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>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
20 #include <linux/res_counter.h>
21 #include <linux/memcontrol.h>
22 #include <linux/cgroup.h>
24 #include <linux/smp.h>
25 #include <linux/page-flags.h>
26 #include <linux/backing-dev.h>
27 #include <linux/bit_spinlock.h>
28 #include <linux/rcupdate.h>
29 #include <linux/swap.h>
30 #include <linux/spinlock.h>
32 #include <linux/seq_file.h>
34 #include <asm/uaccess.h>
36 struct cgroup_subsys mem_cgroup_subsys
;
37 static const int MEM_CGROUP_RECLAIM_RETRIES
= 5;
40 * Statistics for memory cgroup.
42 enum mem_cgroup_stat_index
{
44 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
46 MEM_CGROUP_STAT_CACHE
, /* # of pages charged as cache */
47 MEM_CGROUP_STAT_RSS
, /* # of pages charged as rss */
49 MEM_CGROUP_STAT_NSTATS
,
52 struct mem_cgroup_stat_cpu
{
53 s64 count
[MEM_CGROUP_STAT_NSTATS
];
54 } ____cacheline_aligned_in_smp
;
56 struct mem_cgroup_stat
{
57 struct mem_cgroup_stat_cpu cpustat
[NR_CPUS
];
61 * For accounting under irq disable, no need for increment preempt count.
63 static void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat
*stat
,
64 enum mem_cgroup_stat_index idx
, int val
)
66 int cpu
= smp_processor_id();
67 stat
->cpustat
[cpu
].count
[idx
] += val
;
70 static s64
mem_cgroup_read_stat(struct mem_cgroup_stat
*stat
,
71 enum mem_cgroup_stat_index idx
)
75 for_each_possible_cpu(cpu
)
76 ret
+= stat
->cpustat
[cpu
].count
[idx
];
81 * per-zone information in memory controller.
84 enum mem_cgroup_zstat_index
{
85 MEM_CGROUP_ZSTAT_ACTIVE
,
86 MEM_CGROUP_ZSTAT_INACTIVE
,
91 struct mem_cgroup_per_zone
{
93 * spin_lock to protect the per cgroup LRU
96 struct list_head active_list
;
97 struct list_head inactive_list
;
98 unsigned long count
[NR_MEM_CGROUP_ZSTAT
];
100 /* Macro for accessing counter */
101 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
103 struct mem_cgroup_per_node
{
104 struct mem_cgroup_per_zone zoneinfo
[MAX_NR_ZONES
];
107 struct mem_cgroup_lru_info
{
108 struct mem_cgroup_per_node
*nodeinfo
[MAX_NUMNODES
];
112 * The memory controller data structure. The memory controller controls both
113 * page cache and RSS per cgroup. We would eventually like to provide
114 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
115 * to help the administrator determine what knobs to tune.
117 * TODO: Add a water mark for the memory controller. Reclaim will begin when
118 * we hit the water mark. May be even add a low water mark, such that
119 * no reclaim occurs from a cgroup at it's low water mark, this is
120 * a feature that will be implemented much later in the future.
123 struct cgroup_subsys_state css
;
125 * the counter to account for memory usage
127 struct res_counter res
;
129 * Per cgroup active and inactive list, similar to the
130 * per zone LRU lists.
