL if the whole root's subtree has been visited.
882 * helper function to be used by mem_cgroup_iter
884 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
885 struct mem_cgroup
*last_visited
)
887 struct cgroup_subsys_state
*prev_css
, *next_css
;
889 prev_css
= last_visited
? &last_visited
->css
: NULL
;
891 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
894 * Even if we found a group we have to make sure it is
895 * alive. css && !memcg means that the groups should be
896 * skipped and we should continue the tree walk.
897 * last_visited css is safe to use because it is
898 * protected by css_get and the tree walk is rcu safe.
901 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
903 if (css_tryget(&mem
->css
))
914 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
917 * When a group in the hierarchy below root is destroyed, the
918 * hierarchy iterator can no longer be trusted since it might
919 * have pointed to the destroyed group. Invalidate it.
921 atomic_inc(&root
->dead_count
);
924 static struct mem_cgroup
*
925 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
926 struct mem_cgroup
*root
,
929 struct mem_cgroup
*position
= NULL
;
931 * A cgroup destruction happens in two stages: offlining and
932 * release. They are separated by a RCU grace period.
934 * If the iterator is valid, we may still race with an
935 * offlining. The RCU lock ensures the object won't be
936 * released, tryget will fail if we lost the race.
938 *sequence
= atomic_read(&root
->dead_count
);
939 if (iter
->last_dead_count
== *sequence
) {
941 position
= iter
->last_visited
;
942 if (position
&& !css_tryget(&position
->css
))
948 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
949 struct mem_cgroup
*last_visited
,
950 struct mem_cgroup
*new_position
,
954 css_put(&last_visited
->css
);
956 * We store the sequence count from the time @last_visited was
957 * loaded successfully instead of rereading it here so that we
958 * don't lose destruction events in between. We could have
959 * raced with the destruction of @new_position after all.
961 iter
->last_visited
= new_position
;
963 iter
->last_dead_count
= sequence
;
967 * mem_cgroup_iter - iterate over memory cgroup hierarchy
968 * @root: hierarchy root
969 * @prev: previously returned memcg, NULL on first invocation
970 * @reclaim: cookie for shared reclaim walks, NULL for full walks
972 * Returns references to children of the hierarchy below @root, or
973 * @root itself, or %NULL after a full round-trip.
975 * Caller must pass the return value in @prev on subsequent
976 * invocations for reference counting, or use mem_cgroup_iter_break()
977 * to cancel a hierarchy walk before the round-trip is complete.
979 * Reclaimers can specify a zone and a priority level in @reclaim to
980 * divide up the memcgs in the hierarchy among all concurrent
981 * reclaimers operating on the same zone and priority.
983 struct mem_cgroup
*mem_cgroup_iter(struct mem_cgroup
*root
,
984 struct mem_cgroup
*prev
,
985 struct mem_cgroup_reclaim_cookie
*reclaim
)
987 struct mem_cgroup
*memcg
= NULL
;
988 struct mem_cgroup
*last_visited
= NULL
;
990 if (mem_cgroup_disabled())
994 root
= root_mem_cgroup
;
996 if (prev
&& !reclaim
)
999 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1007 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1008 int uninitialized_var(seq
);
1011 int nid
= zone_to_nid(reclaim
->zone
);
1012 int zid
= zone_idx(reclaim
->zone
);
1013 struct mem_cgroup_per_zone
*mz
;
1015 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1016 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1017 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1018 iter
->last_visited
= NULL
;
1022 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1025 memcg
= __mem_cgroup_iter_next(root
, last_visited
);
1028 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1032 else if (!prev
&& memcg
)
1033 reclaim
->generation
= iter
->generation
;
1042 if (prev
&& prev
!= root
)
1043 css_put(&prev
->css
);
1049 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1050 * @root: hierarchy root
1051 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1053 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1054 struct mem_cgroup
*prev
)
1057 root
= root_mem_cgroup
;
1058 if (prev
&& prev
!= root
)
1059 css_put(&prev
->css
);
1063 * Iteration constructs for visiting all cgroups (under a tree). If
1064 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1065 * be used for reference counting.
1067 #define for_each_mem_cgroup_tree(iter, root) \
1068 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1070 iter = mem_cgroup_iter(root, iter, NULL))
1072 #define for_each_mem_cgroup(iter) \
1073 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1075 iter = mem_cgroup_iter(NULL, iter, NULL))
1077 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1079 struct mem_cgroup
*memcg
;
1082 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1083 if (unlikely(!memcg
))
1088 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1091 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1099 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1102 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1103 * @zone: zone of the wanted lruvec
1104 * @memcg: memcg of the wanted lruvec
1106 * Returns the lru list vector holding pages for the given @zone and
1107 * @mem. This can be the global zone lruvec, if the memory controller
1110 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1111 struct mem_cgroup
*memcg
)
1113 struct mem_cgroup_per_zone
*mz
;
1114 struct lruvec
*lruvec
;
1116 if (mem_cgroup_disabled()) {
1117 lruvec
= &zone
->lruvec
;
1121 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1122 lruvec
= &mz
->lruvec
;
1125 * Since a node can be onlined after the mem_cgroup was created,
1126 * we have to be prepared to initialize lruvec->zone here;
1127 * and if offlined then reonlined, we need to reinitialize it.
1129 if (unlikely(lruvec
->zone
!= zone
))
1130 lruvec
->zone
= zone
;
1135 * Following LRU functions are allowed to be used without PCG_LOCK.
1136 * Operations are called by routine of global LRU independently from memcg.
1137 * What we have to take care of here is validness of pc->mem_cgroup.
1139 * Changes to pc->mem_cgroup happens when
1142 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1143 * It is added to LRU before charge.
1144 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1145 * When moving account, the page is not on LRU. It's isolated.
1149 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1151 * @zone: zone of the page
1153 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1155 struct mem_cgroup_per_zone
*mz
;
1156 struct mem_cgroup
*memcg
;
1157 struct page_cgroup
*pc
;
1158 struct lruvec
*lruvec
;
1160 if (mem_cgroup_disabled()) {
1161 lruvec
= &zone
->lruvec
;
1165 pc
= lookup_page_cgroup(page
);
1166 memcg
= pc
->mem_cgroup
;
1169 * Surreptitiously switch any uncharged offlist page to root:
1170 * an uncharged page off lru does nothing to secure
1171 * its former mem_cgroup from sudden removal.
1173 * Our caller holds lru_lock, and PageCgroupUsed is updated
1174 * under page_cgroup lock: between them, they make all uses
1175 * of pc->mem_cgroup safe.
1177 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1178 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1180 mz
= page_cgroup_zoneinfo(memcg
, page
);
1181 lruvec
= &mz
->lruvec
;
1184 * Since a node can be onlined after the mem_cgroup was created,
1185 * we have to be prepared to initialize lruvec->zone here;
1186 * and if offlined then reonlined, we need to reinitialize it.
1188 if (unlikely(lruvec
->zone
!= zone
))
1189 lruvec
->zone
= zone
;
1194 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1195 * @lruvec: mem_cgroup per zone lru vector
1196 * @lru: index of lru list the page is sitting on
1197 * @nr_pages: positive when adding or negative when removing
1199 * This function must be called when a page is added to or removed from an
1202 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1205 struct mem_cgroup_per_zone
*mz
;
1206 unsigned long *lru_size
;
1208 if (mem_cgroup_disabled())
1211 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1212 lru_size
= mz
->lru_size
+ lru
;
1213 *lru_size
+= nr_pages
;
1214 VM_BUG_ON((long)(*lru_size
) < 0);
1218 * Checks whether given mem is same or in the root_mem_cgroup's
1221 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1222 struct mem_cgroup
*memcg
)
1224 if (root_memcg
== memcg
)
1226 if (!root_memcg
->use_hierarchy
|| !memcg
)
1228 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1231 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1232 struct mem_cgroup
*memcg
)
1237 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1242 bool task_in_mem_cgroup(struct task_struct
*task
,
1243 const struct mem_cgroup
*memcg
)
1245 struct mem_cgroup
*curr
= NULL
;
1246 struct task_struct
*p
;
1249 p
= find_lock_task_mm(task
);
1251 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1255 * All threads may have already detached their mm's, but the oom
1256 * killer still needs to detect if they have already been oom
1257 * killed to prevent needlessly killing additional tasks.
1260 curr
= mem_cgroup_from_task(task
);
1262 css_get(&curr
->css
);
1268 * We should check use_hierarchy of "memcg" not "curr". Because checking
1269 * use_hierarchy of "curr" here make this function true if hierarchy is
1270 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1271 * hierarchy(even if use_hierarchy is disabled in "memcg").
1273 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1274 css_put(&curr
->css
);
1278 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1280 unsigned long inactive_ratio
;
1281 unsigned long inactive
;
1282 unsigned long active
;
1285 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1286 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1288 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1290 inactive_ratio
= int_sqrt(10 * gb
);
1294 return inactive
* inactive_ratio
< active
;
1297 #define mem_cgroup_from_res_counter(counter, member) \
1298 container_of(counter, struct mem_cgroup, member)
1301 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1302 * @memcg: the memory cgroup
1304 * Returns the maximum amount of memory @mem can be charged with, in
1307 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1309 unsigned long long margin
;
1311 margin
= res_counter_margin(&memcg
->res
);
1312 if (do_swap_account
)
1313 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1314 return margin
>> PAGE_SHIFT
;
1317 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1320 if (!css_parent(&memcg
->css
))
1321 return vm_swappiness
;
1323 return memcg
->swappiness
;
1327 * memcg->moving_account is used for checking possibility that some thread is
1328 * calling move_account(). When a thread on CPU-A starts moving pages under
1329 * a memcg, other threads should check memcg->moving_account under
1330 * rcu_read_lock(), like this:
1334 * memcg->moving_account+1 if (memcg->mocing_account)
1336 * synchronize_rcu() update something.
1341 /* for quick checking without looking up memcg */
1342 atomic_t memcg_moving __read_mostly
;
1344 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1346 atomic_inc(&memcg_moving
);
1347 atomic_inc(&memcg
->moving_account
);
1351 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1354 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1355 * We check NULL in callee rather than caller.
1358 atomic_dec(&memcg_moving
);
1359 atomic_dec(&memcg
->moving_account
);
1364 * 2 routines for checking "mem" is under move_account() or not.
1366 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1367 * is used for avoiding races in accounting. If true,
1368 * pc->mem_cgroup may be overwritten.
1370 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1371 * under hierarchy of moving cgroups. This is for
1372 * waiting at hith-memory prressure caused by "move".
1375 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1377 VM_BUG_ON(!rcu_read_lock_held());
1378 return atomic_read(&memcg
->moving_account
) > 0;
1381 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1383 struct mem_cgroup
*from
;
1384 struct mem_cgroup
*to
;
1387 * Unlike task_move routines, we access mc.to, mc.from not under
1388 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1390 spin_lock(&mc
.lock
);
1396 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1397 || mem_cgroup_same_or_subtree(memcg
, to
);
1399 spin_unlock(&mc
.lock
);
1403 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1405 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1406 if (mem_cgroup_under_move(memcg
)) {
1408 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1409 /* moving charge context might have finished. */
1412 finish_wait(&mc
.waitq
, &wait
);
1420 * Take this lock when
1421 * - a code tries to modify page's memcg while it's USED.
1422 * - a code tries to modify page state accounting in a memcg.
1423 * see mem_cgroup_stolen(), too.
1425 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1426 unsigned long *flags
)
1428 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1431 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1432 unsigned long *flags
)
1434 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1437 #define K(x) ((x) << (PAGE_SHIFT-10))
1439 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1440 * @memcg: The memory cgroup that went over limit
1441 * @p: Task that is going to be killed
1443 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1446 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1448 struct cgroup
*task_cgrp
;
1449 struct cgroup
*mem_cgrp
;
1451 * Need a buffer in BSS, can't rely on allocations. The code relies
1452 * on the assumption that OOM is serialized for memory controller.
1453 * If this assumption is broken, revisit this code.
1455 static char memcg_name
[PATH_MAX
];
1457 struct mem_cgroup
*iter
;
1465 mem_cgrp
= memcg
->css
.cgroup
;
1466 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1468 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1471 * Unfortunately, we are unable to convert to a useful name
1472 * But we'll still print out the usage information
1479 pr_info("Task in %s killed", memcg_name
);
1482 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1490 * Continues from above, so we don't need an KERN_ level
1492 pr_cont(" as a result of limit of %s\n", memcg_name
);
1495 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1496 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1497 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1498 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1499 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1500 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1501 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1502 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1503 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1504 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1505 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1506 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1508 for_each_mem_cgroup_tree(iter
, memcg
) {
1509 pr_info("Memory cgroup stats");
1512 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1514 pr_cont(" for %s", memcg_name
);
1518 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1519 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1521 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1522 K(mem_cgroup_read_stat(iter
, i
)));
1525 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1526 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1527 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1534 * This function returns the number of memcg under hierarchy tree. Returns
1535 * 1(self count) if no children.
1537 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1540 struct mem_cgroup
*iter
;
1542 for_each_mem_cgroup_tree(iter
, memcg
)
1548 * Return the memory (and swap, if configured) limit for a memcg.
1550 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1554 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1557 * Do not consider swap space if we cannot swap due to swappiness
1559 if (mem_cgroup_swappiness(memcg
)) {
1562 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1563 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1566 * If memsw is finite and limits the amount of swap space
1567 * available to this memcg, return that limit.
1569 limit
= min(limit
, memsw
);
1575 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1578 struct mem_cgroup
*iter
;
1579 unsigned long chosen_points
= 0;
1580 unsigned long totalpages
;
1581 unsigned int points
= 0;
1582 struct task_struct
*chosen
= NULL
;
1585 * If current has a pending SIGKILL or is exiting, then automatically
1586 * select it. The goal is to allow it to allocate so that it may
1587 * quickly exit and free its memory.
1589 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1590 set_thread_flag(TIF_MEMDIE
);
1594 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1595 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1596 for_each_mem_cgroup_tree(iter
, memcg
) {
1597 struct css_task_iter it
;
1598 struct task_struct
*task
;
1600 css_task_iter_start(&iter
->css
, &it
);
1601 while ((task
= css_task_iter_next(&it
))) {
1602 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1604 case OOM_SCAN_SELECT
:
1606 put_task_struct(chosen
);
1608 chosen_points
= ULONG_MAX
;
1609 get_task_struct(chosen
);
1611 case OOM_SCAN_CONTINUE
:
1613 case OOM_SCAN_ABORT
:
1614 css_task_iter_end(&it
);
1615 mem_cgroup_iter_break(memcg
, iter
);
1617 put_task_struct(chosen
);
1622 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1623 if (points
> chosen_points
) {
1625 put_task_struct(chosen
);
1627 chosen_points
= points
;
1628 get_task_struct(chosen
);
1631 css_task_iter_end(&it
);
1636 points
= chosen_points
* 1000 / totalpages
;
1637 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1638 NULL
, "Memory cgroup out of memory");
1641 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1643 unsigned long flags
)
1645 unsigned long total
= 0;
1646 bool noswap
= false;
1649 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1651 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1654 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1656 drain_all_stock_async(memcg
);
1657 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1659 * Allow limit shrinkers, which are triggered directly
1660 * by userspace, to catch signals and stop reclaim
1661 * after minimal progress, regardless of the margin.
1663 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1665 if (mem_cgroup_margin(memcg
))
1668 * If nothing was reclaimed after two attempts, there
1669 * may be no reclaimable pages in this hierarchy.
1677 #if MAX_NUMNODES > 1
1679 * test_mem_cgroup_node_reclaimable
1680 * @memcg: the target memcg
1681 * @nid: the node ID to be checked.
