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;
140 static struct mem_cgroup init_mem_cgroup;
143 * We use the lower bit of the page->page_cgroup pointer as a bit spin
144 * lock. We need to ensure that page->page_cgroup is at least two
145 * byte aligned (based on comments from Nick Piggin). But since
146 * bit_spin_lock doesn't actually set that lock bit in a non-debug
147 * uniprocessor kernel, we should avoid setting it here too.
149 #define PAGE_CGROUP_LOCK_BIT 0x0
150 #if defined(CONFIG_SMP) || defined(CONFIG_DEBUG_SPINLOCK)
151 #define PAGE_CGROUP_LOCK (1 << PAGE_CGROUP_LOCK_BIT)
153 #define PAGE_CGROUP_LOCK 0x0
157 * A page_cgroup page is associated with every page descriptor. The
158 * page_cgroup helps us identify information about the cgroup
161 struct list_head lru; /* per cgroup LRU list */
163 struct mem_cgroup *mem_cgroup;
164 atomic_t ref_cnt; /* Helpful when pages move b/w */
165 /* mapped and cached states */
168 #define PAGE_CGROUP_FLAG_CACHE (0x1) /* charged as cache */
169 #define PAGE_CGROUP_FLAG_ACTIVE (0x2) /* page is active in this cgroup */
171 static inline int page_cgroup_nid(struct page_cgroup *pc)
173 return page_to_nid(pc->page);
176 static inline enum zone_type page_cgroup_zid(struct page_cgroup *pc)
178 return page_zonenum(pc->page);
182 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
183 MEM_CGROUP_CHARGE_TYPE_MAPPED,
187 * Always modified under lru lock. Then, not necessary to preempt_disable()
189 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, int flags,
192 int val = (charge)? 1 : -1;
193 struct mem_cgroup_stat *stat = &mem->stat;
195 VM_BUG_ON(!irqs_disabled());
196 if (flags & PAGE_CGROUP_FLAG_CACHE)
197 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_CACHE, val);
199 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val);
202 static inline struct mem_cgroup_per_zone *
203 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
205 BUG_ON(!mem->info.nodeinfo[nid]);
206 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
209 static inline struct mem_cgroup_per_zone *
210 page_cgroup_zoneinfo(struct page_cgroup *pc)
212 struct mem_cgroup *mem = pc->mem_cgroup;
213 int nid = page_cgroup_nid(pc);
214 int zid = page_cgroup_zid(pc);
216 return mem_cgroup_zoneinfo(mem, nid, zid);
219 static unsigned long mem_cgroup_get_all_zonestat(struct mem_cgroup *mem,
220 enum mem_cgroup_zstat_index idx)
223 struct mem_cgroup_per_zone *mz;
226 for_each_online_node(nid)
227 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
228 mz = mem_cgroup_zoneinfo(mem, nid, zid);
229 total += MEM_CGROUP_ZSTAT(mz, idx);
235 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
237 return container_of(cgroup_subsys_state(cont,
238 mem_cgroup_subsys_id), struct mem_cgroup,
243 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
245 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
246 struct mem_cgroup, css);
249 void mm_init_cgroup(struct mm_struct *mm, struct task_struct *p)
251 struct mem_cgroup *mem;
253 mem = mem_cgroup_from_task(p);
255 mm->mem_cgroup = mem;
258 void mm_free_cgroup(struct mm_struct *mm)
260 css_put(&mm->mem_cgroup->css);
263 static inline int page_cgroup_locked(struct page *page)
265 return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
268 static void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
270 VM_BUG_ON(!page_cgroup_locked(page));
271 page->page_cgroup = ((unsigned long)pc | PAGE_CGROUP_LOCK);
274 struct page_cgroup *page_get_page_cgroup(struct page *page)
276 return (struct page_cgroup *) (page->page_cgroup & ~PAGE_CGROUP_LOCK);
279 static void __always_inline lock_page_cgroup(struct page *page)
281 bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
282 VM_BUG_ON(!page_cgroup_locked(page));
285 static void __always_inline unlock_page_cgroup(struct page *page)
287 bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
291 * Clear page->page_cgroup member under lock_page_cgroup().