132 struct mem_cgroup_lru_info info
;
134 int prev_priority
; /* for recording reclaim priority */
138 struct mem_cgroup_stat stat
;
142 * We use the lower bit of the page->page_cgroup pointer as a bit spin
143 * lock. We need to ensure that page->page_cgroup is atleast two
144 * byte aligned (based on comments from Nick Piggin)
146 #define PAGE_CGROUP_LOCK_BIT 0x0
147 #define PAGE_CGROUP_LOCK (1 << PAGE_CGROUP_LOCK_BIT)
150 * A page_cgroup page is associated with every page descriptor. The
151 * page_cgroup helps us identify information about the cgroup
154 struct list_head lru
; /* per cgroup LRU list */
156 struct mem_cgroup
*mem_cgroup
;
157 atomic_t ref_cnt
; /* Helpful when pages move b/w */
158 /* mapped and cached states */
161 #define PAGE_CGROUP_FLAG_CACHE (0x1) /* charged as cache */
162 #define PAGE_CGROUP_FLAG_ACTIVE (0x2) /* page is active in this cgroup */
164 static inline int page_cgroup_nid(struct page_cgroup
*pc
)
166 return page_to_nid(pc
->page
);
169 static inline enum zone_type
page_cgroup_zid(struct page_cgroup
*pc
)
171 return page_zonenum(pc
->page
);
175 MEM_CGROUP_TYPE_UNSPEC
= 0,
176 MEM_CGROUP_TYPE_MAPPED
,
177 MEM_CGROUP_TYPE_CACHED
,
183 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
184 MEM_CGROUP_CHARGE_TYPE_MAPPED
,
189 * Always modified under lru lock. Then, not necessary to preempt_disable()
191 static void mem_cgroup_charge_statistics(struct mem_cgroup
*mem
, int flags
,
194 int val
= (charge
)? 1 : -1;
195 struct mem_cgroup_stat
*stat
= &mem
->stat
;
196 VM_BUG_ON(!irqs_disabled());
198 if (flags
& PAGE_CGROUP_FLAG_CACHE
)
199 __mem_cgroup_stat_add_safe(stat
,
200 MEM_CGROUP_STAT_CACHE
, val
);
202 __mem_cgroup_stat_add_safe(stat
, MEM_CGROUP_STAT_RSS
, val
);
205 static inline struct mem_cgroup_per_zone
*
206 mem_cgroup_zoneinfo(struct mem_cgroup
*mem
, int nid
, int zid
)
208 BUG_ON(!mem
->info
.nodeinfo
[nid
]);
209 return &mem
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
212 static inline struct mem_cgroup_per_zone
*
213 page_cgroup_zoneinfo(struct page_cgroup
*pc
)
215 struct mem_cgroup
*mem
= pc
->mem_cgroup
;
216 int nid
= page_cgroup_nid(pc
);
217 int zid
= page_cgroup_zid(pc
);
219 return mem_cgroup_zoneinfo(mem
, nid
, zid
);
222 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup
*mem
,
223 enum mem_cgroup_zstat_index idx
)
226 struct mem_cgroup_per_zone
*mz
;
229 for_each_online_node(nid
)
230 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
231 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
232 total
+= MEM_CGROUP_ZSTAT(mz
, idx
);
237 static struct mem_cgroup init_mem_cgroup
;
240 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
242 return container_of(cgroup_subsys_state(cont
,
243 mem_cgroup_subsys_id
), struct mem_cgroup
,
248 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
250 return container_of(task_subsys_state(p
, mem_cgroup_subsys_id
),
251 struct mem_cgroup
, css
);
254 void mm_init_cgroup(struct mm_struct
*mm
, struct task_struct
*p
)
256 struct mem_cgroup
*mem
;
258 mem
= mem_cgroup_from_task(p
);
260 mm
->mem_cgroup
= mem
;
263 void mm_free_cgroup(struct mm_struct
*mm
)
265 css_put(&mm
->mem_cgroup
->css
);
268 static inline int page_cgroup_locked(struct page
*page
)
270 return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT
,
274 void page_assign_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
279 * While resetting the page_cgroup we might not hold the
280 * page_cgroup lock. free_hot_cold_page() is an example
284 VM_BUG_ON(!page_cgroup_locked(page
));
285 locked
= (page
->page_cgroup
& PAGE_CGROUP_LOCK
);
286 page
->page_cgroup
= ((unsigned long)pc
| locked
);
289 struct page_cgroup
*page_get_page_cgroup(struct page
*page
)
291 return (struct page_cgroup
*)
292 (page
->page_cgroup
& ~PAGE_CGROUP_LOCK
);
295 static void __always_inline
lock_page_cgroup(struct page
*page
)
297 bit_spin_lock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
298 VM_BUG_ON(!page_cgroup_locked(page
));
301 static void __always_inline
unlock_page_cgroup(struct page
*page
)
303 bit_spin_unlock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
307 * Tie new page_cgroup to struct page under lock_page_cgroup()
308 * This can fail if the page has been tied to a page_cgroup.
