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 /* Kernel memory accounting disabled? */
87 static bool cgroup_memory_nokmem;
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 int do_swap_account __read_mostly;
93 #define do_swap_account 0
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
102 static const char * const mem_cgroup_stat_names[] = {
112 static const char * const mem_cgroup_events_names[] = {
119 static const char * const mem_cgroup_lru_names[] = {
127 #define THRESHOLDS_EVENTS_TARGET 128
128 #define SOFTLIMIT_EVENTS_TARGET 1024
129 #define NUMAINFO_EVENTS_TARGET 1024
132 * Cgroups above their limits are maintained in a RB-Tree, independent of
133 * their hierarchy representation
136 struct mem_cgroup_tree_per_zone {
137 struct rb_root rb_root;
141 struct mem_cgroup_tree_per_node {
142 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
145 struct mem_cgroup_tree {
146 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
149 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
152 struct mem_cgroup_eventfd_list {
153 struct list_head list;
154 struct eventfd_ctx *eventfd;
158 * cgroup_event represents events which userspace want to receive.
160 struct mem_cgroup_event {
162 * memcg which the event belongs to.
164 struct mem_cgroup *memcg;
166 * eventfd to signal userspace about the event.
168 struct eventfd_ctx *eventfd;
170 * Each of these stored in a list by the cgroup.
172 struct list_head list;
174 * register_event() callback will be used to add new userspace
175 * waiter for changes related to this event. Use eventfd_signal()
176 * on eventfd to send notification to userspace.
178 int (*register_event)(struct mem_cgroup *memcg,
179 struct eventfd_ctx *eventfd, const char *args);
181 * unregister_event() callback will be called when userspace closes
182 * the eventfd or on cgroup removing. This callback must be set,
183 * if you want provide notification functionality.
185 void (*unregister_event)(struct mem_cgroup *memcg,
186 struct eventfd_ctx *eventfd);
188 * All fields below needed to unregister event when
189 * userspace closes eventfd.
192 wait_queue_head_t *wqh;
194 struct work_struct remove;
197 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
198 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
200 /* Stuffs for move charges at task migration. */
202 * Types of charges to be moved.
204 #define MOVE_ANON 0x1U
205 #define MOVE_FILE 0x2U
206 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
208 /* "mc" and its members are protected by cgroup_mutex */
209 static struct move_charge_struct {
210 spinlock_t lock; /* for from, to */
211 struct mem_cgroup *from;
212 struct mem_cgroup *to;
214 unsigned long precharge;
215 unsigned long moved_charge;
216 unsigned long moved_swap;
217 struct task_struct *moving_task; /* a task moving charges */
218 wait_queue_head_t waitq; /* a waitq for other context */
220 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
221 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
225 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
226 * limit reclaim to prevent infinite loops, if they ever occur.
228 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
229 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
232 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
233 MEM_CGROUP_CHARGE_TYPE_ANON,
234 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
235 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
239 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
254 * The memcg_create_mutex will be held whenever a new cgroup is created.
255 * As a consequence, any change that needs to protect against new child cgroups
256 * appearing has to hold it as well.
258 static DEFINE_MUTEX(memcg_create_mutex);
260 /* Some nice accessors for the vmpressure. */
261 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
264 memcg = root_mem_cgroup;
265 return &memcg->vmpressure;
268 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
270 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
273 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
275 return (memcg == root_mem_cgroup);
279 * We restrict the id in the range of [1, 65535], so it can fit into
282 #define MEM_CGROUP_ID_MAX USHRT_MAX
284 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
286 return memcg->css.id;
290 * A helper function to get mem_cgroup from ID. must be called under
291 * rcu_read_lock(). The caller is responsible for calling
292 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
293 * refcnt from swap can be called against removed memcg.)
295 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
297 struct cgroup_subsys_state *css;
299 css = css_from_id(id, &memory_cgrp_subsys);
300 return mem_cgroup_from_css(css);
305 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
306 * The main reason for not using cgroup id for this:
307 * this works better in sparse environments, where we have a lot of memcgs,
308 * but only a few kmem-limited. Or also, if we have, for instance, 200
309 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
310 * 200 entry array for that.
312 * The current size of the caches array is stored in memcg_nr_cache_ids. It
313 * will double each time we have to increase it.
315 static DEFINE_IDA(memcg_cache_ida);
316 int memcg_nr_cache_ids;
318 /* Protects memcg_nr_cache_ids */
319 static DECLARE_RWSEM(memcg_cache_ids_sem);
321 void memcg_get_cache_ids(void)
323 down_read(&memcg_cache_ids_sem);
326 void memcg_put_cache_ids(void)
328 up_read(&memcg_cache_ids_sem);
332 * MIN_SIZE is different than 1, because we would like to avoid going through
333 * the alloc/free process all the time. In a small machine, 4 kmem-limited
334 * cgroups is a reasonable guess. In the future, it could be a parameter or
335 * tunable, but that is strictly not necessary.
337 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
338 * this constant directly from cgroup, but it is understandable that this is
339 * better kept as an internal representation in cgroup.c. In any case, the
340 * cgrp_id space is not getting any smaller, and we don't have to necessarily
341 * increase ours as well if it increases.
343 #define MEMCG_CACHES_MIN_SIZE 4
344 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
347 * A lot of the calls to the cache allocation functions are expected to be
348 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
349 * conditional to this static branch, we'll have to allow modules that does
350 * kmem_cache_alloc and the such to see this symbol as well
352 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
353 EXPORT_SYMBOL(memcg_kmem_enabled_key);
355 #endif /* !CONFIG_SLOB */
357 static struct mem_cgroup_per_zone *
358 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
360 int nid = zone_to_nid(zone);
361 int zid = zone_idx(zone);
363 return &memcg->nodeinfo[nid]->zoneinfo[zid];
367 * mem_cgroup_css_from_page - css of the memcg associated with a page
368 * @page: page of interest
370 * If memcg is bound to the default hierarchy, css of the memcg associated
371 * with @page is returned. The returned css remains associated with @page
372 * until it is released.
374 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
377 * XXX: The above description of behavior on the default hierarchy isn't
378 * strictly true yet as replace_page_cache_page() can modify the
379 * association before @page is released even on the default hierarchy;
380 * however, the current and planned usages don't mix the the two functions
381 * and replace_page_cache_page() will soon be updated to make the invariant
384 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
386 struct mem_cgroup *memcg;
388 memcg = page->mem_cgroup;
390 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
391 memcg = root_mem_cgroup;
397 * page_cgroup_ino - return inode number of the memcg a page is charged to
400 * Look up the closest online ancestor of the memory cgroup @page is charged to
401 * and return its inode number or 0 if @page is not charged to any cgroup. It
402 * is safe to call this function without holding a reference to @page.
404 * Note, this function is inherently racy, because there is nothing to prevent
405 * the cgroup inode from getting torn down and potentially reallocated a moment
406 * after page_cgroup_ino() returns, so it only should be used by callers that
407 * do not care (such as procfs interfaces).
409 ino_t page_cgroup_ino(struct page *page)
411 struct mem_cgroup *memcg;
412 unsigned long ino = 0;
415 memcg = READ_ONCE(page->mem_cgroup);
416 while (memcg && !(memcg->css.flags & CSS_ONLINE))
417 memcg = parent_mem_cgroup(memcg);
419 ino = cgroup_ino(memcg->css.cgroup);
424 static struct mem_cgroup_per_zone *
425 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
427 int nid = page_to_nid(page);
428 int zid = page_zonenum(page);
430 return &memcg->nodeinfo[nid]->zoneinfo[zid];
433 static struct mem_cgroup_tree_per_zone *
434 soft_limit_tree_node_zone(int nid, int zid)
436 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
439 static struct mem_cgroup_tree_per_zone *
440 soft_limit_tree_from_page(struct page *page)
442 int nid = page_to_nid(page);
443 int zid = page_zonenum(page);
445 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
448 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
449 struct mem_cgroup_tree_per_zone *mctz,
450 unsigned long new_usage_in_excess)
452 struct rb_node **p = &mctz->rb_root.rb_node;
453 struct rb_node *parent = NULL;
454 struct mem_cgroup_per_zone *mz_node;
459 mz->usage_in_excess = new_usage_in_excess;
460 if (!mz->usage_in_excess)
464 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
466 if (mz->usage_in_excess < mz_node->usage_in_excess)
469 * We can't avoid mem cgroups that are over their soft
470 * limit by the same amount
472 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
475 rb_link_node(&mz->tree_node, parent, p);
476 rb_insert_color(&mz->tree_node, &mctz->rb_root);
480 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
481 struct mem_cgroup_tree_per_zone *mctz)
485 rb_erase(&mz->tree_node, &mctz->rb_root);
489 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
490 struct mem_cgroup_tree_per_zone *mctz)
494 spin_lock_irqsave(&mctz->lock, flags);
495 __mem_cgroup_remove_exceeded(mz, mctz);
496 spin_unlock_irqrestore(&mctz->lock, flags);
499 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
501 unsigned long nr_pages = page_counter_read(&memcg->memory);
502 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
503 unsigned long excess = 0;
505 if (nr_pages > soft_limit)
506 excess = nr_pages - soft_limit;
511 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
513 unsigned long excess;
514 struct mem_cgroup_per_zone *mz;
515 struct mem_cgroup_tree_per_zone *mctz;
517 mctz = soft_limit_tree_from_page(page);
519 * Necessary to update all ancestors when hierarchy is used.
520 * because their event counter is not touched.
522 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
523 mz = mem_cgroup_page_zoneinfo(memcg, page);
524 excess = soft_limit_excess(memcg);
526 * We have to update the tree if mz is on RB-tree or
527 * mem is over its softlimit.
529 if (excess || mz->on_tree) {
532 spin_lock_irqsave(&mctz->lock, flags);
533 /* if on-tree, remove it */
535 __mem_cgroup_remove_exceeded(mz, mctz);
537 * Insert again. mz->usage_in_excess will be updated.
538 * If excess is 0, no tree ops.
540 __mem_cgroup_insert_exceeded(mz, mctz, excess);
541 spin_unlock_irqrestore(&mctz->lock, flags);
546 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
548 struct mem_cgroup_tree_per_zone *mctz;
549 struct mem_cgroup_per_zone *mz;
553 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
554 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
555 mctz = soft_limit_tree_node_zone(nid, zid);
556 mem_cgroup_remove_exceeded(mz, mctz);
561 static struct mem_cgroup_per_zone *
562 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
564 struct rb_node *rightmost = NULL;
565 struct mem_cgroup_per_zone *mz;
569 rightmost = rb_last(&mctz->rb_root);
571 goto done; /* Nothing to reclaim from */
573 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
575 * Remove the node now but someone else can add it back,
576 * we will to add it back at the end of reclaim to its correct
577 * position in the tree.
579 __mem_cgroup_remove_exceeded(mz, mctz);
580 if (!soft_limit_excess(mz->memcg) ||
581 !css_tryget_online(&mz->memcg->css))
587 static struct mem_cgroup_per_zone *
588 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
590 struct mem_cgroup_per_zone *mz;
592 spin_lock_irq(&mctz->lock);
593 mz = __mem_cgroup_largest_soft_limit_node(mctz);
594 spin_unlock_irq(&mctz->lock);
599 * Return page count for single (non recursive) @memcg.
601 * Implementation Note: reading percpu statistics for memcg.
603 * Both of vmstat[] and percpu_counter has threshold and do periodic
604 * synchronization to implement "quick" read. There are trade-off between
605 * reading cost and precision of value. Then, we may have a chance to implement
606 * a periodic synchronization of counter in memcg's counter.
608 * But this _read() function is used for user interface now. The user accounts
609 * memory usage by memory cgroup and he _always_ requires exact value because
610 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
611 * have to visit all online cpus and make sum. So, for now, unnecessary
612 * synchronization is not implemented. (just implemented for cpu hotplug)
614 * If there are kernel internal actions which can make use of some not-exact
615 * value, and reading all cpu value can be performance bottleneck in some
616 * common workload, threshold and synchronization as vmstat[] should be
620 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
625 /* Per-cpu values can be negative, use a signed accumulator */
626 for_each_possible_cpu(cpu)
627 val += per_cpu(memcg->stat->count[idx], cpu);
629 * Summing races with updates, so val may be negative. Avoid exposing
630 * transient negative values.
637 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
638 enum mem_cgroup_events_index idx)
640 unsigned long val = 0;
643 for_each_possible_cpu(cpu)
644 val += per_cpu(memcg->stat->events[idx], cpu);
648 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
650 bool compound, int nr_pages)
653 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
654 * counted as CACHE even if it's on ANON LRU.
657 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
660 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
664 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
665 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
669 /* pagein of a big page is an event. So, ignore page size */
671 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
673 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
674 nr_pages = -nr_pages; /* for event */
677 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
680 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
682 unsigned int lru_mask)
684 unsigned long nr = 0;
687 VM_BUG_ON((unsigned)nid >= nr_node_ids);
689 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
690 struct mem_cgroup_per_zone *mz;
694 if (!(BIT(lru) & lru_mask))
696 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
697 nr += mz->lru_size[lru];
703 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
704 unsigned int lru_mask)
706 unsigned long nr = 0;
709 for_each_node_state(nid, N_MEMORY)
710 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
714 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
715 enum mem_cgroup_events_target target)
717 unsigned long val, next;
719 val = __this_cpu_read(memcg->stat->nr_page_events);
720 next = __this_cpu_read(memcg->stat->targets[target]);
721 /* from time_after() in jiffies.h */
722 if ((long)next - (long)val < 0) {
724 case MEM_CGROUP_TARGET_THRESH:
725 next = val + THRESHOLDS_EVENTS_TARGET;
727 case MEM_CGROUP_TARGET_SOFTLIMIT:
728 next = val + SOFTLIMIT_EVENTS_TARGET;
730 case MEM_CGROUP_TARGET_NUMAINFO:
731 next = val + NUMAINFO_EVENTS_TARGET;
736 __this_cpu_write(memcg->stat->targets[target], next);
743 * Check events in order.
746 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
748 /* threshold event is triggered in finer grain than soft limit */
749 if (unlikely(mem_cgroup_event_ratelimit(memcg,
750 MEM_CGROUP_TARGET_THRESH))) {
752 bool do_numainfo __maybe_unused;
754 do_softlimit = mem_cgroup_event_ratelimit(memcg,
755 MEM_CGROUP_TARGET_SOFTLIMIT);
757 do_numainfo = mem_cgroup_event_ratelimit(memcg,
758 MEM_CGROUP_TARGET_NUMAINFO);
760 mem_cgroup_threshold(memcg);
761 if (unlikely(do_softlimit))
762 mem_cgroup_update_tree(memcg, page);
764 if (unlikely(do_numainfo))
765 atomic_inc(&memcg->numainfo_events);
770 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
773 * mm_update_next_owner() may clear mm->owner to NULL
774 * if it races with swapoff, page migration, etc.
775 * So this can be called with p == NULL.
780 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
782 EXPORT_SYMBOL(mem_cgroup_from_task);
784 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
786 struct mem_cgroup *memcg = NULL;
791 * Page cache insertions can happen withou an
792 * actual mm context, e.g. during disk probing
793 * on boot, loopback IO, acct() writes etc.
796 memcg = root_mem_cgroup;
798 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
799 if (unlikely(!memcg))
800 memcg = root_mem_cgroup;
802 } while (!css_tryget_online(&memcg->css));
808 * mem_cgroup_iter - iterate over memory cgroup hierarchy
809 * @root: hierarchy root
810 * @prev: previously returned memcg, NULL on first invocation
811 * @reclaim: cookie for shared reclaim walks, NULL for full walks
813 * Returns references to children of the hierarchy below @root, or
814 * @root itself, or %NULL after a full round-trip.