1682 * @noswap : specify true here if the user wants flle only information.
1684 * This function returns whether the specified memcg contains any
1685 * reclaimable pages on a node. Returns true if there are any reclaimable
1686 * pages in the node.
1688 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1689 int nid
, bool noswap
)
1691 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1693 if (noswap
|| !total_swap_pages
)
1695 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1702 * Always updating the nodemask is not very good - even if we have an empty
1703 * list or the wrong list here, we can start from some node and traverse all
1704 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1707 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1711 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1712 * pagein/pageout changes since the last update.
1714 if (!atomic_read(&memcg
->numainfo_events
))
1716 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1719 /* make a nodemask where this memcg uses memory from */
1720 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1722 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1724 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1725 node_clear(nid
, memcg
->scan_nodes
);
1728 atomic_set(&memcg
->numainfo_events
, 0);
1729 atomic_set(&memcg
->numainfo_updating
, 0);
1733 * Selecting a node where we start reclaim from. Because what we need is just
1734 * reducing usage counter, start from anywhere is O,K. Considering
1735 * memory reclaim from current node, there are pros. and cons.
1737 * Freeing memory from current node means freeing memory from a node which
1738 * we'll use or we've used. So, it may make LRU bad. And if several threads
1739 * hit limits, it will see a contention on a node. But freeing from remote
1740 * node means more costs for memory reclaim because of memory latency.
1742 * Now, we use round-robin. Better algorithm is welcomed.
1744 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1748 mem_cgroup_may_update_nodemask(memcg
);
1749 node
= memcg
->last_scanned_node
;
1751 node
= next_node(node
, memcg
->scan_nodes
);
1752 if (node
== MAX_NUMNODES
)
1753 node
= first_node(memcg
->scan_nodes
);
1755 * We call this when we hit limit, not when pages are added to LRU.
1756 * No LRU may hold pages because all pages are UNEVICTABLE or
1757 * memcg is too small and all pages are not on LRU. In that case,
1758 * we use curret node.
1760 if (unlikely(node
== MAX_NUMNODES
))
1761 node
= numa_node_id();
1763 memcg
->last_scanned_node
= node
;
1768 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1776 * A group is eligible for the soft limit reclaim if it is
1777 * a) is over its soft limit
1778 * b) any parent up the hierarchy is over its soft limit
1780 bool mem_cgroup_soft_reclaim_eligible(struct mem_cgroup
*memcg
)
1782 struct mem_cgroup
*parent
= memcg
;
1784 if (res_counter_soft_limit_excess(&memcg
->res
))
1788 * If any parent up the hierarchy is over its soft limit then we
1789 * have to obey and reclaim from this group as well.
1791 while((parent
= parent_mem_cgroup(parent
))) {
1792 if (res_counter_soft_limit_excess(&parent
->res
))
1800 * Check OOM-Killer is already running under our hierarchy.
1801 * If someone is running, return false.
1802 * Has to be called with memcg_oom_lock
1804 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
1806 struct mem_cgroup
*iter
, *failed
= NULL
;
1808 for_each_mem_cgroup_tree(iter
, memcg
) {
1809 if (iter
->oom_lock
) {
1811 * this subtree of our hierarchy is already locked
1812 * so we cannot give a lock.
1815 mem_cgroup_iter_break(memcg
, iter
);
1818 iter
->oom_lock
= true;
1825 * OK, we failed to lock the whole subtree so we have to clean up
1826 * what we set up to the failing subtree
1828 for_each_mem_cgroup_tree(iter
, memcg
) {
1829 if (iter
== failed
) {
1830 mem_cgroup_iter_break(memcg
, iter
);
1833 iter
->oom_lock
= false;
1839 * Has to be called with memcg_oom_lock
1841 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1843 struct mem_cgroup
*iter
;
1845 for_each_mem_cgroup_tree(iter
, memcg
)
1846 iter
->oom_lock
= false;
1850 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1852 struct mem_cgroup
*iter
;
1854 for_each_mem_cgroup_tree(iter
, memcg
)
1855 atomic_inc(&iter
->under_oom
);
1858 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1860 struct mem_cgroup
*iter
;
1863 * When a new child is created while the hierarchy is under oom,
1864 * mem_cgroup_oom_lock() may not be called. We have to use
1865 * atomic_add_unless() here.
1867 for_each_mem_cgroup_tree(iter
, memcg
)
1868 atomic_add_unless(&iter
->under_oom
, -1, 0);
1871 static DEFINE_SPINLOCK(memcg_oom_lock
);
1872 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1874 struct oom_wait_info
{
1875 struct mem_cgroup
*memcg
;
1879 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1880 unsigned mode
, int sync
, void *arg
)
1882 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1883 struct mem_cgroup
*oom_wait_memcg
;
1884 struct oom_wait_info
*oom_wait_info
;
1886 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1887 oom_wait_memcg
= oom_wait_info
->memcg
;
1890 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1891 * Then we can use css_is_ancestor without taking care of RCU.
1893 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
1894 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
1896 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1899 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
1901 /* for filtering, pass "memcg" as argument. */
1902 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1905 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1907 if (memcg
&& atomic_read(&memcg
->under_oom
))
1908 memcg_wakeup_oom(memcg
);
1912 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1914 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
1917 struct oom_wait_info owait
;
1918 bool locked
, need_to_kill
;
1920 owait
.memcg
= memcg
;
1921 owait
.wait
.flags
= 0;
1922 owait
.wait
.func
= memcg_oom_wake_function
;
1923 owait
.wait
.private = current
;
1924 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1925 need_to_kill
= true;
1926 mem_cgroup_mark_under_oom(memcg
);
1928 /* At first, try to OOM lock hierarchy under memcg.*/
1929 spin_lock(&memcg_oom_lock
);
1930 locked
= mem_cgroup_oom_lock(memcg
);
1932 * Even if signal_pending(), we can't quit charge() loop without
1933 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1934 * under OOM is always welcomed, use TASK_KILLABLE here.
1936 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1937 if (!locked
|| memcg
->oom_kill_disable
)
1938 need_to_kill
= false;
1940 mem_cgroup_oom_notify(memcg
);
1941 spin_unlock(&memcg_oom_lock
);
1944 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1945 mem_cgroup_out_of_memory(memcg
, mask
, order
);
1948 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1950 spin_lock(&memcg_oom_lock
);
1952 mem_cgroup_oom_unlock(memcg
);
1953 memcg_wakeup_oom(memcg
);
1954 spin_unlock(&memcg_oom_lock
);
1956 mem_cgroup_unmark_under_oom(memcg
);
1958 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
1960 /* Give chance to dying process */
1961 schedule_timeout_uninterruptible(1);
1966 * Currently used to update mapped file statistics, but the routine can be
1967 * generalized to update other statistics as well.
1969 * Notes: Race condition
1971 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1972 * it tends to be costly. But considering some conditions, we doesn't need
1973 * to do so _always_.
1975 * Considering "charge", lock_page_cgroup() is not required because all
1976 * file-stat operations happen after a page is attached to radix-tree. There
1977 * are no race with "charge".
1979 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1980 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1981 * if there are race with "uncharge". Statistics itself is properly handled
1984 * Considering "move", this is an only case we see a race. To make the race
1985 * small, we check mm->moving_account and detect there are possibility of race
1986 * If there is, we take a lock.
1989 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
1990 bool *locked
, unsigned long *flags
)
1992 struct mem_cgroup
*memcg
;
1993 struct page_cgroup
*pc
;
1995 pc
= lookup_page_cgroup(page
);
1997 memcg
= pc
->mem_cgroup
;
1998 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2001 * If this memory cgroup is not under account moving, we don't
2002 * need to take move_lock_mem_cgroup(). Because we already hold
2003 * rcu_read_lock(), any calls to move_account will be delayed until
2004 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2006 if (!mem_cgroup_stolen(memcg
))
2009 move_lock_mem_cgroup(memcg
, flags
);
2010 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2011 move_unlock_mem_cgroup(memcg
, flags
);
2017 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2019 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2022 * It's guaranteed that pc->mem_cgroup never changes while
2023 * lock is held because a routine modifies pc->mem_cgroup
2024 * should take move_lock_mem_cgroup().
2026 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2029 void mem_cgroup_update_page_stat(struct page
*page
,
2030 enum mem_cgroup_page_stat_item idx
, int val
)
2032 struct mem_cgroup
*memcg
;
2033 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2034 unsigned long uninitialized_var(flags
);
2036 if (mem_cgroup_disabled())
2039 memcg
= pc
->mem_cgroup
;
2040 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2044 case MEMCG_NR_FILE_MAPPED
:
2045 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2051 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2055 * size of first charge trial. "32" comes from vmscan.c's magic value.
2056 * TODO: maybe necessary to use big numbers in big irons.
2058 #define CHARGE_BATCH 32U
2059 struct memcg_stock_pcp
{
2060 struct mem_cgroup
*cached
; /* this never be root cgroup */
2061 unsigned int nr_pages
;
2062 struct work_struct work
;
2063 unsigned long flags
;
2064 #define FLUSHING_CACHED_CHARGE 0
2066 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2067 static DEFINE_MUTEX(percpu_charge_mutex
);
2070 * consume_stock: Try to consume stocked charge on this cpu.
2071 * @memcg: memcg to consume from.
2072 * @nr_pages: how many pages to charge.
2074 * The charges will only happen if @memcg matches the current cpu's memcg
2075 * stock, and at least @nr_pages are available in that stock. Failure to
2076 * service an allocation will refill the stock.
2078 * returns true if successful, false otherwise.
2080 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2082 struct memcg_stock_pcp
*stock
;
2085 if (nr_pages
> CHARGE_BATCH
)
2088 stock
= &get_cpu_var(memcg_stock
);
2089 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2090 stock
->nr_pages
-= nr_pages
;
2091 else /* need to call res_counter_charge */
2093 put_cpu_var(memcg_stock
);
2098 * Returns stocks cached in percpu to res_counter and reset cached information.
2100 static void drain_stock(struct memcg_stock_pcp
*stock
)
2102 struct mem_cgroup
*old
= stock
->cached
;
2104 if (stock
->nr_pages
) {
2105 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2107 res_counter_uncharge(&old
->res
, bytes
);
2108 if (do_swap_account
)
2109 res_counter_uncharge(&old
->memsw
, bytes
);
2110 stock
->nr_pages
= 0;
2112 stock
->cached
= NULL
;
2116 * This must be called under preempt disabled or must be called by
2117 * a thread which is pinned to local cpu.
2119 static void drain_local_stock(struct work_struct
*dummy
)
2121 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2123 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2126 static void __init
memcg_stock_init(void)
2130 for_each_possible_cpu(cpu
) {
2131 struct memcg_stock_pcp
*stock
=
2132 &per_cpu(memcg_stock
, cpu
);
2133 INIT_WORK(&stock
->work
, drain_local_stock
);
2138 * Cache charges(val) which is from res_counter, to local per_cpu area.
2139 * This will be consumed by consume_stock() function, later.
2141 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2143 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2145 if (stock
->cached
!= memcg
) { /* reset if necessary */
2147 stock
->cached
= memcg
;
2149 stock
->nr_pages
+= nr_pages
;
2150 put_cpu_var(memcg_stock
);
2154 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2155 * of the hierarchy under it. sync flag says whether we should block
2156 * until the work is done.
2158 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2162 /* Notify other cpus that system-wide "drain" is running */
2165 for_each_online_cpu(cpu
) {
2166 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2167 struct mem_cgroup
*memcg
;
2169 memcg
= stock
->cached
;
2170 if (!memcg
|| !stock
->nr_pages
)
2172 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2174 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2176 drain_local_stock(&stock
->work
);
2178 schedule_work_on(cpu
, &stock
->work
);
2186 for_each_online_cpu(cpu
) {
2187 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2188 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2189 flush_work(&stock
->work
);
2196 * Tries to drain stocked charges in other cpus. This function is asynchronous
2197 * and just put a work per cpu for draining localy on each cpu. Caller can
2198 * expects some charges will be back to res_counter later but cannot wait for
2201 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2204 * If someone calls draining, avoid adding more kworker runs.
2206 if (!mutex_trylock(&percpu_charge_mutex
))
2208 drain_all_stock(root_memcg
, false);
2209 mutex_unlock(&percpu_charge_mutex
);
2212 /* This is a synchronous drain interface. */
2213 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2215 /* called when force_empty is called */
2216 mutex_lock(&percpu_charge_mutex
);
2217 drain_all_stock(root_memcg
, true);
2218 mutex_unlock(&percpu_charge_mutex
);
2222 * This function drains percpu counter value from DEAD cpu and
2223 * move it to local cpu. Note that this function can be preempted.
2225 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2229 spin_lock(&memcg
->pcp_counter_lock
);
2230 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2231 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2233 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2234 memcg
->nocpu_base
.count
[i
] += x
;
2236 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2237 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2239 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2240 memcg
->nocpu_base
.events
[i
] += x
;
2242 spin_unlock(&memcg
->pcp_counter_lock
);
2245 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2246 unsigned long action
,
2249 int cpu
= (unsigned long)hcpu
;
2250 struct memcg_stock_pcp
*stock
;
2251 struct mem_cgroup
*iter
;
2253 if (action
== CPU_ONLINE
)
2256 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2259 for_each_mem_cgroup(iter
)
2260 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2262 stock
= &per_cpu(memcg_stock
, cpu
);
2268 /* See __mem_cgroup_try_charge() for details */
2270 CHARGE_OK
, /* success */
2271 CHARGE_RETRY
, /* need to retry but retry is not bad */
2272 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2273 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2274 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2277 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2278 unsigned int nr_pages
, unsigned int min_pages
,
2281 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2282 struct mem_cgroup
*mem_over_limit
;
2283 struct res_counter
*fail_res
;
2284 unsigned long flags
= 0;
2287 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2290 if (!do_swap_account
)
2292 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2296 res_counter_uncharge(&memcg
->res
, csize
);
2297 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2298 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2300 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2302 * Never reclaim on behalf of optional batching, retry with a
2303 * single page instead.
2305 if (nr_pages
> min_pages
)
2306 return CHARGE_RETRY
;
2308 if (!(gfp_mask
& __GFP_WAIT
))
2309 return CHARGE_WOULDBLOCK
;
2311 if (gfp_mask
& __GFP_NORETRY
)
2312 return CHARGE_NOMEM
;
2314 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2315 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2316 return CHARGE_RETRY
;
2318 * Even though the limit is exceeded at this point, reclaim
2319 * may have been able to free some pages. Retry the charge
2320 * before killing the task.
2322 * Only for regular pages, though: huge pages are rather
2323 * unlikely to succeed so close to the limit, and we fall back
2324 * to regular pages anyway in case of failure.
2326 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2327 return CHARGE_RETRY
;
2330 * At task move, charge accounts can be doubly counted. So, it's
2331 * better to wait until the end of task_move if something is going on.
2333 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2334 return CHARGE_RETRY
;
2336 /* If we don't need to call oom-killer at el, return immediately */
2338 return CHARGE_NOMEM
;
2340 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2341 return CHARGE_OOM_DIE
;
2343 return CHARGE_RETRY
;
2347 * __mem_cgroup_try_charge() does
2348 * 1. detect memcg to be charged against from passed *mm and *ptr,
2349 * 2. update res_counter
2350 * 3. call memory reclaim if necessary.
2352 * In some special case, if the task is fatal, fatal_signal_pending() or
2353 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2354 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2355 * as possible without any hazards. 2: all pages should have a valid
2356 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2357 * pointer, that is treated as a charge to root_mem_cgroup.
2359 * So __mem_cgroup_try_charge() will return
2360 * 0 ... on success, filling *ptr with a valid memcg pointer.
