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 */
89 MEM_CGROUP_TYPE_UNSPEC
= 0,
90 MEM_CGROUP_TYPE_MAPPED
,
91 MEM_CGROUP_TYPE_CACHED
,
96 static struct mem_cgroup init_mem_cgroup
;
99 struct mem_cgroup
*mem_cgroup_from_cont(struct cgroup
*cont
)
101 return container_of(cgroup_subsys_state(cont
,
102 mem_cgroup_subsys_id
), struct mem_cgroup
,
107 struct mem_cgroup
*mem_cgroup_from_task(struct task_struct
*p
)
109 return container_of(task_subsys_state(p
, mem_cgroup_subsys_id
),
110 struct mem_cgroup
, css
);
113 void mm_init_cgroup(struct mm_struct
*mm
, struct task_struct
*p
)
115 struct mem_cgroup
*mem
;
117 mem
= mem_cgroup_from_task(p
);
119 mm
->mem_cgroup
= mem
;
122 void mm_free_cgroup(struct mm_struct
*mm
)
124 css_put(&mm
->mem_cgroup
->css
);
127 static inline int page_cgroup_locked(struct page
*page
)
129 return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT
,
133 void page_assign_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
138 * While resetting the page_cgroup we might not hold the
139 * page_cgroup lock. free_hot_cold_page() is an example
143 VM_BUG_ON(!page_cgroup_locked(page
));
144 locked
= (page
->page_cgroup
& PAGE_CGROUP_LOCK
);
145 page
->page_cgroup
= ((unsigned long)pc
| locked
);
148 struct page_cgroup
*page_get_page_cgroup(struct page
*page
)
150 return (struct page_cgroup
*)
151 (page
->page_cgroup
& ~PAGE_CGROUP_LOCK
);
154 static void __always_inline
lock_page_cgroup(struct page
*page
)
156 bit_spin_lock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
157 VM_BUG_ON(!page_cgroup_locked(page
));
160 static void __always_inline
unlock_page_cgroup(struct page
*page
)
162 bit_spin_unlock(PAGE_CGROUP_LOCK_BIT
, &page
->page_cgroup
);
166 * Tie new page_cgroup to struct page under lock_page_cgroup()
167 * This can fail if the page has been tied to a page_cgroup.
168 * If success, returns 0.
171 page_cgroup_assign_new_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
175 lock_page_cgroup(page
);
176 if (!page_get_page_cgroup(page
))
177 page_assign_page_cgroup(page
, pc
);
178 else /* A page is tied to other pc. */
180 unlock_page_cgroup(page
);
185 * Clear page->page_cgroup member under lock_page_cgroup().
186 * If given "pc" value is different from one page->page_cgroup,
187 * page->cgroup is not cleared.
188 * Returns a value of page->page_cgroup at lock taken.
189 * A can can detect failure of clearing by following
190 * clear_page_cgroup(page, pc) == pc
193 static inline struct page_cgroup
*
194 clear_page_cgroup(struct page
*page
, struct page_cgroup
*pc
)
196 struct page_cgroup
*ret
;
198 lock_page_cgroup(page
);
199 ret
= page_get_page_cgroup(page
);
200 if (likely(ret
== pc
))
201 page_assign_page_cgroup(page
, NULL
);
202 unlock_page_cgroup(page
);
207 static void __mem_cgroup_move_lists(struct page_cgroup
*pc
, bool active
)
210 list_move(&pc
->lru
, &pc
->mem_cgroup
->active_list
);
212 list_move(&pc
->lru
, &pc
->mem_cgroup
->inactive_list
);
215 int task_in_mem_cgroup(struct task_struct
*task
, const struct mem_cgroup
*mem
)
220 ret
= task
->mm
&& mm_cgroup(task
->mm
) == mem
;
226 * This routine assumes that the appropriate zone's lru lock is already held
228 void mem_cgroup_move_lists(struct page_cgroup
*pc
, bool active
)
230 struct mem_cgroup
*mem
;
234 mem
= pc
->mem_cgroup
;
236 spin_lock(&mem
->lru_lock
);
237 __mem_cgroup_move_lists(pc
, active
);
238 spin_unlock(&mem
->lru_lock
);
241 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan
,
242 struct list_head
*dst
,
243 unsigned long *scanned
, int order
,
244 int mode
, struct zone
*z
,
245 struct mem_cgroup
*mem_cont
,
248 unsigned long nr_taken
= 0;
252 struct list_head
*src
;
253 struct page_cgroup
*pc
, *tmp
;
256 src
= &mem_cont
->active_list
;
258 src
= &mem_cont
->inactive_list
;
260 spin_lock(&mem_cont
->lru_lock
);
262 list_for_each_entry_safe_reverse(pc
, tmp
, src
, lru
) {
263 if (scan
>= nr_to_scan
)
268 if (unlikely(!PageLRU(page
)))
271 if (PageActive(page
) && !active
) {
272 __mem_cgroup_move_lists(pc
, true);
275 if (!PageActive(page
) && active
) {
276 __mem_cgroup_move_lists(pc
, false);
282 * TODO: make the active/inactive lists per zone
284 if (page_zone(page
) != z
)
288 list_move(&pc
->lru
, &pc_list
);
290 if (__isolate_lru_page(page
, mode
) == 0) {
291 list_move(&page
->lru
, dst
);
296 list_splice(&pc_list
, src
);
297 spin_unlock(&mem_cont
->lru_lock
);
304 * Charge the memory controller for page usage.
306 * 0 if the charge was successful
307 * < 0 if the cgroup is over its limit
309 int mem_cgroup_charge(struct page
*page
, struct mm_struct
*mm
,
312 struct mem_cgroup
*mem
;
313 struct page_cgroup
*pc
;
315 unsigned long nr_retries
= MEM_CGROUP_RECLAIM_RETRIES
;
318 * Should page_cgroup's go to their own slab?
319 * One could optimize the performance of the charging routine
320 * by saving a bit in the page_flags and using it as a lock
321 * to see if the cgroup page already has a page_cgroup associated
325 lock_page_cgroup(page
);
326 pc
= page_get_page_cgroup(page
);
328 * The page_cgroup exists and the page has already been accounted
331 if (unlikely(!atomic_inc_not_zero(&pc
->ref_cnt
))) {
332 /* this page is under being uncharged ? */
333 unlock_page_cgroup(page
);
337 unlock_page_cgroup(page
);
342 unlock_page_cgroup(page
);
344 pc
= kzalloc(sizeof(struct page_cgroup
), gfp_mask
);
350 * We always charge the cgroup the mm_struct belongs to
351 * the mm_struct's mem_cgroup changes on task migration if the
352 * thread group leader migrates. It's possible that mm is not
353 * set, if so charge the init_mm (happens for pagecache usage).
358 mem
= rcu_dereference(mm
->mem_cgroup
);
360 * For every charge from the cgroup, increment reference
367 * If we created the page_cgroup, we should free it on exceeding
370 while (res_counter_charge(&mem
->res
, PAGE_SIZE
)) {
371 bool is_atomic
= gfp_mask
& GFP_ATOMIC
;
373 * We cannot reclaim under GFP_ATOMIC, fail the charge
378 if (try_to_free_mem_cgroup_pages(mem
, gfp_mask
))
382 * try_to_free_mem_cgroup_pages() might not give us a full
383 * picture of reclaim. Some pages are reclaimed and might be
384 * moved to swap cache or just unmapped from the cgroup.
385 * Check the limit again to see if the reclaim reduced the
386 * current usage of the cgroup before giving up
388 if (res_counter_check_under_limit(&mem
->res
))
391 * Since we control both RSS and cache, we end up with a
392 * very interesting scenario where we end up reclaiming
393 * memory (essentially RSS), since the memory is pushed
394 * to swap cache, we eventually end up adding those
395 * pages back to our list. Hence we give ourselves a
396 * few chances before we fail
398 else if (nr_retries
--) {
399 congestion_wait(WRITE
, HZ
/10);
405 mem_cgroup_out_of_memory(mem
, GFP_KERNEL
);
409 atomic_set(&pc
->ref_cnt
, 1);
410 pc
->mem_cgroup
= mem
;
412 if (page_cgroup_assign_new_page_cgroup(page
, pc
)) {
414 * an another charge is added to this page already.
