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 */
258 atomic_t oom_wakeups
;
261 /* OOM-Killer disable */
262 int oom_kill_disable
;
264 /* set when res.limit == memsw.limit */
265 bool memsw_is_minimum
;
267 /* protect arrays of thresholds */
268 struct mutex thresholds_lock
;
270 /* thresholds for memory usage. RCU-protected */
271 struct mem_cgroup_thresholds thresholds
;
273 /* thresholds for mem+swap usage. RCU-protected */
274 struct mem_cgroup_thresholds memsw_thresholds
;
276 /* For oom notifier event fd */
277 struct list_head oom_notify
;
280 * Should we move charges of a task when a task is moved into this
281 * mem_cgroup ? And what type of charges should we move ?
283 unsigned long move_charge_at_immigrate
;
285 * set > 0 if pages under this cgroup are moving to other cgroup.
287 atomic_t moving_account
;
288 /* taken only while moving_account > 0 */
289 spinlock_t move_lock
;
293 struct mem_cgroup_stat_cpu __percpu
*stat
;
295 * used when a cpu is offlined or other synchronizations
296 * See mem_cgroup_read_stat().
298 struct mem_cgroup_stat_cpu nocpu_base
;
299 spinlock_t pcp_counter_lock
;
302 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
303 struct tcp_memcontrol tcp_mem
;
305 #if defined(CONFIG_MEMCG_KMEM)
306 /* analogous to slab_common's slab_caches list. per-memcg */
307 struct list_head memcg_slab_caches
;
308 /* Not a spinlock, we can take a lot of time walking the list */
309 struct mutex slab_caches_mutex
;
310 /* Index in the kmem_cache->memcg_params->memcg_caches array */
314 int last_scanned_node
;
316 nodemask_t scan_nodes
;
317 atomic_t numainfo_events
;
318 atomic_t numainfo_updating
;
321 * Protects soft_contributed transitions.
322 * See mem_cgroup_update_soft_limit
324 spinlock_t soft_lock
;
327 * If true then this group has increased parents' children_in_excess
328 * when it got over the soft limit.
329 * When a group falls bellow the soft limit, parents' children_in_excess
330 * is decreased and soft_contributed changed to false.
332 bool soft_contributed
;
334 /* Number of children that are in soft limit excess */
335 atomic_t children_in_excess
;
337 struct mem_cgroup_per_node
*nodeinfo
[0];
338 /* WARNING: nodeinfo must be the last member here */
341 static size_t memcg_size(void)
343 return sizeof(struct mem_cgroup
) +
344 nr_node_ids
* sizeof(struct mem_cgroup_per_node
);
347 /* internal only representation about the status of kmem accounting. */
349 KMEM_ACCOUNTED_ACTIVE
= 0, /* accounted by this cgroup itself */
350 KMEM_ACCOUNTED_ACTIVATED
, /* static key enabled. */
351 KMEM_ACCOUNTED_DEAD
, /* dead memcg with pending kmem charges */
354 /* We account when limit is on, but only after call sites are patched */
355 #define KMEM_ACCOUNTED_MASK \
356 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
358 #ifdef CONFIG_MEMCG_KMEM
359 static inline void memcg_kmem_set_active(struct mem_cgroup
*memcg
)
361 set_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
364 static bool memcg_kmem_is_active(struct mem_cgroup
*memcg
)
366 return test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
);
369 static void memcg_kmem_set_activated(struct mem_cgroup
*memcg
)
371 set_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
374 static void memcg_kmem_clear_activated(struct mem_cgroup
*memcg
)
376 clear_bit(KMEM_ACCOUNTED_ACTIVATED
, &memcg
->kmem_account_flags
);
379 static void memcg_kmem_mark_dead(struct mem_cgroup
*memcg
)
382 * Our caller must use css_get() first, because memcg_uncharge_kmem()
383 * will call css_put() if it sees the memcg is dead.
386 if (test_bit(KMEM_ACCOUNTED_ACTIVE
, &memcg
->kmem_account_flags
))
387 set_bit(KMEM_ACCOUNTED_DEAD
, &memcg
->kmem_account_flags
);
390 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup
*memcg
)
392 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD
,
393 &memcg
->kmem_account_flags
);
397 /* Stuffs for move charges at task migration. */
399 * Types of charges to be moved. "move_charge_at_immitgrate" and
400 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
403 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
404 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
408 /* "mc" and its members are protected by cgroup_mutex */
409 static struct move_charge_struct
{
410 spinlock_t lock
; /* for from, to */
411 struct mem_cgroup
*from
;
412 struct mem_cgroup
*to
;
413 unsigned long immigrate_flags
;
414 unsigned long precharge
;
415 unsigned long moved_charge
;
416 unsigned long moved_swap
;
417 struct task_struct
*moving_task
; /* a task moving charges */
418 wait_queue_head_t waitq
; /* a waitq for other context */
420 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
421 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
424 static bool move_anon(void)
426 return test_bit(MOVE_CHARGE_TYPE_ANON
, &mc
.immigrate_flags
);
429 static bool move_file(void)
431 return test_bit(MOVE_CHARGE_TYPE_FILE
, &mc
.immigrate_flags
);
435 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
436 * limit reclaim to prevent infinite loops, if they ever occur.
438 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
441 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
442 MEM_CGROUP_CHARGE_TYPE_ANON
,
443 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
444 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
448 /* for encoding cft->private value on file */
456 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
457 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
458 #define MEMFILE_ATTR(val) ((val) & 0xffff)
459 /* Used for OOM nofiier */
460 #define OOM_CONTROL (0)
463 * Reclaim flags for mem_cgroup_hierarchical_reclaim
465 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
466 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
467 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
468 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
471 * The memcg_create_mutex will be held whenever a new cgroup is created.
472 * As a consequence, any change that needs to protect against new child cgroups
473 * appearing has to hold it as well.
475 static DEFINE_MUTEX(memcg_create_mutex
);
477 struct mem_cgroup
*mem_cgroup_from_css(struct cgroup_subsys_state
*s
)
479 return s
? container_of(s
, struct mem_cgroup
, css
) : NULL
;
482 /* Some nice accessors for the vmpressure. */
483 struct vmpressure
*memcg_to_vmpressure(struct mem_cgroup
*memcg
)
486 memcg
= root_mem_cgroup
;
487 return &memcg
->vmpressure
;
490 struct cgroup_subsys_state
*vmpressure_to_css(struct vmpressure
*vmpr
)
492 return &container_of(vmpr
, struct mem_cgroup
, vmpressure
)->css
;
495 struct vmpressure
*css_to_vmpressure(struct cgroup_subsys_state
*css
)
497 return &mem_cgroup_from_css(css
)->vmpressure
;
500 static inline bool mem_cgroup_is_root(struct mem_cgroup
*memcg
)
502 return (memcg
== root_mem_cgroup
);
505 /* Writing them here to avoid exposing memcg's inner layout */
506 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
508 void sock_update_memcg(struct sock
*sk
)
510 if (mem_cgroup_sockets_enabled
) {
511 struct mem_cgroup
*memcg
;
512 struct cg_proto
*cg_proto
;
514 BUG_ON(!sk
->sk_prot
->proto_cgroup
);
516 /* Socket cloning can throw us here with sk_cgrp already
517 * filled. It won't however, necessarily happen from
518 * process context. So the test for root memcg given
519 * the current task's memcg won't help us in this case.
521 * Respecting the original socket's memcg is a better
522 * decision in this case.
525 BUG_ON(mem_cgroup_is_root(sk
->sk_cgrp
->memcg
));
526 css_get(&sk
->sk_cgrp
->memcg
->css
);
531 memcg
= mem_cgroup_from_task(current
);
532 cg_proto
= sk
->sk_prot
->proto_cgroup(memcg
);
533 if (!mem_cgroup_is_root(memcg
) &&
534 memcg_proto_active(cg_proto
) && css_tryget(&memcg
->css
)) {
535 sk
->sk_cgrp
= cg_proto
;
540 EXPORT_SYMBOL(sock_update_memcg
);
542 void sock_release_memcg(struct sock
*sk
)
544 if (mem_cgroup_sockets_enabled
&& sk
->sk_cgrp
) {
545 struct mem_cgroup
*memcg
;
546 WARN_ON(!sk
->sk_cgrp
->memcg
);
547 memcg
= sk
->sk_cgrp
->memcg
;
548 css_put(&sk
->sk_cgrp
->memcg
->css
);
552 struct cg_proto
*tcp_proto_cgroup(struct mem_cgroup
*memcg
)
554 if (!memcg
|| mem_cgroup_is_root(memcg
))
557 return &memcg
->tcp_mem
.cg_proto
;
559 EXPORT_SYMBOL(tcp_proto_cgroup
);
561 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
563 if (!memcg_proto_activated(&memcg
->tcp_mem
.cg_proto
))
565 static_key_slow_dec(&memcg_socket_limit_enabled
);
568 static void disarm_sock_keys(struct mem_cgroup
*memcg
)
573 #ifdef CONFIG_MEMCG_KMEM
575 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
576 * There are two main reasons for not using the css_id for this:
577 * 1) this works better in sparse environments, where we have a lot of memcgs,
578 * but only a few kmem-limited. Or also, if we have, for instance, 200
579 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
580 * 200 entry array for that.
582 * 2) In order not to violate the cgroup API, we would like to do all memory
583 * allocation in ->create(). At that point, we haven't yet allocated the
584 * css_id. Having a separate index prevents us from messing with the cgroup
587 * The current size of the caches array is stored in
588 * memcg_limited_groups_array_size. It will double each time we have to
591 static DEFINE_IDA(kmem_limited_groups
);
592 int memcg_limited_groups_array_size
;
595 * MIN_SIZE is different than 1, because we would like to avoid going through
596 * the alloc/free process all the time. In a small machine, 4 kmem-limited
597 * cgroups is a reasonable guess. In the future, it could be a parameter or
598 * tunable, but that is strictly not necessary.
600 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
601 * this constant directly from cgroup, but it is understandable that this is
602 * better kept as an internal representation in cgroup.c. In any case, the
603 * css_id space is not getting any smaller, and we don't have to necessarily
604 * increase ours as well if it increases.
606 #define MEMCG_CACHES_MIN_SIZE 4
607 #define MEMCG_CACHES_MAX_SIZE 65535
610 * A lot of the calls to the cache allocation functions are expected to be
611 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
612 * conditional to this static branch, we'll have to allow modules that does
613 * kmem_cache_alloc and the such to see this symbol as well
615 struct static_key memcg_kmem_enabled_key
;
616 EXPORT_SYMBOL(memcg_kmem_enabled_key
);
618 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
620 if (memcg_kmem_is_active(memcg
)) {
621 static_key_slow_dec(&memcg_kmem_enabled_key
);
622 ida_simple_remove(&kmem_limited_groups
, memcg
->kmemcg_id
);
625 * This check can't live in kmem destruction function,
626 * since the charges will outlive the cgroup
628 WARN_ON(res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0);
631 static void disarm_kmem_keys(struct mem_cgroup
*memcg
)
634 #endif /* CONFIG_MEMCG_KMEM */
636 static void disarm_static_keys(struct mem_cgroup
*memcg
)
638 disarm_sock_keys(memcg
);
639 disarm_kmem_keys(memcg
);
642 static void drain_all_stock_async(struct mem_cgroup
*memcg
);
644 static struct mem_cgroup_per_zone
*
645 mem_cgroup_zoneinfo(struct mem_cgroup
*memcg
, int nid
, int zid
)
647 VM_BUG_ON((unsigned)nid
>= nr_node_ids
);
648 return &memcg
->nodeinfo
[nid
]->zoneinfo
[zid
];
651 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*memcg
)
656 static struct mem_cgroup_per_zone
*
657 page_cgroup_zoneinfo(struct mem_cgroup
*memcg
, struct page
*page
)
659 int nid
= page_to_nid(page
);
660 int zid
= page_zonenum(page
);
662 return mem_cgroup_zoneinfo(memcg
, nid
, zid
);
666 * Implementation Note: reading percpu statistics for memcg.
668 * Both of vmstat[] and percpu_counter has threshold and do periodic
669 * synchronization to implement "quick" read. There are trade-off between
670 * reading cost and precision of value. Then, we may have a chance to implement
671 * a periodic synchronizion of counter in memcg's counter.
673 * But this _read() function is used for user interface now. The user accounts
674 * memory usage by memory cgroup and he _always_ requires exact value because
675 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
676 * have to visit all online cpus and make sum. So, for now, unnecessary
677 * synchronization is not implemented. (just implemented for cpu hotplug)
679 * If there are kernel internal actions which can make use of some not-exact
680 * value, and reading all cpu value can be performance bottleneck in some
681 * common workload, threashold and synchonization as vmstat[] should be
684 static long mem_cgroup_read_stat(struct mem_cgroup
*memcg
,
685 enum mem_cgroup_stat_index idx
)
691 for_each_online_cpu(cpu
)
692 val
+= per_cpu(memcg
->stat
->count
[idx
], cpu
);
693 #ifdef CONFIG_HOTPLUG_CPU
694 spin_lock(&memcg
->pcp_counter_lock
);
695 val
+= memcg
->nocpu_base
.count
[idx
];
696 spin_unlock(&memcg
->pcp_counter_lock
);
702 static void mem_cgroup_swap_statistics(struct mem_cgroup
*memcg
,
705 int val
= (charge
) ? 1 : -1;
706 this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_SWAP
], val
);
709 static unsigned long mem_cgroup_read_events(struct mem_cgroup
*memcg
,
710 enum mem_cgroup_events_index idx
)
712 unsigned long val
= 0;
715 for_each_online_cpu(cpu
)
716 val
+= per_cpu(memcg
->stat
->events
[idx
], cpu
);
717 #ifdef CONFIG_HOTPLUG_CPU
718 spin_lock(&memcg
->pcp_counter_lock
);
719 val
+= memcg
->nocpu_base
.events
[idx
];
720 spin_unlock(&memcg
->pcp_counter_lock
);
725 static void mem_cgroup_charge_statistics(struct mem_cgroup
*memcg
,
727 bool anon
, int nr_pages
)
732 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
733 * counted as CACHE even if it's on ANON LRU.
736 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS
],
739 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_CACHE
],
742 if (PageTransHuge(page
))
743 __this_cpu_add(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
746 /* pagein of a big page is an event. So, ignore page size */
748 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGIN
]);
750 __this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGPGOUT
]);
751 nr_pages
= -nr_pages
; /* for event */
754 __this_cpu_add(memcg
->stat
->nr_page_events
, nr_pages
);
760 mem_cgroup_get_lru_size(struct lruvec
*lruvec
, enum lru_list lru
)
762 struct mem_cgroup_per_zone
*mz
;
764 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
765 return mz
->lru_size
[lru
];
769 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup
*memcg
, int nid
, int zid
,
770 unsigned int lru_mask
)
772 struct mem_cgroup_per_zone
*mz
;
774 unsigned long ret
= 0;
776 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
779 if (BIT(lru
) & lru_mask
)
780 ret
+= mz
->lru_size
[lru
];
786 mem_cgroup_node_nr_lru_pages(struct mem_cgroup
*memcg
,
787 int nid
, unsigned int lru_mask
)
792 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++)
793 total
+= mem_cgroup_zone_nr_lru_pages(memcg
,
799 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup
*memcg
,
800 unsigned int lru_mask
)
805 for_each_node_state(nid
, N_MEMORY
)
806 total
+= mem_cgroup_node_nr_lru_pages(memcg
, nid
, lru_mask
);
810 static bool mem_cgroup_event_ratelimit(struct mem_cgroup
*memcg
,
811 enum mem_cgroup_events_target target
)
813 unsigned long val
, next
;
815 val
= __this_cpu_read(memcg
->stat
->nr_page_events
);
816 next
= __this_cpu_read(memcg
->stat
->targets
[target
]);
817 /* from time_after() in jiffies.h */
818 if ((long)next
- (long)val
< 0) {
820 case MEM_CGROUP_TARGET_THRESH
:
821 next
= val
+ THRESHOLDS_EVENTS_TARGET
;
823 case MEM_CGROUP_TARGET_SOFTLIMIT
:
824 next
= val
+ SOFTLIMIT_EVENTS_TARGET
;
826 case MEM_CGROUP_TARGET_NUMAINFO
:
827 next
= val
+ NUMAINFO_EVENTS_TARGET
;
832 __this_cpu_write(memcg
->stat
->targets
[target
], next
);
839 * Called from rate-limited memcg_check_events when enough
840 * MEM_CGROUP_TARGET_SOFTLIMIT events are accumulated and it makes sure
841 * that all the parents up the hierarchy will be notified that this group
842 * is in excess or that it is not in excess anymore. mmecg->soft_contributed
843 * makes the transition a single action whenever the state flips from one to
846 static void mem_cgroup_update_soft_limit(struct mem_cgroup
*memcg
)
848 unsigned long long excess
= res_counter_soft_limit_excess(&memcg
->res
);
849 struct mem_cgroup
*parent
= memcg
;
852 spin_lock(&memcg
->soft_lock
);
854 if (!memcg
->soft_contributed
) {
856 memcg
->soft_contributed
= true;
859 if (memcg
->soft_contributed
) {
861 memcg
->soft_contributed
= false;
866 * Necessary to update all ancestors when hierarchy is used
867 * because their event counter is not touched.
868 * We track children even outside the hierarchy for the root
869 * cgroup because tree walk starting at root should visit
870 * all cgroups and we want to prevent from pointless tree
871 * walk if no children is below the limit.
873 while (delta
&& (parent
= parent_mem_cgroup(parent
)))
874 atomic_add(delta
, &parent
->children_in_excess
);
875 if (memcg
!= root_mem_cgroup
&& !root_mem_cgroup
->use_hierarchy
)
876 atomic_add(delta
, &root_mem_cgroup
->children_in_excess
);
877 spin_unlock(&memcg
->soft_lock
);
881 * Check events in order.
884 static void memcg_check_events(struct mem_cgroup
*memcg
, struct page
*page
)
887 /* threshold event is triggered in finer grain than soft limit */
888 if (unlikely(mem_cgroup_event_ratelimit(memcg
,
889 MEM_CGROUP_TARGET_THRESH
))) {
891 bool do_numainfo __maybe_unused
;
893 do_softlimit
= mem_cgroup_event_ratelimit(memcg
,
894 MEM_CGROUP_TARGET_SOFTLIMIT
);
896 do_numainfo
= mem_cgroup_event_ratelimit(memcg
,
897 MEM_CGROUP_TARGET_NUMAINFO
);
901 mem_cgroup_threshold(memcg
);
902 if (unlikely(do_softlimit
))
903 mem_cgroup_update_soft_limit(memcg
);
905 if (unlikely(do_numainfo
))
906 atomic_inc(&memcg
->numainfo_events
);
912 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
915 * mm_update_next_owner() may clear mm->owner to NULL
916 * if it races with swapoff, page migration, etc.
917 * So this can be called with p == NULL.
922 return mem_cgroup_from_css(task_css(p
, mem_cgroup_subsys_id
));
925 struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
927 struct mem_cgroup
*memcg
= NULL
;
932 * Because we have no locks, mm->owner's may be being moved to other
933 * cgroup. We use css_tryget() here even if this looks
934 * pessimistic (rather than adding locks here).
938 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
939 if (unlikely(!memcg
))
941 } while (!css_tryget(&memcg
->css
));
946 static enum mem_cgroup_filter_t
947 mem_cgroup_filter(struct mem_cgroup
*memcg
, struct mem_cgroup
*root
,
948 mem_cgroup_iter_filter cond
)
952 return cond(memcg
, root
);
956 * Returns a next (in a pre-order walk) alive memcg (with elevated css
957 * ref. count) or NULL if the whole root's subtree has been visited.
959 * helper function to be used by mem_cgroup_iter
961 static struct mem_cgroup
*__mem_cgroup_iter_next(struct mem_cgroup
*root
,
962 struct mem_cgroup
*last_visited
, mem_cgroup_iter_filter cond
)
964 struct cgroup_subsys_state
*prev_css
, *next_css
;
966 prev_css
= last_visited
? &last_visited
->css
: NULL
;
968 next_css
= css_next_descendant_pre(prev_css
, &root
->css
);
971 * Even if we found a group we have to make sure it is
972 * alive. css && !memcg means that the groups should be
973 * skipped and we should continue the tree walk.
974 * last_visited css is safe to use because it is
975 * protected by css_get and the tree walk is rcu safe.
978 struct mem_cgroup
*mem
= mem_cgroup_from_css(next_css
);
980 switch (mem_cgroup_filter(mem
, root
, cond
)) {
988 * css_rightmost_descendant is not an optimal way to
989 * skip through a subtree (especially for imbalanced
990 * trees leaning to right) but that's what we have right
991 * now. More effective solution would be traversing
992 * right-up for first non-NULL without calling
993 * css_next_descendant_pre afterwards.
995 prev_css
= css_rightmost_descendant(next_css
);
998 if (css_tryget(&mem
->css
))
1001 prev_css
= next_css
;
1011 static void mem_cgroup_iter_invalidate(struct mem_cgroup
*root
)
1014 * When a group in the hierarchy below root is destroyed, the
1015 * hierarchy iterator can no longer be trusted since it might
1016 * have pointed to the destroyed group. Invalidate it.
1018 atomic_inc(&root
->dead_count
);
1021 static struct mem_cgroup
*
1022 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter
*iter
,
1023 struct mem_cgroup
*root
,
1026 struct mem_cgroup
*position
= NULL
;
1028 * A cgroup destruction happens in two stages: offlining and
1029 * release. They are separated by a RCU grace period.
1031 * If the iterator is valid, we may still race with an
1032 * offlining. The RCU lock ensures the object won't be
1033 * released, tryget will fail if we lost the race.
1035 *sequence
= atomic_read(&root
->dead_count
);
1036 if (iter
->last_dead_count
== *sequence
) {
1038 position
= iter
->last_visited
;
1039 if (position
&& !css_tryget(&position
->css
))
1045 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter
*iter
,
1046 struct mem_cgroup
*last_visited
,
1047 struct mem_cgroup
*new_position
,
1051 css_put(&last_visited
->css
);
1053 * We store the sequence count from the time @last_visited was
1054 * loaded successfully instead of rereading it here so that we
1055 * don't lose destruction events in between. We could have
1056 * raced with the destruction of @new_position after all.
