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/page-flags.h>
25 #include <linux/backing-dev.h>
26 #include <linux/bit_spinlock.h>
27 #include <linux/rcupdate.h>
28 #include <linux/swap.h>
29 #include <linux/spinlock.h>
32 #include <asm/uaccess.h>
34 struct cgroup_subsys mem_cgroup_subsys
;
35 static const int MEM_CGROUP_RECLAIM_RETRIES
= 5;
38 * The memory controller data structure. The memory controller controls both
39 * page cache and RSS per cgroup. We would eventually like to provide
40 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
41 * to help the administrator determine what knobs to tune.
43 * TODO: Add a water mark for the memory controller. Reclaim will begin when
44 * we hit the water mark. May be even add a low water mark, such that
45 * no reclaim occurs from a cgroup at it's low water mark, this is
46 * a feature that will be implemented much later in the future.
49 struct cgroup_subsys_state css
;
51 * the counter to account for memory usage
53 struct res_counter res
;
55 * Per cgroup active and inactive list, similar to the
57 * TODO: Consider making these lists per zone
59 struct list_head active_list
;
60 struct list_head inactive_list
;
62 * spin_lock to protect the per cgroup LRU
65 unsigned long control_type
; /* control RSS or RSS+Pagecache */
69 * We use the lower bit of the page->page_cgroup pointer as a bit spin
70 * lock. We need to ensure that page->page_cgroup is atleast two
71 * byte aligned (based on comments from Nick Piggin)
73 #define PAGE_CGROUP_LOCK_BIT 0x0
74 #define PAGE_CGROUP_LOCK (1 << PAGE_CGROUP_LOCK_BIT)
77 * A page_cgroup page is associated with every page descriptor. The
78 * page_cgroup helps us identify information about the cgroup
81 struct list_head lru
; /* per cgroup LRU list */
83 struct mem_cgroup
*mem_cgroup
;
84 atomic_t ref_cnt
; /* Helpful when pages move b/w */
85 /* mapped and cached states */
88 #define PAGE_CGROUP_FLAG_CACHE (0x1) /* charged as cache */
89 #define PAGE_CGROUP_FLAG_ACTIVE (0x2) /* page is active in this cgroup */
92 MEM_CGROUP_TYPE_UNSPEC
= 0,
93 MEM_CGROUP_TYPE_MAPPED
,
94 MEM_CGROUP_TYPE_CACHED
,
100 MEM_CGROUP_CHARGE_TYPE_CACHE
= 0,
101 MEM_CGROUP_CHARGE_TYPE_MAPPED
,
104 static struct mem_cgroup init_mem_cgroup
;
107 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
109 return container_of(cgroup_subsys_state(cont
,
110 mem_cgroup_subsys_id
), struct mem_cgroup
,
115 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
117 return container_of(task_subsys_state(p
, mem_cgroup_subsys_id
),
118 struct mem_cgroup
, css
);
121 void mm_init_cgroup(struct mm_struct
*mm
, struct task_struct
*p
)
123 struct mem_cgroup
*mem
;
125 mem
= mem_cgroup_from_task(p
);
127 mm
->mem_cgroup
= mem
;
130 void mm_free_cgroup(struct mm_struct
*mm
)
132 css_put(&mm
->mem_cgroup
->css
);
135 static inline int page_cgroup_locked(struct page
*page
)
137 return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT
,
141 void page_assign_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
146 * While resetting the page_cgroup we might not hold the
147 * page_cgroup lock. free_hot_cold_page() is an example
151 VM_BUG_ON(!page_cgroup_locked(page
));
152 locked
= (page
->page_cgroup
& PAGE_CGROUP_LOCK
);
153 page
->page_cgroup
= ((unsigned long)pc
| locked
);
156 struct page_cgroup
*page_get_page_cgroup(struct page
*page
)
158 return (struct page_cgroup
*)
159 (page
->page_cgroup
& ~PAGE_CGROUP_LOCK
);
162 static void __always_inline
lock_page_cgroup(struct page
*page
)
164 bit_spin_lock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
165 VM_BUG_ON(!page_cgroup_locked(page
));
168 static void __always_inline
unlock_page_cgroup(struct page
*page
)
170 bit_spin_unlock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
174 * Tie new page_cgroup to struct page under lock_page_cgroup()
175 * This can fail if the page has been tied to a page_cgroup.
