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>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
69 #include <net/tcp_memcontrol.h>
72 #include <asm/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
77 EXPORT_SYMBOL(memory_cgrp_subsys);
79 struct mem_cgroup *root_mem_cgroup __read_mostly;
81 #define MEM_CGROUP_RECLAIM_RETRIES 5
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket;
86 /* Whether the swap controller is active */
87 #ifdef CONFIG_MEMCG_SWAP
88 int do_swap_account __read_mostly;
90 #define do_swap_account 0
93 /* Whether legacy memory+swap accounting is active */
94 static bool do_memsw_account(void)
96 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
99 static const char * const mem_cgroup_stat_names[] = {
109 static const char * const mem_cgroup_events_names[] = {
116 static const char * const mem_cgroup_lru_names[] = {
124 #define THRESHOLDS_EVENTS_TARGET 128
125 #define SOFTLIMIT_EVENTS_TARGET 1024
126 #define NUMAINFO_EVENTS_TARGET 1024
129 * Cgroups above their limits are maintained in a RB-Tree, independent of
130 * their hierarchy representation
133 struct mem_cgroup_tree_per_zone {
134 struct rb_root rb_root;
138 struct mem_cgroup_tree_per_node {
139 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
142 struct mem_cgroup_tree {
143 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
146 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
149 struct mem_cgroup_eventfd_list {
150 struct list_head list;
151 struct eventfd_ctx *eventfd;
155 * cgroup_event represents events which userspace want to receive.
157 struct mem_cgroup_event {
159 * memcg which the event belongs to.
161 struct mem_cgroup *memcg;
163 * eventfd to signal userspace about the event.
165 struct eventfd_ctx *eventfd;
167 * Each of these stored in a list by the cgroup.
169 struct list_head list;
171 * register_event() callback will be used to add new userspace
172 * waiter for changes related to this event. Use eventfd_signal()
173 * on eventfd to send notification to userspace.
175 int (*register_event)(struct mem_cgroup *memcg,
176 struct eventfd_ctx *eventfd, const char *args);
178 * unregister_event() callback will be called when userspace closes
179 * the eventfd or on cgroup removing. This callback must be set,
180 * if you want provide notification functionality.
182 void (*unregister_event)(struct mem_cgroup *memcg,
183 struct eventfd_ctx *eventfd);
185 * All fields below needed to unregister event when
186 * userspace closes eventfd.
189 wait_queue_head_t *wqh;
191 struct work_struct remove;
194 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
195 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
197 /* Stuffs for move charges at task migration. */
199 * Types of charges to be moved.
201 #define MOVE_ANON 0x1U
202 #define MOVE_FILE 0x2U
203 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
205 /* "mc" and its members are protected by cgroup_mutex */
206 static struct move_charge_struct {
207 spinlock_t lock; /* for from, to */
208 struct mem_cgroup *from;
209 struct mem_cgroup *to;
211 unsigned long precharge;
212 unsigned long moved_charge;
213 unsigned long moved_swap;
214 struct task_struct *moving_task; /* a task moving charges */
215 wait_queue_head_t waitq; /* a waitq for other context */
217 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
218 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
222 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
223 * limit reclaim to prevent infinite loops, if they ever occur.
225 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
226 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
229 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
230 MEM_CGROUP_CHARGE_TYPE_ANON,
231 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
232 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
236 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
251 * The memcg_create_mutex will be held whenever a new cgroup is created.
252 * As a consequence, any change that needs to protect against new child cgroups
253 * appearing has to hold it as well.
255 static DEFINE_MUTEX(memcg_create_mutex);
257 /* Some nice accessors for the vmpressure. */
258 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
261 memcg = root_mem_cgroup;
262 return &memcg->vmpressure;
265 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
267 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
270 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
272 return (memcg == root_mem_cgroup);
276 * We restrict the id in the range of [1, 65535], so it can fit into
279 #define MEM_CGROUP_ID_MAX USHRT_MAX
281 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
283 return memcg->css.id;
287 * A helper function to get mem_cgroup from ID. must be called under
288 * rcu_read_lock(). The caller is responsible for calling
289 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
290 * refcnt from swap can be called against removed memcg.)
292 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
294 struct cgroup_subsys_state *css;
296 css = css_from_id(id, &memory_cgrp_subsys);
297 return mem_cgroup_from_css(css);
300 #ifdef CONFIG_MEMCG_KMEM
302 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
303 * The main reason for not using cgroup id for this:
304 * this works better in sparse environments, where we have a lot of memcgs,
305 * but only a few kmem-limited. Or also, if we have, for instance, 200
306 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
307 * 200 entry array for that.
309 * The current size of the caches array is stored in memcg_nr_cache_ids. It
310 * will double each time we have to increase it.
312 static DEFINE_IDA(memcg_cache_ida);
313 int memcg_nr_cache_ids;
315 /* Protects memcg_nr_cache_ids */
316 static DECLARE_RWSEM(memcg_cache_ids_sem);
318 void memcg_get_cache_ids(void)
320 down_read(&memcg_cache_ids_sem);
323 void memcg_put_cache_ids(void)
325 up_read(&memcg_cache_ids_sem);
329 * MIN_SIZE is different than 1, because we would like to avoid going through
330 * the alloc/free process all the time. In a small machine, 4 kmem-limited
331 * cgroups is a reasonable guess. In the future, it could be a parameter or
332 * tunable, but that is strictly not necessary.
334 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
335 * this constant directly from cgroup, but it is understandable that this is
336 * better kept as an internal representation in cgroup.c. In any case, the
337 * cgrp_id space is not getting any smaller, and we don't have to necessarily
338 * increase ours as well if it increases.
340 #define MEMCG_CACHES_MIN_SIZE 4
341 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
344 * A lot of the calls to the cache allocation functions are expected to be
345 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
346 * conditional to this static branch, we'll have to allow modules that does
347 * kmem_cache_alloc and the such to see this symbol as well
349 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
350 EXPORT_SYMBOL(memcg_kmem_enabled_key);
352 #endif /* CONFIG_MEMCG_KMEM */
354 static struct mem_cgroup_per_zone *
355 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
357 int nid = zone_to_nid(zone);
358 int zid = zone_idx(zone);
360 return &memcg->nodeinfo[nid]->zoneinfo[zid];
364 * mem_cgroup_css_from_page - css of the memcg associated with a page
365 * @page: page of interest
367 * If memcg is bound to the default hierarchy, css of the memcg associated
368 * with @page is returned. The returned css remains associated with @page
369 * until it is released.
371 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
374 * XXX: The above description of behavior on the default hierarchy isn't
375 * strictly true yet as replace_page_cache_page() can modify the
376 * association before @page is released even on the default hierarchy;
377 * however, the current and planned usages don't mix the the two functions
378 * and replace_page_cache_page() will soon be updated to make the invariant
381 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
383 struct mem_cgroup *memcg;
385 memcg = page->mem_cgroup;
387 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
388 memcg = root_mem_cgroup;
394 * page_cgroup_ino - return inode number of the memcg a page is charged to
397 * Look up the closest online ancestor of the memory cgroup @page is charged to
398 * and return its inode number or 0 if @page is not charged to any cgroup. It
399 * is safe to call this function without holding a reference to @page.
401 * Note, this function is inherently racy, because there is nothing to prevent
402 * the cgroup inode from getting torn down and potentially reallocated a moment
403 * after page_cgroup_ino() returns, so it only should be used by callers that
404 * do not care (such as procfs interfaces).
406 ino_t page_cgroup_ino(struct page *page)
408 struct mem_cgroup *memcg;
409 unsigned long ino = 0;
412 memcg = READ_ONCE(page->mem_cgroup);
413 while (memcg && !(memcg->css.flags & CSS_ONLINE))
414 memcg = parent_mem_cgroup(memcg);
416 ino = cgroup_ino(memcg->css.cgroup);
421 static struct mem_cgroup_per_zone *
422 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
424 int nid = page_to_nid(page);
425 int zid = page_zonenum(page);
427 return &memcg->nodeinfo[nid]->zoneinfo[zid];
430 static struct mem_cgroup_tree_per_zone *
431 soft_limit_tree_node_zone(int nid, int zid)
433 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
436 static struct mem_cgroup_tree_per_zone *
437 soft_limit_tree_from_page(struct page *page)
439 int nid = page_to_nid(page);
440 int zid = page_zonenum(page);
442 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
445 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
446 struct mem_cgroup_tree_per_zone *mctz,
447 unsigned long new_usage_in_excess)
449 struct rb_node **p = &mctz->rb_root.rb_node;
450 struct rb_node *parent = NULL;
451 struct mem_cgroup_per_zone *mz_node;
456 mz->usage_in_excess = new_usage_in_excess;
457 if (!mz->usage_in_excess)
461 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
463 if (mz->usage_in_excess < mz_node->usage_in_excess)
466 * We can't avoid mem cgroups that are over their soft
467 * limit by the same amount
469 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
472 rb_link_node(&mz->tree_node, parent, p);
473 rb_insert_color(&mz->tree_node, &mctz->rb_root);
477 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
478 struct mem_cgroup_tree_per_zone *mctz)
482 rb_erase(&mz->tree_node, &mctz->rb_root);
486 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
487 struct mem_cgroup_tree_per_zone *mctz)
491 spin_lock_irqsave(&mctz->lock, flags);
492 __mem_cgroup_remove_exceeded(mz, mctz);
493 spin_unlock_irqrestore(&mctz->lock, flags);
496 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
498 unsigned long nr_pages = page_counter_read(&memcg->memory);
499 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
500 unsigned long excess = 0;
502 if (nr_pages > soft_limit)
503 excess = nr_pages - soft_limit;
508 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
510 unsigned long excess;
511 struct mem_cgroup_per_zone *mz;
512 struct mem_cgroup_tree_per_zone *mctz;
514 mctz = soft_limit_tree_from_page(page);
516 * Necessary to update all ancestors when hierarchy is used.
517 * because their event counter is not touched.
519 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
520 mz = mem_cgroup_page_zoneinfo(memcg, page);
521 excess = soft_limit_excess(memcg);
523 * We have to update the tree if mz is on RB-tree or
524 * mem is over its softlimit.
526 if (excess || mz->on_tree) {
529 spin_lock_irqsave(&mctz->lock, flags);
530 /* if on-tree, remove it */
532 __mem_cgroup_remove_exceeded(mz, mctz);
534 * Insert again. mz->usage_in_excess will be updated.
535 * If excess is 0, no tree ops.
537 __mem_cgroup_insert_exceeded(mz, mctz, excess);
538 spin_unlock_irqrestore(&mctz->lock, flags);
543 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
545 struct mem_cgroup_tree_per_zone *mctz;
546 struct mem_cgroup_per_zone *mz;
550 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
551 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
552 mctz = soft_limit_tree_node_zone(nid, zid);
553 mem_cgroup_remove_exceeded(mz, mctz);
558 static struct mem_cgroup_per_zone *
559 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
561 struct rb_node *rightmost = NULL;
562 struct mem_cgroup_per_zone *mz;
566 rightmost = rb_last(&mctz->rb_root);
568 goto done; /* Nothing to reclaim from */
570 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
572 * Remove the node now but someone else can add it back,
573 * we will to add it back at the end of reclaim to its correct
574 * position in the tree.
576 __mem_cgroup_remove_exceeded(mz, mctz);
577 if (!soft_limit_excess(mz->memcg) ||
578 !css_tryget_online(&mz->memcg->css))
584 static struct mem_cgroup_per_zone *
585 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
587 struct mem_cgroup_per_zone *mz;
589 spin_lock_irq(&mctz->lock);
590 mz = __mem_cgroup_largest_soft_limit_node(mctz);
591 spin_unlock_irq(&mctz->lock);
596 * Return page count for single (non recursive) @memcg.
598 * Implementation Note: reading percpu statistics for memcg.
600 * Both of vmstat[] and percpu_counter has threshold and do periodic
601 * synchronization to implement "quick" read. There are trade-off between
602 * reading cost and precision of value. Then, we may have a chance to implement
603 * a periodic synchronization of counter in memcg's counter.
605 * But this _read() function is used for user interface now. The user accounts
606 * memory usage by memory cgroup and he _always_ requires exact value because
607 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
608 * have to visit all online cpus and make sum. So, for now, unnecessary
609 * synchronization is not implemented. (just implemented for cpu hotplug)
611 * If there are kernel internal actions which can make use of some not-exact
612 * value, and reading all cpu value can be performance bottleneck in some
613 * common workload, threshold and synchronization as vmstat[] should be
617 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
622 /* Per-cpu values can be negative, use a signed accumulator */
623 for_each_possible_cpu(cpu)
624 val += per_cpu(memcg->stat->count[idx], cpu);
626 * Summing races with updates, so val may be negative. Avoid exposing
627 * transient negative values.
634 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
635 enum mem_cgroup_events_index idx)
637 unsigned long val = 0;
640 for_each_possible_cpu(cpu)
641 val += per_cpu(memcg->stat->events[idx], cpu);
645 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
647 bool compound, int nr_pages)
650 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
651 * counted as CACHE even if it's on ANON LRU.
654 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
657 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
661 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
662 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
666 /* pagein of a big page is an event. So, ignore page size */
668 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
670 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
671 nr_pages = -nr_pages; /* for event */
674 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
677 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
679 unsigned int lru_mask)
681 unsigned long nr = 0;
684 VM_BUG_ON((unsigned)nid >= nr_node_ids);
686 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
687 struct mem_cgroup_per_zone *mz;
691 if (!(BIT(lru) & lru_mask))
693 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
694 nr += mz->lru_size[lru];
700 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
701 unsigned int lru_mask)
703 unsigned long nr = 0;
706 for_each_node_state(nid, N_MEMORY)
707 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
711 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
712 enum mem_cgroup_events_target target)
714 unsigned long val, next;
716 val = __this_cpu_read(memcg->stat->nr_page_events);
717 next = __this_cpu_read(memcg->stat->targets[target]);
718 /* from time_after() in jiffies.h */
719 if ((long)next - (long)val < 0) {
721 case MEM_CGROUP_TARGET_THRESH:
722 next = val + THRESHOLDS_EVENTS_TARGET;
724 case MEM_CGROUP_TARGET_SOFTLIMIT:
725 next = val + SOFTLIMIT_EVENTS_TARGET;
727 case MEM_CGROUP_TARGET_NUMAINFO:
728 next = val + NUMAINFO_EVENTS_TARGET;
733 __this_cpu_write(memcg->stat->targets[target], next);
740 * Check events in order.
743 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
745 /* threshold event is triggered in finer grain than soft limit */
746 if (unlikely(mem_cgroup_event_ratelimit(memcg,
747 MEM_CGROUP_TARGET_THRESH))) {
749 bool do_numainfo __maybe_unused;
751 do_softlimit = mem_cgroup_event_ratelimit(memcg,
752 MEM_CGROUP_TARGET_SOFTLIMIT);
754 do_numainfo = mem_cgroup_event_ratelimit(memcg,
755 MEM_CGROUP_TARGET_NUMAINFO);
757 mem_cgroup_threshold(memcg);
758 if (unlikely(do_softlimit))
759 mem_cgroup_update_tree(memcg, page);
761 if (unlikely(do_numainfo))
762 atomic_inc(&memcg->numainfo_events);
767 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
770 * mm_update_next_owner() may clear mm->owner to NULL
771 * if it races with swapoff, page migration, etc.
772 * So this can be called with p == NULL.
777 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
779 EXPORT_SYMBOL(mem_cgroup_from_task);
781 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
783 struct mem_cgroup *memcg = NULL;
788 * Page cache insertions can happen withou an
789 * actual mm context, e.g. during disk probing
790 * on boot, loopback IO, acct() writes etc.