1058 iter
->last_visited
= new_position
;
1060 iter
->last_dead_count
= sequence
;
1064 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1065 * @root: hierarchy root
1066 * @prev: previously returned memcg, NULL on first invocation
1067 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1068 * @cond: filter for visited nodes, NULL for no filter
1070 * Returns references to children of the hierarchy below @root, or
1071 * @root itself, or %NULL after a full round-trip.
1073 * Caller must pass the return value in @prev on subsequent
1074 * invocations for reference counting, or use mem_cgroup_iter_break()
1075 * to cancel a hierarchy walk before the round-trip is complete.
1077 * Reclaimers can specify a zone and a priority level in @reclaim to
1078 * divide up the memcgs in the hierarchy among all concurrent
1079 * reclaimers operating on the same zone and priority.
1081 struct mem_cgroup
*mem_cgroup_iter_cond(struct mem_cgroup
*root
,
1082 struct mem_cgroup
*prev
,
1083 struct mem_cgroup_reclaim_cookie
*reclaim
,
1084 mem_cgroup_iter_filter cond
)
1086 struct mem_cgroup
*memcg
= NULL
;
1087 struct mem_cgroup
*last_visited
= NULL
;
1089 if (mem_cgroup_disabled()) {
1090 /* first call must return non-NULL, second return NULL */
1091 return (struct mem_cgroup
*)(unsigned long)!prev
;
1095 root
= root_mem_cgroup
;
1097 if (prev
&& !reclaim
)
1098 last_visited
= prev
;
1100 if (!root
->use_hierarchy
&& root
!= root_mem_cgroup
) {
1103 if (mem_cgroup_filter(root
, root
, cond
) == VISIT
)
1110 struct mem_cgroup_reclaim_iter
*uninitialized_var(iter
);
1111 int uninitialized_var(seq
);
1114 int nid
= zone_to_nid(reclaim
->zone
);
1115 int zid
= zone_idx(reclaim
->zone
);
1116 struct mem_cgroup_per_zone
*mz
;
1118 mz
= mem_cgroup_zoneinfo(root
, nid
, zid
);
1119 iter
= &mz
->reclaim_iter
[reclaim
->priority
];
1120 if (prev
&& reclaim
->generation
!= iter
->generation
) {
1121 iter
->last_visited
= NULL
;
1125 last_visited
= mem_cgroup_iter_load(iter
, root
, &seq
);
1128 memcg
= __mem_cgroup_iter_next(root
, last_visited
, cond
);
1131 mem_cgroup_iter_update(iter
, last_visited
, memcg
, seq
);
1135 else if (!prev
&& memcg
)
1136 reclaim
->generation
= iter
->generation
;
1140 * We have finished the whole tree walk or no group has been
1141 * visited because filter told us to skip the root node.
1143 if (!memcg
&& (prev
|| (cond
&& !last_visited
)))
1149 if (prev
&& prev
!= root
)
1150 css_put(&prev
->css
);
1156 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1157 * @root: hierarchy root
1158 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1160 void mem_cgroup_iter_break(struct mem_cgroup
*root
,
1161 struct mem_cgroup
*prev
)
1164 root
= root_mem_cgroup
;
1165 if (prev
&& prev
!= root
)
1166 css_put(&prev
->css
);
1170 * Iteration constructs for visiting all cgroups (under a tree). If
1171 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1172 * be used for reference counting.
1174 #define for_each_mem_cgroup_tree(iter, root) \
1175 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1177 iter = mem_cgroup_iter(root, iter, NULL))
1179 #define for_each_mem_cgroup(iter) \
1180 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1182 iter = mem_cgroup_iter(NULL, iter, NULL))
1184 void __mem_cgroup_count_vm_event(struct mm_struct
*mm
, enum vm_event_item idx
)
1186 struct mem_cgroup
*memcg
;
1189 memcg
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
1190 if (unlikely(!memcg
))
1195 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGFAULT
]);
1198 this_cpu_inc(memcg
->stat
->events
[MEM_CGROUP_EVENTS_PGMAJFAULT
]);
1206 EXPORT_SYMBOL(__mem_cgroup_count_vm_event
);
1209 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1210 * @zone: zone of the wanted lruvec
1211 * @memcg: memcg of the wanted lruvec
1213 * Returns the lru list vector holding pages for the given @zone and
1214 * @mem. This can be the global zone lruvec, if the memory controller
1217 struct lruvec
*mem_cgroup_zone_lruvec(struct zone
*zone
,
1218 struct mem_cgroup
*memcg
)
1220 struct mem_cgroup_per_zone
*mz
;
1221 struct lruvec
*lruvec
;
1223 if (mem_cgroup_disabled()) {
1224 lruvec
= &zone
->lruvec
;
1228 mz
= mem_cgroup_zoneinfo(memcg
, zone_to_nid(zone
), zone_idx(zone
));
1229 lruvec
= &mz
->lruvec
;
1232 * Since a node can be onlined after the mem_cgroup was created,
1233 * we have to be prepared to initialize lruvec->zone here;
1234 * and if offlined then reonlined, we need to reinitialize it.
1236 if (unlikely(lruvec
->zone
!= zone
))
1237 lruvec
->zone
= zone
;
1242 * Following LRU functions are allowed to be used without PCG_LOCK.
1243 * Operations are called by routine of global LRU independently from memcg.
1244 * What we have to take care of here is validness of pc->mem_cgroup.
1246 * Changes to pc->mem_cgroup happens when
1249 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1250 * It is added to LRU before charge.
1251 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1252 * When moving account, the page is not on LRU. It's isolated.
1256 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1258 * @zone: zone of the page
1260 struct lruvec
*mem_cgroup_page_lruvec(struct page
*page
, struct zone
*zone
)
1262 struct mem_cgroup_per_zone
*mz
;
1263 struct mem_cgroup
*memcg
;
1264 struct page_cgroup
*pc
;
1265 struct lruvec
*lruvec
;
1267 if (mem_cgroup_disabled()) {
1268 lruvec
= &zone
->lruvec
;
1272 pc
= lookup_page_cgroup(page
);
1273 memcg
= pc
->mem_cgroup
;
1276 * Surreptitiously switch any uncharged offlist page to root:
1277 * an uncharged page off lru does nothing to secure
1278 * its former mem_cgroup from sudden removal.
1280 * Our caller holds lru_lock, and PageCgroupUsed is updated
1281 * under page_cgroup lock: between them, they make all uses
1282 * of pc->mem_cgroup safe.
1284 if (!PageLRU(page
) && !PageCgroupUsed(pc
) && memcg
!= root_mem_cgroup
)
1285 pc
->mem_cgroup
= memcg
= root_mem_cgroup
;
1287 mz
= page_cgroup_zoneinfo(memcg
, page
);
1288 lruvec
= &mz
->lruvec
;
1291 * Since a node can be onlined after the mem_cgroup was created,
1292 * we have to be prepared to initialize lruvec->zone here;
1293 * and if offlined then reonlined, we need to reinitialize it.
1295 if (unlikely(lruvec
->zone
!= zone
))
1296 lruvec
->zone
= zone
;
1301 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1302 * @lruvec: mem_cgroup per zone lru vector
1303 * @lru: index of lru list the page is sitting on
1304 * @nr_pages: positive when adding or negative when removing
1306 * This function must be called when a page is added to or removed from an
1309 void mem_cgroup_update_lru_size(struct lruvec
*lruvec
, enum lru_list lru
,
1312 struct mem_cgroup_per_zone
*mz
;
1313 unsigned long *lru_size
;
1315 if (mem_cgroup_disabled())
1318 mz
= container_of(lruvec
, struct mem_cgroup_per_zone
, lruvec
);
1319 lru_size
= mz
->lru_size
+ lru
;
1320 *lru_size
+= nr_pages
;
1321 VM_BUG_ON((long)(*lru_size
) < 0);
1325 * Checks whether given mem is same or in the root_mem_cgroup's
1328 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1329 struct mem_cgroup
*memcg
)
1331 if (root_memcg
== memcg
)
1333 if (!root_memcg
->use_hierarchy
|| !memcg
)
1335 return css_is_ancestor(&memcg
->css
, &root_memcg
->css
);
1338 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup
*root_memcg
,
1339 struct mem_cgroup
*memcg
)
1344 ret
= __mem_cgroup_same_or_subtree(root_memcg
, memcg
);
1349 bool task_in_mem_cgroup(struct task_struct
*task
,
1350 const struct mem_cgroup
*memcg
)
1352 struct mem_cgroup
*curr
= NULL
;
1353 struct task_struct
*p
;
1356 p
= find_lock_task_mm(task
);
1358 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
1362 * All threads may have already detached their mm's, but the oom
1363 * killer still needs to detect if they have already been oom
1364 * killed to prevent needlessly killing additional tasks.
1367 curr
= mem_cgroup_from_task(task
);
1369 css_get(&curr
->css
);
1375 * We should check use_hierarchy of "memcg" not "curr". Because checking
1376 * use_hierarchy of "curr" here make this function true if hierarchy is
1377 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1378 * hierarchy(even if use_hierarchy is disabled in "memcg").
1380 ret
= mem_cgroup_same_or_subtree(memcg
, curr
);
1381 css_put(&curr
->css
);
1385 int mem_cgroup_inactive_anon_is_low(struct lruvec
*lruvec
)
1387 unsigned long inactive_ratio
;
1388 unsigned long inactive
;
1389 unsigned long active
;
1392 inactive
= mem_cgroup_get_lru_size(lruvec
, LRU_INACTIVE_ANON
);
1393 active
= mem_cgroup_get_lru_size(lruvec
, LRU_ACTIVE_ANON
);
1395 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
1397 inactive_ratio
= int_sqrt(10 * gb
);
1401 return inactive
* inactive_ratio
< active
;
1404 #define mem_cgroup_from_res_counter(counter, member) \
1405 container_of(counter, struct mem_cgroup, member)
1408 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1409 * @memcg: the memory cgroup
1411 * Returns the maximum amount of memory @mem can be charged with, in
1414 static unsigned long mem_cgroup_margin(struct mem_cgroup
*memcg
)
1416 unsigned long long margin
;
1418 margin
= res_counter_margin(&memcg
->res
);
1419 if (do_swap_account
)
1420 margin
= min(margin
, res_counter_margin(&memcg
->memsw
));
1421 return margin
>> PAGE_SHIFT
;
1424 int mem_cgroup_swappiness(struct mem_cgroup
*memcg
)
1427 if (!css_parent(&memcg
->css
))
1428 return vm_swappiness
;
1430 return memcg
->swappiness
;
1434 * memcg->moving_account is used for checking possibility that some thread is
1435 * calling move_account(). When a thread on CPU-A starts moving pages under
1436 * a memcg, other threads should check memcg->moving_account under
1437 * rcu_read_lock(), like this:
1441 * memcg->moving_account+1 if (memcg->mocing_account)
1443 * synchronize_rcu() update something.
1448 /* for quick checking without looking up memcg */
1449 atomic_t memcg_moving __read_mostly
;
1451 static void mem_cgroup_start_move(struct mem_cgroup
*memcg
)
1453 atomic_inc(&memcg_moving
);
1454 atomic_inc(&memcg
->moving_account
);
1458 static void mem_cgroup_end_move(struct mem_cgroup
*memcg
)
1461 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1462 * We check NULL in callee rather than caller.
1465 atomic_dec(&memcg_moving
);
1466 atomic_dec(&memcg
->moving_account
);
1471 * 2 routines for checking "mem" is under move_account() or not.
1473 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1474 * is used for avoiding races in accounting. If true,
1475 * pc->mem_cgroup may be overwritten.
1477 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1478 * under hierarchy of moving cgroups. This is for
1479 * waiting at hith-memory prressure caused by "move".
1482 static bool mem_cgroup_stolen(struct mem_cgroup
*memcg
)
1484 VM_BUG_ON(!rcu_read_lock_held());
1485 return atomic_read(&memcg
->moving_account
) > 0;
1488 static bool mem_cgroup_under_move(struct mem_cgroup
*memcg
)
1490 struct mem_cgroup
*from
;
1491 struct mem_cgroup
*to
;
1494 * Unlike task_move routines, we access mc.to, mc.from not under
1495 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1497 spin_lock(&mc
.lock
);
1503 ret
= mem_cgroup_same_or_subtree(memcg
, from
)
1504 || mem_cgroup_same_or_subtree(memcg
, to
);
1506 spin_unlock(&mc
.lock
);
1510 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*memcg
)
1512 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1513 if (mem_cgroup_under_move(memcg
)) {
1515 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1516 /* moving charge context might have finished. */
1519 finish_wait(&mc
.waitq
, &wait
);
1527 * Take this lock when
1528 * - a code tries to modify page's memcg while it's USED.
1529 * - a code tries to modify page state accounting in a memcg.
1530 * see mem_cgroup_stolen(), too.
1532 static void move_lock_mem_cgroup(struct mem_cgroup
*memcg
,
1533 unsigned long *flags
)
1535 spin_lock_irqsave(&memcg
->move_lock
, *flags
);
1538 static void move_unlock_mem_cgroup(struct mem_cgroup
*memcg
,
1539 unsigned long *flags
)
1541 spin_unlock_irqrestore(&memcg
->move_lock
, *flags
);
1544 #define K(x) ((x) << (PAGE_SHIFT-10))
1546 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1547 * @memcg: The memory cgroup that went over limit
1548 * @p: Task that is going to be killed
1550 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1553 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1555 struct cgroup
*task_cgrp
;
1556 struct cgroup
*mem_cgrp
;
1558 * Need a buffer in BSS, can't rely on allocations. The code relies
1559 * on the assumption that OOM is serialized for memory controller.
1560 * If this assumption is broken, revisit this code.
1562 static char memcg_name
[PATH_MAX
];
1564 struct mem_cgroup
*iter
;
1572 mem_cgrp
= memcg
->css
.cgroup
;
1573 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1575 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1578 * Unfortunately, we are unable to convert to a useful name
1579 * But we'll still print out the usage information
1586 pr_info("Task in %s killed", memcg_name
);
1589 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1597 * Continues from above, so we don't need an KERN_ level
1599 pr_cont(" as a result of limit of %s\n", memcg_name
);
1602 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1603 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1604 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1605 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1606 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1607 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1608 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1609 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1610 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1611 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) >> 10,
1612 res_counter_read_u64(&memcg
->kmem
, RES_LIMIT
) >> 10,
1613 res_counter_read_u64(&memcg
->kmem
, RES_FAILCNT
));
1615 for_each_mem_cgroup_tree(iter
, memcg
) {
1616 pr_info("Memory cgroup stats");
1619 ret
= cgroup_path(iter
->css
.cgroup
, memcg_name
, PATH_MAX
);
1621 pr_cont(" for %s", memcg_name
);
1625 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
1626 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
1628 pr_cont(" %s:%ldKB", mem_cgroup_stat_names
[i
],
1629 K(mem_cgroup_read_stat(iter
, i
)));
1632 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
1633 pr_cont(" %s:%luKB", mem_cgroup_lru_names
[i
],
1634 K(mem_cgroup_nr_lru_pages(iter
, BIT(i
))));
1641 * This function returns the number of memcg under hierarchy tree. Returns
1642 * 1(self count) if no children.
1644 static int mem_cgroup_count_children(struct mem_cgroup
*memcg
)
1647 struct mem_cgroup
*iter
;
1649 for_each_mem_cgroup_tree(iter
, memcg
)
1655 * Return the memory (and swap, if configured) limit for a memcg.
1657 static u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1661 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1664 * Do not consider swap space if we cannot swap due to swappiness
1666 if (mem_cgroup_swappiness(memcg
)) {
1669 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1670 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1673 * If memsw is finite and limits the amount of swap space
1674 * available to this memcg, return that limit.
1676 limit
= min(limit
, memsw
);
1682 static void mem_cgroup_out_of_memory(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
1685 struct mem_cgroup
*iter
;
1686 unsigned long chosen_points
= 0;
1687 unsigned long totalpages
;
1688 unsigned int points
= 0;
1689 struct task_struct
*chosen
= NULL
;
1692 * If current has a pending SIGKILL or is exiting, then automatically
1693 * select it. The goal is to allow it to allocate so that it may
1694 * quickly exit and free its memory.
1696 if (fatal_signal_pending(current
) || current
->flags
& PF_EXITING
) {
1697 set_thread_flag(TIF_MEMDIE
);
1701 check_panic_on_oom(CONSTRAINT_MEMCG
, gfp_mask
, order
, NULL
);
1702 totalpages
= mem_cgroup_get_limit(memcg
) >> PAGE_SHIFT
? : 1;
1703 for_each_mem_cgroup_tree(iter
, memcg
) {
1704 struct css_task_iter it
;
1705 struct task_struct
*task
;
1707 css_task_iter_start(&iter
->css
, &it
);
1708 while ((task
= css_task_iter_next(&it
))) {
1709 switch (oom_scan_process_thread(task
, totalpages
, NULL
,
1711 case OOM_SCAN_SELECT
:
1713 put_task_struct(chosen
);
1715 chosen_points
= ULONG_MAX
;
1716 get_task_struct(chosen
);
1718 case OOM_SCAN_CONTINUE
:
1720 case OOM_SCAN_ABORT
:
1721 css_task_iter_end(&it
);
1722 mem_cgroup_iter_break(memcg
, iter
);
1724 put_task_struct(chosen
);
1729 points
= oom_badness(task
, memcg
, NULL
, totalpages
);
1730 if (points
> chosen_points
) {
1732 put_task_struct(chosen
);
1734 chosen_points
= points
;
1735 get_task_struct(chosen
);
1738 css_task_iter_end(&it
);
1743 points
= chosen_points
* 1000 / totalpages
;
1744 oom_kill_process(chosen
, gfp_mask
, order
, points
, totalpages
, memcg
,
1745 NULL
, "Memory cgroup out of memory");
1748 static unsigned long mem_cgroup_reclaim(struct mem_cgroup
*memcg
,
1750 unsigned long flags
)
1752 unsigned long total
= 0;
1753 bool noswap
= false;
1756 if (flags
& MEM_CGROUP_RECLAIM_NOSWAP
)
1758 if (!(flags
& MEM_CGROUP_RECLAIM_SHRINK
) && memcg
->memsw_is_minimum
)
1761 for (loop
= 0; loop
< MEM_CGROUP_MAX_RECLAIM_LOOPS
; loop
++) {
1763 drain_all_stock_async(memcg
);
1764 total
+= try_to_free_mem_cgroup_pages(memcg
, gfp_mask
, noswap
);
1766 * Allow limit shrinkers, which are triggered directly
1767 * by userspace, to catch signals and stop reclaim
1768 * after minimal progress, regardless of the margin.
1770 if (total
&& (flags
& MEM_CGROUP_RECLAIM_SHRINK
))
1772 if (mem_cgroup_margin(memcg
))
1775 * If nothing was reclaimed after two attempts, there
1776 * may be no reclaimable pages in this hierarchy.
1784 #if MAX_NUMNODES > 1
1786 * test_mem_cgroup_node_reclaimable
1787 * @memcg: the target memcg
1788 * @nid: the node ID to be checked.
1789 * @noswap : specify true here if the user wants flle only information.
1791 * This function returns whether the specified memcg contains any
1792 * reclaimable pages on a node. Returns true if there are any reclaimable
1793 * pages in the node.
1795 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup
*memcg
,
1796 int nid
, bool noswap
)
1798 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_FILE
))
1800 if (noswap
|| !total_swap_pages
)
1802 if (mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL_ANON
))
1809 * Always updating the nodemask is not very good - even if we have an empty
1810 * list or the wrong list here, we can start from some node and traverse all
1811 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1814 static void mem_cgroup_may_update_nodemask(struct mem_cgroup
*memcg
)
1818 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1819 * pagein/pageout changes since the last update.
1821 if (!atomic_read(&memcg
->numainfo_events
))
1823 if (atomic_inc_return(&memcg
->numainfo_updating
) > 1)
1826 /* make a nodemask where this memcg uses memory from */
1827 memcg
->scan_nodes
= node_states
[N_MEMORY
];
1829 for_each_node_mask(nid
, node_states
[N_MEMORY
]) {
1831 if (!test_mem_cgroup_node_reclaimable(memcg
, nid
, false))
1832 node_clear(nid
, memcg
->scan_nodes
);
1835 atomic_set(&memcg
->numainfo_events
, 0);
1836 atomic_set(&memcg
->numainfo_updating
, 0);
1840 * Selecting a node where we start reclaim from. Because what we need is just
1841 * reducing usage counter, start from anywhere is O,K. Considering
1842 * memory reclaim from current node, there are pros. and cons.
1844 * Freeing memory from current node means freeing memory from a node which
1845 * we'll use or we've used. So, it may make LRU bad. And if several threads
1846 * hit limits, it will see a contention on a node. But freeing from remote
1847 * node means more costs for memory reclaim because of memory latency.
1849 * Now, we use round-robin. Better algorithm is welcomed.
1851 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1855 mem_cgroup_may_update_nodemask(memcg
);
1856 node
= memcg
->last_scanned_node
;
1858 node
= next_node(node
, memcg
->scan_nodes
);
1859 if (node
== MAX_NUMNODES
)
1860 node
= first_node(memcg
->scan_nodes
);
1862 * We call this when we hit limit, not when pages are added to LRU.
1863 * No LRU may hold pages because all pages are UNEVICTABLE or
1864 * memcg is too small and all pages are not on LRU. In that case,
1865 * we use curret node.
1867 if (unlikely(node
== MAX_NUMNODES
))
1868 node
= numa_node_id();
1870 memcg
->last_scanned_node
= node
;
1875 int mem_cgroup_select_victim_node(struct mem_cgroup
*memcg
)
1883 * A group is eligible for the soft limit reclaim under the given root
1885 * a) it is over its soft limit
1886 * b) any parent up the hierarchy is over its soft limit
1888 * If the given group doesn't have any children over the limit then it
1889 * doesn't make any sense to iterate its subtree.
