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 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
50 #include <linux/oom.h>
53 #include <asm/uaccess.h>
55 #include <trace/events/vmscan.h>
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
58 #define MEM_CGROUP_RECLAIM_RETRIES 5
59 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly
;
65 /* for remember boot option*/
66 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
67 static int really_do_swap_account __initdata
= 1;
69 static int really_do_swap_account __initdata
= 0;
73 #define do_swap_account (0)
77 * Per memcg event counter is incremented at every pagein/pageout. This counter
78 * is used for trigger some periodic events. This is straightforward and better
79 * than using jiffies etc. to handle periodic memcg event.
81 * These values will be used as !((event) & ((1 <<(thresh)) - 1))
83 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
84 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
87 * Statistics for memory cgroup.
89 enum mem_cgroup_stat_index
{
91 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
93 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
94 MEM_CGROUP_STAT_RSS
, /* # of pages charged as anon rss */
95 MEM_CGROUP_STAT_FILE_MAPPED
, /* # of pages charged as file rss */
96 MEM_CGROUP_STAT_PGPGIN_COUNT
, /* # of pages paged in */
97 MEM_CGROUP_STAT_PGPGOUT_COUNT
, /* # of pages paged out */
98 MEM_CGROUP_STAT_SWAPOUT
, /* # of pages, swapped out */
99 MEM_CGROUP_STAT_DATA
, /* end of data requires synchronization */
100 /* incremented at every pagein/pageout */
101 MEM_CGROUP_EVENTS
= MEM_CGROUP_STAT_DATA
,
102 MEM_CGROUP_ON_MOVE
, /* someone is moving account between groups */
104 MEM_CGROUP_STAT_NSTATS
,
107 struct mem_cgroup_stat_cpu
{
108 s64 count
[MEM_CGROUP_STAT_NSTATS
];
112 * per-zone information in memory controller.
114 struct mem_cgroup_per_zone
{
116 * spin_lock to protect the per cgroup LRU
118 struct list_head lists
[NR_LRU_LISTS
];
119 unsigned long count
[NR_LRU_LISTS
];
121 struct zone_reclaim_stat reclaim_stat
;
122 struct rb_node tree_node
; /* RB tree node */
123 unsigned long long usage_in_excess
;/* Set to the value by which */
124 /* the soft limit is exceeded*/
126 struct mem_cgroup
*mem
; /* Back pointer, we cannot */
127 /* use container_of */
129 /* Macro for accessing counter */
130 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
132 struct mem_cgroup_per_node
{
133 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
136 struct mem_cgroup_lru_info
{
137 struct mem_cgroup_per_node
*nodeinfo
[MAX_NUMNODES
];
141 * Cgroups above their limits are maintained in a RB-Tree, independent of
142 * their hierarchy representation
145 struct mem_cgroup_tree_per_zone
{
146 struct rb_root rb_root
;
150 struct mem_cgroup_tree_per_node
{
151 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
154 struct mem_cgroup_tree
{
155 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
158 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
160 struct mem_cgroup_threshold
{
161 struct eventfd_ctx
*eventfd
;
166 struct mem_cgroup_threshold_ary
{
167 /* An array index points to threshold just below usage. */
168 int current_threshold
;
169 /* Size of entries[] */
171 /* Array of thresholds */
172 struct mem_cgroup_threshold entries
[0];
175 struct mem_cgroup_thresholds
{
176 /* Primary thresholds array */
177 struct mem_cgroup_threshold_ary
*primary
;
179 * Spare threshold array.
180 * This is needed to make mem_cgroup_unregister_event() "never fail".
181 * It must be able to store at least primary->size - 1 entries.
183 struct mem_cgroup_threshold_ary
*spare
;
187 struct mem_cgroup_eventfd_list
{
188 struct list_head list
;
189 struct eventfd_ctx
*eventfd
;
192 static void mem_cgroup_threshold(struct mem_cgroup
*mem
);
193 static void mem_cgroup_oom_notify(struct mem_cgroup
*mem
);
196 * The memory controller data structure. The memory controller controls both
197 * page cache and RSS per cgroup. We would eventually like to provide
198 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
199 * to help the administrator determine what knobs to tune.
201 * TODO: Add a water mark for the memory controller. Reclaim will begin when
202 * we hit the water mark. May be even add a low water mark, such that
203 * no reclaim occurs from a cgroup at it's low water mark, this is
204 * a feature that will be implemented much later in the future.
207 struct cgroup_subsys_state css
;
209 * the counter to account for memory usage
211 struct res_counter res
;
213 * the counter to account for mem+swap usage.
215 struct res_counter memsw
;
217 * Per cgroup active and inactive list, similar to the
218 * per zone LRU lists.
220 struct mem_cgroup_lru_info info
;
223 protect against reclaim related member.
225 spinlock_t reclaim_param_lock
;
228 * While reclaiming in a hierarchy, we cache the last child we
231 int last_scanned_child
;
233 * Should the accounting and control be hierarchical, per subtree?
239 unsigned int swappiness
;
240 /* OOM-Killer disable */
241 int oom_kill_disable
;
243 /* set when res.limit == memsw.limit */
244 bool memsw_is_minimum
;
246 /* protect arrays of thresholds */
247 struct mutex thresholds_lock
;
249 /* thresholds for memory usage. RCU-protected */
250 struct mem_cgroup_thresholds thresholds
;
252 /* thresholds for mem+swap usage. RCU-protected */
253 struct mem_cgroup_thresholds memsw_thresholds
;
255 /* For oom notifier event fd */
256 struct list_head oom_notify
;
259 * Should we move charges of a task when a task is moved into this
260 * mem_cgroup ? And what type of charges should we move ?
262 unsigned long move_charge_at_immigrate
;
266 struct mem_cgroup_stat_cpu
*stat
;
268 * used when a cpu is offlined or other synchronizations
269 * See mem_cgroup_read_stat().
271 struct mem_cgroup_stat_cpu nocpu_base
;
272 spinlock_t pcp_counter_lock
;
275 /* Stuffs for move charges at task migration. */
277 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
278 * left-shifted bitmap of these types.
281 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
282 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
286 /* "mc" and its members are protected by cgroup_mutex */
287 static struct move_charge_struct
{
288 spinlock_t lock
; /* for from, to */
289 struct mem_cgroup
*from
;
290 struct mem_cgroup
*to
;
291 unsigned long precharge
;
292 unsigned long moved_charge
;
293 unsigned long moved_swap
;
294 struct task_struct
*moving_task
; /* a task moving charges */
295 wait_queue_head_t waitq
; /* a waitq for other context */
297 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
298 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
301 static bool move_anon(void)
303 return test_bit(MOVE_CHARGE_TYPE_ANON
,
304 &mc
.to
->move_charge_at_immigrate
);
307 static bool move_file(void)
309 return test_bit(MOVE_CHARGE_TYPE_FILE
,
310 &mc
.to
->move_charge_at_immigrate
);
314 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
315 * limit reclaim to prevent infinite loops, if they ever occur.
317 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
318 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
321 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
322 MEM_CGROUP_CHARGE_TYPE_MAPPED
,
323 MEM_CGROUP_CHARGE_TYPE_SHMEM
, /* used by page migration of shmem */
324 MEM_CGROUP_CHARGE_TYPE_FORCE
, /* used by force_empty */
325 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
326 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
330 /* for encoding cft->private value on file */
333 #define _OOM_TYPE (2)
334 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
335 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
336 #define MEMFILE_ATTR(val) ((val) & 0xffff)
337 /* Used for OOM nofiier */
338 #define OOM_CONTROL (0)
341 * Reclaim flags for mem_cgroup_hierarchical_reclaim
343 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
344 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
345 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
346 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
347 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
348 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
350 static void mem_cgroup_get(struct mem_cgroup
*mem
);
351 static void mem_cgroup_put(struct mem_cgroup
*mem
);
352 static struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*mem
);
353 static void drain_all_stock_async(void);
355 static struct mem_cgroup_per_zone
*
356 mem_cgroup_zoneinfo(struct mem_cgroup
*mem
, int nid
, int zid
)
358 return &mem
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
361 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*mem
)
366 static struct mem_cgroup_per_zone
*
367 page_cgroup_zoneinfo(struct page_cgroup
*pc
)
369 struct mem_cgroup
*mem
= pc
->mem_cgroup
;
370 int nid
= page_cgroup_nid(pc
);
371 int zid
= page_cgroup_zid(pc
);
376 return mem_cgroup_zoneinfo(mem
, nid
, zid
);
379 static struct mem_cgroup_tree_per_zone
*
380 soft_limit_tree_node_zone(int nid
, int zid
)
382 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
385 static struct mem_cgroup_tree_per_zone
*
386 soft_limit_tree_from_page(struct page
*page
)
388 int nid
= page_to_nid(page
);
389 int zid
= page_zonenum(page
);
391 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
395 __mem_cgroup_insert_exceeded(struct mem_cgroup
*mem
,
396 struct mem_cgroup_per_zone
*mz
,
397 struct mem_cgroup_tree_per_zone
*mctz
,
398 unsigned long long new_usage_in_excess
)
400 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
401 struct rb_node
*parent
= NULL
;
402 struct mem_cgroup_per_zone
*mz_node
;
407 mz
->usage_in_excess
= new_usage_in_excess
;
408 if (!mz
->usage_in_excess
)
412 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
414 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
417 * We can't avoid mem cgroups that are over their soft
418 * limit by the same amount
420 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
423 rb_link_node(&mz
->tree_node
, parent
, p
);
424 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
429 __mem_cgroup_remove_exceeded(struct mem_cgroup
*mem
,
430 struct mem_cgroup_per_zone
*mz
,
431 struct mem_cgroup_tree_per_zone
*mctz
)
435 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
440 mem_cgroup_remove_exceeded(struct mem_cgroup
*mem
,
441 struct mem_cgroup_per_zone
*mz
,
442 struct mem_cgroup_tree_per_zone
*mctz
)
444 spin_lock(&mctz
->lock
);
445 __mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
446 spin_unlock(&mctz
->lock
);
450 static void mem_cgroup_update_tree(struct mem_cgroup
*mem
, struct page
*page
)
452 unsigned long long excess
;
453 struct mem_cgroup_per_zone
*mz
;
454 struct mem_cgroup_tree_per_zone
*mctz
;
455 int nid
= page_to_nid(page
);
456 int zid
= page_zonenum(page
);
457 mctz
= soft_limit_tree_from_page(page
);
460 * Necessary to update all ancestors when hierarchy is used.
461 * because their event counter is not touched.
463 for (; mem
; mem
= parent_mem_cgroup(mem
)) {
464 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
465 excess
= res_counter_soft_limit_excess(&mem
->res
);
467 * We have to update the tree if mz is on RB-tree or
468 * mem is over its softlimit.
470 if (excess
|| mz
->on_tree
) {
471 spin_lock(&mctz
->lock
);
472 /* if on-tree, remove it */
474 __mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
476 * Insert again. mz->usage_in_excess will be updated.
477 * If excess is 0, no tree ops.
479 __mem_cgroup_insert_exceeded(mem
, mz
, mctz
, excess
);
480 spin_unlock(&mctz
->lock
);
485 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*mem
)
488 struct mem_cgroup_per_zone
*mz
;
489 struct mem_cgroup_tree_per_zone
*mctz
;
491 for_each_node_state(node
, N_POSSIBLE
) {
492 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
493 mz
= mem_cgroup_zoneinfo(mem
, node
, zone
);
494 mctz
= soft_limit_tree_node_zone(node
, zone
);
495 mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
500 static struct mem_cgroup_per_zone
*
501 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
503 struct rb_node
*rightmost
= NULL
;
504 struct mem_cgroup_per_zone
*mz
;
508 rightmost
= rb_last(&mctz
->rb_root
);
510 goto done
; /* Nothing to reclaim from */
512 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
514 * Remove the node now but someone else can add it back,
515 * we will to add it back at the end of reclaim to its correct
516 * position in the tree.
518 __mem_cgroup_remove_exceeded(mz
->mem
, mz
, mctz
);
519 if (!res_counter_soft_limit_excess(&mz
->mem
->res
) ||
520 !css_tryget(&mz
->mem
->css
))
526 static struct mem_cgroup_per_zone
*
527 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
529 struct mem_cgroup_per_zone
*mz
;
531 spin_lock(&mctz
->lock
);
532 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
533 spin_unlock(&mctz
->lock
);
538 * Implementation Note: reading percpu statistics for memcg.
540 * Both of vmstat[] and percpu_counter has threshold and do periodic
541 * synchronization to implement "quick" read. There are trade-off between
542 * reading cost and precision of value. Then, we may have a chance to implement
543 * a periodic synchronizion of counter in memcg's counter.
545 * But this _read() function is used for user interface now. The user accounts
546 * memory usage by memory cgroup and he _always_ requires exact value because
547 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
548 * have to visit all online cpus and make sum. So, for now, unnecessary
549 * synchronization is not implemented. (just implemented for cpu hotplug)
551 * If there are kernel internal actions which can make use of some not-exact
552 * value, and reading all cpu value can be performance bottleneck in some
553 * common workload, threashold and synchonization as vmstat[] should be
556 static s64
mem_cgroup_read_stat(struct mem_cgroup
*mem
,
557 enum mem_cgroup_stat_index idx
)
563 for_each_online_cpu(cpu
)
564 val
+= per_cpu(mem
->stat
->count
[idx
], cpu
);
565 #ifdef CONFIG_HOTPLUG_CPU
566 spin_lock(&mem
->pcp_counter_lock
);
567 val
+= mem
->nocpu_base
.count
[idx
];
568 spin_unlock(&mem
->pcp_counter_lock
);
574 static s64
mem_cgroup_local_usage(struct mem_cgroup
*mem
)
578 ret
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_RSS
);
579 ret
+= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_CACHE
);
583 static void mem_cgroup_swap_statistics(struct mem_cgroup
*mem
,
586 int val
= (charge
) ? 1 : -1;
587 this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_SWAPOUT
], val
);
590 static void mem_cgroup_charge_statistics(struct mem_cgroup
*mem
,
591 bool file
, int nr_pages
)
596 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_pages
);
598 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_pages
);
600 /* pagein of a big page is an event. So, ignore page size */
602 __this_cpu_inc(mem
->stat
->count
[MEM_CGROUP_STAT_PGPGIN_COUNT
]);
604 __this_cpu_inc(mem
->stat
->count
[MEM_CGROUP_STAT_PGPGOUT_COUNT
]);
605 nr_pages
= -nr_pages
; /* for event */
608 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_EVENTS
], nr_pages
);
613 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup
*mem
,
617 struct mem_cgroup_per_zone
*mz
;
620 for_each_online_node(nid
)
621 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
622 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
623 total
+= MEM_CGROUP_ZSTAT(mz
, idx
);
628 static bool __memcg_event_check(struct mem_cgroup
*mem
, int event_mask_shift
)
632 val
= this_cpu_read(mem
->stat
->count
[MEM_CGROUP_EVENTS
]);
634 return !(val
& ((1 << event_mask_shift
) - 1));
638 * Check events in order.
641 static void memcg_check_events(struct mem_cgroup
*mem
, struct page
*page
)
643 /* threshold event is triggered in finer grain than soft limit */
644 if (unlikely(__memcg_event_check(mem
, THRESHOLDS_EVENTS_THRESH
))) {
645 mem_cgroup_threshold(mem
);
646 if (unlikely(__memcg_event_check(mem
, SOFTLIMIT_EVENTS_THRESH
)))
647 mem_cgroup_update_tree(mem
, page
);
651 static struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
653 return container_of(cgroup_subsys_state(cont
,
654 mem_cgroup_subsys_id
), struct mem_cgroup
,
658 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
661 * mm_update_next_owner() may clear mm->owner to NULL
662 * if it races with swapoff, page migration, etc.
