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 * The memory controller data structure. The memory controller controls both
82 * page cache and RSS per cgroup. We would eventually like to provide
83 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
84 * to help the administrator determine what knobs to tune.
86 * TODO: Add a water mark for the memory controller. Reclaim will begin when
87 * we hit the water mark. May be even add a low water mark, such that
88 * no reclaim occurs from a cgroup at it's low water mark, this is
89 * a feature that will be implemented much later in the future.
92 struct cgroup_subsys_state css;
94 * the counter to account for memory usage
96 struct res_counter res;
98 * Per cgroup active and inactive list, similar to the
100 * TODO: Consider making these lists per zone
102 struct list_head active_list;
103 struct list_head inactive_list;
105 * spin_lock to protect the per cgroup LRU
108 unsigned long control_type; /* control RSS or RSS+Pagecache */
112 struct mem_cgroup_stat stat;
116 * We use the lower bit of the page->page_cgroup pointer as a bit spin
117 * lock. We need to ensure that page->page_cgroup is atleast two
118 * byte aligned (based on comments from Nick Piggin)
120 #define PAGE_CGROUP_LOCK_BIT 0x0
121 #define PAGE_CGROUP_LOCK (1 << PAGE_CGROUP_LOCK_BIT)
124 * A page_cgroup page is associated with every page descriptor. The
125 * page_cgroup helps us identify information about the cgroup
128 struct list_head lru; /* per cgroup LRU list */
130 struct mem_cgroup *mem_cgroup;
131 atomic_t ref_cnt; /* Helpful when pages move b/w */
132 /* mapped and cached states */
135 #define PAGE_CGROUP_FLAG_CACHE (0x1) /* charged as cache */
136 #define PAGE_CGROUP_FLAG_ACTIVE (0x2) /* page is active in this cgroup */
139 MEM_CGROUP_TYPE_UNSPEC = 0,
140 MEM_CGROUP_TYPE_MAPPED,
141 MEM_CGROUP_TYPE_CACHED,
147 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
148 MEM_CGROUP_CHARGE_TYPE_MAPPED,
152 * Always modified under lru lock. Then, not necessary to preempt_disable()
154 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem, int flags,
157 int val = (charge)? 1 : -1;
158 struct mem_cgroup_stat *stat = &mem->stat;
159 VM_BUG_ON(!irqs_disabled());
161 if (flags & PAGE_CGROUP_FLAG_CACHE)
162 __mem_cgroup_stat_add_safe(stat,
163 MEM_CGROUP_STAT_CACHE, val);
165 __mem_cgroup_stat_add_safe(stat, MEM_CGROUP_STAT_RSS, val);
169 static struct mem_cgroup init_mem_cgroup;
172 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
174 return container_of(cgroup_subsys_state(cont,
175 mem_cgroup_subsys_id), struct mem_cgroup,
180 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
182 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
183 struct mem_cgroup, css);
186 void mm_init_cgroup(struct mm_struct *mm, struct task_struct *p)
188 struct mem_cgroup *mem;
190 mem = mem_cgroup_from_task(p);
192 mm->mem_cgroup = mem;
195 void mm_free_cgroup(struct mm_struct *mm)
197 css_put(&mm->mem_cgroup->css);
200 static inline int page_cgroup_locked(struct page *page)
202 return bit_spin_is_locked(PAGE_CGROUP_LOCK_BIT,
206 void page_assign_page_cgroup(struct page *page, struct page_cgroup *pc)
211 * While resetting the page_cgroup we might not hold the
212 * page_cgroup lock. free_hot_cold_page() is an example
216 VM_BUG_ON(!page_cgroup_locked(page));
217 locked = (page->page_cgroup & PAGE_CGROUP_LOCK);
218 page->page_cgroup = ((unsigned long)pc | locked);
221 struct page_cgroup *page_get_page_cgroup(struct page *page)
223 return (struct page_cgroup *)
224 (page->page_cgroup & ~PAGE_CGROUP_LOCK);
227 static void __always_inline lock_page_cgroup(struct page *page)
229 bit_spin_lock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
230 VM_BUG_ON(!page_cgroup_locked(page));
233 static void __always_inline unlock_page_cgroup(struct page *page)
235 bit_spin_unlock(PAGE_CGROUP_LOCK_BIT, &page->page_cgroup);
239 * Tie new page_cgroup to struct page under lock_page_cgroup()
240 * This can fail if the page has been tied to a page_cgroup.
