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 MEM_CGROUP_ZSTAT(mz
, lru
) -= 1;
818 if (mem_cgroup_is_root(pc
->mem_cgroup
))
820 VM_BUG_ON(list_empty(&pc
->lru
));
821 list_del_init(&pc
->lru
);
824 void mem_cgroup_del_lru(struct page
*page
)
826 mem_cgroup_del_lru_list(page
, page_lru(page
));
829 void mem_cgroup_rotate_lru_list(struct page
*page
, enum lru_list lru
)
831 struct mem_cgroup_per_zone
*mz
;
832 struct page_cgroup
*pc
;
834 if (mem_cgroup_disabled())
837 pc
= lookup_page_cgroup(page
);
839 * Used bit is set without atomic ops but after smp_wmb().
840 * For making pc->mem_cgroup visible, insert smp_rmb() here.
843 /* unused or root page is not rotated. */
844 if (!PageCgroupUsed(pc
) || 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
));
860 * Used bit is set without atomic ops but after smp_wmb().
861 * For making pc->mem_cgroup visible, insert smp_rmb() here.
864 if (!PageCgroupUsed(pc
))
867 mz
= page_cgroup_zoneinfo(pc
);
868 MEM_CGROUP_ZSTAT(mz
, lru
) += 1;
869 SetPageCgroupAcctLRU(pc
);
870 if (mem_cgroup_is_root(pc
->mem_cgroup
))
872 list_add(&pc
->lru
, &mz
->lists
[lru
]);
876 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
877 * lru because the page may.be reused after it's fully uncharged (because of
878 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
879 * it again. This function is only used to charge SwapCache. It's done under
880 * lock_page and expected that zone->lru_lock is never held.
882 static void mem_cgroup_lru_del_before_commit_swapcache(struct page
*page
)
885 struct zone
*zone
= page_zone(page
);
886 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
888 spin_lock_irqsave(&zone
->lru_lock
, flags
);
890 * Forget old LRU when this page_cgroup is *not* used. This Used bit
891 * is guarded by lock_page() because the page is SwapCache.
893 if (!PageCgroupUsed(pc
))
894 mem_cgroup_del_lru_list(page
, page_lru(page
));
895 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
898 static void mem_cgroup_lru_add_after_commit_swapcache(struct page
*page
)
901 struct zone
*zone
= page_zone(page
);
902 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
904 spin_lock_irqsave(&zone
->lru_lock
, flags
);
905 /* link when the page is linked to LRU but page_cgroup isn't */
906 if (PageLRU(page
) && !PageCgroupAcctLRU(pc
))
907 mem_cgroup_add_lru_list(page
, page_lru(page
));
908 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
912 void mem_cgroup_move_lists(struct page
*page
,
913 enum lru_list from
, enum lru_list to
)
915 if (mem_cgroup_disabled())
917 mem_cgroup_del_lru_list(page
, from
);
918 mem_cgroup_add_lru_list(page
, to
);
921 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
924 struct mem_cgroup
*curr
= NULL
;
925 struct task_struct
*p
;
927 p
= find_lock_task_mm(task
);
930 curr
= try_get_mem_cgroup_from_mm(p
->mm
);
935 * We should check use_hierarchy of "mem" not "curr". Because checking
936 * use_hierarchy of "curr" here make this function true if hierarchy is
937 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
938 * hierarchy(even if use_hierarchy is disabled in "mem").
940 if (mem
->use_hierarchy
)
941 ret
= css_is_ancestor(&curr
->css
, &mem
->css
);
948 static int calc_inactive_ratio(struct mem_cgroup
*memcg
, unsigned long *present_pages
)
950 unsigned long active
;
951 unsigned long inactive
;
953 unsigned long inactive_ratio
;
955 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_ANON
);
956 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_ANON
);
958 gb
= (inactive
+ active
) >> (30 - PAGE_SHIFT
);
960 inactive_ratio
= int_sqrt(10 * gb
);
965 present_pages
[0] = inactive
;
966 present_pages
[1] = active
;
969 return inactive_ratio
;
972 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup
*memcg
)
974 unsigned long active
;
975 unsigned long inactive
;
976 unsigned long present_pages
[2];
977 unsigned long inactive_ratio
;
979 inactive_ratio
= calc_inactive_ratio(memcg
, present_pages
);
981 inactive
= present_pages
[0];
982 active
= present_pages
[1];
984 if (inactive
* inactive_ratio
< active
)
990 int mem_cgroup_inactive_file_is_low(struct mem_cgroup
*memcg
)
992 unsigned long active
;
993 unsigned long inactive
;
995 inactive
= mem_cgroup_get_local_zonestat(memcg
, LRU_INACTIVE_FILE
);
996 active
= mem_cgroup_get_local_zonestat(memcg
, LRU_ACTIVE_FILE
);
998 return (active
> inactive
);
1001 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup
*memcg
,
1005 int nid
= zone_to_nid(zone
);
1006 int zid
= zone_idx(zone
);
1007 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
1009 return MEM_CGROUP_ZSTAT(mz
, lru
);
1012 struct zone_reclaim_stat
*mem_cgroup_get_reclaim_stat(struct mem_cgroup
*memcg
,
1015 int nid
= zone_to_nid(zone
);
1016 int zid
= zone_idx(zone
);
1017 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(memcg
, nid
, zid
);
1019 return &mz
->reclaim_stat
;
1022 struct zone_reclaim_stat
*
1023 mem_cgroup_get_reclaim_stat_from_page(struct page
*page
)
1025 struct page_cgroup
*pc
;
1026 struct mem_cgroup_per_zone
*mz
;
1028 if (mem_cgroup_disabled())
1031 pc
= lookup_page_cgroup(page
);
1033 * Used bit is set without atomic ops but after smp_wmb().
1034 * For making pc->mem_cgroup visible, insert smp_rmb() here.
1037 if (!PageCgroupUsed(pc
))
1040 mz
= page_cgroup_zoneinfo(pc
);
1044 return &mz
->reclaim_stat
;
1047 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
1048 struct list_head
*dst
,
1049 unsigned long *scanned
, int order
,
1050 int mode
, struct zone
*z
,
1051 struct mem_cgroup
*mem_cont
,
1052 int active
, int file
)
1054 unsigned long nr_taken
= 0;
1058 struct list_head
*src
;
1059 struct page_cgroup
*pc
, *tmp
;
1060 int nid
= zone_to_nid(z
);
1061 int zid
= zone_idx(z
);
1062 struct mem_cgroup_per_zone
*mz
;
1063 int lru
= LRU_FILE
* file
+ active
;
1067 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
1068 src
= &mz
->lists
[lru
];
1071 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
1072 if (scan
>= nr_to_scan
)
1076 if (unlikely(!PageCgroupUsed(pc
)))
1078 if (unlikely(!PageLRU(page
)))
1082 ret
= __isolate_lru_page(page
, mode
, file
);
1085 list_move(&page
->lru
, dst
);
1086 mem_cgroup_del_lru(page
);
1087 nr_taken
+= hpage_nr_pages(page
);
1090 /* we don't affect global LRU but rotate in our LRU */
1091 mem_cgroup_rotate_lru_list(page
, page_lru(page
));
1100 trace_mm_vmscan_memcg_isolate(0, nr_to_scan
, scan
, nr_taken
,
1106 #define mem_cgroup_from_res_counter(counter, member) \
1107 container_of(counter, struct mem_cgroup, member)
1109 static bool mem_cgroup_check_under_limit(struct mem_cgroup
*mem
)
1111 if (do_swap_account
) {
1112 if (res_counter_check_under_limit(&mem
->res
) &&
1113 res_counter_check_under_limit(&mem
->memsw
))
1116 if (res_counter_check_under_limit(&mem
->res
))
1121 static unsigned int get_swappiness(struct mem_cgroup
*memcg
)
1123 struct cgroup
*cgrp
= memcg
->css
.cgroup
;
1124 unsigned int swappiness
;
1127 if (cgrp
->parent
== NULL
)
1128 return vm_swappiness
;
1130 spin_lock(&memcg
->reclaim_param_lock
);
1131 swappiness
= memcg
->swappiness
;
1132 spin_unlock(&memcg
->reclaim_param_lock
);
1137 static void mem_cgroup_start_move(struct mem_cgroup
*mem
)
1142 spin_lock(&mem
->pcp_counter_lock
);
1143 for_each_online_cpu(cpu
)
1144 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) += 1;
1145 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] += 1;
1146 spin_unlock(&mem
->pcp_counter_lock
);
1152 static void mem_cgroup_end_move(struct mem_cgroup
*mem
)
1159 spin_lock(&mem
->pcp_counter_lock
);
1160 for_each_online_cpu(cpu
)
1161 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) -= 1;
1162 mem
->nocpu_base
.count
[MEM_CGROUP_ON_MOVE
] -= 1;
1163 spin_unlock(&mem
->pcp_counter_lock
);
1167 * 2 routines for checking "mem" is under move_account() or not.
1169 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1170 * for avoiding race in accounting. If true,
1171 * pc->mem_cgroup may be overwritten.
1173 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1174 * under hierarchy of moving cgroups. This is for
1175 * waiting at hith-memory prressure caused by "move".
1178 static bool mem_cgroup_stealed(struct mem_cgroup
*mem
)
1180 VM_BUG_ON(!rcu_read_lock_held());
1181 return this_cpu_read(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
]) > 0;
1184 static bool mem_cgroup_under_move(struct mem_cgroup
*mem
)
1186 struct mem_cgroup
*from
;
1187 struct mem_cgroup
*to
;
1190 * Unlike task_move routines, we access mc.to, mc.from not under
1191 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1193 spin_lock(&mc
.lock
);
1198 if (from
== mem
|| to
== mem
1199 || (mem
->use_hierarchy
&& css_is_ancestor(&from
->css
, &mem
->css
))
1200 || (mem
->use_hierarchy
&& css_is_ancestor(&to
->css
, &mem
->css
)))
1203 spin_unlock(&mc
.lock
);
1207 static bool mem_cgroup_wait_acct_move(struct mem_cgroup
*mem
)
1209 if (mc
.moving_task
&& current
!= mc
.moving_task
) {
1210 if (mem_cgroup_under_move(mem
)) {
1212 prepare_to_wait(&mc
.waitq
, &wait
, TASK_INTERRUPTIBLE
);
1213 /* moving charge context might have finished. */
1216 finish_wait(&mc
.waitq
, &wait
);
1224 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1225 * @memcg: The memory cgroup that went over limit
1226 * @p: Task that is going to be killed
1228 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1231 void mem_cgroup_print_oom_info(struct mem_cgroup
*memcg
, struct task_struct
*p
)
1233 struct cgroup
*task_cgrp
;
1234 struct cgroup
*mem_cgrp
;
1236 * Need a buffer in BSS, can't rely on allocations. The code relies
1237 * on the assumption that OOM is serialized for memory controller.
1238 * If this assumption is broken, revisit this code.
1240 static char memcg_name
[PATH_MAX
];
1249 mem_cgrp
= memcg
->css
.cgroup
;
1250 task_cgrp
= task_cgroup(p
, mem_cgroup_subsys_id
);
1252 ret
= cgroup_path(task_cgrp
, memcg_name
, PATH_MAX
);
1255 * Unfortunately, we are unable to convert to a useful name
1256 * But we'll still print out the usage information
1263 printk(KERN_INFO
"Task in %s killed", memcg_name
);
1266 ret
= cgroup_path(mem_cgrp
, memcg_name
, PATH_MAX
);
1274 * Continues from above, so we don't need an KERN_ level
1276 printk(KERN_CONT
" as a result of limit of %s\n", memcg_name
);
1279 printk(KERN_INFO
"memory: usage %llukB, limit %llukB, failcnt %llu\n",
1280 res_counter_read_u64(&memcg
->res
, RES_USAGE
) >> 10,
1281 res_counter_read_u64(&memcg
->res
, RES_LIMIT
) >> 10,
1282 res_counter_read_u64(&memcg
->res
, RES_FAILCNT
));
1283 printk(KERN_INFO
"memory+swap: usage %llukB, limit %llukB, "
1285 res_counter_read_u64(&memcg
->memsw
, RES_USAGE
) >> 10,
1286 res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
) >> 10,
1287 res_counter_read_u64(&memcg
->memsw
, RES_FAILCNT
));
1291 * This function returns the number of memcg under hierarchy tree. Returns
1292 * 1(self count) if no children.
1294 static int mem_cgroup_count_children(struct mem_cgroup
*mem
)
1297 struct mem_cgroup
*iter
;
1299 for_each_mem_cgroup_tree(iter
, mem
)
1305 * Return the memory (and swap, if configured) limit for a memcg.
1307 u64
mem_cgroup_get_limit(struct mem_cgroup
*memcg
)
1312 limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
1313 limit
+= total_swap_pages
<< PAGE_SHIFT
;
1315 memsw
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
1317 * If memsw is finite and limits the amount of swap space available
1318 * to this memcg, return that limit.
1320 return min(limit
, memsw
);
1324 * Visit the first child (need not be the first child as per the ordering
1325 * of the cgroup list, since we track last_scanned_child) of @mem and use
1326 * that to reclaim free pages from.
1328 static struct mem_cgroup
*
1329 mem_cgroup_select_victim(struct mem_cgroup
*root_mem
)
1331 struct mem_cgroup
*ret
= NULL
;
1332 struct cgroup_subsys_state
*css
;
1335 if (!root_mem
->use_hierarchy
) {
1336 css_get(&root_mem
->css
);
1342 nextid
= root_mem
->last_scanned_child
+ 1;
1343 css
= css_get_next(&mem_cgroup_subsys
, nextid
, &root_mem
->css
,
1345 if (css
&& css_tryget(css
))
1346 ret
= container_of(css
, struct mem_cgroup
, css
);
1349 /* Updates scanning parameter */
1350 spin_lock(&root_mem
->reclaim_param_lock
);
1352 /* this means start scan from ID:1 */
1353 root_mem
->last_scanned_child
= 0;
1355 root_mem
->last_scanned_child
= found
;
1356 spin_unlock(&root_mem
->reclaim_param_lock
);
1363 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1364 * we reclaimed from, so that we don't end up penalizing one child extensively
1365 * based on its position in the children list.
1367 * root_mem is the original ancestor that we've been reclaim from.
1369 * We give up and return to the caller when we visit root_mem twice.
1370 * (other groups can be removed while we're walking....)
1372 * If shrink==true, for avoiding to free too much, this returns immedieately.
