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_SOFTLIMIT
,
142 MEM_CGROUP_TARGET_NUMAINFO
,
145 #define THRESHOLDS_EVENTS_TARGET 128
146 #define SOFTLIMIT_EVENTS_TARGET 1024
147 #define NUMAINFO_EVENTS_TARGET 1024
149 struct mem_cgroup_stat_cpu
{
150 long count
[MEM_CGROUP_STAT_NSTATS
];
151 unsigned long events
[MEM_CGROUP_EVENTS_NSTATS
];
152 unsigned long nr_page_events
;
153 unsigned long targets
[MEM_CGROUP_NTARGETS
];
156 struct mem_cgroup_reclaim_iter
{
158 * last scanned hierarchy member. Valid only if last_dead_count
159 * matches memcg->dead_count of the hierarchy root group.
161 struct mem_cgroup
*last_visited
;
162 unsigned long last_dead_count
;
164 /* scan generation, increased every round-trip */
165 unsigned int generation
;
169 * per-zone information in memory controller.
171 struct mem_cgroup_per_zone
{
172 struct lruvec lruvec
;
173 unsigned long lru_size
[NR_LRU_LISTS
];
175 struct mem_cgroup_reclaim_iter reclaim_iter
[DEF_PRIORITY
+ 1];
177 struct mem_cgroup
*memcg
; /* Back pointer, we cannot */
178 /* use container_of */
181 struct mem_cgroup_per_node
{
182 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
185 struct mem_cgroup_threshold
{
186 struct eventfd_ctx
*eventfd
;
191 struct mem_cgroup_threshold_ary
{
192 /* An array index points to threshold just below or equal to usage. */
193 int current_threshold
;
194 /* Size of entries[] */
196 /* Array of thresholds */
197 struct mem_cgroup_threshold entries
[0];
200 struct mem_cgroup_thresholds
{
201 /* Primary thresholds array */
202 struct mem_cgroup_threshold_ary
*primary
;
204 * Spare threshold array.
205 * This is needed to make mem_cgroup_unregister_event() "never fail".
206 * It must be able to store at least primary->size - 1 entries.
208 struct mem_cgroup_threshold_ary
*spare
;
212 struct mem_cgroup_eventfd_list
{
213 struct list_head list
;
214 struct eventfd_ctx
*eventfd
;
217 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
);
218 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
);
221 * The memory controller data structure. The memory controller controls both
222 * page cache and RSS per cgroup. We would eventually like to provide
223 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
224 * to help the administrator determine what knobs to tune.
226 * TODO: Add a water mark for the memory controller. Reclaim will begin when
227 * we hit the water mark. May be even add a low water mark, such that
228 * no reclaim occurs from a cgroup at it's low water mark, this is
229 * a feature that will be implemented much later in the future.
232 struct cgroup_subsys_state css
;
234 * the counter to account for memory usage
236 struct res_counter res
;
238 /* vmpressure notifications */
239 struct vmpressure vmpressure
;
242 * the counter to account for mem+swap usage.
244 struct res_counter memsw
;
247 * the counter to account for kernel memory usage.
249 struct res_counter kmem
;
251 * Should the accounting and control be hierarchical, per subtree?
254 unsigned long kmem_account_flags
; /* See KMEM_ACCOUNTED_*, below */
260 /* OOM-Killer disable */
261 int oom_kill_disable
;
263 /* set when res.limit == memsw.limit */
264 bool memsw_is_minimum
;
266 /* protect arrays of thresholds */
267 struct mutex thresholds_lock
;
269 /* thresholds for memory usage. RCU-protected */
270 struct mem_cgroup_thresholds thresholds
;
272 /* thresholds for mem+swap usage. RCU-protected */
273 struct mem_cgroup_thresholds memsw_thresholds
;
275 /* For oom notifier event fd */
276 struct list_head oom_notify
;
279 * Should we move charges of a task when a task is moved into this
280 * mem_cgroup ? And what type of charges should we move ?
282 unsigned long move_charge_at_immigrate
;
284 * set > 0 if pages under this cgroup are moving to other cgroup.
286 atomic_t moving_account
;
287 /* taken only while moving_account > 0 */
288 spinlock_t move_lock
;
292 struct mem_cgroup_stat_cpu __percpu
*stat
;
294 * used when a cpu is offlined or other synchronizations
295 * See mem_cgroup_read_stat().
297 struct mem_cgroup_stat_cpu nocpu_base
;
298 spinlock_t pcp_counter_lock
;
301 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
302 struct tcp_memcontrol tcp_mem
;
304 #if defined(CONFIG_MEMCG_KMEM)
305 /* analogous to slab_common's slab_caches list. per-memcg */
306 struct list_head memcg_slab_caches
;
307 /* Not a spinlock, we can take a lot of time walking the list */
308 struct mutex slab_caches_mutex
;
309 /* Index in the kmem_cache->memcg_params->memcg_caches array */
313 int last_scanned_node
;
315 nodemask_t scan_nodes
;
316 atomic_t numainfo_events
;
317 atomic_t numainfo_updating
;
320 * Protects soft_contributed transitions.
321 * See mem_cgroup_update_soft_limit
323 spinlock_t soft_lock
;
326 * If true then this group has increased parents' children_in_excess
327 * when it got over the soft limit.
328 * When a group falls bellow the soft limit, parents' children_in_excess
329 * is decreased and soft_contributed changed to false.
331 bool soft_contributed
;
333 /* Number of children that are in soft limit excess */
334 atomic_t children_in_excess
;
336 struct mem_cgroup_per_node
*nodeinfo
[0];
337 /* WARNING: nodeinfo must be the last member here */
340 static size_t memcg_size(void)
342 return sizeof(struct mem_cgroup
) +
343 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
346 /* internal only representation about the status of kmem accounting. */
348 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
349 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
350 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
353 /* We account when limit is on, but only after call sites are patched */
354 #define KMEM_ACCOUNTED_MASK \
355 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
357 #ifdef CONFIG_MEMCG_KMEM
358 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
360 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
363 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
365 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
368 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
370 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
373 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
375 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
378 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
381 * Our caller must use css_get() first, because memcg_uncharge_kmem()
382 * will call css_put() if it sees the memcg is dead.
385 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
386 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
389 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
391 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
392 &memcg
->kmem_account_flags
);
396 /* Stuffs for move charges at task migration. */
398 * Types of charges to be moved. "move_charge_at_immitgrate" and
399 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
402 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
403 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
407 /* "mc" and its members are protected by cgroup_mutex */
408 static struct move_charge_struct
{
409 spinlock_t lock
; /* for from, to */
410 struct mem_cgroup
*from
;
411 struct mem_cgroup
*to
;
412 unsigned long immigrate_flags
;
413 unsigned long precharge
;
414 unsigned long moved_charge
;
415 unsigned long moved_swap
;
416 struct task_struct
*moving_task
; /* a task moving charges */
417 wait_queue_head_t waitq
; /* a waitq for other context */
419 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
420 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
423 static bool move_anon(void)
425 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
428 static bool move_file(void)
430 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
434 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
435 * limit reclaim to prevent infinite loops, if they ever occur.
437 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
440 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
441 MEM_CGROUP_CHARGE_TYPE_ANON
,
442 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
443 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
447 /* for encoding cft->private value on file */
455 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
456 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
457 #define MEMFILE_ATTR(val) ((val) & 0xffff)
458 /* Used for OOM nofiier */
459 #define OOM_CONTROL (0)
462 * Reclaim flags for mem_cgroup_hierarchical_reclaim
464 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
465 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
466 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
467 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
470 * The memcg_create_mutex will be held whenever a new cgroup is created.
471 * As a consequence, any change that needs to protect against new child cgroups
472 * appearing has to hold it as well.
474 static DEFINE_MUTEX(memcg_create_mutex
);
476 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
478 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
481 /* Some nice accessors for the vmpressure. */
482 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
485 memcg
= root_mem_cgroup
;
486 return &memcg
->vmpressure
;
489 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
491 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
494 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
496 return &mem_cgroup_from_css(css
)->vmpressure
;
499 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
501 return (memcg
== root_mem_cgroup
);
504 /* Writing them here to avoid exposing memcg's inner layout */
505 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
507 void sock_update_memcg(struct sock
*sk
)
509 if (mem_cgroup_sockets_enabled
) {
510 struct mem_cgroup
*memcg
;
511 struct cg_proto
*cg_proto
;
513 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
515 /* Socket cloning can throw us here with sk_cgrp already
516 * filled. It won't however, necessarily happen from
517 * process context. So the test for root memcg given
518 * the current task's memcg won't help us in this case.
520 * Respecting the original socket's memcg is a better
521 * decision in this case.
524 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
525 css_get(&sk
->sk_cgrp
->memcg
->css
);
530 memcg
= mem_cgroup_from_task(current
);
531 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
532 if (!mem_cgroup_is_root(memcg
) &&
533 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
534 sk
->sk_cgrp
= cg_proto
;
539 EXPORT_SYMBOL(sock_update_memcg
);
541 void sock_release_memcg(struct sock
*sk
)
543 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
544 struct mem_cgroup
*memcg
;
545 WARN_ON(!sk
->sk_cgrp
->memcg
);
546 memcg
= sk
->sk_cgrp
->memcg
;
547 css_put(&sk
->sk_cgrp
->memcg
->css
);
551 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
553 if (!memcg
|| mem_cgroup_is_root(memcg
))
556 return &memcg
->tcp_mem
.cg_proto
;
558 EXPORT_SYMBOL(tcp_proto_cgroup
);
560 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
562 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
564 static_key_slow_dec(&memcg_socket_limit_enabled
);
567 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
572 #ifdef CONFIG_MEMCG_KMEM
574 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
575 * There are two main reasons for not using the css_id for this:
576 * 1) this works better in sparse environments, where we have a lot of memcgs,
577 * but only a few kmem-limited. Or also, if we have, for instance, 200
578 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
579 * 200 entry array for that.
581 * 2) In order not to violate the cgroup API, we would like to do all memory
582 * allocation in ->create(). At that point, we haven't yet allocated the
583 * css_id. Having a separate index prevents us from messing with the cgroup
586 * The current size of the caches array is stored in
587 * memcg_limited_groups_array_size. It will double each time we have to
590 static DEFINE_IDA(kmem_limited_groups
);
591 int memcg_limited_groups_array_size
;
594 * MIN_SIZE is different than 1, because we would like to avoid going through
595 * the alloc/free process all the time. In a small machine, 4 kmem-limited
596 * cgroups is a reasonable guess. In the future, it could be a parameter or
597 * tunable, but that is strictly not necessary.
599 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
600 * this constant directly from cgroup, but it is understandable that this is
601 * better kept as an internal representation in cgroup.c. In any case, the
602 * css_id space is not getting any smaller, and we don't have to necessarily
603 * increase ours as well if it increases.
605 #define MEMCG_CACHES_MIN_SIZE 4
606 #define MEMCG_CACHES_MAX_SIZE 65535
609 * A lot of the calls to the cache allocation functions are expected to be
610 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
611 * conditional to this static branch, we'll have to allow modules that does
612 * kmem_cache_alloc and the such to see this symbol as well
614 struct static_key memcg_kmem_enabled_key
;
615 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
617 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
619 if (memcg_kmem_is_active(memcg
)) {
620 static_key_slow_dec(&memcg_kmem_enabled_key
);
621 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
624 * This check can't live in kmem destruction function,
625 * since the charges will outlive the cgroup
627 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
630 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
633 #endif /* CONFIG_MEMCG_KMEM */
635 static void disarm_static_keys(struct mem_cgroup
*memcg
)
637 disarm_sock_keys(memcg
);
638 disarm_kmem_keys(memcg
);
641 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
643 static struct mem_cgroup_per_zone
*
644 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
646 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
647 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
650 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
655 static struct mem_cgroup_per_zone
*
656 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
658 int nid
= page_to_nid(page
);
659 int zid
= page_zonenum(page
);
661 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
665 * Implementation Note: reading percpu statistics for memcg.
667 * Both of vmstat[] and percpu_counter has threshold and do periodic
668 * synchronization to implement "quick" read. There are trade-off between
669 * reading cost and precision of value. Then, we may have a chance to implement
670 * a periodic synchronizion of counter in memcg's counter.
672 * But this _read() function is used for user interface now. The user accounts
673 * memory usage by memory cgroup and he _always_ requires exact value because
674 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
675 * have to visit all online cpus and make sum. So, for now, unnecessary
676 * synchronization is not implemented. (just implemented for cpu hotplug)
678 * If there are kernel internal actions which can make use of some not-exact
679 * value, and reading all cpu value can be performance bottleneck in some
680 * common workload, threashold and synchonization as vmstat[] should be
683 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
684 enum mem_cgroup_stat_index idx
)
690 for_each_online_cpu(cpu
)
691 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
692 #ifdef CONFIG_HOTPLUG_CPU
693 spin_lock(&memcg
->pcp_counter_lock
);
694 val
+= memcg
->nocpu_base
.count
[idx
];
695 spin_unlock(&memcg
->pcp_counter_lock
);
701 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
704 int val
= (charge
) ? 1 : -1;
705 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
708 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
709 enum mem_cgroup_events_index idx
)
711 unsigned long val
= 0;
714 for_each_online_cpu(cpu
)
715 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
716 #ifdef CONFIG_HOTPLUG_CPU
717 spin_lock(&memcg
->pcp_counter_lock
);
718 val
+= memcg
->nocpu_base
.events
[idx
];
719 spin_unlock(&memcg
->pcp_counter_lock
);
724 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
726 bool anon
, int nr_pages
)
731 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
732 * counted as CACHE even if it's on ANON LRU.
735 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
738 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
741 if (PageTransHuge(page
))
742 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
745 /* pagein of a big page is an event. So, ignore page size */
747 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
749 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
750 nr_pages
= -nr_pages
; /* for event */
753 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
759 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
761 struct mem_cgroup_per_zone
*mz
;
763 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
764 return mz
->lru_size
[lru
];
768 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
769 unsigned int lru_mask
)
771 struct mem_cgroup_per_zone
*mz
;
773 unsigned long ret
= 0;
775 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
778 if (BIT(lru
) & lru_mask
)
779 ret
+= mz
->lru_size
[lru
];
785 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
786 int nid
, unsigned int lru_mask
)
791 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
792 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
798 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
799 unsigned int lru_mask
)
804 for_each_node_state(nid
, N_MEMORY
)
805 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
809 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
810 enum mem_cgroup_events_target target
)
812 unsigned long val
, next
;
814 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
815 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
816 /* from time_after() in jiffies.h */
817 if ((long)next
- (long)val
< 0) {
819 case MEM_CGROUP_TARGET_THRESH
:
820 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
822 case MEM_CGROUP_TARGET_SOFTLIMIT
:
823 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
825 case MEM_CGROUP_TARGET_NUMAINFO
:
826 next
= val
+ NUMAINFO_EVENTS_TARGET
;
831 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
838 * Called from rate-limited memcg_check_events when enough
839 * MEM_CGROUP_TARGET_SOFTLIMIT events are accumulated and it makes sure
840 * that all the parents up the hierarchy will be notified that this group
841 * is in excess or that it is not in excess anymore. mmecg->soft_contributed
842 * makes the transition a single action whenever the state flips from one to
845 static void mem_cgroup_update_soft_limit(struct mem_cgroup
*memcg
)
847 unsigned long long excess
= res_counter_soft_limit_excess(&memcg
->res
);
848 struct mem_cgroup
*parent
= memcg
;
851 spin_lock(&memcg
->soft_lock
);
853 if (!memcg
->soft_contributed
) {
855 memcg
->soft_contributed
= true;
858 if (memcg
->soft_contributed
) {
860 memcg
->soft_contributed
= false;
865 * Necessary to update all ancestors when hierarchy is used
866 * because their event counter is not touched.
867 * We track children even outside the hierarchy for the root
868 * cgroup because tree walk starting at root should visit
869 * all cgroups and we want to prevent from pointless tree
870 * walk if no children is below the limit.
872 while (delta
&& (parent
= parent_mem_cgroup(parent
)))
873 atomic_add(delta
, &parent
->children_in_excess
);
874 if (memcg
!= root_mem_cgroup
&& !root_mem_cgroup
->use_hierarchy
)
875 atomic_add(delta
, &root_mem_cgroup
->children_in_excess
);
876 spin_unlock(&memcg
->soft_lock
);
880 * Check events in order.
883 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
886 /* threshold event is triggered in finer grain than soft limit */
887 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
888 MEM_CGROUP_TARGET_THRESH
))) {
890 bool do_numainfo __maybe_unused
;
892 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
893 MEM_CGROUP_TARGET_SOFTLIMIT
);
895 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
896 MEM_CGROUP_TARGET_NUMAINFO
);
900 mem_cgroup_threshold(memcg
);
901 if (unlikely(do_softlimit
))
902 mem_cgroup_update_soft_limit(memcg
);
904 if (unlikely(do_numainfo
))
905 atomic_inc(&memcg
->numainfo_events
);
911 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
914 * mm_update_next_owner() may clear mm->owner to NULL
915 * if it races with swapoff, page migration, etc.
916 * So this can be called with p == NULL.
921 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
924 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
926 struct mem_cgroup
*memcg
= NULL
;
931 * Because we have no locks, mm->owner's may be being moved to other
932 * cgroup. We use css_tryget() here even if this looks
933 * pessimistic (rather than adding locks here).
937 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
938 if (unlikely(!memcg
))
940 } while (!css_tryget(&memcg
->css
));
945 static enum mem_cgroup_filter_t
946 mem_cgroup_filter(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
,
947 mem_cgroup_iter_filter cond
)
951 return cond(memcg
, root
);
955 * Returns a next (in a pre-order walk) alive memcg (with elevated css
956 * ref. count) or NULL if the whole root's subtree has been visited.
958 * helper function to be used by mem_cgroup_iter
960 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
961 struct mem_cgroup
*last_visited
, mem_cgroup_iter_filter cond
)
963 struct cgroup_subsys_state
*prev_css
, *next_css
;
965 prev_css
= last_visited
? &last_visited
->css
: NULL
;
967 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
970 * Even if we found a group we have to make sure it is
971 * alive. css && !memcg means that the groups should be
972 * skipped and we should continue the tree walk.