176 * If success, returns 0.
179 page_cgroup_assign_new_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
183 lock_page_cgroup(page
);
184 if (!page_get_page_cgroup(page
))
185 page_assign_page_cgroup(page
, pc
);
186 else /* A page is tied to other pc. */
188 unlock_page_cgroup(page
);
193 * Clear page->page_cgroup member under lock_page_cgroup().
194 * If given "pc" value is different from one page->page_cgroup,
195 * page->cgroup is not cleared.
196 * Returns a value of page->page_cgroup at lock taken.
197 * A can can detect failure of clearing by following
198 * clear_page_cgroup(page, pc) == pc
201 static inline struct page_cgroup
*
202 clear_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
204 struct page_cgroup
*ret
;
206 lock_page_cgroup(page
);
207 ret
= page_get_page_cgroup(page
);
208 if (likely(ret
== pc
))
209 page_assign_page_cgroup(page
, NULL
);
210 unlock_page_cgroup(page
);
215 static void __mem_cgroup_move_lists(struct page_cgroup
*pc
, bool active
)
218 pc
->flags
|= PAGE_CGROUP_FLAG_ACTIVE
;
219 list_move(&pc
->lru
, &pc
->mem_cgroup
->active_list
);
221 pc
->flags
&= ~PAGE_CGROUP_FLAG_ACTIVE
;
222 list_move(&pc
->lru
, &pc
->mem_cgroup
->inactive_list
);
226 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
231 ret
= task
->mm
&& mm_cgroup(task
->mm
) == mem
;
237 * This routine assumes that the appropriate zone's lru lock is already held
239 void mem_cgroup_move_lists(struct page_cgroup
*pc
, bool active
)
241 struct mem_cgroup
*mem
;
245 mem
= pc
->mem_cgroup
;
247 spin_lock(&mem
->lru_lock
);
248 __mem_cgroup_move_lists(pc
, active
);
249 spin_unlock(&mem
->lru_lock
);
252 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
253 struct list_head
*dst
,
254 unsigned long *scanned
, int order
,
255 int mode
, struct zone
*z
,
256 struct mem_cgroup
*mem_cont
,
259 unsigned long nr_taken
= 0;
263 struct list_head
*src
;
264 struct page_cgroup
*pc
, *tmp
;
267 src
= &mem_cont
->active_list
;
269 src
= &mem_cont
->inactive_list
;
271 spin_lock(&mem_cont
->lru_lock
);
273 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
274 if (scan
>= nr_to_scan
)
279 if (unlikely(!PageLRU(page
)))
282 if (PageActive(page
) && !active
) {
283 __mem_cgroup_move_lists(pc
, true);
286 if (!PageActive(page
) && active
) {
287 __mem_cgroup_move_lists(pc
, false);
293 * TODO: make the active/inactive lists per zone
295 if (page_zone(page
) != z
)
299 list_move(&pc
->lru
, &pc_list
);
301 if (__isolate_lru_page(page
, mode
) == 0) {
302 list_move(&page
->lru
, dst
);
307 list_splice(&pc_list
, src
);
308 spin_unlock(&mem_cont
->lru_lock
);
315 * Charge the memory controller for page usage.
317 * 0 if the charge was successful
318 * < 0 if the cgroup is over its limit
320 static int mem_cgroup_charge_common(struct page
*page
, struct mm_struct
*mm
,
321 gfp_t gfp_mask
, enum charge_type ctype
)
323 struct mem_cgroup
*mem
;
324 struct page_cgroup
*pc
;
326 unsigned long nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
329 * Should page_cgroup's go to their own slab?
330 * One could optimize the performance of the charging routine
331 * by saving a bit in the page_flags and using it as a lock
332 * to see if the cgroup page already has a page_cgroup associated
337 lock_page_cgroup(page
);
338 pc
= page_get_page_cgroup(page
);
340 * The page_cgroup exists and
341 * the page has already been accounted.
