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
;
222 * While reclaiming in a hierarchy, we cache the last child we
225 int last_scanned_child
;
227 * Should the accounting and control be hierarchical, per subtree?
233 unsigned int swappiness
;
234 /* OOM-Killer disable */
235 int oom_kill_disable
;
237 /* set when res.limit == memsw.limit */
238 bool memsw_is_minimum
;
240 /* protect arrays of thresholds */
241 struct mutex thresholds_lock
;
243 /* thresholds for memory usage. RCU-protected */
244 struct mem_cgroup_thresholds thresholds
;
246 /* thresholds for mem+swap usage. RCU-protected */
247 struct mem_cgroup_thresholds memsw_thresholds
;
249 /* For oom notifier event fd */
250 struct list_head oom_notify
;
253 * Should we move charges of a task when a task is moved into this
254 * mem_cgroup ? And what type of charges should we move ?
256 unsigned long move_charge_at_immigrate
;
260 struct mem_cgroup_stat_cpu
*stat
;
262 * used when a cpu is offlined or other synchronizations
263 * See mem_cgroup_read_stat().
265 struct mem_cgroup_stat_cpu nocpu_base
;
266 spinlock_t pcp_counter_lock
;
269 /* Stuffs for move charges at task migration. */
271 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
272 * left-shifted bitmap of these types.
275 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
276 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
280 /* "mc" and its members are protected by cgroup_mutex */
281 static struct move_charge_struct
{
282 spinlock_t lock
; /* for from, to */
283 struct mem_cgroup
*from
;
284 struct mem_cgroup
*to
;
285 unsigned long precharge
;
286 unsigned long moved_charge
;
287 unsigned long moved_swap
;
288 struct task_struct
*moving_task
; /* a task moving charges */
289 wait_queue_head_t waitq
; /* a waitq for other context */
291 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
292 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
295 static bool move_anon(void)
297 return test_bit(MOVE_CHARGE_TYPE_ANON
,
298 &mc
.to
->move_charge_at_immigrate
);
301 static bool move_file(void)
303 return test_bit(MOVE_CHARGE_TYPE_FILE
,
304 &mc
.to
->move_charge_at_immigrate
);
308 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
309 * limit reclaim to prevent infinite loops, if they ever occur.
311 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
312 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
315 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
316 MEM_CGROUP_CHARGE_TYPE_MAPPED
,
317 MEM_CGROUP_CHARGE_TYPE_SHMEM
, /* used by page migration of shmem */
318 MEM_CGROUP_CHARGE_TYPE_FORCE
, /* used by force_empty */
319 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
320 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
324 /* for encoding cft->private value on file */
327 #define _OOM_TYPE (2)
328 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
329 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
330 #define MEMFILE_ATTR(val) ((val) & 0xffff)
331 /* Used for OOM nofiier */
332 #define OOM_CONTROL (0)
335 * Reclaim flags for mem_cgroup_hierarchical_reclaim
337 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
338 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
339 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
340 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
341 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
342 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
344 static void mem_cgroup_get(struct mem_cgroup
*mem
);
345 static void mem_cgroup_put(struct mem_cgroup
*mem
);
346 static struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*mem
);
347 static void drain_all_stock_async(void);
349 static struct mem_cgroup_per_zone
*
350 mem_cgroup_zoneinfo(struct mem_cgroup
*mem
, int nid
, int zid
)
352 return &mem
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
355 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*mem
)
360 static struct mem_cgroup_per_zone
*
361 page_cgroup_zoneinfo(struct mem_cgroup
*mem
, struct page
*page
)
363 int nid
= page_to_nid(page
);
364 int zid
= page_zonenum(page
);
366 return mem_cgroup_zoneinfo(mem
, nid
, zid
);
369 static struct mem_cgroup_tree_per_zone
*
370 soft_limit_tree_node_zone(int nid
, int zid
)
372 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
375 static struct mem_cgroup_tree_per_zone
*
376 soft_limit_tree_from_page(struct page
*page
)
378 int nid
= page_to_nid(page
);
379 int zid
= page_zonenum(page
);
381 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
385 __mem_cgroup_insert_exceeded(struct mem_cgroup
*mem
,
386 struct mem_cgroup_per_zone
*mz
,
387 struct mem_cgroup_tree_per_zone
*mctz
,
388 unsigned long long new_usage_in_excess
)
390 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
391 struct rb_node
*parent
= NULL
;
392 struct mem_cgroup_per_zone
*mz_node
;
397 mz
->usage_in_excess
= new_usage_in_excess
;
398 if (!mz
->usage_in_excess
)
402 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
404 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
407 * We can't avoid mem cgroups that are over their soft
408 * limit by the same amount
410 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
413 rb_link_node(&mz
->tree_node
, parent
, p
);
414 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
419 __mem_cgroup_remove_exceeded(struct mem_cgroup
*mem
,
420 struct mem_cgroup_per_zone
*mz
,
421 struct mem_cgroup_tree_per_zone
*mctz
)
425 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
430 mem_cgroup_remove_exceeded(struct mem_cgroup
*mem
,
431 struct mem_cgroup_per_zone
*mz
,
432 struct mem_cgroup_tree_per_zone
*mctz
)
434 spin_lock(&mctz
->lock
);
435 __mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
436 spin_unlock(&mctz
->lock
);
440 static void mem_cgroup_update_tree(struct mem_cgroup
*mem
, struct page
*page
)
442 unsigned long long excess
;
443 struct mem_cgroup_per_zone
*mz
;
444 struct mem_cgroup_tree_per_zone
*mctz
;
445 int nid
= page_to_nid(page
);
446 int zid
= page_zonenum(page
);
447 mctz
= soft_limit_tree_from_page(page
);
450 * Necessary to update all ancestors when hierarchy is used.
451 * because their event counter is not touched.
453 for (; mem
; mem
= parent_mem_cgroup(mem
)) {
454 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
455 excess
= res_counter_soft_limit_excess(&mem
->res
);
457 * We have to update the tree if mz is on RB-tree or
458 * mem is over its softlimit.
460 if (excess
|| mz
->on_tree
) {
461 spin_lock(&mctz
->lock
);
462 /* if on-tree, remove it */
464 __mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
466 * Insert again. mz->usage_in_excess will be updated.
467 * If excess is 0, no tree ops.
469 __mem_cgroup_insert_exceeded(mem
, mz
, mctz
, excess
);
470 spin_unlock(&mctz
->lock
);
475 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*mem
)
478 struct mem_cgroup_per_zone
*mz
;
479 struct mem_cgroup_tree_per_zone
*mctz
;
481 for_each_node_state(node
, N_POSSIBLE
) {
482 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
483 mz
= mem_cgroup_zoneinfo(mem
, node
, zone
);
484 mctz
= soft_limit_tree_node_zone(node
, zone
);
485 mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
490 static struct mem_cgroup_per_zone
*
491 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
493 struct rb_node
*rightmost
= NULL
;
494 struct mem_cgroup_per_zone
*mz
;
498 rightmost
= rb_last(&mctz
->rb_root
);
500 goto done
; /* Nothing to reclaim from */
502 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
504 * Remove the node now but someone else can add it back,
505 * we will to add it back at the end of reclaim to its correct
506 * position in the tree.
508 __mem_cgroup_remove_exceeded(mz
->mem
, mz
, mctz
);
509 if (!res_counter_soft_limit_excess(&mz
->mem
->res
) ||
510 !css_tryget(&mz
->mem
->css
))
516 static struct mem_cgroup_per_zone
*
517 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
519 struct mem_cgroup_per_zone
*mz
;
521 spin_lock(&mctz
->lock
);
522 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
523 spin_unlock(&mctz
->lock
);
528 * Implementation Note: reading percpu statistics for memcg.
530 * Both of vmstat[] and percpu_counter has threshold and do periodic
531 * synchronization to implement "quick" read. There are trade-off between
532 * reading cost and precision of value. Then, we may have a chance to implement
533 * a periodic synchronizion of counter in memcg's counter.
535 * But this _read() function is used for user interface now. The user accounts
536 * memory usage by memory cgroup and he _always_ requires exact value because
537 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
538 * have to visit all online cpus and make sum. So, for now, unnecessary
539 * synchronization is not implemented. (just implemented for cpu hotplug)
541 * If there are kernel internal actions which can make use of some not-exact
542 * value, and reading all cpu value can be performance bottleneck in some
543 * common workload, threashold and synchonization as vmstat[] should be
546 static s64
mem_cgroup_read_stat(struct mem_cgroup
*mem
,
547 enum mem_cgroup_stat_index idx
)
553 for_each_online_cpu(cpu
)
554 val
+= per_cpu(mem
->stat
->count
[idx
], cpu
);
555 #ifdef CONFIG_HOTPLUG_CPU
556 spin_lock(&mem
->pcp_counter_lock
);
557 val
+= mem
->nocpu_base
.count
[idx
];
558 spin_unlock(&mem
->pcp_counter_lock
);
564 static s64
mem_cgroup_local_usage(struct mem_cgroup
*mem
)
568 ret
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_RSS
);
569 ret
+= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_CACHE
);
573 static void mem_cgroup_swap_statistics(struct mem_cgroup
*mem
,
576 int val
= (charge
) ? 1 : -1;
577 this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_SWAPOUT
], val
);
580 static void mem_cgroup_charge_statistics(struct mem_cgroup
*mem
,
581 bool file
, int nr_pages
)
586 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_pages
);
588 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_pages
);
590 /* pagein of a big page is an event. So, ignore page size */
592 __this_cpu_inc(mem
->stat
->count
[MEM_CGROUP_STAT_PGPGIN_COUNT
]);
594 __this_cpu_inc(mem
->stat
->count
[MEM_CGROUP_STAT_PGPGOUT_COUNT
]);
595 nr_pages
= -nr_pages
; /* for event */
598 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_EVENTS
], nr_pages
);
603 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup
*mem
,
607 struct mem_cgroup_per_zone
*mz
;
610 for_each_online_node(nid
)
611 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
612 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
613 total
+= MEM_CGROUP_ZSTAT(mz
, idx
);
618 static bool __memcg_event_check(struct mem_cgroup
*mem
, int event_mask_shift
)
622 val
= this_cpu_read(mem
->stat
->count
[MEM_CGROUP_EVENTS
]);
624 return !(val
& ((1 << event_mask_shift
) - 1));
628 * Check events in order.
631 static void memcg_check_events(struct mem_cgroup
*mem
, struct page
*page
)
633 /* threshold event is triggered in finer grain than soft limit */
634 if (unlikely(__memcg_event_check(mem
, THRESHOLDS_EVENTS_THRESH
))) {
635 mem_cgroup_threshold(mem
);
636 if (unlikely(__memcg_event_check(mem
, SOFTLIMIT_EVENTS_THRESH
)))
637 mem_cgroup_update_tree(mem
, page
);
641 static struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
643 return container_of(cgroup_subsys_state(cont
,
644 mem_cgroup_subsys_id
), struct mem_cgroup
,
648 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
651 * mm_update_next_owner() may clear mm->owner to NULL
652 * if it races with swapoff, page migration, etc.
653 * So this can be called with p == NULL.
658 return container_of(task_subsys_state(p
, mem_cgroup_subsys_id
),
659 struct mem_cgroup
, css
);
662 static struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
664 struct mem_cgroup
*mem
= NULL
;
669 * Because we have no locks, mm->owner's may be being moved to other
670 * cgroup. We use css_tryget() here even if this looks
671 * pessimistic (rather than adding locks here).
675 mem
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
678 } while (!css_tryget(&mem
->css
));
683 /* The caller has to guarantee "mem" exists before calling this */
684 static struct mem_cgroup
*mem_cgroup_start_loop(struct mem_cgroup
*mem
)
686 struct cgroup_subsys_state
*css
;
689 if (!mem
) /* ROOT cgroup has the smallest ID */
690 return root_mem_cgroup
; /*css_put/get against root is ignored*/
691 if (!mem
->use_hierarchy
) {
692 if (css_tryget(&mem
->css
))
698 * searching a memory cgroup which has the smallest ID under given
699 * ROOT cgroup. (ID >= 1)
701 css
= css_get_next(&mem_cgroup_subsys
, 1, &mem
->css
, &found
);
702 if (css
&& css_tryget(css
))
703 mem
= container_of(css
, struct mem_cgroup
, css
);
710 static struct mem_cgroup
*mem_cgroup_get_next(struct mem_cgroup
*iter
,
711 struct mem_cgroup
*root
,
714 int nextid
= css_id(&iter
->css
) + 1;
717 struct cgroup_subsys_state
*css
;
719 hierarchy_used
= iter
->use_hierarchy
;
722 /* If no ROOT, walk all, ignore hierarchy */
723 if (!cond
|| (root
&& !hierarchy_used
))
727 root
= root_mem_cgroup
;
733 css
= css_get_next(&mem_cgroup_subsys
, nextid
,
735 if (css
&& css_tryget(css
))
736 iter
= container_of(css
, struct mem_cgroup
, css
);
738 /* If css is NULL, no more cgroups will be found */
740 } while (css
&& !iter
);
745 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
746 * be careful that "break" loop is not allowed. We have reference count.
747 * Instead of that modify "cond" to be false and "continue" to exit the loop.
749 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
750 for (iter = mem_cgroup_start_loop(root);\
752 iter = mem_cgroup_get_next(iter, root, cond))
754 #define for_each_mem_cgroup_tree(iter, root) \
755 for_each_mem_cgroup_tree_cond(iter, root, true)
757 #define for_each_mem_cgroup_all(iter) \
758 for_each_mem_cgroup_tree_cond(iter, NULL, true)
761 static inline bool mem_cgroup_is_root(struct mem_cgroup
*mem
)
763 return (mem
== root_mem_cgroup
);
767 * Following LRU functions are allowed to be used without PCG_LOCK.
768 * Operations are called by routine of global LRU independently from memcg.
769 * What we have to take care of here is validness of pc->mem_cgroup.
771 * Changes to pc->mem_cgroup happens when
774 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
775 * It is added to LRU before charge.
776 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
777 * When moving account, the page is not on LRU. It's isolated.
780 void mem_cgroup_del_lru_list(struct page
*page
, enum lru_list lru
)
782 struct page_cgroup
*pc
;
783 struct mem_cgroup_per_zone
*mz
;
785 if (mem_cgroup_disabled())
787 pc
= lookup_page_cgroup(page
);
788 /* can happen while we handle swapcache. */
789 if (!TestClearPageCgroupAcctLRU(pc
))
791 VM_BUG_ON(!pc
->mem_cgroup
);
793 * We don't check PCG_USED bit. It's cleared when the "page" is finally
794 * removed from global LRU.