309 * If success, returns 0.
311 static int page_cgroup_assign_new_page_cgroup(struct page
*page
,
312 struct page_cgroup
*pc
)
316 lock_page_cgroup(page
);
317 if (!page_get_page_cgroup(page
))
318 page_assign_page_cgroup(page
, pc
);
319 else /* A page is tied to other pc. */
321 unlock_page_cgroup(page
);
326 * Clear page->page_cgroup member under lock_page_cgroup().
327 * If given "pc" value is different from one page->page_cgroup,
328 * page->cgroup is not cleared.
329 * Returns a value of page->page_cgroup at lock taken.
330 * A can can detect failure of clearing by following
331 * clear_page_cgroup(page, pc) == pc
334 static struct page_cgroup
*clear_page_cgroup(struct page
*page
,
335 struct page_cgroup
*pc
)
337 struct page_cgroup
*ret
;
339 lock_page_cgroup(page
);
340 ret
= page_get_page_cgroup(page
);
341 if (likely(ret
== pc
))
342 page_assign_page_cgroup(page
, NULL
);
343 unlock_page_cgroup(page
);
347 static void __mem_cgroup_remove_list(struct page_cgroup
*pc
)
349 int from
= pc
->flags
& PAGE_CGROUP_FLAG_ACTIVE
;
350 struct mem_cgroup_per_zone
*mz
= page_cgroup_zoneinfo(pc
);
353 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) -= 1;
355 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) -= 1;
357 mem_cgroup_charge_statistics(pc
->mem_cgroup
, pc
->flags
, false);
358 list_del_init(&pc
->lru
);
361 static void __mem_cgroup_add_list(struct page_cgroup
*pc
)
363 int to
= pc
->flags
& PAGE_CGROUP_FLAG_ACTIVE
;
364 struct mem_cgroup_per_zone
*mz
= page_cgroup_zoneinfo(pc
);
367 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) += 1;
368 list_add(&pc
->lru
, &mz
->inactive_list
);
370 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) += 1;
371 list_add(&pc
->lru
, &mz
->active_list
);
373 mem_cgroup_charge_statistics(pc
->mem_cgroup
, pc
->flags
, true);
376 static void __mem_cgroup_move_lists(struct page_cgroup
*pc
, bool active
)
378 int from
= pc
->flags
& PAGE_CGROUP_FLAG_ACTIVE
;
379 struct mem_cgroup_per_zone
*mz
= page_cgroup_zoneinfo(pc
);
382 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) -= 1;
384 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) -= 1;
387 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) += 1;
388 pc
->flags
|= PAGE_CGROUP_FLAG_ACTIVE
;
389 list_move(&pc
->lru
, &mz
->active_list
);
391 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) += 1;
392 pc
->flags
&= ~PAGE_CGROUP_FLAG_ACTIVE
;
393 list_move(&pc
->lru
, &mz
->inactive_list
);
397 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
402 ret
= task
->mm
&& mm_match_cgroup(task
->mm
, mem
);
408 * This routine assumes that the appropriate zone's lru lock is already held
410 void mem_cgroup_move_lists(struct page
*page
, bool active
)
412 struct page_cgroup
*pc
;
413 struct mem_cgroup_per_zone
*mz
;
416 pc
= page_get_page_cgroup(page
);
420 mz
= page_cgroup_zoneinfo(pc
);
421 spin_lock_irqsave(&mz
->lru_lock
, flags
);
422 __mem_cgroup_move_lists(pc
, active
);
423 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
427 * Calculate mapped_ratio under memory controller. This will be used in
428 * vmscan.c for deteremining we have to reclaim mapped pages.
430 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup
*mem
)
435 * usage is recorded in bytes. But, here, we assume the number of
436 * physical pages can be represented by "long" on any arch.
