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>
70 #include <linux/locallock.h>
74 #include <asm/uaccess.h>
76 #include <trace/events/vmscan.h>
78 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
79 EXPORT_SYMBOL(memory_cgrp_subsys);
81 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 static struct mem_cgroup *root_mem_cgroup __read_mostly;
83 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 int do_swap_account __read_mostly;
89 #define do_swap_account 0
92 static DEFINE_LOCAL_IRQ_LOCK(event_lock);
93 static const char * const mem_cgroup_stat_names[] = {
103 static const char * const mem_cgroup_events_names[] = {
110 static const char * const mem_cgroup_lru_names[] = {
118 #define THRESHOLDS_EVENTS_TARGET 128
119 #define SOFTLIMIT_EVENTS_TARGET 1024
120 #define NUMAINFO_EVENTS_TARGET 1024
123 * Cgroups above their limits are maintained in a RB-Tree, independent of
124 * their hierarchy representation
127 struct mem_cgroup_tree_per_zone {
128 struct rb_root rb_root;
132 struct mem_cgroup_tree_per_node {
133 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
136 struct mem_cgroup_tree {
137 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
140 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
143 struct mem_cgroup_eventfd_list {
144 struct list_head list;
145 struct eventfd_ctx *eventfd;
149 * cgroup_event represents events which userspace want to receive.
151 struct mem_cgroup_event {
153 * memcg which the event belongs to.
155 struct mem_cgroup *memcg;
157 * eventfd to signal userspace about the event.
159 struct eventfd_ctx *eventfd;
161 * Each of these stored in a list by the cgroup.
163 struct list_head list;
165 * register_event() callback will be used to add new userspace
166 * waiter for changes related to this event. Use eventfd_signal()
167 * on eventfd to send notification to userspace.
169 int (*register_event)(struct mem_cgroup *memcg,
170 struct eventfd_ctx *eventfd, const char *args);
172 * unregister_event() callback will be called when userspace closes
173 * the eventfd or on cgroup removing. This callback must be set,
174 * if you want provide notification functionality.
176 void (*unregister_event)(struct mem_cgroup *memcg,
177 struct eventfd_ctx *eventfd);
179 * All fields below needed to unregister event when
180 * userspace closes eventfd.
183 wait_queue_head_t *wqh;
185 struct work_struct remove;
188 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
189 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
191 /* Stuffs for move charges at task migration. */
193 * Types of charges to be moved.
195 #define MOVE_ANON 0x1U
196 #define MOVE_FILE 0x2U
197 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
199 /* "mc" and its members are protected by cgroup_mutex */
200 static struct move_charge_struct {
201 spinlock_t lock; /* for from, to */
202 struct mm_struct *mm;
203 struct mem_cgroup *from;
204 struct mem_cgroup *to;
206 unsigned long precharge;
207 unsigned long moved_charge;
208 unsigned long moved_swap;
209 struct task_struct *moving_task; /* a task moving charges */
210 wait_queue_head_t waitq; /* a waitq for other context */
212 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
213 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
217 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
218 * limit reclaim to prevent infinite loops, if they ever occur.
220 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
221 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
224 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
225 MEM_CGROUP_CHARGE_TYPE_ANON,
226 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
227 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
231 /* for encoding cft->private value on file */
239 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
240 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
241 #define MEMFILE_ATTR(val) ((val) & 0xffff)
242 /* Used for OOM nofiier */
243 #define OOM_CONTROL (0)
246 * The memcg_create_mutex will be held whenever a new cgroup is created.
247 * As a consequence, any change that needs to protect against new child cgroups
248 * appearing has to hold it as well.
250 static DEFINE_MUTEX(memcg_create_mutex);
252 /* Some nice accessors for the vmpressure. */
253 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
256 memcg = root_mem_cgroup;
257 return &memcg->vmpressure;
260 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
262 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
265 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
267 return (memcg == root_mem_cgroup);
271 * We restrict the id in the range of [1, 65535], so it can fit into
274 #define MEM_CGROUP_ID_MAX USHRT_MAX
276 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
281 /* Writing them here to avoid exposing memcg's inner layout */
282 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
284 void sock_update_memcg(struct sock *sk)
286 if (mem_cgroup_sockets_enabled) {
287 struct mem_cgroup *memcg;
288 struct cg_proto *cg_proto;
290 BUG_ON(!sk->sk_prot->proto_cgroup);
292 /* Socket cloning can throw us here with sk_cgrp already
293 * filled. It won't however, necessarily happen from
294 * process context. So the test for root memcg given
295 * the current task's memcg won't help us in this case.
297 * Respecting the original socket's memcg is a better
298 * decision in this case.
301 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
302 css_get(&sk->sk_cgrp->memcg->css);
307 memcg = mem_cgroup_from_task(current);
308 cg_proto = sk->sk_prot->proto_cgroup(memcg);
309 if (cg_proto && test_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags) &&
310 css_tryget_online(&memcg->css)) {
311 sk->sk_cgrp = cg_proto;
316 EXPORT_SYMBOL(sock_update_memcg);
318 void sock_release_memcg(struct sock *sk)
320 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
321 struct mem_cgroup *memcg;
322 WARN_ON(!sk->sk_cgrp->memcg);
323 memcg = sk->sk_cgrp->memcg;
324 css_put(&sk->sk_cgrp->memcg->css);
328 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
330 if (!memcg || mem_cgroup_is_root(memcg))
333 return &memcg->tcp_mem;
335 EXPORT_SYMBOL(tcp_proto_cgroup);
339 #ifdef CONFIG_MEMCG_KMEM
341 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
342 * The main reason for not using cgroup id for this:
343 * this works better in sparse environments, where we have a lot of memcgs,
344 * but only a few kmem-limited. Or also, if we have, for instance, 200
345 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
346 * 200 entry array for that.
348 * The current size of the caches array is stored in memcg_nr_cache_ids. It
349 * will double each time we have to increase it.
351 static DEFINE_IDA(memcg_cache_ida);
352 int memcg_nr_cache_ids;
354 /* Protects memcg_nr_cache_ids */
355 static DECLARE_RWSEM(memcg_cache_ids_sem);
357 void memcg_get_cache_ids(void)
359 down_read(&memcg_cache_ids_sem);
362 void memcg_put_cache_ids(void)
364 up_read(&memcg_cache_ids_sem);
368 * MIN_SIZE is different than 1, because we would like to avoid going through
369 * the alloc/free process all the time. In a small machine, 4 kmem-limited
370 * cgroups is a reasonable guess. In the future, it could be a parameter or
371 * tunable, but that is strictly not necessary.
373 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
374 * this constant directly from cgroup, but it is understandable that this is
375 * better kept as an internal representation in cgroup.c. In any case, the
376 * cgrp_id space is not getting any smaller, and we don't have to necessarily
377 * increase ours as well if it increases.
379 #define MEMCG_CACHES_MIN_SIZE 4
380 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
383 * A lot of the calls to the cache allocation functions are expected to be
384 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
385 * conditional to this static branch, we'll have to allow modules that does
386 * kmem_cache_alloc and the such to see this symbol as well
388 struct static_key memcg_kmem_enabled_key;
389 EXPORT_SYMBOL(memcg_kmem_enabled_key);
391 #endif /* CONFIG_MEMCG_KMEM */
393 static struct mem_cgroup_per_zone *
394 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
396 int nid = zone_to_nid(zone);
397 int zid = zone_idx(zone);
399 return &memcg->nodeinfo[nid]->zoneinfo[zid];
403 * mem_cgroup_css_from_page - css of the memcg associated with a page
404 * @page: page of interest
406 * If memcg is bound to the default hierarchy, css of the memcg associated
407 * with @page is returned. The returned css remains associated with @page
408 * until it is released.
410 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
413 * XXX: The above description of behavior on the default hierarchy isn't
414 * strictly true yet as replace_page_cache_page() can modify the
415 * association before @page is released even on the default hierarchy;
416 * however, the current and planned usages don't mix the the two functions
417 * and replace_page_cache_page() will soon be updated to make the invariant
420 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
422 struct mem_cgroup *memcg;
426 memcg = page->mem_cgroup;
428 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
429 memcg = root_mem_cgroup;
436 * page_cgroup_ino - return inode number of the memcg a page is charged to
439 * Look up the closest online ancestor of the memory cgroup @page is charged to
440 * and return its inode number or 0 if @page is not charged to any cgroup. It
441 * is safe to call this function without holding a reference to @page.
443 * Note, this function is inherently racy, because there is nothing to prevent
444 * the cgroup inode from getting torn down and potentially reallocated a moment
445 * after page_cgroup_ino() returns, so it only should be used by callers that
446 * do not care (such as procfs interfaces).
448 ino_t page_cgroup_ino(struct page *page)
450 struct mem_cgroup *memcg;
451 unsigned long ino = 0;
454 memcg = READ_ONCE(page->mem_cgroup);
455 while (memcg && !(memcg->css.flags & CSS_ONLINE))
456 memcg = parent_mem_cgroup(memcg);
458 ino = cgroup_ino(memcg->css.cgroup);
463 static struct mem_cgroup_per_zone *
464 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
466 int nid = page_to_nid(page);
467 int zid = page_zonenum(page);
469 return &memcg->nodeinfo[nid]->zoneinfo[zid];
472 static struct mem_cgroup_tree_per_zone *
473 soft_limit_tree_node_zone(int nid, int zid)
475 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
478 static struct mem_cgroup_tree_per_zone *
479 soft_limit_tree_from_page(struct page *page)
481 int nid = page_to_nid(page);
482 int zid = page_zonenum(page);
484 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
487 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
488 struct mem_cgroup_tree_per_zone *mctz,
489 unsigned long new_usage_in_excess)
491 struct rb_node **p = &mctz->rb_root.rb_node;
492 struct rb_node *parent = NULL;
493 struct mem_cgroup_per_zone *mz_node;
498 mz->usage_in_excess = new_usage_in_excess;
499 if (!mz->usage_in_excess)
503 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
505 if (mz->usage_in_excess < mz_node->usage_in_excess)
508 * We can't avoid mem cgroups that are over their soft
509 * limit by the same amount
511 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
514 rb_link_node(&mz->tree_node, parent, p);
515 rb_insert_color(&mz->tree_node, &mctz->rb_root);
519 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
520 struct mem_cgroup_tree_per_zone *mctz)
524 rb_erase(&mz->tree_node, &mctz->rb_root);
528 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
529 struct mem_cgroup_tree_per_zone *mctz)
533 spin_lock_irqsave(&mctz->lock, flags);
534 __mem_cgroup_remove_exceeded(mz, mctz);
535 spin_unlock_irqrestore(&mctz->lock, flags);
538 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
540 unsigned long nr_pages = page_counter_read(&memcg->memory);
541 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
542 unsigned long excess = 0;
544 if (nr_pages > soft_limit)
545 excess = nr_pages - soft_limit;
550 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
552 unsigned long excess;
553 struct mem_cgroup_per_zone *mz;
554 struct mem_cgroup_tree_per_zone *mctz;
556 mctz = soft_limit_tree_from_page(page);
558 * Necessary to update all ancestors when hierarchy is used.
559 * because their event counter is not touched.
561 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
562 mz = mem_cgroup_page_zoneinfo(memcg, page);
563 excess = soft_limit_excess(memcg);
565 * We have to update the tree if mz is on RB-tree or
566 * mem is over its softlimit.
568 if (excess || mz->on_tree) {
571 spin_lock_irqsave(&mctz->lock, flags);
572 /* if on-tree, remove it */
574 __mem_cgroup_remove_exceeded(mz, mctz);
576 * Insert again. mz->usage_in_excess will be updated.
577 * If excess is 0, no tree ops.
579 __mem_cgroup_insert_exceeded(mz, mctz, excess);
580 spin_unlock_irqrestore(&mctz->lock, flags);
585 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
587 struct mem_cgroup_tree_per_zone *mctz;
588 struct mem_cgroup_per_zone *mz;
592 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
593 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
594 mctz = soft_limit_tree_node_zone(nid, zid);
595 mem_cgroup_remove_exceeded(mz, mctz);
600 static struct mem_cgroup_per_zone *
601 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
603 struct rb_node *rightmost = NULL;
604 struct mem_cgroup_per_zone *mz;
608 rightmost = rb_last(&mctz->rb_root);
610 goto done; /* Nothing to reclaim from */
612 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
614 * Remove the node now but someone else can add it back,
615 * we will to add it back at the end of reclaim to its correct
616 * position in the tree.
618 __mem_cgroup_remove_exceeded(mz, mctz);
619 if (!soft_limit_excess(mz->memcg) ||
620 !css_tryget_online(&mz->memcg->css))
626 static struct mem_cgroup_per_zone *
627 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
629 struct mem_cgroup_per_zone *mz;
631 spin_lock_irq(&mctz->lock);
632 mz = __mem_cgroup_largest_soft_limit_node(mctz);
633 spin_unlock_irq(&mctz->lock);
638 * Return page count for single (non recursive) @memcg.
640 * Implementation Note: reading percpu statistics for memcg.
642 * Both of vmstat[] and percpu_counter has threshold and do periodic
643 * synchronization to implement "quick" read. There are trade-off between
644 * reading cost and precision of value. Then, we may have a chance to implement
645 * a periodic synchronization of counter in memcg's counter.
647 * But this _read() function is used for user interface now. The user accounts
648 * memory usage by memory cgroup and he _always_ requires exact value because
649 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
650 * have to visit all online cpus and make sum. So, for now, unnecessary
651 * synchronization is not implemented. (just implemented for cpu hotplug)
653 * If there are kernel internal actions which can make use of some not-exact
654 * value, and reading all cpu value can be performance bottleneck in some
655 * common workload, threshold and synchronization as vmstat[] should be
659 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
664 /* Per-cpu values can be negative, use a signed accumulator */
665 for_each_possible_cpu(cpu)
666 val += per_cpu(memcg->stat->count[idx], cpu);
668 * Summing races with updates, so val may be negative. Avoid exposing
669 * transient negative values.
676 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
677 enum mem_cgroup_events_index idx)
679 unsigned long val = 0;
682 for_each_possible_cpu(cpu)
683 val += per_cpu(memcg->stat->events[idx], cpu);
687 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
692 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
693 * counted as CACHE even if it's on ANON LRU.
696 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
699 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
702 if (PageTransHuge(page))
703 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
706 /* pagein of a big page is an event. So, ignore page size */
708 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
710 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
711 nr_pages = -nr_pages; /* for event */
714 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
717 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
719 unsigned int lru_mask)
721 unsigned long nr = 0;
724 VM_BUG_ON((unsigned)nid >= nr_node_ids);
726 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
727 struct mem_cgroup_per_zone *mz;
731 if (!(BIT(lru) & lru_mask))
733 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
734 nr += mz->lru_size[lru];
740 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
741 unsigned int lru_mask)
743 unsigned long nr = 0;
746 for_each_node_state(nid, N_MEMORY)
747 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
751 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
752 enum mem_cgroup_events_target target)
754 unsigned long val, next;
756 val = __this_cpu_read(memcg->stat->nr_page_events);
757 next = __this_cpu_read(memcg->stat->targets[target]);
758 /* from time_after() in jiffies.h */
759 if ((long)next - (long)val < 0) {
761 case MEM_CGROUP_TARGET_THRESH:
762 next = val + THRESHOLDS_EVENTS_TARGET;
764 case MEM_CGROUP_TARGET_SOFTLIMIT:
765 next = val + SOFTLIMIT_EVENTS_TARGET;
767 case MEM_CGROUP_TARGET_NUMAINFO:
768 next = val + NUMAINFO_EVENTS_TARGET;
773 __this_cpu_write(memcg->stat->targets[target], next);
780 * Check events in order.