663 * So this can be called with p == NULL.
668 return container_of(task_subsys_state(p
, mem_cgroup_subsys_id
),
669 struct mem_cgroup
, css
);
672 static struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
674 struct mem_cgroup
*mem
= NULL
;
679 * Because we have no locks, mm->owner's may be being moved to other
680 * cgroup. We use css_tryget() here even if this looks
681 * pessimistic (rather than adding locks here).
685 mem
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
688 } while (!css_tryget(&mem
->css
));
693 /* The caller has to guarantee "mem" exists before calling this */
694 static struct mem_cgroup
*mem_cgroup_start_loop(struct mem_cgroup
*mem
)
696 struct cgroup_subsys_state
*css
;
699 if (!mem
) /* ROOT cgroup has the smallest ID */
700 return root_mem_cgroup
; /*css_put/get against root is ignored*/
701 if (!mem
->use_hierarchy
) {
702 if (css_tryget(&mem
->css
))
708 * searching a memory cgroup which has the smallest ID under given
709 * ROOT cgroup. (ID >= 1)
711 css
= css_get_next(&mem_cgroup_subsys
, 1, &mem
->css
, &found
);
712 if (css
&& css_tryget(css
))
713 mem
= container_of(css
, struct mem_cgroup
, css
);
720 static struct mem_cgroup
*mem_cgroup_get_next(struct mem_cgroup
*iter
,
721 struct mem_cgroup
*root
,
724 int nextid
= css_id(&iter
->css
) + 1;
727 struct cgroup_subsys_state
*css
;
729 hierarchy_used
= iter
->use_hierarchy
;
732 /* If no ROOT, walk all, ignore hierarchy */
733 if (!cond
|| (root
&& !hierarchy_used
))
737 root
= root_mem_cgroup
;
743 css
= css_get_next(&mem_cgroup_subsys
, nextid
,
745 if (css
&& css_tryget(css
))
746 iter
= container_of(css
, struct mem_cgroup
, css
);
748 /* If css is NULL, no more cgroups will be found */
750 } while (css
&& !iter
);
755 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
756 * be careful that "break" loop is not allowed. We have reference count.
757 * Instead of that modify "cond" to be false and "continue" to exit the loop.
759 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
760 for (iter = mem_cgroup_start_loop(root);\
762 iter = mem_cgroup_get_next(iter, root, cond))
764 #define for_each_mem_cgroup_tree(iter, root) \
765 for_each_mem_cgroup_tree_cond(iter, root, true)
767 #define for_each_mem_cgroup_all(iter) \
768 for_each_mem_cgroup_tree_cond(iter, NULL, true)
771 static inline bool mem_cgroup_is_root(struct mem_cgroup
*mem
)
773 return (mem
== root_mem_cgroup
);
777 * Following LRU functions are allowed to be used without PCG_LOCK.
778 * Operations are called by routine of global LRU independently from memcg.
779 * What we have to take care of here is validness of pc->mem_cgroup.
781 * Changes to pc->mem_cgroup happens when
784 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
785 * It is added to LRU before charge.
786 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
787 * When moving account, the page is not on LRU. It's isolated.
790 void mem_cgroup_del_lru_list(struct page
*page
, enum lru_list lru
)
792 struct page_cgroup
*pc
;
793 struct mem_cgroup_per_zone
*mz
;
795 if (mem_cgroup_disabled())
797 pc
= lookup_page_cgroup(page
);
798 /* can happen while we handle swapcache. */
799 if (!TestClearPageCgroupAcctLRU(pc
))
801 VM_BUG_ON(!pc
->mem_cgroup
);
803 * We don't check PCG_USED bit. It's cleared when the "page" is finally
804 * removed from global LRU.
806 mz
= page_cgroup_zoneinfo(pc
);
807 /* huge page split is done under lru_lock. so, we have no races. */
808 MEM_CGROUP_ZSTAT(mz
, lru
) -= 1 << compound_order(page
);
809 if (mem_cgroup_is_root(pc
->mem_cgroup
))
811 VM_BUG_ON(list_empty(&pc
->lru
));
812 list_del_init(&pc
->lru
);
815 void mem_cgroup_del_lru(struct page
*page
)
817 mem_cgroup_del_lru_list(page
, page_lru(page
));
821 * Writeback is about to end against a page which has been marked for immediate
822 * reclaim. If it still appears to be reclaimable, move it to the tail of the
825 void mem_cgroup_rotate_reclaimable_page(struct page
*page
)
827 struct mem_cgroup_per_zone
*mz
;
828 struct page_cgroup
*pc
;
829 enum lru_list lru
= page_lru(page
);
831 if (mem_cgroup_disabled())
834 pc
= lookup_page_cgroup(page
);
835 /* unused or root page is not rotated. */
836 if (!PageCgroupUsed(pc
))
838 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
840 if (mem_cgroup_is_root(pc
->mem_cgroup
))
842 mz
= page_cgroup_zoneinfo(pc
);
843 list_move_tail(&pc
->lru
, &mz
->lists
[lru
]);
846 void mem_cgroup_rotate_lru_list(struct page
*page
, enum lru_list lru
)
848 struct mem_cgroup_per_zone
*mz
;
849 struct page_cgroup
*pc
;
851 if (mem_cgroup_disabled())
854 pc
= lookup_page_cgroup(page
);
855 /* unused or root page is not rotated. */
856 if (!PageCgroupUsed(pc
))
858 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
860 if (mem_cgroup_is_root(pc
->mem_cgroup
))
862 mz
= page_cgroup_zoneinfo(pc
);
863 list_move(&pc
->lru
, &mz
->lists
[lru
]);
866 void mem_cgroup_add_lru_list(struct page
*page
, enum lru_list lru
)
868 struct page_cgroup
*pc
;
869 struct mem_cgroup_per_zone
*mz
;
871 if (mem_cgroup_disabled())
873 pc
= lookup_page_cgroup(page
);
874 VM_BUG_ON(PageCgroupAcctLRU(pc
));
875 if (!PageCgroupUsed(pc
))
877 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
879 mz
= page_cgroup_zoneinfo(pc
);
880 /* huge page split is done under lru_lock. so, we have no races. */
881 MEM_CGROUP_ZSTAT(mz
, lru
) += 1 << compound_order(page
);
882 SetPageCgroupAcctLRU(pc
);
883 if (mem_cgroup_is_root(pc
->mem_cgroup
))
885 list_add(&pc
->lru
, &mz
->lists
[lru
]);
889 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
890 * lru because the page may.be reused after it's fully uncharged (because of
891 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
892 * it again. This function is only used to charge SwapCache. It's done under
893 * lock_page and expected that zone->lru_lock is never held.
895 static void mem_cgroup_lru_del_before_commit_swapcache(struct page
*page
)
898 struct zone
*zone
= page_zone(page
);
899 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
901 spin_lock_irqsave(&zone
->lru_lock
, flags
);
903 * Forget old LRU when this page_cgroup is *not* used. This Used bit
904 * is guarded by lock_page() because the page is SwapCache.
906 if (!PageCgroupUsed(pc
))
907 mem_cgroup_del_lru_list(page
, page_lru(page
));
908 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
911 static void mem_cgroup_lru_add_after_commit_swapcache(struct page
*page
)
914 struct zone
*zone
= page_zone(page
);
915 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
917 spin_lock_irqsave(&zone
->lru_lock
, flags
);
918 /* link when the page is linked to LRU but page_cgroup isn't */
919 if (PageLRU(page
) && !PageCgroupAcctLRU(pc
))
920 mem_cgroup_add_lru_list(page
, page_lru(page
));
921 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
925 void mem_cgroup_move_lists(struct page
*page
,
926 enum lru_list from
, enum lru_list to
)
928 if (mem_cgroup_disabled())
930 mem_cgroup_del_lru_list(page
, from
);
931 mem_cgroup_add_lru_list(page
, to
);
934 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
937 struct mem_cgroup
*curr
= NULL
;
938 struct task_struct
*p
;
940 p
= find_lock_task_mm(task
);
943 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
948 * We should check use_hierarchy of "mem" not "curr". Because checking
949 * use_hierarchy of "curr" here make this function true if hierarchy is
950 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
951 * hierarchy(even if use_hierarchy is disabled in "mem").
953 if (mem
->use_hierarchy
)
954 ret
= css_is_ancestor(&curr
->css
, &mem
->css
);
961 static int calc_inactive_ratio(struct mem_cgroup
*memcg
, unsigned long *present_pages
)
963 unsigned long active
;
964 unsigned long inactive
;
966 unsigned long inactive_ratio
;
968 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_ANON
);
969 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_ANON
);
971 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
973 inactive_ratio
= int_sqrt(10 * gb
);
978 present_pages
[0] = inactive
;
979 present_pages
[1] = active
;
982 return inactive_ratio
;
985 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup
*memcg
)
987 unsigned long active
;
988 unsigned long inactive
;
989 unsigned long present_pages
[2];
990 unsigned long inactive_ratio
;
992 inactive_ratio
= calc_inactive_ratio(memcg
, present_pages
);
994 inactive
= present_pages
[0];
995 active
= present_pages
[1];
997 if (inactive
* inactive_ratio
< active
)
1003 int mem_cgroup_inactive_file_is_low(struct mem_cgroup
*memcg
)
1005 unsigned long active
;
1006 unsigned long inactive
;
1008 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_FILE
);
1009 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_FILE
);
1011 return (active
> inactive
);
1014 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup
*memcg
,
1018 int nid
= zone_to_nid(zone
);
1019 int zid
= zone_idx(zone
);
1020 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
1022 return MEM_CGROUP_ZSTAT(mz
, lru
);
1025 struct zone_reclaim_stat
*mem_cgroup_get_reclaim_stat(struct mem_cgroup
*memcg
,
1028 int nid
= zone_to_nid(zone
);
1029 int zid
= zone_idx(zone
);
1030 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
1032 return &mz
->reclaim_stat
;
1035 struct zone_reclaim_stat
*
1036 mem_cgroup_get_reclaim_stat_from_page(struct page
*page
)
1038 struct page_cgroup
*pc
;
1039 struct mem_cgroup_per_zone
*mz
;
1041 if (mem_cgroup_disabled())
1044 pc
= lookup_page_cgroup(page
);
1045 if (!PageCgroupUsed(pc
))
1047 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1049 mz
= page_cgroup_zoneinfo(pc
);
1053 return &mz
->reclaim_stat
;
1056 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
1057 struct list_head
*dst
,
1058 unsigned long *scanned
, int order
,
1059 int mode
, struct zone
*z
,
1060 struct mem_cgroup
*mem_cont
,
1061 int active
, int file
)
1063 unsigned long nr_taken
= 0;
1067 struct list_head
*src
;
1068 struct page_cgroup
*pc
, *tmp
;
1069 int nid
= zone_to_nid(z
);
1070 int zid
= zone_idx(z
);
1071 struct mem_cgroup_per_zone
*mz
;
1072 int lru
= LRU_FILE
* file
+ active
;
1076 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
1077 src
= &mz
->lists
[lru
];
1080 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
1081 if (scan
>= nr_to_scan
)
1085 if (unlikely(!PageCgroupUsed(pc
)))
1087 if (unlikely(!PageLRU(page
)))
1091 ret
= __isolate_lru_page(page
, mode
, file
);
1094 list_move(&page
->lru
, dst
);
1095 mem_cgroup_del_lru(page
);
1096 nr_taken
+= hpage_nr_pages(page
);
1099 /* we don't affect global LRU but rotate in our LRU */
1100 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
1109 trace_mm_vmscan_memcg_isolate(0, nr_to_scan
, scan
, nr_taken
,
1115 #define mem_cgroup_from_res_counter(counter, member) \
1116 container_of(counter, struct mem_cgroup, member)
1119 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1120 * @mem: the memory cgroup
1122 * Returns the maximum amount of memory @mem can be charged with, in
1125 static unsigned long long mem_cgroup_margin(struct mem_cgroup
*mem
)
1127 unsigned long long margin
;
1129 margin
= res_counter_margin(&mem
->res
);
1130 if (do_swap_account
)
1131 margin
= min(margin
, res_counter_margin(&mem
->memsw
));
1135 static unsigned int get_swappiness(struct mem_cgroup
*memcg
)
1137 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1138 unsigned int swappiness
;
1141 if (cgrp
->parent
== NULL
)
1142 return vm_swappiness
;
1144 spin_lock(&memcg
->reclaim_param_lock
);
1145 swappiness
= memcg
->swappiness
;
1146 spin_unlock(&memcg
->reclaim_param_lock
);
1151 static void mem_cgroup_start_move(struct mem_cgroup
*mem
)
1156 spin_lock(&mem
->pcp_counter_lock
);
1157 for_each_online_cpu(cpu
)
1158 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) += 1;
1159 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] += 1;
1160 spin_unlock(&mem
->pcp_counter_lock
);
1166 static void mem_cgroup_end_move(struct mem_cgroup
*mem
)
1173 spin_lock(&mem
->pcp_counter_lock
);
1174 for_each_online_cpu(cpu
)
1175 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) -= 1;
1176 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] -= 1;
1177 spin_unlock(&mem
->pcp_counter_lock
);
1181 * 2 routines for checking "mem" is under move_account() or not.
1183 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1184 * for avoiding race in accounting. If true,
1185 * pc->mem_cgroup may be overwritten.
1187 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1188 * under hierarchy of moving cgroups. This is for
1189 * waiting at hith-memory prressure caused by "move".
1192 static bool mem_cgroup_stealed(struct mem_cgroup
*mem
)
1194 VM_BUG_ON(!rcu_read_lock_held());
1195 return this_cpu_read(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
]) > 0;
1198 static bool mem_cgroup_under_move(struct mem_cgroup
*mem
)
1200 struct mem_cgroup
*from
;
1201 struct mem_cgroup
*to
;
1204 * Unlike task_move routines, we access mc.to, mc.from not under
1205 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1207 spin_lock(&mc
.lock
);
1212 if (from
== mem
|| to
== mem
1213 || (mem
->use_hierarchy
&& css_is_ancestor(&from
->css
, &mem
->css
))
1214 || (mem
->use_hierarchy
&& css_is_ancestor(&to
->css
, &mem
->css
)))
1217 spin_unlock(&mc
.lock
);
1221 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*mem
)
1223 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1224 if (mem_cgroup_under_move(mem
)) {
1226 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1227 /* moving charge context might have finished. */
1230 finish_wait(&mc
.waitq
, &wait
);
1238 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1239 * @memcg: The memory cgroup that went over limit
1240 * @p: Task that is going to be killed
1242 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1245 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1247 struct cgroup
*task_cgrp
;
1248 struct cgroup
*mem_cgrp
;
1250 * Need a buffer in BSS, can't rely on allocations. The code relies
1251 * on the assumption that OOM is serialized for memory controller.
1252 * If this assumption is broken, revisit this code.
1254 static char memcg_name
[PATH_MAX
];
1263 mem_cgrp
= memcg
->css
.cgroup
;
1264 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1266 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1269 * Unfortunately, we are unable to convert to a useful name
1270 * But we'll still print out the usage information
1277 printk(KERN_INFO
"Task in %s killed", memcg_name
);
1280 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1288 * Continues from above, so we don't need an KERN_ level
1290 printk(KERN_CONT
" as a result of limit of %s\n", memcg_name
);
1293 printk(KERN_INFO
"memory: usage %llukB, limit %llukB, failcnt %llu\n",
1294 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1295 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1296 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1297 printk(KERN_INFO
"memory+swap: usage %llukB, limit %llukB, "
1299 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1300 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1301 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1305 * This function returns the number of memcg under hierarchy tree. Returns
1306 * 1(self count) if no children.
