1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
50 #include <linux/oom.h>
53 #include <asm/uaccess.h>
55 #include <trace/events/vmscan.h>
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly
;
58 #define MEM_CGROUP_RECLAIM_RETRIES 5
59 struct mem_cgroup
*root_mem_cgroup __read_mostly
;
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly
;
65 /* for remember boot option*/
66 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
67 static int really_do_swap_account __initdata
= 1;
69 static int really_do_swap_account __initdata
= 0;
73 #define do_swap_account (0)
77 * Per memcg event counter is incremented at every pagein/pageout. This counter
78 * is used for trigger some periodic events. This is straightforward and better
79 * than using jiffies etc. to handle periodic memcg event.
81 * These values will be used as !((event) & ((1 <<(thresh)) - 1))
83 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
84 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
87 * Statistics for memory cgroup.
89 enum mem_cgroup_stat_index
{
91 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
93 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
94 MEM_CGROUP_STAT_RSS
, /* # of pages charged as anon rss */
95 MEM_CGROUP_STAT_FILE_MAPPED
, /* # of pages charged as file rss */
96 MEM_CGROUP_STAT_PGPGIN_COUNT
, /* # of pages paged in */
97 MEM_CGROUP_STAT_PGPGOUT_COUNT
, /* # of pages paged out */
98 MEM_CGROUP_STAT_SWAPOUT
, /* # of pages, swapped out */
99 MEM_CGROUP_STAT_DATA
, /* end of data requires synchronization */
100 /* incremented at every pagein/pageout */
101 MEM_CGROUP_EVENTS
= MEM_CGROUP_STAT_DATA
,
102 MEM_CGROUP_ON_MOVE
, /* someone is moving account between groups */
104 MEM_CGROUP_STAT_NSTATS
,
107 struct mem_cgroup_stat_cpu
{
108 s64 count
[MEM_CGROUP_STAT_NSTATS
];
112 * per-zone information in memory controller.
114 struct mem_cgroup_per_zone
{
116 * spin_lock to protect the per cgroup LRU
118 struct list_head lists
[NR_LRU_LISTS
];
119 unsigned long count
[NR_LRU_LISTS
];
121 struct zone_reclaim_stat reclaim_stat
;
122 struct rb_node tree_node
; /* RB tree node */
123 unsigned long long usage_in_excess
;/* Set to the value by which */
124 /* the soft limit is exceeded*/
126 struct mem_cgroup
*mem
; /* Back pointer, we cannot */
127 /* use container_of */
129 /* Macro for accessing counter */
130 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
132 struct mem_cgroup_per_node
{
133 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
136 struct mem_cgroup_lru_info
{
137 struct mem_cgroup_per_node
*nodeinfo
[MAX_NUMNODES
];
141 * Cgroups above their limits are maintained in a RB-Tree, independent of
142 * their hierarchy representation
145 struct mem_cgroup_tree_per_zone
{
146 struct rb_root rb_root
;
150 struct mem_cgroup_tree_per_node
{
151 struct mem_cgroup_tree_per_zone rb_tree_per_zone
[MAX_NR_ZONES
];
154 struct mem_cgroup_tree
{
155 struct mem_cgroup_tree_per_node
*rb_tree_per_node
[MAX_NUMNODES
];
158 static struct mem_cgroup_tree soft_limit_tree __read_mostly
;
160 struct mem_cgroup_threshold
{
161 struct eventfd_ctx
*eventfd
;
166 struct mem_cgroup_threshold_ary
{
167 /* An array index points to threshold just below usage. */
168 int current_threshold
;
169 /* Size of entries[] */
171 /* Array of thresholds */
172 struct mem_cgroup_threshold entries
[0];
175 struct mem_cgroup_thresholds
{
176 /* Primary thresholds array */
177 struct mem_cgroup_threshold_ary
*primary
;
179 * Spare threshold array.
180 * This is needed to make mem_cgroup_unregister_event() "never fail".
181 * It must be able to store at least primary->size - 1 entries.
183 struct mem_cgroup_threshold_ary
*spare
;
187 struct mem_cgroup_eventfd_list
{
188 struct list_head list
;
189 struct eventfd_ctx
*eventfd
;
192 static void mem_cgroup_threshold(struct mem_cgroup
*mem
);
193 static void mem_cgroup_oom_notify(struct mem_cgroup
*mem
);
196 * The memory controller data structure. The memory controller controls both
197 * page cache and RSS per cgroup. We would eventually like to provide
198 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
199 * to help the administrator determine what knobs to tune.
201 * TODO: Add a water mark for the memory controller. Reclaim will begin when
202 * we hit the water mark. May be even add a low water mark, such that
203 * no reclaim occurs from a cgroup at it's low water mark, this is
204 * a feature that will be implemented much later in the future.
207 struct cgroup_subsys_state css
;
209 * the counter to account for memory usage
211 struct res_counter res
;
213 * the counter to account for mem+swap usage.
215 struct res_counter memsw
;
217 * Per cgroup active and inactive list, similar to the
218 * per zone LRU lists.
220 struct mem_cgroup_lru_info info
;
223 protect against reclaim related member.
225 spinlock_t reclaim_param_lock
;
228 * While reclaiming in a hierarchy, we cache the last child we
231 int last_scanned_child
;
233 * Should the accounting and control be hierarchical, per subtree?
239 unsigned int swappiness
;
240 /* OOM-Killer disable */
241 int oom_kill_disable
;
243 /* set when res.limit == memsw.limit */
244 bool memsw_is_minimum
;
246 /* protect arrays of thresholds */
247 struct mutex thresholds_lock
;
249 /* thresholds for memory usage. RCU-protected */
250 struct mem_cgroup_thresholds thresholds
;
252 /* thresholds for mem+swap usage. RCU-protected */
253 struct mem_cgroup_thresholds memsw_thresholds
;
255 /* For oom notifier event fd */
256 struct list_head oom_notify
;
259 * Should we move charges of a task when a task is moved into this
260 * mem_cgroup ? And what type of charges should we move ?
262 unsigned long move_charge_at_immigrate
;
266 struct mem_cgroup_stat_cpu
*stat
;
268 * used when a cpu is offlined or other synchronizations
269 * See mem_cgroup_read_stat().
271 struct mem_cgroup_stat_cpu nocpu_base
;
272 spinlock_t pcp_counter_lock
;
275 /* Stuffs for move charges at task migration. */
277 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
278 * left-shifted bitmap of these types.
281 MOVE_CHARGE_TYPE_ANON
, /* private anonymous page and swap of it */
282 MOVE_CHARGE_TYPE_FILE
, /* file page(including tmpfs) and swap of it */
286 /* "mc" and its members are protected by cgroup_mutex */
287 static struct move_charge_struct
{
288 spinlock_t lock
; /* for from, to */
289 struct mem_cgroup
*from
;
290 struct mem_cgroup
*to
;
291 unsigned long precharge
;
292 unsigned long moved_charge
;
293 unsigned long moved_swap
;
294 struct task_struct
*moving_task
; /* a task moving charges */
295 wait_queue_head_t waitq
; /* a waitq for other context */
297 .lock
= __SPIN_LOCK_UNLOCKED(mc
.lock
),
298 .waitq
= __WAIT_QUEUE_HEAD_INITIALIZER(mc
.waitq
),
301 static bool move_anon(void)
303 return test_bit(MOVE_CHARGE_TYPE_ANON
,
304 &mc
.to
->move_charge_at_immigrate
);
307 static bool move_file(void)
309 return test_bit(MOVE_CHARGE_TYPE_FILE
,
310 &mc
.to
->move_charge_at_immigrate
);
314 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
315 * limit reclaim to prevent infinite loops, if they ever occur.
317 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
318 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
321 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
322 MEM_CGROUP_CHARGE_TYPE_MAPPED
,
323 MEM_CGROUP_CHARGE_TYPE_SHMEM
, /* used by page migration of shmem */
324 MEM_CGROUP_CHARGE_TYPE_FORCE
, /* used by force_empty */
325 MEM_CGROUP_CHARGE_TYPE_SWAPOUT
, /* for accounting swapcache */
326 MEM_CGROUP_CHARGE_TYPE_DROP
, /* a page was unused swap cache */
330 /* only for here (for easy reading.) */
331 #define PCGF_CACHE (1UL << PCG_CACHE)
332 #define PCGF_USED (1UL << PCG_USED)
333 #define PCGF_LOCK (1UL << PCG_LOCK)
334 /* Not used, but added here for completeness */
335 #define PCGF_ACCT (1UL << PCG_ACCT)
337 /* for encoding cft->private value on file */
340 #define _OOM_TYPE (2)
341 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
342 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
343 #define MEMFILE_ATTR(val) ((val) & 0xffff)
344 /* Used for OOM nofiier */
345 #define OOM_CONTROL (0)
348 * Reclaim flags for mem_cgroup_hierarchical_reclaim
350 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
351 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
352 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
353 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
354 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
355 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
357 static void mem_cgroup_get(struct mem_cgroup
*mem
);
358 static void mem_cgroup_put(struct mem_cgroup
*mem
);
359 static struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*mem
);
360 static void drain_all_stock_async(void);
362 static struct mem_cgroup_per_zone
*
363 mem_cgroup_zoneinfo(struct mem_cgroup
*mem
, int nid
, int zid
)
365 return &mem
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
368 struct cgroup_subsys_state
*mem_cgroup_css(struct mem_cgroup
*mem
)
373 static struct mem_cgroup_per_zone
*
374 page_cgroup_zoneinfo(struct page_cgroup
*pc
)
376 struct mem_cgroup
*mem
= pc
->mem_cgroup
;
377 int nid
= page_cgroup_nid(pc
);
378 int zid
= page_cgroup_zid(pc
);
383 return mem_cgroup_zoneinfo(mem
, nid
, zid
);
386 static struct mem_cgroup_tree_per_zone
*
387 soft_limit_tree_node_zone(int nid
, int zid
)
389 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
392 static struct mem_cgroup_tree_per_zone
*
393 soft_limit_tree_from_page(struct page
*page
)
395 int nid
= page_to_nid(page
);
396 int zid
= page_zonenum(page
);
398 return &soft_limit_tree
.rb_tree_per_node
[nid
]->rb_tree_per_zone
[zid
];
402 __mem_cgroup_insert_exceeded(struct mem_cgroup
*mem
,
403 struct mem_cgroup_per_zone
*mz
,
404 struct mem_cgroup_tree_per_zone
*mctz
,
405 unsigned long long new_usage_in_excess
)
407 struct rb_node
**p
= &mctz
->rb_root
.rb_node
;
408 struct rb_node
*parent
= NULL
;
409 struct mem_cgroup_per_zone
*mz_node
;
414 mz
->usage_in_excess
= new_usage_in_excess
;
415 if (!mz
->usage_in_excess
)
419 mz_node
= rb_entry(parent
, struct mem_cgroup_per_zone
,
421 if (mz
->usage_in_excess
< mz_node
->usage_in_excess
)
424 * We can't avoid mem cgroups that are over their soft
425 * limit by the same amount
427 else if (mz
->usage_in_excess
>= mz_node
->usage_in_excess
)
430 rb_link_node(&mz
->tree_node
, parent
, p
);
431 rb_insert_color(&mz
->tree_node
, &mctz
->rb_root
);
436 __mem_cgroup_remove_exceeded(struct mem_cgroup
*mem
,
437 struct mem_cgroup_per_zone
*mz
,
438 struct mem_cgroup_tree_per_zone
*mctz
)
442 rb_erase(&mz
->tree_node
, &mctz
->rb_root
);
447 mem_cgroup_remove_exceeded(struct mem_cgroup
*mem
,
448 struct mem_cgroup_per_zone
*mz
,
449 struct mem_cgroup_tree_per_zone
*mctz
)
451 spin_lock(&mctz
->lock
);
452 __mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
453 spin_unlock(&mctz
->lock
);
457 static void mem_cgroup_update_tree(struct mem_cgroup
*mem
, struct page
*page
)
459 unsigned long long excess
;
460 struct mem_cgroup_per_zone
*mz
;
461 struct mem_cgroup_tree_per_zone
*mctz
;
462 int nid
= page_to_nid(page
);
463 int zid
= page_zonenum(page
);
464 mctz
= soft_limit_tree_from_page(page
);
467 * Necessary to update all ancestors when hierarchy is used.
468 * because their event counter is not touched.
470 for (; mem
; mem
= parent_mem_cgroup(mem
)) {
471 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
472 excess
= res_counter_soft_limit_excess(&mem
->res
);
474 * We have to update the tree if mz is on RB-tree or
475 * mem is over its softlimit.
477 if (excess
|| mz
->on_tree
) {
478 spin_lock(&mctz
->lock
);
479 /* if on-tree, remove it */
481 __mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
483 * Insert again. mz->usage_in_excess will be updated.
484 * If excess is 0, no tree ops.
486 __mem_cgroup_insert_exceeded(mem
, mz
, mctz
, excess
);
487 spin_unlock(&mctz
->lock
);
492 static void mem_cgroup_remove_from_trees(struct mem_cgroup
*mem
)
495 struct mem_cgroup_per_zone
*mz
;
496 struct mem_cgroup_tree_per_zone
*mctz
;
498 for_each_node_state(node
, N_POSSIBLE
) {
499 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
500 mz
= mem_cgroup_zoneinfo(mem
, node
, zone
);
501 mctz
= soft_limit_tree_node_zone(node
, zone
);
502 mem_cgroup_remove_exceeded(mem
, mz
, mctz
);
507 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup
*mem
)
509 return res_counter_soft_limit_excess(&mem
->res
) >> PAGE_SHIFT
;
512 static struct mem_cgroup_per_zone
*
513 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
515 struct rb_node
*rightmost
= NULL
;
516 struct mem_cgroup_per_zone
*mz
;
520 rightmost
= rb_last(&mctz
->rb_root
);
522 goto done
; /* Nothing to reclaim from */
524 mz
= rb_entry(rightmost
, struct mem_cgroup_per_zone
, tree_node
);
526 * Remove the node now but someone else can add it back,
527 * we will to add it back at the end of reclaim to its correct
528 * position in the tree.
530 __mem_cgroup_remove_exceeded(mz
->mem
, mz
, mctz
);
531 if (!res_counter_soft_limit_excess(&mz
->mem
->res
) ||
532 !css_tryget(&mz
->mem
->css
))
538 static struct mem_cgroup_per_zone
*
539 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone
*mctz
)
541 struct mem_cgroup_per_zone
*mz
;
543 spin_lock(&mctz
->lock
);
544 mz
= __mem_cgroup_largest_soft_limit_node(mctz
);
545 spin_unlock(&mctz
->lock
);
550 * Implementation Note: reading percpu statistics for memcg.
552 * Both of vmstat[] and percpu_counter has threshold and do periodic
553 * synchronization to implement "quick" read. There are trade-off between
554 * reading cost and precision of value. Then, we may have a chance to implement
555 * a periodic synchronizion of counter in memcg's counter.