816 * Caller must pass the return value in @prev on subsequent
817 * invocations for reference counting, or use mem_cgroup_iter_break()
818 * to cancel a hierarchy walk before the round-trip is complete.
820 * Reclaimers can specify a zone and a priority level in @reclaim to
821 * divide up the memcgs in the hierarchy among all concurrent
822 * reclaimers operating on the same zone and priority.
824 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
825 struct mem_cgroup *prev,
826 struct mem_cgroup_reclaim_cookie *reclaim)
828 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
829 struct cgroup_subsys_state *css = NULL;
830 struct mem_cgroup *memcg = NULL;
831 struct mem_cgroup *pos = NULL;
833 if (mem_cgroup_disabled())
837 root = root_mem_cgroup;
839 if (prev && !reclaim)
842 if (!root->use_hierarchy && root != root_mem_cgroup) {
851 struct mem_cgroup_per_zone *mz;
853 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
854 iter = &mz->iter[reclaim->priority];
856 if (prev && reclaim->generation != iter->generation)
860 pos = READ_ONCE(iter->position);
861 if (!pos || css_tryget(&pos->css))
864 * css reference reached zero, so iter->position will
865 * be cleared by ->css_released. However, we should not
866 * rely on this happening soon, because ->css_released
867 * is called from a work queue, and by busy-waiting we
868 * might block it. So we clear iter->position right
871 (void)cmpxchg(&iter->position, pos, NULL);
879 css = css_next_descendant_pre(css, &root->css);
882 * Reclaimers share the hierarchy walk, and a
883 * new one might jump in right at the end of
884 * the hierarchy - make sure they see at least
885 * one group and restart from the beginning.
893 * Verify the css and acquire a reference. The root
894 * is provided by the caller, so we know it's alive
895 * and kicking, and don't take an extra reference.
897 memcg = mem_cgroup_from_css(css);
899 if (css == &root->css)
902 if (css_tryget(css)) {
904 * Make sure the memcg is initialized:
905 * mem_cgroup_css_online() orders the the
906 * initialization against setting the flag.
908 if (smp_load_acquire(&memcg->initialized))
919 * The position could have already been updated by a competing
920 * thread, so check that the value hasn't changed since we read
921 * it to avoid reclaiming from the same cgroup twice.
923 (void)cmpxchg(&iter->position, pos, memcg);
931 reclaim->generation = iter->generation;
937 if (prev && prev != root)
944 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
945 * @root: hierarchy root
946 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
948 void mem_cgroup_iter_break(struct mem_cgroup *root,
949 struct mem_cgroup *prev)
952 root = root_mem_cgroup;
953 if (prev && prev != root)
957 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
959 struct mem_cgroup *memcg = dead_memcg;
960 struct mem_cgroup_reclaim_iter *iter;
961 struct mem_cgroup_per_zone *mz;
965 while ((memcg = parent_mem_cgroup(memcg))) {
967 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
968 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
969 for (i = 0; i <= DEF_PRIORITY; i++) {
971 cmpxchg(&iter->position,
980 * Iteration constructs for visiting all cgroups (under a tree). If
981 * loops are exited prematurely (break), mem_cgroup_iter_break() must
982 * be used for reference counting.
984 #define for_each_mem_cgroup_tree(iter, root) \
985 for (iter = mem_cgroup_iter(root, NULL, NULL); \
987 iter = mem_cgroup_iter(root, iter, NULL))
989 #define for_each_mem_cgroup(iter) \
990 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
992 iter = mem_cgroup_iter(NULL, iter, NULL))
995 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
996 * @zone: zone of the wanted lruvec
997 * @memcg: memcg of the wanted lruvec
999 * Returns the lru list vector holding pages for the given @zone and
1000 * @mem. This can be the global zone lruvec, if the memory controller
1003 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1004 struct mem_cgroup *memcg)
1006 struct mem_cgroup_per_zone *mz;
1007 struct lruvec *lruvec;
1009 if (mem_cgroup_disabled()) {
1010 lruvec = &zone->lruvec;
1014 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1015 lruvec = &mz->lruvec;
1018 * Since a node can be onlined after the mem_cgroup was created,
1019 * we have to be prepared to initialize lruvec->zone here;
1020 * and if offlined then reonlined, we need to reinitialize it.
1022 if (unlikely(lruvec->zone != zone))
1023 lruvec->zone = zone;
1028 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1030 * @zone: zone of the page
1032 * This function is only safe when following the LRU page isolation
1033 * and putback protocol: the LRU lock must be held, and the page must
1034 * either be PageLRU() or the caller must have isolated/allocated it.
1036 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1038 struct mem_cgroup_per_zone *mz;
1039 struct mem_cgroup *memcg;
1040 struct lruvec *lruvec;
1042 if (mem_cgroup_disabled()) {
1043 lruvec = &zone->lruvec;
1047 memcg = page->mem_cgroup;
1049 * Swapcache readahead pages are added to the LRU - and
1050 * possibly migrated - before they are charged.
1053 memcg = root_mem_cgroup;
1055 mz = mem_cgroup_page_zoneinfo(memcg, page);
1056 lruvec = &mz->lruvec;
1059 * Since a node can be onlined after the mem_cgroup was created,
1060 * we have to be prepared to initialize lruvec->zone here;
1061 * and if offlined then reonlined, we need to reinitialize it.
1063 if (unlikely(lruvec->zone != zone))
1064 lruvec->zone = zone;
1069 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1070 * @lruvec: mem_cgroup per zone lru vector
1071 * @lru: index of lru list the page is sitting on
1072 * @nr_pages: positive when adding or negative when removing
1074 * This function must be called when a page is added to or removed from an
1077 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1080 struct mem_cgroup_per_zone *mz;
1081 unsigned long *lru_size;
1083 if (mem_cgroup_disabled())
1086 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1087 lru_size = mz->lru_size + lru;
1088 *lru_size += nr_pages;
1089 VM_BUG_ON((long)(*lru_size) < 0);
1092 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1094 struct mem_cgroup *task_memcg;
1095 struct task_struct *p;
1098 p = find_lock_task_mm(task);
1100 task_memcg = get_mem_cgroup_from_mm(p->mm);
1104 * All threads may have already detached their mm's, but the oom
1105 * killer still needs to detect if they have already been oom
1106 * killed to prevent needlessly killing additional tasks.
1109 task_memcg = mem_cgroup_from_task(task);
1110 css_get(&task_memcg->css);
1113 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1114 css_put(&task_memcg->css);
1119 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1120 * @memcg: the memory cgroup
1122 * Returns the maximum amount of memory @mem can be charged with, in
1125 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1127 unsigned long margin = 0;
1128 unsigned long count;
1129 unsigned long limit;
1131 count = page_counter_read(&memcg->memory);
1132 limit = READ_ONCE(memcg->memory.limit);
1134 margin = limit - count;
1136 if (do_memsw_account()) {
1137 count = page_counter_read(&memcg->memsw);
1138 limit = READ_ONCE(memcg->memsw.limit);
1140 margin = min(margin, limit - count);
1147 * A routine for checking "mem" is under move_account() or not.
1149 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1150 * moving cgroups. This is for waiting at high-memory pressure
1153 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1155 struct mem_cgroup *from;
1156 struct mem_cgroup *to;
1159 * Unlike task_move routines, we access mc.to, mc.from not under
1160 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1162 spin_lock(&mc.lock);
1168 ret = mem_cgroup_is_descendant(from, memcg) ||
1169 mem_cgroup_is_descendant(to, memcg);
1171 spin_unlock(&mc.lock);
1175 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1177 if (mc.moving_task && current != mc.moving_task) {
1178 if (mem_cgroup_under_move(memcg)) {
1180 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1181 /* moving charge context might have finished. */
1184 finish_wait(&mc.waitq, &wait);
1191 #define K(x) ((x) << (PAGE_SHIFT-10))
1193 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1194 * @memcg: The memory cgroup that went over limit
1195 * @p: Task that is going to be killed
1197 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1200 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1202 /* oom_info_lock ensures that parallel ooms do not interleave */
1203 static DEFINE_MUTEX(oom_info_lock);
1204 struct mem_cgroup *iter;
1207 mutex_lock(&oom_info_lock);
1211 pr_info("Task in ");
1212 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1213 pr_cont(" killed as a result of limit of ");
1215 pr_info("Memory limit reached of cgroup ");
1218 pr_cont_cgroup_path(memcg->css.cgroup);
1223 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1224 K((u64)page_counter_read(&memcg->memory)),
1225 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1226 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1227 K((u64)page_counter_read(&memcg->memsw)),
1228 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1229 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1230 K((u64)page_counter_read(&memcg->kmem)),
1231 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1233 for_each_mem_cgroup_tree(iter, memcg) {
1234 pr_info("Memory cgroup stats for ");
1235 pr_cont_cgroup_path(iter->css.cgroup);
1238 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1239 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
1241 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1242 K(mem_cgroup_read_stat(iter, i)));
1245 for (i = 0; i < NR_LRU_LISTS; i++)
1246 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1247 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1251 mutex_unlock(&oom_info_lock);
1255 * This function returns the number of memcg under hierarchy tree. Returns
1256 * 1(self count) if no children.
1258 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1261 struct mem_cgroup *iter;
1263 for_each_mem_cgroup_tree(iter, memcg)
1269 * Return the memory (and swap, if configured) limit for a memcg.
1271 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1273 unsigned long limit;
1275 limit = memcg->memory.limit;
1276 if (mem_cgroup_swappiness(memcg)) {
1277 unsigned long memsw_limit;
1279 memsw_limit = memcg->memsw.limit;
1280 limit = min(limit + total_swap_pages, memsw_limit);
1285 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1288 struct oom_control oc = {
1291 .gfp_mask = gfp_mask,
1294 struct mem_cgroup *iter;
1295 unsigned long chosen_points = 0;
1296 unsigned long totalpages;
1297 unsigned int points = 0;
1298 struct task_struct *chosen = NULL;
1300 mutex_lock(&oom_lock);
1303 * If current has a pending SIGKILL or is exiting, then automatically
1304 * select it. The goal is to allow it to allocate so that it may
1305 * quickly exit and free its memory.
1307 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1308 mark_oom_victim(current);
1312 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1313 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1314 for_each_mem_cgroup_tree(iter, memcg) {
1315 struct css_task_iter it;
1316 struct task_struct *task;
1318 css_task_iter_start(&iter->css, &it);
1319 while ((task = css_task_iter_next(&it))) {
1320 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1321 case OOM_SCAN_SELECT:
1323 put_task_struct(chosen);
1325 chosen_points = ULONG_MAX;
1326 get_task_struct(chosen);
1328 case OOM_SCAN_CONTINUE:
1330 case OOM_SCAN_ABORT:
1331 css_task_iter_end(&it);
1332 mem_cgroup_iter_break(memcg, iter);
1334 put_task_struct(chosen);
1339 points = oom_badness(task, memcg, NULL, totalpages);
1340 if (!points || points < chosen_points)
1342 /* Prefer thread group leaders for display purposes */
1343 if (points == chosen_points &&
1344 thread_group_leader(chosen))
1348 put_task_struct(chosen);
1350 chosen_points = points;
1351 get_task_struct(chosen);
1353 css_task_iter_end(&it);
1357 points = chosen_points * 1000 / totalpages;
1358 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1359 "Memory cgroup out of memory");
1362 mutex_unlock(&oom_lock);
1365 #if MAX_NUMNODES > 1
1368 * test_mem_cgroup_node_reclaimable
1369 * @memcg: the target memcg
1370 * @nid: the node ID to be checked.
1371 * @noswap : specify true here if the user wants flle only information.
1373 * This function returns whether the specified memcg contains any
1374 * reclaimable pages on a node. Returns true if there are any reclaimable
1375 * pages in the node.
1377 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1378 int nid, bool noswap)
1380 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1382 if (noswap || !total_swap_pages)
1384 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1391 * Always updating the nodemask is not very good - even if we have an empty
1392 * list or the wrong list here, we can start from some node and traverse all
1393 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1396 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1400 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1401 * pagein/pageout changes since the last update.
1403 if (!atomic_read(&memcg->numainfo_events))
1405 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1408 /* make a nodemask where this memcg uses memory from */
1409 memcg->scan_nodes = node_states[N_MEMORY];
1411 for_each_node_mask(nid, node_states[N_MEMORY]) {
1413 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1414 node_clear(nid, memcg->scan_nodes);
1417 atomic_set(&memcg->numainfo_events, 0);
1418 atomic_set(&memcg->numainfo_updating, 0);
1422 * Selecting a node where we start reclaim from. Because what we need is just
1423 * reducing usage counter, start from anywhere is O,K. Considering
1424 * memory reclaim from current node, there are pros. and cons.
1426 * Freeing memory from current node means freeing memory from a node which
1427 * we'll use or we've used. So, it may make LRU bad. And if several threads
1428 * hit limits, it will see a contention on a node. But freeing from remote
1429 * node means more costs for memory reclaim because of memory latency.
1431 * Now, we use round-robin. Better algorithm is welcomed.
1433 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1437 mem_cgroup_may_update_nodemask(memcg);
1438 node = memcg->last_scanned_node;
1440 node = next_node(node, memcg->scan_nodes);
1441 if (node == MAX_NUMNODES)
1442 node = first_node(memcg->scan_nodes);
1444 * We call this when we hit limit, not when pages are added to LRU.
1445 * No LRU may hold pages because all pages are UNEVICTABLE or
1446 * memcg is too small and all pages are not on LRU. In that case,
1447 * we use curret node.
1449 if (unlikely(node == MAX_NUMNODES))
1450 node = numa_node_id();
1452 memcg->last_scanned_node = node;
1456 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1462 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1465 unsigned long *total_scanned)
1467 struct mem_cgroup *victim = NULL;
1470 unsigned long excess;
1471 unsigned long nr_scanned;
1472 struct mem_cgroup_reclaim_cookie reclaim = {
1477 excess = soft_limit_excess(root_memcg);
1480 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1485 * If we have not been able to reclaim
1486 * anything, it might because there are
1487 * no reclaimable pages under this hierarchy
1492 * We want to do more targeted reclaim.
1493 * excess >> 2 is not to excessive so as to
1494 * reclaim too much, nor too less that we keep
1495 * coming back to reclaim from this cgroup
1497 if (total >= (excess >> 2) ||
1498 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1503 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1505 *total_scanned += nr_scanned;
1506 if (!soft_limit_excess(root_memcg))
1509 mem_cgroup_iter_break(root_memcg, victim);
1513 #ifdef CONFIG_LOCKDEP
1514 static struct lockdep_map memcg_oom_lock_dep_map = {
1515 .name = "memcg_oom_lock",
1519 static DEFINE_SPINLOCK(memcg_oom_lock);
1522 * Check OOM-Killer is already running under our hierarchy.
1523 * If someone is running, return false.
1525 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1527 struct mem_cgroup *iter, *failed = NULL;
1529 spin_lock(&memcg_oom_lock);
1531 for_each_mem_cgroup_tree(iter, memcg) {
1532 if (iter->oom_lock) {
1534 * this subtree of our hierarchy is already locked
1535 * so we cannot give a lock.