1374 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup
*root_mem
,
1377 unsigned long reclaim_options
)
1379 struct mem_cgroup
*victim
;
1382 bool noswap
= reclaim_options
& MEM_CGROUP_RECLAIM_NOSWAP
;
1383 bool shrink
= reclaim_options
& MEM_CGROUP_RECLAIM_SHRINK
;
1384 bool check_soft
= reclaim_options
& MEM_CGROUP_RECLAIM_SOFT
;
1385 unsigned long excess
= mem_cgroup_get_excess(root_mem
);
1387 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1388 if (root_mem
->memsw_is_minimum
)
1392 victim
= mem_cgroup_select_victim(root_mem
);
1393 if (victim
== root_mem
) {
1396 drain_all_stock_async();
1399 * If we have not been able to reclaim
1400 * anything, it might because there are
1401 * no reclaimable pages under this hierarchy
1403 if (!check_soft
|| !total
) {
1404 css_put(&victim
->css
);
1408 * We want to do more targetted reclaim.
1409 * excess >> 2 is not to excessive so as to
1410 * reclaim too much, nor too less that we keep
1411 * coming back to reclaim from this cgroup
1413 if (total
>= (excess
>> 2) ||
1414 (loop
> MEM_CGROUP_MAX_RECLAIM_LOOPS
)) {
1415 css_put(&victim
->css
);
1420 if (!mem_cgroup_local_usage(victim
)) {
1421 /* this cgroup's local usage == 0 */
1422 css_put(&victim
->css
);
1425 /* we use swappiness of local cgroup */
1427 ret
= mem_cgroup_shrink_node_zone(victim
, gfp_mask
,
1428 noswap
, get_swappiness(victim
), zone
);
1430 ret
= try_to_free_mem_cgroup_pages(victim
, gfp_mask
,
1431 noswap
, get_swappiness(victim
));
1432 css_put(&victim
->css
);
1434 * At shrinking usage, we can't check we should stop here or
1435 * reclaim more. It's depends on callers. last_scanned_child
1436 * will work enough for keeping fairness under tree.
1442 if (res_counter_check_under_soft_limit(&root_mem
->res
))
1444 } else if (mem_cgroup_check_under_limit(root_mem
))
1451 * Check OOM-Killer is already running under our hierarchy.
1452 * If someone is running, return false.
1454 static bool mem_cgroup_oom_lock(struct mem_cgroup
*mem
)
1456 int x
, lock_count
= 0;
1457 struct mem_cgroup
*iter
;
1459 for_each_mem_cgroup_tree(iter
, mem
) {
1460 x
= atomic_inc_return(&iter
->oom_lock
);
1461 lock_count
= max(x
, lock_count
);
1464 if (lock_count
== 1)
1469 static int mem_cgroup_oom_unlock(struct mem_cgroup
*mem
)
1471 struct mem_cgroup
*iter
;
1474 * When a new child is created while the hierarchy is under oom,
1475 * mem_cgroup_oom_lock() may not be called. We have to use
1476 * atomic_add_unless() here.
1478 for_each_mem_cgroup_tree(iter
, mem
)
1479 atomic_add_unless(&iter
->oom_lock
, -1, 0);
1484 static DEFINE_MUTEX(memcg_oom_mutex
);
1485 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq
);
1487 struct oom_wait_info
{
1488 struct mem_cgroup
*mem
;
1492 static int memcg_oom_wake_function(wait_queue_t
*wait
,
1493 unsigned mode
, int sync
, void *arg
)
1495 struct mem_cgroup
*wake_mem
= (struct mem_cgroup
*)arg
;
1496 struct oom_wait_info
*oom_wait_info
;
1498 oom_wait_info
= container_of(wait
, struct oom_wait_info
, wait
);
1500 if (oom_wait_info
->mem
== wake_mem
)
1502 /* if no hierarchy, no match */
1503 if (!oom_wait_info
->mem
->use_hierarchy
|| !wake_mem
->use_hierarchy
)
1506 * Both of oom_wait_info->mem and wake_mem are stable under us.
1507 * Then we can use css_is_ancestor without taking care of RCU.
1509 if (!css_is_ancestor(&oom_wait_info
->mem
->css
, &wake_mem
->css
) &&
1510 !css_is_ancestor(&wake_mem
->css
, &oom_wait_info
->mem
->css
))
1514 return autoremove_wake_function(wait
, mode
, sync
, arg
);
1517 static void memcg_wakeup_oom(struct mem_cgroup
*mem
)
1519 /* for filtering, pass "mem" as argument. */
1520 __wake_up(&memcg_oom_waitq
, TASK_NORMAL
, 0, mem
);
1523 static void memcg_oom_recover(struct mem_cgroup
*mem
)
1525 if (mem
&& atomic_read(&mem
->oom_lock
))
1526 memcg_wakeup_oom(mem
);
1530 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1532 bool mem_cgroup_handle_oom(struct mem_cgroup
*mem
, gfp_t mask
)
1534 struct oom_wait_info owait
;
1535 bool locked
, need_to_kill
;
1538 owait
.wait
.flags
= 0;
1539 owait
.wait
.func
= memcg_oom_wake_function
;
1540 owait
.wait
.private = current
;
1541 INIT_LIST_HEAD(&owait
.wait
.task_list
);
1542 need_to_kill
= true;
1543 /* At first, try to OOM lock hierarchy under mem.*/
1544 mutex_lock(&memcg_oom_mutex
);
1545 locked
= mem_cgroup_oom_lock(mem
);
1547 * Even if signal_pending(), we can't quit charge() loop without
1548 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1549 * under OOM is always welcomed, use TASK_KILLABLE here.
1551 prepare_to_wait(&memcg_oom_waitq
, &owait
.wait
, TASK_KILLABLE
);
1552 if (!locked
|| mem
->oom_kill_disable
)
1553 need_to_kill
= false;
1555 mem_cgroup_oom_notify(mem
);
1556 mutex_unlock(&memcg_oom_mutex
);
1559 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1560 mem_cgroup_out_of_memory(mem
, mask
);
1563 finish_wait(&memcg_oom_waitq
, &owait
.wait
);
1565 mutex_lock(&memcg_oom_mutex
);
1566 mem_cgroup_oom_unlock(mem
);
1567 memcg_wakeup_oom(mem
);
1568 mutex_unlock(&memcg_oom_mutex
);
1570 if (test_thread_flag(TIF_MEMDIE
) || fatal_signal_pending(current
))
1572 /* Give chance to dying process */
1573 schedule_timeout(1);
1578 * Currently used to update mapped file statistics, but the routine can be
1579 * generalized to update other statistics as well.
1581 * Notes: Race condition
1583 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1584 * it tends to be costly. But considering some conditions, we doesn't need
1585 * to do so _always_.
1587 * Considering "charge", lock_page_cgroup() is not required because all
1588 * file-stat operations happen after a page is attached to radix-tree. There
1589 * are no race with "charge".
1591 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1592 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1593 * if there are race with "uncharge". Statistics itself is properly handled
1596 * Considering "move", this is an only case we see a race. To make the race
1597 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1598 * possibility of race condition. If there is, we take a lock.
1601 void mem_cgroup_update_page_stat(struct page
*page
,
1602 enum mem_cgroup_page_stat_item idx
, int val
)
1604 struct mem_cgroup
*mem
;
1605 struct page_cgroup
*pc
= lookup_page_cgroup(page
);
1606 bool need_unlock
= false;
1607 unsigned long uninitialized_var(flags
);
1613 mem
= pc
->mem_cgroup
;
1614 if (unlikely(!mem
|| !PageCgroupUsed(pc
)))
1616 /* pc->mem_cgroup is unstable ? */
1617 if (unlikely(mem_cgroup_stealed(mem
)) || PageTransHuge(page
)) {
1618 /* take a lock against to access pc->mem_cgroup */
1619 move_lock_page_cgroup(pc
, &flags
);
1621 mem
= pc
->mem_cgroup
;
1622 if (!mem
|| !PageCgroupUsed(pc
))
1627 case MEMCG_NR_FILE_MAPPED
:
1629 SetPageCgroupFileMapped(pc
);
1630 else if (!page_mapped(page
))
1631 ClearPageCgroupFileMapped(pc
);
1632 idx
= MEM_CGROUP_STAT_FILE_MAPPED
;
1638 this_cpu_add(mem
->stat
->count
[idx
], val
);
1641 if (unlikely(need_unlock
))
1642 move_unlock_page_cgroup(pc
, &flags
);
1646 EXPORT_SYMBOL(mem_cgroup_update_page_stat
);
1649 * size of first charge trial. "32" comes from vmscan.c's magic value.
1650 * TODO: maybe necessary to use big numbers in big irons.
1652 #define CHARGE_SIZE (32 * PAGE_SIZE)
1653 struct memcg_stock_pcp
{
1654 struct mem_cgroup
*cached
; /* this never be root cgroup */
1656 struct work_struct work
;
1658 static DEFINE_PER_CPU(struct memcg_stock_pcp
, memcg_stock
);
1659 static atomic_t memcg_drain_count
;
1662 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1663 * from local stock and true is returned. If the stock is 0 or charges from a
1664 * cgroup which is not current target, returns false. This stock will be
1667 static bool consume_stock(struct mem_cgroup
*mem
)
1669 struct memcg_stock_pcp
*stock
;
1672 stock
= &get_cpu_var(memcg_stock
);
1673 if (mem
== stock
->cached
&& stock
->charge
)
1674 stock
->charge
-= PAGE_SIZE
;
1675 else /* need to call res_counter_charge */
1677 put_cpu_var(memcg_stock
);
1682 * Returns stocks cached in percpu to res_counter and reset cached information.
1684 static void drain_stock(struct memcg_stock_pcp
*stock
)
1686 struct mem_cgroup
*old
= stock
->cached
;
1688 if (stock
->charge
) {
1689 res_counter_uncharge(&old
->res
, stock
->charge
);
1690 if (do_swap_account
)
1691 res_counter_uncharge(&old
->memsw
, stock
->charge
);
1693 stock
->cached
= NULL
;
1698 * This must be called under preempt disabled or must be called by
1699 * a thread which is pinned to local cpu.
1701 static void drain_local_stock(struct work_struct
*dummy
)
1703 struct memcg_stock_pcp
*stock
= &__get_cpu_var(memcg_stock
);
1708 * Cache charges(val) which is from res_counter, to local per_cpu area.
1709 * This will be consumed by consume_stock() function, later.
1711 static void refill_stock(struct mem_cgroup
*mem
, int val
)
1713 struct memcg_stock_pcp
*stock
= &get_cpu_var(memcg_stock
);
1715 if (stock
->cached
!= mem
) { /* reset if necessary */
1717 stock
->cached
= mem
;
1719 stock
->charge
+= val
;
1720 put_cpu_var(memcg_stock
);
1724 * Tries to drain stocked charges in other cpus. This function is asynchronous
1725 * and just put a work per cpu for draining localy on each cpu. Caller can
1726 * expects some charges will be back to res_counter later but cannot wait for
1729 static void drain_all_stock_async(void)
1732 /* This function is for scheduling "drain" in asynchronous way.
1733 * The result of "drain" is not directly handled by callers. Then,
1734 * if someone is calling drain, we don't have to call drain more.
1735 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1736 * there is a race. We just do loose check here.
1738 if (atomic_read(&memcg_drain_count
))
1740 /* Notify other cpus that system-wide "drain" is running */
1741 atomic_inc(&memcg_drain_count
);
1743 for_each_online_cpu(cpu
) {
1744 struct memcg_stock_pcp
*stock
= &per_cpu(memcg_stock
, cpu
);
1745 schedule_work_on(cpu
, &stock
->work
);
1748 atomic_dec(&memcg_drain_count
);
1749 /* We don't wait for flush_work */
1752 /* This is a synchronous drain interface. */
1753 static void drain_all_stock_sync(void)
1755 /* called when force_empty is called */
1756 atomic_inc(&memcg_drain_count
);
1757 schedule_on_each_cpu(drain_local_stock
);
1758 atomic_dec(&memcg_drain_count
);
1762 * This function drains percpu counter value from DEAD cpu and
1763 * move it to local cpu. Note that this function can be preempted.
1765 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup
*mem
, int cpu
)
1769 spin_lock(&mem
->pcp_counter_lock
);
1770 for (i
= 0; i
< MEM_CGROUP_STAT_DATA
; i
++) {
1771 s64 x
= per_cpu(mem
->stat
->count
[i
], cpu
);
1773 per_cpu(mem
->stat
->count
[i
], cpu
) = 0;
1774 mem
->nocpu_base
.count
[i
] += x
;
1776 /* need to clear ON_MOVE value, works as a kind of lock. */
1777 per_cpu(mem
->stat
->count
[MEM_CGROUP_ON_MOVE
], cpu
) = 0;
1778 spin_unlock(&mem
->pcp_counter_lock
);
1781 static void synchronize_mem_cgroup_on_move(struct mem_cgroup
*mem
, int cpu
)
1783 int idx
= MEM_CGROUP_ON_MOVE
;
1785 spin_lock(&mem
->pcp_counter_lock
);
1786 per_cpu(mem
->stat
->count
[idx
], cpu
) = mem
->nocpu_base
.count
[idx
];
1787 spin_unlock(&mem
->pcp_counter_lock
);
1790 static int __cpuinit
memcg_cpu_hotplug_callback(struct notifier_block
*nb
,
1791 unsigned long action
,
1794 int cpu
= (unsigned long)hcpu
;
1795 struct memcg_stock_pcp
*stock
;
1796 struct mem_cgroup
*iter
;
1798 if ((action
== CPU_ONLINE
)) {
1799 for_each_mem_cgroup_all(iter
)
1800 synchronize_mem_cgroup_on_move(iter
, cpu
);
1804 if ((action
!= CPU_DEAD
) || action
!= CPU_DEAD_FROZEN
)
1807 for_each_mem_cgroup_all(iter
)
1808 mem_cgroup_drain_pcp_counter(iter
, cpu
);
1810 stock
= &per_cpu(memcg_stock
, cpu
);
1816 /* See __mem_cgroup_try_charge() for details */
1818 CHARGE_OK
, /* success */
1819 CHARGE_RETRY
, /* need to retry but retry is not bad */
1820 CHARGE_NOMEM
, /* we can't do more. return -ENOMEM */
1821 CHARGE_WOULDBLOCK
, /* GFP_WAIT wasn't set and no enough res. */
1822 CHARGE_OOM_DIE
, /* the current is killed because of OOM */
1825 static int __mem_cgroup_do_charge(struct mem_cgroup
*mem
, gfp_t gfp_mask
,
1826 int csize
, bool oom_check
)
1828 struct mem_cgroup
*mem_over_limit
;
1829 struct res_counter
*fail_res
;
1830 unsigned long flags
= 0;
1833 ret
= res_counter_charge(&mem
->res
, csize
, &fail_res
);
1836 if (!do_swap_account
)
1838 ret
= res_counter_charge(&mem
->memsw
, csize
, &fail_res
);
1842 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, memsw
);
1843 flags
|= MEM_CGROUP_RECLAIM_NOSWAP
;
1845 mem_over_limit
= mem_cgroup_from_res_counter(fail_res
, res
);
1847 if (csize
> PAGE_SIZE
) /* change csize and retry */
1848 return CHARGE_RETRY
;
1850 if (!(gfp_mask
& __GFP_WAIT
))
1851 return CHARGE_WOULDBLOCK
;
1853 ret
= mem_cgroup_hierarchical_reclaim(mem_over_limit
, NULL
,
1856 * try_to_free_mem_cgroup_pages() might not give us a full
1857 * picture of reclaim. Some pages are reclaimed and might be
1858 * moved to swap cache or just unmapped from the cgroup.