783 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
785 /* threshold event is triggered in finer grain than soft limit */
786 if (unlikely(mem_cgroup_event_ratelimit(memcg,
787 MEM_CGROUP_TARGET_THRESH))) {
789 bool do_numainfo __maybe_unused;
791 do_softlimit = mem_cgroup_event_ratelimit(memcg,
792 MEM_CGROUP_TARGET_SOFTLIMIT);
794 do_numainfo = mem_cgroup_event_ratelimit(memcg,
795 MEM_CGROUP_TARGET_NUMAINFO);
797 mem_cgroup_threshold(memcg);
798 if (unlikely(do_softlimit))
799 mem_cgroup_update_tree(memcg, page);
801 if (unlikely(do_numainfo))
802 atomic_inc(&memcg->numainfo_events);
807 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
810 * mm_update_next_owner() may clear mm->owner to NULL
811 * if it races with swapoff, page migration, etc.
812 * So this can be called with p == NULL.
817 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
819 EXPORT_SYMBOL(mem_cgroup_from_task);
821 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
823 struct mem_cgroup *memcg = NULL;
828 * Page cache insertions can happen withou an
829 * actual mm context, e.g. during disk probing
830 * on boot, loopback IO, acct() writes etc.
833 memcg = root_mem_cgroup;
835 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
836 if (unlikely(!memcg))
837 memcg = root_mem_cgroup;
839 } while (!css_tryget_online(&memcg->css));
845 * mem_cgroup_iter - iterate over memory cgroup hierarchy
846 * @root: hierarchy root
847 * @prev: previously returned memcg, NULL on first invocation
848 * @reclaim: cookie for shared reclaim walks, NULL for full walks
850 * Returns references to children of the hierarchy below @root, or
851 * @root itself, or %NULL after a full round-trip.
853 * Caller must pass the return value in @prev on subsequent
854 * invocations for reference counting, or use mem_cgroup_iter_break()
855 * to cancel a hierarchy walk before the round-trip is complete.
857 * Reclaimers can specify a zone and a priority level in @reclaim to
858 * divide up the memcgs in the hierarchy among all concurrent
859 * reclaimers operating on the same zone and priority.
861 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
862 struct mem_cgroup *prev,
863 struct mem_cgroup_reclaim_cookie *reclaim)
865 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
866 struct cgroup_subsys_state *css = NULL;
867 struct mem_cgroup *memcg = NULL;
868 struct mem_cgroup *pos = NULL;
870 if (mem_cgroup_disabled())
874 root = root_mem_cgroup;
876 if (prev && !reclaim)
879 if (!root->use_hierarchy && root != root_mem_cgroup) {
888 struct mem_cgroup_per_zone *mz;
890 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
891 iter = &mz->iter[reclaim->priority];
893 if (prev && reclaim->generation != iter->generation)
897 pos = READ_ONCE(iter->position);
898 if (!pos || css_tryget(&pos->css))
901 * css reference reached zero, so iter->position will
902 * be cleared by ->css_released. However, we should not
903 * rely on this happening soon, because ->css_released
904 * is called from a work queue, and by busy-waiting we
905 * might block it. So we clear iter->position right
908 (void)cmpxchg(&iter->position, pos, NULL);
916 css = css_next_descendant_pre(css, &root->css);
919 * Reclaimers share the hierarchy walk, and a
920 * new one might jump in right at the end of
921 * the hierarchy - make sure they see at least
922 * one group and restart from the beginning.
930 * Verify the css and acquire a reference. The root
931 * is provided by the caller, so we know it's alive
932 * and kicking, and don't take an extra reference.
934 memcg = mem_cgroup_from_css(css);
936 if (css == &root->css)
939 if (css_tryget(css)) {
941 * Make sure the memcg is initialized:
942 * mem_cgroup_css_online() orders the the
943 * initialization against setting the flag.
945 if (smp_load_acquire(&memcg->initialized))
956 * The position could have already been updated by a competing
957 * thread, so check that the value hasn't changed since we read
958 * it to avoid reclaiming from the same cgroup twice.
960 (void)cmpxchg(&iter->position, pos, memcg);
968 reclaim->generation = iter->generation;
974 if (prev && prev != root)
981 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
982 * @root: hierarchy root
983 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
985 void mem_cgroup_iter_break(struct mem_cgroup *root,
986 struct mem_cgroup *prev)
989 root = root_mem_cgroup;
990 if (prev && prev != root)
994 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
996 struct mem_cgroup *memcg = dead_memcg;
997 struct mem_cgroup_reclaim_iter *iter;
998 struct mem_cgroup_per_zone *mz;
1002 while ((memcg = parent_mem_cgroup(memcg))) {
1003 for_each_node(nid) {
1004 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1005 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1006 for (i = 0; i <= DEF_PRIORITY; i++) {
1007 iter = &mz->iter[i];
1008 cmpxchg(&iter->position,
1017 * Iteration constructs for visiting all cgroups (under a tree). If
1018 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1019 * be used for reference counting.
1021 #define for_each_mem_cgroup_tree(iter, root) \
1022 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1024 iter = mem_cgroup_iter(root, iter, NULL))
1026 #define for_each_mem_cgroup(iter) \
1027 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1029 iter = mem_cgroup_iter(NULL, iter, NULL))
1032 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1033 * @zone: zone of the wanted lruvec
1034 * @memcg: memcg of the wanted lruvec
1036 * Returns the lru list vector holding pages for the given @zone and
1037 * @mem. This can be the global zone lruvec, if the memory controller
1040 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1041 struct mem_cgroup *memcg)
1043 struct mem_cgroup_per_zone *mz;
1044 struct lruvec *lruvec;
1046 if (mem_cgroup_disabled()) {
1047 lruvec = &zone->lruvec;
1051 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1052 lruvec = &mz->lruvec;
1055 * Since a node can be onlined after the mem_cgroup was created,
1056 * we have to be prepared to initialize lruvec->zone here;
1057 * and if offlined then reonlined, we need to reinitialize it.
1059 if (unlikely(lruvec->zone != zone))
1060 lruvec->zone = zone;
1065 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1067 * @zone: zone of the page
1069 * This function is only safe when following the LRU page isolation
1070 * and putback protocol: the LRU lock must be held, and the page must
1071 * either be PageLRU() or the caller must have isolated/allocated it.
1073 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1075 struct mem_cgroup_per_zone *mz;
1076 struct mem_cgroup *memcg;
1077 struct lruvec *lruvec;
1079 if (mem_cgroup_disabled()) {
1080 lruvec = &zone->lruvec;
1084 memcg = page->mem_cgroup;
1086 * Swapcache readahead pages are added to the LRU - and
1087 * possibly migrated - before they are charged.
1090 memcg = root_mem_cgroup;
1092 mz = mem_cgroup_page_zoneinfo(memcg, page);
1093 lruvec = &mz->lruvec;
1096 * Since a node can be onlined after the mem_cgroup was created,
1097 * we have to be prepared to initialize lruvec->zone here;
1098 * and if offlined then reonlined, we need to reinitialize it.
1100 if (unlikely(lruvec->zone != zone))
1101 lruvec->zone = zone;
1106 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1107 * @lruvec: mem_cgroup per zone lru vector
1108 * @lru: index of lru list the page is sitting on
1109 * @nr_pages: positive when adding or negative when removing
1111 * This function must be called when a page is added to or removed from an
1114 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1117 struct mem_cgroup_per_zone *mz;
1118 unsigned long *lru_size;
1120 if (mem_cgroup_disabled())
1123 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1124 lru_size = mz->lru_size + lru;
1125 *lru_size += nr_pages;
1126 VM_BUG_ON((long)(*lru_size) < 0);
1129 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1131 struct mem_cgroup *task_memcg;
1132 struct task_struct *p;
1135 p = find_lock_task_mm(task);
1137 task_memcg = get_mem_cgroup_from_mm(p->mm);
1141 * All threads may have already detached their mm's, but the oom
1142 * killer still needs to detect if they have already been oom
1143 * killed to prevent needlessly killing additional tasks.
1146 task_memcg = mem_cgroup_from_task(task);
1147 css_get(&task_memcg->css);
1150 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1151 css_put(&task_memcg->css);
1155 #define mem_cgroup_from_counter(counter, member) \
1156 container_of(counter, struct mem_cgroup, member)
1159 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1160 * @memcg: the memory cgroup
1162 * Returns the maximum amount of memory @mem can be charged with, in
1165 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1167 unsigned long margin = 0;
1168 unsigned long count;
1169 unsigned long limit;
1171 count = page_counter_read(&memcg->memory);
1172 limit = READ_ONCE(memcg->memory.limit);
1174 margin = limit - count;
1176 if (do_swap_account) {
1177 count = page_counter_read(&memcg->memsw);
1178 limit = READ_ONCE(memcg->memsw.limit);
1180 margin = min(margin, limit - count);
1187 * A routine for checking "mem" is under move_account() or not.
1189 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1190 * moving cgroups. This is for waiting at high-memory pressure
1193 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1195 struct mem_cgroup *from;
1196 struct mem_cgroup *to;
1199 * Unlike task_move routines, we access mc.to, mc.from not under
1200 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1202 spin_lock(&mc.lock);
1208 ret = mem_cgroup_is_descendant(from, memcg) ||
1209 mem_cgroup_is_descendant(to, memcg);
1211 spin_unlock(&mc.lock);
1215 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1217 if (mc.moving_task && current != mc.moving_task) {
1218 if (mem_cgroup_under_move(memcg)) {
1220 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1221 /* moving charge context might have finished. */
1224 finish_wait(&mc.waitq, &wait);
1231 #define K(x) ((x) << (PAGE_SHIFT-10))
1233 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1234 * @memcg: The memory cgroup that went over limit
1235 * @p: Task that is going to be killed
1237 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1240 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1242 /* oom_info_lock ensures that parallel ooms do not interleave */
1243 static DEFINE_MUTEX(oom_info_lock);
1244 struct mem_cgroup *iter;
1247 mutex_lock(&oom_info_lock);
1251 pr_info("Task in ");
1252 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1253 pr_cont(" killed as a result of limit of ");
1255 pr_info("Memory limit reached of cgroup ");
1258 pr_cont_cgroup_path(memcg->css.cgroup);
1263 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1264 K((u64)page_counter_read(&memcg->memory)),
1265 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1266 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1267 K((u64)page_counter_read(&memcg->memsw)),
1268 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1269 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1270 K((u64)page_counter_read(&memcg->kmem)),
1271 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1273 for_each_mem_cgroup_tree(iter, memcg) {
1274 pr_info("Memory cgroup stats for ");
1275 pr_cont_cgroup_path(iter->css.cgroup);
1278 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1279 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1281 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1282 K(mem_cgroup_read_stat(iter, i)));
1285 for (i = 0; i < NR_LRU_LISTS; i++)
1286 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1287 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1291 mutex_unlock(&oom_info_lock);
1295 * This function returns the number of memcg under hierarchy tree. Returns
1296 * 1(self count) if no children.
1298 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1301 struct mem_cgroup *iter;
1303 for_each_mem_cgroup_tree(iter, memcg)
1309 * Return the memory (and swap, if configured) limit for a memcg.
1311 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1313 unsigned long limit;
1315 limit = memcg->memory.limit;
1316 if (mem_cgroup_swappiness(memcg)) {
1317 unsigned long memsw_limit;
1319 memsw_limit = memcg->memsw.limit;
1320 limit = min(limit + total_swap_pages, memsw_limit);
1325 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1328 struct oom_control oc = {
1331 .gfp_mask = gfp_mask,
1334 struct mem_cgroup *iter;
1335 unsigned long chosen_points = 0;
1336 unsigned long totalpages;
1337 unsigned int points = 0;
1338 struct task_struct *chosen = NULL;
1340 mutex_lock(&oom_lock);
1343 * If current has a pending SIGKILL or is exiting, then automatically
1344 * select it. The goal is to allow it to allocate so that it may
1345 * quickly exit and free its memory.
1347 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1348 mark_oom_victim(current);
1352 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1353 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1354 for_each_mem_cgroup_tree(iter, memcg) {
1355 struct css_task_iter it;
1356 struct task_struct *task;
1358 css_task_iter_start(&iter->css, &it);
1359 while ((task = css_task_iter_next(&it))) {
1360 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1361 case OOM_SCAN_SELECT:
1363 put_task_struct(chosen);
1365 chosen_points = ULONG_MAX;
1366 get_task_struct(chosen);
1368 case OOM_SCAN_CONTINUE:
1370 case OOM_SCAN_ABORT:
1371 css_task_iter_end(&it);
1372 mem_cgroup_iter_break(memcg, iter);
1374 put_task_struct(chosen);
1379 points = oom_badness(task, memcg, NULL, totalpages);
1380 if (!points || points < chosen_points)
1382 /* Prefer thread group leaders for display purposes */
1383 if (points == chosen_points &&
1384 thread_group_leader(chosen))
1388 put_task_struct(chosen);
1390 chosen_points = points;
1391 get_task_struct(chosen);
1393 css_task_iter_end(&it);
1397 points = chosen_points * 1000 / totalpages;
1398 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1399 "Memory cgroup out of memory");
1402 mutex_unlock(&oom_lock);
1406 #if MAX_NUMNODES > 1
1409 * test_mem_cgroup_node_reclaimable
1410 * @memcg: the target memcg
1411 * @nid: the node ID to be checked.
1412 * @noswap : specify true here if the user wants flle only information.
1414 * This function returns whether the specified memcg contains any
1415 * reclaimable pages on a node. Returns true if there are any reclaimable
1416 * pages in the node.
1418 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1419 int nid, bool noswap)
1421 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1423 if (noswap || !total_swap_pages)
1425 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1432 * Always updating the nodemask is not very good - even if we have an empty
1433 * list or the wrong list here, we can start from some node and traverse all
1434 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1437 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1441 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1442 * pagein/pageout changes since the last update.
1444 if (!atomic_read(&memcg->numainfo_events))
1446 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1449 /* make a nodemask where this memcg uses memory from */
1450 memcg->scan_nodes = node_states[N_MEMORY];
1452 for_each_node_mask(nid, node_states[N_MEMORY]) {
1454 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1455 node_clear(nid, memcg->scan_nodes);
1458 atomic_set(&memcg->numainfo_events, 0);
1459 atomic_set(&memcg->numainfo_updating, 0);
1463 * Selecting a node where we start reclaim from. Because what we need is just
1464 * reducing usage counter, start from anywhere is O,K. Considering
1465 * memory reclaim from current node, there are pros. and cons.
1467 * Freeing memory from current node means freeing memory from a node which
1468 * we'll use or we've used. So, it may make LRU bad. And if several threads
1469 * hit limits, it will see a contention on a node. But freeing from remote
1470 * node means more costs for memory reclaim because of memory latency.
1472 * Now, we use round-robin. Better algorithm is welcomed.
1474 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1478 mem_cgroup_may_update_nodemask(memcg);
1479 node = memcg->last_scanned_node;
1481 node = next_node(node, memcg->scan_nodes);
1482 if (node == MAX_NUMNODES)
1483 node = first_node(memcg->scan_nodes);
1485 * We call this when we hit limit, not when pages are added to LRU.
1486 * No LRU may hold pages because all pages are UNEVICTABLE or
1487 * memcg is too small and all pages are not on LRU. In that case,
1488 * we use curret node.
1490 if (unlikely(node == MAX_NUMNODES))
1491 node = numa_node_id();
1493 memcg->last_scanned_node = node;
1497 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1503 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1506 unsigned long *total_scanned)
1508 struct mem_cgroup *victim = NULL;
1511 unsigned long excess;
1512 unsigned long nr_scanned;
1513 struct mem_cgroup_reclaim_cookie reclaim = {
1518 excess = soft_limit_excess(root_memcg);
1521 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1526 * If we have not been able to reclaim
1527 * anything, it might because there are
1528 * no reclaimable pages under this hierarchy
1533 * We want to do more targeted reclaim.
