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))) {
273 if (PageActive(page) && !active) {
274 __mem_cgroup_move_lists(pc, true);
278 if (!PageActive(page) && active) {
279 __mem_cgroup_move_lists(pc, false);
286 * TODO: make the active/inactive lists per zone
288 if (page_zone(page) != z)
292 * Check if the meta page went away from under us
294 if (!list_empty(&pc->lru))
295 list_move(&pc->lru, &pc_list);
299 if (__isolate_lru_page(page, mode) == 0) {
300 list_move(&page->lru, dst);
305 list_splice(&pc_list, src);
306 spin_unlock(&mem_cont->lru_lock);
313 * Charge the memory controller for page usage.
315 * 0 if the charge was successful
316 * < 0 if the cgroup is over its limit
318 int mem_cgroup_charge(struct page *page, struct mm_struct *mm,
321 struct mem_cgroup *mem;
322 struct page_cgroup *pc;
324 unsigned long nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
327 * Should page_cgroup's go to their own slab?
328 * One could optimize the performance of the charging routine
329 * by saving a bit in the page_flags and using it as a lock
330 * to see if the cgroup page already has a page_cgroup associated
334 lock_page_cgroup(page);
335 pc = page_get_page_cgroup(page);
337 * The page_cgroup exists and the page has already been accounted
340 if (unlikely(!atomic_inc_not_zero(&pc->ref_cnt))) {
341 /* this page is under being uncharged ? */
342 unlock_page_cgroup(page);
346 unlock_page_cgroup(page);
351 unlock_page_cgroup(page);
353 pc = kzalloc(sizeof(struct page_cgroup), gfp_mask);
359 * We always charge the cgroup the mm_struct belongs to
360 * the mm_struct's mem_cgroup changes on task migration if the
361 * thread group leader migrates. It's possible that mm is not
362 * set, if so charge the init_mm (happens for pagecache usage).
367 mem = rcu_dereference(mm->mem_cgroup);
369 * For every charge from the cgroup, increment reference
376 * If we created the page_cgroup, we should free it on exceeding
379 while (res_counter_charge(&mem->res, PAGE_SIZE)) {
380 bool is_atomic = gfp_mask & GFP_ATOMIC;
382 * We cannot reclaim under GFP_ATOMIC, fail the charge
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))
400 * Since we control both RSS and cache, we end up with a
401 * very interesting scenario where we end up reclaiming
402 * memory (essentially RSS), since the memory is pushed
403 * to swap cache, we eventually end up adding those
404 * pages back to our list. Hence we give ourselves a
405 * few chances before we fail
407 else if (nr_retries--) {
408 congestion_wait(WRITE, HZ/10);
414 mem_cgroup_out_of_memory(mem, GFP_KERNEL);
418 atomic_set(&pc->ref_cnt, 1);
419 pc->mem_cgroup = mem;
421 if (page_cgroup_assign_new_page_cgroup(page, pc)) {
423 * an another charge is added to this page already.
424 * we do take lock_page_cgroup(page) again and read
425 * page->cgroup, increment refcnt.... just retry is OK.
427 res_counter_uncharge(&mem->res, PAGE_SIZE);
433 spin_lock_irqsave(&mem->lru_lock, flags);
434 list_add(&pc->lru, &mem->active_list);
435 spin_unlock_irqrestore(&mem->lru_lock, flags);
446 * See if the cached pages should be charged at all?
448 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
451 struct mem_cgroup *mem;
455 mem = rcu_dereference(mm->mem_cgroup);
456 if (mem->control_type == MEM_CGROUP_TYPE_ALL)
457 return mem_cgroup_charge(page, mm, gfp_mask);
463 * Uncharging is always a welcome operation, we never complain, simply
466 void mem_cgroup_uncharge(struct page_cgroup *pc)
468 struct mem_cgroup *mem;
473 * This can handle cases when a page is not charged at all and we
474 * are switching between handling the control_type.
479 if (atomic_dec_and_test(&pc->ref_cnt)) {
482 * get page->cgroup and clear it under lock.
484 if (clear_page_cgroup(page, pc) == pc) {
485 mem = pc->mem_cgroup;
487 res_counter_uncharge(&mem->res, PAGE_SIZE);
488 spin_lock_irqsave(&mem->lru_lock, flags);
489 list_del_init(&pc->lru);
490 spin_unlock_irqrestore(&mem->lru_lock, flags);
494 * Note:This will be removed when force-empty patch is
495 * applied. just show warning here.
497 printk(KERN_ERR "Race in mem_cgroup_uncharge() ?");
503 * Returns non-zero if a page (under migration) has valid page_cgroup member.
504 * Refcnt of page_cgroup is incremented.
507 int mem_cgroup_prepare_migration(struct page *page)
509 struct page_cgroup *pc;
511 lock_page_cgroup(page);
512 pc = page_get_page_cgroup(page);
513 if (pc && atomic_inc_not_zero(&pc->ref_cnt))
515 unlock_page_cgroup(page);
519 void mem_cgroup_end_migration(struct page *page)
521 struct page_cgroup *pc = page_get_page_cgroup(page);
522 mem_cgroup_uncharge(pc);
525 * We know both *page* and *newpage* are now not-on-LRU and Pg_locked.