1859 * Check the limit again to see if the reclaim reduced the
1860 * current usage of the cgroup before giving up
1862 if (ret
|| mem_cgroup_check_under_limit(mem_over_limit
))
1863 return CHARGE_RETRY
;
1866 * At task move, charge accounts can be doubly counted. So, it's
1867 * better to wait until the end of task_move if something is going on.
1869 if (mem_cgroup_wait_acct_move(mem_over_limit
))
1870 return CHARGE_RETRY
;
1872 /* If we don't need to call oom-killer at el, return immediately */
1874 return CHARGE_NOMEM
;
1876 if (!mem_cgroup_handle_oom(mem_over_limit
, gfp_mask
))
1877 return CHARGE_OOM_DIE
;
1879 return CHARGE_RETRY
;
1883 * Unlike exported interface, "oom" parameter is added. if oom==true,
1884 * oom-killer can be invoked.
1886 static int __mem_cgroup_try_charge(struct mm_struct
*mm
,
1888 struct mem_cgroup
**memcg
, bool oom
,
1891 int nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1892 struct mem_cgroup
*mem
= NULL
;
1894 int csize
= max(CHARGE_SIZE
, (unsigned long) page_size
);
1897 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1898 * in system level. So, allow to go ahead dying process in addition to
1901 if (unlikely(test_thread_flag(TIF_MEMDIE
)
1902 || fatal_signal_pending(current
)))
1906 * We always charge the cgroup the mm_struct belongs to.
1907 * The mm_struct's mem_cgroup changes on task migration if the
1908 * thread group leader migrates. It's possible that mm is not
1909 * set, if so charge the init_mm (happens for pagecache usage).
1914 if (*memcg
) { /* css should be a valid one */
1916 VM_BUG_ON(css_is_removed(&mem
->css
));
1917 if (mem_cgroup_is_root(mem
))
1919 if (page_size
== PAGE_SIZE
&& consume_stock(mem
))
1923 struct task_struct
*p
;
1926 p
= rcu_dereference(mm
->owner
);
1928 * Because we don't have task_lock(), "p" can exit.
1929 * In that case, "mem" can point to root or p can be NULL with
1930 * race with swapoff. Then, we have small risk of mis-accouning.
1931 * But such kind of mis-account by race always happens because
1932 * we don't have cgroup_mutex(). It's overkill and we allo that
1934 * (*) swapoff at el will charge against mm-struct not against
1935 * task-struct. So, mm->owner can be NULL.
1937 mem
= mem_cgroup_from_task(p
);
1938 if (!mem
|| mem_cgroup_is_root(mem
)) {
1942 if (page_size
== PAGE_SIZE
&& consume_stock(mem
)) {
1944 * It seems dagerous to access memcg without css_get().
1945 * But considering how consume_stok works, it's not
1946 * necessary. If consume_stock success, some charges
1947 * from this memcg are cached on this cpu. So, we
1948 * don't need to call css_get()/css_tryget() before
1949 * calling consume_stock().
1954 /* after here, we may be blocked. we need to get refcnt */
1955 if (!css_tryget(&mem
->css
)) {
1965 /* If killed, bypass charge */
1966 if (fatal_signal_pending(current
)) {
1972 if (oom
&& !nr_oom_retries
) {
1974 nr_oom_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
1977 ret
= __mem_cgroup_do_charge(mem
, gfp_mask
, csize
, oom_check
);
1982 case CHARGE_RETRY
: /* not in OOM situation but retry */
1987 case CHARGE_WOULDBLOCK
: /* !__GFP_WAIT */
1990 case CHARGE_NOMEM
: /* OOM routine works */
1995 /* If oom, we never return -ENOMEM */
1998 case CHARGE_OOM_DIE
: /* Killed by OOM Killer */
2002 } while (ret
!= CHARGE_OK
);
2004 if (csize
> page_size
)
2005 refill_stock(mem
, csize
- page_size
);
2019 * Somemtimes we have to undo a charge we got by try_charge().
2020 * This function is for that and do uncharge, put css's refcnt.
2021 * gotten by try_charge().
2023 static void __mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2024 unsigned long count
)
2026 if (!mem_cgroup_is_root(mem
)) {
2027 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
2028 if (do_swap_account
)
2029 res_counter_uncharge(&mem
->memsw
, PAGE_SIZE
* count
);
2033 static void mem_cgroup_cancel_charge(struct mem_cgroup
*mem
,
2036 __mem_cgroup_cancel_charge(mem
, page_size
>> PAGE_SHIFT
);
2040 * A helper function to get mem_cgroup from ID. must be called under
2041 * rcu_read_lock(). The caller must check css_is_removed() or some if
2042 * it's concern. (dropping refcnt from swap can be called against removed
2045 static struct mem_cgroup
*mem_cgroup_lookup(unsigned short id
)
2047 struct cgroup_subsys_state
*css
;
2049 /* ID 0 is unused ID */
2052 css
= css_lookup(&mem_cgroup_subsys
, id
);
2055 return container_of(css
, struct mem_cgroup
, css
);
2058 struct mem_cgroup
*try_get_mem_cgroup_from_page(struct page
*page
)
2060 struct mem_cgroup
*mem
= NULL
;
2061 struct page_cgroup
*pc
;
2065 VM_BUG_ON(!PageLocked(page
));
2067 pc
= lookup_page_cgroup(page
);
2068 lock_page_cgroup(pc
);
2069 if (PageCgroupUsed(pc
)) {
2070 mem
= pc
->mem_cgroup
;
2071 if (mem
&& !css_tryget(&mem
->css
))
2073 } else if (PageSwapCache(page
)) {
2074 ent
.val
= page_private(page
);
2075 id
= lookup_swap_cgroup(ent
);
2077 mem
= mem_cgroup_lookup(id
);
2078 if (mem
&& !css_tryget(&mem
->css
))
2082 unlock_page_cgroup(pc
);
2086 static void __mem_cgroup_commit_charge(struct mem_cgroup
*mem
,
2087 struct page_cgroup
*pc
,
2088 enum charge_type ctype
,
2091 int nr_pages
= page_size
>> PAGE_SHIFT
;
2093 /* try_charge() can return NULL to *memcg, taking care of it. */
2097 lock_page_cgroup(pc
);
2098 if (unlikely(PageCgroupUsed(pc
))) {
2099 unlock_page_cgroup(pc
);
2100 mem_cgroup_cancel_charge(mem
, page_size
);
2104 * we don't need page_cgroup_lock about tail pages, becase they are not
2105 * accessed by any other context at this point.
2107 pc
->mem_cgroup
= mem
;
2109 * We access a page_cgroup asynchronously without lock_page_cgroup().
2110 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2111 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2112 * before USED bit, we need memory barrier here.
2113 * See mem_cgroup_add_lru_list(), etc.
2117 case MEM_CGROUP_CHARGE_TYPE_CACHE
:
2118 case MEM_CGROUP_CHARGE_TYPE_SHMEM
:
2119 SetPageCgroupCache(pc
);
2120 SetPageCgroupUsed(pc
);
2122 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2123 ClearPageCgroupCache(pc
);
2124 SetPageCgroupUsed(pc
);
2130 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), nr_pages
);
2131 unlock_page_cgroup(pc
);
2133 * "charge_statistics" updated event counter. Then, check it.
2134 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2135 * if they exceeds softlimit.
2137 memcg_check_events(mem
, pc
->page
);
2140 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2142 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2143 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2145 * Because tail pages are not marked as "used", set it. We're under
2146 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2148 void mem_cgroup_split_huge_fixup(struct page
*head
, struct page
*tail
)
2150 struct page_cgroup
*head_pc
= lookup_page_cgroup(head
);
2151 struct page_cgroup
*tail_pc
= lookup_page_cgroup(tail
);
2152 unsigned long flags
;
2155 * We have no races witch charge/uncharge but will have races with
2156 * page state accounting.
2158 move_lock_page_cgroup(head_pc
, &flags
);
2160 tail_pc
->mem_cgroup
= head_pc
->mem_cgroup
;
2161 smp_wmb(); /* see __commit_charge() */
2162 /* we don't need to copy all flags...*/
2163 tail_pc
->flags
= head_pc
->flags
& ~PCGF_NOCOPY_AT_SPLIT
;
2164 move_unlock_page_cgroup(head_pc
, &flags
);
2169 * __mem_cgroup_move_account - move account of the page
2170 * @pc: page_cgroup of the page.
2171 * @from: mem_cgroup which the page is moved from.
2172 * @to: mem_cgroup which the page is moved to. @from != @to.
2173 * @uncharge: whether we should call uncharge and css_put against @from.
2175 * The caller must confirm following.
2176 * - page is not on LRU (isolate_page() is useful.)
2177 * - the pc is locked, used, and ->mem_cgroup points to @from.
2179 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2180 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2181 * true, this function does "uncharge" from old cgroup, but it doesn't if
2182 * @uncharge is false, so a caller should do "uncharge".
2185 static void __mem_cgroup_move_account(struct page_cgroup
*pc
,
2186 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool uncharge
)
2188 VM_BUG_ON(from
== to
);
2189 VM_BUG_ON(PageLRU(pc
->page
));
2190 VM_BUG_ON(!page_is_cgroup_locked(pc
));
2191 VM_BUG_ON(!PageCgroupUsed(pc
));
2192 VM_BUG_ON(pc
->mem_cgroup
!= from
);
2194 if (PageCgroupFileMapped(pc
)) {
2195 /* Update mapped_file data for mem_cgroup */
2197 __this_cpu_dec(from
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
2198 __this_cpu_inc(to
->stat
->count
[MEM_CGROUP_STAT_FILE_MAPPED
]);
2201 mem_cgroup_charge_statistics(from
, PageCgroupCache(pc
), -1);
2203 /* This is not "cancel", but cancel_charge does all we need. */
2204 mem_cgroup_cancel_charge(from
, PAGE_SIZE
);
2206 /* caller should have done css_get */
2207 pc
->mem_cgroup
= to
;
2208 mem_cgroup_charge_statistics(to
, PageCgroupCache(pc
), 1);
2210 * We charges against "to" which may not have any tasks. Then, "to"
2211 * can be under rmdir(). But in current implementation, caller of
2212 * this function is just force_empty() and move charge, so it's
2213 * garanteed that "to" is never removed. So, we don't check rmdir
2219 * check whether the @pc is valid for moving account and call
2220 * __mem_cgroup_move_account()
2222 static int mem_cgroup_move_account(struct page_cgroup
*pc
,
2223 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool uncharge
)
2226 unsigned long flags
;
2228 lock_page_cgroup(pc
);
2229 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== from
) {
2230 move_lock_page_cgroup(pc
, &flags
);
2231 __mem_cgroup_move_account(pc
, from
, to
, uncharge
);
2232 move_unlock_page_cgroup(pc
, &flags
);
2235 unlock_page_cgroup(pc
);
2239 memcg_check_events(to
, pc
->page
);
2240 memcg_check_events(from
, pc
->page
);
2245 * move charges to its parent.
2248 static int mem_cgroup_move_parent(struct page_cgroup
*pc
,
2249 struct mem_cgroup
*child
,
2252 struct page
*page
= pc
->page
;
2253 struct cgroup
*cg
= child
->css
.cgroup
;
2254 struct cgroup
*pcg
= cg
->parent
;
2255 struct mem_cgroup
*parent
;
2263 if (!get_page_unless_zero(page
))
2265 if (isolate_lru_page(page
))
2268 parent
= mem_cgroup_from_cont(pcg
);
2269 ret
= __mem_cgroup_try_charge(NULL
, gfp_mask
, &parent
, false,
2274 ret
= mem_cgroup_move_account(pc
, child
, parent
, true);
2276 mem_cgroup_cancel_charge(parent
, PAGE_SIZE
);
2278 putback_lru_page(page
);
2286 * Charge the memory controller for page usage.
2288 * 0 if the charge was successful
2289 * < 0 if the cgroup is over its limit
2291 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
2292 gfp_t gfp_mask
, enum charge_type ctype
)
2294 struct mem_cgroup
*mem
= NULL
;
2295 struct page_cgroup
*pc
;
2297 int page_size
= PAGE_SIZE
;
2299 if (PageTransHuge(page
)) {
2300 page_size
<<= compound_order(page
);
2301 VM_BUG_ON(!PageTransHuge(page
));
2304 pc
= lookup_page_cgroup(page
);
2305 /* can happen at boot */
2310 ret
= __mem_cgroup_try_charge(mm
, gfp_mask
, &mem
, true, page_size
);
2314 __mem_cgroup_commit_charge(mem
, pc
, ctype
, page_size
);
2318 int mem_cgroup_newpage_charge(struct page
*page
,
2319 struct mm_struct
*mm
, gfp_t gfp_mask
)
2321 if (mem_cgroup_disabled())
2324 * If already mapped, we don't have to account.
2325 * If page cache, page->mapping has address_space.
2326 * But page->mapping may have out-of-use anon_vma pointer,
2327 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2330 if (page_mapped(page
) || (page
->mapping
&& !PageAnon(page
)))
2334 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2335 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2339 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2340 enum charge_type ctype
);
2342 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
2347 if (mem_cgroup_disabled())
2349 if (PageCompound(page
))
2352 * Corner case handling. This is called from add_to_page_cache()
2353 * in usual. But some FS (shmem) precharges this page before calling it
2354 * and call add_to_page_cache() with GFP_NOWAIT.
2356 * For GFP_NOWAIT case, the page may be pre-charged before calling
2357 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2358 * charge twice. (It works but has to pay a bit larger cost.)
2359 * And when the page is SwapCache, it should take swap information
2360 * into account. This is under lock_page() now.
2362 if (!(gfp_mask
& __GFP_WAIT
)) {
2363 struct page_cgroup
*pc
;
2365 pc
= lookup_page_cgroup(page
);
2368 lock_page_cgroup(pc
);
2369 if (PageCgroupUsed(pc
)) {
2370 unlock_page_cgroup(pc
);
2373 unlock_page_cgroup(pc
);
2379 if (page_is_file_cache(page
))
2380 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2381 MEM_CGROUP_CHARGE_TYPE_CACHE
);
2384 if (PageSwapCache(page
)) {
2385 struct mem_cgroup
*mem
= NULL
;
2387 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
2389 __mem_cgroup_commit_charge_swapin(page
, mem
,
2390 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2392 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
,
2393 MEM_CGROUP_CHARGE_TYPE_SHMEM
);
2399 * While swap-in, try_charge -> commit or cancel, the page is locked.
