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
));
878 static enum mem_cgroup_filter_t
879 mem_cgroup_filter(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
,
880 mem_cgroup_iter_filter cond
)
884 return cond(memcg
, root
);
888 * Returns a next (in a pre-order walk) alive memcg (with elevated css
889 * ref. count) or NULL if the whole root's subtree has been visited.
891 * helper function to be used by mem_cgroup_iter
893 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
894 struct mem_cgroup
*last_visited
, mem_cgroup_iter_filter cond
)
896 struct cgroup_subsys_state
*prev_css
, *next_css
;
898 prev_css
= last_visited
? &last_visited
->css
: NULL
;
900 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
903 * Even if we found a group we have to make sure it is
904 * alive. css && !memcg means that the groups should be
905 * skipped and we should continue the tree walk.
906 * last_visited css is safe to use because it is
907 * protected by css_get and the tree walk is rcu safe.
910 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
912 switch (mem_cgroup_filter(mem
, root
, cond
)) {
920 * css_rightmost_descendant is not an optimal way to
921 * skip through a subtree (especially for imbalanced
922 * trees leaning to right) but that's what we have right
923 * now. More effective solution would be traversing
924 * right-up for first non-NULL without calling
925 * css_next_descendant_pre afterwards.
927 prev_css
= css_rightmost_descendant(next_css
);
930 if (css_tryget(&mem
->css
))
943 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
946 * When a group in the hierarchy below root is destroyed, the
947 * hierarchy iterator can no longer be trusted since it might
948 * have pointed to the destroyed group. Invalidate it.
950 atomic_inc(&root
->dead_count
);
953 static struct mem_cgroup
*
954 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
955 struct mem_cgroup
*root
,
958 struct mem_cgroup
*position
= NULL
;
960 * A cgroup destruction happens in two stages: offlining and
961 * release. They are separated by a RCU grace period.
963 * If the iterator is valid, we may still race with an
964 * offlining. The RCU lock ensures the object won't be
965 * released, tryget will fail if we lost the race.
967 *sequence
= atomic_read(&root
->dead_count
);
968 if (iter
->last_dead_count
== *sequence
) {
970 position
= iter
->last_visited
;
971 if (position
&& !css_tryget(&position
->css
))
977 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
978 struct mem_cgroup
*last_visited
,
979 struct mem_cgroup
*new_position
,
983 css_put(&last_visited
->css
);
985 * We store the sequence count from the time @last_visited was
986 * loaded successfully instead of rereading it here so that we
987 * don't lose destruction events in between. We could have
988 * raced with the destruction of @new_position after all.
990 iter
->last_visited
= new_position
;
992 iter
->last_dead_count
= sequence
;
996 * mem_cgroup_iter - iterate over memory cgroup hierarchy
997 * @root: hierarchy root
998 * @prev: previously returned memcg, NULL on first invocation
999 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1000 * @cond: filter for visited nodes, NULL for no filter
1002 * Returns references to children of the hierarchy below @root, or
1003 * @root itself, or %NULL after a full round-trip.
1005 * Caller must pass the return value in @prev on subsequent
1006 * invocations for reference counting, or use mem_cgroup_iter_break()
1007 * to cancel a hierarchy walk before the round-trip is complete.
1009 * Reclaimers can specify a zone and a priority level in @reclaim to
1010 * divide up the memcgs in the hierarchy among all concurrent
1011 * reclaimers operating on the same zone and priority.
1013 struct mem_cgroup
*mem_cgroup_iter_cond(struct mem_cgroup
*root
,
1014 struct mem_cgroup
*prev
,
1015 struct mem_cgroup_reclaim_cookie
*reclaim
,
1016 mem_cgroup_iter_filter cond
)
1018 struct mem_cgroup
*memcg
= NULL
;
1019 struct mem_cgroup
*last_visited
= NULL
;
1021 if (mem_cgroup_disabled()) {
1022 /* first call must return non-NULL, second return NULL */
1023 return (struct mem_cgroup
*)(unsigned long)!prev
;
1027 root
= root_mem_cgroup
;
1029 if (prev
&& !reclaim
)
1030 last_visited
= prev
;
1032 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1035 if (mem_cgroup_filter(root
, root
, cond
) == VISIT
)
1042 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1043 int uninitialized_var(seq
);
1046 int nid
= zone_to_nid(reclaim
->zone
);
1047 int zid
= zone_idx(reclaim
->zone
);
1048 struct mem_cgroup_per_zone
*mz
;
1050 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1051 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1052 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1053 iter
->last_visited
= NULL
;
1057 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1060 memcg
= __mem_cgroup_iter_next(root
, last_visited
, cond
);
1063 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1067 else if (!prev
&& memcg
)
1068 reclaim
->generation
= iter
->generation
;
1072 * We have finished the whole tree walk or no group has been
1073 * visited because filter told us to skip the root node.
1075 if (!memcg
&& (prev
|| (cond
&& !last_visited
)))
1081 if (prev
&& prev
!= root
)
1082 css_put(&prev
->css
);
1088 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1089 * @root: hierarchy root
1090 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1092 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1093 struct mem_cgroup
*prev
)
1096 root
= root_mem_cgroup
;
1097 if (prev
&& prev
!= root
)
1098 css_put(&prev
->css
);
1102 * Iteration constructs for visiting all cgroups (under a tree). If
1103 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1104 * be used for reference counting.
1106 #define for_each_mem_cgroup_tree(iter, root) \
1107 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1109 iter = mem_cgroup_iter(root, iter, NULL))
1111 #define for_each_mem_cgroup(iter) \
1112 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1114 iter = mem_cgroup_iter(NULL, iter, NULL))
1116 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1118 struct mem_cgroup
*memcg
;
1121 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1122 if (unlikely(!memcg
))
1127 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1130 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1138 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1141 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1142 * @zone: zone of the wanted lruvec
1143 * @memcg: memcg of the wanted lruvec
1145 * Returns the lru list vector holding pages for the given @zone and
1146 * @mem. This can be the global zone lruvec, if the memory controller
1149 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1150 struct mem_cgroup
*memcg
)
1152 struct mem_cgroup_per_zone
*mz
;
1153 struct lruvec
*lruvec
;
1155 if (mem_cgroup_disabled()) {
1156 lruvec
= &zone
->lruvec
;
1160 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1161 lruvec
= &mz
->lruvec
;
1164 * Since a node can be onlined after the mem_cgroup was created,
1165 * we have to be prepared to initialize lruvec->zone here;
1166 * and if offlined then reonlined, we need to reinitialize it.
1168 if (unlikely(lruvec
->zone
!= zone
))
1169 lruvec
->zone
= zone
;
1174 * Following LRU functions are allowed to be used without PCG_LOCK.
1175 * Operations are called by routine of global LRU independently from memcg.
1176 * What we have to take care of here is validness of pc->mem_cgroup.
1178 * Changes to pc->mem_cgroup happens when
1181 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1182 * It is added to LRU before charge.
1183 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1184 * When moving account, the page is not on LRU. It's isolated.
1188 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1190 * @zone: zone of the page
1192 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1194 struct mem_cgroup_per_zone
*mz
;
1195 struct mem_cgroup
*memcg
;
1196 struct page_cgroup
*pc
;
1197 struct lruvec
*lruvec
;
1199 if (mem_cgroup_disabled()) {
1200 lruvec
= &zone
->lruvec
;
1204 pc
= lookup_page_cgroup(page
);
1205 memcg
= pc
->mem_cgroup
;
1208 * Surreptitiously switch any uncharged offlist page to root:
1209 * an uncharged page off lru does nothing to secure
1210 * its former mem_cgroup from sudden removal.
1212 * Our caller holds lru_lock, and PageCgroupUsed is updated
1213 * under page_cgroup lock: between them, they make all uses
1214 * of pc->mem_cgroup safe.
1216 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1217 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1219 mz
= page_cgroup_zoneinfo(memcg
, page
);
1220 lruvec
= &mz
->lruvec
;
1223 * Since a node can be onlined after the mem_cgroup was created,
1224 * we have to be prepared to initialize lruvec->zone here;
1225 * and if offlined then reonlined, we need to reinitialize it.
1227 if (unlikely(lruvec
->zone
!= zone
))
1228 lruvec
->zone
= zone
;
1233 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1234 * @lruvec: mem_cgroup per zone lru vector
1235 * @lru: index of lru list the page is sitting on
1236 * @nr_pages: positive when adding or negative when removing
1238 * This function must be called when a page is added to or removed from an
1241 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1244 struct mem_cgroup_per_zone
*mz
;
1245 unsigned long *lru_size
;
1247 if (mem_cgroup_disabled())
1250 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1251 lru_size
= mz
->lru_size
+ lru
;
1252 *lru_size
+= nr_pages
;
1253 VM_BUG_ON((long)(*lru_size
) < 0);
1257 * Checks whether given mem is same or in the root_mem_cgroup's
1260 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1261 struct mem_cgroup
*memcg
)
1263 if (root_memcg
== memcg
)
1265 if (!root_memcg
->use_hierarchy
|| !memcg
)
1267 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1270 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1271 struct mem_cgroup
*memcg
)
1276 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1281 bool task_in_mem_cgroup(struct task_struct
*task
,
1282 const struct mem_cgroup
*memcg
)
1284 struct mem_cgroup
*curr
= NULL
;
1285 struct task_struct
*p
;
1288 p
= find_lock_task_mm(task
);
1290 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1294 * All threads may have already detached their mm's, but the oom
1295 * killer still needs to detect if they have already been oom
1296 * killed to prevent needlessly killing additional tasks.
1299 curr
= mem_cgroup_from_task(task
);
1301 css_get(&curr
->css
);
1307 * We should check use_hierarchy of "memcg" not "curr". Because checking
1308 * use_hierarchy of "curr" here make this function true if hierarchy is
1309 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1310 * hierarchy(even if use_hierarchy is disabled in "memcg").
1312 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1313 css_put(&curr
->css
);
1317 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1319 unsigned long inactive_ratio
;
1320 unsigned long inactive
;
1321 unsigned long active
;
1324 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1325 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1327 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1329 inactive_ratio
= int_sqrt(10 * gb
);
1333 return inactive
* inactive_ratio
< active
;
1336 #define mem_cgroup_from_res_counter(counter, member) \
1337 container_of(counter, struct mem_cgroup, member)
1340 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1341 * @memcg: the memory cgroup
1343 * Returns the maximum amount of memory @mem can be charged with, in
1346 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1348 unsigned long long margin
;
1350 margin
= res_counter_margin(&memcg
->res
);
1351 if (do_swap_account
)
1352 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1353 return margin
>> PAGE_SHIFT
;
1356 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1359 if (!css_parent(&memcg
->css
))
1360 return vm_swappiness
;
1362 return memcg
->swappiness
;
1366 * memcg->moving_account is used for checking possibility that some thread is
1367 * calling move_account(). When a thread on CPU-A starts moving pages under
1368 * a memcg, other threads should check memcg->moving_account under
1369 * rcu_read_lock(), like this:
1373 * memcg->moving_account+1 if (memcg->mocing_account)
1375 * synchronize_rcu() update something.
1380 /* for quick checking without looking up memcg */
1381 atomic_t memcg_moving __read_mostly
;
1383 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1385 atomic_inc(&memcg_moving
);
1386 atomic_inc(&memcg
->moving_account
);
1390 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1393 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1394 * We check NULL in callee rather than caller.
1397 atomic_dec(&memcg_moving
);
1398 atomic_dec(&memcg
->moving_account
);
1403 * 2 routines for checking "mem" is under move_account() or not.
1405 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1406 * is used for avoiding races in accounting. If true,
1407 * pc->mem_cgroup may be overwritten.
1409 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1410 * under hierarchy of moving cgroups. This is for
1411 * waiting at hith-memory prressure caused by "move".
1414 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1416 VM_BUG_ON(!rcu_read_lock_held());
1417 return atomic_read(&memcg
->moving_account
) > 0;
1420 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1422 struct mem_cgroup
*from
;
1423 struct mem_cgroup
*to
;
1426 * Unlike task_move routines, we access mc.to, mc.from not under
1427 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1429 spin_lock(&mc
.lock
);
1435 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1436 || mem_cgroup_same_or_subtree(memcg
, to
);
1438 spin_unlock(&mc
.lock
);
1442 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1444 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1445 if (mem_cgroup_under_move(memcg
)) {
1447 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1448 /* moving charge context might have finished. */
1451 finish_wait(&mc
.waitq
, &wait
);
1459 * Take this lock when
1460 * - a code tries to modify page's memcg while it's USED.
1461 * - a code tries to modify page state accounting in a memcg.
1462 * see mem_cgroup_stolen(), too.
1464 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1465 unsigned long *flags
)
1467 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1470 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1471 unsigned long *flags
)
1473 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1476 #define K(x) ((x) << (PAGE_SHIFT-10))
1478 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1479 * @memcg: The memory cgroup that went over limit
1480 * @p: Task that is going to be killed
1482 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1485 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1487 struct cgroup
*task_cgrp
;
1488 struct cgroup
*mem_cgrp
;
1490 * Need a buffer in BSS, can't rely on allocations. The code relies
1491 * on the assumption that OOM is serialized for memory controller.
1492 * If this assumption is broken, revisit this code.
1494 static char memcg_name
[PATH_MAX
];
1496 struct mem_cgroup
*iter
;
1504 mem_cgrp
= memcg
->css
.cgroup
;
1505 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1507 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1510 * Unfortunately, we are unable to convert to a useful name
1511 * But we'll still print out the usage information
1518 pr_info("Task in %s killed", memcg_name
);
1521 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1529 * Continues from above, so we don't need an KERN_ level
1531 pr_cont(" as a result of limit of %s\n", memcg_name
);
1534 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1535 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1536 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1537 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1538 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1539 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1540 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1541 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1542 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1543 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1544 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1545 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1547 for_each_mem_cgroup_tree(iter
, memcg
) {
1548 pr_info("Memory cgroup stats");
1551 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1553 pr_cont(" for %s", memcg_name
);
1557 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1558 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1560 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1561 K(mem_cgroup_read_stat(iter
, i
)));
1564 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1565 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1566 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1573 * This function returns the number of memcg under hierarchy tree. Returns
1574 * 1(self count) if no children.
1576 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1579 struct mem_cgroup
*iter
;
1581 for_each_mem_cgroup_tree(iter
, memcg
)
1587 * Return the memory (and swap, if configured) limit for a memcg.
1589 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1593 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1596 * Do not consider swap space if we cannot swap due to swappiness
1598 if (mem_cgroup_swappiness(memcg
)) {
1601 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1602 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1605 * If memsw is finite and limits the amount of swap space
1606 * available to this memcg, return that limit.
1608 limit
= min(limit
, memsw
);
1614 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1617 struct mem_cgroup
*iter
;
1618 unsigned long chosen_points
= 0;
1619 unsigned long totalpages
;
1620 unsigned int points
= 0;
1621 struct task_struct
*chosen
= NULL
;
1624 * If current has a pending SIGKILL or is exiting, then automatically
1625 * select it. The goal is to allow it to allocate so that it may
1626 * quickly exit and free its memory.
