include: replace linux/module.h with "struct module" wherever possible
[linux-2.6/next.git] / mm / memcontrol.c
blob2fcfbdc45e3081a425944450f2d3115ca3118283
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 * Memory thresholds
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
27 #include <linux/mm.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
45 #include <linux/fs.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
52 #include "internal.h"
54 #include <asm/uaccess.h>
56 #include <trace/events/vmscan.h>
58 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
59 #define MEM_CGROUP_RECLAIM_RETRIES 5
60 struct mem_cgroup *root_mem_cgroup __read_mostly;
62 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
63 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
64 int do_swap_account __read_mostly;
66 /* for remember boot option*/
67 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
68 static int really_do_swap_account __initdata = 1;
69 #else
70 static int really_do_swap_account __initdata = 0;
71 #endif
73 #else
74 #define do_swap_account (0)
75 #endif
79 * Statistics for memory cgroup.
81 enum mem_cgroup_stat_index {
83 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
85 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
86 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
87 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
90 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
91 MEM_CGROUP_STAT_NSTATS,
94 enum mem_cgroup_events_index {
95 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
96 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
97 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
98 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
99 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
100 MEM_CGROUP_EVENTS_NSTATS,
103 * Per memcg event counter is incremented at every pagein/pageout. With THP,
104 * it will be incremated by the number of pages. This counter is used for
105 * for trigger some periodic events. This is straightforward and better
106 * than using jiffies etc. to handle periodic memcg event.
108 enum mem_cgroup_events_target {
109 MEM_CGROUP_TARGET_THRESH,
110 MEM_CGROUP_TARGET_SOFTLIMIT,
111 MEM_CGROUP_TARGET_NUMAINFO,
112 MEM_CGROUP_NTARGETS,
114 #define THRESHOLDS_EVENTS_TARGET (128)
115 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 #define NUMAINFO_EVENTS_TARGET (1024)
118 struct mem_cgroup_stat_cpu {
119 long count[MEM_CGROUP_STAT_NSTATS];
120 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
121 unsigned long targets[MEM_CGROUP_NTARGETS];
125 * per-zone information in memory controller.
127 struct mem_cgroup_per_zone {
129 * spin_lock to protect the per cgroup LRU
131 struct list_head lists[NR_LRU_LISTS];
132 unsigned long count[NR_LRU_LISTS];
134 struct zone_reclaim_stat reclaim_stat;
135 struct rb_node tree_node; /* RB tree node */
136 unsigned long long usage_in_excess;/* Set to the value by which */
137 /* the soft limit is exceeded*/
138 bool on_tree;
139 struct mem_cgroup *mem; /* Back pointer, we cannot */
140 /* use container_of */
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
145 struct mem_cgroup_per_node {
146 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
149 struct mem_cgroup_lru_info {
150 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
154 * Cgroups above their limits are maintained in a RB-Tree, independent of
155 * their hierarchy representation
158 struct mem_cgroup_tree_per_zone {
159 struct rb_root rb_root;
160 spinlock_t lock;
163 struct mem_cgroup_tree_per_node {
164 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
167 struct mem_cgroup_tree {
168 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
173 struct mem_cgroup_threshold {
174 struct eventfd_ctx *eventfd;
175 u64 threshold;
178 /* For threshold */
179 struct mem_cgroup_threshold_ary {
180 /* An array index points to threshold just below usage. */
181 int current_threshold;
182 /* Size of entries[] */
183 unsigned int size;
184 /* Array of thresholds */
185 struct mem_cgroup_threshold entries[0];
188 struct mem_cgroup_thresholds {
189 /* Primary thresholds array */
190 struct mem_cgroup_threshold_ary *primary;
192 * Spare threshold array.
193 * This is needed to make mem_cgroup_unregister_event() "never fail".
194 * It must be able to store at least primary->size - 1 entries.
196 struct mem_cgroup_threshold_ary *spare;
199 /* for OOM */
200 struct mem_cgroup_eventfd_list {
201 struct list_head list;
202 struct eventfd_ctx *eventfd;
205 static void mem_cgroup_threshold(struct mem_cgroup *mem);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
208 enum {
209 SCAN_BY_LIMIT,
210 SCAN_BY_SYSTEM,
211 NR_SCAN_CONTEXT,
212 SCAN_BY_SHRINK, /* not recorded now */
215 enum {
216 SCAN,
217 SCAN_ANON,
218 SCAN_FILE,
219 ROTATE,
220 ROTATE_ANON,
221 ROTATE_FILE,
222 FREED,
223 FREED_ANON,
224 FREED_FILE,
225 ELAPSED,
226 NR_SCANSTATS,
229 struct scanstat {
230 spinlock_t lock;
231 unsigned long stats[NR_SCAN_CONTEXT][NR_SCANSTATS];
232 unsigned long rootstats[NR_SCAN_CONTEXT][NR_SCANSTATS];
235 const char *scanstat_string[NR_SCANSTATS] = {
236 "scanned_pages",
237 "scanned_anon_pages",
238 "scanned_file_pages",
239 "rotated_pages",
240 "rotated_anon_pages",
241 "rotated_file_pages",
242 "freed_pages",
243 "freed_anon_pages",
244 "freed_file_pages",
245 "elapsed_ns",
247 #define SCANSTAT_WORD_LIMIT "_by_limit"
248 #define SCANSTAT_WORD_SYSTEM "_by_system"
249 #define SCANSTAT_WORD_HIERARCHY "_under_hierarchy"
253 * The memory controller data structure. The memory controller controls both
254 * page cache and RSS per cgroup. We would eventually like to provide
255 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
256 * to help the administrator determine what knobs to tune.
258 * TODO: Add a water mark for the memory controller. Reclaim will begin when
259 * we hit the water mark. May be even add a low water mark, such that
260 * no reclaim occurs from a cgroup at it's low water mark, this is
261 * a feature that will be implemented much later in the future.
263 struct mem_cgroup {
264 struct cgroup_subsys_state css;
266 * the counter to account for memory usage
268 struct res_counter res;
270 * the counter to account for mem+swap usage.
272 struct res_counter memsw;
274 * Per cgroup active and inactive list, similar to the
275 * per zone LRU lists.
277 struct mem_cgroup_lru_info info;
279 * While reclaiming in a hierarchy, we cache the last child we
280 * reclaimed from.
282 int last_scanned_child;
283 int last_scanned_node;
284 #if MAX_NUMNODES > 1
285 nodemask_t scan_nodes;
286 atomic_t numainfo_events;
287 atomic_t numainfo_updating;
288 #endif
290 * Should the accounting and control be hierarchical, per subtree?
292 bool use_hierarchy;
294 bool oom_lock;
295 atomic_t under_oom;
297 atomic_t refcnt;
299 int swappiness;
300 /* OOM-Killer disable */
301 int oom_kill_disable;
303 /* set when res.limit == memsw.limit */
304 bool memsw_is_minimum;
306 /* protect arrays of thresholds */
307 struct mutex thresholds_lock;
309 /* thresholds for memory usage. RCU-protected */
310 struct mem_cgroup_thresholds thresholds;
312 /* thresholds for mem+swap usage. RCU-protected */
313 struct mem_cgroup_thresholds memsw_thresholds;
315 /* For oom notifier event fd */
316 struct list_head oom_notify;
317 /* For recording LRU-scan statistics */
318 struct scanstat scanstat;
320 * Should we move charges of a task when a task is moved into this
321 * mem_cgroup ? And what type of charges should we move ?
323 unsigned long move_charge_at_immigrate;
325 * percpu counter.
327 struct mem_cgroup_stat_cpu *stat;
329 * used when a cpu is offlined or other synchronizations
330 * See mem_cgroup_read_stat().
332 struct mem_cgroup_stat_cpu nocpu_base;
333 spinlock_t pcp_counter_lock;
336 /* Stuffs for move charges at task migration. */
338 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
339 * left-shifted bitmap of these types.
341 enum move_type {
342 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
343 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
344 NR_MOVE_TYPE,
347 /* "mc" and its members are protected by cgroup_mutex */
348 static struct move_charge_struct {
349 spinlock_t lock; /* for from, to */
350 struct mem_cgroup *from;
351 struct mem_cgroup *to;
352 unsigned long precharge;
353 unsigned long moved_charge;
354 unsigned long moved_swap;
355 struct task_struct *moving_task; /* a task moving charges */
356 wait_queue_head_t waitq; /* a waitq for other context */
357 } mc = {
358 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
359 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
362 static bool move_anon(void)
364 return test_bit(MOVE_CHARGE_TYPE_ANON,
365 &mc.to->move_charge_at_immigrate);
368 static bool move_file(void)
370 return test_bit(MOVE_CHARGE_TYPE_FILE,
371 &mc.to->move_charge_at_immigrate);
375 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
376 * limit reclaim to prevent infinite loops, if they ever occur.
378 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
379 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
381 enum charge_type {
382 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
383 MEM_CGROUP_CHARGE_TYPE_MAPPED,
384 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
385 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
386 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
387 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
388 NR_CHARGE_TYPE,
391 /* for encoding cft->private value on file */
392 #define _MEM (0)
393 #define _MEMSWAP (1)
394 #define _OOM_TYPE (2)
395 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
396 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
397 #define MEMFILE_ATTR(val) ((val) & 0xffff)
398 /* Used for OOM nofiier */
399 #define OOM_CONTROL (0)
402 * Reclaim flags for mem_cgroup_hierarchical_reclaim
404 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
405 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
406 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
407 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
408 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
409 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
411 static void mem_cgroup_get(struct mem_cgroup *mem);
412 static void mem_cgroup_put(struct mem_cgroup *mem);
413 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
414 static void drain_all_stock_async(struct mem_cgroup *mem);
416 static struct mem_cgroup_per_zone *
417 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
419 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
422 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
424 return &mem->css;
427 static struct mem_cgroup_per_zone *
428 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
430 int nid = page_to_nid(page);
431 int zid = page_zonenum(page);
433 return mem_cgroup_zoneinfo(mem, nid, zid);
436 static struct mem_cgroup_tree_per_zone *
437 soft_limit_tree_node_zone(int nid, int zid)
439 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
442 static struct mem_cgroup_tree_per_zone *
443 soft_limit_tree_from_page(struct page *page)
445 int nid = page_to_nid(page);
446 int zid = page_zonenum(page);
448 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
451 static void
452 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
453 struct mem_cgroup_per_zone *mz,
454 struct mem_cgroup_tree_per_zone *mctz,
455 unsigned long long new_usage_in_excess)
457 struct rb_node **p = &mctz->rb_root.rb_node;
458 struct rb_node *parent = NULL;
459 struct mem_cgroup_per_zone *mz_node;
461 if (mz->on_tree)
462 return;
464 mz->usage_in_excess = new_usage_in_excess;
465 if (!mz->usage_in_excess)
466 return;
467 while (*p) {
468 parent = *p;
469 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
470 tree_node);
471 if (mz->usage_in_excess < mz_node->usage_in_excess)
472 p = &(*p)->rb_left;
474 * We can't avoid mem cgroups that are over their soft
475 * limit by the same amount
477 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
478 p = &(*p)->rb_right;
480 rb_link_node(&mz->tree_node, parent, p);
481 rb_insert_color(&mz->tree_node, &mctz->rb_root);
482 mz->on_tree = true;
485 static void
486 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
487 struct mem_cgroup_per_zone *mz,
488 struct mem_cgroup_tree_per_zone *mctz)
490 if (!mz->on_tree)
491 return;
492 rb_erase(&mz->tree_node, &mctz->rb_root);
493 mz->on_tree = false;
496 static void
497 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
498 struct mem_cgroup_per_zone *mz,
499 struct mem_cgroup_tree_per_zone *mctz)
501 spin_lock(&mctz->lock);
502 __mem_cgroup_remove_exceeded(mem, mz, mctz);
503 spin_unlock(&mctz->lock);
507 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
509 unsigned long long excess;
510 struct mem_cgroup_per_zone *mz;
511 struct mem_cgroup_tree_per_zone *mctz;
512 int nid = page_to_nid(page);
513 int zid = page_zonenum(page);
514 mctz = soft_limit_tree_from_page(page);
517 * Necessary to update all ancestors when hierarchy is used.
518 * because their event counter is not touched.
520 for (; mem; mem = parent_mem_cgroup(mem)) {
521 mz = mem_cgroup_zoneinfo(mem, nid, zid);
522 excess = res_counter_soft_limit_excess(&mem->res);
524 * We have to update the tree if mz is on RB-tree or
525 * mem is over its softlimit.
527 if (excess || mz->on_tree) {
528 spin_lock(&mctz->lock);
529 /* if on-tree, remove it */
530 if (mz->on_tree)
531 __mem_cgroup_remove_exceeded(mem, mz, mctz);
533 * Insert again. mz->usage_in_excess will be updated.
534 * If excess is 0, no tree ops.
536 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
537 spin_unlock(&mctz->lock);
542 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
544 int node, zone;
545 struct mem_cgroup_per_zone *mz;
546 struct mem_cgroup_tree_per_zone *mctz;
548 for_each_node_state(node, N_POSSIBLE) {
549 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
550 mz = mem_cgroup_zoneinfo(mem, node, zone);
551 mctz = soft_limit_tree_node_zone(node, zone);
552 mem_cgroup_remove_exceeded(mem, mz, mctz);
557 static struct mem_cgroup_per_zone *
558 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
560 struct rb_node *rightmost = NULL;
561 struct mem_cgroup_per_zone *mz;
563 retry:
564 mz = NULL;
565 rightmost = rb_last(&mctz->rb_root);
566 if (!rightmost)
567 goto done; /* Nothing to reclaim from */
569 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
571 * Remove the node now but someone else can add it back,
572 * we will to add it back at the end of reclaim to its correct
573 * position in the tree.
575 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
576 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
577 !css_tryget(&mz->mem->css))
578 goto retry;
579 done:
580 return mz;
583 static struct mem_cgroup_per_zone *
584 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
586 struct mem_cgroup_per_zone *mz;
588 spin_lock(&mctz->lock);
589 mz = __mem_cgroup_largest_soft_limit_node(mctz);
590 spin_unlock(&mctz->lock);
591 return mz;
595 * Implementation Note: reading percpu statistics for memcg.
597 * Both of vmstat[] and percpu_counter has threshold and do periodic
598 * synchronization to implement "quick" read. There are trade-off between
599 * reading cost and precision of value. Then, we may have a chance to implement
600 * a periodic synchronizion of counter in memcg's counter.
602 * But this _read() function is used for user interface now. The user accounts
603 * memory usage by memory cgroup and he _always_ requires exact value because
604 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
605 * have to visit all online cpus and make sum. So, for now, unnecessary
606 * synchronization is not implemented. (just implemented for cpu hotplug)
608 * If there are kernel internal actions which can make use of some not-exact
609 * value, and reading all cpu value can be performance bottleneck in some
610 * common workload, threashold and synchonization as vmstat[] should be
611 * implemented.
613 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
614 enum mem_cgroup_stat_index idx)
616 long val = 0;
617 int cpu;
619 get_online_cpus();
620 for_each_online_cpu(cpu)
621 val += per_cpu(mem->stat->count[idx], cpu);
622 #ifdef CONFIG_HOTPLUG_CPU
623 spin_lock(&mem->pcp_counter_lock);
624 val += mem->nocpu_base.count[idx];
625 spin_unlock(&mem->pcp_counter_lock);
626 #endif
627 put_online_cpus();
628 return val;
631 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
632 bool charge)
634 int val = (charge) ? 1 : -1;
635 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
638 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
640 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
643 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
645 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
648 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
649 enum mem_cgroup_events_index idx)
651 unsigned long val = 0;
652 int cpu;
654 for_each_online_cpu(cpu)
655 val += per_cpu(mem->stat->events[idx], cpu);
656 #ifdef CONFIG_HOTPLUG_CPU
657 spin_lock(&mem->pcp_counter_lock);
658 val += mem->nocpu_base.events[idx];
659 spin_unlock(&mem->pcp_counter_lock);
660 #endif
661 return val;
664 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
665 bool file, int nr_pages)
667 preempt_disable();
669 if (file)
670 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
671 else
672 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
674 /* pagein of a big page is an event. So, ignore page size */
675 if (nr_pages > 0)
676 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
677 else {
678 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
679 nr_pages = -nr_pages; /* for event */
682 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
684 preempt_enable();
687 unsigned long
688 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
689 unsigned int lru_mask)
691 struct mem_cgroup_per_zone *mz;
692 enum lru_list l;
693 unsigned long ret = 0;
695 mz = mem_cgroup_zoneinfo(mem, nid, zid);
697 for_each_lru(l) {
698 if (BIT(l) & lru_mask)
699 ret += MEM_CGROUP_ZSTAT(mz, l);
701 return ret;
704 static unsigned long
705 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
706 int nid, unsigned int lru_mask)
708 u64 total = 0;
709 int zid;
711 for (zid = 0; zid < MAX_NR_ZONES; zid++)
712 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
714 return total;
717 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
718 unsigned int lru_mask)
720 int nid;
721 u64 total = 0;
723 for_each_node_state(nid, N_HIGH_MEMORY)
724 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
725 return total;
728 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
730 unsigned long val, next;
732 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
733 next = this_cpu_read(mem->stat->targets[target]);
734 /* from time_after() in jiffies.h */
735 return ((long)next - (long)val < 0);
738 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
740 unsigned long val, next;
742 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
744 switch (target) {
745 case MEM_CGROUP_TARGET_THRESH:
746 next = val + THRESHOLDS_EVENTS_TARGET;
747 break;
748 case MEM_CGROUP_TARGET_SOFTLIMIT:
749 next = val + SOFTLIMIT_EVENTS_TARGET;
750 break;
751 case MEM_CGROUP_TARGET_NUMAINFO:
752 next = val + NUMAINFO_EVENTS_TARGET;
753 break;
754 default:
755 return;
758 this_cpu_write(mem->stat->targets[target], next);
762 * Check events in order.
765 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
767 /* threshold event is triggered in finer grain than soft limit */
768 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
769 mem_cgroup_threshold(mem);
770 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
771 if (unlikely(__memcg_event_check(mem,
772 MEM_CGROUP_TARGET_SOFTLIMIT))) {
773 mem_cgroup_update_tree(mem, page);
774 __mem_cgroup_target_update(mem,
775 MEM_CGROUP_TARGET_SOFTLIMIT);
777 #if MAX_NUMNODES > 1
778 if (unlikely(__memcg_event_check(mem,
779 MEM_CGROUP_TARGET_NUMAINFO))) {
780 atomic_inc(&mem->numainfo_events);
781 __mem_cgroup_target_update(mem,
782 MEM_CGROUP_TARGET_NUMAINFO);
784 #endif
788 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
790 return container_of(cgroup_subsys_state(cont,
791 mem_cgroup_subsys_id), struct mem_cgroup,
792 css);
795 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
798 * mm_update_next_owner() may clear mm->owner to NULL
799 * if it races with swapoff, page migration, etc.