526 * And no race with uncharge() routines because page_cgroup for *page*
527 * has extra one reference by mem_cgroup_prepare_migration.
530 void mem_cgroup_page_migration(struct page *page, struct page *newpage)
532 struct page_cgroup *pc;
534 pc = page_get_page_cgroup(page);
537 if (clear_page_cgroup(page, pc) != pc)
540 lock_page_cgroup(newpage);
541 page_assign_page_cgroup(newpage, pc);
542 unlock_page_cgroup(newpage);
546 int mem_cgroup_write_strategy(char *buf, unsigned long long *tmp)
548 *tmp = memparse(buf, &buf);
553 * Round up the value to the closest page size
555 *tmp = ((*tmp + PAGE_SIZE - 1) >> PAGE_SHIFT) << PAGE_SHIFT;
559 static ssize_t mem_cgroup_read(struct cgroup *cont,
560 struct cftype *cft, struct file *file,
561 char __user *userbuf, size_t nbytes, loff_t *ppos)
563 return res_counter_read(&mem_cgroup_from_cont(cont)->res,
564 cft->private, userbuf, nbytes, ppos,
568 static ssize_t mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
569 struct file *file, const char __user *userbuf,
570 size_t nbytes, loff_t *ppos)
572 return res_counter_write(&mem_cgroup_from_cont(cont)->res,
573 cft->private, userbuf, nbytes, ppos,
574 mem_cgroup_write_strategy);
577 static ssize_t mem_control_type_write(struct cgroup *cont,
578 struct cftype *cft, struct file *file,
579 const char __user *userbuf,
580 size_t nbytes, loff_t *pos)
585 struct mem_cgroup *mem;
587 mem = mem_cgroup_from_cont(cont);
588 buf = kmalloc(nbytes + 1, GFP_KERNEL);
595 if (copy_from_user(buf, userbuf, nbytes))
599 tmp = simple_strtoul(buf, &end, 10);
603 if (tmp <= MEM_CGROUP_TYPE_UNSPEC || tmp >= MEM_CGROUP_TYPE_MAX)
606 mem->control_type = tmp;
614 static ssize_t mem_control_type_read(struct cgroup *cont,
616 struct file *file, char __user *userbuf,
617 size_t nbytes, loff_t *ppos)
621 struct mem_cgroup *mem;
623 mem = mem_cgroup_from_cont(cont);
625 val = mem->control_type;
626 s += sprintf(s, "%lu\n", val);
627 return simple_read_from_buffer((void __user *)userbuf, nbytes,
631 static struct cftype mem_cgroup_files[] = {
633 .name = "usage_in_bytes",
634 .private = RES_USAGE,
635 .read = mem_cgroup_read,
638 .name = "limit_in_bytes",
639 .private = RES_LIMIT,
640 .write = mem_cgroup_write,
641 .read = mem_cgroup_read,
645 .private = RES_FAILCNT,
646 .read = mem_cgroup_read,
649 .name = "control_type",
650 .write = mem_control_type_write,
651 .read = mem_control_type_read,
655 static struct mem_cgroup init_mem_cgroup;
657 static struct cgroup_subsys_state *
658 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
660 struct mem_cgroup *mem;
662 if (unlikely((cont->parent) == NULL)) {
663 mem = &init_mem_cgroup;
664 init_mm.mem_cgroup = mem;
666 mem = kzalloc(sizeof(struct mem_cgroup), GFP_KERNEL);
671 res_counter_init(&mem->res);
672 INIT_LIST_HEAD(&mem->active_list);
673 INIT_LIST_HEAD(&mem->inactive_list);
674 spin_lock_init(&mem->lru_lock);
675 mem->control_type = MEM_CGROUP_TYPE_ALL;
679 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
682 kfree(mem_cgroup_from_cont(cont));
685 static int mem_cgroup_populate(struct cgroup_subsys *ss,
688 return cgroup_add_files(cont, ss, mem_cgroup_files,
689 ARRAY_SIZE(mem_cgroup_files));
692 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
694 struct cgroup *old_cont,
695 struct task_struct *p)
697 struct mm_struct *mm;
698 struct mem_cgroup *mem, *old_mem;
704 mem = mem_cgroup_from_cont(cont);
705 old_mem = mem_cgroup_from_cont(old_cont);
711 * Only thread group leaders are allowed to migrate, the mm_struct is
712 * in effect owned by the leader
714 if (p->tgid != p->pid)
718 rcu_assign_pointer(mm->mem_cgroup, mem);
719 css_put(&old_mem->css);
726 struct cgroup_subsys mem_cgroup_subsys = {
728 .subsys_id = mem_cgroup_subsys_id,
729 .create = mem_cgroup_create,
730 .destroy = mem_cgroup_destroy,
731 .populate = mem_cgroup_populate,
732 .attach = mem_cgroup_move_task,