793 memcg = root_mem_cgroup;
795 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
796 if (unlikely(!memcg))
797 memcg = root_mem_cgroup;
799 } while (!css_tryget_online(&memcg->css));
805 * mem_cgroup_iter - iterate over memory cgroup hierarchy
806 * @root: hierarchy root
807 * @prev: previously returned memcg, NULL on first invocation
808 * @reclaim: cookie for shared reclaim walks, NULL for full walks
810 * Returns references to children of the hierarchy below @root, or
811 * @root itself, or %NULL after a full round-trip.
813 * Caller must pass the return value in @prev on subsequent
814 * invocations for reference counting, or use mem_cgroup_iter_break()
815 * to cancel a hierarchy walk before the round-trip is complete.
817 * Reclaimers can specify a zone and a priority level in @reclaim to
818 * divide up the memcgs in the hierarchy among all concurrent
819 * reclaimers operating on the same zone and priority.
821 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
822 struct mem_cgroup *prev,
823 struct mem_cgroup_reclaim_cookie *reclaim)
825 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
826 struct cgroup_subsys_state *css = NULL;
827 struct mem_cgroup *memcg = NULL;
828 struct mem_cgroup *pos = NULL;
830 if (mem_cgroup_disabled())
834 root = root_mem_cgroup;
836 if (prev && !reclaim)
839 if (!root->use_hierarchy && root != root_mem_cgroup) {
848 struct mem_cgroup_per_zone *mz;
850 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
851 iter = &mz->iter[reclaim->priority];
853 if (prev && reclaim->generation != iter->generation)
857 pos = READ_ONCE(iter->position);
858 if (!pos || css_tryget(&pos->css))
861 * css reference reached zero, so iter->position will
862 * be cleared by ->css_released. However, we should not
863 * rely on this happening soon, because ->css_released
864 * is called from a work queue, and by busy-waiting we
865 * might block it. So we clear iter->position right
868 (void)cmpxchg(&iter->position, pos, NULL);
876 css = css_next_descendant_pre(css, &root->css);
879 * Reclaimers share the hierarchy walk, and a
880 * new one might jump in right at the end of
881 * the hierarchy - make sure they see at least
882 * one group and restart from the beginning.
890 * Verify the css and acquire a reference. The root
891 * is provided by the caller, so we know it's alive
892 * and kicking, and don't take an extra reference.
894 memcg = mem_cgroup_from_css(css);
896 if (css == &root->css)
899 if (css_tryget(css)) {
901 * Make sure the memcg is initialized:
902 * mem_cgroup_css_online() orders the the
903 * initialization against setting the flag.
905 if (smp_load_acquire(&memcg->initialized))
916 * The position could have already been updated by a competing
917 * thread, so check that the value hasn't changed since we read
918 * it to avoid reclaiming from the same cgroup twice.
920 (void)cmpxchg(&iter->position, pos, memcg);
928 reclaim->generation = iter->generation;
934 if (prev && prev != root)
941 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
942 * @root: hierarchy root
943 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
945 void mem_cgroup_iter_break(struct mem_cgroup *root,
946 struct mem_cgroup *prev)
949 root = root_mem_cgroup;
950 if (prev && prev != root)
954 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
956 struct mem_cgroup *memcg = dead_memcg;
957 struct mem_cgroup_reclaim_iter *iter;
958 struct mem_cgroup_per_zone *mz;
962 while ((memcg = parent_mem_cgroup(memcg))) {
964 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
965 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
966 for (i = 0; i <= DEF_PRIORITY; i++) {
968 cmpxchg(&iter->position,
977 * Iteration constructs for visiting all cgroups (under a tree). If
978 * loops are exited prematurely (break), mem_cgroup_iter_break() must
979 * be used for reference counting.
981 #define for_each_mem_cgroup_tree(iter, root) \
982 for (iter = mem_cgroup_iter(root, NULL, NULL); \
984 iter = mem_cgroup_iter(root, iter, NULL))
986 #define for_each_mem_cgroup(iter) \
987 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
989 iter = mem_cgroup_iter(NULL, iter, NULL))
992 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
993 * @zone: zone of the wanted lruvec
994 * @memcg: memcg of the wanted lruvec
996 * Returns the lru list vector holding pages for the given @zone and
997 * @mem. This can be the global zone lruvec, if the memory controller
1000 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1001 struct mem_cgroup *memcg)
1003 struct mem_cgroup_per_zone *mz;
1004 struct lruvec *lruvec;
1006 if (mem_cgroup_disabled()) {
1007 lruvec = &zone->lruvec;
1011 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1012 lruvec = &mz->lruvec;
1015 * Since a node can be onlined after the mem_cgroup was created,
1016 * we have to be prepared to initialize lruvec->zone here;
1017 * and if offlined then reonlined, we need to reinitialize it.
1019 if (unlikely(lruvec->zone != zone))
1020 lruvec->zone = zone;
1025 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1027 * @zone: zone of the page
1029 * This function is only safe when following the LRU page isolation
1030 * and putback protocol: the LRU lock must be held, and the page must
1031 * either be PageLRU() or the caller must have isolated/allocated it.
1033 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1035 struct mem_cgroup_per_zone *mz;
1036 struct mem_cgroup *memcg;
1037 struct lruvec *lruvec;
1039 if (mem_cgroup_disabled()) {
1040 lruvec = &zone->lruvec;
1044 memcg = page->mem_cgroup;
1046 * Swapcache readahead pages are added to the LRU - and
1047 * possibly migrated - before they are charged.
1050 memcg = root_mem_cgroup;
1052 mz = mem_cgroup_page_zoneinfo(memcg, page);
1053 lruvec = &mz->lruvec;
1056 * Since a node can be onlined after the mem_cgroup was created,
1057 * we have to be prepared to initialize lruvec->zone here;
1058 * and if offlined then reonlined, we need to reinitialize it.
1060 if (unlikely(lruvec->zone != zone))
1061 lruvec->zone = zone;
1066 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1067 * @lruvec: mem_cgroup per zone lru vector
1068 * @lru: index of lru list the page is sitting on
1069 * @nr_pages: positive when adding or negative when removing
1071 * This function must be called when a page is added to or removed from an
1074 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1077 struct mem_cgroup_per_zone *mz;
1078 unsigned long *lru_size;
1080 if (mem_cgroup_disabled())
1083 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1084 lru_size = mz->lru_size + lru;
1085 *lru_size += nr_pages;
1086 VM_BUG_ON((long)(*lru_size) < 0);
1089 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1091 struct mem_cgroup *task_memcg;
1092 struct task_struct *p;
1095 p = find_lock_task_mm(task);
1097 task_memcg = get_mem_cgroup_from_mm(p->mm);
1101 * All threads may have already detached their mm's, but the oom
1102 * killer still needs to detect if they have already been oom
1103 * killed to prevent needlessly killing additional tasks.
1106 task_memcg = mem_cgroup_from_task(task);
1107 css_get(&task_memcg->css);
1110 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1111 css_put(&task_memcg->css);
1116 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1117 * @memcg: the memory cgroup
1119 * Returns the maximum amount of memory @mem can be charged with, in
1122 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1124 unsigned long margin = 0;
1125 unsigned long count;
1126 unsigned long limit;
1128 count = page_counter_read(&memcg->memory);
1129 limit = READ_ONCE(memcg->memory.limit);
1131 margin = limit - count;
1133 if (do_memsw_account()) {
1134 count = page_counter_read(&memcg->memsw);
1135 limit = READ_ONCE(memcg->memsw.limit);
1137 margin = min(margin, limit - count);
1144 * A routine for checking "mem" is under move_account() or not.
1146 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1147 * moving cgroups. This is for waiting at high-memory pressure
1150 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1152 struct mem_cgroup *from;
1153 struct mem_cgroup *to;
1156 * Unlike task_move routines, we access mc.to, mc.from not under
1157 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1159 spin_lock(&mc.lock);
1165 ret = mem_cgroup_is_descendant(from, memcg) ||
1166 mem_cgroup_is_descendant(to, memcg);
1168 spin_unlock(&mc.lock);
1172 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1174 if (mc.moving_task && current != mc.moving_task) {
1175 if (mem_cgroup_under_move(memcg)) {
1177 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1178 /* moving charge context might have finished. */
1181 finish_wait(&mc.waitq, &wait);
1188 #define K(x) ((x) << (PAGE_SHIFT-10))
1190 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1191 * @memcg: The memory cgroup that went over limit
1192 * @p: Task that is going to be killed
1194 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1197 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1199 /* oom_info_lock ensures that parallel ooms do not interleave */
1200 static DEFINE_MUTEX(oom_info_lock);
1201 struct mem_cgroup *iter;
1204 mutex_lock(&oom_info_lock);
1208 pr_info("Task in ");
1209 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1210 pr_cont(" killed as a result of limit of ");
1212 pr_info("Memory limit reached of cgroup ");
1215 pr_cont_cgroup_path(memcg->css.cgroup);
1220 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1221 K((u64)page_counter_read(&memcg->memory)),
1222 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1223 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1224 K((u64)page_counter_read(&memcg->memsw)),
1225 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1226 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1227 K((u64)page_counter_read(&memcg->kmem)),
1228 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1230 for_each_mem_cgroup_tree(iter, memcg) {
1231 pr_info("Memory cgroup stats for ");
1232 pr_cont_cgroup_path(iter->css.cgroup);
1235 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1236 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
1238 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1239 K(mem_cgroup_read_stat(iter, i)));
1242 for (i = 0; i < NR_LRU_LISTS; i++)
1243 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1244 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1248 mutex_unlock(&oom_info_lock);
1252 * This function returns the number of memcg under hierarchy tree. Returns
1253 * 1(self count) if no children.
1255 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1258 struct mem_cgroup *iter;
1260 for_each_mem_cgroup_tree(iter, memcg)
1266 * Return the memory (and swap, if configured) limit for a memcg.
1268 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1270 unsigned long limit;
1272 limit = memcg->memory.limit;
1273 if (mem_cgroup_swappiness(memcg)) {
1274 unsigned long memsw_limit;
1276 memsw_limit = memcg->memsw.limit;
1277 limit = min(limit + total_swap_pages, memsw_limit);
1282 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1285 struct oom_control oc = {
1288 .gfp_mask = gfp_mask,
1291 struct mem_cgroup *iter;
1292 unsigned long chosen_points = 0;
1293 unsigned long totalpages;
1294 unsigned int points = 0;
1295 struct task_struct *chosen = NULL;
1297 mutex_lock(&oom_lock);
1300 * If current has a pending SIGKILL or is exiting, then automatically
1301 * select it. The goal is to allow it to allocate so that it may
1302 * quickly exit and free its memory.
1304 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1305 mark_oom_victim(current);
1309 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1310 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1311 for_each_mem_cgroup_tree(iter, memcg) {
1312 struct css_task_iter it;
1313 struct task_struct *task;
1315 css_task_iter_start(&iter->css, &it);
1316 while ((task = css_task_iter_next(&it))) {
1317 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1318 case OOM_SCAN_SELECT:
1320 put_task_struct(chosen);
1322 chosen_points = ULONG_MAX;
1323 get_task_struct(chosen);
1325 case OOM_SCAN_CONTINUE:
1327 case OOM_SCAN_ABORT:
1328 css_task_iter_end(&it);
1329 mem_cgroup_iter_break(memcg, iter);
1331 put_task_struct(chosen);
1336 points = oom_badness(task, memcg, NULL, totalpages);
1337 if (!points || points < chosen_points)
1339 /* Prefer thread group leaders for display purposes */
1340 if (points == chosen_points &&
1341 thread_group_leader(chosen))
1345 put_task_struct(chosen);
1347 chosen_points = points;
1348 get_task_struct(chosen);
1350 css_task_iter_end(&it);
1354 points = chosen_points * 1000 / totalpages;
1355 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1356 "Memory cgroup out of memory");
1359 mutex_unlock(&oom_lock);
1362 #if MAX_NUMNODES > 1
1365 * test_mem_cgroup_node_reclaimable
1366 * @memcg: the target memcg
1367 * @nid: the node ID to be checked.
1368 * @noswap : specify true here if the user wants flle only information.
1370 * This function returns whether the specified memcg contains any
1371 * reclaimable pages on a node. Returns true if there are any reclaimable
1372 * pages in the node.
1374 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1375 int nid, bool noswap)
1377 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1379 if (noswap || !total_swap_pages)
1381 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1388 * Always updating the nodemask is not very good - even if we have an empty
1389 * list or the wrong list here, we can start from some node and traverse all
1390 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1393 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1397 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1398 * pagein/pageout changes since the last update.
1400 if (!atomic_read(&memcg->numainfo_events))
1402 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1405 /* make a nodemask where this memcg uses memory from */
1406 memcg->scan_nodes = node_states[N_MEMORY];
1408 for_each_node_mask(nid, node_states[N_MEMORY]) {
1410 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1411 node_clear(nid, memcg->scan_nodes);
1414 atomic_set(&memcg->numainfo_events, 0);
1415 atomic_set(&memcg->numainfo_updating, 0);
1419 * Selecting a node where we start reclaim from. Because what we need is just
1420 * reducing usage counter, start from anywhere is O,K. Considering
1421 * memory reclaim from current node, there are pros. and cons.
1423 * Freeing memory from current node means freeing memory from a node which
1424 * we'll use or we've used. So, it may make LRU bad. And if several threads
1425 * hit limits, it will see a contention on a node. But freeing from remote
1426 * node means more costs for memory reclaim because of memory latency.
1428 * Now, we use round-robin. Better algorithm is welcomed.
1430 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1434 mem_cgroup_may_update_nodemask(memcg);
1435 node = memcg->last_scanned_node;
1437 node = next_node(node, memcg->scan_nodes);
1438 if (node == MAX_NUMNODES)
1439 node = first_node(memcg->scan_nodes);
1441 * We call this when we hit limit, not when pages are added to LRU.
1442 * No LRU may hold pages because all pages are UNEVICTABLE or
1443 * memcg is too small and all pages are not on LRU. In that case,
1444 * we use curret node.
1446 if (unlikely(node == MAX_NUMNODES))
1447 node = numa_node_id();
1449 memcg->last_scanned_node = node;
1453 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1459 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1462 unsigned long *total_scanned)
1464 struct mem_cgroup *victim = NULL;
1467 unsigned long excess;
1468 unsigned long nr_scanned;
1469 struct mem_cgroup_reclaim_cookie reclaim = {
1474 excess = soft_limit_excess(root_memcg);
1477 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1482 * If we have not been able to reclaim
1483 * anything, it might because there are
1484 * no reclaimable pages under this hierarchy
1489 * We want to do more targeted reclaim.
1490 * excess >> 2 is not to excessive so as to
1491 * reclaim too much, nor too less that we keep
1492 * coming back to reclaim from this cgroup
1494 if (total >= (excess >> 2) ||
1495 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1500 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1502 *total_scanned += nr_scanned;
1503 if (!soft_limit_excess(root_memcg))
1506 mem_cgroup_iter_break(root_memcg, victim);
1510 #ifdef CONFIG_LOCKDEP
1511 static struct lockdep_map memcg_oom_lock_dep_map = {
1512 .name = "memcg_oom_lock",
1516 static DEFINE_SPINLOCK(memcg_oom_lock);
1519 * Check OOM-Killer is already running under our hierarchy.
1520 * If someone is running, return false.
1522 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1524 struct mem_cgroup *iter, *failed = NULL;
1526 spin_lock(&memcg_oom_lock);
1528 for_each_mem_cgroup_tree(iter, memcg) {
1529 if (iter->oom_lock) {
1531 * this subtree of our hierarchy is already locked
1532 * so we cannot give a lock.
