Merge commit 'jwb/next' into next
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / memcontrol.c
blobf99f5991d6bba1b5ec224f0631124144a6c0a8bb
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>
23 #include <linux/mm.h>
24 #include <linux/pagemap.h>
25 #include <linux/smp.h>
26 #include <linux/page-flags.h>
27 #include <linux/backing-dev.h>
28 #include <linux/bit_spinlock.h>
29 #include <linux/rcupdate.h>
30 #include <linux/limits.h>
31 #include <linux/mutex.h>
32 #include <linux/rbtree.h>
33 #include <linux/slab.h>
34 #include <linux/swap.h>
35 #include <linux/spinlock.h>
36 #include <linux/fs.h>
37 #include <linux/seq_file.h>
38 #include <linux/vmalloc.h>
39 #include <linux/mm_inline.h>
40 #include <linux/page_cgroup.h>
41 #include "internal.h"
43 #include <asm/uaccess.h>
45 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
46 #define MEM_CGROUP_RECLAIM_RETRIES 5
47 struct mem_cgroup *root_mem_cgroup __read_mostly;
49 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
50 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
51 int do_swap_account __read_mostly;
52 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
53 #else
54 #define do_swap_account (0)
55 #endif
57 static DEFINE_MUTEX(memcg_tasklist); /* can be hold under cgroup_mutex */
58 #define SOFTLIMIT_EVENTS_THRESH (1000)
61 * Statistics for memory cgroup.
63 enum mem_cgroup_stat_index {
65 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
67 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
68 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
69 MEM_CGROUP_STAT_MAPPED_FILE, /* # of pages charged as file rss */
70 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */
71 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */
72 MEM_CGROUP_STAT_EVENTS, /* sum of pagein + pageout for internal use */
73 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
75 MEM_CGROUP_STAT_NSTATS,
78 struct mem_cgroup_stat_cpu {
79 s64 count[MEM_CGROUP_STAT_NSTATS];
80 } ____cacheline_aligned_in_smp;
82 struct mem_cgroup_stat {
83 struct mem_cgroup_stat_cpu cpustat[0];
86 static inline void
87 __mem_cgroup_stat_reset_safe(struct mem_cgroup_stat_cpu *stat,
88 enum mem_cgroup_stat_index idx)
90 stat->count[idx] = 0;
93 static inline s64
94 __mem_cgroup_stat_read_local(struct mem_cgroup_stat_cpu *stat,
95 enum mem_cgroup_stat_index idx)
97 return stat->count[idx];
101 * For accounting under irq disable, no need for increment preempt count.
103 static inline void __mem_cgroup_stat_add_safe(struct mem_cgroup_stat_cpu *stat,
104 enum mem_cgroup_stat_index idx, int val)
106 stat->count[idx] += val;
109 static s64 mem_cgroup_read_stat(struct mem_cgroup_stat *stat,
110 enum mem_cgroup_stat_index idx)
112 int cpu;
113 s64 ret = 0;
114 for_each_possible_cpu(cpu)
115 ret += stat->cpustat[cpu].count[idx];
116 return ret;
119 static s64 mem_cgroup_local_usage(struct mem_cgroup_stat *stat)
121 s64 ret;
123 ret = mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_CACHE);
124 ret += mem_cgroup_read_stat(stat, MEM_CGROUP_STAT_RSS);
125 return ret;
129 * per-zone information in memory controller.
131 struct mem_cgroup_per_zone {
133 * spin_lock to protect the per cgroup LRU
135 struct list_head lists[NR_LRU_LISTS];
136 unsigned long count[NR_LRU_LISTS];
138 struct zone_reclaim_stat reclaim_stat;
139 struct rb_node tree_node; /* RB tree node */
140 unsigned long long usage_in_excess;/* Set to the value by which */
141 /* the soft limit is exceeded*/
142 bool on_tree;
143 struct mem_cgroup *mem; /* Back pointer, we cannot */
144 /* use container_of */
146 /* Macro for accessing counter */
147 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
149 struct mem_cgroup_per_node {
150 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
153 struct mem_cgroup_lru_info {
154 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
158 * Cgroups above their limits are maintained in a RB-Tree, independent of
159 * their hierarchy representation
162 struct mem_cgroup_tree_per_zone {
163 struct rb_root rb_root;
164 spinlock_t lock;
167 struct mem_cgroup_tree_per_node {
168 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
171 struct mem_cgroup_tree {
172 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
175 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
178 * The memory controller data structure. The memory controller controls both
179 * page cache and RSS per cgroup. We would eventually like to provide
180 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
181 * to help the administrator determine what knobs to tune.
183 * TODO: Add a water mark for the memory controller. Reclaim will begin when
184 * we hit the water mark. May be even add a low water mark, such that
185 * no reclaim occurs from a cgroup at it's low water mark, this is
186 * a feature that will be implemented much later in the future.
188 struct mem_cgroup {
189 struct cgroup_subsys_state css;
191 * the counter to account for memory usage
193 struct res_counter res;
195 * the counter to account for mem+swap usage.
197 struct res_counter memsw;
199 * Per cgroup active and inactive list, similar to the
200 * per zone LRU lists.
202 struct mem_cgroup_lru_info info;
205 protect against reclaim related member.
207 spinlock_t reclaim_param_lock;
209 int prev_priority; /* for recording reclaim priority */
212 * While reclaiming in a hiearchy, we cache the last child we
213 * reclaimed from.
215 int last_scanned_child;
217 * Should the accounting and control be hierarchical, per subtree?
219 bool use_hierarchy;
220 unsigned long last_oom_jiffies;
221 atomic_t refcnt;
223 unsigned int swappiness;
225 /* set when res.limit == memsw.limit */
226 bool memsw_is_minimum;
229 * statistics. This must be placed at the end of memcg.
231 struct mem_cgroup_stat stat;
235 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
236 * limit reclaim to prevent infinite loops, if they ever occur.
238 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
239 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
241 enum charge_type {
242 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
243 MEM_CGROUP_CHARGE_TYPE_MAPPED,
244 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
245 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
246 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
247 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
248 NR_CHARGE_TYPE,
251 /* only for here (for easy reading.) */
252 #define PCGF_CACHE (1UL << PCG_CACHE)
253 #define PCGF_USED (1UL << PCG_USED)
254 #define PCGF_LOCK (1UL << PCG_LOCK)
255 /* Not used, but added here for completeness */
256 #define PCGF_ACCT (1UL << PCG_ACCT)
258 /* for encoding cft->private value on file */
259 #define _MEM (0)
260 #define _MEMSWAP (1)
261 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
262 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
263 #define MEMFILE_ATTR(val) ((val) & 0xffff)
266 * Reclaim flags for mem_cgroup_hierarchical_reclaim
268 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
269 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
270 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
271 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
272 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
273 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
275 static void mem_cgroup_get(struct mem_cgroup *mem);
276 static void mem_cgroup_put(struct mem_cgroup *mem);
277 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
279 static struct mem_cgroup_per_zone *
280 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
282 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
285 static struct mem_cgroup_per_zone *
286 page_cgroup_zoneinfo(struct page_cgroup *pc)
288 struct mem_cgroup *mem = pc->mem_cgroup;
289 int nid = page_cgroup_nid(pc);
290 int zid = page_cgroup_zid(pc);
292 if (!mem)
293 return NULL;
295 return mem_cgroup_zoneinfo(mem, nid, zid);
298 static struct mem_cgroup_tree_per_zone *
299 soft_limit_tree_node_zone(int nid, int zid)
301 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
304 static struct mem_cgroup_tree_per_zone *
305 soft_limit_tree_from_page(struct page *page)
307 int nid = page_to_nid(page);
308 int zid = page_zonenum(page);
310 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
313 static void
314 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
315 struct mem_cgroup_per_zone *mz,
316 struct mem_cgroup_tree_per_zone *mctz,
317 unsigned long long new_usage_in_excess)
319 struct rb_node **p = &mctz->rb_root.rb_node;
320 struct rb_node *parent = NULL;
321 struct mem_cgroup_per_zone *mz_node;
323 if (mz->on_tree)
324 return;
326 mz->usage_in_excess = new_usage_in_excess;
327 if (!mz->usage_in_excess)
328 return;
329 while (*p) {
330 parent = *p;
331 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
332 tree_node);
333 if (mz->usage_in_excess < mz_node->usage_in_excess)
334 p = &(*p)->rb_left;
336 * We can't avoid mem cgroups that are over their soft
337 * limit by the same amount
339 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
340 p = &(*p)->rb_right;
342 rb_link_node(&mz->tree_node, parent, p);
343 rb_insert_color(&mz->tree_node, &mctz->rb_root);
344 mz->on_tree = true;
347 static void
348 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
349 struct mem_cgroup_per_zone *mz,
350 struct mem_cgroup_tree_per_zone *mctz)
352 if (!mz->on_tree)
353 return;
354 rb_erase(&mz->tree_node, &mctz->rb_root);
355 mz->on_tree = false;
358 static void
359 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
360 struct mem_cgroup_per_zone *mz,
361 struct mem_cgroup_tree_per_zone *mctz)
363 spin_lock(&mctz->lock);
364 __mem_cgroup_remove_exceeded(mem, mz, mctz);
365 spin_unlock(&mctz->lock);
368 static bool mem_cgroup_soft_limit_check(struct mem_cgroup *mem)
370 bool ret = false;
371 int cpu;
372 s64 val;
373 struct mem_cgroup_stat_cpu *cpustat;
375 cpu = get_cpu();
376 cpustat = &mem->stat.cpustat[cpu];
377 val = __mem_cgroup_stat_read_local(cpustat, MEM_CGROUP_STAT_EVENTS);
378 if (unlikely(val > SOFTLIMIT_EVENTS_THRESH)) {
379 __mem_cgroup_stat_reset_safe(cpustat, MEM_CGROUP_STAT_EVENTS);
380 ret = true;
382 put_cpu();
383 return ret;
386 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
388 unsigned long long excess;
389 struct mem_cgroup_per_zone *mz;
390 struct mem_cgroup_tree_per_zone *mctz;
391 int nid = page_to_nid(page);
392 int zid = page_zonenum(page);
393 mctz = soft_limit_tree_from_page(page);
396 * Necessary to update all ancestors when hierarchy is used.
397 * because their event counter is not touched.
399 for (; mem; mem = parent_mem_cgroup(mem)) {
400 mz = mem_cgroup_zoneinfo(mem, nid, zid);
401 excess = res_counter_soft_limit_excess(&mem->res);
403 * We have to update the tree if mz is on RB-tree or
404 * mem is over its softlimit.
406 if (excess || mz->on_tree) {
407 spin_lock(&mctz->lock);
408 /* if on-tree, remove it */
409 if (mz->on_tree)
410 __mem_cgroup_remove_exceeded(mem, mz, mctz);
412 * Insert again. mz->usage_in_excess will be updated.