2361 * -ENOMEM ... charge failure because of resource limits.
2362 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2364 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2365 * the oom-killer can be invoked.
2367 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2369 unsigned int nr_pages
,
2370 struct mem_cgroup
**ptr
,
2373 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2374 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2375 struct mem_cgroup
*memcg
= NULL
;
2379 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2380 * in system level. So, allow to go ahead dying process in addition to
2383 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2384 || fatal_signal_pending(current
)))
2388 * We always charge the cgroup the mm_struct belongs to.
2389 * The mm_struct's mem_cgroup changes on task migration if the
2390 * thread group leader migrates. It's possible that mm is not
2391 * set, if so charge the root memcg (happens for pagecache usage).
2394 *ptr
= root_mem_cgroup
;
2396 if (*ptr
) { /* css should be a valid one */
2398 if (mem_cgroup_is_root(memcg
))
2400 if (consume_stock(memcg
, nr_pages
))
2402 css_get(&memcg
->css
);
2404 struct task_struct
*p
;
2407 p
= rcu_dereference(mm
->owner
);
2409 * Because we don't have task_lock(), "p" can exit.
2410 * In that case, "memcg" can point to root or p can be NULL with
2411 * race with swapoff. Then, we have small risk of mis-accouning.
2412 * But such kind of mis-account by race always happens because
2413 * we don't have cgroup_mutex(). It's overkill and we allo that
2415 * (*) swapoff at el will charge against mm-struct not against
2416 * task-struct. So, mm->owner can be NULL.
2418 memcg
= mem_cgroup_from_task(p
);
2420 memcg
= root_mem_cgroup
;
2421 if (mem_cgroup_is_root(memcg
)) {
2425 if (consume_stock(memcg
, nr_pages
)) {
2427 * It seems dagerous to access memcg without css_get().
2428 * But considering how consume_stok works, it's not
2429 * necessary. If consume_stock success, some charges
2430 * from this memcg are cached on this cpu. So, we
2431 * don't need to call css_get()/css_tryget() before
2432 * calling consume_stock().
2437 /* after here, we may be blocked. we need to get refcnt */
2438 if (!css_tryget(&memcg
->css
)) {
2448 /* If killed, bypass charge */
2449 if (fatal_signal_pending(current
)) {
2450 css_put(&memcg
->css
);
2455 if (oom
&& !nr_oom_retries
) {
2457 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2460 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2465 case CHARGE_RETRY
: /* not in OOM situation but retry */
2467 css_put(&memcg
->css
);
2470 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2471 css_put(&memcg
->css
);
2473 case CHARGE_NOMEM
: /* OOM routine works */
2475 css_put(&memcg
->css
);
2478 /* If oom, we never return -ENOMEM */
2481 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2482 css_put(&memcg
->css
);
2485 } while (ret
!= CHARGE_OK
);
2487 if (batch
> nr_pages
)
2488 refill_stock(memcg
, batch
- nr_pages
);
2489 css_put(&memcg
->css
);
2497 *ptr
= root_mem_cgroup
;
2502 * Somemtimes we have to undo a charge we got by try_charge().
2503 * This function is for that and do uncharge, put css's refcnt.
2504 * gotten by try_charge().
2506 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2507 unsigned int nr_pages
)
2509 if (!mem_cgroup_is_root(memcg
)) {
2510 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2512 res_counter_uncharge(&memcg
->res
, bytes
);
2513 if (do_swap_account
)
2514 res_counter_uncharge(&memcg
->memsw
, bytes
);
2519 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2520 * This is useful when moving usage to parent cgroup.
2522 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2523 unsigned int nr_pages
)
2525 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2527 if (mem_cgroup_is_root(memcg
))
2530 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2531 if (do_swap_account
)
2532 res_counter_uncharge_until(&memcg
->memsw
,
2533 memcg
->memsw
.parent
, bytes
);
2537 * A helper function to get mem_cgroup from ID. must be called under
2538 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2539 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2540 * called against removed memcg.)
2542 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2544 struct cgroup_subsys_state
*css
;
2546 /* ID 0 is unused ID */
2549 css
= css_lookup(&mem_cgroup_subsys
, id
);
2552 return mem_cgroup_from_css(css
);
2555 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2557 struct mem_cgroup
*memcg
= NULL
;
2558 struct page_cgroup
*pc
;
2562 VM_BUG_ON(!PageLocked(page
));
2564 pc
= lookup_page_cgroup(page
);
2565 lock_page_cgroup(pc
);
2566 if (PageCgroupUsed(pc
)) {
2567 memcg
= pc
->mem_cgroup
;
2568 if (memcg
&& !css_tryget(&memcg
->css
))
2570 } else if (PageSwapCache(page
)) {
2571 ent
.val
= page_private(page
);
2572 id
= lookup_swap_cgroup_id(ent
);
2574 memcg
= mem_cgroup_lookup(id
);
2575 if (memcg
&& !css_tryget(&memcg
->css
))
2579 unlock_page_cgroup(pc
);
2583 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2585 unsigned int nr_pages
,
2586 enum charge_type ctype
,
2589 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2590 struct zone
*uninitialized_var(zone
);
2591 struct lruvec
*lruvec
;
2592 bool was_on_lru
= false;
2595 lock_page_cgroup(pc
);
2596 VM_BUG_ON(PageCgroupUsed(pc
));
2598 * we don't need page_cgroup_lock about tail pages, becase they are not
2599 * accessed by any other context at this point.
2603 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2604 * may already be on some other mem_cgroup's LRU. Take care of it.
2607 zone
= page_zone(page
);
2608 spin_lock_irq(&zone
->lru_lock
);
2609 if (PageLRU(page
)) {
2610 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2612 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2617 pc
->mem_cgroup
= memcg
;
2619 * We access a page_cgroup asynchronously without lock_page_cgroup().
2620 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2621 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2622 * before USED bit, we need memory barrier here.
2623 * See mem_cgroup_add_lru_list(), etc.
2626 SetPageCgroupUsed(pc
);
2630 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2631 VM_BUG_ON(PageLRU(page
));
2633 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2635 spin_unlock_irq(&zone
->lru_lock
);
2638 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2643 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2644 unlock_page_cgroup(pc
);
2647 * "charge_statistics" updated event counter.
2649 memcg_check_events(memcg
, page
);
2652 static DEFINE_MUTEX(set_limit_mutex
);
2654 #ifdef CONFIG_MEMCG_KMEM
2655 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2657 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2658 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2662 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2663 * in the memcg_cache_params struct.
2665 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2667 struct kmem_cache
*cachep
;
2669 VM_BUG_ON(p
->is_root_cache
);
2670 cachep
= p
->root_cache
;
2671 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2674 #ifdef CONFIG_SLABINFO
2675 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2676 struct cftype
*cft
, struct seq_file
*m
)
2678 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2679 struct memcg_cache_params
*params
;
2681 if (!memcg_can_account_kmem(memcg
))
2684 print_slabinfo_header(m
);
2686 mutex_lock(&memcg
->slab_caches_mutex
);
2687 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2688 cache_show(memcg_params_to_cache(params
), m
);
2689 mutex_unlock(&memcg
->slab_caches_mutex
);
2695 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2697 struct res_counter
*fail_res
;
2698 struct mem_cgroup
*_memcg
;
2702 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2707 * Conditions under which we can wait for the oom_killer. Those are
2708 * the same conditions tested by the core page allocator
2710 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2713 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2716 if (ret
== -EINTR
) {
2718 * __mem_cgroup_try_charge() chosed to bypass to root due to
2719 * OOM kill or fatal signal. Since our only options are to
2720 * either fail the allocation or charge it to this cgroup, do
2721 * it as a temporary condition. But we can't fail. From a
2722 * kmem/slab perspective, the cache has already been selected,
2723 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2726 * This condition will only trigger if the task entered
2727 * memcg_charge_kmem in a sane state, but was OOM-killed during
2728 * __mem_cgroup_try_charge() above. Tasks that were already
2729 * dying when the allocation triggers should have been already
2730 * directed to the root cgroup in memcontrol.h
2732 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2733 if (do_swap_account
)
2734 res_counter_charge_nofail(&memcg
->memsw
, size
,
2738 res_counter_uncharge(&memcg
->kmem
, size
);
2743 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2745 res_counter_uncharge(&memcg
->res
, size
);
2746 if (do_swap_account
)
2747 res_counter_uncharge(&memcg
->memsw
, size
);
2750 if (res_counter_uncharge(&memcg
->kmem
, size
))
2754 * Releases a reference taken in kmem_cgroup_css_offline in case
2755 * this last uncharge is racing with the offlining code or it is
2756 * outliving the memcg existence.
2758 * The memory barrier imposed by test&clear is paired with the
2759 * explicit one in memcg_kmem_mark_dead().
2761 if (memcg_kmem_test_and_clear_dead(memcg
))
2762 css_put(&memcg
->css
);
2765 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2770 mutex_lock(&memcg
->slab_caches_mutex
);
2771 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2772 mutex_unlock(&memcg
->slab_caches_mutex
);
2776 * helper for acessing a memcg's index. It will be used as an index in the
2777 * child cache array in kmem_cache, and also to derive its name. This function
2778 * will return -1 when this is not a kmem-limited memcg.
2780 int memcg_cache_id(struct mem_cgroup
*memcg
)
2782 return memcg
? memcg
->kmemcg_id
: -1;
2786 * This ends up being protected by the set_limit mutex, during normal
2787 * operation, because that is its main call site.
2789 * But when we create a new cache, we can call this as well if its parent
2790 * is kmem-limited. That will have to hold set_limit_mutex as well.
2792 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2796 num
= ida_simple_get(&kmem_limited_groups
,
2797 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2801 * After this point, kmem_accounted (that we test atomically in
2802 * the beginning of this conditional), is no longer 0. This
2803 * guarantees only one process will set the following boolean
2804 * to true. We don't need test_and_set because we're protected
2805 * by the set_limit_mutex anyway.
2807 memcg_kmem_set_activated(memcg
);
2809 ret
= memcg_update_all_caches(num
+1);
2811 ida_simple_remove(&kmem_limited_groups
, num
);
2812 memcg_kmem_clear_activated(memcg
);
2816 memcg
->kmemcg_id
= num
;
2817 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
2818 mutex_init(&memcg
->slab_caches_mutex
);
2822 static size_t memcg_caches_array_size(int num_groups
)
2825 if (num_groups
<= 0)
2828 size
= 2 * num_groups
;
2829 if (size
< MEMCG_CACHES_MIN_SIZE
)
2830 size
= MEMCG_CACHES_MIN_SIZE
;
2831 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2832 size
= MEMCG_CACHES_MAX_SIZE
;
2838 * We should update the current array size iff all caches updates succeed. This
2839 * can only be done from the slab side. The slab mutex needs to be held when
2842 void memcg_update_array_size(int num
)
2844 if (num
> memcg_limited_groups_array_size
)
2845 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2848 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
2850 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2852 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2854 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
2856 if (num_groups
> memcg_limited_groups_array_size
) {
2858 ssize_t size
= memcg_caches_array_size(num_groups
);
2860 size
*= sizeof(void *);
2861 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
2863 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2864 if (!s
->memcg_params
) {
2865 s
->memcg_params
= cur_params
;
2869 s
->memcg_params
->is_root_cache
= true;
2872 * There is the chance it will be bigger than
2873 * memcg_limited_groups_array_size, if we failed an allocation
2874 * in a cache, in which case all caches updated before it, will
2875 * have a bigger array.
2877 * But if that is the case, the data after
2878 * memcg_limited_groups_array_size is certainly unused
2880 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
2881 if (!cur_params
->memcg_caches
[i
])
2883 s
->memcg_params
->memcg_caches
[i
] =
2884 cur_params
->memcg_caches
[i
];
2888 * Ideally, we would wait until all caches succeed, and only
2889 * then free the old one. But this is not worth the extra
2890 * pointer per-cache we'd have to have for this.
2892 * It is not a big deal if some caches are left with a size
2893 * bigger than the others. And all updates will reset this
2901 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
2902 struct kmem_cache
*root_cache
)
2906 if (!memcg_kmem_enabled())
2910 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
2911 size
+= memcg_limited_groups_array_size
* sizeof(void *);
2913 size
= sizeof(struct memcg_cache_params
);
2915 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2916 if (!s
->memcg_params
)
2920 s
->memcg_params
->memcg
= memcg
;
2921 s
->memcg_params
->root_cache
= root_cache
;
2922 INIT_WORK(&s
->memcg_params
->destroy
,
2923 kmem_cache_destroy_work_func
);
2925 s
->memcg_params
->is_root_cache
= true;
2930 void memcg_release_cache(struct kmem_cache
*s
)
2932 struct kmem_cache
*root
;
2933 struct mem_cgroup
*memcg
;
2937 * This happens, for instance, when a root cache goes away before we
2940 if (!s
->memcg_params
)
2943 if (s
->memcg_params
->is_root_cache
)
2946 memcg
= s
->memcg_params
->memcg
;
2947 id
= memcg_cache_id(memcg
);
2949 root
= s
->memcg_params
->root_cache
;
2950 root
->memcg_params
->memcg_caches
[id
] = NULL
;
2952 mutex_lock(&memcg
->slab_caches_mutex
);
2953 list_del(&s
->memcg_params
->list
);
2954 mutex_unlock(&memcg
->slab_caches_mutex
);
2956 css_put(&memcg
->css
);
2958 kfree(s
->memcg_params
);
2962 * During the creation a new cache, we need to disable our accounting mechanism
2963 * altogether. This is true even if we are not creating, but rather just
2964 * enqueing new caches to be created.
2966 * This is because that process will trigger allocations; some visible, like
2967 * explicit kmallocs to auxiliary data structures, name strings and internal
2968 * cache structures; some well concealed, like INIT_WORK() that can allocate
2969 * objects during debug.
2971 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2972 * to it. This may not be a bounded recursion: since the first cache creation
2973 * failed to complete (waiting on the allocation), we'll just try to create the
2974 * cache again, failing at the same point.
2976 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2977 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2978 * inside the following two functions.
2980 static inline void memcg_stop_kmem_account(void)
2982 VM_BUG_ON(!current
->mm
);
2983 current
->memcg_kmem_skip_account
++;
2986 static inline void memcg_resume_kmem_account(void)
2988 VM_BUG_ON(!current
->mm
);
2989 current
->memcg_kmem_skip_account
--;
2992 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
2994 struct kmem_cache
*cachep
;
2995 struct memcg_cache_params
*p
;
2997 p
= container_of(w
, struct memcg_cache_params
, destroy
);
2999 cachep
= memcg_params_to_cache(p
);
3002 * If we get down to 0 after shrink, we could delete right away.
3003 * However, memcg_release_pages() already puts us back in the workqueue
3004 * in that case. If we proceed deleting, we'll get a dangling
3005 * reference, and removing the object from the workqueue in that case
3006 * is unnecessary complication. We are not a fast path.
3008 * Note that this case is fundamentally different from racing with
3009 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3010 * kmem_cache_shrink, not only we would be reinserting a dead cache
3011 * into the queue, but doing so from inside the worker racing to
3014 * So if we aren't down to zero, we'll just schedule a worker and try
3017 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3018 kmem_cache_shrink(cachep
);
3019 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3022 kmem_cache_destroy(cachep
);
3025 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3027 if (!cachep
->memcg_params
->dead
)
3031 * There are many ways in which we can get here.
3033 * We can get to a memory-pressure situation while the delayed work is
3034 * still pending to run. The vmscan shrinkers can then release all
3035 * cache memory and get us to destruction. If this is the case, we'll
3036 * be executed twice, which is a bug (the second time will execute over
3037 * bogus data). In this case, cancelling the work should be fine.