1534 * excess >> 2 is not to excessive so as to
1535 * reclaim too much, nor too less that we keep
1536 * coming back to reclaim from this cgroup
1538 if (total >= (excess >> 2) ||
1539 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1544 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1546 *total_scanned += nr_scanned;
1547 if (!soft_limit_excess(root_memcg))
1550 mem_cgroup_iter_break(root_memcg, victim);
1554 #ifdef CONFIG_LOCKDEP
1555 static struct lockdep_map memcg_oom_lock_dep_map = {
1556 .name = "memcg_oom_lock",
1560 static DEFINE_SPINLOCK(memcg_oom_lock);
1563 * Check OOM-Killer is already running under our hierarchy.
1564 * If someone is running, return false.
1566 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1568 struct mem_cgroup *iter, *failed = NULL;
1570 spin_lock(&memcg_oom_lock);
1572 for_each_mem_cgroup_tree(iter, memcg) {
1573 if (iter->oom_lock) {
1575 * this subtree of our hierarchy is already locked
1576 * so we cannot give a lock.
1579 mem_cgroup_iter_break(memcg, iter);
1582 iter->oom_lock = true;
1587 * OK, we failed to lock the whole subtree so we have
1588 * to clean up what we set up to the failing subtree
1590 for_each_mem_cgroup_tree(iter, memcg) {
1591 if (iter == failed) {
1592 mem_cgroup_iter_break(memcg, iter);
1595 iter->oom_lock = false;
1598 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1600 spin_unlock(&memcg_oom_lock);
1605 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1607 struct mem_cgroup *iter;
1609 spin_lock(&memcg_oom_lock);
1610 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1611 for_each_mem_cgroup_tree(iter, memcg)
1612 iter->oom_lock = false;
1613 spin_unlock(&memcg_oom_lock);
1616 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1618 struct mem_cgroup *iter;
1620 spin_lock(&memcg_oom_lock);
1621 for_each_mem_cgroup_tree(iter, memcg)
1623 spin_unlock(&memcg_oom_lock);
1626 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1628 struct mem_cgroup *iter;
1631 * When a new child is created while the hierarchy is under oom,
1632 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1634 spin_lock(&memcg_oom_lock);
1635 for_each_mem_cgroup_tree(iter, memcg)
1636 if (iter->under_oom > 0)
1638 spin_unlock(&memcg_oom_lock);
1641 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1643 struct oom_wait_info {
1644 struct mem_cgroup *memcg;
1648 static int memcg_oom_wake_function(wait_queue_t *wait,
1649 unsigned mode, int sync, void *arg)
1651 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1652 struct mem_cgroup *oom_wait_memcg;
1653 struct oom_wait_info *oom_wait_info;
1655 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1656 oom_wait_memcg = oom_wait_info->memcg;
1658 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1659 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1661 return autoremove_wake_function(wait, mode, sync, arg);
1664 static void memcg_oom_recover(struct mem_cgroup *memcg)
1667 * For the following lockless ->under_oom test, the only required
1668 * guarantee is that it must see the state asserted by an OOM when
1669 * this function is called as a result of userland actions
1670 * triggered by the notification of the OOM. This is trivially
1671 * achieved by invoking mem_cgroup_mark_under_oom() before
1672 * triggering notification.
1674 if (memcg && memcg->under_oom)
1675 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1678 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1680 if (!current->memcg_may_oom)
1683 * We are in the middle of the charge context here, so we
1684 * don't want to block when potentially sitting on a callstack
1685 * that holds all kinds of filesystem and mm locks.
1687 * Also, the caller may handle a failed allocation gracefully
1688 * (like optional page cache readahead) and so an OOM killer
1689 * invocation might not even be necessary.
1691 * That's why we don't do anything here except remember the
1692 * OOM context and then deal with it at the end of the page
1693 * fault when the stack is unwound, the locks are released,
1694 * and when we know whether the fault was overall successful.
1696 css_get(&memcg->css);
1697 current->memcg_in_oom = memcg;
1698 current->memcg_oom_gfp_mask = mask;
1699 current->memcg_oom_order = order;
1703 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1704 * @handle: actually kill/wait or just clean up the OOM state
1706 * This has to be called at the end of a page fault if the memcg OOM
1707 * handler was enabled.
1709 * Memcg supports userspace OOM handling where failed allocations must
1710 * sleep on a waitqueue until the userspace task resolves the
1711 * situation. Sleeping directly in the charge context with all kinds
1712 * of locks held is not a good idea, instead we remember an OOM state
1713 * in the task and mem_cgroup_oom_synchronize() has to be called at
1714 * the end of the page fault to complete the OOM handling.
1716 * Returns %true if an ongoing memcg OOM situation was detected and
1717 * completed, %false otherwise.
1719 bool mem_cgroup_oom_synchronize(bool handle)
1721 struct mem_cgroup *memcg = current->memcg_in_oom;
1722 struct oom_wait_info owait;
1725 /* OOM is global, do not handle */
1729 if (!handle || oom_killer_disabled)
1732 owait.memcg = memcg;
1733 owait.wait.flags = 0;
1734 owait.wait.func = memcg_oom_wake_function;
1735 owait.wait.private = current;
1736 INIT_LIST_HEAD(&owait.wait.task_list);
1738 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1739 mem_cgroup_mark_under_oom(memcg);
1741 locked = mem_cgroup_oom_trylock(memcg);
1744 mem_cgroup_oom_notify(memcg);
1746 if (locked && !memcg->oom_kill_disable) {
1747 mem_cgroup_unmark_under_oom(memcg);
1748 finish_wait(&memcg_oom_waitq, &owait.wait);
1749 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1750 current->memcg_oom_order);
1753 mem_cgroup_unmark_under_oom(memcg);
1754 finish_wait(&memcg_oom_waitq, &owait.wait);
1758 mem_cgroup_oom_unlock(memcg);
1760 * There is no guarantee that an OOM-lock contender
1761 * sees the wakeups triggered by the OOM kill
1762 * uncharges. Wake any sleepers explicitely.
1764 memcg_oom_recover(memcg);
1767 current->memcg_in_oom = NULL;
1768 css_put(&memcg->css);
1773 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1774 * @page: page that is going to change accounted state
1776 * This function must mark the beginning of an accounted page state
1777 * change to prevent double accounting when the page is concurrently
1778 * being moved to another memcg:
1780 * memcg = mem_cgroup_begin_page_stat(page);
1781 * if (TestClearPageState(page))
1782 * mem_cgroup_update_page_stat(memcg, state, -1);
1783 * mem_cgroup_end_page_stat(memcg);
1785 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1787 struct mem_cgroup *memcg;
1788 unsigned long flags;
1791 * The RCU lock is held throughout the transaction. The fast
1792 * path can get away without acquiring the memcg->move_lock
1793 * because page moving starts with an RCU grace period.
1795 * The RCU lock also protects the memcg from being freed when
1796 * the page state that is going to change is the only thing
1797 * preventing the page from being uncharged.
1798 * E.g. end-writeback clearing PageWriteback(), which allows
1799 * migration to go ahead and uncharge the page before the
1800 * account transaction might be complete.
1804 if (mem_cgroup_disabled())
1807 memcg = page->mem_cgroup;
1808 if (unlikely(!memcg))
1811 if (atomic_read(&memcg->moving_account) <= 0)
1814 spin_lock_irqsave(&memcg->move_lock, flags);
1815 if (memcg != page->mem_cgroup) {
1816 spin_unlock_irqrestore(&memcg->move_lock, flags);
1821 * When charge migration first begins, we can have locked and
1822 * unlocked page stat updates happening concurrently. Track
1823 * the task who has the lock for mem_cgroup_end_page_stat().
1825 memcg->move_lock_task = current;
1826 memcg->move_lock_flags = flags;
1830 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1833 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1834 * @memcg: the memcg that was accounted against
1836 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1838 if (memcg && memcg->move_lock_task == current) {
1839 unsigned long flags = memcg->move_lock_flags;
1841 memcg->move_lock_task = NULL;
1842 memcg->move_lock_flags = 0;
1844 spin_unlock_irqrestore(&memcg->move_lock, flags);
1849 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1852 * size of first charge trial. "32" comes from vmscan.c's magic value.
1853 * TODO: maybe necessary to use big numbers in big irons.
1855 #define CHARGE_BATCH 32U
1856 struct memcg_stock_pcp {
1857 struct mem_cgroup *cached; /* this never be root cgroup */
1858 unsigned int nr_pages;
1859 struct work_struct work;
1860 unsigned long flags;
1861 #define FLUSHING_CACHED_CHARGE 0
1863 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1864 static DEFINE_MUTEX(percpu_charge_mutex);
1867 * consume_stock: Try to consume stocked charge on this cpu.
1868 * @memcg: memcg to consume from.
1869 * @nr_pages: how many pages to charge.
1871 * The charges will only happen if @memcg matches the current cpu's memcg
1872 * stock, and at least @nr_pages are available in that stock. Failure to
1873 * service an allocation will refill the stock.
1875 * returns true if successful, false otherwise.
1877 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1879 struct memcg_stock_pcp *stock;
1882 if (nr_pages > CHARGE_BATCH)
1885 stock = &get_cpu_var(memcg_stock);
1886 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1887 stock->nr_pages -= nr_pages;
1890 put_cpu_var(memcg_stock);
1895 * Returns stocks cached in percpu and reset cached information.
1897 static void drain_stock(struct memcg_stock_pcp *stock)
1899 struct mem_cgroup *old = stock->cached;
1901 if (stock->nr_pages) {
1902 page_counter_uncharge(&old->memory, stock->nr_pages);
1903 if (do_swap_account)
1904 page_counter_uncharge(&old->memsw, stock->nr_pages);
1905 css_put_many(&old->css, stock->nr_pages);
1906 stock->nr_pages = 0;
1908 stock->cached = NULL;
1912 * This must be called under preempt disabled or must be called by
1913 * a thread which is pinned to local cpu.
1915 static void drain_local_stock(struct work_struct *dummy)
1917 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1919 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1923 * Cache charges(val) to local per_cpu area.
1924 * This will be consumed by consume_stock() function, later.
1926 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1928 struct memcg_stock_pcp *stock;
1929 int cpu = get_cpu_light();
1931 stock = &per_cpu(memcg_stock, cpu);
1933 if (stock->cached != memcg) { /* reset if necessary */
1935 stock->cached = memcg;
1937 stock->nr_pages += nr_pages;
1942 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1943 * of the hierarchy under it.
1945 static void drain_all_stock(struct mem_cgroup *root_memcg)
1949 /* If someone's already draining, avoid adding running more workers. */
1950 if (!mutex_trylock(&percpu_charge_mutex))
1952 /* Notify other cpus that system-wide "drain" is running */
1954 curcpu = get_cpu_light();
1955 for_each_online_cpu(cpu) {
1956 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1957 struct mem_cgroup *memcg;
1959 memcg = stock->cached;
1960 if (!memcg || !stock->nr_pages)
1962 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1964 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1966 drain_local_stock(&stock->work);
1968 schedule_work_on(cpu, &stock->work);
1973 mutex_unlock(&percpu_charge_mutex);
1976 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1977 unsigned long action,
1980 int cpu = (unsigned long)hcpu;
1981 struct memcg_stock_pcp *stock;
1983 if (action == CPU_ONLINE)
1986 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1989 stock = &per_cpu(memcg_stock, cpu);
1995 * Scheduled by try_charge() to be executed from the userland return path
1996 * and reclaims memory over the high limit.
1998 void mem_cgroup_handle_over_high(void)
2000 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2001 struct mem_cgroup *memcg, *pos;
2003 if (likely(!nr_pages))
2006 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2009 if (page_counter_read(&pos->memory) <= pos->high)
2011 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2012 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2013 } while ((pos = parent_mem_cgroup(pos)));
2015 css_put(&memcg->css);
2016 current->memcg_nr_pages_over_high = 0;
2019 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2020 unsigned int nr_pages)
2022 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2023 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2024 struct mem_cgroup *mem_over_limit;
2025 struct page_counter *counter;
2026 unsigned long nr_reclaimed;
2027 bool may_swap = true;
2028 bool drained = false;
2030 if (mem_cgroup_is_root(memcg))
2033 if (consume_stock(memcg, nr_pages))
2036 if (!do_swap_account ||
2037 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2038 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2040 if (do_swap_account)
2041 page_counter_uncharge(&memcg->memsw, batch);
2042 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2044 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2048 if (batch > nr_pages) {
2054 * Unlike in global OOM situations, memcg is not in a physical
2055 * memory shortage. Allow dying and OOM-killed tasks to
2056 * bypass the last charges so that they can exit quickly and
2057 * free their memory.
2059 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2060 fatal_signal_pending(current) ||
2061 current->flags & PF_EXITING))
2065 * Prevent unbounded recursion when reclaim operations need to
2066 * allocate memory. This might exceed the limits temporarily,
2067 * but we prefer facilitating memory reclaim and getting back
2068 * under the limit over triggering OOM kills in these cases.
2070 if (unlikely(current->flags & PF_MEMALLOC))
2073 if (unlikely(task_in_memcg_oom(current)))
2076 if (!gfpflags_allow_blocking(gfp_mask))
2079 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2081 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2082 gfp_mask, may_swap);
2084 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2088 drain_all_stock(mem_over_limit);
2093 if (gfp_mask & __GFP_NORETRY)
2096 * Even though the limit is exceeded at this point, reclaim
2097 * may have been able to free some pages. Retry the charge
2098 * before killing the task.
2100 * Only for regular pages, though: huge pages are rather
2101 * unlikely to succeed so close to the limit, and we fall back
2102 * to regular pages anyway in case of failure.
2104 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2107 * At task move, charge accounts can be doubly counted. So, it's
2108 * better to wait until the end of task_move if something is going on.
2110 if (mem_cgroup_wait_acct_move(mem_over_limit))
2116 if (gfp_mask & __GFP_NOFAIL)
2119 if (fatal_signal_pending(current))
2122 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2124 mem_cgroup_oom(mem_over_limit, gfp_mask,
2125 get_order(nr_pages * PAGE_SIZE));
2127 if (!(gfp_mask & __GFP_NOFAIL))
2131 * The allocation either can't fail or will lead to more memory
2132 * being freed very soon. Allow memory usage go over the limit
2133 * temporarily by force charging it.
2135 page_counter_charge(&memcg->memory, nr_pages);
2136 if (do_swap_account)
2137 page_counter_charge(&memcg->memsw, nr_pages);
2138 css_get_many(&memcg->css, nr_pages);
2143 css_get_many(&memcg->css, batch);
2144 if (batch > nr_pages)
2145 refill_stock(memcg, batch - nr_pages);
2148 * If the hierarchy is above the normal consumption range, schedule
2149 * reclaim on returning to userland. We can perform reclaim here
2150 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2151 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2152 * not recorded as it most likely matches current's and won't
2153 * change in the meantime. As high limit is checked again before
2154 * reclaim, the cost of mismatch is negligible.
2157 if (page_counter_read(&memcg->memory) > memcg->high) {
2158 current->memcg_nr_pages_over_high += batch;
2159 set_notify_resume(current);
2162 } while ((memcg = parent_mem_cgroup(memcg)));
2167 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2169 if (mem_cgroup_is_root(memcg))
2172 page_counter_uncharge(&memcg->memory, nr_pages);
2173 if (do_swap_account)
2174 page_counter_uncharge(&memcg->memsw, nr_pages);
2176 css_put_many(&memcg->css, nr_pages);
2179 static void lock_page_lru(struct page *page, int *isolated)
2181 struct zone *zone = page_zone(page);
2183 spin_lock_irq(&zone->lru_lock);
2184 if (PageLRU(page)) {
2185 struct lruvec *lruvec;
2187 lruvec = mem_cgroup_page_lruvec(page, zone);
2189 del_page_from_lru_list(page, lruvec, page_lru(page));
2195 static void unlock_page_lru(struct page *page, int isolated)
2197 struct zone *zone = page_zone(page);
2200 struct lruvec *lruvec;
2202 lruvec = mem_cgroup_page_lruvec(page, zone);
2203 VM_BUG_ON_PAGE(PageLRU(page), page);
2205 add_page_to_lru_list(page, lruvec, page_lru(page));
2207 spin_unlock_irq(&zone->lru_lock);
2210 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2215 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2218 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2219 * may already be on some other mem_cgroup's LRU. Take care of it.