1308 static int mem_cgroup_count_children(struct mem_cgroup
*mem
)
1311 struct mem_cgroup
*iter
;
1313 for_each_mem_cgroup_tree(iter
, mem
)
1319 * Return the memory (and swap, if configured) limit for a memcg.
1321 u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1326 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1327 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1329 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1331 * If memsw is finite and limits the amount of swap space available
1332 * to this memcg, return that limit.
1334 return min(limit
, memsw
);
1338 * Visit the first child (need not be the first child as per the ordering
1339 * of the cgroup list, since we track last_scanned_child) of @mem and use
1340 * that to reclaim free pages from.
1342 static struct mem_cgroup
*
1343 mem_cgroup_select_victim(struct mem_cgroup
*root_mem
)
1345 struct mem_cgroup
*ret
= NULL
;
1346 struct cgroup_subsys_state
*css
;
1349 if (!root_mem
->use_hierarchy
) {
1350 css_get(&root_mem
->css
);
1356 nextid
= root_mem
->last_scanned_child
+ 1;
1357 css
= css_get_next(&mem_cgroup_subsys
, nextid
, &root_mem
->css
,
1359 if (css
&& css_tryget(css
))
1360 ret
= container_of(css
, struct mem_cgroup
, css
);
1363 /* Updates scanning parameter */
1364 spin_lock(&root_mem
->reclaim_param_lock
);
1366 /* this means start scan from ID:1 */
1367 root_mem
->last_scanned_child
= 0;
1369 root_mem
->last_scanned_child
= found
;
1370 spin_unlock(&root_mem
->reclaim_param_lock
);
1377 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1378 * we reclaimed from, so that we don't end up penalizing one child extensively
1379 * based on its position in the children list.
1381 * root_mem is the original ancestor that we've been reclaim from.
1383 * We give up and return to the caller when we visit root_mem twice.
1384 * (other groups can be removed while we're walking....)
1386 * If shrink==true, for avoiding to free too much, this returns immedieately.
1388 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup
*root_mem
,
1391 unsigned long reclaim_options
)
1393 struct mem_cgroup
*victim
;
1396 bool noswap
= reclaim_options
& MEM_CGROUP_RECLAIM_NOSWAP
;
1397 bool shrink
= reclaim_options
& MEM_CGROUP_RECLAIM_SHRINK
;
1398 bool check_soft
= reclaim_options
& MEM_CGROUP_RECLAIM_SOFT
;
1399 unsigned long excess
;
1401 excess
= res_counter_soft_limit_excess(&root_mem
->res
) >> PAGE_SHIFT
;
1403 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1404 if (root_mem
->memsw_is_minimum
)
1408 victim
= mem_cgroup_select_victim(root_mem
);
1409 if (victim
== root_mem
) {
1412 drain_all_stock_async();
1415 * If we have not been able to reclaim
1416 * anything, it might because there are
1417 * no reclaimable pages under this hierarchy
1419 if (!check_soft
|| !total
) {
1420 css_put(&victim
->css
);
1424 * We want to do more targetted reclaim.
1425 * excess >> 2 is not to excessive so as to
1426 * reclaim too much, nor too less that we keep
1427 * coming back to reclaim from this cgroup
1429 if (total
>= (excess
>> 2) ||
1430 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
)) {
1431 css_put(&victim
->css
);
1436 if (!mem_cgroup_local_usage(victim
)) {
1437 /* this cgroup's local usage == 0 */
1438 css_put(&victim
->css
);
1441 /* we use swappiness of local cgroup */
1443 ret
= mem_cgroup_shrink_node_zone(victim
, gfp_mask
,
1444 noswap
, get_swappiness(victim
), zone
);
1446 ret
= try_to_free_mem_cgroup_pages(victim
, gfp_mask
,
1447 noswap
, get_swappiness(victim
));
1448 css_put(&victim
->css
);
1450 * At shrinking usage, we can't check we should stop here or
1451 * reclaim more. It's depends on callers. last_scanned_child
1452 * will work enough for keeping fairness under tree.
1458 if (!res_counter_soft_limit_excess(&root_mem
->res
))
1460 } else if (mem_cgroup_margin(root_mem
))
1467 * Check OOM-Killer is already running under our hierarchy.
1468 * If someone is running, return false.
1470 static bool mem_cgroup_oom_lock(struct mem_cgroup
*mem
)
1472 int x
, lock_count
= 0;
1473 struct mem_cgroup
*iter
;
1475 for_each_mem_cgroup_tree(iter
, mem
) {
1476 x
= atomic_inc_return(&iter
->oom_lock
);
1477 lock_count
= max(x
, lock_count
);
1480 if (lock_count
== 1)
1485 static int mem_cgroup_oom_unlock(struct mem_cgroup
*mem
)
1487 struct mem_cgroup
*iter
;
1490 * When a new child is created while the hierarchy is under oom,
1491 * mem_cgroup_oom_lock() may not be called. We have to use
1492 * atomic_add_unless() here.
1494 for_each_mem_cgroup_tree(iter
, mem
)
1495 atomic_add_unless(&iter
->oom_lock
, -1, 0);
1500 static DEFINE_MUTEX(memcg_oom_mutex
);
1501 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1503 struct oom_wait_info
{
1504 struct mem_cgroup
*mem
;
1508 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1509 unsigned mode
, int sync
, void *arg
)
1511 struct mem_cgroup
*wake_mem
= (struct mem_cgroup
*)arg
;
1512 struct oom_wait_info
*oom_wait_info
;
1514 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1516 if (oom_wait_info
->mem
== wake_mem
)
1518 /* if no hierarchy, no match */
1519 if (!oom_wait_info
->mem
->use_hierarchy
|| !wake_mem
->use_hierarchy
)
1522 * Both of oom_wait_info->mem and wake_mem are stable under us.
1523 * Then we can use css_is_ancestor without taking care of RCU.
1525 if (!css_is_ancestor(&oom_wait_info
->mem
->css
, &wake_mem
->css
) &&
1526 !css_is_ancestor(&wake_mem
->css
, &oom_wait_info
->mem
->css
))
1530 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1533 static void memcg_wakeup_oom(struct mem_cgroup
*mem
)
1535 /* for filtering, pass "mem" as argument. */
1536 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, mem
);
1539 static void memcg_oom_recover(struct mem_cgroup
*mem
)
1541 if (mem
&& atomic_read(&mem
->oom_lock
))
1542 memcg_wakeup_oom(mem
);
1546 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1548 bool mem_cgroup_handle_oom(struct mem_cgroup
*mem
, gfp_t mask
)
1550 struct oom_wait_info owait
;
1551 bool locked
, need_to_kill
;
1554 owait
.wait
.flags
= 0;
1555 owait
.wait
.func
= memcg_oom_wake_function
;
1556 owait
.wait
.private = current
;
1557 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1558 need_to_kill
= true;
1559 /* At first, try to OOM lock hierarchy under mem.*/
1560 mutex_lock(&memcg_oom_mutex
);
1561 locked
= mem_cgroup_oom_lock(mem
);
1563 * Even if signal_pending(), we can't quit charge() loop without
1564 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1565 * under OOM is always welcomed, use TASK_KILLABLE here.
1567 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1568 if (!locked
|| mem
->oom_kill_disable
)
1569 need_to_kill
= false;
1571 mem_cgroup_oom_notify(mem
);
1572 mutex_unlock(&memcg_oom_mutex
);
1575 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1576 mem_cgroup_out_of_memory(mem
, mask
);
1579 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1581 mutex_lock(&memcg_oom_mutex
);
1582 mem_cgroup_oom_unlock(mem
);
1583 memcg_wakeup_oom(mem
);
1584 mutex_unlock(&memcg_oom_mutex
);
1586 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
1588 /* Give chance to dying process */
1589 schedule_timeout(1);
1594 * Currently used to update mapped file statistics, but the routine can be
1595 * generalized to update other statistics as well.
1597 * Notes: Race condition
1599 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1600 * it tends to be costly. But considering some conditions, we doesn't need
1601 * to do so _always_.
1603 * Considering "charge", lock_page_cgroup() is not required because all
1604 * file-stat operations happen after a page is attached to radix-tree. There
1605 * are no race with "charge".
1607 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1608 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1609 * if there are race with "uncharge". Statistics itself is properly handled
1612 * Considering "move", this is an only case we see a race. To make the race
1613 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1614 * possibility of race condition. If there is, we take a lock.
1617 void mem_cgroup_update_page_stat(struct page
*page
,
1618 enum mem_cgroup_page_stat_item idx
, int val
)
1620 struct mem_cgroup
*mem
;
1621 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
1622 bool need_unlock
= false;
1623 unsigned long uninitialized_var(flags
);
1629 mem
= pc
->mem_cgroup
;
1630 if (unlikely(!mem
|| !PageCgroupUsed(pc
)))
1632 /* pc->mem_cgroup is unstable ? */
1633 if (unlikely(mem_cgroup_stealed(mem
)) || PageTransHuge(page
)) {
1634 /* take a lock against to access pc->mem_cgroup */
1635 move_lock_page_cgroup(pc
, &flags
);
1637 mem
= pc
->mem_cgroup
;
1638 if (!mem
|| !PageCgroupUsed(pc
))
1643 case MEMCG_NR_FILE_MAPPED
:
1645 SetPageCgroupFileMapped(pc
);
1646 else if (!page_mapped(page
))
1647 ClearPageCgroupFileMapped(pc
);
1648 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
1654 this_cpu_add(mem
->stat
->count
[idx
], val
);
1657 if (unlikely(need_unlock
))
1658 move_unlock_page_cgroup(pc
, &flags
);
1662 EXPORT_SYMBOL(mem_cgroup_update_page_stat
);
1665 * size of first charge trial. "32" comes from vmscan.c's magic value.
1666 * TODO: maybe necessary to use big numbers in big irons.
1668 #define CHARGE_SIZE (32 * PAGE_SIZE)
1669 struct memcg_stock_pcp
{
1670 struct mem_cgroup
*cached
; /* this never be root cgroup */
1672 struct work_struct work
;
1674 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1675 static atomic_t memcg_drain_count
;
1678 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1679 * from local stock and true is returned. If the stock is 0 or charges from a
1680 * cgroup which is not current target, returns false. This stock will be
1683 static bool consume_stock(struct mem_cgroup
*mem
)
1685 struct memcg_stock_pcp
*stock
;
1688 stock
= &get_cpu_var(memcg_stock
);
1689 if (mem
== stock
->cached
&& stock
->charge
)
1690 stock
->charge
-= PAGE_SIZE
;
1691 else /* need to call res_counter_charge */
1693 put_cpu_var(memcg_stock
);
1698 * Returns stocks cached in percpu to res_counter and reset cached information.
1700 static void drain_stock(struct memcg_stock_pcp
*stock
)
1702 struct mem_cgroup
*old
= stock
->cached
;
1704 if (stock
->charge
) {
1705 res_counter_uncharge(&old
->res
, stock
->charge
);
1706 if (do_swap_account
)
1707 res_counter_uncharge(&old
->memsw
, stock
->charge
);
1709 stock
->cached
= NULL
;
1714 * This must be called under preempt disabled or must be called by
1715 * a thread which is pinned to local cpu.
1717 static void drain_local_stock(struct work_struct
*dummy
)
1719 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
1724 * Cache charges(val) which is from res_counter, to local per_cpu area.
1725 * This will be consumed by consume_stock() function, later.
1727 static void refill_stock(struct mem_cgroup
*mem
, int val
)
1729 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1731 if (stock
->cached
!= mem
) { /* reset if necessary */
1733 stock
->cached
= mem
;
1735 stock
->charge
+= val
;
1736 put_cpu_var(memcg_stock
);
1740 * Tries to drain stocked charges in other cpus. This function is asynchronous
1741 * and just put a work per cpu for draining localy on each cpu. Caller can
1742 * expects some charges will be back to res_counter later but cannot wait for
1745 static void drain_all_stock_async(void)
1748 /* This function is for scheduling "drain" in asynchronous way.
1749 * The result of "drain" is not directly handled by callers. Then,
1750 * if someone is calling drain, we don't have to call drain more.
1751 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1752 * there is a race. We just do loose check here.
1754 if (atomic_read(&memcg_drain_count
))
1756 /* Notify other cpus that system-wide "drain" is running */
1757 atomic_inc(&memcg_drain_count
);
1759 for_each_online_cpu(cpu
) {
1760 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1761 schedule_work_on(cpu
, &stock
->work
);
1764 atomic_dec(&memcg_drain_count
);
1765 /* We don't wait for flush_work */
1768 /* This is a synchronous drain interface. */
1769 static void drain_all_stock_sync(void)
1771 /* called when force_empty is called */
1772 atomic_inc(&memcg_drain_count
);
1773 schedule_on_each_cpu(drain_local_stock
);
1774 atomic_dec(&memcg_drain_count
);
1778 * This function drains percpu counter value from DEAD cpu and
1779 * move it to local cpu. Note that this function can be preempted.
1781 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*mem
, int cpu
)
1785 spin_lock(&mem
->pcp_counter_lock
);
1786 for (i
= 0; i
< MEM_CGROUP_STAT_DATA
; i
++) {
1787 s64 x
= per_cpu(mem
->stat
->count
[i
], cpu
);
1789 per_cpu(mem
->stat
->count
[i
], cpu
) = 0;
1790 mem
->nocpu_base
.count
[i
] += x
;
1792 /* need to clear ON_MOVE value, works as a kind of lock. */
1793 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) = 0;
1794 spin_unlock(&mem
->pcp_counter_lock
);
1797 static void synchronize_mem_cgroup_on_move(struct mem_cgroup
*mem
, int cpu
)
1799 int idx
= MEM_CGROUP_ON_MOVE
;
1801 spin_lock(&mem
->pcp_counter_lock
);
1802 per_cpu(mem
->stat
->count
[idx
], cpu
) = mem
->nocpu_base
.count
[idx
];
1803 spin_unlock(&mem
->pcp_counter_lock
);
1806 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1807 unsigned long action
,
1810 int cpu
= (unsigned long)hcpu
;
1811 struct memcg_stock_pcp
*stock
;
1812 struct mem_cgroup
*iter
;
1814 if ((action
== CPU_ONLINE
)) {
1815 for_each_mem_cgroup_all(iter
)
1816 synchronize_mem_cgroup_on_move(iter
, cpu
);
1820 if ((action
!= CPU_DEAD
) || action
!= CPU_DEAD_FROZEN
)
1823 for_each_mem_cgroup_all(iter
)
1824 mem_cgroup_drain_pcp_counter(iter
, cpu
);
1826 stock
= &per_cpu(memcg_stock
, cpu
);
1832 /* See __mem_cgroup_try_charge() for details */
1834 CHARGE_OK
, /* success */
1835 CHARGE_RETRY
, /* need to retry but retry is not bad */
1836 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
1837 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
1838 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
1841 static int __mem_cgroup_do_charge(struct mem_cgroup
*mem
, gfp_t gfp_mask
,
1842 int csize
, bool oom_check
)
1844 struct mem_cgroup
*mem_over_limit
;
1845 struct res_counter
*fail_res
;
1846 unsigned long flags
= 0;
1849 ret
= res_counter_charge(&mem
->res
, csize
, &fail_res
);
1852 if (!do_swap_account
)
1854 ret
= res_counter_charge(&mem
->memsw
, csize
, &fail_res
);
1858 res_counter_uncharge(&mem
->res
, csize
);
1859 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
1860 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
1862 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
1864 * csize can be either a huge page (HPAGE_SIZE), a batch of
1865 * regular pages (CHARGE_SIZE), or a single regular page
1868 * Never reclaim on behalf of optional batching, retry with a
1869 * single page instead.