1628 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1629 set_thread_flag(TIF_MEMDIE
);
1633 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1634 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1635 for_each_mem_cgroup_tree(iter
, memcg
) {
1636 struct css_task_iter it
;
1637 struct task_struct
*task
;
1639 css_task_iter_start(&iter
->css
, &it
);
1640 while ((task
= css_task_iter_next(&it
))) {
1641 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1643 case OOM_SCAN_SELECT
:
1645 put_task_struct(chosen
);
1647 chosen_points
= ULONG_MAX
;
1648 get_task_struct(chosen
);
1650 case OOM_SCAN_CONTINUE
:
1652 case OOM_SCAN_ABORT
:
1653 css_task_iter_end(&it
);
1654 mem_cgroup_iter_break(memcg
, iter
);
1656 put_task_struct(chosen
);
1661 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1662 if (points
> chosen_points
) {
1664 put_task_struct(chosen
);
1666 chosen_points
= points
;
1667 get_task_struct(chosen
);
1670 css_task_iter_end(&it
);
1675 points
= chosen_points
* 1000 / totalpages
;
1676 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1677 NULL
, "Memory cgroup out of memory");
1680 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1682 unsigned long flags
)
1684 unsigned long total
= 0;
1685 bool noswap
= false;
1688 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1690 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1693 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1695 drain_all_stock_async(memcg
);
1696 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1698 * Allow limit shrinkers, which are triggered directly
1699 * by userspace, to catch signals and stop reclaim
1700 * after minimal progress, regardless of the margin.
1702 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1704 if (mem_cgroup_margin(memcg
))
1707 * If nothing was reclaimed after two attempts, there
1708 * may be no reclaimable pages in this hierarchy.
1716 #if MAX_NUMNODES > 1
1718 * test_mem_cgroup_node_reclaimable
1719 * @memcg: the target memcg
1720 * @nid: the node ID to be checked.
1721 * @noswap : specify true here if the user wants flle only information.
1723 * This function returns whether the specified memcg contains any
1724 * reclaimable pages on a node. Returns true if there are any reclaimable
1725 * pages in the node.
1727 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1728 int nid
, bool noswap
)
1730 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1732 if (noswap
|| !total_swap_pages
)
1734 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1741 * Always updating the nodemask is not very good - even if we have an empty
1742 * list or the wrong list here, we can start from some node and traverse all
1743 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1746 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1750 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1751 * pagein/pageout changes since the last update.
1753 if (!atomic_read(&memcg
->numainfo_events
))
1755 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1758 /* make a nodemask where this memcg uses memory from */
1759 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1761 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1763 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1764 node_clear(nid
, memcg
->scan_nodes
);
1767 atomic_set(&memcg
->numainfo_events
, 0);
1768 atomic_set(&memcg
->numainfo_updating
, 0);
1772 * Selecting a node where we start reclaim from. Because what we need is just
1773 * reducing usage counter, start from anywhere is O,K. Considering
1774 * memory reclaim from current node, there are pros. and cons.
1776 * Freeing memory from current node means freeing memory from a node which
1777 * we'll use or we've used. So, it may make LRU bad. And if several threads
1778 * hit limits, it will see a contention on a node. But freeing from remote
1779 * node means more costs for memory reclaim because of memory latency.
1781 * Now, we use round-robin. Better algorithm is welcomed.
1783 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1787 mem_cgroup_may_update_nodemask(memcg
);
1788 node
= memcg
->last_scanned_node
;
1790 node
= next_node(node
, memcg
->scan_nodes
);
1791 if (node
== MAX_NUMNODES
)
1792 node
= first_node(memcg
->scan_nodes
);
1794 * We call this when we hit limit, not when pages are added to LRU.
1795 * No LRU may hold pages because all pages are UNEVICTABLE or
1796 * memcg is too small and all pages are not on LRU. In that case,
1797 * we use curret node.
1799 if (unlikely(node
== MAX_NUMNODES
))
1800 node
= numa_node_id();
1802 memcg
->last_scanned_node
= node
;
1807 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1815 * A group is eligible for the soft limit reclaim under the given root
1817 * a) it is over its soft limit
1818 * b) any parent up the hierarchy is over its soft limit
1820 enum mem_cgroup_filter_t
1821 mem_cgroup_soft_reclaim_eligible(struct mem_cgroup
*memcg
,
1822 struct mem_cgroup
*root
)
1824 struct mem_cgroup
*parent
= memcg
;
1826 if (res_counter_soft_limit_excess(&memcg
->res
))
1830 * If any parent up to the root in the hierarchy is over its soft limit
1831 * then we have to obey and reclaim from this group as well.
1833 while((parent
= parent_mem_cgroup(parent
))) {
1834 if (res_counter_soft_limit_excess(&parent
->res
))
1844 * Check OOM-Killer is already running under our hierarchy.
1845 * If someone is running, return false.
1846 * Has to be called with memcg_oom_lock
1848 static bool mem_cgroup_oom_lock(struct mem_cgroup
*memcg
)
1850 struct mem_cgroup
*iter
, *failed
= NULL
;
1852 for_each_mem_cgroup_tree(iter
, memcg
) {
1853 if (iter
->oom_lock
) {
1855 * this subtree of our hierarchy is already locked
1856 * so we cannot give a lock.
1859 mem_cgroup_iter_break(memcg
, iter
);
1862 iter
->oom_lock
= true;
1869 * OK, we failed to lock the whole subtree so we have to clean up
1870 * what we set up to the failing subtree
1872 for_each_mem_cgroup_tree(iter
, memcg
) {
1873 if (iter
== failed
) {
1874 mem_cgroup_iter_break(memcg
, iter
);
1877 iter
->oom_lock
= false;
1883 * Has to be called with memcg_oom_lock
1885 static int mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1887 struct mem_cgroup
*iter
;
1889 for_each_mem_cgroup_tree(iter
, memcg
)
1890 iter
->oom_lock
= false;
1894 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1896 struct mem_cgroup
*iter
;
1898 for_each_mem_cgroup_tree(iter
, memcg
)
1899 atomic_inc(&iter
->under_oom
);
1902 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1904 struct mem_cgroup
*iter
;
1907 * When a new child is created while the hierarchy is under oom,
1908 * mem_cgroup_oom_lock() may not be called. We have to use
1909 * atomic_add_unless() here.
1911 for_each_mem_cgroup_tree(iter
, memcg
)
1912 atomic_add_unless(&iter
->under_oom
, -1, 0);
1915 static DEFINE_SPINLOCK(memcg_oom_lock
);
1916 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1918 struct oom_wait_info
{
1919 struct mem_cgroup
*memcg
;
1923 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1924 unsigned mode
, int sync
, void *arg
)
1926 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
1927 struct mem_cgroup
*oom_wait_memcg
;
1928 struct oom_wait_info
*oom_wait_info
;
1930 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1931 oom_wait_memcg
= oom_wait_info
->memcg
;
1934 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1935 * Then we can use css_is_ancestor without taking care of RCU.
1937 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
1938 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
1940 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1943 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
1945 /* for filtering, pass "memcg" as argument. */
1946 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
1949 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
1951 if (memcg
&& atomic_read(&memcg
->under_oom
))
1952 memcg_wakeup_oom(memcg
);
1956 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1958 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
1961 struct oom_wait_info owait
;
1962 bool locked
, need_to_kill
;
1964 owait
.memcg
= memcg
;
1965 owait
.wait
.flags
= 0;
1966 owait
.wait
.func
= memcg_oom_wake_function
;
1967 owait
.wait
.private = current
;
1968 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1969 need_to_kill
= true;
1970 mem_cgroup_mark_under_oom(memcg
);
1972 /* At first, try to OOM lock hierarchy under memcg.*/
1973 spin_lock(&memcg_oom_lock
);
1974 locked
= mem_cgroup_oom_lock(memcg
);
1976 * Even if signal_pending(), we can't quit charge() loop without
1977 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1978 * under OOM is always welcomed, use TASK_KILLABLE here.
1980 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1981 if (!locked
|| memcg
->oom_kill_disable
)
1982 need_to_kill
= false;
1984 mem_cgroup_oom_notify(memcg
);
1985 spin_unlock(&memcg_oom_lock
);
1988 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1989 mem_cgroup_out_of_memory(memcg
, mask
, order
);
1992 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1994 spin_lock(&memcg_oom_lock
);
1996 mem_cgroup_oom_unlock(memcg
);
1997 memcg_wakeup_oom(memcg
);
1998 spin_unlock(&memcg_oom_lock
);
2000 mem_cgroup_unmark_under_oom(memcg
);
2002 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2004 /* Give chance to dying process */
2005 schedule_timeout_uninterruptible(1);
2010 * Currently used to update mapped file statistics, but the routine can be
2011 * generalized to update other statistics as well.
2013 * Notes: Race condition
2015 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2016 * it tends to be costly. But considering some conditions, we doesn't need
2017 * to do so _always_.
2019 * Considering "charge", lock_page_cgroup() is not required because all
2020 * file-stat operations happen after a page is attached to radix-tree. There
2021 * are no race with "charge".
2023 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2024 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2025 * if there are race with "uncharge". Statistics itself is properly handled
2028 * Considering "move", this is an only case we see a race. To make the race
2029 * small, we check mm->moving_account and detect there are possibility of race
2030 * If there is, we take a lock.
2033 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2034 bool *locked
, unsigned long *flags
)
2036 struct mem_cgroup
*memcg
;
2037 struct page_cgroup
*pc
;
2039 pc
= lookup_page_cgroup(page
);
2041 memcg
= pc
->mem_cgroup
;
2042 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2045 * If this memory cgroup is not under account moving, we don't
2046 * need to take move_lock_mem_cgroup(). Because we already hold
2047 * rcu_read_lock(), any calls to move_account will be delayed until
2048 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2050 if (!mem_cgroup_stolen(memcg
))
2053 move_lock_mem_cgroup(memcg
, flags
);
2054 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2055 move_unlock_mem_cgroup(memcg
, flags
);
2061 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2063 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2066 * It's guaranteed that pc->mem_cgroup never changes while
2067 * lock is held because a routine modifies pc->mem_cgroup
2068 * should take move_lock_mem_cgroup().
2070 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2073 void mem_cgroup_update_page_stat(struct page
*page
,
2074 enum mem_cgroup_page_stat_item idx
, int val
)
2076 struct mem_cgroup
*memcg
;
2077 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2078 unsigned long uninitialized_var(flags
);
2080 if (mem_cgroup_disabled())
2083 memcg
= pc
->mem_cgroup
;
2084 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2088 case MEMCG_NR_FILE_MAPPED
:
2089 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2095 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2099 * size of first charge trial. "32" comes from vmscan.c's magic value.
2100 * TODO: maybe necessary to use big numbers in big irons.
2102 #define CHARGE_BATCH 32U
2103 struct memcg_stock_pcp
{
2104 struct mem_cgroup
*cached
; /* this never be root cgroup */
2105 unsigned int nr_pages
;
2106 struct work_struct work
;
2107 unsigned long flags
;
2108 #define FLUSHING_CACHED_CHARGE 0
2110 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2111 static DEFINE_MUTEX(percpu_charge_mutex
);
2114 * consume_stock: Try to consume stocked charge on this cpu.
2115 * @memcg: memcg to consume from.
2116 * @nr_pages: how many pages to charge.
2118 * The charges will only happen if @memcg matches the current cpu's memcg
2119 * stock, and at least @nr_pages are available in that stock. Failure to
2120 * service an allocation will refill the stock.
2122 * returns true if successful, false otherwise.
2124 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2126 struct memcg_stock_pcp
*stock
;
2129 if (nr_pages
> CHARGE_BATCH
)
2132 stock
= &get_cpu_var(memcg_stock
);
2133 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2134 stock
->nr_pages
-= nr_pages
;
2135 else /* need to call res_counter_charge */
2137 put_cpu_var(memcg_stock
);
2142 * Returns stocks cached in percpu to res_counter and reset cached information.
2144 static void drain_stock(struct memcg_stock_pcp
*stock
)
2146 struct mem_cgroup
*old
= stock
->cached
;
2148 if (stock
->nr_pages
) {
2149 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2151 res_counter_uncharge(&old
->res
, bytes
);
2152 if (do_swap_account
)
2153 res_counter_uncharge(&old
->memsw
, bytes
);
2154 stock
->nr_pages
= 0;
2156 stock
->cached
= NULL
;
2160 * This must be called under preempt disabled or must be called by
2161 * a thread which is pinned to local cpu.
2163 static void drain_local_stock(struct work_struct
*dummy
)
2165 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2167 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2170 static void __init
memcg_stock_init(void)
2174 for_each_possible_cpu(cpu
) {
2175 struct memcg_stock_pcp
*stock
=
2176 &per_cpu(memcg_stock
, cpu
);
2177 INIT_WORK(&stock
->work
, drain_local_stock
);
2182 * Cache charges(val) which is from res_counter, to local per_cpu area.
2183 * This will be consumed by consume_stock() function, later.
2185 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2187 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2189 if (stock
->cached
!= memcg
) { /* reset if necessary */
2191 stock
->cached
= memcg
;
2193 stock
->nr_pages
+= nr_pages
;
2194 put_cpu_var(memcg_stock
);
2198 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2199 * of the hierarchy under it. sync flag says whether we should block
2200 * until the work is done.
2202 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2206 /* Notify other cpus that system-wide "drain" is running */
2209 for_each_online_cpu(cpu
) {
2210 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2211 struct mem_cgroup
*memcg
;
2213 memcg
= stock
->cached
;
2214 if (!memcg
|| !stock
->nr_pages
)
2216 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2218 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2220 drain_local_stock(&stock
->work
);
2222 schedule_work_on(cpu
, &stock
->work
);
2230 for_each_online_cpu(cpu
) {
2231 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2232 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2233 flush_work(&stock
->work
);
2240 * Tries to drain stocked charges in other cpus. This function is asynchronous
2241 * and just put a work per cpu for draining localy on each cpu. Caller can
2242 * expects some charges will be back to res_counter later but cannot wait for
2245 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2248 * If someone calls draining, avoid adding more kworker runs.
2250 if (!mutex_trylock(&percpu_charge_mutex
))
2252 drain_all_stock(root_memcg
, false);
2253 mutex_unlock(&percpu_charge_mutex
);
2256 /* This is a synchronous drain interface. */
2257 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2259 /* called when force_empty is called */
2260 mutex_lock(&percpu_charge_mutex
);
2261 drain_all_stock(root_memcg
, true);
2262 mutex_unlock(&percpu_charge_mutex
);
2266 * This function drains percpu counter value from DEAD cpu and
2267 * move it to local cpu. Note that this function can be preempted.
2269 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2273 spin_lock(&memcg
->pcp_counter_lock
);
2274 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2275 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2277 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2278 memcg
->nocpu_base
.count
[i
] += x
;
2280 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2281 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2283 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2284 memcg
->nocpu_base
.events
[i
] += x
;
2286 spin_unlock(&memcg
->pcp_counter_lock
);
2289 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2290 unsigned long action
,
2293 int cpu
= (unsigned long)hcpu
;
2294 struct memcg_stock_pcp
*stock
;
2295 struct mem_cgroup
*iter
;
2297 if (action
== CPU_ONLINE
)
2300 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2303 for_each_mem_cgroup(iter
)
2304 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2306 stock
= &per_cpu(memcg_stock
, cpu
);
2312 /* See __mem_cgroup_try_charge() for details */
2314 CHARGE_OK
, /* success */
2315 CHARGE_RETRY
, /* need to retry but retry is not bad */
2316 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2317 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2318 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2321 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2322 unsigned int nr_pages
, unsigned int min_pages
,
2325 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2326 struct mem_cgroup
*mem_over_limit
;
2327 struct res_counter
*fail_res
;
2328 unsigned long flags
= 0;
2331 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2334 if (!do_swap_account
)
2336 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2340 res_counter_uncharge(&memcg
->res
, csize
);
2341 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2342 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2344 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2346 * Never reclaim on behalf of optional batching, retry with a
2347 * single page instead.
2349 if (nr_pages
> min_pages
)
2350 return CHARGE_RETRY
;
2352 if (!(gfp_mask
& __GFP_WAIT
))
2353 return CHARGE_WOULDBLOCK
;
2355 if (gfp_mask
& __GFP_NORETRY
)
2356 return CHARGE_NOMEM
;
2358 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2359 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2360 return CHARGE_RETRY
;
2362 * Even though the limit is exceeded at this point, reclaim
2363 * may have been able to free some pages. Retry the charge
2364 * before killing the task.