800 * So this can be called with p == NULL.
802 if (unlikely(!p))
803 return NULL;
805 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
806 struct mem_cgroup, css);
809 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
811 struct mem_cgroup *mem = NULL;
813 if (!mm)
814 return NULL;
816 * Because we have no locks, mm->owner's may be being moved to other
817 * cgroup. We use css_tryget() here even if this looks
818 * pessimistic (rather than adding locks here).
820 rcu_read_lock();
821 do {
822 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
823 if (unlikely(!mem))
824 break;
825 } while (!css_tryget(&mem->css));
826 rcu_read_unlock();
827 return mem;
830 /* The caller has to guarantee "mem" exists before calling this */
831 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
833 struct cgroup_subsys_state *css;
834 int found;
836 if (!mem) /* ROOT cgroup has the smallest ID */
837 return root_mem_cgroup; /*css_put/get against root is ignored*/
838 if (!mem->use_hierarchy) {
839 if (css_tryget(&mem->css))
840 return mem;
841 return NULL;
843 rcu_read_lock();
845 * searching a memory cgroup which has the smallest ID under given
846 * ROOT cgroup. (ID >= 1)
848 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
849 if (css && css_tryget(css))
850 mem = container_of(css, struct mem_cgroup, css);
851 else
852 mem = NULL;
853 rcu_read_unlock();
854 return mem;
857 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
858 struct mem_cgroup *root,
859 bool cond)
861 int nextid = css_id(&iter->css) + 1;
862 int found;
863 int hierarchy_used;
864 struct cgroup_subsys_state *css;
866 hierarchy_used = iter->use_hierarchy;
868 css_put(&iter->css);
869 /* If no ROOT, walk all, ignore hierarchy */
870 if (!cond || (root && !hierarchy_used))
871 return NULL;
873 if (!root)
874 root = root_mem_cgroup;
876 do {
877 iter = NULL;
878 rcu_read_lock();
880 css = css_get_next(&mem_cgroup_subsys, nextid,
881 &root->css, &found);
882 if (css && css_tryget(css))
883 iter = container_of(css, struct mem_cgroup, css);
884 rcu_read_unlock();
885 /* If css is NULL, no more cgroups will be found */
886 nextid = found + 1;
887 } while (css && !iter);
889 return iter;
892 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
893 * be careful that "break" loop is not allowed. We have reference count.
894 * Instead of that modify "cond" to be false and "continue" to exit the loop.
896 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
897 for (iter = mem_cgroup_start_loop(root);\
898 iter != NULL;\
899 iter = mem_cgroup_get_next(iter, root, cond))
901 #define for_each_mem_cgroup_tree(iter, root) \
902 for_each_mem_cgroup_tree_cond(iter, root, true)
904 #define for_each_mem_cgroup_all(iter) \
905 for_each_mem_cgroup_tree_cond(iter, NULL, true)
908 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
910 return (mem == root_mem_cgroup);
913 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
915 struct mem_cgroup *mem;
917 if (!mm)
918 return;
920 rcu_read_lock();
921 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
922 if (unlikely(!mem))
923 goto out;
925 switch (idx) {
926 case PGMAJFAULT:
927 mem_cgroup_pgmajfault(mem, 1);
928 break;
929 case PGFAULT:
930 mem_cgroup_pgfault(mem, 1);
931 break;
932 default:
933 BUG();
935 out:
936 rcu_read_unlock();
938 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
941 * Following LRU functions are allowed to be used without PCG_LOCK.
942 * Operations are called by routine of global LRU independently from memcg.
943 * What we have to take care of here is validness of pc->mem_cgroup.
945 * Changes to pc->mem_cgroup happens when
946 * 1. charge
947 * 2. moving account
948 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
949 * It is added to LRU before charge.
950 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
951 * When moving account, the page is not on LRU. It's isolated.
954 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
956 struct page_cgroup *pc;
957 struct mem_cgroup_per_zone *mz;
959 if (mem_cgroup_disabled())
960 return;
961 pc = lookup_page_cgroup(page);
962 /* can happen while we handle swapcache. */
963 if (!TestClearPageCgroupAcctLRU(pc))
964 return;
965 VM_BUG_ON(!pc->mem_cgroup);
967 * We don't check PCG_USED bit. It's cleared when the "page" is finally
968 * removed from global LRU.
970 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
971 /* huge page split is done under lru_lock. so, we have no races. */
972 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
973 if (mem_cgroup_is_root(pc->mem_cgroup))
974 return;
975 VM_BUG_ON(list_empty(&pc->lru));
976 list_del_init(&pc->lru);
979 void mem_cgroup_del_lru(struct page *page)
981 mem_cgroup_del_lru_list(page, page_lru(page));
985 * Writeback is about to end against a page which has been marked for immediate
986 * reclaim. If it still appears to be reclaimable, move it to the tail of the
987 * inactive list.
989 void mem_cgroup_rotate_reclaimable_page(struct page *page)
991 struct mem_cgroup_per_zone *mz;
992 struct page_cgroup *pc;
993 enum lru_list lru = page_lru(page);
995 if (mem_cgroup_disabled())
996 return;
998 pc = lookup_page_cgroup(page);
999 /* unused or root page is not rotated. */
1000 if (!PageCgroupUsed(pc))
1001 return;
1002 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1003 smp_rmb();
1004 if (mem_cgroup_is_root(pc->mem_cgroup))
1005 return;
1006 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1007 list_move_tail(&pc->lru, &mz->lists[lru]);
1010 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1012 struct mem_cgroup_per_zone *mz;
1013 struct page_cgroup *pc;
1015 if (mem_cgroup_disabled())
1016 return;
1018 pc = lookup_page_cgroup(page);
1019 /* unused or root page is not rotated. */
1020 if (!PageCgroupUsed(pc))
1021 return;
1022 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1023 smp_rmb();
1024 if (mem_cgroup_is_root(pc->mem_cgroup))
1025 return;
1026 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1027 list_move(&pc->lru, &mz->lists[lru]);
1030 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1032 struct page_cgroup *pc;
1033 struct mem_cgroup_per_zone *mz;
1035 if (mem_cgroup_disabled())
1036 return;
1037 pc = lookup_page_cgroup(page);
1038 VM_BUG_ON(PageCgroupAcctLRU(pc));
1039 if (!PageCgroupUsed(pc))
1040 return;
1041 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1042 smp_rmb();
1043 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1044 /* huge page split is done under lru_lock. so, we have no races. */
1045 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1046 SetPageCgroupAcctLRU(pc);
1047 if (mem_cgroup_is_root(pc->mem_cgroup))
1048 return;
1049 list_add(&pc->lru, &mz->lists[lru]);
1053 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1054 * while it's linked to lru because the page may be reused after it's fully
1055 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1056 * It's done under lock_page and expected that zone->lru_lock isnever held.
1058 static void mem_cgroup_lru_del_before_commit(struct page *page)
1060 unsigned long flags;
1061 struct zone *zone = page_zone(page);
1062 struct page_cgroup *pc = lookup_page_cgroup(page);
1065 * Doing this check without taking ->lru_lock seems wrong but this
1066 * is safe. Because if page_cgroup's USED bit is unset, the page
1067 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1068 * set, the commit after this will fail, anyway.
1069 * This all charge/uncharge is done under some mutual execustion.
1070 * So, we don't need to taking care of changes in USED bit.
1072 if (likely(!PageLRU(page)))
1073 return;
1075 spin_lock_irqsave(&zone->lru_lock, flags);
1077 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1078 * is guarded by lock_page() because the page is SwapCache.
1080 if (!PageCgroupUsed(pc))
1081 mem_cgroup_del_lru_list(page, page_lru(page));
1082 spin_unlock_irqrestore(&zone->lru_lock, flags);
1085 static void mem_cgroup_lru_add_after_commit(struct page *page)
1087 unsigned long flags;
1088 struct zone *zone = page_zone(page);
1089 struct page_cgroup *pc = lookup_page_cgroup(page);
1091 /* taking care of that the page is added to LRU while we commit it */
1092 if (likely(!PageLRU(page)))
1093 return;
1094 spin_lock_irqsave(&zone->lru_lock, flags);
1095 /* link when the page is linked to LRU but page_cgroup isn't */
1096 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1097 mem_cgroup_add_lru_list(page, page_lru(page));
1098 spin_unlock_irqrestore(&zone->lru_lock, flags);
1102 void mem_cgroup_move_lists(struct page *page,
1103 enum lru_list from, enum lru_list to)
1105 if (mem_cgroup_disabled())
1106 return;
1107 mem_cgroup_del_lru_list(page, from);
1108 mem_cgroup_add_lru_list(page, to);
1112 * Checks whether given mem is same or in the root_mem's
1113 * hierarchy subtree
1115 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_mem,
1116 struct mem_cgroup *mem)
1118 if (root_mem != mem) {
1119 return (root_mem->use_hierarchy &&
1120 css_is_ancestor(&mem->css, &root_mem->css));
1123 return true;
1126 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1128 int ret;
1129 struct mem_cgroup *curr = NULL;
1130 struct task_struct *p;
1132 p = find_lock_task_mm(task);
1133 if (!p)
1134 return 0;
1135 curr = try_get_mem_cgroup_from_mm(p->mm);
1136 task_unlock(p);
1137 if (!curr)
1138 return 0;
1140 * We should check use_hierarchy of "mem" not "curr". Because checking
1141 * use_hierarchy of "curr" here make this function true if hierarchy is
1142 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1143 * hierarchy(even if use_hierarchy is disabled in "mem").
1145 ret = mem_cgroup_same_or_subtree(mem, curr);
1146 css_put(&curr->css);
1147 return ret;
1150 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1152 unsigned long active;
1153 unsigned long inactive;
1154 unsigned long gb;
1155 unsigned long inactive_ratio;
1157 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1158 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1160 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1161 if (gb)
1162 inactive_ratio = int_sqrt(10 * gb);
1163 else
1164 inactive_ratio = 1;
1166 if (present_pages) {
1167 present_pages[0] = inactive;
1168 present_pages[1] = active;
1171 return inactive_ratio;
1174 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1176 unsigned long active;
1177 unsigned long inactive;
1178 unsigned long present_pages[2];
1179 unsigned long inactive_ratio;
1181 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1183 inactive = present_pages[0];
1184 active = present_pages[1];
1186 if (inactive * inactive_ratio < active)
1187 return 1;
1189 return 0;
1192 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1194 unsigned long active;
1195 unsigned long inactive;
1197 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1198 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1200 return (active > inactive);
1203 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1204 struct zone *zone)
1206 int nid = zone_to_nid(zone);
1207 int zid = zone_idx(zone);
1208 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1210 return &mz->reclaim_stat;
1213 struct zone_reclaim_stat *
1214 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1216 struct page_cgroup *pc;
1217 struct mem_cgroup_per_zone *mz;
1219 if (mem_cgroup_disabled())
1220 return NULL;
1222 pc = lookup_page_cgroup(page);
1223 if (!PageCgroupUsed(pc))
1224 return NULL;
1225 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1226 smp_rmb();
1227 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1228 return &mz->reclaim_stat;
1231 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1232 struct list_head *dst,
1233 unsigned long *scanned, int order,
1234 int mode, struct zone *z,
1235 struct mem_cgroup *mem_cont,
1236 int active, int file)
1238 unsigned long nr_taken = 0;
1239 struct page *page;
1240 unsigned long scan;
1241 LIST_HEAD(pc_list);
1242 struct list_head *src;
1243 struct page_cgroup *pc, *tmp;
1244 int nid = zone_to_nid(z);
1245 int zid = zone_idx(z);
1246 struct mem_cgroup_per_zone *mz;
1247 int lru = LRU_FILE * file + active;
1248 int ret;
1250 BUG_ON(!mem_cont);
1251 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1252 src = &mz->lists[lru];
1254 scan = 0;
1255 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1256 if (scan >= nr_to_scan)
1257 break;
1259 if (unlikely(!PageCgroupUsed(pc)))
1260 continue;
1262 page = lookup_cgroup_page(pc);
1264 if (unlikely(!PageLRU(page)))
1265 continue;
1267 scan++;
1268 ret = __isolate_lru_page(page, mode, file);
1269 switch (ret) {
1270 case 0:
1271 list_move(&page->lru, dst);
1272 mem_cgroup_del_lru(page);
1273 nr_taken += hpage_nr_pages(page);
1274 break;
1275 case -EBUSY:
1276 /* we don't affect global LRU but rotate in our LRU */
1277 mem_cgroup_rotate_lru_list(page, page_lru(page));
1278 break;
1279 default:
1280 break;
1284 *scanned = scan;
1286 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1287 0, 0, 0, mode);
1289 return nr_taken;
1292 #define mem_cgroup_from_res_counter(counter, member) \
1293 container_of(counter, struct mem_cgroup, member)
1296 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1297 * @mem: the memory cgroup
1299 * Returns the maximum amount of memory @mem can be charged with, in
1300 * pages.
1302 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1304 unsigned long long margin;
1306 margin = res_counter_margin(&mem->res);
1307 if (do_swap_account)
1308 margin = min(margin, res_counter_margin(&mem->memsw));
1309 return margin >> PAGE_SHIFT;
1312 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1314 struct cgroup *cgrp = memcg->css.cgroup;
1316 /* root ? */
1317 if (cgrp->parent == NULL)
1318 return vm_swappiness;
1320 return memcg->swappiness;
1323 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1325 int cpu;
1327 get_online_cpus();
1328 spin_lock(&mem->pcp_counter_lock);
1329 for_each_online_cpu(cpu)
1330 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1331 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1332 spin_unlock(&mem->pcp_counter_lock);
1333 put_online_cpus();
1335 synchronize_rcu();
1338 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1340 int cpu;
1342 if (!mem)
1343 return;
1344 get_online_cpus();
1345 spin_lock(&mem->pcp_counter_lock);
1346 for_each_online_cpu(cpu)
1347 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1348 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1349 spin_unlock(&mem->pcp_counter_lock);
1350 put_online_cpus();
1353 * 2 routines for checking "mem" is under move_account() or not.
1355 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1356 * for avoiding race in accounting. If true,
1357 * pc->mem_cgroup may be overwritten.
1359 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1360 * under hierarchy of moving cgroups. This is for
1361 * waiting at hith-memory prressure caused by "move".
1364 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1366 VM_BUG_ON(!rcu_read_lock_held());
1367 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1370 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1372 struct mem_cgroup *from;
1373 struct mem_cgroup *to;
1374 bool ret = false;
1376 * Unlike task_move routines, we access mc.to, mc.from not under
1377 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1379 spin_lock(&mc.lock);
1380 from = mc.from;
1381 to = mc.to;
1382 if (!from)
1383 goto unlock;
1385 ret = mem_cgroup_same_or_subtree(mem, from)
1386 || mem_cgroup_same_or_subtree(mem, to);
1387 unlock:
1388 spin_unlock(&mc.lock);
1389 return ret;
1392 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1394 if (mc.moving_task && current != mc.moving_task) {
1395 if (mem_cgroup_under_move(mem)) {
1396 DEFINE_WAIT(wait);
1397 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1398 /* moving charge context might have finished. */
1399 if (mc.moving_task)
1400 schedule();
1401 finish_wait(&mc.waitq, &wait);
1402 return true;
1405 return false;
1409 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1410 * @memcg: The memory cgroup that went over limit
1411 * @p: Task that is going to be killed
1413 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1414 * enabled
1416 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1418 struct cgroup *task_cgrp;
1419 struct cgroup *mem_cgrp;
1421 * Need a buffer in BSS, can't rely on allocations. The code relies
1422 * on the assumption that OOM is serialized for memory controller.
1423 * If this assumption is broken, revisit this code.
1425 static char memcg_name[PATH_MAX];
1426 int ret;
1428 if (!memcg || !p)
1429 return;
1432 rcu_read_lock();
1434 mem_cgrp = memcg->css.cgroup;
1435 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1437 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1438 if (ret < 0) {
1440 * Unfortunately, we are unable to convert to a useful name
1441 * But we'll still print out the usage information
1443 rcu_read_unlock();
1444 goto done;
1446 rcu_read_unlock();
1448 printk(KERN_INFO "Task in %s killed", memcg_name);
1450 rcu_read_lock();
1451 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1452 if (ret < 0) {
1453 rcu_read_unlock();
1454 goto done;
1456 rcu_read_unlock();
1459 * Continues from above, so we don't need an KERN_ level
1461 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1462 done:
1464 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1465 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1466 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1467 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1468 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1469 "failcnt %llu\n",
1470 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1471 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1472 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1476 * This function returns the number of memcg under hierarchy tree. Returns
1477 * 1(self count) if no children.
1479 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1481 int num = 0;
1482 struct mem_cgroup *iter;
1484 for_each_mem_cgroup_tree(iter, mem)
1485 num++;
1486 return num;
1490 * Return the memory (and swap, if configured) limit for a memcg.
1492 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1494 u64 limit;
1495 u64 memsw;
1497 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1498 limit += total_swap_pages << PAGE_SHIFT;
1500 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1502 * If memsw is finite and limits the amount of swap space available
1503 * to this memcg, return that limit.
1505 return min(limit, memsw);
1509 * Visit the first child (need not be the first child as per the ordering
1510 * of the cgroup list, since we track last_scanned_child) of @mem and use
1511 * that to reclaim free pages from.
1513 static struct mem_cgroup *
1514 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1516 struct mem_cgroup *ret = NULL;
1517 struct cgroup_subsys_state *css;
1518 int nextid, found;
1520 if (!root_mem->use_hierarchy) {
1521 css_get(&root_mem->css);
1522 ret = root_mem;
1525 while (!ret) {
1526 rcu_read_lock();
1527 nextid = root_mem->last_scanned_child + 1;
1528 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1529 &found);
1530 if (css && css_tryget(css))
1531 ret = container_of(css, struct mem_cgroup, css);
1533 rcu_read_unlock();
1534 /* Updates scanning parameter */
1535 if (!css) {
1536 /* this means start scan from ID:1 */
1537 root_mem->last_scanned_child = 0;
1538 } else
1539 root_mem->last_scanned_child = found;
1542 return ret;
1546 * test_mem_cgroup_node_reclaimable
1547 * @mem: the target memcg
1548 * @nid: the node ID to be checked.
1549 * @noswap : specify true here if the user wants flle only information.
1551 * This function returns whether the specified memcg contains any
1552 * reclaimable pages on a node. Returns true if there are any reclaimable
1553 * pages in the node.