413 * If excess is 0, no tree ops.
415 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
416 spin_unlock(&mctz->lock);
421 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
423 int node, zone;
424 struct mem_cgroup_per_zone *mz;
425 struct mem_cgroup_tree_per_zone *mctz;
427 for_each_node_state(node, N_POSSIBLE) {
428 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
429 mz = mem_cgroup_zoneinfo(mem, node, zone);
430 mctz = soft_limit_tree_node_zone(node, zone);
431 mem_cgroup_remove_exceeded(mem, mz, mctz);
436 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
438 return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
441 static struct mem_cgroup_per_zone *
442 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
444 struct rb_node *rightmost = NULL;
445 struct mem_cgroup_per_zone *mz;
447 retry:
448 mz = NULL;
449 rightmost = rb_last(&mctz->rb_root);
450 if (!rightmost)
451 goto done; /* Nothing to reclaim from */
453 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
455 * Remove the node now but someone else can add it back,
456 * we will to add it back at the end of reclaim to its correct
457 * position in the tree.
459 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
460 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
461 !css_tryget(&mz->mem->css))
462 goto retry;
463 done:
464 return mz;
467 static struct mem_cgroup_per_zone *
468 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
470 struct mem_cgroup_per_zone *mz;
472 spin_lock(&mctz->lock);
473 mz = __mem_cgroup_largest_soft_limit_node(mctz);
474 spin_unlock(&mctz->lock);
475 return mz;
478 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
479 bool charge)
481 int val = (charge) ? 1 : -1;
482 struct mem_cgroup_stat *stat = &mem->stat;
483 struct mem_cgroup_stat_cpu *cpustat;
484 int cpu = get_cpu();
486 cpustat = &stat->cpustat[cpu];
487 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_SWAPOUT, val);
488 put_cpu();
491 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
492 struct page_cgroup *pc,
493 bool charge)
495 int val = (charge) ? 1 : -1;
496 struct mem_cgroup_stat *stat = &mem->stat;
497 struct mem_cgroup_stat_cpu *cpustat;
498 int cpu = get_cpu();
500 cpustat = &stat->cpustat[cpu];
501 if (PageCgroupCache(pc))
502 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_CACHE, val);
503 else
504 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_RSS, val);
506 if (charge)
507 __mem_cgroup_stat_add_safe(cpustat,
508 MEM_CGROUP_STAT_PGPGIN_COUNT, 1);
509 else
510 __mem_cgroup_stat_add_safe(cpustat,
511 MEM_CGROUP_STAT_PGPGOUT_COUNT, 1);
512 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_EVENTS, 1);
513 put_cpu();
516 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
517 enum lru_list idx)
519 int nid, zid;
520 struct mem_cgroup_per_zone *mz;
521 u64 total = 0;
523 for_each_online_node(nid)
524 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
525 mz = mem_cgroup_zoneinfo(mem, nid, zid);
526 total += MEM_CGROUP_ZSTAT(mz, idx);
528 return total;
531 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
533 return container_of(cgroup_subsys_state(cont,
534 mem_cgroup_subsys_id), struct mem_cgroup,
535 css);
538 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
541 * mm_update_next_owner() may clear mm->owner to NULL
542 * if it races with swapoff, page migration, etc.
543 * So this can be called with p == NULL.
545 if (unlikely(!p))
546 return NULL;
548 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
549 struct mem_cgroup, css);
552 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
554 struct mem_cgroup *mem = NULL;
556 if (!mm)
557 return NULL;
559 * Because we have no locks, mm->owner's may be being moved to other
560 * cgroup. We use css_tryget() here even if this looks
561 * pessimistic (rather than adding locks here).
563 rcu_read_lock();
564 do {
565 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
566 if (unlikely(!mem))
567 break;
568 } while (!css_tryget(&mem->css));
569 rcu_read_unlock();
570 return mem;
574 * Call callback function against all cgroup under hierarchy tree.
576 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
577 int (*func)(struct mem_cgroup *, void *))
579 int found, ret, nextid;
580 struct cgroup_subsys_state *css;
581 struct mem_cgroup *mem;
583 if (!root->use_hierarchy)
584 return (*func)(root, data);
586 nextid = 1;
587 do {
588 ret = 0;
589 mem = NULL;
591 rcu_read_lock();
592 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
593 &found);
594 if (css && css_tryget(css))
595 mem = container_of(css, struct mem_cgroup, css);
596 rcu_read_unlock();
598 if (mem) {
599 ret = (*func)(mem, data);
600 css_put(&mem->css);
602 nextid = found + 1;
603 } while (!ret && css);
605 return ret;
608 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
610 return (mem == root_mem_cgroup);
614 * Following LRU functions are allowed to be used without PCG_LOCK.
615 * Operations are called by routine of global LRU independently from memcg.
616 * What we have to take care of here is validness of pc->mem_cgroup.
618 * Changes to pc->mem_cgroup happens when
619 * 1. charge
620 * 2. moving account
621 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
622 * It is added to LRU before charge.
623 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
624 * When moving account, the page is not on LRU. It's isolated.
627 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
629 struct page_cgroup *pc;
630 struct mem_cgroup_per_zone *mz;
632 if (mem_cgroup_disabled())
633 return;
634 pc = lookup_page_cgroup(page);
635 /* can happen while we handle swapcache. */
636 if (!TestClearPageCgroupAcctLRU(pc))
637 return;
638 VM_BUG_ON(!pc->mem_cgroup);
640 * We don't check PCG_USED bit. It's cleared when the "page" is finally
641 * removed from global LRU.
643 mz = page_cgroup_zoneinfo(pc);
644 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
645 if (mem_cgroup_is_root(pc->mem_cgroup))
646 return;
647 VM_BUG_ON(list_empty(&pc->lru));
648 list_del_init(&pc->lru);
649 return;
652 void mem_cgroup_del_lru(struct page *page)
654 mem_cgroup_del_lru_list(page, page_lru(page));
657 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
659 struct mem_cgroup_per_zone *mz;
660 struct page_cgroup *pc;
662 if (mem_cgroup_disabled())
663 return;
665 pc = lookup_page_cgroup(page);
667 * Used bit is set without atomic ops but after smp_wmb().
668 * For making pc->mem_cgroup visible, insert smp_rmb() here.
670 smp_rmb();
671 /* unused or root page is not rotated. */
672 if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
673 return;
674 mz = page_cgroup_zoneinfo(pc);
675 list_move(&pc->lru, &mz->lists[lru]);
678 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
680 struct page_cgroup *pc;
681 struct mem_cgroup_per_zone *mz;
683 if (mem_cgroup_disabled())
684 return;
685 pc = lookup_page_cgroup(page);
686 VM_BUG_ON(PageCgroupAcctLRU(pc));
688 * Used bit is set without atomic ops but after smp_wmb().
689 * For making pc->mem_cgroup visible, insert smp_rmb() here.
691 smp_rmb();
692 if (!PageCgroupUsed(pc))
693 return;
695 mz = page_cgroup_zoneinfo(pc);
696 MEM_CGROUP_ZSTAT(mz, lru) += 1;
697 SetPageCgroupAcctLRU(pc);
698 if (mem_cgroup_is_root(pc->mem_cgroup))
699 return;
700 list_add(&pc->lru, &mz->lists[lru]);
704 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
705 * lru because the page may.be reused after it's fully uncharged (because of
706 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
707 * it again. This function is only used to charge SwapCache. It's done under
708 * lock_page and expected that zone->lru_lock is never held.
710 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
712 unsigned long flags;
713 struct zone *zone = page_zone(page);
714 struct page_cgroup *pc = lookup_page_cgroup(page);
716 spin_lock_irqsave(&zone->lru_lock, flags);
718 * Forget old LRU when this page_cgroup is *not* used. This Used bit
719 * is guarded by lock_page() because the page is SwapCache.
721 if (!PageCgroupUsed(pc))
722 mem_cgroup_del_lru_list(page, page_lru(page));
723 spin_unlock_irqrestore(&zone->lru_lock, flags);
726 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
728 unsigned long flags;
729 struct zone *zone = page_zone(page);
730 struct page_cgroup *pc = lookup_page_cgroup(page);
732 spin_lock_irqsave(&zone->lru_lock, flags);
733 /* link when the page is linked to LRU but page_cgroup isn't */
734 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
735 mem_cgroup_add_lru_list(page, page_lru(page));
736 spin_unlock_irqrestore(&zone->lru_lock, flags);
740 void mem_cgroup_move_lists(struct page *page,
741 enum lru_list from, enum lru_list to)
743 if (mem_cgroup_disabled())
744 return;
745 mem_cgroup_del_lru_list(page, from);
746 mem_cgroup_add_lru_list(page, to);
749 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
751 int ret;
752 struct mem_cgroup *curr = NULL;
754 task_lock(task);
755 rcu_read_lock();
756 curr = try_get_mem_cgroup_from_mm(task->mm);
757 rcu_read_unlock();
758 task_unlock(task);
759 if (!curr)
760 return 0;
761 if (curr->use_hierarchy)
762 ret = css_is_ancestor(&curr->css, &mem->css);
763 else
764 ret = (curr == mem);
765 css_put(&curr->css);
766 return ret;
770 * prev_priority control...this will be used in memory reclaim path.
772 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
774 int prev_priority;
776 spin_lock(&mem->reclaim_param_lock);
777 prev_priority = mem->prev_priority;
778 spin_unlock(&mem->reclaim_param_lock);
780 return prev_priority;
783 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
785 spin_lock(&mem->reclaim_param_lock);
786 if (priority < mem->prev_priority)
787 mem->prev_priority = priority;
788 spin_unlock(&mem->reclaim_param_lock);
791 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
793 spin_lock(&mem->reclaim_param_lock);
794 mem->prev_priority = priority;
795 spin_unlock(&mem->reclaim_param_lock);
798 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
800 unsigned long active;
801 unsigned long inactive;
802 unsigned long gb;
803 unsigned long inactive_ratio;
805 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
806 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
808 gb = (inactive + active) >> (30 - PAGE_SHIFT);
809 if (gb)
810 inactive_ratio = int_sqrt(10 * gb);
811 else
812 inactive_ratio = 1;
814 if (present_pages) {
815 present_pages[0] = inactive;
816 present_pages[1] = active;
819 return inactive_ratio;
822 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
824 unsigned long active;
825 unsigned long inactive;
826 unsigned long present_pages[2];
827 unsigned long inactive_ratio;
829 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
831 inactive = present_pages[0];
832 active = present_pages[1];
834 if (inactive * inactive_ratio < active)
835 return 1;
837 return 0;
840 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
842 unsigned long active;
843 unsigned long inactive;
845 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
846 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
848 return (active > inactive);
851 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
852 struct zone *zone,
853 enum lru_list lru)
855 int nid = zone->zone_pgdat->node_id;
856 int zid = zone_idx(zone);
857 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
859 return MEM_CGROUP_ZSTAT(mz, lru);
862 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
863 struct zone *zone)
865 int nid = zone->zone_pgdat->node_id;
866 int zid = zone_idx(zone);
867 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
869 return &mz->reclaim_stat;
872 struct zone_reclaim_stat *
873 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
875 struct page_cgroup *pc;
876 struct mem_cgroup_per_zone *mz;
878 if (mem_cgroup_disabled())
879 return NULL;
881 pc = lookup_page_cgroup(page);
883 * Used bit is set without atomic ops but after smp_wmb().