973 * last_visited css is safe to use because it is
974 * protected by css_get and the tree walk is rcu safe.
977 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
979 switch (mem_cgroup_filter(mem
, root
, cond
)) {
987 * css_rightmost_descendant is not an optimal way to
988 * skip through a subtree (especially for imbalanced
989 * trees leaning to right) but that's what we have right
990 * now. More effective solution would be traversing
991 * right-up for first non-NULL without calling
992 * css_next_descendant_pre afterwards.
994 prev_css
= css_rightmost_descendant(next_css
);
997 if (css_tryget(&mem
->css
))
1000 prev_css
= next_css
;
1010 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1013 * When a group in the hierarchy below root is destroyed, the
1014 * hierarchy iterator can no longer be trusted since it might
1015 * have pointed to the destroyed group. Invalidate it.
1017 atomic_inc(&root
->dead_count
);
1020 static struct mem_cgroup
*
1021 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1022 struct mem_cgroup
*root
,
1025 struct mem_cgroup
*position
= NULL
;
1027 * A cgroup destruction happens in two stages: offlining and
1028 * release. They are separated by a RCU grace period.
1030 * If the iterator is valid, we may still race with an
1031 * offlining. The RCU lock ensures the object won't be
1032 * released, tryget will fail if we lost the race.
1034 *sequence
= atomic_read(&root
->dead_count
);
1035 if (iter
->last_dead_count
== *sequence
) {
1037 position
= iter
->last_visited
;
1038 if (position
&& !css_tryget(&position
->css
))
1044 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1045 struct mem_cgroup
*last_visited
,
1046 struct mem_cgroup
*new_position
,
1050 css_put(&last_visited
->css
);
1052 * We store the sequence count from the time @last_visited was
1053 * loaded successfully instead of rereading it here so that we
1054 * don't lose destruction events in between. We could have
1055 * raced with the destruction of @new_position after all.
1057 iter
->last_visited
= new_position
;
1059 iter
->last_dead_count
= sequence
;
1063 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1064 * @root: hierarchy root
1065 * @prev: previously returned memcg, NULL on first invocation
1066 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1067 * @cond: filter for visited nodes, NULL for no filter
1069 * Returns references to children of the hierarchy below @root, or
1070 * @root itself, or %NULL after a full round-trip.
1072 * Caller must pass the return value in @prev on subsequent
1073 * invocations for reference counting, or use mem_cgroup_iter_break()
1074 * to cancel a hierarchy walk before the round-trip is complete.
1076 * Reclaimers can specify a zone and a priority level in @reclaim to
1077 * divide up the memcgs in the hierarchy among all concurrent
1078 * reclaimers operating on the same zone and priority.
1080 struct mem_cgroup
*mem_cgroup_iter_cond(struct mem_cgroup
*root
,
1081 struct mem_cgroup
*prev
,
1082 struct mem_cgroup_reclaim_cookie
*reclaim
,
1083 mem_cgroup_iter_filter cond
)
1085 struct mem_cgroup
*memcg
= NULL
;
1086 struct mem_cgroup
*last_visited
= NULL
;
1088 if (mem_cgroup_disabled()) {
1089 /* first call must return non-NULL, second return NULL */
1090 return (struct mem_cgroup
*)(unsigned long)!prev
;
1094 root
= root_mem_cgroup
;
1096 if (prev
&& !reclaim
)
1097 last_visited
= prev
;
1099 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1102 if (mem_cgroup_filter(root
, root
, cond
) == VISIT
)
1109 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1110 int uninitialized_var(seq
);
1113 int nid
= zone_to_nid(reclaim
->zone
);
1114 int zid
= zone_idx(reclaim
->zone
);
1115 struct mem_cgroup_per_zone
*mz
;
1117 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1118 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1119 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1120 iter
->last_visited
= NULL
;
1124 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1127 memcg
= __mem_cgroup_iter_next(root
, last_visited
, cond
);
1130 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1134 else if (!prev
&& memcg
)
1135 reclaim
->generation
= iter
->generation
;
1139 * We have finished the whole tree walk or no group has been
1140 * visited because filter told us to skip the root node.
1142 if (!memcg
&& (prev
|| (cond
&& !last_visited
)))
1148 if (prev
&& prev
!= root
)
1149 css_put(&prev
->css
);
1155 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1156 * @root: hierarchy root
1157 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1159 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1160 struct mem_cgroup
*prev
)
1163 root
= root_mem_cgroup
;
1164 if (prev
&& prev
!= root
)
1165 css_put(&prev
->css
);
1169 * Iteration constructs for visiting all cgroups (under a tree). If
1170 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1171 * be used for reference counting.
1173 #define for_each_mem_cgroup_tree(iter, root) \
1174 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1176 iter = mem_cgroup_iter(root, iter, NULL))
1178 #define for_each_mem_cgroup(iter) \
1179 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1181 iter = mem_cgroup_iter(NULL, iter, NULL))
1183 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1185 struct mem_cgroup
*memcg
;
1188 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1189 if (unlikely(!memcg
))
1194 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1197 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1205 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1208 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1209 * @zone: zone of the wanted lruvec
1210 * @memcg: memcg of the wanted lruvec
1212 * Returns the lru list vector holding pages for the given @zone and
1213 * @mem. This can be the global zone lruvec, if the memory controller
1216 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1217 struct mem_cgroup
*memcg
)
1219 struct mem_cgroup_per_zone
*mz
;
1220 struct lruvec
*lruvec
;
1222 if (mem_cgroup_disabled()) {
1223 lruvec
= &zone
->lruvec
;
1227 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1228 lruvec
= &mz
->lruvec
;
1231 * Since a node can be onlined after the mem_cgroup was created,
1232 * we have to be prepared to initialize lruvec->zone here;
1233 * and if offlined then reonlined, we need to reinitialize it.
1235 if (unlikely(lruvec
->zone
!= zone
))
1236 lruvec
->zone
= zone
;
1241 * Following LRU functions are allowed to be used without PCG_LOCK.
1242 * Operations are called by routine of global LRU independently from memcg.
1243 * What we have to take care of here is validness of pc->mem_cgroup.
1245 * Changes to pc->mem_cgroup happens when
1248 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1249 * It is added to LRU before charge.
1250 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1251 * When moving account, the page is not on LRU. It's isolated.
1255 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1257 * @zone: zone of the page
1259 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1261 struct mem_cgroup_per_zone
*mz
;
1262 struct mem_cgroup
*memcg
;
1263 struct page_cgroup
*pc
;
1264 struct lruvec
*lruvec
;
1266 if (mem_cgroup_disabled()) {
1267 lruvec
= &zone
->lruvec
;
1271 pc
= lookup_page_cgroup(page
);
1272 memcg
= pc
->mem_cgroup
;
1275 * Surreptitiously switch any uncharged offlist page to root:
1276 * an uncharged page off lru does nothing to secure
1277 * its former mem_cgroup from sudden removal.
1279 * Our caller holds lru_lock, and PageCgroupUsed is updated
1280 * under page_cgroup lock: between them, they make all uses
1281 * of pc->mem_cgroup safe.
1283 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1284 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1286 mz
= page_cgroup_zoneinfo(memcg
, page
);
1287 lruvec
= &mz
->lruvec
;
1290 * Since a node can be onlined after the mem_cgroup was created,
1291 * we have to be prepared to initialize lruvec->zone here;
1292 * and if offlined then reonlined, we need to reinitialize it.
1294 if (unlikely(lruvec
->zone
!= zone
))
1295 lruvec
->zone
= zone
;
1300 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1301 * @lruvec: mem_cgroup per zone lru vector
1302 * @lru: index of lru list the page is sitting on
1303 * @nr_pages: positive when adding or negative when removing
1305 * This function must be called when a page is added to or removed from an
1308 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1311 struct mem_cgroup_per_zone
*mz
;
1312 unsigned long *lru_size
;
1314 if (mem_cgroup_disabled())
1317 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1318 lru_size
= mz
->lru_size
+ lru
;
1319 *lru_size
+= nr_pages
;
1320 VM_BUG_ON((long)(*lru_size
) < 0);
1324 * Checks whether given mem is same or in the root_mem_cgroup's
1327 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1328 struct mem_cgroup
*memcg
)
1330 if (root_memcg
== memcg
)
1332 if (!root_memcg
->use_hierarchy
|| !memcg
)
1334 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1337 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1338 struct mem_cgroup
*memcg
)
1343 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1348 bool task_in_mem_cgroup(struct task_struct
*task
,
1349 const struct mem_cgroup
*memcg
)
1351 struct mem_cgroup
*curr
= NULL
;
1352 struct task_struct
*p
;
1355 p
= find_lock_task_mm(task
);
1357 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1361 * All threads may have already detached their mm's, but the oom
1362 * killer still needs to detect if they have already been oom
1363 * killed to prevent needlessly killing additional tasks.
1366 curr
= mem_cgroup_from_task(task
);
1368 css_get(&curr
->css
);
1374 * We should check use_hierarchy of "memcg" not "curr". Because checking
1375 * use_hierarchy of "curr" here make this function true if hierarchy is
1376 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1377 * hierarchy(even if use_hierarchy is disabled in "memcg").
1379 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1380 css_put(&curr
->css
);
1384 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1386 unsigned long inactive_ratio
;
1387 unsigned long inactive
;
1388 unsigned long active
;
1391 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1392 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1394 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1396 inactive_ratio
= int_sqrt(10 * gb
);
1400 return inactive
* inactive_ratio
< active
;
1403 #define mem_cgroup_from_res_counter(counter, member) \
1404 container_of(counter, struct mem_cgroup, member)
1407 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1408 * @memcg: the memory cgroup
1410 * Returns the maximum amount of memory @mem can be charged with, in
1413 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1415 unsigned long long margin
;
1417 margin
= res_counter_margin(&memcg
->res
);
1418 if (do_swap_account
)
1419 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1420 return margin
>> PAGE_SHIFT
;
1423 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1426 if (!css_parent(&memcg
->css
))
1427 return vm_swappiness
;
1429 return memcg
->swappiness
;
1433 * memcg->moving_account is used for checking possibility that some thread is
1434 * calling move_account(). When a thread on CPU-A starts moving pages under
1435 * a memcg, other threads should check memcg->moving_account under
1436 * rcu_read_lock(), like this:
1440 * memcg->moving_account+1 if (memcg->mocing_account)
1442 * synchronize_rcu() update something.
1447 /* for quick checking without looking up memcg */
1448 atomic_t memcg_moving __read_mostly
;
1450 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1452 atomic_inc(&memcg_moving
);
1453 atomic_inc(&memcg
->moving_account
);
1457 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1460 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1461 * We check NULL in callee rather than caller.
1464 atomic_dec(&memcg_moving
);
1465 atomic_dec(&memcg
->moving_account
);
1470 * 2 routines for checking "mem" is under move_account() or not.
1472 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1473 * is used for avoiding races in accounting. If true,
1474 * pc->mem_cgroup may be overwritten.
1476 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1477 * under hierarchy of moving cgroups. This is for
1478 * waiting at hith-memory prressure caused by "move".
1481 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1483 VM_BUG_ON(!rcu_read_lock_held());
1484 return atomic_read(&memcg
->moving_account
) > 0;
1487 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1489 struct mem_cgroup
*from
;
1490 struct mem_cgroup
*to
;
1493 * Unlike task_move routines, we access mc.to, mc.from not under
1494 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1496 spin_lock(&mc
.lock
);
1502 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1503 || mem_cgroup_same_or_subtree(memcg
, to
);
1505 spin_unlock(&mc
.lock
);
1509 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1511 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1512 if (mem_cgroup_under_move(memcg
)) {
1514 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1515 /* moving charge context might have finished. */
1518 finish_wait(&mc
.waitq
, &wait
);
1526 * Take this lock when
1527 * - a code tries to modify page's memcg while it's USED.
1528 * - a code tries to modify page state accounting in a memcg.
1529 * see mem_cgroup_stolen(), too.
1531 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1532 unsigned long *flags
)
1534 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1537 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1538 unsigned long *flags
)
1540 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1543 #define K(x) ((x) << (PAGE_SHIFT-10))
1545 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1546 * @memcg: The memory cgroup that went over limit
1547 * @p: Task that is going to be killed
1549 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1552 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1554 struct cgroup
*task_cgrp
;
1555 struct cgroup
*mem_cgrp
;
1557 * Need a buffer in BSS, can't rely on allocations. The code relies
1558 * on the assumption that OOM is serialized for memory controller.
1559 * If this assumption is broken, revisit this code.
1561 static char memcg_name
[PATH_MAX
];
1563 struct mem_cgroup
*iter
;
1571 mem_cgrp
= memcg
->css
.cgroup
;
1572 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1574 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1577 * Unfortunately, we are unable to convert to a useful name
1578 * But we'll still print out the usage information
1585 pr_info("Task in %s killed", memcg_name
);
1588 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1596 * Continues from above, so we don't need an KERN_ level
1598 pr_cont(" as a result of limit of %s\n", memcg_name
);
1601 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1602 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1603 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1604 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1605 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1606 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1607 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1608 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1609 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1610 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1611 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1612 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1614 for_each_mem_cgroup_tree(iter
, memcg
) {
1615 pr_info("Memory cgroup stats");
1618 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1620 pr_cont(" for %s", memcg_name
);
1624 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1625 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1627 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1628 K(mem_cgroup_read_stat(iter
, i
)));
1631 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1632 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1633 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1640 * This function returns the number of memcg under hierarchy tree. Returns
1641 * 1(self count) if no children.
1643 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1646 struct mem_cgroup
*iter
;
1648 for_each_mem_cgroup_tree(iter
, memcg
)
1654 * Return the memory (and swap, if configured) limit for a memcg.
1656 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1660 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1663 * Do not consider swap space if we cannot swap due to swappiness
1665 if (mem_cgroup_swappiness(memcg
)) {
1668 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1669 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1672 * If memsw is finite and limits the amount of swap space
1673 * available to this memcg, return that limit.
1675 limit
= min(limit
, memsw
);
1681 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1684 struct mem_cgroup
*iter
;
1685 unsigned long chosen_points
= 0;
1686 unsigned long totalpages
;
1687 unsigned int points
= 0;
1688 struct task_struct
*chosen
= NULL
;
1691 * If current has a pending SIGKILL or is exiting, then automatically
1692 * select it. The goal is to allow it to allocate so that it may
1693 * quickly exit and free its memory.
1695 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1696 set_thread_flag(TIF_MEMDIE
);
1700 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1701 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1702 for_each_mem_cgroup_tree(iter
, memcg
) {
1703 struct css_task_iter it
;
1704 struct task_struct
*task
;
1706 css_task_iter_start(&iter
->css
, &it
);
1707 while ((task
= css_task_iter_next(&it
))) {
1708 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1710 case OOM_SCAN_SELECT
:
1712 put_task_struct(chosen
);
1714 chosen_points
= ULONG_MAX
;
1715 get_task_struct(chosen
);
1717 case OOM_SCAN_CONTINUE
:
1719 case OOM_SCAN_ABORT
:
1720 css_task_iter_end(&it
);
1721 mem_cgroup_iter_break(memcg
, iter
);
1723 put_task_struct(chosen
);
1728 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1729 if (points
> chosen_points
) {
1731 put_task_struct(chosen
);
1733 chosen_points
= points
;
1734 get_task_struct(chosen
);
1737 css_task_iter_end(&it
);
1742 points
= chosen_points
* 1000 / totalpages
;
1743 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1744 NULL
, "Memory cgroup out of memory");
1747 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1749 unsigned long flags
)
1751 unsigned long total
= 0;
1752 bool noswap
= false;
1755 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1757 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1760 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1762 drain_all_stock_async(memcg
);
1763 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1765 * Allow limit shrinkers, which are triggered directly
1766 * by userspace, to catch signals and stop reclaim
1767 * after minimal progress, regardless of the margin.
1769 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1771 if (mem_cgroup_margin(memcg
))
1774 * If nothing was reclaimed after two attempts, there
1775 * may be no reclaimable pages in this hierarchy.
1783 #if MAX_NUMNODES > 1
1785 * test_mem_cgroup_node_reclaimable
1786 * @memcg: the target memcg
1787 * @nid: the node ID to be checked.
1788 * @noswap : specify true here if the user wants flle only information.
1790 * This function returns whether the specified memcg contains any
1791 * reclaimable pages on a node. Returns true if there are any reclaimable
1792 * pages in the node.
1794 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1795 int nid
, bool noswap
)
1797 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1799 if (noswap
|| !total_swap_pages
)
1801 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1808 * Always updating the nodemask is not very good - even if we have an empty
1809 * list or the wrong list here, we can start from some node and traverse all
1810 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1813 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1817 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1818 * pagein/pageout changes since the last update.
1820 if (!atomic_read(&memcg
->numainfo_events
))
1822 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1825 /* make a nodemask where this memcg uses memory from */
1826 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1828 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1830 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1831 node_clear(nid
, memcg
->scan_nodes
);
1834 atomic_set(&memcg
->numainfo_events
, 0);
1835 atomic_set(&memcg
->numainfo_updating
, 0);
1839 * Selecting a node where we start reclaim from. Because what we need is just
1840 * reducing usage counter, start from anywhere is O,K. Considering
1841 * memory reclaim from current node, there are pros. and cons.
1843 * Freeing memory from current node means freeing memory from a node which
1844 * we'll use or we've used. So, it may make LRU bad. And if several threads
1845 * hit limits, it will see a contention on a node. But freeing from remote
1846 * node means more costs for memory reclaim because of memory latency.
1848 * Now, we use round-robin. Better algorithm is welcomed.
1850 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1854 mem_cgroup_may_update_nodemask(memcg
);
1855 node
= memcg
->last_scanned_node
;
1857 node
= next_node(node
, memcg
->scan_nodes
);
1858 if (node
== MAX_NUMNODES
)
1859 node
= first_node(memcg
->scan_nodes
);
1861 * We call this when we hit limit, not when pages are added to LRU.