344 if (unlikely(!atomic_inc_not_zero(&pc
->ref_cnt
))) {
345 /* this page is under being uncharged ? */
346 unlock_page_cgroup(page
);
350 unlock_page_cgroup(page
);
354 unlock_page_cgroup(page
);
357 pc
= kzalloc(sizeof(struct page_cgroup
), gfp_mask
);
362 * We always charge the cgroup the mm_struct belongs to.
363 * The mm_struct's mem_cgroup changes on task migration if the
364 * thread group leader migrates. It's possible that mm is not
365 * set, if so charge the init_mm (happens for pagecache usage).
371 mem
= rcu_dereference(mm
->mem_cgroup
);
373 * For every charge from the cgroup, increment reference
380 * If we created the page_cgroup, we should free it on exceeding
383 while (res_counter_charge(&mem
->res
, PAGE_SIZE
)) {
384 if (!(gfp_mask
& __GFP_WAIT
))
387 if (try_to_free_mem_cgroup_pages(mem
, gfp_mask
))
391 * try_to_free_mem_cgroup_pages() might not give us a full
392 * picture of reclaim. Some pages are reclaimed and might be
393 * moved to swap cache or just unmapped from the cgroup.
394 * Check the limit again to see if the reclaim reduced the
395 * current usage of the cgroup before giving up
397 if (res_counter_check_under_limit(&mem
->res
))
401 mem_cgroup_out_of_memory(mem
, gfp_mask
);
404 congestion_wait(WRITE
, HZ
/10);
407 atomic_set(&pc
->ref_cnt
, 1);
408 pc
->mem_cgroup
= mem
;
410 pc
->flags
= PAGE_CGROUP_FLAG_ACTIVE
;
411 if (ctype
== MEM_CGROUP_CHARGE_TYPE_CACHE
)
412 pc
->flags
|= PAGE_CGROUP_FLAG_CACHE
;
414 if (!page
|| page_cgroup_assign_new_page_cgroup(page
, pc
)) {
416 * Another charge has been added to this page already.
417 * We take lock_page_cgroup(page) again and read
418 * page->cgroup, increment refcnt.... just retry is OK.
420 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
428 spin_lock_irqsave(&mem
->lru_lock
, flags
);
429 list_add(&pc
->lru
, &mem
->active_list
);
430 spin_unlock_irqrestore(&mem
->lru_lock
, flags
);
441 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
,
444 return mem_cgroup_charge_common(page
, mm
, gfp_mask
,
445 MEM_CGROUP_CHARGE_TYPE_MAPPED
);
449 * See if the cached pages should be charged at all?
451 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
455 struct mem_cgroup
*mem
;
460 mem
= rcu_dereference(mm
->mem_cgroup
);
463 if (mem
->control_type
== MEM_CGROUP_TYPE_ALL
)
464 ret
= mem_cgroup_charge_common(page
, mm
, gfp_mask
,
465 MEM_CGROUP_CHARGE_TYPE_CACHE
);
471 * Uncharging is always a welcome operation, we never complain, simply
474 void mem_cgroup_uncharge(struct page_cgroup
*pc
)
476 struct mem_cgroup
*mem
;
481 * This can handle cases when a page is not charged at all and we
482 * are switching between handling the control_type.
487 if (atomic_dec_and_test(&pc
->ref_cnt
)) {
490 * get page->cgroup and clear it under lock.
491 * force_empty can drop page->cgroup without checking refcnt.
493 if (clear_page_cgroup(page
, pc
) == pc
) {
494 mem
= pc
->mem_cgroup
;
496 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
497 spin_lock_irqsave(&mem
->lru_lock
, flags
);
498 list_del_init(&pc
->lru
);
499 spin_unlock_irqrestore(&mem
->lru_lock
, flags
);
505 * Returns non-zero if a page (under migration) has valid page_cgroup member.
506 * Refcnt of page_cgroup is incremented.