1891 enum mem_cgroup_filter_t
1892 mem_cgroup_soft_reclaim_eligible(struct mem_cgroup
*memcg
,
1893 struct mem_cgroup
*root
)
1895 struct mem_cgroup
*parent
;
1898 memcg
= root_mem_cgroup
;
1901 if (res_counter_soft_limit_excess(&memcg
->res
))
1905 * If any parent up to the root in the hierarchy is over its soft limit
1906 * then we have to obey and reclaim from this group as well.
1908 while ((parent
= parent_mem_cgroup(parent
))) {
1909 if (res_counter_soft_limit_excess(&parent
->res
))
1915 if (!atomic_read(&memcg
->children_in_excess
))
1920 static DEFINE_SPINLOCK(memcg_oom_lock
);
1923 * Check OOM-Killer is already running under our hierarchy.
1924 * If someone is running, return false.
1926 static bool mem_cgroup_oom_trylock(struct mem_cgroup
*memcg
)
1928 struct mem_cgroup
*iter
, *failed
= NULL
;
1930 spin_lock(&memcg_oom_lock
);
1932 for_each_mem_cgroup_tree(iter
, memcg
) {
1933 if (iter
->oom_lock
) {
1935 * this subtree of our hierarchy is already locked
1936 * so we cannot give a lock.
1939 mem_cgroup_iter_break(memcg
, iter
);
1942 iter
->oom_lock
= true;
1947 * OK, we failed to lock the whole subtree so we have
1948 * to clean up what we set up to the failing subtree
1950 for_each_mem_cgroup_tree(iter
, memcg
) {
1951 if (iter
== failed
) {
1952 mem_cgroup_iter_break(memcg
, iter
);
1955 iter
->oom_lock
= false;
1959 spin_unlock(&memcg_oom_lock
);
1964 static void mem_cgroup_oom_unlock(struct mem_cgroup
*memcg
)
1966 struct mem_cgroup
*iter
;
1968 spin_lock(&memcg_oom_lock
);
1969 for_each_mem_cgroup_tree(iter
, memcg
)
1970 iter
->oom_lock
= false;
1971 spin_unlock(&memcg_oom_lock
);
1974 static void mem_cgroup_mark_under_oom(struct mem_cgroup
*memcg
)
1976 struct mem_cgroup
*iter
;
1978 for_each_mem_cgroup_tree(iter
, memcg
)
1979 atomic_inc(&iter
->under_oom
);
1982 static void mem_cgroup_unmark_under_oom(struct mem_cgroup
*memcg
)
1984 struct mem_cgroup
*iter
;
1987 * When a new child is created while the hierarchy is under oom,
1988 * mem_cgroup_oom_lock() may not be called. We have to use
1989 * atomic_add_unless() here.
1991 for_each_mem_cgroup_tree(iter
, memcg
)
1992 atomic_add_unless(&iter
->under_oom
, -1, 0);
1995 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1997 struct oom_wait_info
{
1998 struct mem_cgroup
*memcg
;
2002 static int memcg_oom_wake_function(wait_queue_t
*wait
,
2003 unsigned mode
, int sync
, void *arg
)
2005 struct mem_cgroup
*wake_memcg
= (struct mem_cgroup
*)arg
;
2006 struct mem_cgroup
*oom_wait_memcg
;
2007 struct oom_wait_info
*oom_wait_info
;
2009 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
2010 oom_wait_memcg
= oom_wait_info
->memcg
;
2013 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2014 * Then we can use css_is_ancestor without taking care of RCU.
2016 if (!mem_cgroup_same_or_subtree(oom_wait_memcg
, wake_memcg
)
2017 && !mem_cgroup_same_or_subtree(wake_memcg
, oom_wait_memcg
))
2019 return autoremove_wake_function(wait
, mode
, sync
, arg
);
2022 static void memcg_wakeup_oom(struct mem_cgroup
*memcg
)
2024 atomic_inc(&memcg
->oom_wakeups
);
2025 /* for filtering, pass "memcg" as argument. */
2026 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, memcg
);
2029 static void memcg_oom_recover(struct mem_cgroup
*memcg
)
2031 if (memcg
&& atomic_read(&memcg
->under_oom
))
2032 memcg_wakeup_oom(memcg
);
2036 * try to call OOM killer
2038 static void mem_cgroup_oom(struct mem_cgroup
*memcg
, gfp_t mask
, int order
)
2043 if (!current
->memcg_oom
.may_oom
)
2046 current
->memcg_oom
.in_memcg_oom
= 1;
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 wakeups
= atomic_read(&memcg
->oom_wakeups
);
2056 mem_cgroup_mark_under_oom(memcg
);
2058 locked
= mem_cgroup_oom_trylock(memcg
);
2061 mem_cgroup_oom_notify(memcg
);
2063 if (locked
&& !memcg
->oom_kill_disable
) {
2064 mem_cgroup_unmark_under_oom(memcg
);
2065 mem_cgroup_out_of_memory(memcg
, mask
, order
);
2066 mem_cgroup_oom_unlock(memcg
);
2068 * There is no guarantee that an OOM-lock contender
2069 * sees the wakeups triggered by the OOM kill
2070 * uncharges. Wake any sleepers explicitely.
2072 memcg_oom_recover(memcg
);
2075 * A system call can just return -ENOMEM, but if this
2076 * is a page fault and somebody else is handling the
2077 * OOM already, we need to sleep on the OOM waitqueue
2078 * for this memcg until the situation is resolved.
2079 * Which can take some time because it might be
2080 * handled by a userspace task.
2082 * However, this is the charge context, which means
2083 * that we may sit on a large call stack and hold
2084 * various filesystem locks, the mmap_sem etc. and we
2085 * don't want the OOM handler to deadlock on them
2086 * while we sit here and wait. Store the current OOM
2087 * context in the task_struct, then return -ENOMEM.
2088 * At the end of the page fault handler, with the
2089 * stack unwound, pagefault_out_of_memory() will check
2090 * back with us by calling
2091 * mem_cgroup_oom_synchronize(), possibly putting the
2094 current
->memcg_oom
.oom_locked
= locked
;
2095 current
->memcg_oom
.wakeups
= wakeups
;
2096 css_get(&memcg
->css
);
2097 current
->memcg_oom
.wait_on_memcg
= memcg
;
2102 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2104 * This has to be called at the end of a page fault if the the memcg
2105 * OOM handler was enabled and the fault is returning %VM_FAULT_OOM.
2107 * Memcg supports userspace OOM handling, so failed allocations must
2108 * sleep on a waitqueue until the userspace task resolves the
2109 * situation. Sleeping directly in the charge context with all kinds
2110 * of locks held is not a good idea, instead we remember an OOM state
2111 * in the task and mem_cgroup_oom_synchronize() has to be called at
2112 * the end of the page fault to put the task to sleep and clean up the
2115 * Returns %true if an ongoing memcg OOM situation was detected and
2116 * finalized, %false otherwise.
2118 bool mem_cgroup_oom_synchronize(void)
2120 struct oom_wait_info owait
;
2121 struct mem_cgroup
*memcg
;
2123 /* OOM is global, do not handle */
2124 if (!current
->memcg_oom
.in_memcg_oom
)
2128 * We invoked the OOM killer but there is a chance that a kill
2129 * did not free up any charges. Everybody else might already
2130 * be sleeping, so restart the fault and keep the rampage
2131 * going until some charges are released.
2133 memcg
= current
->memcg_oom
.wait_on_memcg
;
2137 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
2140 owait
.memcg
= memcg
;
2141 owait
.wait
.flags
= 0;
2142 owait
.wait
.func
= memcg_oom_wake_function
;
2143 owait
.wait
.private = current
;
2144 INIT_LIST_HEAD(&owait
.wait
.task_list
);
2146 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
2147 /* Only sleep if we didn't miss any wakeups since OOM */
2148 if (atomic_read(&memcg
->oom_wakeups
) == current
->memcg_oom
.wakeups
)
2150 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
2152 mem_cgroup_unmark_under_oom(memcg
);
2153 if (current
->memcg_oom
.oom_locked
) {
2154 mem_cgroup_oom_unlock(memcg
);
2156 * There is no guarantee that an OOM-lock contender
2157 * sees the wakeups triggered by the OOM kill
2158 * uncharges. Wake any sleepers explicitely.
2160 memcg_oom_recover(memcg
);
2162 css_put(&memcg
->css
);
2163 current
->memcg_oom
.wait_on_memcg
= NULL
;
2165 current
->memcg_oom
.in_memcg_oom
= 0;
2170 * Currently used to update mapped file statistics, but the routine can be
2171 * generalized to update other statistics as well.
2173 * Notes: Race condition
2175 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2176 * it tends to be costly. But considering some conditions, we doesn't need
2177 * to do so _always_.
2179 * Considering "charge", lock_page_cgroup() is not required because all
2180 * file-stat operations happen after a page is attached to radix-tree. There
2181 * are no race with "charge".
2183 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2184 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2185 * if there are race with "uncharge". Statistics itself is properly handled
2188 * Considering "move", this is an only case we see a race. To make the race
2189 * small, we check mm->moving_account and detect there are possibility of race
2190 * If there is, we take a lock.
2193 void __mem_cgroup_begin_update_page_stat(struct page
*page
,
2194 bool *locked
, unsigned long *flags
)
2196 struct mem_cgroup
*memcg
;
2197 struct page_cgroup
*pc
;
2199 pc
= lookup_page_cgroup(page
);
2201 memcg
= pc
->mem_cgroup
;
2202 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2205 * If this memory cgroup is not under account moving, we don't
2206 * need to take move_lock_mem_cgroup(). Because we already hold
2207 * rcu_read_lock(), any calls to move_account will be delayed until
2208 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2210 if (!mem_cgroup_stolen(memcg
))
2213 move_lock_mem_cgroup(memcg
, flags
);
2214 if (memcg
!= pc
->mem_cgroup
|| !PageCgroupUsed(pc
)) {
2215 move_unlock_mem_cgroup(memcg
, flags
);
2221 void __mem_cgroup_end_update_page_stat(struct page
*page
, unsigned long *flags
)
2223 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2226 * It's guaranteed that pc->mem_cgroup never changes while
2227 * lock is held because a routine modifies pc->mem_cgroup
2228 * should take move_lock_mem_cgroup().
2230 move_unlock_mem_cgroup(pc
->mem_cgroup
, flags
);
2233 void mem_cgroup_update_page_stat(struct page
*page
,
2234 enum mem_cgroup_page_stat_item idx
, int val
)
2236 struct mem_cgroup
*memcg
;
2237 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2238 unsigned long uninitialized_var(flags
);
2240 if (mem_cgroup_disabled())
2243 memcg
= pc
->mem_cgroup
;
2244 if (unlikely(!memcg
|| !PageCgroupUsed(pc
)))
2248 case MEMCG_NR_FILE_MAPPED
:
2249 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
2255 this_cpu_add(memcg
->stat
->count
[idx
], val
);
2259 * size of first charge trial. "32" comes from vmscan.c's magic value.
2260 * TODO: maybe necessary to use big numbers in big irons.
2262 #define CHARGE_BATCH 32U
2263 struct memcg_stock_pcp
{
2264 struct mem_cgroup
*cached
; /* this never be root cgroup */
2265 unsigned int nr_pages
;
2266 struct work_struct work
;
2267 unsigned long flags
;
2268 #define FLUSHING_CACHED_CHARGE 0
2270 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
2271 static DEFINE_MUTEX(percpu_charge_mutex
);
2274 * consume_stock: Try to consume stocked charge on this cpu.
2275 * @memcg: memcg to consume from.
2276 * @nr_pages: how many pages to charge.
2278 * The charges will only happen if @memcg matches the current cpu's memcg
2279 * stock, and at least @nr_pages are available in that stock. Failure to
2280 * service an allocation will refill the stock.
2282 * returns true if successful, false otherwise.
2284 static bool consume_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2286 struct memcg_stock_pcp
*stock
;
2289 if (nr_pages
> CHARGE_BATCH
)
2292 stock
= &get_cpu_var(memcg_stock
);
2293 if (memcg
== stock
->cached
&& stock
->nr_pages
>= nr_pages
)
2294 stock
->nr_pages
-= nr_pages
;
2295 else /* need to call res_counter_charge */
2297 put_cpu_var(memcg_stock
);
2302 * Returns stocks cached in percpu to res_counter and reset cached information.
2304 static void drain_stock(struct memcg_stock_pcp
*stock
)
2306 struct mem_cgroup
*old
= stock
->cached
;
2308 if (stock
->nr_pages
) {
2309 unsigned long bytes
= stock
->nr_pages
* PAGE_SIZE
;
2311 res_counter_uncharge(&old
->res
, bytes
);
2312 if (do_swap_account
)
2313 res_counter_uncharge(&old
->memsw
, bytes
);
2314 stock
->nr_pages
= 0;
2316 stock
->cached
= NULL
;
2320 * This must be called under preempt disabled or must be called by
2321 * a thread which is pinned to local cpu.
2323 static void drain_local_stock(struct work_struct
*dummy
)
2325 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
2327 clear_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
);
2330 static void __init
memcg_stock_init(void)
2334 for_each_possible_cpu(cpu
) {
2335 struct memcg_stock_pcp
*stock
=
2336 &per_cpu(memcg_stock
, cpu
);
2337 INIT_WORK(&stock
->work
, drain_local_stock
);
2342 * Cache charges(val) which is from res_counter, to local per_cpu area.
2343 * This will be consumed by consume_stock() function, later.
2345 static void refill_stock(struct mem_cgroup
*memcg
, unsigned int nr_pages
)
2347 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
2349 if (stock
->cached
!= memcg
) { /* reset if necessary */
2351 stock
->cached
= memcg
;
2353 stock
->nr_pages
+= nr_pages
;
2354 put_cpu_var(memcg_stock
);
2358 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2359 * of the hierarchy under it. sync flag says whether we should block
2360 * until the work is done.
2362 static void drain_all_stock(struct mem_cgroup
*root_memcg
, bool sync
)
2366 /* Notify other cpus that system-wide "drain" is running */
2369 for_each_online_cpu(cpu
) {
2370 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2371 struct mem_cgroup
*memcg
;
2373 memcg
= stock
->cached
;
2374 if (!memcg
|| !stock
->nr_pages
)
2376 if (!mem_cgroup_same_or_subtree(root_memcg
, memcg
))
2378 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
)) {
2380 drain_local_stock(&stock
->work
);
2382 schedule_work_on(cpu
, &stock
->work
);
2390 for_each_online_cpu(cpu
) {
2391 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
2392 if (test_bit(FLUSHING_CACHED_CHARGE
, &stock
->flags
))
2393 flush_work(&stock
->work
);
2400 * Tries to drain stocked charges in other cpus. This function is asynchronous
2401 * and just put a work per cpu for draining localy on each cpu. Caller can
2402 * expects some charges will be back to res_counter later but cannot wait for
2405 static void drain_all_stock_async(struct mem_cgroup
*root_memcg
)
2408 * If someone calls draining, avoid adding more kworker runs.
2410 if (!mutex_trylock(&percpu_charge_mutex
))
2412 drain_all_stock(root_memcg
, false);
2413 mutex_unlock(&percpu_charge_mutex
);
2416 /* This is a synchronous drain interface. */
2417 static void drain_all_stock_sync(struct mem_cgroup
*root_memcg
)
2419 /* called when force_empty is called */
2420 mutex_lock(&percpu_charge_mutex
);
2421 drain_all_stock(root_memcg
, true);
2422 mutex_unlock(&percpu_charge_mutex
);
2426 * This function drains percpu counter value from DEAD cpu and
2427 * move it to local cpu. Note that this function can be preempted.
2429 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*memcg
, int cpu
)
2433 spin_lock(&memcg
->pcp_counter_lock
);
2434 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
2435 long x
= per_cpu(memcg
->stat
->count
[i
], cpu
);
2437 per_cpu(memcg
->stat
->count
[i
], cpu
) = 0;
2438 memcg
->nocpu_base
.count
[i
] += x
;
2440 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
2441 unsigned long x
= per_cpu(memcg
->stat
->events
[i
], cpu
);
2443 per_cpu(memcg
->stat
->events
[i
], cpu
) = 0;
2444 memcg
->nocpu_base
.events
[i
] += x
;
2446 spin_unlock(&memcg
->pcp_counter_lock
);
2449 static int memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
2450 unsigned long action
,
2453 int cpu
= (unsigned long)hcpu
;
2454 struct memcg_stock_pcp
*stock
;
2455 struct mem_cgroup
*iter
;
2457 if (action
== CPU_ONLINE
)
2460 if (action
!= CPU_DEAD
&& action
!= CPU_DEAD_FROZEN
)
2463 for_each_mem_cgroup(iter
)
2464 mem_cgroup_drain_pcp_counter(iter
, cpu
);
2466 stock
= &per_cpu(memcg_stock
, cpu
);
2472 /* See __mem_cgroup_try_charge() for details */
2474 CHARGE_OK
, /* success */
2475 CHARGE_RETRY
, /* need to retry but retry is not bad */
2476 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
2477 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
2480 static int mem_cgroup_do_charge(struct mem_cgroup
*memcg
, gfp_t gfp_mask
,
2481 unsigned int nr_pages
, unsigned int min_pages
,
2484 unsigned long csize
= nr_pages
* PAGE_SIZE
;
2485 struct mem_cgroup
*mem_over_limit
;
2486 struct res_counter
*fail_res
;
2487 unsigned long flags
= 0;
2490 ret
= res_counter_charge(&memcg
->res
, csize
, &fail_res
);
2493 if (!do_swap_account
)
2495 ret
= res_counter_charge(&memcg
->memsw
, csize
, &fail_res
);
2499 res_counter_uncharge(&memcg
->res
, csize
);
2500 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
2501 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
2503 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
2505 * Never reclaim on behalf of optional batching, retry with a
2506 * single page instead.
2508 if (nr_pages
> min_pages
)
2509 return CHARGE_RETRY
;
2511 if (!(gfp_mask
& __GFP_WAIT
))
2512 return CHARGE_WOULDBLOCK
;
2514 if (gfp_mask
& __GFP_NORETRY
)
2515 return CHARGE_NOMEM
;
2517 ret
= mem_cgroup_reclaim(mem_over_limit
, gfp_mask
, flags
);
2518 if (mem_cgroup_margin(mem_over_limit
) >= nr_pages
)
2519 return CHARGE_RETRY
;
2521 * Even though the limit is exceeded at this point, reclaim
2522 * may have been able to free some pages. Retry the charge
2523 * before killing the task.
2525 * Only for regular pages, though: huge pages are rather
2526 * unlikely to succeed so close to the limit, and we fall back
2527 * to regular pages anyway in case of failure.
2529 if (nr_pages
<= (1 << PAGE_ALLOC_COSTLY_ORDER
) && ret
)
2530 return CHARGE_RETRY
;
2533 * At task move, charge accounts can be doubly counted. So, it's
2534 * better to wait until the end of task_move if something is going on.
2536 if (mem_cgroup_wait_acct_move(mem_over_limit
))
2537 return CHARGE_RETRY
;
2540 mem_cgroup_oom(mem_over_limit
, gfp_mask
, get_order(csize
));
2542 return CHARGE_NOMEM
;
2546 * __mem_cgroup_try_charge() does
2547 * 1. detect memcg to be charged against from passed *mm and *ptr,
2548 * 2. update res_counter
2549 * 3. call memory reclaim if necessary.
2551 * In some special case, if the task is fatal, fatal_signal_pending() or
2552 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2553 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2554 * as possible without any hazards. 2: all pages should have a valid
2555 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2556 * pointer, that is treated as a charge to root_mem_cgroup.
2558 * So __mem_cgroup_try_charge() will return
2559 * 0 ... on success, filling *ptr with a valid memcg pointer.
2560 * -ENOMEM ... charge failure because of resource limits.
2561 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2563 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2564 * the oom-killer can be invoked.
2566 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
2568 unsigned int nr_pages
,
2569 struct mem_cgroup
**ptr
,
2572 unsigned int batch
= max(CHARGE_BATCH
, nr_pages
);
2573 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2574 struct mem_cgroup
*memcg
= NULL
;
2578 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2579 * in system level. So, allow to go ahead dying process in addition to
2582 if (unlikely(test_thread_flag(TIF_MEMDIE
)
2583 || fatal_signal_pending(current
)))
2587 * We always charge the cgroup the mm_struct belongs to.
2588 * The mm_struct's mem_cgroup changes on task migration if the
2589 * thread group leader migrates. It's possible that mm is not
2590 * set, if so charge the root memcg (happens for pagecache usage).
2593 *ptr
= root_mem_cgroup
;
2595 if (*ptr
) { /* css should be a valid one */
2597 if (mem_cgroup_is_root(memcg
))
2599 if (consume_stock(memcg
, nr_pages
))
2601 css_get(&memcg
->css
);
2603 struct task_struct
*p
;
2606 p
= rcu_dereference(mm
->owner
);
2608 * Because we don't have task_lock(), "p" can exit.
2609 * In that case, "memcg" can point to root or p can be NULL with
2610 * race with swapoff. Then, we have small risk of mis-accouning.
2611 * But such kind of mis-account by race always happens because
2612 * we don't have cgroup_mutex(). It's overkill and we allo that
2614 * (*) swapoff at el will charge against mm-struct not against
2615 * task-struct. So, mm->owner can be NULL.
2617 memcg
= mem_cgroup_from_task(p
);
2619 memcg
= root_mem_cgroup
;
2620 if (mem_cgroup_is_root(memcg
)) {
2624 if (consume_stock(memcg
, nr_pages
)) {
2626 * It seems dagerous to access memcg without css_get().
2627 * But considering how consume_stok works, it's not
2628 * necessary. If consume_stock success, some charges
2629 * from this memcg are cached on this cpu. So, we
2630 * don't need to call css_get()/css_tryget() before
2631 * calling consume_stock().