1862 * No LRU may hold pages because all pages are UNEVICTABLE or
1863 * memcg is too small and all pages are not on LRU. In that case,
1864 * we use curret node.
1866 if (unlikely(node
== MAX_NUMNODES
))
1867 node
= numa_node_id();
1869 memcg
->last_scanned_node
= node
;
1874 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1882 * A group is eligible for the soft limit reclaim under the given root
1884 * a) it is over its soft limit
1885 * b) any parent up the hierarchy is over its soft limit
1887 * If the given group doesn't have any children over the limit then it
1888 * doesn't make any sense to iterate its subtree.
1890 enum mem_cgroup_filter_t
1891 mem_cgroup_soft_reclaim_eligible(struct mem_cgroup
*memcg
,
1892 struct mem_cgroup
*root
)
1894 struct mem_cgroup
*parent
;
1897 memcg
= root_mem_cgroup
;
1900 if (res_counter_soft_limit_excess(&memcg
->res
))
1904 * If any parent up to the root in the hierarchy is over its soft limit
1905 * then we have to obey and reclaim from this group as well.
1907 while ((parent
= parent_mem_cgroup(parent
))) {
1908 if (res_counter_soft_limit_excess(&parent
->res
))
1914 if (!atomic_read(&memcg
->children_in_excess
))
1919 static DEFINE_SPINLOCK(memcg_oom_lock
);
1922 * Check OOM-Killer is already running under our hierarchy.
1923 * If someone is running, return false.
1925 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1927 struct mem_cgroup
*iter
, *failed
= NULL
;
1929 spin_lock(&memcg_oom_lock
);
1931 for_each_mem_cgroup_tree(iter
, memcg
) {
1932 if (iter
->oom_lock
) {
1934 * this subtree of our hierarchy is already locked
1935 * so we cannot give a lock.
1938 mem_cgroup_iter_break(memcg
, iter
);
1941 iter
->oom_lock
= true;
1946 * OK, we failed to lock the whole subtree so we have
1947 * to clean up what we set up to the failing subtree
1949 for_each_mem_cgroup_tree(iter
, memcg
) {
1950 if (iter
== failed
) {
1951 mem_cgroup_iter_break(memcg
, iter
);
1954 iter
->oom_lock
= false;
1958 spin_unlock(&memcg_oom_lock
);
1963 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1965 struct mem_cgroup
*iter
;
1967 spin_lock(&memcg_oom_lock
);
1968 for_each_mem_cgroup_tree(iter
, memcg
)
1969 iter
->oom_lock
= false;
1970 spin_unlock(&memcg_oom_lock
);
1973 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1975 struct mem_cgroup
*iter
;
1977 for_each_mem_cgroup_tree(iter
, memcg
)
1978 atomic_inc(&iter
->under_oom
);
1981 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1983 struct mem_cgroup
*iter
;
1986 * When a new child is created while the hierarchy is under oom,
1987 * mem_cgroup_oom_lock() may not be called. We have to use
1988 * atomic_add_unless() here.
1990 for_each_mem_cgroup_tree(iter
, memcg
)
1991 atomic_add_unless(&iter
->under_oom
, -1, 0);
1994 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1996 struct oom_wait_info
{
1997 struct mem_cgroup
*memcg
;
2001 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2002 unsigned mode
, int sync
, void *arg
)
2004 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2005 struct mem_cgroup
*oom_wait_memcg
;
2006 struct oom_wait_info
*oom_wait_info
;
2008 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2009 oom_wait_memcg
= oom_wait_info
->memcg
;
2012 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2013 * Then we can use css_is_ancestor without taking care of RCU.
2015 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2016 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2018 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2021 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2023 /* for filtering, pass "memcg" as argument. */
2024 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2027 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2029 if (memcg
&& atomic_read(&memcg
->under_oom
))
2030 memcg_wakeup_oom(memcg
);
2034 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2036 static bool mem_cgroup_handle_oom(struct mem_cgroup
*memcg
, gfp_t mask
,
2039 struct oom_wait_info owait
;
2042 owait
.memcg
= memcg
;
2043 owait
.wait
.flags
= 0;
2044 owait
.wait
.func
= memcg_oom_wake_function
;
2045 owait
.wait
.private = current
;
2046 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2049 * As with any blocking lock, a contender needs to start
2050 * listening for wakeups before attempting the trylock,
2051 * otherwise it can miss the wakeup from the unlock and sleep
2052 * indefinitely. This is just open-coded because our locking
2053 * is so particular to memcg hierarchies.
2055 * Even if signal_pending(), we can't quit charge() loop without
2056 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2057 * under OOM is always welcomed, use TASK_KILLABLE here.
2059 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2060 mem_cgroup_mark_under_oom(memcg
);
2062 locked
= mem_cgroup_oom_trylock(memcg
);
2065 mem_cgroup_oom_notify(memcg
);
2067 if (locked
&& !memcg
->oom_kill_disable
) {
2068 mem_cgroup_unmark_under_oom(memcg
);
2069 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2070 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2073 mem_cgroup_unmark_under_oom(memcg
);
2074 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2078 mem_cgroup_oom_unlock(memcg
);
2080 * There is no guarantee that an OOM-lock contender
2081 * sees the wakeups triggered by the OOM kill
2082 * uncharges. Wake any sleepers explicitely.
2084 memcg_oom_recover(memcg
);
2087 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2089 /* Give chance to dying process */
2090 schedule_timeout_uninterruptible(1);
2095 * Currently used to update mapped file statistics, but the routine can be
2096 * generalized to update other statistics as well.
2098 * Notes: Race condition
2100 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2101 * it tends to be costly. But considering some conditions, we doesn't need
2102 * to do so _always_.
2104 * Considering "charge", lock_page_cgroup() is not required because all
2105 * file-stat operations happen after a page is attached to radix-tree. There
2106 * are no race with "charge".
2108 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2109 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2110 * if there are race with "uncharge". Statistics itself is properly handled
2113 * Considering "move", this is an only case we see a race. To make the race
2114 * small, we check mm->moving_account and detect there are possibility of race
2115 * If there is, we take a lock.
2118 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2119 bool *locked
, unsigned long *flags
)
2121 struct mem_cgroup
*memcg
;
2122 struct page_cgroup
*pc
;
2124 pc
= lookup_page_cgroup(page
);
2126 memcg
= pc
->mem_cgroup
;
2127 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2130 * If this memory cgroup is not under account moving, we don't
2131 * need to take move_lock_mem_cgroup(). Because we already hold
2132 * rcu_read_lock(), any calls to move_account will be delayed until
2133 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2135 if (!mem_cgroup_stolen(memcg
))
2138 move_lock_mem_cgroup(memcg
, flags
);
2139 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2140 move_unlock_mem_cgroup(memcg
, flags
);
2146 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2148 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2151 * It's guaranteed that pc->mem_cgroup never changes while
2152 * lock is held because a routine modifies pc->mem_cgroup
2153 * should take move_lock_mem_cgroup().
2155 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2158 void mem_cgroup_update_page_stat(struct page
*page
,
2159 enum mem_cgroup_page_stat_item idx
, int val
)
2161 struct mem_cgroup
*memcg
;
2162 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2163 unsigned long uninitialized_var(flags
);
2165 if (mem_cgroup_disabled())
2168 memcg
= pc
->mem_cgroup
;
2169 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2173 case MEMCG_NR_FILE_MAPPED
:
2174 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2180 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2184 * size of first charge trial. "32" comes from vmscan.c's magic value.
2185 * TODO: maybe necessary to use big numbers in big irons.
2187 #define CHARGE_BATCH 32U
2188 struct memcg_stock_pcp
{
2189 struct mem_cgroup
*cached
; /* this never be root cgroup */
2190 unsigned int nr_pages
;
2191 struct work_struct work
;
2192 unsigned long flags
;
2193 #define FLUSHING_CACHED_CHARGE 0
2195 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2196 static DEFINE_MUTEX(percpu_charge_mutex
);
2199 * consume_stock: Try to consume stocked charge on this cpu.
2200 * @memcg: memcg to consume from.
2201 * @nr_pages: how many pages to charge.
2203 * The charges will only happen if @memcg matches the current cpu's memcg
2204 * stock, and at least @nr_pages are available in that stock. Failure to
2205 * service an allocation will refill the stock.
2207 * returns true if successful, false otherwise.
2209 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2211 struct memcg_stock_pcp
*stock
;
2214 if (nr_pages
> CHARGE_BATCH
)
2217 stock
= &get_cpu_var(memcg_stock
);
2218 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2219 stock
->nr_pages
-= nr_pages
;
2220 else /* need to call res_counter_charge */
2222 put_cpu_var(memcg_stock
);
2227 * Returns stocks cached in percpu to res_counter and reset cached information.
2229 static void drain_stock(struct memcg_stock_pcp
*stock
)
2231 struct mem_cgroup
*old
= stock
->cached
;
2233 if (stock
->nr_pages
) {
2234 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2236 res_counter_uncharge(&old
->res
, bytes
);
2237 if (do_swap_account
)
2238 res_counter_uncharge(&old
->memsw
, bytes
);
2239 stock
->nr_pages
= 0;
2241 stock
->cached
= NULL
;
2245 * This must be called under preempt disabled or must be called by
2246 * a thread which is pinned to local cpu.
2248 static void drain_local_stock(struct work_struct
*dummy
)
2250 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2252 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2255 static void __init
memcg_stock_init(void)
2259 for_each_possible_cpu(cpu
) {
2260 struct memcg_stock_pcp
*stock
=
2261 &per_cpu(memcg_stock
, cpu
);
2262 INIT_WORK(&stock
->work
, drain_local_stock
);
2267 * Cache charges(val) which is from res_counter, to local per_cpu area.
2268 * This will be consumed by consume_stock() function, later.
2270 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2272 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2274 if (stock
->cached
!= memcg
) { /* reset if necessary */
2276 stock
->cached
= memcg
;
2278 stock
->nr_pages
+= nr_pages
;
2279 put_cpu_var(memcg_stock
);
2283 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2284 * of the hierarchy under it. sync flag says whether we should block
2285 * until the work is done.
2287 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2291 /* Notify other cpus that system-wide "drain" is running */
2294 for_each_online_cpu(cpu
) {
2295 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2296 struct mem_cgroup
*memcg
;
2298 memcg
= stock
->cached
;
2299 if (!memcg
|| !stock
->nr_pages
)
2301 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2303 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2305 drain_local_stock(&stock
->work
);
2307 schedule_work_on(cpu
, &stock
->work
);
2315 for_each_online_cpu(cpu
) {
2316 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2317 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2318 flush_work(&stock
->work
);
2325 * Tries to drain stocked charges in other cpus. This function is asynchronous
2326 * and just put a work per cpu for draining localy on each cpu. Caller can
2327 * expects some charges will be back to res_counter later but cannot wait for
2330 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2333 * If someone calls draining, avoid adding more kworker runs.
2335 if (!mutex_trylock(&percpu_charge_mutex
))
2337 drain_all_stock(root_memcg
, false);
2338 mutex_unlock(&percpu_charge_mutex
);
2341 /* This is a synchronous drain interface. */
2342 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2344 /* called when force_empty is called */
2345 mutex_lock(&percpu_charge_mutex
);
2346 drain_all_stock(root_memcg
, true);
2347 mutex_unlock(&percpu_charge_mutex
);
2351 * This function drains percpu counter value from DEAD cpu and
2352 * move it to local cpu. Note that this function can be preempted.
2354 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2358 spin_lock(&memcg
->pcp_counter_lock
);
2359 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2360 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2362 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2363 memcg
->nocpu_base
.count
[i
] += x
;
2365 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2366 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2368 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2369 memcg
->nocpu_base
.events
[i
] += x
;
2371 spin_unlock(&memcg
->pcp_counter_lock
);
2374 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2375 unsigned long action
,
2378 int cpu
= (unsigned long)hcpu
;
2379 struct memcg_stock_pcp
*stock
;
2380 struct mem_cgroup
*iter
;
2382 if (action
== CPU_ONLINE
)
2385 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2388 for_each_mem_cgroup(iter
)
2389 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2391 stock
= &per_cpu(memcg_stock
, cpu
);
2397 /* See __mem_cgroup_try_charge() for details */
2399 CHARGE_OK
, /* success */
2400 CHARGE_RETRY
, /* need to retry but retry is not bad */
2401 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2402 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2403 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
2406 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2407 unsigned int nr_pages
, unsigned int min_pages
,
2410 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2411 struct mem_cgroup
*mem_over_limit
;
2412 struct res_counter
*fail_res
;
2413 unsigned long flags
= 0;
2416 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2419 if (!do_swap_account
)
2421 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2425 res_counter_uncharge(&memcg
->res
, csize
);
2426 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2427 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2429 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2431 * Never reclaim on behalf of optional batching, retry with a
2432 * single page instead.
2434 if (nr_pages
> min_pages
)
2435 return CHARGE_RETRY
;
2437 if (!(gfp_mask
& __GFP_WAIT
))
2438 return CHARGE_WOULDBLOCK
;
2440 if (gfp_mask
& __GFP_NORETRY
)
2441 return CHARGE_NOMEM
;
2443 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2444 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2445 return CHARGE_RETRY
;
2447 * Even though the limit is exceeded at this point, reclaim
2448 * may have been able to free some pages. Retry the charge
2449 * before killing the task.
2451 * Only for regular pages, though: huge pages are rather
2452 * unlikely to succeed so close to the limit, and we fall back
2453 * to regular pages anyway in case of failure.
2455 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2456 return CHARGE_RETRY
;
2459 * At task move, charge accounts can be doubly counted. So, it's
2460 * better to wait until the end of task_move if something is going on.
2462 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2463 return CHARGE_RETRY
;
2465 /* If we don't need to call oom-killer at el, return immediately */
2466 if (!oom_check
|| !current
->memcg_oom
.may_oom
)
2467 return CHARGE_NOMEM
;
2469 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
, get_order(csize
)))
2470 return CHARGE_OOM_DIE
;
2472 return CHARGE_RETRY
;
2476 * __mem_cgroup_try_charge() does
2477 * 1. detect memcg to be charged against from passed *mm and *ptr,
2478 * 2. update res_counter
2479 * 3. call memory reclaim if necessary.
2481 * In some special case, if the task is fatal, fatal_signal_pending() or
2482 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2483 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2484 * as possible without any hazards. 2: all pages should have a valid
2485 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2486 * pointer, that is treated as a charge to root_mem_cgroup.
2488 * So __mem_cgroup_try_charge() will return
2489 * 0 ... on success, filling *ptr with a valid memcg pointer.
2490 * -ENOMEM ... charge failure because of resource limits.
2491 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2493 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2494 * the oom-killer can be invoked.
2496 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2498 unsigned int nr_pages
,
2499 struct mem_cgroup
**ptr
,
2502 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2503 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2504 struct mem_cgroup
*memcg
= NULL
;
2508 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2509 * in system level. So, allow to go ahead dying process in addition to
2512 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2513 || fatal_signal_pending(current
)))
2517 * We always charge the cgroup the mm_struct belongs to.
2518 * The mm_struct's mem_cgroup changes on task migration if the
2519 * thread group leader migrates. It's possible that mm is not
2520 * set, if so charge the root memcg (happens for pagecache usage).
2523 *ptr
= root_mem_cgroup
;
2525 if (*ptr
) { /* css should be a valid one */
2527 if (mem_cgroup_is_root(memcg
))
2529 if (consume_stock(memcg
, nr_pages
))
2531 css_get(&memcg
->css
);
2533 struct task_struct
*p
;
2536 p
= rcu_dereference(mm
->owner
);
2538 * Because we don't have task_lock(), "p" can exit.
2539 * In that case, "memcg" can point to root or p can be NULL with
2540 * race with swapoff. Then, we have small risk of mis-accouning.
2541 * But such kind of mis-account by race always happens because
2542 * we don't have cgroup_mutex(). It's overkill and we allo that
2544 * (*) swapoff at el will charge against mm-struct not against
2545 * task-struct. So, mm->owner can be NULL.
2547 memcg
= mem_cgroup_from_task(p
);
2549 memcg
= root_mem_cgroup
;
2550 if (mem_cgroup_is_root(memcg
)) {
2554 if (consume_stock(memcg
, nr_pages
)) {
2556 * It seems dagerous to access memcg without css_get().
2557 * But considering how consume_stok works, it's not
2558 * necessary. If consume_stock success, some charges
2559 * from this memcg are cached on this cpu. So, we
2560 * don't need to call css_get()/css_tryget() before
2561 * calling consume_stock().
2566 /* after here, we may be blocked. we need to get refcnt */
2567 if (!css_tryget(&memcg
->css
)) {
2577 /* If killed, bypass charge */
2578 if (fatal_signal_pending(current
)) {
2579 css_put(&memcg
->css
);
2584 if (oom
&& !nr_oom_retries
) {
2586 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2589 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
, nr_pages
,
2594 case CHARGE_RETRY
: /* not in OOM situation but retry */
2596 css_put(&memcg
->css
);
2599 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2600 css_put(&memcg
->css
);
2602 case CHARGE_NOMEM
: /* OOM routine works */
2604 css_put(&memcg
->css
);
2607 /* If oom, we never return -ENOMEM */
2610 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2611 css_put(&memcg
->css
);
2614 } while (ret
!= CHARGE_OK
);
2616 if (batch
> nr_pages
)
2617 refill_stock(memcg
, batch
- nr_pages
);
2618 css_put(&memcg
->css
);
2626 *ptr
= root_mem_cgroup
;
2631 * Somemtimes we have to undo a charge we got by try_charge().
2632 * This function is for that and do uncharge, put css's refcnt.
2633 * gotten by try_charge().
2635 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2636 unsigned int nr_pages
)
2638 if (!mem_cgroup_is_root(memcg
)) {
2639 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2641 res_counter_uncharge(&memcg
->res
, bytes
);
2642 if (do_swap_account
)
2643 res_counter_uncharge(&memcg
->memsw
, bytes
);
2648 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2649 * This is useful when moving usage to parent cgroup.
