1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
69 #include <net/tcp_memcontrol.h>
72 #include <asm/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
77 EXPORT_SYMBOL(memory_cgrp_subsys);
79 #define MEM_CGROUP_RECLAIM_RETRIES 5
80 static struct mem_cgroup *root_mem_cgroup __read_mostly;
81 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
87 #define do_swap_account 0
90 static const char * const mem_cgroup_stat_names[] = {
100 static const char * const mem_cgroup_events_names[] = {
107 static const char * const mem_cgroup_lru_names[] = {
115 #define THRESHOLDS_EVENTS_TARGET 128
116 #define SOFTLIMIT_EVENTS_TARGET 1024
117 #define NUMAINFO_EVENTS_TARGET 1024
120 * Cgroups above their limits are maintained in a RB-Tree, independent of
121 * their hierarchy representation
124 struct mem_cgroup_tree_per_zone {
125 struct rb_root rb_root;
129 struct mem_cgroup_tree_per_node {
130 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
133 struct mem_cgroup_tree {
134 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
137 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
140 struct mem_cgroup_eventfd_list {
141 struct list_head list;
142 struct eventfd_ctx *eventfd;
146 * cgroup_event represents events which userspace want to receive.
148 struct mem_cgroup_event {
150 * memcg which the event belongs to.
152 struct mem_cgroup *memcg;
154 * eventfd to signal userspace about the event.
156 struct eventfd_ctx *eventfd;
158 * Each of these stored in a list by the cgroup.
160 struct list_head list;
162 * register_event() callback will be used to add new userspace
163 * waiter for changes related to this event. Use eventfd_signal()
164 * on eventfd to send notification to userspace.
166 int (*register_event)(struct mem_cgroup *memcg,
167 struct eventfd_ctx *eventfd, const char *args);
169 * unregister_event() callback will be called when userspace closes
170 * the eventfd or on cgroup removing. This callback must be set,
171 * if you want provide notification functionality.
173 void (*unregister_event)(struct mem_cgroup *memcg,
174 struct eventfd_ctx *eventfd);
176 * All fields below needed to unregister event when
177 * userspace closes eventfd.
180 wait_queue_head_t *wqh;
182 struct work_struct remove;
185 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
186 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
188 /* Stuffs for move charges at task migration. */
190 * Types of charges to be moved.
192 #define MOVE_ANON 0x1U
193 #define MOVE_FILE 0x2U
194 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
196 /* "mc" and its members are protected by cgroup_mutex */
197 static struct move_charge_struct {
198 spinlock_t lock; /* for from, to */
199 struct mm_struct *mm;
200 struct mem_cgroup *from;
201 struct mem_cgroup *to;
203 unsigned long precharge;
204 unsigned long moved_charge;
205 unsigned long moved_swap;
206 struct task_struct *moving_task; /* a task moving charges */
207 wait_queue_head_t waitq; /* a waitq for other context */
209 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
210 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
214 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
215 * limit reclaim to prevent infinite loops, if they ever occur.
217 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
218 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
221 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
222 MEM_CGROUP_CHARGE_TYPE_ANON,
223 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
224 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
228 /* for encoding cft->private value on file */
236 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
237 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
238 #define MEMFILE_ATTR(val) ((val) & 0xffff)
239 /* Used for OOM nofiier */
240 #define OOM_CONTROL (0)
243 * The memcg_create_mutex will be held whenever a new cgroup is created.
244 * As a consequence, any change that needs to protect against new child cgroups
245 * appearing has to hold it as well.
247 static DEFINE_MUTEX(memcg_create_mutex);
249 /* Some nice accessors for the vmpressure. */
250 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
253 memcg = root_mem_cgroup;
254 return &memcg->vmpressure;
257 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
259 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
262 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
264 return (memcg == root_mem_cgroup);
268 * We restrict the id in the range of [1, 65535], so it can fit into
271 #define MEM_CGROUP_ID_MAX USHRT_MAX
273 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
278 /* Writing them here to avoid exposing memcg's inner layout */
279 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
281 void sock_update_memcg(struct sock *sk)
283 if (mem_cgroup_sockets_enabled) {
284 struct mem_cgroup *memcg;
285 struct cg_proto *cg_proto;
287 BUG_ON(!sk->sk_prot->proto_cgroup);
289 /* Socket cloning can throw us here with sk_cgrp already
290 * filled. It won't however, necessarily happen from
291 * process context. So the test for root memcg given
292 * the current task's memcg won't help us in this case.
294 * Respecting the original socket's memcg is a better
295 * decision in this case.
298 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
299 css_get(&sk->sk_cgrp->memcg->css);
304 memcg = mem_cgroup_from_task(current);
305 cg_proto = sk->sk_prot->proto_cgroup(memcg);
306 if (cg_proto && test_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags) &&
307 css_tryget_online(&memcg->css)) {
308 sk->sk_cgrp = cg_proto;
313 EXPORT_SYMBOL(sock_update_memcg);
315 void sock_release_memcg(struct sock *sk)
317 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
318 struct mem_cgroup *memcg;
319 WARN_ON(!sk->sk_cgrp->memcg);
320 memcg = sk->sk_cgrp->memcg;
321 css_put(&sk->sk_cgrp->memcg->css);
325 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
327 if (!memcg || mem_cgroup_is_root(memcg))
330 return &memcg->tcp_mem;
332 EXPORT_SYMBOL(tcp_proto_cgroup);
336 #ifdef CONFIG_MEMCG_KMEM
338 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
339 * The main reason for not using cgroup id for this:
340 * this works better in sparse environments, where we have a lot of memcgs,
341 * but only a few kmem-limited. Or also, if we have, for instance, 200
342 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
343 * 200 entry array for that.
345 * The current size of the caches array is stored in memcg_nr_cache_ids. It
346 * will double each time we have to increase it.
348 static DEFINE_IDA(memcg_cache_ida);
349 int memcg_nr_cache_ids;
351 /* Protects memcg_nr_cache_ids */
352 static DECLARE_RWSEM(memcg_cache_ids_sem);
354 void memcg_get_cache_ids(void)
356 down_read(&memcg_cache_ids_sem);
359 void memcg_put_cache_ids(void)
361 up_read(&memcg_cache_ids_sem);
365 * MIN_SIZE is different than 1, because we would like to avoid going through
366 * the alloc/free process all the time. In a small machine, 4 kmem-limited
367 * cgroups is a reasonable guess. In the future, it could be a parameter or
368 * tunable, but that is strictly not necessary.
370 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
371 * this constant directly from cgroup, but it is understandable that this is
372 * better kept as an internal representation in cgroup.c. In any case, the
373 * cgrp_id space is not getting any smaller, and we don't have to necessarily
374 * increase ours as well if it increases.
376 #define MEMCG_CACHES_MIN_SIZE 4
377 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
380 * A lot of the calls to the cache allocation functions are expected to be
381 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
382 * conditional to this static branch, we'll have to allow modules that does
383 * kmem_cache_alloc and the such to see this symbol as well
385 struct static_key memcg_kmem_enabled_key;
386 EXPORT_SYMBOL(memcg_kmem_enabled_key);
388 #endif /* CONFIG_MEMCG_KMEM */
390 static struct mem_cgroup_per_zone *
391 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
393 int nid = zone_to_nid(zone);
394 int zid = zone_idx(zone);
396 return &memcg->nodeinfo[nid]->zoneinfo[zid];
400 * mem_cgroup_css_from_page - css of the memcg associated with a page
401 * @page: page of interest
403 * If memcg is bound to the default hierarchy, css of the memcg associated
404 * with @page is returned. The returned css remains associated with @page
405 * until it is released.
407 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
410 * XXX: The above description of behavior on the default hierarchy isn't
411 * strictly true yet as replace_page_cache_page() can modify the
412 * association before @page is released even on the default hierarchy;
413 * however, the current and planned usages don't mix the the two functions
414 * and replace_page_cache_page() will soon be updated to make the invariant
417 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
419 struct mem_cgroup *memcg;
423 memcg = page->mem_cgroup;
425 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
426 memcg = root_mem_cgroup;
433 * page_cgroup_ino - return inode number of the memcg a page is charged to
436 * Look up the closest online ancestor of the memory cgroup @page is charged to
437 * and return its inode number or 0 if @page is not charged to any cgroup. It
438 * is safe to call this function without holding a reference to @page.
440 * Note, this function is inherently racy, because there is nothing to prevent
441 * the cgroup inode from getting torn down and potentially reallocated a moment
442 * after page_cgroup_ino() returns, so it only should be used by callers that
443 * do not care (such as procfs interfaces).
445 ino_t page_cgroup_ino(struct page *page)
447 struct mem_cgroup *memcg;
448 unsigned long ino = 0;
451 memcg = READ_ONCE(page->mem_cgroup);
452 while (memcg && !(memcg->css.flags & CSS_ONLINE))
453 memcg = parent_mem_cgroup(memcg);
455 ino = cgroup_ino(memcg->css.cgroup);
460 static struct mem_cgroup_per_zone *
461 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
463 int nid = page_to_nid(page);
464 int zid = page_zonenum(page);
466 return &memcg->nodeinfo[nid]->zoneinfo[zid];
469 static struct mem_cgroup_tree_per_zone *
470 soft_limit_tree_node_zone(int nid, int zid)
472 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
475 static struct mem_cgroup_tree_per_zone *
476 soft_limit_tree_from_page(struct page *page)
478 int nid = page_to_nid(page);
479 int zid = page_zonenum(page);
481 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
484 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
485 struct mem_cgroup_tree_per_zone *mctz,
486 unsigned long new_usage_in_excess)
488 struct rb_node **p = &mctz->rb_root.rb_node;
489 struct rb_node *parent = NULL;
490 struct mem_cgroup_per_zone *mz_node;
495 mz->usage_in_excess = new_usage_in_excess;
496 if (!mz->usage_in_excess)
500 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
502 if (mz->usage_in_excess < mz_node->usage_in_excess)
505 * We can't avoid mem cgroups that are over their soft
506 * limit by the same amount
508 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
511 rb_link_node(&mz->tree_node, parent, p);
512 rb_insert_color(&mz->tree_node, &mctz->rb_root);
516 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
517 struct mem_cgroup_tree_per_zone *mctz)
521 rb_erase(&mz->tree_node, &mctz->rb_root);
525 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
526 struct mem_cgroup_tree_per_zone *mctz)
530 spin_lock_irqsave(&mctz->lock, flags);
531 __mem_cgroup_remove_exceeded(mz, mctz);
532 spin_unlock_irqrestore(&mctz->lock, flags);
535 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
537 unsigned long nr_pages = page_counter_read(&memcg->memory);
538 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
539 unsigned long excess = 0;
541 if (nr_pages > soft_limit)
542 excess = nr_pages - soft_limit;
547 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
549 unsigned long excess;
550 struct mem_cgroup_per_zone *mz;
551 struct mem_cgroup_tree_per_zone *mctz;
553 mctz = soft_limit_tree_from_page(page);
555 * Necessary to update all ancestors when hierarchy is used.
556 * because their event counter is not touched.
558 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
559 mz = mem_cgroup_page_zoneinfo(memcg, page);
560 excess = soft_limit_excess(memcg);
562 * We have to update the tree if mz is on RB-tree or
563 * mem is over its softlimit.
565 if (excess || mz->on_tree) {
568 spin_lock_irqsave(&mctz->lock, flags);
569 /* if on-tree, remove it */
571 __mem_cgroup_remove_exceeded(mz, mctz);
573 * Insert again. mz->usage_in_excess will be updated.
574 * If excess is 0, no tree ops.
576 __mem_cgroup_insert_exceeded(mz, mctz, excess);
577 spin_unlock_irqrestore(&mctz->lock, flags);
582 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
584 struct mem_cgroup_tree_per_zone *mctz;
585 struct mem_cgroup_per_zone *mz;
589 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
590 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
591 mctz = soft_limit_tree_node_zone(nid, zid);
592 mem_cgroup_remove_exceeded(mz, mctz);
597 static struct mem_cgroup_per_zone *
598 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
600 struct rb_node *rightmost = NULL;
601 struct mem_cgroup_per_zone *mz;
605 rightmost = rb_last(&mctz->rb_root);
607 goto done; /* Nothing to reclaim from */
609 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
611 * Remove the node now but someone else can add it back,
612 * we will to add it back at the end of reclaim to its correct
613 * position in the tree.
615 __mem_cgroup_remove_exceeded(mz, mctz);
616 if (!soft_limit_excess(mz->memcg) ||
617 !css_tryget_online(&mz->memcg->css))
623 static struct mem_cgroup_per_zone *
624 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
626 struct mem_cgroup_per_zone *mz;
628 spin_lock_irq(&mctz->lock);
629 mz = __mem_cgroup_largest_soft_limit_node(mctz);
630 spin_unlock_irq(&mctz->lock);
635 * Return page count for single (non recursive) @memcg.
637 * Implementation Note: reading percpu statistics for memcg.
639 * Both of vmstat[] and percpu_counter has threshold and do periodic
640 * synchronization to implement "quick" read. There are trade-off between
641 * reading cost and precision of value. Then, we may have a chance to implement
642 * a periodic synchronization of counter in memcg's counter.