438 total
= (long) (mem
->res
.usage
>> PAGE_SHIFT
) + 1L;
439 rss
= (long)mem_cgroup_read_stat(&mem
->stat
, MEM_CGROUP_STAT_RSS
);
440 return (int)((rss
* 100L) / total
);
443 * This function is called from vmscan.c. In page reclaiming loop. balance
444 * between active and inactive list is calculated. For memory controller
445 * page reclaiming, we should use using mem_cgroup's imbalance rather than
446 * zone's global lru imbalance.
448 long mem_cgroup_reclaim_imbalance(struct mem_cgroup
*mem
)
450 unsigned long active
, inactive
;
451 /* active and inactive are the number of pages. 'long' is ok.*/
452 active
= mem_cgroup_get_all_zonestat(mem
, MEM_CGROUP_ZSTAT_ACTIVE
);
453 inactive
= mem_cgroup_get_all_zonestat(mem
, MEM_CGROUP_ZSTAT_INACTIVE
);
454 return (long) (active
/ (inactive
+ 1));
458 * prev_priority control...this will be used in memory reclaim path.
460 int mem_cgroup_get_reclaim_priority(struct mem_cgroup
*mem
)
462 return mem
->prev_priority
;
465 void mem_cgroup_note_reclaim_priority(struct mem_cgroup
*mem
, int priority
)
467 if (priority
< mem
->prev_priority
)
468 mem
->prev_priority
= priority
;
471 void mem_cgroup_record_reclaim_priority(struct mem_cgroup
*mem
, int priority
)
473 mem
->prev_priority
= priority
;
477 * Calculate # of pages to be scanned in this priority/zone.
480 * priority starts from "DEF_PRIORITY" and decremented in each loop.
481 * (see include/linux/mmzone.h)
484 long mem_cgroup_calc_reclaim_active(struct mem_cgroup
*mem
,
485 struct zone
*zone
, int priority
)
488 int nid
= zone
->zone_pgdat
->node_id
;
489 int zid
= zone_idx(zone
);
490 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
492 nr_active
= MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
);
493 return (nr_active
>> priority
);
496 long mem_cgroup_calc_reclaim_inactive(struct mem_cgroup
*mem
,
497 struct zone
*zone
, int priority
)
500 int nid
= zone
->zone_pgdat
->node_id
;
501 int zid
= zone_idx(zone
);
502 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
504 nr_inactive
= MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
);
506 return (nr_inactive
>> priority
);
509 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
510 struct list_head
*dst
,
511 unsigned long *scanned
, int order
,
512 int mode
, struct zone
*z
,
513 struct mem_cgroup
*mem_cont
,
516 unsigned long nr_taken
= 0;
520 struct list_head
*src
;
521 struct page_cgroup
*pc
, *tmp
;
522 int nid
= z
->zone_pgdat
->node_id
;
523 int zid
= zone_idx(z
);
524 struct mem_cgroup_per_zone
*mz
;
526 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
528 src
= &mz
->active_list
;
530 src
= &mz
->inactive_list
;
533 spin_lock(&mz
->lru_lock
);
535 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
536 if (scan
>= nr_to_scan
)
540 if (unlikely(!PageLRU(page
)))
543 if (PageActive(page
) && !active
) {
544 __mem_cgroup_move_lists(pc
, true);
547 if (!PageActive(page
) && active
) {
548 __mem_cgroup_move_lists(pc
, false);
553 list_move(&pc
->lru
, &pc_list
);
555 if (__isolate_lru_page(page
, mode
) == 0) {
556 list_move(&page
->lru
, dst
);
561 list_splice(&pc_list
, src
);
562 spin_unlock(&mz
->lru_lock
);
569 * Charge the memory controller for page usage.
571 * 0 if the charge was successful
572 * < 0 if the cgroup is over its limit
574 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
575 gfp_t gfp_mask
, enum charge_type ctype
)
577 struct mem_cgroup
*mem
;
578 struct page_cgroup
*pc
;
580 unsigned long nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
581 struct mem_cgroup_per_zone
*mz
;
584 * Should page_cgroup's go to their own slab?