241 * If success, returns 0.
243 static int page_cgroup_assign_new_page_cgroup(struct page *page,
244 struct page_cgroup *pc)
248 lock_page_cgroup(page);
249 if (!page_get_page_cgroup(page))
250 page_assign_page_cgroup(page, pc);
251 else /* A page is tied to other pc. */
253 unlock_page_cgroup(page);
258 * Clear page->page_cgroup member under lock_page_cgroup().
259 * If given "pc" value is different from one page->page_cgroup,
260 * page->cgroup is not cleared.
261 * Returns a value of page->page_cgroup at lock taken.
262 * A can can detect failure of clearing by following
263 * clear_page_cgroup(page, pc) == pc
266 static struct page_cgroup *clear_page_cgroup(struct page *page,
267 struct page_cgroup *pc)
269 struct page_cgroup *ret;
271 lock_page_cgroup(page);
272 ret = page_get_page_cgroup(page);
273 if (likely(ret == pc))
274 page_assign_page_cgroup(page, NULL);
275 unlock_page_cgroup(page);
279 static void __mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
282 pc->flags |= PAGE_CGROUP_FLAG_ACTIVE;
283 list_move(&pc->lru, &pc->mem_cgroup->active_list);
285 pc->flags &= ~PAGE_CGROUP_FLAG_ACTIVE;
286 list_move(&pc->lru, &pc->mem_cgroup->inactive_list);
290 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
295 ret = task->mm && mm_cgroup(task->mm) == mem;
301 * This routine assumes that the appropriate zone's lru lock is already held
303 void mem_cgroup_move_lists(struct page_cgroup *pc, bool active)
305 struct mem_cgroup *mem;
309 mem = pc->mem_cgroup;
311 spin_lock(&mem->lru_lock);
312 __mem_cgroup_move_lists(pc, active);
313 spin_unlock(&mem->lru_lock);
316 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
317 struct list_head *dst,
318 unsigned long *scanned, int order,
319 int mode, struct zone *z,
320 struct mem_cgroup *mem_cont,
323 unsigned long nr_taken = 0;
327 struct list_head *src;
328 struct page_cgroup *pc, *tmp;
331 src = &mem_cont->active_list;
333 src = &mem_cont->inactive_list;
335 spin_lock(&mem_cont->lru_lock);
337 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
338 if (scan >= nr_to_scan)
343 if (unlikely(!PageLRU(page)))
346 if (PageActive(page) && !active) {
347 __mem_cgroup_move_lists(pc, true);
350 if (!PageActive(page) && active) {
351 __mem_cgroup_move_lists(pc, false);
357 * TODO: make the active/inactive lists per zone
359 if (page_zone(page) != z)
363 list_move(&pc->lru, &pc_list);
365 if (__isolate_lru_page(page, mode) == 0) {
366 list_move(&page->lru, dst);
371 list_splice(&pc_list, src);
372 spin_unlock(&mem_cont->lru_lock);
379 * Charge the memory controller for page usage.
381 * 0 if the charge was successful
382 * < 0 if the cgroup is over its limit
384 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
385 gfp_t gfp_mask, enum charge_type ctype)
387 struct mem_cgroup *mem;
388 struct page_cgroup *pc;
390 unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
393 * Should page_cgroup's go to their own slab?
394 * One could optimize the performance of the charging routine
395 * by saving a bit in the page_flags and using it as a lock
396 * to see if the cgroup page already has a page_cgroup associated
401 lock_page_cgroup(page);
402 pc = page_get_page_cgroup(page);
404 * The page_cgroup exists and
405 * the page has already been accounted.