644 * But this _read() function is used for user interface now. The user accounts
645 * memory usage by memory cgroup and he _always_ requires exact value because
646 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
647 * have to visit all online cpus and make sum. So, for now, unnecessary
648 * synchronization is not implemented. (just implemented for cpu hotplug)
650 * If there are kernel internal actions which can make use of some not-exact
651 * value, and reading all cpu value can be performance bottleneck in some
652 * common workload, threshold and synchronization as vmstat[] should be
656 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
661 /* Per-cpu values can be negative, use a signed accumulator */
662 for_each_possible_cpu(cpu)
663 val += per_cpu(memcg->stat->count[idx], cpu);
665 * Summing races with updates, so val may be negative. Avoid exposing
666 * transient negative values.
673 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
674 enum mem_cgroup_events_index idx)
676 unsigned long val = 0;
679 for_each_possible_cpu(cpu)
680 val += per_cpu(memcg->stat->events[idx], cpu);
684 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
689 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
690 * counted as CACHE even if it's on ANON LRU.
693 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
696 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
699 if (PageTransHuge(page))
700 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
703 /* pagein of a big page is an event. So, ignore page size */
705 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
707 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
708 nr_pages = -nr_pages; /* for event */
711 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
714 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
716 unsigned int lru_mask)
718 unsigned long nr = 0;
721 VM_BUG_ON((unsigned)nid >= nr_node_ids);
723 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
724 struct mem_cgroup_per_zone *mz;
728 if (!(BIT(lru) & lru_mask))
730 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
731 nr += mz->lru_size[lru];
737 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
738 unsigned int lru_mask)
740 unsigned long nr = 0;
743 for_each_node_state(nid, N_MEMORY)
744 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
748 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
749 enum mem_cgroup_events_target target)
751 unsigned long val, next;
753 val = __this_cpu_read(memcg->stat->nr_page_events);
754 next = __this_cpu_read(memcg->stat->targets[target]);
755 /* from time_after() in jiffies.h */
756 if ((long)next - (long)val < 0) {
758 case MEM_CGROUP_TARGET_THRESH:
759 next = val + THRESHOLDS_EVENTS_TARGET;
761 case MEM_CGROUP_TARGET_SOFTLIMIT:
762 next = val + SOFTLIMIT_EVENTS_TARGET;
764 case MEM_CGROUP_TARGET_NUMAINFO:
765 next = val + NUMAINFO_EVENTS_TARGET;
770 __this_cpu_write(memcg->stat->targets[target], next);
777 * Check events in order.
780 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
782 /* threshold event is triggered in finer grain than soft limit */
783 if (unlikely(mem_cgroup_event_ratelimit(memcg,
784 MEM_CGROUP_TARGET_THRESH))) {
786 bool do_numainfo __maybe_unused;
788 do_softlimit = mem_cgroup_event_ratelimit(memcg,
789 MEM_CGROUP_TARGET_SOFTLIMIT);
791 do_numainfo = mem_cgroup_event_ratelimit(memcg,
792 MEM_CGROUP_TARGET_NUMAINFO);
794 mem_cgroup_threshold(memcg);
795 if (unlikely(do_softlimit))
796 mem_cgroup_update_tree(memcg, page);
798 if (unlikely(do_numainfo))
799 atomic_inc(&memcg->numainfo_events);
804 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
807 * mm_update_next_owner() may clear mm->owner to NULL
808 * if it races with swapoff, page migration, etc.
809 * So this can be called with p == NULL.
814 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
816 EXPORT_SYMBOL(mem_cgroup_from_task);
818 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
820 struct mem_cgroup *memcg = NULL;
825 * Page cache insertions can happen withou an
826 * actual mm context, e.g. during disk probing
827 * on boot, loopback IO, acct() writes etc.
830 memcg = root_mem_cgroup;
832 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
833 if (unlikely(!memcg))
834 memcg = root_mem_cgroup;
836 } while (!css_tryget_online(&memcg->css));
842 * mem_cgroup_iter - iterate over memory cgroup hierarchy
843 * @root: hierarchy root
844 * @prev: previously returned memcg, NULL on first invocation
845 * @reclaim: cookie for shared reclaim walks, NULL for full walks
847 * Returns references to children of the hierarchy below @root, or
848 * @root itself, or %NULL after a full round-trip.
850 * Caller must pass the return value in @prev on subsequent
851 * invocations for reference counting, or use mem_cgroup_iter_break()
852 * to cancel a hierarchy walk before the round-trip is complete.
854 * Reclaimers can specify a zone and a priority level in @reclaim to
855 * divide up the memcgs in the hierarchy among all concurrent
856 * reclaimers operating on the same zone and priority.
858 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
859 struct mem_cgroup *prev,
860 struct mem_cgroup_reclaim_cookie *reclaim)
862 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
863 struct cgroup_subsys_state *css = NULL;
864 struct mem_cgroup *memcg = NULL;
865 struct mem_cgroup *pos = NULL;
867 if (mem_cgroup_disabled())
871 root = root_mem_cgroup;
873 if (prev && !reclaim)
876 if (!root->use_hierarchy && root != root_mem_cgroup) {
885 struct mem_cgroup_per_zone *mz;
887 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
888 iter = &mz->iter[reclaim->priority];
890 if (prev && reclaim->generation != iter->generation)
894 pos = READ_ONCE(iter->position);
895 if (!pos || css_tryget(&pos->css))
898 * css reference reached zero, so iter->position will
899 * be cleared by ->css_released. However, we should not
900 * rely on this happening soon, because ->css_released
901 * is called from a work queue, and by busy-waiting we
902 * might block it. So we clear iter->position right
905 (void)cmpxchg(&iter->position, pos, NULL);
913 css = css_next_descendant_pre(css, &root->css);
916 * Reclaimers share the hierarchy walk, and a
917 * new one might jump in right at the end of
918 * the hierarchy - make sure they see at least
919 * one group and restart from the beginning.
927 * Verify the css and acquire a reference. The root
928 * is provided by the caller, so we know it's alive
929 * and kicking, and don't take an extra reference.
931 memcg = mem_cgroup_from_css(css);
933 if (css == &root->css)
936 if (css_tryget(css)) {
938 * Make sure the memcg is initialized:
939 * mem_cgroup_css_online() orders the the
940 * initialization against setting the flag.
942 if (smp_load_acquire(&memcg->initialized))
953 * The position could have already been updated by a competing
954 * thread, so check that the value hasn't changed since we read
955 * it to avoid reclaiming from the same cgroup twice.
957 (void)cmpxchg(&iter->position, pos, memcg);
965 reclaim->generation = iter->generation;
971 if (prev && prev != root)
978 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
979 * @root: hierarchy root
980 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
982 void mem_cgroup_iter_break(struct mem_cgroup *root,
983 struct mem_cgroup *prev)
986 root = root_mem_cgroup;
987 if (prev && prev != root)
991 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
993 struct mem_cgroup *memcg = dead_memcg;
994 struct mem_cgroup_reclaim_iter *iter;
995 struct mem_cgroup_per_zone *mz;
999 while ((memcg = parent_mem_cgroup(memcg))) {
1000 for_each_node(nid) {
1001 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1002 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1003 for (i = 0; i <= DEF_PRIORITY; i++) {
1004 iter = &mz->iter[i];
1005 cmpxchg(&iter->position,
1014 * Iteration constructs for visiting all cgroups (under a tree). If
1015 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1016 * be used for reference counting.
1018 #define for_each_mem_cgroup_tree(iter, root) \
1019 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1021 iter = mem_cgroup_iter(root, iter, NULL))
1023 #define for_each_mem_cgroup(iter) \
1024 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1026 iter = mem_cgroup_iter(NULL, iter, NULL))
1029 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1030 * @zone: zone of the wanted lruvec
1031 * @memcg: memcg of the wanted lruvec
1033 * Returns the lru list vector holding pages for the given @zone and
1034 * @mem. This can be the global zone lruvec, if the memory controller
1037 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1038 struct mem_cgroup *memcg)
1040 struct mem_cgroup_per_zone *mz;
1041 struct lruvec *lruvec;
1043 if (mem_cgroup_disabled()) {
1044 lruvec = &zone->lruvec;
1048 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1049 lruvec = &mz->lruvec;
1052 * Since a node can be onlined after the mem_cgroup was created,
1053 * we have to be prepared to initialize lruvec->zone here;
1054 * and if offlined then reonlined, we need to reinitialize it.
1056 if (unlikely(lruvec->zone != zone))
1057 lruvec->zone = zone;
1062 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1064 * @zone: zone of the page
1066 * This function is only safe when following the LRU page isolation
1067 * and putback protocol: the LRU lock must be held, and the page must
1068 * either be PageLRU() or the caller must have isolated/allocated it.
1070 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1072 struct mem_cgroup_per_zone *mz;
1073 struct mem_cgroup *memcg;
1074 struct lruvec *lruvec;
1076 if (mem_cgroup_disabled()) {
1077 lruvec = &zone->lruvec;
1081 memcg = page->mem_cgroup;
1083 * Swapcache readahead pages are added to the LRU - and
1084 * possibly migrated - before they are charged.
1087 memcg = root_mem_cgroup;
1089 mz = mem_cgroup_page_zoneinfo(memcg, page);
1090 lruvec = &mz->lruvec;
1093 * Since a node can be onlined after the mem_cgroup was created,
1094 * we have to be prepared to initialize lruvec->zone here;
1095 * and if offlined then reonlined, we need to reinitialize it.
1097 if (unlikely(lruvec->zone != zone))
1098 lruvec->zone = zone;
1103 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1104 * @lruvec: mem_cgroup per zone lru vector
1105 * @lru: index of lru list the page is sitting on
1106 * @nr_pages: positive when adding or negative when removing
1108 * This function must be called when a page is added to or removed from an
1111 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1114 struct mem_cgroup_per_zone *mz;
1115 unsigned long *lru_size;
1117 if (mem_cgroup_disabled())
1120 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1121 lru_size = mz->lru_size + lru;
1122 *lru_size += nr_pages;
1123 VM_BUG_ON((long)(*lru_size) < 0);
1126 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1128 struct mem_cgroup *task_memcg;
1129 struct task_struct *p;
1132 p = find_lock_task_mm(task);
1134 task_memcg = get_mem_cgroup_from_mm(p->mm);
1138 * All threads may have already detached their mm's, but the oom
1139 * killer still needs to detect if they have already been oom
1140 * killed to prevent needlessly killing additional tasks.
1143 task_memcg = mem_cgroup_from_task(task);
1144 css_get(&task_memcg->css);
1147 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1148 css_put(&task_memcg->css);
1152 #define mem_cgroup_from_counter(counter, member) \
1153 container_of(counter, struct mem_cgroup, member)
1156 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1157 * @memcg: the memory cgroup
1159 * Returns the maximum amount of memory @mem can be charged with, in
1162 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1164 unsigned long margin = 0;
1165 unsigned long count;
1166 unsigned long limit;
1168 count = page_counter_read(&memcg->memory);
1169 limit = READ_ONCE(memcg->memory.limit);
1171 margin = limit - count;
1173 if (do_swap_account) {
1174 count = page_counter_read(&memcg->memsw);
1175 limit = READ_ONCE(memcg->memsw.limit);
1177 margin = min(margin, limit - count);
1184 * A routine for checking "mem" is under move_account() or not.
1186 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1187 * moving cgroups. This is for waiting at high-memory pressure
1190 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1192 struct mem_cgroup *from;
1193 struct mem_cgroup *to;
1196 * Unlike task_move routines, we access mc.to, mc.from not under
1197 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1199 spin_lock(&mc.lock);
1205 ret = mem_cgroup_is_descendant(from, memcg) ||
1206 mem_cgroup_is_descendant(to, memcg);
1208 spin_unlock(&mc.lock);
1212 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1214 if (mc.moving_task && current != mc.moving_task) {
1215 if (mem_cgroup_under_move(memcg)) {
1217 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1218 /* moving charge context might have finished. */
1221 finish_wait(&mc.waitq, &wait);
1228 #define K(x) ((x) << (PAGE_SHIFT-10))
1230 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1231 * @memcg: The memory cgroup that went over limit
1232 * @p: Task that is going to be killed
1234 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1237 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1239 /* oom_info_lock ensures that parallel ooms do not interleave */
1240 static DEFINE_MUTEX(oom_info_lock);
1241 struct mem_cgroup *iter;
1244 mutex_lock(&oom_info_lock);
1248 pr_info("Task in ");
1249 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1250 pr_cont(" killed as a result of limit of ");
1252 pr_info("Memory limit reached of cgroup ");
1255 pr_cont_cgroup_path(memcg->css.cgroup);
1260 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1261 K((u64)page_counter_read(&memcg->memory)),
1262 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1263 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1264 K((u64)page_counter_read(&memcg->memsw)),
1265 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1266 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1267 K((u64)page_counter_read(&memcg->kmem)),
1268 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1270 for_each_mem_cgroup_tree(iter, memcg) {
1271 pr_info("Memory cgroup stats for ");
1272 pr_cont_cgroup_path(iter->css.cgroup);
1275 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1276 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1278 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1279 K(mem_cgroup_read_stat(iter, i)));
1282 for (i = 0; i < NR_LRU_LISTS; i++)
1283 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1284 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1288 mutex_unlock(&oom_info_lock);
1292 * This function returns the number of memcg under hierarchy tree. Returns
1293 * 1(self count) if no children.
1295 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1298 struct mem_cgroup *iter;
1300 for_each_mem_cgroup_tree(iter, memcg)
1306 * Return the memory (and swap, if configured) limit for a memcg.