796 mz
= page_cgroup_zoneinfo(pc
->mem_cgroup
, page
);
797 /* huge page split is done under lru_lock. so, we have no races. */
798 MEM_CGROUP_ZSTAT(mz
, lru
) -= 1 << compound_order(page
);
799 if (mem_cgroup_is_root(pc
->mem_cgroup
))
801 VM_BUG_ON(list_empty(&pc
->lru
));
802 list_del_init(&pc
->lru
);
805 void mem_cgroup_del_lru(struct page
*page
)
807 mem_cgroup_del_lru_list(page
, page_lru(page
));
811 * Writeback is about to end against a page which has been marked for immediate
812 * reclaim. If it still appears to be reclaimable, move it to the tail of the
815 void mem_cgroup_rotate_reclaimable_page(struct page
*page
)
817 struct mem_cgroup_per_zone
*mz
;
818 struct page_cgroup
*pc
;
819 enum lru_list lru
= page_lru(page
);
821 if (mem_cgroup_disabled())
824 pc
= lookup_page_cgroup(page
);
825 /* unused or root page is not rotated. */
826 if (!PageCgroupUsed(pc
))
828 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
830 if (mem_cgroup_is_root(pc
->mem_cgroup
))
832 mz
= page_cgroup_zoneinfo(pc
->mem_cgroup
, page
);
833 list_move_tail(&pc
->lru
, &mz
->lists
[lru
]);
836 void mem_cgroup_rotate_lru_list(struct page
*page
, enum lru_list lru
)
838 struct mem_cgroup_per_zone
*mz
;
839 struct page_cgroup
*pc
;
841 if (mem_cgroup_disabled())
844 pc
= lookup_page_cgroup(page
);
845 /* unused or root page is not rotated. */
846 if (!PageCgroupUsed(pc
))
848 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
850 if (mem_cgroup_is_root(pc
->mem_cgroup
))
852 mz
= page_cgroup_zoneinfo(pc
->mem_cgroup
, page
);
853 list_move(&pc
->lru
, &mz
->lists
[lru
]);
856 void mem_cgroup_add_lru_list(struct page
*page
, enum lru_list lru
)
858 struct page_cgroup
*pc
;
859 struct mem_cgroup_per_zone
*mz
;
861 if (mem_cgroup_disabled())
863 pc
= lookup_page_cgroup(page
);
864 VM_BUG_ON(PageCgroupAcctLRU(pc
));
865 if (!PageCgroupUsed(pc
))
867 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
869 mz
= page_cgroup_zoneinfo(pc
->mem_cgroup
, page
);
870 /* huge page split is done under lru_lock. so, we have no races. */
871 MEM_CGROUP_ZSTAT(mz
, lru
) += 1 << compound_order(page
);
872 SetPageCgroupAcctLRU(pc
);
873 if (mem_cgroup_is_root(pc
->mem_cgroup
))
875 list_add(&pc
->lru
, &mz
->lists
[lru
]);
879 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
880 * lru because the page may.be reused after it's fully uncharged (because of
881 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
882 * it again. This function is only used to charge SwapCache. It's done under
883 * lock_page and expected that zone->lru_lock is never held.
885 static void mem_cgroup_lru_del_before_commit_swapcache(struct page
*page
)
888 struct zone
*zone
= page_zone(page
);
889 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
891 spin_lock_irqsave(&zone
->lru_lock
, flags
);
893 * Forget old LRU when this page_cgroup is *not* used. This Used bit
894 * is guarded by lock_page() because the page is SwapCache.
896 if (!PageCgroupUsed(pc
))
897 mem_cgroup_del_lru_list(page
, page_lru(page
));
898 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
901 static void mem_cgroup_lru_add_after_commit_swapcache(struct page
*page
)
904 struct zone
*zone
= page_zone(page
);
905 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
907 spin_lock_irqsave(&zone
->lru_lock
, flags
);
908 /* link when the page is linked to LRU but page_cgroup isn't */
909 if (PageLRU(page
) && !PageCgroupAcctLRU(pc
))
910 mem_cgroup_add_lru_list(page
, page_lru(page
));
911 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
915 void mem_cgroup_move_lists(struct page
*page
,
916 enum lru_list from
, enum lru_list to
)
918 if (mem_cgroup_disabled())
920 mem_cgroup_del_lru_list(page
, from
);
921 mem_cgroup_add_lru_list(page
, to
);
924 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
927 struct mem_cgroup
*curr
= NULL
;
928 struct task_struct
*p
;
930 p
= find_lock_task_mm(task
);
933 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
938 * We should check use_hierarchy of "mem" not "curr". Because checking
939 * use_hierarchy of "curr" here make this function true if hierarchy is
940 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
941 * hierarchy(even if use_hierarchy is disabled in "mem").
943 if (mem
->use_hierarchy
)
944 ret
= css_is_ancestor(&curr
->css
, &mem
->css
);
951 static int calc_inactive_ratio(struct mem_cgroup
*memcg
, unsigned long *present_pages
)
953 unsigned long active
;
954 unsigned long inactive
;
956 unsigned long inactive_ratio
;
958 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_ANON
);
959 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_ANON
);
961 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
963 inactive_ratio
= int_sqrt(10 * gb
);
968 present_pages
[0] = inactive
;
969 present_pages
[1] = active
;
972 return inactive_ratio
;
975 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup
*memcg
)
977 unsigned long active
;
978 unsigned long inactive
;
979 unsigned long present_pages
[2];
980 unsigned long inactive_ratio
;
982 inactive_ratio
= calc_inactive_ratio(memcg
, present_pages
);
984 inactive
= present_pages
[0];
985 active
= present_pages
[1];
987 if (inactive
* inactive_ratio
< active
)
993 int mem_cgroup_inactive_file_is_low(struct mem_cgroup
*memcg
)
995 unsigned long active
;
996 unsigned long inactive
;
998 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_FILE
);
999 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_FILE
);
1001 return (active
> inactive
);
1004 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup
*memcg
,
1008 int nid
= zone_to_nid(zone
);
1009 int zid
= zone_idx(zone
);
1010 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
1012 return MEM_CGROUP_ZSTAT(mz
, lru
);
1015 struct zone_reclaim_stat
*mem_cgroup_get_reclaim_stat(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 &mz
->reclaim_stat
;
1025 struct zone_reclaim_stat
*
1026 mem_cgroup_get_reclaim_stat_from_page(struct page
*page
)
1028 struct page_cgroup
*pc
;
1029 struct mem_cgroup_per_zone
*mz
;
1031 if (mem_cgroup_disabled())
1034 pc
= lookup_page_cgroup(page
);
1035 if (!PageCgroupUsed(pc
))
1037 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1039 mz
= page_cgroup_zoneinfo(pc
->mem_cgroup
, page
);
1040 return &mz
->reclaim_stat
;
1043 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
1044 struct list_head
*dst
,
1045 unsigned long *scanned
, int order
,
1046 int mode
, struct zone
*z
,
1047 struct mem_cgroup
*mem_cont
,
1048 int active
, int file
)
1050 unsigned long nr_taken
= 0;
1054 struct list_head
*src
;
1055 struct page_cgroup
*pc
, *tmp
;
1056 int nid
= zone_to_nid(z
);
1057 int zid
= zone_idx(z
);
1058 struct mem_cgroup_per_zone
*mz
;
1059 int lru
= LRU_FILE
* file
+ active
;
1063 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
1064 src
= &mz
->lists
[lru
];
1067 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
1068 if (scan
>= nr_to_scan
)
1071 if (unlikely(!PageCgroupUsed(pc
)))
1074 page
= lookup_cgroup_page(pc
);
1076 if (unlikely(!PageLRU(page
)))
1080 ret
= __isolate_lru_page(page
, mode
, file
);
1083 list_move(&page
->lru
, dst
);
1084 mem_cgroup_del_lru(page
);
1085 nr_taken
+= hpage_nr_pages(page
);
1088 /* we don't affect global LRU but rotate in our LRU */
1089 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
1098 trace_mm_vmscan_memcg_isolate(0, nr_to_scan
, scan
, nr_taken
,
1104 #define mem_cgroup_from_res_counter(counter, member) \
1105 container_of(counter, struct mem_cgroup, member)
1108 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1109 * @mem: the memory cgroup
1111 * Returns the maximum amount of memory @mem can be charged with, in
1114 static unsigned long long mem_cgroup_margin(struct mem_cgroup
*mem
)
1116 unsigned long long margin
;
1118 margin
= res_counter_margin(&mem
->res
);
1119 if (do_swap_account
)
1120 margin
= min(margin
, res_counter_margin(&mem
->memsw
));
1124 static unsigned int get_swappiness(struct mem_cgroup
*memcg
)
1126 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1129 if (cgrp
->parent
== NULL
)
1130 return vm_swappiness
;
1132 return memcg
->swappiness
;
1135 static void mem_cgroup_start_move(struct mem_cgroup
*mem
)
1140 spin_lock(&mem
->pcp_counter_lock
);
1141 for_each_online_cpu(cpu
)
1142 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) += 1;
1143 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] += 1;
1144 spin_unlock(&mem
->pcp_counter_lock
);
1150 static void mem_cgroup_end_move(struct mem_cgroup
*mem
)
1157 spin_lock(&mem
->pcp_counter_lock
);
1158 for_each_online_cpu(cpu
)
1159 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) -= 1;
1160 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] -= 1;
1161 spin_unlock(&mem
->pcp_counter_lock
);
1165 * 2 routines for checking "mem" is under move_account() or not.
1167 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1168 * for avoiding race in accounting. If true,
1169 * pc->mem_cgroup may be overwritten.
1171 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1172 * under hierarchy of moving cgroups. This is for
1173 * waiting at hith-memory prressure caused by "move".
1176 static bool mem_cgroup_stealed(struct mem_cgroup
*mem
)
1178 VM_BUG_ON(!rcu_read_lock_held());
1179 return this_cpu_read(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
]) > 0;
1182 static bool mem_cgroup_under_move(struct mem_cgroup
*mem
)
1184 struct mem_cgroup
*from
;
1185 struct mem_cgroup
*to
;
1188 * Unlike task_move routines, we access mc.to, mc.from not under
1189 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1191 spin_lock(&mc
.lock
);
1196 if (from
== mem
|| to
== mem
1197 || (mem
->use_hierarchy
&& css_is_ancestor(&from
->css
, &mem
->css
))
1198 || (mem
->use_hierarchy
&& css_is_ancestor(&to
->css
, &mem
->css
)))
1201 spin_unlock(&mc
.lock
);
1205 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*mem
)
1207 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1208 if (mem_cgroup_under_move(mem
)) {
1210 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1211 /* moving charge context might have finished. */
1214 finish_wait(&mc
.waitq
, &wait
);
1222 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1223 * @memcg: The memory cgroup that went over limit
1224 * @p: Task that is going to be killed
1226 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1229 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1231 struct cgroup
*task_cgrp
;
1232 struct cgroup
*mem_cgrp
;
1234 * Need a buffer in BSS, can't rely on allocations. The code relies
1235 * on the assumption that OOM is serialized for memory controller.
1236 * If this assumption is broken, revisit this code.
1238 static char memcg_name
[PATH_MAX
];
1247 mem_cgrp
= memcg
->css
.cgroup
;
1248 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1250 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1253 * Unfortunately, we are unable to convert to a useful name
1254 * But we'll still print out the usage information
1261 printk(KERN_INFO
"Task in %s killed", memcg_name
);
1264 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1272 * Continues from above, so we don't need an KERN_ level
1274 printk(KERN_CONT
" as a result of limit of %s\n", memcg_name
);
1277 printk(KERN_INFO
"memory: usage %llukB, limit %llukB, failcnt %llu\n",
1278 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1279 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1280 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1281 printk(KERN_INFO
"memory+swap: usage %llukB, limit %llukB, "
1283 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1284 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1285 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1289 * This function returns the number of memcg under hierarchy tree. Returns
1290 * 1(self count) if no children.
1292 static int mem_cgroup_count_children(struct mem_cgroup
*mem
)
1295 struct mem_cgroup
*iter
;
1297 for_each_mem_cgroup_tree(iter
, mem
)
1303 * Return the memory (and swap, if configured) limit for a memcg.
1305 u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1310 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1311 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1313 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1315 * If memsw is finite and limits the amount of swap space available
1316 * to this memcg, return that limit.
1318 return min(limit
, memsw
);
1322 * Visit the first child (need not be the first child as per the ordering
1323 * of the cgroup list, since we track last_scanned_child) of @mem and use
1324 * that to reclaim free pages from.
1326 static struct mem_cgroup
*
1327 mem_cgroup_select_victim(struct mem_cgroup
*root_mem
)
1329 struct mem_cgroup
*ret
= NULL
;
1330 struct cgroup_subsys_state
*css
;
1333 if (!root_mem
->use_hierarchy
) {
1334 css_get(&root_mem
->css
);
1340 nextid
= root_mem
->last_scanned_child
+ 1;
1341 css
= css_get_next(&mem_cgroup_subsys
, nextid
, &root_mem
->css
,
1343 if (css
&& css_tryget(css
))
1344 ret
= container_of(css
, struct mem_cgroup
, css
);
1347 /* Updates scanning parameter */
1349 /* this means start scan from ID:1 */
1350 root_mem
->last_scanned_child
= 0;
1352 root_mem
->last_scanned_child
= found
;
1359 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1360 * we reclaimed from, so that we don't end up penalizing one child extensively
1361 * based on its position in the children list.
1363 * root_mem is the original ancestor that we've been reclaim from.
1365 * We give up and return to the caller when we visit root_mem twice.
1366 * (other groups can be removed while we're walking....)
1368 * If shrink==true, for avoiding to free too much, this returns immedieately.
1370 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup
*root_mem
,
1373 unsigned long reclaim_options
)
1375 struct mem_cgroup
*victim
;
1378 bool noswap
= reclaim_options
& MEM_CGROUP_RECLAIM_NOSWAP
;
1379 bool shrink
= reclaim_options
& MEM_CGROUP_RECLAIM_SHRINK
;
1380 bool check_soft
= reclaim_options
& MEM_CGROUP_RECLAIM_SOFT
;
1381 unsigned long excess
;
1383 excess
= res_counter_soft_limit_excess(&root_mem
->res
) >> PAGE_SHIFT
;
1385 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1386 if (root_mem
->memsw_is_minimum
)
1390 victim
= mem_cgroup_select_victim(root_mem
);
1391 if (victim
== root_mem
) {
1394 drain_all_stock_async();
1397 * If we have not been able to reclaim
1398 * anything, it might because there are
1399 * no reclaimable pages under this hierarchy
1401 if (!check_soft
|| !total
) {
1402 css_put(&victim
->css
);
1406 * We want to do more targetted reclaim.
1407 * excess >> 2 is not to excessive so as to
1408 * reclaim too much, nor too less that we keep
1409 * coming back to reclaim from this cgroup
1411 if (total
>= (excess
>> 2) ||
1412 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
)) {
1413 css_put(&victim
->css
);
1418 if (!mem_cgroup_local_usage(victim
)) {
1419 /* this cgroup's local usage == 0 */
1420 css_put(&victim
->css
);
1423 /* we use swappiness of local cgroup */
1425 ret
= mem_cgroup_shrink_node_zone(victim
, gfp_mask
,
1426 noswap
, get_swappiness(victim
), zone
);
1428 ret
= try_to_free_mem_cgroup_pages(victim
, gfp_mask
,
1429 noswap
, get_swappiness(victim
));
1430 css_put(&victim
->css
);
1432 * At shrinking usage, we can't check we should stop here or
1433 * reclaim more. It's depends on callers. last_scanned_child
1434 * will work enough for keeping fairness under tree.