557 * But this _read() function is used for user interface now. The user accounts
558 * memory usage by memory cgroup and he _always_ requires exact value because
559 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
560 * have to visit all online cpus and make sum. So, for now, unnecessary
561 * synchronization is not implemented. (just implemented for cpu hotplug)
563 * If there are kernel internal actions which can make use of some not-exact
564 * value, and reading all cpu value can be performance bottleneck in some
565 * common workload, threashold and synchonization as vmstat[] should be
568 static s64
mem_cgroup_read_stat(struct mem_cgroup
*mem
,
569 enum mem_cgroup_stat_index idx
)
575 for_each_online_cpu(cpu
)
576 val
+= per_cpu(mem
->stat
->count
[idx
], cpu
);
577 #ifdef CONFIG_HOTPLUG_CPU
578 spin_lock(&mem
->pcp_counter_lock
);
579 val
+= mem
->nocpu_base
.count
[idx
];
580 spin_unlock(&mem
->pcp_counter_lock
);
586 static s64
mem_cgroup_local_usage(struct mem_cgroup
*mem
)
590 ret
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_RSS
);
591 ret
+= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_CACHE
);
595 static void mem_cgroup_swap_statistics(struct mem_cgroup
*mem
,
598 int val
= (charge
) ? 1 : -1;
599 this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_SWAPOUT
], val
);
602 static void mem_cgroup_charge_statistics(struct mem_cgroup
*mem
,
603 bool file
, int nr_pages
)
608 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_CACHE
], nr_pages
);
610 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_STAT_RSS
], nr_pages
);
612 /* pagein of a big page is an event. So, ignore page size */
614 __this_cpu_inc(mem
->stat
->count
[MEM_CGROUP_STAT_PGPGIN_COUNT
]);
616 __this_cpu_inc(mem
->stat
->count
[MEM_CGROUP_STAT_PGPGOUT_COUNT
]);
618 __this_cpu_add(mem
->stat
->count
[MEM_CGROUP_EVENTS
], nr_pages
);
623 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup
*mem
,
627 struct mem_cgroup_per_zone
*mz
;
630 for_each_online_node(nid
)
631 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
632 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
633 total
+= MEM_CGROUP_ZSTAT(mz
, idx
);
638 static bool __memcg_event_check(struct mem_cgroup
*mem
, int event_mask_shift
)
642 val
= this_cpu_read(mem
->stat
->count
[MEM_CGROUP_EVENTS
]);
644 return !(val
& ((1 << event_mask_shift
) - 1));
648 * Check events in order.
651 static void memcg_check_events(struct mem_cgroup
*mem
, struct page
*page
)
653 /* threshold event is triggered in finer grain than soft limit */
654 if (unlikely(__memcg_event_check(mem
, THRESHOLDS_EVENTS_THRESH
))) {
655 mem_cgroup_threshold(mem
);
656 if (unlikely(__memcg_event_check(mem
, SOFTLIMIT_EVENTS_THRESH
)))
657 mem_cgroup_update_tree(mem
, page
);
661 static struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
663 return container_of(cgroup_subsys_state(cont
,
664 mem_cgroup_subsys_id
), struct mem_cgroup
,
668 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
671 * mm_update_next_owner() may clear mm->owner to NULL
672 * if it races with swapoff, page migration, etc.
673 * So this can be called with p == NULL.
678 return container_of(task_subsys_state(p
, mem_cgroup_subsys_id
),
679 struct mem_cgroup
, css
);
682 static struct mem_cgroup
*try_get_mem_cgroup_from_mm(struct mm_struct
*mm
)
684 struct mem_cgroup
*mem
= NULL
;
689 * Because we have no locks, mm->owner's may be being moved to other
690 * cgroup. We use css_tryget() here even if this looks
691 * pessimistic (rather than adding locks here).
695 mem
= mem_cgroup_from_task(rcu_dereference(mm
->owner
));
698 } while (!css_tryget(&mem
->css
));
703 /* The caller has to guarantee "mem" exists before calling this */
704 static struct mem_cgroup
*mem_cgroup_start_loop(struct mem_cgroup
*mem
)
706 struct cgroup_subsys_state
*css
;
709 if (!mem
) /* ROOT cgroup has the smallest ID */
710 return root_mem_cgroup
; /*css_put/get against root is ignored*/
711 if (!mem
->use_hierarchy
) {
712 if (css_tryget(&mem
->css
))
718 * searching a memory cgroup which has the smallest ID under given
719 * ROOT cgroup. (ID >= 1)
721 css
= css_get_next(&mem_cgroup_subsys
, 1, &mem
->css
, &found
);
722 if (css
&& css_tryget(css
))
723 mem
= container_of(css
, struct mem_cgroup
, css
);
730 static struct mem_cgroup
*mem_cgroup_get_next(struct mem_cgroup
*iter
,
731 struct mem_cgroup
*root
,
734 int nextid
= css_id(&iter
->css
) + 1;
737 struct cgroup_subsys_state
*css
;
739 hierarchy_used
= iter
->use_hierarchy
;
742 /* If no ROOT, walk all, ignore hierarchy */
743 if (!cond
|| (root
&& !hierarchy_used
))
747 root
= root_mem_cgroup
;
753 css
= css_get_next(&mem_cgroup_subsys
, nextid
,
755 if (css
&& css_tryget(css
))
756 iter
= container_of(css
, struct mem_cgroup
, css
);
758 /* If css is NULL, no more cgroups will be found */
760 } while (css
&& !iter
);
765 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
766 * be careful that "break" loop is not allowed. We have reference count.
767 * Instead of that modify "cond" to be false and "continue" to exit the loop.
769 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
770 for (iter = mem_cgroup_start_loop(root);\
772 iter = mem_cgroup_get_next(iter, root, cond))
774 #define for_each_mem_cgroup_tree(iter, root) \
775 for_each_mem_cgroup_tree_cond(iter, root, true)
777 #define for_each_mem_cgroup_all(iter) \
778 for_each_mem_cgroup_tree_cond(iter, NULL, true)
781 static inline bool mem_cgroup_is_root(struct mem_cgroup
*mem
)
783 return (mem
== root_mem_cgroup
);
787 * Following LRU functions are allowed to be used without PCG_LOCK.
788 * Operations are called by routine of global LRU independently from memcg.
789 * What we have to take care of here is validness of pc->mem_cgroup.
791 * Changes to pc->mem_cgroup happens when
794 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
795 * It is added to LRU before charge.
796 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
797 * When moving account, the page is not on LRU. It's isolated.
800 void mem_cgroup_del_lru_list(struct page
*page
, enum lru_list lru
)
802 struct page_cgroup
*pc
;
803 struct mem_cgroup_per_zone
*mz
;
805 if (mem_cgroup_disabled())
807 pc
= lookup_page_cgroup(page
);
808 /* can happen while we handle swapcache. */
809 if (!TestClearPageCgroupAcctLRU(pc
))
811 VM_BUG_ON(!pc
->mem_cgroup
);
813 * We don't check PCG_USED bit. It's cleared when the "page" is finally
814 * removed from global LRU.
816 mz
= page_cgroup_zoneinfo(pc
);
817 /* huge page split is done under lru_lock. so, we have no races. */
818 MEM_CGROUP_ZSTAT(mz
, lru
) -= 1 << compound_order(page
);
819 if (mem_cgroup_is_root(pc
->mem_cgroup
))
821 VM_BUG_ON(list_empty(&pc
->lru
));
822 list_del_init(&pc
->lru
);
825 void mem_cgroup_del_lru(struct page
*page
)
827 mem_cgroup_del_lru_list(page
, page_lru(page
));
830 void mem_cgroup_rotate_lru_list(struct page
*page
, enum lru_list lru
)
832 struct mem_cgroup_per_zone
*mz
;
833 struct page_cgroup
*pc
;
835 if (mem_cgroup_disabled())
838 pc
= lookup_page_cgroup(page
);
839 /* unused or root page is not rotated. */
840 if (!PageCgroupUsed(pc
))
842 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
844 if (mem_cgroup_is_root(pc
->mem_cgroup
))
846 mz
= page_cgroup_zoneinfo(pc
);
847 list_move(&pc
->lru
, &mz
->lists
[lru
]);
850 void mem_cgroup_add_lru_list(struct page
*page
, enum lru_list lru
)
852 struct page_cgroup
*pc
;
853 struct mem_cgroup_per_zone
*mz
;
855 if (mem_cgroup_disabled())
857 pc
= lookup_page_cgroup(page
);
858 VM_BUG_ON(PageCgroupAcctLRU(pc
));
859 if (!PageCgroupUsed(pc
))
861 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
863 mz
= page_cgroup_zoneinfo(pc
);
864 /* huge page split is done under lru_lock. so, we have no races. */
865 MEM_CGROUP_ZSTAT(mz
, lru
) += 1 << compound_order(page
);
866 SetPageCgroupAcctLRU(pc
);
867 if (mem_cgroup_is_root(pc
->mem_cgroup
))
869 list_add(&pc
->lru
, &mz
->lists
[lru
]);
873 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
874 * lru because the page may.be reused after it's fully uncharged (because of
875 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
876 * it again. This function is only used to charge SwapCache. It's done under
877 * lock_page and expected that zone->lru_lock is never held.
879 static void mem_cgroup_lru_del_before_commit_swapcache(struct page
*page
)
882 struct zone
*zone
= page_zone(page
);
883 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
885 spin_lock_irqsave(&zone
->lru_lock
, flags
);
887 * Forget old LRU when this page_cgroup is *not* used. This Used bit
888 * is guarded by lock_page() because the page is SwapCache.
890 if (!PageCgroupUsed(pc
))
891 mem_cgroup_del_lru_list(page
, page_lru(page
));
892 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
895 static void mem_cgroup_lru_add_after_commit_swapcache(struct page
*page
)
898 struct zone
*zone
= page_zone(page
);
899 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
901 spin_lock_irqsave(&zone
->lru_lock
, flags
);
902 /* link when the page is linked to LRU but page_cgroup isn't */
903 if (PageLRU(page
) && !PageCgroupAcctLRU(pc
))
904 mem_cgroup_add_lru_list(page
, page_lru(page
));
905 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
909 void mem_cgroup_move_lists(struct page
*page
,
910 enum lru_list from
, enum lru_list to
)
912 if (mem_cgroup_disabled())
914 mem_cgroup_del_lru_list(page
, from
);
915 mem_cgroup_add_lru_list(page
, to
);
918 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
921 struct mem_cgroup
*curr
= NULL
;
922 struct task_struct
*p
;
924 p
= find_lock_task_mm(task
);
927 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
932 * We should check use_hierarchy of "mem" not "curr". Because checking
933 * use_hierarchy of "curr" here make this function true if hierarchy is
934 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
935 * hierarchy(even if use_hierarchy is disabled in "mem").
937 if (mem
->use_hierarchy
)
938 ret
= css_is_ancestor(&curr
->css
, &mem
->css
);
945 static int calc_inactive_ratio(struct mem_cgroup
*memcg
, unsigned long *present_pages
)
947 unsigned long active
;
948 unsigned long inactive
;
950 unsigned long inactive_ratio
;
952 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_ANON
);
953 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_ANON
);
955 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
957 inactive_ratio
= int_sqrt(10 * gb
);
962 present_pages
[0] = inactive
;
963 present_pages
[1] = active
;
966 return inactive_ratio
;
969 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup
*memcg
)
971 unsigned long active
;
972 unsigned long inactive
;
973 unsigned long present_pages
[2];
974 unsigned long inactive_ratio
;
976 inactive_ratio
= calc_inactive_ratio(memcg
, present_pages
);
978 inactive
= present_pages
[0];
979 active
= present_pages
[1];
981 if (inactive
* inactive_ratio
< active
)
987 int mem_cgroup_inactive_file_is_low(struct mem_cgroup
*memcg
)
989 unsigned long active
;
990 unsigned long inactive
;
992 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_FILE
);
993 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_FILE
);
995 return (active
> inactive
);
998 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup
*memcg
,
1002 int nid
= zone_to_nid(zone
);
1003 int zid
= zone_idx(zone
);
1004 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
1006 return MEM_CGROUP_ZSTAT(mz
, lru
);
1009 struct zone_reclaim_stat
*mem_cgroup_get_reclaim_stat(struct mem_cgroup
*memcg
,
1012 int nid
= zone_to_nid(zone
);
1013 int zid
= zone_idx(zone
);
1014 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
1016 return &mz
->reclaim_stat
;
1019 struct zone_reclaim_stat
*
1020 mem_cgroup_get_reclaim_stat_from_page(struct page
*page
)
1022 struct page_cgroup
*pc
;
1023 struct mem_cgroup_per_zone
*mz
;
1025 if (mem_cgroup_disabled())
1028 pc
= lookup_page_cgroup(page
);
1029 if (!PageCgroupUsed(pc
))
1031 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1033 mz
= page_cgroup_zoneinfo(pc
);
1037 return &mz
->reclaim_stat
;
1040 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
1041 struct list_head
*dst
,
1042 unsigned long *scanned
, int order
,
1043 int mode
, struct zone
*z
,
1044 struct mem_cgroup
*mem_cont
,
1045 int active
, int file
)
1047 unsigned long nr_taken
= 0;
1051 struct list_head
*src
;
1052 struct page_cgroup
*pc
, *tmp
;
1053 int nid
= zone_to_nid(z
);
1054 int zid
= zone_idx(z
);
1055 struct mem_cgroup_per_zone
*mz
;
1056 int lru
= LRU_FILE
* file
+ active
;
1060 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
1061 src
= &mz
->lists
[lru
];
1064 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
1065 if (scan
>= nr_to_scan
)
1069 if (unlikely(!PageCgroupUsed(pc
)))
1071 if (unlikely(!PageLRU(page
)))
1075 ret
= __isolate_lru_page(page
, mode
, file
);
1078 list_move(&page
->lru
, dst
);
1079 mem_cgroup_del_lru(page
);
1080 nr_taken
+= hpage_nr_pages(page
);
1083 /* we don't affect global LRU but rotate in our LRU */
1084 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
1093 trace_mm_vmscan_memcg_isolate(0, nr_to_scan
, scan
, nr_taken
,
1099 #define mem_cgroup_from_res_counter(counter, member) \
1100 container_of(counter, struct mem_cgroup, member)
1102 static bool mem_cgroup_check_under_limit(struct mem_cgroup
*mem
)
1104 if (do_swap_account
) {
1105 if (res_counter_check_under_limit(&mem
->res
) &&
1106 res_counter_check_under_limit(&mem
->memsw
))
1109 if (res_counter_check_under_limit(&mem
->res
))
1114 static unsigned int get_swappiness(struct mem_cgroup
*memcg
)
1116 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1117 unsigned int swappiness
;
1120 if (cgrp
->parent
== NULL
)
1121 return vm_swappiness
;
1123 spin_lock(&memcg
->reclaim_param_lock
);
1124 swappiness
= memcg
->swappiness
;
1125 spin_unlock(&memcg
->reclaim_param_lock
);
1130 static void mem_cgroup_start_move(struct mem_cgroup
*mem
)
1135 spin_lock(&mem
->pcp_counter_lock
);
1136 for_each_online_cpu(cpu
)
1137 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) += 1;
1138 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] += 1;
1139 spin_unlock(&mem
->pcp_counter_lock
);
1145 static void mem_cgroup_end_move(struct mem_cgroup
*mem
)
1152 spin_lock(&mem
->pcp_counter_lock
);
1153 for_each_online_cpu(cpu
)
1154 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) -= 1;
1155 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] -= 1;
1156 spin_unlock(&mem
->pcp_counter_lock
);
1160 * 2 routines for checking "mem" is under move_account() or not.
1162 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1163 * for avoiding race in accounting. If true,
1164 * pc->mem_cgroup may be overwritten.
1166 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1167 * under hierarchy of moving cgroups. This is for
1168 * waiting at hith-memory prressure caused by "move".
1171 static bool mem_cgroup_stealed(struct mem_cgroup
*mem
)
1173 VM_BUG_ON(!rcu_read_lock_held());
1174 return this_cpu_read(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
]) > 0;
1177 static bool mem_cgroup_under_move(struct mem_cgroup
*mem
)
1179 struct mem_cgroup
*from
;
1180 struct mem_cgroup
*to
;
1183 * Unlike task_move routines, we access mc.to, mc.from not under
1184 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1186 spin_lock(&mc
.lock
);
1191 if (from
== mem
|| to
== mem
1192 || (mem
->use_hierarchy
&& css_is_ancestor(&from
->css
, &mem
->css
))
1193 || (mem
->use_hierarchy
&& css_is_ancestor(&to
->css
, &mem
->css
)))
1196 spin_unlock(&mc
.lock
);
1200 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*mem
)
1202 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1203 if (mem_cgroup_under_move(mem
)) {
1205 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1206 /* moving charge context might have finished. */
1209 finish_wait(&mc
.waitq
, &wait
);
1217 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1218 * @memcg: The memory cgroup that went over limit
1219 * @p: Task that is going to be killed
1221 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1224 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1226 struct cgroup
*task_cgrp
;
1227 struct cgroup
*mem_cgrp
;
1229 * Need a buffer in BSS, can't rely on allocations. The code relies
1230 * on the assumption that OOM is serialized for memory controller.