1308 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1310 unsigned long limit;
1312 limit = memcg->memory.limit;
1313 if (mem_cgroup_swappiness(memcg)) {
1314 unsigned long memsw_limit;
1316 memsw_limit = memcg->memsw.limit;
1317 limit = min(limit + total_swap_pages, memsw_limit);
1322 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1325 struct oom_control oc = {
1328 .gfp_mask = gfp_mask,
1331 struct mem_cgroup *iter;
1332 unsigned long chosen_points = 0;
1333 unsigned long totalpages;
1334 unsigned int points = 0;
1335 struct task_struct *chosen = NULL;
1337 mutex_lock(&oom_lock);
1340 * If current has a pending SIGKILL or is exiting, then automatically
1341 * select it. The goal is to allow it to allocate so that it may
1342 * quickly exit and free its memory.
1344 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1345 mark_oom_victim(current);
1349 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1350 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1351 for_each_mem_cgroup_tree(iter, memcg) {
1352 struct css_task_iter it;
1353 struct task_struct *task;
1355 css_task_iter_start(&iter->css, &it);
1356 while ((task = css_task_iter_next(&it))) {
1357 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1358 case OOM_SCAN_SELECT:
1360 put_task_struct(chosen);
1362 chosen_points = ULONG_MAX;
1363 get_task_struct(chosen);
1365 case OOM_SCAN_CONTINUE:
1367 case OOM_SCAN_ABORT:
1368 css_task_iter_end(&it);
1369 mem_cgroup_iter_break(memcg, iter);
1371 put_task_struct(chosen);
1376 points = oom_badness(task, memcg, NULL, totalpages);
1377 if (!points || points < chosen_points)
1379 /* Prefer thread group leaders for display purposes */
1380 if (points == chosen_points &&
1381 thread_group_leader(chosen))
1385 put_task_struct(chosen);
1387 chosen_points = points;
1388 get_task_struct(chosen);
1390 css_task_iter_end(&it);
1394 points = chosen_points * 1000 / totalpages;
1395 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1396 "Memory cgroup out of memory");
1399 mutex_unlock(&oom_lock);
1403 #if MAX_NUMNODES > 1
1406 * test_mem_cgroup_node_reclaimable
1407 * @memcg: the target memcg
1408 * @nid: the node ID to be checked.
1409 * @noswap : specify true here if the user wants flle only information.
1411 * This function returns whether the specified memcg contains any
1412 * reclaimable pages on a node. Returns true if there are any reclaimable
1413 * pages in the node.
1415 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1416 int nid, bool noswap)
1418 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1420 if (noswap || !total_swap_pages)
1422 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1429 * Always updating the nodemask is not very good - even if we have an empty
1430 * list or the wrong list here, we can start from some node and traverse all
1431 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1434 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1438 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1439 * pagein/pageout changes since the last update.
1441 if (!atomic_read(&memcg->numainfo_events))
1443 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1446 /* make a nodemask where this memcg uses memory from */
1447 memcg->scan_nodes = node_states[N_MEMORY];
1449 for_each_node_mask(nid, node_states[N_MEMORY]) {
1451 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1452 node_clear(nid, memcg->scan_nodes);
1455 atomic_set(&memcg->numainfo_events, 0);
1456 atomic_set(&memcg->numainfo_updating, 0);
1460 * Selecting a node where we start reclaim from. Because what we need is just
1461 * reducing usage counter, start from anywhere is O,K. Considering
1462 * memory reclaim from current node, there are pros. and cons.
1464 * Freeing memory from current node means freeing memory from a node which
1465 * we'll use or we've used. So, it may make LRU bad. And if several threads
1466 * hit limits, it will see a contention on a node. But freeing from remote
1467 * node means more costs for memory reclaim because of memory latency.
1469 * Now, we use round-robin. Better algorithm is welcomed.
1471 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1475 mem_cgroup_may_update_nodemask(memcg);
1476 node = memcg->last_scanned_node;
1478 node = next_node(node, memcg->scan_nodes);
1479 if (node == MAX_NUMNODES)
1480 node = first_node(memcg->scan_nodes);
1482 * We call this when we hit limit, not when pages are added to LRU.
1483 * No LRU may hold pages because all pages are UNEVICTABLE or
1484 * memcg is too small and all pages are not on LRU. In that case,
1485 * we use curret node.
1487 if (unlikely(node == MAX_NUMNODES))
1488 node = numa_node_id();
1490 memcg->last_scanned_node = node;
1494 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1500 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1503 unsigned long *total_scanned)
1505 struct mem_cgroup *victim = NULL;
1508 unsigned long excess;
1509 unsigned long nr_scanned;
1510 struct mem_cgroup_reclaim_cookie reclaim = {
1515 excess = soft_limit_excess(root_memcg);
1518 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1523 * If we have not been able to reclaim
1524 * anything, it might because there are
1525 * no reclaimable pages under this hierarchy
1530 * We want to do more targeted reclaim.
1531 * excess >> 2 is not to excessive so as to
1532 * reclaim too much, nor too less that we keep
1533 * coming back to reclaim from this cgroup
1535 if (total >= (excess >> 2) ||
1536 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1541 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1543 *total_scanned += nr_scanned;
1544 if (!soft_limit_excess(root_memcg))
1547 mem_cgroup_iter_break(root_memcg, victim);
1551 #ifdef CONFIG_LOCKDEP
1552 static struct lockdep_map memcg_oom_lock_dep_map = {
1553 .name = "memcg_oom_lock",
1557 static DEFINE_SPINLOCK(memcg_oom_lock);
1560 * Check OOM-Killer is already running under our hierarchy.
1561 * If someone is running, return false.
1563 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1565 struct mem_cgroup *iter, *failed = NULL;
1567 spin_lock(&memcg_oom_lock);
1569 for_each_mem_cgroup_tree(iter, memcg) {
1570 if (iter->oom_lock) {
1572 * this subtree of our hierarchy is already locked
1573 * so we cannot give a lock.
1576 mem_cgroup_iter_break(memcg, iter);
1579 iter->oom_lock = true;
1584 * OK, we failed to lock the whole subtree so we have
1585 * to clean up what we set up to the failing subtree
1587 for_each_mem_cgroup_tree(iter, memcg) {
1588 if (iter == failed) {
1589 mem_cgroup_iter_break(memcg, iter);
1592 iter->oom_lock = false;
1595 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1597 spin_unlock(&memcg_oom_lock);
1602 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1604 struct mem_cgroup *iter;
1606 spin_lock(&memcg_oom_lock);
1607 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1608 for_each_mem_cgroup_tree(iter, memcg)
1609 iter->oom_lock = false;
1610 spin_unlock(&memcg_oom_lock);
1613 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1615 struct mem_cgroup *iter;
1617 spin_lock(&memcg_oom_lock);
1618 for_each_mem_cgroup_tree(iter, memcg)
1620 spin_unlock(&memcg_oom_lock);
1623 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1625 struct mem_cgroup *iter;
1628 * When a new child is created while the hierarchy is under oom,
1629 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1631 spin_lock(&memcg_oom_lock);
1632 for_each_mem_cgroup_tree(iter, memcg)
1633 if (iter->under_oom > 0)
1635 spin_unlock(&memcg_oom_lock);
1638 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1640 struct oom_wait_info {
1641 struct mem_cgroup *memcg;
1645 static int memcg_oom_wake_function(wait_queue_t *wait,
1646 unsigned mode, int sync, void *arg)
1648 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1649 struct mem_cgroup *oom_wait_memcg;
1650 struct oom_wait_info *oom_wait_info;
1652 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1653 oom_wait_memcg = oom_wait_info->memcg;
1655 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1656 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1658 return autoremove_wake_function(wait, mode, sync, arg);
1661 static void memcg_oom_recover(struct mem_cgroup *memcg)
1664 * For the following lockless ->under_oom test, the only required
1665 * guarantee is that it must see the state asserted by an OOM when
1666 * this function is called as a result of userland actions
1667 * triggered by the notification of the OOM. This is trivially
1668 * achieved by invoking mem_cgroup_mark_under_oom() before
1669 * triggering notification.
1671 if (memcg && memcg->under_oom)
1672 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1675 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1677 if (!current->memcg_may_oom)
1680 * We are in the middle of the charge context here, so we
1681 * don't want to block when potentially sitting on a callstack
1682 * that holds all kinds of filesystem and mm locks.
1684 * Also, the caller may handle a failed allocation gracefully
1685 * (like optional page cache readahead) and so an OOM killer
1686 * invocation might not even be necessary.
1688 * That's why we don't do anything here except remember the
1689 * OOM context and then deal with it at the end of the page
1690 * fault when the stack is unwound, the locks are released,
1691 * and when we know whether the fault was overall successful.
1693 css_get(&memcg->css);
1694 current->memcg_in_oom = memcg;
1695 current->memcg_oom_gfp_mask = mask;
1696 current->memcg_oom_order = order;
1700 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1701 * @handle: actually kill/wait or just clean up the OOM state
1703 * This has to be called at the end of a page fault if the memcg OOM
1704 * handler was enabled.
1706 * Memcg supports userspace OOM handling where failed allocations must
1707 * sleep on a waitqueue until the userspace task resolves the
1708 * situation. Sleeping directly in the charge context with all kinds
1709 * of locks held is not a good idea, instead we remember an OOM state
1710 * in the task and mem_cgroup_oom_synchronize() has to be called at
1711 * the end of the page fault to complete the OOM handling.
1713 * Returns %true if an ongoing memcg OOM situation was detected and
1714 * completed, %false otherwise.
1716 bool mem_cgroup_oom_synchronize(bool handle)
1718 struct mem_cgroup *memcg = current->memcg_in_oom;
1719 struct oom_wait_info owait;
1722 /* OOM is global, do not handle */
1726 if (!handle || oom_killer_disabled)
1729 owait.memcg = memcg;
1730 owait.wait.flags = 0;
1731 owait.wait.func = memcg_oom_wake_function;
1732 owait.wait.private = current;
1733 INIT_LIST_HEAD(&owait.wait.task_list);
1735 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1736 mem_cgroup_mark_under_oom(memcg);
1738 locked = mem_cgroup_oom_trylock(memcg);
1741 mem_cgroup_oom_notify(memcg);
1743 if (locked && !memcg->oom_kill_disable) {
1744 mem_cgroup_unmark_under_oom(memcg);
1745 finish_wait(&memcg_oom_waitq, &owait.wait);
1746 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1747 current->memcg_oom_order);
1750 mem_cgroup_unmark_under_oom(memcg);
1751 finish_wait(&memcg_oom_waitq, &owait.wait);
1755 mem_cgroup_oom_unlock(memcg);
1757 * There is no guarantee that an OOM-lock contender
1758 * sees the wakeups triggered by the OOM kill
1759 * uncharges. Wake any sleepers explicitely.
1761 memcg_oom_recover(memcg);
1764 current->memcg_in_oom = NULL;
1765 css_put(&memcg->css);
1770 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1771 * @page: page that is going to change accounted state
1773 * This function must mark the beginning of an accounted page state
1774 * change to prevent double accounting when the page is concurrently
1775 * being moved to another memcg:
1777 * memcg = mem_cgroup_begin_page_stat(page);
1778 * if (TestClearPageState(page))
1779 * mem_cgroup_update_page_stat(memcg, state, -1);
1780 * mem_cgroup_end_page_stat(memcg);
1782 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1784 struct mem_cgroup *memcg;
1785 unsigned long flags;
1788 * The RCU lock is held throughout the transaction. The fast
1789 * path can get away without acquiring the memcg->move_lock
1790 * because page moving starts with an RCU grace period.
1792 * The RCU lock also protects the memcg from being freed when
1793 * the page state that is going to change is the only thing
1794 * preventing the page from being uncharged.
1795 * E.g. end-writeback clearing PageWriteback(), which allows
1796 * migration to go ahead and uncharge the page before the
1797 * account transaction might be complete.
1801 if (mem_cgroup_disabled())
1804 memcg = page->mem_cgroup;
1805 if (unlikely(!memcg))
1808 if (atomic_read(&memcg->moving_account) <= 0)
1811 spin_lock_irqsave(&memcg->move_lock, flags);
1812 if (memcg != page->mem_cgroup) {
1813 spin_unlock_irqrestore(&memcg->move_lock, flags);
1818 * When charge migration first begins, we can have locked and
1819 * unlocked page stat updates happening concurrently. Track
1820 * the task who has the lock for mem_cgroup_end_page_stat().
1822 memcg->move_lock_task = current;
1823 memcg->move_lock_flags = flags;
1827 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1830 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1831 * @memcg: the memcg that was accounted against
1833 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1835 if (memcg && memcg->move_lock_task == current) {
1836 unsigned long flags = memcg->move_lock_flags;
1838 memcg->move_lock_task = NULL;
1839 memcg->move_lock_flags = 0;
1841 spin_unlock_irqrestore(&memcg->move_lock, flags);
1846 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1849 * size of first charge trial. "32" comes from vmscan.c's magic value.
1850 * TODO: maybe necessary to use big numbers in big irons.
1852 #define CHARGE_BATCH 32U
1853 struct memcg_stock_pcp {
1854 struct mem_cgroup *cached; /* this never be root cgroup */
1855 unsigned int nr_pages;
1856 struct work_struct work;
1857 unsigned long flags;
1858 #define FLUSHING_CACHED_CHARGE 0
1860 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1861 static DEFINE_MUTEX(percpu_charge_mutex);
1864 * consume_stock: Try to consume stocked charge on this cpu.