1440 if (!res_counter_soft_limit_excess(&root_mem
->res
))
1442 } else if (mem_cgroup_margin(root_mem
))
1449 * Check OOM-Killer is already running under our hierarchy.
1450 * If someone is running, return false.
1452 static bool mem_cgroup_oom_lock(struct mem_cgroup
*mem
)
1454 int x
, lock_count
= 0;
1455 struct mem_cgroup
*iter
;
1457 for_each_mem_cgroup_tree(iter
, mem
) {
1458 x
= atomic_inc_return(&iter
->oom_lock
);
1459 lock_count
= max(x
, lock_count
);
1462 if (lock_count
== 1)
1467 static int mem_cgroup_oom_unlock(struct mem_cgroup
*mem
)
1469 struct mem_cgroup
*iter
;
1472 * When a new child is created while the hierarchy is under oom,
1473 * mem_cgroup_oom_lock() may not be called. We have to use
1474 * atomic_add_unless() here.
1476 for_each_mem_cgroup_tree(iter
, mem
)
1477 atomic_add_unless(&iter
->oom_lock
, -1, 0);
1482 static DEFINE_MUTEX(memcg_oom_mutex
);
1483 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1485 struct oom_wait_info
{
1486 struct mem_cgroup
*mem
;
1490 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1491 unsigned mode
, int sync
, void *arg
)
1493 struct mem_cgroup
*wake_mem
= (struct mem_cgroup
*)arg
;
1494 struct oom_wait_info
*oom_wait_info
;
1496 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1498 if (oom_wait_info
->mem
== wake_mem
)
1500 /* if no hierarchy, no match */
1501 if (!oom_wait_info
->mem
->use_hierarchy
|| !wake_mem
->use_hierarchy
)
1504 * Both of oom_wait_info->mem and wake_mem are stable under us.
1505 * Then we can use css_is_ancestor without taking care of RCU.
1507 if (!css_is_ancestor(&oom_wait_info
->mem
->css
, &wake_mem
->css
) &&
1508 !css_is_ancestor(&wake_mem
->css
, &oom_wait_info
->mem
->css
))
1512 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1515 static void memcg_wakeup_oom(struct mem_cgroup
*mem
)
1517 /* for filtering, pass "mem" as argument. */
1518 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, mem
);
1521 static void memcg_oom_recover(struct mem_cgroup
*mem
)
1523 if (mem
&& atomic_read(&mem
->oom_lock
))
1524 memcg_wakeup_oom(mem
);
1528 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1530 bool mem_cgroup_handle_oom(struct mem_cgroup
*mem
, gfp_t mask
)
1532 struct oom_wait_info owait
;
1533 bool locked
, need_to_kill
;
1536 owait
.wait
.flags
= 0;
1537 owait
.wait
.func
= memcg_oom_wake_function
;
1538 owait
.wait
.private = current
;
1539 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1540 need_to_kill
= true;
1541 /* At first, try to OOM lock hierarchy under mem.*/
1542 mutex_lock(&memcg_oom_mutex
);
1543 locked
= mem_cgroup_oom_lock(mem
);
1545 * Even if signal_pending(), we can't quit charge() loop without
1546 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1547 * under OOM is always welcomed, use TASK_KILLABLE here.
1549 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1550 if (!locked
|| mem
->oom_kill_disable
)
1551 need_to_kill
= false;
1553 mem_cgroup_oom_notify(mem
);
1554 mutex_unlock(&memcg_oom_mutex
);
1557 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1558 mem_cgroup_out_of_memory(mem
, mask
);
1561 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1563 mutex_lock(&memcg_oom_mutex
);
1564 mem_cgroup_oom_unlock(mem
);
1565 memcg_wakeup_oom(mem
);
1566 mutex_unlock(&memcg_oom_mutex
);
1568 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
1570 /* Give chance to dying process */
1571 schedule_timeout(1);
1576 * Currently used to update mapped file statistics, but the routine can be
1577 * generalized to update other statistics as well.
1579 * Notes: Race condition
1581 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1582 * it tends to be costly. But considering some conditions, we doesn't need
1583 * to do so _always_.
1585 * Considering "charge", lock_page_cgroup() is not required because all
1586 * file-stat operations happen after a page is attached to radix-tree. There
1587 * are no race with "charge".
1589 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1590 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1591 * if there are race with "uncharge". Statistics itself is properly handled
1594 * Considering "move", this is an only case we see a race. To make the race
1595 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1596 * possibility of race condition. If there is, we take a lock.
1599 void mem_cgroup_update_page_stat(struct page
*page
,
1600 enum mem_cgroup_page_stat_item idx
, int val
)
1602 struct mem_cgroup
*mem
;
1603 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
1604 bool need_unlock
= false;
1605 unsigned long uninitialized_var(flags
);
1611 mem
= pc
->mem_cgroup
;
1612 if (unlikely(!mem
|| !PageCgroupUsed(pc
)))
1614 /* pc->mem_cgroup is unstable ? */
1615 if (unlikely(mem_cgroup_stealed(mem
)) || PageTransHuge(page
)) {
1616 /* take a lock against to access pc->mem_cgroup */
1617 move_lock_page_cgroup(pc
, &flags
);
1619 mem
= pc
->mem_cgroup
;
1620 if (!mem
|| !PageCgroupUsed(pc
))
1625 case MEMCG_NR_FILE_MAPPED
:
1627 SetPageCgroupFileMapped(pc
);
1628 else if (!page_mapped(page
))
1629 ClearPageCgroupFileMapped(pc
);
1630 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
1636 this_cpu_add(mem
->stat
->count
[idx
], val
);
1639 if (unlikely(need_unlock
))
1640 move_unlock_page_cgroup(pc
, &flags
);
1644 EXPORT_SYMBOL(mem_cgroup_update_page_stat
);
1647 * size of first charge trial. "32" comes from vmscan.c's magic value.
1648 * TODO: maybe necessary to use big numbers in big irons.
1650 #define CHARGE_SIZE (32 * PAGE_SIZE)
1651 struct memcg_stock_pcp
{
1652 struct mem_cgroup
*cached
; /* this never be root cgroup */
1654 struct work_struct work
;
1656 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1657 static atomic_t memcg_drain_count
;
1660 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1661 * from local stock and true is returned. If the stock is 0 or charges from a
1662 * cgroup which is not current target, returns false. This stock will be
1665 static bool consume_stock(struct mem_cgroup
*mem
)
1667 struct memcg_stock_pcp
*stock
;
1670 stock
= &get_cpu_var(memcg_stock
);
1671 if (mem
== stock
->cached
&& stock
->charge
)
1672 stock
->charge
-= PAGE_SIZE
;
1673 else /* need to call res_counter_charge */
1675 put_cpu_var(memcg_stock
);
1680 * Returns stocks cached in percpu to res_counter and reset cached information.
1682 static void drain_stock(struct memcg_stock_pcp
*stock
)
1684 struct mem_cgroup
*old
= stock
->cached
;
1686 if (stock
->charge
) {
1687 res_counter_uncharge(&old
->res
, stock
->charge
);
1688 if (do_swap_account
)
1689 res_counter_uncharge(&old
->memsw
, stock
->charge
);
1691 stock
->cached
= NULL
;
1696 * This must be called under preempt disabled or must be called by
1697 * a thread which is pinned to local cpu.
1699 static void drain_local_stock(struct work_struct
*dummy
)
1701 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
1706 * Cache charges(val) which is from res_counter, to local per_cpu area.
1707 * This will be consumed by consume_stock() function, later.
1709 static void refill_stock(struct mem_cgroup
*mem
, int val
)
1711 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1713 if (stock
->cached
!= mem
) { /* reset if necessary */
1715 stock
->cached
= mem
;
1717 stock
->charge
+= val
;
1718 put_cpu_var(memcg_stock
);
1722 * Tries to drain stocked charges in other cpus. This function is asynchronous
1723 * and just put a work per cpu for draining localy on each cpu. Caller can
1724 * expects some charges will be back to res_counter later but cannot wait for
1727 static void drain_all_stock_async(void)
1730 /* This function is for scheduling "drain" in asynchronous way.
1731 * The result of "drain" is not directly handled by callers. Then,
1732 * if someone is calling drain, we don't have to call drain more.
1733 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1734 * there is a race. We just do loose check here.
1736 if (atomic_read(&memcg_drain_count
))
1738 /* Notify other cpus that system-wide "drain" is running */
1739 atomic_inc(&memcg_drain_count
);
1741 for_each_online_cpu(cpu
) {
1742 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1743 schedule_work_on(cpu
, &stock
->work
);
1746 atomic_dec(&memcg_drain_count
);
1747 /* We don't wait for flush_work */
1750 /* This is a synchronous drain interface. */
1751 static void drain_all_stock_sync(void)
1753 /* called when force_empty is called */
1754 atomic_inc(&memcg_drain_count
);
1755 schedule_on_each_cpu(drain_local_stock
);
1756 atomic_dec(&memcg_drain_count
);
1760 * This function drains percpu counter value from DEAD cpu and
1761 * move it to local cpu. Note that this function can be preempted.
1763 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*mem
, int cpu
)
1767 spin_lock(&mem
->pcp_counter_lock
);
1768 for (i
= 0; i
< MEM_CGROUP_STAT_DATA
; i
++) {
1769 s64 x
= per_cpu(mem
->stat
->count
[i
], cpu
);
1771 per_cpu(mem
->stat
->count
[i
], cpu
) = 0;
1772 mem
->nocpu_base
.count
[i
] += x
;
1774 /* need to clear ON_MOVE value, works as a kind of lock. */
1775 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) = 0;
1776 spin_unlock(&mem
->pcp_counter_lock
);
1779 static void synchronize_mem_cgroup_on_move(struct mem_cgroup
*mem
, int cpu
)
1781 int idx
= MEM_CGROUP_ON_MOVE
;
1783 spin_lock(&mem
->pcp_counter_lock
);
1784 per_cpu(mem
->stat
->count
[idx
], cpu
) = mem
->nocpu_base
.count
[idx
];
1785 spin_unlock(&mem
->pcp_counter_lock
);
1788 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1789 unsigned long action
,
1792 int cpu
= (unsigned long)hcpu
;
1793 struct memcg_stock_pcp
*stock
;
1794 struct mem_cgroup
*iter
;
1796 if ((action
== CPU_ONLINE
)) {
1797 for_each_mem_cgroup_all(iter
)
1798 synchronize_mem_cgroup_on_move(iter
, cpu
);
1802 if ((action
!= CPU_DEAD
) || action
!= CPU_DEAD_FROZEN
)
1805 for_each_mem_cgroup_all(iter
)
1806 mem_cgroup_drain_pcp_counter(iter
, cpu
);
1808 stock
= &per_cpu(memcg_stock
, cpu
);
1814 /* See __mem_cgroup_try_charge() for details */
1816 CHARGE_OK
, /* success */
1817 CHARGE_RETRY
, /* need to retry but retry is not bad */
1818 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
1819 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
1820 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
1823 static int __mem_cgroup_do_charge(struct mem_cgroup
*mem
, gfp_t gfp_mask
,
1824 int csize
, bool oom_check
)
1826 struct mem_cgroup
*mem_over_limit
;
1827 struct res_counter
*fail_res
;
1828 unsigned long flags
= 0;
1831 ret
= res_counter_charge(&mem
->res
, csize
, &fail_res
);
1834 if (!do_swap_account
)
1836 ret
= res_counter_charge(&mem
->memsw
, csize
, &fail_res
);
1840 res_counter_uncharge(&mem
->res
, csize
);
1841 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
1842 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
1844 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
1846 * csize can be either a huge page (HPAGE_SIZE), a batch of
1847 * regular pages (CHARGE_SIZE), or a single regular page
1850 * Never reclaim on behalf of optional batching, retry with a
1851 * single page instead.
1853 if (csize
== CHARGE_SIZE
)
1854 return CHARGE_RETRY
;
1856 if (!(gfp_mask
& __GFP_WAIT
))
1857 return CHARGE_WOULDBLOCK
;
1859 ret
= mem_cgroup_hierarchical_reclaim(mem_over_limit
, NULL
,
1861 if (mem_cgroup_margin(mem_over_limit
) >= csize
)
1862 return CHARGE_RETRY
;
1864 * Even though the limit is exceeded at this point, reclaim
1865 * may have been able to free some pages. Retry the charge
1866 * before killing the task.
1868 * Only for regular pages, though: huge pages are rather
1869 * unlikely to succeed so close to the limit, and we fall back
1870 * to regular pages anyway in case of failure.
1872 if (csize
== PAGE_SIZE
&& ret
)
1873 return CHARGE_RETRY
;
1876 * At task move, charge accounts can be doubly counted. So, it's
1877 * better to wait until the end of task_move if something is going on.
1879 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1880 return CHARGE_RETRY
;
1882 /* If we don't need to call oom-killer at el, return immediately */
1884 return CHARGE_NOMEM
;
1886 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
))
1887 return CHARGE_OOM_DIE
;
1889 return CHARGE_RETRY
;
1893 * Unlike exported interface, "oom" parameter is added. if oom==true,
1894 * oom-killer can be invoked.
1896 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
1898 struct mem_cgroup
**memcg
, bool oom
,
1901 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1902 struct mem_cgroup
*mem
= NULL
;
1904 int csize
= max(CHARGE_SIZE
, (unsigned long) page_size
);
1907 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1908 * in system level. So, allow to go ahead dying process in addition to
1911 if (unlikely(test_thread_flag(TIF_MEMDIE
)
1912 || fatal_signal_pending(current
)))
1916 * We always charge the cgroup the mm_struct belongs to.
1917 * The mm_struct's mem_cgroup changes on task migration if the
1918 * thread group leader migrates. It's possible that mm is not
1919 * set, if so charge the init_mm (happens for pagecache usage).
1924 if (*memcg
) { /* css should be a valid one */
1926 VM_BUG_ON(css_is_removed(&mem
->css
));
1927 if (mem_cgroup_is_root(mem
))
1929 if (page_size
== PAGE_SIZE
&& consume_stock(mem
))
1933 struct task_struct
*p
;
1936 p
= rcu_dereference(mm
->owner
);
1938 * Because we don't have task_lock(), "p" can exit.
1939 * In that case, "mem" can point to root or p can be NULL with
1940 * race with swapoff. Then, we have small risk of mis-accouning.
1941 * But such kind of mis-account by race always happens because
1942 * we don't have cgroup_mutex(). It's overkill and we allo that
1944 * (*) swapoff at el will charge against mm-struct not against
1945 * task-struct. So, mm->owner can be NULL.
1947 mem
= mem_cgroup_from_task(p
);
1948 if (!mem
|| mem_cgroup_is_root(mem
)) {
1952 if (page_size
== PAGE_SIZE
&& consume_stock(mem
)) {
1954 * It seems dagerous to access memcg without css_get().