1538 mem_cgroup_iter_break(memcg, iter);
1541 iter->oom_lock = true;
1546 * OK, we failed to lock the whole subtree so we have
1547 * to clean up what we set up to the failing subtree
1549 for_each_mem_cgroup_tree(iter, memcg) {
1550 if (iter == failed) {
1551 mem_cgroup_iter_break(memcg, iter);
1554 iter->oom_lock = false;
1557 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1559 spin_unlock(&memcg_oom_lock);
1564 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1566 struct mem_cgroup *iter;
1568 spin_lock(&memcg_oom_lock);
1569 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1570 for_each_mem_cgroup_tree(iter, memcg)
1571 iter->oom_lock = false;
1572 spin_unlock(&memcg_oom_lock);
1575 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1577 struct mem_cgroup *iter;
1579 spin_lock(&memcg_oom_lock);
1580 for_each_mem_cgroup_tree(iter, memcg)
1582 spin_unlock(&memcg_oom_lock);
1585 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1587 struct mem_cgroup *iter;
1590 * When a new child is created while the hierarchy is under oom,
1591 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1593 spin_lock(&memcg_oom_lock);
1594 for_each_mem_cgroup_tree(iter, memcg)
1595 if (iter->under_oom > 0)
1597 spin_unlock(&memcg_oom_lock);
1600 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1602 struct oom_wait_info {
1603 struct mem_cgroup *memcg;
1607 static int memcg_oom_wake_function(wait_queue_t *wait,
1608 unsigned mode, int sync, void *arg)
1610 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1611 struct mem_cgroup *oom_wait_memcg;
1612 struct oom_wait_info *oom_wait_info;
1614 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1615 oom_wait_memcg = oom_wait_info->memcg;
1617 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1618 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1620 return autoremove_wake_function(wait, mode, sync, arg);
1623 static void memcg_oom_recover(struct mem_cgroup *memcg)
1626 * For the following lockless ->under_oom test, the only required
1627 * guarantee is that it must see the state asserted by an OOM when
1628 * this function is called as a result of userland actions
1629 * triggered by the notification of the OOM. This is trivially
1630 * achieved by invoking mem_cgroup_mark_under_oom() before
1631 * triggering notification.
1633 if (memcg && memcg->under_oom)
1634 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1637 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1639 if (!current->memcg_may_oom)
1642 * We are in the middle of the charge context here, so we
1643 * don't want to block when potentially sitting on a callstack
1644 * that holds all kinds of filesystem and mm locks.
1646 * Also, the caller may handle a failed allocation gracefully
1647 * (like optional page cache readahead) and so an OOM killer
1648 * invocation might not even be necessary.
1650 * That's why we don't do anything here except remember the
1651 * OOM context and then deal with it at the end of the page
1652 * fault when the stack is unwound, the locks are released,
1653 * and when we know whether the fault was overall successful.
1655 css_get(&memcg->css);
1656 current->memcg_in_oom = memcg;
1657 current->memcg_oom_gfp_mask = mask;
1658 current->memcg_oom_order = order;
1662 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1663 * @handle: actually kill/wait or just clean up the OOM state
1665 * This has to be called at the end of a page fault if the memcg OOM
1666 * handler was enabled.
1668 * Memcg supports userspace OOM handling where failed allocations must
1669 * sleep on a waitqueue until the userspace task resolves the
1670 * situation. Sleeping directly in the charge context with all kinds
1671 * of locks held is not a good idea, instead we remember an OOM state
1672 * in the task and mem_cgroup_oom_synchronize() has to be called at
1673 * the end of the page fault to complete the OOM handling.
1675 * Returns %true if an ongoing memcg OOM situation was detected and
1676 * completed, %false otherwise.
1678 bool mem_cgroup_oom_synchronize(bool handle)
1680 struct mem_cgroup *memcg = current->memcg_in_oom;
1681 struct oom_wait_info owait;
1684 /* OOM is global, do not handle */
1688 if (!handle || oom_killer_disabled)
1691 owait.memcg = memcg;
1692 owait.wait.flags = 0;
1693 owait.wait.func = memcg_oom_wake_function;
1694 owait.wait.private = current;
1695 INIT_LIST_HEAD(&owait.wait.task_list);
1697 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1698 mem_cgroup_mark_under_oom(memcg);
1700 locked = mem_cgroup_oom_trylock(memcg);
1703 mem_cgroup_oom_notify(memcg);
1705 if (locked && !memcg->oom_kill_disable) {
1706 mem_cgroup_unmark_under_oom(memcg);
1707 finish_wait(&memcg_oom_waitq, &owait.wait);
1708 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1709 current->memcg_oom_order);
1712 mem_cgroup_unmark_under_oom(memcg);
1713 finish_wait(&memcg_oom_waitq, &owait.wait);
1717 mem_cgroup_oom_unlock(memcg);
1719 * There is no guarantee that an OOM-lock contender
1720 * sees the wakeups triggered by the OOM kill
1721 * uncharges. Wake any sleepers explicitely.
1723 memcg_oom_recover(memcg);
1726 current->memcg_in_oom = NULL;
1727 css_put(&memcg->css);
1732 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1733 * @page: page that is going to change accounted state
1735 * This function must mark the beginning of an accounted page state
1736 * change to prevent double accounting when the page is concurrently
1737 * being moved to another memcg:
1739 * memcg = mem_cgroup_begin_page_stat(page);
1740 * if (TestClearPageState(page))
1741 * mem_cgroup_update_page_stat(memcg, state, -1);
1742 * mem_cgroup_end_page_stat(memcg);
1744 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1746 struct mem_cgroup *memcg;
1747 unsigned long flags;
1750 * The RCU lock is held throughout the transaction. The fast
1751 * path can get away without acquiring the memcg->move_lock
1752 * because page moving starts with an RCU grace period.
1754 * The RCU lock also protects the memcg from being freed when
1755 * the page state that is going to change is the only thing
1756 * preventing the page from being uncharged.
1757 * E.g. end-writeback clearing PageWriteback(), which allows
1758 * migration to go ahead and uncharge the page before the
1759 * account transaction might be complete.
1763 if (mem_cgroup_disabled())
1766 memcg = page->mem_cgroup;
1767 if (unlikely(!memcg))
1770 if (atomic_read(&memcg->moving_account) <= 0)
1773 spin_lock_irqsave(&memcg->move_lock, flags);
1774 if (memcg != page->mem_cgroup) {
1775 spin_unlock_irqrestore(&memcg->move_lock, flags);
1780 * When charge migration first begins, we can have locked and
1781 * unlocked page stat updates happening concurrently. Track
1782 * the task who has the lock for mem_cgroup_end_page_stat().
1784 memcg->move_lock_task = current;
1785 memcg->move_lock_flags = flags;
1789 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1792 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1793 * @memcg: the memcg that was accounted against
1795 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1797 if (memcg && memcg->move_lock_task == current) {
1798 unsigned long flags = memcg->move_lock_flags;
1800 memcg->move_lock_task = NULL;
1801 memcg->move_lock_flags = 0;
1803 spin_unlock_irqrestore(&memcg->move_lock, flags);
1808 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1811 * size of first charge trial. "32" comes from vmscan.c's magic value.
1812 * TODO: maybe necessary to use big numbers in big irons.
1814 #define CHARGE_BATCH 32U
1815 struct memcg_stock_pcp {
1816 struct mem_cgroup *cached; /* this never be root cgroup */
1817 unsigned int nr_pages;
1818 struct work_struct work;
1819 unsigned long flags;
1820 #define FLUSHING_CACHED_CHARGE 0
1822 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1823 static DEFINE_MUTEX(percpu_charge_mutex);
1826 * consume_stock: Try to consume stocked charge on this cpu.
1827 * @memcg: memcg to consume from.
1828 * @nr_pages: how many pages to charge.
1830 * The charges will only happen if @memcg matches the current cpu's memcg
1831 * stock, and at least @nr_pages are available in that stock. Failure to
1832 * service an allocation will refill the stock.
1834 * returns true if successful, false otherwise.
1836 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1838 struct memcg_stock_pcp *stock;
1841 if (nr_pages > CHARGE_BATCH)
1844 stock = &get_cpu_var(memcg_stock);
1845 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1846 stock->nr_pages -= nr_pages;
1849 put_cpu_var(memcg_stock);
1854 * Returns stocks cached in percpu and reset cached information.
1856 static void drain_stock(struct memcg_stock_pcp *stock)
1858 struct mem_cgroup *old = stock->cached;
1860 if (stock->nr_pages) {
1861 page_counter_uncharge(&old->memory, stock->nr_pages);
1862 if (do_memsw_account())
1863 page_counter_uncharge(&old->memsw, stock->nr_pages);
1864 css_put_many(&old->css, stock->nr_pages);
1865 stock->nr_pages = 0;
1867 stock->cached = NULL;
1871 * This must be called under preempt disabled or must be called by
1872 * a thread which is pinned to local cpu.
1874 static void drain_local_stock(struct work_struct *dummy)
1876 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1878 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1882 * Cache charges(val) to local per_cpu area.
1883 * This will be consumed by consume_stock() function, later.
1885 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1887 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1889 if (stock->cached != memcg) { /* reset if necessary */
1891 stock->cached = memcg;
1893 stock->nr_pages += nr_pages;
1894 put_cpu_var(memcg_stock);
1898 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1899 * of the hierarchy under it.
1901 static void drain_all_stock(struct mem_cgroup *root_memcg)
1905 /* If someone's already draining, avoid adding running more workers. */
1906 if (!mutex_trylock(&percpu_charge_mutex))
1908 /* Notify other cpus that system-wide "drain" is running */
1911 for_each_online_cpu(cpu) {
1912 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1913 struct mem_cgroup *memcg;
1915 memcg = stock->cached;
1916 if (!memcg || !stock->nr_pages)
1918 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1920 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1922 drain_local_stock(&stock->work);
1924 schedule_work_on(cpu, &stock->work);
1929 mutex_unlock(&percpu_charge_mutex);
1932 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1933 unsigned long action,
1936 int cpu = (unsigned long)hcpu;
1937 struct memcg_stock_pcp *stock;
1939 if (action == CPU_ONLINE)
1942 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1945 stock = &per_cpu(memcg_stock, cpu);
1950 static void reclaim_high(struct mem_cgroup *memcg,
1951 unsigned int nr_pages,
1955 if (page_counter_read(&memcg->memory) <= memcg->high)
1957 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1958 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1959 } while ((memcg = parent_mem_cgroup(memcg)));
1962 static void high_work_func(struct work_struct *work)
1964 struct mem_cgroup *memcg;
1966 memcg = container_of(work, struct mem_cgroup, high_work);
1967 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1971 * Scheduled by try_charge() to be executed from the userland return path
1972 * and reclaims memory over the high limit.
1974 void mem_cgroup_handle_over_high(void)
1976 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1977 struct mem_cgroup *memcg;
1979 if (likely(!nr_pages))
1982 memcg = get_mem_cgroup_from_mm(current->mm);
1983 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1984 css_put(&memcg->css);
1985 current->memcg_nr_pages_over_high = 0;
1988 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1989 unsigned int nr_pages)
1991 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1992 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1993 struct mem_cgroup *mem_over_limit;
1994 struct page_counter *counter;
1995 unsigned long nr_reclaimed;
1996 bool may_swap = true;
1997 bool drained = false;
1999 if (mem_cgroup_is_root(memcg))
2002 if (consume_stock(memcg, nr_pages))
2005 if (!do_memsw_account() ||
2006 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2007 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2009 if (do_memsw_account())
2010 page_counter_uncharge(&memcg->memsw, batch);
2011 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2013 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2017 if (batch > nr_pages) {
2023 * Unlike in global OOM situations, memcg is not in a physical
2024 * memory shortage. Allow dying and OOM-killed tasks to
2025 * bypass the last charges so that they can exit quickly and
2026 * free their memory.
2028 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2029 fatal_signal_pending(current) ||
2030 current->flags & PF_EXITING))
2033 if (unlikely(task_in_memcg_oom(current)))
2036 if (!gfpflags_allow_blocking(gfp_mask))
2039 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2041 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2042 gfp_mask, may_swap);
2044 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2048 drain_all_stock(mem_over_limit);
2053 if (gfp_mask & __GFP_NORETRY)
2056 * Even though the limit is exceeded at this point, reclaim
2057 * may have been able to free some pages. Retry the charge
2058 * before killing the task.
2060 * Only for regular pages, though: huge pages are rather
2061 * unlikely to succeed so close to the limit, and we fall back
2062 * to regular pages anyway in case of failure.
2064 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2067 * At task move, charge accounts can be doubly counted. So, it's
2068 * better to wait until the end of task_move if something is going on.
2070 if (mem_cgroup_wait_acct_move(mem_over_limit))
2076 if (gfp_mask & __GFP_NOFAIL)
2079 if (fatal_signal_pending(current))
2082 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2084 mem_cgroup_oom(mem_over_limit, gfp_mask,
2085 get_order(nr_pages * PAGE_SIZE));
2087 if (!(gfp_mask & __GFP_NOFAIL))
2091 * The allocation either can't fail or will lead to more memory
2092 * being freed very soon. Allow memory usage go over the limit
2093 * temporarily by force charging it.
2095 page_counter_charge(&memcg->memory, nr_pages);
2096 if (do_memsw_account())
2097 page_counter_charge(&memcg->memsw, nr_pages);
2098 css_get_many(&memcg->css, nr_pages);
2103 css_get_many(&memcg->css, batch);
2104 if (batch > nr_pages)
2105 refill_stock(memcg, batch - nr_pages);
2108 * If the hierarchy is above the normal consumption range, schedule
2109 * reclaim on returning to userland. We can perform reclaim here
2110 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2111 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2112 * not recorded as it most likely matches current's and won't
2113 * change in the meantime. As high limit is checked again before
2114 * reclaim, the cost of mismatch is negligible.
2117 if (page_counter_read(&memcg->memory) > memcg->high) {
2118 /* Don't bother a random interrupted task */
2119 if (in_interrupt()) {
2120 schedule_work(&memcg->high_work);
2123 current->memcg_nr_pages_over_high += batch;
2124 set_notify_resume(current);
2127 } while ((memcg = parent_mem_cgroup(memcg)));
2132 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2134 if (mem_cgroup_is_root(memcg))
2137 page_counter_uncharge(&memcg->memory, nr_pages);
2138 if (do_memsw_account())
2139 page_counter_uncharge(&memcg->memsw, nr_pages);
2141 css_put_many(&memcg->css, nr_pages);
2144 static void lock_page_lru(struct page *page, int *isolated)
2146 struct zone *zone = page_zone(page);
2148 spin_lock_irq(&zone->lru_lock);
2149 if (PageLRU(page)) {
2150 struct lruvec *lruvec;
2152 lruvec = mem_cgroup_page_lruvec(page, zone);
2154 del_page_from_lru_list(page, lruvec, page_lru(page));
2160 static void unlock_page_lru(struct page *page, int isolated)
2162 struct zone *zone = page_zone(page);
2165 struct lruvec *lruvec;
2167 lruvec = mem_cgroup_page_lruvec(page, zone);
2168 VM_BUG_ON_PAGE(PageLRU(page), page);
2170 add_page_to_lru_list(page, lruvec, page_lru(page));
2172 spin_unlock_irq(&zone->lru_lock);
2175 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2180 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2183 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2184 * may already be on some other mem_cgroup's LRU. Take care of it.
2187 lock_page_lru(page, &isolated);
2190 * Nobody should be changing or seriously looking at
2191 * page->mem_cgroup at this point:
2193 * - the page is uncharged
2195 * - the page is off-LRU
2197 * - an anonymous fault has exclusive page access, except for
2198 * a locked page table
2200 * - a page cache insertion, a swapin fault, or a migration
2201 * have the page locked
2203 page->mem_cgroup = memcg;
2206 unlock_page_lru(page, isolated);
2210 static int memcg_alloc_cache_id(void)
2215 id = ida_simple_get(&memcg_cache_ida,
2216 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2220 if (id < memcg_nr_cache_ids)
2224 * There's no space for the new id in memcg_caches arrays,
2225 * so we have to grow them.