408 if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
409 /* this page is under being uncharged ? */
410 unlock_page_cgroup(page);
414 unlock_page_cgroup(page);
418 unlock_page_cgroup(page);
421 pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
426 * We always charge the cgroup the mm_struct belongs to.
427 * The mm_struct's mem_cgroup changes on task migration if the
428 * thread group leader migrates. It's possible that mm is not
429 * set, if so charge the init_mm (happens for pagecache usage).
435 mem = rcu_dereference(mm->mem_cgroup);
437 * For every charge from the cgroup, increment reference
444 * If we created the page_cgroup, we should free it on exceeding
447 while (res_counter_charge(&mem->res, PAGE_SIZE)) {
448 if (!(gfp_mask & __GFP_WAIT))
451 if (try_to_free_mem_cgroup_pages(mem, gfp_mask))
455 * try_to_free_mem_cgroup_pages() might not give us a full
456 * picture of reclaim. Some pages are reclaimed and might be
457 * moved to swap cache or just unmapped from the cgroup.
458 * Check the limit again to see if the reclaim reduced the
459 * current usage of the cgroup before giving up
461 if (res_counter_check_under_limit(&mem->res))
465 mem_cgroup_out_of_memory(mem, gfp_mask);
468 congestion_wait(WRITE, HZ/10);
471 atomic_set(&pc->ref_cnt, 1);
472 pc->mem_cgroup = mem;
474 pc->flags = PAGE_CGROUP_FLAG_ACTIVE;
475 if (ctype == MEM_CGROUP_CHARGE_TYPE_CACHE)
476 pc->flags |= PAGE_CGROUP_FLAG_CACHE;
478 if (!page || page_cgroup_assign_new_page_cgroup(page, pc)) {
480 * Another charge has been added to this page already.
481 * We take lock_page_cgroup(page) again and read
482 * page->cgroup, increment refcnt.... just retry is OK.
484 res_counter_uncharge(&mem->res, PAGE_SIZE);
492 spin_lock_irqsave(&mem->lru_lock, flags);
493 /* Update statistics vector */
494 mem_cgroup_charge_statistics(mem, pc->flags, true);
495 list_add(&pc->lru, &mem->active_list);
496 spin_unlock_irqrestore(&mem->lru_lock, flags);
507 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
510 return mem_cgroup_charge_common(page, mm, gfp_mask,
511 MEM_CGROUP_CHARGE_TYPE_MAPPED);
515 * See if the cached pages should be charged at all?
517 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
521 struct mem_cgroup *mem;
526 mem = rcu_dereference(mm->mem_cgroup);
529 if (mem->control_type == MEM_CGROUP_TYPE_ALL)
530 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
531 MEM_CGROUP_CHARGE_TYPE_CACHE);
537 * Uncharging is always a welcome operation, we never complain, simply
540 void mem_cgroup_uncharge(struct page_cgroup *pc)
542 struct mem_cgroup *mem;
547 * This can handle cases when a page is not charged at all and we
548 * are switching between handling the control_type.
553 if (atomic_dec_and_test(&pc->ref_cnt)) {
556 * get page->cgroup and clear it under lock.
557 * force_empty can drop page->cgroup without checking refcnt.
559 if (clear_page_cgroup(page, pc) == pc) {
560 mem = pc->mem_cgroup;
562 res_counter_uncharge(&mem->res, PAGE_SIZE);
563 spin_lock_irqsave(&mem->lru_lock, flags);
564 list_del_init(&pc->lru);
565 mem_cgroup_charge_statistics(mem, pc->flags, false);
566 spin_unlock_irqrestore(&mem->lru_lock, flags);
572 * Returns non-zero if a page (under migration) has valid page_cgroup member.
573 * Refcnt of page_cgroup is incremented.
576 int mem_cgroup_prepare_migration(struct page *page)
578 struct page_cgroup *pc;
580 lock_page_cgroup(page);
581 pc = page_get_page_cgroup(page);
582 if (pc && atomic_inc_not_zero(&pc->ref_cnt))
584 unlock_page_cgroup(page);
588 void mem_cgroup_end_migration(struct page *page)
590 struct page_cgroup *pc = page_get_page_cgroup(page);
591 mem_cgroup_uncharge(pc);
594 * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
595 * And no race with uncharge() routines because page_cgroup for *page*
596 * has extra one reference by mem_cgroup_prepare_migration.