1555 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1556 int nid, bool noswap)
1558 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1559 return true;
1560 if (noswap || !total_swap_pages)
1561 return false;
1562 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1563 return true;
1564 return false;
1567 #if MAX_NUMNODES > 1
1570 * Always updating the nodemask is not very good - even if we have an empty
1571 * list or the wrong list here, we can start from some node and traverse all
1572 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1575 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1577 int nid;
1579 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1580 * pagein/pageout changes since the last update.
1582 if (!atomic_read(&mem->numainfo_events))
1583 return;
1584 if (atomic_inc_return(&mem->numainfo_updating) > 1)
1585 return;
1587 /* make a nodemask where this memcg uses memory from */
1588 mem->scan_nodes = node_states[N_HIGH_MEMORY];
1590 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1592 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1593 node_clear(nid, mem->scan_nodes);
1596 atomic_set(&mem->numainfo_events, 0);
1597 atomic_set(&mem->numainfo_updating, 0);
1601 * Selecting a node where we start reclaim from. Because what we need is just
1602 * reducing usage counter, start from anywhere is O,K. Considering
1603 * memory reclaim from current node, there are pros. and cons.
1605 * Freeing memory from current node means freeing memory from a node which
1606 * we'll use or we've used. So, it may make LRU bad. And if several threads
1607 * hit limits, it will see a contention on a node. But freeing from remote
1608 * node means more costs for memory reclaim because of memory latency.
1610 * Now, we use round-robin. Better algorithm is welcomed.
1612 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1614 int node;
1616 mem_cgroup_may_update_nodemask(mem);
1617 node = mem->last_scanned_node;
1619 node = next_node(node, mem->scan_nodes);
1620 if (node == MAX_NUMNODES)
1621 node = first_node(mem->scan_nodes);
1623 * We call this when we hit limit, not when pages are added to LRU.
1624 * No LRU may hold pages because all pages are UNEVICTABLE or
1625 * memcg is too small and all pages are not on LRU. In that case,
1626 * we use curret node.
1628 if (unlikely(node == MAX_NUMNODES))
1629 node = numa_node_id();
1631 mem->last_scanned_node = node;
1632 return node;
1636 * Check all nodes whether it contains reclaimable pages or not.
1637 * For quick scan, we make use of scan_nodes. This will allow us to skip
1638 * unused nodes. But scan_nodes is lazily updated and may not cotain
1639 * enough new information. We need to do double check.
1641 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1643 int nid;
1646 * quick check...making use of scan_node.
1647 * We can skip unused nodes.
1649 if (!nodes_empty(mem->scan_nodes)) {
1650 for (nid = first_node(mem->scan_nodes);
1651 nid < MAX_NUMNODES;
1652 nid = next_node(nid, mem->scan_nodes)) {
1654 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1655 return true;
1659 * Check rest of nodes.
1661 for_each_node_state(nid, N_HIGH_MEMORY) {
1662 if (node_isset(nid, mem->scan_nodes))
1663 continue;
1664 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1665 return true;
1667 return false;
1670 #else
1671 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1673 return 0;
1676 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1678 return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1680 #endif
1682 static void __mem_cgroup_record_scanstat(unsigned long *stats,
1683 struct memcg_scanrecord *rec)
1686 stats[SCAN] += rec->nr_scanned[0] + rec->nr_scanned[1];
1687 stats[SCAN_ANON] += rec->nr_scanned[0];
1688 stats[SCAN_FILE] += rec->nr_scanned[1];
1690 stats[ROTATE] += rec->nr_rotated[0] + rec->nr_rotated[1];
1691 stats[ROTATE_ANON] += rec->nr_rotated[0];
1692 stats[ROTATE_FILE] += rec->nr_rotated[1];
1694 stats[FREED] += rec->nr_freed[0] + rec->nr_freed[1];
1695 stats[FREED_ANON] += rec->nr_freed[0];
1696 stats[FREED_FILE] += rec->nr_freed[1];
1698 stats[ELAPSED] += rec->elapsed;
1701 static void mem_cgroup_record_scanstat(struct memcg_scanrecord *rec)
1703 struct mem_cgroup *mem;
1704 int context = rec->context;
1706 if (context >= NR_SCAN_CONTEXT)
1707 return;
1709 mem = rec->mem;
1710 spin_lock(&mem->scanstat.lock);
1711 __mem_cgroup_record_scanstat(mem->scanstat.stats[context], rec);
1712 spin_unlock(&mem->scanstat.lock);
1714 mem = rec->root;
1715 spin_lock(&mem->scanstat.lock);
1716 __mem_cgroup_record_scanstat(mem->scanstat.rootstats[context], rec);
1717 spin_unlock(&mem->scanstat.lock);
1721 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1722 * we reclaimed from, so that we don't end up penalizing one child extensively
1723 * based on its position in the children list.
1725 * root_mem is the original ancestor that we've been reclaim from.
1727 * We give up and return to the caller when we visit root_mem twice.
1728 * (other groups can be removed while we're walking....)
1730 * If shrink==true, for avoiding to free too much, this returns immedieately.
1732 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1733 struct zone *zone,
1734 gfp_t gfp_mask,
1735 unsigned long reclaim_options,
1736 unsigned long *total_scanned)
1738 struct mem_cgroup *victim;
1739 int ret, total = 0;
1740 int loop = 0;
1741 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1742 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1743 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1744 struct memcg_scanrecord rec;
1745 unsigned long excess;
1746 unsigned long scanned;
1748 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1750 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1751 if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1752 noswap = true;
1754 if (shrink)
1755 rec.context = SCAN_BY_SHRINK;
1756 else if (check_soft)
1757 rec.context = SCAN_BY_SYSTEM;
1758 else
1759 rec.context = SCAN_BY_LIMIT;
1761 rec.root = root_mem;
1763 while (1) {
1764 victim = mem_cgroup_select_victim(root_mem);
1765 if (victim == root_mem) {
1766 loop++;
1768 * We are not draining per cpu cached charges during
1769 * soft limit reclaim because global reclaim doesn't
1770 * care about charges. It tries to free some memory and
1771 * charges will not give any.
1773 if (!check_soft && loop >= 1)
1774 drain_all_stock_async(root_mem);
1775 if (loop >= 2) {
1777 * If we have not been able to reclaim
1778 * anything, it might because there are
1779 * no reclaimable pages under this hierarchy
1781 if (!check_soft || !total) {
1782 css_put(&victim->css);
1783 break;
1786 * We want to do more targeted reclaim.
1787 * excess >> 2 is not to excessive so as to
1788 * reclaim too much, nor too less that we keep
1789 * coming back to reclaim from this cgroup
1791 if (total >= (excess >> 2) ||
1792 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1793 css_put(&victim->css);
1794 break;
1798 if (!mem_cgroup_reclaimable(victim, noswap)) {
1799 /* this cgroup's local usage == 0 */
1800 css_put(&victim->css);
1801 continue;
1803 rec.mem = victim;
1804 rec.nr_scanned[0] = 0;
1805 rec.nr_scanned[1] = 0;
1806 rec.nr_rotated[0] = 0;
1807 rec.nr_rotated[1] = 0;
1808 rec.nr_freed[0] = 0;
1809 rec.nr_freed[1] = 0;
1810 rec.elapsed = 0;
1811 /* we use swappiness of local cgroup */
1812 if (check_soft) {
1813 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1814 noswap, zone, &rec, &scanned);
1815 *total_scanned += scanned;
1816 } else
1817 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1818 noswap, &rec);
1819 mem_cgroup_record_scanstat(&rec);
1820 css_put(&victim->css);
1822 * At shrinking usage, we can't check we should stop here or
1823 * reclaim more. It's depends on callers. last_scanned_child
1824 * will work enough for keeping fairness under tree.
1826 if (shrink)
1827 return ret;
1828 total += ret;
1829 if (check_soft) {
1830 if (!res_counter_soft_limit_excess(&root_mem->res))
1831 return total;
1832 } else if (mem_cgroup_margin(root_mem))
1833 return total;
1835 return total;
1839 * Check OOM-Killer is already running under our hierarchy.
1840 * If someone is running, return false.
1841 * Has to be called with memcg_oom_lock
1843 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1845 int lock_count = -1;
1846 struct mem_cgroup *iter, *failed = NULL;
1847 bool cond = true;
1849 for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1850 bool locked = iter->oom_lock;
1852 iter->oom_lock = true;
1853 if (lock_count == -1)
1854 lock_count = iter->oom_lock;
1855 else if (lock_count != locked) {
1857 * this subtree of our hierarchy is already locked
1858 * so we cannot give a lock.
1860 lock_count = 0;
1861 failed = iter;
1862 cond = false;
1866 if (!failed)
1867 goto done;
1870 * OK, we failed to lock the whole subtree so we have to clean up
1871 * what we set up to the failing subtree
1873 cond = true;
1874 for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1875 if (iter == failed) {
1876 cond = false;
1877 continue;
1879 iter->oom_lock = false;
1881 done:
1882 return lock_count;
1886 * Has to be called with memcg_oom_lock
1888 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1890 struct mem_cgroup *iter;
1892 for_each_mem_cgroup_tree(iter, mem)
1893 iter->oom_lock = false;
1894 return 0;
1897 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1899 struct mem_cgroup *iter;
1901 for_each_mem_cgroup_tree(iter, mem)
1902 atomic_inc(&iter->under_oom);
1905 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1907 struct mem_cgroup *iter;
1910 * When a new child is created while the hierarchy is under oom,
1911 * mem_cgroup_oom_lock() may not be called. We have to use
1912 * atomic_add_unless() here.
1914 for_each_mem_cgroup_tree(iter, mem)
1915 atomic_add_unless(&iter->under_oom, -1, 0);
1918 static DEFINE_SPINLOCK(memcg_oom_lock);
1919 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1921 struct oom_wait_info {
1922 struct mem_cgroup *mem;
1923 wait_queue_t wait;
1926 static int memcg_oom_wake_function(wait_queue_t *wait,
1927 unsigned mode, int sync, void *arg)
1929 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg,
1930 *oom_wait_mem;
1931 struct oom_wait_info *oom_wait_info;
1933 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1934 oom_wait_mem = oom_wait_info->mem;
1937 * Both of oom_wait_info->mem and wake_mem are stable under us.
1938 * Then we can use css_is_ancestor without taking care of RCU.
1940 if (!mem_cgroup_same_or_subtree(oom_wait_mem, wake_mem)
1941 && !mem_cgroup_same_or_subtree(wake_mem, oom_wait_mem))
1942 return 0;
1943 return autoremove_wake_function(wait, mode, sync, arg);
1946 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1948 /* for filtering, pass "mem" as argument. */
1949 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1952 static void memcg_oom_recover(struct mem_cgroup *mem)
1954 if (mem && atomic_read(&mem->under_oom))
1955 memcg_wakeup_oom(mem);
1959 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1961 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1963 struct oom_wait_info owait;
1964 bool locked, need_to_kill;
1966 owait.mem = mem;
1967 owait.wait.flags = 0;
1968 owait.wait.func = memcg_oom_wake_function;
1969 owait.wait.private = current;
1970 INIT_LIST_HEAD(&owait.wait.task_list);
1971 need_to_kill = true;
1972 mem_cgroup_mark_under_oom(mem);
1974 /* At first, try to OOM lock hierarchy under mem.*/
1975 spin_lock(&memcg_oom_lock);
1976 locked = mem_cgroup_oom_lock(mem);
1978 * Even if signal_pending(), we can't quit charge() loop without
1979 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1980 * under OOM is always welcomed, use TASK_KILLABLE here.
1982 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1983 if (!locked || mem->oom_kill_disable)
1984 need_to_kill = false;
1985 if (locked)
1986 mem_cgroup_oom_notify(mem);
1987 spin_unlock(&memcg_oom_lock);
1989 if (need_to_kill) {
1990 finish_wait(&memcg_oom_waitq, &owait.wait);
1991 mem_cgroup_out_of_memory(mem, mask);
1992 } else {
1993 schedule();
1994 finish_wait(&memcg_oom_waitq, &owait.wait);
1996 spin_lock(&memcg_oom_lock);
1997 if (locked)
1998 mem_cgroup_oom_unlock(mem);
1999 memcg_wakeup_oom(mem);
2000 spin_unlock(&memcg_oom_lock);
2002 mem_cgroup_unmark_under_oom(mem);
2004 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2005 return false;
2006 /* Give chance to dying process */
2007 schedule_timeout(1);
2008 return true;
2012 * Currently used to update mapped file statistics, but the routine can be
2013 * generalized to update other statistics as well.
2015 * Notes: Race condition
2017 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2018 * it tends to be costly. But considering some conditions, we doesn't need
2019 * to do so _always_.
2021 * Considering "charge", lock_page_cgroup() is not required because all
2022 * file-stat operations happen after a page is attached to radix-tree. There
2023 * are no race with "charge".
2025 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2026 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2027 * if there are race with "uncharge". Statistics itself is properly handled
2028 * by flags.
2030 * Considering "move", this is an only case we see a race. To make the race
2031 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
2032 * possibility of race condition. If there is, we take a lock.
2035 void mem_cgroup_update_page_stat(struct page *page,
2036 enum mem_cgroup_page_stat_item idx, int val)
2038 struct mem_cgroup *mem;
2039 struct page_cgroup *pc = lookup_page_cgroup(page);
2040 bool need_unlock = false;
2041 unsigned long uninitialized_var(flags);
2043 if (unlikely(!pc))
2044 return;
2046 rcu_read_lock();
2047 mem = pc->mem_cgroup;
2048 if (unlikely(!mem || !PageCgroupUsed(pc)))
2049 goto out;
2050 /* pc->mem_cgroup is unstable ? */
2051 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
2052 /* take a lock against to access pc->mem_cgroup */
2053 move_lock_page_cgroup(pc, &flags);
2054 need_unlock = true;
2055 mem = pc->mem_cgroup;
2056 if (!mem || !PageCgroupUsed(pc))
2057 goto out;
2060 switch (idx) {
2061 case MEMCG_NR_FILE_MAPPED:
2062 if (val > 0)
2063 SetPageCgroupFileMapped(pc);
2064 else if (!page_mapped(page))
2065 ClearPageCgroupFileMapped(pc);
2066 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2067 break;
2068 default:
2069 BUG();
2072 this_cpu_add(mem->stat->count[idx], val);
2074 out:
2075 if (unlikely(need_unlock))
2076 move_unlock_page_cgroup(pc, &flags);
2077 rcu_read_unlock();
2078 return;
2080 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2083 * size of first charge trial. "32" comes from vmscan.c's magic value.
2084 * TODO: maybe necessary to use big numbers in big irons.
2086 #define CHARGE_BATCH 32U
2087 struct memcg_stock_pcp {
2088 struct mem_cgroup *cached; /* this never be root cgroup */
2089 unsigned int nr_pages;
2090 struct work_struct work;
2091 unsigned long flags;
2092 #define FLUSHING_CACHED_CHARGE (0)
2094 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2095 static DEFINE_MUTEX(percpu_charge_mutex);
2098 * Try to consume stocked charge on this cpu. If success, one page is consumed
2099 * from local stock and true is returned. If the stock is 0 or charges from a
2100 * cgroup which is not current target, returns false. This stock will be
2101 * refilled.
2103 static bool consume_stock(struct mem_cgroup *mem)
2105 struct memcg_stock_pcp *stock;
2106 bool ret = true;
2108 stock = &get_cpu_var(memcg_stock);
2109 if (mem == stock->cached && stock->nr_pages)
2110 stock->nr_pages--;
2111 else /* need to call res_counter_charge */
2112 ret = false;
2113 put_cpu_var(memcg_stock);
2114 return ret;
2118 * Returns stocks cached in percpu to res_counter and reset cached information.
2120 static void drain_stock(struct memcg_stock_pcp *stock)
2122 struct mem_cgroup *old = stock->cached;
2124 if (stock->nr_pages) {
2125 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2127 res_counter_uncharge(&old->res, bytes);
2128 if (do_swap_account)
2129 res_counter_uncharge(&old->memsw, bytes);
2130 stock->nr_pages = 0;
2132 stock->cached = NULL;
2136 * This must be called under preempt disabled or must be called by
2137 * a thread which is pinned to local cpu.
2139 static void drain_local_stock(struct work_struct *dummy)
2141 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2142 drain_stock(stock);
2143 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2147 * Cache charges(val) which is from res_counter, to local per_cpu area.
2148 * This will be consumed by consume_stock() function, later.
2150 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2152 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2154 if (stock->cached != mem) { /* reset if necessary */
2155 drain_stock(stock);
2156 stock->cached = mem;
2158 stock->nr_pages += nr_pages;
2159 put_cpu_var(memcg_stock);
2163 * Drains all per-CPU charge caches for given root_mem resp. subtree
2164 * of the hierarchy under it. sync flag says whether we should block
2165 * until the work is done.
2167 static void drain_all_stock(struct mem_cgroup *root_mem, bool sync)
2169 int cpu, curcpu;
2171 /* Notify other cpus that system-wide "drain" is running */
2172 get_online_cpus();
2174 * Get a hint for avoiding draining charges on the current cpu,
2175 * which must be exhausted by our charging. It is not required that
2176 * this be a precise check, so we use raw_smp_processor_id() instead of
2177 * getcpu()/putcpu().
2179 curcpu = raw_smp_processor_id();
2180 for_each_online_cpu(cpu) {
2181 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2182 struct mem_cgroup *mem;
2184 mem = stock->cached;
2185 if (!mem || !stock->nr_pages)
2186 continue;
2187 if (!mem_cgroup_same_or_subtree(root_mem, mem))
2188 continue;
2189 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2190 if (cpu == curcpu)
2191 drain_local_stock(&stock->work);
2192 else
2193 schedule_work_on(cpu, &stock->work);
2197 if (!sync)
2198 goto out;
2200 for_each_online_cpu(cpu) {
2201 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2202 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2203 flush_work(&stock->work);
2205 out:
2206 put_online_cpus();
2210 * Tries to drain stocked charges in other cpus. This function is asynchronous
2211 * and just put a work per cpu for draining localy on each cpu. Caller can
2212 * expects some charges will be back to res_counter later but cannot wait for
2213 * it.
2215 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2218 * If someone calls draining, avoid adding more kworker runs.
2220 if (!mutex_trylock(&percpu_charge_mutex))
2221 return;
2222 drain_all_stock(root_mem, false);
2223 mutex_unlock(&percpu_charge_mutex);
2226 /* This is a synchronous drain interface. */
2227 static void drain_all_stock_sync(struct mem_cgroup *root_mem)
2229 /* called when force_empty is called */
2230 mutex_lock(&percpu_charge_mutex);
2231 drain_all_stock(root_mem, true);
2232 mutex_unlock(&percpu_charge_mutex);
2236 * This function drains percpu counter value from DEAD cpu and
2237 * move it to local cpu. Note that this function can be preempted.
