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 at least two
144 * byte aligned (based on comments from Nick Piggin). But since
145 * bit_spin_lock doesn't actually set that lock bit in a non-debug
146 * uniprocessor kernel, we should avoid setting it here too.
148 #define PAGE_CGROUP_LOCK_BIT 0x0
149 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
150 #define PAGE_CGROUP_LOCK (1 << PAGE_CGROUP_LOCK_BIT)
152 #define PAGE_CGROUP_LOCK 0x0
156 * A page_cgroup page is associated with every page descriptor. The
157 * page_cgroup helps us identify information about the cgroup
160 struct list_head lru
; /* per cgroup LRU list */
162 struct mem_cgroup
*mem_cgroup
;
163 atomic_t ref_cnt
; /* Helpful when pages move b/w */
164 /* mapped and cached states */
167 #define PAGE_CGROUP_FLAG_CACHE (0x1) /* charged as cache */
168 #define PAGE_CGROUP_FLAG_ACTIVE (0x2) /* page is active in this cgroup */
170 static inline int page_cgroup_nid(struct page_cgroup
*pc
)
172 return page_to_nid(pc
->page
);
175 static inline enum zone_type
page_cgroup_zid(struct page_cgroup
*pc
)
177 return page_zonenum(pc
->page
);
181 MEM_CGROUP_TYPE_UNSPEC
= 0,
182 MEM_CGROUP_TYPE_MAPPED
,
183 MEM_CGROUP_TYPE_CACHED
,
189 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
190 MEM_CGROUP_CHARGE_TYPE_MAPPED
,
195 * Always modified under lru lock. Then, not necessary to preempt_disable()
197 static void mem_cgroup_charge_statistics(struct mem_cgroup
*mem
, int flags
,
200 int val
= (charge
)? 1 : -1;
201 struct mem_cgroup_stat
*stat
= &mem
->stat
;
202 VM_BUG_ON(!irqs_disabled());
204 if (flags
& PAGE_CGROUP_FLAG_CACHE
)
205 __mem_cgroup_stat_add_safe(stat
,
206 MEM_CGROUP_STAT_CACHE
, val
);
208 __mem_cgroup_stat_add_safe(stat
, MEM_CGROUP_STAT_RSS
, val
);
211 static inline struct mem_cgroup_per_zone
*
212 mem_cgroup_zoneinfo(struct mem_cgroup
*mem
, int nid
, int zid
)
214 BUG_ON(!mem
->info
.nodeinfo
[nid
]);
215 return &mem
->info
.nodeinfo
[nid
]->zoneinfo
[zid
];
218 static inline struct mem_cgroup_per_zone
*
219 page_cgroup_zoneinfo(struct page_cgroup
*pc
)
221 struct mem_cgroup
*mem
= pc
->mem_cgroup
;
222 int nid
= page_cgroup_nid(pc
);
223 int zid
= page_cgroup_zid(pc
);
225 return mem_cgroup_zoneinfo(mem
, nid
, zid
);
228 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup
*mem
,
229 enum mem_cgroup_zstat_index idx
)
232 struct mem_cgroup_per_zone
*mz
;
235 for_each_online_node(nid
)
236 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
237 mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
238 total
+= MEM_CGROUP_ZSTAT(mz
, idx
);
243 static struct mem_cgroup init_mem_cgroup
;
246 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
248 return container_of(cgroup_subsys_state(cont
,
249 mem_cgroup_subsys_id
), struct mem_cgroup
,
254 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
256 return container_of(task_subsys_state(p
, mem_cgroup_subsys_id
),
257 struct mem_cgroup
, css
);
260 void mm_init_cgroup(struct mm_struct
*mm
, struct task_struct
*p
)
262 struct mem_cgroup
*mem
;
264 mem
= mem_cgroup_from_task(p
);
266 mm
->mem_cgroup
= mem
;
269 void mm_free_cgroup(struct mm_struct
*mm
)
271 css_put(&mm
->mem_cgroup
->css
);
274 static inline int page_cgroup_locked(struct page
*page
)
276 return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT
,
280 static void page_assign_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
282 VM_BUG_ON(!page_cgroup_locked(page
));
283 page
->page_cgroup
= ((unsigned long)pc
| PAGE_CGROUP_LOCK
);
286 struct page_cgroup
*page_get_page_cgroup(struct page
*page
)
288 return (struct page_cgroup
*)
289 (page
->page_cgroup
& ~PAGE_CGROUP_LOCK
);
292 static void __always_inline
lock_page_cgroup(struct page
*page
)
294 bit_spin_lock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
295 VM_BUG_ON(!page_cgroup_locked(page
));
298 static void __always_inline
unlock_page_cgroup(struct page
*page
)
300 bit_spin_unlock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
304 * Tie new page_cgroup to struct page under lock_page_cgroup()
305 * This can fail if the page has been tied to a page_cgroup.