292 * If given "pc" value is different from one page->page_cgroup,
293 * page->cgroup is not cleared.
294 * Returns a value of page->page_cgroup at lock taken.
295 * A can can detect failure of clearing by following
296 * clear_page_cgroup(page, pc) == pc
298 static struct page_cgroup *clear_page_cgroup(struct page *page,
299 struct page_cgroup *pc)
301 struct page_cgroup *ret;
303 lock_page_cgroup(page);
304 ret = page_get_page_cgroup(page);
305 if (likely(ret == pc))
306 page_assign_page_cgroup(page, NULL);
307 unlock_page_cgroup(page);
311 static void __mem_cgroup_remove_list(struct page_cgroup *pc)
313 int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
314 struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
317 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
319 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
321 mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, false);
322 list_del_init(&pc->lru);
325 static void __mem_cgroup_add_list(struct page_cgroup *pc)
327 int to = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
328 struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
331 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
332 list_add(&pc->lru, &mz->inactive_list);
334 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
335 list_add(&pc->lru, &mz->active_list);
337 mem_cgroup_charge_statistics(pc->mem_cgroup, pc->flags, true);
340 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
342 int from = pc->flags & PAGE_CGROUP_FLAG_ACTIVE;
343 struct mem_cgroup_per_zone *mz = page_cgroup_zoneinfo(pc);
346 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) -= 1;
348 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) -= 1;
351 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE) += 1;
352 pc->flags |= PAGE_CGROUP_FLAG_ACTIVE;
353 list_move(&pc->lru, &mz->active_list);
355 MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE) += 1;
356 pc->flags &= ~PAGE_CGROUP_FLAG_ACTIVE;
357 list_move(&pc->lru, &mz->inactive_list);
361 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
366 ret = task->mm && mm_match_cgroup(task->mm, mem);
372 * This routine assumes that the appropriate zone's lru lock is already held
374 void mem_cgroup_move_lists(struct page *page, bool active)
376 struct page_cgroup *pc;
377 struct mem_cgroup_per_zone *mz;
380 pc = page_get_page_cgroup(page);
384 mz = page_cgroup_zoneinfo(pc);
385 spin_lock_irqsave(&mz->lru_lock, flags);
386 __mem_cgroup_move_lists(pc, active);
387 spin_unlock_irqrestore(&mz->lru_lock, flags);
391 * Calculate mapped_ratio under memory controller. This will be used in
392 * vmscan.c for deteremining we have to reclaim mapped pages.
394 int mem_cgroup_calc_mapped_ratio(struct mem_cgroup *mem)
399 * usage is recorded in bytes. But, here, we assume the number of
400 * physical pages can be represented by "long" on any arch.
402 total = (long) (mem->res.usage >> PAGE_SHIFT) + 1L;
403 rss = (long)mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
404 return (int)((rss * 100L) / total);
408 * This function is called from vmscan.c. In page reclaiming loop. balance
409 * between active and inactive list is calculated. For memory controller
410 * page reclaiming, we should use using mem_cgroup's imbalance rather than
411 * zone's global lru imbalance.
413 long mem_cgroup_reclaim_imbalance(struct mem_cgroup *mem)
415 unsigned long active, inactive;
416 /* active and inactive are the number of pages. 'long' is ok.*/
417 active = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_ACTIVE);
418 inactive = mem_cgroup_get_all_zonestat(mem, MEM_CGROUP_ZSTAT_INACTIVE);
419 return (long) (active / (inactive + 1));
423 * prev_priority control...this will be used in memory reclaim path.
425 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
427 return mem->prev_priority;
430 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
432 if (priority < mem->prev_priority)
433 mem->prev_priority = priority;
436 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
438 mem->prev_priority = priority;
442 * Calculate # of pages to be scanned in this priority/zone.
445 * priority starts from "DEF_PRIORITY" and decremented in each loop.