3039 * But we can also get here from the worker itself, if
3040 * kmem_cache_shrink is enough to shake all the remaining objects and
3041 * get the page count to 0. In this case, we'll deadlock if we try to
3042 * cancel the work (the worker runs with an internal lock held, which
3043 * is the same lock we would hold for cancel_work_sync().)
3045 * Since we can't possibly know who got us here, just refrain from
3046 * running if there is already work pending
3048 if (work_pending(&cachep
->memcg_params
->destroy
))
3051 * We have to defer the actual destroying to a workqueue, because
3052 * we might currently be in a context that cannot sleep.
3054 schedule_work(&cachep
->memcg_params
->destroy
);
3058 * This lock protects updaters, not readers. We want readers to be as fast as
3059 * they can, and they will either see NULL or a valid cache value. Our model
3060 * allow them to see NULL, in which case the root memcg will be selected.
3062 * We need this lock because multiple allocations to the same cache from a non
3063 * will span more than one worker. Only one of them can create the cache.
3065 static DEFINE_MUTEX(memcg_cache_mutex
);
3068 * Called with memcg_cache_mutex held
3070 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3071 struct kmem_cache
*s
)
3073 struct kmem_cache
*new;
3074 static char *tmp_name
= NULL
;
3076 lockdep_assert_held(&memcg_cache_mutex
);
3079 * kmem_cache_create_memcg duplicates the given name and
3080 * cgroup_name for this name requires RCU context.
3081 * This static temporary buffer is used to prevent from
3082 * pointless shortliving allocation.
3085 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3091 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3092 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3095 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3096 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3099 new->allocflags
|= __GFP_KMEMCG
;
3104 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3105 struct kmem_cache
*cachep
)
3107 struct kmem_cache
*new_cachep
;
3110 BUG_ON(!memcg_can_account_kmem(memcg
));
3112 idx
= memcg_cache_id(memcg
);
3114 mutex_lock(&memcg_cache_mutex
);
3115 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3117 css_put(&memcg
->css
);
3121 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3122 if (new_cachep
== NULL
) {
3123 new_cachep
= cachep
;
3124 css_put(&memcg
->css
);
3128 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3130 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3132 * the readers won't lock, make sure everybody sees the updated value,
3133 * so they won't put stuff in the queue again for no reason
3137 mutex_unlock(&memcg_cache_mutex
);
3141 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3143 struct kmem_cache
*c
;
3146 if (!s
->memcg_params
)
3148 if (!s
->memcg_params
->is_root_cache
)
3152 * If the cache is being destroyed, we trust that there is no one else
3153 * requesting objects from it. Even if there are, the sanity checks in
3154 * kmem_cache_destroy should caught this ill-case.
3156 * Still, we don't want anyone else freeing memcg_caches under our
3157 * noses, which can happen if a new memcg comes to life. As usual,
3158 * we'll take the set_limit_mutex to protect ourselves against this.
3160 mutex_lock(&set_limit_mutex
);
3161 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3162 c
= s
->memcg_params
->memcg_caches
[i
];
3167 * We will now manually delete the caches, so to avoid races
3168 * we need to cancel all pending destruction workers and
3169 * proceed with destruction ourselves.
3171 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3172 * and that could spawn the workers again: it is likely that
3173 * the cache still have active pages until this very moment.
3174 * This would lead us back to mem_cgroup_destroy_cache.
3176 * But that will not execute at all if the "dead" flag is not
3177 * set, so flip it down to guarantee we are in control.
3179 c
->memcg_params
->dead
= false;
3180 cancel_work_sync(&c
->memcg_params
->destroy
);
3181 kmem_cache_destroy(c
);
3183 mutex_unlock(&set_limit_mutex
);
3186 struct create_work
{
3187 struct mem_cgroup
*memcg
;
3188 struct kmem_cache
*cachep
;
3189 struct work_struct work
;
3192 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3194 struct kmem_cache
*cachep
;
3195 struct memcg_cache_params
*params
;
3197 if (!memcg_kmem_is_active(memcg
))
3200 mutex_lock(&memcg
->slab_caches_mutex
);
3201 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3202 cachep
= memcg_params_to_cache(params
);
3203 cachep
->memcg_params
->dead
= true;
3204 schedule_work(&cachep
->memcg_params
->destroy
);
3206 mutex_unlock(&memcg
->slab_caches_mutex
);
3209 static void memcg_create_cache_work_func(struct work_struct
*w
)
3211 struct create_work
*cw
;
3213 cw
= container_of(w
, struct create_work
, work
);
3214 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3219 * Enqueue the creation of a per-memcg kmem_cache.
3221 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3222 struct kmem_cache
*cachep
)
3224 struct create_work
*cw
;
3226 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3228 css_put(&memcg
->css
);
3233 cw
->cachep
= cachep
;
3235 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3236 schedule_work(&cw
->work
);
3239 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3240 struct kmem_cache
*cachep
)
3243 * We need to stop accounting when we kmalloc, because if the
3244 * corresponding kmalloc cache is not yet created, the first allocation
3245 * in __memcg_create_cache_enqueue will recurse.
3247 * However, it is better to enclose the whole function. Depending on
3248 * the debugging options enabled, INIT_WORK(), for instance, can
3249 * trigger an allocation. This too, will make us recurse. Because at
3250 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3251 * the safest choice is to do it like this, wrapping the whole function.
3253 memcg_stop_kmem_account();
3254 __memcg_create_cache_enqueue(memcg
, cachep
);
3255 memcg_resume_kmem_account();
3258 * Return the kmem_cache we're supposed to use for a slab allocation.
3259 * We try to use the current memcg's version of the cache.
3261 * If the cache does not exist yet, if we are the first user of it,
3262 * we either create it immediately, if possible, or create it asynchronously
3264 * In the latter case, we will let the current allocation go through with
3265 * the original cache.
3267 * Can't be called in interrupt context or from kernel threads.
3268 * This function needs to be called with rcu_read_lock() held.
3270 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3273 struct mem_cgroup
*memcg
;
3276 VM_BUG_ON(!cachep
->memcg_params
);
3277 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3279 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3283 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3285 if (!memcg_can_account_kmem(memcg
))
3288 idx
= memcg_cache_id(memcg
);
3291 * barrier to mare sure we're always seeing the up to date value. The
3292 * code updating memcg_caches will issue a write barrier to match this.
3294 read_barrier_depends();
3295 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3296 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3300 /* The corresponding put will be done in the workqueue. */
3301 if (!css_tryget(&memcg
->css
))
3306 * If we are in a safe context (can wait, and not in interrupt
3307 * context), we could be be predictable and return right away.
3308 * This would guarantee that the allocation being performed
3309 * already belongs in the new cache.
3311 * However, there are some clashes that can arrive from locking.
3312 * For instance, because we acquire the slab_mutex while doing
3313 * kmem_cache_dup, this means no further allocation could happen
3314 * with the slab_mutex held.
3316 * Also, because cache creation issue get_online_cpus(), this
3317 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3318 * that ends up reversed during cpu hotplug. (cpuset allocates
3319 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3320 * better to defer everything.
3322 memcg_create_cache_enqueue(memcg
, cachep
);
3328 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3331 * We need to verify if the allocation against current->mm->owner's memcg is
3332 * possible for the given order. But the page is not allocated yet, so we'll
3333 * need a further commit step to do the final arrangements.
3335 * It is possible for the task to switch cgroups in this mean time, so at
3336 * commit time, we can't rely on task conversion any longer. We'll then use
3337 * the handle argument to return to the caller which cgroup we should commit
3338 * against. We could also return the memcg directly and avoid the pointer
3339 * passing, but a boolean return value gives better semantics considering
3340 * the compiled-out case as well.
3342 * Returning true means the allocation is possible.
3345 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3347 struct mem_cgroup
*memcg
;
3353 * Disabling accounting is only relevant for some specific memcg
3354 * internal allocations. Therefore we would initially not have such
3355 * check here, since direct calls to the page allocator that are marked
3356 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3357 * concerned with cache allocations, and by having this test at
3358 * memcg_kmem_get_cache, we are already able to relay the allocation to
3359 * the root cache and bypass the memcg cache altogether.
3361 * There is one exception, though: the SLUB allocator does not create
3362 * large order caches, but rather service large kmallocs directly from
3363 * the page allocator. Therefore, the following sequence when backed by
3364 * the SLUB allocator:
3366 * memcg_stop_kmem_account();
3367 * kmalloc(<large_number>)
3368 * memcg_resume_kmem_account();
3370 * would effectively ignore the fact that we should skip accounting,
3371 * since it will drive us directly to this function without passing
3372 * through the cache selector memcg_kmem_get_cache. Such large
3373 * allocations are extremely rare but can happen, for instance, for the
3374 * cache arrays. We bring this test here.
3376 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3379 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3382 * very rare case described in mem_cgroup_from_task. Unfortunately there
3383 * isn't much we can do without complicating this too much, and it would
3384 * be gfp-dependent anyway. Just let it go
3386 if (unlikely(!memcg
))
3389 if (!memcg_can_account_kmem(memcg
)) {
3390 css_put(&memcg
->css
);
3394 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3398 css_put(&memcg
->css
);
3402 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3405 struct page_cgroup
*pc
;
3407 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3409 /* The page allocation failed. Revert */
3411 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3415 pc
= lookup_page_cgroup(page
);
3416 lock_page_cgroup(pc
);
3417 pc
->mem_cgroup
= memcg
;
3418 SetPageCgroupUsed(pc
);
3419 unlock_page_cgroup(pc
);
3422 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3424 struct mem_cgroup
*memcg
= NULL
;
3425 struct page_cgroup
*pc
;
3428 pc
= lookup_page_cgroup(page
);
3430 * Fast unlocked return. Theoretically might have changed, have to
3431 * check again after locking.
3433 if (!PageCgroupUsed(pc
))
3436 lock_page_cgroup(pc
);
3437 if (PageCgroupUsed(pc
)) {
3438 memcg
= pc
->mem_cgroup
;
3439 ClearPageCgroupUsed(pc
);
3441 unlock_page_cgroup(pc
);
3444 * We trust that only if there is a memcg associated with the page, it
3445 * is a valid allocation
3450 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3451 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3454 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3457 #endif /* CONFIG_MEMCG_KMEM */
3459 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3461 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3463 * Because tail pages are not marked as "used", set it. We're under
3464 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3465 * charge/uncharge will be never happen and move_account() is done under
3466 * compound_lock(), so we don't have to take care of races.
3468 void mem_cgroup_split_huge_fixup(struct page
*head
)
3470 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3471 struct page_cgroup
*pc
;
3472 struct mem_cgroup
*memcg
;
3475 if (mem_cgroup_disabled())
3478 memcg
= head_pc
->mem_cgroup
;
3479 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3481 pc
->mem_cgroup
= memcg
;
3482 smp_wmb();/* see __commit_charge() */
3483 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3485 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3488 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3491 * mem_cgroup_move_account - move account of the page
3493 * @nr_pages: number of regular pages (>1 for huge pages)
3494 * @pc: page_cgroup of the page.
3495 * @from: mem_cgroup which the page is moved from.
3496 * @to: mem_cgroup which the page is moved to. @from != @to.
3498 * The caller must confirm following.
3499 * - page is not on LRU (isolate_page() is useful.)
3500 * - compound_lock is held when nr_pages > 1
3502 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3505 static int mem_cgroup_move_account(struct page
*page
,
3506 unsigned int nr_pages
,
3507 struct page_cgroup
*pc
,
3508 struct mem_cgroup
*from
,
3509 struct mem_cgroup
*to
)
3511 unsigned long flags
;
3513 bool anon
= PageAnon(page
);
3515 VM_BUG_ON(from
== to
);
3516 VM_BUG_ON(PageLRU(page
));
3518 * The page is isolated from LRU. So, collapse function
3519 * will not handle this page. But page splitting can happen.
3520 * Do this check under compound_page_lock(). The caller should
3524 if (nr_pages
> 1 && !PageTransHuge(page
))
3527 lock_page_cgroup(pc
);
3530 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3533 move_lock_mem_cgroup(from
, &flags
);
3535 if (!anon
&& page_mapped(page
)) {
3536 /* Update mapped_file data for mem_cgroup */
3538 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3539 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3542 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3544 /* caller should have done css_get */
3545 pc
->mem_cgroup
= to
;
3546 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3547 move_unlock_mem_cgroup(from
, &flags
);
3550 unlock_page_cgroup(pc
);
3554 memcg_check_events(to
, page
);
3555 memcg_check_events(from
, page
);
3561 * mem_cgroup_move_parent - moves page to the parent group
3562 * @page: the page to move
3563 * @pc: page_cgroup of the page
3564 * @child: page's cgroup
3566 * move charges to its parent or the root cgroup if the group has no
3567 * parent (aka use_hierarchy==0).
3568 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3569 * mem_cgroup_move_account fails) the failure is always temporary and
3570 * it signals a race with a page removal/uncharge or migration. In the
3571 * first case the page is on the way out and it will vanish from the LRU
3572 * on the next attempt and the call should be retried later.
3573 * Isolation from the LRU fails only if page has been isolated from
3574 * the LRU since we looked at it and that usually means either global
3575 * reclaim or migration going on. The page will either get back to the
3577 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3578 * (!PageCgroupUsed) or moved to a different group. The page will
3579 * disappear in the next attempt.
3581 static int mem_cgroup_move_parent(struct page
*page
,
3582 struct page_cgroup
*pc
,
3583 struct mem_cgroup
*child
)
3585 struct mem_cgroup
*parent
;
3586 unsigned int nr_pages
;
3587 unsigned long uninitialized_var(flags
);
3590 VM_BUG_ON(mem_cgroup_is_root(child
));
3593 if (!get_page_unless_zero(page
))
3595 if (isolate_lru_page(page
))
3598 nr_pages
= hpage_nr_pages(page
);
3600 parent
= parent_mem_cgroup(child
);
3602 * If no parent, move charges to root cgroup.
3605 parent
= root_mem_cgroup
;
3608 VM_BUG_ON(!PageTransHuge(page
));
3609 flags
= compound_lock_irqsave(page
);
3612 ret
= mem_cgroup_move_account(page
, nr_pages
,
3615 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3618 compound_unlock_irqrestore(page
, flags
);
3619 putback_lru_page(page
);
3627 * Charge the memory controller for page usage.
3629 * 0 if the charge was successful
3630 * < 0 if the cgroup is over its limit
3632 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3633 gfp_t gfp_mask
, enum charge_type ctype
)
3635 struct mem_cgroup
*memcg
= NULL
;
3636 unsigned int nr_pages
= 1;
3640 if (PageTransHuge(page
)) {
3641 nr_pages
<<= compound_order(page
);
3642 VM_BUG_ON(!PageTransHuge(page
));
3644 * Never OOM-kill a process for a huge page. The
3645 * fault handler will fall back to regular pages.
3650 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3653 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3657 int mem_cgroup_newpage_charge(struct page
*page
,
3658 struct mm_struct
*mm
, gfp_t gfp_mask
)
3660 if (mem_cgroup_disabled())
3662 VM_BUG_ON(page_mapped(page
));
3663 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3665 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3666 MEM_CGROUP_CHARGE_TYPE_ANON
);
3670 * While swap-in, try_charge -> commit or cancel, the page is locked.