306 * If success, returns 0.
308 static int page_cgroup_assign_new_page_cgroup(struct page
*page
,
309 struct page_cgroup
*pc
)
313 lock_page_cgroup(page
);
314 if (!page_get_page_cgroup(page
))
315 page_assign_page_cgroup(page
, pc
);
316 else /* A page is tied to other pc. */
318 unlock_page_cgroup(page
);
323 * Clear page->page_cgroup member under lock_page_cgroup().
324 * If given "pc" value is different from one page->page_cgroup,
325 * page->cgroup is not cleared.
326 * Returns a value of page->page_cgroup at lock taken.
327 * A can can detect failure of clearing by following
328 * clear_page_cgroup(page, pc) == pc
331 static struct page_cgroup
*clear_page_cgroup(struct page
*page
,
332 struct page_cgroup
*pc
)
334 struct page_cgroup
*ret
;
336 lock_page_cgroup(page
);
337 ret
= page_get_page_cgroup(page
);
338 if (likely(ret
== pc
))
339 page_assign_page_cgroup(page
, NULL
);
340 unlock_page_cgroup(page
);
344 static void __mem_cgroup_remove_list(struct page_cgroup
*pc
)
346 int from
= pc
->flags
& PAGE_CGROUP_FLAG_ACTIVE
;
347 struct mem_cgroup_per_zone
*mz
= page_cgroup_zoneinfo(pc
);
350 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) -= 1;
352 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) -= 1;
354 mem_cgroup_charge_statistics(pc
->mem_cgroup
, pc
->flags
, false);
355 list_del_init(&pc
->lru
);
358 static void __mem_cgroup_add_list(struct page_cgroup
*pc
)
360 int to
= pc
->flags
& PAGE_CGROUP_FLAG_ACTIVE
;
361 struct mem_cgroup_per_zone
*mz
= page_cgroup_zoneinfo(pc
);
364 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) += 1;
365 list_add(&pc
->lru
, &mz
->inactive_list
);
367 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) += 1;
368 list_add(&pc
->lru
, &mz
->active_list
);
370 mem_cgroup_charge_statistics(pc
->mem_cgroup
, pc
->flags
, true);
373 static void __mem_cgroup_move_lists(struct page_cgroup
*pc
, bool active
)
375 int from
= pc
->flags
& PAGE_CGROUP_FLAG_ACTIVE
;
376 struct mem_cgroup_per_zone
*mz
= page_cgroup_zoneinfo(pc
);
379 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) -= 1;
381 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) -= 1;
384 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
) += 1;
385 pc
->flags
|= PAGE_CGROUP_FLAG_ACTIVE
;
386 list_move(&pc
->lru
, &mz
->active_list
);
388 MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
) += 1;
389 pc
->flags
&= ~PAGE_CGROUP_FLAG_ACTIVE
;
390 list_move(&pc
->lru
, &mz
->inactive_list
);
394 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
399 ret
= task
->mm
&& mm_match_cgroup(task
->mm
, mem
);
405 * This routine assumes that the appropriate zone's lru lock is already held
407 void mem_cgroup_move_lists(struct page
*page
, bool active
)
409 struct page_cgroup
*pc
;
410 struct mem_cgroup_per_zone
*mz
;
413 pc
= page_get_page_cgroup(page
);
417 mz
= page_cgroup_zoneinfo(pc
);
418 spin_lock_irqsave(&mz
->lru_lock
, flags
);
419 __mem_cgroup_move_lists(pc
, active
);
420 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
424 * Calculate mapped_ratio under memory controller. This will be used in
425 * vmscan.c for deteremining we have to reclaim mapped pages.