446 * (see include/linux/mmzone.h)
449 long mem_cgroup_calc_reclaim_active(struct mem_cgroup *mem,
450 struct zone *zone, int priority)
453 int nid = zone->zone_pgdat->node_id;
454 int zid = zone_idx(zone);
455 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid);
457 nr_active = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_ACTIVE);
458 return (nr_active >> priority);
461 long mem_cgroup_calc_reclaim_inactive(struct mem_cgroup *mem,
462 struct zone *zone, int priority)
465 int nid = zone->zone_pgdat->node_id;
466 int zid = zone_idx(zone);
467 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(mem, nid, zid);
469 nr_inactive = MEM_CGROUP_ZSTAT(mz, MEM_CGROUP_ZSTAT_INACTIVE);
470 return (nr_inactive >> priority);
473 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
474 struct list_head *dst,
475 unsigned long *scanned, int order,
476 int mode, struct zone *z,
477 struct mem_cgroup *mem_cont,
480 unsigned long nr_taken = 0;
484 struct list_head *src;
485 struct page_cgroup *pc, *tmp;
486 int nid = z->zone_pgdat->node_id;
487 int zid = zone_idx(z);
488 struct mem_cgroup_per_zone *mz;
490 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
492 src = &mz->active_list;
494 src = &mz->inactive_list;
497 spin_lock(&mz->lru_lock);
499 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
500 if (scan >= nr_to_scan)
504 if (unlikely(!PageLRU(page)))
507 if (PageActive(page) && !active) {
508 __mem_cgroup_move_lists(pc, true);
511 if (!PageActive(page) && active) {
512 __mem_cgroup_move_lists(pc, false);
517 list_move(&pc->lru, &pc_list);
519 if (__isolate_lru_page(page, mode) == 0) {
520 list_move(&page->lru, dst);
525 list_splice(&pc_list, src);
526 spin_unlock(&mz->lru_lock);
533 * Charge the memory controller for page usage.
535 * 0 if the charge was successful
536 * < 0 if the cgroup is over its limit
538 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
539 gfp_t gfp_mask, enum charge_type ctype)
541 struct mem_cgroup *mem;
542 struct page_cgroup *pc;
544 unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
545 struct mem_cgroup_per_zone *mz;
548 * Should page_cgroup's go to their own slab?
549 * One could optimize the performance of the charging routine
550 * by saving a bit in the page_flags and using it as a lock
551 * to see if the cgroup page already has a page_cgroup associated
555 lock_page_cgroup(page);
556 pc = page_get_page_cgroup(page);
558 * The page_cgroup exists and
559 * the page has already been accounted.
562 if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
563 /* this page is under being uncharged ? */
564 unlock_page_cgroup(page);
568 unlock_page_cgroup(page);
572 unlock_page_cgroup(page);
574 pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
579 * We always charge the cgroup the mm_struct belongs to.
580 * The mm_struct's mem_cgroup changes on task migration if the
581 * thread group leader migrates. It's possible that mm is not
582 * set, if so charge the init_mm (happens for pagecache usage).
588 mem = rcu_dereference(mm->mem_cgroup);
590 * For every charge from the cgroup, increment reference count
595 while (res_counter_charge(&mem->res, PAGE_SIZE)) {
596 if (!(gfp_mask & __GFP_WAIT))
599 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
603 * try_to_free_mem_cgroup_pages() might not give us a full
604 * picture of reclaim. Some pages are reclaimed and might be
605 * moved to swap cache or just unmapped from the cgroup.
606 * Check the limit again to see if the reclaim reduced the
607 * current usage of the cgroup before giving up
609 if (res_counter_check_under_limit(&mem->res))
613 mem_cgroup_out_of_memory(mem, gfp_mask);
616 congestion_wait(WRITE, HZ/10);
619 atomic_set(&pc->ref_cnt, 1);
620 pc->mem_cgroup = mem;
622 pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
623 if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
624 pc->flags |= PAGE_CGROUP_FLAG_CACHE;
626 lock_page_cgroup(page);
627 if (page_get_page_cgroup(page)) {
628 unlock_page_cgroup(page);
630 * Another charge has been added to this page already.