1871 if (csize
== CHARGE_SIZE
)
1872 return CHARGE_RETRY
;
1874 if (!(gfp_mask
& __GFP_WAIT
))
1875 return CHARGE_WOULDBLOCK
;
1877 ret
= mem_cgroup_hierarchical_reclaim(mem_over_limit
, NULL
,
1879 if (mem_cgroup_margin(mem_over_limit
) >= csize
)
1880 return CHARGE_RETRY
;
1882 * Even though the limit is exceeded at this point, reclaim
1883 * may have been able to free some pages. Retry the charge
1884 * before killing the task.
1886 * Only for regular pages, though: huge pages are rather
1887 * unlikely to succeed so close to the limit, and we fall back
1888 * to regular pages anyway in case of failure.
1890 if (csize
== PAGE_SIZE
&& ret
)
1891 return CHARGE_RETRY
;
1894 * At task move, charge accounts can be doubly counted. So, it's
1895 * better to wait until the end of task_move if something is going on.
1897 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1898 return CHARGE_RETRY
;
1900 /* If we don't need to call oom-killer at el, return immediately */
1902 return CHARGE_NOMEM
;
1904 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
))
1905 return CHARGE_OOM_DIE
;
1907 return CHARGE_RETRY
;
1911 * Unlike exported interface, "oom" parameter is added. if oom==true,
1912 * oom-killer can be invoked.
1914 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
1916 struct mem_cgroup
**memcg
, bool oom
,
1919 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1920 struct mem_cgroup
*mem
= NULL
;
1922 int csize
= max(CHARGE_SIZE
, (unsigned long) page_size
);
1925 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1926 * in system level. So, allow to go ahead dying process in addition to
1929 if (unlikely(test_thread_flag(TIF_MEMDIE
)
1930 || fatal_signal_pending(current
)))
1934 * We always charge the cgroup the mm_struct belongs to.
1935 * The mm_struct's mem_cgroup changes on task migration if the
1936 * thread group leader migrates. It's possible that mm is not
1937 * set, if so charge the init_mm (happens for pagecache usage).
1942 if (*memcg
) { /* css should be a valid one */
1944 VM_BUG_ON(css_is_removed(&mem
->css
));
1945 if (mem_cgroup_is_root(mem
))
1947 if (page_size
== PAGE_SIZE
&& consume_stock(mem
))
1951 struct task_struct
*p
;
1954 p
= rcu_dereference(mm
->owner
);
1956 * Because we don't have task_lock(), "p" can exit.
1957 * In that case, "mem" can point to root or p can be NULL with
1958 * race with swapoff. Then, we have small risk of mis-accouning.
1959 * But such kind of mis-account by race always happens because
1960 * we don't have cgroup_mutex(). It's overkill and we allo that
1962 * (*) swapoff at el will charge against mm-struct not against
1963 * task-struct. So, mm->owner can be NULL.
1965 mem
= mem_cgroup_from_task(p
);
1966 if (!mem
|| mem_cgroup_is_root(mem
)) {
1970 if (page_size
== PAGE_SIZE
&& consume_stock(mem
)) {
1972 * It seems dagerous to access memcg without css_get().
1973 * But considering how consume_stok works, it's not
1974 * necessary. If consume_stock success, some charges
1975 * from this memcg are cached on this cpu. So, we
1976 * don't need to call css_get()/css_tryget() before
1977 * calling consume_stock().
1982 /* after here, we may be blocked. we need to get refcnt */
1983 if (!css_tryget(&mem
->css
)) {
1993 /* If killed, bypass charge */
1994 if (fatal_signal_pending(current
)) {
2000 if (oom
&& !nr_oom_retries
) {
2002 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
2005 ret
= __mem_cgroup_do_charge(mem
, gfp_mask
, csize
, oom_check
);
2010 case CHARGE_RETRY
: /* not in OOM situation but retry */
2015 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2018 case CHARGE_NOMEM
: /* OOM routine works */
2023 /* If oom, we never return -ENOMEM */
2026 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2030 } while (ret
!= CHARGE_OK
);
2032 if (csize
> page_size
)
2033 refill_stock(mem
, csize
- page_size
);
2047 * Somemtimes we have to undo a charge we got by try_charge().
2048 * This function is for that and do uncharge, put css's refcnt.
2049 * gotten by try_charge().
2051 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2052 unsigned long count
)
2054 if (!mem_cgroup_is_root(mem
)) {
2055 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
2056 if (do_swap_account
)
2057 res_counter_uncharge(&mem
->memsw
, PAGE_SIZE
* count
);
2061 static void mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2064 __mem_cgroup_cancel_charge(mem
, page_size
>> PAGE_SHIFT
);
2068 * A helper function to get mem_cgroup from ID. must be called under
2069 * rcu_read_lock(). The caller must check css_is_removed() or some if
2070 * it's concern. (dropping refcnt from swap can be called against removed
2073 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2075 struct cgroup_subsys_state
*css
;
2077 /* ID 0 is unused ID */
2080 css
= css_lookup(&mem_cgroup_subsys
, id
);
2083 return container_of(css
, struct mem_cgroup
, css
);
2086 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2088 struct mem_cgroup
*mem
= NULL
;
2089 struct page_cgroup
*pc
;
2093 VM_BUG_ON(!PageLocked(page
));
2095 pc
= lookup_page_cgroup(page
);
2096 lock_page_cgroup(pc
);
2097 if (PageCgroupUsed(pc
)) {
2098 mem
= pc
->mem_cgroup
;
2099 if (mem
&& !css_tryget(&mem
->css
))
2101 } else if (PageSwapCache(page
)) {
2102 ent
.val
= page_private(page
);
2103 id
= lookup_swap_cgroup(ent
);
2105 mem
= mem_cgroup_lookup(id
);
2106 if (mem
&& !css_tryget(&mem
->css
))
2110 unlock_page_cgroup(pc
);
2114 static void __mem_cgroup_commit_charge(struct mem_cgroup
*mem
,
2115 struct page_cgroup
*pc
,
2116 enum charge_type ctype
,
2119 int nr_pages
= page_size
>> PAGE_SHIFT
;
2121 lock_page_cgroup(pc
);
2122 if (unlikely(PageCgroupUsed(pc
))) {
2123 unlock_page_cgroup(pc
);
2124 mem_cgroup_cancel_charge(mem
, page_size
);
2128 * we don't need page_cgroup_lock about tail pages, becase they are not
2129 * accessed by any other context at this point.
2131 pc
->mem_cgroup
= mem
;
2133 * We access a page_cgroup asynchronously without lock_page_cgroup().
2134 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2135 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2136 * before USED bit, we need memory barrier here.
2137 * See mem_cgroup_add_lru_list(), etc.
2141 case MEM_CGROUP_CHARGE_TYPE_CACHE
:
2142 case MEM_CGROUP_CHARGE_TYPE_SHMEM
:
2143 SetPageCgroupCache(pc
);
2144 SetPageCgroupUsed(pc
);
2146 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2147 ClearPageCgroupCache(pc
);
2148 SetPageCgroupUsed(pc
);
2154 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), nr_pages
);
2155 unlock_page_cgroup(pc
);
2157 * "charge_statistics" updated event counter. Then, check it.
2158 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2159 * if they exceeds softlimit.
2161 memcg_check_events(mem
, pc
->page
);
2164 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2166 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2167 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2169 * Because tail pages are not marked as "used", set it. We're under
2170 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2172 void mem_cgroup_split_huge_fixup(struct page
*head
, struct page
*tail
)
2174 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
2175 struct page_cgroup
*tail_pc
= lookup_page_cgroup(tail
);
2176 unsigned long flags
;
2178 if (mem_cgroup_disabled())
2181 * We have no races with charge/uncharge but will have races with
2182 * page state accounting.
2184 move_lock_page_cgroup(head_pc
, &flags
);
2186 tail_pc
->mem_cgroup
= head_pc
->mem_cgroup
;
2187 smp_wmb(); /* see __commit_charge() */
2188 if (PageCgroupAcctLRU(head_pc
)) {
2190 struct mem_cgroup_per_zone
*mz
;
2193 * LRU flags cannot be copied because we need to add tail
2194 *.page to LRU by generic call and our hook will be called.
2195 * We hold lru_lock, then, reduce counter directly.
2197 lru
= page_lru(head
);
2198 mz
= page_cgroup_zoneinfo(head_pc
);
2199 MEM_CGROUP_ZSTAT(mz
, lru
) -= 1;
2201 tail_pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
2202 move_unlock_page_cgroup(head_pc
, &flags
);
2207 * __mem_cgroup_move_account - move account of the page
2208 * @pc: page_cgroup of the page.
2209 * @from: mem_cgroup which the page is moved from.
2210 * @to: mem_cgroup which the page is moved to. @from != @to.
2211 * @uncharge: whether we should call uncharge and css_put against @from.
2213 * The caller must confirm following.
2214 * - page is not on LRU (isolate_page() is useful.)
2215 * - the pc is locked, used, and ->mem_cgroup points to @from.
2217 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2218 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2219 * true, this function does "uncharge" from old cgroup, but it doesn't if
2220 * @uncharge is false, so a caller should do "uncharge".
2223 static void __mem_cgroup_move_account(struct page_cgroup
*pc
,
2224 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool uncharge
,
2227 int nr_pages
= charge_size
>> PAGE_SHIFT
;
2229 VM_BUG_ON(from
== to
);
2230 VM_BUG_ON(PageLRU(pc
->page
));
2231 VM_BUG_ON(!page_is_cgroup_locked(pc
));
2232 VM_BUG_ON(!PageCgroupUsed(pc
));
2233 VM_BUG_ON(pc
->mem_cgroup
!= from
);
2235 if (PageCgroupFileMapped(pc
)) {
2236 /* Update mapped_file data for mem_cgroup */
2238 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
2239 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
2242 mem_cgroup_charge_statistics(from
, PageCgroupCache(pc
), -nr_pages
);
2244 /* This is not "cancel", but cancel_charge does all we need. */
2245 mem_cgroup_cancel_charge(from
, charge_size
);
2247 /* caller should have done css_get */
2248 pc
->mem_cgroup
= to
;
2249 mem_cgroup_charge_statistics(to
, PageCgroupCache(pc
), nr_pages
);
2251 * We charges against "to" which may not have any tasks. Then, "to"
2252 * can be under rmdir(). But in current implementation, caller of
2253 * this function is just force_empty() and move charge, so it's
2254 * garanteed that "to" is never removed. So, we don't check rmdir
2260 * check whether the @pc is valid for moving account and call
2261 * __mem_cgroup_move_account()
2263 static int mem_cgroup_move_account(struct page_cgroup
*pc
,
2264 struct mem_cgroup
*from
, struct mem_cgroup
*to
,
2265 bool uncharge
, int charge_size
)
2268 unsigned long flags
;
2270 * The page is isolated from LRU. So, collapse function
2271 * will not handle this page. But page splitting can happen.
2272 * Do this check under compound_page_lock(). The caller should
2275 if ((charge_size
> PAGE_SIZE
) && !PageTransHuge(pc
->page
))
2278 lock_page_cgroup(pc
);
2279 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== from
) {
2280 move_lock_page_cgroup(pc
, &flags
);
2281 __mem_cgroup_move_account(pc
, from
, to
, uncharge
, charge_size
);
2282 move_unlock_page_cgroup(pc
, &flags
);
2285 unlock_page_cgroup(pc
);
2289 memcg_check_events(to
, pc
->page
);
2290 memcg_check_events(from
, pc
->page
);
2295 * move charges to its parent.
2298 static int mem_cgroup_move_parent(struct page_cgroup
*pc
,
2299 struct mem_cgroup
*child
,
2302 struct page
*page
= pc
->page
;
2303 struct cgroup
*cg
= child
->css
.cgroup
;
2304 struct cgroup
*pcg
= cg
->parent
;
2305 struct mem_cgroup
*parent
;
2306 int page_size
= PAGE_SIZE
;
2307 unsigned long flags
;
2315 if (!get_page_unless_zero(page
))
2317 if (isolate_lru_page(page
))
2320 if (PageTransHuge(page
))
2321 page_size
= HPAGE_SIZE
;
2323 parent
= mem_cgroup_from_cont(pcg
);
2324 ret
= __mem_cgroup_try_charge(NULL
, gfp_mask
,
2325 &parent
, false, page_size
);
2329 if (page_size
> PAGE_SIZE
)
2330 flags
= compound_lock_irqsave(page
);
2332 ret
= mem_cgroup_move_account(pc
, child
, parent
, true, page_size
);
2334 mem_cgroup_cancel_charge(parent
, page_size
);
2336 if (page_size
> PAGE_SIZE
)
2337 compound_unlock_irqrestore(page
, flags
);
2339 putback_lru_page(page
);
2347 * Charge the memory controller for page usage.
2349 * 0 if the charge was successful
2350 * < 0 if the cgroup is over its limit
2352 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
2353 gfp_t gfp_mask
, enum charge_type ctype
)
2355 struct mem_cgroup
*mem
= NULL
;
2356 int page_size
= PAGE_SIZE
;
2357 struct page_cgroup
*pc
;
2361 if (PageTransHuge(page
)) {
2362 page_size
<<= compound_order(page
);
2363 VM_BUG_ON(!PageTransHuge(page
));
2365 * Never OOM-kill a process for a huge page. The
2366 * fault handler will fall back to regular pages.
2371 pc
= lookup_page_cgroup(page
);
2372 BUG_ON(!pc
); /* XXX: remove this and move pc lookup into commit */
2374 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, &mem
, oom
, page_size
);
2378 __mem_cgroup_commit_charge(mem
, pc
, ctype
, page_size
);
2382 int mem_cgroup_newpage_charge(struct page
*page
,
2383 struct mm_struct
*mm
, gfp_t gfp_mask
)
2385 if (mem_cgroup_disabled())
2388 * If already mapped, we don't have to account.
2389 * If page cache, page->mapping has address_space.
2390 * But page->mapping may have out-of-use anon_vma pointer,
2391 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2394 if (page_mapped(page
) || (page
->mapping
&& !PageAnon(page
)))
2398 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2399 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2403 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2404 enum charge_type ctype
);
2406 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
2411 if (mem_cgroup_disabled())
2413 if (PageCompound(page
))
2416 * Corner case handling. This is called from add_to_page_cache()
2417 * in usual. But some FS (shmem) precharges this page before calling it
2418 * and call add_to_page_cache() with GFP_NOWAIT.
2420 * For GFP_NOWAIT case, the page may be pre-charged before calling
2421 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2422 * charge twice. (It works but has to pay a bit larger cost.)
2423 * And when the page is SwapCache, it should take swap information
2424 * into account. This is under lock_page() now.
2426 if (!(gfp_mask
& __GFP_WAIT
)) {
2427 struct page_cgroup
*pc
;
2429 pc
= lookup_page_cgroup(page
);
2432 lock_page_cgroup(pc
);
2433 if (PageCgroupUsed(pc
)) {
2434 unlock_page_cgroup(pc
);
2437 unlock_page_cgroup(pc
);
2443 if (page_is_file_cache(page
))
2444 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2445 MEM_CGROUP_CHARGE_TYPE_CACHE
);
2448 if (PageSwapCache(page
)) {
2449 struct mem_cgroup
*mem
;
2451 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
2453 __mem_cgroup_commit_charge_swapin(page
, mem
,
2454 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2456 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2457 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2463 * While swap-in, try_charge -> commit or cancel, the page is locked.