2366 * Only for regular pages, though: huge pages are rather
2367 * unlikely to succeed so close to the limit, and we fall back
2368 * to regular pages anyway in case of failure.
2370 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2371 return CHARGE_RETRY
;
2374 * At task move, charge accounts can be doubly counted. So, it's
2375 * better to wait until the end of task_move if something is going on.
2377 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2378 return CHARGE_RETRY
;
2380 /* If we don't need to call oom-killer at el, return immediately */
2382 return CHARGE_NOMEM
;
2384 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2385 return CHARGE_OOM_DIE
;
2387 return CHARGE_RETRY
;
2391 * __mem_cgroup_try_charge() does
2392 * 1. detect memcg to be charged against from passed *mm and *ptr,
2393 * 2. update res_counter
2394 * 3. call memory reclaim if necessary.
2396 * In some special case, if the task is fatal, fatal_signal_pending() or
2397 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2398 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2399 * as possible without any hazards. 2: all pages should have a valid
2400 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2401 * pointer, that is treated as a charge to root_mem_cgroup.
2403 * So __mem_cgroup_try_charge() will return
2404 * 0 ... on success, filling *ptr with a valid memcg pointer.
2405 * -ENOMEM ... charge failure because of resource limits.
2406 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2408 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2409 * the oom-killer can be invoked.
2411 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2413 unsigned int nr_pages
,
2414 struct mem_cgroup
**ptr
,
2417 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2418 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2419 struct mem_cgroup
*memcg
= NULL
;
2423 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2424 * in system level. So, allow to go ahead dying process in addition to
2427 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2428 || fatal_signal_pending(current
)))
2432 * We always charge the cgroup the mm_struct belongs to.
2433 * The mm_struct's mem_cgroup changes on task migration if the
2434 * thread group leader migrates. It's possible that mm is not
2435 * set, if so charge the root memcg (happens for pagecache usage).
2438 *ptr
= root_mem_cgroup
;
2440 if (*ptr
) { /* css should be a valid one */
2442 if (mem_cgroup_is_root(memcg
))
2444 if (consume_stock(memcg
, nr_pages
))
2446 css_get(&memcg
->css
);
2448 struct task_struct
*p
;
2451 p
= rcu_dereference(mm
->owner
);
2453 * Because we don't have task_lock(), "p" can exit.
2454 * In that case, "memcg" can point to root or p can be NULL with
2455 * race with swapoff. Then, we have small risk of mis-accouning.
2456 * But such kind of mis-account by race always happens because
2457 * we don't have cgroup_mutex(). It's overkill and we allo that
2459 * (*) swapoff at el will charge against mm-struct not against
2460 * task-struct. So, mm->owner can be NULL.
2462 memcg
= mem_cgroup_from_task(p
);
2464 memcg
= root_mem_cgroup
;
2465 if (mem_cgroup_is_root(memcg
)) {
2469 if (consume_stock(memcg
, nr_pages
)) {
2471 * It seems dagerous to access memcg without css_get().
2472 * But considering how consume_stok works, it's not
2473 * necessary. If consume_stock success, some charges
2474 * from this memcg are cached on this cpu. So, we
2475 * don't need to call css_get()/css_tryget() before
2476 * calling consume_stock().
2481 /* after here, we may be blocked. we need to get refcnt */
2482 if (!css_tryget(&memcg
->css
)) {
2492 /* If killed, bypass charge */
2493 if (fatal_signal_pending(current
)) {
2494 css_put(&memcg
->css
);
2499 if (oom
&& !nr_oom_retries
) {
2501 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2504 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2509 case CHARGE_RETRY
: /* not in OOM situation but retry */
2511 css_put(&memcg
->css
);
2514 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2515 css_put(&memcg
->css
);
2517 case CHARGE_NOMEM
: /* OOM routine works */
2519 css_put(&memcg
->css
);
2522 /* If oom, we never return -ENOMEM */
2525 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2526 css_put(&memcg
->css
);
2529 } while (ret
!= CHARGE_OK
);
2531 if (batch
> nr_pages
)
2532 refill_stock(memcg
, batch
- nr_pages
);
2533 css_put(&memcg
->css
);
2541 *ptr
= root_mem_cgroup
;
2546 * Somemtimes we have to undo a charge we got by try_charge().
2547 * This function is for that and do uncharge, put css's refcnt.
2548 * gotten by try_charge().
2550 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2551 unsigned int nr_pages
)
2553 if (!mem_cgroup_is_root(memcg
)) {
2554 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2556 res_counter_uncharge(&memcg
->res
, bytes
);
2557 if (do_swap_account
)
2558 res_counter_uncharge(&memcg
->memsw
, bytes
);
2563 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2564 * This is useful when moving usage to parent cgroup.
2566 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2567 unsigned int nr_pages
)
2569 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2571 if (mem_cgroup_is_root(memcg
))
2574 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2575 if (do_swap_account
)
2576 res_counter_uncharge_until(&memcg
->memsw
,
2577 memcg
->memsw
.parent
, bytes
);
2581 * A helper function to get mem_cgroup from ID. must be called under
2582 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2583 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2584 * called against removed memcg.)
2586 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2588 struct cgroup_subsys_state
*css
;
2590 /* ID 0 is unused ID */
2593 css
= css_lookup(&mem_cgroup_subsys
, id
);
2596 return mem_cgroup_from_css(css
);
2599 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2601 struct mem_cgroup
*memcg
= NULL
;
2602 struct page_cgroup
*pc
;
2606 VM_BUG_ON(!PageLocked(page
));
2608 pc
= lookup_page_cgroup(page
);
2609 lock_page_cgroup(pc
);
2610 if (PageCgroupUsed(pc
)) {
2611 memcg
= pc
->mem_cgroup
;
2612 if (memcg
&& !css_tryget(&memcg
->css
))
2614 } else if (PageSwapCache(page
)) {
2615 ent
.val
= page_private(page
);
2616 id
= lookup_swap_cgroup_id(ent
);
2618 memcg
= mem_cgroup_lookup(id
);
2619 if (memcg
&& !css_tryget(&memcg
->css
))
2623 unlock_page_cgroup(pc
);
2627 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2629 unsigned int nr_pages
,
2630 enum charge_type ctype
,
2633 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2634 struct zone
*uninitialized_var(zone
);
2635 struct lruvec
*lruvec
;
2636 bool was_on_lru
= false;
2639 lock_page_cgroup(pc
);
2640 VM_BUG_ON(PageCgroupUsed(pc
));
2642 * we don't need page_cgroup_lock about tail pages, becase they are not
2643 * accessed by any other context at this point.
2647 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2648 * may already be on some other mem_cgroup's LRU. Take care of it.
2651 zone
= page_zone(page
);
2652 spin_lock_irq(&zone
->lru_lock
);
2653 if (PageLRU(page
)) {
2654 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2656 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2661 pc
->mem_cgroup
= memcg
;
2663 * We access a page_cgroup asynchronously without lock_page_cgroup().
2664 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2665 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2666 * before USED bit, we need memory barrier here.
2667 * See mem_cgroup_add_lru_list(), etc.
2670 SetPageCgroupUsed(pc
);
2674 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2675 VM_BUG_ON(PageLRU(page
));
2677 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2679 spin_unlock_irq(&zone
->lru_lock
);
2682 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2687 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2688 unlock_page_cgroup(pc
);
2691 * "charge_statistics" updated event counter.
2693 memcg_check_events(memcg
, page
);
2696 static DEFINE_MUTEX(set_limit_mutex
);
2698 #ifdef CONFIG_MEMCG_KMEM
2699 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2701 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2702 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2706 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2707 * in the memcg_cache_params struct.
2709 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2711 struct kmem_cache
*cachep
;
2713 VM_BUG_ON(p
->is_root_cache
);
2714 cachep
= p
->root_cache
;
2715 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2718 #ifdef CONFIG_SLABINFO
2719 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2720 struct cftype
*cft
, struct seq_file
*m
)
2722 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2723 struct memcg_cache_params
*params
;
2725 if (!memcg_can_account_kmem(memcg
))
2728 print_slabinfo_header(m
);
2730 mutex_lock(&memcg
->slab_caches_mutex
);
2731 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2732 cache_show(memcg_params_to_cache(params
), m
);
2733 mutex_unlock(&memcg
->slab_caches_mutex
);
2739 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2741 struct res_counter
*fail_res
;
2742 struct mem_cgroup
*_memcg
;
2746 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2751 * Conditions under which we can wait for the oom_killer. Those are
2752 * the same conditions tested by the core page allocator
2754 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2757 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2760 if (ret
== -EINTR
) {
2762 * __mem_cgroup_try_charge() chosed to bypass to root due to
2763 * OOM kill or fatal signal. Since our only options are to
2764 * either fail the allocation or charge it to this cgroup, do
2765 * it as a temporary condition. But we can't fail. From a
2766 * kmem/slab perspective, the cache has already been selected,
2767 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2770 * This condition will only trigger if the task entered
2771 * memcg_charge_kmem in a sane state, but was OOM-killed during
2772 * __mem_cgroup_try_charge() above. Tasks that were already
2773 * dying when the allocation triggers should have been already
2774 * directed to the root cgroup in memcontrol.h
2776 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2777 if (do_swap_account
)
2778 res_counter_charge_nofail(&memcg
->memsw
, size
,
2782 res_counter_uncharge(&memcg
->kmem
, size
);
2787 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2789 res_counter_uncharge(&memcg
->res
, size
);
2790 if (do_swap_account
)
2791 res_counter_uncharge(&memcg
->memsw
, size
);
2794 if (res_counter_uncharge(&memcg
->kmem
, size
))
2798 * Releases a reference taken in kmem_cgroup_css_offline in case
2799 * this last uncharge is racing with the offlining code or it is
2800 * outliving the memcg existence.
2802 * The memory barrier imposed by test&clear is paired with the
2803 * explicit one in memcg_kmem_mark_dead().
2805 if (memcg_kmem_test_and_clear_dead(memcg
))
2806 css_put(&memcg
->css
);
2809 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2814 mutex_lock(&memcg
->slab_caches_mutex
);
2815 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2816 mutex_unlock(&memcg
->slab_caches_mutex
);
2820 * helper for acessing a memcg's index. It will be used as an index in the
2821 * child cache array in kmem_cache, and also to derive its name. This function
2822 * will return -1 when this is not a kmem-limited memcg.
2824 int memcg_cache_id(struct mem_cgroup
*memcg
)
2826 return memcg
? memcg
->kmemcg_id
: -1;
2830 * This ends up being protected by the set_limit mutex, during normal
2831 * operation, because that is its main call site.
2833 * But when we create a new cache, we can call this as well if its parent
2834 * is kmem-limited. That will have to hold set_limit_mutex as well.
2836 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2840 num
= ida_simple_get(&kmem_limited_groups
,
2841 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2845 * After this point, kmem_accounted (that we test atomically in
2846 * the beginning of this conditional), is no longer 0. This
2847 * guarantees only one process will set the following boolean
2848 * to true. We don't need test_and_set because we're protected
2849 * by the set_limit_mutex anyway.
2851 memcg_kmem_set_activated(memcg
);
2853 ret
= memcg_update_all_caches(num
+1);
2855 ida_simple_remove(&kmem_limited_groups
, num
);
2856 memcg_kmem_clear_activated(memcg
);
2860 memcg
->kmemcg_id
= num
;
2861 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
2862 mutex_init(&memcg
->slab_caches_mutex
);
2866 static size_t memcg_caches_array_size(int num_groups
)
2869 if (num_groups
<= 0)
2872 size
= 2 * num_groups
;
2873 if (size
< MEMCG_CACHES_MIN_SIZE
)
2874 size
= MEMCG_CACHES_MIN_SIZE
;
2875 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2876 size
= MEMCG_CACHES_MAX_SIZE
;
2882 * We should update the current array size iff all caches updates succeed. This
2883 * can only be done from the slab side. The slab mutex needs to be held when
2886 void memcg_update_array_size(int num
)
2888 if (num
> memcg_limited_groups_array_size
)
2889 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2892 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
2894 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2896 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2898 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
2900 if (num_groups
> memcg_limited_groups_array_size
) {
2902 ssize_t size
= memcg_caches_array_size(num_groups
);
2904 size
*= sizeof(void *);
2905 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
2907 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2908 if (!s
->memcg_params
) {
2909 s
->memcg_params
= cur_params
;
2913 s
->memcg_params
->is_root_cache
= true;
2916 * There is the chance it will be bigger than
2917 * memcg_limited_groups_array_size, if we failed an allocation
2918 * in a cache, in which case all caches updated before it, will
2919 * have a bigger array.
2921 * But if that is the case, the data after
2922 * memcg_limited_groups_array_size is certainly unused
2924 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
2925 if (!cur_params
->memcg_caches
[i
])
2927 s
->memcg_params
->memcg_caches
[i
] =
2928 cur_params
->memcg_caches
[i
];
2932 * Ideally, we would wait until all caches succeed, and only
2933 * then free the old one. But this is not worth the extra
2934 * pointer per-cache we'd have to have for this.
2936 * It is not a big deal if some caches are left with a size
2937 * bigger than the others. And all updates will reset this
2945 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
2946 struct kmem_cache
*root_cache
)
2950 if (!memcg_kmem_enabled())
2954 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
2955 size
+= memcg_limited_groups_array_size
* sizeof(void *);
2957 size
= sizeof(struct memcg_cache_params
);
2959 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2960 if (!s
->memcg_params
)
2964 s
->memcg_params
->memcg
= memcg
;
2965 s
->memcg_params
->root_cache
= root_cache
;
2966 INIT_WORK(&s
->memcg_params
->destroy
,
2967 kmem_cache_destroy_work_func
);
2969 s
->memcg_params
->is_root_cache
= true;
2974 void memcg_release_cache(struct kmem_cache
*s
)
2976 struct kmem_cache
*root
;
2977 struct mem_cgroup
*memcg
;
2981 * This happens, for instance, when a root cache goes away before we
2984 if (!s
->memcg_params
)
2987 if (s
->memcg_params
->is_root_cache
)
2990 memcg
= s
->memcg_params
->memcg
;
2991 id
= memcg_cache_id(memcg
);
2993 root
= s
->memcg_params
->root_cache
;
2994 root
->memcg_params
->memcg_caches
[id
] = NULL
;
2996 mutex_lock(&memcg
->slab_caches_mutex
);
2997 list_del(&s
->memcg_params
->list
);
2998 mutex_unlock(&memcg
->slab_caches_mutex
);
3000 css_put(&memcg
->css
);
3002 kfree(s
->memcg_params
);
3006 * During the creation a new cache, we need to disable our accounting mechanism
3007 * altogether. This is true even if we are not creating, but rather just
3008 * enqueing new caches to be created.
3010 * This is because that process will trigger allocations; some visible, like
3011 * explicit kmallocs to auxiliary data structures, name strings and internal
3012 * cache structures; some well concealed, like INIT_WORK() that can allocate
3013 * objects during debug.
3015 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3016 * to it. This may not be a bounded recursion: since the first cache creation
3017 * failed to complete (waiting on the allocation), we'll just try to create the
3018 * cache again, failing at the same point.
3020 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3021 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3022 * inside the following two functions.
3024 static inline void memcg_stop_kmem_account(void)
3026 VM_BUG_ON(!current
->mm
);
3027 current
->memcg_kmem_skip_account
++;
3030 static inline void memcg_resume_kmem_account(void)
3032 VM_BUG_ON(!current
->mm
);
3033 current
->memcg_kmem_skip_account
--;
3036 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3038 struct kmem_cache
*cachep
;
3039 struct memcg_cache_params
*p
;
3041 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3043 cachep
= memcg_params_to_cache(p
);
3046 * If we get down to 0 after shrink, we could delete right away.