509 int mem_cgroup_prepare_migration(struct page
*page
)
511 struct page_cgroup
*pc
;
513 lock_page_cgroup(page
);
514 pc
= page_get_page_cgroup(page
);
515 if (pc
&& atomic_inc_not_zero(&pc
->ref_cnt
))
517 unlock_page_cgroup(page
);
521 void mem_cgroup_end_migration(struct page
*page
)
523 struct page_cgroup
*pc
= page_get_page_cgroup(page
);
524 mem_cgroup_uncharge(pc
);
527 * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
528 * And no race with uncharge() routines because page_cgroup for *page*
529 * has extra one reference by mem_cgroup_prepare_migration.
532 void mem_cgroup_page_migration(struct page
*page
, struct page
*newpage
)
534 struct page_cgroup
*pc
;
536 pc
= page_get_page_cgroup(page
);
539 if (clear_page_cgroup(page
, pc
) != pc
)
542 lock_page_cgroup(newpage
);
543 page_assign_page_cgroup(newpage
, pc
);
544 unlock_page_cgroup(newpage
);
549 * This routine traverse page_cgroup in given list and drop them all.
550 * This routine ignores page_cgroup->ref_cnt.
551 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
553 #define FORCE_UNCHARGE_BATCH (128)
555 mem_cgroup_force_empty_list(struct mem_cgroup
*mem
, struct list_head
*list
)
557 struct page_cgroup
*pc
;
563 count
= FORCE_UNCHARGE_BATCH
;
564 spin_lock_irqsave(&mem
->lru_lock
, flags
);
566 while (--count
&& !list_empty(list
)) {
567 pc
= list_entry(list
->prev
, struct page_cgroup
, lru
);
569 /* Avoid race with charge */
570 atomic_set(&pc
->ref_cnt
, 0);
571 if (clear_page_cgroup(page
, pc
) == pc
) {
573 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
574 list_del_init(&pc
->lru
);
576 } else /* being uncharged ? ...do relax */
579 spin_unlock_irqrestore(&mem
->lru_lock
, flags
);
580 if (!list_empty(list
)) {
588 * make mem_cgroup's charge to be 0 if there is no task.
589 * This enables deleting this mem_cgroup.
592 int mem_cgroup_force_empty(struct mem_cgroup
*mem
)
597 * page reclaim code (kswapd etc..) will move pages between
598 ` * active_list <-> inactive_list while we don't take a lock.
599 * So, we have to do loop here until all lists are empty.
601 while (!(list_empty(&mem
->active_list
) &&
602 list_empty(&mem
->inactive_list
))) {
603 if (atomic_read(&mem
->css
.cgroup
->count
) > 0)
605 /* drop all page_cgroup in active_list */
606 mem_cgroup_force_empty_list(mem
, &mem
->active_list
);
607 /* drop all page_cgroup in inactive_list */
608 mem_cgroup_force_empty_list(mem
, &mem
->inactive_list
);
618 int mem_cgroup_write_strategy(char *buf
, unsigned long long *tmp
)
620 *tmp
= memparse(buf
, &buf
);
625 * Round up the value to the closest page size
627 *tmp
= ((*tmp
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
) << PAGE_SHIFT
;
631 static ssize_t
mem_cgroup_read(struct cgroup
*cont
,
632 struct cftype
*cft
, struct file
*file
,
633 char __user
*userbuf
, size_t nbytes
, loff_t
*ppos
)
635 return res_counter_read(&mem_cgroup_from_cont(cont
)->res
,
636 cft
->private, userbuf
, nbytes
, ppos
,
640 static ssize_t
mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
641 struct file
*file
, const char __user
*userbuf
,
642 size_t nbytes
, loff_t
*ppos
)
644 return res_counter_write(&mem_cgroup_from_cont(cont
)->res
,
645 cft
->private, userbuf
, nbytes
, ppos
,
646 mem_cgroup_write_strategy
);
649 static ssize_t
mem_control_type_write(struct cgroup