2227 down_write(&memcg_cache_ids_sem);
2229 size = 2 * (id + 1);
2230 if (size < MEMCG_CACHES_MIN_SIZE)
2231 size = MEMCG_CACHES_MIN_SIZE;
2232 else if (size > MEMCG_CACHES_MAX_SIZE)
2233 size = MEMCG_CACHES_MAX_SIZE;
2235 err = memcg_update_all_caches(size);
2237 err = memcg_update_all_list_lrus(size);
2239 memcg_nr_cache_ids = size;
2241 up_write(&memcg_cache_ids_sem);
2244 ida_simple_remove(&memcg_cache_ida, id);
2250 static void memcg_free_cache_id(int id)
2252 ida_simple_remove(&memcg_cache_ida, id);
2255 struct memcg_kmem_cache_create_work {
2256 struct mem_cgroup *memcg;
2257 struct kmem_cache *cachep;
2258 struct work_struct work;
2261 static void memcg_kmem_cache_create_func(struct work_struct *w)
2263 struct memcg_kmem_cache_create_work *cw =
2264 container_of(w, struct memcg_kmem_cache_create_work, work);
2265 struct mem_cgroup *memcg = cw->memcg;
2266 struct kmem_cache *cachep = cw->cachep;
2268 memcg_create_kmem_cache(memcg, cachep);
2270 css_put(&memcg->css);
2275 * Enqueue the creation of a per-memcg kmem_cache.
2277 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2278 struct kmem_cache *cachep)
2280 struct memcg_kmem_cache_create_work *cw;
2282 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2286 css_get(&memcg->css);
2289 cw->cachep = cachep;
2290 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2292 schedule_work(&cw->work);
2295 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2296 struct kmem_cache *cachep)
2299 * We need to stop accounting when we kmalloc, because if the
2300 * corresponding kmalloc cache is not yet created, the first allocation
2301 * in __memcg_schedule_kmem_cache_create will recurse.
2303 * However, it is better to enclose the whole function. Depending on
2304 * the debugging options enabled, INIT_WORK(), for instance, can
2305 * trigger an allocation. This too, will make us recurse. Because at
2306 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2307 * the safest choice is to do it like this, wrapping the whole function.
2309 current->memcg_kmem_skip_account = 1;
2310 __memcg_schedule_kmem_cache_create(memcg, cachep);
2311 current->memcg_kmem_skip_account = 0;
2315 * Return the kmem_cache we're supposed to use for a slab allocation.
2316 * We try to use the current memcg's version of the cache.
2318 * If the cache does not exist yet, if we are the first user of it,
2319 * we either create it immediately, if possible, or create it asynchronously
2321 * In the latter case, we will let the current allocation go through with
2322 * the original cache.
2324 * Can't be called in interrupt context or from kernel threads.
2325 * This function needs to be called with rcu_read_lock() held.
2327 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2329 struct mem_cgroup *memcg;
2330 struct kmem_cache *memcg_cachep;
2333 VM_BUG_ON(!is_root_cache(cachep));
2335 if (cachep->flags & SLAB_ACCOUNT)
2336 gfp |= __GFP_ACCOUNT;
2338 if (!(gfp & __GFP_ACCOUNT))
2341 if (current->memcg_kmem_skip_account)
2344 memcg = get_mem_cgroup_from_mm(current->mm);
2345 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2349 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2350 if (likely(memcg_cachep))
2351 return memcg_cachep;
2354 * If we are in a safe context (can wait, and not in interrupt
2355 * context), we could be be predictable and return right away.
2356 * This would guarantee that the allocation being performed
2357 * already belongs in the new cache.
2359 * However, there are some clashes that can arrive from locking.
2360 * For instance, because we acquire the slab_mutex while doing
2361 * memcg_create_kmem_cache, this means no further allocation
2362 * could happen with the slab_mutex held. So it's better to
2365 memcg_schedule_kmem_cache_create(memcg, cachep);
2367 css_put(&memcg->css);
2371 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2373 if (!is_root_cache(cachep))
2374 css_put(&cachep->memcg_params.memcg->css);
2377 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2378 struct mem_cgroup *memcg)
2380 unsigned int nr_pages = 1 << order;
2381 struct page_counter *counter;
2384 if (!memcg_kmem_online(memcg))
2387 ret = try_charge(memcg, gfp, nr_pages);
2391 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2392 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2393 cancel_charge(memcg, nr_pages);
2397 page->mem_cgroup = memcg;
2402 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2404 struct mem_cgroup *memcg;
2407 memcg = get_mem_cgroup_from_mm(current->mm);
2408 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2409 css_put(&memcg->css);
2413 void __memcg_kmem_uncharge(struct page *page, int order)
2415 struct mem_cgroup *memcg = page->mem_cgroup;
2416 unsigned int nr_pages = 1 << order;
2421 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2423 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2424 page_counter_uncharge(&memcg->kmem, nr_pages);
2426 page_counter_uncharge(&memcg->memory, nr_pages);
2427 if (do_memsw_account())
2428 page_counter_uncharge(&memcg->memsw, nr_pages);
2430 page->mem_cgroup = NULL;
2431 css_put_many(&memcg->css, nr_pages);
2433 #endif /* !CONFIG_SLOB */
2435 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2438 * Because tail pages are not marked as "used", set it. We're under
2439 * zone->lru_lock and migration entries setup in all page mappings.
2441 void mem_cgroup_split_huge_fixup(struct page *head)
2445 if (mem_cgroup_disabled())
2448 for (i = 1; i < HPAGE_PMD_NR; i++)
2449 head[i].mem_cgroup = head->mem_cgroup;
2451 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2454 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2456 #ifdef CONFIG_MEMCG_SWAP
2457 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2460 int val = (charge) ? 1 : -1;
2461 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2465 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2466 * @entry: swap entry to be moved
2467 * @from: mem_cgroup which the entry is moved from
2468 * @to: mem_cgroup which the entry is moved to
2470 * It succeeds only when the swap_cgroup's record for this entry is the same
2471 * as the mem_cgroup's id of @from.
2473 * Returns 0 on success, -EINVAL on failure.
2475 * The caller must have charged to @to, IOW, called page_counter_charge() about
2476 * both res and memsw, and called css_get().
2478 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2479 struct mem_cgroup *from, struct mem_cgroup *to)
2481 unsigned short old_id, new_id;
2483 old_id = mem_cgroup_id(from);
2484 new_id = mem_cgroup_id(to);
2486 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2487 mem_cgroup_swap_statistics(from, false);
2488 mem_cgroup_swap_statistics(to, true);
2494 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2495 struct mem_cgroup *from, struct mem_cgroup *to)
2501 static DEFINE_MUTEX(memcg_limit_mutex);
2503 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2504 unsigned long limit)
2506 unsigned long curusage;
2507 unsigned long oldusage;
2508 bool enlarge = false;
2513 * For keeping hierarchical_reclaim simple, how long we should retry
2514 * is depends on callers. We set our retry-count to be function
2515 * of # of children which we should visit in this loop.
2517 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2518 mem_cgroup_count_children(memcg);
2520 oldusage = page_counter_read(&memcg->memory);
2523 if (signal_pending(current)) {
2528 mutex_lock(&memcg_limit_mutex);
2529 if (limit > memcg->memsw.limit) {
2530 mutex_unlock(&memcg_limit_mutex);
2534 if (limit > memcg->memory.limit)
2536 ret = page_counter_limit(&memcg->memory, limit);
2537 mutex_unlock(&memcg_limit_mutex);
2542 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2544 curusage = page_counter_read(&memcg->memory);
2545 /* Usage is reduced ? */
2546 if (curusage >= oldusage)
2549 oldusage = curusage;
2550 } while (retry_count);
2552 if (!ret && enlarge)
2553 memcg_oom_recover(memcg);
2558 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2559 unsigned long limit)
2561 unsigned long curusage;
2562 unsigned long oldusage;
2563 bool enlarge = false;
2567 /* see mem_cgroup_resize_res_limit */
2568 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2569 mem_cgroup_count_children(memcg);
2571 oldusage = page_counter_read(&memcg->memsw);
2574 if (signal_pending(current)) {
2579 mutex_lock(&memcg_limit_mutex);
2580 if (limit < memcg->memory.limit) {
2581 mutex_unlock(&memcg_limit_mutex);
2585 if (limit > memcg->memsw.limit)
2587 ret = page_counter_limit(&memcg->memsw, limit);
2588 mutex_unlock(&memcg_limit_mutex);
2593 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2595 curusage = page_counter_read(&memcg->memsw);
2596 /* Usage is reduced ? */
2597 if (curusage >= oldusage)
2600 oldusage = curusage;
2601 } while (retry_count);
2603 if (!ret && enlarge)
2604 memcg_oom_recover(memcg);
2609 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2611 unsigned long *total_scanned)
2613 unsigned long nr_reclaimed = 0;
2614 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2615 unsigned long reclaimed;
2617 struct mem_cgroup_tree_per_zone *mctz;
2618 unsigned long excess;
2619 unsigned long nr_scanned;
2624 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2626 * This loop can run a while, specially if mem_cgroup's continuously
2627 * keep exceeding their soft limit and putting the system under
2634 mz = mem_cgroup_largest_soft_limit_node(mctz);
2639 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2640 gfp_mask, &nr_scanned);
2641 nr_reclaimed += reclaimed;
2642 *total_scanned += nr_scanned;
2643 spin_lock_irq(&mctz->lock);
2644 __mem_cgroup_remove_exceeded(mz, mctz);
2647 * If we failed to reclaim anything from this memory cgroup
2648 * it is time to move on to the next cgroup
2652 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2654 excess = soft_limit_excess(mz->memcg);
2656 * One school of thought says that we should not add
2657 * back the node to the tree if reclaim returns 0.
2658 * But our reclaim could return 0, simply because due
2659 * to priority we are exposing a smaller subset of
2660 * memory to reclaim from. Consider this as a longer
2663 /* If excess == 0, no tree ops */
2664 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2665 spin_unlock_irq(&mctz->lock);
2666 css_put(&mz->memcg->css);
2669 * Could not reclaim anything and there are no more
2670 * mem cgroups to try or we seem to be looping without
2671 * reclaiming anything.
2673 if (!nr_reclaimed &&
2675 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2677 } while (!nr_reclaimed);
2679 css_put(&next_mz->memcg->css);
2680 return nr_reclaimed;
2684 * Test whether @memcg has children, dead or alive. Note that this
2685 * function doesn't care whether @memcg has use_hierarchy enabled and
2686 * returns %true if there are child csses according to the cgroup
2687 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2689 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2694 * The lock does not prevent addition or deletion of children, but
2695 * it prevents a new child from being initialized based on this
2696 * parent in css_online(), so it's enough to decide whether
2697 * hierarchically inherited attributes can still be changed or not.
2699 lockdep_assert_held(&memcg_create_mutex);
2702 ret = css_next_child(NULL, &memcg->css);
2708 * Reclaims as many pages from the given memcg as possible and moves
2709 * the rest to the parent.
2711 * Caller is responsible for holding css reference for memcg.
2713 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2715 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2717 /* we call try-to-free pages for make this cgroup empty */
2718 lru_add_drain_all();
2719 /* try to free all pages in this cgroup */
2720 while (nr_retries && page_counter_read(&memcg->memory)) {
2723 if (signal_pending(current))
2726 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2730 /* maybe some writeback is necessary */
2731 congestion_wait(BLK_RW_ASYNC, HZ/10);
2739 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2740 char *buf, size_t nbytes,
2743 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2745 if (mem_cgroup_is_root(memcg))
2747 return mem_cgroup_force_empty(memcg) ?: nbytes;
2750 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2753 return mem_cgroup_from_css(css)->use_hierarchy;
2756 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2757 struct cftype *cft, u64 val)
2760 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2761 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2763 mutex_lock(&memcg_create_mutex);
2765 if (memcg->use_hierarchy == val)
2769 * If parent's use_hierarchy is set, we can't make any modifications
2770 * in the child subtrees. If it is unset, then the change can
2771 * occur, provided the current cgroup has no children.
2773 * For the root cgroup, parent_mem is NULL, we allow value to be
2774 * set if there are no children.
2776 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2777 (val == 1 || val == 0)) {
2778 if (!memcg_has_children(memcg))
2779 memcg->use_hierarchy = val;
2786 mutex_unlock(&memcg_create_mutex);
2791 static unsigned long tree_stat(struct mem_cgroup *memcg,
2792 enum mem_cgroup_stat_index idx)
2794 struct mem_cgroup *iter;
2795 unsigned long val = 0;
2797 for_each_mem_cgroup_tree(iter, memcg)
2798 val += mem_cgroup_read_stat(iter, idx);
2803 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2807 if (mem_cgroup_is_root(memcg)) {
2808 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2809 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2811 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2814 val = page_counter_read(&memcg->memory);
2816 val = page_counter_read(&memcg->memsw);
2829 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2832 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2833 struct page_counter *counter;
2835 switch (MEMFILE_TYPE(cft->private)) {
2837 counter = &memcg->memory;
2840 counter = &memcg->memsw;
2843 counter = &memcg->kmem;
2849 switch (MEMFILE_ATTR(cft->private)) {
2851 if (counter == &memcg->memory)
2852 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2853 if (counter == &memcg->memsw)
2854 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2855 return (u64)page_counter_read(counter) * PAGE_SIZE;
2857 return (u64)counter->limit * PAGE_SIZE;
2859 return (u64)counter->watermark * PAGE_SIZE;
2861 return counter->failcnt;
2862 case RES_SOFT_LIMIT:
2863 return (u64)memcg->soft_limit * PAGE_SIZE;
2870 static int memcg_online_kmem(struct mem_cgroup *memcg)
2875 BUG_ON(memcg->kmemcg_id >= 0);
2876 BUG_ON(memcg->kmem_state);
2879 * For simplicity, we won't allow this to be disabled. It also can't
2880 * be changed if the cgroup has children already, or if tasks had
2883 * If tasks join before we set the limit, a person looking at
2884 * kmem.usage_in_bytes will have no way to determine when it took
2885 * place, which makes the value quite meaningless.
2887 * After it first became limited, changes in the value of the limit are
2888 * of course permitted.
2890 mutex_lock(&memcg_create_mutex);
2891 if (cgroup_is_populated(memcg->css.cgroup) ||
2892 (memcg->use_hierarchy && memcg_has_children(memcg)))
2894 mutex_unlock(&memcg_create_mutex);
2898 memcg_id = memcg_alloc_cache_id();
2904 static_branch_inc(&memcg_kmem_enabled_key);
2906 * A memory cgroup is considered kmem-online as soon as it gets
2907 * kmemcg_id. Setting the id after enabling static branching will
2908 * guarantee no one starts accounting before all call sites are
2911 memcg->kmemcg_id = memcg_id;
2912 memcg->kmem_state = KMEM_ONLINE;
2917 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2920 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2925 mutex_lock(&memcg_limit_mutex);
2927 * If the parent cgroup is not kmem-online now, it cannot be
2928 * onlined after this point, because it has at least one child
2931 if (memcg_kmem_online(parent) ||
2932 (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nokmem))
2933 ret = memcg_online_kmem(memcg);
2934 mutex_unlock(&memcg_limit_mutex);
2938 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2940 struct cgroup_subsys_state *css;
2941 struct mem_cgroup *parent, *child;
2944 if (memcg->kmem_state != KMEM_ONLINE)
2947 * Clear the online state before clearing memcg_caches array
2948 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2949 * guarantees that no cache will be created for this cgroup
2950 * after we are done (see memcg_create_kmem_cache()).