1535 mem_cgroup_iter_break(memcg, iter);
1538 iter->oom_lock = true;
1543 * OK, we failed to lock the whole subtree so we have
1544 * to clean up what we set up to the failing subtree
1546 for_each_mem_cgroup_tree(iter, memcg) {
1547 if (iter == failed) {
1548 mem_cgroup_iter_break(memcg, iter);
1551 iter->oom_lock = false;
1554 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1556 spin_unlock(&memcg_oom_lock);
1561 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1563 struct mem_cgroup *iter;
1565 spin_lock(&memcg_oom_lock);
1566 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1567 for_each_mem_cgroup_tree(iter, memcg)
1568 iter->oom_lock = false;
1569 spin_unlock(&memcg_oom_lock);
1572 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1574 struct mem_cgroup *iter;
1576 spin_lock(&memcg_oom_lock);
1577 for_each_mem_cgroup_tree(iter, memcg)
1579 spin_unlock(&memcg_oom_lock);
1582 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1584 struct mem_cgroup *iter;
1587 * When a new child is created while the hierarchy is under oom,
1588 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1590 spin_lock(&memcg_oom_lock);
1591 for_each_mem_cgroup_tree(iter, memcg)
1592 if (iter->under_oom > 0)
1594 spin_unlock(&memcg_oom_lock);
1597 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1599 struct oom_wait_info {
1600 struct mem_cgroup *memcg;
1604 static int memcg_oom_wake_function(wait_queue_t *wait,
1605 unsigned mode, int sync, void *arg)
1607 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1608 struct mem_cgroup *oom_wait_memcg;
1609 struct oom_wait_info *oom_wait_info;
1611 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1612 oom_wait_memcg = oom_wait_info->memcg;
1614 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1615 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1617 return autoremove_wake_function(wait, mode, sync, arg);
1620 static void memcg_oom_recover(struct mem_cgroup *memcg)
1623 * For the following lockless ->under_oom test, the only required
1624 * guarantee is that it must see the state asserted by an OOM when
1625 * this function is called as a result of userland actions
1626 * triggered by the notification of the OOM. This is trivially
1627 * achieved by invoking mem_cgroup_mark_under_oom() before
1628 * triggering notification.
1630 if (memcg && memcg->under_oom)
1631 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1634 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1636 if (!current->memcg_may_oom)
1639 * We are in the middle of the charge context here, so we
1640 * don't want to block when potentially sitting on a callstack
1641 * that holds all kinds of filesystem and mm locks.
1643 * Also, the caller may handle a failed allocation gracefully
1644 * (like optional page cache readahead) and so an OOM killer
1645 * invocation might not even be necessary.
1647 * That's why we don't do anything here except remember the
1648 * OOM context and then deal with it at the end of the page
1649 * fault when the stack is unwound, the locks are released,
1650 * and when we know whether the fault was overall successful.
1652 css_get(&memcg->css);
1653 current->memcg_in_oom = memcg;
1654 current->memcg_oom_gfp_mask = mask;
1655 current->memcg_oom_order = order;
1659 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1660 * @handle: actually kill/wait or just clean up the OOM state
1662 * This has to be called at the end of a page fault if the memcg OOM
1663 * handler was enabled.
1665 * Memcg supports userspace OOM handling where failed allocations must
1666 * sleep on a waitqueue until the userspace task resolves the
1667 * situation. Sleeping directly in the charge context with all kinds
1668 * of locks held is not a good idea, instead we remember an OOM state
1669 * in the task and mem_cgroup_oom_synchronize() has to be called at
1670 * the end of the page fault to complete the OOM handling.
1672 * Returns %true if an ongoing memcg OOM situation was detected and
1673 * completed, %false otherwise.
1675 bool mem_cgroup_oom_synchronize(bool handle)
1677 struct mem_cgroup *memcg = current->memcg_in_oom;
1678 struct oom_wait_info owait;
1681 /* OOM is global, do not handle */
1685 if (!handle || oom_killer_disabled)
1688 owait.memcg = memcg;
1689 owait.wait.flags = 0;
1690 owait.wait.func = memcg_oom_wake_function;
1691 owait.wait.private = current;
1692 INIT_LIST_HEAD(&owait.wait.task_list);
1694 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1695 mem_cgroup_mark_under_oom(memcg);
1697 locked = mem_cgroup_oom_trylock(memcg);
1700 mem_cgroup_oom_notify(memcg);
1702 if (locked && !memcg->oom_kill_disable) {
1703 mem_cgroup_unmark_under_oom(memcg);
1704 finish_wait(&memcg_oom_waitq, &owait.wait);
1705 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1706 current->memcg_oom_order);
1709 mem_cgroup_unmark_under_oom(memcg);
1710 finish_wait(&memcg_oom_waitq, &owait.wait);
1714 mem_cgroup_oom_unlock(memcg);
1716 * There is no guarantee that an OOM-lock contender
1717 * sees the wakeups triggered by the OOM kill
1718 * uncharges. Wake any sleepers explicitely.
1720 memcg_oom_recover(memcg);
1723 current->memcg_in_oom = NULL;
1724 css_put(&memcg->css);
1729 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1730 * @page: page that is going to change accounted state
1732 * This function must mark the beginning of an accounted page state
1733 * change to prevent double accounting when the page is concurrently
1734 * being moved to another memcg:
1736 * memcg = mem_cgroup_begin_page_stat(page);
1737 * if (TestClearPageState(page))
1738 * mem_cgroup_update_page_stat(memcg, state, -1);
1739 * mem_cgroup_end_page_stat(memcg);
1741 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1743 struct mem_cgroup *memcg;
1744 unsigned long flags;
1747 * The RCU lock is held throughout the transaction. The fast
1748 * path can get away without acquiring the memcg->move_lock
1749 * because page moving starts with an RCU grace period.
1751 * The RCU lock also protects the memcg from being freed when
1752 * the page state that is going to change is the only thing
1753 * preventing the page from being uncharged.
1754 * E.g. end-writeback clearing PageWriteback(), which allows
1755 * migration to go ahead and uncharge the page before the
1756 * account transaction might be complete.
1760 if (mem_cgroup_disabled())
1763 memcg = page->mem_cgroup;
1764 if (unlikely(!memcg))
1767 if (atomic_read(&memcg->moving_account) <= 0)
1770 spin_lock_irqsave(&memcg->move_lock, flags);
1771 if (memcg != page->mem_cgroup) {
1772 spin_unlock_irqrestore(&memcg->move_lock, flags);
1777 * When charge migration first begins, we can have locked and
1778 * unlocked page stat updates happening concurrently. Track
1779 * the task who has the lock for mem_cgroup_end_page_stat().
1781 memcg->move_lock_task = current;
1782 memcg->move_lock_flags = flags;
1786 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1789 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1790 * @memcg: the memcg that was accounted against
1792 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1794 if (memcg && memcg->move_lock_task == current) {
1795 unsigned long flags = memcg->move_lock_flags;
1797 memcg->move_lock_task = NULL;
1798 memcg->move_lock_flags = 0;
1800 spin_unlock_irqrestore(&memcg->move_lock, flags);
1805 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1808 * size of first charge trial. "32" comes from vmscan.c's magic value.
1809 * TODO: maybe necessary to use big numbers in big irons.
1811 #define CHARGE_BATCH 32U
1812 struct memcg_stock_pcp {
1813 struct mem_cgroup *cached; /* this never be root cgroup */
1814 unsigned int nr_pages;
1815 struct work_struct work;
1816 unsigned long flags;
1817 #define FLUSHING_CACHED_CHARGE 0
1819 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1820 static DEFINE_MUTEX(percpu_charge_mutex);
1823 * consume_stock: Try to consume stocked charge on this cpu.
1824 * @memcg: memcg to consume from.
1825 * @nr_pages: how many pages to charge.
1827 * The charges will only happen if @memcg matches the current cpu's memcg
1828 * stock, and at least @nr_pages are available in that stock. Failure to
1829 * service an allocation will refill the stock.
1831 * returns true if successful, false otherwise.
1833 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1835 struct memcg_stock_pcp *stock;
1838 if (nr_pages > CHARGE_BATCH)
1841 stock = &get_cpu_var(memcg_stock);
1842 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1843 stock->nr_pages -= nr_pages;
1846 put_cpu_var(memcg_stock);
1851 * Returns stocks cached in percpu and reset cached information.
1853 static void drain_stock(struct memcg_stock_pcp *stock)
1855 struct mem_cgroup *old = stock->cached;
1857 if (stock->nr_pages) {
1858 page_counter_uncharge(&old->memory, stock->nr_pages);
1859 if (do_memsw_account())
1860 page_counter_uncharge(&old->memsw, stock->nr_pages);
1861 css_put_many(&old->css, stock->nr_pages);
1862 stock->nr_pages = 0;
1864 stock->cached = NULL;
1868 * This must be called under preempt disabled or must be called by
1869 * a thread which is pinned to local cpu.
1871 static void drain_local_stock(struct work_struct *dummy)
1873 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1875 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1879 * Cache charges(val) to local per_cpu area.
1880 * This will be consumed by consume_stock() function, later.
1882 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1884 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1886 if (stock->cached != memcg) { /* reset if necessary */
1888 stock->cached = memcg;
1890 stock->nr_pages += nr_pages;
1891 put_cpu_var(memcg_stock);
1895 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1896 * of the hierarchy under it.
1898 static void drain_all_stock(struct mem_cgroup *root_memcg)
1902 /* If someone's already draining, avoid adding running more workers. */
1903 if (!mutex_trylock(&percpu_charge_mutex))
1905 /* Notify other cpus that system-wide "drain" is running */
1908 for_each_online_cpu(cpu) {
1909 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1910 struct mem_cgroup *memcg;
1912 memcg = stock->cached;
1913 if (!memcg || !stock->nr_pages)
1915 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1917 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1919 drain_local_stock(&stock->work);
1921 schedule_work_on(cpu, &stock->work);
1926 mutex_unlock(&percpu_charge_mutex);
1929 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1930 unsigned long action,
1933 int cpu = (unsigned long)hcpu;
1934 struct memcg_stock_pcp *stock;
1936 if (action == CPU_ONLINE)
1939 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1942 stock = &per_cpu(memcg_stock, cpu);
1947 static void reclaim_high(struct mem_cgroup *memcg,
1948 unsigned int nr_pages,
1952 if (page_counter_read(&memcg->memory) <= memcg->high)
1954 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1955 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1956 } while ((memcg = parent_mem_cgroup(memcg)));
1959 static void high_work_func(struct work_struct *work)
1961 struct mem_cgroup *memcg;
1963 memcg = container_of(work, struct mem_cgroup, high_work);
1964 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1968 * Scheduled by try_charge() to be executed from the userland return path
1969 * and reclaims memory over the high limit.
1971 void mem_cgroup_handle_over_high(void)
1973 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1974 struct mem_cgroup *memcg;
1976 if (likely(!nr_pages))
1979 memcg = get_mem_cgroup_from_mm(current->mm);
1980 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1981 css_put(&memcg->css);
1982 current->memcg_nr_pages_over_high = 0;
1985 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1986 unsigned int nr_pages)
1988 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1989 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1990 struct mem_cgroup *mem_over_limit;
1991 struct page_counter *counter;
1992 unsigned long nr_reclaimed;
1993 bool may_swap = true;
1994 bool drained = false;
1996 if (mem_cgroup_is_root(memcg))
1999 if (consume_stock(memcg, nr_pages))
2002 if (!do_memsw_account() ||
2003 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2004 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2006 if (do_memsw_account())
2007 page_counter_uncharge(&memcg->memsw, batch);
2008 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2010 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2014 if (batch > nr_pages) {
2020 * Unlike in global OOM situations, memcg is not in a physical
2021 * memory shortage. Allow dying and OOM-killed tasks to
2022 * bypass the last charges so that they can exit quickly and
2023 * free their memory.
2025 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2026 fatal_signal_pending(current) ||
2027 current->flags & PF_EXITING))
2030 if (unlikely(task_in_memcg_oom(current)))
2033 if (!gfpflags_allow_blocking(gfp_mask))
2036 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2038 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2039 gfp_mask, may_swap);
2041 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2045 drain_all_stock(mem_over_limit);
2050 if (gfp_mask & __GFP_NORETRY)
2053 * Even though the limit is exceeded at this point, reclaim
2054 * may have been able to free some pages. Retry the charge
2055 * before killing the task.
2057 * Only for regular pages, though: huge pages are rather
2058 * unlikely to succeed so close to the limit, and we fall back
2059 * to regular pages anyway in case of failure.
2061 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2064 * At task move, charge accounts can be doubly counted. So, it's
2065 * better to wait until the end of task_move if something is going on.
2067 if (mem_cgroup_wait_acct_move(mem_over_limit))
2073 if (gfp_mask & __GFP_NOFAIL)
2076 if (fatal_signal_pending(current))
2079 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2081 mem_cgroup_oom(mem_over_limit, gfp_mask,
2082 get_order(nr_pages * PAGE_SIZE));
2084 if (!(gfp_mask & __GFP_NOFAIL))
2088 * The allocation either can't fail or will lead to more memory
2089 * being freed very soon. Allow memory usage go over the limit
2090 * temporarily by force charging it.
2092 page_counter_charge(&memcg->memory, nr_pages);
2093 if (do_memsw_account())
2094 page_counter_charge(&memcg->memsw, nr_pages);
2095 css_get_many(&memcg->css, nr_pages);
2100 css_get_many(&memcg->css, batch);
2101 if (batch > nr_pages)
2102 refill_stock(memcg, batch - nr_pages);
2105 * If the hierarchy is above the normal consumption range, schedule
2106 * reclaim on returning to userland. We can perform reclaim here
2107 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2108 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2109 * not recorded as it most likely matches current's and won't
2110 * change in the meantime. As high limit is checked again before
2111 * reclaim, the cost of mismatch is negligible.
2114 if (page_counter_read(&memcg->memory) > memcg->high) {
2115 /* Don't bother a random interrupted task */
2116 if (in_interrupt()) {
2117 schedule_work(&memcg->high_work);
2120 current->memcg_nr_pages_over_high += batch;
2121 set_notify_resume(current);
2124 } while ((memcg = parent_mem_cgroup(memcg)));
2129 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2131 if (mem_cgroup_is_root(memcg))
2134 page_counter_uncharge(&memcg->memory, nr_pages);
2135 if (do_memsw_account())
2136 page_counter_uncharge(&memcg->memsw, nr_pages);
2138 css_put_many(&memcg->css, nr_pages);
2141 static void lock_page_lru(struct page *page, int *isolated)
2143 struct zone *zone = page_zone(page);
2145 spin_lock_irq(&zone->lru_lock);
2146 if (PageLRU(page)) {
2147 struct lruvec *lruvec;
2149 lruvec = mem_cgroup_page_lruvec(page, zone);
2151 del_page_from_lru_list(page, lruvec, page_lru(page));
2157 static void unlock_page_lru(struct page *page, int isolated)
2159 struct zone *zone = page_zone(page);
2162 struct lruvec *lruvec;
2164 lruvec = mem_cgroup_page_lruvec(page, zone);
2165 VM_BUG_ON_PAGE(PageLRU(page), page);
2167 add_page_to_lru_list(page, lruvec, page_lru(page));
2169 spin_unlock_irq(&zone->lru_lock);
2172 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2177 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2180 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2181 * may already be on some other mem_cgroup's LRU. Take care of it.