2239 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2241 int i;
2243 spin_lock(&mem->pcp_counter_lock);
2244 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2245 long x = per_cpu(mem->stat->count[i], cpu);
2247 per_cpu(mem->stat->count[i], cpu) = 0;
2248 mem->nocpu_base.count[i] += x;
2250 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2251 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2253 per_cpu(mem->stat->events[i], cpu) = 0;
2254 mem->nocpu_base.events[i] += x;
2256 /* need to clear ON_MOVE value, works as a kind of lock. */
2257 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2258 spin_unlock(&mem->pcp_counter_lock);
2261 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2263 int idx = MEM_CGROUP_ON_MOVE;
2265 spin_lock(&mem->pcp_counter_lock);
2266 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2267 spin_unlock(&mem->pcp_counter_lock);
2270 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2271 unsigned long action,
2272 void *hcpu)
2274 int cpu = (unsigned long)hcpu;
2275 struct memcg_stock_pcp *stock;
2276 struct mem_cgroup *iter;
2278 if ((action == CPU_ONLINE)) {
2279 for_each_mem_cgroup_all(iter)
2280 synchronize_mem_cgroup_on_move(iter, cpu);
2281 return NOTIFY_OK;
2284 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2285 return NOTIFY_OK;
2287 for_each_mem_cgroup_all(iter)
2288 mem_cgroup_drain_pcp_counter(iter, cpu);
2290 stock = &per_cpu(memcg_stock, cpu);
2291 drain_stock(stock);
2292 return NOTIFY_OK;
2296 /* See __mem_cgroup_try_charge() for details */
2297 enum {
2298 CHARGE_OK, /* success */
2299 CHARGE_RETRY, /* need to retry but retry is not bad */
2300 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2301 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2302 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2305 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2306 unsigned int nr_pages, bool oom_check)
2308 unsigned long csize = nr_pages * PAGE_SIZE;
2309 struct mem_cgroup *mem_over_limit;
2310 struct res_counter *fail_res;
2311 unsigned long flags = 0;
2312 int ret;
2314 ret = res_counter_charge(&mem->res, csize, &fail_res);
2316 if (likely(!ret)) {
2317 if (!do_swap_account)
2318 return CHARGE_OK;
2319 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2320 if (likely(!ret))
2321 return CHARGE_OK;
2323 res_counter_uncharge(&mem->res, csize);
2324 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2325 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2326 } else
2327 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2329 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2330 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2332 * Never reclaim on behalf of optional batching, retry with a
2333 * single page instead.
2335 if (nr_pages == CHARGE_BATCH)
2336 return CHARGE_RETRY;
2338 if (!(gfp_mask & __GFP_WAIT))
2339 return CHARGE_WOULDBLOCK;
2341 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2342 gfp_mask, flags, NULL);
2343 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2344 return CHARGE_RETRY;
2346 * Even though the limit is exceeded at this point, reclaim
2347 * may have been able to free some pages. Retry the charge
2348 * before killing the task.
2350 * Only for regular pages, though: huge pages are rather
2351 * unlikely to succeed so close to the limit, and we fall back
2352 * to regular pages anyway in case of failure.
2354 if (nr_pages == 1 && ret)
2355 return CHARGE_RETRY;
2358 * At task move, charge accounts can be doubly counted. So, it's
2359 * better to wait until the end of task_move if something is going on.
2361 if (mem_cgroup_wait_acct_move(mem_over_limit))
2362 return CHARGE_RETRY;
2364 /* If we don't need to call oom-killer at el, return immediately */
2365 if (!oom_check)
2366 return CHARGE_NOMEM;
2367 /* check OOM */
2368 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2369 return CHARGE_OOM_DIE;
2371 return CHARGE_RETRY;
2375 * Unlike exported interface, "oom" parameter is added. if oom==true,
2376 * oom-killer can be invoked.
2378 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2379 gfp_t gfp_mask,
2380 unsigned int nr_pages,
2381 struct mem_cgroup **memcg,
2382 bool oom)
2384 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2385 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2386 struct mem_cgroup *mem = NULL;
2387 int ret;
2390 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2391 * in system level. So, allow to go ahead dying process in addition to
2392 * MEMDIE process.
2394 if (unlikely(test_thread_flag(TIF_MEMDIE)
2395 || fatal_signal_pending(current)))
2396 goto bypass;
2399 * We always charge the cgroup the mm_struct belongs to.
2400 * The mm_struct's mem_cgroup changes on task migration if the
2401 * thread group leader migrates. It's possible that mm is not
2402 * set, if so charge the init_mm (happens for pagecache usage).
2404 if (!*memcg && !mm)
2405 goto bypass;
2406 again:
2407 if (*memcg) { /* css should be a valid one */
2408 mem = *memcg;
2409 VM_BUG_ON(css_is_removed(&mem->css));
2410 if (mem_cgroup_is_root(mem))
2411 goto done;
2412 if (nr_pages == 1 && consume_stock(mem))
2413 goto done;
2414 css_get(&mem->css);
2415 } else {
2416 struct task_struct *p;
2418 rcu_read_lock();
2419 p = rcu_dereference(mm->owner);
2421 * Because we don't have task_lock(), "p" can exit.
2422 * In that case, "mem" can point to root or p can be NULL with
2423 * race with swapoff. Then, we have small risk of mis-accouning.
2424 * But such kind of mis-account by race always happens because
2425 * we don't have cgroup_mutex(). It's overkill and we allo that
2426 * small race, here.
2427 * (*) swapoff at el will charge against mm-struct not against
2428 * task-struct. So, mm->owner can be NULL.
2430 mem = mem_cgroup_from_task(p);
2431 if (!mem || mem_cgroup_is_root(mem)) {
2432 rcu_read_unlock();
2433 goto done;
2435 if (nr_pages == 1 && consume_stock(mem)) {
2437 * It seems dagerous to access memcg without css_get().
2438 * But considering how consume_stok works, it's not
2439 * necessary. If consume_stock success, some charges
2440 * from this memcg are cached on this cpu. So, we
2441 * don't need to call css_get()/css_tryget() before
2442 * calling consume_stock().
2444 rcu_read_unlock();
2445 goto done;
2447 /* after here, we may be blocked. we need to get refcnt */
2448 if (!css_tryget(&mem->css)) {
2449 rcu_read_unlock();
2450 goto again;
2452 rcu_read_unlock();
2455 do {
2456 bool oom_check;
2458 /* If killed, bypass charge */
2459 if (fatal_signal_pending(current)) {
2460 css_put(&mem->css);
2461 goto bypass;
2464 oom_check = false;
2465 if (oom && !nr_oom_retries) {
2466 oom_check = true;
2467 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2470 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2471 switch (ret) {
2472 case CHARGE_OK:
2473 break;
2474 case CHARGE_RETRY: /* not in OOM situation but retry */
2475 batch = nr_pages;
2476 css_put(&mem->css);
2477 mem = NULL;
2478 goto again;
2479 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2480 css_put(&mem->css);
2481 goto nomem;
2482 case CHARGE_NOMEM: /* OOM routine works */
2483 if (!oom) {
2484 css_put(&mem->css);
2485 goto nomem;
2487 /* If oom, we never return -ENOMEM */
2488 nr_oom_retries--;
2489 break;
2490 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2491 css_put(&mem->css);
2492 goto bypass;
2494 } while (ret != CHARGE_OK);
2496 if (batch > nr_pages)
2497 refill_stock(mem, batch - nr_pages);
2498 css_put(&mem->css);
2499 done:
2500 *memcg = mem;
2501 return 0;
2502 nomem:
2503 *memcg = NULL;
2504 return -ENOMEM;
2505 bypass:
2506 *memcg = NULL;
2507 return 0;
2511 * Somemtimes we have to undo a charge we got by try_charge().
2512 * This function is for that and do uncharge, put css's refcnt.
2513 * gotten by try_charge().
2515 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2516 unsigned int nr_pages)
2518 if (!mem_cgroup_is_root(mem)) {
2519 unsigned long bytes = nr_pages * PAGE_SIZE;
2521 res_counter_uncharge(&mem->res, bytes);
2522 if (do_swap_account)
2523 res_counter_uncharge(&mem->memsw, bytes);
2528 * A helper function to get mem_cgroup from ID. must be called under
2529 * rcu_read_lock(). The caller must check css_is_removed() or some if
2530 * it's concern. (dropping refcnt from swap can be called against removed
2531 * memcg.)
2533 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2535 struct cgroup_subsys_state *css;
2537 /* ID 0 is unused ID */
2538 if (!id)
2539 return NULL;
2540 css = css_lookup(&mem_cgroup_subsys, id);
2541 if (!css)
2542 return NULL;
2543 return container_of(css, struct mem_cgroup, css);
2546 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2548 struct mem_cgroup *mem = NULL;
2549 struct page_cgroup *pc;
2550 unsigned short id;
2551 swp_entry_t ent;
2553 VM_BUG_ON(!PageLocked(page));
2555 pc = lookup_page_cgroup(page);
2556 lock_page_cgroup(pc);
2557 if (PageCgroupUsed(pc)) {
2558 mem = pc->mem_cgroup;
2559 if (mem && !css_tryget(&mem->css))
2560 mem = NULL;
2561 } else if (PageSwapCache(page)) {
2562 ent.val = page_private(page);
2563 id = lookup_swap_cgroup(ent);
2564 rcu_read_lock();
2565 mem = mem_cgroup_lookup(id);
2566 if (mem && !css_tryget(&mem->css))
2567 mem = NULL;
2568 rcu_read_unlock();
2570 unlock_page_cgroup(pc);
2571 return mem;
2574 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2575 struct page *page,
2576 unsigned int nr_pages,
2577 struct page_cgroup *pc,
2578 enum charge_type ctype)
2580 lock_page_cgroup(pc);
2581 if (unlikely(PageCgroupUsed(pc))) {
2582 unlock_page_cgroup(pc);
2583 __mem_cgroup_cancel_charge(mem, nr_pages);
2584 return;
2587 * we don't need page_cgroup_lock about tail pages, becase they are not
2588 * accessed by any other context at this point.
2590 pc->mem_cgroup = mem;
2592 * We access a page_cgroup asynchronously without lock_page_cgroup().
2593 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2594 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2595 * before USED bit, we need memory barrier here.
2596 * See mem_cgroup_add_lru_list(), etc.
2598 smp_wmb();
2599 switch (ctype) {
2600 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2601 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2602 SetPageCgroupCache(pc);
2603 SetPageCgroupUsed(pc);
2604 break;
2605 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2606 ClearPageCgroupCache(pc);
2607 SetPageCgroupUsed(pc);
2608 break;
2609 default:
2610 break;
2613 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2614 unlock_page_cgroup(pc);
2616 * "charge_statistics" updated event counter. Then, check it.
2617 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2618 * if they exceeds softlimit.
2620 memcg_check_events(mem, page);
2623 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2625 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2626 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2628 * Because tail pages are not marked as "used", set it. We're under
2629 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2631 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2633 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2634 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2635 unsigned long flags;
2637 if (mem_cgroup_disabled())
2638 return;
2640 * We have no races with charge/uncharge but will have races with
2641 * page state accounting.
2643 move_lock_page_cgroup(head_pc, &flags);
2645 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2646 smp_wmb(); /* see __commit_charge() */
2647 if (PageCgroupAcctLRU(head_pc)) {
2648 enum lru_list lru;
2649 struct mem_cgroup_per_zone *mz;
2652 * LRU flags cannot be copied because we need to add tail
2653 *.page to LRU by generic call and our hook will be called.
2654 * We hold lru_lock, then, reduce counter directly.
2656 lru = page_lru(head);
2657 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2658 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2660 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2661 move_unlock_page_cgroup(head_pc, &flags);
2663 #endif
2666 * mem_cgroup_move_account - move account of the page
2667 * @page: the page
2668 * @nr_pages: number of regular pages (>1 for huge pages)
2669 * @pc: page_cgroup of the page.
2670 * @from: mem_cgroup which the page is moved from.
2671 * @to: mem_cgroup which the page is moved to. @from != @to.
2672 * @uncharge: whether we should call uncharge and css_put against @from.
2674 * The caller must confirm following.
2675 * - page is not on LRU (isolate_page() is useful.)
2676 * - compound_lock is held when nr_pages > 1
2678 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2679 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2680 * true, this function does "uncharge" from old cgroup, but it doesn't if
2681 * @uncharge is false, so a caller should do "uncharge".
2683 static int mem_cgroup_move_account(struct page *page,
2684 unsigned int nr_pages,
2685 struct page_cgroup *pc,
2686 struct mem_cgroup *from,
2687 struct mem_cgroup *to,
2688 bool uncharge)
2690 unsigned long flags;
2691 int ret;
2693 VM_BUG_ON(from == to);
2694 VM_BUG_ON(PageLRU(page));
2696 * The page is isolated from LRU. So, collapse function
2697 * will not handle this page. But page splitting can happen.
2698 * Do this check under compound_page_lock(). The caller should
2699 * hold it.
2701 ret = -EBUSY;
2702 if (nr_pages > 1 && !PageTransHuge(page))
2703 goto out;
2705 lock_page_cgroup(pc);
2707 ret = -EINVAL;
2708 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2709 goto unlock;
2711 move_lock_page_cgroup(pc, &flags);
2713 if (PageCgroupFileMapped(pc)) {
2714 /* Update mapped_file data for mem_cgroup */
2715 preempt_disable();
2716 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2717 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2718 preempt_enable();
2720 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2721 if (uncharge)
2722 /* This is not "cancel", but cancel_charge does all we need. */
2723 __mem_cgroup_cancel_charge(from, nr_pages);
2725 /* caller should have done css_get */
2726 pc->mem_cgroup = to;
2727 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2729 * We charges against "to" which may not have any tasks. Then, "to"
2730 * can be under rmdir(). But in current implementation, caller of
2731 * this function is just force_empty() and move charge, so it's
2732 * guaranteed that "to" is never removed. So, we don't check rmdir
2733 * status here.
2735 move_unlock_page_cgroup(pc, &flags);
2736 ret = 0;
2737 unlock:
2738 unlock_page_cgroup(pc);
2740 * check events
2742 memcg_check_events(to, page);
2743 memcg_check_events(from, page);
2744 out:
2745 return ret;
2749 * move charges to its parent.
2752 static int mem_cgroup_move_parent(struct page *page,
2753 struct page_cgroup *pc,
2754 struct mem_cgroup *child,
2755 gfp_t gfp_mask)
2757 struct cgroup *cg = child->css.cgroup;
2758 struct cgroup *pcg = cg->parent;
2759 struct mem_cgroup *parent;
2760 unsigned int nr_pages;
2761 unsigned long uninitialized_var(flags);
2762 int ret;
2764 /* Is ROOT ? */
2765 if (!pcg)
2766 return -EINVAL;
2768 ret = -EBUSY;
2769 if (!get_page_unless_zero(page))
2770 goto out;
2771 if (isolate_lru_page(page))
2772 goto put;
2774 nr_pages = hpage_nr_pages(page);
2776 parent = mem_cgroup_from_cont(pcg);
2777 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2778 if (ret || !parent)
2779 goto put_back;
2781 if (nr_pages > 1)
2782 flags = compound_lock_irqsave(page);
2784 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2785 if (ret)
2786 __mem_cgroup_cancel_charge(parent, nr_pages);
2788 if (nr_pages > 1)
2789 compound_unlock_irqrestore(page, flags);
2790 put_back:
2791 putback_lru_page(page);
2792 put:
2793 put_page(page);
2794 out:
2795 return ret;
2799 * Charge the memory controller for page usage.
2800 * Return
2801 * 0 if the charge was successful
2802 * < 0 if the cgroup is over its limit
2804 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2805 gfp_t gfp_mask, enum charge_type ctype)
2807 struct mem_cgroup *mem = NULL;
2808 unsigned int nr_pages = 1;
2809 struct page_cgroup *pc;
2810 bool oom = true;
2811 int ret;
2813 if (PageTransHuge(page)) {
2814 nr_pages <<= compound_order(page);
2815 VM_BUG_ON(!PageTransHuge(page));
2817 * Never OOM-kill a process for a huge page. The
2818 * fault handler will fall back to regular pages.
2820 oom = false;
2823 pc = lookup_page_cgroup(page);
2824 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2826 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2827 if (ret || !mem)
2828 return ret;
2830 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2831 return 0;
2834 int mem_cgroup_newpage_charge(struct page *page,
2835 struct mm_struct *mm, gfp_t gfp_mask)
2837 if (mem_cgroup_disabled())
2838 return 0;
2840 * If already mapped, we don't have to account.
2841 * If page cache, page->mapping has address_space.
2842 * But page->mapping may have out-of-use anon_vma pointer,
2843 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2844 * is NULL.
2846 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2847 return 0;
2848 if (unlikely(!mm))
2849 mm = &init_mm;
2850 return mem_cgroup_charge_common(page, mm, gfp_mask,
2851 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2854 static void
2855 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2856 enum charge_type ctype);
2858 static void
2859 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2860 enum charge_type ctype)
2862 struct page_cgroup *pc = lookup_page_cgroup(page);
2864 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2865 * is already on LRU. It means the page may on some other page_cgroup's
2866 * LRU. Take care of it.
2868 mem_cgroup_lru_del_before_commit(page);
2869 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2870 mem_cgroup_lru_add_after_commit(page);
2871 return;
2874 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2875 gfp_t gfp_mask)
2877 struct mem_cgroup *mem = NULL;
2878 int ret;
2880 if (mem_cgroup_disabled())
2881 return 0;
2882 if (PageCompound(page))
2883 return 0;
2885 if (unlikely(!mm))
2886 mm = &init_mm;
2888 if (page_is_file_cache(page)) {
2889 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2890 if (ret || !mem)
2891 return ret;
2894 * FUSE reuses pages without going through the final
2895 * put that would remove them from the LRU list, make
2896 * sure that they get relinked properly.
2898 __mem_cgroup_commit_charge_lrucare(page, mem,
2899 MEM_CGROUP_CHARGE_TYPE_CACHE);
2900 return ret;
2902 /* shmem */
2903 if (PageSwapCache(page)) {
2904 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2905 if (!ret)
2906 __mem_cgroup_commit_charge_swapin(page, mem,
2907 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2908 } else
2909 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2910 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2912 return ret;
2916 * While swap-in, try_charge -> commit or cancel, the page is locked.