2651 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2652 unsigned int nr_pages
)
2654 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2656 if (mem_cgroup_is_root(memcg
))
2659 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2660 if (do_swap_account
)
2661 res_counter_uncharge_until(&memcg
->memsw
,
2662 memcg
->memsw
.parent
, bytes
);
2666 * A helper function to get mem_cgroup from ID. must be called under
2667 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2668 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2669 * called against removed memcg.)
2671 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2673 struct cgroup_subsys_state
*css
;
2675 /* ID 0 is unused ID */
2678 css
= css_lookup(&mem_cgroup_subsys
, id
);
2681 return mem_cgroup_from_css(css
);
2684 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2686 struct mem_cgroup
*memcg
= NULL
;
2687 struct page_cgroup
*pc
;
2691 VM_BUG_ON(!PageLocked(page
));
2693 pc
= lookup_page_cgroup(page
);
2694 lock_page_cgroup(pc
);
2695 if (PageCgroupUsed(pc
)) {
2696 memcg
= pc
->mem_cgroup
;
2697 if (memcg
&& !css_tryget(&memcg
->css
))
2699 } else if (PageSwapCache(page
)) {
2700 ent
.val
= page_private(page
);
2701 id
= lookup_swap_cgroup_id(ent
);
2703 memcg
= mem_cgroup_lookup(id
);
2704 if (memcg
&& !css_tryget(&memcg
->css
))
2708 unlock_page_cgroup(pc
);
2712 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2714 unsigned int nr_pages
,
2715 enum charge_type ctype
,
2718 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2719 struct zone
*uninitialized_var(zone
);
2720 struct lruvec
*lruvec
;
2721 bool was_on_lru
= false;
2724 lock_page_cgroup(pc
);
2725 VM_BUG_ON(PageCgroupUsed(pc
));
2727 * we don't need page_cgroup_lock about tail pages, becase they are not
2728 * accessed by any other context at this point.
2732 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2733 * may already be on some other mem_cgroup's LRU. Take care of it.
2736 zone
= page_zone(page
);
2737 spin_lock_irq(&zone
->lru_lock
);
2738 if (PageLRU(page
)) {
2739 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2741 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2746 pc
->mem_cgroup
= memcg
;
2748 * We access a page_cgroup asynchronously without lock_page_cgroup().
2749 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2750 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2751 * before USED bit, we need memory barrier here.
2752 * See mem_cgroup_add_lru_list(), etc.
2755 SetPageCgroupUsed(pc
);
2759 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2760 VM_BUG_ON(PageLRU(page
));
2762 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2764 spin_unlock_irq(&zone
->lru_lock
);
2767 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2772 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2773 unlock_page_cgroup(pc
);
2776 * "charge_statistics" updated event counter.
2778 memcg_check_events(memcg
, page
);
2781 static DEFINE_MUTEX(set_limit_mutex
);
2783 #ifdef CONFIG_MEMCG_KMEM
2784 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2786 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2787 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2791 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2792 * in the memcg_cache_params struct.
2794 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2796 struct kmem_cache
*cachep
;
2798 VM_BUG_ON(p
->is_root_cache
);
2799 cachep
= p
->root_cache
;
2800 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2803 #ifdef CONFIG_SLABINFO
2804 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2805 struct cftype
*cft
, struct seq_file
*m
)
2807 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2808 struct memcg_cache_params
*params
;
2810 if (!memcg_can_account_kmem(memcg
))
2813 print_slabinfo_header(m
);
2815 mutex_lock(&memcg
->slab_caches_mutex
);
2816 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2817 cache_show(memcg_params_to_cache(params
), m
);
2818 mutex_unlock(&memcg
->slab_caches_mutex
);
2824 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2826 struct res_counter
*fail_res
;
2827 struct mem_cgroup
*_memcg
;
2831 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2836 * Conditions under which we can wait for the oom_killer. Those are
2837 * the same conditions tested by the core page allocator
2839 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2842 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2845 if (ret
== -EINTR
) {
2847 * __mem_cgroup_try_charge() chosed to bypass to root due to
2848 * OOM kill or fatal signal. Since our only options are to
2849 * either fail the allocation or charge it to this cgroup, do
2850 * it as a temporary condition. But we can't fail. From a
2851 * kmem/slab perspective, the cache has already been selected,
2852 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2855 * This condition will only trigger if the task entered
2856 * memcg_charge_kmem in a sane state, but was OOM-killed during
2857 * __mem_cgroup_try_charge() above. Tasks that were already
2858 * dying when the allocation triggers should have been already
2859 * directed to the root cgroup in memcontrol.h
2861 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2862 if (do_swap_account
)
2863 res_counter_charge_nofail(&memcg
->memsw
, size
,
2867 res_counter_uncharge(&memcg
->kmem
, size
);
2872 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2874 res_counter_uncharge(&memcg
->res
, size
);
2875 if (do_swap_account
)
2876 res_counter_uncharge(&memcg
->memsw
, size
);
2879 if (res_counter_uncharge(&memcg
->kmem
, size
))
2883 * Releases a reference taken in kmem_cgroup_css_offline in case
2884 * this last uncharge is racing with the offlining code or it is
2885 * outliving the memcg existence.
2887 * The memory barrier imposed by test&clear is paired with the
2888 * explicit one in memcg_kmem_mark_dead().
2890 if (memcg_kmem_test_and_clear_dead(memcg
))
2891 css_put(&memcg
->css
);
2894 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2899 mutex_lock(&memcg
->slab_caches_mutex
);
2900 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2901 mutex_unlock(&memcg
->slab_caches_mutex
);
2905 * helper for acessing a memcg's index. It will be used as an index in the
2906 * child cache array in kmem_cache, and also to derive its name. This function
2907 * will return -1 when this is not a kmem-limited memcg.
2909 int memcg_cache_id(struct mem_cgroup
*memcg
)
2911 return memcg
? memcg
->kmemcg_id
: -1;
2915 * This ends up being protected by the set_limit mutex, during normal
2916 * operation, because that is its main call site.
2918 * But when we create a new cache, we can call this as well if its parent
2919 * is kmem-limited. That will have to hold set_limit_mutex as well.
2921 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2925 num
= ida_simple_get(&kmem_limited_groups
,
2926 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2930 * After this point, kmem_accounted (that we test atomically in
2931 * the beginning of this conditional), is no longer 0. This
2932 * guarantees only one process will set the following boolean
2933 * to true. We don't need test_and_set because we're protected
2934 * by the set_limit_mutex anyway.
2936 memcg_kmem_set_activated(memcg
);
2938 ret
= memcg_update_all_caches(num
+1);
2940 ida_simple_remove(&kmem_limited_groups
, num
);
2941 memcg_kmem_clear_activated(memcg
);
2945 memcg
->kmemcg_id
= num
;
2946 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
2947 mutex_init(&memcg
->slab_caches_mutex
);
2951 static size_t memcg_caches_array_size(int num_groups
)
2954 if (num_groups
<= 0)
2957 size
= 2 * num_groups
;
2958 if (size
< MEMCG_CACHES_MIN_SIZE
)
2959 size
= MEMCG_CACHES_MIN_SIZE
;
2960 else if (size
> MEMCG_CACHES_MAX_SIZE
)
2961 size
= MEMCG_CACHES_MAX_SIZE
;
2967 * We should update the current array size iff all caches updates succeed. This
2968 * can only be done from the slab side. The slab mutex needs to be held when
2971 void memcg_update_array_size(int num
)
2973 if (num
> memcg_limited_groups_array_size
)
2974 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
2977 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
2979 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
2981 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
2983 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
2985 if (num_groups
> memcg_limited_groups_array_size
) {
2987 ssize_t size
= memcg_caches_array_size(num_groups
);
2989 size
*= sizeof(void *);
2990 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
2992 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
2993 if (!s
->memcg_params
) {
2994 s
->memcg_params
= cur_params
;
2998 s
->memcg_params
->is_root_cache
= true;
3001 * There is the chance it will be bigger than
3002 * memcg_limited_groups_array_size, if we failed an allocation
3003 * in a cache, in which case all caches updated before it, will
3004 * have a bigger array.
3006 * But if that is the case, the data after
3007 * memcg_limited_groups_array_size is certainly unused
3009 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3010 if (!cur_params
->memcg_caches
[i
])
3012 s
->memcg_params
->memcg_caches
[i
] =
3013 cur_params
->memcg_caches
[i
];
3017 * Ideally, we would wait until all caches succeed, and only
3018 * then free the old one. But this is not worth the extra
3019 * pointer per-cache we'd have to have for this.
3021 * It is not a big deal if some caches are left with a size
3022 * bigger than the others. And all updates will reset this
3030 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3031 struct kmem_cache
*root_cache
)
3035 if (!memcg_kmem_enabled())
3039 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3040 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3042 size
= sizeof(struct memcg_cache_params
);
3044 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3045 if (!s
->memcg_params
)
3049 s
->memcg_params
->memcg
= memcg
;
3050 s
->memcg_params
->root_cache
= root_cache
;
3051 INIT_WORK(&s
->memcg_params
->destroy
,
3052 kmem_cache_destroy_work_func
);
3054 s
->memcg_params
->is_root_cache
= true;
3059 void memcg_release_cache(struct kmem_cache
*s
)
3061 struct kmem_cache
*root
;
3062 struct mem_cgroup
*memcg
;
3066 * This happens, for instance, when a root cache goes away before we
3069 if (!s
->memcg_params
)
3072 if (s
->memcg_params
->is_root_cache
)
3075 memcg
= s
->memcg_params
->memcg
;
3076 id
= memcg_cache_id(memcg
);
3078 root
= s
->memcg_params
->root_cache
;
3079 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3081 mutex_lock(&memcg
->slab_caches_mutex
);
3082 list_del(&s
->memcg_params
->list
);
3083 mutex_unlock(&memcg
->slab_caches_mutex
);
3085 css_put(&memcg
->css
);
3087 kfree(s
->memcg_params
);
3091 * During the creation a new cache, we need to disable our accounting mechanism
3092 * altogether. This is true even if we are not creating, but rather just
3093 * enqueing new caches to be created.
3095 * This is because that process will trigger allocations; some visible, like
3096 * explicit kmallocs to auxiliary data structures, name strings and internal
3097 * cache structures; some well concealed, like INIT_WORK() that can allocate
3098 * objects during debug.
3100 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3101 * to it. This may not be a bounded recursion: since the first cache creation
3102 * failed to complete (waiting on the allocation), we'll just try to create the
3103 * cache again, failing at the same point.
3105 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3106 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3107 * inside the following two functions.
3109 static inline void memcg_stop_kmem_account(void)
3111 VM_BUG_ON(!current
->mm
);
3112 current
->memcg_kmem_skip_account
++;
3115 static inline void memcg_resume_kmem_account(void)
3117 VM_BUG_ON(!current
->mm
);
3118 current
->memcg_kmem_skip_account
--;
3121 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3123 struct kmem_cache
*cachep
;
3124 struct memcg_cache_params
*p
;
3126 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3128 cachep
= memcg_params_to_cache(p
);
3131 * If we get down to 0 after shrink, we could delete right away.
3132 * However, memcg_release_pages() already puts us back in the workqueue
3133 * in that case. If we proceed deleting, we'll get a dangling
3134 * reference, and removing the object from the workqueue in that case
3135 * is unnecessary complication. We are not a fast path.
3137 * Note that this case is fundamentally different from racing with
3138 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3139 * kmem_cache_shrink, not only we would be reinserting a dead cache
3140 * into the queue, but doing so from inside the worker racing to
3143 * So if we aren't down to zero, we'll just schedule a worker and try
3146 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3147 kmem_cache_shrink(cachep
);
3148 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3151 kmem_cache_destroy(cachep
);
3154 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3156 if (!cachep
->memcg_params
->dead
)
3160 * There are many ways in which we can get here.
3162 * We can get to a memory-pressure situation while the delayed work is
3163 * still pending to run. The vmscan shrinkers can then release all
3164 * cache memory and get us to destruction. If this is the case, we'll
3165 * be executed twice, which is a bug (the second time will execute over
3166 * bogus data). In this case, cancelling the work should be fine.
3168 * But we can also get here from the worker itself, if
3169 * kmem_cache_shrink is enough to shake all the remaining objects and
3170 * get the page count to 0. In this case, we'll deadlock if we try to
3171 * cancel the work (the worker runs with an internal lock held, which
3172 * is the same lock we would hold for cancel_work_sync().)
3174 * Since we can't possibly know who got us here, just refrain from
3175 * running if there is already work pending
3177 if (work_pending(&cachep
->memcg_params
->destroy
))
3180 * We have to defer the actual destroying to a workqueue, because
3181 * we might currently be in a context that cannot sleep.
3183 schedule_work(&cachep
->memcg_params
->destroy
);
3187 * This lock protects updaters, not readers. We want readers to be as fast as
3188 * they can, and they will either see NULL or a valid cache value. Our model
3189 * allow them to see NULL, in which case the root memcg will be selected.
3191 * We need this lock because multiple allocations to the same cache from a non
3192 * will span more than one worker. Only one of them can create the cache.
3194 static DEFINE_MUTEX(memcg_cache_mutex
);
3197 * Called with memcg_cache_mutex held
3199 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3200 struct kmem_cache
*s
)
3202 struct kmem_cache
*new;
3203 static char *tmp_name
= NULL
;
3205 lockdep_assert_held(&memcg_cache_mutex
);
3208 * kmem_cache_create_memcg duplicates the given name and
3209 * cgroup_name for this name requires RCU context.
3210 * This static temporary buffer is used to prevent from
3211 * pointless shortliving allocation.
3214 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3220 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3221 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3224 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3225 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3228 new->allocflags
|= __GFP_KMEMCG
;
3233 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3234 struct kmem_cache
*cachep
)
3236 struct kmem_cache
*new_cachep
;
3239 BUG_ON(!memcg_can_account_kmem(memcg
));
3241 idx
= memcg_cache_id(memcg
);
3243 mutex_lock(&memcg_cache_mutex
);
3244 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3246 css_put(&memcg
->css
);
3250 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3251 if (new_cachep
== NULL
) {
3252 new_cachep
= cachep
;
3253 css_put(&memcg
->css
);
3257 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3259 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3261 * the readers won't lock, make sure everybody sees the updated value,
3262 * so they won't put stuff in the queue again for no reason
3266 mutex_unlock(&memcg_cache_mutex
);
3270 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3272 struct kmem_cache
*c
;
3275 if (!s
->memcg_params
)
3277 if (!s
->memcg_params
->is_root_cache
)
3281 * If the cache is being destroyed, we trust that there is no one else
3282 * requesting objects from it. Even if there are, the sanity checks in
3283 * kmem_cache_destroy should caught this ill-case.
3285 * Still, we don't want anyone else freeing memcg_caches under our
3286 * noses, which can happen if a new memcg comes to life. As usual,
3287 * we'll take the set_limit_mutex to protect ourselves against this.
3289 mutex_lock(&set_limit_mutex
);
3290 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3291 c
= s
->memcg_params
->memcg_caches
[i
];
3296 * We will now manually delete the caches, so to avoid races
3297 * we need to cancel all pending destruction workers and
3298 * proceed with destruction ourselves.
3300 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3301 * and that could spawn the workers again: it is likely that
3302 * the cache still have active pages until this very moment.
3303 * This would lead us back to mem_cgroup_destroy_cache.
3305 * But that will not execute at all if the "dead" flag is not
3306 * set, so flip it down to guarantee we are in control.
3308 c
->memcg_params
->dead
= false;
3309 cancel_work_sync(&c
->memcg_params
->destroy
);
3310 kmem_cache_destroy(c
);
3312 mutex_unlock(&set_limit_mutex
);
3315 struct create_work
{
3316 struct mem_cgroup
*memcg
;
3317 struct kmem_cache
*cachep
;
3318 struct work_struct work
;
3321 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3323 struct kmem_cache
*cachep
;
3324 struct memcg_cache_params
*params
;
3326 if (!memcg_kmem_is_active(memcg
))
3329 mutex_lock(&memcg
->slab_caches_mutex
);
3330 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3331 cachep
= memcg_params_to_cache(params
);
3332 cachep
->memcg_params
->dead
= true;
3333 schedule_work(&cachep
->memcg_params
->destroy
);
3335 mutex_unlock(&memcg
->slab_caches_mutex
);
3338 static void memcg_create_cache_work_func(struct work_struct
*w
)
3340 struct create_work
*cw
;
3342 cw
= container_of(w
, struct create_work
, work
);
3343 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3348 * Enqueue the creation of a per-memcg kmem_cache.
3350 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3351 struct kmem_cache
*cachep
)
3353 struct create_work
*cw
;
3355 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3357 css_put(&memcg
->css
);
3362 cw
->cachep
= cachep
;
3364 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3365 schedule_work(&cw
->work
);
3368 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3369 struct kmem_cache
*cachep
)
3372 * We need to stop accounting when we kmalloc, because if the
3373 * corresponding kmalloc cache is not yet created, the first allocation
3374 * in __memcg_create_cache_enqueue will recurse.
3376 * However, it is better to enclose the whole function. Depending on
3377 * the debugging options enabled, INIT_WORK(), for instance, can
3378 * trigger an allocation. This too, will make us recurse. Because at
3379 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3380 * the safest choice is to do it like this, wrapping the whole function.
3382 memcg_stop_kmem_account();
3383 __memcg_create_cache_enqueue(memcg
, cachep
);
3384 memcg_resume_kmem_account();
3387 * Return the kmem_cache we're supposed to use for a slab allocation.
3388 * We try to use the current memcg's version of the cache.
3390 * If the cache does not exist yet, if we are the first user of it,
3391 * we either create it immediately, if possible, or create it asynchronously
3393 * In the latter case, we will let the current allocation go through with
3394 * the original cache.
3396 * Can't be called in interrupt context or from kernel threads.
3397 * This function needs to be called with rcu_read_lock() held.