2184 lock_page_lru(page, &isolated);
2187 * Nobody should be changing or seriously looking at
2188 * page->mem_cgroup at this point:
2190 * - the page is uncharged
2192 * - the page is off-LRU
2194 * - an anonymous fault has exclusive page access, except for
2195 * a locked page table
2197 * - a page cache insertion, a swapin fault, or a migration
2198 * have the page locked
2200 page->mem_cgroup = memcg;
2203 unlock_page_lru(page, isolated);
2206 #ifdef CONFIG_MEMCG_KMEM
2207 static int memcg_alloc_cache_id(void)
2212 id = ida_simple_get(&memcg_cache_ida,
2213 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2217 if (id < memcg_nr_cache_ids)
2221 * There's no space for the new id in memcg_caches arrays,
2222 * so we have to grow them.
2224 down_write(&memcg_cache_ids_sem);
2226 size = 2 * (id + 1);
2227 if (size < MEMCG_CACHES_MIN_SIZE)
2228 size = MEMCG_CACHES_MIN_SIZE;
2229 else if (size > MEMCG_CACHES_MAX_SIZE)
2230 size = MEMCG_CACHES_MAX_SIZE;
2232 err = memcg_update_all_caches(size);
2234 err = memcg_update_all_list_lrus(size);
2236 memcg_nr_cache_ids = size;
2238 up_write(&memcg_cache_ids_sem);
2241 ida_simple_remove(&memcg_cache_ida, id);
2247 static void memcg_free_cache_id(int id)
2249 ida_simple_remove(&memcg_cache_ida, id);
2252 struct memcg_kmem_cache_create_work {
2253 struct mem_cgroup *memcg;
2254 struct kmem_cache *cachep;
2255 struct work_struct work;
2258 static void memcg_kmem_cache_create_func(struct work_struct *w)
2260 struct memcg_kmem_cache_create_work *cw =
2261 container_of(w, struct memcg_kmem_cache_create_work, work);
2262 struct mem_cgroup *memcg = cw->memcg;
2263 struct kmem_cache *cachep = cw->cachep;
2265 memcg_create_kmem_cache(memcg, cachep);
2267 css_put(&memcg->css);
2272 * Enqueue the creation of a per-memcg kmem_cache.
2274 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2275 struct kmem_cache *cachep)
2277 struct memcg_kmem_cache_create_work *cw;
2279 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2283 css_get(&memcg->css);
2286 cw->cachep = cachep;
2287 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2289 schedule_work(&cw->work);
2292 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2293 struct kmem_cache *cachep)
2296 * We need to stop accounting when we kmalloc, because if the
2297 * corresponding kmalloc cache is not yet created, the first allocation
2298 * in __memcg_schedule_kmem_cache_create will recurse.
2300 * However, it is better to enclose the whole function. Depending on
2301 * the debugging options enabled, INIT_WORK(), for instance, can
2302 * trigger an allocation. This too, will make us recurse. Because at
2303 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2304 * the safest choice is to do it like this, wrapping the whole function.
2306 current->memcg_kmem_skip_account = 1;
2307 __memcg_schedule_kmem_cache_create(memcg, cachep);
2308 current->memcg_kmem_skip_account = 0;
2312 * Return the kmem_cache we're supposed to use for a slab allocation.
2313 * We try to use the current memcg's version of the cache.
2315 * If the cache does not exist yet, if we are the first user of it,
2316 * we either create it immediately, if possible, or create it asynchronously
2318 * In the latter case, we will let the current allocation go through with
2319 * the original cache.
2321 * Can't be called in interrupt context or from kernel threads.
2322 * This function needs to be called with rcu_read_lock() held.
2324 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2326 struct mem_cgroup *memcg;
2327 struct kmem_cache *memcg_cachep;
2330 VM_BUG_ON(!is_root_cache(cachep));
2332 if (cachep->flags & SLAB_ACCOUNT)
2333 gfp |= __GFP_ACCOUNT;
2335 if (!(gfp & __GFP_ACCOUNT))
2338 if (current->memcg_kmem_skip_account)
2341 memcg = get_mem_cgroup_from_mm(current->mm);
2342 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2346 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2347 if (likely(memcg_cachep))
2348 return memcg_cachep;
2351 * If we are in a safe context (can wait, and not in interrupt
2352 * context), we could be be predictable and return right away.
2353 * This would guarantee that the allocation being performed
2354 * already belongs in the new cache.
2356 * However, there are some clashes that can arrive from locking.
2357 * For instance, because we acquire the slab_mutex while doing
2358 * memcg_create_kmem_cache, this means no further allocation
2359 * could happen with the slab_mutex held. So it's better to
2362 memcg_schedule_kmem_cache_create(memcg, cachep);
2364 css_put(&memcg->css);
2368 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2370 if (!is_root_cache(cachep))
2371 css_put(&cachep->memcg_params.memcg->css);
2374 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2375 struct mem_cgroup *memcg)
2377 unsigned int nr_pages = 1 << order;
2378 struct page_counter *counter;
2381 if (!memcg_kmem_online(memcg))
2384 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2387 ret = try_charge(memcg, gfp, nr_pages);
2389 page_counter_uncharge(&memcg->kmem, nr_pages);
2393 page->mem_cgroup = memcg;
2398 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2400 struct mem_cgroup *memcg;
2403 memcg = get_mem_cgroup_from_mm(current->mm);
2404 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2405 css_put(&memcg->css);
2409 void __memcg_kmem_uncharge(struct page *page, int order)
2411 struct mem_cgroup *memcg = page->mem_cgroup;
2412 unsigned int nr_pages = 1 << order;
2417 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2419 page_counter_uncharge(&memcg->kmem, nr_pages);
2420 page_counter_uncharge(&memcg->memory, nr_pages);
2421 if (do_memsw_account())
2422 page_counter_uncharge(&memcg->memsw, nr_pages);
2424 page->mem_cgroup = NULL;
2425 css_put_many(&memcg->css, nr_pages);
2427 #endif /* CONFIG_MEMCG_KMEM */
2429 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2432 * Because tail pages are not marked as "used", set it. We're under
2433 * zone->lru_lock and migration entries setup in all page mappings.
2435 void mem_cgroup_split_huge_fixup(struct page *head)
2439 if (mem_cgroup_disabled())
2442 for (i = 1; i < HPAGE_PMD_NR; i++)
2443 head[i].mem_cgroup = head->mem_cgroup;
2445 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2448 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2450 #ifdef CONFIG_MEMCG_SWAP
2451 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2454 int val = (charge) ? 1 : -1;
2455 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2459 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2460 * @entry: swap entry to be moved
2461 * @from: mem_cgroup which the entry is moved from
2462 * @to: mem_cgroup which the entry is moved to
2464 * It succeeds only when the swap_cgroup's record for this entry is the same
2465 * as the mem_cgroup's id of @from.
2467 * Returns 0 on success, -EINVAL on failure.
2469 * The caller must have charged to @to, IOW, called page_counter_charge() about
2470 * both res and memsw, and called css_get().
2472 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2473 struct mem_cgroup *from, struct mem_cgroup *to)
2475 unsigned short old_id, new_id;
2477 old_id = mem_cgroup_id(from);
2478 new_id = mem_cgroup_id(to);
2480 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2481 mem_cgroup_swap_statistics(from, false);
2482 mem_cgroup_swap_statistics(to, true);
2488 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2489 struct mem_cgroup *from, struct mem_cgroup *to)
2495 static DEFINE_MUTEX(memcg_limit_mutex);
2497 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2498 unsigned long limit)
2500 unsigned long curusage;
2501 unsigned long oldusage;
2502 bool enlarge = false;
2507 * For keeping hierarchical_reclaim simple, how long we should retry
2508 * is depends on callers. We set our retry-count to be function
2509 * of # of children which we should visit in this loop.
2511 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2512 mem_cgroup_count_children(memcg);
2514 oldusage = page_counter_read(&memcg->memory);
2517 if (signal_pending(current)) {
2522 mutex_lock(&memcg_limit_mutex);
2523 if (limit > memcg->memsw.limit) {
2524 mutex_unlock(&memcg_limit_mutex);
2528 if (limit > memcg->memory.limit)
2530 ret = page_counter_limit(&memcg->memory, limit);
2531 mutex_unlock(&memcg_limit_mutex);
2536 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2538 curusage = page_counter_read(&memcg->memory);
2539 /* Usage is reduced ? */
2540 if (curusage >= oldusage)
2543 oldusage = curusage;
2544 } while (retry_count);
2546 if (!ret && enlarge)
2547 memcg_oom_recover(memcg);
2552 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2553 unsigned long limit)
2555 unsigned long curusage;
2556 unsigned long oldusage;
2557 bool enlarge = false;
2561 /* see mem_cgroup_resize_res_limit */
2562 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2563 mem_cgroup_count_children(memcg);
2565 oldusage = page_counter_read(&memcg->memsw);
2568 if (signal_pending(current)) {
2573 mutex_lock(&memcg_limit_mutex);
2574 if (limit < memcg->memory.limit) {
2575 mutex_unlock(&memcg_limit_mutex);
2579 if (limit > memcg->memsw.limit)
2581 ret = page_counter_limit(&memcg->memsw, limit);
2582 mutex_unlock(&memcg_limit_mutex);
2587 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2589 curusage = page_counter_read(&memcg->memsw);
2590 /* Usage is reduced ? */
2591 if (curusage >= oldusage)
2594 oldusage = curusage;
2595 } while (retry_count);
2597 if (!ret && enlarge)
2598 memcg_oom_recover(memcg);
2603 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2605 unsigned long *total_scanned)
2607 unsigned long nr_reclaimed = 0;
2608 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2609 unsigned long reclaimed;
2611 struct mem_cgroup_tree_per_zone *mctz;
2612 unsigned long excess;
2613 unsigned long nr_scanned;
2618 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2620 * This loop can run a while, specially if mem_cgroup's continuously
2621 * keep exceeding their soft limit and putting the system under
2628 mz = mem_cgroup_largest_soft_limit_node(mctz);
2633 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2634 gfp_mask, &nr_scanned);
2635 nr_reclaimed += reclaimed;
2636 *total_scanned += nr_scanned;
2637 spin_lock_irq(&mctz->lock);
2638 __mem_cgroup_remove_exceeded(mz, mctz);
2641 * If we failed to reclaim anything from this memory cgroup
2642 * it is time to move on to the next cgroup
2646 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2648 excess = soft_limit_excess(mz->memcg);
2650 * One school of thought says that we should not add
2651 * back the node to the tree if reclaim returns 0.
2652 * But our reclaim could return 0, simply because due
2653 * to priority we are exposing a smaller subset of
2654 * memory to reclaim from. Consider this as a longer
2657 /* If excess == 0, no tree ops */
2658 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2659 spin_unlock_irq(&mctz->lock);
2660 css_put(&mz->memcg->css);
2663 * Could not reclaim anything and there are no more
2664 * mem cgroups to try or we seem to be looping without
2665 * reclaiming anything.
2667 if (!nr_reclaimed &&
2669 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2671 } while (!nr_reclaimed);
2673 css_put(&next_mz->memcg->css);
2674 return nr_reclaimed;
2678 * Test whether @memcg has children, dead or alive. Note that this
2679 * function doesn't care whether @memcg has use_hierarchy enabled and
2680 * returns %true if there are child csses according to the cgroup
2681 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2683 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2688 * The lock does not prevent addition or deletion of children, but
2689 * it prevents a new child from being initialized based on this
2690 * parent in css_online(), so it's enough to decide whether
2691 * hierarchically inherited attributes can still be changed or not.
2693 lockdep_assert_held(&memcg_create_mutex);
2696 ret = css_next_child(NULL, &memcg->css);
2702 * Reclaims as many pages from the given memcg as possible and moves
2703 * the rest to the parent.
2705 * Caller is responsible for holding css reference for memcg.
2707 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2709 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2711 /* we call try-to-free pages for make this cgroup empty */
2712 lru_add_drain_all();
2713 /* try to free all pages in this cgroup */
2714 while (nr_retries && page_counter_read(&memcg->memory)) {
2717 if (signal_pending(current))
2720 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2724 /* maybe some writeback is necessary */
2725 congestion_wait(BLK_RW_ASYNC, HZ/10);
2733 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2734 char *buf, size_t nbytes,
2737 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2739 if (mem_cgroup_is_root(memcg))
2741 return mem_cgroup_force_empty(memcg) ?: nbytes;
2744 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2747 return mem_cgroup_from_css(css)->use_hierarchy;
2750 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2751 struct cftype *cft, u64 val)
2754 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2755 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2757 mutex_lock(&memcg_create_mutex);
2759 if (memcg->use_hierarchy == val)
2763 * If parent's use_hierarchy is set, we can't make any modifications
2764 * in the child subtrees. If it is unset, then the change can
2765 * occur, provided the current cgroup has no children.
2767 * For the root cgroup, parent_mem is NULL, we allow value to be
2768 * set if there are no children.
2770 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2771 (val == 1 || val == 0)) {
2772 if (!memcg_has_children(memcg))
2773 memcg->use_hierarchy = val;
2780 mutex_unlock(&memcg_create_mutex);
2785 static unsigned long tree_stat(struct mem_cgroup *memcg,
2786 enum mem_cgroup_stat_index idx)
2788 struct mem_cgroup *iter;
2789 unsigned long val = 0;
2791 for_each_mem_cgroup_tree(iter, memcg)
2792 val += mem_cgroup_read_stat(iter, idx);
2797 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2801 if (mem_cgroup_is_root(memcg)) {
2802 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2803 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2805 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2808 val = page_counter_read(&memcg->memory);
2810 val = page_counter_read(&memcg->memsw);
2823 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2826 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2827 struct page_counter *counter;
2829 switch (MEMFILE_TYPE(cft->private)) {
2831 counter = &memcg->memory;
2834 counter = &memcg->memsw;
2837 counter = &memcg->kmem;
2843 switch (MEMFILE_ATTR(cft->private)) {
2845 if (counter == &memcg->memory)
2846 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2847 if (counter == &memcg->memsw)
2848 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2849 return (u64)page_counter_read(counter) * PAGE_SIZE;
2851 return (u64)counter->limit * PAGE_SIZE;
2853 return (u64)counter->watermark * PAGE_SIZE;
2855 return counter->failcnt;
2856 case RES_SOFT_LIMIT:
2857 return (u64)memcg->soft_limit * PAGE_SIZE;
2863 #ifdef CONFIG_MEMCG_KMEM
2864 static int memcg_online_kmem(struct mem_cgroup *memcg)
2869 BUG_ON(memcg->kmemcg_id >= 0);
2870 BUG_ON(memcg->kmem_state);
2873 * For simplicity, we won't allow this to be disabled. It also can't
2874 * be changed if the cgroup has children already, or if tasks had
2877 * If tasks join before we set the limit, a person looking at
2878 * kmem.usage_in_bytes will have no way to determine when it took
2879 * place, which makes the value quite meaningless.
2881 * After it first became limited, changes in the value of the limit are
2882 * of course permitted.