2917 * And when try_charge() successfully returns, one refcnt to memcg without
2918 * struct page_cgroup is acquired. This refcnt will be consumed by
2919 * "commit()" or removed by "cancel()"
2921 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2922 struct page *page,
2923 gfp_t mask, struct mem_cgroup **ptr)
2925 struct mem_cgroup *mem;
2926 int ret;
2928 *ptr = NULL;
2930 if (mem_cgroup_disabled())
2931 return 0;
2933 if (!do_swap_account)
2934 goto charge_cur_mm;
2936 * A racing thread's fault, or swapoff, may have already updated
2937 * the pte, and even removed page from swap cache: in those cases
2938 * do_swap_page()'s pte_same() test will fail; but there's also a
2939 * KSM case which does need to charge the page.
2941 if (!PageSwapCache(page))
2942 goto charge_cur_mm;
2943 mem = try_get_mem_cgroup_from_page(page);
2944 if (!mem)
2945 goto charge_cur_mm;
2946 *ptr = mem;
2947 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2948 css_put(&mem->css);
2949 return ret;
2950 charge_cur_mm:
2951 if (unlikely(!mm))
2952 mm = &init_mm;
2953 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2956 static void
2957 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2958 enum charge_type ctype)
2960 if (mem_cgroup_disabled())
2961 return;
2962 if (!ptr)
2963 return;
2964 cgroup_exclude_rmdir(&ptr->css);
2966 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2968 * Now swap is on-memory. This means this page may be
2969 * counted both as mem and swap....double count.
2970 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2971 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2972 * may call delete_from_swap_cache() before reach here.
2974 if (do_swap_account && PageSwapCache(page)) {
2975 swp_entry_t ent = {.val = page_private(page)};
2976 unsigned short id;
2977 struct mem_cgroup *memcg;
2979 id = swap_cgroup_record(ent, 0);
2980 rcu_read_lock();
2981 memcg = mem_cgroup_lookup(id);
2982 if (memcg) {
2984 * This recorded memcg can be obsolete one. So, avoid
2985 * calling css_tryget
2987 if (!mem_cgroup_is_root(memcg))
2988 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2989 mem_cgroup_swap_statistics(memcg, false);
2990 mem_cgroup_put(memcg);
2992 rcu_read_unlock();
2995 * At swapin, we may charge account against cgroup which has no tasks.
2996 * So, rmdir()->pre_destroy() can be called while we do this charge.
2997 * In that case, we need to call pre_destroy() again. check it here.
2999 cgroup_release_and_wakeup_rmdir(&ptr->css);
3002 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
3004 __mem_cgroup_commit_charge_swapin(page, ptr,
3005 MEM_CGROUP_CHARGE_TYPE_MAPPED);
3008 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
3010 if (mem_cgroup_disabled())
3011 return;
3012 if (!mem)
3013 return;
3014 __mem_cgroup_cancel_charge(mem, 1);
3017 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
3018 unsigned int nr_pages,
3019 const enum charge_type ctype)
3021 struct memcg_batch_info *batch = NULL;
3022 bool uncharge_memsw = true;
3024 /* If swapout, usage of swap doesn't decrease */
3025 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3026 uncharge_memsw = false;
3028 batch = &current->memcg_batch;
3030 * In usual, we do css_get() when we remember memcg pointer.
3031 * But in this case, we keep res->usage until end of a series of
3032 * uncharges. Then, it's ok to ignore memcg's refcnt.
3034 if (!batch->memcg)
3035 batch->memcg = mem;
3037 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3038 * In those cases, all pages freed continuously can be expected to be in
3039 * the same cgroup and we have chance to coalesce uncharges.
3040 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3041 * because we want to do uncharge as soon as possible.
3044 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3045 goto direct_uncharge;
3047 if (nr_pages > 1)
3048 goto direct_uncharge;
3051 * In typical case, batch->memcg == mem. This means we can
3052 * merge a series of uncharges to an uncharge of res_counter.
3053 * If not, we uncharge res_counter ony by one.
3055 if (batch->memcg != mem)
3056 goto direct_uncharge;
3057 /* remember freed charge and uncharge it later */
3058 batch->nr_pages++;
3059 if (uncharge_memsw)
3060 batch->memsw_nr_pages++;
3061 return;
3062 direct_uncharge:
3063 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
3064 if (uncharge_memsw)
3065 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
3066 if (unlikely(batch->memcg != mem))
3067 memcg_oom_recover(mem);
3068 return;
3072 * uncharge if !page_mapped(page)
3074 static struct mem_cgroup *
3075 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3077 struct mem_cgroup *mem = NULL;
3078 unsigned int nr_pages = 1;
3079 struct page_cgroup *pc;
3081 if (mem_cgroup_disabled())
3082 return NULL;
3084 if (PageSwapCache(page))
3085 return NULL;
3087 if (PageTransHuge(page)) {
3088 nr_pages <<= compound_order(page);
3089 VM_BUG_ON(!PageTransHuge(page));
3092 * Check if our page_cgroup is valid
3094 pc = lookup_page_cgroup(page);
3095 if (unlikely(!pc || !PageCgroupUsed(pc)))
3096 return NULL;
3098 lock_page_cgroup(pc);
3100 mem = pc->mem_cgroup;
3102 if (!PageCgroupUsed(pc))
3103 goto unlock_out;
3105 switch (ctype) {
3106 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3107 case MEM_CGROUP_CHARGE_TYPE_DROP:
3108 /* See mem_cgroup_prepare_migration() */
3109 if (page_mapped(page) || PageCgroupMigration(pc))
3110 goto unlock_out;
3111 break;
3112 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3113 if (!PageAnon(page)) { /* Shared memory */
3114 if (page->mapping && !page_is_file_cache(page))
3115 goto unlock_out;
3116 } else if (page_mapped(page)) /* Anon */
3117 goto unlock_out;
3118 break;
3119 default:
3120 break;
3123 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3125 ClearPageCgroupUsed(pc);
3127 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3128 * freed from LRU. This is safe because uncharged page is expected not
3129 * to be reused (freed soon). Exception is SwapCache, it's handled by
3130 * special functions.
3133 unlock_page_cgroup(pc);
3135 * even after unlock, we have mem->res.usage here and this memcg
3136 * will never be freed.
3138 memcg_check_events(mem, page);
3139 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3140 mem_cgroup_swap_statistics(mem, true);
3141 mem_cgroup_get(mem);
3143 if (!mem_cgroup_is_root(mem))
3144 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3146 return mem;
3148 unlock_out:
3149 unlock_page_cgroup(pc);
3150 return NULL;
3153 void mem_cgroup_uncharge_page(struct page *page)
3155 /* early check. */
3156 if (page_mapped(page))
3157 return;
3158 if (page->mapping && !PageAnon(page))
3159 return;
3160 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3163 void mem_cgroup_uncharge_cache_page(struct page *page)
3165 VM_BUG_ON(page_mapped(page));
3166 VM_BUG_ON(page->mapping);
3167 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3171 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3172 * In that cases, pages are freed continuously and we can expect pages
3173 * are in the same memcg. All these calls itself limits the number of
3174 * pages freed at once, then uncharge_start/end() is called properly.
3175 * This may be called prural(2) times in a context,
3178 void mem_cgroup_uncharge_start(void)
3180 current->memcg_batch.do_batch++;
3181 /* We can do nest. */
3182 if (current->memcg_batch.do_batch == 1) {
3183 current->memcg_batch.memcg = NULL;
3184 current->memcg_batch.nr_pages = 0;
3185 current->memcg_batch.memsw_nr_pages = 0;
3189 void mem_cgroup_uncharge_end(void)
3191 struct memcg_batch_info *batch = &current->memcg_batch;
3193 if (!batch->do_batch)
3194 return;
3196 batch->do_batch--;
3197 if (batch->do_batch) /* If stacked, do nothing. */
3198 return;
3200 if (!batch->memcg)
3201 return;
3203 * This "batch->memcg" is valid without any css_get/put etc...
3204 * bacause we hide charges behind us.
3206 if (batch->nr_pages)
3207 res_counter_uncharge(&batch->memcg->res,
3208 batch->nr_pages * PAGE_SIZE);
3209 if (batch->memsw_nr_pages)
3210 res_counter_uncharge(&batch->memcg->memsw,
3211 batch->memsw_nr_pages * PAGE_SIZE);
3212 memcg_oom_recover(batch->memcg);
3213 /* forget this pointer (for sanity check) */
3214 batch->memcg = NULL;
3217 #ifdef CONFIG_SWAP
3219 * called after __delete_from_swap_cache() and drop "page" account.
3220 * memcg information is recorded to swap_cgroup of "ent"
3222 void
3223 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3225 struct mem_cgroup *memcg;
3226 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3228 if (!swapout) /* this was a swap cache but the swap is unused ! */
3229 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3231 memcg = __mem_cgroup_uncharge_common(page, ctype);
3234 * record memcg information, if swapout && memcg != NULL,
3235 * mem_cgroup_get() was called in uncharge().
3237 if (do_swap_account && swapout && memcg)
3238 swap_cgroup_record(ent, css_id(&memcg->css));
3240 #endif
3242 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3244 * called from swap_entry_free(). remove record in swap_cgroup and
3245 * uncharge "memsw" account.
3247 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3249 struct mem_cgroup *memcg;
3250 unsigned short id;
3252 if (!do_swap_account)
3253 return;
3255 id = swap_cgroup_record(ent, 0);
3256 rcu_read_lock();
3257 memcg = mem_cgroup_lookup(id);
3258 if (memcg) {
3260 * We uncharge this because swap is freed.
3261 * This memcg can be obsolete one. We avoid calling css_tryget
3263 if (!mem_cgroup_is_root(memcg))
3264 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3265 mem_cgroup_swap_statistics(memcg, false);
3266 mem_cgroup_put(memcg);
3268 rcu_read_unlock();
3272 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3273 * @entry: swap entry to be moved
3274 * @from: mem_cgroup which the entry is moved from
3275 * @to: mem_cgroup which the entry is moved to
3276 * @need_fixup: whether we should fixup res_counters and refcounts.
3278 * It succeeds only when the swap_cgroup's record for this entry is the same
3279 * as the mem_cgroup's id of @from.
3281 * Returns 0 on success, -EINVAL on failure.
3283 * The caller must have charged to @to, IOW, called res_counter_charge() about
3284 * both res and memsw, and called css_get().
3286 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3287 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3289 unsigned short old_id, new_id;
3291 old_id = css_id(&from->css);
3292 new_id = css_id(&to->css);
3294 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3295 mem_cgroup_swap_statistics(from, false);
3296 mem_cgroup_swap_statistics(to, true);
3298 * This function is only called from task migration context now.
3299 * It postpones res_counter and refcount handling till the end
3300 * of task migration(mem_cgroup_clear_mc()) for performance
3301 * improvement. But we cannot postpone mem_cgroup_get(to)
3302 * because if the process that has been moved to @to does
3303 * swap-in, the refcount of @to might be decreased to 0.
3305 mem_cgroup_get(to);
3306 if (need_fixup) {
3307 if (!mem_cgroup_is_root(from))
3308 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3309 mem_cgroup_put(from);
3311 * we charged both to->res and to->memsw, so we should
3312 * uncharge to->res.
3314 if (!mem_cgroup_is_root(to))
3315 res_counter_uncharge(&to->res, PAGE_SIZE);
3317 return 0;
3319 return -EINVAL;
3321 #else
3322 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3323 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3325 return -EINVAL;
3327 #endif
3330 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3331 * page belongs to.
3333 int mem_cgroup_prepare_migration(struct page *page,
3334 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3336 struct mem_cgroup *mem = NULL;
3337 struct page_cgroup *pc;
3338 enum charge_type ctype;
3339 int ret = 0;
3341 *ptr = NULL;
3343 VM_BUG_ON(PageTransHuge(page));
3344 if (mem_cgroup_disabled())
3345 return 0;
3347 pc = lookup_page_cgroup(page);
3348 lock_page_cgroup(pc);
3349 if (PageCgroupUsed(pc)) {
3350 mem = pc->mem_cgroup;
3351 css_get(&mem->css);
3353 * At migrating an anonymous page, its mapcount goes down
3354 * to 0 and uncharge() will be called. But, even if it's fully
3355 * unmapped, migration may fail and this page has to be
3356 * charged again. We set MIGRATION flag here and delay uncharge
3357 * until end_migration() is called
3359 * Corner Case Thinking
3360 * A)
3361 * When the old page was mapped as Anon and it's unmap-and-freed
3362 * while migration was ongoing.
3363 * If unmap finds the old page, uncharge() of it will be delayed
3364 * until end_migration(). If unmap finds a new page, it's
3365 * uncharged when it make mapcount to be 1->0. If unmap code
3366 * finds swap_migration_entry, the new page will not be mapped
3367 * and end_migration() will find it(mapcount==0).
3369 * B)
3370 * When the old page was mapped but migraion fails, the kernel
3371 * remaps it. A charge for it is kept by MIGRATION flag even
3372 * if mapcount goes down to 0. We can do remap successfully
3373 * without charging it again.
3375 * C)
3376 * The "old" page is under lock_page() until the end of
3377 * migration, so, the old page itself will not be swapped-out.
3378 * If the new page is swapped out before end_migraton, our
3379 * hook to usual swap-out path will catch the event.
3381 if (PageAnon(page))
3382 SetPageCgroupMigration(pc);
3384 unlock_page_cgroup(pc);
3386 * If the page is not charged at this point,
3387 * we return here.
3389 if (!mem)
3390 return 0;
3392 *ptr = mem;
3393 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3394 css_put(&mem->css);/* drop extra refcnt */
3395 if (ret || *ptr == NULL) {
3396 if (PageAnon(page)) {
3397 lock_page_cgroup(pc);
3398 ClearPageCgroupMigration(pc);
3399 unlock_page_cgroup(pc);
3401 * The old page may be fully unmapped while we kept it.
3403 mem_cgroup_uncharge_page(page);
3405 return -ENOMEM;
3408 * We charge new page before it's used/mapped. So, even if unlock_page()
3409 * is called before end_migration, we can catch all events on this new
3410 * page. In the case new page is migrated but not remapped, new page's
3411 * mapcount will be finally 0 and we call uncharge in end_migration().
3413 pc = lookup_page_cgroup(newpage);
3414 if (PageAnon(page))
3415 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3416 else if (page_is_file_cache(page))
3417 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3418 else
3419 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3420 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3421 return ret;
3424 /* remove redundant charge if migration failed*/
3425 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3426 struct page *oldpage, struct page *newpage, bool migration_ok)
3428 struct page *used, *unused;
3429 struct page_cgroup *pc;
3431 if (!mem)
3432 return;
3433 /* blocks rmdir() */
3434 cgroup_exclude_rmdir(&mem->css);
3435 if (!migration_ok) {
3436 used = oldpage;
3437 unused = newpage;
3438 } else {
3439 used = newpage;
3440 unused = oldpage;
3443 * We disallowed uncharge of pages under migration because mapcount
3444 * of the page goes down to zero, temporarly.
3445 * Clear the flag and check the page should be charged.
3447 pc = lookup_page_cgroup(oldpage);
3448 lock_page_cgroup(pc);
3449 ClearPageCgroupMigration(pc);
3450 unlock_page_cgroup(pc);
3452 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3455 * If a page is a file cache, radix-tree replacement is very atomic
3456 * and we can skip this check. When it was an Anon page, its mapcount
3457 * goes down to 0. But because we added MIGRATION flage, it's not
3458 * uncharged yet. There are several case but page->mapcount check
3459 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3460 * check. (see prepare_charge() also)
3462 if (PageAnon(used))
3463 mem_cgroup_uncharge_page(used);
3465 * At migration, we may charge account against cgroup which has no
3466 * tasks.
3467 * So, rmdir()->pre_destroy() can be called while we do this charge.
3468 * In that case, we need to call pre_destroy() again. check it here.
3470 cgroup_release_and_wakeup_rmdir(&mem->css);
3473 #ifdef CONFIG_DEBUG_VM
3474 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3476 struct page_cgroup *pc;
3478 pc = lookup_page_cgroup(page);
3479 if (likely(pc) && PageCgroupUsed(pc))
3480 return pc;
3481 return NULL;
3484 bool mem_cgroup_bad_page_check(struct page *page)
3486 if (mem_cgroup_disabled())
3487 return false;
3489 return lookup_page_cgroup_used(page) != NULL;
3492 void mem_cgroup_print_bad_page(struct page *page)
3494 struct page_cgroup *pc;
3496 pc = lookup_page_cgroup_used(page);
3497 if (pc) {
3498 int ret = -1;
3499 char *path;
3501 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3502 pc, pc->flags, pc->mem_cgroup);
3504 path = kmalloc(PATH_MAX, GFP_KERNEL);
3505 if (path) {
3506 rcu_read_lock();
3507 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3508 path, PATH_MAX);
3509 rcu_read_unlock();
3512 printk(KERN_CONT "(%s)\n",
3513 (ret < 0) ? "cannot get the path" : path);
3514 kfree(path);
3517 #endif
3519 static DEFINE_MUTEX(set_limit_mutex);
3521 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3522 unsigned long long val)
3524 int retry_count;
3525 u64 memswlimit, memlimit;
3526 int ret = 0;
3527 int children = mem_cgroup_count_children(memcg);
3528 u64 curusage, oldusage;
3529 int enlarge;
3532 * For keeping hierarchical_reclaim simple, how long we should retry
3533 * is depends on callers. We set our retry-count to be function
3534 * of # of children which we should visit in this loop.
3536 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3538 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3540 enlarge = 0;
3541 while (retry_count) {
3542 if (signal_pending(current)) {
3543 ret = -EINTR;
3544 break;
3547 * Rather than hide all in some function, I do this in
3548 * open coded manner. You see what this really does.
3549 * We have to guarantee mem->res.limit < mem->memsw.limit.
3551 mutex_lock(&set_limit_mutex);
3552 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3553 if (memswlimit < val) {
3554 ret = -EINVAL;
3555 mutex_unlock(&set_limit_mutex);
3556 break;
3559 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3560 if (memlimit < val)
3561 enlarge = 1;
3563 ret = res_counter_set_limit(&memcg->res, val);
3564 if (!ret) {
3565 if (memswlimit == val)
3566 memcg->memsw_is_minimum = true;
3567 else
3568 memcg->memsw_is_minimum = false;
3570 mutex_unlock(&set_limit_mutex);
3572 if (!ret)
3573 break;
3575 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3576 MEM_CGROUP_RECLAIM_SHRINK,
3577 NULL);
3578 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3579 /* Usage is reduced ? */
3580 if (curusage >= oldusage)
3581 retry_count--;
3582 else
3583 oldusage = curusage;
3585 if (!ret && enlarge)
3586 memcg_oom_recover(memcg);
3588 return ret;
3591 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3592 unsigned long long val)
3594 int retry_count;
3595 u64 memlimit, memswlimit, oldusage, curusage;
3596 int children = mem_cgroup_count_children(memcg);
3597 int ret = -EBUSY;
3598 int enlarge = 0;
3600 /* see mem_cgroup_resize_res_limit */
3601 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3602 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3603 while (retry_count) {
3604 if (signal_pending(current)) {
3605 ret = -EINTR;
3606 break;
3609 * Rather than hide all in some function, I do this in
3610 * open coded manner. You see what this really does.