*cont
,
650 struct cftype
*cft
, struct file
*file
,
651 const char __user
*userbuf
,
652 size_t nbytes
, loff_t
*pos
)
657 struct mem_cgroup
*mem
;
659 mem
= mem_cgroup_from_cont(cont
);
660 buf
= kmalloc(nbytes
+ 1, GFP_KERNEL
);
667 if (copy_from_user(buf
, userbuf
, nbytes
))
671 tmp
= simple_strtoul(buf
, &end
, 10);
675 if (tmp
<= MEM_CGROUP_TYPE_UNSPEC
|| tmp
>= MEM_CGROUP_TYPE_MAX
)
678 mem
->control_type
= tmp
;
686 static ssize_t
mem_control_type_read(struct cgroup
*cont
,
688 struct file
*file
, char __user
*userbuf
,
689 size_t nbytes
, loff_t
*ppos
)
693 struct mem_cgroup
*mem
;
695 mem
= mem_cgroup_from_cont(cont
);
697 val
= mem
->control_type
;
698 s
+= sprintf(s
, "%lu\n", val
);
699 return simple_read_from_buffer((void __user
*)userbuf
, nbytes
,
704 static ssize_t
mem_force_empty_write(struct cgroup
*cont
,
705 struct cftype
*cft
, struct file
*file
,
706 const char __user
*userbuf
,
707 size_t nbytes
, loff_t
*ppos
)
709 struct mem_cgroup
*mem
= mem_cgroup_from_cont(cont
);
711 ret
= mem_cgroup_force_empty(mem
);
718 * Note: This should be removed if cgroup supports write-only file.
721 static ssize_t
mem_force_empty_read(struct cgroup
*cont
,
723 struct file
*file
, char __user
*userbuf
,
724 size_t nbytes
, loff_t
*ppos
)
730 static struct cftype mem_cgroup_files
[] = {
732 .name
= "usage_in_bytes",
733 .private = RES_USAGE
,
734 .read
= mem_cgroup_read
,
737 .name
= "limit_in_bytes",
738 .private = RES_LIMIT
,
739 .write
= mem_cgroup_write
,
740 .read
= mem_cgroup_read
,
744 .private = RES_FAILCNT
,
745 .read
= mem_cgroup_read
,
748 .name
= "control_type",
749 .write
= mem_control_type_write
,
750 .read
= mem_control_type_read
,
753 .name
= "force_empty",
754 .write
= mem_force_empty_write
,
755 .read
= mem_force_empty_read
,
759 static struct mem_cgroup init_mem_cgroup
;
761 static struct cgroup_subsys_state
*
762 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
764 struct mem_cgroup
*mem
;
766 if (unlikely((cont
->parent
) == NULL
)) {
767 mem
= &init_mem_cgroup
;
768 init_mm
.mem_cgroup
= mem
;
770 mem
= kzalloc(sizeof(struct mem_cgroup
), GFP_KERNEL
);
775 res_counter_init(&mem
->res
);
776 INIT_LIST_HEAD(&mem
->active_list
);
777 INIT_LIST_HEAD(&mem
->inactive_list
);
778 spin_lock_init(&mem
->lru_lock
);
779 mem
->control_type
= MEM_CGROUP_TYPE_ALL
;
783 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
786 kfree(mem_cgroup_from_cont(cont
));
789 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
792 return cgroup_add_files(cont
, ss
, mem_cgroup_files
,
793 ARRAY_SIZE(mem_cgroup_files
));
796 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
798 struct cgroup
*old_cont
,
799 struct task_struct
*p
)
801 struct mm_struct
*mm
;
802 struct mem_cgroup
*mem
, *old_mem
;
808 mem
= mem_cgroup_from_cont(cont
);
809 old_mem
= mem_cgroup_from_cont(old_cont
);
815 * Only thread group leaders are allowed to migrate, the mm_struct is
816 * in effect owned by the leader
818 if (p
->tgid
!= p
->pid
)
822 rcu_assign_pointer(mm
->mem_cgroup
, mem
);
823 css_put(&old_mem
->css
);
830 struct cgroup_subsys mem_cgroup_subsys
= {
832 .subsys_id
= mem_cgroup_subsys_id
,
833 .create
= mem_cgroup_create
,
834 .destroy
= mem_cgroup_destroy
,
835 .populate
= mem_cgroup_populate
,
836 .attach
= mem_cgroup_move_task
,