3671 * And when try_charge() successfully returns, one refcnt to memcg without
3672 * struct page_cgroup is acquired. This refcnt will be consumed by
3673 * "commit()" or removed by "cancel()"
3675 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3678 struct mem_cgroup
**memcgp
)
3680 struct mem_cgroup
*memcg
;
3681 struct page_cgroup
*pc
;
3684 pc
= lookup_page_cgroup(page
);
3686 * Every swap fault against a single page tries to charge the
3687 * page, bail as early as possible. shmem_unuse() encounters
3688 * already charged pages, too. The USED bit is protected by
3689 * the page lock, which serializes swap cache removal, which
3690 * in turn serializes uncharging.
3692 if (PageCgroupUsed(pc
))
3694 if (!do_swap_account
)
3696 memcg
= try_get_mem_cgroup_from_page(page
);
3700 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3701 css_put(&memcg
->css
);
3706 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3712 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3713 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3716 if (mem_cgroup_disabled())
3719 * A racing thread's fault, or swapoff, may have already
3720 * updated the pte, and even removed page from swap cache: in
3721 * those cases unuse_pte()'s pte_same() test will fail; but
3722 * there's also a KSM case which does need to charge the page.
3724 if (!PageSwapCache(page
)) {
3727 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3732 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3735 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3737 if (mem_cgroup_disabled())
3741 __mem_cgroup_cancel_charge(memcg
, 1);
3745 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3746 enum charge_type ctype
)
3748 if (mem_cgroup_disabled())
3753 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3755 * Now swap is on-memory. This means this page may be
3756 * counted both as mem and swap....double count.
3757 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3758 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3759 * may call delete_from_swap_cache() before reach here.
3761 if (do_swap_account
&& PageSwapCache(page
)) {
3762 swp_entry_t ent
= {.val
= page_private(page
)};
3763 mem_cgroup_uncharge_swap(ent
);
3767 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3768 struct mem_cgroup
*memcg
)
3770 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3771 MEM_CGROUP_CHARGE_TYPE_ANON
);
3774 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3777 struct mem_cgroup
*memcg
= NULL
;
3778 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3781 if (mem_cgroup_disabled())
3783 if (PageCompound(page
))
3786 if (!PageSwapCache(page
))
3787 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3788 else { /* page is swapcache/shmem */
3789 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3792 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3797 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3798 unsigned int nr_pages
,
3799 const enum charge_type ctype
)
3801 struct memcg_batch_info
*batch
= NULL
;
3802 bool uncharge_memsw
= true;
3804 /* If swapout, usage of swap doesn't decrease */
3805 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3806 uncharge_memsw
= false;
3808 batch
= ¤t
->memcg_batch
;
3810 * In usual, we do css_get() when we remember memcg pointer.
3811 * But in this case, we keep res->usage until end of a series of
3812 * uncharges. Then, it's ok to ignore memcg's refcnt.
3815 batch
->memcg
= memcg
;
3817 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3818 * In those cases, all pages freed continuously can be expected to be in
3819 * the same cgroup and we have chance to coalesce uncharges.
3820 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3821 * because we want to do uncharge as soon as possible.
3824 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3825 goto direct_uncharge
;
3828 goto direct_uncharge
;
3831 * In typical case, batch->memcg == mem. This means we can
3832 * merge a series of uncharges to an uncharge of res_counter.
3833 * If not, we uncharge res_counter ony by one.
3835 if (batch
->memcg
!= memcg
)
3836 goto direct_uncharge
;
3837 /* remember freed charge and uncharge it later */
3840 batch
->memsw_nr_pages
++;
3843 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3845 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3846 if (unlikely(batch
->memcg
!= memcg
))
3847 memcg_oom_recover(memcg
);
3851 * uncharge if !page_mapped(page)
3853 static struct mem_cgroup
*
3854 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3857 struct mem_cgroup
*memcg
= NULL
;
3858 unsigned int nr_pages
= 1;
3859 struct page_cgroup
*pc
;
3862 if (mem_cgroup_disabled())
3865 if (PageTransHuge(page
)) {
3866 nr_pages
<<= compound_order(page
);
3867 VM_BUG_ON(!PageTransHuge(page
));
3870 * Check if our page_cgroup is valid
3872 pc
= lookup_page_cgroup(page
);
3873 if (unlikely(!PageCgroupUsed(pc
)))
3876 lock_page_cgroup(pc
);
3878 memcg
= pc
->mem_cgroup
;
3880 if (!PageCgroupUsed(pc
))
3883 anon
= PageAnon(page
);
3886 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3888 * Generally PageAnon tells if it's the anon statistics to be
3889 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3890 * used before page reached the stage of being marked PageAnon.
3894 case MEM_CGROUP_CHARGE_TYPE_DROP
:
3895 /* See mem_cgroup_prepare_migration() */
3896 if (page_mapped(page
))
3899 * Pages under migration may not be uncharged. But
3900 * end_migration() /must/ be the one uncharging the
3901 * unused post-migration page and so it has to call
3902 * here with the migration bit still set. See the
3903 * res_counter handling below.
3905 if (!end_migration
&& PageCgroupMigration(pc
))
3908 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
3909 if (!PageAnon(page
)) { /* Shared memory */
3910 if (page
->mapping
&& !page_is_file_cache(page
))
3912 } else if (page_mapped(page
)) /* Anon */
3919 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
3921 ClearPageCgroupUsed(pc
);
3923 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3924 * freed from LRU. This is safe because uncharged page is expected not
3925 * to be reused (freed soon). Exception is SwapCache, it's handled by
3926 * special functions.
3929 unlock_page_cgroup(pc
);
3931 * even after unlock, we have memcg->res.usage here and this memcg
3932 * will never be freed, so it's safe to call css_get().
3934 memcg_check_events(memcg
, page
);
3935 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
3936 mem_cgroup_swap_statistics(memcg
, true);
3937 css_get(&memcg
->css
);
3940 * Migration does not charge the res_counter for the
3941 * replacement page, so leave it alone when phasing out the
3942 * page that is unused after the migration.
3944 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
3945 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
3950 unlock_page_cgroup(pc
);
3954 void mem_cgroup_uncharge_page(struct page
*page
)
3957 if (page_mapped(page
))
3959 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3961 * If the page is in swap cache, uncharge should be deferred
3962 * to the swap path, which also properly accounts swap usage
3963 * and handles memcg lifetime.
3965 * Note that this check is not stable and reclaim may add the
3966 * page to swap cache at any time after this. However, if the
3967 * page is not in swap cache by the time page->mapcount hits
3968 * 0, there won't be any page table references to the swap
3969 * slot, and reclaim will free it and not actually write the
3972 if (PageSwapCache(page
))
3974 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
3977 void mem_cgroup_uncharge_cache_page(struct page
*page
)
3979 VM_BUG_ON(page_mapped(page
));
3980 VM_BUG_ON(page
->mapping
);
3981 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
3985 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3986 * In that cases, pages are freed continuously and we can expect pages
3987 * are in the same memcg. All these calls itself limits the number of
3988 * pages freed at once, then uncharge_start/end() is called properly.
3989 * This may be called prural(2) times in a context,
3992 void mem_cgroup_uncharge_start(void)
3994 current
->memcg_batch
.do_batch
++;
3995 /* We can do nest. */
3996 if (current
->memcg_batch
.do_batch
== 1) {
3997 current
->memcg_batch
.memcg
= NULL
;
3998 current
->memcg_batch
.nr_pages
= 0;
3999 current
->memcg_batch
.memsw_nr_pages
= 0;
4003 void mem_cgroup_uncharge_end(void)
4005 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4007 if (!batch
->do_batch
)
4011 if (batch
->do_batch
) /* If stacked, do nothing. */
4017 * This "batch->memcg" is valid without any css_get/put etc...
4018 * bacause we hide charges behind us.
4020 if (batch
->nr_pages
)
4021 res_counter_uncharge(&batch
->memcg
->res
,
4022 batch
->nr_pages
* PAGE_SIZE
);
4023 if (batch
->memsw_nr_pages
)
4024 res_counter_uncharge(&batch
->memcg
->memsw
,
4025 batch
->memsw_nr_pages
* PAGE_SIZE
);
4026 memcg_oom_recover(batch
->memcg
);
4027 /* forget this pointer (for sanity check) */
4028 batch
->memcg
= NULL
;
4033 * called after __delete_from_swap_cache() and drop "page" account.
4034 * memcg information is recorded to swap_cgroup of "ent"
4037 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4039 struct mem_cgroup
*memcg
;
4040 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4042 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4043 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4045 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4048 * record memcg information, if swapout && memcg != NULL,
4049 * css_get() was called in uncharge().
4051 if (do_swap_account
&& swapout
&& memcg
)
4052 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4056 #ifdef CONFIG_MEMCG_SWAP
4058 * called from swap_entry_free(). remove record in swap_cgroup and
4059 * uncharge "memsw" account.
4061 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4063 struct mem_cgroup
*memcg
;
4066 if (!do_swap_account
)
4069 id
= swap_cgroup_record(ent
, 0);
4071 memcg
= mem_cgroup_lookup(id
);
4074 * We uncharge this because swap is freed.
4075 * This memcg can be obsolete one. We avoid calling css_tryget
4077 if (!mem_cgroup_is_root(memcg
))
4078 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4079 mem_cgroup_swap_statistics(memcg
, false);
4080 css_put(&memcg
->css
);
4086 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4087 * @entry: swap entry to be moved
4088 * @from: mem_cgroup which the entry is moved from
4089 * @to: mem_cgroup which the entry is moved to
4091 * It succeeds only when the swap_cgroup's record for this entry is the same
4092 * as the mem_cgroup's id of @from.
4094 * Returns 0 on success, -EINVAL on failure.
4096 * The caller must have charged to @to, IOW, called res_counter_charge() about
4097 * both res and memsw, and called css_get().
4099 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4100 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4102 unsigned short old_id
, new_id
;
4104 old_id
= css_id(&from
->css
);
4105 new_id
= css_id(&to
->css
);
4107 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4108 mem_cgroup_swap_statistics(from
, false);
4109 mem_cgroup_swap_statistics(to
, true);
4111 * This function is only called from task migration context now.
4112 * It postpones res_counter and refcount handling till the end
4113 * of task migration(mem_cgroup_clear_mc()) for performance
4114 * improvement. But we cannot postpone css_get(to) because if
4115 * the process that has been moved to @to does swap-in, the
4116 * refcount of @to might be decreased to 0.
4118 * We are in attach() phase, so the cgroup is guaranteed to be
4119 * alive, so we can just call css_get().
4127 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4128 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4135 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4138 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4139 struct mem_cgroup
**memcgp
)
4141 struct mem_cgroup
*memcg
= NULL
;
4142 unsigned int nr_pages
= 1;
4143 struct page_cgroup
*pc
;
4144 enum charge_type ctype
;
4148 if (mem_cgroup_disabled())
4151 if (PageTransHuge(page
))
4152 nr_pages
<<= compound_order(page
);
4154 pc
= lookup_page_cgroup(page
);
4155 lock_page_cgroup(pc
);
4156 if (PageCgroupUsed(pc
)) {
4157 memcg
= pc
->mem_cgroup
;
4158 css_get(&memcg
->css
);
4160 * At migrating an anonymous page, its mapcount goes down
4161 * to 0 and uncharge() will be called. But, even if it's fully
4162 * unmapped, migration may fail and this page has to be
4163 * charged again. We set MIGRATION flag here and delay uncharge
4164 * until end_migration() is called
4166 * Corner Case Thinking
4168 * When the old page was mapped as Anon and it's unmap-and-freed
4169 * while migration was ongoing.
4170 * If unmap finds the old page, uncharge() of it will be delayed
4171 * until end_migration(). If unmap finds a new page, it's
4172 * uncharged when it make mapcount to be 1->0. If unmap code
4173 * finds swap_migration_entry, the new page will not be mapped
4174 * and end_migration() will find it(mapcount==0).
4177 * When the old page was mapped but migraion fails, the kernel
4178 * remaps it. A charge for it is kept by MIGRATION flag even
4179 * if mapcount goes down to 0. We can do remap successfully
4180 * without charging it again.
4183 * The "old" page is under lock_page() until the end of
4184 * migration, so, the old page itself will not be swapped-out.
4185 * If the new page is swapped out before end_migraton, our
4186 * hook to usual swap-out path will catch the event.
4189 SetPageCgroupMigration(pc
);
4191 unlock_page_cgroup(pc
);
4193 * If the page is not charged at this point,
4201 * We charge new page before it's used/mapped. So, even if unlock_page()
4202 * is called before end_migration, we can catch all events on this new
4203 * page. In the case new page is migrated but not remapped, new page's
4204 * mapcount will be finally 0 and we call uncharge in end_migration().
4207 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4209 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4211 * The page is committed to the memcg, but it's not actually
4212 * charged to the res_counter since we plan on replacing the
4213 * old one and only one page is going to be left afterwards.
4215 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4218 /* remove redundant charge if migration failed*/
4219 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4220 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4222 struct page
*used
, *unused
;
4223 struct page_cgroup
*pc
;
4229 if (!migration_ok
) {
4236 anon
= PageAnon(used
);
4237 __mem_cgroup_uncharge_common(unused
,
4238 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4239 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4241 css_put(&memcg
->css
);
4243 * We disallowed uncharge of pages under migration because mapcount
4244 * of the page goes down to zero, temporarly.
4245 * Clear the flag and check the page should be charged.
4247 pc
= lookup_page_cgroup(oldpage
);
4248 lock_page_cgroup(pc
);
4249 ClearPageCgroupMigration(pc
);
4250 unlock_page_cgroup(pc
);
4253 * If a page is a file cache, radix-tree replacement is very atomic
4254 * and we can skip this check. When it was an Anon page, its mapcount
4255 * goes down to 0. But because we added MIGRATION flage, it's not
4256 * uncharged yet. There are several case but page->mapcount check
4257 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4258 * check. (see prepare_charge() also)
4261 mem_cgroup_uncharge_page(used
);
4265 * At replace page cache, newpage is not under any memcg but it's on
4266 * LRU. So, this function doesn't touch res_counter but handles LRU
4267 * in correct way. Both pages are locked so we cannot race with uncharge.
4269 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4270 struct page
*newpage
)
4272 struct mem_cgroup
*memcg
= NULL
;
4273 struct page_cgroup
*pc
;
4274 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4276 if (mem_cgroup_disabled())
4279 pc
= lookup_page_cgroup(oldpage
);
4280 /* fix accounting on old pages */
4281 lock_page_cgroup(pc
);
4282 if (PageCgroupUsed(pc
)) {
4283 memcg
= pc
->mem_cgroup
;
4284 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4285 ClearPageCgroupUsed(pc
);
4287 unlock_page_cgroup(pc
);
4290 * When called from shmem_replace_page(), in some cases the
4291 * oldpage has already been charged, and in some cases not.
4296 * Even if newpage->mapping was NULL before starting replacement,
4297 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4298 * LRU while we overwrite pc->mem_cgroup.
4300 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4303 #ifdef CONFIG_DEBUG_VM
4304 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4306 struct page_cgroup
*pc
;
4308 pc
= lookup_page_cgroup(page
);
4310 * Can be NULL while feeding pages into the page allocator for
4311 * the first time, i.e. during boot or memory hotplug;
4312 * or when mem_cgroup_disabled().
4314 if (likely(pc
) && PageCgroupUsed(pc
))
4319 bool mem_cgroup_bad_page_check(struct page
*page
)
4321 if (mem_cgroup_disabled())
4324 return lookup_page_cgroup_used(page
) != NULL
;
4327 void mem_cgroup_print_bad_page(struct page
*page
)
4329 struct page_cgroup
*pc
;
4331 pc
= lookup_page_cgroup_used(page
);
4333 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4334 pc
, pc
->flags
, pc
->mem_cgroup
);
4339 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4340 unsigned long long val
)
4343 u64 memswlimit
, memlimit
;
4345 int children
= mem_cgroup_count_children(memcg
);
4346 u64 curusage
, oldusage
;
4350 * For keeping hierarchical_reclaim simple, how long we should retry
4351 * is depends on callers. We set our retry-count to be function
4352 * of # of children which we should visit in this loop.