2952 memcg->kmem_state = KMEM_ALLOCATED;
2954 memcg_deactivate_kmem_caches(memcg);
2956 kmemcg_id = memcg->kmemcg_id;
2957 BUG_ON(kmemcg_id < 0);
2959 parent = parent_mem_cgroup(memcg);
2961 parent = root_mem_cgroup;
2964 * Change kmemcg_id of this cgroup and all its descendants to the
2965 * parent's id, and then move all entries from this cgroup's list_lrus
2966 * to ones of the parent. After we have finished, all list_lrus
2967 * corresponding to this cgroup are guaranteed to remain empty. The
2968 * ordering is imposed by list_lru_node->lock taken by
2969 * memcg_drain_all_list_lrus().
2971 css_for_each_descendant_pre(css, &memcg->css) {
2972 child = mem_cgroup_from_css(css);
2973 BUG_ON(child->kmemcg_id != kmemcg_id);
2974 child->kmemcg_id = parent->kmemcg_id;
2975 if (!memcg->use_hierarchy)
2978 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2980 memcg_free_cache_id(kmemcg_id);
2983 static void memcg_free_kmem(struct mem_cgroup *memcg)
2985 if (memcg->kmem_state == KMEM_ALLOCATED) {
2986 memcg_destroy_kmem_caches(memcg);
2987 static_branch_dec(&memcg_kmem_enabled_key);
2988 WARN_ON(page_counter_read(&memcg->kmem));
2992 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2996 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2999 static void memcg_free_kmem(struct mem_cgroup *memcg)
3002 #endif /* !CONFIG_SLOB */
3004 #ifdef CONFIG_MEMCG_LEGACY_KMEM
3005 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3006 unsigned long limit)
3010 mutex_lock(&memcg_limit_mutex);
3011 /* Top-level cgroup doesn't propagate from root */
3012 if (!memcg_kmem_online(memcg)) {
3013 ret = memcg_online_kmem(memcg);
3017 ret = page_counter_limit(&memcg->kmem, limit);
3019 mutex_unlock(&memcg_limit_mutex);
3023 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3024 unsigned long limit)
3028 #endif /* CONFIG_MEMCG_LEGACY_KMEM */
3032 * The user of this function is...
3035 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3036 char *buf, size_t nbytes, loff_t off)
3038 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3039 unsigned long nr_pages;
3042 buf = strstrip(buf);
3043 ret = page_counter_memparse(buf, "-1", &nr_pages);
3047 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3049 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3053 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3055 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3058 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3061 ret = memcg_update_kmem_limit(memcg, nr_pages);
3065 case RES_SOFT_LIMIT:
3066 memcg->soft_limit = nr_pages;
3070 return ret ?: nbytes;
3073 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3074 size_t nbytes, loff_t off)
3076 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3077 struct page_counter *counter;
3079 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3081 counter = &memcg->memory;
3084 counter = &memcg->memsw;
3087 counter = &memcg->kmem;
3093 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3095 page_counter_reset_watermark(counter);
3098 counter->failcnt = 0;
3107 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3110 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3114 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3115 struct cftype *cft, u64 val)
3117 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3119 if (val & ~MOVE_MASK)
3123 * No kind of locking is needed in here, because ->can_attach() will
3124 * check this value once in the beginning of the process, and then carry
3125 * on with stale data. This means that changes to this value will only
3126 * affect task migrations starting after the change.
3128 memcg->move_charge_at_immigrate = val;
3132 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3133 struct cftype *cft, u64 val)
3140 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3144 unsigned int lru_mask;
3147 static const struct numa_stat stats[] = {
3148 { "total", LRU_ALL },
3149 { "file", LRU_ALL_FILE },
3150 { "anon", LRU_ALL_ANON },
3151 { "unevictable", BIT(LRU_UNEVICTABLE) },
3153 const struct numa_stat *stat;
3156 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3158 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3159 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3160 seq_printf(m, "%s=%lu", stat->name, nr);
3161 for_each_node_state(nid, N_MEMORY) {
3162 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3164 seq_printf(m, " N%d=%lu", nid, nr);
3169 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3170 struct mem_cgroup *iter;
3173 for_each_mem_cgroup_tree(iter, memcg)
3174 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3175 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3176 for_each_node_state(nid, N_MEMORY) {
3178 for_each_mem_cgroup_tree(iter, memcg)
3179 nr += mem_cgroup_node_nr_lru_pages(
3180 iter, nid, stat->lru_mask);
3181 seq_printf(m, " N%d=%lu", nid, nr);
3188 #endif /* CONFIG_NUMA */
3190 static int memcg_stat_show(struct seq_file *m, void *v)
3192 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3193 unsigned long memory, memsw;
3194 struct mem_cgroup *mi;
3197 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3198 MEM_CGROUP_STAT_NSTATS);
3199 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3200 MEM_CGROUP_EVENTS_NSTATS);
3201 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3203 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3204 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3206 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3207 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3210 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3211 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3212 mem_cgroup_read_events(memcg, i));
3214 for (i = 0; i < NR_LRU_LISTS; i++)
3215 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3216 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3218 /* Hierarchical information */
3219 memory = memsw = PAGE_COUNTER_MAX;
3220 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3221 memory = min(memory, mi->memory.limit);
3222 memsw = min(memsw, mi->memsw.limit);
3224 seq_printf(m, "hierarchical_memory_limit %llu\n",
3225 (u64)memory * PAGE_SIZE);
3226 if (do_memsw_account())
3227 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3228 (u64)memsw * PAGE_SIZE);
3230 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3231 unsigned long long val = 0;
3233 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3235 for_each_mem_cgroup_tree(mi, memcg)
3236 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3237 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3240 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3241 unsigned long long val = 0;
3243 for_each_mem_cgroup_tree(mi, memcg)
3244 val += mem_cgroup_read_events(mi, i);
3245 seq_printf(m, "total_%s %llu\n",
3246 mem_cgroup_events_names[i], val);
3249 for (i = 0; i < NR_LRU_LISTS; i++) {
3250 unsigned long long val = 0;
3252 for_each_mem_cgroup_tree(mi, memcg)
3253 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3254 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3257 #ifdef CONFIG_DEBUG_VM
3260 struct mem_cgroup_per_zone *mz;
3261 struct zone_reclaim_stat *rstat;
3262 unsigned long recent_rotated[2] = {0, 0};
3263 unsigned long recent_scanned[2] = {0, 0};
3265 for_each_online_node(nid)
3266 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3267 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3268 rstat = &mz->lruvec.reclaim_stat;
3270 recent_rotated[0] += rstat->recent_rotated[0];
3271 recent_rotated[1] += rstat->recent_rotated[1];
3272 recent_scanned[0] += rstat->recent_scanned[0];
3273 recent_scanned[1] += rstat->recent_scanned[1];
3275 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3276 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3277 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3278 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3285 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3288 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3290 return mem_cgroup_swappiness(memcg);
3293 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3294 struct cftype *cft, u64 val)
3296 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3302 memcg->swappiness = val;
3304 vm_swappiness = val;
3309 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3311 struct mem_cgroup_threshold_ary *t;
3312 unsigned long usage;
3317 t = rcu_dereference(memcg->thresholds.primary);
3319 t = rcu_dereference(memcg->memsw_thresholds.primary);
3324 usage = mem_cgroup_usage(memcg, swap);
3327 * current_threshold points to threshold just below or equal to usage.
3328 * If it's not true, a threshold was crossed after last
3329 * call of __mem_cgroup_threshold().
3331 i = t->current_threshold;
3334 * Iterate backward over array of thresholds starting from
3335 * current_threshold and check if a threshold is crossed.
3336 * If none of thresholds below usage is crossed, we read
3337 * only one element of the array here.
3339 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3340 eventfd_signal(t->entries[i].eventfd, 1);
3342 /* i = current_threshold + 1 */
3346 * Iterate forward over array of thresholds starting from
3347 * current_threshold+1 and check if a threshold is crossed.
3348 * If none of thresholds above usage is crossed, we read
3349 * only one element of the array here.
3351 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3352 eventfd_signal(t->entries[i].eventfd, 1);
3354 /* Update current_threshold */
3355 t->current_threshold = i - 1;
3360 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3363 __mem_cgroup_threshold(memcg, false);
3364 if (do_memsw_account())
3365 __mem_cgroup_threshold(memcg, true);
3367 memcg = parent_mem_cgroup(memcg);
3371 static int compare_thresholds(const void *a, const void *b)
3373 const struct mem_cgroup_threshold *_a = a;
3374 const struct mem_cgroup_threshold *_b = b;
3376 if (_a->threshold > _b->threshold)
3379 if (_a->threshold < _b->threshold)
3385 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3387 struct mem_cgroup_eventfd_list *ev;
3389 spin_lock(&memcg_oom_lock);
3391 list_for_each_entry(ev, &memcg->oom_notify, list)
3392 eventfd_signal(ev->eventfd, 1);
3394 spin_unlock(&memcg_oom_lock);
3398 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3400 struct mem_cgroup *iter;
3402 for_each_mem_cgroup_tree(iter, memcg)
3403 mem_cgroup_oom_notify_cb(iter);
3406 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3407 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3409 struct mem_cgroup_thresholds *thresholds;
3410 struct mem_cgroup_threshold_ary *new;
3411 unsigned long threshold;
3412 unsigned long usage;
3415 ret = page_counter_memparse(args, "-1", &threshold);
3419 mutex_lock(&memcg->thresholds_lock);
3422 thresholds = &memcg->thresholds;
3423 usage = mem_cgroup_usage(memcg, false);
3424 } else if (type == _MEMSWAP) {
3425 thresholds = &memcg->memsw_thresholds;
3426 usage = mem_cgroup_usage(memcg, true);
3430 /* Check if a threshold crossed before adding a new one */
3431 if (thresholds->primary)
3432 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3434 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3436 /* Allocate memory for new array of thresholds */
3437 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3445 /* Copy thresholds (if any) to new array */
3446 if (thresholds->primary) {
3447 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3448 sizeof(struct mem_cgroup_threshold));
3451 /* Add new threshold */
3452 new->entries[size - 1].eventfd = eventfd;
3453 new->entries[size - 1].threshold = threshold;
3455 /* Sort thresholds. Registering of new threshold isn't time-critical */
3456 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3457 compare_thresholds, NULL);
3459 /* Find current threshold */
3460 new->current_threshold = -1;
3461 for (i = 0; i < size; i++) {
3462 if (new->entries[i].threshold <= usage) {
3464 * new->current_threshold will not be used until
3465 * rcu_assign_pointer(), so it's safe to increment
3468 ++new->current_threshold;
3473 /* Free old spare buffer and save old primary buffer as spare */
3474 kfree(thresholds->spare);
3475 thresholds->spare = thresholds->primary;
3477 rcu_assign_pointer(thresholds->primary, new);
3479 /* To be sure that nobody uses thresholds */
3483 mutex_unlock(&memcg->thresholds_lock);
3488 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3489 struct eventfd_ctx *eventfd, const char *args)
3491 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3494 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3495 struct eventfd_ctx *eventfd, const char *args)
3497 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3500 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3501 struct eventfd_ctx *eventfd, enum res_type type)
3503 struct mem_cgroup_thresholds *thresholds;
3504 struct mem_cgroup_threshold_ary *new;
3505 unsigned long usage;
3508 mutex_lock(&memcg->thresholds_lock);
3511 thresholds = &memcg->thresholds;
3512 usage = mem_cgroup_usage(memcg, false);
3513 } else if (type == _MEMSWAP) {
3514 thresholds = &memcg->memsw_thresholds;
3515 usage = mem_cgroup_usage(memcg, true);
3519 if (!thresholds->primary)
3522 /* Check if a threshold crossed before removing */
3523 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3525 /* Calculate new number of threshold */
3527 for (i = 0; i < thresholds->primary->size; i++) {
3528 if (thresholds->primary->entries[i].eventfd != eventfd)
3532 new = thresholds->spare;
3534 /* Set thresholds array to NULL if we don't have thresholds */
3543 /* Copy thresholds and find current threshold */
3544 new->current_threshold = -1;
3545 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3546 if (thresholds->primary->entries[i].eventfd == eventfd)
3549 new->entries[j] = thresholds->primary->entries[i];
3550 if (new->entries[j].threshold <= usage) {
3552 * new->current_threshold will not be used
3553 * until rcu_assign_pointer(), so it's safe to increment
3556 ++new->current_threshold;
3562 /* Swap primary and spare array */
3563 thresholds->spare = thresholds->primary;
3565 rcu_assign_pointer(thresholds->primary, new);
3567 /* To be sure that nobody uses thresholds */
3570 /* If all events are unregistered, free the spare array */
3572 kfree(thresholds->spare);
3573 thresholds->spare = NULL;
3576 mutex_unlock(&memcg->thresholds_lock);
3579 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3580 struct eventfd_ctx *eventfd)
3582 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3585 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3586 struct eventfd_ctx *eventfd)
3588 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3591 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3592 struct eventfd_ctx *eventfd, const char *args)
3594 struct mem_cgroup_eventfd_list *event;
3596 event = kmalloc(sizeof(*event), GFP_KERNEL);
3600 spin_lock(&memcg_oom_lock);
3602 event->eventfd = eventfd;
3603 list_add(&event->list, &memcg->oom_notify);
3605 /* already in OOM ? */
3606 if (memcg->under_oom)
3607 eventfd_signal(eventfd, 1);
3608 spin_unlock(&memcg_oom_lock);
3613 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3614 struct eventfd_ctx *eventfd)
3616 struct mem_cgroup_eventfd_list *ev, *tmp;
3618 spin_lock(&memcg_oom_lock);
3620 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3621 if (ev->eventfd == eventfd) {
3622 list_del(&ev->list);
3627 spin_unlock(&memcg_oom_lock);
3630 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3632 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3634 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3635 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3639 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3640 struct cftype *cft, u64 val)
3642 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3644 /* cannot set to root cgroup and only 0 and 1 are allowed */
3645 if (!css->parent || !((val == 0) || (val == 1)))
3648 memcg->oom_kill_disable = val;
3650 memcg_oom_recover(memcg);
3655 #ifdef CONFIG_CGROUP_WRITEBACK
3657 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3659 return &memcg->cgwb_list;
3662 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3664 return wb_domain_init(&memcg->cgwb_domain, gfp);
3667 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3669 wb_domain_exit(&memcg->cgwb_domain);
3672 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3674 wb_domain_size_changed(&memcg->cgwb_domain);
3677 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3679 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3681 if (!memcg->css.parent)
3684 return &memcg->cgwb_domain;
3688 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3689 * @wb: bdi_writeback in question
3690 * @pfilepages: out parameter for number of file pages
3691 * @pheadroom: out parameter for number of allocatable pages according to memcg
3692 * @pdirty: out parameter for number of dirty pages
3693 * @pwriteback: out parameter for number of pages under writeback
3695 * Determine the numbers of file, headroom, dirty, and writeback pages in
3696 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3697 * is a bit more involved.
3699 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3700 * headroom is calculated as the lowest headroom of itself and the
3701 * ancestors. Note that this doesn't consider the actual amount of
3702 * available memory in the system. The caller should further cap
3703 * *@pheadroom accordingly.