3047 * However, memcg_release_pages() already puts us back in the workqueue
3048 * in that case. If we proceed deleting, we'll get a dangling
3049 * reference, and removing the object from the workqueue in that case
3050 * is unnecessary complication. We are not a fast path.
3052 * Note that this case is fundamentally different from racing with
3053 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3054 * kmem_cache_shrink, not only we would be reinserting a dead cache
3055 * into the queue, but doing so from inside the worker racing to
3058 * So if we aren't down to zero, we'll just schedule a worker and try
3061 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3062 kmem_cache_shrink(cachep
);
3063 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3066 kmem_cache_destroy(cachep
);
3069 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3071 if (!cachep
->memcg_params
->dead
)
3075 * There are many ways in which we can get here.
3077 * We can get to a memory-pressure situation while the delayed work is
3078 * still pending to run. The vmscan shrinkers can then release all
3079 * cache memory and get us to destruction. If this is the case, we'll
3080 * be executed twice, which is a bug (the second time will execute over
3081 * bogus data). In this case, cancelling the work should be fine.
3083 * But we can also get here from the worker itself, if
3084 * kmem_cache_shrink is enough to shake all the remaining objects and
3085 * get the page count to 0. In this case, we'll deadlock if we try to
3086 * cancel the work (the worker runs with an internal lock held, which
3087 * is the same lock we would hold for cancel_work_sync().)
3089 * Since we can't possibly know who got us here, just refrain from
3090 * running if there is already work pending
3092 if (work_pending(&cachep
->memcg_params
->destroy
))
3095 * We have to defer the actual destroying to a workqueue, because
3096 * we might currently be in a context that cannot sleep.
3098 schedule_work(&cachep
->memcg_params
->destroy
);
3102 * This lock protects updaters, not readers. We want readers to be as fast as
3103 * they can, and they will either see NULL or a valid cache value. Our model
3104 * allow them to see NULL, in which case the root memcg will be selected.
3106 * We need this lock because multiple allocations to the same cache from a non
3107 * will span more than one worker. Only one of them can create the cache.
3109 static DEFINE_MUTEX(memcg_cache_mutex
);
3112 * Called with memcg_cache_mutex held
3114 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3115 struct kmem_cache
*s
)
3117 struct kmem_cache
*new;
3118 static char *tmp_name
= NULL
;
3120 lockdep_assert_held(&memcg_cache_mutex
);
3123 * kmem_cache_create_memcg duplicates the given name and
3124 * cgroup_name for this name requires RCU context.
3125 * This static temporary buffer is used to prevent from
3126 * pointless shortliving allocation.
3129 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3135 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3136 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3139 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3140 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3143 new->allocflags
|= __GFP_KMEMCG
;
3148 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3149 struct kmem_cache
*cachep
)
3151 struct kmem_cache
*new_cachep
;
3154 BUG_ON(!memcg_can_account_kmem(memcg
));
3156 idx
= memcg_cache_id(memcg
);
3158 mutex_lock(&memcg_cache_mutex
);
3159 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3161 css_put(&memcg
->css
);
3165 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3166 if (new_cachep
== NULL
) {
3167 new_cachep
= cachep
;
3168 css_put(&memcg
->css
);
3172 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3174 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3176 * the readers won't lock, make sure everybody sees the updated value,
3177 * so they won't put stuff in the queue again for no reason
3181 mutex_unlock(&memcg_cache_mutex
);
3185 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3187 struct kmem_cache
*c
;
3190 if (!s
->memcg_params
)
3192 if (!s
->memcg_params
->is_root_cache
)
3196 * If the cache is being destroyed, we trust that there is no one else
3197 * requesting objects from it. Even if there are, the sanity checks in
3198 * kmem_cache_destroy should caught this ill-case.
3200 * Still, we don't want anyone else freeing memcg_caches under our
3201 * noses, which can happen if a new memcg comes to life. As usual,
3202 * we'll take the set_limit_mutex to protect ourselves against this.
3204 mutex_lock(&set_limit_mutex
);
3205 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3206 c
= s
->memcg_params
->memcg_caches
[i
];
3211 * We will now manually delete the caches, so to avoid races
3212 * we need to cancel all pending destruction workers and
3213 * proceed with destruction ourselves.
3215 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3216 * and that could spawn the workers again: it is likely that
3217 * the cache still have active pages until this very moment.
3218 * This would lead us back to mem_cgroup_destroy_cache.
3220 * But that will not execute at all if the "dead" flag is not
3221 * set, so flip it down to guarantee we are in control.
3223 c
->memcg_params
->dead
= false;
3224 cancel_work_sync(&c
->memcg_params
->destroy
);
3225 kmem_cache_destroy(c
);
3227 mutex_unlock(&set_limit_mutex
);
3230 struct create_work
{
3231 struct mem_cgroup
*memcg
;
3232 struct kmem_cache
*cachep
;
3233 struct work_struct work
;
3236 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3238 struct kmem_cache
*cachep
;
3239 struct memcg_cache_params
*params
;
3241 if (!memcg_kmem_is_active(memcg
))
3244 mutex_lock(&memcg
->slab_caches_mutex
);
3245 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3246 cachep
= memcg_params_to_cache(params
);
3247 cachep
->memcg_params
->dead
= true;
3248 schedule_work(&cachep
->memcg_params
->destroy
);
3250 mutex_unlock(&memcg
->slab_caches_mutex
);
3253 static void memcg_create_cache_work_func(struct work_struct
*w
)
3255 struct create_work
*cw
;
3257 cw
= container_of(w
, struct create_work
, work
);
3258 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3263 * Enqueue the creation of a per-memcg kmem_cache.
3265 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3266 struct kmem_cache
*cachep
)
3268 struct create_work
*cw
;
3270 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3272 css_put(&memcg
->css
);
3277 cw
->cachep
= cachep
;
3279 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3280 schedule_work(&cw
->work
);
3283 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3284 struct kmem_cache
*cachep
)
3287 * We need to stop accounting when we kmalloc, because if the
3288 * corresponding kmalloc cache is not yet created, the first allocation
3289 * in __memcg_create_cache_enqueue will recurse.
3291 * However, it is better to enclose the whole function. Depending on
3292 * the debugging options enabled, INIT_WORK(), for instance, can
3293 * trigger an allocation. This too, will make us recurse. Because at
3294 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3295 * the safest choice is to do it like this, wrapping the whole function.
3297 memcg_stop_kmem_account();
3298 __memcg_create_cache_enqueue(memcg
, cachep
);
3299 memcg_resume_kmem_account();
3302 * Return the kmem_cache we're supposed to use for a slab allocation.
3303 * We try to use the current memcg's version of the cache.
3305 * If the cache does not exist yet, if we are the first user of it,
3306 * we either create it immediately, if possible, or create it asynchronously
3308 * In the latter case, we will let the current allocation go through with
3309 * the original cache.
3311 * Can't be called in interrupt context or from kernel threads.
3312 * This function needs to be called with rcu_read_lock() held.
3314 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3317 struct mem_cgroup
*memcg
;
3320 VM_BUG_ON(!cachep
->memcg_params
);
3321 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3323 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3327 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3329 if (!memcg_can_account_kmem(memcg
))
3332 idx
= memcg_cache_id(memcg
);
3335 * barrier to mare sure we're always seeing the up to date value. The
3336 * code updating memcg_caches will issue a write barrier to match this.
3338 read_barrier_depends();
3339 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3340 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3344 /* The corresponding put will be done in the workqueue. */
3345 if (!css_tryget(&memcg
->css
))
3350 * If we are in a safe context (can wait, and not in interrupt
3351 * context), we could be be predictable and return right away.
3352 * This would guarantee that the allocation being performed
3353 * already belongs in the new cache.
3355 * However, there are some clashes that can arrive from locking.
3356 * For instance, because we acquire the slab_mutex while doing
3357 * kmem_cache_dup, this means no further allocation could happen
3358 * with the slab_mutex held.
3360 * Also, because cache creation issue get_online_cpus(), this
3361 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3362 * that ends up reversed during cpu hotplug. (cpuset allocates
3363 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3364 * better to defer everything.
3366 memcg_create_cache_enqueue(memcg
, cachep
);
3372 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3375 * We need to verify if the allocation against current->mm->owner's memcg is
3376 * possible for the given order. But the page is not allocated yet, so we'll
3377 * need a further commit step to do the final arrangements.
3379 * It is possible for the task to switch cgroups in this mean time, so at
3380 * commit time, we can't rely on task conversion any longer. We'll then use
3381 * the handle argument to return to the caller which cgroup we should commit
3382 * against. We could also return the memcg directly and avoid the pointer
3383 * passing, but a boolean return value gives better semantics considering
3384 * the compiled-out case as well.
3386 * Returning true means the allocation is possible.
3389 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3391 struct mem_cgroup
*memcg
;
3397 * Disabling accounting is only relevant for some specific memcg
3398 * internal allocations. Therefore we would initially not have such
3399 * check here, since direct calls to the page allocator that are marked
3400 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3401 * concerned with cache allocations, and by having this test at
3402 * memcg_kmem_get_cache, we are already able to relay the allocation to
3403 * the root cache and bypass the memcg cache altogether.
3405 * There is one exception, though: the SLUB allocator does not create
3406 * large order caches, but rather service large kmallocs directly from
3407 * the page allocator. Therefore, the following sequence when backed by
3408 * the SLUB allocator:
3410 * memcg_stop_kmem_account();
3411 * kmalloc(<large_number>)
3412 * memcg_resume_kmem_account();
3414 * would effectively ignore the fact that we should skip accounting,
3415 * since it will drive us directly to this function without passing
3416 * through the cache selector memcg_kmem_get_cache. Such large
3417 * allocations are extremely rare but can happen, for instance, for the
3418 * cache arrays. We bring this test here.
3420 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3423 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3426 * very rare case described in mem_cgroup_from_task. Unfortunately there
3427 * isn't much we can do without complicating this too much, and it would
3428 * be gfp-dependent anyway. Just let it go
3430 if (unlikely(!memcg
))
3433 if (!memcg_can_account_kmem(memcg
)) {
3434 css_put(&memcg
->css
);
3438 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3442 css_put(&memcg
->css
);
3446 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3449 struct page_cgroup
*pc
;
3451 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3453 /* The page allocation failed. Revert */
3455 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3459 pc
= lookup_page_cgroup(page
);
3460 lock_page_cgroup(pc
);
3461 pc
->mem_cgroup
= memcg
;
3462 SetPageCgroupUsed(pc
);
3463 unlock_page_cgroup(pc
);
3466 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3468 struct mem_cgroup
*memcg
= NULL
;
3469 struct page_cgroup
*pc
;
3472 pc
= lookup_page_cgroup(page
);
3474 * Fast unlocked return. Theoretically might have changed, have to
3475 * check again after locking.
3477 if (!PageCgroupUsed(pc
))
3480 lock_page_cgroup(pc
);
3481 if (PageCgroupUsed(pc
)) {
3482 memcg
= pc
->mem_cgroup
;
3483 ClearPageCgroupUsed(pc
);
3485 unlock_page_cgroup(pc
);
3488 * We trust that only if there is a memcg associated with the page, it
3489 * is a valid allocation
3494 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3495 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3498 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3501 #endif /* CONFIG_MEMCG_KMEM */
3503 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3505 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3507 * Because tail pages are not marked as "used", set it. We're under
3508 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3509 * charge/uncharge will be never happen and move_account() is done under
3510 * compound_lock(), so we don't have to take care of races.
3512 void mem_cgroup_split_huge_fixup(struct page
*head
)
3514 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3515 struct page_cgroup
*pc
;
3516 struct mem_cgroup
*memcg
;
3519 if (mem_cgroup_disabled())
3522 memcg
= head_pc
->mem_cgroup
;
3523 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3525 pc
->mem_cgroup
= memcg
;
3526 smp_wmb();/* see __commit_charge() */
3527 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3529 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3532 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3535 * mem_cgroup_move_account - move account of the page
3537 * @nr_pages: number of regular pages (>1 for huge pages)
3538 * @pc: page_cgroup of the page.
3539 * @from: mem_cgroup which the page is moved from.
3540 * @to: mem_cgroup which the page is moved to. @from != @to.
3542 * The caller must confirm following.
3543 * - page is not on LRU (isolate_page() is useful.)
3544 * - compound_lock is held when nr_pages > 1
3546 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3549 static int mem_cgroup_move_account(struct page
*page
,
3550 unsigned int nr_pages
,
3551 struct page_cgroup
*pc
,
3552 struct mem_cgroup
*from
,
3553 struct mem_cgroup
*to
)
3555 unsigned long flags
;
3557 bool anon
= PageAnon(page
);
3559 VM_BUG_ON(from
== to
);
3560 VM_BUG_ON(PageLRU(page
));
3562 * The page is isolated from LRU. So, collapse function
3563 * will not handle this page. But page splitting can happen.
3564 * Do this check under compound_page_lock(). The caller should
3568 if (nr_pages
> 1 && !PageTransHuge(page
))
3571 lock_page_cgroup(pc
);
3574 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3577 move_lock_mem_cgroup(from
, &flags
);
3579 if (!anon
&& page_mapped(page
)) {
3580 /* Update mapped_file data for mem_cgroup */
3582 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3583 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3586 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3588 /* caller should have done css_get */
3589 pc
->mem_cgroup
= to
;
3590 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3591 move_unlock_mem_cgroup(from
, &flags
);
3594 unlock_page_cgroup(pc
);
3598 memcg_check_events(to
, page
);
3599 memcg_check_events(from
, page
);
3605 * mem_cgroup_move_parent - moves page to the parent group
3606 * @page: the page to move
3607 * @pc: page_cgroup of the page
3608 * @child: page's cgroup
3610 * move charges to its parent or the root cgroup if the group has no
3611 * parent (aka use_hierarchy==0).
3612 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3613 * mem_cgroup_move_account fails) the failure is always temporary and
3614 * it signals a race with a page removal/uncharge or migration. In the
3615 * first case the page is on the way out and it will vanish from the LRU
3616 * on the next attempt and the call should be retried later.
3617 * Isolation from the LRU fails only if page has been isolated from
3618 * the LRU since we looked at it and that usually means either global
3619 * reclaim or migration going on. The page will either get back to the
3621 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3622 * (!PageCgroupUsed) or moved to a different group. The page will
3623 * disappear in the next attempt.
3625 static int mem_cgroup_move_parent(struct page
*page
,
3626 struct page_cgroup
*pc
,
3627 struct mem_cgroup
*child
)
3629 struct mem_cgroup
*parent
;
3630 unsigned int nr_pages
;
3631 unsigned long uninitialized_var(flags
);
3634 VM_BUG_ON(mem_cgroup_is_root(child
));
3637 if (!get_page_unless_zero(page
))
3639 if (isolate_lru_page(page
))
3642 nr_pages
= hpage_nr_pages(page
);
3644 parent
= parent_mem_cgroup(child
);
3646 * If no parent, move charges to root cgroup.
3649 parent
= root_mem_cgroup
;
3652 VM_BUG_ON(!PageTransHuge(page
));
3653 flags
= compound_lock_irqsave(page
);
3656 ret
= mem_cgroup_move_account(page
, nr_pages
,
3659 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3662 compound_unlock_irqrestore(page
, flags
);
3663 putback_lru_page(page
);
3671 * Charge the memory controller for page usage.
3673 * 0 if the charge was successful
3674 * < 0 if the cgroup is over its limit
3676 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3677 gfp_t gfp_mask
, enum charge_type ctype
)
3679 struct mem_cgroup
*memcg
= NULL
;
3680 unsigned int nr_pages
= 1;
3684 if (PageTransHuge(page
)) {
3685 nr_pages
<<= compound_order(page
);
3686 VM_BUG_ON(!PageTransHuge(page
));
3688 * Never OOM-kill a process for a huge page. The
3689 * fault handler will fall back to regular pages.