1231 * If this assumption is broken, revisit this code.
1233 static char memcg_name
[PATH_MAX
];
1242 mem_cgrp
= memcg
->css
.cgroup
;
1243 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1245 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1248 * Unfortunately, we are unable to convert to a useful name
1249 * But we'll still print out the usage information
1256 printk(KERN_INFO
"Task in %s killed", memcg_name
);
1259 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1267 * Continues from above, so we don't need an KERN_ level
1269 printk(KERN_CONT
" as a result of limit of %s\n", memcg_name
);
1272 printk(KERN_INFO
"memory: usage %llukB, limit %llukB, failcnt %llu\n",
1273 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1274 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1275 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1276 printk(KERN_INFO
"memory+swap: usage %llukB, limit %llukB, "
1278 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1279 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1280 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1284 * This function returns the number of memcg under hierarchy tree. Returns
1285 * 1(self count) if no children.
1287 static int mem_cgroup_count_children(struct mem_cgroup
*mem
)
1290 struct mem_cgroup
*iter
;
1292 for_each_mem_cgroup_tree(iter
, mem
)
1298 * Return the memory (and swap, if configured) limit for a memcg.
1300 u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1305 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1306 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1308 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1310 * If memsw is finite and limits the amount of swap space available
1311 * to this memcg, return that limit.
1313 return min(limit
, memsw
);
1317 * Visit the first child (need not be the first child as per the ordering
1318 * of the cgroup list, since we track last_scanned_child) of @mem and use
1319 * that to reclaim free pages from.
1321 static struct mem_cgroup
*
1322 mem_cgroup_select_victim(struct mem_cgroup
*root_mem
)
1324 struct mem_cgroup
*ret
= NULL
;
1325 struct cgroup_subsys_state
*css
;
1328 if (!root_mem
->use_hierarchy
) {
1329 css_get(&root_mem
->css
);
1335 nextid
= root_mem
->last_scanned_child
+ 1;
1336 css
= css_get_next(&mem_cgroup_subsys
, nextid
, &root_mem
->css
,
1338 if (css
&& css_tryget(css
))
1339 ret
= container_of(css
, struct mem_cgroup
, css
);
1342 /* Updates scanning parameter */
1343 spin_lock(&root_mem
->reclaim_param_lock
);
1345 /* this means start scan from ID:1 */
1346 root_mem
->last_scanned_child
= 0;
1348 root_mem
->last_scanned_child
= found
;
1349 spin_unlock(&root_mem
->reclaim_param_lock
);
1356 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1357 * we reclaimed from, so that we don't end up penalizing one child extensively
1358 * based on its position in the children list.
1360 * root_mem is the original ancestor that we've been reclaim from.
1362 * We give up and return to the caller when we visit root_mem twice.
1363 * (other groups can be removed while we're walking....)
1365 * If shrink==true, for avoiding to free too much, this returns immedieately.
1367 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup
*root_mem
,
1370 unsigned long reclaim_options
)
1372 struct mem_cgroup
*victim
;
1375 bool noswap
= reclaim_options
& MEM_CGROUP_RECLAIM_NOSWAP
;
1376 bool shrink
= reclaim_options
& MEM_CGROUP_RECLAIM_SHRINK
;
1377 bool check_soft
= reclaim_options
& MEM_CGROUP_RECLAIM_SOFT
;
1378 unsigned long excess
= mem_cgroup_get_excess(root_mem
);
1380 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1381 if (root_mem
->memsw_is_minimum
)
1385 victim
= mem_cgroup_select_victim(root_mem
);
1386 if (victim
== root_mem
) {
1389 drain_all_stock_async();
1392 * If we have not been able to reclaim
1393 * anything, it might because there are
1394 * no reclaimable pages under this hierarchy
1396 if (!check_soft
|| !total
) {
1397 css_put(&victim
->css
);
1401 * We want to do more targetted reclaim.
1402 * excess >> 2 is not to excessive so as to
1403 * reclaim too much, nor too less that we keep
1404 * coming back to reclaim from this cgroup
1406 if (total
>= (excess
>> 2) ||
1407 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
)) {
1408 css_put(&victim
->css
);
1413 if (!mem_cgroup_local_usage(victim
)) {
1414 /* this cgroup's local usage == 0 */
1415 css_put(&victim
->css
);
1418 /* we use swappiness of local cgroup */
1420 ret
= mem_cgroup_shrink_node_zone(victim
, gfp_mask
,
1421 noswap
, get_swappiness(victim
), zone
);
1423 ret
= try_to_free_mem_cgroup_pages(victim
, gfp_mask
,
1424 noswap
, get_swappiness(victim
));
1425 css_put(&victim
->css
);
1427 * At shrinking usage, we can't check we should stop here or
1428 * reclaim more. It's depends on callers. last_scanned_child
1429 * will work enough for keeping fairness under tree.
1435 if (res_counter_check_under_soft_limit(&root_mem
->res
))
1437 } else if (mem_cgroup_check_under_limit(root_mem
))
1444 * Check OOM-Killer is already running under our hierarchy.
1445 * If someone is running, return false.
1447 static bool mem_cgroup_oom_lock(struct mem_cgroup
*mem
)
1449 int x
, lock_count
= 0;
1450 struct mem_cgroup
*iter
;
1452 for_each_mem_cgroup_tree(iter
, mem
) {
1453 x
= atomic_inc_return(&iter
->oom_lock
);
1454 lock_count
= max(x
, lock_count
);
1457 if (lock_count
== 1)
1462 static int mem_cgroup_oom_unlock(struct mem_cgroup
*mem
)
1464 struct mem_cgroup
*iter
;
1467 * When a new child is created while the hierarchy is under oom,
1468 * mem_cgroup_oom_lock() may not be called. We have to use
1469 * atomic_add_unless() here.
1471 for_each_mem_cgroup_tree(iter
, mem
)
1472 atomic_add_unless(&iter
->oom_lock
, -1, 0);
1477 static DEFINE_MUTEX(memcg_oom_mutex
);
1478 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1480 struct oom_wait_info
{
1481 struct mem_cgroup
*mem
;
1485 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1486 unsigned mode
, int sync
, void *arg
)
1488 struct mem_cgroup
*wake_mem
= (struct mem_cgroup
*)arg
;
1489 struct oom_wait_info
*oom_wait_info
;
1491 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1493 if (oom_wait_info
->mem
== wake_mem
)
1495 /* if no hierarchy, no match */
1496 if (!oom_wait_info
->mem
->use_hierarchy
|| !wake_mem
->use_hierarchy
)
1499 * Both of oom_wait_info->mem and wake_mem are stable under us.
1500 * Then we can use css_is_ancestor without taking care of RCU.
1502 if (!css_is_ancestor(&oom_wait_info
->mem
->css
, &wake_mem
->css
) &&
1503 !css_is_ancestor(&wake_mem
->css
, &oom_wait_info
->mem
->css
))
1507 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1510 static void memcg_wakeup_oom(struct mem_cgroup
*mem
)
1512 /* for filtering, pass "mem" as argument. */
1513 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, mem
);
1516 static void memcg_oom_recover(struct mem_cgroup
*mem
)
1518 if (mem
&& atomic_read(&mem
->oom_lock
))
1519 memcg_wakeup_oom(mem
);
1523 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1525 bool mem_cgroup_handle_oom(struct mem_cgroup
*mem
, gfp_t mask
)
1527 struct oom_wait_info owait
;
1528 bool locked
, need_to_kill
;
1531 owait
.wait
.flags
= 0;
1532 owait
.wait
.func
= memcg_oom_wake_function
;
1533 owait
.wait
.private = current
;
1534 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1535 need_to_kill
= true;
1536 /* At first, try to OOM lock hierarchy under mem.*/
1537 mutex_lock(&memcg_oom_mutex
);
1538 locked
= mem_cgroup_oom_lock(mem
);
1540 * Even if signal_pending(), we can't quit charge() loop without
1541 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1542 * under OOM is always welcomed, use TASK_KILLABLE here.
1544 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1545 if (!locked
|| mem
->oom_kill_disable
)
1546 need_to_kill
= false;
1548 mem_cgroup_oom_notify(mem
);
1549 mutex_unlock(&memcg_oom_mutex
);
1552 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1553 mem_cgroup_out_of_memory(mem
, mask
);
1556 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1558 mutex_lock(&memcg_oom_mutex
);
1559 mem_cgroup_oom_unlock(mem
);
1560 memcg_wakeup_oom(mem
);
1561 mutex_unlock(&memcg_oom_mutex
);
1563 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
1565 /* Give chance to dying process */
1566 schedule_timeout(1);
1571 * Currently used to update mapped file statistics, but the routine can be
1572 * generalized to update other statistics as well.
1574 * Notes: Race condition
1576 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1577 * it tends to be costly. But considering some conditions, we doesn't need
1578 * to do so _always_.
1580 * Considering "charge", lock_page_cgroup() is not required because all
1581 * file-stat operations happen after a page is attached to radix-tree. There
1582 * are no race with "charge".
1584 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1585 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1586 * if there are race with "uncharge". Statistics itself is properly handled
1589 * Considering "move", this is an only case we see a race. To make the race
1590 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1591 * possibility of race condition. If there is, we take a lock.
1594 void mem_cgroup_update_page_stat(struct page
*page
,
1595 enum mem_cgroup_page_stat_item idx
, int val
)
1597 struct mem_cgroup
*mem
;
1598 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
1599 bool need_unlock
= false;
1600 unsigned long uninitialized_var(flags
);
1606 mem
= pc
->mem_cgroup
;
1607 if (unlikely(!mem
|| !PageCgroupUsed(pc
)))
1609 /* pc->mem_cgroup is unstable ? */
1610 if (unlikely(mem_cgroup_stealed(mem
)) || PageTransHuge(page
)) {
1611 /* take a lock against to access pc->mem_cgroup */
1612 move_lock_page_cgroup(pc
, &flags
);
1614 mem
= pc
->mem_cgroup
;
1615 if (!mem
|| !PageCgroupUsed(pc
))
1620 case MEMCG_NR_FILE_MAPPED
:
1622 SetPageCgroupFileMapped(pc
);
1623 else if (!page_mapped(page
))
1624 ClearPageCgroupFileMapped(pc
);
1625 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
1631 this_cpu_add(mem
->stat
->count
[idx
], val
);
1634 if (unlikely(need_unlock
))
1635 move_unlock_page_cgroup(pc
, &flags
);
1639 EXPORT_SYMBOL(mem_cgroup_update_page_stat
);
1642 * size of first charge trial. "32" comes from vmscan.c's magic value.
1643 * TODO: maybe necessary to use big numbers in big irons.
1645 #define CHARGE_SIZE (32 * PAGE_SIZE)
1646 struct memcg_stock_pcp
{
1647 struct mem_cgroup
*cached
; /* this never be root cgroup */
1649 struct work_struct work
;
1651 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1652 static atomic_t memcg_drain_count
;
1655 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1656 * from local stock and true is returned. If the stock is 0 or charges from a
1657 * cgroup which is not current target, returns false. This stock will be
1660 static bool consume_stock(struct mem_cgroup
*mem
)
1662 struct memcg_stock_pcp
*stock
;
1665 stock
= &get_cpu_var(memcg_stock
);
1666 if (mem
== stock
->cached
&& stock
->charge
)
1667 stock
->charge
-= PAGE_SIZE
;
1668 else /* need to call res_counter_charge */
1670 put_cpu_var(memcg_stock
);
1675 * Returns stocks cached in percpu to res_counter and reset cached information.
1677 static void drain_stock(struct memcg_stock_pcp
*stock
)
1679 struct mem_cgroup
*old
= stock
->cached
;
1681 if (stock
->charge
) {
1682 res_counter_uncharge(&old
->res
, stock
->charge
);
1683 if (do_swap_account
)
1684 res_counter_uncharge(&old
->memsw
, stock
->charge
);
1686 stock
->cached
= NULL
;
1691 * This must be called under preempt disabled or must be called by
1692 * a thread which is pinned to local cpu.
1694 static void drain_local_stock(struct work_struct
*dummy
)
1696 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
1701 * Cache charges(val) which is from res_counter, to local per_cpu area.
1702 * This will be consumed by consume_stock() function, later.
1704 static void refill_stock(struct mem_cgroup
*mem
, int val
)
1706 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1708 if (stock
->cached
!= mem
) { /* reset if necessary */
1710 stock
->cached
= mem
;
1712 stock
->charge
+= val
;
1713 put_cpu_var(memcg_stock
);
1717 * Tries to drain stocked charges in other cpus. This function is asynchronous
1718 * and just put a work per cpu for draining localy on each cpu. Caller can
1719 * expects some charges will be back to res_counter later but cannot wait for
1722 static void drain_all_stock_async(void)
1725 /* This function is for scheduling "drain" in asynchronous way.
1726 * The result of "drain" is not directly handled by callers. Then,
1727 * if someone is calling drain, we don't have to call drain more.
1728 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1729 * there is a race. We just do loose check here.
1731 if (atomic_read(&memcg_drain_count
))
1733 /* Notify other cpus that system-wide "drain" is running */
1734 atomic_inc(&memcg_drain_count
);
1736 for_each_online_cpu(cpu
) {
1737 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1738 schedule_work_on(cpu
, &stock
->work
);
1741 atomic_dec(&memcg_drain_count
);
1742 /* We don't wait for flush_work */
1745 /* This is a synchronous drain interface. */
1746 static void drain_all_stock_sync(void)
1748 /* called when force_empty is called */
1749 atomic_inc(&memcg_drain_count
);
1750 schedule_on_each_cpu(drain_local_stock
);
1751 atomic_dec(&memcg_drain_count
);
1755 * This function drains percpu counter value from DEAD cpu and
1756 * move it to local cpu. Note that this function can be preempted.
1758 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*mem
, int cpu
)
1762 spin_lock(&mem
->pcp_counter_lock
);
1763 for (i
= 0; i
< MEM_CGROUP_STAT_DATA
; i
++) {
1764 s64 x
= per_cpu(mem
->stat
->count
[i
], cpu
);
1766 per_cpu(mem
->stat
->count
[i
], cpu
) = 0;
1767 mem
->nocpu_base
.count
[i
] += x
;
1769 /* need to clear ON_MOVE value, works as a kind of lock. */
1770 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) = 0;
1771 spin_unlock(&mem
->pcp_counter_lock
);
1774 static void synchronize_mem_cgroup_on_move(struct mem_cgroup
*mem
, int cpu
)
1776 int idx
= MEM_CGROUP_ON_MOVE
;
1778 spin_lock(&mem
->pcp_counter_lock
);
1779 per_cpu(mem
->stat
->count
[idx
], cpu
) = mem
->nocpu_base
.count
[idx
];
1780 spin_unlock(&mem
->pcp_counter_lock
);
1783 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1784 unsigned long action
,
1787 int cpu
= (unsigned long)hcpu
;
1788 struct memcg_stock_pcp
*stock
;
1789 struct mem_cgroup
*iter
;
1791 if ((action
== CPU_ONLINE
)) {
1792 for_each_mem_cgroup_all(iter
)
1793 synchronize_mem_cgroup_on_move(iter
, cpu
);
1797 if ((action
!= CPU_DEAD
) || action
!= CPU_DEAD_FROZEN
)
1800 for_each_mem_cgroup_all(iter
)
1801 mem_cgroup_drain_pcp_counter(iter
, cpu
);
1803 stock
= &per_cpu(memcg_stock
, cpu
);
1809 /* See __mem_cgroup_try_charge() for details */
1811 CHARGE_OK
, /* success */
1812 CHARGE_RETRY
, /* need to retry but retry is not bad */
1813 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
1814 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
1815 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
1818 static int __mem_cgroup_do_charge(struct mem_cgroup
*mem
, gfp_t gfp_mask
,
1819 int csize
, bool oom_check
)
1821 struct mem_cgroup
*mem_over_limit
;
1822 struct res_counter
*fail_res
;
1823 unsigned long flags
= 0;
1826 ret
= res_counter_charge(&mem
->res
, csize
, &fail_res
);
1829 if (!do_swap_account
)
1831 ret
= res_counter_charge(&mem
->memsw
, csize
, &fail_res
);
1835 res_counter_uncharge(&mem
->res
, csize
);
1836 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
1837 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
1839 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
1841 * csize can be either a huge page (HPAGE_SIZE), a batch of
1842 * regular pages (CHARGE_SIZE), or a single regular page
1845 * Never reclaim on behalf of optional batching, retry with a
1846 * single page instead.