1865 * @memcg: memcg to consume from.
1866 * @nr_pages: how many pages to charge.
1868 * The charges will only happen if @memcg matches the current cpu's memcg
1869 * stock, and at least @nr_pages are available in that stock. Failure to
1870 * service an allocation will refill the stock.
1872 * returns true if successful, false otherwise.
1874 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1876 struct memcg_stock_pcp *stock;
1879 if (nr_pages > CHARGE_BATCH)
1882 stock = &get_cpu_var(memcg_stock);
1883 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1884 stock->nr_pages -= nr_pages;
1887 put_cpu_var(memcg_stock);
1892 * Returns stocks cached in percpu and reset cached information.
1894 static void drain_stock(struct memcg_stock_pcp *stock)
1896 struct mem_cgroup *old = stock->cached;
1898 if (stock->nr_pages) {
1899 page_counter_uncharge(&old->memory, stock->nr_pages);
1900 if (do_swap_account)
1901 page_counter_uncharge(&old->memsw, stock->nr_pages);
1902 css_put_many(&old->css, stock->nr_pages);
1903 stock->nr_pages = 0;
1905 stock->cached = NULL;
1909 * This must be called under preempt disabled or must be called by
1910 * a thread which is pinned to local cpu.
1912 static void drain_local_stock(struct work_struct *dummy)
1914 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1916 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1920 * Cache charges(val) to local per_cpu area.
1921 * This will be consumed by consume_stock() function, later.
1923 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1925 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1927 if (stock->cached != memcg) { /* reset if necessary */
1929 stock->cached = memcg;
1931 stock->nr_pages += nr_pages;
1932 put_cpu_var(memcg_stock);
1936 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1937 * of the hierarchy under it.
1939 static void drain_all_stock(struct mem_cgroup *root_memcg)
1943 /* If someone's already draining, avoid adding running more workers. */
1944 if (!mutex_trylock(&percpu_charge_mutex))
1946 /* Notify other cpus that system-wide "drain" is running */
1949 for_each_online_cpu(cpu) {
1950 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1951 struct mem_cgroup *memcg;
1953 memcg = stock->cached;
1954 if (!memcg || !stock->nr_pages)
1956 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1958 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1960 drain_local_stock(&stock->work);
1962 schedule_work_on(cpu, &stock->work);
1967 mutex_unlock(&percpu_charge_mutex);
1970 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1971 unsigned long action,
1974 int cpu = (unsigned long)hcpu;
1975 struct memcg_stock_pcp *stock;
1977 if (action == CPU_ONLINE)
1980 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1983 stock = &per_cpu(memcg_stock, cpu);
1989 * Scheduled by try_charge() to be executed from the userland return path
1990 * and reclaims memory over the high limit.
1992 void mem_cgroup_handle_over_high(void)
1994 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1995 struct mem_cgroup *memcg, *pos;
1997 if (likely(!nr_pages))
2000 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2003 if (page_counter_read(&pos->memory) <= pos->high)
2005 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2006 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2007 } while ((pos = parent_mem_cgroup(pos)));
2009 css_put(&memcg->css);
2010 current->memcg_nr_pages_over_high = 0;
2013 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2014 unsigned int nr_pages)
2016 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2017 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2018 struct mem_cgroup *mem_over_limit;
2019 struct page_counter *counter;
2020 unsigned long nr_reclaimed;
2021 bool may_swap = true;
2022 bool drained = false;
2024 if (mem_cgroup_is_root(memcg))
2027 if (consume_stock(memcg, nr_pages))
2030 if (!do_swap_account ||
2031 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2032 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2034 if (do_swap_account)
2035 page_counter_uncharge(&memcg->memsw, batch);
2036 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2038 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2042 if (batch > nr_pages) {
2048 * Unlike in global OOM situations, memcg is not in a physical
2049 * memory shortage. Allow dying and OOM-killed tasks to
2050 * bypass the last charges so that they can exit quickly and
2051 * free their memory.
2053 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2054 fatal_signal_pending(current) ||
2055 current->flags & PF_EXITING))
2058 if (unlikely(task_in_memcg_oom(current)))
2061 if (!gfpflags_allow_blocking(gfp_mask))
2064 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2066 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2067 gfp_mask, may_swap);
2069 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2073 drain_all_stock(mem_over_limit);
2078 if (gfp_mask & __GFP_NORETRY)
2081 * Even though the limit is exceeded at this point, reclaim
2082 * may have been able to free some pages. Retry the charge
2083 * before killing the task.
2085 * Only for regular pages, though: huge pages are rather
2086 * unlikely to succeed so close to the limit, and we fall back
2087 * to regular pages anyway in case of failure.
2089 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2092 * At task move, charge accounts can be doubly counted. So, it's
2093 * better to wait until the end of task_move if something is going on.
2095 if (mem_cgroup_wait_acct_move(mem_over_limit))
2101 if (gfp_mask & __GFP_NOFAIL)
2104 if (fatal_signal_pending(current))
2107 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2109 mem_cgroup_oom(mem_over_limit, gfp_mask,
2110 get_order(nr_pages * PAGE_SIZE));
2112 if (!(gfp_mask & __GFP_NOFAIL))
2116 * The allocation either can't fail or will lead to more memory
2117 * being freed very soon. Allow memory usage go over the limit
2118 * temporarily by force charging it.
2120 page_counter_charge(&memcg->memory, nr_pages);
2121 if (do_swap_account)
2122 page_counter_charge(&memcg->memsw, nr_pages);
2123 css_get_many(&memcg->css, nr_pages);
2128 css_get_many(&memcg->css, batch);
2129 if (batch > nr_pages)
2130 refill_stock(memcg, batch - nr_pages);
2133 * If the hierarchy is above the normal consumption range, schedule
2134 * reclaim on returning to userland. We can perform reclaim here
2135 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2136 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2137 * not recorded as it most likely matches current's and won't
2138 * change in the meantime. As high limit is checked again before
2139 * reclaim, the cost of mismatch is negligible.
2142 if (page_counter_read(&memcg->memory) > memcg->high) {
2143 current->memcg_nr_pages_over_high += batch;
2144 set_notify_resume(current);
2147 } while ((memcg = parent_mem_cgroup(memcg)));
2152 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2154 if (mem_cgroup_is_root(memcg))
2157 page_counter_uncharge(&memcg->memory, nr_pages);
2158 if (do_swap_account)
2159 page_counter_uncharge(&memcg->memsw, nr_pages);
2161 css_put_many(&memcg->css, nr_pages);
2164 static void lock_page_lru(struct page *page, int *isolated)
2166 struct zone *zone = page_zone(page);
2168 spin_lock_irq(&zone->lru_lock);
2169 if (PageLRU(page)) {
2170 struct lruvec *lruvec;
2172 lruvec = mem_cgroup_page_lruvec(page, zone);
2174 del_page_from_lru_list(page, lruvec, page_lru(page));
2180 static void unlock_page_lru(struct page *page, int isolated)
2182 struct zone *zone = page_zone(page);
2185 struct lruvec *lruvec;
2187 lruvec = mem_cgroup_page_lruvec(page, zone);
2188 VM_BUG_ON_PAGE(PageLRU(page), page);
2190 add_page_to_lru_list(page, lruvec, page_lru(page));
2192 spin_unlock_irq(&zone->lru_lock);
2195 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2200 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2203 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2204 * may already be on some other mem_cgroup's LRU. Take care of it.
2207 lock_page_lru(page, &isolated);
2210 * Nobody should be changing or seriously looking at
2211 * page->mem_cgroup at this point:
2213 * - the page is uncharged
2215 * - the page is off-LRU
2217 * - an anonymous fault has exclusive page access, except for
2218 * a locked page table
2220 * - a page cache insertion, a swapin fault, or a migration
2221 * have the page locked
2223 page->mem_cgroup = memcg;
2226 unlock_page_lru(page, isolated);
2229 #ifdef CONFIG_MEMCG_KMEM
2230 static int memcg_alloc_cache_id(void)
2235 id = ida_simple_get(&memcg_cache_ida,
2236 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2240 if (id < memcg_nr_cache_ids)
2244 * There's no space for the new id in memcg_caches arrays,
2245 * so we have to grow them.
2247 down_write(&memcg_cache_ids_sem);
2249 size = 2 * (id + 1);
2250 if (size < MEMCG_CACHES_MIN_SIZE)
2251 size = MEMCG_CACHES_MIN_SIZE;
2252 else if (size > MEMCG_CACHES_MAX_SIZE)
2253 size = MEMCG_CACHES_MAX_SIZE;
2255 err = memcg_update_all_caches(size);
2257 err = memcg_update_all_list_lrus(size);
2259 memcg_nr_cache_ids = size;
2261 up_write(&memcg_cache_ids_sem);
2264 ida_simple_remove(&memcg_cache_ida, id);
2270 static void memcg_free_cache_id(int id)
2272 ida_simple_remove(&memcg_cache_ida, id);
2275 struct memcg_kmem_cache_create_work {
2276 struct mem_cgroup *memcg;
2277 struct kmem_cache *cachep;
2278 struct work_struct work;
2281 static void memcg_kmem_cache_create_func(struct work_struct *w)
2283 struct memcg_kmem_cache_create_work *cw =
2284 container_of(w, struct memcg_kmem_cache_create_work, work);
2285 struct mem_cgroup *memcg = cw->memcg;
2286 struct kmem_cache *cachep = cw->cachep;
2288 memcg_create_kmem_cache(memcg, cachep);
2290 css_put(&memcg->css);
2295 * Enqueue the creation of a per-memcg kmem_cache.
2297 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2298 struct kmem_cache *cachep)
2300 struct memcg_kmem_cache_create_work *cw;
2302 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2306 css_get(&memcg->css);
2309 cw->cachep = cachep;
2310 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2312 schedule_work(&cw->work);
2315 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2316 struct kmem_cache *cachep)
2319 * We need to stop accounting when we kmalloc, because if the
2320 * corresponding kmalloc cache is not yet created, the first allocation
2321 * in __memcg_schedule_kmem_cache_create will recurse.
2323 * However, it is better to enclose the whole function. Depending on
2324 * the debugging options enabled, INIT_WORK(), for instance, can
2325 * trigger an allocation. This too, will make us recurse. Because at
2326 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2327 * the safest choice is to do it like this, wrapping the whole function.
2329 current->memcg_kmem_skip_account = 1;
2330 __memcg_schedule_kmem_cache_create(memcg, cachep);
2331 current->memcg_kmem_skip_account = 0;
2335 * Return the kmem_cache we're supposed to use for a slab allocation.
2336 * We try to use the current memcg's version of the cache.
2338 * If the cache does not exist yet, if we are the first user of it,
2339 * we either create it immediately, if possible, or create it asynchronously
2341 * In the latter case, we will let the current allocation go through with
2342 * the original cache.
2344 * Can't be called in interrupt context or from kernel threads.
2345 * This function needs to be called with rcu_read_lock() held.
2347 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2349 struct mem_cgroup *memcg;
2350 struct kmem_cache *memcg_cachep;
2353 VM_BUG_ON(!is_root_cache(cachep));
2355 if (current->memcg_kmem_skip_account)
2358 memcg = get_mem_cgroup_from_mm(current->mm);
2359 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2363 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2364 if (likely(memcg_cachep))
2365 return memcg_cachep;
2368 * If we are in a safe context (can wait, and not in interrupt
2369 * context), we could be be predictable and return right away.
2370 * This would guarantee that the allocation being performed
2371 * already belongs in the new cache.
2373 * However, there are some clashes that can arrive from locking.
2374 * For instance, because we acquire the slab_mutex while doing
2375 * memcg_create_kmem_cache, this means no further allocation
2376 * could happen with the slab_mutex held. So it's better to
2379 memcg_schedule_kmem_cache_create(memcg, cachep);
2381 css_put(&memcg->css);
2385 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2387 if (!is_root_cache(cachep))
2388 css_put(&cachep->memcg_params.memcg->css);
2391 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2392 struct mem_cgroup *memcg)
2394 unsigned int nr_pages = 1 << order;
2395 struct page_counter *counter;
2398 if (!memcg_kmem_is_active(memcg))
2401 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2404 ret = try_charge(memcg, gfp, nr_pages);
2406 page_counter_uncharge(&memcg->kmem, nr_pages);
2410 page->mem_cgroup = memcg;
2415 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2417 struct mem_cgroup *memcg;
2420 memcg = get_mem_cgroup_from_mm(current->mm);
2421 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2422 css_put(&memcg->css);
2426 void __memcg_kmem_uncharge(struct page *page, int order)
2428 struct mem_cgroup *memcg = page->mem_cgroup;
2429 unsigned int nr_pages = 1 << order;
2434 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2436 page_counter_uncharge(&memcg->kmem, nr_pages);
2437 page_counter_uncharge(&memcg->memory, nr_pages);
2438 if (do_swap_account)
2439 page_counter_uncharge(&memcg->memsw, nr_pages);
2441 page->mem_cgroup = NULL;
2442 css_put_many(&memcg->css, nr_pages);
2444 #endif /* CONFIG_MEMCG_KMEM */
2446 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2449 * Because tail pages are not marked as "used", set it. We're under
2450 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2451 * charge/uncharge will be never happen and move_account() is done under
2452 * compound_lock(), so we don't have to take care of races.