2222 lock_page_lru(page, &isolated);
2225 * Nobody should be changing or seriously looking at
2226 * page->mem_cgroup at this point:
2228 * - the page is uncharged
2230 * - the page is off-LRU
2232 * - an anonymous fault has exclusive page access, except for
2233 * a locked page table
2235 * - a page cache insertion, a swapin fault, or a migration
2236 * have the page locked
2238 page->mem_cgroup = memcg;
2241 unlock_page_lru(page, isolated);
2244 #ifdef CONFIG_MEMCG_KMEM
2245 static int memcg_alloc_cache_id(void)
2250 id = ida_simple_get(&memcg_cache_ida,
2251 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2255 if (id < memcg_nr_cache_ids)
2259 * There's no space for the new id in memcg_caches arrays,
2260 * so we have to grow them.
2262 down_write(&memcg_cache_ids_sem);
2264 size = 2 * (id + 1);
2265 if (size < MEMCG_CACHES_MIN_SIZE)
2266 size = MEMCG_CACHES_MIN_SIZE;
2267 else if (size > MEMCG_CACHES_MAX_SIZE)
2268 size = MEMCG_CACHES_MAX_SIZE;
2270 err = memcg_update_all_caches(size);
2272 err = memcg_update_all_list_lrus(size);
2274 memcg_nr_cache_ids = size;
2276 up_write(&memcg_cache_ids_sem);
2279 ida_simple_remove(&memcg_cache_ida, id);
2285 static void memcg_free_cache_id(int id)
2287 ida_simple_remove(&memcg_cache_ida, id);
2290 struct memcg_kmem_cache_create_work {
2291 struct mem_cgroup *memcg;
2292 struct kmem_cache *cachep;
2293 struct work_struct work;
2296 static void memcg_kmem_cache_create_func(struct work_struct *w)
2298 struct memcg_kmem_cache_create_work *cw =
2299 container_of(w, struct memcg_kmem_cache_create_work, work);
2300 struct mem_cgroup *memcg = cw->memcg;
2301 struct kmem_cache *cachep = cw->cachep;
2303 memcg_create_kmem_cache(memcg, cachep);
2305 css_put(&memcg->css);
2310 * Enqueue the creation of a per-memcg kmem_cache.
2312 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2313 struct kmem_cache *cachep)
2315 struct memcg_kmem_cache_create_work *cw;
2317 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2321 css_get(&memcg->css);
2324 cw->cachep = cachep;
2325 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2327 schedule_work(&cw->work);
2330 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2331 struct kmem_cache *cachep)
2334 * We need to stop accounting when we kmalloc, because if the
2335 * corresponding kmalloc cache is not yet created, the first allocation
2336 * in __memcg_schedule_kmem_cache_create will recurse.
2338 * However, it is better to enclose the whole function. Depending on
2339 * the debugging options enabled, INIT_WORK(), for instance, can
2340 * trigger an allocation. This too, will make us recurse. Because at
2341 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2342 * the safest choice is to do it like this, wrapping the whole function.
2344 current->memcg_kmem_skip_account = 1;
2345 __memcg_schedule_kmem_cache_create(memcg, cachep);
2346 current->memcg_kmem_skip_account = 0;
2350 * Return the kmem_cache we're supposed to use for a slab allocation.
2351 * We try to use the current memcg's version of the cache.
2353 * If the cache does not exist yet, if we are the first user of it,
2354 * we either create it immediately, if possible, or create it asynchronously
2356 * In the latter case, we will let the current allocation go through with
2357 * the original cache.
2359 * Can't be called in interrupt context or from kernel threads.
2360 * This function needs to be called with rcu_read_lock() held.
2362 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2364 struct mem_cgroup *memcg;
2365 struct kmem_cache *memcg_cachep;
2368 VM_BUG_ON(!is_root_cache(cachep));
2370 if (current->memcg_kmem_skip_account)
2373 memcg = get_mem_cgroup_from_mm(current->mm);
2374 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2378 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2379 if (likely(memcg_cachep))
2380 return memcg_cachep;
2383 * If we are in a safe context (can wait, and not in interrupt
2384 * context), we could be be predictable and return right away.
2385 * This would guarantee that the allocation being performed
2386 * already belongs in the new cache.
2388 * However, there are some clashes that can arrive from locking.
2389 * For instance, because we acquire the slab_mutex while doing
2390 * memcg_create_kmem_cache, this means no further allocation
2391 * could happen with the slab_mutex held. So it's better to
2394 memcg_schedule_kmem_cache_create(memcg, cachep);
2396 css_put(&memcg->css);
2400 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2402 if (!is_root_cache(cachep))
2403 css_put(&cachep->memcg_params.memcg->css);
2406 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2407 struct mem_cgroup *memcg)
2409 unsigned int nr_pages = 1 << order;
2410 struct page_counter *counter;
2413 if (!memcg_kmem_is_active(memcg))
2416 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2419 ret = try_charge(memcg, gfp, nr_pages);
2421 page_counter_uncharge(&memcg->kmem, nr_pages);
2425 page->mem_cgroup = memcg;
2430 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2432 struct mem_cgroup *memcg;
2435 memcg = get_mem_cgroup_from_mm(current->mm);
2436 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2437 css_put(&memcg->css);
2441 void __memcg_kmem_uncharge(struct page *page, int order)
2443 struct mem_cgroup *memcg = page->mem_cgroup;
2444 unsigned int nr_pages = 1 << order;
2449 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2451 page_counter_uncharge(&memcg->kmem, nr_pages);
2452 page_counter_uncharge(&memcg->memory, nr_pages);
2453 if (do_swap_account)
2454 page_counter_uncharge(&memcg->memsw, nr_pages);
2456 page->mem_cgroup = NULL;
2457 css_put_many(&memcg->css, nr_pages);
2459 #endif /* CONFIG_MEMCG_KMEM */
2461 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2464 * Because tail pages are not marked as "used", set it. We're under
2465 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2466 * charge/uncharge will be never happen and move_account() is done under
2467 * compound_lock(), so we don't have to take care of races.
2469 void mem_cgroup_split_huge_fixup(struct page *head)
2473 if (mem_cgroup_disabled())
2476 for (i = 1; i < HPAGE_PMD_NR; i++)
2477 head[i].mem_cgroup = head->mem_cgroup;
2479 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2482 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2484 #ifdef CONFIG_MEMCG_SWAP
2485 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2488 int val = (charge) ? 1 : -1;
2489 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2493 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2494 * @entry: swap entry to be moved
2495 * @from: mem_cgroup which the entry is moved from
2496 * @to: mem_cgroup which the entry is moved to
2498 * It succeeds only when the swap_cgroup's record for this entry is the same
2499 * as the mem_cgroup's id of @from.
2501 * Returns 0 on success, -EINVAL on failure.
2503 * The caller must have charged to @to, IOW, called page_counter_charge() about
2504 * both res and memsw, and called css_get().
2506 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2507 struct mem_cgroup *from, struct mem_cgroup *to)
2509 unsigned short old_id, new_id;
2511 old_id = mem_cgroup_id(from);
2512 new_id = mem_cgroup_id(to);
2514 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2515 mem_cgroup_swap_statistics(from, false);
2516 mem_cgroup_swap_statistics(to, true);
2522 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2523 struct mem_cgroup *from, struct mem_cgroup *to)
2529 static DEFINE_MUTEX(memcg_limit_mutex);
2531 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2532 unsigned long limit)
2534 unsigned long curusage;
2535 unsigned long oldusage;
2536 bool enlarge = false;
2541 * For keeping hierarchical_reclaim simple, how long we should retry
2542 * is depends on callers. We set our retry-count to be function
2543 * of # of children which we should visit in this loop.
2545 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2546 mem_cgroup_count_children(memcg);
2548 oldusage = page_counter_read(&memcg->memory);
2551 if (signal_pending(current)) {
2556 mutex_lock(&memcg_limit_mutex);
2557 if (limit > memcg->memsw.limit) {
2558 mutex_unlock(&memcg_limit_mutex);
2562 if (limit > memcg->memory.limit)
2564 ret = page_counter_limit(&memcg->memory, limit);
2565 mutex_unlock(&memcg_limit_mutex);
2570 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2572 curusage = page_counter_read(&memcg->memory);
2573 /* Usage is reduced ? */
2574 if (curusage >= oldusage)
2577 oldusage = curusage;
2578 } while (retry_count);
2580 if (!ret && enlarge)
2581 memcg_oom_recover(memcg);
2586 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2587 unsigned long limit)
2589 unsigned long curusage;
2590 unsigned long oldusage;
2591 bool enlarge = false;
2595 /* see mem_cgroup_resize_res_limit */
2596 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2597 mem_cgroup_count_children(memcg);
2599 oldusage = page_counter_read(&memcg->memsw);
2602 if (signal_pending(current)) {
2607 mutex_lock(&memcg_limit_mutex);
2608 if (limit < memcg->memory.limit) {
2609 mutex_unlock(&memcg_limit_mutex);
2613 if (limit > memcg->memsw.limit)
2615 ret = page_counter_limit(&memcg->memsw, limit);
2616 mutex_unlock(&memcg_limit_mutex);
2621 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2623 curusage = page_counter_read(&memcg->memsw);
2624 /* Usage is reduced ? */
2625 if (curusage >= oldusage)
2628 oldusage = curusage;
2629 } while (retry_count);
2631 if (!ret && enlarge)
2632 memcg_oom_recover(memcg);
2637 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2639 unsigned long *total_scanned)
2641 unsigned long nr_reclaimed = 0;
2642 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2643 unsigned long reclaimed;
2645 struct mem_cgroup_tree_per_zone *mctz;
2646 unsigned long excess;
2647 unsigned long nr_scanned;
2652 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2654 * This loop can run a while, specially if mem_cgroup's continuously
2655 * keep exceeding their soft limit and putting the system under
2662 mz = mem_cgroup_largest_soft_limit_node(mctz);
2667 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2668 gfp_mask, &nr_scanned);
2669 nr_reclaimed += reclaimed;
2670 *total_scanned += nr_scanned;
2671 spin_lock_irq(&mctz->lock);
2672 __mem_cgroup_remove_exceeded(mz, mctz);
2675 * If we failed to reclaim anything from this memory cgroup
2676 * it is time to move on to the next cgroup
2680 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2682 excess = soft_limit_excess(mz->memcg);
2684 * One school of thought says that we should not add
2685 * back the node to the tree if reclaim returns 0.
2686 * But our reclaim could return 0, simply because due
2687 * to priority we are exposing a smaller subset of
2688 * memory to reclaim from. Consider this as a longer
2691 /* If excess == 0, no tree ops */
2692 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2693 spin_unlock_irq(&mctz->lock);
2694 css_put(&mz->memcg->css);
2697 * Could not reclaim anything and there are no more
2698 * mem cgroups to try or we seem to be looping without
2699 * reclaiming anything.
2701 if (!nr_reclaimed &&
2703 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2705 } while (!nr_reclaimed);
2707 css_put(&next_mz->memcg->css);
2708 return nr_reclaimed;
2712 * Test whether @memcg has children, dead or alive. Note that this
2713 * function doesn't care whether @memcg has use_hierarchy enabled and
2714 * returns %true if there are child csses according to the cgroup
2715 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2717 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2722 * The lock does not prevent addition or deletion of children, but
2723 * it prevents a new child from being initialized based on this
2724 * parent in css_online(), so it's enough to decide whether
2725 * hierarchically inherited attributes can still be changed or not.
2727 lockdep_assert_held(&memcg_create_mutex);
2730 ret = css_next_child(NULL, &memcg->css);
2736 * Reclaims as many pages from the given memcg as possible and moves
2737 * the rest to the parent.
2739 * Caller is responsible for holding css reference for memcg.
2741 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2743 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2745 /* we call try-to-free pages for make this cgroup empty */
2746 lru_add_drain_all();
2747 /* try to free all pages in this cgroup */
2748 while (nr_retries && page_counter_read(&memcg->memory)) {
2751 if (signal_pending(current))
2754 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2758 /* maybe some writeback is necessary */
2759 congestion_wait(BLK_RW_ASYNC, HZ/10);
2767 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2768 char *buf, size_t nbytes,
2771 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2773 if (mem_cgroup_is_root(memcg))
2775 return mem_cgroup_force_empty(memcg) ?: nbytes;
2778 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2781 return mem_cgroup_from_css(css)->use_hierarchy;
2784 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2785 struct cftype *cft, u64 val)
2788 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2789 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2791 mutex_lock(&memcg_create_mutex);
2793 if (memcg->use_hierarchy == val)
2797 * If parent's use_hierarchy is set, we can't make any modifications
2798 * in the child subtrees. If it is unset, then the change can
2799 * occur, provided the current cgroup has no children.
2801 * For the root cgroup, parent_mem is NULL, we allow value to be
2802 * set if there are no children.
2804 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2805 (val == 1 || val == 0)) {
2806 if (!memcg_has_children(memcg))
2807 memcg->use_hierarchy = val;
2814 mutex_unlock(&memcg_create_mutex);
2819 static unsigned long tree_stat(struct mem_cgroup *memcg,
2820 enum mem_cgroup_stat_index idx)
2822 struct mem_cgroup *iter;
2823 unsigned long val = 0;
2825 for_each_mem_cgroup_tree(iter, memcg)
2826 val += mem_cgroup_read_stat(iter, idx);
2831 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2835 if (mem_cgroup_is_root(memcg)) {
2836 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2837 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2839 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2842 val = page_counter_read(&memcg->memory);
2844 val = page_counter_read(&memcg->memsw);
2857 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2860 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2861 struct page_counter *counter;
2863 switch (MEMFILE_TYPE(cft->private)) {
2865 counter = &memcg->memory;
2868 counter = &memcg->memsw;
2871 counter = &memcg->kmem;
2877 switch (MEMFILE_ATTR(cft->private)) {
2879 if (counter == &memcg->memory)
2880 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2881 if (counter == &memcg->memsw)
2882 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2883 return (u64)page_counter_read(counter) * PAGE_SIZE;
2885 return (u64)counter->limit * PAGE_SIZE;
2887 return (u64)counter->watermark * PAGE_SIZE;
2889 return counter->failcnt;
2890 case RES_SOFT_LIMIT:
2891 return (u64)memcg->soft_limit * PAGE_SIZE;
2897 #ifdef CONFIG_MEMCG_KMEM
2898 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2899 unsigned long nr_pages)
2904 BUG_ON(memcg->kmemcg_id >= 0);
2905 BUG_ON(memcg->kmem_acct_activated);
2906 BUG_ON(memcg->kmem_acct_active);
2909 * For simplicity, we won't allow this to be disabled. It also can't
2910 * be changed if the cgroup has children already, or if tasks had
2913 * If tasks join before we set the limit, a person looking at
2914 * kmem.usage_in_bytes will have no way to determine when it took
2915 * place, which makes the value quite meaningless.
2917 * After it first became limited, changes in the value of the limit are
2918 * of course permitted.
2920 mutex_lock(&memcg_create_mutex);
2921 if (cgroup_is_populated(memcg->css.cgroup) ||
2922 (memcg->use_hierarchy && memcg_has_children(memcg)))
2924 mutex_unlock(&memcg_create_mutex);
2928 memcg_id = memcg_alloc_cache_id();
2935 * We couldn't have accounted to this cgroup, because it hasn't got
2936 * activated yet, so this should succeed.