3705 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3706 unsigned long *pheadroom, unsigned long *pdirty,
3707 unsigned long *pwriteback)
3709 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3710 struct mem_cgroup *parent;
3712 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3714 /* this should eventually include NR_UNSTABLE_NFS */
3715 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3716 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3717 (1 << LRU_ACTIVE_FILE));
3718 *pheadroom = PAGE_COUNTER_MAX;
3720 while ((parent = parent_mem_cgroup(memcg))) {
3721 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3722 unsigned long used = page_counter_read(&memcg->memory);
3724 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3729 #else /* CONFIG_CGROUP_WRITEBACK */
3731 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3736 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3740 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3744 #endif /* CONFIG_CGROUP_WRITEBACK */
3747 * DO NOT USE IN NEW FILES.
3749 * "cgroup.event_control" implementation.
3751 * This is way over-engineered. It tries to support fully configurable
3752 * events for each user. Such level of flexibility is completely
3753 * unnecessary especially in the light of the planned unified hierarchy.
3755 * Please deprecate this and replace with something simpler if at all
3760 * Unregister event and free resources.
3762 * Gets called from workqueue.
3764 static void memcg_event_remove(struct work_struct *work)
3766 struct mem_cgroup_event *event =
3767 container_of(work, struct mem_cgroup_event, remove);
3768 struct mem_cgroup *memcg = event->memcg;
3770 remove_wait_queue(event->wqh, &event->wait);
3772 event->unregister_event(memcg, event->eventfd);
3774 /* Notify userspace the event is going away. */
3775 eventfd_signal(event->eventfd, 1);
3777 eventfd_ctx_put(event->eventfd);
3779 css_put(&memcg->css);
3783 * Gets called on POLLHUP on eventfd when user closes it.
3785 * Called with wqh->lock held and interrupts disabled.
3787 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3788 int sync, void *key)
3790 struct mem_cgroup_event *event =
3791 container_of(wait, struct mem_cgroup_event, wait);
3792 struct mem_cgroup *memcg = event->memcg;
3793 unsigned long flags = (unsigned long)key;
3795 if (flags & POLLHUP) {
3797 * If the event has been detached at cgroup removal, we
3798 * can simply return knowing the other side will cleanup
3801 * We can't race against event freeing since the other
3802 * side will require wqh->lock via remove_wait_queue(),
3805 spin_lock(&memcg->event_list_lock);
3806 if (!list_empty(&event->list)) {
3807 list_del_init(&event->list);
3809 * We are in atomic context, but cgroup_event_remove()
3810 * may sleep, so we have to call it in workqueue.
3812 schedule_work(&event->remove);
3814 spin_unlock(&memcg->event_list_lock);
3820 static void memcg_event_ptable_queue_proc(struct file *file,
3821 wait_queue_head_t *wqh, poll_table *pt)
3823 struct mem_cgroup_event *event =
3824 container_of(pt, struct mem_cgroup_event, pt);
3827 add_wait_queue(wqh, &event->wait);
3831 * DO NOT USE IN NEW FILES.
3833 * Parse input and register new cgroup event handler.
3835 * Input must be in format '<event_fd> <control_fd> <args>'.
3836 * Interpretation of args is defined by control file implementation.
3838 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3839 char *buf, size_t nbytes, loff_t off)
3841 struct cgroup_subsys_state *css = of_css(of);
3842 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3843 struct mem_cgroup_event *event;
3844 struct cgroup_subsys_state *cfile_css;
3845 unsigned int efd, cfd;
3852 buf = strstrip(buf);
3854 efd = simple_strtoul(buf, &endp, 10);
3859 cfd = simple_strtoul(buf, &endp, 10);
3860 if ((*endp != ' ') && (*endp != '\0'))
3864 event = kzalloc(sizeof(*event), GFP_KERNEL);
3868 event->memcg = memcg;
3869 INIT_LIST_HEAD(&event->list);
3870 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3871 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3872 INIT_WORK(&event->remove, memcg_event_remove);
3880 event->eventfd = eventfd_ctx_fileget(efile.file);
3881 if (IS_ERR(event->eventfd)) {
3882 ret = PTR_ERR(event->eventfd);
3889 goto out_put_eventfd;
3892 /* the process need read permission on control file */
3893 /* AV: shouldn't we check that it's been opened for read instead? */
3894 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3899 * Determine the event callbacks and set them in @event. This used
3900 * to be done via struct cftype but cgroup core no longer knows
3901 * about these events. The following is crude but the whole thing
3902 * is for compatibility anyway.
3904 * DO NOT ADD NEW FILES.
3906 name = cfile.file->f_path.dentry->d_name.name;
3908 if (!strcmp(name, "memory.usage_in_bytes")) {
3909 event->register_event = mem_cgroup_usage_register_event;
3910 event->unregister_event = mem_cgroup_usage_unregister_event;
3911 } else if (!strcmp(name, "memory.oom_control")) {
3912 event->register_event = mem_cgroup_oom_register_event;
3913 event->unregister_event = mem_cgroup_oom_unregister_event;
3914 } else if (!strcmp(name, "memory.pressure_level")) {
3915 event->register_event = vmpressure_register_event;
3916 event->unregister_event = vmpressure_unregister_event;
3917 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3918 event->register_event = memsw_cgroup_usage_register_event;
3919 event->unregister_event = memsw_cgroup_usage_unregister_event;
3926 * Verify @cfile should belong to @css. Also, remaining events are
3927 * automatically removed on cgroup destruction but the removal is
3928 * asynchronous, so take an extra ref on @css.
3930 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3931 &memory_cgrp_subsys);
3933 if (IS_ERR(cfile_css))
3935 if (cfile_css != css) {
3940 ret = event->register_event(memcg, event->eventfd, buf);
3944 efile.file->f_op->poll(efile.file, &event->pt);
3946 spin_lock(&memcg->event_list_lock);
3947 list_add(&event->list, &memcg->event_list);
3948 spin_unlock(&memcg->event_list_lock);
3960 eventfd_ctx_put(event->eventfd);
3969 static struct cftype mem_cgroup_legacy_files[] = {
3971 .name = "usage_in_bytes",
3972 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3973 .read_u64 = mem_cgroup_read_u64,
3976 .name = "max_usage_in_bytes",
3977 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3978 .write = mem_cgroup_reset,
3979 .read_u64 = mem_cgroup_read_u64,
3982 .name = "limit_in_bytes",
3983 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3984 .write = mem_cgroup_write,
3985 .read_u64 = mem_cgroup_read_u64,
3988 .name = "soft_limit_in_bytes",
3989 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3990 .write = mem_cgroup_write,
3991 .read_u64 = mem_cgroup_read_u64,
3995 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3996 .write = mem_cgroup_reset,
3997 .read_u64 = mem_cgroup_read_u64,
4001 .seq_show = memcg_stat_show,
4004 .name = "force_empty",
4005 .write = mem_cgroup_force_empty_write,
4008 .name = "use_hierarchy",
4009 .write_u64 = mem_cgroup_hierarchy_write,
4010 .read_u64 = mem_cgroup_hierarchy_read,
4013 .name = "cgroup.event_control", /* XXX: for compat */
4014 .write = memcg_write_event_control,
4015 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4018 .name = "swappiness",
4019 .read_u64 = mem_cgroup_swappiness_read,
4020 .write_u64 = mem_cgroup_swappiness_write,
4023 .name = "move_charge_at_immigrate",
4024 .read_u64 = mem_cgroup_move_charge_read,
4025 .write_u64 = mem_cgroup_move_charge_write,
4028 .name = "oom_control",
4029 .seq_show = mem_cgroup_oom_control_read,
4030 .write_u64 = mem_cgroup_oom_control_write,
4031 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4034 .name = "pressure_level",
4038 .name = "numa_stat",
4039 .seq_show = memcg_numa_stat_show,
4042 #ifdef CONFIG_MEMCG_LEGACY_KMEM
4044 .name = "kmem.limit_in_bytes",
4045 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4046 .write = mem_cgroup_write,
4047 .read_u64 = mem_cgroup_read_u64,
4050 .name = "kmem.usage_in_bytes",
4051 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4052 .read_u64 = mem_cgroup_read_u64,
4055 .name = "kmem.failcnt",
4056 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4057 .write = mem_cgroup_reset,
4058 .read_u64 = mem_cgroup_read_u64,
4061 .name = "kmem.max_usage_in_bytes",
4062 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4063 .write = mem_cgroup_reset,
4064 .read_u64 = mem_cgroup_read_u64,
4066 #ifdef CONFIG_SLABINFO
4068 .name = "kmem.slabinfo",
4069 .seq_start = slab_start,
4070 .seq_next = slab_next,
4071 .seq_stop = slab_stop,
4072 .seq_show = memcg_slab_show,
4076 { }, /* terminate */
4079 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4081 struct mem_cgroup_per_node *pn;
4082 struct mem_cgroup_per_zone *mz;
4083 int zone, tmp = node;
4085 * This routine is called against possible nodes.
4086 * But it's BUG to call kmalloc() against offline node.
4088 * TODO: this routine can waste much memory for nodes which will
4089 * never be onlined. It's better to use memory hotplug callback
4092 if (!node_state(node, N_NORMAL_MEMORY))
4094 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4098 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4099 mz = &pn->zoneinfo[zone];
4100 lruvec_init(&mz->lruvec);
4101 mz->usage_in_excess = 0;
4102 mz->on_tree = false;
4105 memcg->nodeinfo[node] = pn;
4109 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4111 kfree(memcg->nodeinfo[node]);
4114 static struct mem_cgroup *mem_cgroup_alloc(void)
4116 struct mem_cgroup *memcg;
4119 size = sizeof(struct mem_cgroup);
4120 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4122 memcg = kzalloc(size, GFP_KERNEL);
4126 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4130 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4136 free_percpu(memcg->stat);
4143 * At destroying mem_cgroup, references from swap_cgroup can remain.
4144 * (scanning all at force_empty is too costly...)
4146 * Instead of clearing all references at force_empty, we remember
4147 * the number of reference from swap_cgroup and free mem_cgroup when
4148 * it goes down to 0.
4150 * Removal of cgroup itself succeeds regardless of refs from swap.
4153 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4157 cancel_work_sync(&memcg->high_work);
4159 mem_cgroup_remove_from_trees(memcg);
4162 free_mem_cgroup_per_zone_info(memcg, node);
4164 free_percpu(memcg->stat);
4165 memcg_wb_domain_exit(memcg);
4169 static struct cgroup_subsys_state * __ref
4170 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4172 struct mem_cgroup *memcg;
4173 long error = -ENOMEM;
4176 memcg = mem_cgroup_alloc();
4178 return ERR_PTR(error);
4181 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4185 if (parent_css == NULL) {
4186 root_mem_cgroup = memcg;
4187 page_counter_init(&memcg->memory, NULL);
4188 memcg->high = PAGE_COUNTER_MAX;
4189 memcg->soft_limit = PAGE_COUNTER_MAX;
4190 page_counter_init(&memcg->memsw, NULL);
4191 page_counter_init(&memcg->kmem, NULL);
4194 INIT_WORK(&memcg->high_work, high_work_func);
4195 memcg->last_scanned_node = MAX_NUMNODES;
4196 INIT_LIST_HEAD(&memcg->oom_notify);
4197 memcg->move_charge_at_immigrate = 0;
4198 mutex_init(&memcg->thresholds_lock);
4199 spin_lock_init(&memcg->move_lock);
4200 vmpressure_init(&memcg->vmpressure);
4201 INIT_LIST_HEAD(&memcg->event_list);
4202 spin_lock_init(&memcg->event_list_lock);
4204 memcg->kmemcg_id = -1;
4206 #ifdef CONFIG_CGROUP_WRITEBACK
4207 INIT_LIST_HEAD(&memcg->cgwb_list);
4210 memcg->socket_pressure = jiffies;
4215 __mem_cgroup_free(memcg);
4216 return ERR_PTR(error);
4220 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4222 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4223 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4226 if (css->id > MEM_CGROUP_ID_MAX)
4232 mutex_lock(&memcg_create_mutex);
4234 memcg->use_hierarchy = parent->use_hierarchy;
4235 memcg->oom_kill_disable = parent->oom_kill_disable;
4236 memcg->swappiness = mem_cgroup_swappiness(parent);
4238 if (parent->use_hierarchy) {
4239 page_counter_init(&memcg->memory, &parent->memory);
4240 memcg->high = PAGE_COUNTER_MAX;
4241 memcg->soft_limit = PAGE_COUNTER_MAX;
4242 page_counter_init(&memcg->memsw, &parent->memsw);
4243 page_counter_init(&memcg->kmem, &parent->kmem);
4246 * No need to take a reference to the parent because cgroup
4247 * core guarantees its existence.
4250 page_counter_init(&memcg->memory, NULL);
4251 memcg->high = PAGE_COUNTER_MAX;
4252 memcg->soft_limit = PAGE_COUNTER_MAX;
4253 page_counter_init(&memcg->memsw, NULL);
4254 page_counter_init(&memcg->kmem, NULL);
4256 * Deeper hierachy with use_hierarchy == false doesn't make
4257 * much sense so let cgroup subsystem know about this
4258 * unfortunate state in our controller.
4260 if (parent != root_mem_cgroup)
4261 memory_cgrp_subsys.broken_hierarchy = true;
4263 mutex_unlock(&memcg_create_mutex);
4265 ret = memcg_propagate_kmem(memcg);
4270 #ifdef CONFIG_MEMCG_LEGACY_KMEM
4271 ret = tcp_init_cgroup(memcg);
4276 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4277 static_branch_inc(&memcg_sockets_enabled_key);
4281 * Make sure the memcg is initialized: mem_cgroup_iter()
4282 * orders reading memcg->initialized against its callers
4283 * reading the memcg members.
4285 smp_store_release(&memcg->initialized, 1);
4290 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4292 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4293 struct mem_cgroup_event *event, *tmp;
4296 * Unregister events and notify userspace.
4297 * Notify userspace about cgroup removing only after rmdir of cgroup
4298 * directory to avoid race between userspace and kernelspace.
4300 spin_lock(&memcg->event_list_lock);
4301 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4302 list_del_init(&event->list);
4303 schedule_work(&event->remove);
4305 spin_unlock(&memcg->event_list_lock);
4307 vmpressure_cleanup(&memcg->vmpressure);
4309 memcg_offline_kmem(memcg);
4311 wb_memcg_offline(memcg);
4314 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4316 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4318 invalidate_reclaim_iterators(memcg);
4321 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4323 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4326 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4327 static_branch_dec(&memcg_sockets_enabled_key);
4330 memcg_free_kmem(memcg);
4332 #if defined(CONFIG_MEMCG_LEGACY_KMEM) && defined(CONFIG_INET)
4333 tcp_destroy_cgroup(memcg);
4336 __mem_cgroup_free(memcg);
4340 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4341 * @css: the target css
4343 * Reset the states of the mem_cgroup associated with @css. This is
4344 * invoked when the userland requests disabling on the default hierarchy
4345 * but the memcg is pinned through dependency. The memcg should stop
4346 * applying policies and should revert to the vanilla state as it may be
4347 * made visible again.
4349 * The current implementation only resets the essential configurations.
4350 * This needs to be expanded to cover all the visible parts.