427 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup
*mem
)
432 * usage is recorded in bytes. But, here, we assume the number of
433 * physical pages can be represented by "long" on any arch.
435 total
= (long) (mem
->res
.usage
>> PAGE_SHIFT
) + 1L;
436 rss
= (long)mem_cgroup_read_stat(&mem
->stat
, MEM_CGROUP_STAT_RSS
);
437 return (int)((rss
* 100L) / total
);
440 * This function is called from vmscan.c. In page reclaiming loop. balance
441 * between active and inactive list is calculated. For memory controller
442 * page reclaiming, we should use using mem_cgroup's imbalance rather than
443 * zone's global lru imbalance.
445 long mem_cgroup_reclaim_imbalance(struct mem_cgroup
*mem
)
447 unsigned long active
, inactive
;
448 /* active and inactive are the number of pages. 'long' is ok.*/
449 active
= mem_cgroup_get_all_zonestat(mem
, MEM_CGROUP_ZSTAT_ACTIVE
);
450 inactive
= mem_cgroup_get_all_zonestat(mem
, MEM_CGROUP_ZSTAT_INACTIVE
);
451 return (long) (active
/ (inactive
+ 1));
455 * prev_priority control...this will be used in memory reclaim path.
457 int mem_cgroup_get_reclaim_priority(struct mem_cgroup
*mem
)
459 return mem
->prev_priority
;
462 void mem_cgroup_note_reclaim_priority(struct mem_cgroup
*mem
, int priority
)
464 if (priority
< mem
->prev_priority
)
465 mem
->prev_priority
= priority
;
468 void mem_cgroup_record_reclaim_priority(struct mem_cgroup
*mem
, int priority
)
470 mem
->prev_priority
= priority
;
474 * Calculate # of pages to be scanned in this priority/zone.
477 * priority starts from "DEF_PRIORITY" and decremented in each loop.
478 * (see include/linux/mmzone.h)
481 long mem_cgroup_calc_reclaim_active(struct mem_cgroup
*mem
,
482 struct zone
*zone
, int priority
)
485 int nid
= zone
->zone_pgdat
->node_id
;
486 int zid
= zone_idx(zone
);
487 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
489 nr_active
= MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_ACTIVE
);
490 return (nr_active
>> priority
);
493 long mem_cgroup_calc_reclaim_inactive(struct mem_cgroup
*mem
,
494 struct zone
*zone
, int priority
)
497 int nid
= zone
->zone_pgdat
->node_id
;
498 int zid
= zone_idx(zone
);
499 struct mem_cgroup_per_zone
*mz
= mem_cgroup_zoneinfo(mem
, nid
, zid
);
501 nr_inactive
= MEM_CGROUP_ZSTAT(mz
, MEM_CGROUP_ZSTAT_INACTIVE
);
503 return (nr_inactive
>> priority
);
506 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
507 struct list_head
*dst
,
508 unsigned long *scanned
, int order
,
509 int mode
, struct zone
*z
,
510 struct mem_cgroup
*mem_cont
,
513 unsigned long nr_taken
= 0;
517 struct list_head
*src
;
518 struct page_cgroup
*pc
, *tmp
;
519 int nid
= z
->zone_pgdat
->node_id
;
520 int zid
= zone_idx(z
);
521 struct mem_cgroup_per_zone
*mz
;
523 mz
= mem_cgroup_zoneinfo(mem_cont
, nid
, zid
);
525 src
= &mz
->active_list
;
527 src
= &mz
->inactive_list
;
530 spin_lock(&mz
->lru_lock
);
532 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
533 if (scan
>= nr_to_scan
)
537 if (unlikely(!PageLRU(page
)))
540 if (PageActive(page
) && !active
) {
541 __mem_cgroup_move_lists(pc
, true);
544 if (!PageActive(page
) && active
) {
545 __mem_cgroup_move_lists(pc
, false);
550 list_move(&pc
->lru
, &pc_list
);
552 if (__isolate_lru_page(page
, mode
) == 0) {
553 list_move(&page
->lru
, dst
);
558 list_splice(&pc_list
, src
);
559 spin_unlock(&mz
->lru_lock
);
566 * Charge the memory controller for page usage.