631 * We take lock_page_cgroup(page) again and read
632 * page->cgroup, increment refcnt.... just retry is OK.
634 res_counter_uncharge(&mem->res, PAGE_SIZE);
639 page_assign_page_cgroup(page, pc);
640 unlock_page_cgroup(page);
642 mz = page_cgroup_zoneinfo(pc);
643 spin_lock_irqsave(&mz->lru_lock, flags);
644 __mem_cgroup_add_list(pc);
645 spin_unlock_irqrestore(&mz->lru_lock, flags);
656 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
658 return mem_cgroup_charge_common(page, mm, gfp_mask,
659 MEM_CGROUP_CHARGE_TYPE_MAPPED);
662 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
667 return mem_cgroup_charge_common(page, mm, gfp_mask,
668 MEM_CGROUP_CHARGE_TYPE_CACHE);
672 * Uncharging is always a welcome operation, we never complain, simply
675 void mem_cgroup_uncharge_page(struct page *page)
677 struct page_cgroup *pc;
678 struct mem_cgroup *mem;
679 struct mem_cgroup_per_zone *mz;
683 * Check if our page_cgroup is valid
685 lock_page_cgroup(page);
686 pc = page_get_page_cgroup(page);
690 if (atomic_dec_and_test(&pc->ref_cnt)) {
692 mz = page_cgroup_zoneinfo(pc);
694 * get page->cgroup and clear it under lock.
695 * force_empty can drop page->cgroup without checking refcnt.
697 unlock_page_cgroup(page);
698 if (clear_page_cgroup(page, pc) == pc) {
699 mem = pc->mem_cgroup;
701 res_counter_uncharge(&mem->res, PAGE_SIZE);
702 spin_lock_irqsave(&mz->lru_lock, flags);
703 __mem_cgroup_remove_list(pc);
704 spin_unlock_irqrestore(&mz->lru_lock, flags);
707 lock_page_cgroup(page);
711 unlock_page_cgroup(page);
715 * Returns non-zero if a page (under migration) has valid page_cgroup member.
716 * Refcnt of page_cgroup is incremented.
718 int mem_cgroup_prepare_migration(struct page *page)
720 struct page_cgroup *pc;
723 lock_page_cgroup(page);
724 pc = page_get_page_cgroup(page);
725 if (pc && atomic_inc_not_zero(&pc->ref_cnt))
727 unlock_page_cgroup(page);
731 void mem_cgroup_end_migration(struct page *page)
733 mem_cgroup_uncharge_page(page);
737 * We know both *page* and *newpage* are now not-on-LRU and PG_locked.
738 * And no race with uncharge() routines because page_cgroup for *page*
739 * has extra one reference by mem_cgroup_prepare_migration.
741 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
743 struct page_cgroup *pc;
744 struct mem_cgroup *mem;
746 struct mem_cgroup_per_zone *mz;
749 pc = page_get_page_cgroup(page);
753 mem = pc->mem_cgroup;
754 mz = page_cgroup_zoneinfo(pc);
755 if (clear_page_cgroup(page, pc) != pc)
758 spin_lock_irqsave(&mz->lru_lock, flags);
759 __mem_cgroup_remove_list(pc);
760 spin_unlock_irqrestore(&mz->lru_lock, flags);
763 lock_page_cgroup(newpage);
764 page_assign_page_cgroup(newpage, pc);
765 unlock_page_cgroup(newpage);
767 mz = page_cgroup_zoneinfo(pc);
768 spin_lock_irqsave(&mz->lru_lock, flags);
769 __mem_cgroup_add_list(pc);
770 spin_unlock_irqrestore(&mz->lru_lock, flags);
774 * This routine traverse page_cgroup in given list and drop them all.
775 * This routine ignores page_cgroup->ref_cnt.