3694 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3697 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3701 int mem_cgroup_newpage_charge(struct page
*page
,
3702 struct mm_struct
*mm
, gfp_t gfp_mask
)
3704 if (mem_cgroup_disabled())
3706 VM_BUG_ON(page_mapped(page
));
3707 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3709 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3710 MEM_CGROUP_CHARGE_TYPE_ANON
);
3714 * While swap-in, try_charge -> commit or cancel, the page is locked.
3715 * And when try_charge() successfully returns, one refcnt to memcg without
3716 * struct page_cgroup is acquired. This refcnt will be consumed by
3717 * "commit()" or removed by "cancel()"
3719 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3722 struct mem_cgroup
**memcgp
)
3724 struct mem_cgroup
*memcg
;
3725 struct page_cgroup
*pc
;
3728 pc
= lookup_page_cgroup(page
);
3730 * Every swap fault against a single page tries to charge the
3731 * page, bail as early as possible. shmem_unuse() encounters
3732 * already charged pages, too. The USED bit is protected by
3733 * the page lock, which serializes swap cache removal, which
3734 * in turn serializes uncharging.
3736 if (PageCgroupUsed(pc
))
3738 if (!do_swap_account
)
3740 memcg
= try_get_mem_cgroup_from_page(page
);
3744 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3745 css_put(&memcg
->css
);
3750 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3756 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3757 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3760 if (mem_cgroup_disabled())
3763 * A racing thread's fault, or swapoff, may have already
3764 * updated the pte, and even removed page from swap cache: in
3765 * those cases unuse_pte()'s pte_same() test will fail; but
3766 * there's also a KSM case which does need to charge the page.
3768 if (!PageSwapCache(page
)) {
3771 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3776 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3779 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3781 if (mem_cgroup_disabled())
3785 __mem_cgroup_cancel_charge(memcg
, 1);
3789 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3790 enum charge_type ctype
)
3792 if (mem_cgroup_disabled())
3797 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3799 * Now swap is on-memory. This means this page may be
3800 * counted both as mem and swap....double count.
3801 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3802 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3803 * may call delete_from_swap_cache() before reach here.
3805 if (do_swap_account
&& PageSwapCache(page
)) {
3806 swp_entry_t ent
= {.val
= page_private(page
)};
3807 mem_cgroup_uncharge_swap(ent
);
3811 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3812 struct mem_cgroup
*memcg
)
3814 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3815 MEM_CGROUP_CHARGE_TYPE_ANON
);
3818 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3821 struct mem_cgroup
*memcg
= NULL
;
3822 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3825 if (mem_cgroup_disabled())
3827 if (PageCompound(page
))
3830 if (!PageSwapCache(page
))
3831 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3832 else { /* page is swapcache/shmem */
3833 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3836 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3841 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3842 unsigned int nr_pages
,
3843 const enum charge_type ctype
)
3845 struct memcg_batch_info
*batch
= NULL
;
3846 bool uncharge_memsw
= true;
3848 /* If swapout, usage of swap doesn't decrease */
3849 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3850 uncharge_memsw
= false;
3852 batch
= ¤t
->memcg_batch
;
3854 * In usual, we do css_get() when we remember memcg pointer.
3855 * But in this case, we keep res->usage until end of a series of
3856 * uncharges. Then, it's ok to ignore memcg's refcnt.
3859 batch
->memcg
= memcg
;
3861 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3862 * In those cases, all pages freed continuously can be expected to be in
3863 * the same cgroup and we have chance to coalesce uncharges.
3864 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3865 * because we want to do uncharge as soon as possible.
3868 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3869 goto direct_uncharge
;
3872 goto direct_uncharge
;
3875 * In typical case, batch->memcg == mem. This means we can
3876 * merge a series of uncharges to an uncharge of res_counter.
3877 * If not, we uncharge res_counter ony by one.
3879 if (batch
->memcg
!= memcg
)
3880 goto direct_uncharge
;
3881 /* remember freed charge and uncharge it later */
3884 batch
->memsw_nr_pages
++;
3887 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3889 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3890 if (unlikely(batch
->memcg
!= memcg
))
3891 memcg_oom_recover(memcg
);
3895 * uncharge if !page_mapped(page)
3897 static struct mem_cgroup
*
3898 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3901 struct mem_cgroup
*memcg
= NULL
;
3902 unsigned int nr_pages
= 1;
3903 struct page_cgroup
*pc
;
3906 if (mem_cgroup_disabled())
3909 if (PageTransHuge(page
)) {
3910 nr_pages
<<= compound_order(page
);
3911 VM_BUG_ON(!PageTransHuge(page
));
3914 * Check if our page_cgroup is valid
3916 pc
= lookup_page_cgroup(page
);
3917 if (unlikely(!PageCgroupUsed(pc
)))
3920 lock_page_cgroup(pc
);
3922 memcg
= pc
->mem_cgroup
;
3924 if (!PageCgroupUsed(pc
))
3927 anon
= PageAnon(page
);
3930 case MEM_CGROUP_CHARGE_TYPE_ANON
:
3932 * Generally PageAnon tells if it's the anon statistics to be
3933 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3934 * used before page reached the stage of being marked PageAnon.
3938 case MEM_CGROUP_CHARGE_TYPE_DROP
:
3939 /* See mem_cgroup_prepare_migration() */
3940 if (page_mapped(page
))
3943 * Pages under migration may not be uncharged. But
3944 * end_migration() /must/ be the one uncharging the
3945 * unused post-migration page and so it has to call
3946 * here with the migration bit still set. See the
3947 * res_counter handling below.
3949 if (!end_migration
&& PageCgroupMigration(pc
))
3952 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
3953 if (!PageAnon(page
)) { /* Shared memory */
3954 if (page
->mapping
&& !page_is_file_cache(page
))
3956 } else if (page_mapped(page
)) /* Anon */
3963 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
3965 ClearPageCgroupUsed(pc
);
3967 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3968 * freed from LRU. This is safe because uncharged page is expected not
3969 * to be reused (freed soon). Exception is SwapCache, it's handled by
3970 * special functions.
3973 unlock_page_cgroup(pc
);
3975 * even after unlock, we have memcg->res.usage here and this memcg
3976 * will never be freed, so it's safe to call css_get().
3978 memcg_check_events(memcg
, page
);
3979 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
3980 mem_cgroup_swap_statistics(memcg
, true);
3981 css_get(&memcg
->css
);
3984 * Migration does not charge the res_counter for the
3985 * replacement page, so leave it alone when phasing out the
3986 * page that is unused after the migration.
3988 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
3989 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
3994 unlock_page_cgroup(pc
);
3998 void mem_cgroup_uncharge_page(struct page
*page
)
4001 if (page_mapped(page
))
4003 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4005 * If the page is in swap cache, uncharge should be deferred
4006 * to the swap path, which also properly accounts swap usage
4007 * and handles memcg lifetime.
4009 * Note that this check is not stable and reclaim may add the
4010 * page to swap cache at any time after this. However, if the
4011 * page is not in swap cache by the time page->mapcount hits
4012 * 0, there won't be any page table references to the swap
4013 * slot, and reclaim will free it and not actually write the
4016 if (PageSwapCache(page
))
4018 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4021 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4023 VM_BUG_ON(page_mapped(page
));
4024 VM_BUG_ON(page
->mapping
);
4025 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4029 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4030 * In that cases, pages are freed continuously and we can expect pages
4031 * are in the same memcg. All these calls itself limits the number of
4032 * pages freed at once, then uncharge_start/end() is called properly.
4033 * This may be called prural(2) times in a context,
4036 void mem_cgroup_uncharge_start(void)
4038 current
->memcg_batch
.do_batch
++;
4039 /* We can do nest. */
4040 if (current
->memcg_batch
.do_batch
== 1) {
4041 current
->memcg_batch
.memcg
= NULL
;
4042 current
->memcg_batch
.nr_pages
= 0;
4043 current
->memcg_batch
.memsw_nr_pages
= 0;
4047 void mem_cgroup_uncharge_end(void)
4049 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4051 if (!batch
->do_batch
)
4055 if (batch
->do_batch
) /* If stacked, do nothing. */
4061 * This "batch->memcg" is valid without any css_get/put etc...
4062 * bacause we hide charges behind us.
4064 if (batch
->nr_pages
)
4065 res_counter_uncharge(&batch
->memcg
->res
,
4066 batch
->nr_pages
* PAGE_SIZE
);
4067 if (batch
->memsw_nr_pages
)
4068 res_counter_uncharge(&batch
->memcg
->memsw
,
4069 batch
->memsw_nr_pages
* PAGE_SIZE
);
4070 memcg_oom_recover(batch
->memcg
);
4071 /* forget this pointer (for sanity check) */
4072 batch
->memcg
= NULL
;
4077 * called after __delete_from_swap_cache() and drop "page" account.
4078 * memcg information is recorded to swap_cgroup of "ent"
4081 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4083 struct mem_cgroup
*memcg
;
4084 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4086 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4087 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4089 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4092 * record memcg information, if swapout && memcg != NULL,
4093 * css_get() was called in uncharge().
4095 if (do_swap_account
&& swapout
&& memcg
)
4096 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4100 #ifdef CONFIG_MEMCG_SWAP
4102 * called from swap_entry_free(). remove record in swap_cgroup and
4103 * uncharge "memsw" account.
4105 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4107 struct mem_cgroup
*memcg
;
4110 if (!do_swap_account
)
4113 id
= swap_cgroup_record(ent
, 0);
4115 memcg
= mem_cgroup_lookup(id
);
4118 * We uncharge this because swap is freed.
4119 * This memcg can be obsolete one. We avoid calling css_tryget
4121 if (!mem_cgroup_is_root(memcg
))
4122 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4123 mem_cgroup_swap_statistics(memcg
, false);
4124 css_put(&memcg
->css
);
4130 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4131 * @entry: swap entry to be moved
4132 * @from: mem_cgroup which the entry is moved from
4133 * @to: mem_cgroup which the entry is moved to
4135 * It succeeds only when the swap_cgroup's record for this entry is the same
4136 * as the mem_cgroup's id of @from.
4138 * Returns 0 on success, -EINVAL on failure.
4140 * The caller must have charged to @to, IOW, called res_counter_charge() about
4141 * both res and memsw, and called css_get().
4143 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4144 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4146 unsigned short old_id
, new_id
;
4148 old_id
= css_id(&from
->css
);
4149 new_id
= css_id(&to
->css
);
4151 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4152 mem_cgroup_swap_statistics(from
, false);
4153 mem_cgroup_swap_statistics(to
, true);
4155 * This function is only called from task migration context now.
4156 * It postpones res_counter and refcount handling till the end
4157 * of task migration(mem_cgroup_clear_mc()) for performance
4158 * improvement. But we cannot postpone css_get(to) because if
4159 * the process that has been moved to @to does swap-in, the
4160 * refcount of @to might be decreased to 0.
4162 * We are in attach() phase, so the cgroup is guaranteed to be
4163 * alive, so we can just call css_get().
4171 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4172 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4179 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4182 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4183 struct mem_cgroup
**memcgp
)
4185 struct mem_cgroup
*memcg
= NULL
;
4186 unsigned int nr_pages
= 1;
4187 struct page_cgroup
*pc
;
4188 enum charge_type ctype
;
4192 if (mem_cgroup_disabled())
4195 if (PageTransHuge(page
))
4196 nr_pages
<<= compound_order(page
);
4198 pc
= lookup_page_cgroup(page
);
4199 lock_page_cgroup(pc
);
4200 if (PageCgroupUsed(pc
)) {
4201 memcg
= pc
->mem_cgroup
;
4202 css_get(&memcg
->css
);
4204 * At migrating an anonymous page, its mapcount goes down
4205 * to 0 and uncharge() will be called. But, even if it's fully
4206 * unmapped, migration may fail and this page has to be
4207 * charged again. We set MIGRATION flag here and delay uncharge
4208 * until end_migration() is called
4210 * Corner Case Thinking
4212 * When the old page was mapped as Anon and it's unmap-and-freed
4213 * while migration was ongoing.
4214 * If unmap finds the old page, uncharge() of it will be delayed
4215 * until end_migration(). If unmap finds a new page, it's
4216 * uncharged when it make mapcount to be 1->0. If unmap code
4217 * finds swap_migration_entry, the new page will not be mapped
4218 * and end_migration() will find it(mapcount==0).
4221 * When the old page was mapped but migraion fails, the kernel
4222 * remaps it. A charge for it is kept by MIGRATION flag even
4223 * if mapcount goes down to 0. We can do remap successfully
4224 * without charging it again.
4227 * The "old" page is under lock_page() until the end of
4228 * migration, so, the old page itself will not be swapped-out.
4229 * If the new page is swapped out before end_migraton, our
4230 * hook to usual swap-out path will catch the event.
4233 SetPageCgroupMigration(pc
);
4235 unlock_page_cgroup(pc
);
4237 * If the page is not charged at this point,
4245 * We charge new page before it's used/mapped. So, even if unlock_page()
4246 * is called before end_migration, we can catch all events on this new
4247 * page. In the case new page is migrated but not remapped, new page's
4248 * mapcount will be finally 0 and we call uncharge in end_migration().
4251 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4253 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4255 * The page is committed to the memcg, but it's not actually
4256 * charged to the res_counter since we plan on replacing the
4257 * old one and only one page is going to be left afterwards.
4259 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4262 /* remove redundant charge if migration failed*/
4263 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4264 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4266 struct page
*used
, *unused
;
4267 struct page_cgroup
*pc
;
4273 if (!migration_ok
) {
4280 anon
= PageAnon(used
);
4281 __mem_cgroup_uncharge_common(unused
,
4282 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4283 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4285 css_put(&memcg
->css
);
4287 * We disallowed uncharge of pages under migration because mapcount
4288 * of the page goes down to zero, temporarly.
4289 * Clear the flag and check the page should be charged.
4291 pc
= lookup_page_cgroup(oldpage
);
4292 lock_page_cgroup(pc
);
4293 ClearPageCgroupMigration(pc
);
4294 unlock_page_cgroup(pc
);
4297 * If a page is a file cache, radix-tree replacement is very atomic
4298 * and we can skip this check. When it was an Anon page, its mapcount
4299 * goes down to 0. But because we added MIGRATION flage, it's not
4300 * uncharged yet. There are several case but page->mapcount check
4301 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4302 * check. (see prepare_charge() also)
4305 mem_cgroup_uncharge_page(used
);
4309 * At replace page cache, newpage is not under any memcg but it's on
4310 * LRU. So, this function doesn't touch res_counter but handles LRU
4311 * in correct way. Both pages are locked so we cannot race with uncharge.
4313 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4314 struct page
*newpage
)
4316 struct mem_cgroup
*memcg
= NULL
;
4317 struct page_cgroup
*pc
;
4318 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4320 if (mem_cgroup_disabled())
4323 pc
= lookup_page_cgroup(oldpage
);
4324 /* fix accounting on old pages */
4325 lock_page_cgroup(pc
);
4326 if (PageCgroupUsed(pc
)) {
4327 memcg
= pc
->mem_cgroup
;
4328 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4329 ClearPageCgroupUsed(pc
);
4331 unlock_page_cgroup(pc
);
4334 * When called from shmem_replace_page(), in some cases the
4335 * oldpage has already been charged, and in some cases not.
4340 * Even if newpage->mapping was NULL before starting replacement,
4341 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4342 * LRU while we overwrite pc->mem_cgroup.
4344 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4347 #ifdef CONFIG_DEBUG_VM
4348 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4350 struct page_cgroup
*pc
;
4352 pc
= lookup_page_cgroup(page
);
4354 * Can be NULL while feeding pages into the page allocator for
4355 * the first time, i.e. during boot or memory hotplug;
4356 * or when mem_cgroup_disabled().