2454 void mem_cgroup_split_huge_fixup(struct page *head)
2458 if (mem_cgroup_disabled())
2461 for (i = 1; i < HPAGE_PMD_NR; i++)
2462 head[i].mem_cgroup = head->mem_cgroup;
2464 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2467 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2469 #ifdef CONFIG_MEMCG_SWAP
2470 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2473 int val = (charge) ? 1 : -1;
2474 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2478 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2479 * @entry: swap entry to be moved
2480 * @from: mem_cgroup which the entry is moved from
2481 * @to: mem_cgroup which the entry is moved to
2483 * It succeeds only when the swap_cgroup's record for this entry is the same
2484 * as the mem_cgroup's id of @from.
2486 * Returns 0 on success, -EINVAL on failure.
2488 * The caller must have charged to @to, IOW, called page_counter_charge() about
2489 * both res and memsw, and called css_get().
2491 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2492 struct mem_cgroup *from, struct mem_cgroup *to)
2494 unsigned short old_id, new_id;
2496 old_id = mem_cgroup_id(from);
2497 new_id = mem_cgroup_id(to);
2499 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2500 mem_cgroup_swap_statistics(from, false);
2501 mem_cgroup_swap_statistics(to, true);
2507 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2508 struct mem_cgroup *from, struct mem_cgroup *to)
2514 static DEFINE_MUTEX(memcg_limit_mutex);
2516 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2517 unsigned long limit)
2519 unsigned long curusage;
2520 unsigned long oldusage;
2521 bool enlarge = false;
2526 * For keeping hierarchical_reclaim simple, how long we should retry
2527 * is depends on callers. We set our retry-count to be function
2528 * of # of children which we should visit in this loop.
2530 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2531 mem_cgroup_count_children(memcg);
2533 oldusage = page_counter_read(&memcg->memory);
2536 if (signal_pending(current)) {
2541 mutex_lock(&memcg_limit_mutex);
2542 if (limit > memcg->memsw.limit) {
2543 mutex_unlock(&memcg_limit_mutex);
2547 if (limit > memcg->memory.limit)
2549 ret = page_counter_limit(&memcg->memory, limit);
2550 mutex_unlock(&memcg_limit_mutex);
2555 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2557 curusage = page_counter_read(&memcg->memory);
2558 /* Usage is reduced ? */
2559 if (curusage >= oldusage)
2562 oldusage = curusage;
2563 } while (retry_count);
2565 if (!ret && enlarge)
2566 memcg_oom_recover(memcg);
2571 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2572 unsigned long limit)
2574 unsigned long curusage;
2575 unsigned long oldusage;
2576 bool enlarge = false;
2580 /* see mem_cgroup_resize_res_limit */
2581 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2582 mem_cgroup_count_children(memcg);
2584 oldusage = page_counter_read(&memcg->memsw);
2587 if (signal_pending(current)) {
2592 mutex_lock(&memcg_limit_mutex);
2593 if (limit < memcg->memory.limit) {
2594 mutex_unlock(&memcg_limit_mutex);
2598 if (limit > memcg->memsw.limit)
2600 ret = page_counter_limit(&memcg->memsw, limit);
2601 mutex_unlock(&memcg_limit_mutex);
2606 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2608 curusage = page_counter_read(&memcg->memsw);
2609 /* Usage is reduced ? */
2610 if (curusage >= oldusage)
2613 oldusage = curusage;
2614 } while (retry_count);
2616 if (!ret && enlarge)
2617 memcg_oom_recover(memcg);
2622 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2624 unsigned long *total_scanned)
2626 unsigned long nr_reclaimed = 0;
2627 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2628 unsigned long reclaimed;
2630 struct mem_cgroup_tree_per_zone *mctz;
2631 unsigned long excess;
2632 unsigned long nr_scanned;
2637 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2639 * This loop can run a while, specially if mem_cgroup's continuously
2640 * keep exceeding their soft limit and putting the system under
2647 mz = mem_cgroup_largest_soft_limit_node(mctz);
2652 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2653 gfp_mask, &nr_scanned);
2654 nr_reclaimed += reclaimed;
2655 *total_scanned += nr_scanned;
2656 spin_lock_irq(&mctz->lock);
2657 __mem_cgroup_remove_exceeded(mz, mctz);
2660 * If we failed to reclaim anything from this memory cgroup
2661 * it is time to move on to the next cgroup
2665 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2667 excess = soft_limit_excess(mz->memcg);
2669 * One school of thought says that we should not add
2670 * back the node to the tree if reclaim returns 0.
2671 * But our reclaim could return 0, simply because due
2672 * to priority we are exposing a smaller subset of
2673 * memory to reclaim from. Consider this as a longer
2676 /* If excess == 0, no tree ops */
2677 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2678 spin_unlock_irq(&mctz->lock);
2679 css_put(&mz->memcg->css);
2682 * Could not reclaim anything and there are no more
2683 * mem cgroups to try or we seem to be looping without
2684 * reclaiming anything.
2686 if (!nr_reclaimed &&
2688 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2690 } while (!nr_reclaimed);
2692 css_put(&next_mz->memcg->css);
2693 return nr_reclaimed;
2697 * Test whether @memcg has children, dead or alive. Note that this
2698 * function doesn't care whether @memcg has use_hierarchy enabled and
2699 * returns %true if there are child csses according to the cgroup
2700 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2702 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2707 * The lock does not prevent addition or deletion of children, but
2708 * it prevents a new child from being initialized based on this
2709 * parent in css_online(), so it's enough to decide whether
2710 * hierarchically inherited attributes can still be changed or not.
2712 lockdep_assert_held(&memcg_create_mutex);
2715 ret = css_next_child(NULL, &memcg->css);
2721 * Reclaims as many pages from the given memcg as possible and moves
2722 * the rest to the parent.
2724 * Caller is responsible for holding css reference for memcg.
2726 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2728 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2730 /* we call try-to-free pages for make this cgroup empty */
2731 lru_add_drain_all();
2732 /* try to free all pages in this cgroup */
2733 while (nr_retries && page_counter_read(&memcg->memory)) {
2736 if (signal_pending(current))
2739 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2743 /* maybe some writeback is necessary */
2744 congestion_wait(BLK_RW_ASYNC, HZ/10);
2752 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2753 char *buf, size_t nbytes,
2756 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2758 if (mem_cgroup_is_root(memcg))
2760 return mem_cgroup_force_empty(memcg) ?: nbytes;
2763 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2766 return mem_cgroup_from_css(css)->use_hierarchy;
2769 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2770 struct cftype *cft, u64 val)
2773 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2774 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2776 mutex_lock(&memcg_create_mutex);
2778 if (memcg->use_hierarchy == val)
2782 * If parent's use_hierarchy is set, we can't make any modifications
2783 * in the child subtrees. If it is unset, then the change can
2784 * occur, provided the current cgroup has no children.
2786 * For the root cgroup, parent_mem is NULL, we allow value to be
2787 * set if there are no children.
2789 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2790 (val == 1 || val == 0)) {
2791 if (!memcg_has_children(memcg))
2792 memcg->use_hierarchy = val;
2799 mutex_unlock(&memcg_create_mutex);
2804 static unsigned long tree_stat(struct mem_cgroup *memcg,
2805 enum mem_cgroup_stat_index idx)
2807 struct mem_cgroup *iter;
2808 unsigned long val = 0;
2810 for_each_mem_cgroup_tree(iter, memcg)
2811 val += mem_cgroup_read_stat(iter, idx);
2816 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2820 if (mem_cgroup_is_root(memcg)) {
2821 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2822 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2824 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2827 val = page_counter_read(&memcg->memory);
2829 val = page_counter_read(&memcg->memsw);
2842 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2845 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2846 struct page_counter *counter;
2848 switch (MEMFILE_TYPE(cft->private)) {
2850 counter = &memcg->memory;
2853 counter = &memcg->memsw;
2856 counter = &memcg->kmem;
2862 switch (MEMFILE_ATTR(cft->private)) {
2864 if (counter == &memcg->memory)
2865 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2866 if (counter == &memcg->memsw)
2867 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2868 return (u64)page_counter_read(counter) * PAGE_SIZE;
2870 return (u64)counter->limit * PAGE_SIZE;
2872 return (u64)counter->watermark * PAGE_SIZE;
2874 return counter->failcnt;
2875 case RES_SOFT_LIMIT:
2876 return (u64)memcg->soft_limit * PAGE_SIZE;
2882 #ifdef CONFIG_MEMCG_KMEM
2883 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2884 unsigned long nr_pages)
2889 BUG_ON(memcg->kmemcg_id >= 0);
2890 BUG_ON(memcg->kmem_acct_activated);
2891 BUG_ON(memcg->kmem_acct_active);
2894 * For simplicity, we won't allow this to be disabled. It also can't
2895 * be changed if the cgroup has children already, or if tasks had
2898 * If tasks join before we set the limit, a person looking at
2899 * kmem.usage_in_bytes will have no way to determine when it took
2900 * place, which makes the value quite meaningless.
2902 * After it first became limited, changes in the value of the limit are
2903 * of course permitted.
2905 mutex_lock(&memcg_create_mutex);
2906 if (cgroup_is_populated(memcg->css.cgroup) ||
2907 (memcg->use_hierarchy && memcg_has_children(memcg)))
2909 mutex_unlock(&memcg_create_mutex);
2913 memcg_id = memcg_alloc_cache_id();
2920 * We couldn't have accounted to this cgroup, because it hasn't got
2921 * activated yet, so this should succeed.
2923 err = page_counter_limit(&memcg->kmem, nr_pages);
2926 static_key_slow_inc(&memcg_kmem_enabled_key);
2928 * A memory cgroup is considered kmem-active as soon as it gets
2929 * kmemcg_id. Setting the id after enabling static branching will
2930 * guarantee no one starts accounting before all call sites are
2933 memcg->kmemcg_id = memcg_id;
2934 memcg->kmem_acct_activated = true;
2935 memcg->kmem_acct_active = true;
2940 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2941 unsigned long limit)
2945 mutex_lock(&memcg_limit_mutex);
2946 if (!memcg_kmem_is_active(memcg))
2947 ret = memcg_activate_kmem(memcg, limit);
2949 ret = page_counter_limit(&memcg->kmem, limit);
2950 mutex_unlock(&memcg_limit_mutex);
2954 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2957 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2962 mutex_lock(&memcg_limit_mutex);
2964 * If the parent cgroup is not kmem-active now, it cannot be activated
2965 * after this point, because it has at least one child already.
2967 if (memcg_kmem_is_active(parent))
2968 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2969 mutex_unlock(&memcg_limit_mutex);
2973 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2974 unsigned long limit)
2978 #endif /* CONFIG_MEMCG_KMEM */
2981 * The user of this function is...
2984 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2985 char *buf, size_t nbytes, loff_t off)
2987 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2988 unsigned long nr_pages;
2991 buf = strstrip(buf);
2992 ret = page_counter_memparse(buf, "-1", &nr_pages);
2996 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2998 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3002 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3004 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3007 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3010 ret = memcg_update_kmem_limit(memcg, nr_pages);
3014 case RES_SOFT_LIMIT:
3015 memcg->soft_limit = nr_pages;
3019 return ret ?: nbytes;
3022 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3023 size_t nbytes, loff_t off)
3025 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3026 struct page_counter *counter;
3028 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3030 counter = &memcg->memory;
3033 counter = &memcg->memsw;
3036 counter = &memcg->kmem;
3042 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3044 page_counter_reset_watermark(counter);
3047 counter->failcnt = 0;
3056 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3059 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3063 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3064 struct cftype *cft, u64 val)
3066 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3068 if (val & ~MOVE_MASK)
3072 * No kind of locking is needed in here, because ->can_attach() will
3073 * check this value once in the beginning of the process, and then carry
3074 * on with stale data. This means that changes to this value will only
3075 * affect task migrations starting after the change.