1955 * But considering how consume_stok works, it's not
1956 * necessary. If consume_stock success, some charges
1957 * from this memcg are cached on this cpu. So, we
1958 * don't need to call css_get()/css_tryget() before
1959 * calling consume_stock().
1964 /* after here, we may be blocked. we need to get refcnt */
1965 if (!css_tryget(&mem
->css
)) {
1975 /* If killed, bypass charge */
1976 if (fatal_signal_pending(current
)) {
1982 if (oom
&& !nr_oom_retries
) {
1984 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1987 ret
= __mem_cgroup_do_charge(mem
, gfp_mask
, csize
, oom_check
);
1992 case CHARGE_RETRY
: /* not in OOM situation but retry */
1997 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
2000 case CHARGE_NOMEM
: /* OOM routine works */
2005 /* If oom, we never return -ENOMEM */
2008 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2012 } while (ret
!= CHARGE_OK
);
2014 if (csize
> page_size
)
2015 refill_stock(mem
, csize
- page_size
);
2029 * Somemtimes we have to undo a charge we got by try_charge().
2030 * This function is for that and do uncharge, put css's refcnt.
2031 * gotten by try_charge().
2033 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2034 unsigned long count
)
2036 if (!mem_cgroup_is_root(mem
)) {
2037 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
2038 if (do_swap_account
)
2039 res_counter_uncharge(&mem
->memsw
, PAGE_SIZE
* count
);
2043 static void mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2046 __mem_cgroup_cancel_charge(mem
, page_size
>> PAGE_SHIFT
);
2050 * A helper function to get mem_cgroup from ID. must be called under
2051 * rcu_read_lock(). The caller must check css_is_removed() or some if
2052 * it's concern. (dropping refcnt from swap can be called against removed
2055 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2057 struct cgroup_subsys_state
*css
;
2059 /* ID 0 is unused ID */
2062 css
= css_lookup(&mem_cgroup_subsys
, id
);
2065 return container_of(css
, struct mem_cgroup
, css
);
2068 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2070 struct mem_cgroup
*mem
= NULL
;
2071 struct page_cgroup
*pc
;
2075 VM_BUG_ON(!PageLocked(page
));
2077 pc
= lookup_page_cgroup(page
);
2078 lock_page_cgroup(pc
);
2079 if (PageCgroupUsed(pc
)) {
2080 mem
= pc
->mem_cgroup
;
2081 if (mem
&& !css_tryget(&mem
->css
))
2083 } else if (PageSwapCache(page
)) {
2084 ent
.val
= page_private(page
);
2085 id
= lookup_swap_cgroup(ent
);
2087 mem
= mem_cgroup_lookup(id
);
2088 if (mem
&& !css_tryget(&mem
->css
))
2092 unlock_page_cgroup(pc
);
2096 static void __mem_cgroup_commit_charge(struct mem_cgroup
*mem
,
2098 struct page_cgroup
*pc
,
2099 enum charge_type ctype
,
2102 int nr_pages
= page_size
>> PAGE_SHIFT
;
2104 lock_page_cgroup(pc
);
2105 if (unlikely(PageCgroupUsed(pc
))) {
2106 unlock_page_cgroup(pc
);
2107 mem_cgroup_cancel_charge(mem
, page_size
);
2111 * we don't need page_cgroup_lock about tail pages, becase they are not
2112 * accessed by any other context at this point.
2114 pc
->mem_cgroup
= mem
;
2116 * We access a page_cgroup asynchronously without lock_page_cgroup().
2117 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2118 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2119 * before USED bit, we need memory barrier here.
2120 * See mem_cgroup_add_lru_list(), etc.
2124 case MEM_CGROUP_CHARGE_TYPE_CACHE
:
2125 case MEM_CGROUP_CHARGE_TYPE_SHMEM
:
2126 SetPageCgroupCache(pc
);
2127 SetPageCgroupUsed(pc
);
2129 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2130 ClearPageCgroupCache(pc
);
2131 SetPageCgroupUsed(pc
);
2137 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), nr_pages
);
2138 unlock_page_cgroup(pc
);
2140 * "charge_statistics" updated event counter. Then, check it.
2141 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2142 * if they exceeds softlimit.
2144 memcg_check_events(mem
, page
);
2147 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2149 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2150 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2152 * Because tail pages are not marked as "used", set it. We're under
2153 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2155 void mem_cgroup_split_huge_fixup(struct page
*head
, struct page
*tail
)
2157 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
2158 struct page_cgroup
*tail_pc
= lookup_page_cgroup(tail
);
2159 unsigned long flags
;
2161 if (mem_cgroup_disabled())
2164 * We have no races with charge/uncharge but will have races with
2165 * page state accounting.
2167 move_lock_page_cgroup(head_pc
, &flags
);
2169 tail_pc
->mem_cgroup
= head_pc
->mem_cgroup
;
2170 smp_wmb(); /* see __commit_charge() */
2171 if (PageCgroupAcctLRU(head_pc
)) {
2173 struct mem_cgroup_per_zone
*mz
;
2176 * LRU flags cannot be copied because we need to add tail
2177 *.page to LRU by generic call and our hook will be called.
2178 * We hold lru_lock, then, reduce counter directly.
2180 lru
= page_lru(head
);
2181 mz
= page_cgroup_zoneinfo(head_pc
->mem_cgroup
, head
);
2182 MEM_CGROUP_ZSTAT(mz
, lru
) -= 1;
2184 tail_pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
2185 move_unlock_page_cgroup(head_pc
, &flags
);
2190 * mem_cgroup_move_account - move account of the page
2192 * @pc: page_cgroup of the page.
2193 * @from: mem_cgroup which the page is moved from.
2194 * @to: mem_cgroup which the page is moved to. @from != @to.
2195 * @uncharge: whether we should call uncharge and css_put against @from.
2196 * @charge_size: number of bytes to charge (regular or huge page)
2198 * The caller must confirm following.
2199 * - page is not on LRU (isolate_page() is useful.)
2200 * - compound_lock is held when charge_size > PAGE_SIZE
2202 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2203 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2204 * true, this function does "uncharge" from old cgroup, but it doesn't if
2205 * @uncharge is false, so a caller should do "uncharge".
2207 static int mem_cgroup_move_account(struct page
*page
, struct page_cgroup
*pc
,
2208 struct mem_cgroup
*from
, struct mem_cgroup
*to
,
2209 bool uncharge
, int charge_size
)
2211 int nr_pages
= charge_size
>> PAGE_SHIFT
;
2212 unsigned long flags
;
2215 VM_BUG_ON(from
== to
);
2216 VM_BUG_ON(PageLRU(page
));
2218 * The page is isolated from LRU. So, collapse function
2219 * will not handle this page. But page splitting can happen.
2220 * Do this check under compound_page_lock(). The caller should
2224 if (charge_size
> PAGE_SIZE
&& !PageTransHuge(page
))
2227 lock_page_cgroup(pc
);
2230 if (!PageCgroupUsed(pc
) || pc
->mem_cgroup
!= from
)
2233 move_lock_page_cgroup(pc
, &flags
);
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
2257 move_unlock_page_cgroup(pc
, &flags
);
2260 unlock_page_cgroup(pc
);
2264 memcg_check_events(to
, page
);
2265 memcg_check_events(from
, page
);
2271 * move charges to its parent.
2274 static int mem_cgroup_move_parent(struct page
*page
,
2275 struct page_cgroup
*pc
,
2276 struct mem_cgroup
*child
,
2279 struct cgroup
*cg
= child
->css
.cgroup
;
2280 struct cgroup
*pcg
= cg
->parent
;
2281 struct mem_cgroup
*parent
;
2282 int page_size
= PAGE_SIZE
;
2283 unsigned long flags
;
2291 if (!get_page_unless_zero(page
))
2293 if (isolate_lru_page(page
))
2296 if (PageTransHuge(page
))
2297 page_size
= HPAGE_SIZE
;
2299 parent
= mem_cgroup_from_cont(pcg
);
2300 ret
= __mem_cgroup_try_charge(NULL
, gfp_mask
,
2301 &parent
, false, page_size
);
2305 if (page_size
> PAGE_SIZE
)
2306 flags
= compound_lock_irqsave(page
);
2308 ret
= mem_cgroup_move_account(page
, pc
, child
, parent
, true, page_size
);
2310 mem_cgroup_cancel_charge(parent
, page_size
);
2312 if (page_size
> PAGE_SIZE
)
2313 compound_unlock_irqrestore(page
, flags
);
2315 putback_lru_page(page
);
2323 * Charge the memory controller for page usage.
2325 * 0 if the charge was successful
2326 * < 0 if the cgroup is over its limit
2328 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
2329 gfp_t gfp_mask
, enum charge_type ctype
)
2331 struct mem_cgroup
*mem
= NULL
;
2332 int page_size
= PAGE_SIZE
;
2333 struct page_cgroup
*pc
;
2337 if (PageTransHuge(page
)) {
2338 page_size
<<= compound_order(page
);
2339 VM_BUG_ON(!PageTransHuge(page
));
2341 * Never OOM-kill a process for a huge page. The
2342 * fault handler will fall back to regular pages.
2347 pc
= lookup_page_cgroup(page
);
2348 BUG_ON(!pc
); /* XXX: remove this and move pc lookup into commit */
2350 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, &mem
, oom
, page_size
);
2354 __mem_cgroup_commit_charge(mem
, page
, pc
, ctype
, page_size
);
2358 int mem_cgroup_newpage_charge(struct page
*page
,
2359 struct mm_struct
*mm
, gfp_t gfp_mask
)
2361 if (mem_cgroup_disabled())
2364 * If already mapped, we don't have to account.
2365 * If page cache, page->mapping has address_space.
2366 * But page->mapping may have out-of-use anon_vma pointer,
2367 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2370 if (page_mapped(page
) || (page
->mapping
&& !PageAnon(page
)))
2374 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2375 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2379 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2380 enum charge_type ctype
);
2382 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
2387 if (mem_cgroup_disabled())
2389 if (PageCompound(page
))
2392 * Corner case handling. This is called from add_to_page_cache()
2393 * in usual. But some FS (shmem) precharges this page before calling it
2394 * and call add_to_page_cache() with GFP_NOWAIT.
2396 * For GFP_NOWAIT case, the page may be pre-charged before calling
2397 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2398 * charge twice. (It works but has to pay a bit larger cost.)
2399 * And when the page is SwapCache, it should take swap information
2400 * into account. This is under lock_page() now.
2402 if (!(gfp_mask
& __GFP_WAIT
)) {
2403 struct page_cgroup
*pc
;
2405 pc
= lookup_page_cgroup(page
);
2408 lock_page_cgroup(pc
);
2409 if (PageCgroupUsed(pc
)) {
2410 unlock_page_cgroup(pc
);
2413 unlock_page_cgroup(pc
);
2419 if (page_is_file_cache(page
))
2420 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2421 MEM_CGROUP_CHARGE_TYPE_CACHE
);
2424 if (PageSwapCache(page
)) {
2425 struct mem_cgroup
*mem
;
2427 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
2429 __mem_cgroup_commit_charge_swapin(page
, mem
,
2430 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2432 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2433 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2439 * While swap-in, try_charge -> commit or cancel, the page is locked.
2440 * And when try_charge() successfully returns, one refcnt to memcg without
2441 * struct page_cgroup is acquired. This refcnt will be consumed by
2442 * "commit()" or removed by "cancel()"
2444 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
2446 gfp_t mask
, struct mem_cgroup
**ptr
)
2448 struct mem_cgroup
*mem
;
2453 if (mem_cgroup_disabled())
2456 if (!do_swap_account
)
2459 * A racing thread's fault, or swapoff, may have already updated
2460 * the pte, and even removed page from swap cache: in those cases
2461 * do_swap_page()'s pte_same() test will fail; but there's also a
2462 * KSM case which does need to charge the page.
2464 if (!PageSwapCache(page
))
2466 mem
= try_get_mem_cgroup_from_page(page
);
2470 ret
= __mem_cgroup_try_charge(NULL
, mask
, ptr
, true, PAGE_SIZE
);
2476 return __mem_cgroup_try_charge(mm
, mask
, ptr
, true, PAGE_SIZE
);
2480 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2481 enum charge_type ctype
)
2483 struct page_cgroup
*pc
;
2485 if (mem_cgroup_disabled())
2489 cgroup_exclude_rmdir(&ptr
->css
);
2490 pc
= lookup_page_cgroup(page
);
2491 mem_cgroup_lru_del_before_commit_swapcache(page
);
2492 __mem_cgroup_commit_charge(ptr
, page
, pc
, ctype
, PAGE_SIZE
);
2493 mem_cgroup_lru_add_after_commit_swapcache(page
);
2495 * Now swap is on-memory. This means this page may be
2496 * counted both as mem and swap....double count.
2497 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2498 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2499 * may call delete_from_swap_cache() before reach here.
2501 if (do_swap_account
&& PageSwapCache(page
)) {
2502 swp_entry_t ent
= {.val
= page_private(page
)};
2504 struct mem_cgroup
*memcg
;
2506 id
= swap_cgroup_record(ent
, 0);
2508 memcg
= mem_cgroup_lookup(id
);
2511 * This recorded memcg can be obsolete one. So, avoid
2512 * calling css_tryget
2514 if (!mem_cgroup_is_root(memcg
))
2515 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2516 mem_cgroup_swap_statistics(memcg
, false);
2517 mem_cgroup_put(memcg
);
2522 * At swapin, we may charge account against cgroup which has no tasks.
2523 * So, rmdir()->pre_destroy() can be called while we do this charge.
2524 * In that case, we need to call pre_destroy() again. check it here.
2526 cgroup_release_and_wakeup_rmdir(&ptr
->css
);
2529 void mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
)
2531 __mem_cgroup_commit_charge_swapin(page
, ptr
,
2532 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2535 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*mem
)
2537 if (mem_cgroup_disabled())
2541 mem_cgroup_cancel_charge(mem
, PAGE_SIZE
);
2545 __do_uncharge(struct mem_cgroup
*mem
, const enum charge_type ctype
,
2548 struct memcg_batch_info
*batch
= NULL
;
2549 bool uncharge_memsw
= true;
2550 /* If swapout, usage of swap doesn't decrease */
2551 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
2552 uncharge_memsw
= false;
2554 batch
= ¤t
->memcg_batch
;
2556 * In usual, we do css_get() when we remember memcg pointer.
2557 * But in this case, we keep res->usage until end of a series of
2558 * uncharges. Then, it's ok to ignore memcg's refcnt.
2563 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2564 * In those cases, all pages freed continously can be expected to be in
2565 * the same cgroup and we have chance to coalesce uncharges.
2566 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2567 * because we want to do uncharge as soon as possible.
2570 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
2571 goto direct_uncharge
;
2573 if (page_size
!= PAGE_SIZE
)
2574 goto direct_uncharge
;
2577 * In typical case, batch->memcg == mem. This means we can
2578 * merge a series of uncharges to an uncharge of res_counter.
2579 * If not, we uncharge res_counter ony by one.