4358 if (likely(pc
) && PageCgroupUsed(pc
))
4363 bool mem_cgroup_bad_page_check(struct page
*page
)
4365 if (mem_cgroup_disabled())
4368 return lookup_page_cgroup_used(page
) != NULL
;
4371 void mem_cgroup_print_bad_page(struct page
*page
)
4373 struct page_cgroup
*pc
;
4375 pc
= lookup_page_cgroup_used(page
);
4377 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4378 pc
, pc
->flags
, pc
->mem_cgroup
);
4383 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4384 unsigned long long val
)
4387 u64 memswlimit
, memlimit
;
4389 int children
= mem_cgroup_count_children(memcg
);
4390 u64 curusage
, oldusage
;
4394 * For keeping hierarchical_reclaim simple, how long we should retry
4395 * is depends on callers. We set our retry-count to be function
4396 * of # of children which we should visit in this loop.
4398 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4400 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4403 while (retry_count
) {
4404 if (signal_pending(current
)) {
4409 * Rather than hide all in some function, I do this in
4410 * open coded manner. You see what this really does.
4411 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4413 mutex_lock(&set_limit_mutex
);
4414 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4415 if (memswlimit
< val
) {
4417 mutex_unlock(&set_limit_mutex
);
4421 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4425 ret
= res_counter_set_limit(&memcg
->res
, val
);
4427 if (memswlimit
== val
)
4428 memcg
->memsw_is_minimum
= true;
4430 memcg
->memsw_is_minimum
= false;
4432 mutex_unlock(&set_limit_mutex
);
4437 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4438 MEM_CGROUP_RECLAIM_SHRINK
);
4439 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4440 /* Usage is reduced ? */
4441 if (curusage
>= oldusage
)
4444 oldusage
= curusage
;
4446 if (!ret
&& enlarge
)
4447 memcg_oom_recover(memcg
);
4452 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4453 unsigned long long val
)
4456 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4457 int children
= mem_cgroup_count_children(memcg
);
4461 /* see mem_cgroup_resize_res_limit */
4462 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4463 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4464 while (retry_count
) {
4465 if (signal_pending(current
)) {
4470 * Rather than hide all in some function, I do this in
4471 * open coded manner. You see what this really does.
4472 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4474 mutex_lock(&set_limit_mutex
);
4475 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4476 if (memlimit
> val
) {
4478 mutex_unlock(&set_limit_mutex
);
4481 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4482 if (memswlimit
< val
)
4484 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4486 if (memlimit
== val
)
4487 memcg
->memsw_is_minimum
= true;
4489 memcg
->memsw_is_minimum
= false;
4491 mutex_unlock(&set_limit_mutex
);
4496 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4497 MEM_CGROUP_RECLAIM_NOSWAP
|
4498 MEM_CGROUP_RECLAIM_SHRINK
);
4499 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4500 /* Usage is reduced ? */
4501 if (curusage
>= oldusage
)
4504 oldusage
= curusage
;
4506 if (!ret
&& enlarge
)
4507 memcg_oom_recover(memcg
);
4512 * mem_cgroup_force_empty_list - clears LRU of a group
4513 * @memcg: group to clear
4516 * @lru: lru to to clear
4518 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4519 * reclaim the pages page themselves - pages are moved to the parent (or root)
4522 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4523 int node
, int zid
, enum lru_list lru
)
4525 struct lruvec
*lruvec
;
4526 unsigned long flags
;
4527 struct list_head
*list
;
4531 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4532 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4533 list
= &lruvec
->lists
[lru
];
4537 struct page_cgroup
*pc
;
4540 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4541 if (list_empty(list
)) {
4542 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4545 page
= list_entry(list
->prev
, struct page
, lru
);
4547 list_move(&page
->lru
, list
);
4549 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4552 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4554 pc
= lookup_page_cgroup(page
);
4556 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4557 /* found lock contention or "pc" is obsolete. */
4562 } while (!list_empty(list
));
4566 * make mem_cgroup's charge to be 0 if there is no task by moving
4567 * all the charges and pages to the parent.
4568 * This enables deleting this mem_cgroup.
4570 * Caller is responsible for holding css reference on the memcg.
4572 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4578 /* This is for making all *used* pages to be on LRU. */
4579 lru_add_drain_all();
4580 drain_all_stock_sync(memcg
);
4581 mem_cgroup_start_move(memcg
);
4582 for_each_node_state(node
, N_MEMORY
) {
4583 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4586 mem_cgroup_force_empty_list(memcg
,
4591 mem_cgroup_end_move(memcg
);
4592 memcg_oom_recover(memcg
);
4596 * Kernel memory may not necessarily be trackable to a specific
4597 * process. So they are not migrated, and therefore we can't
4598 * expect their value to drop to 0 here.
4599 * Having res filled up with kmem only is enough.
4601 * This is a safety check because mem_cgroup_force_empty_list
4602 * could have raced with mem_cgroup_replace_page_cache callers
4603 * so the lru seemed empty but the page could have been added
4604 * right after the check. RES_USAGE should be safe as we always
4605 * charge before adding to the LRU.
4607 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4608 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4609 } while (usage
> 0);
4613 * This mainly exists for tests during the setting of set of use_hierarchy.
4614 * Since this is the very setting we are changing, the current hierarchy value
4617 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4619 struct cgroup_subsys_state
*pos
;
4621 /* bounce at first found */
4622 css_for_each_child(pos
, &memcg
->css
)
4628 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4629 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4630 * from mem_cgroup_count_children(), in the sense that we don't really care how
4631 * many children we have; we only need to know if we have any. It also counts
4632 * any memcg without hierarchy as infertile.
4634 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4636 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4640 * Reclaims as many pages from the given memcg as possible and moves
4641 * the rest to the parent.
4643 * Caller is responsible for holding css reference for memcg.
4645 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4647 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4648 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4650 /* returns EBUSY if there is a task or if we come here twice. */
4651 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4654 /* we call try-to-free pages for make this cgroup empty */
4655 lru_add_drain_all();
4656 /* try to free all pages in this cgroup */
4657 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4660 if (signal_pending(current
))
4663 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4667 /* maybe some writeback is necessary */
4668 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4673 mem_cgroup_reparent_charges(memcg
);
4678 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4681 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4683 if (mem_cgroup_is_root(memcg
))
4685 return mem_cgroup_force_empty(memcg
);
4688 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4691 return mem_cgroup_from_css(css
)->use_hierarchy
;
4694 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4695 struct cftype
*cft
, u64 val
)
4698 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4699 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4701 mutex_lock(&memcg_create_mutex
);
4703 if (memcg
->use_hierarchy
== val
)
4707 * If parent's use_hierarchy is set, we can't make any modifications
4708 * in the child subtrees. If it is unset, then the change can
4709 * occur, provided the current cgroup has no children.
4711 * For the root cgroup, parent_mem is NULL, we allow value to be
4712 * set if there are no children.
4714 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4715 (val
== 1 || val
== 0)) {
4716 if (!__memcg_has_children(memcg
))
4717 memcg
->use_hierarchy
= val
;
4724 mutex_unlock(&memcg_create_mutex
);
4730 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4731 enum mem_cgroup_stat_index idx
)
4733 struct mem_cgroup
*iter
;
4736 /* Per-cpu values can be negative, use a signed accumulator */
4737 for_each_mem_cgroup_tree(iter
, memcg
)
4738 val
+= mem_cgroup_read_stat(iter
, idx
);
4740 if (val
< 0) /* race ? */
4745 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4749 if (!mem_cgroup_is_root(memcg
)) {
4751 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4753 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4757 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4758 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4760 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4761 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4764 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4766 return val
<< PAGE_SHIFT
;
4769 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
4770 struct cftype
*cft
, struct file
*file
,
4771 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
4773 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4779 type
= MEMFILE_TYPE(cft
->private);
4780 name
= MEMFILE_ATTR(cft
->private);
4784 if (name
== RES_USAGE
)
4785 val
= mem_cgroup_usage(memcg
, false);
4787 val
= res_counter_read_u64(&memcg
->res
, name
);
4790 if (name
== RES_USAGE
)
4791 val
= mem_cgroup_usage(memcg
, true);
4793 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4796 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4802 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4803 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4806 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
4809 #ifdef CONFIG_MEMCG_KMEM
4810 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4812 * For simplicity, we won't allow this to be disabled. It also can't
4813 * be changed if the cgroup has children already, or if tasks had
4816 * If tasks join before we set the limit, a person looking at
4817 * kmem.usage_in_bytes will have no way to determine when it took
4818 * place, which makes the value quite meaningless.
4820 * After it first became limited, changes in the value of the limit are
4821 * of course permitted.
4823 mutex_lock(&memcg_create_mutex
);
4824 mutex_lock(&set_limit_mutex
);
4825 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4826 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
4830 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4833 ret
= memcg_update_cache_sizes(memcg
);
4835 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4838 static_key_slow_inc(&memcg_kmem_enabled_key
);
4840 * setting the active bit after the inc will guarantee no one
4841 * starts accounting before all call sites are patched
4843 memcg_kmem_set_active(memcg
);
4845 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4847 mutex_unlock(&set_limit_mutex
);
4848 mutex_unlock(&memcg_create_mutex
);
4853 #ifdef CONFIG_MEMCG_KMEM
4854 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4857 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4861 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
4863 * When that happen, we need to disable the static branch only on those
4864 * memcgs that enabled it. To achieve this, we would be forced to
4865 * complicate the code by keeping track of which memcgs were the ones
4866 * that actually enabled limits, and which ones got it from its
4869 * It is a lot simpler just to do static_key_slow_inc() on every child
4870 * that is accounted.
4872 if (!memcg_kmem_is_active(memcg
))
4876 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4877 * memcg is active already. If the later initialization fails then the
4878 * cgroup core triggers the cleanup so we do not have to do it here.
4880 static_key_slow_inc(&memcg_kmem_enabled_key
);
4882 mutex_lock(&set_limit_mutex
);
4883 memcg_stop_kmem_account();
4884 ret
= memcg_update_cache_sizes(memcg
);
4885 memcg_resume_kmem_account();
4886 mutex_unlock(&set_limit_mutex
);
4890 #endif /* CONFIG_MEMCG_KMEM */
4893 * The user of this function is...
4896 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
4899 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4902 unsigned long long val
;
4905 type
= MEMFILE_TYPE(cft
->private);
4906 name
= MEMFILE_ATTR(cft
->private);
4910 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4914 /* This function does all necessary parse...reuse it */
4915 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
4919 ret
= mem_cgroup_resize_limit(memcg
, val
);
4920 else if (type
== _MEMSWAP
)
4921 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
4922 else if (type
== _KMEM
)
4923 ret
= memcg_update_kmem_limit(css
, val
);
4927 case RES_SOFT_LIMIT
:
4928 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
4932 * For memsw, soft limits are hard to implement in terms
4933 * of semantics, for now, we support soft limits for
4934 * control without swap
4937 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
4942 ret
= -EINVAL
; /* should be BUG() ? */
4948 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
4949 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
4951 unsigned long long min_limit
, min_memsw_limit
, tmp
;
4953 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4954 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4955 if (!memcg
->use_hierarchy
)
4958 while (css_parent(&memcg
->css
)) {
4959 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4960 if (!memcg
->use_hierarchy
)
4962 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4963 min_limit
= min(min_limit
, tmp
);
4964 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4965 min_memsw_limit
= min(min_memsw_limit
, tmp
);
4968 *mem_limit
= min_limit
;
4969 *memsw_limit
= min_memsw_limit
;
4972 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
4974 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4978 type
= MEMFILE_TYPE(event
);
4979 name
= MEMFILE_ATTR(event
);
4984 res_counter_reset_max(&memcg
->res
);
4985 else if (type
== _MEMSWAP
)
4986 res_counter_reset_max(&memcg
->memsw
);
4987 else if (type
== _KMEM
)
4988 res_counter_reset_max(&memcg
->kmem
);
4994 res_counter_reset_failcnt(&memcg
->res
);
4995 else if (type
== _MEMSWAP
)
4996 res_counter_reset_failcnt(&memcg
->memsw
);
4997 else if (type
== _KMEM
)
4998 res_counter_reset_failcnt(&memcg
->kmem
);
5007 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5010 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5014 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5015 struct cftype
*cft
, u64 val
)
5017 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5019 if (val
>= (1 << NR_MOVE_TYPE
))
5023 * No kind of locking is needed in here, because ->can_attach() will
5024 * check this value once in the beginning of the process, and then carry
5025 * on with stale data. This means that changes to this value will only
5026 * affect task migrations starting after the change.
5028 memcg
->move_charge_at_immigrate
= val
;
5032 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5033 struct cftype
*cft
, u64 val
)
5040 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5041 struct cftype
*cft
, struct seq_file
*m
)
5044 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5045 unsigned long node_nr
;
5046 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5048 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5049 seq_printf(m
, "total=%lu", total_nr
);
5050 for_each_node_state(nid
, N_MEMORY
) {
5051 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5052 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5056 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5057 seq_printf(m
, "file=%lu", file_nr
);
5058 for_each_node_state(nid
, N_MEMORY
) {
5059 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5061 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5065 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5066 seq_printf(m
, "anon=%lu", anon_nr
);
5067 for_each_node_state(nid
, N_MEMORY
) {
5068 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5070 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5074 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5075 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5076 for_each_node_state(nid
, N_MEMORY
) {
5077 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5078 BIT(LRU_UNEVICTABLE
));
5079 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5084 #endif /* CONFIG_NUMA */
5086 static inline void mem_cgroup_lru_names_not_uptodate(void)
5088 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5091 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5094 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5095 struct mem_cgroup
*mi
;
5098 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5099 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5101 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5102 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5105 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5106 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5107 mem_cgroup_read_events(memcg
, i
));
5109 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5110 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5111 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5113 /* Hierarchical information */
5115 unsigned long long limit
, memsw_limit
;
5116 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5117 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5118 if (do_swap_account
)
5119 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5123 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5126 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5128 for_each_mem_cgroup_tree(mi
, memcg
)
5129 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5130 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5133 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5134 unsigned long long val
= 0;
5136 for_each_mem_cgroup_tree(mi
, memcg
)
5137 val
+= mem_cgroup_read_events(mi
, i
);
5138 seq_printf(m
, "total_%s %llu\n",
5139 mem_cgroup_events_names
[i
], val
);
5142 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5143 unsigned long long val
= 0;
5145 for_each_mem_cgroup_tree(mi
, memcg
)
5146 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5147 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5150 #ifdef CONFIG_DEBUG_VM
5153 struct mem_cgroup_per_zone
*mz
;
5154 struct zone_reclaim_stat
*rstat
;
5155 unsigned long recent_rotated
[2] = {0, 0};
5156 unsigned long recent_scanned
[2] = {0, 0};
5158 for_each_online_node(nid
)
5159 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5160 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5161 rstat
= &mz
->lruvec
.reclaim_stat
;
5163 recent_rotated
[0] += rstat
->recent_rotated
[0];
5164 recent_rotated
[1] += rstat
->recent_rotated
[1];
5165 recent_scanned
[0] += rstat
->recent_scanned
[0];
5166 recent_scanned
[1] += rstat
->recent_scanned
[1];
5168 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5169 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5170 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5171 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5178 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5181 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5183 return mem_cgroup_swappiness(memcg
);
5186 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5187 struct cftype
*cft
, u64 val
)
5189 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5190 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5192 if (val
> 100 || !parent
)
5195 mutex_lock(&memcg_create_mutex
);
5197 /* If under hierarchy, only empty-root can set this value */
5198 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5199 mutex_unlock(&memcg_create_mutex
);
5203 memcg
->swappiness
= val
;
5205 mutex_unlock(&memcg_create_mutex
);
5210 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5212 struct mem_cgroup_threshold_ary
*t
;
5218 t
= rcu_dereference(memcg
->thresholds
.primary
);
5220 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5225 usage
= mem_cgroup_usage(memcg
, swap
);
5228 * current_threshold points to threshold just below or equal to usage.