884 * For making pc->mem_cgroup visible, insert smp_rmb() here.
886 smp_rmb();
887 if (!PageCgroupUsed(pc))
888 return NULL;
890 mz = page_cgroup_zoneinfo(pc);
891 if (!mz)
892 return NULL;
894 return &mz->reclaim_stat;
897 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
898 struct list_head *dst,
899 unsigned long *scanned, int order,
900 int mode, struct zone *z,
901 struct mem_cgroup *mem_cont,
902 int active, int file)
904 unsigned long nr_taken = 0;
905 struct page *page;
906 unsigned long scan;
907 LIST_HEAD(pc_list);
908 struct list_head *src;
909 struct page_cgroup *pc, *tmp;
910 int nid = z->zone_pgdat->node_id;
911 int zid = zone_idx(z);
912 struct mem_cgroup_per_zone *mz;
913 int lru = LRU_FILE * file + active;
914 int ret;
916 BUG_ON(!mem_cont);
917 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
918 src = &mz->lists[lru];
920 scan = 0;
921 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
922 if (scan >= nr_to_scan)
923 break;
925 page = pc->page;
926 if (unlikely(!PageCgroupUsed(pc)))
927 continue;
928 if (unlikely(!PageLRU(page)))
929 continue;
931 scan++;
932 ret = __isolate_lru_page(page, mode, file);
933 switch (ret) {
934 case 0:
935 list_move(&page->lru, dst);
936 mem_cgroup_del_lru(page);
937 nr_taken++;
938 break;
939 case -EBUSY:
940 /* we don't affect global LRU but rotate in our LRU */
941 mem_cgroup_rotate_lru_list(page, page_lru(page));
942 break;
943 default:
944 break;
948 *scanned = scan;
949 return nr_taken;
952 #define mem_cgroup_from_res_counter(counter, member) \
953 container_of(counter, struct mem_cgroup, member)
955 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
957 if (do_swap_account) {
958 if (res_counter_check_under_limit(&mem->res) &&
959 res_counter_check_under_limit(&mem->memsw))
960 return true;
961 } else
962 if (res_counter_check_under_limit(&mem->res))
963 return true;
964 return false;
967 static unsigned int get_swappiness(struct mem_cgroup *memcg)
969 struct cgroup *cgrp = memcg->css.cgroup;
970 unsigned int swappiness;
972 /* root ? */
973 if (cgrp->parent == NULL)
974 return vm_swappiness;
976 spin_lock(&memcg->reclaim_param_lock);
977 swappiness = memcg->swappiness;
978 spin_unlock(&memcg->reclaim_param_lock);
980 return swappiness;
983 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
985 int *val = data;
986 (*val)++;
987 return 0;
991 * mem_cgroup_print_mem_info: Called from OOM with tasklist_lock held in read mode.
992 * @memcg: The memory cgroup that went over limit
993 * @p: Task that is going to be killed
995 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
996 * enabled
998 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1000 struct cgroup *task_cgrp;
1001 struct cgroup *mem_cgrp;
1003 * Need a buffer in BSS, can't rely on allocations. The code relies
1004 * on the assumption that OOM is serialized for memory controller.
1005 * If this assumption is broken, revisit this code.
1007 static char memcg_name[PATH_MAX];
1008 int ret;
1010 if (!memcg)
1011 return;
1014 rcu_read_lock();
1016 mem_cgrp = memcg->css.cgroup;
1017 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1019 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1020 if (ret < 0) {
1022 * Unfortunately, we are unable to convert to a useful name
1023 * But we'll still print out the usage information
1025 rcu_read_unlock();
1026 goto done;
1028 rcu_read_unlock();
1030 printk(KERN_INFO "Task in %s killed", memcg_name);
1032 rcu_read_lock();
1033 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1034 if (ret < 0) {
1035 rcu_read_unlock();
1036 goto done;
1038 rcu_read_unlock();
1041 * Continues from above, so we don't need an KERN_ level
1043 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1044 done:
1046 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1047 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1048 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1049 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1050 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1051 "failcnt %llu\n",
1052 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1053 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1054 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1058 * This function returns the number of memcg under hierarchy tree. Returns
1059 * 1(self count) if no children.
1061 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1063 int num = 0;
1064 mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1065 return num;
1069 * Visit the first child (need not be the first child as per the ordering
1070 * of the cgroup list, since we track last_scanned_child) of @mem and use
1071 * that to reclaim free pages from.
1073 static struct mem_cgroup *
1074 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1076 struct mem_cgroup *ret = NULL;
1077 struct cgroup_subsys_state *css;
1078 int nextid, found;
1080 if (!root_mem->use_hierarchy) {
1081 css_get(&root_mem->css);
1082 ret = root_mem;
1085 while (!ret) {
1086 rcu_read_lock();
1087 nextid = root_mem->last_scanned_child + 1;
1088 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1089 &found);
1090 if (css && css_tryget(css))
1091 ret = container_of(css, struct mem_cgroup, css);
1093 rcu_read_unlock();
1094 /* Updates scanning parameter */
1095 spin_lock(&root_mem->reclaim_param_lock);
1096 if (!css) {
1097 /* this means start scan from ID:1 */
1098 root_mem->last_scanned_child = 0;
1099 } else
1100 root_mem->last_scanned_child = found;
1101 spin_unlock(&root_mem->reclaim_param_lock);
1104 return ret;
1108 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1109 * we reclaimed from, so that we don't end up penalizing one child extensively
1110 * based on its position in the children list.
1112 * root_mem is the original ancestor that we've been reclaim from.
1114 * We give up and return to the caller when we visit root_mem twice.
1115 * (other groups can be removed while we're walking....)
1117 * If shrink==true, for avoiding to free too much, this returns immedieately.
1119 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1120 struct zone *zone,
1121 gfp_t gfp_mask,
1122 unsigned long reclaim_options)
1124 struct mem_cgroup *victim;
1125 int ret, total = 0;
1126 int loop = 0;
1127 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1128 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1129 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1130 unsigned long excess = mem_cgroup_get_excess(root_mem);
1132 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1133 if (root_mem->memsw_is_minimum)
1134 noswap = true;
1136 while (1) {
1137 victim = mem_cgroup_select_victim(root_mem);
1138 if (victim == root_mem) {
1139 loop++;
1140 if (loop >= 2) {
1142 * If we have not been able to reclaim
1143 * anything, it might because there are
1144 * no reclaimable pages under this hierarchy
1146 if (!check_soft || !total) {
1147 css_put(&victim->css);
1148 break;
1151 * We want to do more targetted reclaim.
1152 * excess >> 2 is not to excessive so as to
1153 * reclaim too much, nor too less that we keep
1154 * coming back to reclaim from this cgroup
1156 if (total >= (excess >> 2) ||
1157 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1158 css_put(&victim->css);
1159 break;
1163 if (!mem_cgroup_local_usage(&victim->stat)) {
1164 /* this cgroup's local usage == 0 */
1165 css_put(&victim->css);
1166 continue;
1168 /* we use swappiness of local cgroup */
1169 if (check_soft)
1170 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1171 noswap, get_swappiness(victim), zone,
1172 zone->zone_pgdat->node_id);
1173 else
1174 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1175 noswap, get_swappiness(victim));
1176 css_put(&victim->css);
1178 * At shrinking usage, we can't check we should stop here or
1179 * reclaim more. It's depends on callers. last_scanned_child
1180 * will work enough for keeping fairness under tree.
1182 if (shrink)
1183 return ret;
1184 total += ret;
1185 if (check_soft) {
1186 if (res_counter_check_under_soft_limit(&root_mem->res))
1187 return total;
1188 } else if (mem_cgroup_check_under_limit(root_mem))
1189 return 1 + total;
1191 return total;
1194 bool mem_cgroup_oom_called(struct task_struct *task)
1196 bool ret = false;
1197 struct mem_cgroup *mem;
1198 struct mm_struct *mm;
1200 rcu_read_lock();
1201 mm = task->mm;
1202 if (!mm)
1203 mm = &init_mm;
1204 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
1205 if (mem && time_before(jiffies, mem->last_oom_jiffies + HZ/10))
1206 ret = true;
1207 rcu_read_unlock();
1208 return ret;
1211 static int record_last_oom_cb(struct mem_cgroup *mem, void *data)
1213 mem->last_oom_jiffies = jiffies;
1214 return 0;
1217 static void record_last_oom(struct mem_cgroup *mem)
1219 mem_cgroup_walk_tree(mem, NULL, record_last_oom_cb);
1223 * Currently used to update mapped file statistics, but the routine can be
1224 * generalized to update other statistics as well.
1226 void mem_cgroup_update_mapped_file_stat(struct page *page, int val)
1228 struct mem_cgroup *mem;
1229 struct mem_cgroup_stat *stat;
1230 struct mem_cgroup_stat_cpu *cpustat;
1231 int cpu;
1232 struct page_cgroup *pc;
1234 if (!page_is_file_cache(page))
1235 return;
1237 pc = lookup_page_cgroup(page);
1238 if (unlikely(!pc))
1239 return;
1241 lock_page_cgroup(pc);
1242 mem = pc->mem_cgroup;
1243 if (!mem)
1244 goto done;
1246 if (!PageCgroupUsed(pc))
1247 goto done;
1250 * Preemption is already disabled, we don't need get_cpu()
1252 cpu = smp_processor_id();
1253 stat = &mem->stat;
1254 cpustat = &stat->cpustat[cpu];
1256 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE, val);
1257 done:
1258 unlock_page_cgroup(pc);
1262 * Unlike exported interface, "oom" parameter is added. if oom==true,
1263 * oom-killer can be invoked.
1265 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1266 gfp_t gfp_mask, struct mem_cgroup **memcg,
1267 bool oom, struct page *page)
1269 struct mem_cgroup *mem, *mem_over_limit;
1270 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1271 struct res_counter *fail_res;
1273 if (unlikely(test_thread_flag(TIF_MEMDIE))) {
1274 /* Don't account this! */
1275 *memcg = NULL;
1276 return 0;
1280 * We always charge the cgroup the mm_struct belongs to.