2581 if (batch
->memcg
!= mem
)
2582 goto direct_uncharge
;
2583 /* remember freed charge and uncharge it later */
2584 batch
->bytes
+= PAGE_SIZE
;
2586 batch
->memsw_bytes
+= PAGE_SIZE
;
2589 res_counter_uncharge(&mem
->res
, page_size
);
2591 res_counter_uncharge(&mem
->memsw
, page_size
);
2592 if (unlikely(batch
->memcg
!= mem
))
2593 memcg_oom_recover(mem
);
2598 * uncharge if !page_mapped(page)
2600 static struct mem_cgroup
*
2601 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
)
2604 struct page_cgroup
*pc
;
2605 struct mem_cgroup
*mem
= NULL
;
2606 int page_size
= PAGE_SIZE
;
2608 if (mem_cgroup_disabled())
2611 if (PageSwapCache(page
))
2614 if (PageTransHuge(page
)) {
2615 page_size
<<= compound_order(page
);
2616 VM_BUG_ON(!PageTransHuge(page
));
2619 count
= page_size
>> PAGE_SHIFT
;
2621 * Check if our page_cgroup is valid
2623 pc
= lookup_page_cgroup(page
);
2624 if (unlikely(!pc
|| !PageCgroupUsed(pc
)))
2627 lock_page_cgroup(pc
);
2629 mem
= pc
->mem_cgroup
;
2631 if (!PageCgroupUsed(pc
))
2635 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2636 case MEM_CGROUP_CHARGE_TYPE_DROP
:
2637 /* See mem_cgroup_prepare_migration() */
2638 if (page_mapped(page
) || PageCgroupMigration(pc
))
2641 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
2642 if (!PageAnon(page
)) { /* Shared memory */
2643 if (page
->mapping
&& !page_is_file_cache(page
))
2645 } else if (page_mapped(page
)) /* Anon */
2652 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), -count
);
2654 ClearPageCgroupUsed(pc
);
2656 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2657 * freed from LRU. This is safe because uncharged page is expected not
2658 * to be reused (freed soon). Exception is SwapCache, it's handled by
2659 * special functions.
2662 unlock_page_cgroup(pc
);
2664 * even after unlock, we have mem->res.usage here and this memcg
2665 * will never be freed.
2667 memcg_check_events(mem
, page
);
2668 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
2669 mem_cgroup_swap_statistics(mem
, true);
2670 mem_cgroup_get(mem
);
2672 if (!mem_cgroup_is_root(mem
))
2673 __do_uncharge(mem
, ctype
, page_size
);
2678 unlock_page_cgroup(pc
);
2682 void mem_cgroup_uncharge_page(struct page
*page
)
2685 if (page_mapped(page
))
2687 if (page
->mapping
&& !PageAnon(page
))
2689 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2692 void mem_cgroup_uncharge_cache_page(struct page
*page
)
2694 VM_BUG_ON(page_mapped(page
));
2695 VM_BUG_ON(page
->mapping
);
2696 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
);
2700 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2701 * In that cases, pages are freed continuously and we can expect pages
2702 * are in the same memcg. All these calls itself limits the number of
2703 * pages freed at once, then uncharge_start/end() is called properly.
2704 * This may be called prural(2) times in a context,
2707 void mem_cgroup_uncharge_start(void)
2709 current
->memcg_batch
.do_batch
++;
2710 /* We can do nest. */
2711 if (current
->memcg_batch
.do_batch
== 1) {
2712 current
->memcg_batch
.memcg
= NULL
;
2713 current
->memcg_batch
.bytes
= 0;
2714 current
->memcg_batch
.memsw_bytes
= 0;
2718 void mem_cgroup_uncharge_end(void)
2720 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
2722 if (!batch
->do_batch
)
2726 if (batch
->do_batch
) /* If stacked, do nothing. */
2732 * This "batch->memcg" is valid without any css_get/put etc...
2733 * bacause we hide charges behind us.
2736 res_counter_uncharge(&batch
->memcg
->res
, batch
->bytes
);
2737 if (batch
->memsw_bytes
)
2738 res_counter_uncharge(&batch
->memcg
->memsw
, batch
->memsw_bytes
);
2739 memcg_oom_recover(batch
->memcg
);
2740 /* forget this pointer (for sanity check) */
2741 batch
->memcg
= NULL
;
2746 * called after __delete_from_swap_cache() and drop "page" account.
2747 * memcg information is recorded to swap_cgroup of "ent"
2750 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
2752 struct mem_cgroup
*memcg
;
2753 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
2755 if (!swapout
) /* this was a swap cache but the swap is unused ! */
2756 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
2758 memcg
= __mem_cgroup_uncharge_common(page
, ctype
);
2761 * record memcg information, if swapout && memcg != NULL,
2762 * mem_cgroup_get() was called in uncharge().
2764 if (do_swap_account
&& swapout
&& memcg
)
2765 swap_cgroup_record(ent
, css_id(&memcg
->css
));
2769 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2771 * called from swap_entry_free(). remove record in swap_cgroup and
2772 * uncharge "memsw" account.
2774 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
2776 struct mem_cgroup
*memcg
;
2779 if (!do_swap_account
)
2782 id
= swap_cgroup_record(ent
, 0);
2784 memcg
= mem_cgroup_lookup(id
);
2787 * We uncharge this because swap is freed.
2788 * This memcg can be obsolete one. We avoid calling css_tryget
2790 if (!mem_cgroup_is_root(memcg
))
2791 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2792 mem_cgroup_swap_statistics(memcg
, false);
2793 mem_cgroup_put(memcg
);
2799 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2800 * @entry: swap entry to be moved
2801 * @from: mem_cgroup which the entry is moved from
2802 * @to: mem_cgroup which the entry is moved to
2803 * @need_fixup: whether we should fixup res_counters and refcounts.
2805 * It succeeds only when the swap_cgroup's record for this entry is the same
2806 * as the mem_cgroup's id of @from.
2808 * Returns 0 on success, -EINVAL on failure.
2810 * The caller must have charged to @to, IOW, called res_counter_charge() about
2811 * both res and memsw, and called css_get().
2813 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2814 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2816 unsigned short old_id
, new_id
;
2818 old_id
= css_id(&from
->css
);
2819 new_id
= css_id(&to
->css
);
2821 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2822 mem_cgroup_swap_statistics(from
, false);
2823 mem_cgroup_swap_statistics(to
, true);
2825 * This function is only called from task migration context now.
2826 * It postpones res_counter and refcount handling till the end
2827 * of task migration(mem_cgroup_clear_mc()) for performance
2828 * improvement. But we cannot postpone mem_cgroup_get(to)
2829 * because if the process that has been moved to @to does
2830 * swap-in, the refcount of @to might be decreased to 0.
2834 if (!mem_cgroup_is_root(from
))
2835 res_counter_uncharge(&from
->memsw
, PAGE_SIZE
);
2836 mem_cgroup_put(from
);
2838 * we charged both to->res and to->memsw, so we should
2841 if (!mem_cgroup_is_root(to
))
2842 res_counter_uncharge(&to
->res
, PAGE_SIZE
);
2849 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2850 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2857 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2860 int mem_cgroup_prepare_migration(struct page
*page
,
2861 struct page
*newpage
, struct mem_cgroup
**ptr
, gfp_t gfp_mask
)
2863 struct page_cgroup
*pc
;
2864 struct mem_cgroup
*mem
= NULL
;
2865 enum charge_type ctype
;
2870 VM_BUG_ON(PageTransHuge(page
));
2871 if (mem_cgroup_disabled())
2874 pc
= lookup_page_cgroup(page
);
2875 lock_page_cgroup(pc
);
2876 if (PageCgroupUsed(pc
)) {
2877 mem
= pc
->mem_cgroup
;
2880 * At migrating an anonymous page, its mapcount goes down
2881 * to 0 and uncharge() will be called. But, even if it's fully
2882 * unmapped, migration may fail and this page has to be
2883 * charged again. We set MIGRATION flag here and delay uncharge
2884 * until end_migration() is called
2886 * Corner Case Thinking
2888 * When the old page was mapped as Anon and it's unmap-and-freed
2889 * while migration was ongoing.
2890 * If unmap finds the old page, uncharge() of it will be delayed
2891 * until end_migration(). If unmap finds a new page, it's
2892 * uncharged when it make mapcount to be 1->0. If unmap code
2893 * finds swap_migration_entry, the new page will not be mapped
2894 * and end_migration() will find it(mapcount==0).
2897 * When the old page was mapped but migraion fails, the kernel
2898 * remaps it. A charge for it is kept by MIGRATION flag even
2899 * if mapcount goes down to 0. We can do remap successfully
2900 * without charging it again.
2903 * The "old" page is under lock_page() until the end of
2904 * migration, so, the old page itself will not be swapped-out.
2905 * If the new page is swapped out before end_migraton, our
2906 * hook to usual swap-out path will catch the event.
2909 SetPageCgroupMigration(pc
);
2911 unlock_page_cgroup(pc
);
2913 * If the page is not charged at this point,
2920 ret
= __mem_cgroup_try_charge(NULL
, gfp_mask
, ptr
, false, PAGE_SIZE
);
2921 css_put(&mem
->css
);/* drop extra refcnt */
2922 if (ret
|| *ptr
== NULL
) {
2923 if (PageAnon(page
)) {
2924 lock_page_cgroup(pc
);
2925 ClearPageCgroupMigration(pc
);
2926 unlock_page_cgroup(pc
);
2928 * The old page may be fully unmapped while we kept it.
2930 mem_cgroup_uncharge_page(page
);
2935 * We charge new page before it's used/mapped. So, even if unlock_page()
2936 * is called before end_migration, we can catch all events on this new
2937 * page. In the case new page is migrated but not remapped, new page's
2938 * mapcount will be finally 0 and we call uncharge in end_migration().
2940 pc
= lookup_page_cgroup(newpage
);
2942 ctype
= MEM_CGROUP_CHARGE_TYPE_MAPPED
;
2943 else if (page_is_file_cache(page
))
2944 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
2946 ctype
= MEM_CGROUP_CHARGE_TYPE_SHMEM
;
2947 __mem_cgroup_commit_charge(mem
, page
, pc
, ctype
, PAGE_SIZE
);
2951 /* remove redundant charge if migration failed*/
2952 void mem_cgroup_end_migration(struct mem_cgroup
*mem
,
2953 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
2955 struct page
*used
, *unused
;
2956 struct page_cgroup
*pc
;
2960 /* blocks rmdir() */
2961 cgroup_exclude_rmdir(&mem
->css
);
2962 if (!migration_ok
) {
2970 * We disallowed uncharge of pages under migration because mapcount
2971 * of the page goes down to zero, temporarly.
2972 * Clear the flag and check the page should be charged.
2974 pc
= lookup_page_cgroup(oldpage
);
2975 lock_page_cgroup(pc
);
2976 ClearPageCgroupMigration(pc
);
2977 unlock_page_cgroup(pc
);
2979 __mem_cgroup_uncharge_common(unused
, MEM_CGROUP_CHARGE_TYPE_FORCE
);
2982 * If a page is a file cache, radix-tree replacement is very atomic
2983 * and we can skip this check. When it was an Anon page, its mapcount
2984 * goes down to 0. But because we added MIGRATION flage, it's not
2985 * uncharged yet. There are several case but page->mapcount check
2986 * and USED bit check in mem_cgroup_uncharge_page() will do enough
2987 * check. (see prepare_charge() also)
2990 mem_cgroup_uncharge_page(used
);
2992 * At migration, we may charge account against cgroup which has no
2994 * So, rmdir()->pre_destroy() can be called while we do this charge.
2995 * In that case, we need to call pre_destroy() again. check it here.
2997 cgroup_release_and_wakeup_rmdir(&mem
->css
);
3001 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3002 * Calling hierarchical_reclaim is not enough because we should update
3003 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3004 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3005 * not from the memcg which this page would be charged to.
3006 * try_charge_swapin does all of these works properly.
3008 int mem_cgroup_shmem_charge_fallback(struct page
*page
,
3009 struct mm_struct
*mm
,
3012 struct mem_cgroup
*mem
;
3015 if (mem_cgroup_disabled())
3018 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
3020 mem_cgroup_cancel_charge_swapin(mem
); /* it does !mem check */
3025 #ifdef CONFIG_DEBUG_VM
3026 static struct page_cgroup
*lookup_page_cgroup_used(struct page
*page
)
3028 struct page_cgroup
*pc
;
3030 pc
= lookup_page_cgroup(page
);
3031 if (likely(pc
) && PageCgroupUsed(pc
))
3036 bool mem_cgroup_bad_page_check(struct page
*page
)
3038 if (mem_cgroup_disabled())
3041 return lookup_page_cgroup_used(page
) != NULL
;
3044 void mem_cgroup_print_bad_page(struct page
*page
)
3046 struct page_cgroup
*pc
;
3048 pc
= lookup_page_cgroup_used(page
);
3053 printk(KERN_ALERT
"pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3054 pc
, pc
->flags
, pc
->mem_cgroup
);
3056 path
= kmalloc(PATH_MAX
, GFP_KERNEL
);
3059 ret
= cgroup_path(pc
->mem_cgroup
->css
.cgroup
,
3064 printk(KERN_CONT
"(%s)\n",
3065 (ret
< 0) ? "cannot get the path" : path
);
3071 static DEFINE_MUTEX(set_limit_mutex
);
3073 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3074 unsigned long long val
)
3077 u64 memswlimit
, memlimit
;
3079 int children
= mem_cgroup_count_children(memcg
);
3080 u64 curusage
, oldusage
;
3084 * For keeping hierarchical_reclaim simple, how long we should retry
3085 * is depends on callers. We set our retry-count to be function
3086 * of # of children which we should visit in this loop.
3088 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
3090 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3093 while (retry_count
) {
3094 if (signal_pending(current
)) {
3099 * Rather than hide all in some function, I do this in
3100 * open coded manner. You see what this really does.
3101 * We have to guarantee mem->res.limit < mem->memsw.limit.
3103 mutex_lock(&set_limit_mutex
);
3104 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3105 if (memswlimit
< val
) {
3107 mutex_unlock(&set_limit_mutex
);
3111 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3115 ret
= res_counter_set_limit(&memcg
->res
, val
);
3117 if (memswlimit
== val
)
3118 memcg
->memsw_is_minimum
= true;
3120 memcg
->memsw_is_minimum
= false;
3122 mutex_unlock(&set_limit_mutex
);
3127 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3128 MEM_CGROUP_RECLAIM_SHRINK
);
3129 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3130 /* Usage is reduced ? */
3131 if (curusage
>= oldusage
)
3134 oldusage
= curusage
;
3136 if (!ret
&& enlarge
)
3137 memcg_oom_recover(memcg
);
3142 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3143 unsigned long long val
)
3146 u64 memlimit
, memswlimit
, oldusage
, curusage
;
3147 int children
= mem_cgroup_count_children(memcg
);
3151 /* see mem_cgroup_resize_res_limit */
3152 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
3153 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3154 while (retry_count
) {
3155 if (signal_pending(current
)) {
3160 * Rather than hide all in some function, I do this in
3161 * open coded manner. You see what this really does.