3399 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3402 struct mem_cgroup
*memcg
;
3405 VM_BUG_ON(!cachep
->memcg_params
);
3406 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3408 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3412 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3414 if (!memcg_can_account_kmem(memcg
))
3417 idx
= memcg_cache_id(memcg
);
3420 * barrier to mare sure we're always seeing the up to date value. The
3421 * code updating memcg_caches will issue a write barrier to match this.
3423 read_barrier_depends();
3424 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3425 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3429 /* The corresponding put will be done in the workqueue. */
3430 if (!css_tryget(&memcg
->css
))
3435 * If we are in a safe context (can wait, and not in interrupt
3436 * context), we could be be predictable and return right away.
3437 * This would guarantee that the allocation being performed
3438 * already belongs in the new cache.
3440 * However, there are some clashes that can arrive from locking.
3441 * For instance, because we acquire the slab_mutex while doing
3442 * kmem_cache_dup, this means no further allocation could happen
3443 * with the slab_mutex held.
3445 * Also, because cache creation issue get_online_cpus(), this
3446 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3447 * that ends up reversed during cpu hotplug. (cpuset allocates
3448 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3449 * better to defer everything.
3451 memcg_create_cache_enqueue(memcg
, cachep
);
3457 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3460 * We need to verify if the allocation against current->mm->owner's memcg is
3461 * possible for the given order. But the page is not allocated yet, so we'll
3462 * need a further commit step to do the final arrangements.
3464 * It is possible for the task to switch cgroups in this mean time, so at
3465 * commit time, we can't rely on task conversion any longer. We'll then use
3466 * the handle argument to return to the caller which cgroup we should commit
3467 * against. We could also return the memcg directly and avoid the pointer
3468 * passing, but a boolean return value gives better semantics considering
3469 * the compiled-out case as well.
3471 * Returning true means the allocation is possible.
3474 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3476 struct mem_cgroup
*memcg
;
3482 * Disabling accounting is only relevant for some specific memcg
3483 * internal allocations. Therefore we would initially not have such
3484 * check here, since direct calls to the page allocator that are marked
3485 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3486 * concerned with cache allocations, and by having this test at
3487 * memcg_kmem_get_cache, we are already able to relay the allocation to
3488 * the root cache and bypass the memcg cache altogether.
3490 * There is one exception, though: the SLUB allocator does not create
3491 * large order caches, but rather service large kmallocs directly from
3492 * the page allocator. Therefore, the following sequence when backed by
3493 * the SLUB allocator:
3495 * memcg_stop_kmem_account();
3496 * kmalloc(<large_number>)
3497 * memcg_resume_kmem_account();
3499 * would effectively ignore the fact that we should skip accounting,
3500 * since it will drive us directly to this function without passing
3501 * through the cache selector memcg_kmem_get_cache. Such large
3502 * allocations are extremely rare but can happen, for instance, for the
3503 * cache arrays. We bring this test here.
3505 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3508 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3511 * very rare case described in mem_cgroup_from_task. Unfortunately there
3512 * isn't much we can do without complicating this too much, and it would
3513 * be gfp-dependent anyway. Just let it go
3515 if (unlikely(!memcg
))
3518 if (!memcg_can_account_kmem(memcg
)) {
3519 css_put(&memcg
->css
);
3523 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3527 css_put(&memcg
->css
);
3531 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3534 struct page_cgroup
*pc
;
3536 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3538 /* The page allocation failed. Revert */
3540 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3544 pc
= lookup_page_cgroup(page
);
3545 lock_page_cgroup(pc
);
3546 pc
->mem_cgroup
= memcg
;
3547 SetPageCgroupUsed(pc
);
3548 unlock_page_cgroup(pc
);
3551 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3553 struct mem_cgroup
*memcg
= NULL
;
3554 struct page_cgroup
*pc
;
3557 pc
= lookup_page_cgroup(page
);
3559 * Fast unlocked return. Theoretically might have changed, have to
3560 * check again after locking.
3562 if (!PageCgroupUsed(pc
))
3565 lock_page_cgroup(pc
);
3566 if (PageCgroupUsed(pc
)) {
3567 memcg
= pc
->mem_cgroup
;
3568 ClearPageCgroupUsed(pc
);
3570 unlock_page_cgroup(pc
);
3573 * We trust that only if there is a memcg associated with the page, it
3574 * is a valid allocation
3579 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3580 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3583 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3586 #endif /* CONFIG_MEMCG_KMEM */
3588 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3590 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3592 * Because tail pages are not marked as "used", set it. We're under
3593 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3594 * charge/uncharge will be never happen and move_account() is done under
3595 * compound_lock(), so we don't have to take care of races.
3597 void mem_cgroup_split_huge_fixup(struct page
*head
)
3599 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3600 struct page_cgroup
*pc
;
3601 struct mem_cgroup
*memcg
;
3604 if (mem_cgroup_disabled())
3607 memcg
= head_pc
->mem_cgroup
;
3608 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3610 pc
->mem_cgroup
= memcg
;
3611 smp_wmb();/* see __commit_charge() */
3612 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3614 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3617 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3620 * mem_cgroup_move_account - move account of the page
3622 * @nr_pages: number of regular pages (>1 for huge pages)
3623 * @pc: page_cgroup of the page.
3624 * @from: mem_cgroup which the page is moved from.
3625 * @to: mem_cgroup which the page is moved to. @from != @to.
3627 * The caller must confirm following.
3628 * - page is not on LRU (isolate_page() is useful.)
3629 * - compound_lock is held when nr_pages > 1
3631 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3634 static int mem_cgroup_move_account(struct page
*page
,
3635 unsigned int nr_pages
,
3636 struct page_cgroup
*pc
,
3637 struct mem_cgroup
*from
,
3638 struct mem_cgroup
*to
)
3640 unsigned long flags
;
3642 bool anon
= PageAnon(page
);
3644 VM_BUG_ON(from
== to
);
3645 VM_BUG_ON(PageLRU(page
));
3647 * The page is isolated from LRU. So, collapse function
3648 * will not handle this page. But page splitting can happen.
3649 * Do this check under compound_page_lock(). The caller should
3653 if (nr_pages
> 1 && !PageTransHuge(page
))
3656 lock_page_cgroup(pc
);
3659 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3662 move_lock_mem_cgroup(from
, &flags
);
3664 if (!anon
&& page_mapped(page
)) {
3665 /* Update mapped_file data for mem_cgroup */
3667 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3668 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3671 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3673 /* caller should have done css_get */
3674 pc
->mem_cgroup
= to
;
3675 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3676 move_unlock_mem_cgroup(from
, &flags
);
3679 unlock_page_cgroup(pc
);
3683 memcg_check_events(to
, page
);
3684 memcg_check_events(from
, page
);
3690 * mem_cgroup_move_parent - moves page to the parent group
3691 * @page: the page to move
3692 * @pc: page_cgroup of the page
3693 * @child: page's cgroup
3695 * move charges to its parent or the root cgroup if the group has no
3696 * parent (aka use_hierarchy==0).
3697 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3698 * mem_cgroup_move_account fails) the failure is always temporary and
3699 * it signals a race with a page removal/uncharge or migration. In the
3700 * first case the page is on the way out and it will vanish from the LRU
3701 * on the next attempt and the call should be retried later.
3702 * Isolation from the LRU fails only if page has been isolated from
3703 * the LRU since we looked at it and that usually means either global
3704 * reclaim or migration going on. The page will either get back to the
3706 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3707 * (!PageCgroupUsed) or moved to a different group. The page will
3708 * disappear in the next attempt.
3710 static int mem_cgroup_move_parent(struct page
*page
,
3711 struct page_cgroup
*pc
,
3712 struct mem_cgroup
*child
)
3714 struct mem_cgroup
*parent
;
3715 unsigned int nr_pages
;
3716 unsigned long uninitialized_var(flags
);
3719 VM_BUG_ON(mem_cgroup_is_root(child
));
3722 if (!get_page_unless_zero(page
))
3724 if (isolate_lru_page(page
))
3727 nr_pages
= hpage_nr_pages(page
);
3729 parent
= parent_mem_cgroup(child
);
3731 * If no parent, move charges to root cgroup.
3734 parent
= root_mem_cgroup
;
3737 VM_BUG_ON(!PageTransHuge(page
));
3738 flags
= compound_lock_irqsave(page
);
3741 ret
= mem_cgroup_move_account(page
, nr_pages
,
3744 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3747 compound_unlock_irqrestore(page
, flags
);
3748 putback_lru_page(page
);
3756 * Charge the memory controller for page usage.
3758 * 0 if the charge was successful
3759 * < 0 if the cgroup is over its limit
3761 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3762 gfp_t gfp_mask
, enum charge_type ctype
)
3764 struct mem_cgroup
*memcg
= NULL
;
3765 unsigned int nr_pages
= 1;
3769 if (PageTransHuge(page
)) {
3770 nr_pages
<<= compound_order(page
);
3771 VM_BUG_ON(!PageTransHuge(page
));
3773 * Never OOM-kill a process for a huge page. The
3774 * fault handler will fall back to regular pages.
3779 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3782 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3786 int mem_cgroup_newpage_charge(struct page
*page
,
3787 struct mm_struct
*mm
, gfp_t gfp_mask
)
3789 if (mem_cgroup_disabled())
3791 VM_BUG_ON(page_mapped(page
));
3792 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3794 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3795 MEM_CGROUP_CHARGE_TYPE_ANON
);
3799 * While swap-in, try_charge -> commit or cancel, the page is locked.
3800 * And when try_charge() successfully returns, one refcnt to memcg without
3801 * struct page_cgroup is acquired. This refcnt will be consumed by
3802 * "commit()" or removed by "cancel()"
3804 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3807 struct mem_cgroup
**memcgp
)
3809 struct mem_cgroup
*memcg
;
3810 struct page_cgroup
*pc
;
3813 pc
= lookup_page_cgroup(page
);
3815 * Every swap fault against a single page tries to charge the
3816 * page, bail as early as possible. shmem_unuse() encounters
3817 * already charged pages, too. The USED bit is protected by
3818 * the page lock, which serializes swap cache removal, which
3819 * in turn serializes uncharging.
3821 if (PageCgroupUsed(pc
))
3823 if (!do_swap_account
)
3825 memcg
= try_get_mem_cgroup_from_page(page
);
3829 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3830 css_put(&memcg
->css
);
3835 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3841 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3842 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3845 if (mem_cgroup_disabled())
3848 * A racing thread's fault, or swapoff, may have already
3849 * updated the pte, and even removed page from swap cache: in
3850 * those cases unuse_pte()'s pte_same() test will fail; but
3851 * there's also a KSM case which does need to charge the page.
3853 if (!PageSwapCache(page
)) {
3856 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3861 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3864 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3866 if (mem_cgroup_disabled())
3870 __mem_cgroup_cancel_charge(memcg
, 1);
3874 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3875 enum charge_type ctype
)
3877 if (mem_cgroup_disabled())
3882 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3884 * Now swap is on-memory. This means this page may be
3885 * counted both as mem and swap....double count.
3886 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3887 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3888 * may call delete_from_swap_cache() before reach here.
3890 if (do_swap_account
&& PageSwapCache(page
)) {
3891 swp_entry_t ent
= {.val
= page_private(page
)};
3892 mem_cgroup_uncharge_swap(ent
);
3896 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3897 struct mem_cgroup
*memcg
)
3899 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3900 MEM_CGROUP_CHARGE_TYPE_ANON
);
3903 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3906 struct mem_cgroup
*memcg
= NULL
;
3907 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3910 if (mem_cgroup_disabled())
3912 if (PageCompound(page
))
3915 if (!PageSwapCache(page
))
3916 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3917 else { /* page is swapcache/shmem */
3918 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3921 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3926 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3927 unsigned int nr_pages
,
3928 const enum charge_type ctype
)
3930 struct memcg_batch_info
*batch
= NULL
;
3931 bool uncharge_memsw
= true;
3933 /* If swapout, usage of swap doesn't decrease */
3934 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3935 uncharge_memsw
= false;
3937 batch
= ¤t
->memcg_batch
;
3939 * In usual, we do css_get() when we remember memcg pointer.
3940 * But in this case, we keep res->usage until end of a series of
3941 * uncharges. Then, it's ok to ignore memcg's refcnt.
3944 batch
->memcg
= memcg
;
3946 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3947 * In those cases, all pages freed continuously can be expected to be in
3948 * the same cgroup and we have chance to coalesce uncharges.
3949 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3950 * because we want to do uncharge as soon as possible.
3953 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
3954 goto direct_uncharge
;
3957 goto direct_uncharge
;
3960 * In typical case, batch->memcg == mem. This means we can
3961 * merge a series of uncharges to an uncharge of res_counter.
3962 * If not, we uncharge res_counter ony by one.
3964 if (batch
->memcg
!= memcg
)
3965 goto direct_uncharge
;
3966 /* remember freed charge and uncharge it later */
3969 batch
->memsw_nr_pages
++;
3972 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
3974 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
3975 if (unlikely(batch
->memcg
!= memcg
))
3976 memcg_oom_recover(memcg
);
3980 * uncharge if !page_mapped(page)
3982 static struct mem_cgroup
*
3983 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
3986 struct mem_cgroup
*memcg
= NULL
;
3987 unsigned int nr_pages
= 1;
3988 struct page_cgroup
*pc
;
3991 if (mem_cgroup_disabled())
3994 if (PageTransHuge(page
)) {
3995 nr_pages
<<= compound_order(page
);
3996 VM_BUG_ON(!PageTransHuge(page
));
3999 * Check if our page_cgroup is valid
4001 pc
= lookup_page_cgroup(page
);
4002 if (unlikely(!PageCgroupUsed(pc
)))
4005 lock_page_cgroup(pc
);
4007 memcg
= pc
->mem_cgroup
;
4009 if (!PageCgroupUsed(pc
))
4012 anon
= PageAnon(page
);
4015 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4017 * Generally PageAnon tells if it's the anon statistics to be
4018 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4019 * used before page reached the stage of being marked PageAnon.
4023 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4024 /* See mem_cgroup_prepare_migration() */
4025 if (page_mapped(page
))
4028 * Pages under migration may not be uncharged. But
4029 * end_migration() /must/ be the one uncharging the
4030 * unused post-migration page and so it has to call
4031 * here with the migration bit still set. See the
4032 * res_counter handling below.
4034 if (!end_migration
&& PageCgroupMigration(pc
))
4037 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4038 if (!PageAnon(page
)) { /* Shared memory */
4039 if (page
->mapping
&& !page_is_file_cache(page
))
4041 } else if (page_mapped(page
)) /* Anon */
4048 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4050 ClearPageCgroupUsed(pc
);
4052 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4053 * freed from LRU. This is safe because uncharged page is expected not
4054 * to be reused (freed soon). Exception is SwapCache, it's handled by
4055 * special functions.
4058 unlock_page_cgroup(pc
);
4060 * even after unlock, we have memcg->res.usage here and this memcg
4061 * will never be freed, so it's safe to call css_get().
4063 memcg_check_events(memcg
, page
);
4064 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4065 mem_cgroup_swap_statistics(memcg
, true);
4066 css_get(&memcg
->css
);
4069 * Migration does not charge the res_counter for the
4070 * replacement page, so leave it alone when phasing out the
4071 * page that is unused after the migration.
4073 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4074 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4079 unlock_page_cgroup(pc
);
4083 void mem_cgroup_uncharge_page(struct page
*page
)
4086 if (page_mapped(page
))
4088 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4090 * If the page is in swap cache, uncharge should be deferred
4091 * to the swap path, which also properly accounts swap usage
4092 * and handles memcg lifetime.
4094 * Note that this check is not stable and reclaim may add the
4095 * page to swap cache at any time after this. However, if the
4096 * page is not in swap cache by the time page->mapcount hits
4097 * 0, there won't be any page table references to the swap
4098 * slot, and reclaim will free it and not actually write the
4101 if (PageSwapCache(page
))
4103 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4106 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4108 VM_BUG_ON(page_mapped(page
));
4109 VM_BUG_ON(page
->mapping
);
4110 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4114 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4115 * In that cases, pages are freed continuously and we can expect pages
4116 * are in the same memcg. All these calls itself limits the number of
4117 * pages freed at once, then uncharge_start/end() is called properly.
4118 * This may be called prural(2) times in a context,
4121 void mem_cgroup_uncharge_start(void)
4123 current
->memcg_batch
.do_batch
++;
4124 /* We can do nest. */
4125 if (current
->memcg_batch
.do_batch
== 1) {
4126 current
->memcg_batch
.memcg
= NULL
;
4127 current
->memcg_batch
.nr_pages
= 0;
4128 current
->memcg_batch
.memsw_nr_pages
= 0;
4132 void mem_cgroup_uncharge_end(void)
4134 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4136 if (!batch
->do_batch
)
4140 if (batch
->do_batch
) /* If stacked, do nothing. */
4146 * This "batch->memcg" is valid without any css_get/put etc...
4147 * bacause we hide charges behind us.
4149 if (batch
->nr_pages
)
4150 res_counter_uncharge(&batch
->memcg
->res
,
4151 batch
->nr_pages
* PAGE_SIZE
);
4152 if (batch
->memsw_nr_pages
)
4153 res_counter_uncharge(&batch
->memcg
->memsw
,
4154 batch
->memsw_nr_pages
* PAGE_SIZE
);
4155 memcg_oom_recover(batch
->memcg
);
4156 /* forget this pointer (for sanity check) */
4157 batch
->memcg
= NULL
;
4162 * called after __delete_from_swap_cache() and drop "page" account.
4163 * memcg information is recorded to swap_cgroup of "ent"
4166 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4168 struct mem_cgroup
*memcg
;
4169 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4171 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4172 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4174 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4177 * record memcg information, if swapout && memcg != NULL,
4178 * css_get() was called in uncharge().
4180 if (do_swap_account
&& swapout
&& memcg
)
4181 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4185 #ifdef CONFIG_MEMCG_SWAP
4187 * called from swap_entry_free(). remove record in swap_cgroup and
4188 * uncharge "memsw" account.