415 * we do take lock_page_cgroup(page) again and read
416 * page->cgroup, increment refcnt.... just retry is OK.
418 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
424 spin_lock_irqsave(&mem
->lru_lock
, flags
);
425 list_add(&pc
->lru
, &mem
->active_list
);
426 spin_unlock_irqrestore(&mem
->lru_lock
, flags
);
437 * See if the cached pages should be charged at all?
439 int mem_cgroup_cache_charge(struct page
*page
, struct mm_struct
*mm
,
442 struct mem_cgroup
*mem
;
446 mem
= rcu_dereference(mm
->mem_cgroup
);
447 if (mem
->control_type
== MEM_CGROUP_TYPE_ALL
)
448 return mem_cgroup_charge(page
, mm
, gfp_mask
);
454 * Uncharging is always a welcome operation, we never complain, simply
457 void mem_cgroup_uncharge(struct page_cgroup
*pc
)
459 struct mem_cgroup
*mem
;
464 * This can handle cases when a page is not charged at all and we
465 * are switching between handling the control_type.
470 if (atomic_dec_and_test(&pc
->ref_cnt
)) {
473 * get page->cgroup and clear it under lock.
475 if (clear_page_cgroup(page
, pc
) == pc
) {
476 mem
= pc
->mem_cgroup
;
478 res_counter_uncharge(&mem
->res
, PAGE_SIZE
);
479 spin_lock_irqsave(&mem
->lru_lock
, flags
);
480 list_del_init(&pc
->lru
);
481 spin_unlock_irqrestore(&mem
->lru_lock
, flags
);
485 * Note:This will be removed when force-empty patch is
486 * applied. just show warning here.
488 printk(KERN_ERR
"Race in mem_cgroup_uncharge() ?");
494 * Returns non-zero if a page (under migration) has valid page_cgroup member.
495 * Refcnt of page_cgroup is incremented.
498 int mem_cgroup_prepare_migration(struct page
*page
)
500 struct page_cgroup
*pc
;
502 lock_page_cgroup(page
);
503 pc
= page_get_page_cgroup(page
);
504 if (pc
&& atomic_inc_not_zero(&pc
->ref_cnt
))
506 unlock_page_cgroup(page
);
510 void mem_cgroup_end_migration(struct page
*page
)
512 struct page_cgroup
*pc
= page_get_page_cgroup(page
);
513 mem_cgroup_uncharge(pc
);
516 * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
517 * And no race with uncharge() routines because page_cgroup for *page*
518 * has extra one reference by mem_cgroup_prepare_migration.
521 void mem_cgroup_page_migration(struct page
*page
, struct page
*newpage
)
523 struct page_cgroup
*pc
;
525 pc
= page_get_page_cgroup(page
);
528 if (clear_page_cgroup(page
, pc
) != pc
)
531 lock_page_cgroup(newpage
);
532 page_assign_page_cgroup(newpage
, pc
);
533 unlock_page_cgroup(newpage
);
537 int mem_cgroup_write_strategy(char *buf
, unsigned long long *tmp
)
539 *tmp
= memparse(buf
, &buf
);
544 * Round up the value to the closest page size
546 *tmp
= ((*tmp
+ PAGE_SIZE
- 1) >> PAGE_SHIFT
) << PAGE_SHIFT
;
550 static ssize_t
mem_cgroup_read(struct cgroup
*cont
,
551 struct cftype
*cft
, struct file
*file
,
552 char __user
*userbuf
, size_t nbytes
, loff_t
*ppos
)
554 return res_counter_read(&mem_cgroup_from_cont(cont
)->res
,
555 cft
->private, userbuf
, nbytes
, ppos
,
559 static ssize_t
mem_cgroup_write(struct cgroup
*cont
, struct cftype
*cft
,
560 struct file