3162 * We have to guarantee mem->res.limit < mem->memsw.limit.
3164 mutex_lock(&set_limit_mutex
);
3165 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3166 if (memlimit
> val
) {
3168 mutex_unlock(&set_limit_mutex
);
3171 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3172 if (memswlimit
< val
)
3174 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
3176 if (memlimit
== val
)
3177 memcg
->memsw_is_minimum
= true;
3179 memcg
->memsw_is_minimum
= false;
3181 mutex_unlock(&set_limit_mutex
);
3186 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3187 MEM_CGROUP_RECLAIM_NOSWAP
|
3188 MEM_CGROUP_RECLAIM_SHRINK
);
3189 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3190 /* Usage is reduced ? */
3191 if (curusage
>= oldusage
)
3194 oldusage
= curusage
;
3196 if (!ret
&& enlarge
)
3197 memcg_oom_recover(memcg
);
3201 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3204 unsigned long nr_reclaimed
= 0;
3205 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3206 unsigned long reclaimed
;
3208 struct mem_cgroup_tree_per_zone
*mctz
;
3209 unsigned long long excess
;
3214 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3216 * This loop can run a while, specially if mem_cgroup's continuously
3217 * keep exceeding their soft limit and putting the system under
3224 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3228 reclaimed
= mem_cgroup_hierarchical_reclaim(mz
->mem
, zone
,
3230 MEM_CGROUP_RECLAIM_SOFT
);
3231 nr_reclaimed
+= reclaimed
;
3232 spin_lock(&mctz
->lock
);
3235 * If we failed to reclaim anything from this memory cgroup
3236 * it is time to move on to the next cgroup
3242 * Loop until we find yet another one.
3244 * By the time we get the soft_limit lock
3245 * again, someone might have aded the
3246 * group back on the RB tree. Iterate to
3247 * make sure we get a different mem.
3248 * mem_cgroup_largest_soft_limit_node returns
3249 * NULL if no other cgroup is present on
3253 __mem_cgroup_largest_soft_limit_node(mctz
);
3254 if (next_mz
== mz
) {
3255 css_put(&next_mz
->mem
->css
);
3257 } else /* next_mz == NULL or other memcg */
3261 __mem_cgroup_remove_exceeded(mz
->mem
, mz
, mctz
);
3262 excess
= res_counter_soft_limit_excess(&mz
->mem
->res
);
3264 * One school of thought says that we should not add
3265 * back the node to the tree if reclaim returns 0.
3266 * But our reclaim could return 0, simply because due
3267 * to priority we are exposing a smaller subset of
3268 * memory to reclaim from. Consider this as a longer
3271 /* If excess == 0, no tree ops */
3272 __mem_cgroup_insert_exceeded(mz
->mem
, mz
, mctz
, excess
);
3273 spin_unlock(&mctz
->lock
);
3274 css_put(&mz
->mem
->css
);
3277 * Could not reclaim anything and there are no more
3278 * mem cgroups to try or we seem to be looping without
3279 * reclaiming anything.
3281 if (!nr_reclaimed
&&
3283 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3285 } while (!nr_reclaimed
);
3287 css_put(&next_mz
->mem
->css
);
3288 return nr_reclaimed
;
3292 * This routine traverse page_cgroup in given list and drop them all.
3293 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3295 static int mem_cgroup_force_empty_list(struct mem_cgroup
*mem
,
3296 int node
, int zid
, enum lru_list lru
)
3299 struct mem_cgroup_per_zone
*mz
;
3300 struct page_cgroup
*pc
, *busy
;
3301 unsigned long flags
, loop
;
3302 struct list_head
*list
;
3305 zone
= &NODE_DATA(node
)->node_zones
[zid
];
3306 mz
= mem_cgroup_zoneinfo(mem
, node
, zid
);
3307 list
= &mz
->lists
[lru
];
3309 loop
= MEM_CGROUP_ZSTAT(mz
, lru
);
3310 /* give some margin against EBUSY etc...*/
3317 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3318 if (list_empty(list
)) {
3319 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3322 pc
= list_entry(list
->prev
, struct page_cgroup
, lru
);
3324 list_move(&pc
->lru
, list
);
3326 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3329 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3331 page
= lookup_cgroup_page(pc
);
3333 ret
= mem_cgroup_move_parent(page
, pc
, mem
, GFP_KERNEL
);
3337 if (ret
== -EBUSY
|| ret
== -EINVAL
) {
3338 /* found lock contention or "pc" is obsolete. */
3345 if (!ret
&& !list_empty(list
))
3351 * make mem_cgroup's charge to be 0 if there is no task.
3352 * This enables deleting this mem_cgroup.
3354 static int mem_cgroup_force_empty(struct mem_cgroup
*mem
, bool free_all
)
3357 int node
, zid
, shrink
;
3358 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3359 struct cgroup
*cgrp
= mem
->css
.cgroup
;
3364 /* should free all ? */
3370 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
3373 if (signal_pending(current
))
3375 /* This is for making all *used* pages to be on LRU. */
3376 lru_add_drain_all();
3377 drain_all_stock_sync();
3379 mem_cgroup_start_move(mem
);
3380 for_each_node_state(node
, N_HIGH_MEMORY
) {
3381 for (zid
= 0; !ret
&& zid
< MAX_NR_ZONES
; zid
++) {
3384 ret
= mem_cgroup_force_empty_list(mem
,
3393 mem_cgroup_end_move(mem
);
3394 memcg_oom_recover(mem
);
3395 /* it seems parent cgroup doesn't have enough mem */
3399 /* "ret" should also be checked to ensure all lists are empty. */
3400 } while (mem
->res
.usage
> 0 || ret
);
3406 /* returns EBUSY if there is a task or if we come here twice. */
3407 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
) || shrink
) {
3411 /* we call try-to-free pages for make this cgroup empty */
3412 lru_add_drain_all();
3413 /* try to free all pages in this cgroup */
3415 while (nr_retries
&& mem
->res
.usage
> 0) {
3418 if (signal_pending(current
)) {
3422 progress
= try_to_free_mem_cgroup_pages(mem
, GFP_KERNEL
,
3423 false, get_swappiness(mem
));
3426 /* maybe some writeback is necessary */
3427 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3432 /* try move_account...there may be some *locked* pages. */
3436 int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
3438 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont
), true);
3442 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
3444 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
3447 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
3451 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3452 struct cgroup
*parent
= cont
->parent
;
3453 struct mem_cgroup
*parent_mem
= NULL
;
3456 parent_mem
= mem_cgroup_from_cont(parent
);
3460 * If parent's use_hierarchy is set, we can't make any modifications
3461 * in the child subtrees. If it is unset, then the change can
3462 * occur, provided the current cgroup has no children.
3464 * For the root cgroup, parent_mem is NULL, we allow value to be
3465 * set if there are no children.
3467 if ((!parent_mem
|| !parent_mem
->use_hierarchy
) &&
3468 (val
== 1 || val
== 0)) {
3469 if (list_empty(&cont
->children
))
3470 mem
->use_hierarchy
= val
;
3481 static u64
mem_cgroup_get_recursive_idx_stat(struct mem_cgroup
*mem
,
3482 enum mem_cgroup_stat_index idx
)
3484 struct mem_cgroup
*iter
;
3487 /* each per cpu's value can be minus.Then, use s64 */
3488 for_each_mem_cgroup_tree(iter
, mem
)
3489 val
+= mem_cgroup_read_stat(iter
, idx
);
3491 if (val
< 0) /* race ? */
3496 static inline u64
mem_cgroup_usage(struct mem_cgroup
*mem
, bool swap
)
3500 if (!mem_cgroup_is_root(mem
)) {
3502 return res_counter_read_u64(&mem
->res
, RES_USAGE
);
3504 return res_counter_read_u64(&mem
->memsw
, RES_USAGE
);
3507 val
= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3508 val
+= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_RSS
);
3511 val
+= mem_cgroup_get_recursive_idx_stat(mem
,
3512 MEM_CGROUP_STAT_SWAPOUT
);
3514 return val
<< PAGE_SHIFT
;
3517 static u64
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
)
3519 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3523 type
= MEMFILE_TYPE(cft
->private);
3524 name
= MEMFILE_ATTR(cft
->private);
3527 if (name
== RES_USAGE
)
3528 val
= mem_cgroup_usage(mem
, false);
3530 val
= res_counter_read_u64(&mem
->res
, name
);
3533 if (name
== RES_USAGE
)
3534 val
= mem_cgroup_usage(mem
, true);
3536 val
= res_counter_read_u64(&mem
->memsw
, name
);
3545 * The user of this function is...
3548 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
3551 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
3553 unsigned long long val
;
3556 type
= MEMFILE_TYPE(cft
->private);
3557 name
= MEMFILE_ATTR(cft
->private);
3560 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3564 /* This function does all necessary parse...reuse it */
3565 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3569 ret
= mem_cgroup_resize_limit(memcg
, val
);
3571 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
3573 case RES_SOFT_LIMIT
:
3574 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3578 * For memsw, soft limits are hard to implement in terms
3579 * of semantics, for now, we support soft limits for
3580 * control without swap
3583 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
3588 ret
= -EINVAL
; /* should be BUG() ? */
3594 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
3595 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
3597 struct cgroup
*cgroup
;
3598 unsigned long long min_limit
, min_memsw_limit
, tmp
;
3600 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3601 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3602 cgroup
= memcg
->css
.cgroup
;
3603 if (!memcg
->use_hierarchy
)
3606 while (cgroup
->parent
) {
3607 cgroup
= cgroup
->parent
;
3608 memcg
= mem_cgroup_from_cont(cgroup
);
3609 if (!memcg
->use_hierarchy
)
3611 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3612 min_limit
= min(min_limit
, tmp
);
3613 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3614 min_memsw_limit
= min(min_memsw_limit
, tmp
);
3617 *mem_limit
= min_limit
;
3618 *memsw_limit
= min_memsw_limit
;
3622 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
3624 struct mem_cgroup
*mem
;
3627 mem
= mem_cgroup_from_cont(cont
);
3628 type
= MEMFILE_TYPE(event
);
3629 name
= MEMFILE_ATTR(event
);
3633 res_counter_reset_max(&mem
->res
);
3635 res_counter_reset_max(&mem
->memsw
);
3639 res_counter_reset_failcnt(&mem
->res
);
3641 res_counter_reset_failcnt(&mem
->memsw
);
3648 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
3651 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
3655 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3656 struct cftype
*cft
, u64 val
)
3658 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
3660 if (val
>= (1 << NR_MOVE_TYPE
))
3663 * We check this value several times in both in can_attach() and
3664 * attach(), so we need cgroup lock to prevent this value from being
3668 mem
->move_charge_at_immigrate
= val
;
3674 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3675 struct cftype
*cft
, u64 val
)
3682 /* For read statistics */
3698 struct mcs_total_stat
{
3699 s64 stat
[NR_MCS_STAT
];
3705 } memcg_stat_strings
[NR_MCS_STAT
] = {
3706 {"cache", "total_cache"},
3707 {"rss", "total_rss"},
3708 {"mapped_file", "total_mapped_file"},
3709 {"pgpgin", "total_pgpgin"},
3710 {"pgpgout", "total_pgpgout"},
3711 {"swap", "total_swap"},
3712 {"inactive_anon", "total_inactive_anon"},
3713 {"active_anon", "total_active_anon"},
3714 {"inactive_file", "total_inactive_file"},
3715 {"active_file", "total_active_file"},
3716 {"unevictable", "total_unevictable"}
3721 mem_cgroup_get_local_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3726 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3727 s
->stat
[MCS_CACHE
] += val
* PAGE_SIZE
;
3728 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_RSS
);
3729 s
->stat
[MCS_RSS
] += val
* PAGE_SIZE
;
3730 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_FILE_MAPPED
);
3731 s
->stat
[MCS_FILE_MAPPED
] += val
* PAGE_SIZE
;
3732 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGIN_COUNT
);
3733 s
->stat
[MCS_PGPGIN
] += val
;
3734 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGOUT_COUNT
);
3735 s
->stat
[MCS_PGPGOUT
] += val
;
3736 if (do_swap_account
) {
3737 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_SWAPOUT
);
3738 s
->stat
[MCS_SWAP
] += val
* PAGE_SIZE
;
3742 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_ANON
);
3743 s
->stat
[MCS_INACTIVE_ANON
] += val
* PAGE_SIZE
;
3744 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_ANON
);
3745 s
->stat
[MCS_ACTIVE_ANON
] += val
* PAGE_SIZE
;
3746 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_FILE
);
3747 s
->stat
[MCS_INACTIVE_FILE
] += val
* PAGE_SIZE
;
3748 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_FILE
);
3749 s
->stat
[MCS_ACTIVE_FILE
] += val
* PAGE_SIZE
;
3750 val
= mem_cgroup_get_local_zonestat(mem
, LRU_UNEVICTABLE
);
3751 s
->stat
[MCS_UNEVICTABLE
] += val
* PAGE_SIZE
;
3755 mem_cgroup_get_total_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3757 struct mem_cgroup
*iter
;
3759 for_each_mem_cgroup_tree(iter
, mem
)
3760 mem_cgroup_get_local_stat(iter
, s
);
3763 static int mem_control_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
3764 struct cgroup_map_cb
*cb
)
3766 struct mem_cgroup
*mem_cont
= mem_cgroup_from_cont(cont
);
3767 struct mcs_total_stat mystat
;
3770 memset(&mystat
, 0, sizeof(mystat
));
3771 mem_cgroup_get_local_stat(mem_cont
, &mystat
);
3773 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3774 if (i
== MCS_SWAP
&& !do_swap_account
)
3776 cb
->fill(cb
, memcg_stat_strings
[i
].local_name
, mystat
.stat
[i
]);
3779 /* Hierarchical information */
3781 unsigned long long limit
, memsw_limit
;
3782 memcg_get_hierarchical_limit(mem_cont
, &limit
, &memsw_limit
);
3783 cb
->fill(cb
, "hierarchical_memory_limit", limit
);
3784 if (do_swap_account
)
3785 cb
->fill(cb
, "hierarchical_memsw_limit", memsw_limit
);
3788 memset(&mystat
, 0, sizeof(mystat
));
3789 mem_cgroup_get_total_stat(mem_cont
, &mystat
);
3790 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3791 if (i
== MCS_SWAP
&& !do_swap_account
)
3793 cb
->fill(cb
, memcg_stat_strings
[i
].total_name
, mystat
.stat
[i
]);
3796 #ifdef CONFIG_DEBUG_VM
3797 cb
->fill(cb
, "inactive_ratio", calc_inactive_ratio(mem_cont
, NULL
));
3801 struct mem_cgroup_per_zone
*mz
;
3802 unsigned long recent_rotated
[2] = {0, 0};
3803 unsigned long recent_scanned
[2] = {0, 0};
3805 for_each_online_node(nid
)
3806 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3807 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
3809 recent_rotated
[0] +=
3810 mz
->reclaim_stat
.recent_rotated
[0];
3811 recent_rotated
[1] +=
3812 mz
->reclaim_stat
.recent_rotated
[1];
3813 recent_scanned
[0] +=
3814 mz
->reclaim_stat
.recent_scanned
[0];
3815 recent_scanned
[1] +=
3816 mz
->reclaim_stat
.recent_scanned
[1];
3818 cb
->fill(cb
, "recent_rotated_anon", recent_rotated
[0]);
3819 cb
->fill(cb
, "recent_rotated_file", recent_rotated
[1]);
3820 cb
->fill(cb
, "recent_scanned_anon", recent_scanned
[0]);
3821 cb
->fill(cb
, "recent_scanned_file", recent_scanned
[1]);
3828 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
3830 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3832 return get_swappiness(memcg
);
3835 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
3838 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3839 struct mem_cgroup
*parent
;
3844 if (cgrp
->parent
== NULL
)
3847 parent
= mem_cgroup_from_cont(cgrp
->parent
);
3851 /* If under hierarchy, only empty-root can set this value */
3852 if ((parent
->use_hierarchy
) ||
3853 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
3858 memcg
->swappiness
= val
;
3865 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3867 struct mem_cgroup_threshold_ary
*t
;
3873 t
= rcu_dereference(memcg
->thresholds
.primary
);
3875 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3880 usage
= mem_cgroup_usage(memcg
, swap
);
3883 * current_threshold points to threshold just below usage.