3077 memcg->move_charge_at_immigrate = val;
3081 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3082 struct cftype *cft, u64 val)
3089 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3093 unsigned int lru_mask;
3096 static const struct numa_stat stats[] = {
3097 { "total", LRU_ALL },
3098 { "file", LRU_ALL_FILE },
3099 { "anon", LRU_ALL_ANON },
3100 { "unevictable", BIT(LRU_UNEVICTABLE) },
3102 const struct numa_stat *stat;
3105 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3107 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3108 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3109 seq_printf(m, "%s=%lu", stat->name, nr);
3110 for_each_node_state(nid, N_MEMORY) {
3111 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3113 seq_printf(m, " N%d=%lu", nid, nr);
3118 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3119 struct mem_cgroup *iter;
3122 for_each_mem_cgroup_tree(iter, memcg)
3123 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3124 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3125 for_each_node_state(nid, N_MEMORY) {
3127 for_each_mem_cgroup_tree(iter, memcg)
3128 nr += mem_cgroup_node_nr_lru_pages(
3129 iter, nid, stat->lru_mask);
3130 seq_printf(m, " N%d=%lu", nid, nr);
3137 #endif /* CONFIG_NUMA */
3139 static int memcg_stat_show(struct seq_file *m, void *v)
3141 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3142 unsigned long memory, memsw;
3143 struct mem_cgroup *mi;
3146 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3147 MEM_CGROUP_STAT_NSTATS);
3148 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3149 MEM_CGROUP_EVENTS_NSTATS);
3150 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3152 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3153 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3155 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3156 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3159 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3160 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3161 mem_cgroup_read_events(memcg, i));
3163 for (i = 0; i < NR_LRU_LISTS; i++)
3164 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3165 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3167 /* Hierarchical information */
3168 memory = memsw = PAGE_COUNTER_MAX;
3169 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3170 memory = min(memory, mi->memory.limit);
3171 memsw = min(memsw, mi->memsw.limit);
3173 seq_printf(m, "hierarchical_memory_limit %llu\n",
3174 (u64)memory * PAGE_SIZE);
3175 if (do_swap_account)
3176 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3177 (u64)memsw * PAGE_SIZE);
3179 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3180 unsigned long long val = 0;
3182 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3184 for_each_mem_cgroup_tree(mi, memcg)
3185 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3186 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3189 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3190 unsigned long long val = 0;
3192 for_each_mem_cgroup_tree(mi, memcg)
3193 val += mem_cgroup_read_events(mi, i);
3194 seq_printf(m, "total_%s %llu\n",
3195 mem_cgroup_events_names[i], val);
3198 for (i = 0; i < NR_LRU_LISTS; i++) {
3199 unsigned long long val = 0;
3201 for_each_mem_cgroup_tree(mi, memcg)
3202 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3203 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3206 #ifdef CONFIG_DEBUG_VM
3209 struct mem_cgroup_per_zone *mz;
3210 struct zone_reclaim_stat *rstat;
3211 unsigned long recent_rotated[2] = {0, 0};
3212 unsigned long recent_scanned[2] = {0, 0};
3214 for_each_online_node(nid)
3215 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3216 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3217 rstat = &mz->lruvec.reclaim_stat;
3219 recent_rotated[0] += rstat->recent_rotated[0];
3220 recent_rotated[1] += rstat->recent_rotated[1];
3221 recent_scanned[0] += rstat->recent_scanned[0];
3222 recent_scanned[1] += rstat->recent_scanned[1];
3224 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3225 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3226 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3227 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3234 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3237 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3239 return mem_cgroup_swappiness(memcg);
3242 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3243 struct cftype *cft, u64 val)
3245 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3251 memcg->swappiness = val;
3253 vm_swappiness = val;
3258 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3260 struct mem_cgroup_threshold_ary *t;
3261 unsigned long usage;
3266 t = rcu_dereference(memcg->thresholds.primary);
3268 t = rcu_dereference(memcg->memsw_thresholds.primary);
3273 usage = mem_cgroup_usage(memcg, swap);
3276 * current_threshold points to threshold just below or equal to usage.
3277 * If it's not true, a threshold was crossed after last
3278 * call of __mem_cgroup_threshold().
3280 i = t->current_threshold;
3283 * Iterate backward over array of thresholds starting from
3284 * current_threshold and check if a threshold is crossed.
3285 * If none of thresholds below usage is crossed, we read
3286 * only one element of the array here.
3288 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3289 eventfd_signal(t->entries[i].eventfd, 1);
3291 /* i = current_threshold + 1 */
3295 * Iterate forward over array of thresholds starting from
3296 * current_threshold+1 and check if a threshold is crossed.
3297 * If none of thresholds above usage is crossed, we read
3298 * only one element of the array here.
3300 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3301 eventfd_signal(t->entries[i].eventfd, 1);
3303 /* Update current_threshold */
3304 t->current_threshold = i - 1;
3309 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3312 __mem_cgroup_threshold(memcg, false);
3313 if (do_swap_account)
3314 __mem_cgroup_threshold(memcg, true);
3316 memcg = parent_mem_cgroup(memcg);
3320 static int compare_thresholds(const void *a, const void *b)
3322 const struct mem_cgroup_threshold *_a = a;
3323 const struct mem_cgroup_threshold *_b = b;
3325 if (_a->threshold > _b->threshold)
3328 if (_a->threshold < _b->threshold)
3334 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3336 struct mem_cgroup_eventfd_list *ev;
3338 spin_lock(&memcg_oom_lock);
3340 list_for_each_entry(ev, &memcg->oom_notify, list)
3341 eventfd_signal(ev->eventfd, 1);
3343 spin_unlock(&memcg_oom_lock);
3347 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3349 struct mem_cgroup *iter;
3351 for_each_mem_cgroup_tree(iter, memcg)
3352 mem_cgroup_oom_notify_cb(iter);
3355 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3356 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3358 struct mem_cgroup_thresholds *thresholds;
3359 struct mem_cgroup_threshold_ary *new;
3360 unsigned long threshold;
3361 unsigned long usage;
3364 ret = page_counter_memparse(args, "-1", &threshold);
3368 mutex_lock(&memcg->thresholds_lock);
3371 thresholds = &memcg->thresholds;
3372 usage = mem_cgroup_usage(memcg, false);
3373 } else if (type == _MEMSWAP) {
3374 thresholds = &memcg->memsw_thresholds;
3375 usage = mem_cgroup_usage(memcg, true);
3379 /* Check if a threshold crossed before adding a new one */
3380 if (thresholds->primary)
3381 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3383 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3385 /* Allocate memory for new array of thresholds */
3386 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3394 /* Copy thresholds (if any) to new array */
3395 if (thresholds->primary) {
3396 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3397 sizeof(struct mem_cgroup_threshold));
3400 /* Add new threshold */
3401 new->entries[size - 1].eventfd = eventfd;
3402 new->entries[size - 1].threshold = threshold;
3404 /* Sort thresholds. Registering of new threshold isn't time-critical */
3405 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3406 compare_thresholds, NULL);
3408 /* Find current threshold */
3409 new->current_threshold = -1;
3410 for (i = 0; i < size; i++) {
3411 if (new->entries[i].threshold <= usage) {
3413 * new->current_threshold will not be used until
3414 * rcu_assign_pointer(), so it's safe to increment
3417 ++new->current_threshold;
3422 /* Free old spare buffer and save old primary buffer as spare */
3423 kfree(thresholds->spare);
3424 thresholds->spare = thresholds->primary;
3426 rcu_assign_pointer(thresholds->primary, new);
3428 /* To be sure that nobody uses thresholds */
3432 mutex_unlock(&memcg->thresholds_lock);
3437 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3438 struct eventfd_ctx *eventfd, const char *args)
3440 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3443 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3444 struct eventfd_ctx *eventfd, const char *args)
3446 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3449 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3450 struct eventfd_ctx *eventfd, enum res_type type)
3452 struct mem_cgroup_thresholds *thresholds;
3453 struct mem_cgroup_threshold_ary *new;
3454 unsigned long usage;
3457 mutex_lock(&memcg->thresholds_lock);
3460 thresholds = &memcg->thresholds;
3461 usage = mem_cgroup_usage(memcg, false);
3462 } else if (type == _MEMSWAP) {
3463 thresholds = &memcg->memsw_thresholds;
3464 usage = mem_cgroup_usage(memcg, true);
3468 if (!thresholds->primary)
3471 /* Check if a threshold crossed before removing */
3472 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3474 /* Calculate new number of threshold */
3476 for (i = 0; i < thresholds->primary->size; i++) {
3477 if (thresholds->primary->entries[i].eventfd != eventfd)
3481 new = thresholds->spare;
3483 /* Set thresholds array to NULL if we don't have thresholds */
3492 /* Copy thresholds and find current threshold */
3493 new->current_threshold = -1;
3494 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3495 if (thresholds->primary->entries[i].eventfd == eventfd)
3498 new->entries[j] = thresholds->primary->entries[i];
3499 if (new->entries[j].threshold <= usage) {
3501 * new->current_threshold will not be used
3502 * until rcu_assign_pointer(), so it's safe to increment
3505 ++new->current_threshold;
3511 /* Swap primary and spare array */
3512 thresholds->spare = thresholds->primary;
3514 rcu_assign_pointer(thresholds->primary, new);
3516 /* To be sure that nobody uses thresholds */
3519 /* If all events are unregistered, free the spare array */
3521 kfree(thresholds->spare);
3522 thresholds->spare = NULL;
3525 mutex_unlock(&memcg->thresholds_lock);
3528 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3529 struct eventfd_ctx *eventfd)
3531 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3534 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3535 struct eventfd_ctx *eventfd)
3537 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3540 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3541 struct eventfd_ctx *eventfd, const char *args)
3543 struct mem_cgroup_eventfd_list *event;
3545 event = kmalloc(sizeof(*event), GFP_KERNEL);
3549 spin_lock(&memcg_oom_lock);
3551 event->eventfd = eventfd;
3552 list_add(&event->list, &memcg->oom_notify);
3554 /* already in OOM ? */
3555 if (memcg->under_oom)
3556 eventfd_signal(eventfd, 1);
3557 spin_unlock(&memcg_oom_lock);
3562 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3563 struct eventfd_ctx *eventfd)
3565 struct mem_cgroup_eventfd_list *ev, *tmp;
3567 spin_lock(&memcg_oom_lock);
3569 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3570 if (ev->eventfd == eventfd) {
3571 list_del(&ev->list);
3576 spin_unlock(&memcg_oom_lock);
3579 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3581 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3583 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3584 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3588 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3589 struct cftype *cft, u64 val)
3591 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3593 /* cannot set to root cgroup and only 0 and 1 are allowed */
3594 if (!css->parent || !((val == 0) || (val == 1)))
3597 memcg->oom_kill_disable = val;
3599 memcg_oom_recover(memcg);
3604 #ifdef CONFIG_MEMCG_KMEM
3605 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3609 ret = memcg_propagate_kmem(memcg);
3613 return mem_cgroup_sockets_init(memcg, ss);
3616 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3618 struct cgroup_subsys_state *css;
3619 struct mem_cgroup *parent, *child;
3622 if (!memcg->kmem_acct_active)
3626 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3627 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3628 * guarantees no cache will be created for this cgroup after we are
3629 * done (see memcg_create_kmem_cache()).
3631 memcg->kmem_acct_active = false;
3633 memcg_deactivate_kmem_caches(memcg);
3635 kmemcg_id = memcg->kmemcg_id;
3636 BUG_ON(kmemcg_id < 0);
3638 parent = parent_mem_cgroup(memcg);
3640 parent = root_mem_cgroup;
3643 * Change kmemcg_id of this cgroup and all its descendants to the
3644 * parent's id, and then move all entries from this cgroup's list_lrus
3645 * to ones of the parent. After we have finished, all list_lrus
3646 * corresponding to this cgroup are guaranteed to remain empty. The
3647 * ordering is imposed by list_lru_node->lock taken by
3648 * memcg_drain_all_list_lrus().
3650 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3651 css_for_each_descendant_pre(css, &memcg->css) {
3652 child = mem_cgroup_from_css(css);
3653 BUG_ON(child->kmemcg_id != kmemcg_id);
3654 child->kmemcg_id = parent->kmemcg_id;
3655 if (!memcg->use_hierarchy)
3660 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3662 memcg_free_cache_id(kmemcg_id);
3665 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3667 if (memcg->kmem_acct_activated) {
3668 memcg_destroy_kmem_caches(memcg);
3669 static_key_slow_dec(&memcg_kmem_enabled_key);
3670 WARN_ON(page_counter_read(&memcg->kmem));
3672 mem_cgroup_sockets_destroy(memcg);
3675 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3680 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3684 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3689 #ifdef CONFIG_CGROUP_WRITEBACK
3691 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3693 return &memcg->cgwb_list;
3696 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3698 return wb_domain_init(&memcg->cgwb_domain, gfp);
3701 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3703 wb_domain_exit(&memcg->cgwb_domain);
3706 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3708 wb_domain_size_changed(&memcg->cgwb_domain);
3711 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3713 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3715 if (!memcg->css.parent)
3718 return &memcg->cgwb_domain;
3722 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3723 * @wb: bdi_writeback in question
3724 * @pfilepages: out parameter for number of file pages
3725 * @pheadroom: out parameter for number of allocatable pages according to memcg
3726 * @pdirty: out parameter for number of dirty pages
3727 * @pwriteback: out parameter for number of pages under writeback
3729 * Determine the numbers of file, headroom, dirty, and writeback pages in
3730 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3731 * is a bit more involved.
3733 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3734 * headroom is calculated as the lowest headroom of itself and the
3735 * ancestors. Note that this doesn't consider the actual amount of
3736 * available memory in the system. The caller should further cap
3737 * *@pheadroom accordingly.
3739 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3740 unsigned long *pheadroom, unsigned long *pdirty,
3741 unsigned long *pwriteback)
3743 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3744 struct mem_cgroup *parent;
3746 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3748 /* this should eventually include NR_UNSTABLE_NFS */
3749 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3750 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3751 (1 << LRU_ACTIVE_FILE));
3752 *pheadroom = PAGE_COUNTER_MAX;
3754 while ((parent = parent_mem_cgroup(memcg))) {
3755 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3756 unsigned long used = page_counter_read(&memcg->memory);
3758 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3763 #else /* CONFIG_CGROUP_WRITEBACK */
3765 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3770 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3774 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3778 #endif /* CONFIG_CGROUP_WRITEBACK */
3781 * DO NOT USE IN NEW FILES.
3783 * "cgroup.event_control" implementation.
3785 * This is way over-engineered. It tries to support fully configurable
3786 * events for each user. Such level of flexibility is completely
3787 * unnecessary especially in the light of the planned unified hierarchy.
3789 * Please deprecate this and replace with something simpler if at all
3794 * Unregister event and free resources.
3796 * Gets called from workqueue.