1281 * The mm_struct's mem_cgroup changes on task migration if the
1282 * thread group leader migrates. It's possible that mm is not
1283 * set, if so charge the init_mm (happens for pagecache usage).
1285 mem = *memcg;
1286 if (likely(!mem)) {
1287 mem = try_get_mem_cgroup_from_mm(mm);
1288 *memcg = mem;
1289 } else {
1290 css_get(&mem->css);
1292 if (unlikely(!mem))
1293 return 0;
1295 VM_BUG_ON(css_is_removed(&mem->css));
1297 while (1) {
1298 int ret = 0;
1299 unsigned long flags = 0;
1301 if (mem_cgroup_is_root(mem))
1302 goto done;
1303 ret = res_counter_charge(&mem->res, PAGE_SIZE, &fail_res);
1304 if (likely(!ret)) {
1305 if (!do_swap_account)
1306 break;
1307 ret = res_counter_charge(&mem->memsw, PAGE_SIZE,
1308 &fail_res);
1309 if (likely(!ret))
1310 break;
1311 /* mem+swap counter fails */
1312 res_counter_uncharge(&mem->res, PAGE_SIZE);
1313 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1314 mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1315 memsw);
1316 } else
1317 /* mem counter fails */
1318 mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1319 res);
1321 if (!(gfp_mask & __GFP_WAIT))
1322 goto nomem;
1324 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1325 gfp_mask, flags);
1326 if (ret)
1327 continue;
1330 * try_to_free_mem_cgroup_pages() might not give us a full
1331 * picture of reclaim. Some pages are reclaimed and might be
1332 * moved to swap cache or just unmapped from the cgroup.
1333 * Check the limit again to see if the reclaim reduced the
1334 * current usage of the cgroup before giving up
1337 if (mem_cgroup_check_under_limit(mem_over_limit))
1338 continue;
1340 if (!nr_retries--) {
1341 if (oom) {
1342 mutex_lock(&memcg_tasklist);
1343 mem_cgroup_out_of_memory(mem_over_limit, gfp_mask);
1344 mutex_unlock(&memcg_tasklist);
1345 record_last_oom(mem_over_limit);
1347 goto nomem;
1351 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1352 * if they exceeds softlimit.
1354 if (mem_cgroup_soft_limit_check(mem))
1355 mem_cgroup_update_tree(mem, page);
1356 done:
1357 return 0;
1358 nomem:
1359 css_put(&mem->css);
1360 return -ENOMEM;
1364 * A helper function to get mem_cgroup from ID. must be called under
1365 * rcu_read_lock(). The caller must check css_is_removed() or some if
1366 * it's concern. (dropping refcnt from swap can be called against removed
1367 * memcg.)
1369 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1371 struct cgroup_subsys_state *css;
1373 /* ID 0 is unused ID */
1374 if (!id)
1375 return NULL;
1376 css = css_lookup(&mem_cgroup_subsys, id);
1377 if (!css)
1378 return NULL;
1379 return container_of(css, struct mem_cgroup, css);
1382 static struct mem_cgroup *try_get_mem_cgroup_from_swapcache(struct page *page)
1384 struct mem_cgroup *mem;
1385 struct page_cgroup *pc;
1386 unsigned short id;
1387 swp_entry_t ent;
1389 VM_BUG_ON(!PageLocked(page));
1391 if (!PageSwapCache(page))
1392 return NULL;
1394 pc = lookup_page_cgroup(page);
1395 lock_page_cgroup(pc);
1396 if (PageCgroupUsed(pc)) {
1397 mem = pc->mem_cgroup;
1398 if (mem && !css_tryget(&mem->css))
1399 mem = NULL;
1400 } else {
1401 ent.val = page_private(page);
1402 id = lookup_swap_cgroup(ent);
1403 rcu_read_lock();
1404 mem = mem_cgroup_lookup(id);
1405 if (mem && !css_tryget(&mem->css))
1406 mem = NULL;
1407 rcu_read_unlock();
1409 unlock_page_cgroup(pc);
1410 return mem;
1414 * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1415 * USED state. If already USED, uncharge and return.
1418 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1419 struct page_cgroup *pc,
1420 enum charge_type ctype)
1422 /* try_charge() can return NULL to *memcg, taking care of it. */
1423 if (!mem)
1424 return;
1426 lock_page_cgroup(pc);
1427 if (unlikely(PageCgroupUsed(pc))) {
1428 unlock_page_cgroup(pc);
1429 if (!mem_cgroup_is_root(mem)) {
1430 res_counter_uncharge(&mem->res, PAGE_SIZE);
1431 if (do_swap_account)
1432 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1434 css_put(&mem->css);
1435 return;
1438 pc->mem_cgroup = mem;
1440 * We access a page_cgroup asynchronously without lock_page_cgroup().
1441 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1442 * is accessed after testing USED bit. To make pc->mem_cgroup visible
1443 * before USED bit, we need memory barrier here.
1444 * See mem_cgroup_add_lru_list(), etc.
1446 smp_wmb();
1447 switch (ctype) {
1448 case MEM_CGROUP_CHARGE_TYPE_CACHE:
1449 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1450 SetPageCgroupCache(pc);
1451 SetPageCgroupUsed(pc);
1452 break;
1453 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1454 ClearPageCgroupCache(pc);
1455 SetPageCgroupUsed(pc);
1456 break;
1457 default:
1458 break;
1461 mem_cgroup_charge_statistics(mem, pc, true);
1463 unlock_page_cgroup(pc);
1467 * mem_cgroup_move_account - move account of the page
1468 * @pc: page_cgroup of the page.
1469 * @from: mem_cgroup which the page is moved from.
1470 * @to: mem_cgroup which the page is moved to. @from != @to.
1472 * The caller must confirm following.
1473 * - page is not on LRU (isolate_page() is useful.)
1475 * returns 0 at success,
1476 * returns -EBUSY when lock is busy or "pc" is unstable.
1478 * This function does "uncharge" from old cgroup but doesn't do "charge" to
1479 * new cgroup. It should be done by a caller.
1482 static int mem_cgroup_move_account(struct page_cgroup *pc,
1483 struct mem_cgroup *from, struct mem_cgroup *to)
1485 struct mem_cgroup_per_zone *from_mz, *to_mz;
1486 int nid, zid;
1487 int ret = -EBUSY;
1488 struct page *page;
1489 int cpu;
1490 struct mem_cgroup_stat *stat;
1491 struct mem_cgroup_stat_cpu *cpustat;
1493 VM_BUG_ON(from == to);
1494 VM_BUG_ON(PageLRU(pc->page));
1496 nid = page_cgroup_nid(pc);
1497 zid = page_cgroup_zid(pc);
1498 from_mz = mem_cgroup_zoneinfo(from, nid, zid);
1499 to_mz = mem_cgroup_zoneinfo(to, nid, zid);
1501 if (!trylock_page_cgroup(pc))
1502 return ret;
1504 if (!PageCgroupUsed(pc))
1505 goto out;
1507 if (pc->mem_cgroup != from)
1508 goto out;
1510 if (!mem_cgroup_is_root(from))
1511 res_counter_uncharge(&from->res, PAGE_SIZE);
1512 mem_cgroup_charge_statistics(from, pc, false);
1514 page = pc->page;
1515 if (page_is_file_cache(page) && page_mapped(page)) {
1516 cpu = smp_processor_id();
1517 /* Update mapped_file data for mem_cgroup "from" */
1518 stat = &from->stat;
1519 cpustat = &stat->cpustat[cpu];
1520 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1521 -1);
1523 /* Update mapped_file data for mem_cgroup "to" */
1524 stat = &to->stat;
1525 cpustat = &stat->cpustat[cpu];
1526 __mem_cgroup_stat_add_safe(cpustat, MEM_CGROUP_STAT_MAPPED_FILE,
1530 if (do_swap_account && !mem_cgroup_is_root(from))
1531 res_counter_uncharge(&from->memsw, PAGE_SIZE);
1532 css_put(&from->css);
1534 css_get(&to->css);
1535 pc->mem_cgroup = to;
1536 mem_cgroup_charge_statistics(to, pc, true);
1537 ret = 0;
1538 out:
1539 unlock_page_cgroup(pc);
1541 * We charges against "to" which may not have any tasks. Then, "to"
1542 * can be under rmdir(). But in current implementation, caller of
1543 * this function is just force_empty() and it's garanteed that
1544 * "to" is never removed. So, we don't check rmdir status here.
1546 return ret;
1550 * move charges to its parent.
1553 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1554 struct mem_cgroup *child,
1555 gfp_t gfp_mask)
1557 struct page *page = pc->page;
1558 struct cgroup *cg = child->css.cgroup;
1559 struct cgroup *pcg = cg->parent;
1560 struct mem_cgroup *parent;
1561 int ret;
1563 /* Is ROOT ? */
1564 if (!pcg)
1565 return -EINVAL;
1568 parent = mem_cgroup_from_cont(pcg);
1571 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false, page);
1572 if (ret || !parent)
1573 return ret;
1575 if (!get_page_unless_zero(page)) {
1576 ret = -EBUSY;
1577 goto uncharge;
1580 ret = isolate_lru_page(page);
1582 if (ret)
1583 goto cancel;
1585 ret = mem_cgroup_move_account(pc, child, parent);
1587 putback_lru_page(page);
1588 if (!ret) {
1589 put_page(page);
1590 /* drop extra refcnt by try_charge() */
1591 css_put(&parent->css);
1592 return 0;
1595 cancel:
1596 put_page(page);
1597 uncharge:
1598 /* drop extra refcnt by try_charge() */
1599 css_put(&parent->css);
1600 /* uncharge if move fails */
1601 if (!mem_cgroup_is_root(parent)) {
1602 res_counter_uncharge(&parent->res, PAGE_SIZE);
1603 if (do_swap_account)
1604 res_counter_uncharge(&parent->memsw, PAGE_SIZE);
1606 return ret;
1610 * Charge the memory controller for page usage.
1611 * Return
1612 * 0 if the charge was successful
1613 * < 0 if the cgroup is over its limit
1615 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1616 gfp_t gfp_mask, enum charge_type ctype,
1617 struct mem_cgroup *memcg)
1619 struct mem_cgroup *mem;
1620 struct page_cgroup *pc;
1621 int ret;
1623 pc = lookup_page_cgroup(page);
1624 /* can happen at boot */
1625 if (unlikely(!pc))
1626 return 0;
1627 prefetchw(pc);
1629 mem = memcg;
1630 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true, page);
1631 if (ret || !mem)
1632 return ret;
1634 __mem_cgroup_commit_charge(mem, pc, ctype);
1635 return 0;
1638 int mem_cgroup_newpage_charge(struct page *page,
1639 struct mm_struct *mm, gfp_t gfp_mask)
1641 if (mem_cgroup_disabled())
1642 return 0;
1643 if (PageCompound(page))
1644 return 0;
1646 * If already mapped, we don't have to account.