585 * One could optimize the performance of the charging routine
586 * by saving a bit in the page_flags and using it as a lock
587 * to see if the cgroup page already has a page_cgroup associated
592 lock_page_cgroup(page
);
593 pc
= page_get_page_cgroup(page
);
595 * The page_cgroup exists and
596 * the page has already been accounted.
599 if (unlikely(!atomic_inc_not_zero(&pc
->ref_cnt
))) {
600 /* this page is under being uncharged ? */
601 unlock_page_cgroup(page
);
605 unlock_page_cgroup(page
);
609 unlock_page_cgroup(page
);
612 pc
= kzalloc(sizeof(struct page_cgroup
), gfp_mask
);
617 * We always charge the cgroup the mm_struct belongs to.
618 * The mm_struct's mem_cgroup changes on task migration if the
619 * thread group leader migrates. It's possible that mm is not
620 * set, if so charge the init_mm (happens for pagecache usage).
626 mem
= rcu_dereference(mm
->mem_cgroup
);
628 * For every charge from the cgroup, increment reference
635 * If we created the page_cgroup, we should free it on exceeding
638 while (res_counter_charge(&mem
->res
, PAGE_SIZE
)) {
639 if (!(gfp_mask
& __GFP_WAIT
))
642 if (try_to_free_mem_cgroup_pages(mem
, gfp_mask
))
646 * try_to_free_mem_cgroup_pages() might not give us a full
647 * picture of reclaim. Some pages are reclaimed and might be
648 * moved to swap cache or just unmapped from the cgroup.
649 * Check the limit again to see if the reclaim reduced the
650 * current usage of the cgroup before giving up
652 if (res_counter_check_under_limit(&mem
->res
))
656 mem_cgroup_out_of_memory(mem
, gfp_mask
);
659 congestion_wait(WRITE
, HZ
/10);
662 atomic_set(&pc
->ref_cnt
, 1);
663 pc
->mem_cgroup
= mem
;
665 pc
->flags
= PAGE_CGROUP_FLAG_ACTIVE
;
666 if (ctype
== MEM_CGROUP_CHARGE_TYPE_CACHE
)
667 pc
->flags
|= PAGE_CGROUP_FLAG_CACHE
;
669 if (!page
|| page_cgroup_assign_new_page_cgroup(page
, pc
)) {
671 * Another charge has been added to this page already.
672 * We take lock_page_cgroup(page) again and read
673 * page->cgroup, increment refcnt.... just retry is OK.
675 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
683 mz
= page_cgroup_zoneinfo(pc
);
684 spin_lock_irqsave(&mz
->lru_lock
, flags
);
685 /* Update statistics vector */
686 __mem_cgroup_add_list(pc
);
687 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
698 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
,
701 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
702 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
706 * See if the cached pages should be charged at all?
708 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
715 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
,
716 MEM_CGROUP_CHARGE_TYPE_CACHE
);
721 * Uncharging is always a welcome operation, we never complain, simply
722 * uncharge. This routine should be called with lock_page_cgroup held
724 void mem_cgroup_uncharge(struct page_cgroup
*pc
)
726 struct mem_cgroup
*mem
;
727 struct mem_cgroup_per_zone
*mz
;
732 * Check if our page_cgroup is valid
737 if (atomic_dec_and_test(&pc
->ref_cnt
)) {
739 mz
= page_cgroup_zoneinfo(pc
);
741 * get page->cgroup and clear it under lock.
742 * force_empty can drop page->cgroup without checking refcnt.
744 unlock_page_cgroup(page
);
745 if (clear_page_cgroup(page
, pc
) == pc
) {
746 mem
= pc
->mem_cgroup
;
748 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
749 spin_lock_irqsave(&mz
->lru_lock
, flags
);
750 __mem_cgroup_remove_list(pc
);
751 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
754 lock_page_cgroup(page
);
758 void mem_cgroup_uncharge_page(struct page
*page
)
760 lock_page_cgroup(page
);
761 mem_cgroup_uncharge(page_get_page_cgroup(page
));
762 unlock_page_cgroup(page
);
766 * Returns non-zero if a page (under migration) has valid page_cgroup member.