3798 static void memcg_event_remove(struct work_struct *work)
3800 struct mem_cgroup_event *event =
3801 container_of(work, struct mem_cgroup_event, remove);
3802 struct mem_cgroup *memcg = event->memcg;
3804 remove_wait_queue(event->wqh, &event->wait);
3806 event->unregister_event(memcg, event->eventfd);
3808 /* Notify userspace the event is going away. */
3809 eventfd_signal(event->eventfd, 1);
3811 eventfd_ctx_put(event->eventfd);
3813 css_put(&memcg->css);
3817 * Gets called on POLLHUP on eventfd when user closes it.
3819 * Called with wqh->lock held and interrupts disabled.
3821 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3822 int sync, void *key)
3824 struct mem_cgroup_event *event =
3825 container_of(wait, struct mem_cgroup_event, wait);
3826 struct mem_cgroup *memcg = event->memcg;
3827 unsigned long flags = (unsigned long)key;
3829 if (flags & POLLHUP) {
3831 * If the event has been detached at cgroup removal, we
3832 * can simply return knowing the other side will cleanup
3835 * We can't race against event freeing since the other
3836 * side will require wqh->lock via remove_wait_queue(),
3839 spin_lock(&memcg->event_list_lock);
3840 if (!list_empty(&event->list)) {
3841 list_del_init(&event->list);
3843 * We are in atomic context, but cgroup_event_remove()
3844 * may sleep, so we have to call it in workqueue.
3846 schedule_work(&event->remove);
3848 spin_unlock(&memcg->event_list_lock);
3854 static void memcg_event_ptable_queue_proc(struct file *file,
3855 wait_queue_head_t *wqh, poll_table *pt)
3857 struct mem_cgroup_event *event =
3858 container_of(pt, struct mem_cgroup_event, pt);
3861 add_wait_queue(wqh, &event->wait);
3865 * DO NOT USE IN NEW FILES.
3867 * Parse input and register new cgroup event handler.
3869 * Input must be in format '<event_fd> <control_fd> <args>'.
3870 * Interpretation of args is defined by control file implementation.
3872 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3873 char *buf, size_t nbytes, loff_t off)
3875 struct cgroup_subsys_state *css = of_css(of);
3876 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3877 struct mem_cgroup_event *event;
3878 struct cgroup_subsys_state *cfile_css;
3879 unsigned int efd, cfd;
3886 buf = strstrip(buf);
3888 efd = simple_strtoul(buf, &endp, 10);
3893 cfd = simple_strtoul(buf, &endp, 10);
3894 if ((*endp != ' ') && (*endp != '\0'))
3898 event = kzalloc(sizeof(*event), GFP_KERNEL);
3902 event->memcg = memcg;
3903 INIT_LIST_HEAD(&event->list);
3904 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3905 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3906 INIT_WORK(&event->remove, memcg_event_remove);
3914 event->eventfd = eventfd_ctx_fileget(efile.file);
3915 if (IS_ERR(event->eventfd)) {
3916 ret = PTR_ERR(event->eventfd);
3923 goto out_put_eventfd;
3926 /* the process need read permission on control file */
3927 /* AV: shouldn't we check that it's been opened for read instead? */
3928 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3933 * Determine the event callbacks and set them in @event. This used
3934 * to be done via struct cftype but cgroup core no longer knows
3935 * about these events. The following is crude but the whole thing
3936 * is for compatibility anyway.
3938 * DO NOT ADD NEW FILES.
3940 name = cfile.file->f_path.dentry->d_name.name;
3942 if (!strcmp(name, "memory.usage_in_bytes")) {
3943 event->register_event = mem_cgroup_usage_register_event;
3944 event->unregister_event = mem_cgroup_usage_unregister_event;
3945 } else if (!strcmp(name, "memory.oom_control")) {
3946 event->register_event = mem_cgroup_oom_register_event;
3947 event->unregister_event = mem_cgroup_oom_unregister_event;
3948 } else if (!strcmp(name, "memory.pressure_level")) {
3949 event->register_event = vmpressure_register_event;
3950 event->unregister_event = vmpressure_unregister_event;
3951 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3952 event->register_event = memsw_cgroup_usage_register_event;
3953 event->unregister_event = memsw_cgroup_usage_unregister_event;
3960 * Verify @cfile should belong to @css. Also, remaining events are
3961 * automatically removed on cgroup destruction but the removal is
3962 * asynchronous, so take an extra ref on @css.
3964 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3965 &memory_cgrp_subsys);
3967 if (IS_ERR(cfile_css))
3969 if (cfile_css != css) {
3974 ret = event->register_event(memcg, event->eventfd, buf);
3978 efile.file->f_op->poll(efile.file, &event->pt);
3980 spin_lock(&memcg->event_list_lock);
3981 list_add(&event->list, &memcg->event_list);
3982 spin_unlock(&memcg->event_list_lock);
3994 eventfd_ctx_put(event->eventfd);
4003 static struct cftype mem_cgroup_legacy_files[] = {
4005 .name = "usage_in_bytes",
4006 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4007 .read_u64 = mem_cgroup_read_u64,
4010 .name = "max_usage_in_bytes",
4011 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4012 .write = mem_cgroup_reset,
4013 .read_u64 = mem_cgroup_read_u64,
4016 .name = "limit_in_bytes",
4017 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4018 .write = mem_cgroup_write,
4019 .read_u64 = mem_cgroup_read_u64,
4022 .name = "soft_limit_in_bytes",
4023 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4024 .write = mem_cgroup_write,
4025 .read_u64 = mem_cgroup_read_u64,
4029 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4030 .write = mem_cgroup_reset,
4031 .read_u64 = mem_cgroup_read_u64,
4035 .seq_show = memcg_stat_show,
4038 .name = "force_empty",
4039 .write = mem_cgroup_force_empty_write,
4042 .name = "use_hierarchy",
4043 .write_u64 = mem_cgroup_hierarchy_write,
4044 .read_u64 = mem_cgroup_hierarchy_read,
4047 .name = "cgroup.event_control", /* XXX: for compat */
4048 .write = memcg_write_event_control,
4049 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4052 .name = "swappiness",
4053 .read_u64 = mem_cgroup_swappiness_read,
4054 .write_u64 = mem_cgroup_swappiness_write,
4057 .name = "move_charge_at_immigrate",
4058 .read_u64 = mem_cgroup_move_charge_read,
4059 .write_u64 = mem_cgroup_move_charge_write,
4062 .name = "oom_control",
4063 .seq_show = mem_cgroup_oom_control_read,
4064 .write_u64 = mem_cgroup_oom_control_write,
4065 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4068 .name = "pressure_level",
4072 .name = "numa_stat",
4073 .seq_show = memcg_numa_stat_show,
4076 #ifdef CONFIG_MEMCG_KMEM
4078 .name = "kmem.limit_in_bytes",
4079 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4080 .write = mem_cgroup_write,
4081 .read_u64 = mem_cgroup_read_u64,
4084 .name = "kmem.usage_in_bytes",
4085 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4086 .read_u64 = mem_cgroup_read_u64,
4089 .name = "kmem.failcnt",
4090 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4091 .write = mem_cgroup_reset,
4092 .read_u64 = mem_cgroup_read_u64,
4095 .name = "kmem.max_usage_in_bytes",
4096 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4097 .write = mem_cgroup_reset,
4098 .read_u64 = mem_cgroup_read_u64,
4100 #ifdef CONFIG_SLABINFO
4102 .name = "kmem.slabinfo",
4103 .seq_start = slab_start,
4104 .seq_next = slab_next,
4105 .seq_stop = slab_stop,
4106 .seq_show = memcg_slab_show,
4110 { }, /* terminate */
4114 * Private memory cgroup IDR
4116 * Swap-out records and page cache shadow entries need to store memcg
4117 * references in constrained space, so we maintain an ID space that is
4118 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4119 * memory-controlled cgroups to 64k.
4121 * However, there usually are many references to the oflline CSS after
4122 * the cgroup has been destroyed, such as page cache or reclaimable
4123 * slab objects, that don't need to hang on to the ID. We want to keep
4124 * those dead CSS from occupying IDs, or we might quickly exhaust the
4125 * relatively small ID space and prevent the creation of new cgroups
4126 * even when there are much fewer than 64k cgroups - possibly none.
4128 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4129 * be freed and recycled when it's no longer needed, which is usually
4130 * when the CSS is offlined.
4132 * The only exception to that are records of swapped out tmpfs/shmem
4133 * pages that need to be attributed to live ancestors on swapin. But
4134 * those references are manageable from userspace.
4137 static DEFINE_IDR(mem_cgroup_idr);
4139 static void mem_cgroup_id_get(struct mem_cgroup *memcg)
4141 atomic_inc(&memcg->id.ref);
4144 static void mem_cgroup_id_put(struct mem_cgroup *memcg)
4146 if (atomic_dec_and_test(&memcg->id.ref)) {
4147 idr_remove(&mem_cgroup_idr, memcg->id.id);
4150 /* Memcg ID pins CSS */
4151 css_put(&memcg->css);
4156 * mem_cgroup_from_id - look up a memcg from a memcg id
4157 * @id: the memcg id to look up
4159 * Caller must hold rcu_read_lock().
4161 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4163 WARN_ON_ONCE(!rcu_read_lock_held());
4164 return idr_find(&mem_cgroup_idr, id);
4167 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4169 struct mem_cgroup_per_node *pn;
4170 struct mem_cgroup_per_zone *mz;
4171 int zone, tmp = node;
4173 * This routine is called against possible nodes.
4174 * But it's BUG to call kmalloc() against offline node.
4176 * TODO: this routine can waste much memory for nodes which will
4177 * never be onlined. It's better to use memory hotplug callback
4180 if (!node_state(node, N_NORMAL_MEMORY))
4182 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4186 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4187 mz = &pn->zoneinfo[zone];
4188 lruvec_init(&mz->lruvec);
4189 mz->usage_in_excess = 0;
4190 mz->on_tree = false;
4193 memcg->nodeinfo[node] = pn;
4197 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4199 kfree(memcg->nodeinfo[node]);
4202 static struct mem_cgroup *mem_cgroup_alloc(void)
4204 struct mem_cgroup *memcg;
4207 size = sizeof(struct mem_cgroup);
4208 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4210 memcg = kzalloc(size, GFP_KERNEL);
4214 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4218 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4221 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4222 1, MEM_CGROUP_ID_MAX,
4224 if (memcg->id.id < 0)
4230 free_percpu(memcg->stat);
4237 * At destroying mem_cgroup, references from swap_cgroup can remain.
4238 * (scanning all at force_empty is too costly...)
4240 * Instead of clearing all references at force_empty, we remember
4241 * the number of reference from swap_cgroup and free mem_cgroup when
4242 * it goes down to 0.
4244 * Removal of cgroup itself succeeds regardless of refs from swap.
4247 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4251 mem_cgroup_remove_from_trees(memcg);
4254 free_mem_cgroup_per_zone_info(memcg, node);
4256 free_percpu(memcg->stat);
4257 memcg_wb_domain_exit(memcg);
4262 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4264 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4266 if (!memcg->memory.parent)
4268 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4270 EXPORT_SYMBOL(parent_mem_cgroup);
4272 static struct cgroup_subsys_state * __ref
4273 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4275 struct mem_cgroup *memcg;
4276 long error = -ENOMEM;
4279 memcg = mem_cgroup_alloc();
4281 return ERR_PTR(error);
4284 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4288 if (parent_css == NULL) {
4289 root_mem_cgroup = memcg;
4290 mem_cgroup_root_css = &memcg->css;
4291 page_counter_init(&memcg->memory, NULL);
4292 memcg->high = PAGE_COUNTER_MAX;
4293 memcg->soft_limit = PAGE_COUNTER_MAX;
4294 page_counter_init(&memcg->memsw, NULL);
4295 page_counter_init(&memcg->kmem, NULL);
4298 memcg->last_scanned_node = MAX_NUMNODES;
4299 INIT_LIST_HEAD(&memcg->oom_notify);
4300 memcg->move_charge_at_immigrate = 0;
4301 mutex_init(&memcg->thresholds_lock);
4302 spin_lock_init(&memcg->move_lock);
4303 vmpressure_init(&memcg->vmpressure);
4304 INIT_LIST_HEAD(&memcg->event_list);
4305 spin_lock_init(&memcg->event_list_lock);
4306 #ifdef CONFIG_MEMCG_KMEM
4307 memcg->kmemcg_id = -1;
4309 #ifdef CONFIG_CGROUP_WRITEBACK
4310 INIT_LIST_HEAD(&memcg->cgwb_list);
4312 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4316 idr_remove(&mem_cgroup_idr, memcg->id.id);
4317 __mem_cgroup_free(memcg);
4318 return ERR_PTR(error);
4322 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4324 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4325 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4328 /* Online state pins memcg ID, memcg ID pins CSS */
4329 mem_cgroup_id_get(mem_cgroup_from_css(css));
4335 mutex_lock(&memcg_create_mutex);
4337 memcg->use_hierarchy = parent->use_hierarchy;
4338 memcg->oom_kill_disable = parent->oom_kill_disable;
4339 memcg->swappiness = mem_cgroup_swappiness(parent);
4341 if (parent->use_hierarchy) {
4342 page_counter_init(&memcg->memory, &parent->memory);
4343 memcg->high = PAGE_COUNTER_MAX;
4344 memcg->soft_limit = PAGE_COUNTER_MAX;
4345 page_counter_init(&memcg->memsw, &parent->memsw);
4346 page_counter_init(&memcg->kmem, &parent->kmem);
4349 * No need to take a reference to the parent because cgroup
4350 * core guarantees its existence.