4352 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4354 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4356 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4357 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4358 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4360 memcg->high = PAGE_COUNTER_MAX;
4361 memcg->soft_limit = PAGE_COUNTER_MAX;
4362 memcg_wb_domain_size_changed(memcg);
4366 /* Handlers for move charge at task migration. */
4367 static int mem_cgroup_do_precharge(unsigned long count)
4371 /* Try a single bulk charge without reclaim first, kswapd may wake */
4372 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4374 mc.precharge += count;
4378 /* Try charges one by one with reclaim */
4380 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4390 * get_mctgt_type - get target type of moving charge
4391 * @vma: the vma the pte to be checked belongs
4392 * @addr: the address corresponding to the pte to be checked
4393 * @ptent: the pte to be checked
4394 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4397 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4398 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4399 * move charge. if @target is not NULL, the page is stored in target->page
4400 * with extra refcnt got(Callers should handle it).
4401 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4402 * target for charge migration. if @target is not NULL, the entry is stored
4405 * Called with pte lock held.
4412 enum mc_target_type {
4418 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4419 unsigned long addr, pte_t ptent)
4421 struct page *page = vm_normal_page(vma, addr, ptent);
4423 if (!page || !page_mapped(page))
4425 if (PageAnon(page)) {
4426 if (!(mc.flags & MOVE_ANON))
4429 if (!(mc.flags & MOVE_FILE))
4432 if (!get_page_unless_zero(page))
4439 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4440 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4442 struct page *page = NULL;
4443 swp_entry_t ent = pte_to_swp_entry(ptent);
4445 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4448 * Because lookup_swap_cache() updates some statistics counter,
4449 * we call find_get_page() with swapper_space directly.
4451 page = find_get_page(swap_address_space(ent), ent.val);
4452 if (do_memsw_account())
4453 entry->val = ent.val;
4458 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4459 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4465 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4466 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4468 struct page *page = NULL;
4469 struct address_space *mapping;
4472 if (!vma->vm_file) /* anonymous vma */
4474 if (!(mc.flags & MOVE_FILE))
4477 mapping = vma->vm_file->f_mapping;
4478 pgoff = linear_page_index(vma, addr);
4480 /* page is moved even if it's not RSS of this task(page-faulted). */
4482 /* shmem/tmpfs may report page out on swap: account for that too. */
4483 if (shmem_mapping(mapping)) {
4484 page = find_get_entry(mapping, pgoff);
4485 if (radix_tree_exceptional_entry(page)) {
4486 swp_entry_t swp = radix_to_swp_entry(page);
4487 if (do_memsw_account())
4489 page = find_get_page(swap_address_space(swp), swp.val);
4492 page = find_get_page(mapping, pgoff);
4494 page = find_get_page(mapping, pgoff);
4500 * mem_cgroup_move_account - move account of the page
4502 * @nr_pages: number of regular pages (>1 for huge pages)
4503 * @from: mem_cgroup which the page is moved from.
4504 * @to: mem_cgroup which the page is moved to. @from != @to.
4506 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4508 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4511 static int mem_cgroup_move_account(struct page *page,
4513 struct mem_cgroup *from,
4514 struct mem_cgroup *to)
4516 unsigned long flags;
4517 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4521 VM_BUG_ON(from == to);
4522 VM_BUG_ON_PAGE(PageLRU(page), page);
4523 VM_BUG_ON(compound && !PageTransHuge(page));
4526 * Prevent mem_cgroup_replace_page() from looking at
4527 * page->mem_cgroup of its source page while we change it.
4530 if (!trylock_page(page))
4534 if (page->mem_cgroup != from)
4537 anon = PageAnon(page);
4539 spin_lock_irqsave(&from->move_lock, flags);
4541 if (!anon && page_mapped(page)) {
4542 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4544 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4549 * move_lock grabbed above and caller set from->moving_account, so
4550 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4551 * So mapping should be stable for dirty pages.
4553 if (!anon && PageDirty(page)) {
4554 struct address_space *mapping = page_mapping(page);
4556 if (mapping_cap_account_dirty(mapping)) {
4557 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4559 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4564 if (PageWriteback(page)) {
4565 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4567 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4572 * It is safe to change page->mem_cgroup here because the page
4573 * is referenced, charged, and isolated - we can't race with
4574 * uncharging, charging, migration, or LRU putback.
4577 /* caller should have done css_get */
4578 page->mem_cgroup = to;
4579 spin_unlock_irqrestore(&from->move_lock, flags);
4583 local_irq_disable();
4584 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4585 memcg_check_events(to, page);
4586 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4587 memcg_check_events(from, page);
4595 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4596 unsigned long addr, pte_t ptent, union mc_target *target)
4598 struct page *page = NULL;
4599 enum mc_target_type ret = MC_TARGET_NONE;
4600 swp_entry_t ent = { .val = 0 };
4602 if (pte_present(ptent))
4603 page = mc_handle_present_pte(vma, addr, ptent);
4604 else if (is_swap_pte(ptent))
4605 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4606 else if (pte_none(ptent))
4607 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4609 if (!page && !ent.val)
4613 * Do only loose check w/o serialization.
4614 * mem_cgroup_move_account() checks the page is valid or
4615 * not under LRU exclusion.
4617 if (page->mem_cgroup == mc.from) {
4618 ret = MC_TARGET_PAGE;
4620 target->page = page;
4622 if (!ret || !target)
4625 /* There is a swap entry and a page doesn't exist or isn't charged */
4626 if (ent.val && !ret &&
4627 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4628 ret = MC_TARGET_SWAP;
4635 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4637 * We don't consider swapping or file mapped pages because THP does not
4638 * support them for now.
4639 * Caller should make sure that pmd_trans_huge(pmd) is true.
4641 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4642 unsigned long addr, pmd_t pmd, union mc_target *target)
4644 struct page *page = NULL;
4645 enum mc_target_type ret = MC_TARGET_NONE;
4647 page = pmd_page(pmd);
4648 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4649 if (!(mc.flags & MOVE_ANON))
4651 if (page->mem_cgroup == mc.from) {
4652 ret = MC_TARGET_PAGE;
4655 target->page = page;
4661 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4662 unsigned long addr, pmd_t pmd, union mc_target *target)
4664 return MC_TARGET_NONE;
4668 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4669 unsigned long addr, unsigned long end,
4670 struct mm_walk *walk)
4672 struct vm_area_struct *vma = walk->vma;
4676 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4677 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4678 mc.precharge += HPAGE_PMD_NR;
4683 if (pmd_trans_unstable(pmd))
4685 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4686 for (; addr != end; pte++, addr += PAGE_SIZE)
4687 if (get_mctgt_type(vma, addr, *pte, NULL))
4688 mc.precharge++; /* increment precharge temporarily */
4689 pte_unmap_unlock(pte - 1, ptl);
4695 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4697 unsigned long precharge;
4699 struct mm_walk mem_cgroup_count_precharge_walk = {
4700 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4703 down_read(&mm->mmap_sem);
4704 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4705 up_read(&mm->mmap_sem);
4707 precharge = mc.precharge;
4713 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4715 unsigned long precharge = mem_cgroup_count_precharge(mm);
4717 VM_BUG_ON(mc.moving_task);
4718 mc.moving_task = current;
4719 return mem_cgroup_do_precharge(precharge);
4722 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4723 static void __mem_cgroup_clear_mc(void)
4725 struct mem_cgroup *from = mc.from;
4726 struct mem_cgroup *to = mc.to;
4728 /* we must uncharge all the leftover precharges from mc.to */
4730 cancel_charge(mc.to, mc.precharge);
4734 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4735 * we must uncharge here.
4737 if (mc.moved_charge) {
4738 cancel_charge(mc.from, mc.moved_charge);
4739 mc.moved_charge = 0;
4741 /* we must fixup refcnts and charges */
4742 if (mc.moved_swap) {
4743 /* uncharge swap account from the old cgroup */
4744 if (!mem_cgroup_is_root(mc.from))
4745 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4748 * we charged both to->memory and to->memsw, so we
4749 * should uncharge to->memory.
4751 if (!mem_cgroup_is_root(mc.to))
4752 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4754 css_put_many(&mc.from->css, mc.moved_swap);
4756 /* we've already done css_get(mc.to) */
4759 memcg_oom_recover(from);
4760 memcg_oom_recover(to);
4761 wake_up_all(&mc.waitq);
4764 static void mem_cgroup_clear_mc(void)
4767 * we must clear moving_task before waking up waiters at the end of
4770 mc.moving_task = NULL;
4771 __mem_cgroup_clear_mc();
4772 spin_lock(&mc.lock);
4775 spin_unlock(&mc.lock);
4778 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4780 struct cgroup_subsys_state *css;
4781 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4782 struct mem_cgroup *from;
4783 struct task_struct *leader, *p;
4784 struct mm_struct *mm;
4785 unsigned long move_flags;
4788 /* charge immigration isn't supported on the default hierarchy */
4789 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4793 * Multi-process migrations only happen on the default hierarchy
4794 * where charge immigration is not used. Perform charge
4795 * immigration if @tset contains a leader and whine if there are
4799 cgroup_taskset_for_each_leader(leader, css, tset) {
4802 memcg = mem_cgroup_from_css(css);
4808 * We are now commited to this value whatever it is. Changes in this
4809 * tunable will only affect upcoming migrations, not the current one.
4810 * So we need to save it, and keep it going.
4812 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4816 from = mem_cgroup_from_task(p);
4818 VM_BUG_ON(from == memcg);
4820 mm = get_task_mm(p);
4823 /* We move charges only when we move a owner of the mm */
4824 if (mm->owner == p) {
4827 VM_BUG_ON(mc.precharge);
4828 VM_BUG_ON(mc.moved_charge);
4829 VM_BUG_ON(mc.moved_swap);
4831 spin_lock(&mc.lock);
4834 mc.flags = move_flags;
4835 spin_unlock(&mc.lock);
4836 /* We set mc.moving_task later */
4838 ret = mem_cgroup_precharge_mc(mm);
4840 mem_cgroup_clear_mc();
4846 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4849 mem_cgroup_clear_mc();
4852 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4853 unsigned long addr, unsigned long end,
4854 struct mm_walk *walk)
4857 struct vm_area_struct *vma = walk->vma;
4860 enum mc_target_type target_type;
4861 union mc_target target;
4864 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4865 if (mc.precharge < HPAGE_PMD_NR) {
4869 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4870 if (target_type == MC_TARGET_PAGE) {
4872 if (!isolate_lru_page(page)) {
4873 if (!mem_cgroup_move_account(page, true,
4875 mc.precharge -= HPAGE_PMD_NR;
4876 mc.moved_charge += HPAGE_PMD_NR;
4878 putback_lru_page(page);
4886 if (pmd_trans_unstable(pmd))
4889 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4890 for (; addr != end; addr += PAGE_SIZE) {
4891 pte_t ptent = *(pte++);
4897 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4898 case MC_TARGET_PAGE:
4901 * We can have a part of the split pmd here. Moving it
4902 * can be done but it would be too convoluted so simply
4903 * ignore such a partial THP and keep it in original
4904 * memcg. There should be somebody mapping the head.
4906 if (PageTransCompound(page))
4908 if (isolate_lru_page(page))
4910 if (!mem_cgroup_move_account(page, false,
4913 /* we uncharge from mc.from later. */
4916 putback_lru_page(page);
4917 put: /* get_mctgt_type() gets the page */
4920 case MC_TARGET_SWAP:
4922 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4924 /* we fixup refcnts and charges later. */
4932 pte_unmap_unlock(pte - 1, ptl);
4937 * We have consumed all precharges we got in can_attach().
4938 * We try charge one by one, but don't do any additional
4939 * charges to mc.to if we have failed in charge once in attach()
4942 ret = mem_cgroup_do_precharge(1);
4950 static void mem_cgroup_move_charge(struct mm_struct *mm)
4952 struct mm_walk mem_cgroup_move_charge_walk = {
4953 .pmd_entry = mem_cgroup_move_charge_pte_range,
4957 lru_add_drain_all();
4959 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4960 * move_lock while we're moving its pages to another memcg.
4961 * Then wait for already started RCU-only updates to finish.
4963 atomic_inc(&mc.from->moving_account);
4966 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4968 * Someone who are holding the mmap_sem might be waiting in
4969 * waitq. So we cancel all extra charges, wake up all waiters,
4970 * and retry. Because we cancel precharges, we might not be able
4971 * to move enough charges, but moving charge is a best-effort
4972 * feature anyway, so it wouldn't be a big problem.
4974 __mem_cgroup_clear_mc();
4979 * When we have consumed all precharges and failed in doing
4980 * additional charge, the page walk just aborts.
4982 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4983 up_read(&mm->mmap_sem);
4984 atomic_dec(&mc.from->moving_account);
4987 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4989 struct cgroup_subsys_state *css;
4990 struct task_struct *p = cgroup_taskset_first(tset, &css);
4991 struct mm_struct *mm = get_task_mm(p);
4995 mem_cgroup_move_charge(mm);
4999 mem_cgroup_clear_mc();
5001 #else /* !CONFIG_MMU */
5002 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5006 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5009 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5015 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5016 * to verify whether we're attached to the default hierarchy on each mount
5019 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5022 * use_hierarchy is forced on the default hierarchy. cgroup core
5023 * guarantees that @root doesn't have any children, so turning it
5024 * on for the root memcg is enough.
5026 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5027 root_mem_cgroup->use_hierarchy = true;
5029 root_mem_cgroup->use_hierarchy = false;
5032 static u64 memory_current_read(struct cgroup_subsys_state *css,
5035 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5037 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5040 static int memory_low_show(struct seq_file *m, void *v)
5042 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5043 unsigned long low = READ_ONCE(memcg->low);
5045 if (low == PAGE_COUNTER_MAX)
5046 seq_puts(m, "max\n");
5048 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5053 static ssize_t memory_low_write(struct kernfs_open_file *of,
5054 char *buf, size_t nbytes, loff_t off)
5056 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5060 buf = strstrip(buf);
5061 err = page_counter_memparse(buf, "max", &low);
5070 static int memory_high_show(struct seq_file *m, void *v)
5072 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5073 unsigned long high = READ_ONCE(memcg->high);
5075 if (high == PAGE_COUNTER_MAX)
5076 seq_puts(m, "max\n");
5078 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5083 static ssize_t memory_high_write(struct kernfs_open_file *of,
5084 char *buf, size_t nbytes, loff_t off)
5086 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5090 buf = strstrip(buf);
5091 err = page_counter_memparse(buf, "max", &high);
5097 memcg_wb_domain_size_changed(memcg);
5101 static int memory_max_show(struct seq_file *m, void *v)
5103 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5104 unsigned long max = READ_ONCE(memcg->memory.limit);
5106 if (max == PAGE_COUNTER_MAX)
5107 seq_puts(m, "max\n");
5109 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5114 static ssize_t memory_max_write(struct kernfs_open_file *of,
5115 char *buf, size_t nbytes, loff_t off)
5117 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5121 buf = strstrip(buf);
5122 err = page_counter_memparse(buf, "max", &max);
5126 err = mem_cgroup_resize_limit(memcg, max);
5130 memcg_wb_domain_size_changed(memcg);
5134 static int memory_events_show(struct seq_file *m, void *v)
5136 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5138 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5139 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5140 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5141 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5146 static struct cftype memory_files[] = {
5149 .flags = CFTYPE_NOT_ON_ROOT,
5150 .read_u64 = memory_current_read,
5154 .flags = CFTYPE_NOT_ON_ROOT,
5155 .seq_show = memory_low_show,
5156 .write = memory_low_write,
5160 .flags = CFTYPE_NOT_ON_ROOT,
5161 .seq_show = memory_high_show,
5162 .write = memory_high_write,
5166 .flags = CFTYPE_NOT_ON_ROOT,
5167 .seq_show = memory_max_show,
5168 .write = memory_max_write,
5172 .flags = CFTYPE_NOT_ON_ROOT,
5173 .file_offset = offsetof(struct mem_cgroup, events_file),
5174 .seq_show = memory_events_show,
5179 struct cgroup_subsys memory_cgrp_subsys = {
5180 .css_alloc = mem_cgroup_css_alloc,
5181 .css_online = mem_cgroup_css_online,
5182 .css_offline = mem_cgroup_css_offline,
5183 .css_released = mem_cgroup_css_released,
5184 .css_free = mem_cgroup_css_free,
5185 .css_reset = mem_cgroup_css_reset,
5186 .can_attach = mem_cgroup_can_attach,
5187 .cancel_attach = mem_cgroup_cancel_attach,
5188 .attach = mem_cgroup_move_task,
5189 .bind = mem_cgroup_bind,
5190 .dfl_cftypes = memory_files,
5191 .legacy_cftypes = mem_cgroup_legacy_files,
5196 * mem_cgroup_low - check if memory consumption is below the normal range
5197 * @root: the highest ancestor to consider
5198 * @memcg: the memory cgroup to check
5200 * Returns %true if memory consumption of @memcg, and that of all
5201 * configurable ancestors up to @root, is below the normal range.