767 * Refcnt of page_cgroup is incremented.
770 int mem_cgroup_prepare_migration(struct page
*page
)
772 struct page_cgroup
*pc
;
774 lock_page_cgroup(page
);
775 pc
= page_get_page_cgroup(page
);
776 if (pc
&& atomic_inc_not_zero(&pc
->ref_cnt
))
778 unlock_page_cgroup(page
);
782 void mem_cgroup_end_migration(struct page
*page
)
784 struct page_cgroup
*pc
;
786 lock_page_cgroup(page
);
787 pc
= page_get_page_cgroup(page
);
788 mem_cgroup_uncharge(pc
);
789 unlock_page_cgroup(page
);
792 * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
793 * And no race with uncharge() routines because page_cgroup for *page*
794 * has extra one reference by mem_cgroup_prepare_migration.
797 void mem_cgroup_page_migration(struct page
*page
, struct page
*newpage
)
799 struct page_cgroup
*pc
;
800 struct mem_cgroup
*mem
;
802 struct mem_cgroup_per_zone
*mz
;
804 pc
= page_get_page_cgroup(page
);
807 mem
= pc
->mem_cgroup
;
808 mz
= page_cgroup_zoneinfo(pc
);
809 if (clear_page_cgroup(page
, pc
) != pc
)
811 spin_lock_irqsave(&mz
->lru_lock
, flags
);
813 __mem_cgroup_remove_list(pc
);
814 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
817 lock_page_cgroup(newpage
);
818 page_assign_page_cgroup(newpage
, pc
);
819 unlock_page_cgroup(newpage
);
821 mz
= page_cgroup_zoneinfo(pc
);
822 spin_lock_irqsave(&mz
->lru_lock
, flags
);
823 __mem_cgroup_add_list(pc
);
824 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
829 * This routine traverse page_cgroup in given list and drop them all.
830 * This routine ignores page_cgroup->ref_cnt.
831 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
833 #define FORCE_UNCHARGE_BATCH (128)
835 mem_cgroup_force_empty_list(struct mem_cgroup
*mem
,
836 struct mem_cgroup_per_zone
*mz
,
839 struct page_cgroup
*pc
;
843 struct list_head
*list
;
846 list
= &mz
->active_list
;
848 list
= &mz
->inactive_list
;
850 if (list_empty(list
))
853 count
= FORCE_UNCHARGE_BATCH
;
854 spin_lock_irqsave(&mz
->lru_lock
, flags
);
856 while (--count
&& !list_empty(list
)) {
857 pc
= list_entry(list
->prev
, struct page_cgroup
, lru
);
859 /* Avoid race with charge */
860 atomic_set(&pc
->ref_cnt
, 0);
861 if (clear_page_cgroup(page
, pc
) == pc
) {
863 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
864 __mem_cgroup_remove_list(pc
);
866 } else /* being uncharged ? ...do relax */
869 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
870 if (!list_empty(list
)) {
878 * make mem_cgroup's charge to be 0 if there is no task.
879 * This enables deleting this mem_cgroup.
882 int mem_cgroup_force_empty(struct mem_cgroup
*mem
)
888 * page reclaim code (kswapd etc..) will move pages between
889 ` * active_list <-> inactive_list while we don't take a lock.
890 * So, we have to do loop here until all lists are empty.