568 * 0 if the charge was successful
569 * < 0 if the cgroup is over its limit
571 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
572 gfp_t gfp_mask
, enum charge_type ctype
)
574 struct mem_cgroup
*mem
;
575 struct page_cgroup
*pc
;
577 unsigned long nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
578 struct mem_cgroup_per_zone
*mz
;
581 * Should page_cgroup's go to their own slab?
582 * One could optimize the performance of the charging routine
583 * by saving a bit in the page_flags and using it as a lock
584 * to see if the cgroup page already has a page_cgroup associated
589 lock_page_cgroup(page
);
590 pc
= page_get_page_cgroup(page
);
592 * The page_cgroup exists and
593 * the page has already been accounted.
596 if (unlikely(!atomic_inc_not_zero(&pc
->ref_cnt
))) {
597 /* this page is under being uncharged ? */
598 unlock_page_cgroup(page
);
602 unlock_page_cgroup(page
);
606 unlock_page_cgroup(page
);
609 pc
= kzalloc(sizeof(struct page_cgroup
), gfp_mask
);
614 * We always charge the cgroup the mm_struct belongs to.
615 * The mm_struct's mem_cgroup changes on task migration if the
616 * thread group leader migrates. It's possible that mm is not
617 * set, if so charge the init_mm (happens for pagecache usage).
623 mem
= rcu_dereference(mm
->mem_cgroup
);
625 * For every charge from the cgroup, increment reference
632 * If we created the page_cgroup, we should free it on exceeding
635 while (res_counter_charge(&mem
->res
, PAGE_SIZE
)) {
636 if (!(gfp_mask
& __GFP_WAIT
))
639 if (try_to_free_mem_cgroup_pages(mem
, gfp_mask
))
643 * try_to_free_mem_cgroup_pages() might not give us a full
644 * picture of reclaim. Some pages are reclaimed and might be
645 * moved to swap cache or just unmapped from the cgroup.
646 * Check the limit again to see if the reclaim reduced the
647 * current usage of the cgroup before giving up
649 if (res_counter_check_under_limit(&mem
->res
))
653 mem_cgroup_out_of_memory(mem
, gfp_mask
);
656 congestion_wait(WRITE
, HZ
/10);
659 atomic_set(&pc
->ref_cnt
, 1);
660 pc
->mem_cgroup
= mem
;
662 pc
->flags
= PAGE_CGROUP_FLAG_ACTIVE
;
663 if (ctype
== MEM_CGROUP_CHARGE_TYPE_CACHE
)
664 pc
->flags
|= PAGE_CGROUP_FLAG_CACHE
;
666 if (!page
|| page_cgroup_assign_new_page_cgroup(page
, pc
)) {
668 * Another charge has been added to this page already.
669 * We take lock_page_cgroup(page) again and read
670 * page->cgroup, increment refcnt.... just retry is OK.
672 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
680 mz
= page_cgroup_zoneinfo(pc
);
681 spin_lock_irqsave(&mz
->lru_lock
, flags
);
682 /* Update statistics vector */
683 __mem_cgroup_add_list(pc
);
684 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
695 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
,
698 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
699 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
703 * See if the cached pages should be charged at all?
705 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
712 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
,
713 MEM_CGROUP_CHARGE_TYPE_CACHE
);
718 * Uncharging is always a welcome operation, we never complain, simply
719 * uncharge. This routine should be called with lock_page_cgroup held
721 void mem_cgroup_uncharge(struct page_cgroup
*pc
)
723 struct mem_cgroup
*mem
;
724 struct mem_cgroup_per_zone
*mz
;
729 * Check if our page_cgroup is valid
734 if (atomic_dec_and_test(&pc
->ref_cnt
)) {
736 mz
= page_cgroup_zoneinfo(pc
);
738 * get page->cgroup and clear it under lock.
739 * force_empty can drop page->cgroup without checking refcnt.