1647 * If page cache, page->mapping has address_space.
1648 * But page->mapping may have out-of-use anon_vma pointer,
1649 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
1650 * is NULL.
1652 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
1653 return 0;
1654 if (unlikely(!mm))
1655 mm = &init_mm;
1656 return mem_cgroup_charge_common(page, mm, gfp_mask,
1657 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
1660 static void
1661 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1662 enum charge_type ctype);
1664 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
1665 gfp_t gfp_mask)
1667 struct mem_cgroup *mem = NULL;
1668 int ret;
1670 if (mem_cgroup_disabled())
1671 return 0;
1672 if (PageCompound(page))
1673 return 0;
1675 * Corner case handling. This is called from add_to_page_cache()
1676 * in usual. But some FS (shmem) precharges this page before calling it
1677 * and call add_to_page_cache() with GFP_NOWAIT.
1679 * For GFP_NOWAIT case, the page may be pre-charged before calling
1680 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
1681 * charge twice. (It works but has to pay a bit larger cost.)
1682 * And when the page is SwapCache, it should take swap information
1683 * into account. This is under lock_page() now.
1685 if (!(gfp_mask & __GFP_WAIT)) {
1686 struct page_cgroup *pc;
1689 pc = lookup_page_cgroup(page);
1690 if (!pc)
1691 return 0;
1692 lock_page_cgroup(pc);
1693 if (PageCgroupUsed(pc)) {
1694 unlock_page_cgroup(pc);
1695 return 0;
1697 unlock_page_cgroup(pc);
1700 if (unlikely(!mm && !mem))
1701 mm = &init_mm;
1703 if (page_is_file_cache(page))
1704 return mem_cgroup_charge_common(page, mm, gfp_mask,
1705 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
1707 /* shmem */
1708 if (PageSwapCache(page)) {
1709 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
1710 if (!ret)
1711 __mem_cgroup_commit_charge_swapin(page, mem,
1712 MEM_CGROUP_CHARGE_TYPE_SHMEM);
1713 } else
1714 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
1715 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
1717 return ret;
1721 * While swap-in, try_charge -> commit or cancel, the page is locked.
1722 * And when try_charge() successfully returns, one refcnt to memcg without
1723 * struct page_cgroup is aquired. This refcnt will be cumsumed by
1724 * "commit()" or removed by "cancel()"
1726 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
1727 struct page *page,
1728 gfp_t mask, struct mem_cgroup **ptr)
1730 struct mem_cgroup *mem;
1731 int ret;
1733 if (mem_cgroup_disabled())
1734 return 0;
1736 if (!do_swap_account)
1737 goto charge_cur_mm;
1739 * A racing thread's fault, or swapoff, may have already updated
1740 * the pte, and even removed page from swap cache: return success
1741 * to go on to do_swap_page()'s pte_same() test, which should fail.
1743 if (!PageSwapCache(page))
1744 return 0;
1745 mem = try_get_mem_cgroup_from_swapcache(page);
1746 if (!mem)
1747 goto charge_cur_mm;
1748 *ptr = mem;
1749 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, page);
1750 /* drop extra refcnt from tryget */
1751 css_put(&mem->css);
1752 return ret;
1753 charge_cur_mm:
1754 if (unlikely(!mm))
1755 mm = &init_mm;
1756 return __mem_cgroup_try_charge(mm, mask, ptr, true, page);
1759 static void
1760 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
1761 enum charge_type ctype)
1763 struct page_cgroup *pc;
1765 if (mem_cgroup_disabled())
1766 return;
1767 if (!ptr)
1768 return;
1769 cgroup_exclude_rmdir(&ptr->css);
1770 pc = lookup_page_cgroup(page);
1771 mem_cgroup_lru_del_before_commit_swapcache(page);
1772 __mem_cgroup_commit_charge(ptr, pc, ctype);
1773 mem_cgroup_lru_add_after_commit_swapcache(page);
1775 * Now swap is on-memory. This means this page may be
1776 * counted both as mem and swap....double count.
1777 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
1778 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
1779 * may call delete_from_swap_cache() before reach here.
1781 if (do_swap_account && PageSwapCache(page)) {
1782 swp_entry_t ent = {.val = page_private(page)};
1783 unsigned short id;
1784 struct mem_cgroup *memcg;
1786 id = swap_cgroup_record(ent, 0);
1787 rcu_read_lock();
1788 memcg = mem_cgroup_lookup(id);
1789 if (memcg) {
1791 * This recorded memcg can be obsolete one. So, avoid
1792 * calling css_tryget
1794 if (!mem_cgroup_is_root(memcg))
1795 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1796 mem_cgroup_swap_statistics(memcg, false);
1797 mem_cgroup_put(memcg);
1799 rcu_read_unlock();
1802 * At swapin, we may charge account against cgroup which has no tasks.
1803 * So, rmdir()->pre_destroy() can be called while we do this charge.
1804 * In that case, we need to call pre_destroy() again. check it here.
1806 cgroup_release_and_wakeup_rmdir(&ptr->css);
1809 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
1811 __mem_cgroup_commit_charge_swapin(page, ptr,
1812 MEM_CGROUP_CHARGE_TYPE_MAPPED);
1815 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
1817 if (mem_cgroup_disabled())
1818 return;
1819 if (!mem)
1820 return;
1821 if (!mem_cgroup_is_root(mem)) {
1822 res_counter_uncharge(&mem->res, PAGE_SIZE);
1823 if (do_swap_account)
1824 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1826 css_put(&mem->css);
1831 * uncharge if !page_mapped(page)
1833 static struct mem_cgroup *
1834 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
1836 struct page_cgroup *pc;
1837 struct mem_cgroup *mem = NULL;
1838 struct mem_cgroup_per_zone *mz;
1840 if (mem_cgroup_disabled())
1841 return NULL;
1843 if (PageSwapCache(page))
1844 return NULL;
1847 * Check if our page_cgroup is valid
1849 pc = lookup_page_cgroup(page);
1850 if (unlikely(!pc || !PageCgroupUsed(pc)))
1851 return NULL;
1853 lock_page_cgroup(pc);
1855 mem = pc->mem_cgroup;
1857 if (!PageCgroupUsed(pc))
1858 goto unlock_out;
1860 switch (ctype) {
1861 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1862 case MEM_CGROUP_CHARGE_TYPE_DROP:
1863 if (page_mapped(page))
1864 goto unlock_out;
1865 break;
1866 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
1867 if (!PageAnon(page)) { /* Shared memory */
1868 if (page->mapping && !page_is_file_cache(page))
1869 goto unlock_out;
1870 } else if (page_mapped(page)) /* Anon */
1871 goto unlock_out;
1872 break;
1873 default:
1874 break;
1877 if (!mem_cgroup_is_root(mem)) {
1878 res_counter_uncharge(&mem->res, PAGE_SIZE);
1879 if (do_swap_account &&
1880 (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT))
1881 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
1883 if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1884 mem_cgroup_swap_statistics(mem, true);
1885 mem_cgroup_charge_statistics(mem, pc, false);
1887 ClearPageCgroupUsed(pc);
1889 * pc->mem_cgroup is not cleared here. It will be accessed when it's
1890 * freed from LRU. This is safe because uncharged page is expected not
1891 * to be reused (freed soon). Exception is SwapCache, it's handled by
1892 * special functions.
1895 mz = page_cgroup_zoneinfo(pc);
1896 unlock_page_cgroup(pc);
1898 if (mem_cgroup_soft_limit_check(mem))
1899 mem_cgroup_update_tree(mem, page);
1900 /* at swapout, this memcg will be accessed to record to swap */
1901 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
1902 css_put(&mem->css);
1904 return mem;
1906 unlock_out:
1907 unlock_page_cgroup(pc);
1908 return NULL;
1911 void mem_cgroup_uncharge_page(struct page *page)
1913 /* early check. */
1914 if (page_mapped(page))
1915 return;
1916 if (page->mapping && !PageAnon(page))
1917 return;
1918 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
1921 void mem_cgroup_uncharge_cache_page(struct page *page)
1923 VM_BUG_ON(page_mapped(page));
1924 VM_BUG_ON(page->mapping);
1925 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
1928 #ifdef CONFIG_SWAP
1930 * called after __delete_from_swap_cache() and drop "page" account.
1931 * memcg information is recorded to swap_cgroup of "ent"
1933 void
1934 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
1936 struct mem_cgroup *memcg;
1937 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
1939 if (!swapout) /* this was a swap cache but the swap is unused ! */
1940 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
1942 memcg = __mem_cgroup_uncharge_common(page, ctype);
1944 /* record memcg information */
1945 if (do_swap_account && swapout && memcg) {
1946 swap_cgroup_record(ent, css_id(&memcg->css));
1947 mem_cgroup_get(memcg);
1949 if (swapout && memcg)
1950 css_put(&memcg->css);
1952 #endif
1954 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
1956 * called from swap_entry_free(). remove record in swap_cgroup and
1957 * uncharge "memsw" account.
1959 void mem_cgroup_uncharge_swap(swp_entry_t ent)
1961 struct mem_cgroup *memcg;
1962 unsigned short id;
1964 if (!do_swap_account)
1965 return;
1967 id = swap_cgroup_record(ent, 0);
1968 rcu_read_lock();
1969 memcg = mem_cgroup_lookup(id);
1970 if (memcg) {
1972 * We uncharge this because swap is freed.
1973 * This memcg can be obsolete one. We avoid calling css_tryget
1975 if (!mem_cgroup_is_root(memcg))
1976 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
1977 mem_cgroup_swap_statistics(memcg, false);
1978 mem_cgroup_put(memcg);
1980 rcu_read_unlock();
1982 #endif
1985 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
1986 * page belongs to.