1848 if (csize
== CHARGE_SIZE
)
1849 return CHARGE_RETRY
;
1851 if (!(gfp_mask
& __GFP_WAIT
))
1852 return CHARGE_WOULDBLOCK
;
1854 ret
= mem_cgroup_hierarchical_reclaim(mem_over_limit
, NULL
,
1857 * try_to_free_mem_cgroup_pages() might not give us a full
1858 * picture of reclaim. Some pages are reclaimed and might be
1859 * moved to swap cache or just unmapped from the cgroup.
1860 * Check the limit again to see if the reclaim reduced the
1861 * current usage of the cgroup before giving up
1863 if (ret
|| mem_cgroup_check_under_limit(mem_over_limit
))
1864 return CHARGE_RETRY
;
1867 * At task move, charge accounts can be doubly counted. So, it's
1868 * better to wait until the end of task_move if something is going on.
1870 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1871 return CHARGE_RETRY
;
1873 /* If we don't need to call oom-killer at el, return immediately */
1875 return CHARGE_NOMEM
;
1877 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
))
1878 return CHARGE_OOM_DIE
;
1880 return CHARGE_RETRY
;
1884 * Unlike exported interface, "oom" parameter is added. if oom==true,
1885 * oom-killer can be invoked.
1887 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
1889 struct mem_cgroup
**memcg
, bool oom
,
1892 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1893 struct mem_cgroup
*mem
= NULL
;
1895 int csize
= max(CHARGE_SIZE
, (unsigned long) page_size
);
1898 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1899 * in system level. So, allow to go ahead dying process in addition to
1902 if (unlikely(test_thread_flag(TIF_MEMDIE
)
1903 || fatal_signal_pending(current
)))
1907 * We always charge the cgroup the mm_struct belongs to.
1908 * The mm_struct's mem_cgroup changes on task migration if the
1909 * thread group leader migrates. It's possible that mm is not
1910 * set, if so charge the init_mm (happens for pagecache usage).
1915 if (*memcg
) { /* css should be a valid one */
1917 VM_BUG_ON(css_is_removed(&mem
->css
));
1918 if (mem_cgroup_is_root(mem
))
1920 if (page_size
== PAGE_SIZE
&& consume_stock(mem
))
1924 struct task_struct
*p
;
1927 p
= rcu_dereference(mm
->owner
);
1929 * Because we don't have task_lock(), "p" can exit.
1930 * In that case, "mem" can point to root or p can be NULL with
1931 * race with swapoff. Then, we have small risk of mis-accouning.
1932 * But such kind of mis-account by race always happens because
1933 * we don't have cgroup_mutex(). It's overkill and we allo that
1935 * (*) swapoff at el will charge against mm-struct not against
1936 * task-struct. So, mm->owner can be NULL.
1938 mem
= mem_cgroup_from_task(p
);
1939 if (!mem
|| mem_cgroup_is_root(mem
)) {
1943 if (page_size
== PAGE_SIZE
&& consume_stock(mem
)) {
1945 * It seems dagerous to access memcg without css_get().
1946 * But considering how consume_stok works, it's not
1947 * necessary. If consume_stock success, some charges
1948 * from this memcg are cached on this cpu. So, we
1949 * don't need to call css_get()/css_tryget() before
1950 * calling consume_stock().
1955 /* after here, we may be blocked. we need to get refcnt */
1956 if (!css_tryget(&mem
->css
)) {
1966 /* If killed, bypass charge */
1967 if (fatal_signal_pending(current
)) {
1973 if (oom
&& !nr_oom_retries
) {
1975 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1978 ret
= __mem_cgroup_do_charge(mem
, gfp_mask
, csize
, oom_check
);
1983 case CHARGE_RETRY
: /* not in OOM situation but retry */
1988 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
1991 case CHARGE_NOMEM
: /* OOM routine works */
1996 /* If oom, we never return -ENOMEM */
1999 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2003 } while (ret
!= CHARGE_OK
);
2005 if (csize
> page_size
)
2006 refill_stock(mem
, csize
- page_size
);
2020 * Somemtimes we have to undo a charge we got by try_charge().
2021 * This function is for that and do uncharge, put css's refcnt.
2022 * gotten by try_charge().
2024 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2025 unsigned long count
)
2027 if (!mem_cgroup_is_root(mem
)) {
2028 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
2029 if (do_swap_account
)
2030 res_counter_uncharge(&mem
->memsw
, PAGE_SIZE
* count
);
2034 static void mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2037 __mem_cgroup_cancel_charge(mem
, page_size
>> PAGE_SHIFT
);
2041 * A helper function to get mem_cgroup from ID. must be called under
2042 * rcu_read_lock(). The caller must check css_is_removed() or some if
2043 * it's concern. (dropping refcnt from swap can be called against removed
2046 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2048 struct cgroup_subsys_state
*css
;
2050 /* ID 0 is unused ID */
2053 css
= css_lookup(&mem_cgroup_subsys
, id
);
2056 return container_of(css
, struct mem_cgroup
, css
);
2059 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2061 struct mem_cgroup
*mem
= NULL
;
2062 struct page_cgroup
*pc
;
2066 VM_BUG_ON(!PageLocked(page
));
2068 pc
= lookup_page_cgroup(page
);
2069 lock_page_cgroup(pc
);
2070 if (PageCgroupUsed(pc
)) {
2071 mem
= pc
->mem_cgroup
;
2072 if (mem
&& !css_tryget(&mem
->css
))
2074 } else if (PageSwapCache(page
)) {
2075 ent
.val
= page_private(page
);
2076 id
= lookup_swap_cgroup(ent
);
2078 mem
= mem_cgroup_lookup(id
);
2079 if (mem
&& !css_tryget(&mem
->css
))
2083 unlock_page_cgroup(pc
);
2087 static void __mem_cgroup_commit_charge(struct mem_cgroup
*mem
,
2088 struct page_cgroup
*pc
,
2089 enum charge_type ctype
,
2092 int nr_pages
= page_size
>> PAGE_SHIFT
;
2094 /* try_charge() can return NULL to *memcg, taking care of it. */
2098 lock_page_cgroup(pc
);
2099 if (unlikely(PageCgroupUsed(pc
))) {
2100 unlock_page_cgroup(pc
);
2101 mem_cgroup_cancel_charge(mem
, page_size
);
2105 * we don't need page_cgroup_lock about tail pages, becase they are not
2106 * accessed by any other context at this point.
2108 pc
->mem_cgroup
= mem
;
2110 * We access a page_cgroup asynchronously without lock_page_cgroup().
2111 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2112 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2113 * before USED bit, we need memory barrier here.
2114 * See mem_cgroup_add_lru_list(), etc.
2118 case MEM_CGROUP_CHARGE_TYPE_CACHE
:
2119 case MEM_CGROUP_CHARGE_TYPE_SHMEM
:
2120 SetPageCgroupCache(pc
);
2121 SetPageCgroupUsed(pc
);
2123 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2124 ClearPageCgroupCache(pc
);
2125 SetPageCgroupUsed(pc
);
2131 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), nr_pages
);
2132 unlock_page_cgroup(pc
);
2134 * "charge_statistics" updated event counter. Then, check it.
2135 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2136 * if they exceeds softlimit.
2138 memcg_check_events(mem
, pc
->page
);
2141 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2143 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2144 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2146 * Because tail pages are not marked as "used", set it. We're under
2147 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2149 void mem_cgroup_split_huge_fixup(struct page
*head
, struct page
*tail
)
2151 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
2152 struct page_cgroup
*tail_pc
= lookup_page_cgroup(tail
);
2153 unsigned long flags
;
2155 if (mem_cgroup_disabled())
2158 * We have no races with charge/uncharge but will have races with
2159 * page state accounting.
2161 move_lock_page_cgroup(head_pc
, &flags
);
2163 tail_pc
->mem_cgroup
= head_pc
->mem_cgroup
;
2164 smp_wmb(); /* see __commit_charge() */
2165 if (PageCgroupAcctLRU(head_pc
)) {
2167 struct mem_cgroup_per_zone
*mz
;
2170 * LRU flags cannot be copied because we need to add tail
2171 *.page to LRU by generic call and our hook will be called.
2172 * We hold lru_lock, then, reduce counter directly.
2174 lru
= page_lru(head
);
2175 mz
= page_cgroup_zoneinfo(head_pc
);
2176 MEM_CGROUP_ZSTAT(mz
, lru
) -= 1;
2178 tail_pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
2179 move_unlock_page_cgroup(head_pc
, &flags
);
2184 * __mem_cgroup_move_account - move account of the page
2185 * @pc: page_cgroup of the page.
2186 * @from: mem_cgroup which the page is moved from.
2187 * @to: mem_cgroup which the page is moved to. @from != @to.
2188 * @uncharge: whether we should call uncharge and css_put against @from.
2190 * The caller must confirm following.
2191 * - page is not on LRU (isolate_page() is useful.)
2192 * - the pc is locked, used, and ->mem_cgroup points to @from.
2194 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2195 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2196 * true, this function does "uncharge" from old cgroup, but it doesn't if
2197 * @uncharge is false, so a caller should do "uncharge".
2200 static void __mem_cgroup_move_account(struct page_cgroup
*pc
,
2201 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool uncharge
,
2204 int nr_pages
= charge_size
>> PAGE_SHIFT
;
2206 VM_BUG_ON(from
== to
);
2207 VM_BUG_ON(PageLRU(pc
->page
));
2208 VM_BUG_ON(!page_is_cgroup_locked(pc
));
2209 VM_BUG_ON(!PageCgroupUsed(pc
));
2210 VM_BUG_ON(pc
->mem_cgroup
!= from
);
2212 if (PageCgroupFileMapped(pc
)) {
2213 /* Update mapped_file data for mem_cgroup */
2215 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
2216 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
2219 mem_cgroup_charge_statistics(from
, PageCgroupCache(pc
), -nr_pages
);
2221 /* This is not "cancel", but cancel_charge does all we need. */
2222 mem_cgroup_cancel_charge(from
, charge_size
);
2224 /* caller should have done css_get */
2225 pc
->mem_cgroup
= to
;
2226 mem_cgroup_charge_statistics(to
, PageCgroupCache(pc
), nr_pages
);
2228 * We charges against "to" which may not have any tasks. Then, "to"
2229 * can be under rmdir(). But in current implementation, caller of
2230 * this function is just force_empty() and move charge, so it's
2231 * garanteed that "to" is never removed. So, we don't check rmdir
2237 * check whether the @pc is valid for moving account and call
2238 * __mem_cgroup_move_account()
2240 static int mem_cgroup_move_account(struct page_cgroup
*pc
,
2241 struct mem_cgroup
*from
, struct mem_cgroup
*to
,
2242 bool uncharge
, int charge_size
)
2245 unsigned long flags
;
2247 * The page is isolated from LRU. So, collapse function
2248 * will not handle this page. But page splitting can happen.
2249 * Do this check under compound_page_lock(). The caller should
2252 if ((charge_size
> PAGE_SIZE
) && !PageTransHuge(pc
->page
))
2255 lock_page_cgroup(pc
);
2256 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== from
) {
2257 move_lock_page_cgroup(pc
, &flags
);
2258 __mem_cgroup_move_account(pc
, from
, to
, uncharge
, charge_size
);
2259 move_unlock_page_cgroup(pc
, &flags
);
2262 unlock_page_cgroup(pc
);
2266 memcg_check_events(to
, pc
->page
);
2267 memcg_check_events(from
, pc
->page
);
2272 * move charges to its parent.
2275 static int mem_cgroup_move_parent(struct page_cgroup
*pc
,
2276 struct mem_cgroup
*child
,
2279 struct page
*page
= pc
->page
;
2280 struct cgroup
*cg
= child
->css
.cgroup
;
2281 struct cgroup
*pcg
= cg
->parent
;
2282 struct mem_cgroup
*parent
;
2283 int page_size
= PAGE_SIZE
;
2284 unsigned long flags
;
2292 if (!get_page_unless_zero(page
))
2294 if (isolate_lru_page(page
))
2297 if (PageTransHuge(page
))
2298 page_size
= HPAGE_SIZE
;
2300 parent
= mem_cgroup_from_cont(pcg
);
2301 ret
= __mem_cgroup_try_charge(NULL
, gfp_mask
,
2302 &parent
, false, page_size
);
2306 if (page_size
> PAGE_SIZE
)
2307 flags
= compound_lock_irqsave(page
);
2309 ret
= mem_cgroup_move_account(pc
, child
, parent
, true, page_size
);
2311 mem_cgroup_cancel_charge(parent
, page_size
);
2313 if (page_size
> PAGE_SIZE
)
2314 compound_unlock_irqrestore(page
, flags
);
2316 putback_lru_page(page
);
2324 * Charge the memory controller for page usage.
2326 * 0 if the charge was successful
2327 * < 0 if the cgroup is over its limit
2329 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
2330 gfp_t gfp_mask
, enum charge_type ctype
)
2332 struct mem_cgroup
*mem
= NULL
;
2333 struct page_cgroup
*pc
;
2335 int page_size
= PAGE_SIZE
;
2337 if (PageTransHuge(page
)) {
2338 page_size
<<= compound_order(page
);
2339 VM_BUG_ON(!PageTransHuge(page
));
2342 pc
= lookup_page_cgroup(page
);
2343 /* can happen at boot */
2348 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, &mem
, true, page_size
);
2352 __mem_cgroup_commit_charge(mem
, pc
, ctype
, page_size
);
2356 int mem_cgroup_newpage_charge(struct page
*page
,
2357 struct mm_struct
*mm
, gfp_t gfp_mask
)
2359 if (mem_cgroup_disabled())
2362 * If already mapped, we don't have to account.
2363 * If page cache, page->mapping has address_space.
2364 * But page->mapping may have out-of-use anon_vma pointer,
2365 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2368 if (page_mapped(page
) || (page
->mapping
&& !PageAnon(page
)))
2372 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2373 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2377 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2378 enum charge_type ctype
);
2380 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
2385 if (mem_cgroup_disabled())
2387 if (PageCompound(page
))
2390 * Corner case handling. This is called from add_to_page_cache()
2391 * in usual. But some FS (shmem) precharges this page before calling it
2392 * and call add_to_page_cache() with GFP_NOWAIT.
2394 * For GFP_NOWAIT case, the page may be pre-charged before calling
2395 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2396 * charge twice. (It works but has to pay a bit larger cost.)
2397 * And when the page is SwapCache, it should take swap information
2398 * into account. This is under lock_page() now.
2400 if (!(gfp_mask
& __GFP_WAIT
)) {
2401 struct page_cgroup
*pc
;
2403 pc
= lookup_page_cgroup(page
);
2406 lock_page_cgroup(pc
);
2407 if (PageCgroupUsed(pc
)) {
2408 unlock_page_cgroup(pc
);
2411 unlock_page_cgroup(pc
);
2417 if (page_is_file_cache(page
))
2418 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2419 MEM_CGROUP_CHARGE_TYPE_CACHE
);
2422 if (PageSwapCache(page
)) {
2423 struct mem_cgroup
*mem
= NULL
;
2425 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
2427 __mem_cgroup_commit_charge_swapin(page
, mem
,
2428 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2430 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2431 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2437 * While swap-in, try_charge -> commit or cancel, the page is locked.