5229 * If it's not true, a threshold was crossed after last
5230 * call of __mem_cgroup_threshold().
5232 i
= t
->current_threshold
;
5235 * Iterate backward over array of thresholds starting from
5236 * current_threshold and check if a threshold is crossed.
5237 * If none of thresholds below usage is crossed, we read
5238 * only one element of the array here.
5240 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5241 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5243 /* i = current_threshold + 1 */
5247 * Iterate forward over array of thresholds starting from
5248 * current_threshold+1 and check if a threshold is crossed.
5249 * If none of thresholds above usage is crossed, we read
5250 * only one element of the array here.
5252 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5253 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5255 /* Update current_threshold */
5256 t
->current_threshold
= i
- 1;
5261 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5264 __mem_cgroup_threshold(memcg
, false);
5265 if (do_swap_account
)
5266 __mem_cgroup_threshold(memcg
, true);
5268 memcg
= parent_mem_cgroup(memcg
);
5272 static int compare_thresholds(const void *a
, const void *b
)
5274 const struct mem_cgroup_threshold
*_a
= a
;
5275 const struct mem_cgroup_threshold
*_b
= b
;
5277 if (_a
->threshold
> _b
->threshold
)
5280 if (_a
->threshold
< _b
->threshold
)
5286 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5288 struct mem_cgroup_eventfd_list
*ev
;
5290 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5291 eventfd_signal(ev
->eventfd
, 1);
5295 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5297 struct mem_cgroup
*iter
;
5299 for_each_mem_cgroup_tree(iter
, memcg
)
5300 mem_cgroup_oom_notify_cb(iter
);
5303 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5304 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5306 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5307 struct mem_cgroup_thresholds
*thresholds
;
5308 struct mem_cgroup_threshold_ary
*new;
5309 enum res_type type
= MEMFILE_TYPE(cft
->private);
5310 u64 threshold
, usage
;
5313 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5317 mutex_lock(&memcg
->thresholds_lock
);
5320 thresholds
= &memcg
->thresholds
;
5321 else if (type
== _MEMSWAP
)
5322 thresholds
= &memcg
->memsw_thresholds
;
5326 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5328 /* Check if a threshold crossed before adding a new one */
5329 if (thresholds
->primary
)
5330 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5332 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5334 /* Allocate memory for new array of thresholds */
5335 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5343 /* Copy thresholds (if any) to new array */
5344 if (thresholds
->primary
) {
5345 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5346 sizeof(struct mem_cgroup_threshold
));
5349 /* Add new threshold */
5350 new->entries
[size
- 1].eventfd
= eventfd
;
5351 new->entries
[size
- 1].threshold
= threshold
;
5353 /* Sort thresholds. Registering of new threshold isn't time-critical */
5354 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5355 compare_thresholds
, NULL
);
5357 /* Find current threshold */
5358 new->current_threshold
= -1;
5359 for (i
= 0; i
< size
; i
++) {
5360 if (new->entries
[i
].threshold
<= usage
) {
5362 * new->current_threshold will not be used until
5363 * rcu_assign_pointer(), so it's safe to increment
5366 ++new->current_threshold
;
5371 /* Free old spare buffer and save old primary buffer as spare */
5372 kfree(thresholds
->spare
);
5373 thresholds
->spare
= thresholds
->primary
;
5375 rcu_assign_pointer(thresholds
->primary
, new);
5377 /* To be sure that nobody uses thresholds */
5381 mutex_unlock(&memcg
->thresholds_lock
);
5386 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5387 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5389 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5390 struct mem_cgroup_thresholds
*thresholds
;
5391 struct mem_cgroup_threshold_ary
*new;
5392 enum res_type type
= MEMFILE_TYPE(cft
->private);
5396 mutex_lock(&memcg
->thresholds_lock
);
5398 thresholds
= &memcg
->thresholds
;
5399 else if (type
== _MEMSWAP
)
5400 thresholds
= &memcg
->memsw_thresholds
;
5404 if (!thresholds
->primary
)
5407 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5409 /* Check if a threshold crossed before removing */
5410 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5412 /* Calculate new number of threshold */
5414 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5415 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5419 new = thresholds
->spare
;
5421 /* Set thresholds array to NULL if we don't have thresholds */
5430 /* Copy thresholds and find current threshold */
5431 new->current_threshold
= -1;
5432 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5433 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5436 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5437 if (new->entries
[j
].threshold
<= usage
) {
5439 * new->current_threshold will not be used
5440 * until rcu_assign_pointer(), so it's safe to increment
5443 ++new->current_threshold
;
5449 /* Swap primary and spare array */
5450 thresholds
->spare
= thresholds
->primary
;
5451 /* If all events are unregistered, free the spare array */
5453 kfree(thresholds
->spare
);
5454 thresholds
->spare
= NULL
;
5457 rcu_assign_pointer(thresholds
->primary
, new);
5459 /* To be sure that nobody uses thresholds */
5462 mutex_unlock(&memcg
->thresholds_lock
);
5465 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5466 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5468 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5469 struct mem_cgroup_eventfd_list
*event
;
5470 enum res_type type
= MEMFILE_TYPE(cft
->private);
5472 BUG_ON(type
!= _OOM_TYPE
);
5473 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5477 spin_lock(&memcg_oom_lock
);
5479 event
->eventfd
= eventfd
;
5480 list_add(&event
->list
, &memcg
->oom_notify
);
5482 /* already in OOM ? */
5483 if (atomic_read(&memcg
->under_oom
))
5484 eventfd_signal(eventfd
, 1);
5485 spin_unlock(&memcg_oom_lock
);
5490 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5491 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5493 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5494 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5495 enum res_type type
= MEMFILE_TYPE(cft
->private);
5497 BUG_ON(type
!= _OOM_TYPE
);
5499 spin_lock(&memcg_oom_lock
);
5501 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5502 if (ev
->eventfd
== eventfd
) {
5503 list_del(&ev
->list
);
5508 spin_unlock(&memcg_oom_lock
);
5511 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5512 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5514 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5516 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5518 if (atomic_read(&memcg
->under_oom
))
5519 cb
->fill(cb
, "under_oom", 1);
5521 cb
->fill(cb
, "under_oom", 0);
5525 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5526 struct cftype
*cft
, u64 val
)
5528 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5529 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5531 /* cannot set to root cgroup and only 0 and 1 are allowed */
5532 if (!parent
|| !((val
== 0) || (val
== 1)))
5535 mutex_lock(&memcg_create_mutex
);
5536 /* oom-kill-disable is a flag for subhierarchy. */
5537 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5538 mutex_unlock(&memcg_create_mutex
);
5541 memcg
->oom_kill_disable
= val
;
5543 memcg_oom_recover(memcg
);
5544 mutex_unlock(&memcg_create_mutex
);
5548 #ifdef CONFIG_MEMCG_KMEM
5549 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5553 memcg
->kmemcg_id
= -1;
5554 ret
= memcg_propagate_kmem(memcg
);
5558 return mem_cgroup_sockets_init(memcg
, ss
);
5561 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5563 mem_cgroup_sockets_destroy(memcg
);
5566 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5568 if (!memcg_kmem_is_active(memcg
))
5572 * kmem charges can outlive the cgroup. In the case of slab
5573 * pages, for instance, a page contain objects from various
5574 * processes. As we prevent from taking a reference for every
5575 * such allocation we have to be careful when doing uncharge
5576 * (see memcg_uncharge_kmem) and here during offlining.
5578 * The idea is that that only the _last_ uncharge which sees
5579 * the dead memcg will drop the last reference. An additional
5580 * reference is taken here before the group is marked dead
5581 * which is then paired with css_put during uncharge resp. here.
5583 * Although this might sound strange as this path is called from
5584 * css_offline() when the referencemight have dropped down to 0
5585 * and shouldn't be incremented anymore (css_tryget would fail)
5586 * we do not have other options because of the kmem allocations
5589 css_get(&memcg
->css
);
5591 memcg_kmem_mark_dead(memcg
);
5593 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5596 if (memcg_kmem_test_and_clear_dead(memcg
))
5597 css_put(&memcg
->css
);
5600 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5605 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5609 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5614 static struct cftype mem_cgroup_files
[] = {
5616 .name
= "usage_in_bytes",
5617 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5618 .read
= mem_cgroup_read
,
5619 .register_event
= mem_cgroup_usage_register_event
,
5620 .unregister_event
= mem_cgroup_usage_unregister_event
,
5623 .name
= "max_usage_in_bytes",
5624 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5625 .trigger
= mem_cgroup_reset
,
5626 .read
= mem_cgroup_read
,
5629 .name
= "limit_in_bytes",
5630 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5631 .write_string
= mem_cgroup_write
,
5632 .read
= mem_cgroup_read
,
5635 .name
= "soft_limit_in_bytes",
5636 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5637 .write_string
= mem_cgroup_write
,
5638 .read
= mem_cgroup_read
,
5642 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5643 .trigger
= mem_cgroup_reset
,
5644 .read
= mem_cgroup_read
,
5648 .read_seq_string
= memcg_stat_show
,
5651 .name
= "force_empty",
5652 .trigger
= mem_cgroup_force_empty_write
,
5655 .name
= "use_hierarchy",
5656 .flags
= CFTYPE_INSANE
,
5657 .write_u64
= mem_cgroup_hierarchy_write
,
5658 .read_u64
= mem_cgroup_hierarchy_read
,
5661 .name
= "swappiness",
5662 .read_u64
= mem_cgroup_swappiness_read
,
5663 .write_u64
= mem_cgroup_swappiness_write
,
5666 .name
= "move_charge_at_immigrate",
5667 .read_u64
= mem_cgroup_move_charge_read
,
5668 .write_u64
= mem_cgroup_move_charge_write
,
5671 .name
= "oom_control",
5672 .read_map
= mem_cgroup_oom_control_read
,
5673 .write_u64
= mem_cgroup_oom_control_write
,
5674 .register_event
= mem_cgroup_oom_register_event
,
5675 .unregister_event
= mem_cgroup_oom_unregister_event
,
5676 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5679 .name
= "pressure_level",
5680 .register_event
= vmpressure_register_event
,
5681 .unregister_event
= vmpressure_unregister_event
,
5685 .name
= "numa_stat",
5686 .read_seq_string
= memcg_numa_stat_show
,
5689 #ifdef CONFIG_MEMCG_KMEM
5691 .name
= "kmem.limit_in_bytes",
5692 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5693 .write_string
= mem_cgroup_write
,
5694 .read
= mem_cgroup_read
,
5697 .name
= "kmem.usage_in_bytes",
5698 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5699 .read
= mem_cgroup_read
,
5702 .name
= "kmem.failcnt",
5703 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5704 .trigger
= mem_cgroup_reset
,
5705 .read
= mem_cgroup_read
,
5708 .name
= "kmem.max_usage_in_bytes",
5709 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5710 .trigger
= mem_cgroup_reset
,
5711 .read
= mem_cgroup_read
,
5713 #ifdef CONFIG_SLABINFO
5715 .name
= "kmem.slabinfo",
5716 .read_seq_string
= mem_cgroup_slabinfo_read
,
5720 { }, /* terminate */
5723 #ifdef CONFIG_MEMCG_SWAP
5724 static struct cftype memsw_cgroup_files
[] = {
5726 .name
= "memsw.usage_in_bytes",
5727 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5728 .read
= mem_cgroup_read
,
5729 .register_event
= mem_cgroup_usage_register_event
,
5730 .unregister_event
= mem_cgroup_usage_unregister_event
,
5733 .name
= "memsw.max_usage_in_bytes",
5734 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5735 .trigger
= mem_cgroup_reset
,
5736 .read
= mem_cgroup_read
,
5739 .name
= "memsw.limit_in_bytes",
5740 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5741 .write_string
= mem_cgroup_write
,
5742 .read
= mem_cgroup_read
,
5745 .name
= "memsw.failcnt",
5746 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5747 .trigger
= mem_cgroup_reset
,
5748 .read
= mem_cgroup_read
,
5750 { }, /* terminate */
5753 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5755 struct mem_cgroup_per_node
*pn
;
5756 struct mem_cgroup_per_zone
*mz
;
5757 int zone
, tmp
= node
;
5759 * This routine is called against possible nodes.
5760 * But it's BUG to call kmalloc() against offline node.
5762 * TODO: this routine can waste much memory for nodes which will
5763 * never be onlined. It's better to use memory hotplug callback
5766 if (!node_state(node
, N_NORMAL_MEMORY
))
5768 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5772 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5773 mz
= &pn
->zoneinfo
[zone
];
5774 lruvec_init(&mz
->lruvec
);
5777 memcg
->nodeinfo
[node
] = pn
;
5781 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5783 kfree(memcg
->nodeinfo
[node
]);
5786 static struct mem_cgroup
*mem_cgroup_alloc(void)
5788 struct mem_cgroup
*memcg
;
5789 size_t size
= memcg_size();
5791 /* Can be very big if nr_node_ids is very big */
5792 if (size
< PAGE_SIZE
)
5793 memcg
= kzalloc(size
, GFP_KERNEL
);
5795 memcg
= vzalloc(size
);
5800 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5803 spin_lock_init(&memcg
->pcp_counter_lock
);
5807 if (size
< PAGE_SIZE
)
5815 * At destroying mem_cgroup, references from swap_cgroup can remain.
5816 * (scanning all at force_empty is too costly...)
5818 * Instead of clearing all references at force_empty, we remember
5819 * the number of reference from swap_cgroup and free mem_cgroup when
5820 * it goes down to 0.
5822 * Removal of cgroup itself succeeds regardless of refs from swap.
5825 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5828 size_t size
= memcg_size();
5830 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5833 free_mem_cgroup_per_zone_info(memcg
, node
);
5835 free_percpu(memcg
->stat
);
5838 * We need to make sure that (at least for now), the jump label
5839 * destruction code runs outside of the cgroup lock. This is because
5840 * get_online_cpus(), which is called from the static_branch update,
5841 * can't be called inside the cgroup_lock. cpusets are the ones
5842 * enforcing this dependency, so if they ever change, we might as well.
5844 * schedule_work() will guarantee this happens. Be careful if you need
5845 * to move this code around, and make sure it is outside
5848 disarm_static_keys(memcg
);
5849 if (size
< PAGE_SIZE
)
5856 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5858 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
5860 if (!memcg
->res
.parent
)
5862 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
5864 EXPORT_SYMBOL(parent_mem_cgroup
);
5866 static struct cgroup_subsys_state
* __ref
5867 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5869 struct mem_cgroup
*memcg
;
5870 long error
= -ENOMEM
;
5873 memcg
= mem_cgroup_alloc();
5875 return ERR_PTR(error
);
5878 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
5882 if (parent_css
== NULL
) {
5883 root_mem_cgroup
= memcg
;
5884 res_counter_init(&memcg
->res
, NULL
);
5885 res_counter_init(&memcg
->memsw
, NULL
);
5886 res_counter_init(&memcg
->kmem
, NULL
);
5889 memcg
->last_scanned_node
= MAX_NUMNODES
;
5890 INIT_LIST_HEAD(&memcg
->oom_notify
);
5891 memcg
->move_charge_at_immigrate
= 0;
5892 mutex_init(&memcg
->thresholds_lock
);
5893 spin_lock_init(&memcg
->move_lock
);
5894 vmpressure_init(&memcg
->vmpressure
);
5899 __mem_cgroup_free(memcg
);
5900 return ERR_PTR(error
);
5904 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5906 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5907 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
5913 mutex_lock(&memcg_create_mutex
);
5915 memcg
->use_hierarchy
= parent
->use_hierarchy
;
5916 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
5917 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
5919 if (parent
->use_hierarchy
) {
5920 res_counter_init(&memcg
->res
, &parent
->res
);
5921 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
5922 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
5925 * No need to take a reference to the parent because cgroup
5926 * core guarantees its existence.