892 while (mem
->res
.usage
> 0) {
893 if (atomic_read(&mem
->css
.cgroup
->count
) > 0)
895 for_each_node_state(node
, N_POSSIBLE
)
896 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
897 struct mem_cgroup_per_zone
*mz
;
898 mz
= mem_cgroup_zoneinfo(mem
, node
, zid
);
899 /* drop all page_cgroup in active_list */
900 mem_cgroup_force_empty_list(mem
, mz
, 1);
901 /* drop all page_cgroup in inactive_list */
902 mem_cgroup_force_empty_list(mem
, mz
, 0);
913 int mem_cgroup_write_strategy(char *buf
, unsigned long long *tmp
)
915 *tmp
= memparse(buf
, &buf
);
920 * Round up the value to the closest page size
922 *tmp
= ((*tmp
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
) << PAGE_SHIFT
;
926 static ssize_t
mem_cgroup_read(struct cgroup
*cont
,
927 struct cftype
*cft
, struct file
*file
,
928 char __user
*userbuf
, size_t nbytes
, loff_t
*ppos
)
930 return res_counter_read(&mem_cgroup_from_cont(cont
)->res
,
931 cft
->private, userbuf
, nbytes
, ppos
,
935 static ssize_t
mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
936 struct file
*file
, const char __user
*userbuf
,
937 size_t nbytes
, loff_t
*ppos
)
939 return res_counter_write(&mem_cgroup_from_cont(cont
)->res
,
940 cft
->private, userbuf
, nbytes
, ppos
,
941 mem_cgroup_write_strategy
);
944 static ssize_t
mem_force_empty_write(struct cgroup
*cont
,
945 struct cftype
*cft
, struct file
*file
,
946 const char __user
*userbuf
,
947 size_t nbytes
, loff_t
*ppos
)
949 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
951 ret
= mem_cgroup_force_empty(mem
);
958 * Note: This should be removed if cgroup supports write-only file.
961 static ssize_t
mem_force_empty_read(struct cgroup
*cont
,
963 struct file
*file
, char __user
*userbuf
,
964 size_t nbytes
, loff_t
*ppos
)
970 static const struct mem_cgroup_stat_desc
{
973 } mem_cgroup_stat_desc
[] = {
974 [MEM_CGROUP_STAT_CACHE
] = { "cache", PAGE_SIZE
, },
975 [MEM_CGROUP_STAT_RSS
] = { "rss", PAGE_SIZE
, },
978 static int mem_control_stat_show(struct seq_file
*m
, void *arg
)
980 struct cgroup
*cont
= m
->private;
981 struct mem_cgroup
*mem_cont
= mem_cgroup_from_cont(cont
);
982 struct mem_cgroup_stat
*stat
= &mem_cont
->stat
;
985 for (i
= 0; i
< ARRAY_SIZE(stat
->cpustat
[0].count
); i
++) {
988 val
= mem_cgroup_read_stat(stat
, i
);
989 val
*= mem_cgroup_stat_desc
[i
].unit
;
990 seq_printf(m
, "%s %lld\n", mem_cgroup_stat_desc
[i
].msg
,
993 /* showing # of active pages */
995 unsigned long active
, inactive
;
997 inactive
= mem_cgroup_get_all_zonestat(mem_cont
,
998 MEM_CGROUP_ZSTAT_INACTIVE
);
999 active
= mem_cgroup_get_all_zonestat(mem_cont
,
1000 MEM_CGROUP_ZSTAT_ACTIVE
);
1001 seq_printf(m
, "active %ld\n", (active
) * PAGE_SIZE
);
1002 seq_printf(m
, "inactive %ld\n", (inactive
) * PAGE_SIZE
);
1007 static const struct file_operations mem_control_stat_file_operations
= {
1009 .llseek
= seq_lseek
,
1010 .release
= single_release
,
1013 static int mem_control_stat_open(struct inode
*unused
, struct file
*file
)
1016 struct cgroup
*cont
= file
->f_dentry
->d_parent
->d_fsdata
;
1018 file
->f_op
= &mem_control_stat_file_operations
;
1019 return single_open(file
, mem_control_stat_show
, cont
);
1024 static struct cftype mem_cgroup_files
[] = {
1026 .name
= "usage_in_bytes",
1027 .private = RES_USAGE
,
1028 .read
= mem_cgroup_read
,
1031 .name
= "limit_in_bytes",
1032 .private = RES_LIMIT
,
1033 .write
= mem_cgroup_write
,
1034 .read
= mem_cgroup_read
,
1038 .private = RES_FAILCNT
,
1039 .read
= mem_cgroup_read
,
1042 .name
= "force_empty",
1043 .write
= mem_force_empty_write
,
1044 .read
= mem_force_empty_read
,
1048 .open
= mem_control_stat_open
,
1052 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
1054 struct mem_cgroup_per_node
*pn
;
1055 struct mem_cgroup_per_zone
*mz
;
1058 * This routine is called against possible nodes.