2884 mutex_lock(&memcg_create_mutex);
2885 if (cgroup_is_populated(memcg->css.cgroup) ||
2886 (memcg->use_hierarchy && memcg_has_children(memcg)))
2888 mutex_unlock(&memcg_create_mutex);
2892 memcg_id = memcg_alloc_cache_id();
2898 static_branch_inc(&memcg_kmem_enabled_key);
2900 * A memory cgroup is considered kmem-online as soon as it gets
2901 * kmemcg_id. Setting the id after enabling static branching will
2902 * guarantee no one starts accounting before all call sites are
2905 memcg->kmemcg_id = memcg_id;
2906 memcg->kmem_state = KMEM_ONLINE;
2911 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2912 unsigned long limit)
2916 mutex_lock(&memcg_limit_mutex);
2917 /* Top-level cgroup doesn't propagate from root */
2918 if (!memcg_kmem_online(memcg)) {
2919 ret = memcg_online_kmem(memcg);
2923 ret = page_counter_limit(&memcg->kmem, limit);
2925 mutex_unlock(&memcg_limit_mutex);
2929 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2932 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2937 mutex_lock(&memcg_limit_mutex);
2939 * If the parent cgroup is not kmem-online now, it cannot be
2940 * onlined after this point, because it has at least one child
2943 if (memcg_kmem_online(parent))
2944 ret = memcg_online_kmem(memcg);
2945 mutex_unlock(&memcg_limit_mutex);
2949 static int memcg_init_kmem(struct mem_cgroup *memcg)
2953 ret = memcg_propagate_kmem(memcg);
2957 return tcp_init_cgroup(memcg);
2960 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2962 struct cgroup_subsys_state *css;
2963 struct mem_cgroup *parent, *child;
2966 if (memcg->kmem_state != KMEM_ONLINE)
2969 * Clear the online state before clearing memcg_caches array
2970 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2971 * guarantees that no cache will be created for this cgroup
2972 * after we are done (see memcg_create_kmem_cache()).
2974 memcg->kmem_state = KMEM_ALLOCATED;
2976 memcg_deactivate_kmem_caches(memcg);
2978 kmemcg_id = memcg->kmemcg_id;
2979 BUG_ON(kmemcg_id < 0);
2981 parent = parent_mem_cgroup(memcg);
2983 parent = root_mem_cgroup;
2986 * Change kmemcg_id of this cgroup and all its descendants to the
2987 * parent's id, and then move all entries from this cgroup's list_lrus
2988 * to ones of the parent. After we have finished, all list_lrus
2989 * corresponding to this cgroup are guaranteed to remain empty. The
2990 * ordering is imposed by list_lru_node->lock taken by
2991 * memcg_drain_all_list_lrus().
2993 css_for_each_descendant_pre(css, &memcg->css) {
2994 child = mem_cgroup_from_css(css);
2995 BUG_ON(child->kmemcg_id != kmemcg_id);
2996 child->kmemcg_id = parent->kmemcg_id;
2997 if (!memcg->use_hierarchy)
3000 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3002 memcg_free_cache_id(kmemcg_id);
3005 static void memcg_free_kmem(struct mem_cgroup *memcg)
3007 if (memcg->kmem_state == KMEM_ALLOCATED) {
3008 memcg_destroy_kmem_caches(memcg);
3009 static_branch_dec(&memcg_kmem_enabled_key);
3010 WARN_ON(page_counter_read(&memcg->kmem));
3012 tcp_destroy_cgroup(memcg);
3015 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3016 unsigned long limit)
3020 static int memcg_init_kmem(struct mem_cgroup *memcg)
3024 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3027 static void memcg_free_kmem(struct mem_cgroup *memcg)
3030 #endif /* CONFIG_MEMCG_KMEM */
3033 * The user of this function is...
3036 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3037 char *buf, size_t nbytes, loff_t off)
3039 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3040 unsigned long nr_pages;
3043 buf = strstrip(buf);
3044 ret = page_counter_memparse(buf, "-1", &nr_pages);
3048 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3050 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3054 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3056 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3059 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3062 ret = memcg_update_kmem_limit(memcg, nr_pages);
3066 case RES_SOFT_LIMIT:
3067 memcg->soft_limit = nr_pages;
3071 return ret ?: nbytes;
3074 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3075 size_t nbytes, loff_t off)
3077 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3078 struct page_counter *counter;
3080 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3082 counter = &memcg->memory;
3085 counter = &memcg->memsw;
3088 counter = &memcg->kmem;
3094 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3096 page_counter_reset_watermark(counter);
3099 counter->failcnt = 0;
3108 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3111 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3115 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3116 struct cftype *cft, u64 val)
3118 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3120 if (val & ~MOVE_MASK)
3124 * No kind of locking is needed in here, because ->can_attach() will
3125 * check this value once in the beginning of the process, and then carry
3126 * on with stale data. This means that changes to this value will only
3127 * affect task migrations starting after the change.
3129 memcg->move_charge_at_immigrate = val;
3133 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3134 struct cftype *cft, u64 val)
3141 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3145 unsigned int lru_mask;
3148 static const struct numa_stat stats[] = {
3149 { "total", LRU_ALL },
3150 { "file", LRU_ALL_FILE },
3151 { "anon", LRU_ALL_ANON },
3152 { "unevictable", BIT(LRU_UNEVICTABLE) },
3154 const struct numa_stat *stat;
3157 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3159 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3160 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3161 seq_printf(m, "%s=%lu", stat->name, nr);
3162 for_each_node_state(nid, N_MEMORY) {
3163 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3165 seq_printf(m, " N%d=%lu", nid, nr);
3170 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3171 struct mem_cgroup *iter;
3174 for_each_mem_cgroup_tree(iter, memcg)
3175 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3176 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3177 for_each_node_state(nid, N_MEMORY) {
3179 for_each_mem_cgroup_tree(iter, memcg)
3180 nr += mem_cgroup_node_nr_lru_pages(
3181 iter, nid, stat->lru_mask);
3182 seq_printf(m, " N%d=%lu", nid, nr);
3189 #endif /* CONFIG_NUMA */
3191 static int memcg_stat_show(struct seq_file *m, void *v)
3193 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3194 unsigned long memory, memsw;
3195 struct mem_cgroup *mi;
3198 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3199 MEM_CGROUP_STAT_NSTATS);
3200 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3201 MEM_CGROUP_EVENTS_NSTATS);
3202 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3204 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3205 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3207 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3208 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3211 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3212 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3213 mem_cgroup_read_events(memcg, i));
3215 for (i = 0; i < NR_LRU_LISTS; i++)
3216 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3217 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3219 /* Hierarchical information */
3220 memory = memsw = PAGE_COUNTER_MAX;
3221 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3222 memory = min(memory, mi->memory.limit);
3223 memsw = min(memsw, mi->memsw.limit);
3225 seq_printf(m, "hierarchical_memory_limit %llu\n",
3226 (u64)memory * PAGE_SIZE);
3227 if (do_memsw_account())
3228 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3229 (u64)memsw * PAGE_SIZE);
3231 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3232 unsigned long long val = 0;
3234 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3236 for_each_mem_cgroup_tree(mi, memcg)
3237 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3238 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3241 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3242 unsigned long long val = 0;
3244 for_each_mem_cgroup_tree(mi, memcg)
3245 val += mem_cgroup_read_events(mi, i);
3246 seq_printf(m, "total_%s %llu\n",
3247 mem_cgroup_events_names[i], val);
3250 for (i = 0; i < NR_LRU_LISTS; i++) {
3251 unsigned long long val = 0;
3253 for_each_mem_cgroup_tree(mi, memcg)
3254 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3255 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3258 #ifdef CONFIG_DEBUG_VM
3261 struct mem_cgroup_per_zone *mz;
3262 struct zone_reclaim_stat *rstat;
3263 unsigned long recent_rotated[2] = {0, 0};
3264 unsigned long recent_scanned[2] = {0, 0};
3266 for_each_online_node(nid)
3267 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3268 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3269 rstat = &mz->lruvec.reclaim_stat;
3271 recent_rotated[0] += rstat->recent_rotated[0];
3272 recent_rotated[1] += rstat->recent_rotated[1];
3273 recent_scanned[0] += rstat->recent_scanned[0];
3274 recent_scanned[1] += rstat->recent_scanned[1];
3276 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3277 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3278 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3279 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3286 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3289 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3291 return mem_cgroup_swappiness(memcg);
3294 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3295 struct cftype *cft, u64 val)
3297 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3303 memcg->swappiness = val;
3305 vm_swappiness = val;
3310 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3312 struct mem_cgroup_threshold_ary *t;
3313 unsigned long usage;
3318 t = rcu_dereference(memcg->thresholds.primary);
3320 t = rcu_dereference(memcg->memsw_thresholds.primary);
3325 usage = mem_cgroup_usage(memcg, swap);
3328 * current_threshold points to threshold just below or equal to usage.
3329 * If it's not true, a threshold was crossed after last
3330 * call of __mem_cgroup_threshold().
3332 i = t->current_threshold;
3335 * Iterate backward over array of thresholds starting from
3336 * current_threshold and check if a threshold is crossed.
3337 * If none of thresholds below usage is crossed, we read
3338 * only one element of the array here.
3340 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3341 eventfd_signal(t->entries[i].eventfd, 1);
3343 /* i = current_threshold + 1 */
3347 * Iterate forward over array of thresholds starting from
3348 * current_threshold+1 and check if a threshold is crossed.
3349 * If none of thresholds above usage is crossed, we read
3350 * only one element of the array here.
3352 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3353 eventfd_signal(t->entries[i].eventfd, 1);
3355 /* Update current_threshold */
3356 t->current_threshold = i - 1;
3361 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3364 __mem_cgroup_threshold(memcg, false);
3365 if (do_memsw_account())
3366 __mem_cgroup_threshold(memcg, true);
3368 memcg = parent_mem_cgroup(memcg);
3372 static int compare_thresholds(const void *a, const void *b)
3374 const struct mem_cgroup_threshold *_a = a;
3375 const struct mem_cgroup_threshold *_b = b;
3377 if (_a->threshold > _b->threshold)
3380 if (_a->threshold < _b->threshold)
3386 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3388 struct mem_cgroup_eventfd_list *ev;
3390 spin_lock(&memcg_oom_lock);
3392 list_for_each_entry(ev, &memcg->oom_notify, list)
3393 eventfd_signal(ev->eventfd, 1);
3395 spin_unlock(&memcg_oom_lock);
3399 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3401 struct mem_cgroup *iter;
3403 for_each_mem_cgroup_tree(iter, memcg)
3404 mem_cgroup_oom_notify_cb(iter);
3407 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3408 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3410 struct mem_cgroup_thresholds *thresholds;
3411 struct mem_cgroup_threshold_ary *new;
3412 unsigned long threshold;
3413 unsigned long usage;
3416 ret = page_counter_memparse(args, "-1", &threshold);
3420 mutex_lock(&memcg->thresholds_lock);
3423 thresholds = &memcg->thresholds;
3424 usage = mem_cgroup_usage(memcg, false);
3425 } else if (type == _MEMSWAP) {
3426 thresholds = &memcg->memsw_thresholds;
3427 usage = mem_cgroup_usage(memcg, true);
3431 /* Check if a threshold crossed before adding a new one */
3432 if (thresholds->primary)
3433 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3435 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3437 /* Allocate memory for new array of thresholds */
3438 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3446 /* Copy thresholds (if any) to new array */
3447 if (thresholds->primary) {
3448 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3449 sizeof(struct mem_cgroup_threshold));
3452 /* Add new threshold */
3453 new->entries[size - 1].eventfd = eventfd;
3454 new->entries[size - 1].threshold = threshold;
3456 /* Sort thresholds. Registering of new threshold isn't time-critical */
3457 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3458 compare_thresholds, NULL);
3460 /* Find current threshold */
3461 new->current_threshold = -1;
3462 for (i = 0; i < size; i++) {
3463 if (new->entries[i].threshold <= usage) {
3465 * new->current_threshold will not be used until
3466 * rcu_assign_pointer(), so it's safe to increment
3469 ++new->current_threshold;
3474 /* Free old spare buffer and save old primary buffer as spare */
3475 kfree(thresholds->spare);
3476 thresholds->spare = thresholds->primary;
3478 rcu_assign_pointer(thresholds->primary, new);
3480 /* To be sure that nobody uses thresholds */
3484 mutex_unlock(&memcg->thresholds_lock);
3489 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3490 struct eventfd_ctx *eventfd, const char *args)
3492 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3495 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3496 struct eventfd_ctx *eventfd, const char *args)
3498 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3501 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3502 struct eventfd_ctx *eventfd, enum res_type type)
3504 struct mem_cgroup_thresholds *thresholds;
3505 struct mem_cgroup_threshold_ary *new;
3506 unsigned long usage;
3509 mutex_lock(&memcg->thresholds_lock);
3512 thresholds = &memcg->thresholds;
3513 usage = mem_cgroup_usage(memcg, false);
3514 } else if (type == _MEMSWAP) {
3515 thresholds = &memcg->memsw_thresholds;
3516 usage = mem_cgroup_usage(memcg, true);
3520 if (!thresholds->primary)
3523 /* Check if a threshold crossed before removing */
3524 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3526 /* Calculate new number of threshold */
3528 for (i = 0; i < thresholds->primary->size; i++) {
3529 if (thresholds->primary->entries[i].eventfd != eventfd)
3533 new = thresholds->spare;
3535 /* Set thresholds array to NULL if we don't have thresholds */
3544 /* Copy thresholds and find current threshold */
3545 new->current_threshold = -1;
3546 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3547 if (thresholds->primary->entries[i].eventfd == eventfd)
3550 new->entries[j] = thresholds->primary->entries[i];
3551 if (new->entries[j].threshold <= usage) {
3553 * new->current_threshold will not be used
3554 * until rcu_assign_pointer(), so it's safe to increment
3557 ++new->current_threshold;
3563 /* Swap primary and spare array */
3564 thresholds->spare = thresholds->primary;
3566 rcu_assign_pointer(thresholds->primary, new);
3568 /* To be sure that nobody uses thresholds */
3571 /* If all events are unregistered, free the spare array */
3573 kfree(thresholds->spare);
3574 thresholds->spare = NULL;
3577 mutex_unlock(&memcg->thresholds_lock);
3580 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3581 struct eventfd_ctx *eventfd)
3583 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3586 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3587 struct eventfd_ctx *eventfd)
3589 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3592 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3593 struct eventfd_ctx *eventfd, const char *args)
3595 struct mem_cgroup_eventfd_list *event;
3597 event = kmalloc(sizeof(*event), GFP_KERNEL);
3601 spin_lock(&memcg_oom_lock);
3603 event->eventfd = eventfd;
3604 list_add(&event->list, &memcg->oom_notify);
3606 /* already in OOM ? */
3607 if (memcg->under_oom)
3608 eventfd_signal(eventfd, 1);
3609 spin_unlock(&memcg_oom_lock);
3614 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3615 struct eventfd_ctx *eventfd)
3617 struct mem_cgroup_eventfd_list *ev, *tmp;
3619 spin_lock(&memcg_oom_lock);
3621 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3622 if (ev->eventfd == eventfd) {
3623 list_del(&ev->list);
3628 spin_unlock(&memcg_oom_lock);
3631 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3633 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3635 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3636 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3640 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3641 struct cftype *cft, u64 val)
3643 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3645 /* cannot set to root cgroup and only 0 and 1 are allowed */
3646 if (!css->parent || !((val == 0) || (val == 1)))
3649 memcg->oom_kill_disable = val;
3651 memcg_oom_recover(memcg);
3656 #ifdef CONFIG_CGROUP_WRITEBACK
3658 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3660 return &memcg->cgwb_list;
3663 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3665 return wb_domain_init(&memcg->cgwb_domain, gfp);
3668 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3670 wb_domain_exit(&memcg->cgwb_domain);
3673 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3675 wb_domain_size_changed(&memcg->cgwb_domain);
3678 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3680 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3682 if (!memcg->css.parent)
3685 return &memcg->cgwb_domain;
3689 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3690 * @wb: bdi_writeback in question
3691 * @pfilepages: out parameter for number of file pages
3692 * @pheadroom: out parameter for number of allocatable pages according to memcg
3693 * @pdirty: out parameter for number of dirty pages
3694 * @pwriteback: out parameter for number of pages under writeback
3696 * Determine the numbers of file, headroom, dirty, and writeback pages in
3697 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3698 * is a bit more involved.