2636 /* after here, we may be blocked. we need to get refcnt */
2637 if (!css_tryget(&memcg
->css
)) {
2645 bool invoke_oom
= oom
&& !nr_oom_retries
;
2647 /* If killed, bypass charge */
2648 if (fatal_signal_pending(current
)) {
2649 css_put(&memcg
->css
);
2653 ret
= mem_cgroup_do_charge(memcg
, gfp_mask
, batch
,
2654 nr_pages
, invoke_oom
);
2658 case CHARGE_RETRY
: /* not in OOM situation but retry */
2660 css_put(&memcg
->css
);
2663 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2664 css_put(&memcg
->css
);
2666 case CHARGE_NOMEM
: /* OOM routine works */
2667 if (!oom
|| invoke_oom
) {
2668 css_put(&memcg
->css
);
2674 } while (ret
!= CHARGE_OK
);
2676 if (batch
> nr_pages
)
2677 refill_stock(memcg
, batch
- nr_pages
);
2678 css_put(&memcg
->css
);
2686 *ptr
= root_mem_cgroup
;
2691 * Somemtimes we have to undo a charge we got by try_charge().
2692 * This function is for that and do uncharge, put css's refcnt.
2693 * gotten by try_charge().
2695 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*memcg
,
2696 unsigned int nr_pages
)
2698 if (!mem_cgroup_is_root(memcg
)) {
2699 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2701 res_counter_uncharge(&memcg
->res
, bytes
);
2702 if (do_swap_account
)
2703 res_counter_uncharge(&memcg
->memsw
, bytes
);
2708 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2709 * This is useful when moving usage to parent cgroup.
2711 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup
*memcg
,
2712 unsigned int nr_pages
)
2714 unsigned long bytes
= nr_pages
* PAGE_SIZE
;
2716 if (mem_cgroup_is_root(memcg
))
2719 res_counter_uncharge_until(&memcg
->res
, memcg
->res
.parent
, bytes
);
2720 if (do_swap_account
)
2721 res_counter_uncharge_until(&memcg
->memsw
,
2722 memcg
->memsw
.parent
, bytes
);
2726 * A helper function to get mem_cgroup from ID. must be called under
2727 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2728 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2729 * called against removed memcg.)
2731 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2733 struct cgroup_subsys_state
*css
;
2735 /* ID 0 is unused ID */
2738 css
= css_lookup(&mem_cgroup_subsys
, id
);
2741 return mem_cgroup_from_css(css
);
2744 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2746 struct mem_cgroup
*memcg
= NULL
;
2747 struct page_cgroup
*pc
;
2751 VM_BUG_ON(!PageLocked(page
));
2753 pc
= lookup_page_cgroup(page
);
2754 lock_page_cgroup(pc
);
2755 if (PageCgroupUsed(pc
)) {
2756 memcg
= pc
->mem_cgroup
;
2757 if (memcg
&& !css_tryget(&memcg
->css
))
2759 } else if (PageSwapCache(page
)) {
2760 ent
.val
= page_private(page
);
2761 id
= lookup_swap_cgroup_id(ent
);
2763 memcg
= mem_cgroup_lookup(id
);
2764 if (memcg
&& !css_tryget(&memcg
->css
))
2768 unlock_page_cgroup(pc
);
2772 static void __mem_cgroup_commit_charge(struct mem_cgroup
*memcg
,
2774 unsigned int nr_pages
,
2775 enum charge_type ctype
,
2778 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
2779 struct zone
*uninitialized_var(zone
);
2780 struct lruvec
*lruvec
;
2781 bool was_on_lru
= false;
2784 lock_page_cgroup(pc
);
2785 VM_BUG_ON(PageCgroupUsed(pc
));
2787 * we don't need page_cgroup_lock about tail pages, becase they are not
2788 * accessed by any other context at this point.
2792 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2793 * may already be on some other mem_cgroup's LRU. Take care of it.
2796 zone
= page_zone(page
);
2797 spin_lock_irq(&zone
->lru_lock
);
2798 if (PageLRU(page
)) {
2799 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2801 del_page_from_lru_list(page
, lruvec
, page_lru(page
));
2806 pc
->mem_cgroup
= memcg
;
2808 * We access a page_cgroup asynchronously without lock_page_cgroup().
2809 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2810 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2811 * before USED bit, we need memory barrier here.
2812 * See mem_cgroup_add_lru_list(), etc.
2815 SetPageCgroupUsed(pc
);
2819 lruvec
= mem_cgroup_zone_lruvec(zone
, pc
->mem_cgroup
);
2820 VM_BUG_ON(PageLRU(page
));
2822 add_page_to_lru_list(page
, lruvec
, page_lru(page
));
2824 spin_unlock_irq(&zone
->lru_lock
);
2827 if (ctype
== MEM_CGROUP_CHARGE_TYPE_ANON
)
2832 mem_cgroup_charge_statistics(memcg
, page
, anon
, nr_pages
);
2833 unlock_page_cgroup(pc
);
2836 * "charge_statistics" updated event counter.
2838 memcg_check_events(memcg
, page
);
2841 static DEFINE_MUTEX(set_limit_mutex
);
2843 #ifdef CONFIG_MEMCG_KMEM
2844 static inline bool memcg_can_account_kmem(struct mem_cgroup
*memcg
)
2846 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg
) &&
2847 (memcg
->kmem_account_flags
& KMEM_ACCOUNTED_MASK
);
2851 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2852 * in the memcg_cache_params struct.
2854 static struct kmem_cache
*memcg_params_to_cache(struct memcg_cache_params
*p
)
2856 struct kmem_cache
*cachep
;
2858 VM_BUG_ON(p
->is_root_cache
);
2859 cachep
= p
->root_cache
;
2860 return cachep
->memcg_params
->memcg_caches
[memcg_cache_id(p
->memcg
)];
2863 #ifdef CONFIG_SLABINFO
2864 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state
*css
,
2865 struct cftype
*cft
, struct seq_file
*m
)
2867 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
2868 struct memcg_cache_params
*params
;
2870 if (!memcg_can_account_kmem(memcg
))
2873 print_slabinfo_header(m
);
2875 mutex_lock(&memcg
->slab_caches_mutex
);
2876 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
)
2877 cache_show(memcg_params_to_cache(params
), m
);
2878 mutex_unlock(&memcg
->slab_caches_mutex
);
2884 static int memcg_charge_kmem(struct mem_cgroup
*memcg
, gfp_t gfp
, u64 size
)
2886 struct res_counter
*fail_res
;
2887 struct mem_cgroup
*_memcg
;
2891 ret
= res_counter_charge(&memcg
->kmem
, size
, &fail_res
);
2896 * Conditions under which we can wait for the oom_killer. Those are
2897 * the same conditions tested by the core page allocator
2899 may_oom
= (gfp
& __GFP_FS
) && !(gfp
& __GFP_NORETRY
);
2902 ret
= __mem_cgroup_try_charge(NULL
, gfp
, size
>> PAGE_SHIFT
,
2905 if (ret
== -EINTR
) {
2907 * __mem_cgroup_try_charge() chosed to bypass to root due to
2908 * OOM kill or fatal signal. Since our only options are to
2909 * either fail the allocation or charge it to this cgroup, do
2910 * it as a temporary condition. But we can't fail. From a
2911 * kmem/slab perspective, the cache has already been selected,
2912 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2915 * This condition will only trigger if the task entered
2916 * memcg_charge_kmem in a sane state, but was OOM-killed during
2917 * __mem_cgroup_try_charge() above. Tasks that were already
2918 * dying when the allocation triggers should have been already
2919 * directed to the root cgroup in memcontrol.h
2921 res_counter_charge_nofail(&memcg
->res
, size
, &fail_res
);
2922 if (do_swap_account
)
2923 res_counter_charge_nofail(&memcg
->memsw
, size
,
2927 res_counter_uncharge(&memcg
->kmem
, size
);
2932 static void memcg_uncharge_kmem(struct mem_cgroup
*memcg
, u64 size
)
2934 res_counter_uncharge(&memcg
->res
, size
);
2935 if (do_swap_account
)
2936 res_counter_uncharge(&memcg
->memsw
, size
);
2939 if (res_counter_uncharge(&memcg
->kmem
, size
))
2943 * Releases a reference taken in kmem_cgroup_css_offline in case
2944 * this last uncharge is racing with the offlining code or it is
2945 * outliving the memcg existence.
2947 * The memory barrier imposed by test&clear is paired with the
2948 * explicit one in memcg_kmem_mark_dead().
2950 if (memcg_kmem_test_and_clear_dead(memcg
))
2951 css_put(&memcg
->css
);
2954 void memcg_cache_list_add(struct mem_cgroup
*memcg
, struct kmem_cache
*cachep
)
2959 mutex_lock(&memcg
->slab_caches_mutex
);
2960 list_add(&cachep
->memcg_params
->list
, &memcg
->memcg_slab_caches
);
2961 mutex_unlock(&memcg
->slab_caches_mutex
);
2965 * helper for acessing a memcg's index. It will be used as an index in the
2966 * child cache array in kmem_cache, and also to derive its name. This function
2967 * will return -1 when this is not a kmem-limited memcg.
2969 int memcg_cache_id(struct mem_cgroup
*memcg
)
2971 return memcg
? memcg
->kmemcg_id
: -1;
2975 * This ends up being protected by the set_limit mutex, during normal
2976 * operation, because that is its main call site.
2978 * But when we create a new cache, we can call this as well if its parent
2979 * is kmem-limited. That will have to hold set_limit_mutex as well.
2981 int memcg_update_cache_sizes(struct mem_cgroup
*memcg
)
2985 num
= ida_simple_get(&kmem_limited_groups
,
2986 0, MEMCG_CACHES_MAX_SIZE
, GFP_KERNEL
);
2990 * After this point, kmem_accounted (that we test atomically in
2991 * the beginning of this conditional), is no longer 0. This
2992 * guarantees only one process will set the following boolean
2993 * to true. We don't need test_and_set because we're protected
2994 * by the set_limit_mutex anyway.
2996 memcg_kmem_set_activated(memcg
);
2998 ret
= memcg_update_all_caches(num
+1);
3000 ida_simple_remove(&kmem_limited_groups
, num
);
3001 memcg_kmem_clear_activated(memcg
);
3005 memcg
->kmemcg_id
= num
;
3006 INIT_LIST_HEAD(&memcg
->memcg_slab_caches
);
3007 mutex_init(&memcg
->slab_caches_mutex
);
3011 static size_t memcg_caches_array_size(int num_groups
)
3014 if (num_groups
<= 0)
3017 size
= 2 * num_groups
;
3018 if (size
< MEMCG_CACHES_MIN_SIZE
)
3019 size
= MEMCG_CACHES_MIN_SIZE
;
3020 else if (size
> MEMCG_CACHES_MAX_SIZE
)
3021 size
= MEMCG_CACHES_MAX_SIZE
;
3027 * We should update the current array size iff all caches updates succeed. This
3028 * can only be done from the slab side. The slab mutex needs to be held when
3031 void memcg_update_array_size(int num
)
3033 if (num
> memcg_limited_groups_array_size
)
3034 memcg_limited_groups_array_size
= memcg_caches_array_size(num
);
3037 static void kmem_cache_destroy_work_func(struct work_struct
*w
);
3039 int memcg_update_cache_size(struct kmem_cache
*s
, int num_groups
)
3041 struct memcg_cache_params
*cur_params
= s
->memcg_params
;
3043 VM_BUG_ON(s
->memcg_params
&& !s
->memcg_params
->is_root_cache
);
3045 if (num_groups
> memcg_limited_groups_array_size
) {
3047 ssize_t size
= memcg_caches_array_size(num_groups
);
3049 size
*= sizeof(void *);
3050 size
+= offsetof(struct memcg_cache_params
, memcg_caches
);
3052 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3053 if (!s
->memcg_params
) {
3054 s
->memcg_params
= cur_params
;
3058 s
->memcg_params
->is_root_cache
= true;
3061 * There is the chance it will be bigger than
3062 * memcg_limited_groups_array_size, if we failed an allocation
3063 * in a cache, in which case all caches updated before it, will
3064 * have a bigger array.
3066 * But if that is the case, the data after
3067 * memcg_limited_groups_array_size is certainly unused
3069 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3070 if (!cur_params
->memcg_caches
[i
])
3072 s
->memcg_params
->memcg_caches
[i
] =
3073 cur_params
->memcg_caches
[i
];
3077 * Ideally, we would wait until all caches succeed, and only
3078 * then free the old one. But this is not worth the extra
3079 * pointer per-cache we'd have to have for this.
3081 * It is not a big deal if some caches are left with a size
3082 * bigger than the others. And all updates will reset this
3090 int memcg_register_cache(struct mem_cgroup
*memcg
, struct kmem_cache
*s
,
3091 struct kmem_cache
*root_cache
)
3095 if (!memcg_kmem_enabled())
3099 size
= offsetof(struct memcg_cache_params
, memcg_caches
);
3100 size
+= memcg_limited_groups_array_size
* sizeof(void *);
3102 size
= sizeof(struct memcg_cache_params
);
3104 s
->memcg_params
= kzalloc(size
, GFP_KERNEL
);
3105 if (!s
->memcg_params
)
3109 s
->memcg_params
->memcg
= memcg
;
3110 s
->memcg_params
->root_cache
= root_cache
;
3111 INIT_WORK(&s
->memcg_params
->destroy
,
3112 kmem_cache_destroy_work_func
);
3114 s
->memcg_params
->is_root_cache
= true;
3119 void memcg_release_cache(struct kmem_cache
*s
)
3121 struct kmem_cache
*root
;
3122 struct mem_cgroup
*memcg
;
3126 * This happens, for instance, when a root cache goes away before we
3129 if (!s
->memcg_params
)
3132 if (s
->memcg_params
->is_root_cache
)
3135 memcg
= s
->memcg_params
->memcg
;
3136 id
= memcg_cache_id(memcg
);
3138 root
= s
->memcg_params
->root_cache
;
3139 root
->memcg_params
->memcg_caches
[id
] = NULL
;
3141 mutex_lock(&memcg
->slab_caches_mutex
);
3142 list_del(&s
->memcg_params
->list
);
3143 mutex_unlock(&memcg
->slab_caches_mutex
);
3145 css_put(&memcg
->css
);
3147 kfree(s
->memcg_params
);
3151 * During the creation a new cache, we need to disable our accounting mechanism
3152 * altogether. This is true even if we are not creating, but rather just
3153 * enqueing new caches to be created.
3155 * This is because that process will trigger allocations; some visible, like
3156 * explicit kmallocs to auxiliary data structures, name strings and internal
3157 * cache structures; some well concealed, like INIT_WORK() that can allocate
3158 * objects during debug.
3160 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3161 * to it. This may not be a bounded recursion: since the first cache creation
3162 * failed to complete (waiting on the allocation), we'll just try to create the
3163 * cache again, failing at the same point.
3165 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3166 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3167 * inside the following two functions.
3169 static inline void memcg_stop_kmem_account(void)
3171 VM_BUG_ON(!current
->mm
);
3172 current
->memcg_kmem_skip_account
++;
3175 static inline void memcg_resume_kmem_account(void)
3177 VM_BUG_ON(!current
->mm
);
3178 current
->memcg_kmem_skip_account
--;
3181 static void kmem_cache_destroy_work_func(struct work_struct
*w
)
3183 struct kmem_cache
*cachep
;
3184 struct memcg_cache_params
*p
;
3186 p
= container_of(w
, struct memcg_cache_params
, destroy
);
3188 cachep
= memcg_params_to_cache(p
);
3191 * If we get down to 0 after shrink, we could delete right away.
3192 * However, memcg_release_pages() already puts us back in the workqueue
3193 * in that case. If we proceed deleting, we'll get a dangling
3194 * reference, and removing the object from the workqueue in that case
3195 * is unnecessary complication. We are not a fast path.
3197 * Note that this case is fundamentally different from racing with
3198 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3199 * kmem_cache_shrink, not only we would be reinserting a dead cache
3200 * into the queue, but doing so from inside the worker racing to
3203 * So if we aren't down to zero, we'll just schedule a worker and try
3206 if (atomic_read(&cachep
->memcg_params
->nr_pages
) != 0) {
3207 kmem_cache_shrink(cachep
);
3208 if (atomic_read(&cachep
->memcg_params
->nr_pages
) == 0)
3211 kmem_cache_destroy(cachep
);
3214 void mem_cgroup_destroy_cache(struct kmem_cache
*cachep
)
3216 if (!cachep
->memcg_params
->dead
)
3220 * There are many ways in which we can get here.
3222 * We can get to a memory-pressure situation while the delayed work is
3223 * still pending to run. The vmscan shrinkers can then release all
3224 * cache memory and get us to destruction. If this is the case, we'll
3225 * be executed twice, which is a bug (the second time will execute over
3226 * bogus data). In this case, cancelling the work should be fine.
3228 * But we can also get here from the worker itself, if
3229 * kmem_cache_shrink is enough to shake all the remaining objects and
3230 * get the page count to 0. In this case, we'll deadlock if we try to
3231 * cancel the work (the worker runs with an internal lock held, which
3232 * is the same lock we would hold for cancel_work_sync().)
3234 * Since we can't possibly know who got us here, just refrain from
3235 * running if there is already work pending
3237 if (work_pending(&cachep
->memcg_params
->destroy
))
3240 * We have to defer the actual destroying to a workqueue, because
3241 * we might currently be in a context that cannot sleep.
3243 schedule_work(&cachep
->memcg_params
->destroy
);
3247 * This lock protects updaters, not readers. We want readers to be as fast as
3248 * they can, and they will either see NULL or a valid cache value. Our model
3249 * allow them to see NULL, in which case the root memcg will be selected.
3251 * We need this lock because multiple allocations to the same cache from a non
3252 * will span more than one worker. Only one of them can create the cache.
3254 static DEFINE_MUTEX(memcg_cache_mutex
);
3257 * Called with memcg_cache_mutex held
3259 static struct kmem_cache
*kmem_cache_dup(struct mem_cgroup
*memcg
,
3260 struct kmem_cache
*s
)
3262 struct kmem_cache
*new;
3263 static char *tmp_name
= NULL
;
3265 lockdep_assert_held(&memcg_cache_mutex
);
3268 * kmem_cache_create_memcg duplicates the given name and
3269 * cgroup_name for this name requires RCU context.
3270 * This static temporary buffer is used to prevent from
3271 * pointless shortliving allocation.
3274 tmp_name
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3280 snprintf(tmp_name
, PATH_MAX
, "%s(%d:%s)", s
->name
,
3281 memcg_cache_id(memcg
), cgroup_name(memcg
->css
.cgroup
));
3284 new = kmem_cache_create_memcg(memcg
, tmp_name
, s
->object_size
, s
->align
,
3285 (s
->flags
& ~SLAB_PANIC
), s
->ctor
, s
);
3288 new->allocflags
|= __GFP_KMEMCG
;
3293 static struct kmem_cache
*memcg_create_kmem_cache(struct mem_cgroup
*memcg
,
3294 struct kmem_cache
*cachep
)
3296 struct kmem_cache
*new_cachep
;
3299 BUG_ON(!memcg_can_account_kmem(memcg
));
3301 idx
= memcg_cache_id(memcg
);
3303 mutex_lock(&memcg_cache_mutex
);
3304 new_cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3306 css_put(&memcg
->css
);
3310 new_cachep
= kmem_cache_dup(memcg
, cachep
);
3311 if (new_cachep
== NULL
) {
3312 new_cachep
= cachep
;
3313 css_put(&memcg
->css
);
3317 atomic_set(&new_cachep
->memcg_params
->nr_pages
, 0);
3319 cachep
->memcg_params
->memcg_caches
[idx
] = new_cachep
;
3321 * the readers won't lock, make sure everybody sees the updated value,
3322 * so they won't put stuff in the queue again for no reason
3326 mutex_unlock(&memcg_cache_mutex
);
3330 void kmem_cache_destroy_memcg_children(struct kmem_cache
*s
)
3332 struct kmem_cache
*c
;
3335 if (!s
->memcg_params
)
3337 if (!s
->memcg_params
->is_root_cache
)
3341 * If the cache is being destroyed, we trust that there is no one else
3342 * requesting objects from it. Even if there are, the sanity checks in
3343 * kmem_cache_destroy should caught this ill-case.
3345 * Still, we don't want anyone else freeing memcg_caches under our
3346 * noses, which can happen if a new memcg comes to life. As usual,
3347 * we'll take the set_limit_mutex to protect ourselves against this.
3349 mutex_lock(&set_limit_mutex
);
3350 for (i
= 0; i
< memcg_limited_groups_array_size
; i
++) {
3351 c
= s
->memcg_params
->memcg_caches
[i
];
3356 * We will now manually delete the caches, so to avoid races
3357 * we need to cancel all pending destruction workers and
3358 * proceed with destruction ourselves.
3360 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3361 * and that could spawn the workers again: it is likely that
3362 * the cache still have active pages until this very moment.
3363 * This would lead us back to mem_cgroup_destroy_cache.
3365 * But that will not execute at all if the "dead" flag is not
3366 * set, so flip it down to guarantee we are in control.
3368 c
->memcg_params
->dead
= false;
3369 cancel_work_sync(&c
->memcg_params
->destroy
);
3370 kmem_cache_destroy(c
);
3372 mutex_unlock(&set_limit_mutex
);
3375 struct create_work
{
3376 struct mem_cgroup
*memcg
;
3377 struct kmem_cache
*cachep
;
3378 struct work_struct work
;
3381 static void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3383 struct kmem_cache
*cachep
;
3384 struct memcg_cache_params
*params
;
3386 if (!memcg_kmem_is_active(memcg
))
3389 mutex_lock(&memcg
->slab_caches_mutex
);
3390 list_for_each_entry(params
, &memcg
->memcg_slab_caches
, list
) {
3391 cachep
= memcg_params_to_cache(params
);
3392 cachep
->memcg_params
->dead
= true;
3393 schedule_work(&cachep
->memcg_params
->destroy
);
3395 mutex_unlock(&memcg
->slab_caches_mutex
);
3398 static void memcg_create_cache_work_func(struct work_struct
*w
)
3400 struct create_work
*cw
;
3402 cw
= container_of(w
, struct create_work
, work
);
3403 memcg_create_kmem_cache(cw
->memcg
, cw
->cachep
);
3408 * Enqueue the creation of a per-memcg kmem_cache.