2464 * And when try_charge() successfully returns, one refcnt to memcg without
2465 * struct page_cgroup is acquired. This refcnt will be consumed by
2466 * "commit()" or removed by "cancel()"
2468 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
2470 gfp_t mask
, struct mem_cgroup
**ptr
)
2472 struct mem_cgroup
*mem
;
2477 if (mem_cgroup_disabled())
2480 if (!do_swap_account
)
2483 * A racing thread's fault, or swapoff, may have already updated
2484 * the pte, and even removed page from swap cache: in those cases
2485 * do_swap_page()'s pte_same() test will fail; but there's also a
2486 * KSM case which does need to charge the page.
2488 if (!PageSwapCache(page
))
2490 mem
= try_get_mem_cgroup_from_page(page
);
2494 ret
= __mem_cgroup_try_charge(NULL
, mask
, ptr
, true, PAGE_SIZE
);
2500 return __mem_cgroup_try_charge(mm
, mask
, ptr
, true, PAGE_SIZE
);
2504 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2505 enum charge_type ctype
)
2507 struct page_cgroup
*pc
;
2509 if (mem_cgroup_disabled())
2513 cgroup_exclude_rmdir(&ptr
->css
);
2514 pc
= lookup_page_cgroup(page
);
2515 mem_cgroup_lru_del_before_commit_swapcache(page
);
2516 __mem_cgroup_commit_charge(ptr
, pc
, ctype
, PAGE_SIZE
);
2517 mem_cgroup_lru_add_after_commit_swapcache(page
);
2519 * Now swap is on-memory. This means this page may be
2520 * counted both as mem and swap....double count.
2521 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2522 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2523 * may call delete_from_swap_cache() before reach here.
2525 if (do_swap_account
&& PageSwapCache(page
)) {
2526 swp_entry_t ent
= {.val
= page_private(page
)};
2528 struct mem_cgroup
*memcg
;
2530 id
= swap_cgroup_record(ent
, 0);
2532 memcg
= mem_cgroup_lookup(id
);
2535 * This recorded memcg can be obsolete one. So, avoid
2536 * calling css_tryget
2538 if (!mem_cgroup_is_root(memcg
))
2539 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2540 mem_cgroup_swap_statistics(memcg
, false);
2541 mem_cgroup_put(memcg
);
2546 * At swapin, we may charge account against cgroup which has no tasks.
2547 * So, rmdir()->pre_destroy() can be called while we do this charge.
2548 * In that case, we need to call pre_destroy() again. check it here.
2550 cgroup_release_and_wakeup_rmdir(&ptr
->css
);
2553 void mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
)
2555 __mem_cgroup_commit_charge_swapin(page
, ptr
,
2556 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2559 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*mem
)
2561 if (mem_cgroup_disabled())
2565 mem_cgroup_cancel_charge(mem
, PAGE_SIZE
);
2569 __do_uncharge(struct mem_cgroup
*mem
, const enum charge_type ctype
,
2572 struct memcg_batch_info
*batch
= NULL
;
2573 bool uncharge_memsw
= true;
2574 /* If swapout, usage of swap doesn't decrease */
2575 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
2576 uncharge_memsw
= false;
2578 batch
= ¤t
->memcg_batch
;
2580 * In usual, we do css_get() when we remember memcg pointer.
2581 * But in this case, we keep res->usage until end of a series of
2582 * uncharges. Then, it's ok to ignore memcg's refcnt.
2587 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2588 * In those cases, all pages freed continously can be expected to be in
2589 * the same cgroup and we have chance to coalesce uncharges.
2590 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2591 * because we want to do uncharge as soon as possible.
2594 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
2595 goto direct_uncharge
;
2597 if (page_size
!= PAGE_SIZE
)
2598 goto direct_uncharge
;
2601 * In typical case, batch->memcg == mem. This means we can
2602 * merge a series of uncharges to an uncharge of res_counter.
2603 * If not, we uncharge res_counter ony by one.
2605 if (batch
->memcg
!= mem
)
2606 goto direct_uncharge
;
2607 /* remember freed charge and uncharge it later */
2608 batch
->bytes
+= PAGE_SIZE
;
2610 batch
->memsw_bytes
+= PAGE_SIZE
;
2613 res_counter_uncharge(&mem
->res
, page_size
);
2615 res_counter_uncharge(&mem
->memsw
, page_size
);
2616 if (unlikely(batch
->memcg
!= mem
))
2617 memcg_oom_recover(mem
);
2622 * uncharge if !page_mapped(page)
2624 static struct mem_cgroup
*
2625 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
)
2628 struct page_cgroup
*pc
;
2629 struct mem_cgroup
*mem
= NULL
;
2630 int page_size
= PAGE_SIZE
;
2632 if (mem_cgroup_disabled())
2635 if (PageSwapCache(page
))
2638 if (PageTransHuge(page
)) {
2639 page_size
<<= compound_order(page
);
2640 VM_BUG_ON(!PageTransHuge(page
));
2643 count
= page_size
>> PAGE_SHIFT
;
2645 * Check if our page_cgroup is valid
2647 pc
= lookup_page_cgroup(page
);
2648 if (unlikely(!pc
|| !PageCgroupUsed(pc
)))
2651 lock_page_cgroup(pc
);
2653 mem
= pc
->mem_cgroup
;
2655 if (!PageCgroupUsed(pc
))
2659 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2660 case MEM_CGROUP_CHARGE_TYPE_DROP
:
2661 /* See mem_cgroup_prepare_migration() */
2662 if (page_mapped(page
) || PageCgroupMigration(pc
))
2665 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
2666 if (!PageAnon(page
)) { /* Shared memory */
2667 if (page
->mapping
&& !page_is_file_cache(page
))
2669 } else if (page_mapped(page
)) /* Anon */
2676 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), -count
);
2678 ClearPageCgroupUsed(pc
);
2680 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2681 * freed from LRU. This is safe because uncharged page is expected not
2682 * to be reused (freed soon). Exception is SwapCache, it's handled by
2683 * special functions.
2686 unlock_page_cgroup(pc
);
2688 * even after unlock, we have mem->res.usage here and this memcg
2689 * will never be freed.
2691 memcg_check_events(mem
, page
);
2692 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
2693 mem_cgroup_swap_statistics(mem
, true);
2694 mem_cgroup_get(mem
);
2696 if (!mem_cgroup_is_root(mem
))
2697 __do_uncharge(mem
, ctype
, page_size
);
2702 unlock_page_cgroup(pc
);
2706 void mem_cgroup_uncharge_page(struct page
*page
)
2709 if (page_mapped(page
))
2711 if (page
->mapping
&& !PageAnon(page
))
2713 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2716 void mem_cgroup_uncharge_cache_page(struct page
*page
)
2718 VM_BUG_ON(page_mapped(page
));
2719 VM_BUG_ON(page
->mapping
);
2720 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
);
2724 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2725 * In that cases, pages are freed continuously and we can expect pages
2726 * are in the same memcg. All these calls itself limits the number of
2727 * pages freed at once, then uncharge_start/end() is called properly.
2728 * This may be called prural(2) times in a context,
2731 void mem_cgroup_uncharge_start(void)
2733 current
->memcg_batch
.do_batch
++;
2734 /* We can do nest. */
2735 if (current
->memcg_batch
.do_batch
== 1) {
2736 current
->memcg_batch
.memcg
= NULL
;
2737 current
->memcg_batch
.bytes
= 0;
2738 current
->memcg_batch
.memsw_bytes
= 0;
2742 void mem_cgroup_uncharge_end(void)
2744 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
2746 if (!batch
->do_batch
)
2750 if (batch
->do_batch
) /* If stacked, do nothing. */
2756 * This "batch->memcg" is valid without any css_get/put etc...
2757 * bacause we hide charges behind us.
2760 res_counter_uncharge(&batch
->memcg
->res
, batch
->bytes
);
2761 if (batch
->memsw_bytes
)
2762 res_counter_uncharge(&batch
->memcg
->memsw
, batch
->memsw_bytes
);
2763 memcg_oom_recover(batch
->memcg
);
2764 /* forget this pointer (for sanity check) */
2765 batch
->memcg
= NULL
;
2770 * called after __delete_from_swap_cache() and drop "page" account.
2771 * memcg information is recorded to swap_cgroup of "ent"
2774 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
2776 struct mem_cgroup
*memcg
;
2777 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
2779 if (!swapout
) /* this was a swap cache but the swap is unused ! */
2780 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
2782 memcg
= __mem_cgroup_uncharge_common(page
, ctype
);
2785 * record memcg information, if swapout && memcg != NULL,
2786 * mem_cgroup_get() was called in uncharge().
2788 if (do_swap_account
&& swapout
&& memcg
)
2789 swap_cgroup_record(ent
, css_id(&memcg
->css
));
2793 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2795 * called from swap_entry_free(). remove record in swap_cgroup and
2796 * uncharge "memsw" account.
2798 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
2800 struct mem_cgroup
*memcg
;
2803 if (!do_swap_account
)
2806 id
= swap_cgroup_record(ent
, 0);
2808 memcg
= mem_cgroup_lookup(id
);
2811 * We uncharge this because swap is freed.
2812 * This memcg can be obsolete one. We avoid calling css_tryget
2814 if (!mem_cgroup_is_root(memcg
))
2815 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2816 mem_cgroup_swap_statistics(memcg
, false);
2817 mem_cgroup_put(memcg
);
2823 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2824 * @entry: swap entry to be moved
2825 * @from: mem_cgroup which the entry is moved from
2826 * @to: mem_cgroup which the entry is moved to
2827 * @need_fixup: whether we should fixup res_counters and refcounts.
2829 * It succeeds only when the swap_cgroup's record for this entry is the same
2830 * as the mem_cgroup's id of @from.
2832 * Returns 0 on success, -EINVAL on failure.
2834 * The caller must have charged to @to, IOW, called res_counter_charge() about
2835 * both res and memsw, and called css_get().
2837 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2838 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2840 unsigned short old_id
, new_id
;
2842 old_id
= css_id(&from
->css
);
2843 new_id
= css_id(&to
->css
);
2845 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2846 mem_cgroup_swap_statistics(from
, false);
2847 mem_cgroup_swap_statistics(to
, true);
2849 * This function is only called from task migration context now.
2850 * It postpones res_counter and refcount handling till the end
2851 * of task migration(mem_cgroup_clear_mc()) for performance
2852 * improvement. But we cannot postpone mem_cgroup_get(to)
2853 * because if the process that has been moved to @to does
2854 * swap-in, the refcount of @to might be decreased to 0.
2858 if (!mem_cgroup_is_root(from
))
2859 res_counter_uncharge(&from
->memsw
, PAGE_SIZE
);
2860 mem_cgroup_put(from
);
2862 * we charged both to->res and to->memsw, so we should
2865 if (!mem_cgroup_is_root(to
))
2866 res_counter_uncharge(&to
->res
, PAGE_SIZE
);
2873 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2874 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2881 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2884 int mem_cgroup_prepare_migration(struct page
*page
,
2885 struct page
*newpage
, struct mem_cgroup
**ptr
, gfp_t gfp_mask
)
2887 struct page_cgroup
*pc
;
2888 struct mem_cgroup
*mem
= NULL
;
2889 enum charge_type ctype
;
2894 VM_BUG_ON(PageTransHuge(page
));
2895 if (mem_cgroup_disabled())
2898 pc
= lookup_page_cgroup(page
);
2899 lock_page_cgroup(pc
);
2900 if (PageCgroupUsed(pc
)) {
2901 mem
= pc
->mem_cgroup
;
2904 * At migrating an anonymous page, its mapcount goes down
2905 * to 0 and uncharge() will be called. But, even if it's fully
2906 * unmapped, migration may fail and this page has to be
2907 * charged again. We set MIGRATION flag here and delay uncharge
2908 * until end_migration() is called
2910 * Corner Case Thinking
2912 * When the old page was mapped as Anon and it's unmap-and-freed
2913 * while migration was ongoing.
2914 * If unmap finds the old page, uncharge() of it will be delayed
2915 * until end_migration(). If unmap finds a new page, it's
2916 * uncharged when it make mapcount to be 1->0. If unmap code
2917 * finds swap_migration_entry, the new page will not be mapped
2918 * and end_migration() will find it(mapcount==0).
2921 * When the old page was mapped but migraion fails, the kernel
2922 * remaps it. A charge for it is kept by MIGRATION flag even
2923 * if mapcount goes down to 0. We can do remap successfully
2924 * without charging it again.
2927 * The "old" page is under lock_page() until the end of
2928 * migration, so, the old page itself will not be swapped-out.
2929 * If the new page is swapped out before end_migraton, our
2930 * hook to usual swap-out path will catch the event.
2933 SetPageCgroupMigration(pc
);
2935 unlock_page_cgroup(pc
);
2937 * If the page is not charged at this point,
2944 ret
= __mem_cgroup_try_charge(NULL
, gfp_mask
, ptr
, false, PAGE_SIZE
);
2945 css_put(&mem
->css
);/* drop extra refcnt */
2946 if (ret
|| *ptr
== NULL
) {
2947 if (PageAnon(page
)) {
2948 lock_page_cgroup(pc
);
2949 ClearPageCgroupMigration(pc
);
2950 unlock_page_cgroup(pc
);
2952 * The old page may be fully unmapped while we kept it.
2954 mem_cgroup_uncharge_page(page
);
2959 * We charge new page before it's used/mapped. So, even if unlock_page()
2960 * is called before end_migration, we can catch all events on this new
2961 * page. In the case new page is migrated but not remapped, new page's
2962 * mapcount will be finally 0 and we call uncharge in end_migration().
2964 pc
= lookup_page_cgroup(newpage
);
2966 ctype
= MEM_CGROUP_CHARGE_TYPE_MAPPED
;
2967 else if (page_is_file_cache(page
))
2968 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
2970 ctype
= MEM_CGROUP_CHARGE_TYPE_SHMEM
;
2971 __mem_cgroup_commit_charge(mem
, pc
, ctype
, PAGE_SIZE
);
2975 /* remove redundant charge if migration failed*/
2976 void mem_cgroup_end_migration(struct mem_cgroup
*mem
,
2977 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
2979 struct page
*used
, *unused
;
2980 struct page_cgroup
*pc
;
2984 /* blocks rmdir() */
2985 cgroup_exclude_rmdir(&mem
->css
);
2986 if (!migration_ok
) {
2994 * We disallowed uncharge of pages under migration because mapcount
2995 * of the page goes down to zero, temporarly.
2996 * Clear the flag and check the page should be charged.
2998 pc
= lookup_page_cgroup(oldpage
);
2999 lock_page_cgroup(pc
);
3000 ClearPageCgroupMigration(pc
);
3001 unlock_page_cgroup(pc
);
3003 __mem_cgroup_uncharge_common(unused
, MEM_CGROUP_CHARGE_TYPE_FORCE
);
3006 * If a page is a file cache, radix-tree replacement is very atomic
3007 * and we can skip this check. When it was an Anon page, its mapcount
3008 * goes down to 0. But because we added MIGRATION flage, it's not
3009 * uncharged yet. There are several case but page->mapcount check
3010 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3011 * check. (see prepare_charge() also)
3014 mem_cgroup_uncharge_page(used
);
3016 * At migration, we may charge account against cgroup which has no
3018 * So, rmdir()->pre_destroy() can be called while we do this charge.
3019 * In that case, we need to call pre_destroy() again. check it here.
3021 cgroup_release_and_wakeup_rmdir(&mem
->css
);
3025 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3026 * Calling hierarchical_reclaim is not enough because we should update
3027 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3028 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3029 * not from the memcg which this page would be charged to.
3030 * try_charge_swapin does all of these works properly.