4190 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4192 struct mem_cgroup
*memcg
;
4195 if (!do_swap_account
)
4198 id
= swap_cgroup_record(ent
, 0);
4200 memcg
= mem_cgroup_lookup(id
);
4203 * We uncharge this because swap is freed.
4204 * This memcg can be obsolete one. We avoid calling css_tryget
4206 if (!mem_cgroup_is_root(memcg
))
4207 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4208 mem_cgroup_swap_statistics(memcg
, false);
4209 css_put(&memcg
->css
);
4215 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4216 * @entry: swap entry to be moved
4217 * @from: mem_cgroup which the entry is moved from
4218 * @to: mem_cgroup which the entry is moved to
4220 * It succeeds only when the swap_cgroup's record for this entry is the same
4221 * as the mem_cgroup's id of @from.
4223 * Returns 0 on success, -EINVAL on failure.
4225 * The caller must have charged to @to, IOW, called res_counter_charge() about
4226 * both res and memsw, and called css_get().
4228 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4229 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4231 unsigned short old_id
, new_id
;
4233 old_id
= css_id(&from
->css
);
4234 new_id
= css_id(&to
->css
);
4236 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4237 mem_cgroup_swap_statistics(from
, false);
4238 mem_cgroup_swap_statistics(to
, true);
4240 * This function is only called from task migration context now.
4241 * It postpones res_counter and refcount handling till the end
4242 * of task migration(mem_cgroup_clear_mc()) for performance
4243 * improvement. But we cannot postpone css_get(to) because if
4244 * the process that has been moved to @to does swap-in, the
4245 * refcount of @to might be decreased to 0.
4247 * We are in attach() phase, so the cgroup is guaranteed to be
4248 * alive, so we can just call css_get().
4256 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4257 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4264 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4267 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4268 struct mem_cgroup
**memcgp
)
4270 struct mem_cgroup
*memcg
= NULL
;
4271 unsigned int nr_pages
= 1;
4272 struct page_cgroup
*pc
;
4273 enum charge_type ctype
;
4277 if (mem_cgroup_disabled())
4280 if (PageTransHuge(page
))
4281 nr_pages
<<= compound_order(page
);
4283 pc
= lookup_page_cgroup(page
);
4284 lock_page_cgroup(pc
);
4285 if (PageCgroupUsed(pc
)) {
4286 memcg
= pc
->mem_cgroup
;
4287 css_get(&memcg
->css
);
4289 * At migrating an anonymous page, its mapcount goes down
4290 * to 0 and uncharge() will be called. But, even if it's fully
4291 * unmapped, migration may fail and this page has to be
4292 * charged again. We set MIGRATION flag here and delay uncharge
4293 * until end_migration() is called
4295 * Corner Case Thinking
4297 * When the old page was mapped as Anon and it's unmap-and-freed
4298 * while migration was ongoing.
4299 * If unmap finds the old page, uncharge() of it will be delayed
4300 * until end_migration(). If unmap finds a new page, it's
4301 * uncharged when it make mapcount to be 1->0. If unmap code
4302 * finds swap_migration_entry, the new page will not be mapped
4303 * and end_migration() will find it(mapcount==0).
4306 * When the old page was mapped but migraion fails, the kernel
4307 * remaps it. A charge for it is kept by MIGRATION flag even
4308 * if mapcount goes down to 0. We can do remap successfully
4309 * without charging it again.
4312 * The "old" page is under lock_page() until the end of
4313 * migration, so, the old page itself will not be swapped-out.
4314 * If the new page is swapped out before end_migraton, our
4315 * hook to usual swap-out path will catch the event.
4318 SetPageCgroupMigration(pc
);
4320 unlock_page_cgroup(pc
);
4322 * If the page is not charged at this point,
4330 * We charge new page before it's used/mapped. So, even if unlock_page()
4331 * is called before end_migration, we can catch all events on this new
4332 * page. In the case new page is migrated but not remapped, new page's
4333 * mapcount will be finally 0 and we call uncharge in end_migration().
4336 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4338 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4340 * The page is committed to the memcg, but it's not actually
4341 * charged to the res_counter since we plan on replacing the
4342 * old one and only one page is going to be left afterwards.
4344 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4347 /* remove redundant charge if migration failed*/
4348 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4349 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4351 struct page
*used
, *unused
;
4352 struct page_cgroup
*pc
;
4358 if (!migration_ok
) {
4365 anon
= PageAnon(used
);
4366 __mem_cgroup_uncharge_common(unused
,
4367 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4368 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4370 css_put(&memcg
->css
);
4372 * We disallowed uncharge of pages under migration because mapcount
4373 * of the page goes down to zero, temporarly.
4374 * Clear the flag and check the page should be charged.
4376 pc
= lookup_page_cgroup(oldpage
);
4377 lock_page_cgroup(pc
);
4378 ClearPageCgroupMigration(pc
);
4379 unlock_page_cgroup(pc
);
4382 * If a page is a file cache, radix-tree replacement is very atomic
4383 * and we can skip this check. When it was an Anon page, its mapcount
4384 * goes down to 0. But because we added MIGRATION flage, it's not
4385 * uncharged yet. There are several case but page->mapcount check
4386 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4387 * check. (see prepare_charge() also)
4390 mem_cgroup_uncharge_page(used
);
4394 * At replace page cache, newpage is not under any memcg but it's on
4395 * LRU. So, this function doesn't touch res_counter but handles LRU
4396 * in correct way. Both pages are locked so we cannot race with uncharge.
4398 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4399 struct page
*newpage
)
4401 struct mem_cgroup
*memcg
= NULL
;
4402 struct page_cgroup
*pc
;
4403 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4405 if (mem_cgroup_disabled())
4408 pc
= lookup_page_cgroup(oldpage
);
4409 /* fix accounting on old pages */
4410 lock_page_cgroup(pc
);
4411 if (PageCgroupUsed(pc
)) {
4412 memcg
= pc
->mem_cgroup
;
4413 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4414 ClearPageCgroupUsed(pc
);
4416 unlock_page_cgroup(pc
);
4419 * When called from shmem_replace_page(), in some cases the
4420 * oldpage has already been charged, and in some cases not.
4425 * Even if newpage->mapping was NULL before starting replacement,
4426 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4427 * LRU while we overwrite pc->mem_cgroup.
4429 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4432 #ifdef CONFIG_DEBUG_VM
4433 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4435 struct page_cgroup
*pc
;
4437 pc
= lookup_page_cgroup(page
);
4439 * Can be NULL while feeding pages into the page allocator for
4440 * the first time, i.e. during boot or memory hotplug;
4441 * or when mem_cgroup_disabled().
4443 if (likely(pc
) && PageCgroupUsed(pc
))
4448 bool mem_cgroup_bad_page_check(struct page
*page
)
4450 if (mem_cgroup_disabled())
4453 return lookup_page_cgroup_used(page
) != NULL
;
4456 void mem_cgroup_print_bad_page(struct page
*page
)
4458 struct page_cgroup
*pc
;
4460 pc
= lookup_page_cgroup_used(page
);
4462 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4463 pc
, pc
->flags
, pc
->mem_cgroup
);
4468 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4469 unsigned long long val
)
4472 u64 memswlimit
, memlimit
;
4474 int children
= mem_cgroup_count_children(memcg
);
4475 u64 curusage
, oldusage
;
4479 * For keeping hierarchical_reclaim simple, how long we should retry
4480 * is depends on callers. We set our retry-count to be function
4481 * of # of children which we should visit in this loop.
4483 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4485 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4488 while (retry_count
) {
4489 if (signal_pending(current
)) {
4494 * Rather than hide all in some function, I do this in
4495 * open coded manner. You see what this really does.
4496 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4498 mutex_lock(&set_limit_mutex
);
4499 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4500 if (memswlimit
< val
) {
4502 mutex_unlock(&set_limit_mutex
);
4506 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4510 ret
= res_counter_set_limit(&memcg
->res
, val
);
4512 if (memswlimit
== val
)
4513 memcg
->memsw_is_minimum
= true;
4515 memcg
->memsw_is_minimum
= false;
4517 mutex_unlock(&set_limit_mutex
);
4522 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4523 MEM_CGROUP_RECLAIM_SHRINK
);
4524 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4525 /* Usage is reduced ? */
4526 if (curusage
>= oldusage
)
4529 oldusage
= curusage
;
4531 if (!ret
&& enlarge
)
4532 memcg_oom_recover(memcg
);
4537 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4538 unsigned long long val
)
4541 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4542 int children
= mem_cgroup_count_children(memcg
);
4546 /* see mem_cgroup_resize_res_limit */
4547 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4548 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4549 while (retry_count
) {
4550 if (signal_pending(current
)) {
4555 * Rather than hide all in some function, I do this in
4556 * open coded manner. You see what this really does.
4557 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4559 mutex_lock(&set_limit_mutex
);
4560 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4561 if (memlimit
> val
) {
4563 mutex_unlock(&set_limit_mutex
);
4566 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4567 if (memswlimit
< val
)
4569 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4571 if (memlimit
== val
)
4572 memcg
->memsw_is_minimum
= true;
4574 memcg
->memsw_is_minimum
= false;
4576 mutex_unlock(&set_limit_mutex
);
4581 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4582 MEM_CGROUP_RECLAIM_NOSWAP
|
4583 MEM_CGROUP_RECLAIM_SHRINK
);
4584 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4585 /* Usage is reduced ? */
4586 if (curusage
>= oldusage
)
4589 oldusage
= curusage
;
4591 if (!ret
&& enlarge
)
4592 memcg_oom_recover(memcg
);
4597 * mem_cgroup_force_empty_list - clears LRU of a group
4598 * @memcg: group to clear
4601 * @lru: lru to to clear
4603 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4604 * reclaim the pages page themselves - pages are moved to the parent (or root)
4607 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4608 int node
, int zid
, enum lru_list lru
)
4610 struct lruvec
*lruvec
;
4611 unsigned long flags
;
4612 struct list_head
*list
;
4616 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4617 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4618 list
= &lruvec
->lists
[lru
];
4622 struct page_cgroup
*pc
;
4625 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4626 if (list_empty(list
)) {
4627 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4630 page
= list_entry(list
->prev
, struct page
, lru
);
4632 list_move(&page
->lru
, list
);
4634 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4637 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4639 pc
= lookup_page_cgroup(page
);
4641 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4642 /* found lock contention or "pc" is obsolete. */
4647 } while (!list_empty(list
));
4651 * make mem_cgroup's charge to be 0 if there is no task by moving
4652 * all the charges and pages to the parent.
4653 * This enables deleting this mem_cgroup.
4655 * Caller is responsible for holding css reference on the memcg.
4657 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4663 /* This is for making all *used* pages to be on LRU. */
4664 lru_add_drain_all();
4665 drain_all_stock_sync(memcg
);
4666 mem_cgroup_start_move(memcg
);
4667 for_each_node_state(node
, N_MEMORY
) {
4668 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4671 mem_cgroup_force_empty_list(memcg
,
4676 mem_cgroup_end_move(memcg
);
4677 memcg_oom_recover(memcg
);
4681 * Kernel memory may not necessarily be trackable to a specific
4682 * process. So they are not migrated, and therefore we can't
4683 * expect their value to drop to 0 here.
4684 * Having res filled up with kmem only is enough.
4686 * This is a safety check because mem_cgroup_force_empty_list
4687 * could have raced with mem_cgroup_replace_page_cache callers
4688 * so the lru seemed empty but the page could have been added
4689 * right after the check. RES_USAGE should be safe as we always
4690 * charge before adding to the LRU.
4692 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4693 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4694 } while (usage
> 0);
4698 * This mainly exists for tests during the setting of set of use_hierarchy.
4699 * Since this is the very setting we are changing, the current hierarchy value
4702 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4704 struct cgroup_subsys_state
*pos
;
4706 /* bounce at first found */
4707 css_for_each_child(pos
, &memcg
->css
)
4713 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4714 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4715 * from mem_cgroup_count_children(), in the sense that we don't really care how
4716 * many children we have; we only need to know if we have any. It also counts
4717 * any memcg without hierarchy as infertile.
4719 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4721 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4725 * Reclaims as many pages from the given memcg as possible and moves
4726 * the rest to the parent.
4728 * Caller is responsible for holding css reference for memcg.
4730 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4732 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4733 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4735 /* returns EBUSY if there is a task or if we come here twice. */
4736 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4739 /* we call try-to-free pages for make this cgroup empty */
4740 lru_add_drain_all();
4741 /* try to free all pages in this cgroup */
4742 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4745 if (signal_pending(current
))
4748 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4752 /* maybe some writeback is necessary */
4753 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4758 mem_cgroup_reparent_charges(memcg
);
4763 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4766 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4768 if (mem_cgroup_is_root(memcg
))
4770 return mem_cgroup_force_empty(memcg
);
4773 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4776 return mem_cgroup_from_css(css
)->use_hierarchy
;
4779 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4780 struct cftype
*cft
, u64 val
)
4783 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4784 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4786 mutex_lock(&memcg_create_mutex
);
4788 if (memcg
->use_hierarchy
== val
)
4792 * If parent's use_hierarchy is set, we can't make any modifications
4793 * in the child subtrees. If it is unset, then the change can
4794 * occur, provided the current cgroup has no children.
4796 * For the root cgroup, parent_mem is NULL, we allow value to be
4797 * set if there are no children.
4799 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4800 (val
== 1 || val
== 0)) {
4801 if (!__memcg_has_children(memcg
))
4802 memcg
->use_hierarchy
= val
;
4809 mutex_unlock(&memcg_create_mutex
);
4815 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4816 enum mem_cgroup_stat_index idx
)
4818 struct mem_cgroup
*iter
;
4821 /* Per-cpu values can be negative, use a signed accumulator */
4822 for_each_mem_cgroup_tree(iter
, memcg
)
4823 val
+= mem_cgroup_read_stat(iter
, idx
);
4825 if (val
< 0) /* race ? */
4830 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4834 if (!mem_cgroup_is_root(memcg
)) {
4836 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4838 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4842 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4843 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4845 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4846 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4849 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4851 return val
<< PAGE_SHIFT
;
4854 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
4855 struct cftype
*cft
, struct file
*file
,
4856 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
4858 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4864 type
= MEMFILE_TYPE(cft
->private);
4865 name
= MEMFILE_ATTR(cft
->private);
4869 if (name
== RES_USAGE
)
4870 val
= mem_cgroup_usage(memcg
, false);
4872 val
= res_counter_read_u64(&memcg
->res
, name
);
4875 if (name
== RES_USAGE
)
4876 val
= mem_cgroup_usage(memcg
, true);
4878 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4881 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4887 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4888 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4891 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
4894 #ifdef CONFIG_MEMCG_KMEM
4895 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4897 * For simplicity, we won't allow this to be disabled. It also can't
4898 * be changed if the cgroup has children already, or if tasks had
4901 * If tasks join before we set the limit, a person looking at
4902 * kmem.usage_in_bytes will have no way to determine when it took
4903 * place, which makes the value quite meaningless.
4905 * After it first became limited, changes in the value of the limit are
4906 * of course permitted.
4908 mutex_lock(&memcg_create_mutex
);
4909 mutex_lock(&set_limit_mutex
);
4910 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4911 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
4915 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4918 ret
= memcg_update_cache_sizes(memcg
);
4920 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4923 static_key_slow_inc(&memcg_kmem_enabled_key
);
4925 * setting the active bit after the inc will guarantee no one
4926 * starts accounting before all call sites are patched
4928 memcg_kmem_set_active(memcg
);
4930 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4932 mutex_unlock(&set_limit_mutex
);
4933 mutex_unlock(&memcg_create_mutex
);
4938 #ifdef CONFIG_MEMCG_KMEM
4939 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
4942 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
4946 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
4948 * When that happen, we need to disable the static branch only on those
4949 * memcgs that enabled it. To achieve this, we would be forced to
4950 * complicate the code by keeping track of which memcgs were the ones
4951 * that actually enabled limits, and which ones got it from its
4954 * It is a lot simpler just to do static_key_slow_inc() on every child
4955 * that is accounted.
4957 if (!memcg_kmem_is_active(memcg
))
4961 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4962 * memcg is active already. If the later initialization fails then the
4963 * cgroup core triggers the cleanup so we do not have to do it here.
4965 static_key_slow_inc(&memcg_kmem_enabled_key
);
4967 mutex_lock(&set_limit_mutex
);
4968 memcg_stop_kmem_account();
4969 ret
= memcg_update_cache_sizes(memcg
);
4970 memcg_resume_kmem_account();
4971 mutex_unlock(&set_limit_mutex
);
4975 #endif /* CONFIG_MEMCG_KMEM */
4978 * The user of this function is...
4981 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
4984 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4987 unsigned long long val
;
4990 type
= MEMFILE_TYPE(cft
->private);
4991 name
= MEMFILE_ATTR(cft
->private);
4995 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
4999 /* This function does all necessary parse...reuse it */
5000 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5004 ret
= mem_cgroup_resize_limit(memcg
, val
);
5005 else if (type
== _MEMSWAP
)
5006 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5007 else if (type
== _KMEM
)
5008 ret
= memcg_update_kmem_limit(css
, val
);
5012 case RES_SOFT_LIMIT
:
5013 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5017 * For memsw, soft limits are hard to implement in terms
5018 * of semantics, for now, we support soft limits for
5019 * control without swap
5022 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5027 ret
= -EINVAL
; /* should be BUG() ? */
5033 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5034 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5036 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5038 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5039 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5040 if (!memcg
->use_hierarchy
)
5043 while (css_parent(&memcg
->css
)) {
5044 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5045 if (!memcg
->use_hierarchy
)
5047 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5048 min_limit
= min(min_limit
, tmp
);
5049 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5050 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5053 *mem_limit
= min_limit
;
5054 *memsw_limit
= min_memsw_limit
;
5057 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5059 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5063 type
= MEMFILE_TYPE(event
);
5064 name
= MEMFILE_ATTR(event
);
5069 res_counter_reset_max(&memcg
->res
);
5070 else if (type
== _MEMSWAP
)
5071 res_counter_reset_max(&memcg
->memsw
);
5072 else if (type
== _KMEM
)
5073 res_counter_reset_max(&memcg
->kmem
);
5079 res_counter_reset_failcnt(&memcg
->res
);
5080 else if (type
== _MEMSWAP
)
5081 res_counter_reset_failcnt(&memcg
->memsw
);
5082 else if (type
== _KMEM
)
5083 res_counter_reset_failcnt(&memcg
->kmem
);
5092 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5095 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5099 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5100 struct cftype
*cft
, u64 val
)
5102 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5104 if (val
>= (1 << NR_MOVE_TYPE
))
5108 * No kind of locking is needed in here, because ->can_attach() will
5109 * check this value once in the beginning of the process, and then carry
5110 * on with stale data. This means that changes to this value will only
5111 * affect task migrations starting after the change.