2938 err = page_counter_limit(&memcg->kmem, nr_pages);
2941 static_key_slow_inc(&memcg_kmem_enabled_key);
2943 * A memory cgroup is considered kmem-active as soon as it gets
2944 * kmemcg_id. Setting the id after enabling static branching will
2945 * guarantee no one starts accounting before all call sites are
2948 memcg->kmemcg_id = memcg_id;
2949 memcg->kmem_acct_activated = true;
2950 memcg->kmem_acct_active = true;
2955 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2956 unsigned long limit)
2960 mutex_lock(&memcg_limit_mutex);
2961 if (!memcg_kmem_is_active(memcg))
2962 ret = memcg_activate_kmem(memcg, limit);
2964 ret = page_counter_limit(&memcg->kmem, limit);
2965 mutex_unlock(&memcg_limit_mutex);
2969 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2972 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2977 mutex_lock(&memcg_limit_mutex);
2979 * If the parent cgroup is not kmem-active now, it cannot be activated
2980 * after this point, because it has at least one child already.
2982 if (memcg_kmem_is_active(parent))
2983 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2984 mutex_unlock(&memcg_limit_mutex);
2988 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2989 unsigned long limit)
2993 #endif /* CONFIG_MEMCG_KMEM */
2996 * The user of this function is...
2999 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3000 char *buf, size_t nbytes, loff_t off)
3002 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3003 unsigned long nr_pages;
3006 buf = strstrip(buf);
3007 ret = page_counter_memparse(buf, "-1", &nr_pages);
3011 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3013 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3017 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3019 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3022 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3025 ret = memcg_update_kmem_limit(memcg, nr_pages);
3029 case RES_SOFT_LIMIT:
3030 memcg->soft_limit = nr_pages;
3034 return ret ?: nbytes;
3037 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3038 size_t nbytes, loff_t off)
3040 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3041 struct page_counter *counter;
3043 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3045 counter = &memcg->memory;
3048 counter = &memcg->memsw;
3051 counter = &memcg->kmem;
3057 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3059 page_counter_reset_watermark(counter);
3062 counter->failcnt = 0;
3071 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3074 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3078 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3079 struct cftype *cft, u64 val)
3081 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3083 if (val & ~MOVE_MASK)
3087 * No kind of locking is needed in here, because ->can_attach() will
3088 * check this value once in the beginning of the process, and then carry
3089 * on with stale data. This means that changes to this value will only
3090 * affect task migrations starting after the change.
3092 memcg->move_charge_at_immigrate = val;
3096 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3097 struct cftype *cft, u64 val)
3104 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3108 unsigned int lru_mask;
3111 static const struct numa_stat stats[] = {
3112 { "total", LRU_ALL },
3113 { "file", LRU_ALL_FILE },
3114 { "anon", LRU_ALL_ANON },
3115 { "unevictable", BIT(LRU_UNEVICTABLE) },
3117 const struct numa_stat *stat;
3120 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3122 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3123 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3124 seq_printf(m, "%s=%lu", stat->name, nr);
3125 for_each_node_state(nid, N_MEMORY) {
3126 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3128 seq_printf(m, " N%d=%lu", nid, nr);
3133 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3134 struct mem_cgroup *iter;
3137 for_each_mem_cgroup_tree(iter, memcg)
3138 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3139 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3140 for_each_node_state(nid, N_MEMORY) {
3142 for_each_mem_cgroup_tree(iter, memcg)
3143 nr += mem_cgroup_node_nr_lru_pages(
3144 iter, nid, stat->lru_mask);
3145 seq_printf(m, " N%d=%lu", nid, nr);
3152 #endif /* CONFIG_NUMA */
3154 static int memcg_stat_show(struct seq_file *m, void *v)
3156 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3157 unsigned long memory, memsw;
3158 struct mem_cgroup *mi;
3161 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3162 MEM_CGROUP_STAT_NSTATS);
3163 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3164 MEM_CGROUP_EVENTS_NSTATS);
3165 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3167 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3168 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3170 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3171 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3174 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3175 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3176 mem_cgroup_read_events(memcg, i));
3178 for (i = 0; i < NR_LRU_LISTS; i++)
3179 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3180 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3182 /* Hierarchical information */
3183 memory = memsw = PAGE_COUNTER_MAX;
3184 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3185 memory = min(memory, mi->memory.limit);
3186 memsw = min(memsw, mi->memsw.limit);
3188 seq_printf(m, "hierarchical_memory_limit %llu\n",
3189 (u64)memory * PAGE_SIZE);
3190 if (do_swap_account)
3191 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3192 (u64)memsw * PAGE_SIZE);
3194 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3195 unsigned long long val = 0;
3197 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3199 for_each_mem_cgroup_tree(mi, memcg)
3200 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3201 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3204 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3205 unsigned long long val = 0;
3207 for_each_mem_cgroup_tree(mi, memcg)
3208 val += mem_cgroup_read_events(mi, i);
3209 seq_printf(m, "total_%s %llu\n",
3210 mem_cgroup_events_names[i], val);
3213 for (i = 0; i < NR_LRU_LISTS; i++) {
3214 unsigned long long val = 0;
3216 for_each_mem_cgroup_tree(mi, memcg)
3217 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3218 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3221 #ifdef CONFIG_DEBUG_VM
3224 struct mem_cgroup_per_zone *mz;
3225 struct zone_reclaim_stat *rstat;
3226 unsigned long recent_rotated[2] = {0, 0};
3227 unsigned long recent_scanned[2] = {0, 0};
3229 for_each_online_node(nid)
3230 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3231 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3232 rstat = &mz->lruvec.reclaim_stat;
3234 recent_rotated[0] += rstat->recent_rotated[0];
3235 recent_rotated[1] += rstat->recent_rotated[1];
3236 recent_scanned[0] += rstat->recent_scanned[0];
3237 recent_scanned[1] += rstat->recent_scanned[1];
3239 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3240 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3241 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3242 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3249 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3252 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3254 return mem_cgroup_swappiness(memcg);
3257 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3258 struct cftype *cft, u64 val)
3260 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3266 memcg->swappiness = val;
3268 vm_swappiness = val;
3273 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3275 struct mem_cgroup_threshold_ary *t;
3276 unsigned long usage;
3281 t = rcu_dereference(memcg->thresholds.primary);
3283 t = rcu_dereference(memcg->memsw_thresholds.primary);
3288 usage = mem_cgroup_usage(memcg, swap);
3291 * current_threshold points to threshold just below or equal to usage.
3292 * If it's not true, a threshold was crossed after last
3293 * call of __mem_cgroup_threshold().
3295 i = t->current_threshold;
3298 * Iterate backward over array of thresholds starting from
3299 * current_threshold and check if a threshold is crossed.
3300 * If none of thresholds below usage is crossed, we read
3301 * only one element of the array here.
3303 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3304 eventfd_signal(t->entries[i].eventfd, 1);
3306 /* i = current_threshold + 1 */
3310 * Iterate forward over array of thresholds starting from
3311 * current_threshold+1 and check if a threshold is crossed.
3312 * If none of thresholds above usage is crossed, we read
3313 * only one element of the array here.
3315 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3316 eventfd_signal(t->entries[i].eventfd, 1);
3318 /* Update current_threshold */
3319 t->current_threshold = i - 1;
3324 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3327 __mem_cgroup_threshold(memcg, false);
3328 if (do_swap_account)
3329 __mem_cgroup_threshold(memcg, true);
3331 memcg = parent_mem_cgroup(memcg);
3335 static int compare_thresholds(const void *a, const void *b)
3337 const struct mem_cgroup_threshold *_a = a;
3338 const struct mem_cgroup_threshold *_b = b;
3340 if (_a->threshold > _b->threshold)
3343 if (_a->threshold < _b->threshold)
3349 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3351 struct mem_cgroup_eventfd_list *ev;
3353 spin_lock(&memcg_oom_lock);
3355 list_for_each_entry(ev, &memcg->oom_notify, list)
3356 eventfd_signal(ev->eventfd, 1);
3358 spin_unlock(&memcg_oom_lock);
3362 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3364 struct mem_cgroup *iter;
3366 for_each_mem_cgroup_tree(iter, memcg)
3367 mem_cgroup_oom_notify_cb(iter);
3370 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3371 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3373 struct mem_cgroup_thresholds *thresholds;
3374 struct mem_cgroup_threshold_ary *new;
3375 unsigned long threshold;
3376 unsigned long usage;
3379 ret = page_counter_memparse(args, "-1", &threshold);
3383 mutex_lock(&memcg->thresholds_lock);
3386 thresholds = &memcg->thresholds;
3387 usage = mem_cgroup_usage(memcg, false);
3388 } else if (type == _MEMSWAP) {
3389 thresholds = &memcg->memsw_thresholds;
3390 usage = mem_cgroup_usage(memcg, true);
3394 /* Check if a threshold crossed before adding a new one */
3395 if (thresholds->primary)
3396 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3398 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3400 /* Allocate memory for new array of thresholds */
3401 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3409 /* Copy thresholds (if any) to new array */
3410 if (thresholds->primary) {
3411 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3412 sizeof(struct mem_cgroup_threshold));
3415 /* Add new threshold */
3416 new->entries[size - 1].eventfd = eventfd;
3417 new->entries[size - 1].threshold = threshold;
3419 /* Sort thresholds. Registering of new threshold isn't time-critical */
3420 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3421 compare_thresholds, NULL);
3423 /* Find current threshold */
3424 new->current_threshold = -1;
3425 for (i = 0; i < size; i++) {
3426 if (new->entries[i].threshold <= usage) {
3428 * new->current_threshold will not be used until
3429 * rcu_assign_pointer(), so it's safe to increment
3432 ++new->current_threshold;
3437 /* Free old spare buffer and save old primary buffer as spare */
3438 kfree(thresholds->spare);
3439 thresholds->spare = thresholds->primary;
3441 rcu_assign_pointer(thresholds->primary, new);
3443 /* To be sure that nobody uses thresholds */
3447 mutex_unlock(&memcg->thresholds_lock);
3452 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3453 struct eventfd_ctx *eventfd, const char *args)
3455 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3458 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3459 struct eventfd_ctx *eventfd, const char *args)
3461 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3464 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3465 struct eventfd_ctx *eventfd, enum res_type type)
3467 struct mem_cgroup_thresholds *thresholds;
3468 struct mem_cgroup_threshold_ary *new;
3469 unsigned long usage;
3472 mutex_lock(&memcg->thresholds_lock);
3475 thresholds = &memcg->thresholds;
3476 usage = mem_cgroup_usage(memcg, false);
3477 } else if (type == _MEMSWAP) {
3478 thresholds = &memcg->memsw_thresholds;
3479 usage = mem_cgroup_usage(memcg, true);
3483 if (!thresholds->primary)
3486 /* Check if a threshold crossed before removing */
3487 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3489 /* Calculate new number of threshold */
3491 for (i = 0; i < thresholds->primary->size; i++) {
3492 if (thresholds->primary->entries[i].eventfd != eventfd)
3496 new = thresholds->spare;
3498 /* Set thresholds array to NULL if we don't have thresholds */
3507 /* Copy thresholds and find current threshold */
3508 new->current_threshold = -1;
3509 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3510 if (thresholds->primary->entries[i].eventfd == eventfd)
3513 new->entries[j] = thresholds->primary->entries[i];
3514 if (new->entries[j].threshold <= usage) {
3516 * new->current_threshold will not be used
3517 * until rcu_assign_pointer(), so it's safe to increment
3520 ++new->current_threshold;
3526 /* Swap primary and spare array */
3527 thresholds->spare = thresholds->primary;
3529 rcu_assign_pointer(thresholds->primary, new);
3531 /* To be sure that nobody uses thresholds */
3534 /* If all events are unregistered, free the spare array */
3536 kfree(thresholds->spare);
3537 thresholds->spare = NULL;
3540 mutex_unlock(&memcg->thresholds_lock);
3543 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3544 struct eventfd_ctx *eventfd)
3546 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3549 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3550 struct eventfd_ctx *eventfd)
3552 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3555 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3556 struct eventfd_ctx *eventfd, const char *args)
3558 struct mem_cgroup_eventfd_list *event;
3560 event = kmalloc(sizeof(*event), GFP_KERNEL);
3564 spin_lock(&memcg_oom_lock);
3566 event->eventfd = eventfd;
3567 list_add(&event->list, &memcg->oom_notify);
3569 /* already in OOM ? */
3570 if (memcg->under_oom)
3571 eventfd_signal(eventfd, 1);
3572 spin_unlock(&memcg_oom_lock);
3577 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3578 struct eventfd_ctx *eventfd)
3580 struct mem_cgroup_eventfd_list *ev, *tmp;
3582 spin_lock(&memcg_oom_lock);
3584 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3585 if (ev->eventfd == eventfd) {
3586 list_del(&ev->list);
3591 spin_unlock(&memcg_oom_lock);
3594 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3596 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3598 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3599 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3603 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3604 struct cftype *cft, u64 val)
3606 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3608 /* cannot set to root cgroup and only 0 and 1 are allowed */
3609 if (!css->parent || !((val == 0) || (val == 1)))
3612 memcg->oom_kill_disable = val;
3614 memcg_oom_recover(memcg);
3619 #ifdef CONFIG_MEMCG_KMEM
3620 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3624 ret = memcg_propagate_kmem(memcg);
3628 return mem_cgroup_sockets_init(memcg, ss);
3631 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3633 struct cgroup_subsys_state *css;
3634 struct mem_cgroup *parent, *child;
3637 if (!memcg->kmem_acct_active)
3641 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3642 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3643 * guarantees no cache will be created for this cgroup after we are
3644 * done (see memcg_create_kmem_cache()).
3646 memcg->kmem_acct_active = false;
3648 memcg_deactivate_kmem_caches(memcg);
3650 kmemcg_id = memcg->kmemcg_id;
3651 BUG_ON(kmemcg_id < 0);
3653 parent = parent_mem_cgroup(memcg);
3655 parent = root_mem_cgroup;
3658 * Change kmemcg_id of this cgroup and all its descendants to the
3659 * parent's id, and then move all entries from this cgroup's list_lrus
3660 * to ones of the parent. After we have finished, all list_lrus
3661 * corresponding to this cgroup are guaranteed to remain empty. The
3662 * ordering is imposed by list_lru_node->lock taken by
3663 * memcg_drain_all_list_lrus().
3665 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3666 css_for_each_descendant_pre(css, &memcg->css) {
3667 child = mem_cgroup_from_css(css);
3668 BUG_ON(child->kmemcg_id != kmemcg_id);
3669 child->kmemcg_id = parent->kmemcg_id;
3670 if (!memcg->use_hierarchy)
3675 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3677 memcg_free_cache_id(kmemcg_id);
3680 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3682 if (memcg->kmem_acct_activated) {
3683 memcg_destroy_kmem_caches(memcg);
3684 static_key_slow_dec(&memcg_kmem_enabled_key);
3685 WARN_ON(page_counter_read(&memcg->kmem));
3687 mem_cgroup_sockets_destroy(memcg);
3690 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3695 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3699 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3704 #ifdef CONFIG_CGROUP_WRITEBACK
3706 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3708 return &memcg->cgwb_list;
3711 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3713 return wb_domain_init(&memcg->cgwb_domain, gfp);
3716 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3718 wb_domain_exit(&memcg->cgwb_domain);
3721 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3723 wb_domain_size_changed(&memcg->cgwb_domain);
3726 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3728 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3730 if (!memcg->css.parent)
3733 return &memcg->cgwb_domain;
3737 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3738 * @wb: bdi_writeback in question
3739 * @pfilepages: out parameter for number of file pages
3740 * @pheadroom: out parameter for number of allocatable pages according to memcg
3741 * @pdirty: out parameter for number of dirty pages
3742 * @pwriteback: out parameter for number of pages under writeback
3744 * Determine the numbers of file, headroom, dirty, and writeback pages in
3745 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3746 * is a bit more involved.
3748 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3749 * headroom is calculated as the lowest headroom of itself and the
3750 * ancestors. Note that this doesn't consider the actual amount of
3751 * available memory in the system. The caller should further cap
3752 * *@pheadroom accordingly.