2438 * And when try_charge() successfully returns, one refcnt to memcg without
2439 * struct page_cgroup is acquired. This refcnt will be consumed by
2440 * "commit()" or removed by "cancel()"
2442 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
2444 gfp_t mask
, struct mem_cgroup
**ptr
)
2446 struct mem_cgroup
*mem
;
2449 if (mem_cgroup_disabled())
2452 if (!do_swap_account
)
2455 * A racing thread's fault, or swapoff, may have already updated
2456 * the pte, and even removed page from swap cache: in those cases
2457 * do_swap_page()'s pte_same() test will fail; but there's also a
2458 * KSM case which does need to charge the page.
2460 if (!PageSwapCache(page
))
2462 mem
= try_get_mem_cgroup_from_page(page
);
2466 ret
= __mem_cgroup_try_charge(NULL
, mask
, ptr
, true, PAGE_SIZE
);
2472 return __mem_cgroup_try_charge(mm
, mask
, ptr
, true, PAGE_SIZE
);
2476 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2477 enum charge_type ctype
)
2479 struct page_cgroup
*pc
;
2481 if (mem_cgroup_disabled())
2485 cgroup_exclude_rmdir(&ptr
->css
);
2486 pc
= lookup_page_cgroup(page
);
2487 mem_cgroup_lru_del_before_commit_swapcache(page
);
2488 __mem_cgroup_commit_charge(ptr
, pc
, ctype
, PAGE_SIZE
);
2489 mem_cgroup_lru_add_after_commit_swapcache(page
);
2491 * Now swap is on-memory. This means this page may be
2492 * counted both as mem and swap....double count.
2493 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2494 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2495 * may call delete_from_swap_cache() before reach here.
2497 if (do_swap_account
&& PageSwapCache(page
)) {
2498 swp_entry_t ent
= {.val
= page_private(page
)};
2500 struct mem_cgroup
*memcg
;
2502 id
= swap_cgroup_record(ent
, 0);
2504 memcg
= mem_cgroup_lookup(id
);
2507 * This recorded memcg can be obsolete one. So, avoid
2508 * calling css_tryget
2510 if (!mem_cgroup_is_root(memcg
))
2511 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2512 mem_cgroup_swap_statistics(memcg
, false);
2513 mem_cgroup_put(memcg
);
2518 * At swapin, we may charge account against cgroup which has no tasks.
2519 * So, rmdir()->pre_destroy() can be called while we do this charge.
2520 * In that case, we need to call pre_destroy() again. check it here.
2522 cgroup_release_and_wakeup_rmdir(&ptr
->css
);
2525 void mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
)
2527 __mem_cgroup_commit_charge_swapin(page
, ptr
,
2528 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2531 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*mem
)
2533 if (mem_cgroup_disabled())
2537 mem_cgroup_cancel_charge(mem
, PAGE_SIZE
);
2541 __do_uncharge(struct mem_cgroup
*mem
, const enum charge_type ctype
,
2544 struct memcg_batch_info
*batch
= NULL
;
2545 bool uncharge_memsw
= true;
2546 /* If swapout, usage of swap doesn't decrease */
2547 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
2548 uncharge_memsw
= false;
2550 batch
= ¤t
->memcg_batch
;
2552 * In usual, we do css_get() when we remember memcg pointer.
2553 * But in this case, we keep res->usage until end of a series of
2554 * uncharges. Then, it's ok to ignore memcg's refcnt.
2559 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2560 * In those cases, all pages freed continously can be expected to be in
2561 * the same cgroup and we have chance to coalesce uncharges.
2562 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2563 * because we want to do uncharge as soon as possible.
2566 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
2567 goto direct_uncharge
;
2569 if (page_size
!= PAGE_SIZE
)
2570 goto direct_uncharge
;
2573 * In typical case, batch->memcg == mem. This means we can
2574 * merge a series of uncharges to an uncharge of res_counter.
2575 * If not, we uncharge res_counter ony by one.
2577 if (batch
->memcg
!= mem
)
2578 goto direct_uncharge
;
2579 /* remember freed charge and uncharge it later */
2580 batch
->bytes
+= PAGE_SIZE
;
2582 batch
->memsw_bytes
+= PAGE_SIZE
;
2585 res_counter_uncharge(&mem
->res
, page_size
);
2587 res_counter_uncharge(&mem
->memsw
, page_size
);
2588 if (unlikely(batch
->memcg
!= mem
))
2589 memcg_oom_recover(mem
);
2594 * uncharge if !page_mapped(page)
2596 static struct mem_cgroup
*
2597 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
)
2600 struct page_cgroup
*pc
;
2601 struct mem_cgroup
*mem
= NULL
;
2602 int page_size
= PAGE_SIZE
;
2604 if (mem_cgroup_disabled())
2607 if (PageSwapCache(page
))
2610 if (PageTransHuge(page
)) {
2611 page_size
<<= compound_order(page
);
2612 VM_BUG_ON(!PageTransHuge(page
));
2615 count
= page_size
>> PAGE_SHIFT
;
2617 * Check if our page_cgroup is valid
2619 pc
= lookup_page_cgroup(page
);
2620 if (unlikely(!pc
|| !PageCgroupUsed(pc
)))
2623 lock_page_cgroup(pc
);
2625 mem
= pc
->mem_cgroup
;
2627 if (!PageCgroupUsed(pc
))
2631 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2632 case MEM_CGROUP_CHARGE_TYPE_DROP
:
2633 /* See mem_cgroup_prepare_migration() */
2634 if (page_mapped(page
) || PageCgroupMigration(pc
))
2637 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
2638 if (!PageAnon(page
)) { /* Shared memory */
2639 if (page
->mapping
&& !page_is_file_cache(page
))
2641 } else if (page_mapped(page
)) /* Anon */
2648 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), -count
);
2650 ClearPageCgroupUsed(pc
);
2652 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2653 * freed from LRU. This is safe because uncharged page is expected not
2654 * to be reused (freed soon). Exception is SwapCache, it's handled by
2655 * special functions.
2658 unlock_page_cgroup(pc
);
2660 * even after unlock, we have mem->res.usage here and this memcg
2661 * will never be freed.
2663 memcg_check_events(mem
, page
);
2664 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
2665 mem_cgroup_swap_statistics(mem
, true);
2666 mem_cgroup_get(mem
);
2668 if (!mem_cgroup_is_root(mem
))
2669 __do_uncharge(mem
, ctype
, page_size
);
2674 unlock_page_cgroup(pc
);
2678 void mem_cgroup_uncharge_page(struct page
*page
)
2681 if (page_mapped(page
))
2683 if (page
->mapping
&& !PageAnon(page
))
2685 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2688 void mem_cgroup_uncharge_cache_page(struct page
*page
)
2690 VM_BUG_ON(page_mapped(page
));
2691 VM_BUG_ON(page
->mapping
);
2692 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
);
2696 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2697 * In that cases, pages are freed continuously and we can expect pages
2698 * are in the same memcg. All these calls itself limits the number of
2699 * pages freed at once, then uncharge_start/end() is called properly.
2700 * This may be called prural(2) times in a context,
2703 void mem_cgroup_uncharge_start(void)
2705 current
->memcg_batch
.do_batch
++;
2706 /* We can do nest. */
2707 if (current
->memcg_batch
.do_batch
== 1) {
2708 current
->memcg_batch
.memcg
= NULL
;
2709 current
->memcg_batch
.bytes
= 0;
2710 current
->memcg_batch
.memsw_bytes
= 0;
2714 void mem_cgroup_uncharge_end(void)
2716 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
2718 if (!batch
->do_batch
)
2722 if (batch
->do_batch
) /* If stacked, do nothing. */
2728 * This "batch->memcg" is valid without any css_get/put etc...
2729 * bacause we hide charges behind us.
2732 res_counter_uncharge(&batch
->memcg
->res
, batch
->bytes
);
2733 if (batch
->memsw_bytes
)
2734 res_counter_uncharge(&batch
->memcg
->memsw
, batch
->memsw_bytes
);
2735 memcg_oom_recover(batch
->memcg
);
2736 /* forget this pointer (for sanity check) */
2737 batch
->memcg
= NULL
;
2742 * called after __delete_from_swap_cache() and drop "page" account.
2743 * memcg information is recorded to swap_cgroup of "ent"
2746 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
2748 struct mem_cgroup
*memcg
;
2749 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
2751 if (!swapout
) /* this was a swap cache but the swap is unused ! */
2752 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
2754 memcg
= __mem_cgroup_uncharge_common(page
, ctype
);
2757 * record memcg information, if swapout && memcg != NULL,
2758 * mem_cgroup_get() was called in uncharge().
2760 if (do_swap_account
&& swapout
&& memcg
)
2761 swap_cgroup_record(ent
, css_id(&memcg
->css
));
2765 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2767 * called from swap_entry_free(). remove record in swap_cgroup and
2768 * uncharge "memsw" account.
2770 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
2772 struct mem_cgroup
*memcg
;
2775 if (!do_swap_account
)
2778 id
= swap_cgroup_record(ent
, 0);
2780 memcg
= mem_cgroup_lookup(id
);
2783 * We uncharge this because swap is freed.
2784 * This memcg can be obsolete one. We avoid calling css_tryget
2786 if (!mem_cgroup_is_root(memcg
))
2787 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2788 mem_cgroup_swap_statistics(memcg
, false);
2789 mem_cgroup_put(memcg
);
2795 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2796 * @entry: swap entry to be moved
2797 * @from: mem_cgroup which the entry is moved from
2798 * @to: mem_cgroup which the entry is moved to
2799 * @need_fixup: whether we should fixup res_counters and refcounts.
2801 * It succeeds only when the swap_cgroup's record for this entry is the same
2802 * as the mem_cgroup's id of @from.
2804 * Returns 0 on success, -EINVAL on failure.
2806 * The caller must have charged to @to, IOW, called res_counter_charge() about
2807 * both res and memsw, and called css_get().
2809 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2810 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2812 unsigned short old_id
, new_id
;
2814 old_id
= css_id(&from
->css
);
2815 new_id
= css_id(&to
->css
);
2817 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2818 mem_cgroup_swap_statistics(from
, false);
2819 mem_cgroup_swap_statistics(to
, true);
2821 * This function is only called from task migration context now.
2822 * It postpones res_counter and refcount handling till the end
2823 * of task migration(mem_cgroup_clear_mc()) for performance
2824 * improvement. But we cannot postpone mem_cgroup_get(to)
2825 * because if the process that has been moved to @to does
2826 * swap-in, the refcount of @to might be decreased to 0.
2830 if (!mem_cgroup_is_root(from
))
2831 res_counter_uncharge(&from
->memsw
, PAGE_SIZE
);
2832 mem_cgroup_put(from
);
2834 * we charged both to->res and to->memsw, so we should
2837 if (!mem_cgroup_is_root(to
))
2838 res_counter_uncharge(&to
->res
, PAGE_SIZE
);
2845 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2846 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2853 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2856 int mem_cgroup_prepare_migration(struct page
*page
,
2857 struct page
*newpage
, struct mem_cgroup
**ptr
)
2859 struct page_cgroup
*pc
;
2860 struct mem_cgroup
*mem
= NULL
;
2861 enum charge_type ctype
;
2864 VM_BUG_ON(PageTransHuge(page
));
2865 if (mem_cgroup_disabled())
2868 pc
= lookup_page_cgroup(page
);
2869 lock_page_cgroup(pc
);
2870 if (PageCgroupUsed(pc
)) {
2871 mem
= pc
->mem_cgroup
;
2874 * At migrating an anonymous page, its mapcount goes down
2875 * to 0 and uncharge() will be called. But, even if it's fully
2876 * unmapped, migration may fail and this page has to be
2877 * charged again. We set MIGRATION flag here and delay uncharge
2878 * until end_migration() is called
2880 * Corner Case Thinking
2882 * When the old page was mapped as Anon and it's unmap-and-freed
2883 * while migration was ongoing.
2884 * If unmap finds the old page, uncharge() of it will be delayed
2885 * until end_migration(). If unmap finds a new page, it's
2886 * uncharged when it make mapcount to be 1->0. If unmap code
2887 * finds swap_migration_entry, the new page will not be mapped
2888 * and end_migration() will find it(mapcount==0).
2891 * When the old page was mapped but migraion fails, the kernel
2892 * remaps it. A charge for it is kept by MIGRATION flag even
2893 * if mapcount goes down to 0. We can do remap successfully
2894 * without charging it again.
2897 * The "old" page is under lock_page() until the end of
2898 * migration, so, the old page itself will not be swapped-out.
2899 * If the new page is swapped out before end_migraton, our
2900 * hook to usual swap-out path will catch the event.
2903 SetPageCgroupMigration(pc
);
2905 unlock_page_cgroup(pc
);
2907 * If the page is not charged at this point,
2914 ret
= __mem_cgroup_try_charge(NULL
, GFP_KERNEL
, ptr
, false, PAGE_SIZE
);
2915 css_put(&mem
->css
);/* drop extra refcnt */
2916 if (ret
|| *ptr
== NULL
) {
2917 if (PageAnon(page
)) {
2918 lock_page_cgroup(pc
);
2919 ClearPageCgroupMigration(pc
);
2920 unlock_page_cgroup(pc
);
2922 * The old page may be fully unmapped while we kept it.
2924 mem_cgroup_uncharge_page(page
);
2929 * We charge new page before it's used/mapped. So, even if unlock_page()
2930 * is called before end_migration, we can catch all events on this new
2931 * page. In the case new page is migrated but not remapped, new page's
2932 * mapcount will be finally 0 and we call uncharge in end_migration().
2934 pc
= lookup_page_cgroup(newpage
);
2936 ctype
= MEM_CGROUP_CHARGE_TYPE_MAPPED
;
2937 else if (page_is_file_cache(page
))
2938 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
2940 ctype
= MEM_CGROUP_CHARGE_TYPE_SHMEM
;
2941 __mem_cgroup_commit_charge(mem
, pc
, ctype
, PAGE_SIZE
);
2945 /* remove redundant charge if migration failed*/
2946 void mem_cgroup_end_migration(struct mem_cgroup
*mem
,
2947 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
2949 struct page
*used
, *unused
;
2950 struct page_cgroup
*pc
;
2954 /* blocks rmdir() */
2955 cgroup_exclude_rmdir(&mem
->css
);
2956 if (!migration_ok
) {
2964 * We disallowed uncharge of pages under migration because mapcount
2965 * of the page goes down to zero, temporarly.
2966 * Clear the flag and check the page should be charged.
2968 pc
= lookup_page_cgroup(oldpage
);
2969 lock_page_cgroup(pc
);
2970 ClearPageCgroupMigration(pc
);
2971 unlock_page_cgroup(pc
);
2973 __mem_cgroup_uncharge_common(unused
, MEM_CGROUP_CHARGE_TYPE_FORCE
);
2976 * If a page is a file cache, radix-tree replacement is very atomic
2977 * and we can skip this check. When it was an Anon page, its mapcount
2978 * goes down to 0. But because we added MIGRATION flage, it's not
2979 * uncharged yet. There are several case but page->mapcount check
2980 * and USED bit check in mem_cgroup_uncharge_page() will do enough
2981 * check. (see prepare_charge() also)
2984 mem_cgroup_uncharge_page(used
);
2986 * At migration, we may charge account against cgroup which has no
2988 * So, rmdir()->pre_destroy() can be called while we do this charge.
2989 * In that case, we need to call pre_destroy() again. check it here.
2991 cgroup_release_and_wakeup_rmdir(&mem
->css
);
2995 * A call to try to shrink memory usage on charge failure at shmem's swapin.