3032 int mem_cgroup_shmem_charge_fallback(struct page
*page
,
3033 struct mm_struct
*mm
,
3036 struct mem_cgroup
*mem
;
3039 if (mem_cgroup_disabled())
3042 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
3044 mem_cgroup_cancel_charge_swapin(mem
); /* it does !mem check */
3049 static DEFINE_MUTEX(set_limit_mutex
);
3051 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3052 unsigned long long val
)
3055 u64 memswlimit
, memlimit
;
3057 int children
= mem_cgroup_count_children(memcg
);
3058 u64 curusage
, oldusage
;
3062 * For keeping hierarchical_reclaim simple, how long we should retry
3063 * is depends on callers. We set our retry-count to be function
3064 * of # of children which we should visit in this loop.
3066 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
3068 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3071 while (retry_count
) {
3072 if (signal_pending(current
)) {
3077 * Rather than hide all in some function, I do this in
3078 * open coded manner. You see what this really does.
3079 * We have to guarantee mem->res.limit < mem->memsw.limit.
3081 mutex_lock(&set_limit_mutex
);
3082 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3083 if (memswlimit
< val
) {
3085 mutex_unlock(&set_limit_mutex
);
3089 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3093 ret
= res_counter_set_limit(&memcg
->res
, val
);
3095 if (memswlimit
== val
)
3096 memcg
->memsw_is_minimum
= true;
3098 memcg
->memsw_is_minimum
= false;
3100 mutex_unlock(&set_limit_mutex
);
3105 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3106 MEM_CGROUP_RECLAIM_SHRINK
);
3107 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3108 /* Usage is reduced ? */
3109 if (curusage
>= oldusage
)
3112 oldusage
= curusage
;
3114 if (!ret
&& enlarge
)
3115 memcg_oom_recover(memcg
);
3120 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3121 unsigned long long val
)
3124 u64 memlimit
, memswlimit
, oldusage
, curusage
;
3125 int children
= mem_cgroup_count_children(memcg
);
3129 /* see mem_cgroup_resize_res_limit */
3130 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
3131 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3132 while (retry_count
) {
3133 if (signal_pending(current
)) {
3138 * Rather than hide all in some function, I do this in
3139 * open coded manner. You see what this really does.
3140 * We have to guarantee mem->res.limit < mem->memsw.limit.
3142 mutex_lock(&set_limit_mutex
);
3143 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3144 if (memlimit
> val
) {
3146 mutex_unlock(&set_limit_mutex
);
3149 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3150 if (memswlimit
< val
)
3152 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
3154 if (memlimit
== val
)
3155 memcg
->memsw_is_minimum
= true;
3157 memcg
->memsw_is_minimum
= false;
3159 mutex_unlock(&set_limit_mutex
);
3164 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3165 MEM_CGROUP_RECLAIM_NOSWAP
|
3166 MEM_CGROUP_RECLAIM_SHRINK
);
3167 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3168 /* Usage is reduced ? */
3169 if (curusage
>= oldusage
)
3172 oldusage
= curusage
;
3174 if (!ret
&& enlarge
)
3175 memcg_oom_recover(memcg
);
3179 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3182 unsigned long nr_reclaimed
= 0;
3183 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3184 unsigned long reclaimed
;
3186 struct mem_cgroup_tree_per_zone
*mctz
;
3187 unsigned long long excess
;
3192 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3194 * This loop can run a while, specially if mem_cgroup's continuously
3195 * keep exceeding their soft limit and putting the system under
3202 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3206 reclaimed
= mem_cgroup_hierarchical_reclaim(mz
->mem
, zone
,
3208 MEM_CGROUP_RECLAIM_SOFT
);
3209 nr_reclaimed
+= reclaimed
;
3210 spin_lock(&mctz
->lock
);
3213 * If we failed to reclaim anything from this memory cgroup
3214 * it is time to move on to the next cgroup
3220 * Loop until we find yet another one.
3222 * By the time we get the soft_limit lock
3223 * again, someone might have aded the
3224 * group back on the RB tree. Iterate to
3225 * make sure we get a different mem.
3226 * mem_cgroup_largest_soft_limit_node returns
3227 * NULL if no other cgroup is present on
3231 __mem_cgroup_largest_soft_limit_node(mctz
);
3232 if (next_mz
== mz
) {
3233 css_put(&next_mz
->mem
->css
);
3235 } else /* next_mz == NULL or other memcg */
3239 __mem_cgroup_remove_exceeded(mz
->mem
, mz
, mctz
);
3240 excess
= res_counter_soft_limit_excess(&mz
->mem
->res
);
3242 * One school of thought says that we should not add
3243 * back the node to the tree if reclaim returns 0.
3244 * But our reclaim could return 0, simply because due
3245 * to priority we are exposing a smaller subset of
3246 * memory to reclaim from. Consider this as a longer
3249 /* If excess == 0, no tree ops */
3250 __mem_cgroup_insert_exceeded(mz
->mem
, mz
, mctz
, excess
);
3251 spin_unlock(&mctz
->lock
);
3252 css_put(&mz
->mem
->css
);
3255 * Could not reclaim anything and there are no more
3256 * mem cgroups to try or we seem to be looping without
3257 * reclaiming anything.
3259 if (!nr_reclaimed
&&
3261 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3263 } while (!nr_reclaimed
);
3265 css_put(&next_mz
->mem
->css
);
3266 return nr_reclaimed
;
3270 * This routine traverse page_cgroup in given list and drop them all.
3271 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3273 static int mem_cgroup_force_empty_list(struct mem_cgroup
*mem
,
3274 int node
, int zid
, enum lru_list lru
)
3277 struct mem_cgroup_per_zone
*mz
;
3278 struct page_cgroup
*pc
, *busy
;
3279 unsigned long flags
, loop
;
3280 struct list_head
*list
;
3283 zone
= &NODE_DATA(node
)->node_zones
[zid
];
3284 mz
= mem_cgroup_zoneinfo(mem
, node
, zid
);
3285 list
= &mz
->lists
[lru
];
3287 loop
= MEM_CGROUP_ZSTAT(mz
, lru
);
3288 /* give some margin against EBUSY etc...*/
3293 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3294 if (list_empty(list
)) {
3295 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3298 pc
= list_entry(list
->prev
, struct page_cgroup
, lru
);
3300 list_move(&pc
->lru
, list
);
3302 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3305 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3307 ret
= mem_cgroup_move_parent(pc
, mem
, GFP_KERNEL
);
3311 if (ret
== -EBUSY
|| ret
== -EINVAL
) {
3312 /* found lock contention or "pc" is obsolete. */
3319 if (!ret
&& !list_empty(list
))
3325 * make mem_cgroup's charge to be 0 if there is no task.
3326 * This enables deleting this mem_cgroup.
3328 static int mem_cgroup_force_empty(struct mem_cgroup
*mem
, bool free_all
)
3331 int node
, zid
, shrink
;
3332 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3333 struct cgroup
*cgrp
= mem
->css
.cgroup
;
3338 /* should free all ? */
3344 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
3347 if (signal_pending(current
))
3349 /* This is for making all *used* pages to be on LRU. */
3350 lru_add_drain_all();
3351 drain_all_stock_sync();
3353 mem_cgroup_start_move(mem
);
3354 for_each_node_state(node
, N_HIGH_MEMORY
) {
3355 for (zid
= 0; !ret
&& zid
< MAX_NR_ZONES
; zid
++) {
3358 ret
= mem_cgroup_force_empty_list(mem
,
3367 mem_cgroup_end_move(mem
);
3368 memcg_oom_recover(mem
);
3369 /* it seems parent cgroup doesn't have enough mem */
3373 /* "ret" should also be checked to ensure all lists are empty. */
3374 } while (mem
->res
.usage
> 0 || ret
);
3380 /* returns EBUSY if there is a task or if we come here twice. */
3381 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
) || shrink
) {
3385 /* we call try-to-free pages for make this cgroup empty */
3386 lru_add_drain_all();
3387 /* try to free all pages in this cgroup */
3389 while (nr_retries
&& mem
->res
.usage
> 0) {
3392 if (signal_pending(current
)) {
3396 progress
= try_to_free_mem_cgroup_pages(mem
, GFP_KERNEL
,
3397 false, get_swappiness(mem
));
3400 /* maybe some writeback is necessary */
3401 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3406 /* try move_account...there may be some *locked* pages. */
3410 int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
3412 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont
), true);
3416 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
3418 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
3421 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
3425 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3426 struct cgroup
*parent
= cont
->parent
;
3427 struct mem_cgroup
*parent_mem
= NULL
;
3430 parent_mem
= mem_cgroup_from_cont(parent
);
3434 * If parent's use_hierarchy is set, we can't make any modifications
3435 * in the child subtrees. If it is unset, then the change can
3436 * occur, provided the current cgroup has no children.
3438 * For the root cgroup, parent_mem is NULL, we allow value to be
3439 * set if there are no children.
3441 if ((!parent_mem
|| !parent_mem
->use_hierarchy
) &&
3442 (val
== 1 || val
== 0)) {
3443 if (list_empty(&cont
->children
))
3444 mem
->use_hierarchy
= val
;
3455 static u64
mem_cgroup_get_recursive_idx_stat(struct mem_cgroup
*mem
,
3456 enum mem_cgroup_stat_index idx
)
3458 struct mem_cgroup
*iter
;
3461 /* each per cpu's value can be minus.Then, use s64 */
3462 for_each_mem_cgroup_tree(iter
, mem
)
3463 val
+= mem_cgroup_read_stat(iter
, idx
);
3465 if (val
< 0) /* race ? */
3470 static inline u64
mem_cgroup_usage(struct mem_cgroup
*mem
, bool swap
)
3474 if (!mem_cgroup_is_root(mem
)) {
3476 return res_counter_read_u64(&mem
->res
, RES_USAGE
);
3478 return res_counter_read_u64(&mem
->memsw
, RES_USAGE
);
3481 val
= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3482 val
+= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_RSS
);
3485 val
+= mem_cgroup_get_recursive_idx_stat(mem
,
3486 MEM_CGROUP_STAT_SWAPOUT
);
3488 return val
<< PAGE_SHIFT
;
3491 static u64
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
)
3493 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3497 type
= MEMFILE_TYPE(cft
->private);
3498 name
= MEMFILE_ATTR(cft
->private);
3501 if (name
== RES_USAGE
)
3502 val
= mem_cgroup_usage(mem
, false);
3504 val
= res_counter_read_u64(&mem
->res
, name
);
3507 if (name
== RES_USAGE
)
3508 val
= mem_cgroup_usage(mem
, true);
3510 val
= res_counter_read_u64(&mem
->memsw
, name
);
3519 * The user of this function is...
3522 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
3525 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
3527 unsigned long long val
;
3530 type
= MEMFILE_TYPE(cft
->private);
3531 name
= MEMFILE_ATTR(cft
->private);
3534 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3538 /* This function does all necessary parse...reuse it */
3539 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3543 ret
= mem_cgroup_resize_limit(memcg
, val
);
3545 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
3547 case RES_SOFT_LIMIT
:
3548 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3552 * For memsw, soft limits are hard to implement in terms
3553 * of semantics, for now, we support soft limits for
3554 * control without swap
3557 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
3562 ret
= -EINVAL
; /* should be BUG() ? */
3568 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
3569 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
3571 struct cgroup
*cgroup
;
3572 unsigned long long min_limit
, min_memsw_limit
, tmp
;
3574 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3575 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3576 cgroup
= memcg
->css
.cgroup
;
3577 if (!memcg
->use_hierarchy
)
3580 while (cgroup
->parent
) {
3581 cgroup
= cgroup
->parent
;
3582 memcg
= mem_cgroup_from_cont(cgroup
);
3583 if (!memcg
->use_hierarchy
)
3585 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3586 min_limit
= min(min_limit
, tmp
);
3587 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3588 min_memsw_limit
= min(min_memsw_limit
, tmp
);
3591 *mem_limit
= min_limit
;
3592 *memsw_limit
= min_memsw_limit
;
3596 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
3598 struct mem_cgroup
*mem
;
3601 mem
= mem_cgroup_from_cont(cont
);
3602 type
= MEMFILE_TYPE(event
);
3603 name
= MEMFILE_ATTR(event
);
3607 res_counter_reset_max(&mem
->res
);
3609 res_counter_reset_max(&mem
->memsw
);
3613 res_counter_reset_failcnt(&mem
->res
);
3615 res_counter_reset_failcnt(&mem
->memsw
);
3622 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
3625 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
3629 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3630 struct cftype
*cft
, u64 val
)
3632 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
3634 if (val
>= (1 << NR_MOVE_TYPE
))
3637 * We check this value several times in both in can_attach() and
3638 * attach(), so we need cgroup lock to prevent this value from being
3642 mem
->move_charge_at_immigrate
= val
;
3648 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3649 struct cftype
*cft
, u64 val
)
3656 /* For read statistics */
3672 struct mcs_total_stat
{
3673 s64 stat
[NR_MCS_STAT
];
3679 } memcg_stat_strings
[NR_MCS_STAT
] = {
3680 {"cache", "total_cache"},
3681 {"rss", "total_rss"},
3682 {"mapped_file", "total_mapped_file"},
3683 {"pgpgin", "total_pgpgin"},
3684 {"pgpgout", "total_pgpgout"},
3685 {"swap", "total_swap"},
3686 {"inactive_anon", "total_inactive_anon"},
3687 {"active_anon", "total_active_anon"},
3688 {"inactive_file", "total_inactive_file"},
3689 {"active_file", "total_active_file"},
3690 {"unevictable", "total_unevictable"}
3695 mem_cgroup_get_local_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3700 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3701 s
->stat
[MCS_CACHE
] += val
* PAGE_SIZE
;
3702 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_RSS
);
3703 s
->stat
[MCS_RSS
] += val
* PAGE_SIZE
;
3704 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_FILE_MAPPED
);
3705 s
->stat
[MCS_FILE_MAPPED
] += val
* PAGE_SIZE
;
3706 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGIN_COUNT
);
3707 s
->stat
[MCS_PGPGIN
] += val
;
3708 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGOUT_COUNT
);
3709 s
->stat
[MCS_PGPGOUT
] += val
;
3710 if (do_swap_account
) {
3711 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_SWAPOUT
);
3712 s
->stat
[MCS_SWAP
] += val
* PAGE_SIZE
;
3716 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_ANON
);
3717 s
->stat
[MCS_INACTIVE_ANON
] += val
* PAGE_SIZE
;
3718 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_ANON
);
3719 s
->stat
[MCS_ACTIVE_ANON
] += val
* PAGE_SIZE
;
3720 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_FILE
);
3721 s
->stat
[MCS_INACTIVE_FILE
] += val
* PAGE_SIZE
;
3722 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_FILE
);
3723 s
->stat
[MCS_ACTIVE_FILE
] += val
* PAGE_SIZE
;
3724 val
= mem_cgroup_get_local_zonestat(mem
, LRU_UNEVICTABLE
);
3725 s
->stat
[MCS_UNEVICTABLE
] += val
* PAGE_SIZE
;
3729 mem_cgroup_get_total_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3731 struct mem_cgroup
*iter
;
3733 for_each_mem_cgroup_tree(iter
, mem
)
3734 mem_cgroup_get_local_stat(iter
, s
);
3737 static int mem_control_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
3738 struct cgroup_map_cb
*cb
)
3740 struct mem_cgroup
*mem_cont
= mem_cgroup_from_cont(cont
);
3741 struct mcs_total_stat mystat
;
3744 memset(&mystat
, 0, sizeof(mystat
));
3745 mem_cgroup_get_local_stat(mem_cont
, &mystat
);
3747 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3748 if (i
== MCS_SWAP
&& !do_swap_account
)
3750 cb
->fill(cb
, memcg_stat_strings
[i
].local_name
, mystat
.stat
[i
]);
3753 /* Hierarchical information */
3755 unsigned long long limit
, memsw_limit
;
3756 memcg_get_hierarchical_limit(mem_cont
, &limit
, &memsw_limit
);
3757 cb
->fill(cb
, "hierarchical_memory_limit", limit
);
3758 if (do_swap_account
)
3759 cb
->fill(cb
, "hierarchical_memsw_limit", memsw_limit
);
3762 memset(&mystat
, 0, sizeof(mystat
));
3763 mem_cgroup_get_total_stat(mem_cont
, &mystat
);
3764 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3765 if (i
== MCS_SWAP
&& !do_swap_account
)
3767 cb
->fill(cb
, memcg_stat_strings
[i
].total_name
, mystat
.stat
[i
]);
3770 #ifdef CONFIG_DEBUG_VM
3771 cb
->fill(cb
, "inactive_ratio", calc_inactive_ratio(mem_cont
, NULL
));
3775 struct mem_cgroup_per_zone
*mz
;
3776 unsigned long recent_rotated
[2] = {0, 0};
3777 unsigned long recent_scanned
[2] = {0, 0};
3779 for_each_online_node(nid
)
3780 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3781 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
3783 recent_rotated
[0] +=
3784 mz
->reclaim_stat
.recent_rotated
[0];
3785 recent_rotated
[1] +=
3786 mz
->reclaim_stat
.recent_rotated
[1];
3787 recent_scanned
[0] +=
3788 mz
->reclaim_stat
.recent_scanned
[0];
3789 recent_scanned
[1] +=
3790 mz
->reclaim_stat
.recent_scanned
[1];
3792 cb
->fill(cb
, "recent_rotated_anon", recent_rotated
[0]);
3793 cb
->fill(cb
, "recent_rotated_file", recent_rotated
[1]);
3794 cb
->fill(cb
, "recent_scanned_anon", recent_scanned
[0]);
3795 cb
->fill(cb
, "recent_scanned_file", recent_scanned
[1]);
3802 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
3804 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3806 return get_swappiness(memcg
);
3809 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
3812 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3813 struct mem_cgroup
*parent
;
3818 if (cgrp
->parent
== NULL
)
3821 parent
= mem_cgroup_from_cont(cgrp
->parent
);
3825 /* If under hierarchy, only empty-root can set this value */
3826 if ((parent
->use_hierarchy
) ||
3827 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
3832 spin_lock(&memcg
->reclaim_param_lock
);
3833 memcg
->swappiness
= val
;
3834 spin_unlock(&memcg
->reclaim_param_lock
);
3841 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3843 struct mem_cgroup_threshold_ary
*t
;
3849 t
= rcu_dereference(memcg
->thresholds
.primary
);
3851 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3856 usage
= mem_cgroup_usage(memcg
, swap
);
3859 * current_threshold points to threshold just below usage.