3884 * If it's not true, a threshold was crossed after last
3885 * call of __mem_cgroup_threshold().
3887 i
= t
->current_threshold
;
3890 * Iterate backward over array of thresholds starting from
3891 * current_threshold and check if a threshold is crossed.
3892 * If none of thresholds below usage is crossed, we read
3893 * only one element of the array here.
3895 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3896 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3898 /* i = current_threshold + 1 */
3902 * Iterate forward over array of thresholds starting from
3903 * current_threshold+1 and check if a threshold is crossed.
3904 * If none of thresholds above usage is crossed, we read
3905 * only one element of the array here.
3907 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3908 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3910 /* Update current_threshold */
3911 t
->current_threshold
= i
- 1;
3916 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3919 __mem_cgroup_threshold(memcg
, false);
3920 if (do_swap_account
)
3921 __mem_cgroup_threshold(memcg
, true);
3923 memcg
= parent_mem_cgroup(memcg
);
3927 static int compare_thresholds(const void *a
, const void *b
)
3929 const struct mem_cgroup_threshold
*_a
= a
;
3930 const struct mem_cgroup_threshold
*_b
= b
;
3932 return _a
->threshold
- _b
->threshold
;
3935 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*mem
)
3937 struct mem_cgroup_eventfd_list
*ev
;
3939 list_for_each_entry(ev
, &mem
->oom_notify
, list
)
3940 eventfd_signal(ev
->eventfd
, 1);
3944 static void mem_cgroup_oom_notify(struct mem_cgroup
*mem
)
3946 struct mem_cgroup
*iter
;
3948 for_each_mem_cgroup_tree(iter
, mem
)
3949 mem_cgroup_oom_notify_cb(iter
);
3952 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
3953 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
3955 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3956 struct mem_cgroup_thresholds
*thresholds
;
3957 struct mem_cgroup_threshold_ary
*new;
3958 int type
= MEMFILE_TYPE(cft
->private);
3959 u64 threshold
, usage
;
3962 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
3966 mutex_lock(&memcg
->thresholds_lock
);
3969 thresholds
= &memcg
->thresholds
;
3970 else if (type
== _MEMSWAP
)
3971 thresholds
= &memcg
->memsw_thresholds
;
3975 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
3977 /* Check if a threshold crossed before adding a new one */
3978 if (thresholds
->primary
)
3979 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3981 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3983 /* Allocate memory for new array of thresholds */
3984 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3992 /* Copy thresholds (if any) to new array */
3993 if (thresholds
->primary
) {
3994 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3995 sizeof(struct mem_cgroup_threshold
));
3998 /* Add new threshold */
3999 new->entries
[size
- 1].eventfd
= eventfd
;
4000 new->entries
[size
- 1].threshold
= threshold
;
4002 /* Sort thresholds. Registering of new threshold isn't time-critical */
4003 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
4004 compare_thresholds
, NULL
);
4006 /* Find current threshold */
4007 new->current_threshold
= -1;
4008 for (i
= 0; i
< size
; i
++) {
4009 if (new->entries
[i
].threshold
< usage
) {
4011 * new->current_threshold will not be used until
4012 * rcu_assign_pointer(), so it's safe to increment
4015 ++new->current_threshold
;
4019 /* Free old spare buffer and save old primary buffer as spare */
4020 kfree(thresholds
->spare
);
4021 thresholds
->spare
= thresholds
->primary
;
4023 rcu_assign_pointer(thresholds
->primary
, new);
4025 /* To be sure that nobody uses thresholds */
4029 mutex_unlock(&memcg
->thresholds_lock
);
4034 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
4035 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
4037 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
4038 struct mem_cgroup_thresholds
*thresholds
;
4039 struct mem_cgroup_threshold_ary
*new;
4040 int type
= MEMFILE_TYPE(cft
->private);
4044 mutex_lock(&memcg
->thresholds_lock
);
4046 thresholds
= &memcg
->thresholds
;
4047 else if (type
== _MEMSWAP
)
4048 thresholds
= &memcg
->memsw_thresholds
;
4053 * Something went wrong if we trying to unregister a threshold
4054 * if we don't have thresholds
4056 BUG_ON(!thresholds
);
4058 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
4060 /* Check if a threshold crossed before removing */
4061 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4063 /* Calculate new number of threshold */
4065 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4066 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4070 new = thresholds
->spare
;
4072 /* Set thresholds array to NULL if we don't have thresholds */
4081 /* Copy thresholds and find current threshold */
4082 new->current_threshold
= -1;
4083 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4084 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4087 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4088 if (new->entries
[j
].threshold
< usage
) {
4090 * new->current_threshold will not be used
4091 * until rcu_assign_pointer(), so it's safe to increment
4094 ++new->current_threshold
;
4100 /* Swap primary and spare array */
4101 thresholds
->spare
= thresholds
->primary
;
4102 rcu_assign_pointer(thresholds
->primary
, new);
4104 /* To be sure that nobody uses thresholds */
4107 mutex_unlock(&memcg
->thresholds_lock
);
4110 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
4111 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
4113 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
4114 struct mem_cgroup_eventfd_list
*event
;
4115 int type
= MEMFILE_TYPE(cft
->private);
4117 BUG_ON(type
!= _OOM_TYPE
);
4118 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4122 mutex_lock(&memcg_oom_mutex
);
4124 event
->eventfd
= eventfd
;
4125 list_add(&event
->list
, &memcg
->oom_notify
);
4127 /* already in OOM ? */
4128 if (atomic_read(&memcg
->oom_lock
))
4129 eventfd_signal(eventfd
, 1);
4130 mutex_unlock(&memcg_oom_mutex
);
4135 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
4136 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
4138 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4139 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4140 int type
= MEMFILE_TYPE(cft
->private);
4142 BUG_ON(type
!= _OOM_TYPE
);
4144 mutex_lock(&memcg_oom_mutex
);
4146 list_for_each_entry_safe(ev
, tmp
, &mem
->oom_notify
, list
) {
4147 if (ev
->eventfd
== eventfd
) {
4148 list_del(&ev
->list
);
4153 mutex_unlock(&memcg_oom_mutex
);
4156 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
4157 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
4159 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4161 cb
->fill(cb
, "oom_kill_disable", mem
->oom_kill_disable
);
4163 if (atomic_read(&mem
->oom_lock
))
4164 cb
->fill(cb
, "under_oom", 1);
4166 cb
->fill(cb
, "under_oom", 0);
4170 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
4171 struct cftype
*cft
, u64 val
)
4173 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4174 struct mem_cgroup
*parent
;
4176 /* cannot set to root cgroup and only 0 and 1 are allowed */
4177 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
4180 parent
= mem_cgroup_from_cont(cgrp
->parent
);
4183 /* oom-kill-disable is a flag for subhierarchy. */
4184 if ((parent
->use_hierarchy
) ||
4185 (mem
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
4189 mem
->oom_kill_disable
= val
;
4191 memcg_oom_recover(mem
);
4196 static struct cftype mem_cgroup_files
[] = {
4198 .name
= "usage_in_bytes",
4199 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4200 .read_u64
= mem_cgroup_read
,
4201 .register_event
= mem_cgroup_usage_register_event
,
4202 .unregister_event
= mem_cgroup_usage_unregister_event
,
4205 .name
= "max_usage_in_bytes",
4206 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4207 .trigger
= mem_cgroup_reset
,
4208 .read_u64
= mem_cgroup_read
,
4211 .name
= "limit_in_bytes",
4212 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4213 .write_string
= mem_cgroup_write
,
4214 .read_u64
= mem_cgroup_read
,
4217 .name
= "soft_limit_in_bytes",
4218 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4219 .write_string
= mem_cgroup_write
,
4220 .read_u64
= mem_cgroup_read
,
4224 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4225 .trigger
= mem_cgroup_reset
,
4226 .read_u64
= mem_cgroup_read
,
4230 .read_map
= mem_control_stat_show
,
4233 .name
= "force_empty",
4234 .trigger
= mem_cgroup_force_empty_write
,
4237 .name
= "use_hierarchy",
4238 .write_u64
= mem_cgroup_hierarchy_write
,
4239 .read_u64
= mem_cgroup_hierarchy_read
,
4242 .name
= "swappiness",
4243 .read_u64
= mem_cgroup_swappiness_read
,
4244 .write_u64
= mem_cgroup_swappiness_write
,
4247 .name
= "move_charge_at_immigrate",
4248 .read_u64
= mem_cgroup_move_charge_read
,
4249 .write_u64
= mem_cgroup_move_charge_write
,
4252 .name
= "oom_control",
4253 .read_map
= mem_cgroup_oom_control_read
,
4254 .write_u64
= mem_cgroup_oom_control_write
,
4255 .register_event
= mem_cgroup_oom_register_event
,
4256 .unregister_event
= mem_cgroup_oom_unregister_event
,
4257 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4261 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4262 static struct cftype memsw_cgroup_files
[] = {
4264 .name
= "memsw.usage_in_bytes",
4265 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
4266 .read_u64
= mem_cgroup_read
,
4267 .register_event
= mem_cgroup_usage_register_event
,
4268 .unregister_event
= mem_cgroup_usage_unregister_event
,
4271 .name
= "memsw.max_usage_in_bytes",
4272 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
4273 .trigger
= mem_cgroup_reset
,
4274 .read_u64
= mem_cgroup_read
,
4277 .name
= "memsw.limit_in_bytes",
4278 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
4279 .write_string
= mem_cgroup_write
,
4280 .read_u64
= mem_cgroup_read
,
4283 .name
= "memsw.failcnt",
4284 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
4285 .trigger
= mem_cgroup_reset
,
4286 .read_u64
= mem_cgroup_read
,
4290 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4292 if (!do_swap_account
)
4294 return cgroup_add_files(cont
, ss
, memsw_cgroup_files
,
4295 ARRAY_SIZE(memsw_cgroup_files
));
4298 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4304 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4306 struct mem_cgroup_per_node
*pn
;
4307 struct mem_cgroup_per_zone
*mz
;
4309 int zone
, tmp
= node
;
4311 * This routine is called against possible nodes.
4312 * But it's BUG to call kmalloc() against offline node.
4314 * TODO: this routine can waste much memory for nodes which will
4315 * never be onlined. It's better to use memory hotplug callback
4318 if (!node_state(node
, N_NORMAL_MEMORY
))
4320 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4324 mem
->info
.nodeinfo
[node
] = pn
;
4325 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4326 mz
= &pn
->zoneinfo
[zone
];
4328 INIT_LIST_HEAD(&mz
->lists
[l
]);
4329 mz
->usage_in_excess
= 0;
4330 mz
->on_tree
= false;
4336 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4338 kfree(mem
->info
.nodeinfo
[node
]);
4341 static struct mem_cgroup
*mem_cgroup_alloc(void)
4343 struct mem_cgroup
*mem
;
4344 int size
= sizeof(struct mem_cgroup
);
4346 /* Can be very big if MAX_NUMNODES is very big */
4347 if (size
< PAGE_SIZE
)
4348 mem
= kzalloc(size
, GFP_KERNEL
);
4350 mem
= vzalloc(size
);
4355 mem
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4358 spin_lock_init(&mem
->pcp_counter_lock
);
4362 if (size
< PAGE_SIZE
)
4370 * At destroying mem_cgroup, references from swap_cgroup can remain.
4371 * (scanning all at force_empty is too costly...)
4373 * Instead of clearing all references at force_empty, we remember
4374 * the number of reference from swap_cgroup and free mem_cgroup when
4375 * it goes down to 0.
4377 * Removal of cgroup itself succeeds regardless of refs from swap.
4380 static void __mem_cgroup_free(struct mem_cgroup
*mem
)
4384 mem_cgroup_remove_from_trees(mem
);
4385 free_css_id(&mem_cgroup_subsys
, &mem
->css
);
4387 for_each_node_state(node
, N_POSSIBLE
)
4388 free_mem_cgroup_per_zone_info(mem
, node
);
4390 free_percpu(mem
->stat
);
4391 if (sizeof(struct mem_cgroup
) < PAGE_SIZE
)
4397 static void mem_cgroup_get(struct mem_cgroup
*mem
)
4399 atomic_inc(&mem
->refcnt
);
4402 static void __mem_cgroup_put(struct mem_cgroup
*mem
, int count
)
4404 if (atomic_sub_and_test(count
, &mem
->refcnt
)) {
4405 struct mem_cgroup
*parent
= parent_mem_cgroup(mem
);
4406 __mem_cgroup_free(mem
);
4408 mem_cgroup_put(parent
);
4412 static void mem_cgroup_put(struct mem_cgroup
*mem
)
4414 __mem_cgroup_put(mem
, 1);
4418 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4420 static struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*mem
)
4422 if (!mem
->res
.parent
)
4424 return mem_cgroup_from_res_counter(mem
->res
.parent
, res
);
4427 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4428 static void __init
enable_swap_cgroup(void)
4430 if (!mem_cgroup_disabled() && really_do_swap_account
)
4431 do_swap_account
= 1;
4434 static void __init
enable_swap_cgroup(void)
4439 static int mem_cgroup_soft_limit_tree_init(void)
4441 struct mem_cgroup_tree_per_node
*rtpn
;
4442 struct mem_cgroup_tree_per_zone
*rtpz
;
4443 int tmp
, node
, zone
;
4445 for_each_node_state(node
, N_POSSIBLE
) {
4447 if (!node_state(node
, N_NORMAL_MEMORY
))
4449 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
4453 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
4455 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4456 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
4457 rtpz
->rb_root
= RB_ROOT
;
4458 spin_lock_init(&rtpz
->lock
);
4464 static struct cgroup_subsys_state
* __ref
4465 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
4467 struct mem_cgroup
*mem
, *parent
;
4468 long error
= -ENOMEM
;
4471 mem
= mem_cgroup_alloc();
4473 return ERR_PTR(error
);
4475 for_each_node_state(node
, N_POSSIBLE
)
4476 if (alloc_mem_cgroup_per_zone_info(mem
, node
))
4480 if (cont
->parent
== NULL
) {
4482 enable_swap_cgroup();
4484 root_mem_cgroup
= mem
;
4485 if (mem_cgroup_soft_limit_tree_init())
4487 for_each_possible_cpu(cpu
) {
4488 struct memcg_stock_pcp
*stock
=
4489 &per_cpu(memcg_stock
, cpu
);
4490 INIT_WORK(&stock
->work
, drain_local_stock
);
4492 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
4494 parent
= mem_cgroup_from_cont(cont
->parent
);
4495 mem
->use_hierarchy
= parent
->use_hierarchy
;
4496 mem
->oom_kill_disable
= parent
->oom_kill_disable
;
4499 if (parent
&& parent
->use_hierarchy
) {
4500 res_counter_init(&mem
->res
, &parent
->res
);
4501 res_counter_init(&mem
->memsw
, &parent
->memsw
);
4503 * We increment refcnt of the parent to ensure that we can
4504 * safely access it on res_counter_charge/uncharge.