1988 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
1990 struct page_cgroup *pc;
1991 struct mem_cgroup *mem = NULL;
1992 int ret = 0;
1994 if (mem_cgroup_disabled())
1995 return 0;
1997 pc = lookup_page_cgroup(page);
1998 lock_page_cgroup(pc);
1999 if (PageCgroupUsed(pc)) {
2000 mem = pc->mem_cgroup;
2001 css_get(&mem->css);
2003 unlock_page_cgroup(pc);
2005 if (mem) {
2006 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
2007 page);
2008 css_put(&mem->css);
2010 *ptr = mem;
2011 return ret;
2014 /* remove redundant charge if migration failed*/
2015 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2016 struct page *oldpage, struct page *newpage)
2018 struct page *target, *unused;
2019 struct page_cgroup *pc;
2020 enum charge_type ctype;
2022 if (!mem)
2023 return;
2024 cgroup_exclude_rmdir(&mem->css);
2025 /* at migration success, oldpage->mapping is NULL. */
2026 if (oldpage->mapping) {
2027 target = oldpage;
2028 unused = NULL;
2029 } else {
2030 target = newpage;
2031 unused = oldpage;
2034 if (PageAnon(target))
2035 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2036 else if (page_is_file_cache(target))
2037 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2038 else
2039 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2041 /* unused page is not on radix-tree now. */
2042 if (unused)
2043 __mem_cgroup_uncharge_common(unused, ctype);
2045 pc = lookup_page_cgroup(target);
2047 * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
2048 * So, double-counting is effectively avoided.
2050 __mem_cgroup_commit_charge(mem, pc, ctype);
2053 * Both of oldpage and newpage are still under lock_page().
2054 * Then, we don't have to care about race in radix-tree.
2055 * But we have to be careful that this page is unmapped or not.
2057 * There is a case for !page_mapped(). At the start of
2058 * migration, oldpage was mapped. But now, it's zapped.
2059 * But we know *target* page is not freed/reused under us.
2060 * mem_cgroup_uncharge_page() does all necessary checks.
2062 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2063 mem_cgroup_uncharge_page(target);
2065 * At migration, we may charge account against cgroup which has no tasks
2066 * So, rmdir()->pre_destroy() can be called while we do this charge.
2067 * In that case, we need to call pre_destroy() again. check it here.
2069 cgroup_release_and_wakeup_rmdir(&mem->css);
2073 * A call to try to shrink memory usage on charge failure at shmem's swapin.
2074 * Calling hierarchical_reclaim is not enough because we should update
2075 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2076 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2077 * not from the memcg which this page would be charged to.
2078 * try_charge_swapin does all of these works properly.
2080 int mem_cgroup_shmem_charge_fallback(struct page *page,
2081 struct mm_struct *mm,
2082 gfp_t gfp_mask)
2084 struct mem_cgroup *mem = NULL;
2085 int ret;
2087 if (mem_cgroup_disabled())
2088 return 0;
2090 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2091 if (!ret)
2092 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2094 return ret;
2097 static DEFINE_MUTEX(set_limit_mutex);
2099 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2100 unsigned long long val)
2102 int retry_count;
2103 int progress;
2104 u64 memswlimit;
2105 int ret = 0;
2106 int children = mem_cgroup_count_children(memcg);
2107 u64 curusage, oldusage;
2110 * For keeping hierarchical_reclaim simple, how long we should retry
2111 * is depends on callers. We set our retry-count to be function
2112 * of # of children which we should visit in this loop.
2114 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2116 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2118 while (retry_count) {
2119 if (signal_pending(current)) {
2120 ret = -EINTR;
2121 break;
2124 * Rather than hide all in some function, I do this in
2125 * open coded manner. You see what this really does.
2126 * We have to guarantee mem->res.limit < mem->memsw.limit.
2128 mutex_lock(&set_limit_mutex);
2129 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2130 if (memswlimit < val) {
2131 ret = -EINVAL;
2132 mutex_unlock(&set_limit_mutex);
2133 break;
2135 ret = res_counter_set_limit(&memcg->res, val);
2136 if (!ret) {
2137 if (memswlimit == val)
2138 memcg->memsw_is_minimum = true;
2139 else
2140 memcg->memsw_is_minimum = false;
2142 mutex_unlock(&set_limit_mutex);
2144 if (!ret)
2145 break;
2147 progress = mem_cgroup_hierarchical_reclaim(memcg, NULL,
2148 GFP_KERNEL,
2149 MEM_CGROUP_RECLAIM_SHRINK);
2150 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2151 /* Usage is reduced ? */
2152 if (curusage >= oldusage)
2153 retry_count--;
2154 else
2155 oldusage = curusage;
2158 return ret;
2161 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2162 unsigned long long val)
2164 int retry_count;
2165 u64 memlimit, oldusage, curusage;
2166 int children = mem_cgroup_count_children(memcg);
2167 int ret = -EBUSY;
2169 /* see mem_cgroup_resize_res_limit */
2170 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2171 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2172 while (retry_count) {
2173 if (signal_pending(current)) {
2174 ret = -EINTR;
2175 break;
2178 * Rather than hide all in some function, I do this in
2179 * open coded manner. You see what this really does.
2180 * We have to guarantee mem->res.limit < mem->memsw.limit.
2182 mutex_lock(&set_limit_mutex);
2183 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2184 if (memlimit > val) {
2185 ret = -EINVAL;
2186 mutex_unlock(&set_limit_mutex);
2187 break;
2189 ret = res_counter_set_limit(&memcg->memsw, val);
2190 if (!ret) {
2191 if (memlimit == val)
2192 memcg->memsw_is_minimum = true;
2193 else
2194 memcg->memsw_is_minimum = false;
2196 mutex_unlock(&set_limit_mutex);
2198 if (!ret)
2199 break;
2201 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2202 MEM_CGROUP_RECLAIM_NOSWAP |
2203 MEM_CGROUP_RECLAIM_SHRINK);
2204 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2205 /* Usage is reduced ? */
2206 if (curusage >= oldusage)
2207 retry_count--;
2208 else
2209 oldusage = curusage;
2211 return ret;
2214 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2215 gfp_t gfp_mask, int nid,
2216 int zid)
2218 unsigned long nr_reclaimed = 0;
2219 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2220 unsigned long reclaimed;
2221 int loop = 0;
2222 struct mem_cgroup_tree_per_zone *mctz;
2223 unsigned long long excess;
2225 if (order > 0)
2226 return 0;
2228 mctz = soft_limit_tree_node_zone(nid, zid);
2230 * This loop can run a while, specially if mem_cgroup's continuously
2231 * keep exceeding their soft limit and putting the system under
2232 * pressure
2234 do {
2235 if (next_mz)
2236 mz = next_mz;
2237 else
2238 mz = mem_cgroup_largest_soft_limit_node(mctz);
2239 if (!mz)
2240 break;
2242 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2243 gfp_mask,
2244 MEM_CGROUP_RECLAIM_SOFT);
2245 nr_reclaimed += reclaimed;
2246 spin_lock(&mctz->lock);
2249 * If we failed to reclaim anything from this memory cgroup
2250 * it is time to move on to the next cgroup
2252 next_mz = NULL;
2253 if (!reclaimed) {
2254 do {
2256 * Loop until we find yet another one.
2258 * By the time we get the soft_limit lock
2259 * again, someone might have aded the
2260 * group back on the RB tree. Iterate to
2261 * make sure we get a different mem.
2262 * mem_cgroup_largest_soft_limit_node returns
2263 * NULL if no other cgroup is present on
2264 * the tree
2266 next_mz =
2267 __mem_cgroup_largest_soft_limit_node(mctz);
2268 if (next_mz == mz) {
2269 css_put(&next_mz->mem->css);
2270 next_mz = NULL;
2271 } else /* next_mz == NULL or other memcg */
2272 break;
2273 } while (1);
2275 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2276 excess = res_counter_soft_limit_excess(&mz->mem->res);
2278 * One school of thought says that we should not add
2279 * back the node to the tree if reclaim returns 0.
2280 * But our reclaim could return 0, simply because due
2281 * to priority we are exposing a smaller subset of
2282 * memory to reclaim from. Consider this as a longer
2283 * term TODO.
2285 /* If excess == 0, no tree ops */
2286 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2287 spin_unlock(&mctz->lock);
2288 css_put(&mz->mem->css);
2289 loop++;
2291 * Could not reclaim anything and there are no more
2292 * mem cgroups to try or we seem to be looping without
2293 * reclaiming anything.
2295 if (!nr_reclaimed &&
2296 (next_mz == NULL ||
2297 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2298 break;
2299 } while (!nr_reclaimed);
2300 if (next_mz)
2301 css_put(&next_mz->mem->css);
2302 return nr_reclaimed;
2306 * This routine traverse page_cgroup in given list and drop them all.
2307 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2309 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2310 int node, int zid, enum lru_list lru)
2312 struct zone *zone;
2313 struct mem_cgroup_per_zone *mz;
2314 struct page_cgroup *pc, *busy;
2315 unsigned long flags, loop;
2316 struct list_head *list;
2317 int ret = 0;
2319 zone = &NODE_DATA(node)->node_zones[zid];
2320 mz = mem_cgroup_zoneinfo(mem, node, zid);
2321 list = &mz->lists[lru];
2323 loop = MEM_CGROUP_ZSTAT(mz, lru);
2324 /* give some margin against EBUSY etc...*/
2325 loop += 256;
2326 busy = NULL;
2327 while (loop--) {
2328 ret = 0;
2329 spin_lock_irqsave(&zone->lru_lock, flags);
2330 if (list_empty(list)) {
2331 spin_unlock_irqrestore(&zone->lru_lock, flags);
2332 break;
2334 pc = list_entry(list->prev, struct page_cgroup, lru);
2335 if (busy == pc) {
2336 list_move(&pc->lru, list);
2337 busy = 0;
2338 spin_unlock_irqrestore(&zone->lru_lock, flags);
2339 continue;
2341 spin_unlock_irqrestore(&zone->lru_lock, flags);
2343 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2344 if (ret == -ENOMEM)
2345 break;
2347 if (ret == -EBUSY || ret == -EINVAL) {
2348 /* found lock contention or "pc" is obsolete. */
2349 busy = pc;
2350 cond_resched();
2351 } else
2352 busy = NULL;
2355 if (!ret && !list_empty(list))
2356 return -EBUSY;
2357 return ret;
2361 * make mem_cgroup's charge to be 0 if there is no task.
2362 * This enables deleting this mem_cgroup.