3611 * We have to guarantee mem->res.limit < mem->memsw.limit.
3613 mutex_lock(&set_limit_mutex);
3614 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3615 if (memlimit > val) {
3616 ret = -EINVAL;
3617 mutex_unlock(&set_limit_mutex);
3618 break;
3620 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3621 if (memswlimit < val)
3622 enlarge = 1;
3623 ret = res_counter_set_limit(&memcg->memsw, val);
3624 if (!ret) {
3625 if (memlimit == val)
3626 memcg->memsw_is_minimum = true;
3627 else
3628 memcg->memsw_is_minimum = false;
3630 mutex_unlock(&set_limit_mutex);
3632 if (!ret)
3633 break;
3635 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3636 MEM_CGROUP_RECLAIM_NOSWAP |
3637 MEM_CGROUP_RECLAIM_SHRINK,
3638 NULL);
3639 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3640 /* Usage is reduced ? */
3641 if (curusage >= oldusage)
3642 retry_count--;
3643 else
3644 oldusage = curusage;
3646 if (!ret && enlarge)
3647 memcg_oom_recover(memcg);
3648 return ret;
3651 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3652 gfp_t gfp_mask,
3653 unsigned long *total_scanned)
3655 unsigned long nr_reclaimed = 0;
3656 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3657 unsigned long reclaimed;
3658 int loop = 0;
3659 struct mem_cgroup_tree_per_zone *mctz;
3660 unsigned long long excess;
3661 unsigned long nr_scanned;
3663 if (order > 0)
3664 return 0;
3666 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3668 * This loop can run a while, specially if mem_cgroup's continuously
3669 * keep exceeding their soft limit and putting the system under
3670 * pressure
3672 do {
3673 if (next_mz)
3674 mz = next_mz;
3675 else
3676 mz = mem_cgroup_largest_soft_limit_node(mctz);
3677 if (!mz)
3678 break;
3680 nr_scanned = 0;
3681 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3682 gfp_mask,
3683 MEM_CGROUP_RECLAIM_SOFT,
3684 &nr_scanned);
3685 nr_reclaimed += reclaimed;
3686 *total_scanned += nr_scanned;
3687 spin_lock(&mctz->lock);
3690 * If we failed to reclaim anything from this memory cgroup
3691 * it is time to move on to the next cgroup
3693 next_mz = NULL;
3694 if (!reclaimed) {
3695 do {
3697 * Loop until we find yet another one.
3699 * By the time we get the soft_limit lock
3700 * again, someone might have aded the
3701 * group back on the RB tree. Iterate to
3702 * make sure we get a different mem.
3703 * mem_cgroup_largest_soft_limit_node returns
3704 * NULL if no other cgroup is present on
3705 * the tree
3707 next_mz =
3708 __mem_cgroup_largest_soft_limit_node(mctz);
3709 if (next_mz == mz)
3710 css_put(&next_mz->mem->css);
3711 else /* next_mz == NULL or other memcg */
3712 break;
3713 } while (1);
3715 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3716 excess = res_counter_soft_limit_excess(&mz->mem->res);
3718 * One school of thought says that we should not add
3719 * back the node to the tree if reclaim returns 0.
3720 * But our reclaim could return 0, simply because due
3721 * to priority we are exposing a smaller subset of
3722 * memory to reclaim from. Consider this as a longer
3723 * term TODO.
3725 /* If excess == 0, no tree ops */
3726 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3727 spin_unlock(&mctz->lock);
3728 css_put(&mz->mem->css);
3729 loop++;
3731 * Could not reclaim anything and there are no more
3732 * mem cgroups to try or we seem to be looping without
3733 * reclaiming anything.
3735 if (!nr_reclaimed &&
3736 (next_mz == NULL ||
3737 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3738 break;
3739 } while (!nr_reclaimed);
3740 if (next_mz)
3741 css_put(&next_mz->mem->css);
3742 return nr_reclaimed;
3746 * This routine traverse page_cgroup in given list and drop them all.
3747 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3749 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3750 int node, int zid, enum lru_list lru)
3752 struct zone *zone;
3753 struct mem_cgroup_per_zone *mz;
3754 struct page_cgroup *pc, *busy;
3755 unsigned long flags, loop;
3756 struct list_head *list;
3757 int ret = 0;
3759 zone = &NODE_DATA(node)->node_zones[zid];
3760 mz = mem_cgroup_zoneinfo(mem, node, zid);
3761 list = &mz->lists[lru];
3763 loop = MEM_CGROUP_ZSTAT(mz, lru);
3764 /* give some margin against EBUSY etc...*/
3765 loop += 256;
3766 busy = NULL;
3767 while (loop--) {
3768 struct page *page;
3770 ret = 0;
3771 spin_lock_irqsave(&zone->lru_lock, flags);
3772 if (list_empty(list)) {
3773 spin_unlock_irqrestore(&zone->lru_lock, flags);
3774 break;
3776 pc = list_entry(list->prev, struct page_cgroup, lru);
3777 if (busy == pc) {
3778 list_move(&pc->lru, list);
3779 busy = NULL;
3780 spin_unlock_irqrestore(&zone->lru_lock, flags);
3781 continue;
3783 spin_unlock_irqrestore(&zone->lru_lock, flags);
3785 page = lookup_cgroup_page(pc);
3787 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3788 if (ret == -ENOMEM)
3789 break;
3791 if (ret == -EBUSY || ret == -EINVAL) {
3792 /* found lock contention or "pc" is obsolete. */
3793 busy = pc;
3794 cond_resched();
3795 } else
3796 busy = NULL;
3799 if (!ret && !list_empty(list))
3800 return -EBUSY;
3801 return ret;
3805 * make mem_cgroup's charge to be 0 if there is no task.
3806 * This enables deleting this mem_cgroup.
3808 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3810 int ret;
3811 int node, zid, shrink;
3812 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3813 struct cgroup *cgrp = mem->css.cgroup;
3815 css_get(&mem->css);
3817 shrink = 0;
3818 /* should free all ? */
3819 if (free_all)
3820 goto try_to_free;
3821 move_account:
3822 do {
3823 ret = -EBUSY;
3824 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3825 goto out;
3826 ret = -EINTR;
3827 if (signal_pending(current))
3828 goto out;
3829 /* This is for making all *used* pages to be on LRU. */
3830 lru_add_drain_all();
3831 drain_all_stock_sync(mem);
3832 ret = 0;
3833 mem_cgroup_start_move(mem);
3834 for_each_node_state(node, N_HIGH_MEMORY) {
3835 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3836 enum lru_list l;
3837 for_each_lru(l) {
3838 ret = mem_cgroup_force_empty_list(mem,
3839 node, zid, l);
3840 if (ret)
3841 break;
3844 if (ret)
3845 break;
3847 mem_cgroup_end_move(mem);
3848 memcg_oom_recover(mem);
3849 /* it seems parent cgroup doesn't have enough mem */
3850 if (ret == -ENOMEM)
3851 goto try_to_free;
3852 cond_resched();
3853 /* "ret" should also be checked to ensure all lists are empty. */
3854 } while (mem->res.usage > 0 || ret);
3855 out:
3856 css_put(&mem->css);
3857 return ret;
3859 try_to_free:
3860 /* returns EBUSY if there is a task or if we come here twice. */
3861 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3862 ret = -EBUSY;
3863 goto out;
3865 /* we call try-to-free pages for make this cgroup empty */
3866 lru_add_drain_all();
3867 /* try to free all pages in this cgroup */
3868 shrink = 1;
3869 while (nr_retries && mem->res.usage > 0) {
3870 struct memcg_scanrecord rec;
3871 int progress;
3873 if (signal_pending(current)) {
3874 ret = -EINTR;
3875 goto out;
3877 rec.context = SCAN_BY_SHRINK;
3878 rec.mem = mem;
3879 rec.root = mem;
3880 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3881 false, &rec);
3882 if (!progress) {
3883 nr_retries--;
3884 /* maybe some writeback is necessary */
3885 congestion_wait(BLK_RW_ASYNC, HZ/10);
3889 lru_add_drain();
3890 /* try move_account...there may be some *locked* pages. */
3891 goto move_account;
3894 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3896 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3900 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3902 return mem_cgroup_from_cont(cont)->use_hierarchy;
3905 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3906 u64 val)
3908 int retval = 0;
3909 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3910 struct cgroup *parent = cont->parent;
3911 struct mem_cgroup *parent_mem = NULL;
3913 if (parent)
3914 parent_mem = mem_cgroup_from_cont(parent);
3916 cgroup_lock();
3918 * If parent's use_hierarchy is set, we can't make any modifications
3919 * in the child subtrees. If it is unset, then the change can
3920 * occur, provided the current cgroup has no children.
3922 * For the root cgroup, parent_mem is NULL, we allow value to be
3923 * set if there are no children.
3925 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3926 (val == 1 || val == 0)) {
3927 if (list_empty(&cont->children))
3928 mem->use_hierarchy = val;
3929 else
3930 retval = -EBUSY;
3931 } else
3932 retval = -EINVAL;
3933 cgroup_unlock();
3935 return retval;
3939 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3940 enum mem_cgroup_stat_index idx)
3942 struct mem_cgroup *iter;
3943 long val = 0;
3945 /* Per-cpu values can be negative, use a signed accumulator */
3946 for_each_mem_cgroup_tree(iter, mem)
3947 val += mem_cgroup_read_stat(iter, idx);
3949 if (val < 0) /* race ? */
3950 val = 0;
3951 return val;
3954 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3956 u64 val;
3958 if (!mem_cgroup_is_root(mem)) {
3959 if (!swap)
3960 return res_counter_read_u64(&mem->res, RES_USAGE);
3961 else
3962 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3965 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3966 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3968 if (swap)
3969 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3971 return val << PAGE_SHIFT;
3974 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3976 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3977 u64 val;
3978 int type, name;
3980 type = MEMFILE_TYPE(cft->private);
3981 name = MEMFILE_ATTR(cft->private);
3982 switch (type) {
3983 case _MEM:
3984 if (name == RES_USAGE)
3985 val = mem_cgroup_usage(mem, false);
3986 else
3987 val = res_counter_read_u64(&mem->res, name);
3988 break;
3989 case _MEMSWAP:
3990 if (name == RES_USAGE)
3991 val = mem_cgroup_usage(mem, true);
3992 else
3993 val = res_counter_read_u64(&mem->memsw, name);
3994 break;
3995 default:
3996 BUG();
3997 break;
3999 return val;
4002 * The user of this function is...
4003 * RES_LIMIT.
4005 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4006 const char *buffer)
4008 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4009 int type, name;
4010 unsigned long long val;
4011 int ret;
4013 type = MEMFILE_TYPE(cft->private);
4014 name = MEMFILE_ATTR(cft->private);
4015 switch (name) {
4016 case RES_LIMIT:
4017 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4018 ret = -EINVAL;
4019 break;
4021 /* This function does all necessary parse...reuse it */
4022 ret = res_counter_memparse_write_strategy(buffer, &val);
4023 if (ret)
4024 break;
4025 if (type == _MEM)
4026 ret = mem_cgroup_resize_limit(memcg, val);
4027 else
4028 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4029 break;
4030 case RES_SOFT_LIMIT:
4031 ret = res_counter_memparse_write_strategy(buffer, &val);
4032 if (ret)
4033 break;
4035 * For memsw, soft limits are hard to implement in terms
4036 * of semantics, for now, we support soft limits for
4037 * control without swap
4039 if (type == _MEM)
4040 ret = res_counter_set_soft_limit(&memcg->res, val);
4041 else
4042 ret = -EINVAL;
4043 break;
4044 default:
4045 ret = -EINVAL; /* should be BUG() ? */
4046 break;
4048 return ret;
4051 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4052 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4054 struct cgroup *cgroup;
4055 unsigned long long min_limit, min_memsw_limit, tmp;
4057 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4058 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4059 cgroup = memcg->css.cgroup;
4060 if (!memcg->use_hierarchy)
4061 goto out;
4063 while (cgroup->parent) {
4064 cgroup = cgroup->parent;
4065 memcg = mem_cgroup_from_cont(cgroup);
4066 if (!memcg->use_hierarchy)
4067 break;
4068 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4069 min_limit = min(min_limit, tmp);
4070 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4071 min_memsw_limit = min(min_memsw_limit, tmp);
4073 out:
4074 *mem_limit = min_limit;
4075 *memsw_limit = min_memsw_limit;
4076 return;
4079 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4081 struct mem_cgroup *mem;
4082 int type, name;
4084 mem = mem_cgroup_from_cont(cont);
4085 type = MEMFILE_TYPE(event);
4086 name = MEMFILE_ATTR(event);
4087 switch (name) {
4088 case RES_MAX_USAGE:
4089 if (type == _MEM)
4090 res_counter_reset_max(&mem->res);
4091 else
4092 res_counter_reset_max(&mem->memsw);
4093 break;
4094 case RES_FAILCNT:
4095 if (type == _MEM)
4096 res_counter_reset_failcnt(&mem->res);
4097 else
4098 res_counter_reset_failcnt(&mem->memsw);
4099 break;
4102 return 0;
4105 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4106 struct cftype *cft)
4108 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4111 #ifdef CONFIG_MMU
4112 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4113 struct cftype *cft, u64 val)
4115 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4117 if (val >= (1 << NR_MOVE_TYPE))
4118 return -EINVAL;
4120 * We check this value several times in both in can_attach() and
4121 * attach(), so we need cgroup lock to prevent this value from being
4122 * inconsistent.
4124 cgroup_lock();
4125 mem->move_charge_at_immigrate = val;
4126 cgroup_unlock();
4128 return 0;
4130 #else
4131 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4132 struct cftype *cft, u64 val)
4134 return -ENOSYS;
4136 #endif
4139 /* For read statistics */
4140 enum {
4141 MCS_CACHE,
4142 MCS_RSS,
4143 MCS_FILE_MAPPED,
4144 MCS_PGPGIN,
4145 MCS_PGPGOUT,
4146 MCS_SWAP,
4147 MCS_PGFAULT,
4148 MCS_PGMAJFAULT,
4149 MCS_INACTIVE_ANON,
4150 MCS_ACTIVE_ANON,
4151 MCS_INACTIVE_FILE,
4152 MCS_ACTIVE_FILE,
4153 MCS_UNEVICTABLE,
4154 NR_MCS_STAT,
4157 struct mcs_total_stat {
4158 s64 stat[NR_MCS_STAT];
4161 struct {
4162 char *local_name;
4163 char *total_name;
4164 } memcg_stat_strings[NR_MCS_STAT] = {
4165 {"cache", "total_cache"},
4166 {"rss", "total_rss"},
4167 {"mapped_file", "total_mapped_file"},
4168 {"pgpgin", "total_pgpgin"},
4169 {"pgpgout", "total_pgpgout"},
4170 {"swap", "total_swap"},
4171 {"pgfault", "total_pgfault"},
4172 {"pgmajfault", "total_pgmajfault"},
4173 {"inactive_anon", "total_inactive_anon"},
4174 {"active_anon", "total_active_anon"},
4175 {"inactive_file", "total_inactive_file"},
4176 {"active_file", "total_active_file"},
4177 {"unevictable", "total_unevictable"}
4181 static void
4182 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4184 s64 val;
4186 /* per cpu stat */
4187 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4188 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4189 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4190 s->stat[MCS_RSS] += val * PAGE_SIZE;
4191 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4192 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4193 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4194 s->stat[MCS_PGPGIN] += val;
4195 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4196 s->stat[MCS_PGPGOUT] += val;
4197 if (do_swap_account) {
4198 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4199 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4201 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4202 s->stat[MCS_PGFAULT] += val;
4203 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4204 s->stat[MCS_PGMAJFAULT] += val;
4206 /* per zone stat */
4207 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4208 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4209 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4210 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4211 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4212 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4213 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4214 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4215 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4216 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4219 static void
4220 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4222 struct mem_cgroup *iter;
4224 for_each_mem_cgroup_tree(iter, mem)
4225 mem_cgroup_get_local_stat(iter, s);
4228 #ifdef CONFIG_NUMA
4229 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4231 int nid;
4232 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4233 unsigned long node_nr;
4234 struct cgroup *cont = m->private;
4235 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4237 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4238 seq_printf(m, "total=%lu", total_nr);
4239 for_each_node_state(nid, N_HIGH_MEMORY) {
4240 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4241 seq_printf(m, " N%d=%lu", nid, node_nr);
4243 seq_putc(m, '\n');
4245 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4246 seq_printf(m, "file=%lu", file_nr);
4247 for_each_node_state(nid, N_HIGH_MEMORY) {
4248 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4249 LRU_ALL_FILE);
4250 seq_printf(m, " N%d=%lu", nid, node_nr);
4252 seq_putc(m, '\n');
4254 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4255 seq_printf(m, "anon=%lu", anon_nr);
4256 for_each_node_state(nid, N_HIGH_MEMORY) {
4257 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4258 LRU_ALL_ANON);
4259 seq_printf(m, " N%d=%lu", nid, node_nr);
4261 seq_putc(m, '\n');
4263 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4264 seq_printf(m, "unevictable=%lu", unevictable_nr);
4265 for_each_node_state(nid, N_HIGH_MEMORY) {
4266 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4267 BIT(LRU_UNEVICTABLE));
4268 seq_printf(m, " N%d=%lu", nid, node_nr);
4270 seq_putc(m, '\n');
4271 return 0;
4273 #endif /* CONFIG_NUMA */
4275 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4276 struct cgroup_map_cb *cb)
4278 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4279 struct mcs_total_stat mystat;
4280 int i;
4282 memset(&mystat, 0, sizeof(mystat));
4283 mem_cgroup_get_local_stat(mem_cont, &mystat);
4286 for (i = 0; i < NR_MCS_STAT; i++) {
4287 if (i == MCS_SWAP && !do_swap_account)
4288 continue;
4289 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4292 /* Hierarchical information */
4294 unsigned long long limit, memsw_limit;
4295 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4296 cb->fill(cb, "hierarchical_memory_limit", limit);
4297 if (do_swap_account)
4298 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4301 memset(&mystat, 0, sizeof(mystat));
4302 mem_cgroup_get_total_stat(mem_cont, &mystat);
4303 for (i = 0; i < NR_MCS_STAT; i++) {
4304 if (i == MCS_SWAP && !do_swap_account)
4305 continue;
4306 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4309 #ifdef CONFIG_DEBUG_VM
4310 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4313 int nid, zid;
4314 struct mem_cgroup_per_zone *mz;
4315 unsigned long recent_rotated[2] = {0, 0};
4316 unsigned long recent_scanned[2] = {0, 0};
4318 for_each_online_node(nid)
4319 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4320 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4322 recent_rotated[0] +=
4323 mz->reclaim_stat.recent_rotated[0];
4324 recent_rotated[1] +=
4325 mz->reclaim_stat.recent_rotated[1];
4326 recent_scanned[0] +=
4327 mz->reclaim_stat.recent_scanned[0];
4328 recent_scanned[1] +=
4329 mz->reclaim_stat.recent_scanned[1];
4331 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4332 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4333 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4334 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4336 #endif
4338 return 0;
4341 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4343 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4345 return mem_cgroup_swappiness(memcg);
4348 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4349 u64 val)
4351 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4352 struct mem_cgroup *parent;
4354 if (val > 100)
4355 return -EINVAL;
4357 if (cgrp->parent == NULL)
4358 return -EINVAL;
4360 parent = mem_cgroup_from_cont(cgrp->parent);
4362 cgroup_lock();
4364 /* If under hierarchy, only empty-root can set this value */
4365 if ((parent->use_hierarchy) ||
4366 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4367 cgroup_unlock();
4368 return -EINVAL;
4371 memcg->swappiness = val;
4373 cgroup_unlock();
4375 return 0;
4378 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4380 struct mem_cgroup_threshold_ary *t;
4381 u64 usage;
4382 int i;
4384 rcu_read_lock();
4385 if (!swap)
4386 t = rcu_dereference(memcg->thresholds.primary);
4387 else
4388 t = rcu_dereference(memcg->memsw_thresholds.primary);
4390 if (!t)
4391 goto unlock;
4393 usage = mem_cgroup_usage(memcg, swap);
4396 * current_threshold points to threshold just below usage.