3860 * If it's not true, a threshold was crossed after last
3861 * call of __mem_cgroup_threshold().
3863 i
= t
->current_threshold
;
3866 * Iterate backward over array of thresholds starting from
3867 * current_threshold and check if a threshold is crossed.
3868 * If none of thresholds below usage is crossed, we read
3869 * only one element of the array here.
3871 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3872 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3874 /* i = current_threshold + 1 */
3878 * Iterate forward over array of thresholds starting from
3879 * current_threshold+1 and check if a threshold is crossed.
3880 * If none of thresholds above usage is crossed, we read
3881 * only one element of the array here.
3883 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3884 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3886 /* Update current_threshold */
3887 t
->current_threshold
= i
- 1;
3892 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3895 __mem_cgroup_threshold(memcg
, false);
3896 if (do_swap_account
)
3897 __mem_cgroup_threshold(memcg
, true);
3899 memcg
= parent_mem_cgroup(memcg
);
3903 static int compare_thresholds(const void *a
, const void *b
)
3905 const struct mem_cgroup_threshold
*_a
= a
;
3906 const struct mem_cgroup_threshold
*_b
= b
;
3908 return _a
->threshold
- _b
->threshold
;
3911 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*mem
)
3913 struct mem_cgroup_eventfd_list
*ev
;
3915 list_for_each_entry(ev
, &mem
->oom_notify
, list
)
3916 eventfd_signal(ev
->eventfd
, 1);
3920 static void mem_cgroup_oom_notify(struct mem_cgroup
*mem
)
3922 struct mem_cgroup
*iter
;
3924 for_each_mem_cgroup_tree(iter
, mem
)
3925 mem_cgroup_oom_notify_cb(iter
);
3928 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
3929 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
3931 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3932 struct mem_cgroup_thresholds
*thresholds
;
3933 struct mem_cgroup_threshold_ary
*new;
3934 int type
= MEMFILE_TYPE(cft
->private);
3935 u64 threshold
, usage
;
3938 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
3942 mutex_lock(&memcg
->thresholds_lock
);
3945 thresholds
= &memcg
->thresholds
;
3946 else if (type
== _MEMSWAP
)
3947 thresholds
= &memcg
->memsw_thresholds
;
3951 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
3953 /* Check if a threshold crossed before adding a new one */
3954 if (thresholds
->primary
)
3955 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3957 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3959 /* Allocate memory for new array of thresholds */
3960 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3968 /* Copy thresholds (if any) to new array */
3969 if (thresholds
->primary
) {
3970 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3971 sizeof(struct mem_cgroup_threshold
));
3974 /* Add new threshold */
3975 new->entries
[size
- 1].eventfd
= eventfd
;
3976 new->entries
[size
- 1].threshold
= threshold
;
3978 /* Sort thresholds. Registering of new threshold isn't time-critical */
3979 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3980 compare_thresholds
, NULL
);
3982 /* Find current threshold */
3983 new->current_threshold
= -1;
3984 for (i
= 0; i
< size
; i
++) {
3985 if (new->entries
[i
].threshold
< usage
) {
3987 * new->current_threshold will not be used until
3988 * rcu_assign_pointer(), so it's safe to increment
3991 ++new->current_threshold
;
3995 /* Free old spare buffer and save old primary buffer as spare */
3996 kfree(thresholds
->spare
);
3997 thresholds
->spare
= thresholds
->primary
;
3999 rcu_assign_pointer(thresholds
->primary
, new);
4001 /* To be sure that nobody uses thresholds */
4005 mutex_unlock(&memcg
->thresholds_lock
);
4010 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
4011 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
4013 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
4014 struct mem_cgroup_thresholds
*thresholds
;
4015 struct mem_cgroup_threshold_ary
*new;
4016 int type
= MEMFILE_TYPE(cft
->private);
4020 mutex_lock(&memcg
->thresholds_lock
);
4022 thresholds
= &memcg
->thresholds
;
4023 else if (type
== _MEMSWAP
)
4024 thresholds
= &memcg
->memsw_thresholds
;
4029 * Something went wrong if we trying to unregister a threshold
4030 * if we don't have thresholds
4032 BUG_ON(!thresholds
);
4034 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
4036 /* Check if a threshold crossed before removing */
4037 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4039 /* Calculate new number of threshold */
4041 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4042 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4046 new = thresholds
->spare
;
4048 /* Set thresholds array to NULL if we don't have thresholds */
4057 /* Copy thresholds and find current threshold */
4058 new->current_threshold
= -1;
4059 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4060 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4063 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4064 if (new->entries
[j
].threshold
< usage
) {
4066 * new->current_threshold will not be used
4067 * until rcu_assign_pointer(), so it's safe to increment
4070 ++new->current_threshold
;
4076 /* Swap primary and spare array */
4077 thresholds
->spare
= thresholds
->primary
;
4078 rcu_assign_pointer(thresholds
->primary
, new);
4080 /* To be sure that nobody uses thresholds */
4083 mutex_unlock(&memcg
->thresholds_lock
);
4086 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
4087 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
4089 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
4090 struct mem_cgroup_eventfd_list
*event
;
4091 int type
= MEMFILE_TYPE(cft
->private);
4093 BUG_ON(type
!= _OOM_TYPE
);
4094 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4098 mutex_lock(&memcg_oom_mutex
);
4100 event
->eventfd
= eventfd
;
4101 list_add(&event
->list
, &memcg
->oom_notify
);
4103 /* already in OOM ? */
4104 if (atomic_read(&memcg
->oom_lock
))
4105 eventfd_signal(eventfd
, 1);
4106 mutex_unlock(&memcg_oom_mutex
);
4111 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
4112 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
4114 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4115 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4116 int type
= MEMFILE_TYPE(cft
->private);
4118 BUG_ON(type
!= _OOM_TYPE
);
4120 mutex_lock(&memcg_oom_mutex
);
4122 list_for_each_entry_safe(ev
, tmp
, &mem
->oom_notify
, list
) {
4123 if (ev
->eventfd
== eventfd
) {
4124 list_del(&ev
->list
);
4129 mutex_unlock(&memcg_oom_mutex
);
4132 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
4133 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
4135 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4137 cb
->fill(cb
, "oom_kill_disable", mem
->oom_kill_disable
);
4139 if (atomic_read(&mem
->oom_lock
))
4140 cb
->fill(cb
, "under_oom", 1);
4142 cb
->fill(cb
, "under_oom", 0);
4146 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
4147 struct cftype
*cft
, u64 val
)
4149 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4150 struct mem_cgroup
*parent
;
4152 /* cannot set to root cgroup and only 0 and 1 are allowed */
4153 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
4156 parent
= mem_cgroup_from_cont(cgrp
->parent
);
4159 /* oom-kill-disable is a flag for subhierarchy. */
4160 if ((parent
->use_hierarchy
) ||
4161 (mem
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
4165 mem
->oom_kill_disable
= val
;
4167 memcg_oom_recover(mem
);
4172 static struct cftype mem_cgroup_files
[] = {
4174 .name
= "usage_in_bytes",
4175 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4176 .read_u64
= mem_cgroup_read
,
4177 .register_event
= mem_cgroup_usage_register_event
,
4178 .unregister_event
= mem_cgroup_usage_unregister_event
,
4181 .name
= "max_usage_in_bytes",
4182 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4183 .trigger
= mem_cgroup_reset
,
4184 .read_u64
= mem_cgroup_read
,
4187 .name
= "limit_in_bytes",
4188 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4189 .write_string
= mem_cgroup_write
,
4190 .read_u64
= mem_cgroup_read
,
4193 .name
= "soft_limit_in_bytes",
4194 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4195 .write_string
= mem_cgroup_write
,
4196 .read_u64
= mem_cgroup_read
,
4200 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4201 .trigger
= mem_cgroup_reset
,
4202 .read_u64
= mem_cgroup_read
,
4206 .read_map
= mem_control_stat_show
,
4209 .name
= "force_empty",
4210 .trigger
= mem_cgroup_force_empty_write
,
4213 .name
= "use_hierarchy",
4214 .write_u64
= mem_cgroup_hierarchy_write
,
4215 .read_u64
= mem_cgroup_hierarchy_read
,
4218 .name
= "swappiness",
4219 .read_u64
= mem_cgroup_swappiness_read
,
4220 .write_u64
= mem_cgroup_swappiness_write
,
4223 .name
= "move_charge_at_immigrate",
4224 .read_u64
= mem_cgroup_move_charge_read
,
4225 .write_u64
= mem_cgroup_move_charge_write
,
4228 .name
= "oom_control",
4229 .read_map
= mem_cgroup_oom_control_read
,
4230 .write_u64
= mem_cgroup_oom_control_write
,
4231 .register_event
= mem_cgroup_oom_register_event
,
4232 .unregister_event
= mem_cgroup_oom_unregister_event
,
4233 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4237 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4238 static struct cftype memsw_cgroup_files
[] = {
4240 .name
= "memsw.usage_in_bytes",
4241 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
4242 .read_u64
= mem_cgroup_read
,
4243 .register_event
= mem_cgroup_usage_register_event
,
4244 .unregister_event
= mem_cgroup_usage_unregister_event
,
4247 .name
= "memsw.max_usage_in_bytes",
4248 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
4249 .trigger
= mem_cgroup_reset
,
4250 .read_u64
= mem_cgroup_read
,
4253 .name
= "memsw.limit_in_bytes",
4254 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
4255 .write_string
= mem_cgroup_write
,
4256 .read_u64
= mem_cgroup_read
,
4259 .name
= "memsw.failcnt",
4260 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
4261 .trigger
= mem_cgroup_reset
,
4262 .read_u64
= mem_cgroup_read
,
4266 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4268 if (!do_swap_account
)
4270 return cgroup_add_files(cont
, ss
, memsw_cgroup_files
,
4271 ARRAY_SIZE(memsw_cgroup_files
));
4274 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4280 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4282 struct mem_cgroup_per_node
*pn
;
4283 struct mem_cgroup_per_zone
*mz
;
4285 int zone
, tmp
= node
;
4287 * This routine is called against possible nodes.
4288 * But it's BUG to call kmalloc() against offline node.
4290 * TODO: this routine can waste much memory for nodes which will
4291 * never be onlined. It's better to use memory hotplug callback
4294 if (!node_state(node
, N_NORMAL_MEMORY
))
4296 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4300 mem
->info
.nodeinfo
[node
] = pn
;
4301 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4302 mz
= &pn
->zoneinfo
[zone
];
4304 INIT_LIST_HEAD(&mz
->lists
[l
]);
4305 mz
->usage_in_excess
= 0;
4306 mz
->on_tree
= false;
4312 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4314 kfree(mem
->info
.nodeinfo
[node
]);
4317 static struct mem_cgroup
*mem_cgroup_alloc(void)
4319 struct mem_cgroup
*mem
;
4320 int size
= sizeof(struct mem_cgroup
);
4322 /* Can be very big if MAX_NUMNODES is very big */
4323 if (size
< PAGE_SIZE
)
4324 mem
= kzalloc(size
, GFP_KERNEL
);
4326 mem
= vzalloc(size
);
4331 mem
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4334 spin_lock_init(&mem
->pcp_counter_lock
);
4338 if (size
< PAGE_SIZE
)
4346 * At destroying mem_cgroup, references from swap_cgroup can remain.
4347 * (scanning all at force_empty is too costly...)
4349 * Instead of clearing all references at force_empty, we remember
4350 * the number of reference from swap_cgroup and free mem_cgroup when
4351 * it goes down to 0.
4353 * Removal of cgroup itself succeeds regardless of refs from swap.