2364 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2366 int ret;
2367 int node, zid, shrink;
2368 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2369 struct cgroup *cgrp = mem->css.cgroup;
2371 css_get(&mem->css);
2373 shrink = 0;
2374 /* should free all ? */
2375 if (free_all)
2376 goto try_to_free;
2377 move_account:
2378 while (mem->res.usage > 0) {
2379 ret = -EBUSY;
2380 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2381 goto out;
2382 ret = -EINTR;
2383 if (signal_pending(current))
2384 goto out;
2385 /* This is for making all *used* pages to be on LRU. */
2386 lru_add_drain_all();
2387 ret = 0;
2388 for_each_node_state(node, N_HIGH_MEMORY) {
2389 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2390 enum lru_list l;
2391 for_each_lru(l) {
2392 ret = mem_cgroup_force_empty_list(mem,
2393 node, zid, l);
2394 if (ret)
2395 break;
2398 if (ret)
2399 break;
2401 /* it seems parent cgroup doesn't have enough mem */
2402 if (ret == -ENOMEM)
2403 goto try_to_free;
2404 cond_resched();
2406 ret = 0;
2407 out:
2408 css_put(&mem->css);
2409 return ret;
2411 try_to_free:
2412 /* returns EBUSY if there is a task or if we come here twice. */
2413 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2414 ret = -EBUSY;
2415 goto out;
2417 /* we call try-to-free pages for make this cgroup empty */
2418 lru_add_drain_all();
2419 /* try to free all pages in this cgroup */
2420 shrink = 1;
2421 while (nr_retries && mem->res.usage > 0) {
2422 int progress;
2424 if (signal_pending(current)) {
2425 ret = -EINTR;
2426 goto out;
2428 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2429 false, get_swappiness(mem));
2430 if (!progress) {
2431 nr_retries--;
2432 /* maybe some writeback is necessary */
2433 congestion_wait(BLK_RW_ASYNC, HZ/10);
2437 lru_add_drain();
2438 /* try move_account...there may be some *locked* pages. */
2439 if (mem->res.usage)
2440 goto move_account;
2441 ret = 0;
2442 goto out;
2445 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2447 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2451 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2453 return mem_cgroup_from_cont(cont)->use_hierarchy;
2456 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2457 u64 val)
2459 int retval = 0;
2460 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2461 struct cgroup *parent = cont->parent;
2462 struct mem_cgroup *parent_mem = NULL;
2464 if (parent)
2465 parent_mem = mem_cgroup_from_cont(parent);
2467 cgroup_lock();
2469 * If parent's use_hiearchy is set, we can't make any modifications
2470 * in the child subtrees. If it is unset, then the change can
2471 * occur, provided the current cgroup has no children.
2473 * For the root cgroup, parent_mem is NULL, we allow value to be
2474 * set if there are no children.
2476 if ((!parent_mem || !parent_mem->use_hierarchy) &&
2477 (val == 1 || val == 0)) {
2478 if (list_empty(&cont->children))
2479 mem->use_hierarchy = val;
2480 else
2481 retval = -EBUSY;
2482 } else
2483 retval = -EINVAL;
2484 cgroup_unlock();
2486 return retval;
2489 struct mem_cgroup_idx_data {
2490 s64 val;
2491 enum mem_cgroup_stat_index idx;
2494 static int
2495 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
2497 struct mem_cgroup_idx_data *d = data;
2498 d->val += mem_cgroup_read_stat(&mem->stat, d->idx);
2499 return 0;
2502 static void
2503 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
2504 enum mem_cgroup_stat_index idx, s64 *val)
2506 struct mem_cgroup_idx_data d;
2507 d.idx = idx;
2508 d.val = 0;
2509 mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
2510 *val = d.val;
2513 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
2515 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2516 u64 idx_val, val;
2517 int type, name;
2519 type = MEMFILE_TYPE(cft->private);
2520 name = MEMFILE_ATTR(cft->private);
2521 switch (type) {
2522 case _MEM:
2523 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2524 mem_cgroup_get_recursive_idx_stat(mem,
2525 MEM_CGROUP_STAT_CACHE, &idx_val);
2526 val = idx_val;
2527 mem_cgroup_get_recursive_idx_stat(mem,
2528 MEM_CGROUP_STAT_RSS, &idx_val);
2529 val += idx_val;
2530 val <<= PAGE_SHIFT;
2531 } else
2532 val = res_counter_read_u64(&mem->res, name);
2533 break;
2534 case _MEMSWAP:
2535 if (name == RES_USAGE && mem_cgroup_is_root(mem)) {
2536 mem_cgroup_get_recursive_idx_stat(mem,
2537 MEM_CGROUP_STAT_CACHE, &idx_val);
2538 val = idx_val;
2539 mem_cgroup_get_recursive_idx_stat(mem,
2540 MEM_CGROUP_STAT_RSS, &idx_val);
2541 val += idx_val;
2542 mem_cgroup_get_recursive_idx_stat(mem,
2543 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
2544 val <<= PAGE_SHIFT;
2545 } else
2546 val = res_counter_read_u64(&mem->memsw, name);
2547 break;
2548 default:
2549 BUG();
2550 break;
2552 return val;
2555 * The user of this function is...
2556 * RES_LIMIT.
2558 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
2559 const char *buffer)
2561 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2562 int type, name;
2563 unsigned long long val;
2564 int ret;
2566 type = MEMFILE_TYPE(cft->private);
2567 name = MEMFILE_ATTR(cft->private);
2568 switch (name) {
2569 case RES_LIMIT:
2570 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2571 ret = -EINVAL;
2572 break;
2574 /* This function does all necessary parse...reuse it */
2575 ret = res_counter_memparse_write_strategy(buffer, &val);
2576 if (ret)
2577 break;
2578 if (type == _MEM)
2579 ret = mem_cgroup_resize_limit(memcg, val);
2580 else
2581 ret = mem_cgroup_resize_memsw_limit(memcg, val);
2582 break;
2583 case RES_SOFT_LIMIT:
2584 ret = res_counter_memparse_write_strategy(buffer, &val);
2585 if (ret)
2586 break;
2588 * For memsw, soft limits are hard to implement in terms
2589 * of semantics, for now, we support soft limits for
2590 * control without swap
2592 if (type == _MEM)
2593 ret = res_counter_set_soft_limit(&memcg->res, val);
2594 else
2595 ret = -EINVAL;
2596 break;
2597 default:
2598 ret = -EINVAL; /* should be BUG() ? */
2599 break;
2601 return ret;
2604 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
2605 unsigned long long *mem_limit, unsigned long long *memsw_limit)
2607 struct cgroup *cgroup;
2608 unsigned long long min_limit, min_memsw_limit, tmp;
2610 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2611 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2612 cgroup = memcg->css.cgroup;
2613 if (!memcg->use_hierarchy)
2614 goto out;
2616 while (cgroup->parent) {
2617 cgroup = cgroup->parent;
2618 memcg = mem_cgroup_from_cont(cgroup);
2619 if (!memcg->use_hierarchy)
2620 break;
2621 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
2622 min_limit = min(min_limit, tmp);
2623 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2624 min_memsw_limit = min(min_memsw_limit, tmp);
2626 out:
2627 *mem_limit = min_limit;
2628 *memsw_limit = min_memsw_limit;
2629 return;
2632 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
2634 struct mem_cgroup *mem;
2635 int type, name;
2637 mem = mem_cgroup_from_cont(cont);
2638 type = MEMFILE_TYPE(event);
2639 name = MEMFILE_ATTR(event);
2640 switch (name) {
2641 case RES_MAX_USAGE:
2642 if (type == _MEM)
2643 res_counter_reset_max(&mem->res);
2644 else
2645 res_counter_reset_max(&mem->memsw);
2646 break;
2647 case RES_FAILCNT:
2648 if (type == _MEM)
2649 res_counter_reset_failcnt(&mem->res);
2650 else
2651 res_counter_reset_failcnt(&mem->memsw);
2652 break;
2655 return 0;
2659 /* For read statistics */
2660 enum {
2661 MCS_CACHE,
2662 MCS_RSS,
2663 MCS_MAPPED_FILE,
2664 MCS_PGPGIN,
2665 MCS_PGPGOUT,
2666 MCS_SWAP,
2667 MCS_INACTIVE_ANON,
2668 MCS_ACTIVE_ANON,
2669 MCS_INACTIVE_FILE,
2670 MCS_ACTIVE_FILE,
2671 MCS_UNEVICTABLE,
2672 NR_MCS_STAT,
2675 struct mcs_total_stat {
2676 s64 stat[NR_MCS_STAT];
2679 struct {
2680 char *local_name;
2681 char *total_name;
2682 } memcg_stat_strings[NR_MCS_STAT] = {
2683 {"cache", "total_cache"},
2684 {"rss", "total_rss"},
2685 {"mapped_file", "total_mapped_file"},
2686 {"pgpgin", "total_pgpgin"},
2687 {"pgpgout", "total_pgpgout"},
2688 {"swap", "total_swap"},
2689 {"inactive_anon", "total_inactive_anon"},
2690 {"active_anon", "total_active_anon"},
2691 {"inactive_file", "total_inactive_file"},
2692 {"active_file", "total_active_file"},
2693 {"unevictable", "total_unevictable"}
2697 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
2699 struct mcs_total_stat *s = data;
2700 s64 val;
2702 /* per cpu stat */
2703 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_CACHE);
2704 s->stat[MCS_CACHE] += val * PAGE_SIZE;
2705 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_RSS);
2706 s->stat[MCS_RSS] += val * PAGE_SIZE;
2707 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_MAPPED_FILE);
2708 s->stat[MCS_MAPPED_FILE] += val * PAGE_SIZE;
2709 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGIN_COUNT);
2710 s->stat[MCS_PGPGIN] += val;
2711 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_PGPGOUT_COUNT);
2712 s->stat[MCS_PGPGOUT] += val;
2713 if (do_swap_account) {
2714 val = mem_cgroup_read_stat(&mem->stat, MEM_CGROUP_STAT_SWAPOUT);
2715 s->stat[MCS_SWAP] += val * PAGE_SIZE;
2718 /* per zone stat */
2719 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
2720 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
2721 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
2722 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
2723 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
2724 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
2725 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
2726 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
2727 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
2728 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
2729 return 0;
2732 static void
2733 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
2735 mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
2738 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
2739 struct cgroup_map_cb *cb)
2741 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
2742 struct mcs_total_stat mystat;
2743 int i;
2745 memset(&mystat, 0, sizeof(mystat));
2746 mem_cgroup_get_local_stat(mem_cont, &mystat);
2748 for (i = 0; i < NR_MCS_STAT; i++) {
2749 if (i == MCS_SWAP && !do_swap_account)
2750 continue;
2751 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
2754 /* Hierarchical information */
2756 unsigned long long limit, memsw_limit;
2757 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
2758 cb->fill(cb, "hierarchical_memory_limit", limit);
2759 if (do_swap_account)
2760 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
2763 memset(&mystat, 0, sizeof(mystat));
2764 mem_cgroup_get_total_stat(mem_cont, &mystat);