4397 * If it's not true, a threshold was crossed after last
4398 * call of __mem_cgroup_threshold().
4400 i = t->current_threshold;
4403 * Iterate backward over array of thresholds starting from
4404 * current_threshold and check if a threshold is crossed.
4405 * If none of thresholds below usage is crossed, we read
4406 * only one element of the array here.
4408 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4409 eventfd_signal(t->entries[i].eventfd, 1);
4411 /* i = current_threshold + 1 */
4412 i++;
4415 * Iterate forward over array of thresholds starting from
4416 * current_threshold+1 and check if a threshold is crossed.
4417 * If none of thresholds above usage is crossed, we read
4418 * only one element of the array here.
4420 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4421 eventfd_signal(t->entries[i].eventfd, 1);
4423 /* Update current_threshold */
4424 t->current_threshold = i - 1;
4425 unlock:
4426 rcu_read_unlock();
4429 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4431 while (memcg) {
4432 __mem_cgroup_threshold(memcg, false);
4433 if (do_swap_account)
4434 __mem_cgroup_threshold(memcg, true);
4436 memcg = parent_mem_cgroup(memcg);
4440 static int compare_thresholds(const void *a, const void *b)
4442 const struct mem_cgroup_threshold *_a = a;
4443 const struct mem_cgroup_threshold *_b = b;
4445 return _a->threshold - _b->threshold;
4448 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4450 struct mem_cgroup_eventfd_list *ev;
4452 list_for_each_entry(ev, &mem->oom_notify, list)
4453 eventfd_signal(ev->eventfd, 1);
4454 return 0;
4457 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4459 struct mem_cgroup *iter;
4461 for_each_mem_cgroup_tree(iter, mem)
4462 mem_cgroup_oom_notify_cb(iter);
4465 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4466 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4468 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4469 struct mem_cgroup_thresholds *thresholds;
4470 struct mem_cgroup_threshold_ary *new;
4471 int type = MEMFILE_TYPE(cft->private);
4472 u64 threshold, usage;
4473 int i, size, ret;
4475 ret = res_counter_memparse_write_strategy(args, &threshold);
4476 if (ret)
4477 return ret;
4479 mutex_lock(&memcg->thresholds_lock);
4481 if (type == _MEM)
4482 thresholds = &memcg->thresholds;
4483 else if (type == _MEMSWAP)
4484 thresholds = &memcg->memsw_thresholds;
4485 else
4486 BUG();
4488 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4490 /* Check if a threshold crossed before adding a new one */
4491 if (thresholds->primary)
4492 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4494 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4496 /* Allocate memory for new array of thresholds */
4497 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4498 GFP_KERNEL);
4499 if (!new) {
4500 ret = -ENOMEM;
4501 goto unlock;
4503 new->size = size;
4505 /* Copy thresholds (if any) to new array */
4506 if (thresholds->primary) {
4507 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4508 sizeof(struct mem_cgroup_threshold));
4511 /* Add new threshold */
4512 new->entries[size - 1].eventfd = eventfd;
4513 new->entries[size - 1].threshold = threshold;
4515 /* Sort thresholds. Registering of new threshold isn't time-critical */
4516 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4517 compare_thresholds, NULL);
4519 /* Find current threshold */
4520 new->current_threshold = -1;
4521 for (i = 0; i < size; i++) {
4522 if (new->entries[i].threshold < usage) {
4524 * new->current_threshold will not be used until
4525 * rcu_assign_pointer(), so it's safe to increment
4526 * it here.
4528 ++new->current_threshold;
4532 /* Free old spare buffer and save old primary buffer as spare */
4533 kfree(thresholds->spare);
4534 thresholds->spare = thresholds->primary;
4536 rcu_assign_pointer(thresholds->primary, new);
4538 /* To be sure that nobody uses thresholds */
4539 synchronize_rcu();
4541 unlock:
4542 mutex_unlock(&memcg->thresholds_lock);
4544 return ret;
4547 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4548 struct cftype *cft, struct eventfd_ctx *eventfd)
4550 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4551 struct mem_cgroup_thresholds *thresholds;
4552 struct mem_cgroup_threshold_ary *new;
4553 int type = MEMFILE_TYPE(cft->private);
4554 u64 usage;
4555 int i, j, size;
4557 mutex_lock(&memcg->thresholds_lock);
4558 if (type == _MEM)
4559 thresholds = &memcg->thresholds;
4560 else if (type == _MEMSWAP)
4561 thresholds = &memcg->memsw_thresholds;
4562 else
4563 BUG();
4566 * Something went wrong if we trying to unregister a threshold
4567 * if we don't have thresholds
4569 BUG_ON(!thresholds);
4571 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4573 /* Check if a threshold crossed before removing */
4574 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4576 /* Calculate new number of threshold */
4577 size = 0;
4578 for (i = 0; i < thresholds->primary->size; i++) {
4579 if (thresholds->primary->entries[i].eventfd != eventfd)
4580 size++;
4583 new = thresholds->spare;
4585 /* Set thresholds array to NULL if we don't have thresholds */
4586 if (!size) {
4587 kfree(new);
4588 new = NULL;
4589 goto swap_buffers;
4592 new->size = size;
4594 /* Copy thresholds and find current threshold */
4595 new->current_threshold = -1;
4596 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4597 if (thresholds->primary->entries[i].eventfd == eventfd)
4598 continue;
4600 new->entries[j] = thresholds->primary->entries[i];
4601 if (new->entries[j].threshold < usage) {
4603 * new->current_threshold will not be used
4604 * until rcu_assign_pointer(), so it's safe to increment
4605 * it here.
4607 ++new->current_threshold;
4609 j++;
4612 swap_buffers:
4613 /* Swap primary and spare array */
4614 thresholds->spare = thresholds->primary;
4615 rcu_assign_pointer(thresholds->primary, new);
4617 /* To be sure that nobody uses thresholds */
4618 synchronize_rcu();
4620 mutex_unlock(&memcg->thresholds_lock);
4623 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4624 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4626 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4627 struct mem_cgroup_eventfd_list *event;
4628 int type = MEMFILE_TYPE(cft->private);
4630 BUG_ON(type != _OOM_TYPE);
4631 event = kmalloc(sizeof(*event), GFP_KERNEL);
4632 if (!event)
4633 return -ENOMEM;
4635 spin_lock(&memcg_oom_lock);
4637 event->eventfd = eventfd;
4638 list_add(&event->list, &memcg->oom_notify);
4640 /* already in OOM ? */
4641 if (atomic_read(&memcg->under_oom))
4642 eventfd_signal(eventfd, 1);
4643 spin_unlock(&memcg_oom_lock);
4645 return 0;
4648 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4649 struct cftype *cft, struct eventfd_ctx *eventfd)
4651 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4652 struct mem_cgroup_eventfd_list *ev, *tmp;
4653 int type = MEMFILE_TYPE(cft->private);
4655 BUG_ON(type != _OOM_TYPE);
4657 spin_lock(&memcg_oom_lock);
4659 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4660 if (ev->eventfd == eventfd) {
4661 list_del(&ev->list);
4662 kfree(ev);
4666 spin_unlock(&memcg_oom_lock);
4669 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4670 struct cftype *cft, struct cgroup_map_cb *cb)
4672 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4674 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4676 if (atomic_read(&mem->under_oom))
4677 cb->fill(cb, "under_oom", 1);
4678 else
4679 cb->fill(cb, "under_oom", 0);
4680 return 0;
4683 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4684 struct cftype *cft, u64 val)
4686 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4687 struct mem_cgroup *parent;
4689 /* cannot set to root cgroup and only 0 and 1 are allowed */
4690 if (!cgrp->parent || !((val == 0) || (val == 1)))
4691 return -EINVAL;
4693 parent = mem_cgroup_from_cont(cgrp->parent);
4695 cgroup_lock();
4696 /* oom-kill-disable is a flag for subhierarchy. */
4697 if ((parent->use_hierarchy) ||
4698 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4699 cgroup_unlock();
4700 return -EINVAL;
4702 mem->oom_kill_disable = val;
4703 if (!val)
4704 memcg_oom_recover(mem);
4705 cgroup_unlock();
4706 return 0;
4709 #ifdef CONFIG_NUMA
4710 static const struct file_operations mem_control_numa_stat_file_operations = {
4711 .read = seq_read,
4712 .llseek = seq_lseek,
4713 .release = single_release,
4716 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4718 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4720 file->f_op = &mem_control_numa_stat_file_operations;
4721 return single_open(file, mem_control_numa_stat_show, cont);
4723 #endif /* CONFIG_NUMA */
4725 static int mem_cgroup_vmscan_stat_read(struct cgroup *cgrp,
4726 struct cftype *cft,
4727 struct cgroup_map_cb *cb)
4729 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4730 char string[64];
4731 int i;
4733 for (i = 0; i < NR_SCANSTATS; i++) {
4734 strcpy(string, scanstat_string[i]);
4735 strcat(string, SCANSTAT_WORD_LIMIT);
4736 cb->fill(cb, string, mem->scanstat.stats[SCAN_BY_LIMIT][i]);
4739 for (i = 0; i < NR_SCANSTATS; i++) {
4740 strcpy(string, scanstat_string[i]);
4741 strcat(string, SCANSTAT_WORD_SYSTEM);
4742 cb->fill(cb, string, mem->scanstat.stats[SCAN_BY_SYSTEM][i]);
4745 for (i = 0; i < NR_SCANSTATS; i++) {
4746 strcpy(string, scanstat_string[i]);
4747 strcat(string, SCANSTAT_WORD_LIMIT);
4748 strcat(string, SCANSTAT_WORD_HIERARCHY);
4749 cb->fill(cb, string, mem->scanstat.rootstats[SCAN_BY_LIMIT][i]);
4751 for (i = 0; i < NR_SCANSTATS; i++) {
4752 strcpy(string, scanstat_string[i]);
4753 strcat(string, SCANSTAT_WORD_SYSTEM);
4754 strcat(string, SCANSTAT_WORD_HIERARCHY);
4755 cb->fill(cb, string, mem->scanstat.rootstats[SCAN_BY_SYSTEM][i]);
4757 return 0;
4760 static int mem_cgroup_reset_vmscan_stat(struct cgroup *cgrp,
4761 unsigned int event)
4763 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4765 spin_lock(&mem->scanstat.lock);
4766 memset(&mem->scanstat.stats, 0, sizeof(mem->scanstat.stats));
4767 memset(&mem->scanstat.rootstats, 0, sizeof(mem->scanstat.rootstats));
4768 spin_unlock(&mem->scanstat.lock);
4769 return 0;
4773 static struct cftype mem_cgroup_files[] = {
4775 .name = "usage_in_bytes",
4776 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4777 .read_u64 = mem_cgroup_read,
4778 .register_event = mem_cgroup_usage_register_event,
4779 .unregister_event = mem_cgroup_usage_unregister_event,
4782 .name = "max_usage_in_bytes",
4783 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4784 .trigger = mem_cgroup_reset,
4785 .read_u64 = mem_cgroup_read,
4788 .name = "limit_in_bytes",
4789 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4790 .write_string = mem_cgroup_write,
4791 .read_u64 = mem_cgroup_read,
4794 .name = "soft_limit_in_bytes",
4795 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4796 .write_string = mem_cgroup_write,
4797 .read_u64 = mem_cgroup_read,
4800 .name = "failcnt",
4801 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4802 .trigger = mem_cgroup_reset,
4803 .read_u64 = mem_cgroup_read,
4806 .name = "stat",
4807 .read_map = mem_control_stat_show,
4810 .name = "force_empty",
4811 .trigger = mem_cgroup_force_empty_write,
4814 .name = "use_hierarchy",
4815 .write_u64 = mem_cgroup_hierarchy_write,
4816 .read_u64 = mem_cgroup_hierarchy_read,
4819 .name = "swappiness",
4820 .read_u64 = mem_cgroup_swappiness_read,
4821 .write_u64 = mem_cgroup_swappiness_write,
4824 .name = "move_charge_at_immigrate",
4825 .read_u64 = mem_cgroup_move_charge_read,
4826 .write_u64 = mem_cgroup_move_charge_write,
4829 .name = "oom_control",
4830 .read_map = mem_cgroup_oom_control_read,
4831 .write_u64 = mem_cgroup_oom_control_write,
4832 .register_event = mem_cgroup_oom_register_event,
4833 .unregister_event = mem_cgroup_oom_unregister_event,
4834 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4836 #ifdef CONFIG_NUMA
4838 .name = "numa_stat",
4839 .open = mem_control_numa_stat_open,
4840 .mode = S_IRUGO,
4842 #endif
4844 .name = "vmscan_stat",
4845 .read_map = mem_cgroup_vmscan_stat_read,
4846 .trigger = mem_cgroup_reset_vmscan_stat,
4850 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4851 static struct cftype memsw_cgroup_files[] = {
4853 .name = "memsw.usage_in_bytes",
4854 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4855 .read_u64 = mem_cgroup_read,
4856 .register_event = mem_cgroup_usage_register_event,
4857 .unregister_event = mem_cgroup_usage_unregister_event,
4860 .name = "memsw.max_usage_in_bytes",
4861 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4862 .trigger = mem_cgroup_reset,
4863 .read_u64 = mem_cgroup_read,
4866 .name = "memsw.limit_in_bytes",
4867 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4868 .write_string = mem_cgroup_write,
4869 .read_u64 = mem_cgroup_read,
4872 .name = "memsw.failcnt",
4873 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4874 .trigger = mem_cgroup_reset,
4875 .read_u64 = mem_cgroup_read,
4879 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4881 if (!do_swap_account)
4882 return 0;
4883 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4884 ARRAY_SIZE(memsw_cgroup_files));
4886 #else
4887 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4889 return 0;
4891 #endif
4893 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4895 struct mem_cgroup_per_node *pn;
4896 struct mem_cgroup_per_zone *mz;
4897 enum lru_list l;
4898 int zone, tmp = node;
4900 * This routine is called against possible nodes.
4901 * But it's BUG to call kmalloc() against offline node.
4903 * TODO: this routine can waste much memory for nodes which will
4904 * never be onlined. It's better to use memory hotplug callback
4905 * function.
4907 if (!node_state(node, N_NORMAL_MEMORY))
4908 tmp = -1;
4909 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4910 if (!pn)
4911 return 1;
4913 mem->info.nodeinfo[node] = pn;
4914 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4915 mz = &pn->zoneinfo[zone];
4916 for_each_lru(l)
4917 INIT_LIST_HEAD(&mz->lists[l]);
4918 mz->usage_in_excess = 0;
4919 mz->on_tree = false;
4920 mz->mem = mem;
4922 return 0;
4925 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4927 kfree(mem->info.nodeinfo[node]);
4930 static struct mem_cgroup *mem_cgroup_alloc(void)
4932 struct mem_cgroup *mem;
4933 int size = sizeof(struct mem_cgroup);
4935 /* Can be very big if MAX_NUMNODES is very big */
4936 if (size < PAGE_SIZE)
4937 mem = kzalloc(size, GFP_KERNEL);
4938 else
4939 mem = vzalloc(size);
4941 if (!mem)
4942 return NULL;
4944 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4945 if (!mem->stat)
4946 goto out_free;
4947 spin_lock_init(&mem->pcp_counter_lock);
4948 return mem;
4950 out_free:
4951 if (size < PAGE_SIZE)
4952 kfree(mem);
4953 else
4954 vfree(mem);
4955 return NULL;
4959 * At destroying mem_cgroup, references from swap_cgroup can remain.
4960 * (scanning all at force_empty is too costly...)
4962 * Instead of clearing all references at force_empty, we remember
4963 * the number of reference from swap_cgroup and free mem_cgroup when
4964 * it goes down to 0.
4966 * Removal of cgroup itself succeeds regardless of refs from swap.