4505 * This refcnt will be decremented when freeing this
4506 * mem_cgroup(see mem_cgroup_put).
4508 mem_cgroup_get(parent
);
4510 res_counter_init(&mem
->res
, NULL
);
4511 res_counter_init(&mem
->memsw
, NULL
);
4513 mem
->last_scanned_child
= 0;
4514 INIT_LIST_HEAD(&mem
->oom_notify
);
4517 mem
->swappiness
= get_swappiness(parent
);
4518 atomic_set(&mem
->refcnt
, 1);
4519 mem
->move_charge_at_immigrate
= 0;
4520 mutex_init(&mem
->thresholds_lock
);
4523 __mem_cgroup_free(mem
);
4524 root_mem_cgroup
= NULL
;
4525 return ERR_PTR(error
);
4528 static int mem_cgroup_pre_destroy(struct cgroup_subsys
*ss
,
4529 struct cgroup
*cont
)
4531 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4533 return mem_cgroup_force_empty(mem
, false);
4536 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
4537 struct cgroup
*cont
)
4539 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4541 mem_cgroup_put(mem
);
4544 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
4545 struct cgroup
*cont
)
4549 ret
= cgroup_add_files(cont
, ss
, mem_cgroup_files
,
4550 ARRAY_SIZE(mem_cgroup_files
));
4553 ret
= register_memsw_files(cont
, ss
);
4558 /* Handlers for move charge at task migration. */
4559 #define PRECHARGE_COUNT_AT_ONCE 256
4560 static int mem_cgroup_do_precharge(unsigned long count
)
4563 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4564 struct mem_cgroup
*mem
= mc
.to
;
4566 if (mem_cgroup_is_root(mem
)) {
4567 mc
.precharge
+= count
;
4568 /* we don't need css_get for root */
4571 /* try to charge at once */
4573 struct res_counter
*dummy
;
4575 * "mem" cannot be under rmdir() because we've already checked
4576 * by cgroup_lock_live_cgroup() that it is not removed and we
4577 * are still under the same cgroup_mutex. So we can postpone
4580 if (res_counter_charge(&mem
->res
, PAGE_SIZE
* count
, &dummy
))
4582 if (do_swap_account
&& res_counter_charge(&mem
->memsw
,
4583 PAGE_SIZE
* count
, &dummy
)) {
4584 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
4587 mc
.precharge
+= count
;
4591 /* fall back to one by one charge */
4593 if (signal_pending(current
)) {
4597 if (!batch_count
--) {
4598 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4601 ret
= __mem_cgroup_try_charge(NULL
, GFP_KERNEL
, &mem
, false,
4604 /* mem_cgroup_clear_mc() will do uncharge later */
4612 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4613 * @vma: the vma the pte to be checked belongs
4614 * @addr: the address corresponding to the pte to be checked
4615 * @ptent: the pte to be checked
4616 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4619 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4620 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4621 * move charge. if @target is not NULL, the page is stored in target->page
4622 * with extra refcnt got(Callers should handle it).
4623 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4624 * target for charge migration. if @target is not NULL, the entry is stored
4627 * Called with pte lock held.
4634 enum mc_target_type
{
4635 MC_TARGET_NONE
, /* not used */
4640 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4641 unsigned long addr
, pte_t ptent
)
4643 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4645 if (!page
|| !page_mapped(page
))
4647 if (PageAnon(page
)) {
4648 /* we don't move shared anon */
4649 if (!move_anon() || page_mapcount(page
) > 2)
4651 } else if (!move_file())
4652 /* we ignore mapcount for file pages */
4654 if (!get_page_unless_zero(page
))
4660 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4661 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4664 struct page
*page
= NULL
;
4665 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4667 if (!move_anon() || non_swap_entry(ent
))
4669 usage_count
= mem_cgroup_count_swap_user(ent
, &page
);
4670 if (usage_count
> 1) { /* we don't move shared anon */
4675 if (do_swap_account
)
4676 entry
->val
= ent
.val
;
4681 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4682 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4684 struct page
*page
= NULL
;
4685 struct inode
*inode
;
4686 struct address_space
*mapping
;
4689 if (!vma
->vm_file
) /* anonymous vma */
4694 inode
= vma
->vm_file
->f_path
.dentry
->d_inode
;
4695 mapping
= vma
->vm_file
->f_mapping
;
4696 if (pte_none(ptent
))
4697 pgoff
= linear_page_index(vma
, addr
);
4698 else /* pte_file(ptent) is true */
4699 pgoff
= pte_to_pgoff(ptent
);
4701 /* page is moved even if it's not RSS of this task(page-faulted). */
4702 if (!mapping_cap_swap_backed(mapping
)) { /* normal file */
4703 page
= find_get_page(mapping
, pgoff
);
4704 } else { /* shmem/tmpfs file. we should take account of swap too. */
4706 mem_cgroup_get_shmem_target(inode
, pgoff
, &page
, &ent
);
4707 if (do_swap_account
)
4708 entry
->val
= ent
.val
;
4714 static int is_target_pte_for_mc(struct vm_area_struct
*vma
,
4715 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4717 struct page
*page
= NULL
;
4718 struct page_cgroup
*pc
;
4720 swp_entry_t ent
= { .val
= 0 };
4722 if (pte_present(ptent
))
4723 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4724 else if (is_swap_pte(ptent
))
4725 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4726 else if (pte_none(ptent
) || pte_file(ptent
))
4727 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4729 if (!page
&& !ent
.val
)
4732 pc
= lookup_page_cgroup(page
);
4734 * Do only loose check w/o page_cgroup lock.
4735 * mem_cgroup_move_account() checks the pc is valid or not under
4738 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
4739 ret
= MC_TARGET_PAGE
;
4741 target
->page
= page
;
4743 if (!ret
|| !target
)
4746 /* There is a swap entry and a page doesn't exist or isn't charged */
4747 if (ent
.val
&& !ret
&&
4748 css_id(&mc
.from
->css
) == lookup_swap_cgroup(ent
)) {
4749 ret
= MC_TARGET_SWAP
;
4756 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4757 unsigned long addr
, unsigned long end
,
4758 struct mm_walk
*walk
)
4760 struct vm_area_struct
*vma
= walk
->private;
4764 split_huge_page_pmd(walk
->mm
, pmd
);
4766 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4767 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4768 if (is_target_pte_for_mc(vma
, addr
, *pte
, NULL
))
4769 mc
.precharge
++; /* increment precharge temporarily */
4770 pte_unmap_unlock(pte
- 1, ptl
);
4776 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4778 unsigned long precharge
;
4779 struct vm_area_struct
*vma
;
4781 down_read(&mm
->mmap_sem
);
4782 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
4783 struct mm_walk mem_cgroup_count_precharge_walk
= {
4784 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4788 if (is_vm_hugetlb_page(vma
))
4790 walk_page_range(vma
->vm_start
, vma
->vm_end
,
4791 &mem_cgroup_count_precharge_walk
);
4793 up_read(&mm
->mmap_sem
);
4795 precharge
= mc
.precharge
;
4801 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4803 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4805 VM_BUG_ON(mc
.moving_task
);
4806 mc
.moving_task
= current
;
4807 return mem_cgroup_do_precharge(precharge
);
4810 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4811 static void __mem_cgroup_clear_mc(void)
4813 struct mem_cgroup
*from
= mc
.from
;
4814 struct mem_cgroup
*to
= mc
.to
;
4816 /* we must uncharge all the leftover precharges from mc.to */
4818 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
4822 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4823 * we must uncharge here.
4825 if (mc
.moved_charge
) {
4826 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
4827 mc
.moved_charge
= 0;
4829 /* we must fixup refcnts and charges */
4830 if (mc
.moved_swap
) {
4831 /* uncharge swap account from the old cgroup */
4832 if (!mem_cgroup_is_root(mc
.from
))
4833 res_counter_uncharge(&mc
.from
->memsw
,
4834 PAGE_SIZE
* mc
.moved_swap
);
4835 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
4837 if (!mem_cgroup_is_root(mc
.to
)) {
4839 * we charged both to->res and to->memsw, so we should
4842 res_counter_uncharge(&mc
.to
->res
,
4843 PAGE_SIZE
* mc
.moved_swap
);
4845 /* we've already done mem_cgroup_get(mc.to) */
4848 memcg_oom_recover(from
);
4849 memcg_oom_recover(to
);
4850 wake_up_all(&mc
.waitq
);
4853 static void mem_cgroup_clear_mc(void)
4855 struct mem_cgroup
*from
= mc
.from
;
4858 * we must clear moving_task before waking up waiters at the end of
4861 mc
.moving_task
= NULL
;
4862 __mem_cgroup_clear_mc();
4863 spin_lock(&mc
.lock
);
4866 spin_unlock(&mc
.lock
);
4867 mem_cgroup_end_move(from
);
4870 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
4871 struct cgroup
*cgroup
,
4872 struct task_struct
*p
,
4876 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgroup
);
4878 if (mem
->move_charge_at_immigrate
) {
4879 struct mm_struct
*mm
;
4880 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
4882 VM_BUG_ON(from
== mem
);
4884 mm
= get_task_mm(p
);
4887 /* We move charges only when we move a owner of the mm */
4888 if (mm
->owner
== p
) {
4891 VM_BUG_ON(mc
.precharge
);
4892 VM_BUG_ON(mc
.moved_charge
);
4893 VM_BUG_ON(mc
.moved_swap
);
4894 mem_cgroup_start_move(from
);
4895 spin_lock(&mc
.lock
);
4898 spin_unlock(&mc
.lock
);
4899 /* We set mc.moving_task later */
4901 ret
= mem_cgroup_precharge_mc(mm
);
4903 mem_cgroup_clear_mc();
4910 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
4911 struct cgroup
*cgroup
,
4912 struct task_struct
*p
,
4915 mem_cgroup_clear_mc();
4918 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4919 unsigned long addr
, unsigned long end
,
4920 struct mm_walk
*walk
)
4923 struct vm_area_struct
*vma
= walk
->private;
4927 split_huge_page_pmd(walk
->mm
, pmd
);
4929 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4930 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4931 pte_t ptent
= *(pte
++);
4932 union mc_target target
;
4935 struct page_cgroup
*pc
;
4941 type
= is_target_pte_for_mc(vma
, addr
, ptent
, &target
);
4943 case MC_TARGET_PAGE
:
4945 if (isolate_lru_page(page
))
4947 pc
= lookup_page_cgroup(page
);
4948 if (!mem_cgroup_move_account(page
, pc
,
4949 mc
.from
, mc
.to
, false, PAGE_SIZE
)) {
4951 /* we uncharge from mc.from later. */
4954 putback_lru_page(page
);
4955 put
: /* is_target_pte_for_mc() gets the page */
4958 case MC_TARGET_SWAP
:
4960 if (!mem_cgroup_move_swap_account(ent
,
4961 mc
.from
, mc
.to
, false)) {
4963 /* we fixup refcnts and charges later. */
4971 pte_unmap_unlock(pte
- 1, ptl
);
4976 * We have consumed all precharges we got in can_attach().
4977 * We try charge one by one, but don't do any additional
4978 * charges to mc.to if we have failed in charge once in attach()
4981 ret
= mem_cgroup_do_precharge(1);
4989 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4991 struct vm_area_struct
*vma
;
4993 lru_add_drain_all();
4995 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4997 * Someone who are holding the mmap_sem might be waiting in
4998 * waitq. So we cancel all extra charges, wake up all waiters,
4999 * and retry. Because we cancel precharges, we might not be able
5000 * to move enough charges, but moving charge is a best-effort
5001 * feature anyway, so it wouldn't be a big problem.
5003 __mem_cgroup_clear_mc();
5007 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
5009 struct mm_walk mem_cgroup_move_charge_walk
= {
5010 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
5014 if (is_vm_hugetlb_page(vma
))
5016 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
5017 &mem_cgroup_move_charge_walk
);
5020 * means we have consumed all precharges and failed in
5021 * doing additional charge. Just abandon here.
5025 up_read(&mm
->mmap_sem
);
5028 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
5029 struct cgroup
*cont
,
5030 struct cgroup
*old_cont
,
5031 struct task_struct
*p
,
5034 struct mm_struct
*mm
;
5037 /* no need to move charge */
5040 mm
= get_task_mm(p
);
5042 mem_cgroup_move_charge(mm
);
5045 mem_cgroup_clear_mc();
5047 #else /* !CONFIG_MMU */
5048 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
5049 struct cgroup
*cgroup
,
5050 struct task_struct
*p
,
5055 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
5056 struct cgroup
*cgroup
,
5057 struct task_struct
*p
,
5061 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
5062 struct cgroup
*cont
,
5063 struct cgroup
*old_cont
,
5064 struct task_struct
*p
,
5070 struct cgroup_subsys mem_cgroup_subsys
= {
5072 .subsys_id
= mem_cgroup_subsys_id
,
5073 .create
= mem_cgroup_create
,
5074 .pre_destroy
= mem_cgroup_pre_destroy
,
5075 .destroy
= mem_cgroup_destroy
,
5076 .populate
= mem_cgroup_populate
,
5077 .can_attach
= mem_cgroup_can_attach
,
5078 .cancel_attach
= mem_cgroup_cancel_attach
,
5079 .attach
= mem_cgroup_move_task
,
5084 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5085 static int __init
enable_swap_account(char *s
)
5087 /* consider enabled if no parameter or 1 is given */
5088 if (!(*s
) || !strcmp(s
, "=1"))
5089 really_do_swap_account
= 1;
5090 else if (!strcmp(s
, "=0"))
5091 really_do_swap_account
= 0;
5094 __setup("swapaccount", enable_swap_account
);
5096 static int __init
disable_swap_account(char *s
)
5098 printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5099 enable_swap_account("=0");
5102 __setup("noswapaccount", disable_swap_account
);