2400 * And when try_charge() successfully returns, one refcnt to memcg without
2401 * struct page_cgroup is acquired. This refcnt will be consumed by
2402 * "commit()" or removed by "cancel()"
2404 int mem_cgroup_try_charge_swapin(struct mm_struct
*mm
,
2406 gfp_t mask
, struct mem_cgroup
**ptr
)
2408 struct mem_cgroup
*mem
;
2411 if (mem_cgroup_disabled())
2414 if (!do_swap_account
)
2417 * A racing thread's fault, or swapoff, may have already updated
2418 * the pte, and even removed page from swap cache: in those cases
2419 * do_swap_page()'s pte_same() test will fail; but there's also a
2420 * KSM case which does need to charge the page.
2422 if (!PageSwapCache(page
))
2424 mem
= try_get_mem_cgroup_from_page(page
);
2428 ret
= __mem_cgroup_try_charge(NULL
, mask
, ptr
, true, PAGE_SIZE
);
2434 return __mem_cgroup_try_charge(mm
, mask
, ptr
, true, PAGE_SIZE
);
2438 __mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
,
2439 enum charge_type ctype
)
2441 struct page_cgroup
*pc
;
2443 if (mem_cgroup_disabled())
2447 cgroup_exclude_rmdir(&ptr
->css
);
2448 pc
= lookup_page_cgroup(page
);
2449 mem_cgroup_lru_del_before_commit_swapcache(page
);
2450 __mem_cgroup_commit_charge(ptr
, pc
, ctype
, PAGE_SIZE
);
2451 mem_cgroup_lru_add_after_commit_swapcache(page
);
2453 * Now swap is on-memory. This means this page may be
2454 * counted both as mem and swap....double count.
2455 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2456 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2457 * may call delete_from_swap_cache() before reach here.
2459 if (do_swap_account
&& PageSwapCache(page
)) {
2460 swp_entry_t ent
= {.val
= page_private(page
)};
2462 struct mem_cgroup
*memcg
;
2464 id
= swap_cgroup_record(ent
, 0);
2466 memcg
= mem_cgroup_lookup(id
);
2469 * This recorded memcg can be obsolete one. So, avoid
2470 * calling css_tryget
2472 if (!mem_cgroup_is_root(memcg
))
2473 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2474 mem_cgroup_swap_statistics(memcg
, false);
2475 mem_cgroup_put(memcg
);
2480 * At swapin, we may charge account against cgroup which has no tasks.
2481 * So, rmdir()->pre_destroy() can be called while we do this charge.
2482 * In that case, we need to call pre_destroy() again. check it here.
2484 cgroup_release_and_wakeup_rmdir(&ptr
->css
);
2487 void mem_cgroup_commit_charge_swapin(struct page
*page
, struct mem_cgroup
*ptr
)
2489 __mem_cgroup_commit_charge_swapin(page
, ptr
,
2490 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2493 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup
*mem
)
2495 if (mem_cgroup_disabled())
2499 mem_cgroup_cancel_charge(mem
, PAGE_SIZE
);
2503 __do_uncharge(struct mem_cgroup
*mem
, const enum charge_type ctype
,
2506 struct memcg_batch_info
*batch
= NULL
;
2507 bool uncharge_memsw
= true;
2508 /* If swapout, usage of swap doesn't decrease */
2509 if (!do_swap_account
|| ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
)
2510 uncharge_memsw
= false;
2512 batch
= ¤t
->memcg_batch
;
2514 * In usual, we do css_get() when we remember memcg pointer.
2515 * But in this case, we keep res->usage until end of a series of
2516 * uncharges. Then, it's ok to ignore memcg's refcnt.
2521 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2522 * In those cases, all pages freed continously can be expected to be in
2523 * the same cgroup and we have chance to coalesce uncharges.
2524 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2525 * because we want to do uncharge as soon as possible.
2528 if (!batch
->do_batch
|| test_thread_flag(TIF_MEMDIE
))
2529 goto direct_uncharge
;
2531 if (page_size
!= PAGE_SIZE
)
2532 goto direct_uncharge
;
2535 * In typical case, batch->memcg == mem. This means we can
2536 * merge a series of uncharges to an uncharge of res_counter.
2537 * If not, we uncharge res_counter ony by one.
2539 if (batch
->memcg
!= mem
)
2540 goto direct_uncharge
;
2541 /* remember freed charge and uncharge it later */
2542 batch
->bytes
+= PAGE_SIZE
;
2544 batch
->memsw_bytes
+= PAGE_SIZE
;
2547 res_counter_uncharge(&mem
->res
, page_size
);
2549 res_counter_uncharge(&mem
->memsw
, page_size
);
2550 if (unlikely(batch
->memcg
!= mem
))
2551 memcg_oom_recover(mem
);
2556 * uncharge if !page_mapped(page)
2558 static struct mem_cgroup
*
2559 __mem_cgroup_uncharge_common(struct page
*page
, enum charge_type ctype
)
2562 struct page_cgroup
*pc
;
2563 struct mem_cgroup
*mem
= NULL
;
2564 int page_size
= PAGE_SIZE
;
2566 if (mem_cgroup_disabled())
2569 if (PageSwapCache(page
))
2572 if (PageTransHuge(page
)) {
2573 page_size
<<= compound_order(page
);
2574 VM_BUG_ON(!PageTransHuge(page
));
2577 count
= page_size
>> PAGE_SHIFT
;
2579 * Check if our page_cgroup is valid
2581 pc
= lookup_page_cgroup(page
);
2582 if (unlikely(!pc
|| !PageCgroupUsed(pc
)))
2585 lock_page_cgroup(pc
);
2587 mem
= pc
->mem_cgroup
;
2589 if (!PageCgroupUsed(pc
))
2593 case MEM_CGROUP_CHARGE_TYPE_MAPPED
:
2594 case MEM_CGROUP_CHARGE_TYPE_DROP
:
2595 /* See mem_cgroup_prepare_migration() */
2596 if (page_mapped(page
) || PageCgroupMigration(pc
))
2599 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT
:
2600 if (!PageAnon(page
)) { /* Shared memory */
2601 if (page
->mapping
&& !page_is_file_cache(page
))
2603 } else if (page_mapped(page
)) /* Anon */
2610 mem_cgroup_charge_statistics(mem
, PageCgroupCache(pc
), -count
);
2612 ClearPageCgroupUsed(pc
);
2614 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2615 * freed from LRU. This is safe because uncharged page is expected not
2616 * to be reused (freed soon). Exception is SwapCache, it's handled by
2617 * special functions.
2620 unlock_page_cgroup(pc
);
2622 * even after unlock, we have mem->res.usage here and this memcg
2623 * will never be freed.
2625 memcg_check_events(mem
, page
);
2626 if (do_swap_account
&& ctype
== MEM_CGROUP_CHARGE_TYPE_SWAPOUT
) {
2627 mem_cgroup_swap_statistics(mem
, true);
2628 mem_cgroup_get(mem
);
2630 if (!mem_cgroup_is_root(mem
))
2631 __do_uncharge(mem
, ctype
, page_size
);
2636 unlock_page_cgroup(pc
);
2640 void mem_cgroup_uncharge_page(struct page
*page
)
2643 if (page_mapped(page
))
2645 if (page
->mapping
&& !PageAnon(page
))
2647 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_MAPPED
);
2650 void mem_cgroup_uncharge_cache_page(struct page
*page
)
2652 VM_BUG_ON(page_mapped(page
));
2653 VM_BUG_ON(page
->mapping
);
2654 __mem_cgroup_uncharge_common(page
, MEM_CGROUP_CHARGE_TYPE_CACHE
);
2658 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2659 * In that cases, pages are freed continuously and we can expect pages
2660 * are in the same memcg. All these calls itself limits the number of
2661 * pages freed at once, then uncharge_start/end() is called properly.
2662 * This may be called prural(2) times in a context,
2665 void mem_cgroup_uncharge_start(void)
2667 current
->memcg_batch
.do_batch
++;
2668 /* We can do nest. */
2669 if (current
->memcg_batch
.do_batch
== 1) {
2670 current
->memcg_batch
.memcg
= NULL
;
2671 current
->memcg_batch
.bytes
= 0;
2672 current
->memcg_batch
.memsw_bytes
= 0;
2676 void mem_cgroup_uncharge_end(void)
2678 struct memcg_batch_info
*batch
= ¤t
->memcg_batch
;
2680 if (!batch
->do_batch
)
2684 if (batch
->do_batch
) /* If stacked, do nothing. */
2690 * This "batch->memcg" is valid without any css_get/put etc...
2691 * bacause we hide charges behind us.
2694 res_counter_uncharge(&batch
->memcg
->res
, batch
->bytes
);
2695 if (batch
->memsw_bytes
)
2696 res_counter_uncharge(&batch
->memcg
->memsw
, batch
->memsw_bytes
);
2697 memcg_oom_recover(batch
->memcg
);
2698 /* forget this pointer (for sanity check) */
2699 batch
->memcg
= NULL
;
2704 * called after __delete_from_swap_cache() and drop "page" account.
2705 * memcg information is recorded to swap_cgroup of "ent"
2708 mem_cgroup_uncharge_swapcache(struct page
*page
, swp_entry_t ent
, bool swapout
)
2710 struct mem_cgroup
*memcg
;
2711 int ctype
= MEM_CGROUP_CHARGE_TYPE_SWAPOUT
;
2713 if (!swapout
) /* this was a swap cache but the swap is unused ! */
2714 ctype
= MEM_CGROUP_CHARGE_TYPE_DROP
;
2716 memcg
= __mem_cgroup_uncharge_common(page
, ctype
);
2719 * record memcg information, if swapout && memcg != NULL,
2720 * mem_cgroup_get() was called in uncharge().
2722 if (do_swap_account
&& swapout
&& memcg
)
2723 swap_cgroup_record(ent
, css_id(&memcg
->css
));
2727 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2729 * called from swap_entry_free(). remove record in swap_cgroup and
2730 * uncharge "memsw" account.
2732 void mem_cgroup_uncharge_swap(swp_entry_t ent
)
2734 struct mem_cgroup
*memcg
;
2737 if (!do_swap_account
)
2740 id
= swap_cgroup_record(ent
, 0);
2742 memcg
= mem_cgroup_lookup(id
);
2745 * We uncharge this because swap is freed.
2746 * This memcg can be obsolete one. We avoid calling css_tryget
2748 if (!mem_cgroup_is_root(memcg
))
2749 res_counter_uncharge(&memcg
->memsw
, PAGE_SIZE
);
2750 mem_cgroup_swap_statistics(memcg
, false);
2751 mem_cgroup_put(memcg
);
2757 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2758 * @entry: swap entry to be moved
2759 * @from: mem_cgroup which the entry is moved from
2760 * @to: mem_cgroup which the entry is moved to
2761 * @need_fixup: whether we should fixup res_counters and refcounts.
2763 * It succeeds only when the swap_cgroup's record for this entry is the same
2764 * as the mem_cgroup's id of @from.
2766 * Returns 0 on success, -EINVAL on failure.
2768 * The caller must have charged to @to, IOW, called res_counter_charge() about
2769 * both res and memsw, and called css_get().
2771 static int mem_cgroup_move_swap_account(swp_entry_t entry
,
2772 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2774 unsigned short old_id
, new_id
;
2776 old_id
= css_id(&from
->css
);
2777 new_id
= css_id(&to
->css
);
2779 if (swap_cgroup_cmpxchg(entry
, old_id
, new_id
) == old_id
) {
2780 mem_cgroup_swap_statistics(from
, false);
2781 mem_cgroup_swap_statistics(to
, true);
2783 * This function is only called from task migration context now.
2784 * It postpones res_counter and refcount handling till the end
2785 * of task migration(mem_cgroup_clear_mc()) for performance
2786 * improvement. But we cannot postpone mem_cgroup_get(to)
2787 * because if the process that has been moved to @to does
2788 * swap-in, the refcount of @to might be decreased to 0.
2792 if (!mem_cgroup_is_root(from
))
2793 res_counter_uncharge(&from
->memsw
, PAGE_SIZE
);
2794 mem_cgroup_put(from
);
2796 * we charged both to->res and to->memsw, so we should
2799 if (!mem_cgroup_is_root(to
))
2800 res_counter_uncharge(&to
->res
, PAGE_SIZE
);
2807 static inline int mem_cgroup_move_swap_account(swp_entry_t entry
,
2808 struct mem_cgroup
*from
, struct mem_cgroup
*to
, bool need_fixup
)
2815 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2818 int mem_cgroup_prepare_migration(struct page
*page
,
2819 struct page
*newpage
, struct mem_cgroup
**ptr
)
2821 struct page_cgroup
*pc
;
2822 struct mem_cgroup
*mem
= NULL
;
2823 enum charge_type ctype
;
2826 VM_BUG_ON(PageTransHuge(page
));
2827 if (mem_cgroup_disabled())
2830 pc
= lookup_page_cgroup(page
);
2831 lock_page_cgroup(pc
);
2832 if (PageCgroupUsed(pc
)) {
2833 mem
= pc
->mem_cgroup
;
2836 * At migrating an anonymous page, its mapcount goes down
2837 * to 0 and uncharge() will be called. But, even if it's fully
2838 * unmapped, migration may fail and this page has to be
2839 * charged again. We set MIGRATION flag here and delay uncharge
2840 * until end_migration() is called
2842 * Corner Case Thinking
2844 * When the old page was mapped as Anon and it's unmap-and-freed
2845 * while migration was ongoing.
2846 * If unmap finds the old page, uncharge() of it will be delayed
2847 * until end_migration(). If unmap finds a new page, it's
2848 * uncharged when it make mapcount to be 1->0. If unmap code
2849 * finds swap_migration_entry, the new page will not be mapped
2850 * and end_migration() will find it(mapcount==0).
2853 * When the old page was mapped but migraion fails, the kernel
2854 * remaps it. A charge for it is kept by MIGRATION flag even
2855 * if mapcount goes down to 0. We can do remap successfully
2856 * without charging it again.
2859 * The "old" page is under lock_page() until the end of
2860 * migration, so, the old page itself will not be swapped-out.
2861 * If the new page is swapped out before end_migraton, our
2862 * hook to usual swap-out path will catch the event.
2865 SetPageCgroupMigration(pc
);
2867 unlock_page_cgroup(pc
);
2869 * If the page is not charged at this point,
2876 ret
= __mem_cgroup_try_charge(NULL
, GFP_KERNEL
, ptr
, false, PAGE_SIZE
);
2877 css_put(&mem
->css
);/* drop extra refcnt */
2878 if (ret
|| *ptr
== NULL
) {
2879 if (PageAnon(page
)) {
2880 lock_page_cgroup(pc
);
2881 ClearPageCgroupMigration(pc
);
2882 unlock_page_cgroup(pc
);
2884 * The old page may be fully unmapped while we kept it.
2886 mem_cgroup_uncharge_page(page
);
2891 * We charge new page before it's used/mapped. So, even if unlock_page()
2892 * is called before end_migration, we can catch all events on this new
2893 * page. In the case new page is migrated but not remapped, new page's
2894 * mapcount will be finally 0 and we call uncharge in end_migration().