4354 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4356 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4359 while (retry_count
) {
4360 if (signal_pending(current
)) {
4365 * Rather than hide all in some function, I do this in
4366 * open coded manner. You see what this really does.
4367 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4369 mutex_lock(&set_limit_mutex
);
4370 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4371 if (memswlimit
< val
) {
4373 mutex_unlock(&set_limit_mutex
);
4377 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4381 ret
= res_counter_set_limit(&memcg
->res
, val
);
4383 if (memswlimit
== val
)
4384 memcg
->memsw_is_minimum
= true;
4386 memcg
->memsw_is_minimum
= false;
4388 mutex_unlock(&set_limit_mutex
);
4393 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4394 MEM_CGROUP_RECLAIM_SHRINK
);
4395 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4396 /* Usage is reduced ? */
4397 if (curusage
>= oldusage
)
4400 oldusage
= curusage
;
4402 if (!ret
&& enlarge
)
4403 memcg_oom_recover(memcg
);
4408 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4409 unsigned long long val
)
4412 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4413 int children
= mem_cgroup_count_children(memcg
);
4417 /* see mem_cgroup_resize_res_limit */
4418 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4419 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4420 while (retry_count
) {
4421 if (signal_pending(current
)) {
4426 * Rather than hide all in some function, I do this in
4427 * open coded manner. You see what this really does.
4428 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4430 mutex_lock(&set_limit_mutex
);
4431 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4432 if (memlimit
> val
) {
4434 mutex_unlock(&set_limit_mutex
);
4437 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4438 if (memswlimit
< val
)
4440 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4442 if (memlimit
== val
)
4443 memcg
->memsw_is_minimum
= true;
4445 memcg
->memsw_is_minimum
= false;
4447 mutex_unlock(&set_limit_mutex
);
4452 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4453 MEM_CGROUP_RECLAIM_NOSWAP
|
4454 MEM_CGROUP_RECLAIM_SHRINK
);
4455 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4456 /* Usage is reduced ? */
4457 if (curusage
>= oldusage
)
4460 oldusage
= curusage
;
4462 if (!ret
&& enlarge
)
4463 memcg_oom_recover(memcg
);
4468 * mem_cgroup_force_empty_list - clears LRU of a group
4469 * @memcg: group to clear
4472 * @lru: lru to to clear
4474 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4475 * reclaim the pages page themselves - pages are moved to the parent (or root)
4478 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4479 int node
, int zid
, enum lru_list lru
)
4481 struct lruvec
*lruvec
;
4482 unsigned long flags
;
4483 struct list_head
*list
;
4487 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4488 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4489 list
= &lruvec
->lists
[lru
];
4493 struct page_cgroup
*pc
;
4496 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4497 if (list_empty(list
)) {
4498 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4501 page
= list_entry(list
->prev
, struct page
, lru
);
4503 list_move(&page
->lru
, list
);
4505 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4508 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4510 pc
= lookup_page_cgroup(page
);
4512 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4513 /* found lock contention or "pc" is obsolete. */
4518 } while (!list_empty(list
));
4522 * make mem_cgroup's charge to be 0 if there is no task by moving
4523 * all the charges and pages to the parent.
4524 * This enables deleting this mem_cgroup.
4526 * Caller is responsible for holding css reference on the memcg.
4528 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4534 /* This is for making all *used* pages to be on LRU. */
4535 lru_add_drain_all();
4536 drain_all_stock_sync(memcg
);
4537 mem_cgroup_start_move(memcg
);
4538 for_each_node_state(node
, N_MEMORY
) {
4539 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4542 mem_cgroup_force_empty_list(memcg
,
4547 mem_cgroup_end_move(memcg
);
4548 memcg_oom_recover(memcg
);
4552 * Kernel memory may not necessarily be trackable to a specific
4553 * process. So they are not migrated, and therefore we can't
4554 * expect their value to drop to 0 here.
4555 * Having res filled up with kmem only is enough.
4557 * This is a safety check because mem_cgroup_force_empty_list
4558 * could have raced with mem_cgroup_replace_page_cache callers
4559 * so the lru seemed empty but the page could have been added
4560 * right after the check. RES_USAGE should be safe as we always
4561 * charge before adding to the LRU.
4563 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4564 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4565 } while (usage
> 0);
4569 * This mainly exists for tests during the setting of set of use_hierarchy.
4570 * Since this is the very setting we are changing, the current hierarchy value
4573 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4575 struct cgroup_subsys_state
*pos
;
4577 /* bounce at first found */
4578 css_for_each_child(pos
, &memcg
->css
)
4584 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4585 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4586 * from mem_cgroup_count_children(), in the sense that we don't really care how
4587 * many children we have; we only need to know if we have any. It also counts
4588 * any memcg without hierarchy as infertile.
4590 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4592 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4596 * Reclaims as many pages from the given memcg as possible and moves
4597 * the rest to the parent.
4599 * Caller is responsible for holding css reference for memcg.
4601 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4603 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4604 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4606 /* returns EBUSY if there is a task or if we come here twice. */
4607 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4610 /* we call try-to-free pages for make this cgroup empty */
4611 lru_add_drain_all();
4612 /* try to free all pages in this cgroup */
4613 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4616 if (signal_pending(current
))
4619 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4623 /* maybe some writeback is necessary */
4624 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4629 mem_cgroup_reparent_charges(memcg
);
4634 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4637 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4639 if (mem_cgroup_is_root(memcg
))
4641 return mem_cgroup_force_empty(memcg
);
4644 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4647 return mem_cgroup_from_css(css
)->use_hierarchy
;
4650 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4651 struct cftype
*cft
, u64 val
)
4654 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4655 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4657 mutex_lock(&memcg_create_mutex
);
4659 if (memcg
->use_hierarchy
== val
)
4663 * If parent's use_hierarchy is set, we can't make any modifications
4664 * in the child subtrees. If it is unset, then the change can
4665 * occur, provided the current cgroup has no children.
4667 * For the root cgroup, parent_mem is NULL, we allow value to be
4668 * set if there are no children.
4670 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4671 (val
== 1 || val
== 0)) {
4672 if (!__memcg_has_children(memcg
))
4673 memcg
->use_hierarchy
= val
;
4680 mutex_unlock(&memcg_create_mutex
);
4686 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4687 enum mem_cgroup_stat_index idx
)
4689 struct mem_cgroup
*iter
;
4692 /* Per-cpu values can be negative, use a signed accumulator */
4693 for_each_mem_cgroup_tree(iter
, memcg
)
4694 val
+= mem_cgroup_read_stat(iter
, idx
);
4696 if (val
< 0) /* race ? */
4701 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4705 if (!mem_cgroup_is_root(memcg
)) {
4707 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4709 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4713 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4714 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4716 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4717 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4720 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4722 return val
<< PAGE_SHIFT
;
4725 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
4726 struct cftype
*cft
, struct file
*file
,
4727 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
4729 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4735 type
= MEMFILE_TYPE(cft
->private);
4736 name
= MEMFILE_ATTR(cft
->private);
4740 if (name
== RES_USAGE
)
4741 val
= mem_cgroup_usage(memcg
, false);
4743 val
= res_counter_read_u64(&memcg
->res
, name
);
4746 if (name
== RES_USAGE
)
4747 val
= mem_cgroup_usage(memcg
, true);
4749 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4752 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4758 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4759 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4762 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
4765 #ifdef CONFIG_MEMCG_KMEM
4766 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4768 * For simplicity, we won't allow this to be disabled. It also can't
4769 * be changed if the cgroup has children already, or if tasks had
4772 * If tasks join before we set the limit, a person looking at
4773 * kmem.usage_in_bytes will have no way to determine when it took
4774 * place, which makes the value quite meaningless.
4776 * After it first became limited, changes in the value of the limit are
4777 * of course permitted.
4779 mutex_lock(&memcg_create_mutex
);
4780 mutex_lock(&set_limit_mutex
);
4781 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4782 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
4786 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4789 ret
= memcg_update_cache_sizes(memcg
);
4791 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4794 static_key_slow_inc(&memcg_kmem_enabled_key
);
4796 * setting the active bit after the inc will guarantee no one
4797 * starts accounting before all call sites are patched
4799 memcg_kmem_set_active(memcg
);
4801 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4803 mutex_unlock(&set_limit_mutex
);
4804 mutex_unlock(&memcg_create_mutex
);
4809 #ifdef CONFIG_MEMCG_KMEM
4810 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4813 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4817 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
4819 * When that happen, we need to disable the static branch only on those
4820 * memcgs that enabled it. To achieve this, we would be forced to
4821 * complicate the code by keeping track of which memcgs were the ones
4822 * that actually enabled limits, and which ones got it from its
4825 * It is a lot simpler just to do static_key_slow_inc() on every child
4826 * that is accounted.
4828 if (!memcg_kmem_is_active(memcg
))
4832 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4833 * memcg is active already. If the later initialization fails then the
4834 * cgroup core triggers the cleanup so we do not have to do it here.
4836 static_key_slow_inc(&memcg_kmem_enabled_key
);
4838 mutex_lock(&set_limit_mutex
);
4839 memcg_stop_kmem_account();
4840 ret
= memcg_update_cache_sizes(memcg
);
4841 memcg_resume_kmem_account();
4842 mutex_unlock(&set_limit_mutex
);
4846 #endif /* CONFIG_MEMCG_KMEM */
4849 * The user of this function is...
4852 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
4855 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4858 unsigned long long val
;
4861 type
= MEMFILE_TYPE(cft
->private);
4862 name
= MEMFILE_ATTR(cft
->private);
4866 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4870 /* This function does all necessary parse...reuse it */
4871 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
4875 ret
= mem_cgroup_resize_limit(memcg
, val
);
4876 else if (type
== _MEMSWAP
)
4877 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
4878 else if (type
== _KMEM
)
4879 ret
= memcg_update_kmem_limit(css
, val
);
4883 case RES_SOFT_LIMIT
:
4884 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
4888 * For memsw, soft limits are hard to implement in terms
4889 * of semantics, for now, we support soft limits for
4890 * control without swap
4893 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
4898 ret
= -EINVAL
; /* should be BUG() ? */
4904 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
4905 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
4907 unsigned long long min_limit
, min_memsw_limit
, tmp
;
4909 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4910 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4911 if (!memcg
->use_hierarchy
)
4914 while (css_parent(&memcg
->css
)) {
4915 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4916 if (!memcg
->use_hierarchy
)
4918 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4919 min_limit
= min(min_limit
, tmp
);
4920 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4921 min_memsw_limit
= min(min_memsw_limit
, tmp
);
4924 *mem_limit
= min_limit
;
4925 *memsw_limit
= min_memsw_limit
;
4928 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
4930 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4934 type
= MEMFILE_TYPE(event
);
4935 name
= MEMFILE_ATTR(event
);
4940 res_counter_reset_max(&memcg
->res
);
4941 else if (type
== _MEMSWAP
)
4942 res_counter_reset_max(&memcg
->memsw
);
4943 else if (type
== _KMEM
)
4944 res_counter_reset_max(&memcg
->kmem
);
4950 res_counter_reset_failcnt(&memcg
->res
);
4951 else if (type
== _MEMSWAP
)
4952 res_counter_reset_failcnt(&memcg
->memsw
);
4953 else if (type
== _KMEM
)
4954 res_counter_reset_failcnt(&memcg
->kmem
);
4963 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
4966 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
4970 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4971 struct cftype
*cft
, u64 val
)
4973 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4975 if (val
>= (1 << NR_MOVE_TYPE
))
4979 * No kind of locking is needed in here, because ->can_attach() will
4980 * check this value once in the beginning of the process, and then carry
4981 * on with stale data. This means that changes to this value will only
4982 * affect task migrations starting after the change.
4984 memcg
->move_charge_at_immigrate
= val
;
4988 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
4989 struct cftype
*cft
, u64 val
)
4996 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
4997 struct cftype
*cft
, struct seq_file
*m
)
5000 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5001 unsigned long node_nr
;
5002 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5004 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5005 seq_printf(m
, "total=%lu", total_nr
);
5006 for_each_node_state(nid
, N_MEMORY
) {
5007 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5008 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5012 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5013 seq_printf(m
, "file=%lu", file_nr
);
5014 for_each_node_state(nid
, N_MEMORY
) {
5015 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5017 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5021 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5022 seq_printf(m
, "anon=%lu", anon_nr
);
5023 for_each_node_state(nid
, N_MEMORY
) {
5024 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5026 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5030 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5031 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5032 for_each_node_state(nid
, N_MEMORY
) {
5033 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5034 BIT(LRU_UNEVICTABLE
));
5035 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5040 #endif /* CONFIG_NUMA */
5042 static inline void mem_cgroup_lru_names_not_uptodate(void)
5044 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5047 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5050 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5051 struct mem_cgroup
*mi
;
5054 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5055 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5057 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5058 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5061 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5062 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5063 mem_cgroup_read_events(memcg
, i
));
5065 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5066 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5067 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5069 /* Hierarchical information */
5071 unsigned long long limit
, memsw_limit
;
5072 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5073 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5074 if (do_swap_account
)
5075 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5079 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5082 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5084 for_each_mem_cgroup_tree(mi
, memcg
)
5085 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5086 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5089 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5090 unsigned long long val
= 0;
5092 for_each_mem_cgroup_tree(mi
, memcg
)
5093 val
+= mem_cgroup_read_events(mi
, i
);
5094 seq_printf(m
, "total_%s %llu\n",
5095 mem_cgroup_events_names
[i
], val
);
5098 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5099 unsigned long long val
= 0;
5101 for_each_mem_cgroup_tree(mi
, memcg
)
5102 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5103 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5106 #ifdef CONFIG_DEBUG_VM
5109 struct mem_cgroup_per_zone
*mz
;
5110 struct zone_reclaim_stat
*rstat
;
5111 unsigned long recent_rotated
[2] = {0, 0};
5112 unsigned long recent_scanned
[2] = {0, 0};
5114 for_each_online_node(nid
)
5115 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5116 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5117 rstat
= &mz
->lruvec
.reclaim_stat
;
5119 recent_rotated
[0] += rstat
->recent_rotated
[0];
5120 recent_rotated
[1] += rstat
->recent_rotated
[1];
5121 recent_scanned
[0] += rstat
->recent_scanned
[0];
5122 recent_scanned
[1] += rstat
->recent_scanned
[1];
5124 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5125 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5126 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5127 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5134 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5137 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5139 return mem_cgroup_swappiness(memcg
);
5142 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5143 struct cftype
*cft
, u64 val
)
5145 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5146 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5148 if (val
> 100 || !parent
)
5151 mutex_lock(&memcg_create_mutex
);
5153 /* If under hierarchy, only empty-root can set this value */
5154 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5155 mutex_unlock(&memcg_create_mutex
);
5159 memcg
->swappiness
= val
;
5161 mutex_unlock(&memcg_create_mutex
);
5166 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5168 struct mem_cgroup_threshold_ary
*t
;
5174 t
= rcu_dereference(memcg
->thresholds
.primary
);
5176 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5181 usage
= mem_cgroup_usage(memcg
, swap
);
5184 * current_threshold points to threshold just below or equal to usage.
5185 * If it's not true, a threshold was crossed after last
5186 * call of __mem_cgroup_threshold().
5188 i
= t
->current_threshold
;
5191 * Iterate backward over array of thresholds starting from
5192 * current_threshold and check if a threshold is crossed.