3754 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3755 unsigned long *pheadroom, unsigned long *pdirty,
3756 unsigned long *pwriteback)
3758 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3759 struct mem_cgroup *parent;
3761 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3763 /* this should eventually include NR_UNSTABLE_NFS */
3764 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3765 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3766 (1 << LRU_ACTIVE_FILE));
3767 *pheadroom = PAGE_COUNTER_MAX;
3769 while ((parent = parent_mem_cgroup(memcg))) {
3770 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3771 unsigned long used = page_counter_read(&memcg->memory);
3773 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3778 #else /* CONFIG_CGROUP_WRITEBACK */
3780 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3785 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3789 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3793 #endif /* CONFIG_CGROUP_WRITEBACK */
3796 * DO NOT USE IN NEW FILES.
3798 * "cgroup.event_control" implementation.
3800 * This is way over-engineered. It tries to support fully configurable
3801 * events for each user. Such level of flexibility is completely
3802 * unnecessary especially in the light of the planned unified hierarchy.
3804 * Please deprecate this and replace with something simpler if at all
3809 * Unregister event and free resources.
3811 * Gets called from workqueue.
3813 static void memcg_event_remove(struct work_struct *work)
3815 struct mem_cgroup_event *event =
3816 container_of(work, struct mem_cgroup_event, remove);
3817 struct mem_cgroup *memcg = event->memcg;
3819 remove_wait_queue(event->wqh, &event->wait);
3821 event->unregister_event(memcg, event->eventfd);
3823 /* Notify userspace the event is going away. */
3824 eventfd_signal(event->eventfd, 1);
3826 eventfd_ctx_put(event->eventfd);
3828 css_put(&memcg->css);
3832 * Gets called on POLLHUP on eventfd when user closes it.
3834 * Called with wqh->lock held and interrupts disabled.
3836 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3837 int sync, void *key)
3839 struct mem_cgroup_event *event =
3840 container_of(wait, struct mem_cgroup_event, wait);
3841 struct mem_cgroup *memcg = event->memcg;
3842 unsigned long flags = (unsigned long)key;
3844 if (flags & POLLHUP) {
3846 * If the event has been detached at cgroup removal, we
3847 * can simply return knowing the other side will cleanup
3850 * We can't race against event freeing since the other
3851 * side will require wqh->lock via remove_wait_queue(),
3854 spin_lock(&memcg->event_list_lock);
3855 if (!list_empty(&event->list)) {
3856 list_del_init(&event->list);
3858 * We are in atomic context, but cgroup_event_remove()
3859 * may sleep, so we have to call it in workqueue.
3861 schedule_work(&event->remove);
3863 spin_unlock(&memcg->event_list_lock);
3869 static void memcg_event_ptable_queue_proc(struct file *file,
3870 wait_queue_head_t *wqh, poll_table *pt)
3872 struct mem_cgroup_event *event =
3873 container_of(pt, struct mem_cgroup_event, pt);
3876 add_wait_queue(wqh, &event->wait);
3880 * DO NOT USE IN NEW FILES.
3882 * Parse input and register new cgroup event handler.
3884 * Input must be in format '<event_fd> <control_fd> <args>'.
3885 * Interpretation of args is defined by control file implementation.
3887 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3888 char *buf, size_t nbytes, loff_t off)
3890 struct cgroup_subsys_state *css = of_css(of);
3891 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3892 struct mem_cgroup_event *event;
3893 struct cgroup_subsys_state *cfile_css;
3894 unsigned int efd, cfd;
3901 buf = strstrip(buf);
3903 efd = simple_strtoul(buf, &endp, 10);
3908 cfd = simple_strtoul(buf, &endp, 10);
3909 if ((*endp != ' ') && (*endp != '\0'))
3913 event = kzalloc(sizeof(*event), GFP_KERNEL);
3917 event->memcg = memcg;
3918 INIT_LIST_HEAD(&event->list);
3919 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3920 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3921 INIT_WORK(&event->remove, memcg_event_remove);
3929 event->eventfd = eventfd_ctx_fileget(efile.file);
3930 if (IS_ERR(event->eventfd)) {
3931 ret = PTR_ERR(event->eventfd);
3938 goto out_put_eventfd;
3941 /* the process need read permission on control file */
3942 /* AV: shouldn't we check that it's been opened for read instead? */
3943 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3948 * Determine the event callbacks and set them in @event. This used
3949 * to be done via struct cftype but cgroup core no longer knows
3950 * about these events. The following is crude but the whole thing
3951 * is for compatibility anyway.
3953 * DO NOT ADD NEW FILES.
3955 name = cfile.file->f_path.dentry->d_name.name;
3957 if (!strcmp(name, "memory.usage_in_bytes")) {
3958 event->register_event = mem_cgroup_usage_register_event;
3959 event->unregister_event = mem_cgroup_usage_unregister_event;
3960 } else if (!strcmp(name, "memory.oom_control")) {
3961 event->register_event = mem_cgroup_oom_register_event;
3962 event->unregister_event = mem_cgroup_oom_unregister_event;
3963 } else if (!strcmp(name, "memory.pressure_level")) {
3964 event->register_event = vmpressure_register_event;
3965 event->unregister_event = vmpressure_unregister_event;
3966 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3967 event->register_event = memsw_cgroup_usage_register_event;
3968 event->unregister_event = memsw_cgroup_usage_unregister_event;
3975 * Verify @cfile should belong to @css. Also, remaining events are
3976 * automatically removed on cgroup destruction but the removal is
3977 * asynchronous, so take an extra ref on @css.
3979 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3980 &memory_cgrp_subsys);
3982 if (IS_ERR(cfile_css))
3984 if (cfile_css != css) {
3989 ret = event->register_event(memcg, event->eventfd, buf);
3993 efile.file->f_op->poll(efile.file, &event->pt);
3995 spin_lock(&memcg->event_list_lock);
3996 list_add(&event->list, &memcg->event_list);
3997 spin_unlock(&memcg->event_list_lock);
4009 eventfd_ctx_put(event->eventfd);
4018 static struct cftype mem_cgroup_legacy_files[] = {
4020 .name = "usage_in_bytes",
4021 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4022 .read_u64 = mem_cgroup_read_u64,
4025 .name = "max_usage_in_bytes",
4026 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4027 .write = mem_cgroup_reset,
4028 .read_u64 = mem_cgroup_read_u64,
4031 .name = "limit_in_bytes",
4032 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4033 .write = mem_cgroup_write,
4034 .read_u64 = mem_cgroup_read_u64,
4037 .name = "soft_limit_in_bytes",
4038 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4039 .write = mem_cgroup_write,
4040 .read_u64 = mem_cgroup_read_u64,
4044 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4045 .write = mem_cgroup_reset,
4046 .read_u64 = mem_cgroup_read_u64,
4050 .seq_show = memcg_stat_show,
4053 .name = "force_empty",
4054 .write = mem_cgroup_force_empty_write,
4057 .name = "use_hierarchy",
4058 .write_u64 = mem_cgroup_hierarchy_write,
4059 .read_u64 = mem_cgroup_hierarchy_read,
4062 .name = "cgroup.event_control", /* XXX: for compat */
4063 .write = memcg_write_event_control,
4064 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4067 .name = "swappiness",
4068 .read_u64 = mem_cgroup_swappiness_read,
4069 .write_u64 = mem_cgroup_swappiness_write,
4072 .name = "move_charge_at_immigrate",
4073 .read_u64 = mem_cgroup_move_charge_read,
4074 .write_u64 = mem_cgroup_move_charge_write,
4077 .name = "oom_control",
4078 .seq_show = mem_cgroup_oom_control_read,
4079 .write_u64 = mem_cgroup_oom_control_write,
4080 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4083 .name = "pressure_level",
4087 .name = "numa_stat",
4088 .seq_show = memcg_numa_stat_show,
4091 #ifdef CONFIG_MEMCG_KMEM
4093 .name = "kmem.limit_in_bytes",
4094 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4095 .write = mem_cgroup_write,
4096 .read_u64 = mem_cgroup_read_u64,
4099 .name = "kmem.usage_in_bytes",
4100 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4101 .read_u64 = mem_cgroup_read_u64,
4104 .name = "kmem.failcnt",
4105 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4106 .write = mem_cgroup_reset,
4107 .read_u64 = mem_cgroup_read_u64,
4110 .name = "kmem.max_usage_in_bytes",
4111 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4112 .write = mem_cgroup_reset,
4113 .read_u64 = mem_cgroup_read_u64,
4115 #ifdef CONFIG_SLABINFO
4117 .name = "kmem.slabinfo",
4118 .seq_start = slab_start,
4119 .seq_next = slab_next,
4120 .seq_stop = slab_stop,
4121 .seq_show = memcg_slab_show,
4125 { }, /* terminate */
4129 * Private memory cgroup IDR
4131 * Swap-out records and page cache shadow entries need to store memcg
4132 * references in constrained space, so we maintain an ID space that is
4133 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4134 * memory-controlled cgroups to 64k.
4136 * However, there usually are many references to the oflline CSS after
4137 * the cgroup has been destroyed, such as page cache or reclaimable
4138 * slab objects, that don't need to hang on to the ID. We want to keep
4139 * those dead CSS from occupying IDs, or we might quickly exhaust the
4140 * relatively small ID space and prevent the creation of new cgroups
4141 * even when there are much fewer than 64k cgroups - possibly none.
4143 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4144 * be freed and recycled when it's no longer needed, which is usually
4145 * when the CSS is offlined.
4147 * The only exception to that are records of swapped out tmpfs/shmem
4148 * pages that need to be attributed to live ancestors on swapin. But
4149 * those references are manageable from userspace.
4152 static DEFINE_IDR(mem_cgroup_idr);
4154 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4156 atomic_add(n, &memcg->id.ref);
4159 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
4161 while (!atomic_inc_not_zero(&memcg->id.ref)) {
4163 * The root cgroup cannot be destroyed, so it's refcount must
4166 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
4170 memcg = parent_mem_cgroup(memcg);
4172 memcg = root_mem_cgroup;
4177 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4179 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4180 idr_remove(&mem_cgroup_idr, memcg->id.id);
4183 /* Memcg ID pins CSS */
4184 css_put(&memcg->css);
4188 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4190 mem_cgroup_id_get_many(memcg, 1);
4193 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4195 mem_cgroup_id_put_many(memcg, 1);
4199 * mem_cgroup_from_id - look up a memcg from a memcg id
4200 * @id: the memcg id to look up
4202 * Caller must hold rcu_read_lock().
4204 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4206 WARN_ON_ONCE(!rcu_read_lock_held());
4207 return idr_find(&mem_cgroup_idr, id);
4210 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4212 struct mem_cgroup_per_node *pn;
4213 struct mem_cgroup_per_zone *mz;
4214 int zone, tmp = node;
4216 * This routine is called against possible nodes.
4217 * But it's BUG to call kmalloc() against offline node.
4219 * TODO: this routine can waste much memory for nodes which will
4220 * never be onlined. It's better to use memory hotplug callback
4223 if (!node_state(node, N_NORMAL_MEMORY))
4225 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4229 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4230 mz = &pn->zoneinfo[zone];
4231 lruvec_init(&mz->lruvec);
4232 mz->usage_in_excess = 0;
4233 mz->on_tree = false;
4236 memcg->nodeinfo[node] = pn;
4240 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4242 kfree(memcg->nodeinfo[node]);
4245 static struct mem_cgroup *mem_cgroup_alloc(void)
4247 struct mem_cgroup *memcg;
4250 size = sizeof(struct mem_cgroup);
4251 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4253 memcg = kzalloc(size, GFP_KERNEL);
4257 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4261 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4264 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4265 1, MEM_CGROUP_ID_MAX,
4267 if (memcg->id.id < 0)
4273 free_percpu(memcg->stat);
4280 * At destroying mem_cgroup, references from swap_cgroup can remain.
4281 * (scanning all at force_empty is too costly...)
4283 * Instead of clearing all references at force_empty, we remember
4284 * the number of reference from swap_cgroup and free mem_cgroup when
4285 * it goes down to 0.
4287 * Removal of cgroup itself succeeds regardless of refs from swap.
4290 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4294 mem_cgroup_remove_from_trees(memcg);
4297 free_mem_cgroup_per_zone_info(memcg, node);
4299 free_percpu(memcg->stat);
4300 memcg_wb_domain_exit(memcg);
4305 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4307 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4309 if (!memcg->memory.parent)
4311 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4313 EXPORT_SYMBOL(parent_mem_cgroup);
4315 static struct cgroup_subsys_state * __ref
4316 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4318 struct mem_cgroup *memcg;
4319 long error = -ENOMEM;
4322 memcg = mem_cgroup_alloc();
4324 return ERR_PTR(error);
4327 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4331 if (parent_css == NULL) {
4332 root_mem_cgroup = memcg;
4333 mem_cgroup_root_css = &memcg->css;
4334 page_counter_init(&memcg->memory, NULL);
4335 memcg->high = PAGE_COUNTER_MAX;
4336 memcg->soft_limit = PAGE_COUNTER_MAX;
4337 page_counter_init(&memcg->memsw, NULL);
4338 page_counter_init(&memcg->kmem, NULL);
4341 memcg->last_scanned_node = MAX_NUMNODES;
4342 INIT_LIST_HEAD(&memcg->oom_notify);
4343 memcg->move_charge_at_immigrate = 0;
4344 mutex_init(&memcg->thresholds_lock);
4345 spin_lock_init(&memcg->move_lock);
4346 vmpressure_init(&memcg->vmpressure);
4347 INIT_LIST_HEAD(&memcg->event_list);
4348 spin_lock_init(&memcg->event_list_lock);
4349 #ifdef CONFIG_MEMCG_KMEM
4350 memcg->kmemcg_id = -1;
4352 #ifdef CONFIG_CGROUP_WRITEBACK
4353 INIT_LIST_HEAD(&memcg->cgwb_list);
4355 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4359 idr_remove(&mem_cgroup_idr, memcg->id.id);
4360 __mem_cgroup_free(memcg);
4361 return ERR_PTR(error);
4365 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4367 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4368 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4371 /* Online state pins memcg ID, memcg ID pins CSS */
4372 mem_cgroup_id_get(mem_cgroup_from_css(css));
4378 mutex_lock(&memcg_create_mutex);
4380 memcg->use_hierarchy = parent->use_hierarchy;
4381 memcg->oom_kill_disable = parent->oom_kill_disable;
4382 memcg->swappiness = mem_cgroup_swappiness(parent);
4384 if (parent->use_hierarchy) {
4385 page_counter_init(&memcg->memory, &parent->memory);
4386 memcg->high = PAGE_COUNTER_MAX;
4387 memcg->soft_limit = PAGE_COUNTER_MAX;
4388 page_counter_init(&memcg->memsw, &parent->memsw);
4389 page_counter_init(&memcg->kmem, &parent->kmem);
4392 * No need to take a reference to the parent because cgroup
4393 * core guarantees its existence.
4396 page_counter_init(&memcg->memory, NULL);
4397 memcg->high = PAGE_COUNTER_MAX;
4398 memcg->soft_limit = PAGE_COUNTER_MAX;
4399 page_counter_init(&memcg->memsw, NULL);
4400 page_counter_init(&memcg->kmem, NULL);
4402 * Deeper hierachy with use_hierarchy == false doesn't make
4403 * much sense so let cgroup subsystem know about this
4404 * unfortunate state in our controller.
4406 if (parent != root_mem_cgroup)
4407 memory_cgrp_subsys.broken_hierarchy = true;
4409 mutex_unlock(&memcg_create_mutex);
4411 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4416 * Make sure the memcg is initialized: mem_cgroup_iter()
4417 * orders reading memcg->initialized against its callers
4418 * reading the memcg members.
4420 smp_store_release(&memcg->initialized, 1);
4425 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4427 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4428 struct mem_cgroup_event *event, *tmp;
4431 * Unregister events and notify userspace.
4432 * Notify userspace about cgroup removing only after rmdir of cgroup
4433 * directory to avoid race between userspace and kernelspace.