2896 pc
= lookup_page_cgroup(newpage
);
2898 ctype
= MEM_CGROUP_CHARGE_TYPE_MAPPED
;
2899 else if (page_is_file_cache(page
))
2900 ctype
= MEM_CGROUP_CHARGE_TYPE_CACHE
;
2902 ctype
= MEM_CGROUP_CHARGE_TYPE_SHMEM
;
2903 __mem_cgroup_commit_charge(mem
, pc
, ctype
, PAGE_SIZE
);
2907 /* remove redundant charge if migration failed*/
2908 void mem_cgroup_end_migration(struct mem_cgroup
*mem
,
2909 struct page
*oldpage
, struct page
*newpage
, bool migration_ok
)
2911 struct page
*used
, *unused
;
2912 struct page_cgroup
*pc
;
2916 /* blocks rmdir() */
2917 cgroup_exclude_rmdir(&mem
->css
);
2918 if (!migration_ok
) {
2926 * We disallowed uncharge of pages under migration because mapcount
2927 * of the page goes down to zero, temporarly.
2928 * Clear the flag and check the page should be charged.
2930 pc
= lookup_page_cgroup(oldpage
);
2931 lock_page_cgroup(pc
);
2932 ClearPageCgroupMigration(pc
);
2933 unlock_page_cgroup(pc
);
2935 __mem_cgroup_uncharge_common(unused
, MEM_CGROUP_CHARGE_TYPE_FORCE
);
2938 * If a page is a file cache, radix-tree replacement is very atomic
2939 * and we can skip this check. When it was an Anon page, its mapcount
2940 * goes down to 0. But because we added MIGRATION flage, it's not
2941 * uncharged yet. There are several case but page->mapcount check
2942 * and USED bit check in mem_cgroup_uncharge_page() will do enough
2943 * check. (see prepare_charge() also)
2946 mem_cgroup_uncharge_page(used
);
2948 * At migration, we may charge account against cgroup which has no
2950 * So, rmdir()->pre_destroy() can be called while we do this charge.
2951 * In that case, we need to call pre_destroy() again. check it here.
2953 cgroup_release_and_wakeup_rmdir(&mem
->css
);
2957 * A call to try to shrink memory usage on charge failure at shmem's swapin.
2958 * Calling hierarchical_reclaim is not enough because we should update
2959 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2960 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2961 * not from the memcg which this page would be charged to.
2962 * try_charge_swapin does all of these works properly.
2964 int mem_cgroup_shmem_charge_fallback(struct page
*page
,
2965 struct mm_struct
*mm
,
2968 struct mem_cgroup
*mem
= NULL
;
2971 if (mem_cgroup_disabled())
2974 ret
= mem_cgroup_try_charge_swapin(mm
, page
, gfp_mask
, &mem
);
2976 mem_cgroup_cancel_charge_swapin(mem
); /* it does !mem check */
2981 static DEFINE_MUTEX(set_limit_mutex
);
2983 static int mem_cgroup_resize_limit(struct mem_cgroup
*memcg
,
2984 unsigned long long val
)
2987 u64 memswlimit
, memlimit
;
2989 int children
= mem_cgroup_count_children(memcg
);
2990 u64 curusage
, oldusage
;
2994 * For keeping hierarchical_reclaim simple, how long we should retry
2995 * is depends on callers. We set our retry-count to be function
2996 * of # of children which we should visit in this loop.
2998 retry_count
= MEM_CGROUP_RECLAIM_RETRIES
* children
;
3000 oldusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3003 while (retry_count
) {
3004 if (signal_pending(current
)) {
3009 * Rather than hide all in some function, I do this in
3010 * open coded manner. You see what this really does.
3011 * We have to guarantee mem->res.limit < mem->memsw.limit.
3013 mutex_lock(&set_limit_mutex
);
3014 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3015 if (memswlimit
< val
) {
3017 mutex_unlock(&set_limit_mutex
);
3021 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3025 ret
= res_counter_set_limit(&memcg
->res
, val
);
3027 if (memswlimit
== val
)
3028 memcg
->memsw_is_minimum
= true;
3030 memcg
->memsw_is_minimum
= false;
3032 mutex_unlock(&set_limit_mutex
);
3037 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3038 MEM_CGROUP_RECLAIM_SHRINK
);
3039 curusage
= res_counter_read_u64(&memcg
->res
, RES_USAGE
);
3040 /* Usage is reduced ? */
3041 if (curusage
>= oldusage
)
3044 oldusage
= curusage
;
3046 if (!ret
&& enlarge
)
3047 memcg_oom_recover(memcg
);
3052 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup
*memcg
,
3053 unsigned long long val
)
3056 u64 memlimit
, memswlimit
, oldusage
, curusage
;
3057 int children
= mem_cgroup_count_children(memcg
);
3061 /* see mem_cgroup_resize_res_limit */
3062 retry_count
= children
* MEM_CGROUP_RECLAIM_RETRIES
;
3063 oldusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3064 while (retry_count
) {
3065 if (signal_pending(current
)) {
3070 * Rather than hide all in some function, I do this in
3071 * open coded manner. You see what this really does.
3072 * We have to guarantee mem->res.limit < mem->memsw.limit.
3074 mutex_lock(&set_limit_mutex
);
3075 memlimit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3076 if (memlimit
> val
) {
3078 mutex_unlock(&set_limit_mutex
);
3081 memswlimit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3082 if (memswlimit
< val
)
3084 ret
= res_counter_set_limit(&memcg
->memsw
, val
);
3086 if (memlimit
== val
)
3087 memcg
->memsw_is_minimum
= true;
3089 memcg
->memsw_is_minimum
= false;
3091 mutex_unlock(&set_limit_mutex
);
3096 mem_cgroup_hierarchical_reclaim(memcg
, NULL
, GFP_KERNEL
,
3097 MEM_CGROUP_RECLAIM_NOSWAP
|
3098 MEM_CGROUP_RECLAIM_SHRINK
);
3099 curusage
= res_counter_read_u64(&memcg
->memsw
, RES_USAGE
);
3100 /* Usage is reduced ? */
3101 if (curusage
>= oldusage
)
3104 oldusage
= curusage
;
3106 if (!ret
&& enlarge
)
3107 memcg_oom_recover(memcg
);
3111 unsigned long mem_cgroup_soft_limit_reclaim(struct zone
*zone
, int order
,
3114 unsigned long nr_reclaimed
= 0;
3115 struct mem_cgroup_per_zone
*mz
, *next_mz
= NULL
;
3116 unsigned long reclaimed
;
3118 struct mem_cgroup_tree_per_zone
*mctz
;
3119 unsigned long long excess
;
3124 mctz
= soft_limit_tree_node_zone(zone_to_nid(zone
), zone_idx(zone
));
3126 * This loop can run a while, specially if mem_cgroup's continuously
3127 * keep exceeding their soft limit and putting the system under
3134 mz
= mem_cgroup_largest_soft_limit_node(mctz
);
3138 reclaimed
= mem_cgroup_hierarchical_reclaim(mz
->mem
, zone
,
3140 MEM_CGROUP_RECLAIM_SOFT
);
3141 nr_reclaimed
+= reclaimed
;
3142 spin_lock(&mctz
->lock
);
3145 * If we failed to reclaim anything from this memory cgroup
3146 * it is time to move on to the next cgroup
3152 * Loop until we find yet another one.
3154 * By the time we get the soft_limit lock
3155 * again, someone might have aded the
3156 * group back on the RB tree. Iterate to
3157 * make sure we get a different mem.
3158 * mem_cgroup_largest_soft_limit_node returns
3159 * NULL if no other cgroup is present on
3163 __mem_cgroup_largest_soft_limit_node(mctz
);
3164 if (next_mz
== mz
) {
3165 css_put(&next_mz
->mem
->css
);
3167 } else /* next_mz == NULL or other memcg */
3171 __mem_cgroup_remove_exceeded(mz
->mem
, mz
, mctz
);
3172 excess
= res_counter_soft_limit_excess(&mz
->mem
->res
);
3174 * One school of thought says that we should not add
3175 * back the node to the tree if reclaim returns 0.
3176 * But our reclaim could return 0, simply because due
3177 * to priority we are exposing a smaller subset of
3178 * memory to reclaim from. Consider this as a longer
3181 /* If excess == 0, no tree ops */
3182 __mem_cgroup_insert_exceeded(mz
->mem
, mz
, mctz
, excess
);
3183 spin_unlock(&mctz
->lock
);
3184 css_put(&mz
->mem
->css
);
3187 * Could not reclaim anything and there are no more
3188 * mem cgroups to try or we seem to be looping without
3189 * reclaiming anything.
3191 if (!nr_reclaimed
&&
3193 loop
> MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS
))
3195 } while (!nr_reclaimed
);
3197 css_put(&next_mz
->mem
->css
);
3198 return nr_reclaimed
;
3202 * This routine traverse page_cgroup in given list and drop them all.
3203 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3205 static int mem_cgroup_force_empty_list(struct mem_cgroup
*mem
,
3206 int node
, int zid
, enum lru_list lru
)
3209 struct mem_cgroup_per_zone
*mz
;
3210 struct page_cgroup
*pc
, *busy
;
3211 unsigned long flags
, loop
;
3212 struct list_head
*list
;
3215 zone
= &NODE_DATA(node
)->node_zones
[zid
];
3216 mz
= mem_cgroup_zoneinfo(mem
, node
, zid
);
3217 list
= &mz
->lists
[lru
];
3219 loop
= MEM_CGROUP_ZSTAT(mz
, lru
);
3220 /* give some margin against EBUSY etc...*/
3225 spin_lock_irqsave(&zone
->lru_lock
, flags
);
3226 if (list_empty(list
)) {
3227 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3230 pc
= list_entry(list
->prev
, struct page_cgroup
, lru
);
3232 list_move(&pc
->lru
, list
);
3234 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3237 spin_unlock_irqrestore(&zone
->lru_lock
, flags
);
3239 ret
= mem_cgroup_move_parent(pc
, mem
, GFP_KERNEL
);
3243 if (ret
== -EBUSY
|| ret
== -EINVAL
) {
3244 /* found lock contention or "pc" is obsolete. */
3251 if (!ret
&& !list_empty(list
))
3257 * make mem_cgroup's charge to be 0 if there is no task.
3258 * This enables deleting this mem_cgroup.
3260 static int mem_cgroup_force_empty(struct mem_cgroup
*mem
, bool free_all
)
3263 int node
, zid
, shrink
;
3264 int nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
3265 struct cgroup
*cgrp
= mem
->css
.cgroup
;
3270 /* should free all ? */
3276 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
))
3279 if (signal_pending(current
))
3281 /* This is for making all *used* pages to be on LRU. */
3282 lru_add_drain_all();
3283 drain_all_stock_sync();
3285 mem_cgroup_start_move(mem
);
3286 for_each_node_state(node
, N_HIGH_MEMORY
) {
3287 for (zid
= 0; !ret
&& zid
< MAX_NR_ZONES
; zid
++) {
3290 ret
= mem_cgroup_force_empty_list(mem
,
3299 mem_cgroup_end_move(mem
);
3300 memcg_oom_recover(mem
);
3301 /* it seems parent cgroup doesn't have enough mem */
3305 /* "ret" should also be checked to ensure all lists are empty. */
3306 } while (mem
->res
.usage
> 0 || ret
);
3312 /* returns EBUSY if there is a task or if we come here twice. */
3313 if (cgroup_task_count(cgrp
) || !list_empty(&cgrp
->children
) || shrink
) {
3317 /* we call try-to-free pages for make this cgroup empty */
3318 lru_add_drain_all();
3319 /* try to free all pages in this cgroup */
3321 while (nr_retries
&& mem
->res
.usage
> 0) {
3324 if (signal_pending(current
)) {
3328 progress
= try_to_free_mem_cgroup_pages(mem
, GFP_KERNEL
,
3329 false, get_swappiness(mem
));
3332 /* maybe some writeback is necessary */
3333 congestion_wait(BLK_RW_ASYNC
, HZ
/10);
3338 /* try move_account...there may be some *locked* pages. */
3342 int mem_cgroup_force_empty_write(struct cgroup
*cont
, unsigned int event
)
3344 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont
), true);
3348 static u64
mem_cgroup_hierarchy_read(struct cgroup
*cont
, struct cftype
*cft
)
3350 return mem_cgroup_from_cont(cont
)->use_hierarchy
;
3353 static int mem_cgroup_hierarchy_write(struct cgroup
*cont
, struct cftype
*cft
,
3357 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3358 struct cgroup
*parent
= cont
->parent
;
3359 struct mem_cgroup
*parent_mem
= NULL
;
3362 parent_mem
= mem_cgroup_from_cont(parent
);
3366 * If parent's use_hierarchy is set, we can't make any modifications
3367 * in the child subtrees. If it is unset, then the change can
3368 * occur, provided the current cgroup has no children.
3370 * For the root cgroup, parent_mem is NULL, we allow value to be
3371 * set if there are no children.
3373 if ((!parent_mem
|| !parent_mem
->use_hierarchy
) &&
3374 (val
== 1 || val
== 0)) {
3375 if (list_empty(&cont
->children
))
3376 mem
->use_hierarchy
= val
;
3387 static u64
mem_cgroup_get_recursive_idx_stat(struct mem_cgroup
*mem
,
3388 enum mem_cgroup_stat_index idx
)
3390 struct mem_cgroup
*iter
;
3393 /* each per cpu's value can be minus.Then, use s64 */
3394 for_each_mem_cgroup_tree(iter
, mem
)
3395 val
+= mem_cgroup_read_stat(iter
, idx
);
3397 if (val
< 0) /* race ? */
3402 static inline u64
mem_cgroup_usage(struct mem_cgroup
*mem
, bool swap
)
3406 if (!mem_cgroup_is_root(mem
)) {
3408 return res_counter_read_u64(&mem
->res
, RES_USAGE
);
3410 return res_counter_read_u64(&mem
->memsw
, RES_USAGE
);
3413 val
= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3414 val
+= mem_cgroup_get_recursive_idx_stat(mem
, MEM_CGROUP_STAT_RSS
);
3417 val
+= mem_cgroup_get_recursive_idx_stat(mem
,
3418 MEM_CGROUP_STAT_SWAPOUT
);
3420 return val
<< PAGE_SHIFT
;
3423 static u64
mem_cgroup_read(struct cgroup
*cont
, struct cftype
*cft
)
3425 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
3429 type
= MEMFILE_TYPE(cft
->private);
3430 name
= MEMFILE_ATTR(cft
->private);
3433 if (name
== RES_USAGE
)
3434 val
= mem_cgroup_usage(mem
, false);
3436 val
= res_counter_read_u64(&mem
->res
, name
);
3439 if (name
== RES_USAGE
)
3440 val
= mem_cgroup_usage(mem
, true);
3442 val
= res_counter_read_u64(&mem
->memsw
, name
);
3451 * The user of this function is...