741 unlock_page_cgroup(page
);
742 if (clear_page_cgroup(page
, pc
) == pc
) {
743 mem
= pc
->mem_cgroup
;
745 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
746 spin_lock_irqsave(&mz
->lru_lock
, flags
);
747 __mem_cgroup_remove_list(pc
);
748 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
751 lock_page_cgroup(page
);
755 void mem_cgroup_uncharge_page(struct page
*page
)
757 lock_page_cgroup(page
);
758 mem_cgroup_uncharge(page_get_page_cgroup(page
));
759 unlock_page_cgroup(page
);
763 * Returns non-zero if a page (under migration) has valid page_cgroup member.
764 * Refcnt of page_cgroup is incremented.
767 int mem_cgroup_prepare_migration(struct page
*page
)
769 struct page_cgroup
*pc
;
771 lock_page_cgroup(page
);
772 pc
= page_get_page_cgroup(page
);
773 if (pc
&& atomic_inc_not_zero(&pc
->ref_cnt
))
775 unlock_page_cgroup(page
);
779 void mem_cgroup_end_migration(struct page
*page
)
781 struct page_cgroup
*pc
;
783 lock_page_cgroup(page
);
784 pc
= page_get_page_cgroup(page
);
785 mem_cgroup_uncharge(pc
);
786 unlock_page_cgroup(page
);
789 * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
790 * And no race with uncharge() routines because page_cgroup for *page*
791 * has extra one reference by mem_cgroup_prepare_migration.
794 void mem_cgroup_page_migration(struct page
*page
, struct page
*newpage
)
796 struct page_cgroup
*pc
;
797 struct mem_cgroup
*mem
;
799 struct mem_cgroup_per_zone
*mz
;
801 pc
= page_get_page_cgroup(page
);
804 mem
= pc
->mem_cgroup
;
805 mz
= page_cgroup_zoneinfo(pc
);
806 if (clear_page_cgroup(page
, pc
) != pc
)
808 spin_lock_irqsave(&mz
->lru_lock
, flags
);
810 __mem_cgroup_remove_list(pc
);
811 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
814 lock_page_cgroup(newpage
);
815 page_assign_page_cgroup(newpage
, pc
);
816 unlock_page_cgroup(newpage
);
818 mz
= page_cgroup_zoneinfo(pc
);
819 spin_lock_irqsave(&mz
->lru_lock
, flags
);
820 __mem_cgroup_add_list(pc
);
821 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
826 * This routine traverse page_cgroup in given list and drop them all.
827 * This routine ignores page_cgroup->ref_cnt.
828 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
830 #define FORCE_UNCHARGE_BATCH (128)
832 mem_cgroup_force_empty_list(struct mem_cgroup
*mem
,
833 struct mem_cgroup_per_zone
*mz
,
836 struct page_cgroup
*pc
;
840 struct list_head
*list
;
843 list
= &mz
->active_list
;
845 list
= &mz
->inactive_list
;
847 if (list_empty(list
))
850 count
= FORCE_UNCHARGE_BATCH
;
851 spin_lock_irqsave(&mz
->lru_lock
, flags
);
853 while (--count
&& !list_empty(list
)) {
854 pc
= list_entry(list
->prev
, struct page_cgroup
, lru
);
856 /* Avoid race with charge */
857 atomic_set(&pc
->ref_cnt
, 0);
858 if (clear_page_cgroup(page
, pc
) == pc
) {
860 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
861 __mem_cgroup_remove_list(pc
);
863 } else /* being uncharged ? ...do relax */
866 spin_unlock_irqrestore(&mz
->lru_lock
, flags
);
867 if (!list_empty(list
)) {
875 * make mem_cgroup's charge to be 0 if there is no task.
876 * This enables deleting this mem_cgroup.
879 int mem_cgroup_force_empty(struct mem_cgroup
*mem
)
885 * page reclaim code (kswapd etc..) will move pages between
886 ` * active_list <-> inactive_list while we don't take a lock.
887 * So, we have to do loop here until all lists are empty.