3410 static void __memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3411 struct kmem_cache
*cachep
)
3413 struct create_work
*cw
;
3415 cw
= kmalloc(sizeof(struct create_work
), GFP_NOWAIT
);
3417 css_put(&memcg
->css
);
3422 cw
->cachep
= cachep
;
3424 INIT_WORK(&cw
->work
, memcg_create_cache_work_func
);
3425 schedule_work(&cw
->work
);
3428 static void memcg_create_cache_enqueue(struct mem_cgroup
*memcg
,
3429 struct kmem_cache
*cachep
)
3432 * We need to stop accounting when we kmalloc, because if the
3433 * corresponding kmalloc cache is not yet created, the first allocation
3434 * in __memcg_create_cache_enqueue will recurse.
3436 * However, it is better to enclose the whole function. Depending on
3437 * the debugging options enabled, INIT_WORK(), for instance, can
3438 * trigger an allocation. This too, will make us recurse. Because at
3439 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3440 * the safest choice is to do it like this, wrapping the whole function.
3442 memcg_stop_kmem_account();
3443 __memcg_create_cache_enqueue(memcg
, cachep
);
3444 memcg_resume_kmem_account();
3447 * Return the kmem_cache we're supposed to use for a slab allocation.
3448 * We try to use the current memcg's version of the cache.
3450 * If the cache does not exist yet, if we are the first user of it,
3451 * we either create it immediately, if possible, or create it asynchronously
3453 * In the latter case, we will let the current allocation go through with
3454 * the original cache.
3456 * Can't be called in interrupt context or from kernel threads.
3457 * This function needs to be called with rcu_read_lock() held.
3459 struct kmem_cache
*__memcg_kmem_get_cache(struct kmem_cache
*cachep
,
3462 struct mem_cgroup
*memcg
;
3465 VM_BUG_ON(!cachep
->memcg_params
);
3466 VM_BUG_ON(!cachep
->memcg_params
->is_root_cache
);
3468 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3472 memcg
= mem_cgroup_from_task(rcu_dereference(current
->mm
->owner
));
3474 if (!memcg_can_account_kmem(memcg
))
3477 idx
= memcg_cache_id(memcg
);
3480 * barrier to mare sure we're always seeing the up to date value. The
3481 * code updating memcg_caches will issue a write barrier to match this.
3483 read_barrier_depends();
3484 if (likely(cachep
->memcg_params
->memcg_caches
[idx
])) {
3485 cachep
= cachep
->memcg_params
->memcg_caches
[idx
];
3489 /* The corresponding put will be done in the workqueue. */
3490 if (!css_tryget(&memcg
->css
))
3495 * If we are in a safe context (can wait, and not in interrupt
3496 * context), we could be be predictable and return right away.
3497 * This would guarantee that the allocation being performed
3498 * already belongs in the new cache.
3500 * However, there are some clashes that can arrive from locking.
3501 * For instance, because we acquire the slab_mutex while doing
3502 * kmem_cache_dup, this means no further allocation could happen
3503 * with the slab_mutex held.
3505 * Also, because cache creation issue get_online_cpus(), this
3506 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3507 * that ends up reversed during cpu hotplug. (cpuset allocates
3508 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3509 * better to defer everything.
3511 memcg_create_cache_enqueue(memcg
, cachep
);
3517 EXPORT_SYMBOL(__memcg_kmem_get_cache
);
3520 * We need to verify if the allocation against current->mm->owner's memcg is
3521 * possible for the given order. But the page is not allocated yet, so we'll
3522 * need a further commit step to do the final arrangements.
3524 * It is possible for the task to switch cgroups in this mean time, so at
3525 * commit time, we can't rely on task conversion any longer. We'll then use
3526 * the handle argument to return to the caller which cgroup we should commit
3527 * against. We could also return the memcg directly and avoid the pointer
3528 * passing, but a boolean return value gives better semantics considering
3529 * the compiled-out case as well.
3531 * Returning true means the allocation is possible.
3534 __memcg_kmem_newpage_charge(gfp_t gfp
, struct mem_cgroup
**_memcg
, int order
)
3536 struct mem_cgroup
*memcg
;
3542 * Disabling accounting is only relevant for some specific memcg
3543 * internal allocations. Therefore we would initially not have such
3544 * check here, since direct calls to the page allocator that are marked
3545 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3546 * concerned with cache allocations, and by having this test at
3547 * memcg_kmem_get_cache, we are already able to relay the allocation to
3548 * the root cache and bypass the memcg cache altogether.
3550 * There is one exception, though: the SLUB allocator does not create
3551 * large order caches, but rather service large kmallocs directly from
3552 * the page allocator. Therefore, the following sequence when backed by
3553 * the SLUB allocator:
3555 * memcg_stop_kmem_account();
3556 * kmalloc(<large_number>)
3557 * memcg_resume_kmem_account();
3559 * would effectively ignore the fact that we should skip accounting,
3560 * since it will drive us directly to this function without passing
3561 * through the cache selector memcg_kmem_get_cache. Such large
3562 * allocations are extremely rare but can happen, for instance, for the
3563 * cache arrays. We bring this test here.
3565 if (!current
->mm
|| current
->memcg_kmem_skip_account
)
3568 memcg
= try_get_mem_cgroup_from_mm(current
->mm
);
3571 * very rare case described in mem_cgroup_from_task. Unfortunately there
3572 * isn't much we can do without complicating this too much, and it would
3573 * be gfp-dependent anyway. Just let it go
3575 if (unlikely(!memcg
))
3578 if (!memcg_can_account_kmem(memcg
)) {
3579 css_put(&memcg
->css
);
3583 ret
= memcg_charge_kmem(memcg
, gfp
, PAGE_SIZE
<< order
);
3587 css_put(&memcg
->css
);
3591 void __memcg_kmem_commit_charge(struct page
*page
, struct mem_cgroup
*memcg
,
3594 struct page_cgroup
*pc
;
3596 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3598 /* The page allocation failed. Revert */
3600 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3604 pc
= lookup_page_cgroup(page
);
3605 lock_page_cgroup(pc
);
3606 pc
->mem_cgroup
= memcg
;
3607 SetPageCgroupUsed(pc
);
3608 unlock_page_cgroup(pc
);
3611 void __memcg_kmem_uncharge_pages(struct page
*page
, int order
)
3613 struct mem_cgroup
*memcg
= NULL
;
3614 struct page_cgroup
*pc
;
3617 pc
= lookup_page_cgroup(page
);
3619 * Fast unlocked return. Theoretically might have changed, have to
3620 * check again after locking.
3622 if (!PageCgroupUsed(pc
))
3625 lock_page_cgroup(pc
);
3626 if (PageCgroupUsed(pc
)) {
3627 memcg
= pc
->mem_cgroup
;
3628 ClearPageCgroupUsed(pc
);
3630 unlock_page_cgroup(pc
);
3633 * We trust that only if there is a memcg associated with the page, it
3634 * is a valid allocation
3639 VM_BUG_ON(mem_cgroup_is_root(memcg
));
3640 memcg_uncharge_kmem(memcg
, PAGE_SIZE
<< order
);
3643 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup
*memcg
)
3646 #endif /* CONFIG_MEMCG_KMEM */
3648 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3650 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3652 * Because tail pages are not marked as "used", set it. We're under
3653 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3654 * charge/uncharge will be never happen and move_account() is done under
3655 * compound_lock(), so we don't have to take care of races.
3657 void mem_cgroup_split_huge_fixup(struct page
*head
)
3659 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
3660 struct page_cgroup
*pc
;
3661 struct mem_cgroup
*memcg
;
3664 if (mem_cgroup_disabled())
3667 memcg
= head_pc
->mem_cgroup
;
3668 for (i
= 1; i
< HPAGE_PMD_NR
; i
++) {
3670 pc
->mem_cgroup
= memcg
;
3671 smp_wmb();/* see __commit_charge() */
3672 pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
3674 __this_cpu_sub(memcg
->stat
->count
[MEM_CGROUP_STAT_RSS_HUGE
],
3677 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3680 * mem_cgroup_move_account - move account of the page
3682 * @nr_pages: number of regular pages (>1 for huge pages)
3683 * @pc: page_cgroup of the page.
3684 * @from: mem_cgroup which the page is moved from.
3685 * @to: mem_cgroup which the page is moved to. @from != @to.
3687 * The caller must confirm following.
3688 * - page is not on LRU (isolate_page() is useful.)
3689 * - compound_lock is held when nr_pages > 1
3691 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3694 static int mem_cgroup_move_account(struct page
*page
,
3695 unsigned int nr_pages
,
3696 struct page_cgroup
*pc
,
3697 struct mem_cgroup
*from
,
3698 struct mem_cgroup
*to
)
3700 unsigned long flags
;
3702 bool anon
= PageAnon(page
);
3704 VM_BUG_ON(from
== to
);
3705 VM_BUG_ON(PageLRU(page
));
3707 * The page is isolated from LRU. So, collapse function
3708 * will not handle this page. But page splitting can happen.
3709 * Do this check under compound_page_lock(). The caller should
3713 if (nr_pages
> 1 && !PageTransHuge(page
))
3716 lock_page_cgroup(pc
);
3719 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
3722 move_lock_mem_cgroup(from
, &flags
);
3724 if (!anon
&& page_mapped(page
)) {
3725 /* Update mapped_file data for mem_cgroup */
3727 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3728 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
3731 mem_cgroup_charge_statistics(from
, page
, anon
, -nr_pages
);
3733 /* caller should have done css_get */
3734 pc
->mem_cgroup
= to
;
3735 mem_cgroup_charge_statistics(to
, page
, anon
, nr_pages
);
3736 move_unlock_mem_cgroup(from
, &flags
);
3739 unlock_page_cgroup(pc
);
3743 memcg_check_events(to
, page
);
3744 memcg_check_events(from
, page
);
3750 * mem_cgroup_move_parent - moves page to the parent group
3751 * @page: the page to move
3752 * @pc: page_cgroup of the page
3753 * @child: page's cgroup
3755 * move charges to its parent or the root cgroup if the group has no
3756 * parent (aka use_hierarchy==0).
3757 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3758 * mem_cgroup_move_account fails) the failure is always temporary and
3759 * it signals a race with a page removal/uncharge or migration. In the
3760 * first case the page is on the way out and it will vanish from the LRU
3761 * on the next attempt and the call should be retried later.
3762 * Isolation from the LRU fails only if page has been isolated from
3763 * the LRU since we looked at it and that usually means either global
3764 * reclaim or migration going on. The page will either get back to the
3766 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3767 * (!PageCgroupUsed) or moved to a different group. The page will
3768 * disappear in the next attempt.
3770 static int mem_cgroup_move_parent(struct page
*page
,
3771 struct page_cgroup
*pc
,
3772 struct mem_cgroup
*child
)
3774 struct mem_cgroup
*parent
;
3775 unsigned int nr_pages
;
3776 unsigned long uninitialized_var(flags
);
3779 VM_BUG_ON(mem_cgroup_is_root(child
));
3782 if (!get_page_unless_zero(page
))
3784 if (isolate_lru_page(page
))
3787 nr_pages
= hpage_nr_pages(page
);
3789 parent
= parent_mem_cgroup(child
);
3791 * If no parent, move charges to root cgroup.
3794 parent
= root_mem_cgroup
;
3797 VM_BUG_ON(!PageTransHuge(page
));
3798 flags
= compound_lock_irqsave(page
);
3801 ret
= mem_cgroup_move_account(page
, nr_pages
,
3804 __mem_cgroup_cancel_local_charge(child
, nr_pages
);
3807 compound_unlock_irqrestore(page
, flags
);
3808 putback_lru_page(page
);
3816 * Charge the memory controller for page usage.
3818 * 0 if the charge was successful
3819 * < 0 if the cgroup is over its limit
3821 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
3822 gfp_t gfp_mask
, enum charge_type ctype
)
3824 struct mem_cgroup
*memcg
= NULL
;
3825 unsigned int nr_pages
= 1;
3829 if (PageTransHuge(page
)) {
3830 nr_pages
<<= compound_order(page
);
3831 VM_BUG_ON(!PageTransHuge(page
));
3833 * Never OOM-kill a process for a huge page. The
3834 * fault handler will fall back to regular pages.
3839 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, nr_pages
, &memcg
, oom
);
3842 __mem_cgroup_commit_charge(memcg
, page
, nr_pages
, ctype
, false);
3846 int mem_cgroup_newpage_charge(struct page
*page
,
3847 struct mm_struct
*mm
, gfp_t gfp_mask
)
3849 if (mem_cgroup_disabled())
3851 VM_BUG_ON(page_mapped(page
));
3852 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
3854 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
3855 MEM_CGROUP_CHARGE_TYPE_ANON
);
3859 * While swap-in, try_charge -> commit or cancel, the page is locked.
3860 * And when try_charge() successfully returns, one refcnt to memcg without
3861 * struct page_cgroup is acquired. This refcnt will be consumed by
3862 * "commit()" or removed by "cancel()"
3864 static int __mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
3867 struct mem_cgroup
**memcgp
)
3869 struct mem_cgroup
*memcg
;
3870 struct page_cgroup
*pc
;
3873 pc
= lookup_page_cgroup(page
);
3875 * Every swap fault against a single page tries to charge the
3876 * page, bail as early as possible. shmem_unuse() encounters
3877 * already charged pages, too. The USED bit is protected by
3878 * the page lock, which serializes swap cache removal, which
3879 * in turn serializes uncharging.
3881 if (PageCgroupUsed(pc
))
3883 if (!do_swap_account
)
3885 memcg
= try_get_mem_cgroup_from_page(page
);
3889 ret
= __mem_cgroup_try_charge(NULL
, mask
, 1, memcgp
, true);
3890 css_put(&memcg
->css
);
3895 ret
= __mem_cgroup_try_charge(mm
, mask
, 1, memcgp
, true);
3901 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
, struct page
*page
,
3902 gfp_t gfp_mask
, struct mem_cgroup
**memcgp
)
3905 if (mem_cgroup_disabled())
3908 * A racing thread's fault, or swapoff, may have already
3909 * updated the pte, and even removed page from swap cache: in
3910 * those cases unuse_pte()'s pte_same() test will fail; but
3911 * there's also a KSM case which does need to charge the page.
3913 if (!PageSwapCache(page
)) {
3916 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, 1, memcgp
, true);
3921 return __mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, memcgp
);
3924 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*memcg
)
3926 if (mem_cgroup_disabled())
3930 __mem_cgroup_cancel_charge(memcg
, 1);
3934 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*memcg
,
3935 enum charge_type ctype
)
3937 if (mem_cgroup_disabled())
3942 __mem_cgroup_commit_charge(memcg
, page
, 1, ctype
, true);
3944 * Now swap is on-memory. This means this page may be
3945 * counted both as mem and swap....double count.
3946 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3947 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3948 * may call delete_from_swap_cache() before reach here.
3950 if (do_swap_account
&& PageSwapCache(page
)) {
3951 swp_entry_t ent
= {.val
= page_private(page
)};
3952 mem_cgroup_uncharge_swap(ent
);
3956 void mem_cgroup_commit_charge_swapin(struct page
*page
,
3957 struct mem_cgroup
*memcg
)
3959 __mem_cgroup_commit_charge_swapin(page
, memcg
,
3960 MEM_CGROUP_CHARGE_TYPE_ANON
);
3963 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
3966 struct mem_cgroup
*memcg
= NULL
;
3967 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
3970 if (mem_cgroup_disabled())
3972 if (PageCompound(page
))
3975 if (!PageSwapCache(page
))
3976 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
, type
);
3977 else { /* page is swapcache/shmem */
3978 ret
= __mem_cgroup_try_charge_swapin(mm
, page
,
3981 __mem_cgroup_commit_charge_swapin(page
, memcg
, type
);
3986 static void mem_cgroup_do_uncharge(struct mem_cgroup
*memcg
,
3987 unsigned int nr_pages
,
3988 const enum charge_type ctype
)
3990 struct memcg_batch_info
*batch
= NULL
;
3991 bool uncharge_memsw
= true;
3993 /* If swapout, usage of swap doesn't decrease */
3994 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
3995 uncharge_memsw
= false;
3997 batch
= ¤t
->memcg_batch
;
3999 * In usual, we do css_get() when we remember memcg pointer.
4000 * But in this case, we keep res->usage until end of a series of
4001 * uncharges. Then, it's ok to ignore memcg's refcnt.
4004 batch
->memcg
= memcg
;
4006 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4007 * In those cases, all pages freed continuously can be expected to be in
4008 * the same cgroup and we have chance to coalesce uncharges.
4009 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4010 * because we want to do uncharge as soon as possible.
4013 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
4014 goto direct_uncharge
;
4017 goto direct_uncharge
;
4020 * In typical case, batch->memcg == mem. This means we can
4021 * merge a series of uncharges to an uncharge of res_counter.
4022 * If not, we uncharge res_counter ony by one.
4024 if (batch
->memcg
!= memcg
)
4025 goto direct_uncharge
;
4026 /* remember freed charge and uncharge it later */
4029 batch
->memsw_nr_pages
++;
4032 res_counter_uncharge(&memcg
->res
, nr_pages
* PAGE_SIZE
);
4034 res_counter_uncharge(&memcg
->memsw
, nr_pages
* PAGE_SIZE
);
4035 if (unlikely(batch
->memcg
!= memcg
))
4036 memcg_oom_recover(memcg
);
4040 * uncharge if !page_mapped(page)
4042 static struct mem_cgroup
*
4043 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
,
4046 struct mem_cgroup
*memcg
= NULL
;
4047 unsigned int nr_pages
= 1;
4048 struct page_cgroup
*pc
;
4051 if (mem_cgroup_disabled())
4054 if (PageTransHuge(page
)) {
4055 nr_pages
<<= compound_order(page
);
4056 VM_BUG_ON(!PageTransHuge(page
));
4059 * Check if our page_cgroup is valid
4061 pc
= lookup_page_cgroup(page
);
4062 if (unlikely(!PageCgroupUsed(pc
)))
4065 lock_page_cgroup(pc
);
4067 memcg
= pc
->mem_cgroup
;
4069 if (!PageCgroupUsed(pc
))
4072 anon
= PageAnon(page
);
4075 case MEM_CGROUP_CHARGE_TYPE_ANON
:
4077 * Generally PageAnon tells if it's the anon statistics to be
4078 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4079 * used before page reached the stage of being marked PageAnon.
4083 case MEM_CGROUP_CHARGE_TYPE_DROP
:
4084 /* See mem_cgroup_prepare_migration() */
4085 if (page_mapped(page
))
4088 * Pages under migration may not be uncharged. But
4089 * end_migration() /must/ be the one uncharging the
4090 * unused post-migration page and so it has to call
4091 * here with the migration bit still set. See the
4092 * res_counter handling below.
4094 if (!end_migration
&& PageCgroupMigration(pc
))
4097 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
4098 if (!PageAnon(page
)) { /* Shared memory */
4099 if (page
->mapping
&& !page_is_file_cache(page
))
4101 } else if (page_mapped(page
)) /* Anon */
4108 mem_cgroup_charge_statistics(memcg
, page
, anon
, -nr_pages
);
4110 ClearPageCgroupUsed(pc
);
4112 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4113 * freed from LRU. This is safe because uncharged page is expected not
4114 * to be reused (freed soon). Exception is SwapCache, it's handled by
4115 * special functions.
4118 unlock_page_cgroup(pc
);
4120 * even after unlock, we have memcg->res.usage here and this memcg
4121 * will never be freed, so it's safe to call css_get().
4123 memcg_check_events(memcg
, page
);
4124 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
4125 mem_cgroup_swap_statistics(memcg
, true);
4126 css_get(&memcg
->css
);
4129 * Migration does not charge the res_counter for the
4130 * replacement page, so leave it alone when phasing out the
4131 * page that is unused after the migration.
4133 if (!end_migration
&& !mem_cgroup_is_root(memcg
))
4134 mem_cgroup_do_uncharge(memcg
, nr_pages
, ctype
);
4139 unlock_page_cgroup(pc
);
4143 void mem_cgroup_uncharge_page(struct page
*page
)
4146 if (page_mapped(page
))
4148 VM_BUG_ON(page
->mapping
&& !PageAnon(page
));
4150 * If the page is in swap cache, uncharge should be deferred
4151 * to the swap path, which also properly accounts swap usage
4152 * and handles memcg lifetime.
4154 * Note that this check is not stable and reclaim may add the
4155 * page to swap cache at any time after this. However, if the
4156 * page is not in swap cache by the time page->mapcount hits
4157 * 0, there won't be any page table references to the swap
4158 * slot, and reclaim will free it and not actually write the
4161 if (PageSwapCache(page
))
4163 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_ANON
, false);
4166 void mem_cgroup_uncharge_cache_page(struct page
*page
)
4168 VM_BUG_ON(page_mapped(page
));
4169 VM_BUG_ON(page
->mapping
);
4170 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
, false);
4174 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4175 * In that cases, pages are freed continuously and we can expect pages
4176 * are in the same memcg. All these calls itself limits the number of
4177 * pages freed at once, then uncharge_start/end() is called properly.
4178 * This may be called prural(2) times in a context,
4181 void mem_cgroup_uncharge_start(void)
4183 current
->memcg_batch
.do_batch
++;
4184 /* We can do nest. */
4185 if (current
->memcg_batch
.do_batch
== 1) {
4186 current
->memcg_batch
.memcg
= NULL
;
4187 current
->memcg_batch
.nr_pages
= 0;
4188 current
->memcg_batch
.memsw_nr_pages
= 0;
4192 void mem_cgroup_uncharge_end(void)
4194 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
4196 if (!batch
->do_batch
)
4200 if (batch
->do_batch
) /* If stacked, do nothing. */
4206 * This "batch->memcg" is valid without any css_get/put etc...
4207 * bacause we hide charges behind us.
4209 if (batch
->nr_pages
)
4210 res_counter_uncharge(&batch
->memcg
->res
,
4211 batch
->nr_pages
* PAGE_SIZE
);
4212 if (batch
->memsw_nr_pages
)
4213 res_counter_uncharge(&batch
->memcg
->memsw
,
4214 batch
->memsw_nr_pages
* PAGE_SIZE
);
4215 memcg_oom_recover(batch
->memcg
);
4216 /* forget this pointer (for sanity check) */
4217 batch
->memcg
= NULL
;
4222 * called after __delete_from_swap_cache() and drop "page" account.