4356 static void __mem_cgroup_free(struct mem_cgroup
*mem
)
4360 mem_cgroup_remove_from_trees(mem
);
4361 free_css_id(&mem_cgroup_subsys
, &mem
->css
);
4363 for_each_node_state(node
, N_POSSIBLE
)
4364 free_mem_cgroup_per_zone_info(mem
, node
);
4366 free_percpu(mem
->stat
);
4367 if (sizeof(struct mem_cgroup
) < PAGE_SIZE
)
4373 static void mem_cgroup_get(struct mem_cgroup
*mem
)
4375 atomic_inc(&mem
->refcnt
);
4378 static void __mem_cgroup_put(struct mem_cgroup
*mem
, int count
)
4380 if (atomic_sub_and_test(count
, &mem
->refcnt
)) {
4381 struct mem_cgroup
*parent
= parent_mem_cgroup(mem
);
4382 __mem_cgroup_free(mem
);
4384 mem_cgroup_put(parent
);
4388 static void mem_cgroup_put(struct mem_cgroup
*mem
)
4390 __mem_cgroup_put(mem
, 1);
4394 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4396 static struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*mem
)
4398 if (!mem
->res
.parent
)
4400 return mem_cgroup_from_res_counter(mem
->res
.parent
, res
);
4403 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4404 static void __init
enable_swap_cgroup(void)
4406 if (!mem_cgroup_disabled() && really_do_swap_account
)
4407 do_swap_account
= 1;
4410 static void __init
enable_swap_cgroup(void)
4415 static int mem_cgroup_soft_limit_tree_init(void)
4417 struct mem_cgroup_tree_per_node
*rtpn
;
4418 struct mem_cgroup_tree_per_zone
*rtpz
;
4419 int tmp
, node
, zone
;
4421 for_each_node_state(node
, N_POSSIBLE
) {
4423 if (!node_state(node
, N_NORMAL_MEMORY
))
4425 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
4429 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
4431 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4432 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
4433 rtpz
->rb_root
= RB_ROOT
;
4434 spin_lock_init(&rtpz
->lock
);
4440 static struct cgroup_subsys_state
* __ref
4441 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
4443 struct mem_cgroup
*mem
, *parent
;
4444 long error
= -ENOMEM
;
4447 mem
= mem_cgroup_alloc();
4449 return ERR_PTR(error
);
4451 for_each_node_state(node
, N_POSSIBLE
)
4452 if (alloc_mem_cgroup_per_zone_info(mem
, node
))
4456 if (cont
->parent
== NULL
) {
4458 enable_swap_cgroup();
4460 root_mem_cgroup
= mem
;
4461 if (mem_cgroup_soft_limit_tree_init())
4463 for_each_possible_cpu(cpu
) {
4464 struct memcg_stock_pcp
*stock
=
4465 &per_cpu(memcg_stock
, cpu
);
4466 INIT_WORK(&stock
->work
, drain_local_stock
);
4468 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
4470 parent
= mem_cgroup_from_cont(cont
->parent
);
4471 mem
->use_hierarchy
= parent
->use_hierarchy
;
4472 mem
->oom_kill_disable
= parent
->oom_kill_disable
;
4475 if (parent
&& parent
->use_hierarchy
) {
4476 res_counter_init(&mem
->res
, &parent
->res
);
4477 res_counter_init(&mem
->memsw
, &parent
->memsw
);
4479 * We increment refcnt of the parent to ensure that we can
4480 * safely access it on res_counter_charge/uncharge.
4481 * This refcnt will be decremented when freeing this
4482 * mem_cgroup(see mem_cgroup_put).
4484 mem_cgroup_get(parent
);
4486 res_counter_init(&mem
->res
, NULL
);
4487 res_counter_init(&mem
->memsw
, NULL
);
4489 mem
->last_scanned_child
= 0;
4490 spin_lock_init(&mem
->reclaim_param_lock
);
4491 INIT_LIST_HEAD(&mem
->oom_notify
);
4494 mem
->swappiness
= get_swappiness(parent
);
4495 atomic_set(&mem
->refcnt
, 1);
4496 mem
->move_charge_at_immigrate
= 0;
4497 mutex_init(&mem
->thresholds_lock
);
4500 __mem_cgroup_free(mem
);
4501 root_mem_cgroup
= NULL
;
4502 return ERR_PTR(error
);
4505 static int mem_cgroup_pre_destroy(struct cgroup_subsys
*ss
,
4506 struct cgroup
*cont
)
4508 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4510 return mem_cgroup_force_empty(mem
, false);
4513 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
4514 struct cgroup
*cont
)
4516 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4518 mem_cgroup_put(mem
);
4521 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
4522 struct cgroup
*cont
)
4526 ret
= cgroup_add_files(cont
, ss
, mem_cgroup_files
,
4527 ARRAY_SIZE(mem_cgroup_files
));
4530 ret
= register_memsw_files(cont
, ss
);
4535 /* Handlers for move charge at task migration. */
4536 #define PRECHARGE_COUNT_AT_ONCE 256
4537 static int mem_cgroup_do_precharge(unsigned long count
)
4540 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4541 struct mem_cgroup
*mem
= mc
.to
;
4543 if (mem_cgroup_is_root(mem
)) {
4544 mc
.precharge
+= count
;
4545 /* we don't need css_get for root */
4548 /* try to charge at once */
4550 struct res_counter
*dummy
;
4552 * "mem" cannot be under rmdir() because we've already checked
4553 * by cgroup_lock_live_cgroup() that it is not removed and we
4554 * are still under the same cgroup_mutex. So we can postpone
4557 if (res_counter_charge(&mem
->res
, PAGE_SIZE
* count
, &dummy
))
4559 if (do_swap_account
&& res_counter_charge(&mem
->memsw
,
4560 PAGE_SIZE
* count
, &dummy
)) {
4561 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
4564 mc
.precharge
+= count
;
4568 /* fall back to one by one charge */
4570 if (signal_pending(current
)) {
4574 if (!batch_count
--) {
4575 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4578 ret
= __mem_cgroup_try_charge(NULL
, GFP_KERNEL
, &mem
, false,
4581 /* mem_cgroup_clear_mc() will do uncharge later */
4589 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4590 * @vma: the vma the pte to be checked belongs
4591 * @addr: the address corresponding to the pte to be checked
4592 * @ptent: the pte to be checked
4593 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4596 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4597 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4598 * move charge. if @target is not NULL, the page is stored in target->page
4599 * with extra refcnt got(Callers should handle it).
4600 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4601 * target for charge migration. if @target is not NULL, the entry is stored
4604 * Called with pte lock held.
4611 enum mc_target_type
{
4612 MC_TARGET_NONE
, /* not used */
4617 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4618 unsigned long addr
, pte_t ptent
)
4620 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4622 if (!page
|| !page_mapped(page
))
4624 if (PageAnon(page
)) {
4625 /* we don't move shared anon */
4626 if (!move_anon() || page_mapcount(page
) > 2)
4628 } else if (!move_file())
4629 /* we ignore mapcount for file pages */
4631 if (!get_page_unless_zero(page
))
4637 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4638 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4641 struct page
*page
= NULL
;
4642 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4644 if (!move_anon() || non_swap_entry(ent
))
4646 usage_count
= mem_cgroup_count_swap_user(ent
, &page
);
4647 if (usage_count
> 1) { /* we don't move shared anon */
4652 if (do_swap_account
)
4653 entry
->val
= ent
.val
;
4658 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4659 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4661 struct page
*page
= NULL
;
4662 struct inode
*inode
;
4663 struct address_space
*mapping
;
4666 if (!vma
->vm_file
) /* anonymous vma */
4671 inode
= vma
->vm_file
->f_path
.dentry
->d_inode
;
4672 mapping
= vma
->vm_file
->f_mapping
;
4673 if (pte_none(ptent
))
4674 pgoff
= linear_page_index(vma
, addr
);
4675 else /* pte_file(ptent) is true */
4676 pgoff
= pte_to_pgoff(ptent
);
4678 /* page is moved even if it's not RSS of this task(page-faulted). */
4679 if (!mapping_cap_swap_backed(mapping
)) { /* normal file */
4680 page
= find_get_page(mapping
, pgoff
);
4681 } else { /* shmem/tmpfs file. we should take account of swap too. */
4683 mem_cgroup_get_shmem_target(inode
, pgoff
, &page
, &ent
);
4684 if (do_swap_account
)
4685 entry
->val
= ent
.val
;
4691 static int is_target_pte_for_mc(struct vm_area_struct
*vma
,
4692 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4694 struct page
*page
= NULL
;
4695 struct page_cgroup
*pc
;
4697 swp_entry_t ent
= { .val
= 0 };
4699 if (pte_present(ptent
))
4700 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4701 else if (is_swap_pte(ptent
))
4702 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4703 else if (pte_none(ptent
) || pte_file(ptent
))
4704 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4706 if (!page
&& !ent
.val
)
4709 pc
= lookup_page_cgroup(page
);
4711 * Do only loose check w/o page_cgroup lock.
4712 * mem_cgroup_move_account() checks the pc is valid or not under
4715 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
4716 ret
= MC_TARGET_PAGE
;
4718 target
->page
= page
;
4720 if (!ret
|| !target
)
4723 /* There is a swap entry and a page doesn't exist or isn't charged */
4724 if (ent
.val
&& !ret
&&
4725 css_id(&mc
.from
->css
) == lookup_swap_cgroup(ent
)) {
4726 ret
= MC_TARGET_SWAP
;
4733 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4734 unsigned long addr
, unsigned long end
,
4735 struct mm_walk
*walk
)
4737 struct vm_area_struct
*vma
= walk
->private;
4741 split_huge_page_pmd(walk
->mm
, pmd
);
4743 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4744 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4745 if (is_target_pte_for_mc(vma
, addr
, *pte
, NULL
))
4746 mc
.precharge
++; /* increment precharge temporarily */
4747 pte_unmap_unlock(pte
- 1, ptl
);
4753 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4755 unsigned long precharge
;
4756 struct vm_area_struct
*vma
;
4758 down_read(&mm
->mmap_sem
);
4759 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
4760 struct mm_walk mem_cgroup_count_precharge_walk
= {
4761 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4765 if (is_vm_hugetlb_page(vma
))
4767 walk_page_range(vma
->vm_start
, vma
->vm_end
,
4768 &mem_cgroup_count_precharge_walk
);
4770 up_read(&mm
->mmap_sem
);
4772 precharge
= mc
.precharge
;
4778 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4780 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4782 VM_BUG_ON(mc
.moving_task
);
4783 mc
.moving_task
= current
;
4784 return mem_cgroup_do_precharge(precharge
);
4787 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4788 static void __mem_cgroup_clear_mc(void)
4790 struct mem_cgroup
*from
= mc
.from
;
4791 struct mem_cgroup
*to
= mc
.to
;
4793 /* we must uncharge all the leftover precharges from mc.to */
4795 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
4799 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4800 * we must uncharge here.
4802 if (mc
.moved_charge
) {
4803 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
4804 mc
.moved_charge
= 0;
4806 /* we must fixup refcnts and charges */
4807 if (mc
.moved_swap
) {
4808 /* uncharge swap account from the old cgroup */
4809 if (!mem_cgroup_is_root(mc
.from
))
4810 res_counter_uncharge(&mc
.from
->memsw
,
4811 PAGE_SIZE
* mc
.moved_swap
);
4812 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
4814 if (!mem_cgroup_is_root(mc
.to
)) {
4816 * we charged both to->res and to->memsw, so we should
4819 res_counter_uncharge(&mc
.to
->res
,
4820 PAGE_SIZE
* mc
.moved_swap
);
4822 /* we've already done mem_cgroup_get(mc.to) */
4825 memcg_oom_recover(from
);
4826 memcg_oom_recover(to
);
4827 wake_up_all(&mc
.waitq
);
4830 static void mem_cgroup_clear_mc(void)
4832 struct mem_cgroup
*from
= mc
.from
;
4835 * we must clear moving_task before waking up waiters at the end of
4838 mc
.moving_task
= NULL
;
4839 __mem_cgroup_clear_mc();
4840 spin_lock(&mc
.lock
);
4843 spin_unlock(&mc
.lock
);
4844 mem_cgroup_end_move(from
);
4847 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
4848 struct cgroup
*cgroup
,
4849 struct task_struct
*p
,
4853 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgroup
);
4855 if (mem
->move_charge_at_immigrate
) {
4856 struct mm_struct
*mm
;
4857 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
4859 VM_BUG_ON(from
== mem
);
4861 mm
= get_task_mm(p
);
4864 /* We move charges only when we move a owner of the mm */
4865 if (mm
->owner
== p
) {
4868 VM_BUG_ON(mc
.precharge
);
4869 VM_BUG_ON(mc
.moved_charge
);
4870 VM_BUG_ON(mc
.moved_swap
);
4871 mem_cgroup_start_move(from
);
4872 spin_lock(&mc
.lock
);
4875 spin_unlock(&mc
.lock
);
4876 /* We set mc.moving_task later */
4878 ret
= mem_cgroup_precharge_mc(mm
);
4880 mem_cgroup_clear_mc();
4887 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
4888 struct cgroup
*cgroup
,
4889 struct task_struct
*p
,
4892 mem_cgroup_clear_mc();
4895 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4896 unsigned long addr
, unsigned long end
,
4897 struct mm_walk
*walk
)
4900 struct vm_area_struct
*vma
= walk
->private;
4904 split_huge_page_pmd(walk
->mm
, pmd
);
4906 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4907 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4908 pte_t ptent
= *(pte
++);
4909 union mc_target target
;
4912 struct page_cgroup
*pc
;
4918 type
= is_target_pte_for_mc(vma
, addr
, ptent
, &target
);
4920 case MC_TARGET_PAGE
:
4922 if (isolate_lru_page(page
))
4924 pc
= lookup_page_cgroup(page
);
4925 if (!mem_cgroup_move_account(pc
,
4926 mc
.from
, mc
.to
, false, PAGE_SIZE
)) {
4928 /* we uncharge from mc.from later. */
4931 putback_lru_page(page
);
4932 put
: /* is_target_pte_for_mc() gets the page */
4935 case MC_TARGET_SWAP
:
4937 if (!mem_cgroup_move_swap_account(ent
,
4938 mc
.from
, mc
.to
, false)) {
4940 /* we fixup refcnts and charges later. */
4948 pte_unmap_unlock(pte
- 1, ptl
);
4953 * We have consumed all precharges we got in can_attach().
4954 * We try charge one by one, but don't do any additional
4955 * charges to mc.to if we have failed in charge once in attach()
4958 ret
= mem_cgroup_do_precharge(1);
4966 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4968 struct vm_area_struct
*vma
;
4970 lru_add_drain_all();
4972 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4974 * Someone who are holding the mmap_sem might be waiting in
4975 * waitq. So we cancel all extra charges, wake up all waiters,
4976 * and retry. Because we cancel precharges, we might not be able
4977 * to move enough charges, but moving charge is a best-effort
4978 * feature anyway, so it wouldn't be a big problem.
4980 __mem_cgroup_clear_mc();
4984 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
4986 struct mm_walk mem_cgroup_move_charge_walk
= {
4987 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4991 if (is_vm_hugetlb_page(vma
))
4993 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
4994 &mem_cgroup_move_charge_walk
);
4997 * means we have consumed all precharges and failed in
4998 * doing additional charge. Just abandon here.
5002 up_read(&mm
->mmap_sem
);
5005 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
5006 struct cgroup
*cont
,
5007 struct cgroup
*old_cont
,
5008 struct task_struct
*p
,
5011 struct mm_struct
*mm
;
5014 /* no need to move charge */
5017 mm
= get_task_mm(p
);
5019 mem_cgroup_move_charge(mm
);
5022 mem_cgroup_clear_mc();
5024 #else /* !CONFIG_MMU */
5025 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
5026 struct cgroup
*cgroup
,
5027 struct task_struct
*p
,
5032 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
5033 struct cgroup
*cgroup
,
5034 struct task_struct
*p
,
5038 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
5039 struct cgroup
*cont
,
5040 struct cgroup
*old_cont
,
5041 struct task_struct
*p
,
5047 struct cgroup_subsys mem_cgroup_subsys
= {
5049 .subsys_id
= mem_cgroup_subsys_id
,
5050 .create
= mem_cgroup_create
,
5051 .pre_destroy
= mem_cgroup_pre_destroy
,
5052 .destroy
= mem_cgroup_destroy
,
5053 .populate
= mem_cgroup_populate
,
5054 .can_attach
= mem_cgroup_can_attach
,
5055 .cancel_attach
= mem_cgroup_cancel_attach
,
5056 .attach
= mem_cgroup_move_task
,
5061 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5062 static int __init
enable_swap_account(char *s
)
5064 /* consider enabled if no parameter or 1 is given */
5065 if (!(*s
) || !strcmp(s
, "=1"))
5066 really_do_swap_account
= 1;
5067 else if (!strcmp(s
, "=0"))
5068 really_do_swap_account
= 0;
5071 __setup("swapaccount", enable_swap_account
);
5073 static int __init
disable_swap_account(char *s
)
5075 printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5076 enable_swap_account("=0");
5079 __setup("noswapaccount", disable_swap_account
);