889 while (mem
->res
.usage
> 0) {
890 if (atomic_read(&mem
->css
.cgroup
->count
) > 0)
892 for_each_node_state(node
, N_POSSIBLE
)
893 for (zid
= 0; zid
< MAX_NR_ZONES
; zid
++) {
894 struct mem_cgroup_per_zone
*mz
;
895 mz
= mem_cgroup_zoneinfo(mem
, node
, zid
);
896 /* drop all page_cgroup in active_list */
897 mem_cgroup_force_empty_list(mem
, mz
, 1);
898 /* drop all page_cgroup in inactive_list */
899 mem_cgroup_force_empty_list(mem
, mz
, 0);
910 int mem_cgroup_write_strategy(char *buf
, unsigned long long *tmp
)
912 *tmp
= memparse(buf
, &buf
);
917 * Round up the value to the closest page size
919 *tmp
= ((*tmp
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
) << PAGE_SHIFT
;
923 static ssize_t
mem_cgroup_read(struct cgroup
*cont
,
924 struct cftype
*cft
, struct file
*file
,
925 char __user
*userbuf
, size_t nbytes
, loff_t
*ppos
)
927 return res_counter_read(&mem_cgroup_from_cont(cont
)->res
,
928 cft
->private, userbuf
, nbytes
, ppos
,
932 static ssize_t
mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
933 struct file
*file
, const char __user
*userbuf
,
934 size_t nbytes
, loff_t
*ppos
)
936 return res_counter_write(&mem_cgroup_from_cont(cont
)->res
,
937 cft
->private, userbuf
, nbytes
, ppos
,
938 mem_cgroup_write_strategy
);
941 static ssize_t
mem_force_empty_write(struct cgroup
*cont
,
942 struct cftype
*cft
, struct file
*file
,
943 const char __user
*userbuf
,
944 size_t nbytes
, loff_t
*ppos
)
946 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
948 ret
= mem_cgroup_force_empty(mem
);
955 * Note: This should be removed if cgroup supports write-only file.
958 static ssize_t
mem_force_empty_read(struct cgroup
*cont
,
960 struct file
*file
, char __user
*userbuf
,
961 size_t nbytes
, loff_t
*ppos
)
967 static const struct mem_cgroup_stat_desc
{
970 } mem_cgroup_stat_desc
[] = {
971 [MEM_CGROUP_STAT_CACHE
] = { "cache", PAGE_SIZE
, },
972 [MEM_CGROUP_STAT_RSS
] = { "rss", PAGE_SIZE
, },
975 static int mem_control_stat_show(struct seq_file
*m
, void *arg
)
977 struct cgroup
*cont
= m
->private;
978 struct mem_cgroup
*mem_cont
= mem_cgroup_from_cont(cont
);
979 struct mem_cgroup_stat
*stat
= &mem_cont
->stat
;
982 for (i
= 0; i
< ARRAY_SIZE(stat
->cpustat
[0].count
); i
++) {
985 val
= mem_cgroup_read_stat(stat
, i
);
986 val
*= mem_cgroup_stat_desc
[i
].unit
;
987 seq_printf(m
, "%s %lld\n", mem_cgroup_stat_desc
[i
].msg
,
990 /* showing # of active pages */
992 unsigned long active
, inactive
;
994 inactive
= mem_cgroup_get_all_zonestat(mem_cont
,
995 MEM_CGROUP_ZSTAT_INACTIVE
);
996 active
= mem_cgroup_get_all_zonestat(mem_cont
,
997 MEM_CGROUP_ZSTAT_ACTIVE
);
998 seq_printf(m
, "active %ld\n", (active
) * PAGE_SIZE
);
999 seq_printf(m
, "inactive %ld\n", (inactive
) * PAGE_SIZE
);
1004 static const struct file_operations mem_control_stat_file_operations
= {
1006 .llseek
= seq_lseek
,
1007 .release
= single_release
,
1010 static int mem_control_stat_open(struct inode
*unused
, struct file
*file
)
1013 struct cgroup
*cont
= file
->f_dentry
->d_parent
->d_fsdata
;
1015 file
->f_op
= &mem_control_stat_file_operations
;
1016 return single_open(file
, mem_control_stat_show
, cont
);
1021 static struct cftype mem_cgroup_files
[] = {
1023 .name
= "usage_in_bytes",
1024 .private = RES_USAGE
,
1025 .read
= mem_cgroup_read
,
1028 .name
= "limit_in_bytes",
1029 .private = RES_LIMIT
,
1030 .write
= mem_cgroup_write
,
1031 .read
= mem_cgroup_read
,
1035 .private = RES_FAILCNT
,
1036 .read
= mem_cgroup_read
,
1039 .name
= "force_empty",
1040 .write
= mem_force_empty_write
,
1041 .read
= mem_force_empty_read
,
1045 .open
= mem_control_stat_open
,
1049 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
1051 struct mem_cgroup_per_node
*pn
;
1052 struct mem_cgroup_per_zone
*mz
;
1055 * This routine is called against possible nodes.