4223 * memcg information is recorded to swap_cgroup of "ent"
4226 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
4228 struct mem_cgroup
*memcg
;
4229 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
4231 if (!swapout
) /* this was a swap cache but the swap is unused ! */
4232 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
4234 memcg
= __mem_cgroup_uncharge_common(page
, ctype
, false);
4237 * record memcg information, if swapout && memcg != NULL,
4238 * css_get() was called in uncharge().
4240 if (do_swap_account
&& swapout
&& memcg
)
4241 swap_cgroup_record(ent
, css_id(&memcg
->css
));
4245 #ifdef CONFIG_MEMCG_SWAP
4247 * called from swap_entry_free(). remove record in swap_cgroup and
4248 * uncharge "memsw" account.
4250 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
4252 struct mem_cgroup
*memcg
;
4255 if (!do_swap_account
)
4258 id
= swap_cgroup_record(ent
, 0);
4260 memcg
= mem_cgroup_lookup(id
);
4263 * We uncharge this because swap is freed.
4264 * This memcg can be obsolete one. We avoid calling css_tryget
4266 if (!mem_cgroup_is_root(memcg
))
4267 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
4268 mem_cgroup_swap_statistics(memcg
, false);
4269 css_put(&memcg
->css
);
4275 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4276 * @entry: swap entry to be moved
4277 * @from: mem_cgroup which the entry is moved from
4278 * @to: mem_cgroup which the entry is moved to
4280 * It succeeds only when the swap_cgroup's record for this entry is the same
4281 * as the mem_cgroup's id of @from.
4283 * Returns 0 on success, -EINVAL on failure.
4285 * The caller must have charged to @to, IOW, called res_counter_charge() about
4286 * both res and memsw, and called css_get().
4288 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
4289 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4291 unsigned short old_id
, new_id
;
4293 old_id
= css_id(&from
->css
);
4294 new_id
= css_id(&to
->css
);
4296 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
4297 mem_cgroup_swap_statistics(from
, false);
4298 mem_cgroup_swap_statistics(to
, true);
4300 * This function is only called from task migration context now.
4301 * It postpones res_counter and refcount handling till the end
4302 * of task migration(mem_cgroup_clear_mc()) for performance
4303 * improvement. But we cannot postpone css_get(to) because if
4304 * the process that has been moved to @to does swap-in, the
4305 * refcount of @to might be decreased to 0.
4307 * We are in attach() phase, so the cgroup is guaranteed to be
4308 * alive, so we can just call css_get().
4316 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
4317 struct mem_cgroup
*from
, struct mem_cgroup
*to
)
4324 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4327 void mem_cgroup_prepare_migration(struct page
*page
, struct page
*newpage
,
4328 struct mem_cgroup
**memcgp
)
4330 struct mem_cgroup
*memcg
= NULL
;
4331 unsigned int nr_pages
= 1;
4332 struct page_cgroup
*pc
;
4333 enum charge_type ctype
;
4337 if (mem_cgroup_disabled())
4340 if (PageTransHuge(page
))
4341 nr_pages
<<= compound_order(page
);
4343 pc
= lookup_page_cgroup(page
);
4344 lock_page_cgroup(pc
);
4345 if (PageCgroupUsed(pc
)) {
4346 memcg
= pc
->mem_cgroup
;
4347 css_get(&memcg
->css
);
4349 * At migrating an anonymous page, its mapcount goes down
4350 * to 0 and uncharge() will be called. But, even if it's fully
4351 * unmapped, migration may fail and this page has to be
4352 * charged again. We set MIGRATION flag here and delay uncharge
4353 * until end_migration() is called
4355 * Corner Case Thinking
4357 * When the old page was mapped as Anon and it's unmap-and-freed
4358 * while migration was ongoing.
4359 * If unmap finds the old page, uncharge() of it will be delayed
4360 * until end_migration(). If unmap finds a new page, it's
4361 * uncharged when it make mapcount to be 1->0. If unmap code
4362 * finds swap_migration_entry, the new page will not be mapped
4363 * and end_migration() will find it(mapcount==0).
4366 * When the old page was mapped but migraion fails, the kernel
4367 * remaps it. A charge for it is kept by MIGRATION flag even
4368 * if mapcount goes down to 0. We can do remap successfully
4369 * without charging it again.
4372 * The "old" page is under lock_page() until the end of
4373 * migration, so, the old page itself will not be swapped-out.
4374 * If the new page is swapped out before end_migraton, our
4375 * hook to usual swap-out path will catch the event.
4378 SetPageCgroupMigration(pc
);
4380 unlock_page_cgroup(pc
);
4382 * If the page is not charged at this point,
4390 * We charge new page before it's used/mapped. So, even if unlock_page()
4391 * is called before end_migration, we can catch all events on this new
4392 * page. In the case new page is migrated but not remapped, new page's
4393 * mapcount will be finally 0 and we call uncharge in end_migration().
4396 ctype
= MEM_CGROUP_CHARGE_TYPE_ANON
;
4398 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4400 * The page is committed to the memcg, but it's not actually
4401 * charged to the res_counter since we plan on replacing the
4402 * old one and only one page is going to be left afterwards.
4404 __mem_cgroup_commit_charge(memcg
, newpage
, nr_pages
, ctype
, false);
4407 /* remove redundant charge if migration failed*/
4408 void mem_cgroup_end_migration(struct mem_cgroup
*memcg
,
4409 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
4411 struct page
*used
, *unused
;
4412 struct page_cgroup
*pc
;
4418 if (!migration_ok
) {
4425 anon
= PageAnon(used
);
4426 __mem_cgroup_uncharge_common(unused
,
4427 anon
? MEM_CGROUP_CHARGE_TYPE_ANON
4428 : MEM_CGROUP_CHARGE_TYPE_CACHE
,
4430 css_put(&memcg
->css
);
4432 * We disallowed uncharge of pages under migration because mapcount
4433 * of the page goes down to zero, temporarly.
4434 * Clear the flag and check the page should be charged.
4436 pc
= lookup_page_cgroup(oldpage
);
4437 lock_page_cgroup(pc
);
4438 ClearPageCgroupMigration(pc
);
4439 unlock_page_cgroup(pc
);
4442 * If a page is a file cache, radix-tree replacement is very atomic
4443 * and we can skip this check. When it was an Anon page, its mapcount
4444 * goes down to 0. But because we added MIGRATION flage, it's not
4445 * uncharged yet. There are several case but page->mapcount check
4446 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4447 * check. (see prepare_charge() also)
4450 mem_cgroup_uncharge_page(used
);
4454 * At replace page cache, newpage is not under any memcg but it's on
4455 * LRU. So, this function doesn't touch res_counter but handles LRU
4456 * in correct way. Both pages are locked so we cannot race with uncharge.
4458 void mem_cgroup_replace_page_cache(struct page
*oldpage
,
4459 struct page
*newpage
)
4461 struct mem_cgroup
*memcg
= NULL
;
4462 struct page_cgroup
*pc
;
4463 enum charge_type type
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
4465 if (mem_cgroup_disabled())
4468 pc
= lookup_page_cgroup(oldpage
);
4469 /* fix accounting on old pages */
4470 lock_page_cgroup(pc
);
4471 if (PageCgroupUsed(pc
)) {
4472 memcg
= pc
->mem_cgroup
;
4473 mem_cgroup_charge_statistics(memcg
, oldpage
, false, -1);
4474 ClearPageCgroupUsed(pc
);
4476 unlock_page_cgroup(pc
);
4479 * When called from shmem_replace_page(), in some cases the
4480 * oldpage has already been charged, and in some cases not.
4485 * Even if newpage->mapping was NULL before starting replacement,
4486 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4487 * LRU while we overwrite pc->mem_cgroup.
4489 __mem_cgroup_commit_charge(memcg
, newpage
, 1, type
, true);
4492 #ifdef CONFIG_DEBUG_VM
4493 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
4495 struct page_cgroup
*pc
;
4497 pc
= lookup_page_cgroup(page
);
4499 * Can be NULL while feeding pages into the page allocator for
4500 * the first time, i.e. during boot or memory hotplug;
4501 * or when mem_cgroup_disabled().
4503 if (likely(pc
) && PageCgroupUsed(pc
))
4508 bool mem_cgroup_bad_page_check(struct page
*page
)
4510 if (mem_cgroup_disabled())
4513 return lookup_page_cgroup_used(page
) != NULL
;
4516 void mem_cgroup_print_bad_page(struct page
*page
)
4518 struct page_cgroup
*pc
;
4520 pc
= lookup_page_cgroup_used(page
);
4522 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4523 pc
, pc
->flags
, pc
->mem_cgroup
);
4528 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
4529 unsigned long long val
)
4532 u64 memswlimit
, memlimit
;
4534 int children
= mem_cgroup_count_children(memcg
);
4535 u64 curusage
, oldusage
;
4539 * For keeping hierarchical_reclaim simple, how long we should retry
4540 * is depends on callers. We set our retry-count to be function
4541 * of # of children which we should visit in this loop.
4543 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
4545 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4548 while (retry_count
) {
4549 if (signal_pending(current
)) {
4554 * Rather than hide all in some function, I do this in
4555 * open coded manner. You see what this really does.
4556 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4558 mutex_lock(&set_limit_mutex
);
4559 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4560 if (memswlimit
< val
) {
4562 mutex_unlock(&set_limit_mutex
);
4566 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4570 ret
= res_counter_set_limit(&memcg
->res
, val
);
4572 if (memswlimit
== val
)
4573 memcg
->memsw_is_minimum
= true;
4575 memcg
->memsw_is_minimum
= false;
4577 mutex_unlock(&set_limit_mutex
);
4582 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4583 MEM_CGROUP_RECLAIM_SHRINK
);
4584 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4585 /* Usage is reduced ? */
4586 if (curusage
>= oldusage
)
4589 oldusage
= curusage
;
4591 if (!ret
&& enlarge
)
4592 memcg_oom_recover(memcg
);
4597 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
4598 unsigned long long val
)
4601 u64 memlimit
, memswlimit
, oldusage
, curusage
;
4602 int children
= mem_cgroup_count_children(memcg
);
4606 /* see mem_cgroup_resize_res_limit */
4607 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
4608 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4609 while (retry_count
) {
4610 if (signal_pending(current
)) {
4615 * Rather than hide all in some function, I do this in
4616 * open coded manner. You see what this really does.
4617 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4619 mutex_lock(&set_limit_mutex
);
4620 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
4621 if (memlimit
> val
) {
4623 mutex_unlock(&set_limit_mutex
);
4626 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
4627 if (memswlimit
< val
)
4629 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
4631 if (memlimit
== val
)
4632 memcg
->memsw_is_minimum
= true;
4634 memcg
->memsw_is_minimum
= false;
4636 mutex_unlock(&set_limit_mutex
);
4641 mem_cgroup_reclaim(memcg
, GFP_KERNEL
,
4642 MEM_CGROUP_RECLAIM_NOSWAP
|
4643 MEM_CGROUP_RECLAIM_SHRINK
);
4644 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4645 /* Usage is reduced ? */
4646 if (curusage
>= oldusage
)
4649 oldusage
= curusage
;
4651 if (!ret
&& enlarge
)
4652 memcg_oom_recover(memcg
);
4657 * mem_cgroup_force_empty_list - clears LRU of a group
4658 * @memcg: group to clear
4661 * @lru: lru to to clear
4663 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4664 * reclaim the pages page themselves - pages are moved to the parent (or root)
4667 static void mem_cgroup_force_empty_list(struct mem_cgroup
*memcg
,
4668 int node
, int zid
, enum lru_list lru
)
4670 struct lruvec
*lruvec
;
4671 unsigned long flags
;
4672 struct list_head
*list
;
4676 zone
= &NODE_DATA(node
)->node_zones
[zid
];
4677 lruvec
= mem_cgroup_zone_lruvec(zone
, memcg
);
4678 list
= &lruvec
->lists
[lru
];
4682 struct page_cgroup
*pc
;
4685 spin_lock_irqsave(&zone
->lru_lock
, flags
);
4686 if (list_empty(list
)) {
4687 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4690 page
= list_entry(list
->prev
, struct page
, lru
);
4692 list_move(&page
->lru
, list
);
4694 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4697 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
4699 pc
= lookup_page_cgroup(page
);
4701 if (mem_cgroup_move_parent(page
, pc
, memcg
)) {
4702 /* found lock contention or "pc" is obsolete. */
4707 } while (!list_empty(list
));
4711 * make mem_cgroup's charge to be 0 if there is no task by moving
4712 * all the charges and pages to the parent.
4713 * This enables deleting this mem_cgroup.
4715 * Caller is responsible for holding css reference on the memcg.
4717 static void mem_cgroup_reparent_charges(struct mem_cgroup
*memcg
)
4723 /* This is for making all *used* pages to be on LRU. */
4724 lru_add_drain_all();
4725 drain_all_stock_sync(memcg
);
4726 mem_cgroup_start_move(memcg
);
4727 for_each_node_state(node
, N_MEMORY
) {
4728 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
4731 mem_cgroup_force_empty_list(memcg
,
4736 mem_cgroup_end_move(memcg
);
4737 memcg_oom_recover(memcg
);
4741 * Kernel memory may not necessarily be trackable to a specific
4742 * process. So they are not migrated, and therefore we can't
4743 * expect their value to drop to 0 here.
4744 * Having res filled up with kmem only is enough.
4746 * This is a safety check because mem_cgroup_force_empty_list
4747 * could have raced with mem_cgroup_replace_page_cache callers
4748 * so the lru seemed empty but the page could have been added
4749 * right after the check. RES_USAGE should be safe as we always
4750 * charge before adding to the LRU.
4752 usage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
) -
4753 res_counter_read_u64(&memcg
->kmem
, RES_USAGE
);
4754 } while (usage
> 0);
4758 * This mainly exists for tests during the setting of set of use_hierarchy.
4759 * Since this is the very setting we are changing, the current hierarchy value
4762 static inline bool __memcg_has_children(struct mem_cgroup
*memcg
)
4764 struct cgroup_subsys_state
*pos
;
4766 /* bounce at first found */
4767 css_for_each_child(pos
, &memcg
->css
)
4773 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4774 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4775 * from mem_cgroup_count_children(), in the sense that we don't really care how
4776 * many children we have; we only need to know if we have any. It also counts
4777 * any memcg without hierarchy as infertile.
4779 static inline bool memcg_has_children(struct mem_cgroup
*memcg
)
4781 return memcg
->use_hierarchy
&& __memcg_has_children(memcg
);
4785 * Reclaims as many pages from the given memcg as possible and moves
4786 * the rest to the parent.
4788 * Caller is responsible for holding css reference for memcg.
4790 static int mem_cgroup_force_empty(struct mem_cgroup
*memcg
)
4792 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
4793 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
4795 /* returns EBUSY if there is a task or if we come here twice. */
4796 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
4799 /* we call try-to-free pages for make this cgroup empty */
4800 lru_add_drain_all();
4801 /* try to free all pages in this cgroup */
4802 while (nr_retries
&& res_counter_read_u64(&memcg
->res
, RES_USAGE
) > 0) {
4805 if (signal_pending(current
))
4808 progress
= try_to_free_mem_cgroup_pages(memcg
, GFP_KERNEL
,
4812 /* maybe some writeback is necessary */
4813 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
4818 mem_cgroup_reparent_charges(memcg
);
4823 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state
*css
,
4826 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4828 if (mem_cgroup_is_root(memcg
))
4830 return mem_cgroup_force_empty(memcg
);
4833 static u64
mem_cgroup_hierarchy_read(struct cgroup_subsys_state
*css
,
4836 return mem_cgroup_from_css(css
)->use_hierarchy
;
4839 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state
*css
,
4840 struct cftype
*cft
, u64 val
)
4843 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4844 struct mem_cgroup
*parent_memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
4846 mutex_lock(&memcg_create_mutex
);
4848 if (memcg
->use_hierarchy
== val
)
4852 * If parent's use_hierarchy is set, we can't make any modifications
4853 * in the child subtrees. If it is unset, then the change can
4854 * occur, provided the current cgroup has no children.
4856 * For the root cgroup, parent_mem is NULL, we allow value to be
4857 * set if there are no children.
4859 if ((!parent_memcg
|| !parent_memcg
->use_hierarchy
) &&
4860 (val
== 1 || val
== 0)) {
4861 if (!__memcg_has_children(memcg
))
4862 memcg
->use_hierarchy
= val
;
4869 mutex_unlock(&memcg_create_mutex
);
4875 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup
*memcg
,
4876 enum mem_cgroup_stat_index idx
)
4878 struct mem_cgroup
*iter
;
4881 /* Per-cpu values can be negative, use a signed accumulator */
4882 for_each_mem_cgroup_tree(iter
, memcg
)
4883 val
+= mem_cgroup_read_stat(iter
, idx
);
4885 if (val
< 0) /* race ? */
4890 static inline u64
mem_cgroup_usage(struct mem_cgroup
*memcg
, bool swap
)
4894 if (!mem_cgroup_is_root(memcg
)) {
4896 return res_counter_read_u64(&memcg
->res
, RES_USAGE
);
4898 return res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
4902 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4903 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4905 val
= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_CACHE
);
4906 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_RSS
);
4909 val
+= mem_cgroup_recursive_stat(memcg
, MEM_CGROUP_STAT_SWAP
);
4911 return val
<< PAGE_SHIFT
;
4914 static ssize_t
mem_cgroup_read(struct cgroup_subsys_state
*css
,
4915 struct cftype
*cft
, struct file
*file
,
4916 char __user
*buf
, size_t nbytes
, loff_t
*ppos
)
4918 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4924 type
= MEMFILE_TYPE(cft
->private);
4925 name
= MEMFILE_ATTR(cft
->private);
4929 if (name
== RES_USAGE
)
4930 val
= mem_cgroup_usage(memcg
, false);
4932 val
= res_counter_read_u64(&memcg
->res
, name
);
4935 if (name
== RES_USAGE
)
4936 val
= mem_cgroup_usage(memcg
, true);
4938 val
= res_counter_read_u64(&memcg
->memsw
, name
);
4941 val
= res_counter_read_u64(&memcg
->kmem
, name
);
4947 len
= scnprintf(str
, sizeof(str
), "%llu\n", (unsigned long long)val
);
4948 return simple_read_from_buffer(buf
, nbytes
, ppos
, str
, len
);
4951 static int memcg_update_kmem_limit(struct cgroup_subsys_state
*css
, u64 val
)
4954 #ifdef CONFIG_MEMCG_KMEM
4955 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
4957 * For simplicity, we won't allow this to be disabled. It also can't
4958 * be changed if the cgroup has children already, or if tasks had
4961 * If tasks join before we set the limit, a person looking at
4962 * kmem.usage_in_bytes will have no way to determine when it took
4963 * place, which makes the value quite meaningless.
4965 * After it first became limited, changes in the value of the limit are
4966 * of course permitted.
4968 mutex_lock(&memcg_create_mutex
);
4969 mutex_lock(&set_limit_mutex
);
4970 if (!memcg
->kmem_account_flags
&& val
!= RESOURCE_MAX
) {
4971 if (cgroup_task_count(css
->cgroup
) || memcg_has_children(memcg
)) {
4975 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4978 ret
= memcg_update_cache_sizes(memcg
);
4980 res_counter_set_limit(&memcg
->kmem
, RESOURCE_MAX
);
4983 static_key_slow_inc(&memcg_kmem_enabled_key
);
4985 * setting the active bit after the inc will guarantee no one
4986 * starts accounting before all call sites are patched
4988 memcg_kmem_set_active(memcg
);
4990 ret
= res_counter_set_limit(&memcg
->kmem
, val
);
4992 mutex_unlock(&set_limit_mutex
);
4993 mutex_unlock(&memcg_create_mutex
);
4998 #ifdef CONFIG_MEMCG_KMEM
4999 static int memcg_propagate_kmem(struct mem_cgroup
*memcg
)
5002 struct mem_cgroup
*parent
= parent_mem_cgroup(memcg
);
5006 memcg
->kmem_account_flags
= parent
->kmem_account_flags
;
5008 * When that happen, we need to disable the static branch only on those
5009 * memcgs that enabled it. To achieve this, we would be forced to
5010 * complicate the code by keeping track of which memcgs were the ones
5011 * that actually enabled limits, and which ones got it from its
5014 * It is a lot simpler just to do static_key_slow_inc() on every child
5015 * that is accounted.
5017 if (!memcg_kmem_is_active(memcg
))
5021 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5022 * memcg is active already. If the later initialization fails then the
5023 * cgroup core triggers the cleanup so we do not have to do it here.
5025 static_key_slow_inc(&memcg_kmem_enabled_key
);
5027 mutex_lock(&set_limit_mutex
);
5028 memcg_stop_kmem_account();
5029 ret
= memcg_update_cache_sizes(memcg
);
5030 memcg_resume_kmem_account();
5031 mutex_unlock(&set_limit_mutex
);
5035 #endif /* CONFIG_MEMCG_KMEM */
5038 * The user of this function is...