3700 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3701 * headroom is calculated as the lowest headroom of itself and the
3702 * ancestors. Note that this doesn't consider the actual amount of
3703 * available memory in the system. The caller should further cap
3704 * *@pheadroom accordingly.
3706 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3707 unsigned long *pheadroom, unsigned long *pdirty,
3708 unsigned long *pwriteback)
3710 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3711 struct mem_cgroup *parent;
3713 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3715 /* this should eventually include NR_UNSTABLE_NFS */
3716 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3717 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3718 (1 << LRU_ACTIVE_FILE));
3719 *pheadroom = PAGE_COUNTER_MAX;
3721 while ((parent = parent_mem_cgroup(memcg))) {
3722 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3723 unsigned long used = page_counter_read(&memcg->memory);
3725 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3730 #else /* CONFIG_CGROUP_WRITEBACK */
3732 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3737 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3741 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3745 #endif /* CONFIG_CGROUP_WRITEBACK */
3748 * DO NOT USE IN NEW FILES.
3750 * "cgroup.event_control" implementation.
3752 * This is way over-engineered. It tries to support fully configurable
3753 * events for each user. Such level of flexibility is completely
3754 * unnecessary especially in the light of the planned unified hierarchy.
3756 * Please deprecate this and replace with something simpler if at all
3761 * Unregister event and free resources.
3763 * Gets called from workqueue.
3765 static void memcg_event_remove(struct work_struct *work)
3767 struct mem_cgroup_event *event =
3768 container_of(work, struct mem_cgroup_event, remove);
3769 struct mem_cgroup *memcg = event->memcg;
3771 remove_wait_queue(event->wqh, &event->wait);
3773 event->unregister_event(memcg, event->eventfd);
3775 /* Notify userspace the event is going away. */
3776 eventfd_signal(event->eventfd, 1);
3778 eventfd_ctx_put(event->eventfd);
3780 css_put(&memcg->css);
3784 * Gets called on POLLHUP on eventfd when user closes it.
3786 * Called with wqh->lock held and interrupts disabled.
3788 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3789 int sync, void *key)
3791 struct mem_cgroup_event *event =
3792 container_of(wait, struct mem_cgroup_event, wait);
3793 struct mem_cgroup *memcg = event->memcg;
3794 unsigned long flags = (unsigned long)key;
3796 if (flags & POLLHUP) {
3798 * If the event has been detached at cgroup removal, we
3799 * can simply return knowing the other side will cleanup
3802 * We can't race against event freeing since the other
3803 * side will require wqh->lock via remove_wait_queue(),
3806 spin_lock(&memcg->event_list_lock);
3807 if (!list_empty(&event->list)) {
3808 list_del_init(&event->list);
3810 * We are in atomic context, but cgroup_event_remove()
3811 * may sleep, so we have to call it in workqueue.
3813 schedule_work(&event->remove);
3815 spin_unlock(&memcg->event_list_lock);
3821 static void memcg_event_ptable_queue_proc(struct file *file,
3822 wait_queue_head_t *wqh, poll_table *pt)
3824 struct mem_cgroup_event *event =
3825 container_of(pt, struct mem_cgroup_event, pt);
3828 add_wait_queue(wqh, &event->wait);
3832 * DO NOT USE IN NEW FILES.
3834 * Parse input and register new cgroup event handler.
3836 * Input must be in format '<event_fd> <control_fd> <args>'.
3837 * Interpretation of args is defined by control file implementation.
3839 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3840 char *buf, size_t nbytes, loff_t off)
3842 struct cgroup_subsys_state *css = of_css(of);
3843 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3844 struct mem_cgroup_event *event;
3845 struct cgroup_subsys_state *cfile_css;
3846 unsigned int efd, cfd;
3853 buf = strstrip(buf);
3855 efd = simple_strtoul(buf, &endp, 10);
3860 cfd = simple_strtoul(buf, &endp, 10);
3861 if ((*endp != ' ') && (*endp != '\0'))
3865 event = kzalloc(sizeof(*event), GFP_KERNEL);
3869 event->memcg = memcg;
3870 INIT_LIST_HEAD(&event->list);
3871 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3872 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3873 INIT_WORK(&event->remove, memcg_event_remove);
3881 event->eventfd = eventfd_ctx_fileget(efile.file);
3882 if (IS_ERR(event->eventfd)) {
3883 ret = PTR_ERR(event->eventfd);
3890 goto out_put_eventfd;
3893 /* the process need read permission on control file */
3894 /* AV: shouldn't we check that it's been opened for read instead? */
3895 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3900 * Determine the event callbacks and set them in @event. This used
3901 * to be done via struct cftype but cgroup core no longer knows
3902 * about these events. The following is crude but the whole thing
3903 * is for compatibility anyway.
3905 * DO NOT ADD NEW FILES.
3907 name = cfile.file->f_path.dentry->d_name.name;
3909 if (!strcmp(name, "memory.usage_in_bytes")) {
3910 event->register_event = mem_cgroup_usage_register_event;
3911 event->unregister_event = mem_cgroup_usage_unregister_event;
3912 } else if (!strcmp(name, "memory.oom_control")) {
3913 event->register_event = mem_cgroup_oom_register_event;
3914 event->unregister_event = mem_cgroup_oom_unregister_event;
3915 } else if (!strcmp(name, "memory.pressure_level")) {
3916 event->register_event = vmpressure_register_event;
3917 event->unregister_event = vmpressure_unregister_event;
3918 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3919 event->register_event = memsw_cgroup_usage_register_event;
3920 event->unregister_event = memsw_cgroup_usage_unregister_event;
3927 * Verify @cfile should belong to @css. Also, remaining events are
3928 * automatically removed on cgroup destruction but the removal is
3929 * asynchronous, so take an extra ref on @css.
3931 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3932 &memory_cgrp_subsys);
3934 if (IS_ERR(cfile_css))
3936 if (cfile_css != css) {
3941 ret = event->register_event(memcg, event->eventfd, buf);
3945 efile.file->f_op->poll(efile.file, &event->pt);
3947 spin_lock(&memcg->event_list_lock);
3948 list_add(&event->list, &memcg->event_list);
3949 spin_unlock(&memcg->event_list_lock);
3961 eventfd_ctx_put(event->eventfd);
3970 static struct cftype mem_cgroup_legacy_files[] = {
3972 .name = "usage_in_bytes",
3973 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3974 .read_u64 = mem_cgroup_read_u64,
3977 .name = "max_usage_in_bytes",
3978 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3979 .write = mem_cgroup_reset,
3980 .read_u64 = mem_cgroup_read_u64,
3983 .name = "limit_in_bytes",
3984 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3985 .write = mem_cgroup_write,
3986 .read_u64 = mem_cgroup_read_u64,
3989 .name = "soft_limit_in_bytes",
3990 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3991 .write = mem_cgroup_write,
3992 .read_u64 = mem_cgroup_read_u64,
3996 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3997 .write = mem_cgroup_reset,
3998 .read_u64 = mem_cgroup_read_u64,
4002 .seq_show = memcg_stat_show,
4005 .name = "force_empty",
4006 .write = mem_cgroup_force_empty_write,
4009 .name = "use_hierarchy",
4010 .write_u64 = mem_cgroup_hierarchy_write,
4011 .read_u64 = mem_cgroup_hierarchy_read,
4014 .name = "cgroup.event_control", /* XXX: for compat */
4015 .write = memcg_write_event_control,
4016 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4019 .name = "swappiness",
4020 .read_u64 = mem_cgroup_swappiness_read,
4021 .write_u64 = mem_cgroup_swappiness_write,
4024 .name = "move_charge_at_immigrate",
4025 .read_u64 = mem_cgroup_move_charge_read,
4026 .write_u64 = mem_cgroup_move_charge_write,
4029 .name = "oom_control",
4030 .seq_show = mem_cgroup_oom_control_read,
4031 .write_u64 = mem_cgroup_oom_control_write,
4032 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4035 .name = "pressure_level",
4039 .name = "numa_stat",
4040 .seq_show = memcg_numa_stat_show,
4043 #ifdef CONFIG_MEMCG_KMEM
4045 .name = "kmem.limit_in_bytes",
4046 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4047 .write = mem_cgroup_write,
4048 .read_u64 = mem_cgroup_read_u64,
4051 .name = "kmem.usage_in_bytes",
4052 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4053 .read_u64 = mem_cgroup_read_u64,
4056 .name = "kmem.failcnt",
4057 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4058 .write = mem_cgroup_reset,
4059 .read_u64 = mem_cgroup_read_u64,
4062 .name = "kmem.max_usage_in_bytes",
4063 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4064 .write = mem_cgroup_reset,
4065 .read_u64 = mem_cgroup_read_u64,
4067 #ifdef CONFIG_SLABINFO
4069 .name = "kmem.slabinfo",
4070 .seq_start = slab_start,
4071 .seq_next = slab_next,
4072 .seq_stop = slab_stop,
4073 .seq_show = memcg_slab_show,
4077 { }, /* terminate */
4080 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4082 struct mem_cgroup_per_node *pn;
4083 struct mem_cgroup_per_zone *mz;
4084 int zone, tmp = node;
4086 * This routine is called against possible nodes.
4087 * But it's BUG to call kmalloc() against offline node.
4089 * TODO: this routine can waste much memory for nodes which will
4090 * never be onlined. It's better to use memory hotplug callback
4093 if (!node_state(node, N_NORMAL_MEMORY))
4095 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4099 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4100 mz = &pn->zoneinfo[zone];
4101 lruvec_init(&mz->lruvec);
4102 mz->usage_in_excess = 0;
4103 mz->on_tree = false;
4106 memcg->nodeinfo[node] = pn;
4110 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4112 kfree(memcg->nodeinfo[node]);
4115 static struct mem_cgroup *mem_cgroup_alloc(void)
4117 struct mem_cgroup *memcg;
4120 size = sizeof(struct mem_cgroup);
4121 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4123 memcg = kzalloc(size, GFP_KERNEL);
4127 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4131 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4137 free_percpu(memcg->stat);
4144 * At destroying mem_cgroup, references from swap_cgroup can remain.
4145 * (scanning all at force_empty is too costly...)
4147 * Instead of clearing all references at force_empty, we remember
4148 * the number of reference from swap_cgroup and free mem_cgroup when
4149 * it goes down to 0.
4151 * Removal of cgroup itself succeeds regardless of refs from swap.
4154 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4158 cancel_work_sync(&memcg->high_work);
4160 mem_cgroup_remove_from_trees(memcg);
4163 free_mem_cgroup_per_zone_info(memcg, node);
4165 free_percpu(memcg->stat);
4166 memcg_wb_domain_exit(memcg);
4170 static struct cgroup_subsys_state * __ref
4171 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4173 struct mem_cgroup *memcg;
4174 long error = -ENOMEM;
4177 memcg = mem_cgroup_alloc();
4179 return ERR_PTR(error);
4182 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4186 if (parent_css == NULL) {
4187 root_mem_cgroup = memcg;
4188 page_counter_init(&memcg->memory, NULL);
4189 memcg->high = PAGE_COUNTER_MAX;
4190 memcg->soft_limit = PAGE_COUNTER_MAX;
4191 page_counter_init(&memcg->memsw, NULL);
4192 page_counter_init(&memcg->kmem, NULL);
4195 INIT_WORK(&memcg->high_work, high_work_func);
4196 memcg->last_scanned_node = MAX_NUMNODES;
4197 INIT_LIST_HEAD(&memcg->oom_notify);
4198 memcg->move_charge_at_immigrate = 0;
4199 mutex_init(&memcg->thresholds_lock);
4200 spin_lock_init(&memcg->move_lock);
4201 vmpressure_init(&memcg->vmpressure);
4202 INIT_LIST_HEAD(&memcg->event_list);
4203 spin_lock_init(&memcg->event_list_lock);
4204 #ifdef CONFIG_MEMCG_KMEM
4205 memcg->kmemcg_id = -1;
4207 #ifdef CONFIG_CGROUP_WRITEBACK
4208 INIT_LIST_HEAD(&memcg->cgwb_list);
4211 memcg->socket_pressure = jiffies;
4216 __mem_cgroup_free(memcg);
4217 return ERR_PTR(error);
4221 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4223 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4224 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4227 if (css->id > MEM_CGROUP_ID_MAX)
4233 mutex_lock(&memcg_create_mutex);
4235 memcg->use_hierarchy = parent->use_hierarchy;
4236 memcg->oom_kill_disable = parent->oom_kill_disable;
4237 memcg->swappiness = mem_cgroup_swappiness(parent);
4239 if (parent->use_hierarchy) {
4240 page_counter_init(&memcg->memory, &parent->memory);
4241 memcg->high = PAGE_COUNTER_MAX;
4242 memcg->soft_limit = PAGE_COUNTER_MAX;
4243 page_counter_init(&memcg->memsw, &parent->memsw);
4244 page_counter_init(&memcg->kmem, &parent->kmem);
4247 * No need to take a reference to the parent because cgroup
4248 * core guarantees its existence.
4251 page_counter_init(&memcg->memory, NULL);
4252 memcg->high = PAGE_COUNTER_MAX;
4253 memcg->soft_limit = PAGE_COUNTER_MAX;
4254 page_counter_init(&memcg->memsw, NULL);
4255 page_counter_init(&memcg->kmem, NULL);
4257 * Deeper hierachy with use_hierarchy == false doesn't make
4258 * much sense so let cgroup subsystem know about this
4259 * unfortunate state in our controller.
4261 if (parent != root_mem_cgroup)
4262 memory_cgrp_subsys.broken_hierarchy = true;
4264 mutex_unlock(&memcg_create_mutex);
4266 ret = memcg_init_kmem(memcg);
4271 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4272 static_branch_inc(&memcg_sockets_enabled_key);
4276 * Make sure the memcg is initialized: mem_cgroup_iter()
4277 * orders reading memcg->initialized against its callers
4278 * reading the memcg members.
4280 smp_store_release(&memcg->initialized, 1);
4285 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4287 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4288 struct mem_cgroup_event *event, *tmp;
4291 * Unregister events and notify userspace.
4292 * Notify userspace about cgroup removing only after rmdir of cgroup
4293 * directory to avoid race between userspace and kernelspace.
4295 spin_lock(&memcg->event_list_lock);
4296 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4297 list_del_init(&event->list);
4298 schedule_work(&event->remove);
4300 spin_unlock(&memcg->event_list_lock);
4302 vmpressure_cleanup(&memcg->vmpressure);
4304 memcg_offline_kmem(memcg);
4306 wb_memcg_offline(memcg);
4309 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4311 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4313 invalidate_reclaim_iterators(memcg);
4316 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4318 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4320 memcg_free_kmem(memcg);
4322 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4323 static_branch_dec(&memcg_sockets_enabled_key);
4325 __mem_cgroup_free(memcg);
4329 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4330 * @css: the target css
4332 * Reset the states of the mem_cgroup associated with @css. This is
4333 * invoked when the userland requests disabling on the default hierarchy
4334 * but the memcg is pinned through dependency. The memcg should stop
4335 * applying policies and should revert to the vanilla state as it may be
4336 * made visible again.
4338 * The current implementation only resets the essential configurations.
4339 * This needs to be expanded to cover all the visible parts.