1056 * But it's BUG to call kmalloc() against offline node.
1058 * TODO: this routine can waste much memory for nodes which will
1059 * never be onlined. It's better to use memory hotplug callback
1062 if (node_state(node
, N_HIGH_MEMORY
))
1063 pn
= kmalloc_node(sizeof(*pn
), GFP_KERNEL
, node
);
1065 pn
= kmalloc(sizeof(*pn
), GFP_KERNEL
);
1069 mem
->info
.nodeinfo
[node
] = pn
;
1070 memset(pn
, 0, sizeof(*pn
));
1072 for (zone
= 0; zone
< MAX_NR_ZONES
; zone
++) {
1073 mz
= &pn
->zoneinfo
[zone
];
1074 INIT_LIST_HEAD(&mz
->active_list
);
1075 INIT_LIST_HEAD(&mz
->inactive_list
);
1076 spin_lock_init(&mz
->lru_lock
);
1081 static void free_mem_cgroup_per_zone_info(struct mem_cgroup
*mem
, int node
)
1083 kfree(mem
->info
.nodeinfo
[node
]);
1087 static struct mem_cgroup init_mem_cgroup
;
1089 static struct cgroup_subsys_state
*
1090 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
1092 struct mem_cgroup
*mem
;
1095 if (unlikely((cont
->parent
) == NULL
)) {
1096 mem
= &init_mem_cgroup
;
1097 init_mm
.mem_cgroup
= mem
;
1099 mem
= kzalloc(sizeof(struct mem_cgroup
), GFP_KERNEL
);
1102 return ERR_PTR(-ENOMEM
);
1104 res_counter_init(&mem
->res
);
1106 memset(&mem
->info
, 0, sizeof(mem
->info
));
1108 for_each_node_state(node
, N_POSSIBLE
)
1109 if (alloc_mem_cgroup_per_zone_info(mem
, node
))
1114 for_each_node_state(node
, N_POSSIBLE
)
1115 free_mem_cgroup_per_zone_info(mem
, node
);
1116 if (cont
->parent
!= NULL
)
1118 return ERR_PTR(-ENOMEM
);
1121 static void mem_cgroup_pre_destroy(struct cgroup_subsys
*ss
,
1122 struct cgroup
*cont
)
1124 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
1125 mem_cgroup_force_empty(mem
);
1128 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
1129 struct cgroup
*cont
)
1132 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
1134 for_each_node_state(node
, N_POSSIBLE
)
1135 free_mem_cgroup_per_zone_info(mem
, node
);
1137 kfree(mem_cgroup_from_cont(cont
));
1140 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
1141 struct cgroup
*cont
)
1143 return cgroup_add_files(cont
, ss
, mem_cgroup_files
,
1144 ARRAY_SIZE(mem_cgroup_files
));
1147 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
1148 struct cgroup
*cont
,
1149 struct cgroup
*old_cont
,
1150 struct task_struct
*p
)
1152 struct mm_struct
*mm
;
1153 struct mem_cgroup
*mem
, *old_mem
;
1155 mm
= get_task_mm(p
);
1159 mem
= mem_cgroup_from_cont(cont
);
1160 old_mem
= mem_cgroup_from_cont(old_cont
);
1166 * Only thread group leaders are allowed to migrate, the mm_struct is
1167 * in effect owned by the leader
1169 if (p
->tgid
!= p
->pid
)
1173 rcu_assign_pointer(mm
->mem_cgroup
, mem
);
1174 css_put(&old_mem
->css
);
1181 struct cgroup_subsys mem_cgroup_subsys
= {
1183 .subsys_id
= mem_cgroup_subsys_id
,
1184 .create
= mem_cgroup_create
,
1185 .pre_destroy
= mem_cgroup_pre_destroy
,
1186 .destroy
= mem_cgroup_destroy
,
1187 .populate
= mem_cgroup_populate
,
1188 .attach
= mem_cgroup_move_task
,