5041 static int mem_cgroup_write(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5044 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5047 unsigned long long val
;
5050 type
= MEMFILE_TYPE(cft
->private);
5051 name
= MEMFILE_ATTR(cft
->private);
5055 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
5059 /* This function does all necessary parse...reuse it */
5060 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5064 ret
= mem_cgroup_resize_limit(memcg
, val
);
5065 else if (type
== _MEMSWAP
)
5066 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
5067 else if (type
== _KMEM
)
5068 ret
= memcg_update_kmem_limit(css
, val
);
5072 case RES_SOFT_LIMIT
:
5073 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
5077 * For memsw, soft limits are hard to implement in terms
5078 * of semantics, for now, we support soft limits for
5079 * control without swap
5082 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
5087 ret
= -EINVAL
; /* should be BUG() ? */
5093 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
5094 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
5096 unsigned long long min_limit
, min_memsw_limit
, tmp
;
5098 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5099 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5100 if (!memcg
->use_hierarchy
)
5103 while (css_parent(&memcg
->css
)) {
5104 memcg
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5105 if (!memcg
->use_hierarchy
)
5107 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
5108 min_limit
= min(min_limit
, tmp
);
5109 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
5110 min_memsw_limit
= min(min_memsw_limit
, tmp
);
5113 *mem_limit
= min_limit
;
5114 *memsw_limit
= min_memsw_limit
;
5117 static int mem_cgroup_reset(struct cgroup_subsys_state
*css
, unsigned int event
)
5119 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5123 type
= MEMFILE_TYPE(event
);
5124 name
= MEMFILE_ATTR(event
);
5129 res_counter_reset_max(&memcg
->res
);
5130 else if (type
== _MEMSWAP
)
5131 res_counter_reset_max(&memcg
->memsw
);
5132 else if (type
== _KMEM
)
5133 res_counter_reset_max(&memcg
->kmem
);
5139 res_counter_reset_failcnt(&memcg
->res
);
5140 else if (type
== _MEMSWAP
)
5141 res_counter_reset_failcnt(&memcg
->memsw
);
5142 else if (type
== _KMEM
)
5143 res_counter_reset_failcnt(&memcg
->kmem
);
5152 static u64
mem_cgroup_move_charge_read(struct cgroup_subsys_state
*css
,
5155 return mem_cgroup_from_css(css
)->move_charge_at_immigrate
;
5159 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5160 struct cftype
*cft
, u64 val
)
5162 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5164 if (val
>= (1 << NR_MOVE_TYPE
))
5168 * No kind of locking is needed in here, because ->can_attach() will
5169 * check this value once in the beginning of the process, and then carry
5170 * on with stale data. This means that changes to this value will only
5171 * affect task migrations starting after the change.
5173 memcg
->move_charge_at_immigrate
= val
;
5177 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state
*css
,
5178 struct cftype
*cft
, u64 val
)
5185 static int memcg_numa_stat_show(struct cgroup_subsys_state
*css
,
5186 struct cftype
*cft
, struct seq_file
*m
)
5189 unsigned long total_nr
, file_nr
, anon_nr
, unevictable_nr
;
5190 unsigned long node_nr
;
5191 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5193 total_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL
);
5194 seq_printf(m
, "total=%lu", total_nr
);
5195 for_each_node_state(nid
, N_MEMORY
) {
5196 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
, LRU_ALL
);
5197 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5201 file_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_FILE
);
5202 seq_printf(m
, "file=%lu", file_nr
);
5203 for_each_node_state(nid
, N_MEMORY
) {
5204 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5206 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5210 anon_nr
= mem_cgroup_nr_lru_pages(memcg
, LRU_ALL_ANON
);
5211 seq_printf(m
, "anon=%lu", anon_nr
);
5212 for_each_node_state(nid
, N_MEMORY
) {
5213 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5215 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5219 unevictable_nr
= mem_cgroup_nr_lru_pages(memcg
, BIT(LRU_UNEVICTABLE
));
5220 seq_printf(m
, "unevictable=%lu", unevictable_nr
);
5221 for_each_node_state(nid
, N_MEMORY
) {
5222 node_nr
= mem_cgroup_node_nr_lru_pages(memcg
, nid
,
5223 BIT(LRU_UNEVICTABLE
));
5224 seq_printf(m
, " N%d=%lu", nid
, node_nr
);
5229 #endif /* CONFIG_NUMA */
5231 static inline void mem_cgroup_lru_names_not_uptodate(void)
5233 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names
) != NR_LRU_LISTS
);
5236 static int memcg_stat_show(struct cgroup_subsys_state
*css
, struct cftype
*cft
,
5239 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5240 struct mem_cgroup
*mi
;
5243 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5244 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5246 seq_printf(m
, "%s %ld\n", mem_cgroup_stat_names
[i
],
5247 mem_cgroup_read_stat(memcg
, i
) * PAGE_SIZE
);
5250 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++)
5251 seq_printf(m
, "%s %lu\n", mem_cgroup_events_names
[i
],
5252 mem_cgroup_read_events(memcg
, i
));
5254 for (i
= 0; i
< NR_LRU_LISTS
; i
++)
5255 seq_printf(m
, "%s %lu\n", mem_cgroup_lru_names
[i
],
5256 mem_cgroup_nr_lru_pages(memcg
, BIT(i
)) * PAGE_SIZE
);
5258 /* Hierarchical information */
5260 unsigned long long limit
, memsw_limit
;
5261 memcg_get_hierarchical_limit(memcg
, &limit
, &memsw_limit
);
5262 seq_printf(m
, "hierarchical_memory_limit %llu\n", limit
);
5263 if (do_swap_account
)
5264 seq_printf(m
, "hierarchical_memsw_limit %llu\n",
5268 for (i
= 0; i
< MEM_CGROUP_STAT_NSTATS
; i
++) {
5271 if (i
== MEM_CGROUP_STAT_SWAP
&& !do_swap_account
)
5273 for_each_mem_cgroup_tree(mi
, memcg
)
5274 val
+= mem_cgroup_read_stat(mi
, i
) * PAGE_SIZE
;
5275 seq_printf(m
, "total_%s %lld\n", mem_cgroup_stat_names
[i
], val
);
5278 for (i
= 0; i
< MEM_CGROUP_EVENTS_NSTATS
; i
++) {
5279 unsigned long long val
= 0;
5281 for_each_mem_cgroup_tree(mi
, memcg
)
5282 val
+= mem_cgroup_read_events(mi
, i
);
5283 seq_printf(m
, "total_%s %llu\n",
5284 mem_cgroup_events_names
[i
], val
);
5287 for (i
= 0; i
< NR_LRU_LISTS
; i
++) {
5288 unsigned long long val
= 0;
5290 for_each_mem_cgroup_tree(mi
, memcg
)
5291 val
+= mem_cgroup_nr_lru_pages(mi
, BIT(i
)) * PAGE_SIZE
;
5292 seq_printf(m
, "total_%s %llu\n", mem_cgroup_lru_names
[i
], val
);
5295 #ifdef CONFIG_DEBUG_VM
5298 struct mem_cgroup_per_zone
*mz
;
5299 struct zone_reclaim_stat
*rstat
;
5300 unsigned long recent_rotated
[2] = {0, 0};
5301 unsigned long recent_scanned
[2] = {0, 0};
5303 for_each_online_node(nid
)
5304 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
5305 mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
5306 rstat
= &mz
->lruvec
.reclaim_stat
;
5308 recent_rotated
[0] += rstat
->recent_rotated
[0];
5309 recent_rotated
[1] += rstat
->recent_rotated
[1];
5310 recent_scanned
[0] += rstat
->recent_scanned
[0];
5311 recent_scanned
[1] += rstat
->recent_scanned
[1];
5313 seq_printf(m
, "recent_rotated_anon %lu\n", recent_rotated
[0]);
5314 seq_printf(m
, "recent_rotated_file %lu\n", recent_rotated
[1]);
5315 seq_printf(m
, "recent_scanned_anon %lu\n", recent_scanned
[0]);
5316 seq_printf(m
, "recent_scanned_file %lu\n", recent_scanned
[1]);
5323 static u64
mem_cgroup_swappiness_read(struct cgroup_subsys_state
*css
,
5326 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5328 return mem_cgroup_swappiness(memcg
);
5331 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state
*css
,
5332 struct cftype
*cft
, u64 val
)
5334 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5335 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5337 if (val
> 100 || !parent
)
5340 mutex_lock(&memcg_create_mutex
);
5342 /* If under hierarchy, only empty-root can set this value */
5343 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5344 mutex_unlock(&memcg_create_mutex
);
5348 memcg
->swappiness
= val
;
5350 mutex_unlock(&memcg_create_mutex
);
5355 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
5357 struct mem_cgroup_threshold_ary
*t
;
5363 t
= rcu_dereference(memcg
->thresholds
.primary
);
5365 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
5370 usage
= mem_cgroup_usage(memcg
, swap
);
5373 * current_threshold points to threshold just below or equal to usage.
5374 * If it's not true, a threshold was crossed after last
5375 * call of __mem_cgroup_threshold().
5377 i
= t
->current_threshold
;
5380 * Iterate backward over array of thresholds starting from
5381 * current_threshold and check if a threshold is crossed.
5382 * If none of thresholds below usage is crossed, we read
5383 * only one element of the array here.
5385 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
5386 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5388 /* i = current_threshold + 1 */
5392 * Iterate forward over array of thresholds starting from
5393 * current_threshold+1 and check if a threshold is crossed.
5394 * If none of thresholds above usage is crossed, we read
5395 * only one element of the array here.
5397 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
5398 eventfd_signal(t
->entries
[i
].eventfd
, 1);
5400 /* Update current_threshold */
5401 t
->current_threshold
= i
- 1;
5406 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
5409 __mem_cgroup_threshold(memcg
, false);
5410 if (do_swap_account
)
5411 __mem_cgroup_threshold(memcg
, true);
5413 memcg
= parent_mem_cgroup(memcg
);
5417 static int compare_thresholds(const void *a
, const void *b
)
5419 const struct mem_cgroup_threshold
*_a
= a
;
5420 const struct mem_cgroup_threshold
*_b
= b
;
5422 if (_a
->threshold
> _b
->threshold
)
5425 if (_a
->threshold
< _b
->threshold
)
5431 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*memcg
)
5433 struct mem_cgroup_eventfd_list
*ev
;
5435 list_for_each_entry(ev
, &memcg
->oom_notify
, list
)
5436 eventfd_signal(ev
->eventfd
, 1);
5440 static void mem_cgroup_oom_notify(struct mem_cgroup
*memcg
)
5442 struct mem_cgroup
*iter
;
5444 for_each_mem_cgroup_tree(iter
, memcg
)
5445 mem_cgroup_oom_notify_cb(iter
);
5448 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state
*css
,
5449 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5451 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5452 struct mem_cgroup_thresholds
*thresholds
;
5453 struct mem_cgroup_threshold_ary
*new;
5454 enum res_type type
= MEMFILE_TYPE(cft
->private);
5455 u64 threshold
, usage
;
5458 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
5462 mutex_lock(&memcg
->thresholds_lock
);
5465 thresholds
= &memcg
->thresholds
;
5466 else if (type
== _MEMSWAP
)
5467 thresholds
= &memcg
->memsw_thresholds
;
5471 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5473 /* Check if a threshold crossed before adding a new one */
5474 if (thresholds
->primary
)
5475 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5477 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
5479 /* Allocate memory for new array of thresholds */
5480 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
5488 /* Copy thresholds (if any) to new array */
5489 if (thresholds
->primary
) {
5490 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
5491 sizeof(struct mem_cgroup_threshold
));
5494 /* Add new threshold */
5495 new->entries
[size
- 1].eventfd
= eventfd
;
5496 new->entries
[size
- 1].threshold
= threshold
;
5498 /* Sort thresholds. Registering of new threshold isn't time-critical */
5499 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
5500 compare_thresholds
, NULL
);
5502 /* Find current threshold */
5503 new->current_threshold
= -1;
5504 for (i
= 0; i
< size
; i
++) {
5505 if (new->entries
[i
].threshold
<= usage
) {
5507 * new->current_threshold will not be used until
5508 * rcu_assign_pointer(), so it's safe to increment
5511 ++new->current_threshold
;
5516 /* Free old spare buffer and save old primary buffer as spare */
5517 kfree(thresholds
->spare
);
5518 thresholds
->spare
= thresholds
->primary
;
5520 rcu_assign_pointer(thresholds
->primary
, new);
5522 /* To be sure that nobody uses thresholds */
5526 mutex_unlock(&memcg
->thresholds_lock
);
5531 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state
*css
,
5532 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5534 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5535 struct mem_cgroup_thresholds
*thresholds
;
5536 struct mem_cgroup_threshold_ary
*new;
5537 enum res_type type
= MEMFILE_TYPE(cft
->private);
5541 mutex_lock(&memcg
->thresholds_lock
);
5543 thresholds
= &memcg
->thresholds
;
5544 else if (type
== _MEMSWAP
)
5545 thresholds
= &memcg
->memsw_thresholds
;
5549 if (!thresholds
->primary
)
5552 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
5554 /* Check if a threshold crossed before removing */
5555 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
5557 /* Calculate new number of threshold */
5559 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
5560 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
5564 new = thresholds
->spare
;
5566 /* Set thresholds array to NULL if we don't have thresholds */
5575 /* Copy thresholds and find current threshold */
5576 new->current_threshold
= -1;
5577 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
5578 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
5581 new->entries
[j
] = thresholds
->primary
->entries
[i
];
5582 if (new->entries
[j
].threshold
<= usage
) {
5584 * new->current_threshold will not be used
5585 * until rcu_assign_pointer(), so it's safe to increment
5588 ++new->current_threshold
;
5594 /* Swap primary and spare array */
5595 thresholds
->spare
= thresholds
->primary
;
5596 /* If all events are unregistered, free the spare array */
5598 kfree(thresholds
->spare
);
5599 thresholds
->spare
= NULL
;
5602 rcu_assign_pointer(thresholds
->primary
, new);
5604 /* To be sure that nobody uses thresholds */
5607 mutex_unlock(&memcg
->thresholds_lock
);
5610 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state
*css
,
5611 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
5613 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5614 struct mem_cgroup_eventfd_list
*event
;
5615 enum res_type type
= MEMFILE_TYPE(cft
->private);
5617 BUG_ON(type
!= _OOM_TYPE
);
5618 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
5622 spin_lock(&memcg_oom_lock
);
5624 event
->eventfd
= eventfd
;
5625 list_add(&event
->list
, &memcg
->oom_notify
);
5627 /* already in OOM ? */
5628 if (atomic_read(&memcg
->under_oom
))
5629 eventfd_signal(eventfd
, 1);
5630 spin_unlock(&memcg_oom_lock
);
5635 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state
*css
,
5636 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
5638 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5639 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
5640 enum res_type type
= MEMFILE_TYPE(cft
->private);
5642 BUG_ON(type
!= _OOM_TYPE
);
5644 spin_lock(&memcg_oom_lock
);
5646 list_for_each_entry_safe(ev
, tmp
, &memcg
->oom_notify
, list
) {
5647 if (ev
->eventfd
== eventfd
) {
5648 list_del(&ev
->list
);
5653 spin_unlock(&memcg_oom_lock
);
5656 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state
*css
,
5657 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
5659 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5661 cb
->fill(cb
, "oom_kill_disable", memcg
->oom_kill_disable
);
5663 if (atomic_read(&memcg
->under_oom
))
5664 cb
->fill(cb
, "under_oom", 1);
5666 cb
->fill(cb
, "under_oom", 0);
5670 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state
*css
,
5671 struct cftype
*cft
, u64 val
)
5673 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
5674 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(&memcg
->css
));
5676 /* cannot set to root cgroup and only 0 and 1 are allowed */
5677 if (!parent
|| !((val
== 0) || (val
== 1)))
5680 mutex_lock(&memcg_create_mutex
);
5681 /* oom-kill-disable is a flag for subhierarchy. */
5682 if ((parent
->use_hierarchy
) || memcg_has_children(memcg
)) {
5683 mutex_unlock(&memcg_create_mutex
);
5686 memcg
->oom_kill_disable
= val
;
5688 memcg_oom_recover(memcg
);
5689 mutex_unlock(&memcg_create_mutex
);
5693 #ifdef CONFIG_MEMCG_KMEM
5694 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5698 memcg
->kmemcg_id
= -1;
5699 ret
= memcg_propagate_kmem(memcg
);
5703 return mem_cgroup_sockets_init(memcg
, ss
);
5706 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5708 mem_cgroup_sockets_destroy(memcg
);
5711 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5713 if (!memcg_kmem_is_active(memcg
))
5717 * kmem charges can outlive the cgroup. In the case of slab
5718 * pages, for instance, a page contain objects from various
5719 * processes. As we prevent from taking a reference for every
5720 * such allocation we have to be careful when doing uncharge
5721 * (see memcg_uncharge_kmem) and here during offlining.
5723 * The idea is that that only the _last_ uncharge which sees
5724 * the dead memcg will drop the last reference. An additional
5725 * reference is taken here before the group is marked dead
5726 * which is then paired with css_put during uncharge resp. here.
5728 * Although this might sound strange as this path is called from
5729 * css_offline() when the referencemight have dropped down to 0
5730 * and shouldn't be incremented anymore (css_tryget would fail)
5731 * we do not have other options because of the kmem allocations
5734 css_get(&memcg
->css
);
5736 memcg_kmem_mark_dead(memcg
);
5738 if (res_counter_read_u64(&memcg
->kmem
, RES_USAGE
) != 0)
5741 if (memcg_kmem_test_and_clear_dead(memcg
))
5742 css_put(&memcg
->css
);
5745 static int memcg_init_kmem(struct mem_cgroup
*memcg
, struct cgroup_subsys
*ss
)
5750 static void memcg_destroy_kmem(struct mem_cgroup
*memcg
)
5754 static void kmem_cgroup_css_offline(struct mem_cgroup
*memcg
)
5759 static struct cftype mem_cgroup_files
[] = {
5761 .name
= "usage_in_bytes",
5762 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
5763 .read
= mem_cgroup_read
,
5764 .register_event
= mem_cgroup_usage_register_event
,
5765 .unregister_event
= mem_cgroup_usage_unregister_event
,
5768 .name
= "max_usage_in_bytes",
5769 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
5770 .trigger
= mem_cgroup_reset
,
5771 .read
= mem_cgroup_read
,
5774 .name
= "limit_in_bytes",
5775 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
5776 .write_string
= mem_cgroup_write
,
5777 .read
= mem_cgroup_read
,
5780 .name
= "soft_limit_in_bytes",
5781 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
5782 .write_string
= mem_cgroup_write
,
5783 .read
= mem_cgroup_read
,
5787 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
5788 .trigger
= mem_cgroup_reset
,
5789 .read
= mem_cgroup_read
,
5793 .read_seq_string
= memcg_stat_show
,
5796 .name
= "force_empty",
5797 .trigger
= mem_cgroup_force_empty_write
,
5800 .name
= "use_hierarchy",
5801 .flags
= CFTYPE_INSANE
,
5802 .write_u64
= mem_cgroup_hierarchy_write
,
5803 .read_u64
= mem_cgroup_hierarchy_read
,
5806 .name
= "swappiness",
5807 .read_u64
= mem_cgroup_swappiness_read
,
5808 .write_u64
= mem_cgroup_swappiness_write
,
5811 .name
= "move_charge_at_immigrate",
5812 .read_u64
= mem_cgroup_move_charge_read
,
5813 .write_u64
= mem_cgroup_move_charge_write
,
5816 .name
= "oom_control",
5817 .read_map
= mem_cgroup_oom_control_read
,
5818 .write_u64
= mem_cgroup_oom_control_write
,
5819 .register_event
= mem_cgroup_oom_register_event
,
5820 .unregister_event
= mem_cgroup_oom_unregister_event
,
5821 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
5824 .name
= "pressure_level",
5825 .register_event
= vmpressure_register_event
,
5826 .unregister_event
= vmpressure_unregister_event
,
5830 .name
= "numa_stat",
5831 .read_seq_string
= memcg_numa_stat_show
,
5834 #ifdef CONFIG_MEMCG_KMEM
5836 .name
= "kmem.limit_in_bytes",
5837 .private = MEMFILE_PRIVATE(_KMEM
, RES_LIMIT
),
5838 .write_string
= mem_cgroup_write
,
5839 .read
= mem_cgroup_read
,
5842 .name
= "kmem.usage_in_bytes",
5843 .private = MEMFILE_PRIVATE(_KMEM
, RES_USAGE
),
5844 .read
= mem_cgroup_read
,
5847 .name
= "kmem.failcnt",
5848 .private = MEMFILE_PRIVATE(_KMEM
, RES_FAILCNT
),
5849 .trigger
= mem_cgroup_reset
,
5850 .read
= mem_cgroup_read
,
5853 .name
= "kmem.max_usage_in_bytes",
5854 .private = MEMFILE_PRIVATE(_KMEM
, RES_MAX_USAGE
),
5855 .trigger
= mem_cgroup_reset
,
5856 .read
= mem_cgroup_read
,
5858 #ifdef CONFIG_SLABINFO
5860 .name
= "kmem.slabinfo",
5861 .read_seq_string
= mem_cgroup_slabinfo_read
,
5865 { }, /* terminate */
5868 #ifdef CONFIG_MEMCG_SWAP
5869 static struct cftype memsw_cgroup_files
[] = {
5871 .name
= "memsw.usage_in_bytes",
5872 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
5873 .read
= mem_cgroup_read
,
5874 .register_event
= mem_cgroup_usage_register_event
,
5875 .unregister_event
= mem_cgroup_usage_unregister_event
,
5878 .name
= "memsw.max_usage_in_bytes",
5879 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
5880 .trigger
= mem_cgroup_reset
,
5881 .read
= mem_cgroup_read
,
5884 .name
= "memsw.limit_in_bytes",
5885 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
5886 .write_string
= mem_cgroup_write
,
5887 .read
= mem_cgroup_read
,
5890 .name
= "memsw.failcnt",
5891 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
5892 .trigger
= mem_cgroup_reset
,
5893 .read
= mem_cgroup_read
,
5895 { }, /* terminate */
5898 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5900 struct mem_cgroup_per_node
*pn
;
5901 struct mem_cgroup_per_zone
*mz
;
5902 int zone
, tmp
= node
;
5904 * This routine is called against possible nodes.
5905 * But it's BUG to call kmalloc() against offline node.
5907 * TODO: this routine can waste much memory for nodes which will
5908 * never be onlined. It's better to use memory hotplug callback
5911 if (!node_state(node
, N_NORMAL_MEMORY
))
5913 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
5917 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
5918 mz
= &pn
->zoneinfo
[zone
];
5919 lruvec_init(&mz
->lruvec
);
5922 memcg
->nodeinfo
[node
] = pn
;
5926 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*memcg
, int node
)
5928 kfree(memcg
->nodeinfo
[node
]);
5931 static struct mem_cgroup
*mem_cgroup_alloc(void)
5933 struct mem_cgroup
*memcg
;
5934 size_t size
= memcg_size();
5936 /* Can be very big if nr_node_ids is very big */
5937 if (size
< PAGE_SIZE
)
5938 memcg
= kzalloc(size
, GFP_KERNEL
);
5940 memcg
= vzalloc(size
);
5945 memcg
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
5948 spin_lock_init(&memcg
->pcp_counter_lock
);
5952 if (size
< PAGE_SIZE
)
5960 * At destroying mem_cgroup, references from swap_cgroup can remain.