5113 memcg
->move_charge_at_immigrate
= val
;
5117 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5118 struct cftype
*cft
, u64 val
)
5125 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5126 struct cftype
*cft
, struct seq_file
*m
)
5129 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5130 unsigned long node_nr
;
5131 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5133 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5134 seq_printf(m
, "total=%lu", total_nr
);
5135 for_each_node_state(nid
, N_MEMORY
) {
5136 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5137 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5141 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5142 seq_printf(m
, "file=%lu", file_nr
);
5143 for_each_node_state(nid
, N_MEMORY
) {
5144 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5146 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5150 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5151 seq_printf(m
, "anon=%lu", anon_nr
);
5152 for_each_node_state(nid
, N_MEMORY
) {
5153 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5155 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5159 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5160 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5161 for_each_node_state(nid
, N_MEMORY
) {
5162 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5163 BIT(LRU_UNEVICTABLE
));
5164 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5169 #endif /* CONFIG_NUMA */
5171 static inline void mem_cgroup_lru_names_not_uptodate(void)
5173 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5176 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5179 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5180 struct mem_cgroup
*mi
;
5183 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5184 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5186 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5187 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5190 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5191 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5192 mem_cgroup_read_events(memcg
, i
));
5194 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5195 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5196 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5198 /* Hierarchical information */
5200 unsigned long long limit
, memsw_limit
;
5201 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5202 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5203 if (do_swap_account
)
5204 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5208 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5211 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5213 for_each_mem_cgroup_tree(mi
, memcg
)
5214 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5215 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5218 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5219 unsigned long long val
= 0;
5221 for_each_mem_cgroup_tree(mi
, memcg
)
5222 val
+= mem_cgroup_read_events(mi
, i
);
5223 seq_printf(m
, "total_%s %llu\n",
5224 mem_cgroup_events_names
[i
], val
);
5227 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5228 unsigned long long val
= 0;
5230 for_each_mem_cgroup_tree(mi
, memcg
)
5231 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5232 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5235 #ifdef CONFIG_DEBUG_VM
5238 struct mem_cgroup_per_zone
*mz
;
5239 struct zone_reclaim_stat
*rstat
;
5240 unsigned long recent_rotated
[2] = {0, 0};
5241 unsigned long recent_scanned
[2] = {0, 0};
5243 for_each_online_node(nid
)
5244 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5245 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5246 rstat
= &mz
->lruvec
.reclaim_stat
;
5248 recent_rotated
[0] += rstat
->recent_rotated
[0];
5249 recent_rotated
[1] += rstat
->recent_rotated
[1];
5250 recent_scanned
[0] += rstat
->recent_scanned
[0];
5251 recent_scanned
[1] += rstat
->recent_scanned
[1];
5253 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5254 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5255 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5256 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5263 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5266 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5268 return mem_cgroup_swappiness(memcg
);
5271 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5272 struct cftype
*cft
, u64 val
)
5274 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5275 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5277 if (val
> 100 || !parent
)
5280 mutex_lock(&memcg_create_mutex
);
5282 /* If under hierarchy, only empty-root can set this value */
5283 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5284 mutex_unlock(&memcg_create_mutex
);
5288 memcg
->swappiness
= val
;
5290 mutex_unlock(&memcg_create_mutex
);
5295 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5297 struct mem_cgroup_threshold_ary
*t
;
5303 t
= rcu_dereference(memcg
->thresholds
.primary
);
5305 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5310 usage
= mem_cgroup_usage(memcg
, swap
);
5313 * current_threshold points to threshold just below or equal to usage.
5314 * If it's not true, a threshold was crossed after last
5315 * call of __mem_cgroup_threshold().
5317 i
= t
->current_threshold
;
5320 * Iterate backward over array of thresholds starting from
5321 * current_threshold and check if a threshold is crossed.
5322 * If none of thresholds below usage is crossed, we read
5323 * only one element of the array here.
5325 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5326 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5328 /* i = current_threshold + 1 */
5332 * Iterate forward over array of thresholds starting from
5333 * current_threshold+1 and check if a threshold is crossed.
5334 * If none of thresholds above usage is crossed, we read
5335 * only one element of the array here.
5337 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5338 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5340 /* Update current_threshold */
5341 t
->current_threshold
= i
- 1;
5346 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5349 __mem_cgroup_threshold(memcg
, false);
5350 if (do_swap_account
)
5351 __mem_cgroup_threshold(memcg
, true);
5353 memcg
= parent_mem_cgroup(memcg
);
5357 static int compare_thresholds(const void *a
, const void *b
)
5359 const struct mem_cgroup_threshold
*_a
= a
;
5360 const struct mem_cgroup_threshold
*_b
= b
;
5362 if (_a
->threshold
> _b
->threshold
)
5365 if (_a
->threshold
< _b
->threshold
)
5371 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5373 struct mem_cgroup_eventfd_list
*ev
;
5375 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5376 eventfd_signal(ev
->eventfd
, 1);
5380 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5382 struct mem_cgroup
*iter
;
5384 for_each_mem_cgroup_tree(iter
, memcg
)
5385 mem_cgroup_oom_notify_cb(iter
);
5388 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5389 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5391 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5392 struct mem_cgroup_thresholds
*thresholds
;
5393 struct mem_cgroup_threshold_ary
*new;
5394 enum res_type type
= MEMFILE_TYPE(cft
->private);
5395 u64 threshold
, usage
;
5398 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5402 mutex_lock(&memcg
->thresholds_lock
);
5405 thresholds
= &memcg
->thresholds
;
5406 else if (type
== _MEMSWAP
)
5407 thresholds
= &memcg
->memsw_thresholds
;
5411 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5413 /* Check if a threshold crossed before adding a new one */
5414 if (thresholds
->primary
)
5415 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5417 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5419 /* Allocate memory for new array of thresholds */
5420 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5428 /* Copy thresholds (if any) to new array */
5429 if (thresholds
->primary
) {
5430 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5431 sizeof(struct mem_cgroup_threshold
));
5434 /* Add new threshold */
5435 new->entries
[size
- 1].eventfd
= eventfd
;
5436 new->entries
[size
- 1].threshold
= threshold
;
5438 /* Sort thresholds. Registering of new threshold isn't time-critical */
5439 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5440 compare_thresholds
, NULL
);
5442 /* Find current threshold */
5443 new->current_threshold
= -1;
5444 for (i
= 0; i
< size
; i
++) {
5445 if (new->entries
[i
].threshold
<= usage
) {
5447 * new->current_threshold will not be used until
5448 * rcu_assign_pointer(), so it's safe to increment
5451 ++new->current_threshold
;
5456 /* Free old spare buffer and save old primary buffer as spare */
5457 kfree(thresholds
->spare
);
5458 thresholds
->spare
= thresholds
->primary
;
5460 rcu_assign_pointer(thresholds
->primary
, new);
5462 /* To be sure that nobody uses thresholds */
5466 mutex_unlock(&memcg
->thresholds_lock
);
5471 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5472 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5474 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5475 struct mem_cgroup_thresholds
*thresholds
;
5476 struct mem_cgroup_threshold_ary
*new;
5477 enum res_type type
= MEMFILE_TYPE(cft
->private);
5481 mutex_lock(&memcg
->thresholds_lock
);
5483 thresholds
= &memcg
->thresholds
;
5484 else if (type
== _MEMSWAP
)
5485 thresholds
= &memcg
->memsw_thresholds
;
5489 if (!thresholds
->primary
)
5492 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5494 /* Check if a threshold crossed before removing */
5495 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5497 /* Calculate new number of threshold */
5499 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5500 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5504 new = thresholds
->spare
;
5506 /* Set thresholds array to NULL if we don't have thresholds */
5515 /* Copy thresholds and find current threshold */
5516 new->current_threshold
= -1;
5517 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5518 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5521 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5522 if (new->entries
[j
].threshold
<= usage
) {
5524 * new->current_threshold will not be used
5525 * until rcu_assign_pointer(), so it's safe to increment
5528 ++new->current_threshold
;
5534 /* Swap primary and spare array */
5535 thresholds
->spare
= thresholds
->primary
;
5536 /* If all events are unregistered, free the spare array */
5538 kfree(thresholds
->spare
);
5539 thresholds
->spare
= NULL
;
5542 rcu_assign_pointer(thresholds
->primary
, new);
5544 /* To be sure that nobody uses thresholds */
5547 mutex_unlock(&memcg
->thresholds_lock
);
5550 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5551 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5553 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5554 struct mem_cgroup_eventfd_list
*event
;
5555 enum res_type type
= MEMFILE_TYPE(cft
->private);
5557 BUG_ON(type
!= _OOM_TYPE
);
5558 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5562 spin_lock(&memcg_oom_lock
);
5564 event
->eventfd
= eventfd
;
5565 list_add(&event
->list
, &memcg
->oom_notify
);
5567 /* already in OOM ? */
5568 if (atomic_read(&memcg
->under_oom
))
5569 eventfd_signal(eventfd
, 1);
5570 spin_unlock(&memcg_oom_lock
);
5575 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5576 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5578 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5579 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5580 enum res_type type
= MEMFILE_TYPE(cft
->private);
5582 BUG_ON(type
!= _OOM_TYPE
);
5584 spin_lock(&memcg_oom_lock
);
5586 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5587 if (ev
->eventfd
== eventfd
) {
5588 list_del(&ev
->list
);
5593 spin_unlock(&memcg_oom_lock
);
5596 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5597 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5599 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5601 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5603 if (atomic_read(&memcg
->under_oom
))
5604 cb
->fill(cb
, "under_oom", 1);
5606 cb
->fill(cb
, "under_oom", 0);
5610 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5611 struct cftype
*cft
, u64 val
)
5613 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5614 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5616 /* cannot set to root cgroup and only 0 and 1 are allowed */
5617 if (!parent
|| !((val
== 0) || (val
== 1)))
5620 mutex_lock(&memcg_create_mutex
);
5621 /* oom-kill-disable is a flag for subhierarchy. */
5622 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5623 mutex_unlock(&memcg_create_mutex
);
5626 memcg
->oom_kill_disable
= val
;
5628 memcg_oom_recover(memcg
);
5629 mutex_unlock(&memcg_create_mutex
);
5633 #ifdef CONFIG_MEMCG_KMEM
5634 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5638 memcg
->kmemcg_id
= -1;
5639 ret
= memcg_propagate_kmem(memcg
);
5643 return mem_cgroup_sockets_init(memcg
, ss
);
5646 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5648 mem_cgroup_sockets_destroy(memcg
);
5651 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5653 if (!memcg_kmem_is_active(memcg
))
5657 * kmem charges can outlive the cgroup. In the case of slab
5658 * pages, for instance, a page contain objects from various
5659 * processes. As we prevent from taking a reference for every
5660 * such allocation we have to be careful when doing uncharge
5661 * (see memcg_uncharge_kmem) and here during offlining.
5663 * The idea is that that only the _last_ uncharge which sees
5664 * the dead memcg will drop the last reference. An additional
5665 * reference is taken here before the group is marked dead
5666 * which is then paired with css_put during uncharge resp. here.
5668 * Although this might sound strange as this path is called from
5669 * css_offline() when the referencemight have dropped down to 0
5670 * and shouldn't be incremented anymore (css_tryget would fail)
5671 * we do not have other options because of the kmem allocations
5674 css_get(&memcg
->css
);
5676 memcg_kmem_mark_dead(memcg
);
5678 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5681 if (memcg_kmem_test_and_clear_dead(memcg
))
5682 css_put(&memcg
->css
);
5685 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5690 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5694 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5699 static struct cftype mem_cgroup_files
[] = {
5701 .name
= "usage_in_bytes",
5702 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5703 .read
= mem_cgroup_read
,
5704 .register_event
= mem_cgroup_usage_register_event
,
5705 .unregister_event
= mem_cgroup_usage_unregister_event
,
5708 .name
= "max_usage_in_bytes",
5709 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5710 .trigger
= mem_cgroup_reset
,
5711 .read
= mem_cgroup_read
,
5714 .name
= "limit_in_bytes",
5715 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5716 .write_string
= mem_cgroup_write
,
5717 .read
= mem_cgroup_read
,
5720 .name
= "soft_limit_in_bytes",
5721 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5722 .write_string
= mem_cgroup_write
,
5723 .read
= mem_cgroup_read
,
5727 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5728 .trigger
= mem_cgroup_reset
,
5729 .read
= mem_cgroup_read
,
5733 .read_seq_string
= memcg_stat_show
,
5736 .name
= "force_empty",
5737 .trigger
= mem_cgroup_force_empty_write
,
5740 .name
= "use_hierarchy",
5741 .flags
= CFTYPE_INSANE
,
5742 .write_u64
= mem_cgroup_hierarchy_write
,
5743 .read_u64
= mem_cgroup_hierarchy_read
,
5746 .name
= "swappiness",
5747 .read_u64
= mem_cgroup_swappiness_read
,
5748 .write_u64
= mem_cgroup_swappiness_write
,
5751 .name
= "move_charge_at_immigrate",
5752 .read_u64
= mem_cgroup_move_charge_read
,
5753 .write_u64
= mem_cgroup_move_charge_write
,
5756 .name
= "oom_control",
5757 .read_map
= mem_cgroup_oom_control_read
,
5758 .write_u64
= mem_cgroup_oom_control_write
,
5759 .register_event
= mem_cgroup_oom_register_event
,
5760 .unregister_event
= mem_cgroup_oom_unregister_event
,
5761 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5764 .name
= "pressure_level",
5765 .register_event
= vmpressure_register_event
,
5766 .unregister_event
= vmpressure_unregister_event
,
5770 .name
= "numa_stat",
5771 .read_seq_string
= memcg_numa_stat_show
,
5774 #ifdef CONFIG_MEMCG_KMEM
5776 .name
= "kmem.limit_in_bytes",
5777 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5778 .write_string
= mem_cgroup_write
,
5779 .read
= mem_cgroup_read
,
5782 .name
= "kmem.usage_in_bytes",
5783 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5784 .read
= mem_cgroup_read
,
5787 .name
= "kmem.failcnt",
5788 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5789 .trigger
= mem_cgroup_reset
,
5790 .read
= mem_cgroup_read
,
5793 .name
= "kmem.max_usage_in_bytes",
5794 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5795 .trigger
= mem_cgroup_reset
,
5796 .read
= mem_cgroup_read
,
5798 #ifdef CONFIG_SLABINFO
5800 .name
= "kmem.slabinfo",
5801 .read_seq_string
= mem_cgroup_slabinfo_read
,
5805 { }, /* terminate */
5808 #ifdef CONFIG_MEMCG_SWAP
5809 static struct cftype memsw_cgroup_files
[] = {
5811 .name
= "memsw.usage_in_bytes",
5812 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5813 .read
= mem_cgroup_read
,
5814 .register_event
= mem_cgroup_usage_register_event
,
5815 .unregister_event
= mem_cgroup_usage_unregister_event
,
5818 .name
= "memsw.max_usage_in_bytes",
5819 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5820 .trigger
= mem_cgroup_reset
,
5821 .read
= mem_cgroup_read
,
5824 .name
= "memsw.limit_in_bytes",
5825 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5826 .write_string
= mem_cgroup_write
,
5827 .read
= mem_cgroup_read
,
5830 .name
= "memsw.failcnt",
5831 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5832 .trigger
= mem_cgroup_reset
,
5833 .read
= mem_cgroup_read
,
5835 { }, /* terminate */
5838 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5840 struct mem_cgroup_per_node
*pn
;
5841 struct mem_cgroup_per_zone
*mz
;
5842 int zone
, tmp
= node
;
5844 * This routine is called against possible nodes.
5845 * But it's BUG to call kmalloc() against offline node.
5847 * TODO: this routine can waste much memory for nodes which will
5848 * never be onlined. It's better to use memory hotplug callback
5851 if (!node_state(node
, N_NORMAL_MEMORY
))
5853 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5857 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5858 mz
= &pn
->zoneinfo
[zone
];
5859 lruvec_init(&mz
->lruvec
);
5862 memcg
->nodeinfo
[node
] = pn
;
5866 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5868 kfree(memcg
->nodeinfo
[node
]);
5871 static struct mem_cgroup
*mem_cgroup_alloc(void)
5873 struct mem_cgroup
*memcg
;
5874 size_t size
= memcg_size();
5876 /* Can be very big if nr_node_ids is very big */
5877 if (size
< PAGE_SIZE
)
5878 memcg
= kzalloc(size
, GFP_KERNEL
);
5880 memcg
= vzalloc(size
);
5885 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5888 spin_lock_init(&memcg
->pcp_counter_lock
);
5892 if (size
< PAGE_SIZE
)
5900 * At destroying mem_cgroup, references from swap_cgroup can remain.
5901 * (scanning all at force_empty is too costly...)
5903 * Instead of clearing all references at force_empty, we remember
5904 * the number of reference from swap_cgroup and free mem_cgroup when
5905 * it goes down to 0.
5907 * Removal of cgroup itself succeeds regardless of refs from swap.