1059 * But it's BUG to call kmalloc() against offline node.
1061 * TODO: this routine can waste much memory for nodes which will
1062 * never be onlined. It's better to use memory hotplug callback
1065 if (node_state(node
, N_HIGH_MEMORY
))
1066 pn
= kmalloc_node(sizeof(*pn
), GFP_KERNEL
, node
);
1068 pn
= kmalloc(sizeof(*pn
), GFP_KERNEL
);
1072 mem
->info
.nodeinfo
[node
] = pn
;
1073 memset(pn
, 0, sizeof(*pn
));
1075 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
1076 mz
= &pn
->zoneinfo
[zone
];
1077 INIT_LIST_HEAD(&mz
->active_list
);
1078 INIT_LIST_HEAD(&mz
->inactive_list
);
1079 spin_lock_init(&mz
->lru_lock
);
1084 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
1086 kfree(mem
->info
.nodeinfo
[node
]);
1090 static struct mem_cgroup init_mem_cgroup
;
1092 static struct cgroup_subsys_state
*
1093 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1095 struct mem_cgroup
*mem
;
1098 if (unlikely((cont
->parent
) == NULL
)) {
1099 mem
= &init_mem_cgroup
;
1100 init_mm
.mem_cgroup
= mem
;
1102 mem
= kzalloc(sizeof(struct mem_cgroup
), GFP_KERNEL
);
1105 return ERR_PTR(-ENOMEM
);
1107 res_counter_init(&mem
->res
);
1109 memset(&mem
->info
, 0, sizeof(mem
->info
));
1111 for_each_node_state(node
, N_POSSIBLE
)
1112 if (alloc_mem_cgroup_per_zone_info(mem
, node
))
1117 for_each_node_state(node
, N_POSSIBLE
)
1118 free_mem_cgroup_per_zone_info(mem
, node
);
1119 if (cont
->parent
!= NULL
)
1121 return ERR_PTR(-ENOMEM
);
1124 static void mem_cgroup_pre_destroy(struct cgroup_subsys
*ss
,
1125 struct cgroup
*cont
)
1127 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
1128 mem_cgroup_force_empty(mem
);
1131 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
1132 struct cgroup
*cont
)
1135 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
1137 for_each_node_state(node
, N_POSSIBLE
)
1138 free_mem_cgroup_per_zone_info(mem
, node
);
1140 kfree(mem_cgroup_from_cont(cont
));
1143 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
1144 struct cgroup
*cont
)
1146 return cgroup_add_files(cont
, ss
, mem_cgroup_files
,
1147 ARRAY_SIZE(mem_cgroup_files
));
1150 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
1151 struct cgroup
*cont
,
1152 struct cgroup
*old_cont
,
1153 struct task_struct
*p
)
1155 struct mm_struct
*mm
;
1156 struct mem_cgroup
*mem
, *old_mem
;
1158 mm
= get_task_mm(p
);
1162 mem
= mem_cgroup_from_cont(cont
);
1163 old_mem
= mem_cgroup_from_cont(old_cont
);
1169 * Only thread group leaders are allowed to migrate, the mm_struct is
1170 * in effect owned by the leader
1172 if (p
->tgid
!= p
->pid
)
1176 rcu_assign_pointer(mm
->mem_cgroup
, mem
);
1177 css_put(&old_mem
->css
);
1184 struct cgroup_subsys mem_cgroup_subsys
= {
1186 .subsys_id
= mem_cgroup_subsys_id
,
1187 .create
= mem_cgroup_create
,
1188 .pre_destroy
= mem_cgroup_pre_destroy
,
1189 .destroy
= mem_cgroup_destroy
,
1190 .populate
= mem_cgroup_populate
,
1191 .attach
= mem_cgroup_move_task
,