5193 * If none of thresholds below usage is crossed, we read
5194 * only one element of the array here.
5196 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5197 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5199 /* i = current_threshold + 1 */
5203 * Iterate forward over array of thresholds starting from
5204 * current_threshold+1 and check if a threshold is crossed.
5205 * If none of thresholds above usage is crossed, we read
5206 * only one element of the array here.
5208 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5209 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5211 /* Update current_threshold */
5212 t
->current_threshold
= i
- 1;
5217 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5220 __mem_cgroup_threshold(memcg
, false);
5221 if (do_swap_account
)
5222 __mem_cgroup_threshold(memcg
, true);
5224 memcg
= parent_mem_cgroup(memcg
);
5228 static int compare_thresholds(const void *a
, const void *b
)
5230 const struct mem_cgroup_threshold
*_a
= a
;
5231 const struct mem_cgroup_threshold
*_b
= b
;
5233 if (_a
->threshold
> _b
->threshold
)
5236 if (_a
->threshold
< _b
->threshold
)
5242 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5244 struct mem_cgroup_eventfd_list
*ev
;
5246 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5247 eventfd_signal(ev
->eventfd
, 1);
5251 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5253 struct mem_cgroup
*iter
;
5255 for_each_mem_cgroup_tree(iter
, memcg
)
5256 mem_cgroup_oom_notify_cb(iter
);
5259 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5260 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5262 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5263 struct mem_cgroup_thresholds
*thresholds
;
5264 struct mem_cgroup_threshold_ary
*new;
5265 enum res_type type
= MEMFILE_TYPE(cft
->private);
5266 u64 threshold
, usage
;
5269 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5273 mutex_lock(&memcg
->thresholds_lock
);
5276 thresholds
= &memcg
->thresholds
;
5277 else if (type
== _MEMSWAP
)
5278 thresholds
= &memcg
->memsw_thresholds
;
5282 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5284 /* Check if a threshold crossed before adding a new one */
5285 if (thresholds
->primary
)
5286 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5288 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5290 /* Allocate memory for new array of thresholds */
5291 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5299 /* Copy thresholds (if any) to new array */
5300 if (thresholds
->primary
) {
5301 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5302 sizeof(struct mem_cgroup_threshold
));
5305 /* Add new threshold */
5306 new->entries
[size
- 1].eventfd
= eventfd
;
5307 new->entries
[size
- 1].threshold
= threshold
;
5309 /* Sort thresholds. Registering of new threshold isn't time-critical */
5310 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5311 compare_thresholds
, NULL
);
5313 /* Find current threshold */
5314 new->current_threshold
= -1;
5315 for (i
= 0; i
< size
; i
++) {
5316 if (new->entries
[i
].threshold
<= usage
) {
5318 * new->current_threshold will not be used until
5319 * rcu_assign_pointer(), so it's safe to increment
5322 ++new->current_threshold
;
5327 /* Free old spare buffer and save old primary buffer as spare */
5328 kfree(thresholds
->spare
);
5329 thresholds
->spare
= thresholds
->primary
;
5331 rcu_assign_pointer(thresholds
->primary
, new);
5333 /* To be sure that nobody uses thresholds */
5337 mutex_unlock(&memcg
->thresholds_lock
);
5342 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5343 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5345 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5346 struct mem_cgroup_thresholds
*thresholds
;
5347 struct mem_cgroup_threshold_ary
*new;
5348 enum res_type type
= MEMFILE_TYPE(cft
->private);
5352 mutex_lock(&memcg
->thresholds_lock
);
5354 thresholds
= &memcg
->thresholds
;
5355 else if (type
== _MEMSWAP
)
5356 thresholds
= &memcg
->memsw_thresholds
;
5360 if (!thresholds
->primary
)
5363 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5365 /* Check if a threshold crossed before removing */
5366 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5368 /* Calculate new number of threshold */
5370 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5371 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5375 new = thresholds
->spare
;
5377 /* Set thresholds array to NULL if we don't have thresholds */
5386 /* Copy thresholds and find current threshold */
5387 new->current_threshold
= -1;
5388 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5389 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5392 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5393 if (new->entries
[j
].threshold
<= usage
) {
5395 * new->current_threshold will not be used
5396 * until rcu_assign_pointer(), so it's safe to increment
5399 ++new->current_threshold
;
5405 /* Swap primary and spare array */
5406 thresholds
->spare
= thresholds
->primary
;
5407 /* If all events are unregistered, free the spare array */
5409 kfree(thresholds
->spare
);
5410 thresholds
->spare
= NULL
;
5413 rcu_assign_pointer(thresholds
->primary
, new);
5415 /* To be sure that nobody uses thresholds */
5418 mutex_unlock(&memcg
->thresholds_lock
);
5421 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5422 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5424 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5425 struct mem_cgroup_eventfd_list
*event
;
5426 enum res_type type
= MEMFILE_TYPE(cft
->private);
5428 BUG_ON(type
!= _OOM_TYPE
);
5429 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5433 spin_lock(&memcg_oom_lock
);
5435 event
->eventfd
= eventfd
;
5436 list_add(&event
->list
, &memcg
->oom_notify
);
5438 /* already in OOM ? */
5439 if (atomic_read(&memcg
->under_oom
))
5440 eventfd_signal(eventfd
, 1);
5441 spin_unlock(&memcg_oom_lock
);
5446 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5447 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5449 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5450 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5451 enum res_type type
= MEMFILE_TYPE(cft
->private);
5453 BUG_ON(type
!= _OOM_TYPE
);
5455 spin_lock(&memcg_oom_lock
);
5457 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5458 if (ev
->eventfd
== eventfd
) {
5459 list_del(&ev
->list
);
5464 spin_unlock(&memcg_oom_lock
);
5467 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5468 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5470 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5472 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5474 if (atomic_read(&memcg
->under_oom
))
5475 cb
->fill(cb
, "under_oom", 1);
5477 cb
->fill(cb
, "under_oom", 0);
5481 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5482 struct cftype
*cft
, u64 val
)
5484 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5485 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5487 /* cannot set to root cgroup and only 0 and 1 are allowed */
5488 if (!parent
|| !((val
== 0) || (val
== 1)))
5491 mutex_lock(&memcg_create_mutex
);
5492 /* oom-kill-disable is a flag for subhierarchy. */
5493 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5494 mutex_unlock(&memcg_create_mutex
);
5497 memcg
->oom_kill_disable
= val
;
5499 memcg_oom_recover(memcg
);
5500 mutex_unlock(&memcg_create_mutex
);
5504 #ifdef CONFIG_MEMCG_KMEM
5505 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5509 memcg
->kmemcg_id
= -1;
5510 ret
= memcg_propagate_kmem(memcg
);
5514 return mem_cgroup_sockets_init(memcg
, ss
);
5517 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5519 mem_cgroup_sockets_destroy(memcg
);
5522 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5524 if (!memcg_kmem_is_active(memcg
))
5528 * kmem charges can outlive the cgroup. In the case of slab
5529 * pages, for instance, a page contain objects from various
5530 * processes. As we prevent from taking a reference for every
5531 * such allocation we have to be careful when doing uncharge
5532 * (see memcg_uncharge_kmem) and here during offlining.
5534 * The idea is that that only the _last_ uncharge which sees
5535 * the dead memcg will drop the last reference. An additional
5536 * reference is taken here before the group is marked dead
5537 * which is then paired with css_put during uncharge resp. here.
5539 * Although this might sound strange as this path is called from
5540 * css_offline() when the referencemight have dropped down to 0
5541 * and shouldn't be incremented anymore (css_tryget would fail)
5542 * we do not have other options because of the kmem allocations
5545 css_get(&memcg
->css
);
5547 memcg_kmem_mark_dead(memcg
);
5549 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5552 if (memcg_kmem_test_and_clear_dead(memcg
))
5553 css_put(&memcg
->css
);
5556 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5561 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5565 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5570 static struct cftype mem_cgroup_files
[] = {
5572 .name
= "usage_in_bytes",
5573 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5574 .read
= mem_cgroup_read
,
5575 .register_event
= mem_cgroup_usage_register_event
,
5576 .unregister_event
= mem_cgroup_usage_unregister_event
,
5579 .name
= "max_usage_in_bytes",
5580 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5581 .trigger
= mem_cgroup_reset
,
5582 .read
= mem_cgroup_read
,
5585 .name
= "limit_in_bytes",
5586 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5587 .write_string
= mem_cgroup_write
,
5588 .read
= mem_cgroup_read
,
5591 .name
= "soft_limit_in_bytes",
5592 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5593 .write_string
= mem_cgroup_write
,
5594 .read
= mem_cgroup_read
,
5598 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5599 .trigger
= mem_cgroup_reset
,
5600 .read
= mem_cgroup_read
,
5604 .read_seq_string
= memcg_stat_show
,
5607 .name
= "force_empty",
5608 .trigger
= mem_cgroup_force_empty_write
,
5611 .name
= "use_hierarchy",
5612 .flags
= CFTYPE_INSANE
,
5613 .write_u64
= mem_cgroup_hierarchy_write
,
5614 .read_u64
= mem_cgroup_hierarchy_read
,
5617 .name
= "swappiness",
5618 .read_u64
= mem_cgroup_swappiness_read
,
5619 .write_u64
= mem_cgroup_swappiness_write
,
5622 .name
= "move_charge_at_immigrate",
5623 .read_u64
= mem_cgroup_move_charge_read
,
5624 .write_u64
= mem_cgroup_move_charge_write
,
5627 .name
= "oom_control",
5628 .read_map
= mem_cgroup_oom_control_read
,
5629 .write_u64
= mem_cgroup_oom_control_write
,
5630 .register_event
= mem_cgroup_oom_register_event
,
5631 .unregister_event
= mem_cgroup_oom_unregister_event
,
5632 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5635 .name
= "pressure_level",
5636 .register_event
= vmpressure_register_event
,
5637 .unregister_event
= vmpressure_unregister_event
,
5641 .name
= "numa_stat",
5642 .read_seq_string
= memcg_numa_stat_show
,
5645 #ifdef CONFIG_MEMCG_KMEM
5647 .name
= "kmem.limit_in_bytes",
5648 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5649 .write_string
= mem_cgroup_write
,
5650 .read
= mem_cgroup_read
,
5653 .name
= "kmem.usage_in_bytes",
5654 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5655 .read
= mem_cgroup_read
,
5658 .name
= "kmem.failcnt",
5659 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5660 .trigger
= mem_cgroup_reset
,
5661 .read
= mem_cgroup_read
,
5664 .name
= "kmem.max_usage_in_bytes",
5665 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5666 .trigger
= mem_cgroup_reset
,
5667 .read
= mem_cgroup_read
,
5669 #ifdef CONFIG_SLABINFO
5671 .name
= "kmem.slabinfo",
5672 .read_seq_string
= mem_cgroup_slabinfo_read
,
5676 { }, /* terminate */
5679 #ifdef CONFIG_MEMCG_SWAP
5680 static struct cftype memsw_cgroup_files
[] = {
5682 .name
= "memsw.usage_in_bytes",
5683 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5684 .read
= mem_cgroup_read
,
5685 .register_event
= mem_cgroup_usage_register_event
,
5686 .unregister_event
= mem_cgroup_usage_unregister_event
,
5689 .name
= "memsw.max_usage_in_bytes",
5690 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5691 .trigger
= mem_cgroup_reset
,
5692 .read
= mem_cgroup_read
,
5695 .name
= "memsw.limit_in_bytes",
5696 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5697 .write_string
= mem_cgroup_write
,
5698 .read
= mem_cgroup_read
,
5701 .name
= "memsw.failcnt",
5702 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5703 .trigger
= mem_cgroup_reset
,
5704 .read
= mem_cgroup_read
,
5706 { }, /* terminate */
5709 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5711 struct mem_cgroup_per_node
*pn
;
5712 struct mem_cgroup_per_zone
*mz
;
5713 int zone
, tmp
= node
;
5715 * This routine is called against possible nodes.
5716 * But it's BUG to call kmalloc() against offline node.
5718 * TODO: this routine can waste much memory for nodes which will
5719 * never be onlined. It's better to use memory hotplug callback
5722 if (!node_state(node
, N_NORMAL_MEMORY
))
5724 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5728 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5729 mz
= &pn
->zoneinfo
[zone
];
5730 lruvec_init(&mz
->lruvec
);
5733 memcg
->nodeinfo
[node
] = pn
;
5737 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5739 kfree(memcg
->nodeinfo
[node
]);
5742 static struct mem_cgroup
*mem_cgroup_alloc(void)
5744 struct mem_cgroup
*memcg
;
5745 size_t size
= memcg_size();
5747 /* Can be very big if nr_node_ids is very big */
5748 if (size
< PAGE_SIZE
)
5749 memcg
= kzalloc(size
, GFP_KERNEL
);
5751 memcg
= vzalloc(size
);
5756 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5759 spin_lock_init(&memcg
->pcp_counter_lock
);
5763 if (size
< PAGE_SIZE
)
5771 * At destroying mem_cgroup, references from swap_cgroup can remain.
5772 * (scanning all at force_empty is too costly...)
5774 * Instead of clearing all references at force_empty, we remember
5775 * the number of reference from swap_cgroup and free mem_cgroup when
5776 * it goes down to 0.
5778 * Removal of cgroup itself succeeds regardless of refs from swap.
5781 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5784 size_t size
= memcg_size();
5786 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5789 free_mem_cgroup_per_zone_info(memcg
, node
);
5791 free_percpu(memcg
->stat
);
5794 * We need to make sure that (at least for now), the jump label
5795 * destruction code runs outside of the cgroup lock. This is because
5796 * get_online_cpus(), which is called from the static_branch update,
5797 * can't be called inside the cgroup_lock. cpusets are the ones
5798 * enforcing this dependency, so if they ever change, we might as well.
5800 * schedule_work() will guarantee this happens. Be careful if you need
5801 * to move this code around, and make sure it is outside
5804 disarm_static_keys(memcg
);
5805 if (size
< PAGE_SIZE
)
5812 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5814 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
5816 if (!memcg
->res
.parent
)
5818 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
5820 EXPORT_SYMBOL(parent_mem_cgroup
);
5822 static struct cgroup_subsys_state
* __ref
5823 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5825 struct mem_cgroup
*memcg
;
5826 long error
= -ENOMEM
;
5829 memcg
= mem_cgroup_alloc();
5831 return ERR_PTR(error
);
5834 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
5838 if (parent_css
== NULL
) {
5839 root_mem_cgroup
= memcg
;
5840 res_counter_init(&memcg
->res
, NULL
);
5841 res_counter_init(&memcg
->memsw
, NULL
);
5842 res_counter_init(&memcg
->kmem
, NULL
);
5845 memcg
->last_scanned_node
= MAX_NUMNODES
;
5846 INIT_LIST_HEAD(&memcg
->oom_notify
);
5847 memcg
->move_charge_at_immigrate
= 0;
5848 mutex_init(&memcg
->thresholds_lock
);
5849 spin_lock_init(&memcg
->move_lock
);
5850 vmpressure_init(&memcg
->vmpressure
);
5855 __mem_cgroup_free(memcg
);
5856 return ERR_PTR(error
);
5860 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5862 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5863 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
5869 mutex_lock(&memcg_create_mutex
);
5871 memcg
->use_hierarchy
= parent
->use_hierarchy
;
5872 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5873 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5875 if (parent
->use_hierarchy
) {
5876 res_counter_init(&memcg
->res
, &parent
->res
);
5877 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
5878 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
5881 * No need to take a reference to the parent because cgroup
5882 * core guarantees its existence.