*file
, const char __user
*userbuf
,
561 size_t nbytes
, loff_t
*ppos
)
563 return res_counter_write(&mem_cgroup_from_cont(cont
)->res
,
564 cft
->private, userbuf
, nbytes
, ppos
,
565 mem_cgroup_write_strategy
);
568 static ssize_t
mem_control_type_write(struct cgroup
*cont
,
569 struct cftype
*cft
, struct file
*file
,
570 const char __user
*userbuf
,
571 size_t nbytes
, loff_t
*pos
)
576 struct mem_cgroup
*mem
;
578 mem
= mem_cgroup_from_cont(cont
);
579 buf
= kmalloc(nbytes
+ 1, GFP_KERNEL
);
586 if (copy_from_user(buf
, userbuf
, nbytes
))
590 tmp
= simple_strtoul(buf
, &end
, 10);
594 if (tmp
<= MEM_CGROUP_TYPE_UNSPEC
|| tmp
>= MEM_CGROUP_TYPE_MAX
)
597 mem
->control_type
= tmp
;
605 static ssize_t
mem_control_type_read(struct cgroup
*cont
,
607 struct file
*file
, char __user
*userbuf
,
608 size_t nbytes
, loff_t
*ppos
)
612 struct mem_cgroup
*mem
;
614 mem
= mem_cgroup_from_cont(cont
);
616 val
= mem
->control_type
;
617 s
+= sprintf(s
, "%lu\n", val
);
618 return simple_read_from_buffer((void __user
*)userbuf
, nbytes
,
622 static struct cftype mem_cgroup_files
[] = {
624 .name
= "usage_in_bytes",
625 .private = RES_USAGE
,
626 .read
= mem_cgroup_read
,
629 .name
= "limit_in_bytes",
630 .private = RES_LIMIT
,
631 .write
= mem_cgroup_write
,
632 .read
= mem_cgroup_read
,
636 .private = RES_FAILCNT
,
637 .read
= mem_cgroup_read
,
640 .name
= "control_type",
641 .write
= mem_control_type_write
,
642 .read
= mem_control_type_read
,
646 static struct mem_cgroup init_mem_cgroup
;
648 static struct cgroup_subsys_state
*
649 mem_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
651 struct mem_cgroup
*mem
;
653 if (unlikely((cont
->parent
) == NULL
)) {
654 mem
= &init_mem_cgroup
;
655 init_mm
.mem_cgroup
= mem
;
657 mem
= kzalloc(sizeof(struct mem_cgroup
), GFP_KERNEL
);
662 res_counter_init(&mem
->res
);
663 INIT_LIST_HEAD(&mem
->active_list
);
664 INIT_LIST_HEAD(&mem
->inactive_list
);
665 spin_lock_init(&mem
->lru_lock
);
666 mem
->control_type
= MEM_CGROUP_TYPE_ALL
;
670 static void mem_cgroup_destroy(struct cgroup_subsys
*ss
,
673 kfree(mem_cgroup_from_cont(cont
));
676 static int mem_cgroup_populate(struct cgroup_subsys
*ss
,
679 return cgroup_add_files(cont
, ss
, mem_cgroup_files
,
680 ARRAY_SIZE(mem_cgroup_files
));
683 static void mem_cgroup_move_task(struct cgroup_subsys
*ss
,
685 struct cgroup
*old_cont
,
686 struct task_struct
*p
)
688 struct mm_struct
*mm
;
689 struct mem_cgroup
*mem
, *old_mem
;
695 mem
= mem_cgroup_from_cont(cont
);
696 old_mem
= mem_cgroup_from_cont(old_cont
);
702 * Only thread group leaders are allowed to migrate, the mm_struct is
703 * in effect owned by the leader
705 if (p
->tgid
!= p
->pid
)
709 rcu_assign_pointer(mm
->mem_cgroup
, mem
);
710 css_put(&old_mem
->css
);
717 struct cgroup_subsys mem_cgroup_subsys
= {
719 .subsys_id
= mem_cgroup_subsys_id
,
720 .create
= mem_cgroup_create
,
721 .destroy
= mem_cgroup_destroy
,
722 .populate
= mem_cgroup_populate
,
723 .attach
= mem_cgroup_move_task
,