2765 for (i = 0; i < NR_MCS_STAT; i++) {
2766 if (i == MCS_SWAP && !do_swap_account)
2767 continue;
2768 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
2771 #ifdef CONFIG_DEBUG_VM
2772 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
2775 int nid, zid;
2776 struct mem_cgroup_per_zone *mz;
2777 unsigned long recent_rotated[2] = {0, 0};
2778 unsigned long recent_scanned[2] = {0, 0};
2780 for_each_online_node(nid)
2781 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
2782 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
2784 recent_rotated[0] +=
2785 mz->reclaim_stat.recent_rotated[0];
2786 recent_rotated[1] +=
2787 mz->reclaim_stat.recent_rotated[1];
2788 recent_scanned[0] +=
2789 mz->reclaim_stat.recent_scanned[0];
2790 recent_scanned[1] +=
2791 mz->reclaim_stat.recent_scanned[1];
2793 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
2794 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
2795 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
2796 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
2798 #endif
2800 return 0;
2803 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
2805 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2807 return get_swappiness(memcg);
2810 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
2811 u64 val)
2813 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
2814 struct mem_cgroup *parent;
2816 if (val > 100)
2817 return -EINVAL;
2819 if (cgrp->parent == NULL)
2820 return -EINVAL;
2822 parent = mem_cgroup_from_cont(cgrp->parent);
2824 cgroup_lock();
2826 /* If under hierarchy, only empty-root can set this value */
2827 if ((parent->use_hierarchy) ||
2828 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
2829 cgroup_unlock();
2830 return -EINVAL;
2833 spin_lock(&memcg->reclaim_param_lock);
2834 memcg->swappiness = val;
2835 spin_unlock(&memcg->reclaim_param_lock);
2837 cgroup_unlock();
2839 return 0;
2843 static struct cftype mem_cgroup_files[] = {
2845 .name = "usage_in_bytes",
2846 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
2847 .read_u64 = mem_cgroup_read,
2850 .name = "max_usage_in_bytes",
2851 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
2852 .trigger = mem_cgroup_reset,
2853 .read_u64 = mem_cgroup_read,
2856 .name = "limit_in_bytes",
2857 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
2858 .write_string = mem_cgroup_write,
2859 .read_u64 = mem_cgroup_read,
2862 .name = "soft_limit_in_bytes",
2863 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
2864 .write_string = mem_cgroup_write,
2865 .read_u64 = mem_cgroup_read,
2868 .name = "failcnt",
2869 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
2870 .trigger = mem_cgroup_reset,
2871 .read_u64 = mem_cgroup_read,
2874 .name = "stat",
2875 .read_map = mem_control_stat_show,
2878 .name = "force_empty",
2879 .trigger = mem_cgroup_force_empty_write,
2882 .name = "use_hierarchy",
2883 .write_u64 = mem_cgroup_hierarchy_write,
2884 .read_u64 = mem_cgroup_hierarchy_read,
2887 .name = "swappiness",
2888 .read_u64 = mem_cgroup_swappiness_read,
2889 .write_u64 = mem_cgroup_swappiness_write,
2893 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2894 static struct cftype memsw_cgroup_files[] = {
2896 .name = "memsw.usage_in_bytes",
2897 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
2898 .read_u64 = mem_cgroup_read,
2901 .name = "memsw.max_usage_in_bytes",
2902 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
2903 .trigger = mem_cgroup_reset,
2904 .read_u64 = mem_cgroup_read,
2907 .name = "memsw.limit_in_bytes",
2908 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
2909 .write_string = mem_cgroup_write,
2910 .read_u64 = mem_cgroup_read,
2913 .name = "memsw.failcnt",
2914 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
2915 .trigger = mem_cgroup_reset,
2916 .read_u64 = mem_cgroup_read,
2920 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2922 if (!do_swap_account)
2923 return 0;
2924 return cgroup_add_files(cont, ss, memsw_cgroup_files,
2925 ARRAY_SIZE(memsw_cgroup_files));
2927 #else
2928 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
2930 return 0;
2932 #endif
2934 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2936 struct mem_cgroup_per_node *pn;
2937 struct mem_cgroup_per_zone *mz;
2938 enum lru_list l;
2939 int zone, tmp = node;
2941 * This routine is called against possible nodes.
2942 * But it's BUG to call kmalloc() against offline node.
2944 * TODO: this routine can waste much memory for nodes which will
2945 * never be onlined. It's better to use memory hotplug callback
2946 * function.
2948 if (!node_state(node, N_NORMAL_MEMORY))
2949 tmp = -1;
2950 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
2951 if (!pn)
2952 return 1;
2954 mem->info.nodeinfo[node] = pn;
2955 memset(pn, 0, sizeof(*pn));
2957 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
2958 mz = &pn->zoneinfo[zone];
2959 for_each_lru(l)
2960 INIT_LIST_HEAD(&mz->lists[l]);
2961 mz->usage_in_excess = 0;
2962 mz->on_tree = false;
2963 mz->mem = mem;
2965 return 0;
2968 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
2970 kfree(mem->info.nodeinfo[node]);
2973 static int mem_cgroup_size(void)
2975 int cpustat_size = nr_cpu_ids * sizeof(struct mem_cgroup_stat_cpu);
2976 return sizeof(struct mem_cgroup) + cpustat_size;
2979 static struct mem_cgroup *mem_cgroup_alloc(void)
2981 struct mem_cgroup *mem;
2982 int size = mem_cgroup_size();
2984 if (size < PAGE_SIZE)
2985 mem = kmalloc(size, GFP_KERNEL);
2986 else
2987 mem = vmalloc(size);
2989 if (mem)
2990 memset(mem, 0, size);
2991 return mem;
2995 * At destroying mem_cgroup, references from swap_cgroup can remain.
2996 * (scanning all at force_empty is too costly...)
2998 * Instead of clearing all references at force_empty, we remember
2999 * the number of reference from swap_cgroup and free mem_cgroup when
3000 * it goes down to 0.
3002 * Removal of cgroup itself succeeds regardless of refs from swap.
3005 static void __mem_cgroup_free(struct mem_cgroup *mem)
3007 int node;
3009 mem_cgroup_remove_from_trees(mem);
3010 free_css_id(&mem_cgroup_subsys, &mem->css);
3012 for_each_node_state(node, N_POSSIBLE)
3013 free_mem_cgroup_per_zone_info(mem, node);
3015 if (mem_cgroup_size() < PAGE_SIZE)
3016 kfree(mem);
3017 else
3018 vfree(mem);
3021 static void mem_cgroup_get(struct mem_cgroup *mem)
3023 atomic_inc(&mem->refcnt);
3026 static void mem_cgroup_put(struct mem_cgroup *mem)
3028 if (atomic_dec_and_test(&mem->refcnt)) {
3029 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3030 __mem_cgroup_free(mem);
3031 if (parent)
3032 mem_cgroup_put(parent);
3037 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3039 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
3041 if (!mem->res.parent)
3042 return NULL;
3043 return mem_cgroup_from_res_counter(mem->res.parent, res);
3046 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3047 static void __init enable_swap_cgroup(void)
3049 if (!mem_cgroup_disabled() && really_do_swap_account)
3050 do_swap_account = 1;
3052 #else
3053 static void __init enable_swap_cgroup(void)
3056 #endif
3058 static int mem_cgroup_soft_limit_tree_init(void)
3060 struct mem_cgroup_tree_per_node *rtpn;
3061 struct mem_cgroup_tree_per_zone *rtpz;
3062 int tmp, node, zone;
3064 for_each_node_state(node, N_POSSIBLE) {
3065 tmp = node;
3066 if (!node_state(node, N_NORMAL_MEMORY))
3067 tmp = -1;
3068 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
3069 if (!rtpn)
3070 return 1;
3072 soft_limit_tree.rb_tree_per_node[node] = rtpn;
3074 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3075 rtpz = &rtpn->rb_tree_per_zone[zone];
3076 rtpz->rb_root = RB_ROOT;
3077 spin_lock_init(&rtpz->lock);
3080 return 0;
3083 static struct cgroup_subsys_state * __ref
3084 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
3086 struct mem_cgroup *mem, *parent;
3087 long error = -ENOMEM;
3088 int node;
3090 mem = mem_cgroup_alloc();
3091 if (!mem)
3092 return ERR_PTR(error);
3094 for_each_node_state(node, N_POSSIBLE)
3095 if (alloc_mem_cgroup_per_zone_info(mem, node))
3096 goto free_out;
3098 /* root ? */
3099 if (cont->parent == NULL) {
3100 enable_swap_cgroup();
3101 parent = NULL;
3102 root_mem_cgroup = mem;
3103 if (mem_cgroup_soft_limit_tree_init())
3104 goto free_out;
3106 } else {
3107 parent = mem_cgroup_from_cont(cont->parent);
3108 mem->use_hierarchy = parent->use_hierarchy;
3111 if (parent && parent->use_hierarchy) {
3112 res_counter_init(&mem->res, &parent->res);
3113 res_counter_init(&mem->memsw, &parent->memsw);
3115 * We increment refcnt of the parent to ensure that we can
3116 * safely access it on res_counter_charge/uncharge.
3117 * This refcnt will be decremented when freeing this
3118 * mem_cgroup(see mem_cgroup_put).
3120 mem_cgroup_get(parent);
3121 } else {
3122 res_counter_init(&mem->res, NULL);
3123 res_counter_init(&mem->memsw, NULL);
3125 mem->last_scanned_child = 0;
3126 spin_lock_init(&mem->reclaim_param_lock);
3128 if (parent)
3129 mem->swappiness = get_swappiness(parent);
3130 atomic_set(&mem->refcnt, 1);
3131 return &mem->css;
3132 free_out:
3133 __mem_cgroup_free(mem);
3134 root_mem_cgroup = NULL;
3135 return ERR_PTR(error);
3138 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
3139 struct cgroup *cont)
3141 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3143 return mem_cgroup_force_empty(mem, false);
3146 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
3147 struct cgroup *cont)
3149 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3151 mem_cgroup_put(mem);
3154 static int mem_cgroup_populate(struct cgroup_subsys *ss,
3155 struct cgroup *cont)
3157 int ret;
3159 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
3160 ARRAY_SIZE(mem_cgroup_files));
3162 if (!ret)
3163 ret = register_memsw_files(cont, ss);
3164 return ret;
3167 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
3168 struct cgroup *cont,
3169 struct cgroup *old_cont,
3170 struct task_struct *p,
3171 bool threadgroup)
3173 mutex_lock(&memcg_tasklist);
3175 * FIXME: It's better to move charges of this process from old
3176 * memcg to new memcg. But it's just on TODO-List now.
3178 mutex_unlock(&memcg_tasklist);
3181 struct cgroup_subsys mem_cgroup_subsys = {
3182 .name = "memory",
3183 .subsys_id = mem_cgroup_subsys_id,
3184 .create = mem_cgroup_create,
3185 .pre_destroy = mem_cgroup_pre_destroy,
3186 .destroy = mem_cgroup_destroy,
3187 .populate = mem_cgroup_populate,
3188 .attach = mem_cgroup_move_task,
3189 .early_init = 0,
3190 .use_id = 1,
3193 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3195 static int __init disable_swap_account(char *s)
3197 really_do_swap_account = 0;
3198 return 1;
3200 __setup("noswapaccount", disable_swap_account);
3201 #endif