4341 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4345 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4346 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4347 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4349 memcg->high = PAGE_COUNTER_MAX;
4350 memcg->soft_limit = PAGE_COUNTER_MAX;
4351 memcg_wb_domain_size_changed(memcg);
4355 /* Handlers for move charge at task migration. */
4356 static int mem_cgroup_do_precharge(unsigned long count)
4360 /* Try a single bulk charge without reclaim first, kswapd may wake */
4361 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4363 mc.precharge += count;
4367 /* Try charges one by one with reclaim */
4369 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4379 * get_mctgt_type - get target type of moving charge
4380 * @vma: the vma the pte to be checked belongs
4381 * @addr: the address corresponding to the pte to be checked
4382 * @ptent: the pte to be checked
4383 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4386 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4387 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4388 * move charge. if @target is not NULL, the page is stored in target->page
4389 * with extra refcnt got(Callers should handle it).
4390 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4391 * target for charge migration. if @target is not NULL, the entry is stored
4394 * Called with pte lock held.
4401 enum mc_target_type {
4407 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4408 unsigned long addr, pte_t ptent)
4410 struct page *page = vm_normal_page(vma, addr, ptent);
4412 if (!page || !page_mapped(page))
4414 if (PageAnon(page)) {
4415 if (!(mc.flags & MOVE_ANON))
4418 if (!(mc.flags & MOVE_FILE))
4421 if (!get_page_unless_zero(page))
4428 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4429 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4431 struct page *page = NULL;
4432 swp_entry_t ent = pte_to_swp_entry(ptent);
4434 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4437 * Because lookup_swap_cache() updates some statistics counter,
4438 * we call find_get_page() with swapper_space directly.
4440 page = find_get_page(swap_address_space(ent), ent.val);
4441 if (do_memsw_account())
4442 entry->val = ent.val;
4447 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4448 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4454 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4455 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4457 struct page *page = NULL;
4458 struct address_space *mapping;
4461 if (!vma->vm_file) /* anonymous vma */
4463 if (!(mc.flags & MOVE_FILE))
4466 mapping = vma->vm_file->f_mapping;
4467 pgoff = linear_page_index(vma, addr);
4469 /* page is moved even if it's not RSS of this task(page-faulted). */
4471 /* shmem/tmpfs may report page out on swap: account for that too. */
4472 if (shmem_mapping(mapping)) {
4473 page = find_get_entry(mapping, pgoff);
4474 if (radix_tree_exceptional_entry(page)) {
4475 swp_entry_t swp = radix_to_swp_entry(page);
4476 if (do_memsw_account())
4478 page = find_get_page(swap_address_space(swp), swp.val);
4481 page = find_get_page(mapping, pgoff);
4483 page = find_get_page(mapping, pgoff);
4489 * mem_cgroup_move_account - move account of the page
4491 * @nr_pages: number of regular pages (>1 for huge pages)
4492 * @from: mem_cgroup which the page is moved from.
4493 * @to: mem_cgroup which the page is moved to. @from != @to.
4495 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4497 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4500 static int mem_cgroup_move_account(struct page *page,
4502 struct mem_cgroup *from,
4503 struct mem_cgroup *to)
4505 unsigned long flags;
4506 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4510 VM_BUG_ON(from == to);
4511 VM_BUG_ON_PAGE(PageLRU(page), page);
4512 VM_BUG_ON(compound && !PageTransHuge(page));
4515 * Prevent mem_cgroup_replace_page() from looking at
4516 * page->mem_cgroup of its source page while we change it.
4519 if (!trylock_page(page))
4523 if (page->mem_cgroup != from)
4526 anon = PageAnon(page);
4528 spin_lock_irqsave(&from->move_lock, flags);
4530 if (!anon && page_mapped(page)) {
4531 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4533 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4538 * move_lock grabbed above and caller set from->moving_account, so
4539 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4540 * So mapping should be stable for dirty pages.
4542 if (!anon && PageDirty(page)) {
4543 struct address_space *mapping = page_mapping(page);
4545 if (mapping_cap_account_dirty(mapping)) {
4546 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4548 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4553 if (PageWriteback(page)) {
4554 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4556 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4561 * It is safe to change page->mem_cgroup here because the page
4562 * is referenced, charged, and isolated - we can't race with
4563 * uncharging, charging, migration, or LRU putback.
4566 /* caller should have done css_get */
4567 page->mem_cgroup = to;
4568 spin_unlock_irqrestore(&from->move_lock, flags);
4572 local_irq_disable();
4573 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4574 memcg_check_events(to, page);
4575 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4576 memcg_check_events(from, page);
4584 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4585 unsigned long addr, pte_t ptent, union mc_target *target)
4587 struct page *page = NULL;
4588 enum mc_target_type ret = MC_TARGET_NONE;
4589 swp_entry_t ent = { .val = 0 };
4591 if (pte_present(ptent))
4592 page = mc_handle_present_pte(vma, addr, ptent);
4593 else if (is_swap_pte(ptent))
4594 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4595 else if (pte_none(ptent))
4596 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4598 if (!page && !ent.val)
4602 * Do only loose check w/o serialization.
4603 * mem_cgroup_move_account() checks the page is valid or
4604 * not under LRU exclusion.
4606 if (page->mem_cgroup == mc.from) {
4607 ret = MC_TARGET_PAGE;
4609 target->page = page;
4611 if (!ret || !target)
4614 /* There is a swap entry and a page doesn't exist or isn't charged */
4615 if (ent.val && !ret &&
4616 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4617 ret = MC_TARGET_SWAP;
4624 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4626 * We don't consider swapping or file mapped pages because THP does not
4627 * support them for now.
4628 * Caller should make sure that pmd_trans_huge(pmd) is true.
4630 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4631 unsigned long addr, pmd_t pmd, union mc_target *target)
4633 struct page *page = NULL;
4634 enum mc_target_type ret = MC_TARGET_NONE;
4636 page = pmd_page(pmd);
4637 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4638 if (!(mc.flags & MOVE_ANON))
4640 if (page->mem_cgroup == mc.from) {
4641 ret = MC_TARGET_PAGE;
4644 target->page = page;
4650 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4651 unsigned long addr, pmd_t pmd, union mc_target *target)
4653 return MC_TARGET_NONE;
4657 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4658 unsigned long addr, unsigned long end,
4659 struct mm_walk *walk)
4661 struct vm_area_struct *vma = walk->vma;
4665 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4666 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4667 mc.precharge += HPAGE_PMD_NR;
4672 if (pmd_trans_unstable(pmd))
4674 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4675 for (; addr != end; pte++, addr += PAGE_SIZE)
4676 if (get_mctgt_type(vma, addr, *pte, NULL))
4677 mc.precharge++; /* increment precharge temporarily */
4678 pte_unmap_unlock(pte - 1, ptl);
4684 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4686 unsigned long precharge;
4688 struct mm_walk mem_cgroup_count_precharge_walk = {
4689 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4692 down_read(&mm->mmap_sem);
4693 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4694 up_read(&mm->mmap_sem);
4696 precharge = mc.precharge;
4702 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4704 unsigned long precharge = mem_cgroup_count_precharge(mm);
4706 VM_BUG_ON(mc.moving_task);
4707 mc.moving_task = current;
4708 return mem_cgroup_do_precharge(precharge);
4711 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4712 static void __mem_cgroup_clear_mc(void)
4714 struct mem_cgroup *from = mc.from;
4715 struct mem_cgroup *to = mc.to;
4717 /* we must uncharge all the leftover precharges from mc.to */
4719 cancel_charge(mc.to, mc.precharge);
4723 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4724 * we must uncharge here.
4726 if (mc.moved_charge) {
4727 cancel_charge(mc.from, mc.moved_charge);
4728 mc.moved_charge = 0;
4730 /* we must fixup refcnts and charges */
4731 if (mc.moved_swap) {
4732 /* uncharge swap account from the old cgroup */
4733 if (!mem_cgroup_is_root(mc.from))
4734 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4737 * we charged both to->memory and to->memsw, so we
4738 * should uncharge to->memory.
4740 if (!mem_cgroup_is_root(mc.to))
4741 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4743 css_put_many(&mc.from->css, mc.moved_swap);
4745 /* we've already done css_get(mc.to) */
4748 memcg_oom_recover(from);
4749 memcg_oom_recover(to);
4750 wake_up_all(&mc.waitq);
4753 static void mem_cgroup_clear_mc(void)
4756 * we must clear moving_task before waking up waiters at the end of
4759 mc.moving_task = NULL;
4760 __mem_cgroup_clear_mc();
4761 spin_lock(&mc.lock);
4764 spin_unlock(&mc.lock);
4767 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4769 struct cgroup_subsys_state *css;
4770 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4771 struct mem_cgroup *from;
4772 struct task_struct *leader, *p;
4773 struct mm_struct *mm;
4774 unsigned long move_flags;
4777 /* charge immigration isn't supported on the default hierarchy */
4778 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4782 * Multi-process migrations only happen on the default hierarchy
4783 * where charge immigration is not used. Perform charge
4784 * immigration if @tset contains a leader and whine if there are
4788 cgroup_taskset_for_each_leader(leader, css, tset) {
4791 memcg = mem_cgroup_from_css(css);
4797 * We are now commited to this value whatever it is. Changes in this
4798 * tunable will only affect upcoming migrations, not the current one.
4799 * So we need to save it, and keep it going.
4801 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4805 from = mem_cgroup_from_task(p);
4807 VM_BUG_ON(from == memcg);
4809 mm = get_task_mm(p);
4812 /* We move charges only when we move a owner of the mm */
4813 if (mm->owner == p) {
4816 VM_BUG_ON(mc.precharge);
4817 VM_BUG_ON(mc.moved_charge);
4818 VM_BUG_ON(mc.moved_swap);
4820 spin_lock(&mc.lock);
4823 mc.flags = move_flags;
4824 spin_unlock(&mc.lock);
4825 /* We set mc.moving_task later */
4827 ret = mem_cgroup_precharge_mc(mm);
4829 mem_cgroup_clear_mc();
4835 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4838 mem_cgroup_clear_mc();
4841 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4842 unsigned long addr, unsigned long end,
4843 struct mm_walk *walk)
4846 struct vm_area_struct *vma = walk->vma;
4849 enum mc_target_type target_type;
4850 union mc_target target;
4853 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4854 if (mc.precharge < HPAGE_PMD_NR) {
4858 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4859 if (target_type == MC_TARGET_PAGE) {
4861 if (!isolate_lru_page(page)) {
4862 if (!mem_cgroup_move_account(page, true,
4864 mc.precharge -= HPAGE_PMD_NR;
4865 mc.moved_charge += HPAGE_PMD_NR;
4867 putback_lru_page(page);
4875 if (pmd_trans_unstable(pmd))
4878 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4879 for (; addr != end; addr += PAGE_SIZE) {
4880 pte_t ptent = *(pte++);
4886 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4887 case MC_TARGET_PAGE:
4890 * We can have a part of the split pmd here. Moving it
4891 * can be done but it would be too convoluted so simply
4892 * ignore such a partial THP and keep it in original
4893 * memcg. There should be somebody mapping the head.
4895 if (PageTransCompound(page))
4897 if (isolate_lru_page(page))
4899 if (!mem_cgroup_move_account(page, false,
4902 /* we uncharge from mc.from later. */
4905 putback_lru_page(page);
4906 put: /* get_mctgt_type() gets the page */
4909 case MC_TARGET_SWAP:
4911 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4913 /* we fixup refcnts and charges later. */
4921 pte_unmap_unlock(pte - 1, ptl);
4926 * We have consumed all precharges we got in can_attach().
4927 * We try charge one by one, but don't do any additional
4928 * charges to mc.to if we have failed in charge once in attach()
4931 ret = mem_cgroup_do_precharge(1);
4939 static void mem_cgroup_move_charge(struct mm_struct *mm)
4941 struct mm_walk mem_cgroup_move_charge_walk = {
4942 .pmd_entry = mem_cgroup_move_charge_pte_range,
4946 lru_add_drain_all();
4948 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4949 * move_lock while we're moving its pages to another memcg.
4950 * Then wait for already started RCU-only updates to finish.
4952 atomic_inc(&mc.from->moving_account);
4955 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4957 * Someone who are holding the mmap_sem might be waiting in
4958 * waitq. So we cancel all extra charges, wake up all waiters,
4959 * and retry. Because we cancel precharges, we might not be able
4960 * to move enough charges, but moving charge is a best-effort
4961 * feature anyway, so it wouldn't be a big problem.
4963 __mem_cgroup_clear_mc();
4968 * When we have consumed all precharges and failed in doing
4969 * additional charge, the page walk just aborts.
4971 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4972 up_read(&mm->mmap_sem);
4973 atomic_dec(&mc.from->moving_account);
4976 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4978 struct cgroup_subsys_state *css;
4979 struct task_struct *p = cgroup_taskset_first(tset, &css);
4980 struct mm_struct *mm = get_task_mm(p);
4984 mem_cgroup_move_charge(mm);
4988 mem_cgroup_clear_mc();
4990 #else /* !CONFIG_MMU */
4991 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4995 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4998 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5004 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5005 * to verify whether we're attached to the default hierarchy on each mount
5008 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5011 * use_hierarchy is forced on the default hierarchy. cgroup core
5012 * guarantees that @root doesn't have any children, so turning it
5013 * on for the root memcg is enough.
5015 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5016 root_mem_cgroup->use_hierarchy = true;
5018 root_mem_cgroup->use_hierarchy = false;
5021 static u64 memory_current_read(struct cgroup_subsys_state *css,
5024 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5026 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5029 static int memory_low_show(struct seq_file *m, void *v)
5031 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5032 unsigned long low = READ_ONCE(memcg->low);
5034 if (low == PAGE_COUNTER_MAX)
5035 seq_puts(m, "max\n");
5037 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5042 static ssize_t memory_low_write(struct kernfs_open_file *of,
5043 char *buf, size_t nbytes, loff_t off)
5045 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5049 buf = strstrip(buf);
5050 err = page_counter_memparse(buf, "max", &low);
5059 static int memory_high_show(struct seq_file *m, void *v)
5061 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5062 unsigned long high = READ_ONCE(memcg->high);
5064 if (high == PAGE_COUNTER_MAX)
5065 seq_puts(m, "max\n");
5067 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5072 static ssize_t memory_high_write(struct kernfs_open_file *of,
5073 char *buf, size_t nbytes, loff_t off)
5075 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5079 buf = strstrip(buf);
5080 err = page_counter_memparse(buf, "max", &high);
5086 memcg_wb_domain_size_changed(memcg);
5090 static int memory_max_show(struct seq_file *m, void *v)
5092 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5093 unsigned long max = READ_ONCE(memcg->memory.limit);
5095 if (max == PAGE_COUNTER_MAX)
5096 seq_puts(m, "max\n");
5098 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5103 static ssize_t memory_max_write(struct kernfs_open_file *of,
5104 char *buf, size_t nbytes, loff_t off)
5106 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5110 buf = strstrip(buf);
5111 err = page_counter_memparse(buf, "max", &max);
5115 err = mem_cgroup_resize_limit(memcg, max);
5119 memcg_wb_domain_size_changed(memcg);
5123 static int memory_events_show(struct seq_file *m, void *v)
5125 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5127 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5128 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5129 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5130 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5135 static struct cftype memory_files[] = {
5138 .flags = CFTYPE_NOT_ON_ROOT,
5139 .read_u64 = memory_current_read,
5143 .flags = CFTYPE_NOT_ON_ROOT,
5144 .seq_show = memory_low_show,
5145 .write = memory_low_write,
5149 .flags = CFTYPE_NOT_ON_ROOT,
5150 .seq_show = memory_high_show,
5151 .write = memory_high_write,
5155 .flags = CFTYPE_NOT_ON_ROOT,
5156 .seq_show = memory_max_show,
5157 .write = memory_max_write,
5161 .flags = CFTYPE_NOT_ON_ROOT,
5162 .file_offset = offsetof(struct mem_cgroup, events_file),
5163 .seq_show = memory_events_show,
5168 struct cgroup_subsys memory_cgrp_subsys = {
5169 .css_alloc = mem_cgroup_css_alloc,
5170 .css_online = mem_cgroup_css_online,
5171 .css_offline = mem_cgroup_css_offline,
5172 .css_released = mem_cgroup_css_released,
5173 .css_free = mem_cgroup_css_free,
5174 .css_reset = mem_cgroup_css_reset,
5175 .can_attach = mem_cgroup_can_attach,
5176 .cancel_attach = mem_cgroup_cancel_attach,
5177 .attach = mem_cgroup_move_task,
5178 .bind = mem_cgroup_bind,
5179 .dfl_cftypes = memory_files,
5180 .legacy_cftypes = mem_cgroup_legacy_files,
5185 * mem_cgroup_low - check if memory consumption is below the normal range
5186 * @root: the highest ancestor to consider
5187 * @memcg: the memory cgroup to check
5189 * Returns %true if memory consumption of @memcg, and that of all
5190 * configurable ancestors up to @root, is below the normal range.