5929 res_counter_init(&memcg
->res
, NULL
);
5930 res_counter_init(&memcg
->memsw
, NULL
);
5931 res_counter_init(&memcg
->kmem
, NULL
);
5933 * Deeper hierachy with use_hierarchy == false doesn't make
5934 * much sense so let cgroup subsystem know about this
5935 * unfortunate state in our controller.
5937 if (parent
!= root_mem_cgroup
)
5938 mem_cgroup_subsys
.broken_hierarchy
= true;
5941 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
5942 mutex_unlock(&memcg_create_mutex
);
5947 * Announce all parents that a group from their hierarchy is gone.
5949 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
5951 struct mem_cgroup
*parent
= memcg
;
5953 while ((parent
= parent_mem_cgroup(parent
)))
5954 mem_cgroup_iter_invalidate(parent
);
5957 * if the root memcg is not hierarchical we have to check it
5960 if (!root_mem_cgroup
->use_hierarchy
)
5961 mem_cgroup_iter_invalidate(root_mem_cgroup
);
5964 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
5966 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5968 kmem_cgroup_css_offline(memcg
);
5970 mem_cgroup_invalidate_reclaim_iterators(memcg
);
5971 mem_cgroup_reparent_charges(memcg
);
5972 mem_cgroup_destroy_all_caches(memcg
);
5973 vmpressure_cleanup(&memcg
->vmpressure
);
5976 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
5978 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5980 memcg_destroy_kmem(memcg
);
5981 __mem_cgroup_free(memcg
);
5985 /* Handlers for move charge at task migration. */
5986 #define PRECHARGE_COUNT_AT_ONCE 256
5987 static int mem_cgroup_do_precharge(unsigned long count
)
5990 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
5991 struct mem_cgroup
*memcg
= mc
.to
;
5993 if (mem_cgroup_is_root(memcg
)) {
5994 mc
.precharge
+= count
;
5995 /* we don't need css_get for root */
5998 /* try to charge at once */
6000 struct res_counter
*dummy
;
6002 * "memcg" cannot be under rmdir() because we've already checked
6003 * by cgroup_lock_live_cgroup() that it is not removed and we
6004 * are still under the same cgroup_mutex. So we can postpone
6007 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6009 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6010 PAGE_SIZE
* count
, &dummy
)) {
6011 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6014 mc
.precharge
+= count
;
6018 /* fall back to one by one charge */
6020 if (signal_pending(current
)) {
6024 if (!batch_count
--) {
6025 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6028 ret
= __mem_cgroup_try_charge(NULL
,
6029 GFP_KERNEL
, 1, &memcg
, false);
6031 /* mem_cgroup_clear_mc() will do uncharge later */
6039 * get_mctgt_type - get target type of moving charge
6040 * @vma: the vma the pte to be checked belongs
6041 * @addr: the address corresponding to the pte to be checked
6042 * @ptent: the pte to be checked
6043 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6046 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6047 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6048 * move charge. if @target is not NULL, the page is stored in target->page
6049 * with extra refcnt got(Callers should handle it).
6050 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6051 * target for charge migration. if @target is not NULL, the entry is stored
6054 * Called with pte lock held.
6061 enum mc_target_type
{
6067 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6068 unsigned long addr
, pte_t ptent
)
6070 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6072 if (!page
|| !page_mapped(page
))
6074 if (PageAnon(page
)) {
6075 /* we don't move shared anon */
6078 } else if (!move_file())
6079 /* we ignore mapcount for file pages */
6081 if (!get_page_unless_zero(page
))
6088 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6089 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6091 struct page
*page
= NULL
;
6092 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6094 if (!move_anon() || non_swap_entry(ent
))
6097 * Because lookup_swap_cache() updates some statistics counter,
6098 * we call find_get_page() with swapper_space directly.
6100 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6101 if (do_swap_account
)
6102 entry
->val
= ent
.val
;
6107 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6108 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6114 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6115 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6117 struct page
*page
= NULL
;
6118 struct address_space
*mapping
;
6121 if (!vma
->vm_file
) /* anonymous vma */
6126 mapping
= vma
->vm_file
->f_mapping
;
6127 if (pte_none(ptent
))
6128 pgoff
= linear_page_index(vma
, addr
);
6129 else /* pte_file(ptent) is true */
6130 pgoff
= pte_to_pgoff(ptent
);
6132 /* page is moved even if it's not RSS of this task(page-faulted). */
6133 page
= find_get_page(mapping
, pgoff
);
6136 /* shmem/tmpfs may report page out on swap: account for that too. */
6137 if (radix_tree_exceptional_entry(page
)) {
6138 swp_entry_t swap
= radix_to_swp_entry(page
);
6139 if (do_swap_account
)
6141 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6147 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6148 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6150 struct page
*page
= NULL
;
6151 struct page_cgroup
*pc
;
6152 enum mc_target_type ret
= MC_TARGET_NONE
;
6153 swp_entry_t ent
= { .val
= 0 };
6155 if (pte_present(ptent
))
6156 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6157 else if (is_swap_pte(ptent
))
6158 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6159 else if (pte_none(ptent
) || pte_file(ptent
))
6160 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6162 if (!page
&& !ent
.val
)
6165 pc
= lookup_page_cgroup(page
);
6167 * Do only loose check w/o page_cgroup lock.
6168 * mem_cgroup_move_account() checks the pc is valid or not under
6171 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6172 ret
= MC_TARGET_PAGE
;
6174 target
->page
= page
;
6176 if (!ret
|| !target
)
6179 /* There is a swap entry and a page doesn't exist or isn't charged */
6180 if (ent
.val
&& !ret
&&
6181 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6182 ret
= MC_TARGET_SWAP
;
6189 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6191 * We don't consider swapping or file mapped pages because THP does not
6192 * support them for now.
6193 * Caller should make sure that pmd_trans_huge(pmd) is true.
6195 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6196 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6198 struct page
*page
= NULL
;
6199 struct page_cgroup
*pc
;
6200 enum mc_target_type ret
= MC_TARGET_NONE
;
6202 page
= pmd_page(pmd
);
6203 VM_BUG_ON(!page
|| !PageHead(page
));
6206 pc
= lookup_page_cgroup(page
);
6207 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6208 ret
= MC_TARGET_PAGE
;
6211 target
->page
= page
;
6217 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6218 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6220 return MC_TARGET_NONE
;
6224 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6225 unsigned long addr
, unsigned long end
,
6226 struct mm_walk
*walk
)
6228 struct vm_area_struct
*vma
= walk
->private;
6232 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6233 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6234 mc
.precharge
+= HPAGE_PMD_NR
;
6235 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6239 if (pmd_trans_unstable(pmd
))
6241 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6242 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6243 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6244 mc
.precharge
++; /* increment precharge temporarily */
6245 pte_unmap_unlock(pte
- 1, ptl
);
6251 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6253 unsigned long precharge
;
6254 struct vm_area_struct
*vma
;
6256 down_read(&mm
->mmap_sem
);
6257 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6258 struct mm_walk mem_cgroup_count_precharge_walk
= {
6259 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6263 if (is_vm_hugetlb_page(vma
))
6265 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6266 &mem_cgroup_count_precharge_walk
);
6268 up_read(&mm
->mmap_sem
);
6270 precharge
= mc
.precharge
;
6276 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6278 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6280 VM_BUG_ON(mc
.moving_task
);
6281 mc
.moving_task
= current
;
6282 return mem_cgroup_do_precharge(precharge
);
6285 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6286 static void __mem_cgroup_clear_mc(void)
6288 struct mem_cgroup
*from
= mc
.from
;
6289 struct mem_cgroup
*to
= mc
.to
;
6292 /* we must uncharge all the leftover precharges from mc.to */
6294 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6298 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6299 * we must uncharge here.
6301 if (mc
.moved_charge
) {
6302 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6303 mc
.moved_charge
= 0;
6305 /* we must fixup refcnts and charges */
6306 if (mc
.moved_swap
) {
6307 /* uncharge swap account from the old cgroup */
6308 if (!mem_cgroup_is_root(mc
.from
))
6309 res_counter_uncharge(&mc
.from
->memsw
,
6310 PAGE_SIZE
* mc
.moved_swap
);
6312 for (i
= 0; i
< mc
.moved_swap
; i
++)
6313 css_put(&mc
.from
->css
);
6315 if (!mem_cgroup_is_root(mc
.to
)) {
6317 * we charged both to->res and to->memsw, so we should
6320 res_counter_uncharge(&mc
.to
->res
,
6321 PAGE_SIZE
* mc
.moved_swap
);
6323 /* we've already done css_get(mc.to) */
6326 memcg_oom_recover(from
);
6327 memcg_oom_recover(to
);
6328 wake_up_all(&mc
.waitq
);
6331 static void mem_cgroup_clear_mc(void)
6333 struct mem_cgroup
*from
= mc
.from
;
6336 * we must clear moving_task before waking up waiters at the end of
6339 mc
.moving_task
= NULL
;
6340 __mem_cgroup_clear_mc();
6341 spin_lock(&mc
.lock
);
6344 spin_unlock(&mc
.lock
);
6345 mem_cgroup_end_move(from
);
6348 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6349 struct cgroup_taskset
*tset
)
6351 struct task_struct
*p
= cgroup_taskset_first(tset
);
6353 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6354 unsigned long move_charge_at_immigrate
;
6357 * We are now commited to this value whatever it is. Changes in this
6358 * tunable will only affect upcoming migrations, not the current one.
6359 * So we need to save it, and keep it going.
6361 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6362 if (move_charge_at_immigrate
) {
6363 struct mm_struct
*mm
;
6364 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6366 VM_BUG_ON(from
== memcg
);
6368 mm
= get_task_mm(p
);
6371 /* We move charges only when we move a owner of the mm */
6372 if (mm
->owner
== p
) {
6375 VM_BUG_ON(mc
.precharge
);
6376 VM_BUG_ON(mc
.moved_charge
);
6377 VM_BUG_ON(mc
.moved_swap
);
6378 mem_cgroup_start_move(from
);
6379 spin_lock(&mc
.lock
);
6382 mc
.immigrate_flags
= move_charge_at_immigrate
;
6383 spin_unlock(&mc
.lock
);
6384 /* We set mc.moving_task later */
6386 ret
= mem_cgroup_precharge_mc(mm
);
6388 mem_cgroup_clear_mc();
6395 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6396 struct cgroup_taskset
*tset
)
6398 mem_cgroup_clear_mc();
6401 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6402 unsigned long addr
, unsigned long end
,
6403 struct mm_walk
*walk
)
6406 struct vm_area_struct
*vma
= walk
->private;
6409 enum mc_target_type target_type
;
6410 union mc_target target
;
6412 struct page_cgroup
*pc
;
6415 * We don't take compound_lock() here but no race with splitting thp
6417 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6418 * under splitting, which means there's no concurrent thp split,
6419 * - if another thread runs into split_huge_page() just after we
6420 * entered this if-block, the thread must wait for page table lock
6421 * to be unlocked in __split_huge_page_splitting(), where the main
6422 * part of thp split is not executed yet.
6424 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6425 if (mc
.precharge
< HPAGE_PMD_NR
) {
6426 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6429 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6430 if (target_type
== MC_TARGET_PAGE
) {
6432 if (!isolate_lru_page(page
)) {
6433 pc
= lookup_page_cgroup(page
);
6434 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6435 pc
, mc
.from
, mc
.to
)) {
6436 mc
.precharge
-= HPAGE_PMD_NR
;
6437 mc
.moved_charge
+= HPAGE_PMD_NR
;
6439 putback_lru_page(page
);
6443 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6447 if (pmd_trans_unstable(pmd
))
6450 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6451 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6452 pte_t ptent
= *(pte
++);
6458 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6459 case MC_TARGET_PAGE
:
6461 if (isolate_lru_page(page
))
6463 pc
= lookup_page_cgroup(page
);
6464 if (!mem_cgroup_move_account(page
, 1, pc
,
6467 /* we uncharge from mc.from later. */
6470 putback_lru_page(page
);
6471 put
: /* get_mctgt_type() gets the page */
6474 case MC_TARGET_SWAP
:
6476 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6478 /* we fixup refcnts and charges later. */
6486 pte_unmap_unlock(pte
- 1, ptl
);
6491 * We have consumed all precharges we got in can_attach().
6492 * We try charge one by one, but don't do any additional
6493 * charges to mc.to if we have failed in charge once in attach()
6496 ret
= mem_cgroup_do_precharge(1);
6504 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6506 struct vm_area_struct
*vma
;
6508 lru_add_drain_all();
6510 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6512 * Someone who are holding the mmap_sem might be waiting in
6513 * waitq. So we cancel all extra charges, wake up all waiters,
6514 * and retry. Because we cancel precharges, we might not be able
6515 * to move enough charges, but moving charge is a best-effort
6516 * feature anyway, so it wouldn't be a big problem.
6518 __mem_cgroup_clear_mc();
6522 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6524 struct mm_walk mem_cgroup_move_charge_walk
= {
6525 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6529 if (is_vm_hugetlb_page(vma
))
6531 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6532 &mem_cgroup_move_charge_walk
);
6535 * means we have consumed all precharges and failed in
6536 * doing additional charge. Just abandon here.
6540 up_read(&mm
->mmap_sem
);
6543 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6544 struct cgroup_taskset
*tset
)
6546 struct task_struct
*p
= cgroup_taskset_first(tset
);
6547 struct mm_struct
*mm
= get_task_mm(p
);
6551 mem_cgroup_move_charge(mm
);
6555 mem_cgroup_clear_mc();
6557 #else /* !CONFIG_MMU */
6558 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6559 struct cgroup_taskset
*tset
)
6563 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6564 struct cgroup_taskset
*tset
)
6567 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6568 struct cgroup_taskset
*tset
)
6574 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6575 * to verify sane_behavior flag on each mount attempt.
6577 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6580 * use_hierarchy is forced with sane_behavior. cgroup core
6581 * guarantees that @root doesn't have any children, so turning it
6582 * on for the root memcg is enough.
6584 if (cgroup_sane_behavior(root_css
->cgroup
))
6585 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6588 struct cgroup_subsys mem_cgroup_subsys
= {
6590 .subsys_id
= mem_cgroup_subsys_id
,
6591 .css_alloc
= mem_cgroup_css_alloc
,
6592 .css_online
= mem_cgroup_css_online
,
6593 .css_offline
= mem_cgroup_css_offline
,
6594 .css_free
= mem_cgroup_css_free
,
6595 .can_attach
= mem_cgroup_can_attach
,
6596 .cancel_attach
= mem_cgroup_cancel_attach
,
6597 .attach
= mem_cgroup_move_task
,
6598 .bind
= mem_cgroup_bind
,
6599 .base_cftypes
= mem_cgroup_files
,
6604 #ifdef CONFIG_MEMCG_SWAP
6605 static int __init
enable_swap_account(char *s
)
6607 if (!strcmp(s
, "1"))
6608 really_do_swap_account
= 1;
6609 else if (!strcmp(s
, "0"))
6610 really_do_swap_account
= 0;
6613 __setup("swapaccount=", enable_swap_account
);
6615 static void __init
memsw_file_init(void)
6617 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6620 static void __init
enable_swap_cgroup(void)
6622 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6623 do_swap_account
= 1;
6629 static void __init
enable_swap_cgroup(void)
6635 * subsys_initcall() for memory controller.
6637 * Some parts like hotcpu_notifier() have to be initialized from this context
6638 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6639 * everything that doesn't depend on a specific mem_cgroup structure should
6640 * be initialized from here.
6642 static int __init
mem_cgroup_init(void)
6644 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6645 enable_swap_cgroup();
6649 subsys_initcall(mem_cgroup_init
);