599 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
601 struct page_cgroup *pc;
603 pc = page_get_page_cgroup(page);
606 if (clear_page_cgroup(page, pc) != pc)
609 lock_page_cgroup(newpage);
610 page_assign_page_cgroup(newpage, pc);
611 unlock_page_cgroup(newpage);
616 * This routine traverse page_cgroup in given list and drop them all.
617 * This routine ignores page_cgroup->ref_cnt.
618 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
620 #define FORCE_UNCHARGE_BATCH (128)
622 mem_cgroup_force_empty_list(struct mem_cgroup *mem, struct list_head *list)
624 struct page_cgroup *pc;
630 count = FORCE_UNCHARGE_BATCH;
631 spin_lock_irqsave(&mem->lru_lock, flags);
633 while (--count && !list_empty(list)) {
634 pc = list_entry(list->prev, struct page_cgroup, lru);
636 /* Avoid race with charge */
637 atomic_set(&pc->ref_cnt, 0);
638 if (clear_page_cgroup(page, pc) == pc) {
640 res_counter_uncharge(&mem->res, PAGE_SIZE);
641 list_del_init(&pc->lru);
642 mem_cgroup_charge_statistics(mem, pc->flags, false);
644 } else /* being uncharged ? ...do relax */
647 spin_unlock_irqrestore(&mem->lru_lock, flags);
648 if (!list_empty(list)) {
656 * make mem_cgroup's charge to be 0 if there is no task.
657 * This enables deleting this mem_cgroup.
660 int mem_cgroup_force_empty(struct mem_cgroup *mem)
665 * page reclaim code (kswapd etc..) will move pages between
666 ` * active_list <-> inactive_list while we don't take a lock.
667 * So, we have to do loop here until all lists are empty.
669 while (!(list_empty(&mem->active_list) &&
670 list_empty(&mem->inactive_list))) {
671 if (atomic_read(&mem->css.cgroup->count) > 0)
673 /* drop all page_cgroup in active_list */
674 mem_cgroup_force_empty_list(mem, &mem->active_list);
675 /* drop all page_cgroup in inactive_list */
676 mem_cgroup_force_empty_list(mem, &mem->inactive_list);
686 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
688 *tmp = memparse(buf, &buf);
693 * Round up the value to the closest page size
695 *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
699 static ssize_t mem_cgroup_read(struct cgroup *cont,
700 struct cftype *cft, struct file *file,
701 char __user *userbuf, size_t nbytes, loff_t *ppos)
703 return res_counter_read(&mem_cgroup_from_cont(cont)->res,
704 cft->private, userbuf, nbytes, ppos,
708 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
709 struct file *file, const char __user *userbuf,
710 size_t nbytes, loff_t *ppos)
712 return res_counter_write(&mem_cgroup_from_cont(cont)->res,
713 cft->private, userbuf, nbytes, ppos,
714 mem_cgroup_write_strategy);
717 static ssize_t mem_control_type_write(struct cgroup *cont,
718 struct cftype *cft, struct file *file,
719 const char __user *userbuf,
720 size_t nbytes, loff_t *pos)
725 struct mem_cgroup *mem;
727 mem = mem_cgroup_from_cont(cont);
728 buf = kmalloc(nbytes + 1, GFP_KERNEL);
735 if (copy_from_user(buf, userbuf, nbytes))
739 tmp = simple_strtoul(buf, &end, 10);
743 if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
746 mem->control_type = tmp;
754 static ssize_t mem_control_type_read(struct cgroup *cont,
756 struct file *file, char __user *userbuf,
757 size_t nbytes, loff_t *ppos)
761 struct mem_cgroup *mem;
763 mem = mem_cgroup_from_cont(cont);
765 val = mem->control_type;
766 s += sprintf(s, "%lu\n", val);
767 return simple_read_from_buffer((void __user *)userbuf, nbytes,
772 static ssize_t mem_force_empty_write(struct cgroup *cont,
773 struct cftype *cft, struct file *file,
774 const char __user *userbuf,
775 size_t nbytes, loff_t *ppos)
777 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
779 ret = mem_cgroup_force_empty(mem);
786 * Note: This should be removed if cgroup supports write-only file.