2996 * Calling hierarchical_reclaim is not enough because we should update
2997 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2998 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2999 * not from the memcg which this page would be charged to.
3000 * try_charge_swapin does all of these works properly.
3002 int mem_cgroup_shmem_charge_fallback(struct page
*page
,
3003 struct mm_struct
*mm
,
3006 struct mem_cgroup
*mem
= NULL
;
3009 if (mem_cgroup_disabled())
3012 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
3014 mem_cgroup_cancel_charge_swapin(mem
); /* it does !mem check */
3019 static DEFINE_MUTEX(set_limit_mutex
);
3021 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
3022 unsigned long long val
)
3025 u64 memswlimit
, memlimit
;
3027 int children
= mem_cgroup_count_children(memcg
);
3028 u64 curusage
, oldusage
;
3032 * For keeping hierarchical_reclaim simple, how long we should retry
3033 * is depends on callers. We set our retry-count to be function
3034 * of # of children which we should visit in this loop.
3036 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
3038 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3041 while (retry_count
) {
3042 if (signal_pending(current
)) {
3047 * Rather than hide all in some function, I do this in
3048 * open coded manner. You see what this really does.
3049 * We have to guarantee mem->res.limit < mem->memsw.limit.
3051 mutex_lock(&set_limit_mutex
);
3052 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3053 if (memswlimit
< val
) {
3055 mutex_unlock(&set_limit_mutex
);
3059 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3063 ret
= res_counter_set_limit(&memcg
->res
, val
);
3065 if (memswlimit
== val
)
3066 memcg
->memsw_is_minimum
= true;
3068 memcg
->memsw_is_minimum
= false;
3070 mutex_unlock(&set_limit_mutex
);
3075 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3076 MEM_CGROUP_RECLAIM_SHRINK
);
3077 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3078 /* Usage is reduced ? */
3079 if (curusage
>= oldusage
)
3082 oldusage
= curusage
;
3084 if (!ret
&& enlarge
)
3085 memcg_oom_recover(memcg
);
3090 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3091 unsigned long long val
)
3094 u64 memlimit
, memswlimit
, oldusage
, curusage
;
3095 int children
= mem_cgroup_count_children(memcg
);
3099 /* see mem_cgroup_resize_res_limit */
3100 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
3101 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3102 while (retry_count
) {
3103 if (signal_pending(current
)) {
3108 * Rather than hide all in some function, I do this in
3109 * open coded manner. You see what this really does.
3110 * We have to guarantee mem->res.limit < mem->memsw.limit.
3112 mutex_lock(&set_limit_mutex
);
3113 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3114 if (memlimit
> val
) {
3116 mutex_unlock(&set_limit_mutex
);
3119 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3120 if (memswlimit
< val
)
3122 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
3124 if (memlimit
== val
)
3125 memcg
->memsw_is_minimum
= true;
3127 memcg
->memsw_is_minimum
= false;
3129 mutex_unlock(&set_limit_mutex
);
3134 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3135 MEM_CGROUP_RECLAIM_NOSWAP
|
3136 MEM_CGROUP_RECLAIM_SHRINK
);
3137 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3138 /* Usage is reduced ? */
3139 if (curusage
>= oldusage
)
3142 oldusage
= curusage
;
3144 if (!ret
&& enlarge
)
3145 memcg_oom_recover(memcg
);
3149 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3152 unsigned long nr_reclaimed
= 0;
3153 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3154 unsigned long reclaimed
;
3156 struct mem_cgroup_tree_per_zone
*mctz
;
3157 unsigned long long excess
;
3162 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3164 * This loop can run a while, specially if mem_cgroup's continuously
3165 * keep exceeding their soft limit and putting the system under
3172 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3176 reclaimed
= mem_cgroup_hierarchical_reclaim(mz
->mem
, zone
,
3178 MEM_CGROUP_RECLAIM_SOFT
);
3179 nr_reclaimed
+= reclaimed
;
3180 spin_lock(&mctz
->lock
);
3183 * If we failed to reclaim anything from this memory cgroup
3184 * it is time to move on to the next cgroup
3190 * Loop until we find yet another one.
3192 * By the time we get the soft_limit lock
3193 * again, someone might have aded the
3194 * group back on the RB tree. Iterate to
3195 * make sure we get a different mem.
3196 * mem_cgroup_largest_soft_limit_node returns
3197 * NULL if no other cgroup is present on
3201 __mem_cgroup_largest_soft_limit_node(mctz
);
3202 if (next_mz
== mz
) {
3203 css_put(&next_mz
->mem
->css
);
3205 } else /* next_mz == NULL or other memcg */
3209 __mem_cgroup_remove_exceeded(mz
->mem
, mz
, mctz
);
3210 excess
= res_counter_soft_limit_excess(&mz
->mem
->res
);
3212 * One school of thought says that we should not add
3213 * back the node to the tree if reclaim returns 0.
3214 * But our reclaim could return 0, simply because due
3215 * to priority we are exposing a smaller subset of
3216 * memory to reclaim from. Consider this as a longer
3219 /* If excess == 0, no tree ops */
3220 __mem_cgroup_insert_exceeded(mz
->mem
, mz
, mctz
, excess
);
3221 spin_unlock(&mctz
->lock
);
3222 css_put(&mz
->mem
->css
);
3225 * Could not reclaim anything and there are no more
3226 * mem cgroups to try or we seem to be looping without
3227 * reclaiming anything.
3229 if (!nr_reclaimed
&&
3231 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3233 } while (!nr_reclaimed
);
3235 css_put(&next_mz
->mem
->css
);
3236 return nr_reclaimed
;
3240 * This routine traverse page_cgroup in given list and drop them all.
3241 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3243 static int mem_cgroup_force_empty_list(struct mem_cgroup
*mem
,
3244 int node
, int zid
, enum lru_list lru
)
3247 struct mem_cgroup_per_zone
*mz
;
3248 struct page_cgroup
*pc
, *busy
;
3249 unsigned long flags
, loop
;
3250 struct list_head
*list
;
3253 zone
= &NODE_DATA(node
)->node_zones
[zid
];
3254 mz
= mem_cgroup_zoneinfo(mem
, node
, zid
);
3255 list
= &mz
->lists
[lru
];
3257 loop
= MEM_CGROUP_ZSTAT(mz
, lru
);
3258 /* give some margin against EBUSY etc...*/
3263 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3264 if (list_empty(list
)) {
3265 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3268 pc
= list_entry(list
->prev
, struct page_cgroup
, lru
);
3270 list_move(&pc
->lru
, list
);
3272 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3275 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3277 ret
= mem_cgroup_move_parent(pc
, mem
, GFP_KERNEL
);
3281 if (ret
== -EBUSY
|| ret
== -EINVAL
) {
3282 /* found lock contention or "pc" is obsolete. */
3289 if (!ret
&& !list_empty(list
))
3295 * make mem_cgroup's charge to be 0 if there is no task.
3296 * This enables deleting this mem_cgroup.
3298 static int mem_cgroup_force_empty(struct mem_cgroup
*mem
, bool free_all
)
3301 int node
, zid
, shrink
;
3302 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3303 struct cgroup
*cgrp
= mem
->css
.cgroup
;
3308 /* should free all ? */
3314 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
3317 if (signal_pending(current
))
3319 /* This is for making all *used* pages to be on LRU. */
3320 lru_add_drain_all();
3321 drain_all_stock_sync();
3323 mem_cgroup_start_move(mem
);
3324 for_each_node_state(node
, N_HIGH_MEMORY
) {
3325 for (zid
= 0; !ret
&& zid
< MAX_NR_ZONES
; zid
++) {
3328 ret
= mem_cgroup_force_empty_list(mem
,
3337 mem_cgroup_end_move(mem
);
3338 memcg_oom_recover(mem
);
3339 /* it seems parent cgroup doesn't have enough mem */
3343 /* "ret" should also be checked to ensure all lists are empty. */
3344 } while (mem
->res
.usage
> 0 || ret
);
3350 /* returns EBUSY if there is a task or if we come here twice. */
3351 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
) || shrink
) {
3355 /* we call try-to-free pages for make this cgroup empty */
3356 lru_add_drain_all();
3357 /* try to free all pages in this cgroup */
3359 while (nr_retries
&& mem
->res
.usage
> 0) {
3362 if (signal_pending(current
)) {
3366 progress
= try_to_free_mem_cgroup_pages(mem
, GFP_KERNEL
,
3367 false, get_swappiness(mem
));
3370 /* maybe some writeback is necessary */
3371 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3376 /* try move_account...there may be some *locked* pages. */
3380 int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
3382 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont
), true);
3386 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
3388 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
3391 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
3395 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3396 struct cgroup
*parent
= cont
->parent
;
3397 struct mem_cgroup
*parent_mem
= NULL
;
3400 parent_mem
= mem_cgroup_from_cont(parent
);
3404 * If parent's use_hierarchy is set, we can't make any modifications
3405 * in the child subtrees. If it is unset, then the change can
3406 * occur, provided the current cgroup has no children.
3408 * For the root cgroup, parent_mem is NULL, we allow value to be
3409 * set if there are no children.
3411 if ((!parent_mem
|| !parent_mem
->use_hierarchy
) &&
3412 (val
== 1 || val
== 0)) {
3413 if (list_empty(&cont
->children
))
3414 mem
->use_hierarchy
= val
;
3425 static u64
mem_cgroup_get_recursive_idx_stat(struct mem_cgroup
*mem
,
3426 enum mem_cgroup_stat_index idx
)
3428 struct mem_cgroup
*iter
;
3431 /* each per cpu's value can be minus.Then, use s64 */
3432 for_each_mem_cgroup_tree(iter
, mem
)
3433 val
+= mem_cgroup_read_stat(iter
, idx
);
3435 if (val
< 0) /* race ? */
3440 static inline u64
mem_cgroup_usage(struct mem_cgroup
*mem
, bool swap
)
3444 if (!mem_cgroup_is_root(mem
)) {
3446 return res_counter_read_u64(&mem
->res
, RES_USAGE
);
3448 return res_counter_read_u64(&mem
->memsw
, RES_USAGE
);
3451 val
= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3452 val
+= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_RSS
);
3455 val
+= mem_cgroup_get_recursive_idx_stat(mem
,
3456 MEM_CGROUP_STAT_SWAPOUT
);
3458 return val
<< PAGE_SHIFT
;
3461 static u64
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
)
3463 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3467 type
= MEMFILE_TYPE(cft
->private);
3468 name
= MEMFILE_ATTR(cft
->private);
3471 if (name
== RES_USAGE
)
3472 val
= mem_cgroup_usage(mem
, false);
3474 val
= res_counter_read_u64(&mem
->res
, name
);
3477 if (name
== RES_USAGE
)
3478 val
= mem_cgroup_usage(mem
, true);
3480 val
= res_counter_read_u64(&mem
->memsw
, name
);
3489 * The user of this function is...
3492 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
3495 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
3497 unsigned long long val
;
3500 type
= MEMFILE_TYPE(cft
->private);
3501 name
= MEMFILE_ATTR(cft
->private);
3504 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3508 /* This function does all necessary parse...reuse it */
3509 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3513 ret
= mem_cgroup_resize_limit(memcg
, val
);
3515 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
3517 case RES_SOFT_LIMIT
:
3518 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3522 * For memsw, soft limits are hard to implement in terms
3523 * of semantics, for now, we support soft limits for
3524 * control without swap
3527 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
3532 ret
= -EINVAL
; /* should be BUG() ? */
3538 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
3539 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
3541 struct cgroup
*cgroup
;
3542 unsigned long long min_limit
, min_memsw_limit
, tmp
;
3544 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3545 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3546 cgroup
= memcg
->css
.cgroup
;
3547 if (!memcg
->use_hierarchy
)
3550 while (cgroup
->parent
) {
3551 cgroup
= cgroup
->parent
;
3552 memcg
= mem_cgroup_from_cont(cgroup
);
3553 if (!memcg
->use_hierarchy
)
3555 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3556 min_limit
= min(min_limit
, tmp
);
3557 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3558 min_memsw_limit
= min(min_memsw_limit
, tmp
);
3561 *mem_limit
= min_limit
;
3562 *memsw_limit
= min_memsw_limit
;
3566 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
3568 struct mem_cgroup
*mem
;
3571 mem
= mem_cgroup_from_cont(cont
);
3572 type
= MEMFILE_TYPE(event
);
3573 name
= MEMFILE_ATTR(event
);
3577 res_counter_reset_max(&mem
->res
);
3579 res_counter_reset_max(&mem
->memsw
);
3583 res_counter_reset_failcnt(&mem
->res
);
3585 res_counter_reset_failcnt(&mem
->memsw
);
3592 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
3595 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
3599 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3600 struct cftype
*cft
, u64 val
)
3602 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
3604 if (val
>= (1 << NR_MOVE_TYPE
))
3607 * We check this value several times in both in can_attach() and
3608 * attach(), so we need cgroup lock to prevent this value from being
3612 mem
->move_charge_at_immigrate
= val
;
3618 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3619 struct cftype
*cft
, u64 val
)
3626 /* For read statistics */
3642 struct mcs_total_stat
{
3643 s64 stat
[NR_MCS_STAT
];
3649 } memcg_stat_strings
[NR_MCS_STAT
] = {
3650 {"cache", "total_cache"},
3651 {"rss", "total_rss"},
3652 {"mapped_file", "total_mapped_file"},
3653 {"pgpgin", "total_pgpgin"},
3654 {"pgpgout", "total_pgpgout"},
3655 {"swap", "total_swap"},
3656 {"inactive_anon", "total_inactive_anon"},
3657 {"active_anon", "total_active_anon"},
3658 {"inactive_file", "total_inactive_file"},
3659 {"active_file", "total_active_file"},
3660 {"unevictable", "total_unevictable"}
3665 mem_cgroup_get_local_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3670 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3671 s
->stat
[MCS_CACHE
] += val
* PAGE_SIZE
;
3672 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_RSS
);
3673 s
->stat
[MCS_RSS
] += val
* PAGE_SIZE
;
3674 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_FILE_MAPPED
);
3675 s
->stat
[MCS_FILE_MAPPED
] += val
* PAGE_SIZE
;
3676 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGIN_COUNT
);
3677 s
->stat
[MCS_PGPGIN
] += val
;
3678 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGOUT_COUNT
);
3679 s
->stat
[MCS_PGPGOUT
] += val
;
3680 if (do_swap_account
) {
3681 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_SWAPOUT
);
3682 s
->stat
[MCS_SWAP
] += val
* PAGE_SIZE
;
3686 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_ANON
);
3687 s
->stat
[MCS_INACTIVE_ANON
] += val
* PAGE_SIZE
;
3688 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_ANON
);
3689 s
->stat
[MCS_ACTIVE_ANON
] += val
* PAGE_SIZE
;
3690 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_FILE
);
3691 s
->stat
[MCS_INACTIVE_FILE
] += val
* PAGE_SIZE
;
3692 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_FILE
);
3693 s
->stat
[MCS_ACTIVE_FILE
] += val
* PAGE_SIZE
;
3694 val
= mem_cgroup_get_local_zonestat(mem
, LRU_UNEVICTABLE
);
3695 s
->stat
[MCS_UNEVICTABLE
] += val
* PAGE_SIZE
;
3699 mem_cgroup_get_total_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3701 struct mem_cgroup
*iter
;
3703 for_each_mem_cgroup_tree(iter
, mem
)
3704 mem_cgroup_get_local_stat(iter
, s
);
3707 static int mem_control_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
3708 struct cgroup_map_cb
*cb
)
3710 struct mem_cgroup
*mem_cont
= mem_cgroup_from_cont(cont
);
3711 struct mcs_total_stat mystat
;
3714 memset(&mystat
, 0, sizeof(mystat
));
3715 mem_cgroup_get_local_stat(mem_cont
, &mystat
);
3717 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3718 if (i
== MCS_SWAP
&& !do_swap_account
)
3720 cb
->fill(cb
, memcg_stat_strings
[i
].local_name
, mystat
.stat
[i
]);
3723 /* Hierarchical information */
3725 unsigned long long limit
, memsw_limit
;
3726 memcg_get_hierarchical_limit(mem_cont
, &limit
, &memsw_limit
);
3727 cb
->fill(cb
, "hierarchical_memory_limit", limit
);
3728 if (do_swap_account
)
3729 cb
->fill(cb
, "hierarchical_memsw_limit", memsw_limit
);
3732 memset(&mystat
, 0, sizeof(mystat
));
3733 mem_cgroup_get_total_stat(mem_cont
, &mystat
);
3734 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3735 if (i
== MCS_SWAP
&& !do_swap_account
)
3737 cb
->fill(cb
, memcg_stat_strings
[i
].total_name
, mystat
.stat
[i
]);
3740 #ifdef CONFIG_DEBUG_VM
3741 cb
->fill(cb
, "inactive_ratio", calc_inactive_ratio(mem_cont
, NULL
));
3745 struct mem_cgroup_per_zone
*mz
;
3746 unsigned long recent_rotated
[2] = {0, 0};
3747 unsigned long recent_scanned
[2] = {0, 0};
3749 for_each_online_node(nid
)
3750 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3751 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
3753 recent_rotated
[0] +=
3754 mz
->reclaim_stat
.recent_rotated
[0];
3755 recent_rotated
[1] +=
3756 mz
->reclaim_stat
.recent_rotated
[1];
3757 recent_scanned
[0] +=
3758 mz
->reclaim_stat
.recent_scanned
[0];
3759 recent_scanned
[1] +=
3760 mz
->reclaim_stat
.recent_scanned
[1];
3762 cb
->fill(cb
, "recent_rotated_anon", recent_rotated
[0]);
3763 cb
->fill(cb
, "recent_rotated_file", recent_rotated
[1]);
3764 cb
->fill(cb
, "recent_scanned_anon", recent_scanned
[0]);
3765 cb
->fill(cb
, "recent_scanned_file", recent_scanned
[1]);
3772 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
3774 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3776 return get_swappiness(memcg
);
3779 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
3782 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3783 struct mem_cgroup
*parent
;
3788 if (cgrp
->parent
== NULL
)
3791 parent
= mem_cgroup_from_cont(cgrp
->parent
);
3795 /* If under hierarchy, only empty-root can set this value */
3796 if ((parent
->use_hierarchy
) ||
3797 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
3802 spin_lock(&memcg
->reclaim_param_lock
);
3803 memcg
->swappiness
= val
;
3804 spin_unlock(&memcg
->reclaim_param_lock
);
3811 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3813 struct mem_cgroup_threshold_ary
*t
;
3819 t
= rcu_dereference(memcg
->thresholds
.primary
);
3821 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3826 usage
= mem_cgroup_usage(memcg
, swap
);
3829 * current_threshold points to threshold just below usage.