4435 spin_lock(&memcg->event_list_lock);
4436 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4437 list_del_init(&event->list);
4438 schedule_work(&event->remove);
4440 spin_unlock(&memcg->event_list_lock);
4442 vmpressure_cleanup(&memcg->vmpressure);
4444 memcg_deactivate_kmem(memcg);
4446 wb_memcg_offline(memcg);
4448 mem_cgroup_id_put(memcg);
4451 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4453 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4455 invalidate_reclaim_iterators(memcg);
4458 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4460 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4462 memcg_destroy_kmem(memcg);
4463 __mem_cgroup_free(memcg);
4467 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4468 * @css: the target css
4470 * Reset the states of the mem_cgroup associated with @css. This is
4471 * invoked when the userland requests disabling on the default hierarchy
4472 * but the memcg is pinned through dependency. The memcg should stop
4473 * applying policies and should revert to the vanilla state as it may be
4474 * made visible again.
4476 * The current implementation only resets the essential configurations.
4477 * This needs to be expanded to cover all the visible parts.
4479 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4481 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4483 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4484 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4485 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4487 memcg->high = PAGE_COUNTER_MAX;
4488 memcg->soft_limit = PAGE_COUNTER_MAX;
4489 memcg_wb_domain_size_changed(memcg);
4493 /* Handlers for move charge at task migration. */
4494 static int mem_cgroup_do_precharge(unsigned long count)
4498 /* Try a single bulk charge without reclaim first, kswapd may wake */
4499 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4501 mc.precharge += count;
4505 /* Try charges one by one with reclaim, but do not retry */
4507 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4517 * get_mctgt_type - get target type of moving charge
4518 * @vma: the vma the pte to be checked belongs
4519 * @addr: the address corresponding to the pte to be checked
4520 * @ptent: the pte to be checked
4521 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4524 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4525 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4526 * move charge. if @target is not NULL, the page is stored in target->page
4527 * with extra refcnt got(Callers should handle it).
4528 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4529 * target for charge migration. if @target is not NULL, the entry is stored
4532 * Called with pte lock held.
4539 enum mc_target_type {
4545 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4546 unsigned long addr, pte_t ptent)
4548 struct page *page = vm_normal_page(vma, addr, ptent);
4550 if (!page || !page_mapped(page))
4552 if (PageAnon(page)) {
4553 if (!(mc.flags & MOVE_ANON))
4556 if (!(mc.flags & MOVE_FILE))
4559 if (!get_page_unless_zero(page))
4566 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4567 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4569 struct page *page = NULL;
4570 swp_entry_t ent = pte_to_swp_entry(ptent);
4572 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4575 * Because lookup_swap_cache() updates some statistics counter,
4576 * we call find_get_page() with swapper_space directly.
4578 page = find_get_page(swap_address_space(ent), ent.val);
4579 if (do_swap_account)
4580 entry->val = ent.val;
4585 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4586 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4592 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4593 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4595 struct page *page = NULL;
4596 struct address_space *mapping;
4599 if (!vma->vm_file) /* anonymous vma */
4601 if (!(mc.flags & MOVE_FILE))
4604 mapping = vma->vm_file->f_mapping;
4605 pgoff = linear_page_index(vma, addr);
4607 /* page is moved even if it's not RSS of this task(page-faulted). */
4609 /* shmem/tmpfs may report page out on swap: account for that too. */
4610 if (shmem_mapping(mapping)) {
4611 page = find_get_entry(mapping, pgoff);
4612 if (radix_tree_exceptional_entry(page)) {
4613 swp_entry_t swp = radix_to_swp_entry(page);
4614 if (do_swap_account)
4616 page = find_get_page(swap_address_space(swp), swp.val);
4619 page = find_get_page(mapping, pgoff);
4621 page = find_get_page(mapping, pgoff);
4627 * mem_cgroup_move_account - move account of the page
4629 * @nr_pages: number of regular pages (>1 for huge pages)
4630 * @from: mem_cgroup which the page is moved from.
4631 * @to: mem_cgroup which the page is moved to. @from != @to.
4633 * The caller must confirm following.
4634 * - page is not on LRU (isolate_page() is useful.)
4635 * - compound_lock is held when nr_pages > 1
4637 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4640 static int mem_cgroup_move_account(struct page *page,
4641 unsigned int nr_pages,
4642 struct mem_cgroup *from,
4643 struct mem_cgroup *to)
4645 unsigned long flags;
4649 VM_BUG_ON(from == to);
4650 VM_BUG_ON_PAGE(PageLRU(page), page);
4652 * The page is isolated from LRU. So, collapse function
4653 * will not handle this page. But page splitting can happen.
4654 * Do this check under compound_page_lock(). The caller should
4658 if (nr_pages > 1 && !PageTransHuge(page))
4662 * Prevent mem_cgroup_replace_page() from looking at
4663 * page->mem_cgroup of its source page while we change it.
4665 if (!trylock_page(page))
4669 if (page->mem_cgroup != from)
4672 anon = PageAnon(page);
4674 spin_lock_irqsave(&from->move_lock, flags);
4676 if (!anon && page_mapped(page)) {
4677 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4679 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4684 * move_lock grabbed above and caller set from->moving_account, so
4685 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4686 * So mapping should be stable for dirty pages.
4688 if (!anon && PageDirty(page)) {
4689 struct address_space *mapping = page_mapping(page);
4691 if (mapping_cap_account_dirty(mapping)) {
4692 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4694 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4699 if (PageWriteback(page)) {
4700 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4702 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4707 * It is safe to change page->mem_cgroup here because the page
4708 * is referenced, charged, and isolated - we can't race with
4709 * uncharging, charging, migration, or LRU putback.
4712 /* caller should have done css_get */
4713 page->mem_cgroup = to;
4714 spin_unlock_irqrestore(&from->move_lock, flags);
4718 local_lock_irq(event_lock);
4719 mem_cgroup_charge_statistics(to, page, nr_pages);
4720 memcg_check_events(to, page);
4721 mem_cgroup_charge_statistics(from, page, -nr_pages);
4722 memcg_check_events(from, page);
4723 local_unlock_irq(event_lock);
4730 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4731 unsigned long addr, pte_t ptent, union mc_target *target)
4733 struct page *page = NULL;
4734 enum mc_target_type ret = MC_TARGET_NONE;
4735 swp_entry_t ent = { .val = 0 };
4737 if (pte_present(ptent))
4738 page = mc_handle_present_pte(vma, addr, ptent);
4739 else if (is_swap_pte(ptent))
4740 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4741 else if (pte_none(ptent))
4742 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4744 if (!page && !ent.val)
4748 * Do only loose check w/o serialization.
4749 * mem_cgroup_move_account() checks the page is valid or
4750 * not under LRU exclusion.
4752 if (page->mem_cgroup == mc.from) {
4753 ret = MC_TARGET_PAGE;
4755 target->page = page;
4757 if (!ret || !target)
4760 /* There is a swap entry and a page doesn't exist or isn't charged */
4761 if (ent.val && !ret &&
4762 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4763 ret = MC_TARGET_SWAP;
4770 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4772 * We don't consider swapping or file mapped pages because THP does not
4773 * support them for now.
4774 * Caller should make sure that pmd_trans_huge(pmd) is true.
4776 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4777 unsigned long addr, pmd_t pmd, union mc_target *target)
4779 struct page *page = NULL;
4780 enum mc_target_type ret = MC_TARGET_NONE;
4782 page = pmd_page(pmd);
4783 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4784 if (!(mc.flags & MOVE_ANON))
4786 if (page->mem_cgroup == mc.from) {
4787 ret = MC_TARGET_PAGE;
4790 target->page = page;
4796 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4797 unsigned long addr, pmd_t pmd, union mc_target *target)
4799 return MC_TARGET_NONE;
4803 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4804 unsigned long addr, unsigned long end,
4805 struct mm_walk *walk)
4807 struct vm_area_struct *vma = walk->vma;
4811 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4812 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4813 mc.precharge += HPAGE_PMD_NR;
4818 if (pmd_trans_unstable(pmd))
4820 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4821 for (; addr != end; pte++, addr += PAGE_SIZE)
4822 if (get_mctgt_type(vma, addr, *pte, NULL))
4823 mc.precharge++; /* increment precharge temporarily */
4824 pte_unmap_unlock(pte - 1, ptl);
4830 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4832 unsigned long precharge;
4834 struct mm_walk mem_cgroup_count_precharge_walk = {
4835 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4838 down_read(&mm->mmap_sem);
4839 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4840 up_read(&mm->mmap_sem);
4842 precharge = mc.precharge;
4848 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4850 unsigned long precharge = mem_cgroup_count_precharge(mm);
4852 VM_BUG_ON(mc.moving_task);
4853 mc.moving_task = current;
4854 return mem_cgroup_do_precharge(precharge);
4857 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4858 static void __mem_cgroup_clear_mc(void)
4860 struct mem_cgroup *from = mc.from;
4861 struct mem_cgroup *to = mc.to;
4863 /* we must uncharge all the leftover precharges from mc.to */
4865 cancel_charge(mc.to, mc.precharge);
4869 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4870 * we must uncharge here.
4872 if (mc.moved_charge) {
4873 cancel_charge(mc.from, mc.moved_charge);
4874 mc.moved_charge = 0;
4876 /* we must fixup refcnts and charges */
4877 if (mc.moved_swap) {
4878 /* uncharge swap account from the old cgroup */
4879 if (!mem_cgroup_is_root(mc.from))
4880 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4882 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4885 * we charged both to->memory and to->memsw, so we
4886 * should uncharge to->memory.
4888 if (!mem_cgroup_is_root(mc.to))
4889 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4891 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4892 css_put_many(&mc.to->css, mc.moved_swap);
4896 memcg_oom_recover(from);
4897 memcg_oom_recover(to);
4898 wake_up_all(&mc.waitq);
4901 static void mem_cgroup_clear_mc(void)
4903 struct mm_struct *mm = mc.mm;
4906 * we must clear moving_task before waking up waiters at the end of
4909 mc.moving_task = NULL;
4910 __mem_cgroup_clear_mc();
4911 spin_lock(&mc.lock);
4915 spin_unlock(&mc.lock);
4920 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4922 struct cgroup_subsys_state *css;
4923 struct mem_cgroup *memcg;
4924 struct mem_cgroup *from;
4925 struct task_struct *leader, *p;
4926 struct mm_struct *mm;
4927 unsigned long move_flags;
4930 /* charge immigration isn't supported on the default hierarchy */
4931 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4935 * Multi-process migrations only happen on the default hierarchy
4936 * where charge immigration is not used. Perform charge
4937 * immigration if @tset contains a leader and whine if there are
4941 cgroup_taskset_for_each_leader(leader, css, tset) {
4944 memcg = mem_cgroup_from_css(css);
4950 * We are now commited to this value whatever it is. Changes in this
4951 * tunable will only affect upcoming migrations, not the current one.
4952 * So we need to save it, and keep it going.
4954 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4958 from = mem_cgroup_from_task(p);
4960 VM_BUG_ON(from == memcg);
4962 mm = get_task_mm(p);
4965 /* We move charges only when we move a owner of the mm */
4966 if (mm->owner == p) {
4969 VM_BUG_ON(mc.precharge);
4970 VM_BUG_ON(mc.moved_charge);
4971 VM_BUG_ON(mc.moved_swap);
4973 spin_lock(&mc.lock);
4977 mc.flags = move_flags;
4978 spin_unlock(&mc.lock);
4979 /* We set mc.moving_task later */
4981 ret = mem_cgroup_precharge_mc(mm);
4983 mem_cgroup_clear_mc();
4990 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4993 mem_cgroup_clear_mc();
4996 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4997 unsigned long addr, unsigned long end,
4998 struct mm_walk *walk)
5001 struct vm_area_struct *vma = walk->vma;
5004 enum mc_target_type target_type;
5005 union mc_target target;
5009 * We don't take compound_lock() here but no race with splitting thp
5011 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5012 * under splitting, which means there's no concurrent thp split,
5013 * - if another thread runs into split_huge_page() just after we
5014 * entered this if-block, the thread must wait for page table lock
5015 * to be unlocked in __split_huge_page_splitting(), where the main
5016 * part of thp split is not executed yet.
5018 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5019 if (mc.precharge < HPAGE_PMD_NR) {
5023 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5024 if (target_type == MC_TARGET_PAGE) {
5026 if (!isolate_lru_page(page)) {
5027 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5029 mc.precharge -= HPAGE_PMD_NR;
5030 mc.moved_charge += HPAGE_PMD_NR;
5032 putback_lru_page(page);
5040 if (pmd_trans_unstable(pmd))
5043 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5044 for (; addr != end; addr += PAGE_SIZE) {
5045 pte_t ptent = *(pte++);
5051 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5052 case MC_TARGET_PAGE:
5054 if (isolate_lru_page(page))
5056 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5058 /* we uncharge from mc.from later. */
5061 putback_lru_page(page);
5062 put: /* get_mctgt_type() gets the page */
5065 case MC_TARGET_SWAP:
5067 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5069 /* we fixup refcnts and charges later. */
5077 pte_unmap_unlock(pte - 1, ptl);
5082 * We have consumed all precharges we got in can_attach().
5083 * We try charge one by one, but don't do any additional
5084 * charges to mc.to if we have failed in charge once in attach()
5087 ret = mem_cgroup_do_precharge(1);
5095 static void mem_cgroup_move_charge(void)
5097 struct mm_walk mem_cgroup_move_charge_walk = {
5098 .pmd_entry = mem_cgroup_move_charge_pte_range,
5102 lru_add_drain_all();
5104 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5105 * move_lock while we're moving its pages to another memcg.
5106 * Then wait for already started RCU-only updates to finish.
5108 atomic_inc(&mc.from->moving_account);
5111 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5113 * Someone who are holding the mmap_sem might be waiting in
5114 * waitq. So we cancel all extra charges, wake up all waiters,
5115 * and retry. Because we cancel precharges, we might not be able
5116 * to move enough charges, but moving charge is a best-effort
5117 * feature anyway, so it wouldn't be a big problem.
5119 __mem_cgroup_clear_mc();
5124 * When we have consumed all precharges and failed in doing
5125 * additional charge, the page walk just aborts.
5127 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5128 up_read(&mc.mm->mmap_sem);
5129 atomic_dec(&mc.from->moving_account);
5132 static void mem_cgroup_move_task(void)
5135 mem_cgroup_move_charge();
5136 mem_cgroup_clear_mc();
5139 #else /* !CONFIG_MMU */
5140 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5144 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5147 static void mem_cgroup_move_task(void)
5153 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5154 * to verify whether we're attached to the default hierarchy on each mount
5157 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5160 * use_hierarchy is forced on the default hierarchy. cgroup core
5161 * guarantees that @root doesn't have any children, so turning it
5162 * on for the root memcg is enough.