5192 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5194 if (mem_cgroup_disabled())
5198 * The toplevel group doesn't have a configurable range, so
5199 * it's never low when looked at directly, and it is not
5200 * considered an ancestor when assessing the hierarchy.
5203 if (memcg == root_mem_cgroup)
5206 if (page_counter_read(&memcg->memory) >= memcg->low)
5209 while (memcg != root) {
5210 memcg = parent_mem_cgroup(memcg);
5212 if (memcg == root_mem_cgroup)
5215 if (page_counter_read(&memcg->memory) >= memcg->low)
5222 * mem_cgroup_try_charge - try charging a page
5223 * @page: page to charge
5224 * @mm: mm context of the victim
5225 * @gfp_mask: reclaim mode
5226 * @memcgp: charged memcg return
5228 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5229 * pages according to @gfp_mask if necessary.
5231 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5232 * Otherwise, an error code is returned.
5234 * After page->mapping has been set up, the caller must finalize the
5235 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5236 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5238 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5239 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5242 struct mem_cgroup *memcg = NULL;
5243 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5246 if (mem_cgroup_disabled())
5249 if (PageSwapCache(page)) {
5251 * Every swap fault against a single page tries to charge the
5252 * page, bail as early as possible. shmem_unuse() encounters
5253 * already charged pages, too. The USED bit is protected by
5254 * the page lock, which serializes swap cache removal, which
5255 * in turn serializes uncharging.
5257 VM_BUG_ON_PAGE(!PageLocked(page), page);
5258 if (page->mem_cgroup)
5261 if (do_memsw_account()) {
5262 swp_entry_t ent = { .val = page_private(page), };
5263 unsigned short id = lookup_swap_cgroup_id(ent);
5266 memcg = mem_cgroup_from_id(id);
5267 if (memcg && !css_tryget_online(&memcg->css))
5274 memcg = get_mem_cgroup_from_mm(mm);
5276 ret = try_charge(memcg, gfp_mask, nr_pages);
5278 css_put(&memcg->css);
5285 * mem_cgroup_commit_charge - commit a page charge
5286 * @page: page to charge
5287 * @memcg: memcg to charge the page to
5288 * @lrucare: page might be on LRU already
5290 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5291 * after page->mapping has been set up. This must happen atomically
5292 * as part of the page instantiation, i.e. under the page table lock
5293 * for anonymous pages, under the page lock for page and swap cache.
5295 * In addition, the page must not be on the LRU during the commit, to
5296 * prevent racing with task migration. If it might be, use @lrucare.
5298 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5300 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5301 bool lrucare, bool compound)
5303 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5305 VM_BUG_ON_PAGE(!page->mapping, page);
5306 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5308 if (mem_cgroup_disabled())
5311 * Swap faults will attempt to charge the same page multiple
5312 * times. But reuse_swap_page() might have removed the page
5313 * from swapcache already, so we can't check PageSwapCache().
5318 commit_charge(page, memcg, lrucare);
5320 local_irq_disable();
5321 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5322 memcg_check_events(memcg, page);
5325 if (do_memsw_account() && PageSwapCache(page)) {
5326 swp_entry_t entry = { .val = page_private(page) };
5328 * The swap entry might not get freed for a long time,
5329 * let's not wait for it. The page already received a
5330 * memory+swap charge, drop the swap entry duplicate.
5332 mem_cgroup_uncharge_swap(entry);
5337 * mem_cgroup_cancel_charge - cancel a page charge
5338 * @page: page to charge
5339 * @memcg: memcg to charge the page to
5341 * Cancel a charge transaction started by mem_cgroup_try_charge().
5343 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5346 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5348 if (mem_cgroup_disabled())
5351 * Swap faults will attempt to charge the same page multiple
5352 * times. But reuse_swap_page() might have removed the page
5353 * from swapcache already, so we can't check PageSwapCache().
5358 cancel_charge(memcg, nr_pages);
5361 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5362 unsigned long nr_anon, unsigned long nr_file,
5363 unsigned long nr_huge, struct page *dummy_page)
5365 unsigned long nr_pages = nr_anon + nr_file;
5366 unsigned long flags;
5368 if (!mem_cgroup_is_root(memcg)) {
5369 page_counter_uncharge(&memcg->memory, nr_pages);
5370 if (do_memsw_account())
5371 page_counter_uncharge(&memcg->memsw, nr_pages);
5372 memcg_oom_recover(memcg);
5375 local_irq_save(flags);
5376 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5377 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5378 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5379 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5380 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5381 memcg_check_events(memcg, dummy_page);
5382 local_irq_restore(flags);
5384 if (!mem_cgroup_is_root(memcg))
5385 css_put_many(&memcg->css, nr_pages);
5388 static void uncharge_list(struct list_head *page_list)
5390 struct mem_cgroup *memcg = NULL;
5391 unsigned long nr_anon = 0;
5392 unsigned long nr_file = 0;
5393 unsigned long nr_huge = 0;
5394 unsigned long pgpgout = 0;
5395 struct list_head *next;
5398 next = page_list->next;
5400 unsigned int nr_pages = 1;
5402 page = list_entry(next, struct page, lru);
5403 next = page->lru.next;
5405 VM_BUG_ON_PAGE(PageLRU(page), page);
5406 VM_BUG_ON_PAGE(page_count(page), page);
5408 if (!page->mem_cgroup)
5412 * Nobody should be changing or seriously looking at
5413 * page->mem_cgroup at this point, we have fully
5414 * exclusive access to the page.
5417 if (memcg != page->mem_cgroup) {
5419 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5421 pgpgout = nr_anon = nr_file = nr_huge = 0;
5423 memcg = page->mem_cgroup;
5426 if (PageTransHuge(page)) {
5427 nr_pages <<= compound_order(page);
5428 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5429 nr_huge += nr_pages;
5433 nr_anon += nr_pages;
5435 nr_file += nr_pages;
5437 page->mem_cgroup = NULL;
5440 } while (next != page_list);
5443 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5448 * mem_cgroup_uncharge - uncharge a page
5449 * @page: page to uncharge
5451 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5452 * mem_cgroup_commit_charge().
5454 void mem_cgroup_uncharge(struct page *page)
5456 if (mem_cgroup_disabled())
5459 /* Don't touch page->lru of any random page, pre-check: */
5460 if (!page->mem_cgroup)
5463 INIT_LIST_HEAD(&page->lru);
5464 uncharge_list(&page->lru);
5468 * mem_cgroup_uncharge_list - uncharge a list of page
5469 * @page_list: list of pages to uncharge
5471 * Uncharge a list of pages previously charged with
5472 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5474 void mem_cgroup_uncharge_list(struct list_head *page_list)
5476 if (mem_cgroup_disabled())
5479 if (!list_empty(page_list))
5480 uncharge_list(page_list);
5484 * mem_cgroup_replace_page - migrate a charge to another page
5485 * @oldpage: currently charged page
5486 * @newpage: page to transfer the charge to
5488 * Migrate the charge from @oldpage to @newpage.
5490 * Both pages must be locked, @newpage->mapping must be set up.
5491 * Either or both pages might be on the LRU already.
5493 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5495 struct mem_cgroup *memcg;
5498 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5499 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5500 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5501 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5504 if (mem_cgroup_disabled())
5507 /* Page cache replacement: new page already charged? */
5508 if (newpage->mem_cgroup)
5511 /* Swapcache readahead pages can get replaced before being charged */
5512 memcg = oldpage->mem_cgroup;
5516 lock_page_lru(oldpage, &isolated);
5517 oldpage->mem_cgroup = NULL;
5518 unlock_page_lru(oldpage, isolated);
5520 commit_charge(newpage, memcg, true);
5525 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5526 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5528 void sock_update_memcg(struct sock *sk)
5530 struct mem_cgroup *memcg;
5532 /* Socket cloning can throw us here with sk_cgrp already
5533 * filled. It won't however, necessarily happen from
5534 * process context. So the test for root memcg given
5535 * the current task's memcg won't help us in this case.
5537 * Respecting the original socket's memcg is a better
5538 * decision in this case.
5541 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5542 css_get(&sk->sk_memcg->css);
5547 memcg = mem_cgroup_from_task(current);
5548 if (memcg == root_mem_cgroup)
5550 #ifdef CONFIG_MEMCG_KMEM
5551 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcp_mem.active)
5554 if (css_tryget_online(&memcg->css))
5555 sk->sk_memcg = memcg;
5559 EXPORT_SYMBOL(sock_update_memcg);
5561 void sock_release_memcg(struct sock *sk)
5563 WARN_ON(!sk->sk_memcg);
5564 css_put(&sk->sk_memcg->css);
5568 * mem_cgroup_charge_skmem - charge socket memory
5569 * @memcg: memcg to charge
5570 * @nr_pages: number of pages to charge
5572 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5573 * @memcg's configured limit, %false if the charge had to be forced.
5575 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5577 gfp_t gfp_mask = GFP_KERNEL;
5579 #ifdef CONFIG_MEMCG_KMEM
5580 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5581 struct page_counter *counter;
5583 if (page_counter_try_charge(&memcg->tcp_mem.memory_allocated,
5584 nr_pages, &counter)) {
5585 memcg->tcp_mem.memory_pressure = 0;
5588 page_counter_charge(&memcg->tcp_mem.memory_allocated, nr_pages);
5589 memcg->tcp_mem.memory_pressure = 1;
5593 /* Don't block in the packet receive path */
5595 gfp_mask = GFP_NOWAIT;
5597 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5600 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5605 * mem_cgroup_uncharge_skmem - uncharge socket memory
5606 * @memcg - memcg to uncharge
5607 * @nr_pages - number of pages to uncharge
5609 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5611 #ifdef CONFIG_MEMCG_KMEM
5612 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5613 page_counter_uncharge(&memcg->tcp_mem.memory_allocated,
5618 page_counter_uncharge(&memcg->memory, nr_pages);
5619 css_put_many(&memcg->css, nr_pages);
5622 #endif /* CONFIG_INET */
5624 static int __init cgroup_memory(char *s)
5628 while ((token = strsep(&s, ",")) != NULL) {
5631 if (!strcmp(token, "nosocket"))
5632 cgroup_memory_nosocket = true;
5636 __setup("cgroup.memory=", cgroup_memory);
5639 * subsys_initcall() for memory controller.
5641 * Some parts like hotcpu_notifier() have to be initialized from this context
5642 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5643 * everything that doesn't depend on a specific mem_cgroup structure should
5644 * be initialized from here.
5646 static int __init mem_cgroup_init(void)
5650 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5652 for_each_possible_cpu(cpu)
5653 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5656 for_each_node(node) {
5657 struct mem_cgroup_tree_per_node *rtpn;
5660 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5661 node_online(node) ? node : NUMA_NO_NODE);
5663 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5664 struct mem_cgroup_tree_per_zone *rtpz;
5666 rtpz = &rtpn->rb_tree_per_zone[zone];
5667 rtpz->rb_root = RB_ROOT;
5668 spin_lock_init(&rtpz->lock);
5670 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5675 subsys_initcall(mem_cgroup_init);
5677 #ifdef CONFIG_MEMCG_SWAP
5679 * mem_cgroup_swapout - transfer a memsw charge to swap
5680 * @page: page whose memsw charge to transfer
5681 * @entry: swap entry to move the charge to
5683 * Transfer the memsw charge of @page to @entry.
5685 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5687 struct mem_cgroup *memcg;
5688 unsigned short oldid;
5690 VM_BUG_ON_PAGE(PageLRU(page), page);
5691 VM_BUG_ON_PAGE(page_count(page), page);
5693 if (!do_memsw_account())
5696 memcg = page->mem_cgroup;
5698 /* Readahead page, never charged */
5702 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5703 VM_BUG_ON_PAGE(oldid, page);
5704 mem_cgroup_swap_statistics(memcg, true);
5706 page->mem_cgroup = NULL;
5708 if (!mem_cgroup_is_root(memcg))
5709 page_counter_uncharge(&memcg->memory, 1);
5712 * Interrupts should be disabled here because the caller holds the
5713 * mapping->tree_lock lock which is taken with interrupts-off. It is
5714 * important here to have the interrupts disabled because it is the
5715 * only synchronisation we have for udpating the per-CPU variables.
5717 VM_BUG_ON(!irqs_disabled());
5718 mem_cgroup_charge_statistics(memcg, page, false, -1);
5719 memcg_check_events(memcg, page);
5723 * mem_cgroup_uncharge_swap - uncharge a swap entry
5724 * @entry: swap entry to uncharge
5726 * Drop the memsw charge associated with @entry.
5728 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5730 struct mem_cgroup *memcg;
5733 if (!do_memsw_account())
5736 id = swap_cgroup_record(entry, 0);
5738 memcg = mem_cgroup_from_id(id);
5740 if (!mem_cgroup_is_root(memcg))
5741 page_counter_uncharge(&memcg->memsw, 1);
5742 mem_cgroup_swap_statistics(memcg, false);
5743 css_put(&memcg->css);
5748 /* for remember boot option*/
5749 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5750 static int really_do_swap_account __initdata = 1;
5752 static int really_do_swap_account __initdata;
5755 static int __init enable_swap_account(char *s)
5757 if (!strcmp(s, "1"))
5758 really_do_swap_account = 1;
5759 else if (!strcmp(s, "0"))
5760 really_do_swap_account = 0;
5763 __setup("swapaccount=", enable_swap_account);
5765 static struct cftype memsw_cgroup_files[] = {
5767 .name = "memsw.usage_in_bytes",
5768 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5769 .read_u64 = mem_cgroup_read_u64,
5772 .name = "memsw.max_usage_in_bytes",
5773 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5774 .write = mem_cgroup_reset,
5775 .read_u64 = mem_cgroup_read_u64,
5778 .name = "memsw.limit_in_bytes",
5779 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5780 .write = mem_cgroup_write,
5781 .read_u64 = mem_cgroup_read_u64,
5784 .name = "memsw.failcnt",
5785 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5786 .write = mem_cgroup_reset,
5787 .read_u64 = mem_cgroup_read_u64,
5789 { }, /* terminate */
5792 static int __init mem_cgroup_swap_init(void)
5794 if (!mem_cgroup_disabled() && really_do_swap_account) {
5795 do_swap_account = 1;
5796 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5797 memsw_cgroup_files));
5801 subsys_initcall(mem_cgroup_swap_init);
5803 #endif /* CONFIG_MEMCG_SWAP */