3830 * If it's not true, a threshold was crossed after last
3831 * call of __mem_cgroup_threshold().
3833 i
= t
->current_threshold
;
3836 * Iterate backward over array of thresholds starting from
3837 * current_threshold and check if a threshold is crossed.
3838 * If none of thresholds below usage is crossed, we read
3839 * only one element of the array here.
3841 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3842 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3844 /* i = current_threshold + 1 */
3848 * Iterate forward over array of thresholds starting from
3849 * current_threshold+1 and check if a threshold is crossed.
3850 * If none of thresholds above usage is crossed, we read
3851 * only one element of the array here.
3853 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3854 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3856 /* Update current_threshold */
3857 t
->current_threshold
= i
- 1;
3862 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3865 __mem_cgroup_threshold(memcg
, false);
3866 if (do_swap_account
)
3867 __mem_cgroup_threshold(memcg
, true);
3869 memcg
= parent_mem_cgroup(memcg
);
3873 static int compare_thresholds(const void *a
, const void *b
)
3875 const struct mem_cgroup_threshold
*_a
= a
;
3876 const struct mem_cgroup_threshold
*_b
= b
;
3878 return _a
->threshold
- _b
->threshold
;
3881 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*mem
)
3883 struct mem_cgroup_eventfd_list
*ev
;
3885 list_for_each_entry(ev
, &mem
->oom_notify
, list
)
3886 eventfd_signal(ev
->eventfd
, 1);
3890 static void mem_cgroup_oom_notify(struct mem_cgroup
*mem
)
3892 struct mem_cgroup
*iter
;
3894 for_each_mem_cgroup_tree(iter
, mem
)
3895 mem_cgroup_oom_notify_cb(iter
);
3898 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
3899 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
3901 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3902 struct mem_cgroup_thresholds
*thresholds
;
3903 struct mem_cgroup_threshold_ary
*new;
3904 int type
= MEMFILE_TYPE(cft
->private);
3905 u64 threshold
, usage
;
3908 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
3912 mutex_lock(&memcg
->thresholds_lock
);
3915 thresholds
= &memcg
->thresholds
;
3916 else if (type
== _MEMSWAP
)
3917 thresholds
= &memcg
->memsw_thresholds
;
3921 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
3923 /* Check if a threshold crossed before adding a new one */
3924 if (thresholds
->primary
)
3925 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3927 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3929 /* Allocate memory for new array of thresholds */
3930 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3938 /* Copy thresholds (if any) to new array */
3939 if (thresholds
->primary
) {
3940 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3941 sizeof(struct mem_cgroup_threshold
));
3944 /* Add new threshold */
3945 new->entries
[size
- 1].eventfd
= eventfd
;
3946 new->entries
[size
- 1].threshold
= threshold
;
3948 /* Sort thresholds. Registering of new threshold isn't time-critical */
3949 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3950 compare_thresholds
, NULL
);
3952 /* Find current threshold */
3953 new->current_threshold
= -1;
3954 for (i
= 0; i
< size
; i
++) {
3955 if (new->entries
[i
].threshold
< usage
) {
3957 * new->current_threshold will not be used until
3958 * rcu_assign_pointer(), so it's safe to increment
3961 ++new->current_threshold
;
3965 /* Free old spare buffer and save old primary buffer as spare */
3966 kfree(thresholds
->spare
);
3967 thresholds
->spare
= thresholds
->primary
;
3969 rcu_assign_pointer(thresholds
->primary
, new);
3971 /* To be sure that nobody uses thresholds */
3975 mutex_unlock(&memcg
->thresholds_lock
);
3980 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
3981 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
3983 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3984 struct mem_cgroup_thresholds
*thresholds
;
3985 struct mem_cgroup_threshold_ary
*new;
3986 int type
= MEMFILE_TYPE(cft
->private);
3990 mutex_lock(&memcg
->thresholds_lock
);
3992 thresholds
= &memcg
->thresholds
;
3993 else if (type
== _MEMSWAP
)
3994 thresholds
= &memcg
->memsw_thresholds
;
3999 * Something went wrong if we trying to unregister a threshold
4000 * if we don't have thresholds
4002 BUG_ON(!thresholds
);
4004 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
4006 /* Check if a threshold crossed before removing */
4007 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
4009 /* Calculate new number of threshold */
4011 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
4012 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
4016 new = thresholds
->spare
;
4018 /* Set thresholds array to NULL if we don't have thresholds */
4027 /* Copy thresholds and find current threshold */
4028 new->current_threshold
= -1;
4029 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
4030 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
4033 new->entries
[j
] = thresholds
->primary
->entries
[i
];
4034 if (new->entries
[j
].threshold
< usage
) {
4036 * new->current_threshold will not be used
4037 * until rcu_assign_pointer(), so it's safe to increment
4040 ++new->current_threshold
;
4046 /* Swap primary and spare array */
4047 thresholds
->spare
= thresholds
->primary
;
4048 rcu_assign_pointer(thresholds
->primary
, new);
4050 /* To be sure that nobody uses thresholds */
4053 mutex_unlock(&memcg
->thresholds_lock
);
4056 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
4057 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
4059 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
4060 struct mem_cgroup_eventfd_list
*event
;
4061 int type
= MEMFILE_TYPE(cft
->private);
4063 BUG_ON(type
!= _OOM_TYPE
);
4064 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4068 mutex_lock(&memcg_oom_mutex
);
4070 event
->eventfd
= eventfd
;
4071 list_add(&event
->list
, &memcg
->oom_notify
);
4073 /* already in OOM ? */
4074 if (atomic_read(&memcg
->oom_lock
))
4075 eventfd_signal(eventfd
, 1);
4076 mutex_unlock(&memcg_oom_mutex
);
4081 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
4082 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
4084 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4085 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4086 int type
= MEMFILE_TYPE(cft
->private);
4088 BUG_ON(type
!= _OOM_TYPE
);
4090 mutex_lock(&memcg_oom_mutex
);
4092 list_for_each_entry_safe(ev
, tmp
, &mem
->oom_notify
, list
) {
4093 if (ev
->eventfd
== eventfd
) {
4094 list_del(&ev
->list
);
4099 mutex_unlock(&memcg_oom_mutex
);
4102 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
4103 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
4105 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4107 cb
->fill(cb
, "oom_kill_disable", mem
->oom_kill_disable
);
4109 if (atomic_read(&mem
->oom_lock
))
4110 cb
->fill(cb
, "under_oom", 1);
4112 cb
->fill(cb
, "under_oom", 0);
4116 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
4117 struct cftype
*cft
, u64 val
)
4119 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4120 struct mem_cgroup
*parent
;
4122 /* cannot set to root cgroup and only 0 and 1 are allowed */
4123 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
4126 parent
= mem_cgroup_from_cont(cgrp
->parent
);
4129 /* oom-kill-disable is a flag for subhierarchy. */
4130 if ((parent
->use_hierarchy
) ||
4131 (mem
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
4135 mem
->oom_kill_disable
= val
;
4137 memcg_oom_recover(mem
);
4142 static struct cftype mem_cgroup_files
[] = {
4144 .name
= "usage_in_bytes",
4145 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4146 .read_u64
= mem_cgroup_read
,
4147 .register_event
= mem_cgroup_usage_register_event
,
4148 .unregister_event
= mem_cgroup_usage_unregister_event
,
4151 .name
= "max_usage_in_bytes",
4152 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4153 .trigger
= mem_cgroup_reset
,
4154 .read_u64
= mem_cgroup_read
,
4157 .name
= "limit_in_bytes",
4158 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4159 .write_string
= mem_cgroup_write
,
4160 .read_u64
= mem_cgroup_read
,
4163 .name
= "soft_limit_in_bytes",
4164 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4165 .write_string
= mem_cgroup_write
,
4166 .read_u64
= mem_cgroup_read
,
4170 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4171 .trigger
= mem_cgroup_reset
,
4172 .read_u64
= mem_cgroup_read
,
4176 .read_map
= mem_control_stat_show
,
4179 .name
= "force_empty",
4180 .trigger
= mem_cgroup_force_empty_write
,
4183 .name
= "use_hierarchy",
4184 .write_u64
= mem_cgroup_hierarchy_write
,
4185 .read_u64
= mem_cgroup_hierarchy_read
,
4188 .name
= "swappiness",
4189 .read_u64
= mem_cgroup_swappiness_read
,
4190 .write_u64
= mem_cgroup_swappiness_write
,
4193 .name
= "move_charge_at_immigrate",
4194 .read_u64
= mem_cgroup_move_charge_read
,
4195 .write_u64
= mem_cgroup_move_charge_write
,
4198 .name
= "oom_control",
4199 .read_map
= mem_cgroup_oom_control_read
,
4200 .write_u64
= mem_cgroup_oom_control_write
,
4201 .register_event
= mem_cgroup_oom_register_event
,
4202 .unregister_event
= mem_cgroup_oom_unregister_event
,
4203 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4207 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4208 static struct cftype memsw_cgroup_files
[] = {
4210 .name
= "memsw.usage_in_bytes",
4211 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
4212 .read_u64
= mem_cgroup_read
,
4213 .register_event
= mem_cgroup_usage_register_event
,
4214 .unregister_event
= mem_cgroup_usage_unregister_event
,
4217 .name
= "memsw.max_usage_in_bytes",
4218 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
4219 .trigger
= mem_cgroup_reset
,
4220 .read_u64
= mem_cgroup_read
,
4223 .name
= "memsw.limit_in_bytes",
4224 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
4225 .write_string
= mem_cgroup_write
,
4226 .read_u64
= mem_cgroup_read
,
4229 .name
= "memsw.failcnt",
4230 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
4231 .trigger
= mem_cgroup_reset
,
4232 .read_u64
= mem_cgroup_read
,
4236 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4238 if (!do_swap_account
)
4240 return cgroup_add_files(cont
, ss
, memsw_cgroup_files
,
4241 ARRAY_SIZE(memsw_cgroup_files
));
4244 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4250 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4252 struct mem_cgroup_per_node
*pn
;
4253 struct mem_cgroup_per_zone
*mz
;
4255 int zone
, tmp
= node
;
4257 * This routine is called against possible nodes.
4258 * But it's BUG to call kmalloc() against offline node.
4260 * TODO: this routine can waste much memory for nodes which will
4261 * never be onlined. It's better to use memory hotplug callback
4264 if (!node_state(node
, N_NORMAL_MEMORY
))
4266 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4270 mem
->info
.nodeinfo
[node
] = pn
;
4271 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4272 mz
= &pn
->zoneinfo
[zone
];
4274 INIT_LIST_HEAD(&mz
->lists
[l
]);
4275 mz
->usage_in_excess
= 0;
4276 mz
->on_tree
= false;
4282 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4284 kfree(mem
->info
.nodeinfo
[node
]);
4287 static struct mem_cgroup
*mem_cgroup_alloc(void)
4289 struct mem_cgroup
*mem
;
4290 int size
= sizeof(struct mem_cgroup
);
4292 /* Can be very big if MAX_NUMNODES is very big */
4293 if (size
< PAGE_SIZE
)
4294 mem
= kzalloc(size
, GFP_KERNEL
);
4296 mem
= vzalloc(size
);
4301 mem
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4304 spin_lock_init(&mem
->pcp_counter_lock
);
4308 if (size
< PAGE_SIZE
)
4316 * At destroying mem_cgroup, references from swap_cgroup can remain.
4317 * (scanning all at force_empty is too costly...)
4319 * Instead of clearing all references at force_empty, we remember
4320 * the number of reference from swap_cgroup and free mem_cgroup when
4321 * it goes down to 0.
4323 * Removal of cgroup itself succeeds regardless of refs from swap.