5910 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5913 size_t size
= memcg_size();
5915 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5918 free_mem_cgroup_per_zone_info(memcg
, node
);
5920 free_percpu(memcg
->stat
);
5923 * We need to make sure that (at least for now), the jump label
5924 * destruction code runs outside of the cgroup lock. This is because
5925 * get_online_cpus(), which is called from the static_branch update,
5926 * can't be called inside the cgroup_lock. cpusets are the ones
5927 * enforcing this dependency, so if they ever change, we might as well.
5929 * schedule_work() will guarantee this happens. Be careful if you need
5930 * to move this code around, and make sure it is outside
5933 disarm_static_keys(memcg
);
5934 if (size
< PAGE_SIZE
)
5941 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5943 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
5945 if (!memcg
->res
.parent
)
5947 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
5949 EXPORT_SYMBOL(parent_mem_cgroup
);
5951 static struct cgroup_subsys_state
* __ref
5952 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
5954 struct mem_cgroup
*memcg
;
5955 long error
= -ENOMEM
;
5958 memcg
= mem_cgroup_alloc();
5960 return ERR_PTR(error
);
5963 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
5967 if (parent_css
== NULL
) {
5968 root_mem_cgroup
= memcg
;
5969 res_counter_init(&memcg
->res
, NULL
);
5970 res_counter_init(&memcg
->memsw
, NULL
);
5971 res_counter_init(&memcg
->kmem
, NULL
);
5974 memcg
->last_scanned_node
= MAX_NUMNODES
;
5975 INIT_LIST_HEAD(&memcg
->oom_notify
);
5976 memcg
->move_charge_at_immigrate
= 0;
5977 mutex_init(&memcg
->thresholds_lock
);
5978 spin_lock_init(&memcg
->move_lock
);
5979 vmpressure_init(&memcg
->vmpressure
);
5980 spin_lock_init(&memcg
->soft_lock
);
5985 __mem_cgroup_free(memcg
);
5986 return ERR_PTR(error
);
5990 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
5992 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5993 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
5999 mutex_lock(&memcg_create_mutex
);
6001 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6002 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6003 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6005 if (parent
->use_hierarchy
) {
6006 res_counter_init(&memcg
->res
, &parent
->res
);
6007 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6008 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6011 * No need to take a reference to the parent because cgroup
6012 * core guarantees its existence.
6015 res_counter_init(&memcg
->res
, NULL
);
6016 res_counter_init(&memcg
->memsw
, NULL
);
6017 res_counter_init(&memcg
->kmem
, NULL
);
6019 * Deeper hierachy with use_hierarchy == false doesn't make
6020 * much sense so let cgroup subsystem know about this
6021 * unfortunate state in our controller.
6023 if (parent
!= root_mem_cgroup
)
6024 mem_cgroup_subsys
.broken_hierarchy
= true;
6027 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6028 mutex_unlock(&memcg_create_mutex
);
6033 * Announce all parents that a group from their hierarchy is gone.
6035 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6037 struct mem_cgroup
*parent
= memcg
;
6039 while ((parent
= parent_mem_cgroup(parent
)))
6040 mem_cgroup_iter_invalidate(parent
);
6043 * if the root memcg is not hierarchical we have to check it
6046 if (!root_mem_cgroup
->use_hierarchy
)
6047 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6050 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6052 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6054 kmem_cgroup_css_offline(memcg
);
6056 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6057 mem_cgroup_reparent_charges(memcg
);
6058 if (memcg
->soft_contributed
) {
6059 while ((memcg
= parent_mem_cgroup(memcg
)))
6060 atomic_dec(&memcg
->children_in_excess
);
6062 if (memcg
!= root_mem_cgroup
&& !root_mem_cgroup
->use_hierarchy
)
6063 atomic_dec(&root_mem_cgroup
->children_in_excess
);
6065 mem_cgroup_destroy_all_caches(memcg
);
6066 vmpressure_cleanup(&memcg
->vmpressure
);
6069 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6071 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6073 memcg_destroy_kmem(memcg
);
6074 __mem_cgroup_free(memcg
);
6078 /* Handlers for move charge at task migration. */
6079 #define PRECHARGE_COUNT_AT_ONCE 256
6080 static int mem_cgroup_do_precharge(unsigned long count
)
6083 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6084 struct mem_cgroup
*memcg
= mc
.to
;
6086 if (mem_cgroup_is_root(memcg
)) {
6087 mc
.precharge
+= count
;
6088 /* we don't need css_get for root */
6091 /* try to charge at once */
6093 struct res_counter
*dummy
;
6095 * "memcg" cannot be under rmdir() because we've already checked
6096 * by cgroup_lock_live_cgroup() that it is not removed and we
6097 * are still under the same cgroup_mutex. So we can postpone
6100 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6102 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6103 PAGE_SIZE
* count
, &dummy
)) {
6104 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6107 mc
.precharge
+= count
;
6111 /* fall back to one by one charge */
6113 if (signal_pending(current
)) {
6117 if (!batch_count
--) {
6118 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6121 ret
= __mem_cgroup_try_charge(NULL
,
6122 GFP_KERNEL
, 1, &memcg
, false);
6124 /* mem_cgroup_clear_mc() will do uncharge later */
6132 * get_mctgt_type - get target type of moving charge
6133 * @vma: the vma the pte to be checked belongs
6134 * @addr: the address corresponding to the pte to be checked
6135 * @ptent: the pte to be checked
6136 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6139 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6140 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6141 * move charge. if @target is not NULL, the page is stored in target->page
6142 * with extra refcnt got(Callers should handle it).
6143 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6144 * target for charge migration. if @target is not NULL, the entry is stored
6147 * Called with pte lock held.
6154 enum mc_target_type
{
6160 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6161 unsigned long addr
, pte_t ptent
)
6163 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6165 if (!page
|| !page_mapped(page
))
6167 if (PageAnon(page
)) {
6168 /* we don't move shared anon */
6171 } else if (!move_file())
6172 /* we ignore mapcount for file pages */
6174 if (!get_page_unless_zero(page
))
6181 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6182 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6184 struct page
*page
= NULL
;
6185 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6187 if (!move_anon() || non_swap_entry(ent
))
6190 * Because lookup_swap_cache() updates some statistics counter,
6191 * we call find_get_page() with swapper_space directly.
6193 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6194 if (do_swap_account
)
6195 entry
->val
= ent
.val
;
6200 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6201 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6207 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6208 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6210 struct page
*page
= NULL
;
6211 struct address_space
*mapping
;
6214 if (!vma
->vm_file
) /* anonymous vma */
6219 mapping
= vma
->vm_file
->f_mapping
;
6220 if (pte_none(ptent
))
6221 pgoff
= linear_page_index(vma
, addr
);
6222 else /* pte_file(ptent) is true */
6223 pgoff
= pte_to_pgoff(ptent
);
6225 /* page is moved even if it's not RSS of this task(page-faulted). */
6226 page
= find_get_page(mapping
, pgoff
);
6229 /* shmem/tmpfs may report page out on swap: account for that too. */
6230 if (radix_tree_exceptional_entry(page
)) {
6231 swp_entry_t swap
= radix_to_swp_entry(page
);
6232 if (do_swap_account
)
6234 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6240 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6241 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6243 struct page
*page
= NULL
;
6244 struct page_cgroup
*pc
;
6245 enum mc_target_type ret
= MC_TARGET_NONE
;
6246 swp_entry_t ent
= { .val
= 0 };
6248 if (pte_present(ptent
))
6249 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6250 else if (is_swap_pte(ptent
))
6251 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6252 else if (pte_none(ptent
) || pte_file(ptent
))
6253 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6255 if (!page
&& !ent
.val
)
6258 pc
= lookup_page_cgroup(page
);
6260 * Do only loose check w/o page_cgroup lock.
6261 * mem_cgroup_move_account() checks the pc is valid or not under
6264 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6265 ret
= MC_TARGET_PAGE
;
6267 target
->page
= page
;
6269 if (!ret
|| !target
)
6272 /* There is a swap entry and a page doesn't exist or isn't charged */
6273 if (ent
.val
&& !ret
&&
6274 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6275 ret
= MC_TARGET_SWAP
;
6282 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6284 * We don't consider swapping or file mapped pages because THP does not
6285 * support them for now.
6286 * Caller should make sure that pmd_trans_huge(pmd) is true.
6288 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6289 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6291 struct page
*page
= NULL
;
6292 struct page_cgroup
*pc
;
6293 enum mc_target_type ret
= MC_TARGET_NONE
;
6295 page
= pmd_page(pmd
);
6296 VM_BUG_ON(!page
|| !PageHead(page
));
6299 pc
= lookup_page_cgroup(page
);
6300 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6301 ret
= MC_TARGET_PAGE
;
6304 target
->page
= page
;
6310 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6311 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6313 return MC_TARGET_NONE
;
6317 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6318 unsigned long addr
, unsigned long end
,
6319 struct mm_walk
*walk
)
6321 struct vm_area_struct
*vma
= walk
->private;
6325 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6326 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6327 mc
.precharge
+= HPAGE_PMD_NR
;
6328 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6332 if (pmd_trans_unstable(pmd
))
6334 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6335 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6336 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6337 mc
.precharge
++; /* increment precharge temporarily */
6338 pte_unmap_unlock(pte
- 1, ptl
);
6344 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6346 unsigned long precharge
;
6347 struct vm_area_struct
*vma
;
6349 down_read(&mm
->mmap_sem
);
6350 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6351 struct mm_walk mem_cgroup_count_precharge_walk
= {
6352 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6356 if (is_vm_hugetlb_page(vma
))
6358 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6359 &mem_cgroup_count_precharge_walk
);
6361 up_read(&mm
->mmap_sem
);
6363 precharge
= mc
.precharge
;
6369 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6371 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6373 VM_BUG_ON(mc
.moving_task
);
6374 mc
.moving_task
= current
;
6375 return mem_cgroup_do_precharge(precharge
);
6378 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6379 static void __mem_cgroup_clear_mc(void)
6381 struct mem_cgroup
*from
= mc
.from
;
6382 struct mem_cgroup
*to
= mc
.to
;
6385 /* we must uncharge all the leftover precharges from mc.to */
6387 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6391 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6392 * we must uncharge here.
6394 if (mc
.moved_charge
) {
6395 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6396 mc
.moved_charge
= 0;
6398 /* we must fixup refcnts and charges */
6399 if (mc
.moved_swap
) {
6400 /* uncharge swap account from the old cgroup */
6401 if (!mem_cgroup_is_root(mc
.from
))
6402 res_counter_uncharge(&mc
.from
->memsw
,
6403 PAGE_SIZE
* mc
.moved_swap
);
6405 for (i
= 0; i
< mc
.moved_swap
; i
++)
6406 css_put(&mc
.from
->css
);
6408 if (!mem_cgroup_is_root(mc
.to
)) {
6410 * we charged both to->res and to->memsw, so we should
6413 res_counter_uncharge(&mc
.to
->res
,
6414 PAGE_SIZE
* mc
.moved_swap
);
6416 /* we've already done css_get(mc.to) */
6419 memcg_oom_recover(from
);
6420 memcg_oom_recover(to
);
6421 wake_up_all(&mc
.waitq
);
6424 static void mem_cgroup_clear_mc(void)
6426 struct mem_cgroup
*from
= mc
.from
;
6429 * we must clear moving_task before waking up waiters at the end of
6432 mc
.moving_task
= NULL
;
6433 __mem_cgroup_clear_mc();
6434 spin_lock(&mc
.lock
);
6437 spin_unlock(&mc
.lock
);
6438 mem_cgroup_end_move(from
);
6441 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6442 struct cgroup_taskset
*tset
)
6444 struct task_struct
*p
= cgroup_taskset_first(tset
);
6446 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6447 unsigned long move_charge_at_immigrate
;
6450 * We are now commited to this value whatever it is. Changes in this
6451 * tunable will only affect upcoming migrations, not the current one.
6452 * So we need to save it, and keep it going.
6454 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6455 if (move_charge_at_immigrate
) {
6456 struct mm_struct
*mm
;
6457 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6459 VM_BUG_ON(from
== memcg
);
6461 mm
= get_task_mm(p
);
6464 /* We move charges only when we move a owner of the mm */
6465 if (mm
->owner
== p
) {
6468 VM_BUG_ON(mc
.precharge
);
6469 VM_BUG_ON(mc
.moved_charge
);
6470 VM_BUG_ON(mc
.moved_swap
);
6471 mem_cgroup_start_move(from
);
6472 spin_lock(&mc
.lock
);
6475 mc
.immigrate_flags
= move_charge_at_immigrate
;
6476 spin_unlock(&mc
.lock
);
6477 /* We set mc.moving_task later */
6479 ret
= mem_cgroup_precharge_mc(mm
);
6481 mem_cgroup_clear_mc();
6488 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6489 struct cgroup_taskset
*tset
)
6491 mem_cgroup_clear_mc();
6494 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6495 unsigned long addr
, unsigned long end
,
6496 struct mm_walk
*walk
)
6499 struct vm_area_struct
*vma
= walk
->private;
6502 enum mc_target_type target_type
;
6503 union mc_target target
;
6505 struct page_cgroup
*pc
;
6508 * We don't take compound_lock() here but no race with splitting thp
6510 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6511 * under splitting, which means there's no concurrent thp split,
6512 * - if another thread runs into split_huge_page() just after we
6513 * entered this if-block, the thread must wait for page table lock
6514 * to be unlocked in __split_huge_page_splitting(), where the main
6515 * part of thp split is not executed yet.
6517 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6518 if (mc
.precharge
< HPAGE_PMD_NR
) {
6519 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6522 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6523 if (target_type
== MC_TARGET_PAGE
) {
6525 if (!isolate_lru_page(page
)) {
6526 pc
= lookup_page_cgroup(page
);
6527 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6528 pc
, mc
.from
, mc
.to
)) {
6529 mc
.precharge
-= HPAGE_PMD_NR
;
6530 mc
.moved_charge
+= HPAGE_PMD_NR
;
6532 putback_lru_page(page
);
6536 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6540 if (pmd_trans_unstable(pmd
))
6543 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6544 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6545 pte_t ptent
= *(pte
++);
6551 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6552 case MC_TARGET_PAGE
:
6554 if (isolate_lru_page(page
))
6556 pc
= lookup_page_cgroup(page
);
6557 if (!mem_cgroup_move_account(page
, 1, pc
,
6560 /* we uncharge from mc.from later. */
6563 putback_lru_page(page
);
6564 put
: /* get_mctgt_type() gets the page */
6567 case MC_TARGET_SWAP
:
6569 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6571 /* we fixup refcnts and charges later. */
6579 pte_unmap_unlock(pte
- 1, ptl
);
6584 * We have consumed all precharges we got in can_attach().
6585 * We try charge one by one, but don't do any additional
6586 * charges to mc.to if we have failed in charge once in attach()
6589 ret
= mem_cgroup_do_precharge(1);
6597 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6599 struct vm_area_struct
*vma
;
6601 lru_add_drain_all();
6603 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6605 * Someone who are holding the mmap_sem might be waiting in
6606 * waitq. So we cancel all extra charges, wake up all waiters,
6607 * and retry. Because we cancel precharges, we might not be able
6608 * to move enough charges, but moving charge is a best-effort
6609 * feature anyway, so it wouldn't be a big problem.
6611 __mem_cgroup_clear_mc();
6615 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6617 struct mm_walk mem_cgroup_move_charge_walk
= {
6618 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6622 if (is_vm_hugetlb_page(vma
))
6624 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6625 &mem_cgroup_move_charge_walk
);
6628 * means we have consumed all precharges and failed in
6629 * doing additional charge. Just abandon here.
6633 up_read(&mm
->mmap_sem
);
6636 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6637 struct cgroup_taskset
*tset
)
6639 struct task_struct
*p
= cgroup_taskset_first(tset
);
6640 struct mm_struct
*mm
= get_task_mm(p
);
6644 mem_cgroup_move_charge(mm
);
6648 mem_cgroup_clear_mc();
6650 #else /* !CONFIG_MMU */
6651 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6652 struct cgroup_taskset
*tset
)
6656 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6657 struct cgroup_taskset
*tset
)
6660 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6661 struct cgroup_taskset
*tset
)
6667 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6668 * to verify sane_behavior flag on each mount attempt.
6670 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6673 * use_hierarchy is forced with sane_behavior. cgroup core
6674 * guarantees that @root doesn't have any children, so turning it
6675 * on for the root memcg is enough.
6677 if (cgroup_sane_behavior(root_css
->cgroup
))
6678 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6681 struct cgroup_subsys mem_cgroup_subsys
= {
6683 .subsys_id
= mem_cgroup_subsys_id
,
6684 .css_alloc
= mem_cgroup_css_alloc
,
6685 .css_online
= mem_cgroup_css_online
,
6686 .css_offline
= mem_cgroup_css_offline
,
6687 .css_free
= mem_cgroup_css_free
,
6688 .can_attach
= mem_cgroup_can_attach
,
6689 .cancel_attach
= mem_cgroup_cancel_attach
,
6690 .attach
= mem_cgroup_move_task
,
6691 .bind
= mem_cgroup_bind
,
6692 .base_cftypes
= mem_cgroup_files
,
6697 #ifdef CONFIG_MEMCG_SWAP
6698 static int __init
enable_swap_account(char *s
)
6700 if (!strcmp(s
, "1"))
6701 really_do_swap_account
= 1;
6702 else if (!strcmp(s
, "0"))
6703 really_do_swap_account
= 0;
6706 __setup("swapaccount=", enable_swap_account
);
6708 static void __init
memsw_file_init(void)
6710 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6713 static void __init
enable_swap_cgroup(void)
6715 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6716 do_swap_account
= 1;
6722 static void __init
enable_swap_cgroup(void)
6728 * subsys_initcall() for memory controller.
6730 * Some parts like hotcpu_notifier() have to be initialized from this context
6731 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6732 * everything that doesn't depend on a specific mem_cgroup structure should
6733 * be initialized from here.
6735 static int __init
mem_cgroup_init(void)
6737 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6738 enable_swap_cgroup();
6742 subsys_initcall(mem_cgroup_init
);