4969 static void __mem_cgroup_free(struct mem_cgroup *mem)
4971 int node;
4973 mem_cgroup_remove_from_trees(mem);
4974 free_css_id(&mem_cgroup_subsys, &mem->css);
4976 for_each_node_state(node, N_POSSIBLE)
4977 free_mem_cgroup_per_zone_info(mem, node);
4979 free_percpu(mem->stat);
4980 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4981 kfree(mem);
4982 else
4983 vfree(mem);
4986 static void mem_cgroup_get(struct mem_cgroup *mem)
4988 atomic_inc(&mem->refcnt);
4991 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4993 if (atomic_sub_and_test(count, &mem->refcnt)) {
4994 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4995 __mem_cgroup_free(mem);
4996 if (parent)
4997 mem_cgroup_put(parent);
5001 static void mem_cgroup_put(struct mem_cgroup *mem)
5003 __mem_cgroup_put(mem, 1);
5007 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5009 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
5011 if (!mem->res.parent)
5012 return NULL;
5013 return mem_cgroup_from_res_counter(mem->res.parent, res);
5016 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5017 static void __init enable_swap_cgroup(void)
5019 if (!mem_cgroup_disabled() && really_do_swap_account)
5020 do_swap_account = 1;
5022 #else
5023 static void __init enable_swap_cgroup(void)
5026 #endif
5028 static int mem_cgroup_soft_limit_tree_init(void)
5030 struct mem_cgroup_tree_per_node *rtpn;
5031 struct mem_cgroup_tree_per_zone *rtpz;
5032 int tmp, node, zone;
5034 for_each_node_state(node, N_POSSIBLE) {
5035 tmp = node;
5036 if (!node_state(node, N_NORMAL_MEMORY))
5037 tmp = -1;
5038 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5039 if (!rtpn)
5040 return 1;
5042 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5044 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5045 rtpz = &rtpn->rb_tree_per_zone[zone];
5046 rtpz->rb_root = RB_ROOT;
5047 spin_lock_init(&rtpz->lock);
5050 return 0;
5053 static struct cgroup_subsys_state * __ref
5054 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5056 struct mem_cgroup *mem, *parent;
5057 long error = -ENOMEM;
5058 int node;
5060 mem = mem_cgroup_alloc();
5061 if (!mem)
5062 return ERR_PTR(error);
5064 for_each_node_state(node, N_POSSIBLE)
5065 if (alloc_mem_cgroup_per_zone_info(mem, node))
5066 goto free_out;
5068 /* root ? */
5069 if (cont->parent == NULL) {
5070 int cpu;
5071 enable_swap_cgroup();
5072 parent = NULL;
5073 root_mem_cgroup = mem;
5074 if (mem_cgroup_soft_limit_tree_init())
5075 goto free_out;
5076 for_each_possible_cpu(cpu) {
5077 struct memcg_stock_pcp *stock =
5078 &per_cpu(memcg_stock, cpu);
5079 INIT_WORK(&stock->work, drain_local_stock);
5081 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5082 } else {
5083 parent = mem_cgroup_from_cont(cont->parent);
5084 mem->use_hierarchy = parent->use_hierarchy;
5085 mem->oom_kill_disable = parent->oom_kill_disable;
5088 if (parent && parent->use_hierarchy) {
5089 res_counter_init(&mem->res, &parent->res);
5090 res_counter_init(&mem->memsw, &parent->memsw);
5092 * We increment refcnt of the parent to ensure that we can
5093 * safely access it on res_counter_charge/uncharge.
5094 * This refcnt will be decremented when freeing this
5095 * mem_cgroup(see mem_cgroup_put).
5097 mem_cgroup_get(parent);
5098 } else {
5099 res_counter_init(&mem->res, NULL);
5100 res_counter_init(&mem->memsw, NULL);
5102 mem->last_scanned_child = 0;
5103 mem->last_scanned_node = MAX_NUMNODES;
5104 INIT_LIST_HEAD(&mem->oom_notify);
5106 if (parent)
5107 mem->swappiness = mem_cgroup_swappiness(parent);
5108 atomic_set(&mem->refcnt, 1);
5109 mem->move_charge_at_immigrate = 0;
5110 mutex_init(&mem->thresholds_lock);
5111 spin_lock_init(&mem->scanstat.lock);
5112 return &mem->css;
5113 free_out:
5114 __mem_cgroup_free(mem);
5115 root_mem_cgroup = NULL;
5116 return ERR_PTR(error);
5119 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5120 struct cgroup *cont)
5122 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5124 return mem_cgroup_force_empty(mem, false);
5127 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5128 struct cgroup *cont)
5130 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5132 mem_cgroup_put(mem);
5135 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5136 struct cgroup *cont)
5138 int ret;
5140 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5141 ARRAY_SIZE(mem_cgroup_files));
5143 if (!ret)
5144 ret = register_memsw_files(cont, ss);
5145 return ret;
5148 #ifdef CONFIG_MMU
5149 /* Handlers for move charge at task migration. */
5150 #define PRECHARGE_COUNT_AT_ONCE 256
5151 static int mem_cgroup_do_precharge(unsigned long count)
5153 int ret = 0;
5154 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5155 struct mem_cgroup *mem = mc.to;
5157 if (mem_cgroup_is_root(mem)) {
5158 mc.precharge += count;
5159 /* we don't need css_get for root */
5160 return ret;
5162 /* try to charge at once */
5163 if (count > 1) {
5164 struct res_counter *dummy;
5166 * "mem" cannot be under rmdir() because we've already checked
5167 * by cgroup_lock_live_cgroup() that it is not removed and we
5168 * are still under the same cgroup_mutex. So we can postpone
5169 * css_get().
5171 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5172 goto one_by_one;
5173 if (do_swap_account && res_counter_charge(&mem->memsw,
5174 PAGE_SIZE * count, &dummy)) {
5175 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5176 goto one_by_one;
5178 mc.precharge += count;
5179 return ret;
5181 one_by_one:
5182 /* fall back to one by one charge */
5183 while (count--) {
5184 if (signal_pending(current)) {
5185 ret = -EINTR;
5186 break;
5188 if (!batch_count--) {
5189 batch_count = PRECHARGE_COUNT_AT_ONCE;
5190 cond_resched();
5192 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5193 if (ret || !mem)
5194 /* mem_cgroup_clear_mc() will do uncharge later */
5195 return -ENOMEM;
5196 mc.precharge++;
5198 return ret;
5202 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5203 * @vma: the vma the pte to be checked belongs
5204 * @addr: the address corresponding to the pte to be checked
5205 * @ptent: the pte to be checked
5206 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5208 * Returns
5209 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5210 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5211 * move charge. if @target is not NULL, the page is stored in target->page
5212 * with extra refcnt got(Callers should handle it).
5213 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5214 * target for charge migration. if @target is not NULL, the entry is stored
5215 * in target->ent.
5217 * Called with pte lock held.
5219 union mc_target {
5220 struct page *page;
5221 swp_entry_t ent;
5224 enum mc_target_type {
5225 MC_TARGET_NONE, /* not used */
5226 MC_TARGET_PAGE,
5227 MC_TARGET_SWAP,
5230 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5231 unsigned long addr, pte_t ptent)
5233 struct page *page = vm_normal_page(vma, addr, ptent);
5235 if (!page || !page_mapped(page))
5236 return NULL;
5237 if (PageAnon(page)) {
5238 /* we don't move shared anon */
5239 if (!move_anon() || page_mapcount(page) > 2)
5240 return NULL;
5241 } else if (!move_file())
5242 /* we ignore mapcount for file pages */
5243 return NULL;
5244 if (!get_page_unless_zero(page))
5245 return NULL;
5247 return page;
5250 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5251 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5253 int usage_count;
5254 struct page *page = NULL;
5255 swp_entry_t ent = pte_to_swp_entry(ptent);
5257 if (!move_anon() || non_swap_entry(ent))
5258 return NULL;
5259 usage_count = mem_cgroup_count_swap_user(ent, &page);
5260 if (usage_count > 1) { /* we don't move shared anon */
5261 if (page)
5262 put_page(page);
5263 return NULL;
5265 if (do_swap_account)
5266 entry->val = ent.val;
5268 return page;
5271 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5272 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5274 struct page *page = NULL;
5275 struct inode *inode;
5276 struct address_space *mapping;
5277 pgoff_t pgoff;
5279 if (!vma->vm_file) /* anonymous vma */
5280 return NULL;
5281 if (!move_file())
5282 return NULL;
5284 inode = vma->vm_file->f_path.dentry->d_inode;
5285 mapping = vma->vm_file->f_mapping;
5286 if (pte_none(ptent))
5287 pgoff = linear_page_index(vma, addr);
5288 else /* pte_file(ptent) is true */
5289 pgoff = pte_to_pgoff(ptent);
5291 /* page is moved even if it's not RSS of this task(page-faulted). */
5292 page = find_get_page(mapping, pgoff);
5294 #ifdef CONFIG_SWAP
5295 /* shmem/tmpfs may report page out on swap: account for that too. */
5296 if (radix_tree_exceptional_entry(page)) {
5297 swp_entry_t swap = radix_to_swp_entry(page);
5298 if (do_swap_account)
5299 *entry = swap;
5300 page = find_get_page(&swapper_space, swap.val);
5302 #endif
5303 return page;
5306 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5307 unsigned long addr, pte_t ptent, union mc_target *target)
5309 struct page *page = NULL;
5310 struct page_cgroup *pc;
5311 int ret = 0;
5312 swp_entry_t ent = { .val = 0 };
5314 if (pte_present(ptent))
5315 page = mc_handle_present_pte(vma, addr, ptent);
5316 else if (is_swap_pte(ptent))
5317 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5318 else if (pte_none(ptent) || pte_file(ptent))
5319 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5321 if (!page && !ent.val)
5322 return 0;
5323 if (page) {
5324 pc = lookup_page_cgroup(page);
5326 * Do only loose check w/o page_cgroup lock.
5327 * mem_cgroup_move_account() checks the pc is valid or not under
5328 * the lock.
5330 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5331 ret = MC_TARGET_PAGE;
5332 if (target)
5333 target->page = page;
5335 if (!ret || !target)
5336 put_page(page);
5338 /* There is a swap entry and a page doesn't exist or isn't charged */
5339 if (ent.val && !ret &&
5340 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5341 ret = MC_TARGET_SWAP;
5342 if (target)
5343 target->ent = ent;
5345 return ret;
5348 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5349 unsigned long addr, unsigned long end,
5350 struct mm_walk *walk)
5352 struct vm_area_struct *vma = walk->private;
5353 pte_t *pte;
5354 spinlock_t *ptl;
5356 split_huge_page_pmd(walk->mm, pmd);
5358 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5359 for (; addr != end; pte++, addr += PAGE_SIZE)
5360 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5361 mc.precharge++; /* increment precharge temporarily */
5362 pte_unmap_unlock(pte - 1, ptl);
5363 cond_resched();
5365 return 0;
5368 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5370 unsigned long precharge;
5371 struct vm_area_struct *vma;
5373 down_read(&mm->mmap_sem);
5374 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5375 struct mm_walk mem_cgroup_count_precharge_walk = {
5376 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5377 .mm = mm,
5378 .private = vma,
5380 if (is_vm_hugetlb_page(vma))
5381 continue;
5382 walk_page_range(vma->vm_start, vma->vm_end,
5383 &mem_cgroup_count_precharge_walk);
5385 up_read(&mm->mmap_sem);
5387 precharge = mc.precharge;
5388 mc.precharge = 0;
5390 return precharge;
5393 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5395 unsigned long precharge = mem_cgroup_count_precharge(mm);
5397 VM_BUG_ON(mc.moving_task);
5398 mc.moving_task = current;
5399 return mem_cgroup_do_precharge(precharge);
5402 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5403 static void __mem_cgroup_clear_mc(void)
5405 struct mem_cgroup *from = mc.from;
5406 struct mem_cgroup *to = mc.to;
5408 /* we must uncharge all the leftover precharges from mc.to */
5409 if (mc.precharge) {
5410 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5411 mc.precharge = 0;
5414 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5415 * we must uncharge here.
5417 if (mc.moved_charge) {
5418 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5419 mc.moved_charge = 0;
5421 /* we must fixup refcnts and charges */
5422 if (mc.moved_swap) {
5423 /* uncharge swap account from the old cgroup */
5424 if (!mem_cgroup_is_root(mc.from))
5425 res_counter_uncharge(&mc.from->memsw,
5426 PAGE_SIZE * mc.moved_swap);
5427 __mem_cgroup_put(mc.from, mc.moved_swap);
5429 if (!mem_cgroup_is_root(mc.to)) {
5431 * we charged both to->res and to->memsw, so we should
5432 * uncharge to->res.
5434 res_counter_uncharge(&mc.to->res,
5435 PAGE_SIZE * mc.moved_swap);
5437 /* we've already done mem_cgroup_get(mc.to) */
5438 mc.moved_swap = 0;
5440 memcg_oom_recover(from);
5441 memcg_oom_recover(to);
5442 wake_up_all(&mc.waitq);
5445 static void mem_cgroup_clear_mc(void)
5447 struct mem_cgroup *from = mc.from;
5450 * we must clear moving_task before waking up waiters at the end of
5451 * task migration.
5453 mc.moving_task = NULL;
5454 __mem_cgroup_clear_mc();
5455 spin_lock(&mc.lock);
5456 mc.from = NULL;
5457 mc.to = NULL;
5458 spin_unlock(&mc.lock);
5459 mem_cgroup_end_move(from);
5462 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5463 struct cgroup *cgroup,
5464 struct task_struct *p)
5466 int ret = 0;
5467 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5469 if (mem->move_charge_at_immigrate) {
5470 struct mm_struct *mm;
5471 struct mem_cgroup *from = mem_cgroup_from_task(p);
5473 VM_BUG_ON(from == mem);
5475 mm = get_task_mm(p);
5476 if (!mm)
5477 return 0;
5478 /* We move charges only when we move a owner of the mm */
5479 if (mm->owner == p) {
5480 VM_BUG_ON(mc.from);
5481 VM_BUG_ON(mc.to);
5482 VM_BUG_ON(mc.precharge);
5483 VM_BUG_ON(mc.moved_charge);
5484 VM_BUG_ON(mc.moved_swap);
5485 mem_cgroup_start_move(from);
5486 spin_lock(&mc.lock);
5487 mc.from = from;
5488 mc.to = mem;
5489 spin_unlock(&mc.lock);
5490 /* We set mc.moving_task later */
5492 ret = mem_cgroup_precharge_mc(mm);
5493 if (ret)
5494 mem_cgroup_clear_mc();
5496 mmput(mm);
5498 return ret;
5501 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5502 struct cgroup *cgroup,
5503 struct task_struct *p)
5505 mem_cgroup_clear_mc();
5508 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5509 unsigned long addr, unsigned long end,
5510 struct mm_walk *walk)
5512 int ret = 0;
5513 struct vm_area_struct *vma = walk->private;
5514 pte_t *pte;
5515 spinlock_t *ptl;
5517 split_huge_page_pmd(walk->mm, pmd);
5518 retry:
5519 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5520 for (; addr != end; addr += PAGE_SIZE) {
5521 pte_t ptent = *(pte++);
5522 union mc_target target;
5523 int type;
5524 struct page *page;
5525 struct page_cgroup *pc;
5526 swp_entry_t ent;
5528 if (!mc.precharge)
5529 break;
5531 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5532 switch (type) {
5533 case MC_TARGET_PAGE:
5534 page = target.page;
5535 if (isolate_lru_page(page))
5536 goto put;
5537 pc = lookup_page_cgroup(page);
5538 if (!mem_cgroup_move_account(page, 1, pc,
5539 mc.from, mc.to, false)) {
5540 mc.precharge--;
5541 /* we uncharge from mc.from later. */
5542 mc.moved_charge++;
5544 putback_lru_page(page);
5545 put: /* is_target_pte_for_mc() gets the page */
5546 put_page(page);
5547 break;
5548 case MC_TARGET_SWAP:
5549 ent = target.ent;
5550 if (!mem_cgroup_move_swap_account(ent,
5551 mc.from, mc.to, false)) {
5552 mc.precharge--;
5553 /* we fixup refcnts and charges later. */
5554 mc.moved_swap++;
5556 break;
5557 default:
5558 break;
5561 pte_unmap_unlock(pte - 1, ptl);
5562 cond_resched();
5564 if (addr != end) {
5566 * We have consumed all precharges we got in can_attach().
5567 * We try charge one by one, but don't do any additional
5568 * charges to mc.to if we have failed in charge once in attach()
5569 * phase.
5571 ret = mem_cgroup_do_precharge(1);
5572 if (!ret)
5573 goto retry;
5576 return ret;
5579 static void mem_cgroup_move_charge(struct mm_struct *mm)
5581 struct vm_area_struct *vma;
5583 lru_add_drain_all();
5584 retry:
5585 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5587 * Someone who are holding the mmap_sem might be waiting in
5588 * waitq. So we cancel all extra charges, wake up all waiters,
5589 * and retry. Because we cancel precharges, we might not be able
5590 * to move enough charges, but moving charge is a best-effort
5591 * feature anyway, so it wouldn't be a big problem.
5593 __mem_cgroup_clear_mc();
5594 cond_resched();
5595 goto retry;
5597 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5598 int ret;
5599 struct mm_walk mem_cgroup_move_charge_walk = {
5600 .pmd_entry = mem_cgroup_move_charge_pte_range,
5601 .mm = mm,
5602 .private = vma,
5604 if (is_vm_hugetlb_page(vma))
5605 continue;
5606 ret = walk_page_range(vma->vm_start, vma->vm_end,
5607 &mem_cgroup_move_charge_walk);
5608 if (ret)
5610 * means we have consumed all precharges and failed in
5611 * doing additional charge. Just abandon here.
5613 break;
5615 up_read(&mm->mmap_sem);
5618 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5619 struct cgroup *cont,
5620 struct cgroup *old_cont,
5621 struct task_struct *p)
5623 struct mm_struct *mm = get_task_mm(p);
5625 if (mm) {
5626 if (mc.to)
5627 mem_cgroup_move_charge(mm);
5628 put_swap_token(mm);
5629 mmput(mm);
5631 if (mc.to)
5632 mem_cgroup_clear_mc();
5634 #else /* !CONFIG_MMU */
5635 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5636 struct cgroup *cgroup,
5637 struct task_struct *p)
5639 return 0;
5641 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5642 struct cgroup *cgroup,
5643 struct task_struct *p)
5646 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5647 struct cgroup *cont,
5648 struct cgroup *old_cont,
5649 struct task_struct *p)
5652 #endif
5654 struct cgroup_subsys mem_cgroup_subsys = {
5655 .name = "memory",
5656 .subsys_id = mem_cgroup_subsys_id,
5657 .create = mem_cgroup_create,
5658 .pre_destroy = mem_cgroup_pre_destroy,
5659 .destroy = mem_cgroup_destroy,
5660 .populate = mem_cgroup_populate,
5661 .can_attach = mem_cgroup_can_attach,
5662 .cancel_attach = mem_cgroup_cancel_attach,
5663 .attach = mem_cgroup_move_task,
5664 .early_init = 0,
5665 .use_id = 1,
5668 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5669 static int __init enable_swap_account(char *s)
5671 /* consider enabled if no parameter or 1 is given */
5672 if (!strcmp(s, "1"))
5673 really_do_swap_account = 1;
5674 else if (!strcmp(s, "0"))
5675 really_do_swap_account = 0;
5676 return 1;
5678 __setup("swapaccount=", enable_swap_account);
5680 #endif