776 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
778 #define FORCE_UNCHARGE_BATCH (128)
779 static void mem_cgroup_force_empty_list(struct mem_cgroup *mem,
780 struct mem_cgroup_per_zone *mz,
783 struct page_cgroup *pc;
787 struct list_head *list;
790 list = &mz->active_list;
792 list = &mz->inactive_list;
794 if (list_empty(list))
797 count = FORCE_UNCHARGE_BATCH;
798 spin_lock_irqsave(&mz->lru_lock, flags);
800 while (--count && !list_empty(list)) {
801 pc = list_entry(list->prev, struct page_cgroup, lru);
803 /* Avoid race with charge */
804 atomic_set(&pc->ref_cnt, 0);
805 if (clear_page_cgroup(page, pc) == pc) {
807 res_counter_uncharge(&mem->res, PAGE_SIZE);
808 __mem_cgroup_remove_list(pc);
810 } else /* being uncharged ? ...do relax */
814 spin_unlock_irqrestore(&mz->lru_lock, flags);
815 if (!list_empty(list)) {
822 * make mem_cgroup's charge to be 0 if there is no task.
823 * This enables deleting this mem_cgroup.
825 int mem_cgroup_force_empty(struct mem_cgroup *mem)
832 * page reclaim code (kswapd etc..) will move pages between
833 * active_list <-> inactive_list while we don't take a lock.
834 * So, we have to do loop here until all lists are empty.
836 while (mem->res.usage > 0) {
837 if (atomic_read(&mem->css.cgroup->count) > 0)
839 for_each_node_state(node, N_POSSIBLE)
840 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
841 struct mem_cgroup_per_zone *mz;
842 mz = mem_cgroup_zoneinfo(mem, node, zid);
843 /* drop all page_cgroup in active_list */
844 mem_cgroup_force_empty_list(mem, mz, 1);
845 /* drop all page_cgroup in inactive_list */
846 mem_cgroup_force_empty_list(mem, mz, 0);
855 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
857 *tmp = memparse(buf, &buf);
862 * Round up the value to the closest page size
864 *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
868 static ssize_t mem_cgroup_read(struct cgroup *cont,
869 struct cftype *cft, struct file *file,
870 char __user *userbuf, size_t nbytes, loff_t *ppos)
872 return res_counter_read(&mem_cgroup_from_cont(cont)->res,
873 cft->private, userbuf, nbytes, ppos,
877 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
878 struct file *file, const char __user *userbuf,
879 size_t nbytes, loff_t *ppos)
881 return res_counter_write(&mem_cgroup_from_cont(cont)->res,
882 cft->private, userbuf, nbytes, ppos,
883 mem_cgroup_write_strategy);
886 static ssize_t mem_force_empty_write(struct cgroup *cont,
887 struct cftype *cft, struct file *file,
888 const char __user *userbuf,
889 size_t nbytes, loff_t *ppos)
891 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
892 int ret = mem_cgroup_force_empty(mem);
899 * Note: This should be removed if cgroup supports write-only file.
901 static ssize_t mem_force_empty_read(struct cgroup *cont,
903 struct file *file, char __user *userbuf,
904 size_t nbytes, loff_t *ppos)
909 static const struct mem_cgroup_stat_desc {
912 } mem_cgroup_stat_desc[] = {
913 [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
914 [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
917 static int mem_control_stat_show(struct seq_file *m, void *arg)
919 struct cgroup *cont = m->private;
920 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
921 struct mem_cgroup_stat *stat = &mem_cont->stat;
924 for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
927 val = mem_cgroup_read_stat(stat, i);
928 val *= mem_cgroup_stat_desc[i].unit;
929 seq_printf(m, "%s %lld\n", mem_cgroup_stat_desc[i].msg,
932 /* showing # of active pages */
934 unsigned long active, inactive;
936 inactive = mem_cgroup_get_all_zonestat(mem_cont,
937 MEM_CGROUP_ZSTAT_INACTIVE);
938 active = mem_cgroup_get_all_zonestat(mem_cont,
939 MEM_CGROUP_ZSTAT_ACTIVE);
940 seq_printf(m, "active %ld\n", (active) * PAGE_SIZE);
941 seq_printf(m, "inactive %ld\n", (inactive) * PAGE_SIZE);
946 static const struct file_operations mem_control_stat_file_operations = {
949 .release = single_release,
952 static int mem_control_stat_open(struct inode *unused, struct file *file)
955 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
957 file->f_op = &mem_control_stat_file_operations;
958 return single_open(file, mem_control_stat_show, cont);
961 static struct cftype mem_cgroup_files[] = {
963 .name = "usage_in_bytes",
964 .private = RES_USAGE,
965 .read = mem_cgroup_read,
968 .name = "limit_in_bytes",
969 .private = RES_LIMIT,
970 .write = mem_cgroup_write,
971 .read = mem_cgroup_read,
975 .private = RES_FAILCNT,
976 .read = mem_cgroup_read,
979 .name = "force_empty",
980 .write = mem_force_empty_write,
981 .read = mem_force_empty_read,
985 .open = mem_control_stat_open,
989 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
991 struct mem_cgroup_per_node *pn;
992 struct mem_cgroup_per_zone *mz;
995 * This routine is called against possible nodes.