3454 static int mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
3457 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cont
);
3459 unsigned long long val
;
3462 type
= MEMFILE_TYPE(cft
->private);
3463 name
= MEMFILE_ATTR(cft
->private);
3466 if (mem_cgroup_is_root(memcg
)) { /* Can't set limit on root */
3470 /* This function does all necessary parse...reuse it */
3471 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3475 ret
= mem_cgroup_resize_limit(memcg
, val
);
3477 ret
= mem_cgroup_resize_memsw_limit(memcg
, val
);
3479 case RES_SOFT_LIMIT
:
3480 ret
= res_counter_memparse_write_strategy(buffer
, &val
);
3484 * For memsw, soft limits are hard to implement in terms
3485 * of semantics, for now, we support soft limits for
3486 * control without swap
3489 ret
= res_counter_set_soft_limit(&memcg
->res
, val
);
3494 ret
= -EINVAL
; /* should be BUG() ? */
3500 static void memcg_get_hierarchical_limit(struct mem_cgroup
*memcg
,
3501 unsigned long long *mem_limit
, unsigned long long *memsw_limit
)
3503 struct cgroup
*cgroup
;
3504 unsigned long long min_limit
, min_memsw_limit
, tmp
;
3506 min_limit
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3507 min_memsw_limit
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3508 cgroup
= memcg
->css
.cgroup
;
3509 if (!memcg
->use_hierarchy
)
3512 while (cgroup
->parent
) {
3513 cgroup
= cgroup
->parent
;
3514 memcg
= mem_cgroup_from_cont(cgroup
);
3515 if (!memcg
->use_hierarchy
)
3517 tmp
= res_counter_read_u64(&memcg
->res
, RES_LIMIT
);
3518 min_limit
= min(min_limit
, tmp
);
3519 tmp
= res_counter_read_u64(&memcg
->memsw
, RES_LIMIT
);
3520 min_memsw_limit
= min(min_memsw_limit
, tmp
);
3523 *mem_limit
= min_limit
;
3524 *memsw_limit
= min_memsw_limit
;
3528 static int mem_cgroup_reset(struct cgroup
*cont
, unsigned int event
)
3530 struct mem_cgroup
*mem
;
3533 mem
= mem_cgroup_from_cont(cont
);
3534 type
= MEMFILE_TYPE(event
);
3535 name
= MEMFILE_ATTR(event
);
3539 res_counter_reset_max(&mem
->res
);
3541 res_counter_reset_max(&mem
->memsw
);
3545 res_counter_reset_failcnt(&mem
->res
);
3547 res_counter_reset_failcnt(&mem
->memsw
);
3554 static u64
mem_cgroup_move_charge_read(struct cgroup
*cgrp
,
3557 return mem_cgroup_from_cont(cgrp
)->move_charge_at_immigrate
;
3561 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3562 struct cftype
*cft
, u64 val
)
3564 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
3566 if (val
>= (1 << NR_MOVE_TYPE
))
3569 * We check this value several times in both in can_attach() and
3570 * attach(), so we need cgroup lock to prevent this value from being
3574 mem
->move_charge_at_immigrate
= val
;
3580 static int mem_cgroup_move_charge_write(struct cgroup
*cgrp
,
3581 struct cftype
*cft
, u64 val
)
3588 /* For read statistics */
3604 struct mcs_total_stat
{
3605 s64 stat
[NR_MCS_STAT
];
3611 } memcg_stat_strings
[NR_MCS_STAT
] = {
3612 {"cache", "total_cache"},
3613 {"rss", "total_rss"},
3614 {"mapped_file", "total_mapped_file"},
3615 {"pgpgin", "total_pgpgin"},
3616 {"pgpgout", "total_pgpgout"},
3617 {"swap", "total_swap"},
3618 {"inactive_anon", "total_inactive_anon"},
3619 {"active_anon", "total_active_anon"},
3620 {"inactive_file", "total_inactive_file"},
3621 {"active_file", "total_active_file"},
3622 {"unevictable", "total_unevictable"}
3627 mem_cgroup_get_local_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3632 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_CACHE
);
3633 s
->stat
[MCS_CACHE
] += val
* PAGE_SIZE
;
3634 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_RSS
);
3635 s
->stat
[MCS_RSS
] += val
* PAGE_SIZE
;
3636 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_FILE_MAPPED
);
3637 s
->stat
[MCS_FILE_MAPPED
] += val
* PAGE_SIZE
;
3638 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGIN_COUNT
);
3639 s
->stat
[MCS_PGPGIN
] += val
;
3640 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_PGPGOUT_COUNT
);
3641 s
->stat
[MCS_PGPGOUT
] += val
;
3642 if (do_swap_account
) {
3643 val
= mem_cgroup_read_stat(mem
, MEM_CGROUP_STAT_SWAPOUT
);
3644 s
->stat
[MCS_SWAP
] += val
* PAGE_SIZE
;
3648 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_ANON
);
3649 s
->stat
[MCS_INACTIVE_ANON
] += val
* PAGE_SIZE
;
3650 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_ANON
);
3651 s
->stat
[MCS_ACTIVE_ANON
] += val
* PAGE_SIZE
;
3652 val
= mem_cgroup_get_local_zonestat(mem
, LRU_INACTIVE_FILE
);
3653 s
->stat
[MCS_INACTIVE_FILE
] += val
* PAGE_SIZE
;
3654 val
= mem_cgroup_get_local_zonestat(mem
, LRU_ACTIVE_FILE
);
3655 s
->stat
[MCS_ACTIVE_FILE
] += val
* PAGE_SIZE
;
3656 val
= mem_cgroup_get_local_zonestat(mem
, LRU_UNEVICTABLE
);
3657 s
->stat
[MCS_UNEVICTABLE
] += val
* PAGE_SIZE
;
3661 mem_cgroup_get_total_stat(struct mem_cgroup
*mem
, struct mcs_total_stat
*s
)
3663 struct mem_cgroup
*iter
;
3665 for_each_mem_cgroup_tree(iter
, mem
)
3666 mem_cgroup_get_local_stat(iter
, s
);
3669 static int mem_control_stat_show(struct cgroup
*cont
, struct cftype
*cft
,
3670 struct cgroup_map_cb
*cb
)
3672 struct mem_cgroup
*mem_cont
= mem_cgroup_from_cont(cont
);
3673 struct mcs_total_stat mystat
;
3676 memset(&mystat
, 0, sizeof(mystat
));
3677 mem_cgroup_get_local_stat(mem_cont
, &mystat
);
3679 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3680 if (i
== MCS_SWAP
&& !do_swap_account
)
3682 cb
->fill(cb
, memcg_stat_strings
[i
].local_name
, mystat
.stat
[i
]);
3685 /* Hierarchical information */
3687 unsigned long long limit
, memsw_limit
;
3688 memcg_get_hierarchical_limit(mem_cont
, &limit
, &memsw_limit
);
3689 cb
->fill(cb
, "hierarchical_memory_limit", limit
);
3690 if (do_swap_account
)
3691 cb
->fill(cb
, "hierarchical_memsw_limit", memsw_limit
);
3694 memset(&mystat
, 0, sizeof(mystat
));
3695 mem_cgroup_get_total_stat(mem_cont
, &mystat
);
3696 for (i
= 0; i
< NR_MCS_STAT
; i
++) {
3697 if (i
== MCS_SWAP
&& !do_swap_account
)
3699 cb
->fill(cb
, memcg_stat_strings
[i
].total_name
, mystat
.stat
[i
]);
3702 #ifdef CONFIG_DEBUG_VM
3703 cb
->fill(cb
, "inactive_ratio", calc_inactive_ratio(mem_cont
, NULL
));
3707 struct mem_cgroup_per_zone
*mz
;
3708 unsigned long recent_rotated
[2] = {0, 0};
3709 unsigned long recent_scanned
[2] = {0, 0};
3711 for_each_online_node(nid
)
3712 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
3713 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
3715 recent_rotated
[0] +=
3716 mz
->reclaim_stat
.recent_rotated
[0];
3717 recent_rotated
[1] +=
3718 mz
->reclaim_stat
.recent_rotated
[1];
3719 recent_scanned
[0] +=
3720 mz
->reclaim_stat
.recent_scanned
[0];
3721 recent_scanned
[1] +=
3722 mz
->reclaim_stat
.recent_scanned
[1];
3724 cb
->fill(cb
, "recent_rotated_anon", recent_rotated
[0]);
3725 cb
->fill(cb
, "recent_rotated_file", recent_rotated
[1]);
3726 cb
->fill(cb
, "recent_scanned_anon", recent_scanned
[0]);
3727 cb
->fill(cb
, "recent_scanned_file", recent_scanned
[1]);
3734 static u64
mem_cgroup_swappiness_read(struct cgroup
*cgrp
, struct cftype
*cft
)
3736 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3738 return get_swappiness(memcg
);
3741 static int mem_cgroup_swappiness_write(struct cgroup
*cgrp
, struct cftype
*cft
,
3744 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3745 struct mem_cgroup
*parent
;
3750 if (cgrp
->parent
== NULL
)
3753 parent
= mem_cgroup_from_cont(cgrp
->parent
);
3757 /* If under hierarchy, only empty-root can set this value */
3758 if ((parent
->use_hierarchy
) ||
3759 (memcg
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
3764 spin_lock(&memcg
->reclaim_param_lock
);
3765 memcg
->swappiness
= val
;
3766 spin_unlock(&memcg
->reclaim_param_lock
);
3773 static void __mem_cgroup_threshold(struct mem_cgroup
*memcg
, bool swap
)
3775 struct mem_cgroup_threshold_ary
*t
;
3781 t
= rcu_dereference(memcg
->thresholds
.primary
);
3783 t
= rcu_dereference(memcg
->memsw_thresholds
.primary
);
3788 usage
= mem_cgroup_usage(memcg
, swap
);
3791 * current_threshold points to threshold just below usage.
3792 * If it's not true, a threshold was crossed after last
3793 * call of __mem_cgroup_threshold().
3795 i
= t
->current_threshold
;
3798 * Iterate backward over array of thresholds starting from
3799 * current_threshold and check if a threshold is crossed.
3800 * If none of thresholds below usage is crossed, we read
3801 * only one element of the array here.
3803 for (; i
>= 0 && unlikely(t
->entries
[i
].threshold
> usage
); i
--)
3804 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3806 /* i = current_threshold + 1 */
3810 * Iterate forward over array of thresholds starting from
3811 * current_threshold+1 and check if a threshold is crossed.
3812 * If none of thresholds above usage is crossed, we read
3813 * only one element of the array here.