5203 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5205 if (mem_cgroup_disabled())
5209 * The toplevel group doesn't have a configurable range, so
5210 * it's never low when looked at directly, and it is not
5211 * considered an ancestor when assessing the hierarchy.
5214 if (memcg == root_mem_cgroup)
5217 if (page_counter_read(&memcg->memory) >= memcg->low)
5220 while (memcg != root) {
5221 memcg = parent_mem_cgroup(memcg);
5223 if (memcg == root_mem_cgroup)
5226 if (page_counter_read(&memcg->memory) >= memcg->low)
5233 * mem_cgroup_try_charge - try charging a page
5234 * @page: page to charge
5235 * @mm: mm context of the victim
5236 * @gfp_mask: reclaim mode
5237 * @memcgp: charged memcg return
5239 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5240 * pages according to @gfp_mask if necessary.
5242 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5243 * Otherwise, an error code is returned.
5245 * After page->mapping has been set up, the caller must finalize the
5246 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5247 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5249 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5250 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5253 struct mem_cgroup *memcg = NULL;
5254 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5257 if (mem_cgroup_disabled())
5260 if (PageSwapCache(page)) {
5262 * Every swap fault against a single page tries to charge the
5263 * page, bail as early as possible. shmem_unuse() encounters
5264 * already charged pages, too. The USED bit is protected by
5265 * the page lock, which serializes swap cache removal, which
5266 * in turn serializes uncharging.
5268 VM_BUG_ON_PAGE(!PageLocked(page), page);
5269 if (page->mem_cgroup)
5272 if (do_memsw_account()) {
5273 swp_entry_t ent = { .val = page_private(page), };
5274 unsigned short id = lookup_swap_cgroup_id(ent);
5277 memcg = mem_cgroup_from_id(id);
5278 if (memcg && !css_tryget_online(&memcg->css))
5285 memcg = get_mem_cgroup_from_mm(mm);
5287 ret = try_charge(memcg, gfp_mask, nr_pages);
5289 css_put(&memcg->css);
5296 * mem_cgroup_commit_charge - commit a page charge
5297 * @page: page to charge
5298 * @memcg: memcg to charge the page to
5299 * @lrucare: page might be on LRU already
5301 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5302 * after page->mapping has been set up. This must happen atomically
5303 * as part of the page instantiation, i.e. under the page table lock
5304 * for anonymous pages, under the page lock for page and swap cache.
5306 * In addition, the page must not be on the LRU during the commit, to
5307 * prevent racing with task migration. If it might be, use @lrucare.
5309 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5311 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5312 bool lrucare, bool compound)
5314 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5316 VM_BUG_ON_PAGE(!page->mapping, page);
5317 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5319 if (mem_cgroup_disabled())
5322 * Swap faults will attempt to charge the same page multiple
5323 * times. But reuse_swap_page() might have removed the page
5324 * from swapcache already, so we can't check PageSwapCache().
5329 commit_charge(page, memcg, lrucare);
5331 local_irq_disable();
5332 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5333 memcg_check_events(memcg, page);
5336 if (do_memsw_account() && PageSwapCache(page)) {
5337 swp_entry_t entry = { .val = page_private(page) };
5339 * The swap entry might not get freed for a long time,
5340 * let's not wait for it. The page already received a
5341 * memory+swap charge, drop the swap entry duplicate.
5343 mem_cgroup_uncharge_swap(entry);
5348 * mem_cgroup_cancel_charge - cancel a page charge
5349 * @page: page to charge
5350 * @memcg: memcg to charge the page to
5352 * Cancel a charge transaction started by mem_cgroup_try_charge().
5354 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5357 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5359 if (mem_cgroup_disabled())
5362 * Swap faults will attempt to charge the same page multiple
5363 * times. But reuse_swap_page() might have removed the page
5364 * from swapcache already, so we can't check PageSwapCache().
5369 cancel_charge(memcg, nr_pages);
5372 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5373 unsigned long nr_anon, unsigned long nr_file,
5374 unsigned long nr_huge, struct page *dummy_page)
5376 unsigned long nr_pages = nr_anon + nr_file;
5377 unsigned long flags;
5379 if (!mem_cgroup_is_root(memcg)) {
5380 page_counter_uncharge(&memcg->memory, nr_pages);
5381 if (do_memsw_account())
5382 page_counter_uncharge(&memcg->memsw, nr_pages);
5383 memcg_oom_recover(memcg);
5386 local_irq_save(flags);
5387 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5388 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5389 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5390 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5391 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5392 memcg_check_events(memcg, dummy_page);
5393 local_irq_restore(flags);
5395 if (!mem_cgroup_is_root(memcg))
5396 css_put_many(&memcg->css, nr_pages);
5399 static void uncharge_list(struct list_head *page_list)
5401 struct mem_cgroup *memcg = NULL;
5402 unsigned long nr_anon = 0;
5403 unsigned long nr_file = 0;
5404 unsigned long nr_huge = 0;
5405 unsigned long pgpgout = 0;
5406 struct list_head *next;
5409 next = page_list->next;
5411 unsigned int nr_pages = 1;
5413 page = list_entry(next, struct page, lru);
5414 next = page->lru.next;
5416 VM_BUG_ON_PAGE(PageLRU(page), page);
5417 VM_BUG_ON_PAGE(page_count(page), page);
5419 if (!page->mem_cgroup)
5423 * Nobody should be changing or seriously looking at
5424 * page->mem_cgroup at this point, we have fully
5425 * exclusive access to the page.
5428 if (memcg != page->mem_cgroup) {
5430 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5432 pgpgout = nr_anon = nr_file = nr_huge = 0;
5434 memcg = page->mem_cgroup;
5437 if (PageTransHuge(page)) {
5438 nr_pages <<= compound_order(page);
5439 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5440 nr_huge += nr_pages;
5444 nr_anon += nr_pages;
5446 nr_file += nr_pages;
5448 page->mem_cgroup = NULL;
5451 } while (next != page_list);
5454 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5459 * mem_cgroup_uncharge - uncharge a page
5460 * @page: page to uncharge
5462 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5463 * mem_cgroup_commit_charge().
5465 void mem_cgroup_uncharge(struct page *page)
5467 if (mem_cgroup_disabled())
5470 /* Don't touch page->lru of any random page, pre-check: */
5471 if (!page->mem_cgroup)
5474 INIT_LIST_HEAD(&page->lru);
5475 uncharge_list(&page->lru);
5479 * mem_cgroup_uncharge_list - uncharge a list of page
5480 * @page_list: list of pages to uncharge
5482 * Uncharge a list of pages previously charged with
5483 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5485 void mem_cgroup_uncharge_list(struct list_head *page_list)
5487 if (mem_cgroup_disabled())
5490 if (!list_empty(page_list))
5491 uncharge_list(page_list);
5495 * mem_cgroup_replace_page - migrate a charge to another page
5496 * @oldpage: currently charged page
5497 * @newpage: page to transfer the charge to
5499 * Migrate the charge from @oldpage to @newpage.
5501 * Both pages must be locked, @newpage->mapping must be set up.
5502 * Either or both pages might be on the LRU already.
5504 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5506 struct mem_cgroup *memcg;
5509 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5510 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5511 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5512 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5515 if (mem_cgroup_disabled())
5518 /* Page cache replacement: new page already charged? */
5519 if (newpage->mem_cgroup)
5522 /* Swapcache readahead pages can get replaced before being charged */
5523 memcg = oldpage->mem_cgroup;
5527 lock_page_lru(oldpage, &isolated);
5528 oldpage->mem_cgroup = NULL;
5529 unlock_page_lru(oldpage, isolated);
5531 commit_charge(newpage, memcg, true);
5536 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5537 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5539 void sock_update_memcg(struct sock *sk)
5541 struct mem_cgroup *memcg;
5543 /* Socket cloning can throw us here with sk_cgrp already
5544 * filled. It won't however, necessarily happen from
5545 * process context. So the test for root memcg given
5546 * the current task's memcg won't help us in this case.
5548 * Respecting the original socket's memcg is a better
5549 * decision in this case.
5552 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5553 css_get(&sk->sk_memcg->css);
5558 memcg = mem_cgroup_from_task(current);
5559 if (memcg == root_mem_cgroup)
5561 #ifdef CONFIG_MEMCG_LEGACY_KMEM
5562 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcp_mem.active)
5565 if (css_tryget_online(&memcg->css))
5566 sk->sk_memcg = memcg;
5570 EXPORT_SYMBOL(sock_update_memcg);
5572 void sock_release_memcg(struct sock *sk)
5574 WARN_ON(!sk->sk_memcg);
5575 css_put(&sk->sk_memcg->css);
5579 * mem_cgroup_charge_skmem - charge socket memory
5580 * @memcg: memcg to charge
5581 * @nr_pages: number of pages to charge
5583 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5584 * @memcg's configured limit, %false if the charge had to be forced.
5586 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5588 gfp_t gfp_mask = GFP_KERNEL;
5590 #ifdef CONFIG_MEMCG_LEGACY_KMEM
5591 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5592 struct page_counter *counter;
5594 if (page_counter_try_charge(&memcg->tcp_mem.memory_allocated,
5595 nr_pages, &counter)) {
5596 memcg->tcp_mem.memory_pressure = 0;
5599 page_counter_charge(&memcg->tcp_mem.memory_allocated, nr_pages);
5600 memcg->tcp_mem.memory_pressure = 1;
5604 /* Don't block in the packet receive path */
5606 gfp_mask = GFP_NOWAIT;
5608 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5611 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5616 * mem_cgroup_uncharge_skmem - uncharge socket memory
5617 * @memcg - memcg to uncharge
5618 * @nr_pages - number of pages to uncharge
5620 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5622 #ifdef CONFIG_MEMCG_LEGACY_KMEM
5623 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5624 page_counter_uncharge(&memcg->tcp_mem.memory_allocated,
5629 page_counter_uncharge(&memcg->memory, nr_pages);
5630 css_put_many(&memcg->css, nr_pages);
5633 #endif /* CONFIG_INET */
5635 static int __init cgroup_memory(char *s)
5639 while ((token = strsep(&s, ",")) != NULL) {
5642 if (!strcmp(token, "nosocket"))
5643 cgroup_memory_nosocket = true;
5644 if (!strcmp(token, "nokmem"))
5645 cgroup_memory_nokmem = true;
5649 __setup("cgroup.memory=", cgroup_memory);
5652 * subsys_initcall() for memory controller.
5654 * Some parts like hotcpu_notifier() have to be initialized from this context
5655 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5656 * everything that doesn't depend on a specific mem_cgroup structure should
5657 * be initialized from here.
5659 static int __init mem_cgroup_init(void)
5663 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5665 for_each_possible_cpu(cpu)
5666 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5669 for_each_node(node) {
5670 struct mem_cgroup_tree_per_node *rtpn;
5673 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5674 node_online(node) ? node : NUMA_NO_NODE);
5676 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5677 struct mem_cgroup_tree_per_zone *rtpz;
5679 rtpz = &rtpn->rb_tree_per_zone[zone];
5680 rtpz->rb_root = RB_ROOT;
5681 spin_lock_init(&rtpz->lock);
5683 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5688 subsys_initcall(mem_cgroup_init);
5690 #ifdef CONFIG_MEMCG_SWAP
5692 * mem_cgroup_swapout - transfer a memsw charge to swap
5693 * @page: page whose memsw charge to transfer
5694 * @entry: swap entry to move the charge to
5696 * Transfer the memsw charge of @page to @entry.
5698 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5700 struct mem_cgroup *memcg;
5701 unsigned short oldid;
5703 VM_BUG_ON_PAGE(PageLRU(page), page);
5704 VM_BUG_ON_PAGE(page_count(page), page);
5706 if (!do_memsw_account())
5709 memcg = page->mem_cgroup;
5711 /* Readahead page, never charged */
5715 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5716 VM_BUG_ON_PAGE(oldid, page);
5717 mem_cgroup_swap_statistics(memcg, true);
5719 page->mem_cgroup = NULL;
5721 if (!mem_cgroup_is_root(memcg))
5722 page_counter_uncharge(&memcg->memory, 1);
5725 * Interrupts should be disabled here because the caller holds the
5726 * mapping->tree_lock lock which is taken with interrupts-off. It is
5727 * important here to have the interrupts disabled because it is the
5728 * only synchronisation we have for udpating the per-CPU variables.
5730 VM_BUG_ON(!irqs_disabled());
5731 mem_cgroup_charge_statistics(memcg, page, false, -1);
5732 memcg_check_events(memcg, page);
5736 * mem_cgroup_uncharge_swap - uncharge a swap entry
5737 * @entry: swap entry to uncharge
5739 * Drop the memsw charge associated with @entry.
5741 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5743 struct mem_cgroup *memcg;
5746 if (!do_memsw_account())
5749 id = swap_cgroup_record(entry, 0);
5751 memcg = mem_cgroup_from_id(id);
5753 if (!mem_cgroup_is_root(memcg))
5754 page_counter_uncharge(&memcg->memsw, 1);
5755 mem_cgroup_swap_statistics(memcg, false);
5756 css_put(&memcg->css);
5761 /* for remember boot option*/
5762 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5763 static int really_do_swap_account __initdata = 1;
5765 static int really_do_swap_account __initdata;
5768 static int __init enable_swap_account(char *s)
5770 if (!strcmp(s, "1"))
5771 really_do_swap_account = 1;
5772 else if (!strcmp(s, "0"))
5773 really_do_swap_account = 0;
5776 __setup("swapaccount=", enable_swap_account);
5778 static struct cftype memsw_cgroup_files[] = {
5780 .name = "memsw.usage_in_bytes",
5781 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5782 .read_u64 = mem_cgroup_read_u64,
5785 .name = "memsw.max_usage_in_bytes",
5786 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5787 .write = mem_cgroup_reset,
5788 .read_u64 = mem_cgroup_read_u64,
5791 .name = "memsw.limit_in_bytes",
5792 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5793 .write = mem_cgroup_write,
5794 .read_u64 = mem_cgroup_read_u64,
5797 .name = "memsw.failcnt",
5798 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5799 .write = mem_cgroup_reset,
5800 .read_u64 = mem_cgroup_read_u64,
5802 { }, /* terminate */
5805 static int __init mem_cgroup_swap_init(void)
5807 if (!mem_cgroup_disabled() && really_do_swap_account) {
5808 do_swap_account = 1;
5809 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5810 memsw_cgroup_files));
5814 subsys_initcall(mem_cgroup_swap_init);
5816 #endif /* CONFIG_MEMCG_SWAP */