789 static ssize_t mem_force_empty_read(struct cgroup *cont,
791 struct file *file, char __user *userbuf,
792 size_t nbytes, loff_t *ppos)
798 static const struct mem_cgroup_stat_desc {
801 } mem_cgroup_stat_desc[] = {
802 [MEM_CGROUP_STAT_CACHE] = { "cache", PAGE_SIZE, },
803 [MEM_CGROUP_STAT_RSS] = { "rss", PAGE_SIZE, },
806 static int mem_control_stat_show(struct seq_file *m, void *arg)
808 struct cgroup *cont = m->private;
809 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
810 struct mem_cgroup_stat *stat = &mem_cont->stat;
813 for (i = 0; i < ARRAY_SIZE(stat->cpustat[0].count); i++) {
816 val = mem_cgroup_read_stat(stat, i);
817 val *= mem_cgroup_stat_desc[i].unit;
818 seq_printf(m, "%s %lld\n", mem_cgroup_stat_desc[i].msg,
824 static const struct file_operations mem_control_stat_file_operations = {
827 .release = single_release,
830 static int mem_control_stat_open(struct inode *unused, struct file *file)
833 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
835 file->f_op = &mem_control_stat_file_operations;
836 return single_open(file, mem_control_stat_show, cont);
841 static struct cftype mem_cgroup_files[] = {
843 .name = "usage_in_bytes",
844 .private = RES_USAGE,
845 .read = mem_cgroup_read,
848 .name = "limit_in_bytes",
849 .private = RES_LIMIT,
850 .write = mem_cgroup_write,
851 .read = mem_cgroup_read,
855 .private = RES_FAILCNT,
856 .read = mem_cgroup_read,
859 .name = "control_type",
860 .write = mem_control_type_write,
861 .read = mem_control_type_read,
864 .name = "force_empty",
865 .write = mem_force_empty_write,
866 .read = mem_force_empty_read,
870 .open = mem_control_stat_open,
874 static struct mem_cgroup init_mem_cgroup;
876 static struct cgroup_subsys_state *
877 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
879 struct mem_cgroup *mem;
881 if (unlikely((cont->parent) == NULL)) {
882 mem = &init_mem_cgroup;
883 init_mm.mem_cgroup = mem;
885 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
890 res_counter_init(&mem->res);
891 INIT_LIST_HEAD(&mem->active_list);
892 INIT_LIST_HEAD(&mem->inactive_list);
893 spin_lock_init(&mem->lru_lock);
894 mem->control_type = MEM_CGROUP_TYPE_ALL;
898 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
901 kfree(mem_cgroup_from_cont(cont));
904 static int mem_cgroup_populate(struct cgroup_subsys *ss,
907 return cgroup_add_files(cont, ss, mem_cgroup_files,
908 ARRAY_SIZE(mem_cgroup_files));
911 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
913 struct cgroup *old_cont,
914 struct task_struct *p)
916 struct mm_struct *mm;
917 struct mem_cgroup *mem, *old_mem;
923 mem = mem_cgroup_from_cont(cont);
924 old_mem = mem_cgroup_from_cont(old_cont);
930 * Only thread group leaders are allowed to migrate, the mm_struct is
931 * in effect owned by the leader
933 if (p->tgid != p->pid)
937 rcu_assign_pointer(mm->mem_cgroup, mem);
938 css_put(&old_mem->css);
945 struct cgroup_subsys mem_cgroup_subsys = {
947 .subsys_id = mem_cgroup_subsys_id,
948 .create = mem_cgroup_create,
949 .destroy = mem_cgroup_destroy,
950 .populate = mem_cgroup_populate,
951 .attach = mem_cgroup_move_task,