3815 for (; i
< t
->size
&& unlikely(t
->entries
[i
].threshold
<= usage
); i
++)
3816 eventfd_signal(t
->entries
[i
].eventfd
, 1);
3818 /* Update current_threshold */
3819 t
->current_threshold
= i
- 1;
3824 static void mem_cgroup_threshold(struct mem_cgroup
*memcg
)
3827 __mem_cgroup_threshold(memcg
, false);
3828 if (do_swap_account
)
3829 __mem_cgroup_threshold(memcg
, true);
3831 memcg
= parent_mem_cgroup(memcg
);
3835 static int compare_thresholds(const void *a
, const void *b
)
3837 const struct mem_cgroup_threshold
*_a
= a
;
3838 const struct mem_cgroup_threshold
*_b
= b
;
3840 return _a
->threshold
- _b
->threshold
;
3843 static int mem_cgroup_oom_notify_cb(struct mem_cgroup
*mem
)
3845 struct mem_cgroup_eventfd_list
*ev
;
3847 list_for_each_entry(ev
, &mem
->oom_notify
, list
)
3848 eventfd_signal(ev
->eventfd
, 1);
3852 static void mem_cgroup_oom_notify(struct mem_cgroup
*mem
)
3854 struct mem_cgroup
*iter
;
3856 for_each_mem_cgroup_tree(iter
, mem
)
3857 mem_cgroup_oom_notify_cb(iter
);
3860 static int mem_cgroup_usage_register_event(struct cgroup
*cgrp
,
3861 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
3863 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3864 struct mem_cgroup_thresholds
*thresholds
;
3865 struct mem_cgroup_threshold_ary
*new;
3866 int type
= MEMFILE_TYPE(cft
->private);
3867 u64 threshold
, usage
;
3870 ret
= res_counter_memparse_write_strategy(args
, &threshold
);
3874 mutex_lock(&memcg
->thresholds_lock
);
3877 thresholds
= &memcg
->thresholds
;
3878 else if (type
== _MEMSWAP
)
3879 thresholds
= &memcg
->memsw_thresholds
;
3883 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
3885 /* Check if a threshold crossed before adding a new one */
3886 if (thresholds
->primary
)
3887 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3889 size
= thresholds
->primary
? thresholds
->primary
->size
+ 1 : 1;
3891 /* Allocate memory for new array of thresholds */
3892 new = kmalloc(sizeof(*new) + size
* sizeof(struct mem_cgroup_threshold
),
3900 /* Copy thresholds (if any) to new array */
3901 if (thresholds
->primary
) {
3902 memcpy(new->entries
, thresholds
->primary
->entries
, (size
- 1) *
3903 sizeof(struct mem_cgroup_threshold
));
3906 /* Add new threshold */
3907 new->entries
[size
- 1].eventfd
= eventfd
;
3908 new->entries
[size
- 1].threshold
= threshold
;
3910 /* Sort thresholds. Registering of new threshold isn't time-critical */
3911 sort(new->entries
, size
, sizeof(struct mem_cgroup_threshold
),
3912 compare_thresholds
, NULL
);
3914 /* Find current threshold */
3915 new->current_threshold
= -1;
3916 for (i
= 0; i
< size
; i
++) {
3917 if (new->entries
[i
].threshold
< usage
) {
3919 * new->current_threshold will not be used until
3920 * rcu_assign_pointer(), so it's safe to increment
3923 ++new->current_threshold
;
3927 /* Free old spare buffer and save old primary buffer as spare */
3928 kfree(thresholds
->spare
);
3929 thresholds
->spare
= thresholds
->primary
;
3931 rcu_assign_pointer(thresholds
->primary
, new);
3933 /* To be sure that nobody uses thresholds */
3937 mutex_unlock(&memcg
->thresholds_lock
);
3942 static void mem_cgroup_usage_unregister_event(struct cgroup
*cgrp
,
3943 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
3945 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
3946 struct mem_cgroup_thresholds
*thresholds
;
3947 struct mem_cgroup_threshold_ary
*new;
3948 int type
= MEMFILE_TYPE(cft
->private);
3952 mutex_lock(&memcg
->thresholds_lock
);
3954 thresholds
= &memcg
->thresholds
;
3955 else if (type
== _MEMSWAP
)
3956 thresholds
= &memcg
->memsw_thresholds
;
3961 * Something went wrong if we trying to unregister a threshold
3962 * if we don't have thresholds
3964 BUG_ON(!thresholds
);
3966 usage
= mem_cgroup_usage(memcg
, type
== _MEMSWAP
);
3968 /* Check if a threshold crossed before removing */
3969 __mem_cgroup_threshold(memcg
, type
== _MEMSWAP
);
3971 /* Calculate new number of threshold */
3973 for (i
= 0; i
< thresholds
->primary
->size
; i
++) {
3974 if (thresholds
->primary
->entries
[i
].eventfd
!= eventfd
)
3978 new = thresholds
->spare
;
3980 /* Set thresholds array to NULL if we don't have thresholds */
3989 /* Copy thresholds and find current threshold */
3990 new->current_threshold
= -1;
3991 for (i
= 0, j
= 0; i
< thresholds
->primary
->size
; i
++) {
3992 if (thresholds
->primary
->entries
[i
].eventfd
== eventfd
)
3995 new->entries
[j
] = thresholds
->primary
->entries
[i
];
3996 if (new->entries
[j
].threshold
< usage
) {
3998 * new->current_threshold will not be used
3999 * until rcu_assign_pointer(), so it's safe to increment
4002 ++new->current_threshold
;
4008 /* Swap primary and spare array */
4009 thresholds
->spare
= thresholds
->primary
;
4010 rcu_assign_pointer(thresholds
->primary
, new);
4012 /* To be sure that nobody uses thresholds */
4015 mutex_unlock(&memcg
->thresholds_lock
);
4018 static int mem_cgroup_oom_register_event(struct cgroup
*cgrp
,
4019 struct cftype
*cft
, struct eventfd_ctx
*eventfd
, const char *args
)
4021 struct mem_cgroup
*memcg
= mem_cgroup_from_cont(cgrp
);
4022 struct mem_cgroup_eventfd_list
*event
;
4023 int type
= MEMFILE_TYPE(cft
->private);
4025 BUG_ON(type
!= _OOM_TYPE
);
4026 event
= kmalloc(sizeof(*event
), GFP_KERNEL
);
4030 mutex_lock(&memcg_oom_mutex
);
4032 event
->eventfd
= eventfd
;
4033 list_add(&event
->list
, &memcg
->oom_notify
);
4035 /* already in OOM ? */
4036 if (atomic_read(&memcg
->oom_lock
))
4037 eventfd_signal(eventfd
, 1);
4038 mutex_unlock(&memcg_oom_mutex
);
4043 static void mem_cgroup_oom_unregister_event(struct cgroup
*cgrp
,
4044 struct cftype
*cft
, struct eventfd_ctx
*eventfd
)
4046 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4047 struct mem_cgroup_eventfd_list
*ev
, *tmp
;
4048 int type
= MEMFILE_TYPE(cft
->private);
4050 BUG_ON(type
!= _OOM_TYPE
);
4052 mutex_lock(&memcg_oom_mutex
);
4054 list_for_each_entry_safe(ev
, tmp
, &mem
->oom_notify
, list
) {
4055 if (ev
->eventfd
== eventfd
) {
4056 list_del(&ev
->list
);
4061 mutex_unlock(&memcg_oom_mutex
);
4064 static int mem_cgroup_oom_control_read(struct cgroup
*cgrp
,
4065 struct cftype
*cft
, struct cgroup_map_cb
*cb
)
4067 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4069 cb
->fill(cb
, "oom_kill_disable", mem
->oom_kill_disable
);
4071 if (atomic_read(&mem
->oom_lock
))
4072 cb
->fill(cb
, "under_oom", 1);
4074 cb
->fill(cb
, "under_oom", 0);
4078 static int mem_cgroup_oom_control_write(struct cgroup
*cgrp
,
4079 struct cftype
*cft
, u64 val
)
4081 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgrp
);
4082 struct mem_cgroup
*parent
;
4084 /* cannot set to root cgroup and only 0 and 1 are allowed */
4085 if (!cgrp
->parent
|| !((val
== 0) || (val
== 1)))
4088 parent
= mem_cgroup_from_cont(cgrp
->parent
);
4091 /* oom-kill-disable is a flag for subhierarchy. */
4092 if ((parent
->use_hierarchy
) ||
4093 (mem
->use_hierarchy
&& !list_empty(&cgrp
->children
))) {
4097 mem
->oom_kill_disable
= val
;
4099 memcg_oom_recover(mem
);
4104 static struct cftype mem_cgroup_files
[] = {
4106 .name
= "usage_in_bytes",
4107 .private = MEMFILE_PRIVATE(_MEM
, RES_USAGE
),
4108 .read_u64
= mem_cgroup_read
,
4109 .register_event
= mem_cgroup_usage_register_event
,
4110 .unregister_event
= mem_cgroup_usage_unregister_event
,
4113 .name
= "max_usage_in_bytes",
4114 .private = MEMFILE_PRIVATE(_MEM
, RES_MAX_USAGE
),
4115 .trigger
= mem_cgroup_reset
,
4116 .read_u64
= mem_cgroup_read
,
4119 .name
= "limit_in_bytes",
4120 .private = MEMFILE_PRIVATE(_MEM
, RES_LIMIT
),
4121 .write_string
= mem_cgroup_write
,
4122 .read_u64
= mem_cgroup_read
,
4125 .name
= "soft_limit_in_bytes",
4126 .private = MEMFILE_PRIVATE(_MEM
, RES_SOFT_LIMIT
),
4127 .write_string
= mem_cgroup_write
,
4128 .read_u64
= mem_cgroup_read
,
4132 .private = MEMFILE_PRIVATE(_MEM
, RES_FAILCNT
),
4133 .trigger
= mem_cgroup_reset
,
4134 .read_u64
= mem_cgroup_read
,
4138 .read_map
= mem_control_stat_show
,
4141 .name
= "force_empty",
4142 .trigger
= mem_cgroup_force_empty_write
,
4145 .name
= "use_hierarchy",
4146 .write_u64
= mem_cgroup_hierarchy_write
,
4147 .read_u64
= mem_cgroup_hierarchy_read
,
4150 .name
= "swappiness",
4151 .read_u64
= mem_cgroup_swappiness_read
,
4152 .write_u64
= mem_cgroup_swappiness_write
,
4155 .name
= "move_charge_at_immigrate",
4156 .read_u64
= mem_cgroup_move_charge_read
,
4157 .write_u64
= mem_cgroup_move_charge_write
,
4160 .name
= "oom_control",
4161 .read_map
= mem_cgroup_oom_control_read
,
4162 .write_u64
= mem_cgroup_oom_control_write
,
4163 .register_event
= mem_cgroup_oom_register_event
,
4164 .unregister_event
= mem_cgroup_oom_unregister_event
,
4165 .private = MEMFILE_PRIVATE(_OOM_TYPE
, OOM_CONTROL
),
4169 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4170 static struct cftype memsw_cgroup_files
[] = {
4172 .name
= "memsw.usage_in_bytes",
4173 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_USAGE
),
4174 .read_u64
= mem_cgroup_read
,
4175 .register_event
= mem_cgroup_usage_register_event
,
4176 .unregister_event
= mem_cgroup_usage_unregister_event
,
4179 .name
= "memsw.max_usage_in_bytes",
4180 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_MAX_USAGE
),
4181 .trigger
= mem_cgroup_reset
,
4182 .read_u64
= mem_cgroup_read
,
4185 .name
= "memsw.limit_in_bytes",
4186 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_LIMIT
),
4187 .write_string
= mem_cgroup_write
,
4188 .read_u64
= mem_cgroup_read
,
4191 .name
= "memsw.failcnt",
4192 .private = MEMFILE_PRIVATE(_MEMSWAP
, RES_FAILCNT
),
4193 .trigger
= mem_cgroup_reset
,
4194 .read_u64
= mem_cgroup_read
,
4198 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4200 if (!do_swap_account
)
4202 return cgroup_add_files(cont
, ss
, memsw_cgroup_files
,
4203 ARRAY_SIZE(memsw_cgroup_files
));
4206 static int register_memsw_files(struct cgroup
*cont
, struct cgroup_subsys
*ss
)
4212 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4214 struct mem_cgroup_per_node
*pn
;
4215 struct mem_cgroup_per_zone
*mz
;
4217 int zone
, tmp
= node
;
4219 * This routine is called against possible nodes.
4220 * But it's BUG to call kmalloc() against offline node.
4222 * TODO: this routine can waste much memory for nodes which will
4223 * never be onlined. It's better to use memory hotplug callback
4226 if (!node_state(node
, N_NORMAL_MEMORY
))
4228 pn
= kzalloc_node(sizeof(*pn
), GFP_KERNEL
, tmp
);
4232 mem
->info
.nodeinfo
[node
] = pn
;
4233 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4234 mz
= &pn
->zoneinfo
[zone
];
4236 INIT_LIST_HEAD(&mz
->lists
[l
]);
4237 mz
->usage_in_excess
= 0;
4238 mz
->on_tree
= false;
4244 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
4246 kfree(mem
->info
.nodeinfo
[node
]);
4249 static struct mem_cgroup
*mem_cgroup_alloc(void)
4251 struct mem_cgroup
*mem
;
4252 int size
= sizeof(struct mem_cgroup
);
4254 /* Can be very big if MAX_NUMNODES is very big */
4255 if (size
< PAGE_SIZE
)
4256 mem
= kzalloc(size
, GFP_KERNEL
);
4258 mem
= vzalloc(size
);
4263 mem
->stat
= alloc_percpu(struct mem_cgroup_stat_cpu
);
4266 spin_lock_init(&mem
->pcp_counter_lock
);
4270 if (size
< PAGE_SIZE
)
4278 * At destroying mem_cgroup, references from swap_cgroup can remain.
4279 * (scanning all at force_empty is too costly...)
4281 * Instead of clearing all references at force_empty, we remember
4282 * the number of reference from swap_cgroup and free mem_cgroup when
4283 * it goes down to 0.
4285 * Removal of cgroup itself succeeds regardless of refs from swap.
4288 static void __mem_cgroup_free(struct mem_cgroup
*mem
)
4292 mem_cgroup_remove_from_trees(mem
);
4293 free_css_id(&mem_cgroup_subsys
, &mem
->css
);
4295 for_each_node_state(node
, N_POSSIBLE
)
4296 free_mem_cgroup_per_zone_info(mem
, node
);
4298 free_percpu(mem
->stat
);
4299 if (sizeof(struct mem_cgroup
) < PAGE_SIZE
)
4305 static void mem_cgroup_get(struct mem_cgroup
*mem
)
4307 atomic_inc(&mem
->refcnt
);
4310 static void __mem_cgroup_put(struct mem_cgroup
*mem
, int count
)
4312 if (atomic_sub_and_test(count
, &mem
->refcnt
)) {
4313 struct mem_cgroup
*parent
= parent_mem_cgroup(mem
);
4314 __mem_cgroup_free(mem
);
4316 mem_cgroup_put(parent
);
4320 static void mem_cgroup_put(struct mem_cgroup
*mem
)
4322 __mem_cgroup_put(mem
, 1);
4326 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4328 static struct mem_cgroup
*parent_mem_cgroup(struct mem_cgroup
*mem
)
4330 if (!mem
->res
.parent
)
4332 return mem_cgroup_from_res_counter(mem
->res
.parent
, res
);
4335 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4336 static void __init
enable_swap_cgroup(void)
4338 if (!mem_cgroup_disabled() && really_do_swap_account
)
4339 do_swap_account
= 1;
4342 static void __init
enable_swap_cgroup(void)
4347 static int mem_cgroup_soft_limit_tree_init(void)
4349 struct mem_cgroup_tree_per_node
*rtpn
;
4350 struct mem_cgroup_tree_per_zone
*rtpz
;
4351 int tmp
, node
, zone
;
4353 for_each_node_state(node
, N_POSSIBLE
) {
4355 if (!node_state(node
, N_NORMAL_MEMORY
))
4357 rtpn
= kzalloc_node(sizeof(*rtpn
), GFP_KERNEL
, tmp
);
4361 soft_limit_tree
.rb_tree_per_node
[node
] = rtpn
;
4363 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
4364 rtpz
= &rtpn
->rb_tree_per_zone
[zone
];
4365 rtpz
->rb_root
= RB_ROOT
;
4366 spin_lock_init(&rtpz
->lock
);
4372 static struct cgroup_subsys_state
* __ref
4373 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
4375 struct mem_cgroup
*mem
, *parent
;
4376 long error
= -ENOMEM
;
4379 mem
= mem_cgroup_alloc();
4381 return ERR_PTR(error
);
4383 for_each_node_state(node
, N_POSSIBLE
)
4384 if (alloc_mem_cgroup_per_zone_info(mem
, node
))
4388 if (cont
->parent
== NULL
) {
4390 enable_swap_cgroup();
4392 root_mem_cgroup
= mem
;
4393 if (mem_cgroup_soft_limit_tree_init())
4395 for_each_possible_cpu(cpu
) {
4396 struct memcg_stock_pcp
*stock
=
4397 &per_cpu(memcg_stock
, cpu
);
4398 INIT_WORK(&stock
->work
, drain_local_stock
);
4400 hotcpu_notifier(memcg_cpu_hotplug_callback
, 0);
4402 parent
= mem_cgroup_from_cont(cont
->parent
);
4403 mem
->use_hierarchy
= parent
->use_hierarchy
;
4404 mem
->oom_kill_disable
= parent
->oom_kill_disable
;
4407 if (parent
&& parent
->use_hierarchy
) {
4408 res_counter_init(&mem
->res
, &parent
->res
);
4409 res_counter_init(&mem
->memsw
, &parent
->memsw
);
4411 * We increment refcnt of the parent to ensure that we can
4412 * safely access it on res_counter_charge/uncharge.
4413 * This refcnt will be decremented when freeing this
4414 * mem_cgroup(see mem_cgroup_put).