5164 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5165 root_mem_cgroup->use_hierarchy = true;
5167 root_mem_cgroup->use_hierarchy = false;
5170 static u64 memory_current_read(struct cgroup_subsys_state *css,
5173 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5175 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5178 static int memory_low_show(struct seq_file *m, void *v)
5180 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5181 unsigned long low = READ_ONCE(memcg->low);
5183 if (low == PAGE_COUNTER_MAX)
5184 seq_puts(m, "max\n");
5186 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5191 static ssize_t memory_low_write(struct kernfs_open_file *of,
5192 char *buf, size_t nbytes, loff_t off)
5194 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5198 buf = strstrip(buf);
5199 err = page_counter_memparse(buf, "max", &low);
5208 static int memory_high_show(struct seq_file *m, void *v)
5210 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5211 unsigned long high = READ_ONCE(memcg->high);
5213 if (high == PAGE_COUNTER_MAX)
5214 seq_puts(m, "max\n");
5216 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5221 static ssize_t memory_high_write(struct kernfs_open_file *of,
5222 char *buf, size_t nbytes, loff_t off)
5224 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5225 unsigned long nr_pages;
5229 buf = strstrip(buf);
5230 err = page_counter_memparse(buf, "max", &high);
5236 nr_pages = page_counter_read(&memcg->memory);
5237 if (nr_pages > high)
5238 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5241 memcg_wb_domain_size_changed(memcg);
5245 static int memory_max_show(struct seq_file *m, void *v)
5247 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5248 unsigned long max = READ_ONCE(memcg->memory.limit);
5250 if (max == PAGE_COUNTER_MAX)
5251 seq_puts(m, "max\n");
5253 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5258 static ssize_t memory_max_write(struct kernfs_open_file *of,
5259 char *buf, size_t nbytes, loff_t off)
5261 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5262 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5263 bool drained = false;
5267 buf = strstrip(buf);
5268 err = page_counter_memparse(buf, "max", &max);
5272 xchg(&memcg->memory.limit, max);
5275 unsigned long nr_pages = page_counter_read(&memcg->memory);
5277 if (nr_pages <= max)
5280 if (signal_pending(current)) {
5286 drain_all_stock(memcg);
5292 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5298 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5299 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5303 memcg_wb_domain_size_changed(memcg);
5307 static int memory_events_show(struct seq_file *m, void *v)
5309 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5311 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5312 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5313 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5314 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5319 static struct cftype memory_files[] = {
5322 .flags = CFTYPE_NOT_ON_ROOT,
5323 .read_u64 = memory_current_read,
5327 .flags = CFTYPE_NOT_ON_ROOT,
5328 .seq_show = memory_low_show,
5329 .write = memory_low_write,
5333 .flags = CFTYPE_NOT_ON_ROOT,
5334 .seq_show = memory_high_show,
5335 .write = memory_high_write,
5339 .flags = CFTYPE_NOT_ON_ROOT,
5340 .seq_show = memory_max_show,
5341 .write = memory_max_write,
5345 .flags = CFTYPE_NOT_ON_ROOT,
5346 .file_offset = offsetof(struct mem_cgroup, events_file),
5347 .seq_show = memory_events_show,
5352 struct cgroup_subsys memory_cgrp_subsys = {
5353 .css_alloc = mem_cgroup_css_alloc,
5354 .css_online = mem_cgroup_css_online,
5355 .css_offline = mem_cgroup_css_offline,
5356 .css_released = mem_cgroup_css_released,
5357 .css_free = mem_cgroup_css_free,
5358 .css_reset = mem_cgroup_css_reset,
5359 .can_attach = mem_cgroup_can_attach,
5360 .cancel_attach = mem_cgroup_cancel_attach,
5361 .post_attach = mem_cgroup_move_task,
5362 .bind = mem_cgroup_bind,
5363 .dfl_cftypes = memory_files,
5364 .legacy_cftypes = mem_cgroup_legacy_files,
5369 * mem_cgroup_low - check if memory consumption is below the normal range
5370 * @root: the highest ancestor to consider
5371 * @memcg: the memory cgroup to check
5373 * Returns %true if memory consumption of @memcg, and that of all
5374 * configurable ancestors up to @root, is below the normal range.
5376 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5378 if (mem_cgroup_disabled())
5382 * The toplevel group doesn't have a configurable range, so
5383 * it's never low when looked at directly, and it is not
5384 * considered an ancestor when assessing the hierarchy.
5387 if (memcg == root_mem_cgroup)
5390 if (page_counter_read(&memcg->memory) >= memcg->low)
5393 while (memcg != root) {
5394 memcg = parent_mem_cgroup(memcg);
5396 if (memcg == root_mem_cgroup)
5399 if (page_counter_read(&memcg->memory) >= memcg->low)
5406 * mem_cgroup_try_charge - try charging a page
5407 * @page: page to charge
5408 * @mm: mm context of the victim
5409 * @gfp_mask: reclaim mode
5410 * @memcgp: charged memcg return
5412 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5413 * pages according to @gfp_mask if necessary.
5415 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5416 * Otherwise, an error code is returned.
5418 * After page->mapping has been set up, the caller must finalize the
5419 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5420 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5422 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5423 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5425 struct mem_cgroup *memcg = NULL;
5426 unsigned int nr_pages = 1;
5429 if (mem_cgroup_disabled())
5432 if (PageSwapCache(page)) {
5434 * Every swap fault against a single page tries to charge the
5435 * page, bail as early as possible. shmem_unuse() encounters
5436 * already charged pages, too. The USED bit is protected by
5437 * the page lock, which serializes swap cache removal, which
5438 * in turn serializes uncharging.
5440 VM_BUG_ON_PAGE(!PageLocked(page), page);
5441 if (page->mem_cgroup)
5444 if (do_swap_account) {
5445 swp_entry_t ent = { .val = page_private(page), };
5446 unsigned short id = lookup_swap_cgroup_id(ent);
5449 memcg = mem_cgroup_from_id(id);
5450 if (memcg && !css_tryget_online(&memcg->css))
5456 if (PageTransHuge(page)) {
5457 nr_pages <<= compound_order(page);
5458 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5462 memcg = get_mem_cgroup_from_mm(mm);
5464 ret = try_charge(memcg, gfp_mask, nr_pages);
5466 css_put(&memcg->css);
5473 * mem_cgroup_commit_charge - commit a page charge
5474 * @page: page to charge
5475 * @memcg: memcg to charge the page to
5476 * @lrucare: page might be on LRU already
5478 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5479 * after page->mapping has been set up. This must happen atomically
5480 * as part of the page instantiation, i.e. under the page table lock
5481 * for anonymous pages, under the page lock for page and swap cache.
5483 * In addition, the page must not be on the LRU during the commit, to
5484 * prevent racing with task migration. If it might be, use @lrucare.
5486 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5488 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5491 unsigned int nr_pages = 1;
5493 VM_BUG_ON_PAGE(!page->mapping, page);
5494 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5496 if (mem_cgroup_disabled())
5499 * Swap faults will attempt to charge the same page multiple
5500 * times. But reuse_swap_page() might have removed the page
5501 * from swapcache already, so we can't check PageSwapCache().
5506 commit_charge(page, memcg, lrucare);
5508 if (PageTransHuge(page)) {
5509 nr_pages <<= compound_order(page);
5510 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5513 local_lock_irq(event_lock);
5514 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5515 memcg_check_events(memcg, page);
5516 local_unlock_irq(event_lock);
5518 if (do_swap_account && PageSwapCache(page)) {
5519 swp_entry_t entry = { .val = page_private(page) };
5521 * The swap entry might not get freed for a long time,
5522 * let's not wait for it. The page already received a
5523 * memory+swap charge, drop the swap entry duplicate.
5525 mem_cgroup_uncharge_swap(entry);
5530 * mem_cgroup_cancel_charge - cancel a page charge
5531 * @page: page to charge
5532 * @memcg: memcg to charge the page to
5534 * Cancel a charge transaction started by mem_cgroup_try_charge().
5536 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5538 unsigned int nr_pages = 1;
5540 if (mem_cgroup_disabled())
5543 * Swap faults will attempt to charge the same page multiple
5544 * times. But reuse_swap_page() might have removed the page
5545 * from swapcache already, so we can't check PageSwapCache().
5550 if (PageTransHuge(page)) {
5551 nr_pages <<= compound_order(page);
5552 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5555 cancel_charge(memcg, nr_pages);
5558 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5559 unsigned long nr_anon, unsigned long nr_file,
5560 unsigned long nr_huge, struct page *dummy_page)
5562 unsigned long nr_pages = nr_anon + nr_file;
5563 unsigned long flags;
5565 if (!mem_cgroup_is_root(memcg)) {
5566 page_counter_uncharge(&memcg->memory, nr_pages);
5567 if (do_swap_account)
5568 page_counter_uncharge(&memcg->memsw, nr_pages);
5569 memcg_oom_recover(memcg);
5572 local_lock_irqsave(event_lock, flags);
5573 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5574 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5575 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5576 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5577 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5578 memcg_check_events(memcg, dummy_page);
5579 local_unlock_irqrestore(event_lock, flags);
5581 if (!mem_cgroup_is_root(memcg))
5582 css_put_many(&memcg->css, nr_pages);
5585 static void uncharge_list(struct list_head *page_list)
5587 struct mem_cgroup *memcg = NULL;
5588 unsigned long nr_anon = 0;
5589 unsigned long nr_file = 0;
5590 unsigned long nr_huge = 0;
5591 unsigned long pgpgout = 0;
5592 struct list_head *next;
5595 next = page_list->next;
5597 unsigned int nr_pages = 1;
5599 page = list_entry(next, struct page, lru);
5600 next = page->lru.next;
5602 VM_BUG_ON_PAGE(PageLRU(page), page);
5603 VM_BUG_ON_PAGE(page_count(page), page);
5605 if (!page->mem_cgroup)
5609 * Nobody should be changing or seriously looking at
5610 * page->mem_cgroup at this point, we have fully
5611 * exclusive access to the page.
5614 if (memcg != page->mem_cgroup) {
5616 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5618 pgpgout = nr_anon = nr_file = nr_huge = 0;
5620 memcg = page->mem_cgroup;
5623 if (PageTransHuge(page)) {
5624 nr_pages <<= compound_order(page);
5625 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5626 nr_huge += nr_pages;
5630 nr_anon += nr_pages;
5632 nr_file += nr_pages;
5634 page->mem_cgroup = NULL;
5637 } while (next != page_list);
5640 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5645 * mem_cgroup_uncharge - uncharge a page
5646 * @page: page to uncharge
5648 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5649 * mem_cgroup_commit_charge().
5651 void mem_cgroup_uncharge(struct page *page)
5653 if (mem_cgroup_disabled())
5656 /* Don't touch page->lru of any random page, pre-check: */
5657 if (!page->mem_cgroup)
5660 INIT_LIST_HEAD(&page->lru);
5661 uncharge_list(&page->lru);
5665 * mem_cgroup_uncharge_list - uncharge a list of page
5666 * @page_list: list of pages to uncharge
5668 * Uncharge a list of pages previously charged with
5669 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5671 void mem_cgroup_uncharge_list(struct list_head *page_list)
5673 if (mem_cgroup_disabled())
5676 if (!list_empty(page_list))
5677 uncharge_list(page_list);
5681 * mem_cgroup_replace_page - migrate a charge to another page
5682 * @oldpage: currently charged page
5683 * @newpage: page to transfer the charge to
5685 * Migrate the charge from @oldpage to @newpage.
5687 * Both pages must be locked, @newpage->mapping must be set up.
5688 * Either or both pages might be on the LRU already.
5690 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5692 struct mem_cgroup *memcg;
5695 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5696 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5697 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5698 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5701 if (mem_cgroup_disabled())
5704 /* Page cache replacement: new page already charged? */
5705 if (newpage->mem_cgroup)
5708 /* Swapcache readahead pages can get replaced before being charged */
5709 memcg = oldpage->mem_cgroup;
5713 lock_page_lru(oldpage, &isolated);
5714 oldpage->mem_cgroup = NULL;
5715 unlock_page_lru(oldpage, isolated);
5717 commit_charge(newpage, memcg, true);
5721 * subsys_initcall() for memory controller.
5723 * Some parts like hotcpu_notifier() have to be initialized from this context
5724 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5725 * everything that doesn't depend on a specific mem_cgroup structure should
5726 * be initialized from here.
5728 static int __init mem_cgroup_init(void)
5732 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5734 for_each_possible_cpu(cpu)
5735 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5738 for_each_node(node) {
5739 struct mem_cgroup_tree_per_node *rtpn;
5742 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5743 node_online(node) ? node : NUMA_NO_NODE);
5745 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5746 struct mem_cgroup_tree_per_zone *rtpz;
5748 rtpz = &rtpn->rb_tree_per_zone[zone];
5749 rtpz->rb_root = RB_ROOT;
5750 spin_lock_init(&rtpz->lock);
5752 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5757 subsys_initcall(mem_cgroup_init);
5759 #ifdef CONFIG_MEMCG_SWAP
5761 * mem_cgroup_swapout - transfer a memsw charge to swap
5762 * @page: page whose memsw charge to transfer
5763 * @entry: swap entry to move the charge to
5765 * Transfer the memsw charge of @page to @entry.
5767 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5769 struct mem_cgroup *memcg, *swap_memcg;
5770 unsigned short oldid;
5771 unsigned long flags;
5773 VM_BUG_ON_PAGE(PageLRU(page), page);
5774 VM_BUG_ON_PAGE(page_count(page), page);
5776 if (!do_swap_account)
5779 memcg = page->mem_cgroup;
5781 /* Readahead page, never charged */
5786 * In case the memcg owning these pages has been offlined and doesn't
5787 * have an ID allocated to it anymore, charge the closest online
5788 * ancestor for the swap instead and transfer the memory+swap charge.
5790 swap_memcg = mem_cgroup_id_get_online(memcg);
5791 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5792 VM_BUG_ON_PAGE(oldid, page);
5793 mem_cgroup_swap_statistics(swap_memcg, true);
5795 page->mem_cgroup = NULL;
5797 if (!mem_cgroup_is_root(memcg))
5798 page_counter_uncharge(&memcg->memory, 1);
5800 if (memcg != swap_memcg) {
5801 if (!mem_cgroup_is_root(swap_memcg))
5802 page_counter_charge(&swap_memcg->memsw, 1);
5803 page_counter_uncharge(&memcg->memsw, 1);
5807 * Interrupts should be disabled here because the caller holds the
5808 * mapping->tree_lock lock which is taken with interrupts-off. It is
5809 * important here to have the interrupts disabled because it is the
5810 * only synchronisation we have for udpating the per-CPU variables.
5812 local_lock_irqsave(event_lock, flags);
5813 #ifndef CONFIG_PREEMPT_RT_BASE
5814 VM_BUG_ON(!irqs_disabled());
5816 mem_cgroup_charge_statistics(memcg, page, -1);
5817 memcg_check_events(memcg, page);
5819 if (!mem_cgroup_is_root(memcg))
5820 css_put(&memcg->css);
5821 local_unlock_irqrestore(event_lock, flags);
5825 * mem_cgroup_uncharge_swap - uncharge a swap entry
5826 * @entry: swap entry to uncharge
5828 * Drop the memsw charge associated with @entry.
5830 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5832 struct mem_cgroup *memcg;
5835 if (!do_swap_account)
5838 id = swap_cgroup_record(entry, 0);
5840 memcg = mem_cgroup_from_id(id);
5842 if (!mem_cgroup_is_root(memcg))
5843 page_counter_uncharge(&memcg->memsw, 1);
5844 mem_cgroup_swap_statistics(memcg, false);
5845 mem_cgroup_id_put(memcg);
5850 /* for remember boot option*/
5851 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5852 static int really_do_swap_account __initdata = 1;
5854 static int really_do_swap_account __initdata;
5857 static int __init enable_swap_account(char *s)
5859 if (!strcmp(s, "1"))
5860 really_do_swap_account = 1;
5861 else if (!strcmp(s, "0"))
5862 really_do_swap_account = 0;
5865 __setup("swapaccount=", enable_swap_account);
5867 static struct cftype memsw_cgroup_files[] = {
5869 .name = "memsw.usage_in_bytes",
5870 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5871 .read_u64 = mem_cgroup_read_u64,
5874 .name = "memsw.max_usage_in_bytes",
5875 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5876 .write = mem_cgroup_reset,
5877 .read_u64 = mem_cgroup_read_u64,
5880 .name = "memsw.limit_in_bytes",
5881 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5882 .write = mem_cgroup_write,
5883 .read_u64 = mem_cgroup_read_u64,
5886 .name = "memsw.failcnt",
5887 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5888 .write = mem_cgroup_reset,
5889 .read_u64 = mem_cgroup_read_u64,
5891 { }, /* terminate */
5894 static int __init mem_cgroup_swap_init(void)
5896 if (!mem_cgroup_disabled() && really_do_swap_account) {
5897 do_swap_account = 1;
5898 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5899 memsw_cgroup_files));
5903 subsys_initcall(mem_cgroup_swap_init);
5905 #endif /* CONFIG_MEMCG_SWAP */