4326 static void __mem_cgroup_free(struct mem_cgroup
*mem
)
4330 mem_cgroup_remove_from_trees(mem
);
4331 free_css_id(&mem_cgroup_subsys
, &mem
->css
);
4333 for_each_node_state(node
, N_POSSIBLE
)
4334 free_mem_cgroup_per_zone_info(mem
, node
);
4336 free_percpu(mem
->stat
);
4337 if (sizeof(struct mem_cgroup
) < PAGE_SIZE
)
4343 static void mem_cgroup_get(struct mem_cgroup
*mem
)
4345 atomic_inc(&mem
->refcnt
);
4348 static void __mem_cgroup_put(struct mem_cgroup
*mem
, int count
)
4350 if (atomic_sub_and_test(count
, &mem
->refcnt
)) {
4351 struct mem_cgroup
*parent
= parent_mem_cgroup(mem
);
4352 __mem_cgroup_free(mem
);
4354 mem_cgroup_put(parent
);
4358 static void mem_cgroup_put(struct mem_cgroup
*mem
)
4360 __mem_cgroup_put(mem
, 1);
4364 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4366 static struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*mem
)
4368 if (!mem
->res
.parent
)
4370 return mem_cgroup_from_res_counter(mem
->res
.parent
, res
);
4373 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4374 static void __init
enable_swap_cgroup(void)
4376 if (!mem_cgroup_disabled() && really_do_swap_account
)
4377 do_swap_account
= 1;
4380 static void __init
enable_swap_cgroup(void)
4385 static int mem_cgroup_soft_limit_tree_init(void)
4387 struct mem_cgroup_tree_per_node
*rtpn
;
4388 struct mem_cgroup_tree_per_zone
*rtpz
;
4389 int tmp
, node
, zone
;
4391 for_each_node_state(node
, N_POSSIBLE
) {
4393 if (!node_state(node
, N_NORMAL_MEMORY
))
4395 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
4399 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
4401 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4402 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
4403 rtpz
->rb_root
= RB_ROOT
;
4404 spin_lock_init(&rtpz
->lock
);
4410 static struct cgroup_subsys_state
* __ref
4411 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
4413 struct mem_cgroup
*mem
, *parent
;
4414 long error
= -ENOMEM
;
4417 mem
= mem_cgroup_alloc();
4419 return ERR_PTR(error
);
4421 for_each_node_state(node
, N_POSSIBLE
)
4422 if (alloc_mem_cgroup_per_zone_info(mem
, node
))
4426 if (cont
->parent
== NULL
) {
4428 enable_swap_cgroup();
4430 root_mem_cgroup
= mem
;
4431 if (mem_cgroup_soft_limit_tree_init())
4433 for_each_possible_cpu(cpu
) {
4434 struct memcg_stock_pcp
*stock
=
4435 &per_cpu(memcg_stock
, cpu
);
4436 INIT_WORK(&stock
->work
, drain_local_stock
);
4438 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
4440 parent
= mem_cgroup_from_cont(cont
->parent
);
4441 mem
->use_hierarchy
= parent
->use_hierarchy
;
4442 mem
->oom_kill_disable
= parent
->oom_kill_disable
;
4445 if (parent
&& parent
->use_hierarchy
) {
4446 res_counter_init(&mem
->res
, &parent
->res
);
4447 res_counter_init(&mem
->memsw
, &parent
->memsw
);
4449 * We increment refcnt of the parent to ensure that we can
4450 * safely access it on res_counter_charge/uncharge.
4451 * This refcnt will be decremented when freeing this
4452 * mem_cgroup(see mem_cgroup_put).
4454 mem_cgroup_get(parent
);
4456 res_counter_init(&mem
->res
, NULL
);
4457 res_counter_init(&mem
->memsw
, NULL
);
4459 mem
->last_scanned_child
= 0;
4460 spin_lock_init(&mem
->reclaim_param_lock
);
4461 INIT_LIST_HEAD(&mem
->oom_notify
);
4464 mem
->swappiness
= get_swappiness(parent
);
4465 atomic_set(&mem
->refcnt
, 1);
4466 mem
->move_charge_at_immigrate
= 0;
4467 mutex_init(&mem
->thresholds_lock
);
4470 __mem_cgroup_free(mem
);
4471 root_mem_cgroup
= NULL
;
4472 return ERR_PTR(error
);
4475 static int mem_cgroup_pre_destroy(struct cgroup_subsys
*ss
,
4476 struct cgroup
*cont
)
4478 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4480 return mem_cgroup_force_empty(mem
, false);
4483 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
4484 struct cgroup
*cont
)
4486 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4488 mem_cgroup_put(mem
);
4491 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
4492 struct cgroup
*cont
)
4496 ret
= cgroup_add_files(cont
, ss
, mem_cgroup_files
,
4497 ARRAY_SIZE(mem_cgroup_files
));
4500 ret
= register_memsw_files(cont
, ss
);
4505 /* Handlers for move charge at task migration. */
4506 #define PRECHARGE_COUNT_AT_ONCE 256
4507 static int mem_cgroup_do_precharge(unsigned long count
)
4510 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4511 struct mem_cgroup
*mem
= mc
.to
;
4513 if (mem_cgroup_is_root(mem
)) {
4514 mc
.precharge
+= count
;
4515 /* we don't need css_get for root */
4518 /* try to charge at once */
4520 struct res_counter
*dummy
;
4522 * "mem" cannot be under rmdir() because we've already checked
4523 * by cgroup_lock_live_cgroup() that it is not removed and we
4524 * are still under the same cgroup_mutex. So we can postpone
4527 if (res_counter_charge(&mem
->res
, PAGE_SIZE
* count
, &dummy
))
4529 if (do_swap_account
&& res_counter_charge(&mem
->memsw
,
4530 PAGE_SIZE
* count
, &dummy
)) {
4531 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
4534 mc
.precharge
+= count
;
4538 /* fall back to one by one charge */
4540 if (signal_pending(current
)) {
4544 if (!batch_count
--) {
4545 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4548 ret
= __mem_cgroup_try_charge(NULL
, GFP_KERNEL
, &mem
, false,
4551 /* mem_cgroup_clear_mc() will do uncharge later */
4559 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4560 * @vma: the vma the pte to be checked belongs
4561 * @addr: the address corresponding to the pte to be checked
4562 * @ptent: the pte to be checked
4563 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4566 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4567 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4568 * move charge. if @target is not NULL, the page is stored in target->page
4569 * with extra refcnt got(Callers should handle it).
4570 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4571 * target for charge migration. if @target is not NULL, the entry is stored
4574 * Called with pte lock held.
4581 enum mc_target_type
{
4582 MC_TARGET_NONE
, /* not used */
4587 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4588 unsigned long addr
, pte_t ptent
)
4590 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4592 if (!page
|| !page_mapped(page
))
4594 if (PageAnon(page
)) {
4595 /* we don't move shared anon */
4596 if (!move_anon() || page_mapcount(page
) > 2)
4598 } else if (!move_file())
4599 /* we ignore mapcount for file pages */
4601 if (!get_page_unless_zero(page
))
4607 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4608 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4611 struct page
*page
= NULL
;
4612 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4614 if (!move_anon() || non_swap_entry(ent
))
4616 usage_count
= mem_cgroup_count_swap_user(ent
, &page
);
4617 if (usage_count
> 1) { /* we don't move shared anon */
4622 if (do_swap_account
)
4623 entry
->val
= ent
.val
;
4628 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4629 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4631 struct page
*page
= NULL
;
4632 struct inode
*inode
;
4633 struct address_space
*mapping
;
4636 if (!vma
->vm_file
) /* anonymous vma */
4641 inode
= vma
->vm_file
->f_path
.dentry
->d_inode
;
4642 mapping
= vma
->vm_file
->f_mapping
;
4643 if (pte_none(ptent
))
4644 pgoff
= linear_page_index(vma
, addr
);
4645 else /* pte_file(ptent) is true */
4646 pgoff
= pte_to_pgoff(ptent
);
4648 /* page is moved even if it's not RSS of this task(page-faulted). */
4649 if (!mapping_cap_swap_backed(mapping
)) { /* normal file */
4650 page
= find_get_page(mapping
, pgoff
);
4651 } else { /* shmem/tmpfs file. we should take account of swap too. */
4653 mem_cgroup_get_shmem_target(inode
, pgoff
, &page
, &ent
);
4654 if (do_swap_account
)
4655 entry
->val
= ent
.val
;
4661 static int is_target_pte_for_mc(struct vm_area_struct
*vma
,
4662 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4664 struct page
*page
= NULL
;
4665 struct page_cgroup
*pc
;
4667 swp_entry_t ent
= { .val
= 0 };
4669 if (pte_present(ptent
))
4670 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4671 else if (is_swap_pte(ptent
))
4672 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4673 else if (pte_none(ptent
) || pte_file(ptent
))
4674 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4676 if (!page
&& !ent
.val
)
4679 pc
= lookup_page_cgroup(page
);
4681 * Do only loose check w/o page_cgroup lock.
4682 * mem_cgroup_move_account() checks the pc is valid or not under
4685 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
4686 ret
= MC_TARGET_PAGE
;
4688 target
->page
= page
;
4690 if (!ret
|| !target
)
4693 /* There is a swap entry and a page doesn't exist or isn't charged */
4694 if (ent
.val
&& !ret
&&
4695 css_id(&mc
.from
->css
) == lookup_swap_cgroup(ent
)) {
4696 ret
= MC_TARGET_SWAP
;
4703 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4704 unsigned long addr
, unsigned long end
,
4705 struct mm_walk
*walk
)
4707 struct vm_area_struct
*vma
= walk
->private;
4711 VM_BUG_ON(pmd_trans_huge(*pmd
));
4712 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4713 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4714 if (is_target_pte_for_mc(vma
, addr
, *pte
, NULL
))
4715 mc
.precharge
++; /* increment precharge temporarily */
4716 pte_unmap_unlock(pte
- 1, ptl
);
4722 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4724 unsigned long precharge
;
4725 struct vm_area_struct
*vma
;
4727 down_read(&mm
->mmap_sem
);
4728 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
4729 struct mm_walk mem_cgroup_count_precharge_walk
= {
4730 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4734 if (is_vm_hugetlb_page(vma
))
4736 walk_page_range(vma
->vm_start
, vma
->vm_end
,
4737 &mem_cgroup_count_precharge_walk
);
4739 up_read(&mm
->mmap_sem
);
4741 precharge
= mc
.precharge
;
4747 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4749 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4751 VM_BUG_ON(mc
.moving_task
);
4752 mc
.moving_task
= current
;
4753 return mem_cgroup_do_precharge(precharge
);
4756 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4757 static void __mem_cgroup_clear_mc(void)
4759 struct mem_cgroup
*from
= mc
.from
;
4760 struct mem_cgroup
*to
= mc
.to
;
4762 /* we must uncharge all the leftover precharges from mc.to */
4764 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
4768 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4769 * we must uncharge here.
4771 if (mc
.moved_charge
) {
4772 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
4773 mc
.moved_charge
= 0;
4775 /* we must fixup refcnts and charges */
4776 if (mc
.moved_swap
) {
4777 /* uncharge swap account from the old cgroup */
4778 if (!mem_cgroup_is_root(mc
.from
))
4779 res_counter_uncharge(&mc
.from
->memsw
,
4780 PAGE_SIZE
* mc
.moved_swap
);
4781 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
4783 if (!mem_cgroup_is_root(mc
.to
)) {
4785 * we charged both to->res and to->memsw, so we should
4788 res_counter_uncharge(&mc
.to
->res
,
4789 PAGE_SIZE
* mc
.moved_swap
);
4791 /* we've already done mem_cgroup_get(mc.to) */
4794 memcg_oom_recover(from
);
4795 memcg_oom_recover(to
);
4796 wake_up_all(&mc
.waitq
);
4799 static void mem_cgroup_clear_mc(void)
4801 struct mem_cgroup
*from
= mc
.from
;
4804 * we must clear moving_task before waking up waiters at the end of
4807 mc
.moving_task
= NULL
;
4808 __mem_cgroup_clear_mc();
4809 spin_lock(&mc
.lock
);
4812 spin_unlock(&mc
.lock
);
4813 mem_cgroup_end_move(from
);
4816 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
4817 struct cgroup
*cgroup
,
4818 struct task_struct
*p
,
4822 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgroup
);
4824 if (mem
->move_charge_at_immigrate
) {
4825 struct mm_struct
*mm
;
4826 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
4828 VM_BUG_ON(from
== mem
);
4830 mm
= get_task_mm(p
);
4833 /* We move charges only when we move a owner of the mm */
4834 if (mm
->owner
== p
) {
4837 VM_BUG_ON(mc
.precharge
);
4838 VM_BUG_ON(mc
.moved_charge
);
4839 VM_BUG_ON(mc
.moved_swap
);
4840 mem_cgroup_start_move(from
);
4841 spin_lock(&mc
.lock
);
4844 spin_unlock(&mc
.lock
);
4845 /* We set mc.moving_task later */
4847 ret
= mem_cgroup_precharge_mc(mm
);
4849 mem_cgroup_clear_mc();
4856 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
4857 struct cgroup
*cgroup
,
4858 struct task_struct
*p
,
4861 mem_cgroup_clear_mc();
4864 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4865 unsigned long addr
, unsigned long end
,
4866 struct mm_walk
*walk
)
4869 struct vm_area_struct
*vma
= walk
->private;
4874 VM_BUG_ON(pmd_trans_huge(*pmd
));
4875 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4876 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4877 pte_t ptent
= *(pte
++);
4878 union mc_target target
;
4881 struct page_cgroup
*pc
;
4887 type
= is_target_pte_for_mc(vma
, addr
, ptent
, &target
);
4889 case MC_TARGET_PAGE
:
4891 if (isolate_lru_page(page
))
4893 pc
= lookup_page_cgroup(page
);
4894 if (!mem_cgroup_move_account(pc
,
4895 mc
.from
, mc
.to
, false, PAGE_SIZE
)) {
4897 /* we uncharge from mc.from later. */
4900 putback_lru_page(page
);
4901 put
: /* is_target_pte_for_mc() gets the page */
4904 case MC_TARGET_SWAP
:
4906 if (!mem_cgroup_move_swap_account(ent
,
4907 mc
.from
, mc
.to
, false)) {
4909 /* we fixup refcnts and charges later. */
4917 pte_unmap_unlock(pte
- 1, ptl
);
4922 * We have consumed all precharges we got in can_attach().
4923 * We try charge one by one, but don't do any additional
4924 * charges to mc.to if we have failed in charge once in attach()
4927 ret
= mem_cgroup_do_precharge(1);
4935 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4937 struct vm_area_struct
*vma
;
4939 lru_add_drain_all();
4941 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4943 * Someone who are holding the mmap_sem might be waiting in
4944 * waitq. So we cancel all extra charges, wake up all waiters,
4945 * and retry. Because we cancel precharges, we might not be able
4946 * to move enough charges, but moving charge is a best-effort
4947 * feature anyway, so it wouldn't be a big problem.
4949 __mem_cgroup_clear_mc();
4953 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
4955 struct mm_walk mem_cgroup_move_charge_walk
= {
4956 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4960 if (is_vm_hugetlb_page(vma
))
4962 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
4963 &mem_cgroup_move_charge_walk
);
4966 * means we have consumed all precharges and failed in
4967 * doing additional charge. Just abandon here.
4971 up_read(&mm
->mmap_sem
);
4974 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
4975 struct cgroup
*cont
,
4976 struct cgroup
*old_cont
,
4977 struct task_struct
*p
,
4980 struct mm_struct
*mm
;
4983 /* no need to move charge */
4986 mm
= get_task_mm(p
);
4988 mem_cgroup_move_charge(mm
);
4991 mem_cgroup_clear_mc();
4993 #else /* !CONFIG_MMU */
4994 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
4995 struct cgroup
*cgroup
,
4996 struct task_struct
*p
,
5001 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
5002 struct cgroup
*cgroup
,
5003 struct task_struct
*p
,
5007 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
5008 struct cgroup
*cont
,
5009 struct cgroup
*old_cont
,
5010 struct task_struct
*p
,
5016 struct cgroup_subsys mem_cgroup_subsys
= {
5018 .subsys_id
= mem_cgroup_subsys_id
,
5019 .create
= mem_cgroup_create
,
5020 .pre_destroy
= mem_cgroup_pre_destroy
,
5021 .destroy
= mem_cgroup_destroy
,
5022 .populate
= mem_cgroup_populate
,
5023 .can_attach
= mem_cgroup_can_attach
,
5024 .cancel_attach
= mem_cgroup_cancel_attach
,
5025 .attach
= mem_cgroup_move_task
,
5030 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5031 static int __init
enable_swap_account(char *s
)
5033 /* consider enabled if no parameter or 1 is given */
5034 if (!(*s
) || !strcmp(s
, "=1"))
5035 really_do_swap_account
= 1;
5036 else if (!strcmp(s
, "=0"))
5037 really_do_swap_account
= 0;
5040 __setup("swapaccount", enable_swap_account
);
5042 static int __init
disable_swap_account(char *s
)
5044 printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5045 enable_swap_account("=0");
5048 __setup("noswapaccount", disable_swap_account
);