5885 res_counter_init(&memcg
->res
, NULL
);
5886 res_counter_init(&memcg
->memsw
, NULL
);
5887 res_counter_init(&memcg
->kmem
, NULL
);
5889 * Deeper hierachy with use_hierarchy == false doesn't make
5890 * much sense so let cgroup subsystem know about this
5891 * unfortunate state in our controller.
5893 if (parent
!= root_mem_cgroup
)
5894 mem_cgroup_subsys
.broken_hierarchy
= true;
5897 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
5898 mutex_unlock(&memcg_create_mutex
);
5903 * Announce all parents that a group from their hierarchy is gone.
5905 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
5907 struct mem_cgroup
*parent
= memcg
;
5909 while ((parent
= parent_mem_cgroup(parent
)))
5910 mem_cgroup_iter_invalidate(parent
);
5913 * if the root memcg is not hierarchical we have to check it
5916 if (!root_mem_cgroup
->use_hierarchy
)
5917 mem_cgroup_iter_invalidate(root_mem_cgroup
);
5920 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5922 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5924 kmem_cgroup_css_offline(memcg
);
5926 mem_cgroup_invalidate_reclaim_iterators(memcg
);
5927 mem_cgroup_reparent_charges(memcg
);
5928 mem_cgroup_destroy_all_caches(memcg
);
5929 vmpressure_cleanup(&memcg
->vmpressure
);
5932 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5934 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5936 memcg_destroy_kmem(memcg
);
5937 __mem_cgroup_free(memcg
);
5941 /* Handlers for move charge at task migration. */
5942 #define PRECHARGE_COUNT_AT_ONCE 256
5943 static int mem_cgroup_do_precharge(unsigned long count
)
5946 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
5947 struct mem_cgroup
*memcg
= mc
.to
;
5949 if (mem_cgroup_is_root(memcg
)) {
5950 mc
.precharge
+= count
;
5951 /* we don't need css_get for root */
5954 /* try to charge at once */
5956 struct res_counter
*dummy
;
5958 * "memcg" cannot be under rmdir() because we've already checked
5959 * by cgroup_lock_live_cgroup() that it is not removed and we
5960 * are still under the same cgroup_mutex. So we can postpone
5963 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
5965 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
5966 PAGE_SIZE
* count
, &dummy
)) {
5967 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
5970 mc
.precharge
+= count
;
5974 /* fall back to one by one charge */
5976 if (signal_pending(current
)) {
5980 if (!batch_count
--) {
5981 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
5984 ret
= __mem_cgroup_try_charge(NULL
,
5985 GFP_KERNEL
, 1, &memcg
, false);
5987 /* mem_cgroup_clear_mc() will do uncharge later */
5995 * get_mctgt_type - get target type of moving charge
5996 * @vma: the vma the pte to be checked belongs
5997 * @addr: the address corresponding to the pte to be checked
5998 * @ptent: the pte to be checked
5999 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6002 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6003 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6004 * move charge. if @target is not NULL, the page is stored in target->page
6005 * with extra refcnt got(Callers should handle it).
6006 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6007 * target for charge migration. if @target is not NULL, the entry is stored
6010 * Called with pte lock held.
6017 enum mc_target_type
{
6023 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6024 unsigned long addr
, pte_t ptent
)
6026 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6028 if (!page
|| !page_mapped(page
))
6030 if (PageAnon(page
)) {
6031 /* we don't move shared anon */
6034 } else if (!move_file())
6035 /* we ignore mapcount for file pages */
6037 if (!get_page_unless_zero(page
))
6044 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6045 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6047 struct page
*page
= NULL
;
6048 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6050 if (!move_anon() || non_swap_entry(ent
))
6053 * Because lookup_swap_cache() updates some statistics counter,
6054 * we call find_get_page() with swapper_space directly.
6056 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6057 if (do_swap_account
)
6058 entry
->val
= ent
.val
;
6063 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6064 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6070 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6071 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6073 struct page
*page
= NULL
;
6074 struct address_space
*mapping
;
6077 if (!vma
->vm_file
) /* anonymous vma */
6082 mapping
= vma
->vm_file
->f_mapping
;
6083 if (pte_none(ptent
))
6084 pgoff
= linear_page_index(vma
, addr
);
6085 else /* pte_file(ptent) is true */
6086 pgoff
= pte_to_pgoff(ptent
);
6088 /* page is moved even if it's not RSS of this task(page-faulted). */
6089 page
= find_get_page(mapping
, pgoff
);
6092 /* shmem/tmpfs may report page out on swap: account for that too. */
6093 if (radix_tree_exceptional_entry(page
)) {
6094 swp_entry_t swap
= radix_to_swp_entry(page
);
6095 if (do_swap_account
)
6097 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6103 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6104 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6106 struct page
*page
= NULL
;
6107 struct page_cgroup
*pc
;
6108 enum mc_target_type ret
= MC_TARGET_NONE
;
6109 swp_entry_t ent
= { .val
= 0 };
6111 if (pte_present(ptent
))
6112 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6113 else if (is_swap_pte(ptent
))
6114 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6115 else if (pte_none(ptent
) || pte_file(ptent
))
6116 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6118 if (!page
&& !ent
.val
)
6121 pc
= lookup_page_cgroup(page
);
6123 * Do only loose check w/o page_cgroup lock.
6124 * mem_cgroup_move_account() checks the pc is valid or not under
6127 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6128 ret
= MC_TARGET_PAGE
;
6130 target
->page
= page
;
6132 if (!ret
|| !target
)
6135 /* There is a swap entry and a page doesn't exist or isn't charged */
6136 if (ent
.val
&& !ret
&&
6137 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6138 ret
= MC_TARGET_SWAP
;
6145 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6147 * We don't consider swapping or file mapped pages because THP does not
6148 * support them for now.
6149 * Caller should make sure that pmd_trans_huge(pmd) is true.
6151 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6152 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6154 struct page
*page
= NULL
;
6155 struct page_cgroup
*pc
;
6156 enum mc_target_type ret
= MC_TARGET_NONE
;
6158 page
= pmd_page(pmd
);
6159 VM_BUG_ON(!page
|| !PageHead(page
));
6162 pc
= lookup_page_cgroup(page
);
6163 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6164 ret
= MC_TARGET_PAGE
;
6167 target
->page
= page
;
6173 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6174 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6176 return MC_TARGET_NONE
;
6180 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6181 unsigned long addr
, unsigned long end
,
6182 struct mm_walk
*walk
)
6184 struct vm_area_struct
*vma
= walk
->private;
6188 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6189 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6190 mc
.precharge
+= HPAGE_PMD_NR
;
6191 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6195 if (pmd_trans_unstable(pmd
))
6197 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6198 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6199 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6200 mc
.precharge
++; /* increment precharge temporarily */
6201 pte_unmap_unlock(pte
- 1, ptl
);
6207 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6209 unsigned long precharge
;
6210 struct vm_area_struct
*vma
;
6212 down_read(&mm
->mmap_sem
);
6213 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6214 struct mm_walk mem_cgroup_count_precharge_walk
= {
6215 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6219 if (is_vm_hugetlb_page(vma
))
6221 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6222 &mem_cgroup_count_precharge_walk
);
6224 up_read(&mm
->mmap_sem
);
6226 precharge
= mc
.precharge
;
6232 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6234 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6236 VM_BUG_ON(mc
.moving_task
);
6237 mc
.moving_task
= current
;
6238 return mem_cgroup_do_precharge(precharge
);
6241 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6242 static void __mem_cgroup_clear_mc(void)
6244 struct mem_cgroup
*from
= mc
.from
;
6245 struct mem_cgroup
*to
= mc
.to
;
6248 /* we must uncharge all the leftover precharges from mc.to */
6250 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6254 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6255 * we must uncharge here.
6257 if (mc
.moved_charge
) {
6258 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6259 mc
.moved_charge
= 0;
6261 /* we must fixup refcnts and charges */
6262 if (mc
.moved_swap
) {
6263 /* uncharge swap account from the old cgroup */
6264 if (!mem_cgroup_is_root(mc
.from
))
6265 res_counter_uncharge(&mc
.from
->memsw
,
6266 PAGE_SIZE
* mc
.moved_swap
);
6268 for (i
= 0; i
< mc
.moved_swap
; i
++)
6269 css_put(&mc
.from
->css
);
6271 if (!mem_cgroup_is_root(mc
.to
)) {
6273 * we charged both to->res and to->memsw, so we should
6276 res_counter_uncharge(&mc
.to
->res
,
6277 PAGE_SIZE
* mc
.moved_swap
);
6279 /* we've already done css_get(mc.to) */
6282 memcg_oom_recover(from
);
6283 memcg_oom_recover(to
);
6284 wake_up_all(&mc
.waitq
);
6287 static void mem_cgroup_clear_mc(void)
6289 struct mem_cgroup
*from
= mc
.from
;
6292 * we must clear moving_task before waking up waiters at the end of
6295 mc
.moving_task
= NULL
;
6296 __mem_cgroup_clear_mc();
6297 spin_lock(&mc
.lock
);
6300 spin_unlock(&mc
.lock
);
6301 mem_cgroup_end_move(from
);
6304 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6305 struct cgroup_taskset
*tset
)
6307 struct task_struct
*p
= cgroup_taskset_first(tset
);
6309 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6310 unsigned long move_charge_at_immigrate
;
6313 * We are now commited to this value whatever it is. Changes in this
6314 * tunable will only affect upcoming migrations, not the current one.
6315 * So we need to save it, and keep it going.
6317 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6318 if (move_charge_at_immigrate
) {
6319 struct mm_struct
*mm
;
6320 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6322 VM_BUG_ON(from
== memcg
);
6324 mm
= get_task_mm(p
);
6327 /* We move charges only when we move a owner of the mm */
6328 if (mm
->owner
== p
) {
6331 VM_BUG_ON(mc
.precharge
);
6332 VM_BUG_ON(mc
.moved_charge
);
6333 VM_BUG_ON(mc
.moved_swap
);
6334 mem_cgroup_start_move(from
);
6335 spin_lock(&mc
.lock
);
6338 mc
.immigrate_flags
= move_charge_at_immigrate
;
6339 spin_unlock(&mc
.lock
);
6340 /* We set mc.moving_task later */
6342 ret
= mem_cgroup_precharge_mc(mm
);
6344 mem_cgroup_clear_mc();
6351 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6352 struct cgroup_taskset
*tset
)
6354 mem_cgroup_clear_mc();
6357 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6358 unsigned long addr
, unsigned long end
,
6359 struct mm_walk
*walk
)
6362 struct vm_area_struct
*vma
= walk
->private;
6365 enum mc_target_type target_type
;
6366 union mc_target target
;
6368 struct page_cgroup
*pc
;
6371 * We don't take compound_lock() here but no race with splitting thp
6373 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6374 * under splitting, which means there's no concurrent thp split,
6375 * - if another thread runs into split_huge_page() just after we
6376 * entered this if-block, the thread must wait for page table lock
6377 * to be unlocked in __split_huge_page_splitting(), where the main
6378 * part of thp split is not executed yet.
6380 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6381 if (mc
.precharge
< HPAGE_PMD_NR
) {
6382 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6385 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6386 if (target_type
== MC_TARGET_PAGE
) {
6388 if (!isolate_lru_page(page
)) {
6389 pc
= lookup_page_cgroup(page
);
6390 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6391 pc
, mc
.from
, mc
.to
)) {
6392 mc
.precharge
-= HPAGE_PMD_NR
;
6393 mc
.moved_charge
+= HPAGE_PMD_NR
;
6395 putback_lru_page(page
);
6399 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6403 if (pmd_trans_unstable(pmd
))
6406 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6407 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6408 pte_t ptent
= *(pte
++);
6414 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6415 case MC_TARGET_PAGE
:
6417 if (isolate_lru_page(page
))
6419 pc
= lookup_page_cgroup(page
);
6420 if (!mem_cgroup_move_account(page
, 1, pc
,
6423 /* we uncharge from mc.from later. */
6426 putback_lru_page(page
);
6427 put
: /* get_mctgt_type() gets the page */
6430 case MC_TARGET_SWAP
:
6432 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6434 /* we fixup refcnts and charges later. */
6442 pte_unmap_unlock(pte
- 1, ptl
);
6447 * We have consumed all precharges we got in can_attach().
6448 * We try charge one by one, but don't do any additional
6449 * charges to mc.to if we have failed in charge once in attach()
6452 ret
= mem_cgroup_do_precharge(1);
6460 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6462 struct vm_area_struct
*vma
;
6464 lru_add_drain_all();
6466 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6468 * Someone who are holding the mmap_sem might be waiting in
6469 * waitq. So we cancel all extra charges, wake up all waiters,
6470 * and retry. Because we cancel precharges, we might not be able
6471 * to move enough charges, but moving charge is a best-effort
6472 * feature anyway, so it wouldn't be a big problem.
6474 __mem_cgroup_clear_mc();
6478 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6480 struct mm_walk mem_cgroup_move_charge_walk
= {
6481 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6485 if (is_vm_hugetlb_page(vma
))
6487 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6488 &mem_cgroup_move_charge_walk
);
6491 * means we have consumed all precharges and failed in
6492 * doing additional charge. Just abandon here.
6496 up_read(&mm
->mmap_sem
);
6499 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6500 struct cgroup_taskset
*tset
)
6502 struct task_struct
*p
= cgroup_taskset_first(tset
);
6503 struct mm_struct
*mm
= get_task_mm(p
);
6507 mem_cgroup_move_charge(mm
);
6511 mem_cgroup_clear_mc();
6513 #else /* !CONFIG_MMU */
6514 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6515 struct cgroup_taskset
*tset
)
6519 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6520 struct cgroup_taskset
*tset
)
6523 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6524 struct cgroup_taskset
*tset
)
6530 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6531 * to verify sane_behavior flag on each mount attempt.
6533 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6536 * use_hierarchy is forced with sane_behavior. cgroup core
6537 * guarantees that @root doesn't have any children, so turning it
6538 * on for the root memcg is enough.
6540 if (cgroup_sane_behavior(root_css
->cgroup
))
6541 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6544 struct cgroup_subsys mem_cgroup_subsys
= {
6546 .subsys_id
= mem_cgroup_subsys_id
,
6547 .css_alloc
= mem_cgroup_css_alloc
,
6548 .css_online
= mem_cgroup_css_online
,
6549 .css_offline
= mem_cgroup_css_offline
,
6550 .css_free
= mem_cgroup_css_free
,
6551 .can_attach
= mem_cgroup_can_attach
,
6552 .cancel_attach
= mem_cgroup_cancel_attach
,
6553 .attach
= mem_cgroup_move_task
,
6554 .bind
= mem_cgroup_bind
,
6555 .base_cftypes
= mem_cgroup_files
,
6560 #ifdef CONFIG_MEMCG_SWAP
6561 static int __init
enable_swap_account(char *s
)
6563 if (!strcmp(s
, "1"))
6564 really_do_swap_account
= 1;
6565 else if (!strcmp(s
, "0"))
6566 really_do_swap_account
= 0;
6569 __setup("swapaccount=", enable_swap_account
);
6571 static void __init
memsw_file_init(void)
6573 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6576 static void __init
enable_swap_cgroup(void)
6578 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6579 do_swap_account
= 1;
6585 static void __init
enable_swap_cgroup(void)
6591 * subsys_initcall() for memory controller.
6593 * Some parts like hotcpu_notifier() have to be initialized from this context
6594 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6595 * everything that doesn't depend on a specific mem_cgroup structure should
6596 * be initialized from here.
6598 static int __init
mem_cgroup_init(void)
6600 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6601 enable_swap_cgroup();
6605 subsys_initcall(mem_cgroup_init
);