5961 * (scanning all at force_empty is too costly...)
5963 * Instead of clearing all references at force_empty, we remember
5964 * the number of reference from swap_cgroup and free mem_cgroup when
5965 * it goes down to 0.
5967 * Removal of cgroup itself succeeds regardless of refs from swap.
5970 static void __mem_cgroup_free(struct mem_cgroup
*memcg
)
5973 size_t size
= memcg_size();
5975 free_css_id(&mem_cgroup_subsys
, &memcg
->css
);
5978 free_mem_cgroup_per_zone_info(memcg
, node
);
5980 free_percpu(memcg
->stat
);
5983 * We need to make sure that (at least for now), the jump label
5984 * destruction code runs outside of the cgroup lock. This is because
5985 * get_online_cpus(), which is called from the static_branch update,
5986 * can't be called inside the cgroup_lock. cpusets are the ones
5987 * enforcing this dependency, so if they ever change, we might as well.
5989 * schedule_work() will guarantee this happens. Be careful if you need
5990 * to move this code around, and make sure it is outside
5993 disarm_static_keys(memcg
);
5994 if (size
< PAGE_SIZE
)
6001 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6003 struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*memcg
)
6005 if (!memcg
->res
.parent
)
6007 return mem_cgroup_from_res_counter(memcg
->res
.parent
, res
);
6009 EXPORT_SYMBOL(parent_mem_cgroup
);
6011 static struct cgroup_subsys_state
* __ref
6012 mem_cgroup_css_alloc(struct cgroup_subsys_state
*parent_css
)
6014 struct mem_cgroup
*memcg
;
6015 long error
= -ENOMEM
;
6018 memcg
= mem_cgroup_alloc();
6020 return ERR_PTR(error
);
6023 if (alloc_mem_cgroup_per_zone_info(memcg
, node
))
6027 if (parent_css
== NULL
) {
6028 root_mem_cgroup
= memcg
;
6029 res_counter_init(&memcg
->res
, NULL
);
6030 res_counter_init(&memcg
->memsw
, NULL
);
6031 res_counter_init(&memcg
->kmem
, NULL
);
6034 memcg
->last_scanned_node
= MAX_NUMNODES
;
6035 INIT_LIST_HEAD(&memcg
->oom_notify
);
6036 memcg
->move_charge_at_immigrate
= 0;
6037 mutex_init(&memcg
->thresholds_lock
);
6038 spin_lock_init(&memcg
->move_lock
);
6039 vmpressure_init(&memcg
->vmpressure
);
6040 spin_lock_init(&memcg
->soft_lock
);
6045 __mem_cgroup_free(memcg
);
6046 return ERR_PTR(error
);
6050 mem_cgroup_css_online(struct cgroup_subsys_state
*css
)
6052 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6053 struct mem_cgroup
*parent
= mem_cgroup_from_css(css_parent(css
));
6059 mutex_lock(&memcg_create_mutex
);
6061 memcg
->use_hierarchy
= parent
->use_hierarchy
;
6062 memcg
->oom_kill_disable
= parent
->oom_kill_disable
;
6063 memcg
->swappiness
= mem_cgroup_swappiness(parent
);
6065 if (parent
->use_hierarchy
) {
6066 res_counter_init(&memcg
->res
, &parent
->res
);
6067 res_counter_init(&memcg
->memsw
, &parent
->memsw
);
6068 res_counter_init(&memcg
->kmem
, &parent
->kmem
);
6071 * No need to take a reference to the parent because cgroup
6072 * core guarantees its existence.
6075 res_counter_init(&memcg
->res
, NULL
);
6076 res_counter_init(&memcg
->memsw
, NULL
);
6077 res_counter_init(&memcg
->kmem
, NULL
);
6079 * Deeper hierachy with use_hierarchy == false doesn't make
6080 * much sense so let cgroup subsystem know about this
6081 * unfortunate state in our controller.
6083 if (parent
!= root_mem_cgroup
)
6084 mem_cgroup_subsys
.broken_hierarchy
= true;
6087 error
= memcg_init_kmem(memcg
, &mem_cgroup_subsys
);
6088 mutex_unlock(&memcg_create_mutex
);
6093 * Announce all parents that a group from their hierarchy is gone.
6095 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup
*memcg
)
6097 struct mem_cgroup
*parent
= memcg
;
6099 while ((parent
= parent_mem_cgroup(parent
)))
6100 mem_cgroup_iter_invalidate(parent
);
6103 * if the root memcg is not hierarchical we have to check it
6106 if (!root_mem_cgroup
->use_hierarchy
)
6107 mem_cgroup_iter_invalidate(root_mem_cgroup
);
6110 static void mem_cgroup_css_offline(struct cgroup_subsys_state
*css
)
6112 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6114 kmem_cgroup_css_offline(memcg
);
6116 mem_cgroup_invalidate_reclaim_iterators(memcg
);
6117 mem_cgroup_reparent_charges(memcg
);
6118 if (memcg
->soft_contributed
) {
6119 while ((memcg
= parent_mem_cgroup(memcg
)))
6120 atomic_dec(&memcg
->children_in_excess
);
6122 if (memcg
!= root_mem_cgroup
&& !root_mem_cgroup
->use_hierarchy
)
6123 atomic_dec(&root_mem_cgroup
->children_in_excess
);
6125 mem_cgroup_destroy_all_caches(memcg
);
6126 vmpressure_cleanup(&memcg
->vmpressure
);
6129 static void mem_cgroup_css_free(struct cgroup_subsys_state
*css
)
6131 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6133 memcg_destroy_kmem(memcg
);
6134 __mem_cgroup_free(memcg
);
6138 /* Handlers for move charge at task migration. */
6139 #define PRECHARGE_COUNT_AT_ONCE 256
6140 static int mem_cgroup_do_precharge(unsigned long count
)
6143 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6144 struct mem_cgroup
*memcg
= mc
.to
;
6146 if (mem_cgroup_is_root(memcg
)) {
6147 mc
.precharge
+= count
;
6148 /* we don't need css_get for root */
6151 /* try to charge at once */
6153 struct res_counter
*dummy
;
6155 * "memcg" cannot be under rmdir() because we've already checked
6156 * by cgroup_lock_live_cgroup() that it is not removed and we
6157 * are still under the same cgroup_mutex. So we can postpone
6160 if (res_counter_charge(&memcg
->res
, PAGE_SIZE
* count
, &dummy
))
6162 if (do_swap_account
&& res_counter_charge(&memcg
->memsw
,
6163 PAGE_SIZE
* count
, &dummy
)) {
6164 res_counter_uncharge(&memcg
->res
, PAGE_SIZE
* count
);
6167 mc
.precharge
+= count
;
6171 /* fall back to one by one charge */
6173 if (signal_pending(current
)) {
6177 if (!batch_count
--) {
6178 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
6181 ret
= __mem_cgroup_try_charge(NULL
,
6182 GFP_KERNEL
, 1, &memcg
, false);
6184 /* mem_cgroup_clear_mc() will do uncharge later */
6192 * get_mctgt_type - get target type of moving charge
6193 * @vma: the vma the pte to be checked belongs
6194 * @addr: the address corresponding to the pte to be checked
6195 * @ptent: the pte to be checked
6196 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6199 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6200 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6201 * move charge. if @target is not NULL, the page is stored in target->page
6202 * with extra refcnt got(Callers should handle it).
6203 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6204 * target for charge migration. if @target is not NULL, the entry is stored
6207 * Called with pte lock held.
6214 enum mc_target_type
{
6220 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
6221 unsigned long addr
, pte_t ptent
)
6223 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
6225 if (!page
|| !page_mapped(page
))
6227 if (PageAnon(page
)) {
6228 /* we don't move shared anon */
6231 } else if (!move_file())
6232 /* we ignore mapcount for file pages */
6234 if (!get_page_unless_zero(page
))
6241 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6242 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6244 struct page
*page
= NULL
;
6245 swp_entry_t ent
= pte_to_swp_entry(ptent
);
6247 if (!move_anon() || non_swap_entry(ent
))
6250 * Because lookup_swap_cache() updates some statistics counter,
6251 * we call find_get_page() with swapper_space directly.
6253 page
= find_get_page(swap_address_space(ent
), ent
.val
);
6254 if (do_swap_account
)
6255 entry
->val
= ent
.val
;
6260 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
6261 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6267 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
6268 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
6270 struct page
*page
= NULL
;
6271 struct address_space
*mapping
;
6274 if (!vma
->vm_file
) /* anonymous vma */
6279 mapping
= vma
->vm_file
->f_mapping
;
6280 if (pte_none(ptent
))
6281 pgoff
= linear_page_index(vma
, addr
);
6282 else /* pte_file(ptent) is true */
6283 pgoff
= pte_to_pgoff(ptent
);
6285 /* page is moved even if it's not RSS of this task(page-faulted). */
6286 page
= find_get_page(mapping
, pgoff
);
6289 /* shmem/tmpfs may report page out on swap: account for that too. */
6290 if (radix_tree_exceptional_entry(page
)) {
6291 swp_entry_t swap
= radix_to_swp_entry(page
);
6292 if (do_swap_account
)
6294 page
= find_get_page(swap_address_space(swap
), swap
.val
);
6300 static enum mc_target_type
get_mctgt_type(struct vm_area_struct
*vma
,
6301 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
6303 struct page
*page
= NULL
;
6304 struct page_cgroup
*pc
;
6305 enum mc_target_type ret
= MC_TARGET_NONE
;
6306 swp_entry_t ent
= { .val
= 0 };
6308 if (pte_present(ptent
))
6309 page
= mc_handle_present_pte(vma
, addr
, ptent
);
6310 else if (is_swap_pte(ptent
))
6311 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
6312 else if (pte_none(ptent
) || pte_file(ptent
))
6313 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
6315 if (!page
&& !ent
.val
)
6318 pc
= lookup_page_cgroup(page
);
6320 * Do only loose check w/o page_cgroup lock.
6321 * mem_cgroup_move_account() checks the pc is valid or not under
6324 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6325 ret
= MC_TARGET_PAGE
;
6327 target
->page
= page
;
6329 if (!ret
|| !target
)
6332 /* There is a swap entry and a page doesn't exist or isn't charged */
6333 if (ent
.val
&& !ret
&&
6334 css_id(&mc
.from
->css
) == lookup_swap_cgroup_id(ent
)) {
6335 ret
= MC_TARGET_SWAP
;
6342 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6344 * We don't consider swapping or file mapped pages because THP does not
6345 * support them for now.
6346 * Caller should make sure that pmd_trans_huge(pmd) is true.
6348 static enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6349 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6351 struct page
*page
= NULL
;
6352 struct page_cgroup
*pc
;
6353 enum mc_target_type ret
= MC_TARGET_NONE
;
6355 page
= pmd_page(pmd
);
6356 VM_BUG_ON(!page
|| !PageHead(page
));
6359 pc
= lookup_page_cgroup(page
);
6360 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
6361 ret
= MC_TARGET_PAGE
;
6364 target
->page
= page
;
6370 static inline enum mc_target_type
get_mctgt_type_thp(struct vm_area_struct
*vma
,
6371 unsigned long addr
, pmd_t pmd
, union mc_target
*target
)
6373 return MC_TARGET_NONE
;
6377 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
6378 unsigned long addr
, unsigned long end
,
6379 struct mm_walk
*walk
)
6381 struct vm_area_struct
*vma
= walk
->private;
6385 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6386 if (get_mctgt_type_thp(vma
, addr
, *pmd
, NULL
) == MC_TARGET_PAGE
)
6387 mc
.precharge
+= HPAGE_PMD_NR
;
6388 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6392 if (pmd_trans_unstable(pmd
))
6394 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6395 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
6396 if (get_mctgt_type(vma
, addr
, *pte
, NULL
))
6397 mc
.precharge
++; /* increment precharge temporarily */
6398 pte_unmap_unlock(pte
- 1, ptl
);
6404 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
6406 unsigned long precharge
;
6407 struct vm_area_struct
*vma
;
6409 down_read(&mm
->mmap_sem
);
6410 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6411 struct mm_walk mem_cgroup_count_precharge_walk
= {
6412 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
6416 if (is_vm_hugetlb_page(vma
))
6418 walk_page_range(vma
->vm_start
, vma
->vm_end
,
6419 &mem_cgroup_count_precharge_walk
);
6421 up_read(&mm
->mmap_sem
);
6423 precharge
= mc
.precharge
;
6429 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
6431 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
6433 VM_BUG_ON(mc
.moving_task
);
6434 mc
.moving_task
= current
;
6435 return mem_cgroup_do_precharge(precharge
);
6438 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6439 static void __mem_cgroup_clear_mc(void)
6441 struct mem_cgroup
*from
= mc
.from
;
6442 struct mem_cgroup
*to
= mc
.to
;
6445 /* we must uncharge all the leftover precharges from mc.to */
6447 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
6451 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6452 * we must uncharge here.
6454 if (mc
.moved_charge
) {
6455 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
6456 mc
.moved_charge
= 0;
6458 /* we must fixup refcnts and charges */
6459 if (mc
.moved_swap
) {
6460 /* uncharge swap account from the old cgroup */
6461 if (!mem_cgroup_is_root(mc
.from
))
6462 res_counter_uncharge(&mc
.from
->memsw
,
6463 PAGE_SIZE
* mc
.moved_swap
);
6465 for (i
= 0; i
< mc
.moved_swap
; i
++)
6466 css_put(&mc
.from
->css
);
6468 if (!mem_cgroup_is_root(mc
.to
)) {
6470 * we charged both to->res and to->memsw, so we should
6473 res_counter_uncharge(&mc
.to
->res
,
6474 PAGE_SIZE
* mc
.moved_swap
);
6476 /* we've already done css_get(mc.to) */
6479 memcg_oom_recover(from
);
6480 memcg_oom_recover(to
);
6481 wake_up_all(&mc
.waitq
);
6484 static void mem_cgroup_clear_mc(void)
6486 struct mem_cgroup
*from
= mc
.from
;
6489 * we must clear moving_task before waking up waiters at the end of
6492 mc
.moving_task
= NULL
;
6493 __mem_cgroup_clear_mc();
6494 spin_lock(&mc
.lock
);
6497 spin_unlock(&mc
.lock
);
6498 mem_cgroup_end_move(from
);
6501 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6502 struct cgroup_taskset
*tset
)
6504 struct task_struct
*p
= cgroup_taskset_first(tset
);
6506 struct mem_cgroup
*memcg
= mem_cgroup_from_css(css
);
6507 unsigned long move_charge_at_immigrate
;
6510 * We are now commited to this value whatever it is. Changes in this
6511 * tunable will only affect upcoming migrations, not the current one.
6512 * So we need to save it, and keep it going.
6514 move_charge_at_immigrate
= memcg
->move_charge_at_immigrate
;
6515 if (move_charge_at_immigrate
) {
6516 struct mm_struct
*mm
;
6517 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
6519 VM_BUG_ON(from
== memcg
);
6521 mm
= get_task_mm(p
);
6524 /* We move charges only when we move a owner of the mm */
6525 if (mm
->owner
== p
) {
6528 VM_BUG_ON(mc
.precharge
);
6529 VM_BUG_ON(mc
.moved_charge
);
6530 VM_BUG_ON(mc
.moved_swap
);
6531 mem_cgroup_start_move(from
);
6532 spin_lock(&mc
.lock
);
6535 mc
.immigrate_flags
= move_charge_at_immigrate
;
6536 spin_unlock(&mc
.lock
);
6537 /* We set mc.moving_task later */
6539 ret
= mem_cgroup_precharge_mc(mm
);
6541 mem_cgroup_clear_mc();
6548 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6549 struct cgroup_taskset
*tset
)
6551 mem_cgroup_clear_mc();
6554 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
6555 unsigned long addr
, unsigned long end
,
6556 struct mm_walk
*walk
)
6559 struct vm_area_struct
*vma
= walk
->private;
6562 enum mc_target_type target_type
;
6563 union mc_target target
;
6565 struct page_cgroup
*pc
;
6568 * We don't take compound_lock() here but no race with splitting thp
6570 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6571 * under splitting, which means there's no concurrent thp split,
6572 * - if another thread runs into split_huge_page() just after we
6573 * entered this if-block, the thread must wait for page table lock
6574 * to be unlocked in __split_huge_page_splitting(), where the main
6575 * part of thp split is not executed yet.
6577 if (pmd_trans_huge_lock(pmd
, vma
) == 1) {
6578 if (mc
.precharge
< HPAGE_PMD_NR
) {
6579 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6582 target_type
= get_mctgt_type_thp(vma
, addr
, *pmd
, &target
);
6583 if (target_type
== MC_TARGET_PAGE
) {
6585 if (!isolate_lru_page(page
)) {
6586 pc
= lookup_page_cgroup(page
);
6587 if (!mem_cgroup_move_account(page
, HPAGE_PMD_NR
,
6588 pc
, mc
.from
, mc
.to
)) {
6589 mc
.precharge
-= HPAGE_PMD_NR
;
6590 mc
.moved_charge
+= HPAGE_PMD_NR
;
6592 putback_lru_page(page
);
6596 spin_unlock(&vma
->vm_mm
->page_table_lock
);
6600 if (pmd_trans_unstable(pmd
))
6603 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
6604 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
6605 pte_t ptent
= *(pte
++);
6611 switch (get_mctgt_type(vma
, addr
, ptent
, &target
)) {
6612 case MC_TARGET_PAGE
:
6614 if (isolate_lru_page(page
))
6616 pc
= lookup_page_cgroup(page
);
6617 if (!mem_cgroup_move_account(page
, 1, pc
,
6620 /* we uncharge from mc.from later. */
6623 putback_lru_page(page
);
6624 put
: /* get_mctgt_type() gets the page */
6627 case MC_TARGET_SWAP
:
6629 if (!mem_cgroup_move_swap_account(ent
, mc
.from
, mc
.to
)) {
6631 /* we fixup refcnts and charges later. */
6639 pte_unmap_unlock(pte
- 1, ptl
);
6644 * We have consumed all precharges we got in can_attach().
6645 * We try charge one by one, but don't do any additional
6646 * charges to mc.to if we have failed in charge once in attach()
6649 ret
= mem_cgroup_do_precharge(1);
6657 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
6659 struct vm_area_struct
*vma
;
6661 lru_add_drain_all();
6663 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
6665 * Someone who are holding the mmap_sem might be waiting in
6666 * waitq. So we cancel all extra charges, wake up all waiters,
6667 * and retry. Because we cancel precharges, we might not be able
6668 * to move enough charges, but moving charge is a best-effort
6669 * feature anyway, so it wouldn't be a big problem.
6671 __mem_cgroup_clear_mc();
6675 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
6677 struct mm_walk mem_cgroup_move_charge_walk
= {
6678 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
6682 if (is_vm_hugetlb_page(vma
))
6684 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
6685 &mem_cgroup_move_charge_walk
);
6688 * means we have consumed all precharges and failed in
6689 * doing additional charge. Just abandon here.
6693 up_read(&mm
->mmap_sem
);
6696 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6697 struct cgroup_taskset
*tset
)
6699 struct task_struct
*p
= cgroup_taskset_first(tset
);
6700 struct mm_struct
*mm
= get_task_mm(p
);
6704 mem_cgroup_move_charge(mm
);
6708 mem_cgroup_clear_mc();
6710 #else /* !CONFIG_MMU */
6711 static int mem_cgroup_can_attach(struct cgroup_subsys_state
*css
,
6712 struct cgroup_taskset
*tset
)
6716 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state
*css
,
6717 struct cgroup_taskset
*tset
)
6720 static void mem_cgroup_move_task(struct cgroup_subsys_state
*css
,
6721 struct cgroup_taskset
*tset
)
6727 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6728 * to verify sane_behavior flag on each mount attempt.
6730 static void mem_cgroup_bind(struct cgroup_subsys_state
*root_css
)
6733 * use_hierarchy is forced with sane_behavior. cgroup core
6734 * guarantees that @root doesn't have any children, so turning it
6735 * on for the root memcg is enough.
6737 if (cgroup_sane_behavior(root_css
->cgroup
))
6738 mem_cgroup_from_css(root_css
)->use_hierarchy
= true;
6741 struct cgroup_subsys mem_cgroup_subsys
= {
6743 .subsys_id
= mem_cgroup_subsys_id
,
6744 .css_alloc
= mem_cgroup_css_alloc
,
6745 .css_online
= mem_cgroup_css_online
,
6746 .css_offline
= mem_cgroup_css_offline
,
6747 .css_free
= mem_cgroup_css_free
,
6748 .can_attach
= mem_cgroup_can_attach
,
6749 .cancel_attach
= mem_cgroup_cancel_attach
,
6750 .attach
= mem_cgroup_move_task
,
6751 .bind
= mem_cgroup_bind
,
6752 .base_cftypes
= mem_cgroup_files
,
6757 #ifdef CONFIG_MEMCG_SWAP
6758 static int __init
enable_swap_account(char *s
)
6760 if (!strcmp(s
, "1"))
6761 really_do_swap_account
= 1;
6762 else if (!strcmp(s
, "0"))
6763 really_do_swap_account
= 0;
6766 __setup("swapaccount=", enable_swap_account
);
6768 static void __init
memsw_file_init(void)
6770 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys
, memsw_cgroup_files
));
6773 static void __init
enable_swap_cgroup(void)
6775 if (!mem_cgroup_disabled() && really_do_swap_account
) {
6776 do_swap_account
= 1;
6782 static void __init
enable_swap_cgroup(void)
6788 * subsys_initcall() for memory controller.
6790 * Some parts like hotcpu_notifier() have to be initialized from this context
6791 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6792 * everything that doesn't depend on a specific mem_cgroup structure should
6793 * be initialized from here.
6795 static int __init
mem_cgroup_init(void)
6797 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
6798 enable_swap_cgroup();
6802 subsys_initcall(mem_cgroup_init
);