4416 mem_cgroup_get(parent
);
4418 res_counter_init(&mem
->res
, NULL
);
4419 res_counter_init(&mem
->memsw
, NULL
);
4421 mem
->last_scanned_child
= 0;
4422 spin_lock_init(&mem
->reclaim_param_lock
);
4423 INIT_LIST_HEAD(&mem
->oom_notify
);
4426 mem
->swappiness
= get_swappiness(parent
);
4427 atomic_set(&mem
->refcnt
, 1);
4428 mem
->move_charge_at_immigrate
= 0;
4429 mutex_init(&mem
->thresholds_lock
);
4432 __mem_cgroup_free(mem
);
4433 root_mem_cgroup
= NULL
;
4434 return ERR_PTR(error
);
4437 static int mem_cgroup_pre_destroy(struct cgroup_subsys
*ss
,
4438 struct cgroup
*cont
)
4440 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4442 return mem_cgroup_force_empty(mem
, false);
4445 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
4446 struct cgroup
*cont
)
4448 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
4450 mem_cgroup_put(mem
);
4453 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
4454 struct cgroup
*cont
)
4458 ret
= cgroup_add_files(cont
, ss
, mem_cgroup_files
,
4459 ARRAY_SIZE(mem_cgroup_files
));
4462 ret
= register_memsw_files(cont
, ss
);
4467 /* Handlers for move charge at task migration. */
4468 #define PRECHARGE_COUNT_AT_ONCE 256
4469 static int mem_cgroup_do_precharge(unsigned long count
)
4472 int batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4473 struct mem_cgroup
*mem
= mc
.to
;
4475 if (mem_cgroup_is_root(mem
)) {
4476 mc
.precharge
+= count
;
4477 /* we don't need css_get for root */
4480 /* try to charge at once */
4482 struct res_counter
*dummy
;
4484 * "mem" cannot be under rmdir() because we've already checked
4485 * by cgroup_lock_live_cgroup() that it is not removed and we
4486 * are still under the same cgroup_mutex. So we can postpone
4489 if (res_counter_charge(&mem
->res
, PAGE_SIZE
* count
, &dummy
))
4491 if (do_swap_account
&& res_counter_charge(&mem
->memsw
,
4492 PAGE_SIZE
* count
, &dummy
)) {
4493 res_counter_uncharge(&mem
->res
, PAGE_SIZE
* count
);
4496 mc
.precharge
+= count
;
4500 /* fall back to one by one charge */
4502 if (signal_pending(current
)) {
4506 if (!batch_count
--) {
4507 batch_count
= PRECHARGE_COUNT_AT_ONCE
;
4510 ret
= __mem_cgroup_try_charge(NULL
, GFP_KERNEL
, &mem
, false,
4513 /* mem_cgroup_clear_mc() will do uncharge later */
4521 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4522 * @vma: the vma the pte to be checked belongs
4523 * @addr: the address corresponding to the pte to be checked
4524 * @ptent: the pte to be checked
4525 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4528 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4529 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4530 * move charge. if @target is not NULL, the page is stored in target->page
4531 * with extra refcnt got(Callers should handle it).
4532 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4533 * target for charge migration. if @target is not NULL, the entry is stored
4536 * Called with pte lock held.
4543 enum mc_target_type
{
4544 MC_TARGET_NONE
, /* not used */
4549 static struct page
*mc_handle_present_pte(struct vm_area_struct
*vma
,
4550 unsigned long addr
, pte_t ptent
)
4552 struct page
*page
= vm_normal_page(vma
, addr
, ptent
);
4554 if (!page
|| !page_mapped(page
))
4556 if (PageAnon(page
)) {
4557 /* we don't move shared anon */
4558 if (!move_anon() || page_mapcount(page
) > 2)
4560 } else if (!move_file())
4561 /* we ignore mapcount for file pages */
4563 if (!get_page_unless_zero(page
))
4569 static struct page
*mc_handle_swap_pte(struct vm_area_struct
*vma
,
4570 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4573 struct page
*page
= NULL
;
4574 swp_entry_t ent
= pte_to_swp_entry(ptent
);
4576 if (!move_anon() || non_swap_entry(ent
))
4578 usage_count
= mem_cgroup_count_swap_user(ent
, &page
);
4579 if (usage_count
> 1) { /* we don't move shared anon */
4584 if (do_swap_account
)
4585 entry
->val
= ent
.val
;
4590 static struct page
*mc_handle_file_pte(struct vm_area_struct
*vma
,
4591 unsigned long addr
, pte_t ptent
, swp_entry_t
*entry
)
4593 struct page
*page
= NULL
;
4594 struct inode
*inode
;
4595 struct address_space
*mapping
;
4598 if (!vma
->vm_file
) /* anonymous vma */
4603 inode
= vma
->vm_file
->f_path
.dentry
->d_inode
;
4604 mapping
= vma
->vm_file
->f_mapping
;
4605 if (pte_none(ptent
))
4606 pgoff
= linear_page_index(vma
, addr
);
4607 else /* pte_file(ptent) is true */
4608 pgoff
= pte_to_pgoff(ptent
);
4610 /* page is moved even if it's not RSS of this task(page-faulted). */
4611 if (!mapping_cap_swap_backed(mapping
)) { /* normal file */
4612 page
= find_get_page(mapping
, pgoff
);
4613 } else { /* shmem/tmpfs file. we should take account of swap too. */
4615 mem_cgroup_get_shmem_target(inode
, pgoff
, &page
, &ent
);
4616 if (do_swap_account
)
4617 entry
->val
= ent
.val
;
4623 static int is_target_pte_for_mc(struct vm_area_struct
*vma
,
4624 unsigned long addr
, pte_t ptent
, union mc_target
*target
)
4626 struct page
*page
= NULL
;
4627 struct page_cgroup
*pc
;
4629 swp_entry_t ent
= { .val
= 0 };
4631 if (pte_present(ptent
))
4632 page
= mc_handle_present_pte(vma
, addr
, ptent
);
4633 else if (is_swap_pte(ptent
))
4634 page
= mc_handle_swap_pte(vma
, addr
, ptent
, &ent
);
4635 else if (pte_none(ptent
) || pte_file(ptent
))
4636 page
= mc_handle_file_pte(vma
, addr
, ptent
, &ent
);
4638 if (!page
&& !ent
.val
)
4641 pc
= lookup_page_cgroup(page
);
4643 * Do only loose check w/o page_cgroup lock.
4644 * mem_cgroup_move_account() checks the pc is valid or not under
4647 if (PageCgroupUsed(pc
) && pc
->mem_cgroup
== mc
.from
) {
4648 ret
= MC_TARGET_PAGE
;
4650 target
->page
= page
;
4652 if (!ret
|| !target
)
4655 /* There is a swap entry and a page doesn't exist or isn't charged */
4656 if (ent
.val
&& !ret
&&
4657 css_id(&mc
.from
->css
) == lookup_swap_cgroup(ent
)) {
4658 ret
= MC_TARGET_SWAP
;
4665 static int mem_cgroup_count_precharge_pte_range(pmd_t
*pmd
,
4666 unsigned long addr
, unsigned long end
,
4667 struct mm_walk
*walk
)
4669 struct vm_area_struct
*vma
= walk
->private;
4673 VM_BUG_ON(pmd_trans_huge(*pmd
));
4674 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4675 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
)
4676 if (is_target_pte_for_mc(vma
, addr
, *pte
, NULL
))
4677 mc
.precharge
++; /* increment precharge temporarily */
4678 pte_unmap_unlock(pte
- 1, ptl
);
4684 static unsigned long mem_cgroup_count_precharge(struct mm_struct
*mm
)
4686 unsigned long precharge
;
4687 struct vm_area_struct
*vma
;
4689 down_read(&mm
->mmap_sem
);
4690 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
4691 struct mm_walk mem_cgroup_count_precharge_walk
= {
4692 .pmd_entry
= mem_cgroup_count_precharge_pte_range
,
4696 if (is_vm_hugetlb_page(vma
))
4698 walk_page_range(vma
->vm_start
, vma
->vm_end
,
4699 &mem_cgroup_count_precharge_walk
);
4701 up_read(&mm
->mmap_sem
);
4703 precharge
= mc
.precharge
;
4709 static int mem_cgroup_precharge_mc(struct mm_struct
*mm
)
4711 unsigned long precharge
= mem_cgroup_count_precharge(mm
);
4713 VM_BUG_ON(mc
.moving_task
);
4714 mc
.moving_task
= current
;
4715 return mem_cgroup_do_precharge(precharge
);
4718 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4719 static void __mem_cgroup_clear_mc(void)
4721 struct mem_cgroup
*from
= mc
.from
;
4722 struct mem_cgroup
*to
= mc
.to
;
4724 /* we must uncharge all the leftover precharges from mc.to */
4726 __mem_cgroup_cancel_charge(mc
.to
, mc
.precharge
);
4730 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4731 * we must uncharge here.
4733 if (mc
.moved_charge
) {
4734 __mem_cgroup_cancel_charge(mc
.from
, mc
.moved_charge
);
4735 mc
.moved_charge
= 0;
4737 /* we must fixup refcnts and charges */
4738 if (mc
.moved_swap
) {
4739 /* uncharge swap account from the old cgroup */
4740 if (!mem_cgroup_is_root(mc
.from
))
4741 res_counter_uncharge(&mc
.from
->memsw
,
4742 PAGE_SIZE
* mc
.moved_swap
);
4743 __mem_cgroup_put(mc
.from
, mc
.moved_swap
);
4745 if (!mem_cgroup_is_root(mc
.to
)) {
4747 * we charged both to->res and to->memsw, so we should
4750 res_counter_uncharge(&mc
.to
->res
,
4751 PAGE_SIZE
* mc
.moved_swap
);
4753 /* we've already done mem_cgroup_get(mc.to) */
4756 memcg_oom_recover(from
);
4757 memcg_oom_recover(to
);
4758 wake_up_all(&mc
.waitq
);
4761 static void mem_cgroup_clear_mc(void)
4763 struct mem_cgroup
*from
= mc
.from
;
4766 * we must clear moving_task before waking up waiters at the end of
4769 mc
.moving_task
= NULL
;
4770 __mem_cgroup_clear_mc();
4771 spin_lock(&mc
.lock
);
4774 spin_unlock(&mc
.lock
);
4775 mem_cgroup_end_move(from
);
4778 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
4779 struct cgroup
*cgroup
,
4780 struct task_struct
*p
,
4784 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cgroup
);
4786 if (mem
->move_charge_at_immigrate
) {
4787 struct mm_struct
*mm
;
4788 struct mem_cgroup
*from
= mem_cgroup_from_task(p
);
4790 VM_BUG_ON(from
== mem
);
4792 mm
= get_task_mm(p
);
4795 /* We move charges only when we move a owner of the mm */
4796 if (mm
->owner
== p
) {
4799 VM_BUG_ON(mc
.precharge
);
4800 VM_BUG_ON(mc
.moved_charge
);
4801 VM_BUG_ON(mc
.moved_swap
);
4802 mem_cgroup_start_move(from
);
4803 spin_lock(&mc
.lock
);
4806 spin_unlock(&mc
.lock
);
4807 /* We set mc.moving_task later */
4809 ret
= mem_cgroup_precharge_mc(mm
);
4811 mem_cgroup_clear_mc();
4818 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
4819 struct cgroup
*cgroup
,
4820 struct task_struct
*p
,
4823 mem_cgroup_clear_mc();
4826 static int mem_cgroup_move_charge_pte_range(pmd_t
*pmd
,
4827 unsigned long addr
, unsigned long end
,
4828 struct mm_walk
*walk
)
4831 struct vm_area_struct
*vma
= walk
->private;
4836 VM_BUG_ON(pmd_trans_huge(*pmd
));
4837 pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
4838 for (; addr
!= end
; addr
+= PAGE_SIZE
) {
4839 pte_t ptent
= *(pte
++);
4840 union mc_target target
;
4843 struct page_cgroup
*pc
;
4849 type
= is_target_pte_for_mc(vma
, addr
, ptent
, &target
);
4851 case MC_TARGET_PAGE
:
4853 if (isolate_lru_page(page
))
4855 pc
= lookup_page_cgroup(page
);
4856 if (!mem_cgroup_move_account(pc
,
4857 mc
.from
, mc
.to
, false)) {
4859 /* we uncharge from mc.from later. */
4862 putback_lru_page(page
);
4863 put
: /* is_target_pte_for_mc() gets the page */
4866 case MC_TARGET_SWAP
:
4868 if (!mem_cgroup_move_swap_account(ent
,
4869 mc
.from
, mc
.to
, false)) {
4871 /* we fixup refcnts and charges later. */
4879 pte_unmap_unlock(pte
- 1, ptl
);
4884 * We have consumed all precharges we got in can_attach().
4885 * We try charge one by one, but don't do any additional
4886 * charges to mc.to if we have failed in charge once in attach()
4889 ret
= mem_cgroup_do_precharge(1);
4897 static void mem_cgroup_move_charge(struct mm_struct
*mm
)
4899 struct vm_area_struct
*vma
;
4901 lru_add_drain_all();
4903 if (unlikely(!down_read_trylock(&mm
->mmap_sem
))) {
4905 * Someone who are holding the mmap_sem might be waiting in
4906 * waitq. So we cancel all extra charges, wake up all waiters,
4907 * and retry. Because we cancel precharges, we might not be able
4908 * to move enough charges, but moving charge is a best-effort
4909 * feature anyway, so it wouldn't be a big problem.
4911 __mem_cgroup_clear_mc();
4915 for (vma
= mm
->mmap
; vma
; vma
= vma
->vm_next
) {
4917 struct mm_walk mem_cgroup_move_charge_walk
= {
4918 .pmd_entry
= mem_cgroup_move_charge_pte_range
,
4922 if (is_vm_hugetlb_page(vma
))
4924 ret
= walk_page_range(vma
->vm_start
, vma
->vm_end
,
4925 &mem_cgroup_move_charge_walk
);
4928 * means we have consumed all precharges and failed in
4929 * doing additional charge. Just abandon here.
4933 up_read(&mm
->mmap_sem
);
4936 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
4937 struct cgroup
*cont
,
4938 struct cgroup
*old_cont
,
4939 struct task_struct
*p
,
4942 struct mm_struct
*mm
;
4945 /* no need to move charge */
4948 mm
= get_task_mm(p
);
4950 mem_cgroup_move_charge(mm
);
4953 mem_cgroup_clear_mc();
4955 #else /* !CONFIG_MMU */
4956 static int mem_cgroup_can_attach(struct cgroup_subsys
*ss
,
4957 struct cgroup
*cgroup
,
4958 struct task_struct
*p
,
4963 static void mem_cgroup_cancel_attach(struct cgroup_subsys
*ss
,
4964 struct cgroup
*cgroup
,
4965 struct task_struct
*p
,
4969 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
4970 struct cgroup
*cont
,
4971 struct cgroup
*old_cont
,
4972 struct task_struct
*p
,
4978 struct cgroup_subsys mem_cgroup_subsys
= {
4980 .subsys_id
= mem_cgroup_subsys_id
,
4981 .create
= mem_cgroup_create
,
4982 .pre_destroy
= mem_cgroup_pre_destroy
,
4983 .destroy
= mem_cgroup_destroy
,
4984 .populate
= mem_cgroup_populate
,
4985 .can_attach
= mem_cgroup_can_attach
,
4986 .cancel_attach
= mem_cgroup_cancel_attach
,
4987 .attach
= mem_cgroup_move_task
,
4992 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4993 static int __init
enable_swap_account(char *s
)
4995 /* consider enabled if no parameter or 1 is given */
4996 if (!s
|| !strcmp(s
, "1"))
4997 really_do_swap_account
= 1;
4998 else if (!strcmp(s
, "0"))
4999 really_do_swap_account
= 0;
5002 __setup("swapaccount", enable_swap_account
);
5004 static int __init
disable_swap_account(char *s
)
5006 enable_swap_account("0");
5009 __setup("noswapaccount", disable_swap_account
);