996 * But it's BUG to call kmalloc() against offline node.
998 * TODO: this routine can waste much memory for nodes which will
999 * never be onlined. It's better to use memory hotplug callback
1002 if (node_state(node, N_HIGH_MEMORY))
1003 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, node);
1005 pn = kmalloc(sizeof(*pn), GFP_KERNEL);
1009 mem->info.nodeinfo[node] = pn;
1010 memset(pn, 0, sizeof(*pn));
1012 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
1013 mz = &pn->zoneinfo[zone];
1014 INIT_LIST_HEAD(&mz->active_list);
1015 INIT_LIST_HEAD(&mz->inactive_list);
1016 spin_lock_init(&mz->lru_lock);
1021 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
1023 kfree(mem->info.nodeinfo[node]);
1026 static struct cgroup_subsys_state *
1027 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
1029 struct mem_cgroup *mem;
1032 if (unlikely((cont->parent) == NULL)) {
1033 mem = &init_mem_cgroup;
1034 init_mm.mem_cgroup = mem;
1036 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
1039 return ERR_PTR(-ENOMEM);
1041 res_counter_init(&mem->res);
1043 memset(&mem->info, 0, sizeof(mem->info));
1045 for_each_node_state(node, N_POSSIBLE)
1046 if (alloc_mem_cgroup_per_zone_info(mem, node))
1051 for_each_node_state(node, N_POSSIBLE)
1052 free_mem_cgroup_per_zone_info(mem, node);
1053 if (cont->parent != NULL)
1055 return ERR_PTR(-ENOMEM);
1058 static void mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
1059 struct cgroup *cont)
1061 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1062 mem_cgroup_force_empty(mem);
1065 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
1066 struct cgroup *cont)
1069 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
1071 for_each_node_state(node, N_POSSIBLE)
1072 free_mem_cgroup_per_zone_info(mem, node);
1074 kfree(mem_cgroup_from_cont(cont));
1077 static int mem_cgroup_populate(struct cgroup_subsys *ss,
1078 struct cgroup *cont)
1080 return cgroup_add_files(cont, ss, mem_cgroup_files,
1081 ARRAY_SIZE(mem_cgroup_files));
1084 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
1085 struct cgroup *cont,
1086 struct cgroup *old_cont,
1087 struct task_struct *p)
1089 struct mm_struct *mm;
1090 struct mem_cgroup *mem, *old_mem;
1092 mm = get_task_mm(p);
1096 mem = mem_cgroup_from_cont(cont);
1097 old_mem = mem_cgroup_from_cont(old_cont);
1103 * Only thread group leaders are allowed to migrate, the mm_struct is
1104 * in effect owned by the leader
1106 if (p->tgid != p->pid)
1110 rcu_assign_pointer(mm->mem_cgroup, mem);
1111 css_put(&old_mem->css);
1117 struct cgroup_subsys mem_cgroup_subsys = {
1119 .subsys_id = mem_cgroup_subsys_id,
1120 .create = mem_cgroup_create,
1121 .pre_destroy = mem_cgroup_pre_destroy,
1122 .destroy = mem_cgroup_destroy,
1123 .populate = mem_cgroup_populate,
1124 .attach = mem_cgroup_move_task,