4353 page_counter_init(&memcg->memory, NULL);
4354 memcg->high = PAGE_COUNTER_MAX;
4355 memcg->soft_limit = PAGE_COUNTER_MAX;
4356 page_counter_init(&memcg->memsw, NULL);
4357 page_counter_init(&memcg->kmem, NULL);
4359 * Deeper hierachy with use_hierarchy == false doesn't make
4360 * much sense so let cgroup subsystem know about this
4361 * unfortunate state in our controller.
4363 if (parent != root_mem_cgroup)
4364 memory_cgrp_subsys.broken_hierarchy = true;
4366 mutex_unlock(&memcg_create_mutex);
4368 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4373 * Make sure the memcg is initialized: mem_cgroup_iter()
4374 * orders reading memcg->initialized against its callers
4375 * reading the memcg members.
4377 smp_store_release(&memcg->initialized, 1);
4382 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4384 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4385 struct mem_cgroup_event *event, *tmp;
4388 * Unregister events and notify userspace.
4389 * Notify userspace about cgroup removing only after rmdir of cgroup
4390 * directory to avoid race between userspace and kernelspace.
4392 spin_lock(&memcg->event_list_lock);
4393 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4394 list_del_init(&event->list);
4395 schedule_work(&event->remove);
4397 spin_unlock(&memcg->event_list_lock);
4399 vmpressure_cleanup(&memcg->vmpressure);
4401 memcg_deactivate_kmem(memcg);
4403 wb_memcg_offline(memcg);
4405 mem_cgroup_id_put(memcg);
4408 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4410 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4412 invalidate_reclaim_iterators(memcg);
4415 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4417 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4419 memcg_destroy_kmem(memcg);
4420 __mem_cgroup_free(memcg);
4424 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4425 * @css: the target css
4427 * Reset the states of the mem_cgroup associated with @css. This is
4428 * invoked when the userland requests disabling on the default hierarchy
4429 * but the memcg is pinned through dependency. The memcg should stop
4430 * applying policies and should revert to the vanilla state as it may be
4431 * made visible again.
4433 * The current implementation only resets the essential configurations.
4434 * This needs to be expanded to cover all the visible parts.
4436 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4438 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4440 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4441 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4442 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4444 memcg->high = PAGE_COUNTER_MAX;
4445 memcg->soft_limit = PAGE_COUNTER_MAX;
4446 memcg_wb_domain_size_changed(memcg);
4450 /* Handlers for move charge at task migration. */
4451 static int mem_cgroup_do_precharge(unsigned long count)
4455 /* Try a single bulk charge without reclaim first, kswapd may wake */
4456 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4458 mc.precharge += count;
4462 /* Try charges one by one with reclaim */
4464 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4474 * get_mctgt_type - get target type of moving charge
4475 * @vma: the vma the pte to be checked belongs
4476 * @addr: the address corresponding to the pte to be checked
4477 * @ptent: the pte to be checked
4478 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4481 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4482 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4483 * move charge. if @target is not NULL, the page is stored in target->page
4484 * with extra refcnt got(Callers should handle it).
4485 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4486 * target for charge migration. if @target is not NULL, the entry is stored
4489 * Called with pte lock held.
4496 enum mc_target_type {
4502 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4503 unsigned long addr, pte_t ptent)
4505 struct page *page = vm_normal_page(vma, addr, ptent);
4507 if (!page || !page_mapped(page))
4509 if (PageAnon(page)) {
4510 if (!(mc.flags & MOVE_ANON))
4513 if (!(mc.flags & MOVE_FILE))
4516 if (!get_page_unless_zero(page))
4523 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4524 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4526 struct page *page = NULL;
4527 swp_entry_t ent = pte_to_swp_entry(ptent);
4529 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4532 * Because lookup_swap_cache() updates some statistics counter,
4533 * we call find_get_page() with swapper_space directly.
4535 page = find_get_page(swap_address_space(ent), ent.val);
4536 if (do_swap_account)
4537 entry->val = ent.val;
4542 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4543 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4549 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4550 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4552 struct page *page = NULL;
4553 struct address_space *mapping;
4556 if (!vma->vm_file) /* anonymous vma */
4558 if (!(mc.flags & MOVE_FILE))
4561 mapping = vma->vm_file->f_mapping;
4562 pgoff = linear_page_index(vma, addr);
4564 /* page is moved even if it's not RSS of this task(page-faulted). */
4566 /* shmem/tmpfs may report page out on swap: account for that too. */
4567 if (shmem_mapping(mapping)) {
4568 page = find_get_entry(mapping, pgoff);
4569 if (radix_tree_exceptional_entry(page)) {
4570 swp_entry_t swp = radix_to_swp_entry(page);
4571 if (do_swap_account)
4573 page = find_get_page(swap_address_space(swp), swp.val);
4576 page = find_get_page(mapping, pgoff);
4578 page = find_get_page(mapping, pgoff);
4584 * mem_cgroup_move_account - move account of the page
4586 * @nr_pages: number of regular pages (>1 for huge pages)
4587 * @from: mem_cgroup which the page is moved from.
4588 * @to: mem_cgroup which the page is moved to. @from != @to.
4590 * The caller must confirm following.
4591 * - page is not on LRU (isolate_page() is useful.)
4592 * - compound_lock is held when nr_pages > 1
4594 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4597 static int mem_cgroup_move_account(struct page *page,
4598 unsigned int nr_pages,
4599 struct mem_cgroup *from,
4600 struct mem_cgroup *to)
4602 unsigned long flags;
4606 VM_BUG_ON(from == to);
4607 VM_BUG_ON_PAGE(PageLRU(page), page);
4609 * The page is isolated from LRU. So, collapse function
4610 * will not handle this page. But page splitting can happen.
4611 * Do this check under compound_page_lock(). The caller should
4615 if (nr_pages > 1 && !PageTransHuge(page))
4619 * Prevent mem_cgroup_replace_page() from looking at
4620 * page->mem_cgroup of its source page while we change it.
4622 if (!trylock_page(page))
4626 if (page->mem_cgroup != from)
4629 anon = PageAnon(page);
4631 spin_lock_irqsave(&from->move_lock, flags);
4633 if (!anon && page_mapped(page)) {
4634 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4636 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4641 * move_lock grabbed above and caller set from->moving_account, so
4642 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4643 * So mapping should be stable for dirty pages.
4645 if (!anon && PageDirty(page)) {
4646 struct address_space *mapping = page_mapping(page);
4648 if (mapping_cap_account_dirty(mapping)) {
4649 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4651 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4656 if (PageWriteback(page)) {
4657 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4659 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4664 * It is safe to change page->mem_cgroup here because the page
4665 * is referenced, charged, and isolated - we can't race with
4666 * uncharging, charging, migration, or LRU putback.
4669 /* caller should have done css_get */
4670 page->mem_cgroup = to;
4671 spin_unlock_irqrestore(&from->move_lock, flags);
4675 local_irq_disable();
4676 mem_cgroup_charge_statistics(to, page, nr_pages);
4677 memcg_check_events(to, page);
4678 mem_cgroup_charge_statistics(from, page, -nr_pages);
4679 memcg_check_events(from, page);
4687 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4688 unsigned long addr, pte_t ptent, union mc_target *target)
4690 struct page *page = NULL;
4691 enum mc_target_type ret = MC_TARGET_NONE;
4692 swp_entry_t ent = { .val = 0 };
4694 if (pte_present(ptent))
4695 page = mc_handle_present_pte(vma, addr, ptent);
4696 else if (is_swap_pte(ptent))
4697 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4698 else if (pte_none(ptent))
4699 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4701 if (!page && !ent.val)
4705 * Do only loose check w/o serialization.
4706 * mem_cgroup_move_account() checks the page is valid or
4707 * not under LRU exclusion.
4709 if (page->mem_cgroup == mc.from) {
4710 ret = MC_TARGET_PAGE;
4712 target->page = page;
4714 if (!ret || !target)
4717 /* There is a swap entry and a page doesn't exist or isn't charged */
4718 if (ent.val && !ret &&
4719 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4720 ret = MC_TARGET_SWAP;
4727 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4729 * We don't consider swapping or file mapped pages because THP does not
4730 * support them for now.
4731 * Caller should make sure that pmd_trans_huge(pmd) is true.
4733 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4734 unsigned long addr, pmd_t pmd, union mc_target *target)
4736 struct page *page = NULL;
4737 enum mc_target_type ret = MC_TARGET_NONE;
4739 page = pmd_page(pmd);
4740 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4741 if (!(mc.flags & MOVE_ANON))
4743 if (page->mem_cgroup == mc.from) {
4744 ret = MC_TARGET_PAGE;
4747 target->page = page;
4753 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4754 unsigned long addr, pmd_t pmd, union mc_target *target)
4756 return MC_TARGET_NONE;
4760 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4761 unsigned long addr, unsigned long end,
4762 struct mm_walk *walk)
4764 struct vm_area_struct *vma = walk->vma;
4768 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4769 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4770 mc.precharge += HPAGE_PMD_NR;
4775 if (pmd_trans_unstable(pmd))
4777 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4778 for (; addr != end; pte++, addr += PAGE_SIZE)
4779 if (get_mctgt_type(vma, addr, *pte, NULL))
4780 mc.precharge++; /* increment precharge temporarily */
4781 pte_unmap_unlock(pte - 1, ptl);
4787 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4789 unsigned long precharge;
4791 struct mm_walk mem_cgroup_count_precharge_walk = {
4792 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4795 down_read(&mm->mmap_sem);
4796 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4797 up_read(&mm->mmap_sem);
4799 precharge = mc.precharge;
4805 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4807 unsigned long precharge = mem_cgroup_count_precharge(mm);
4809 VM_BUG_ON(mc.moving_task);
4810 mc.moving_task = current;
4811 return mem_cgroup_do_precharge(precharge);
4814 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4815 static void __mem_cgroup_clear_mc(void)
4817 struct mem_cgroup *from = mc.from;
4818 struct mem_cgroup *to = mc.to;
4820 /* we must uncharge all the leftover precharges from mc.to */
4822 cancel_charge(mc.to, mc.precharge);
4826 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4827 * we must uncharge here.
4829 if (mc.moved_charge) {
4830 cancel_charge(mc.from, mc.moved_charge);
4831 mc.moved_charge = 0;
4833 /* we must fixup refcnts and charges */
4834 if (mc.moved_swap) {
4835 /* uncharge swap account from the old cgroup */
4836 if (!mem_cgroup_is_root(mc.from))
4837 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4840 * we charged both to->memory and to->memsw, so we
4841 * should uncharge to->memory.
4843 if (!mem_cgroup_is_root(mc.to))
4844 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4846 css_put_many(&mc.from->css, mc.moved_swap);
4848 /* we've already done css_get(mc.to) */
4851 memcg_oom_recover(from);
4852 memcg_oom_recover(to);
4853 wake_up_all(&mc.waitq);
4856 static void mem_cgroup_clear_mc(void)
4858 struct mm_struct *mm = mc.mm;
4861 * we must clear moving_task before waking up waiters at the end of
4864 mc.moving_task = NULL;
4865 __mem_cgroup_clear_mc();
4866 spin_lock(&mc.lock);
4870 spin_unlock(&mc.lock);
4875 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4877 struct cgroup_subsys_state *css;
4878 struct mem_cgroup *memcg;
4879 struct mem_cgroup *from;
4880 struct task_struct *leader, *p;
4881 struct mm_struct *mm;
4882 unsigned long move_flags;
4885 /* charge immigration isn't supported on the default hierarchy */
4886 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4890 * Multi-process migrations only happen on the default hierarchy
4891 * where charge immigration is not used. Perform charge
4892 * immigration if @tset contains a leader and whine if there are
4896 cgroup_taskset_for_each_leader(leader, css, tset) {
4899 memcg = mem_cgroup_from_css(css);
4905 * We are now commited to this value whatever it is. Changes in this
4906 * tunable will only affect upcoming migrations, not the current one.
4907 * So we need to save it, and keep it going.
4909 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4913 from = mem_cgroup_from_task(p);
4915 VM_BUG_ON(from == memcg);
4917 mm = get_task_mm(p);
4920 /* We move charges only when we move a owner of the mm */
4921 if (mm->owner == p) {
4924 VM_BUG_ON(mc.precharge);
4925 VM_BUG_ON(mc.moved_charge);
4926 VM_BUG_ON(mc.moved_swap);
4928 spin_lock(&mc.lock);
4932 mc.flags = move_flags;
4933 spin_unlock(&mc.lock);
4934 /* We set mc.moving_task later */
4936 ret = mem_cgroup_precharge_mc(mm);
4938 mem_cgroup_clear_mc();
4945 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4948 mem_cgroup_clear_mc();
4951 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4952 unsigned long addr, unsigned long end,
4953 struct mm_walk *walk)
4956 struct vm_area_struct *vma = walk->vma;
4959 enum mc_target_type target_type;
4960 union mc_target target;
4964 * We don't take compound_lock() here but no race with splitting thp
4966 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4967 * under splitting, which means there's no concurrent thp split,
4968 * - if another thread runs into split_huge_page() just after we
4969 * entered this if-block, the thread must wait for page table lock
4970 * to be unlocked in __split_huge_page_splitting(), where the main
4971 * part of thp split is not executed yet.
4973 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4974 if (mc.precharge < HPAGE_PMD_NR) {
4978 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4979 if (target_type == MC_TARGET_PAGE) {
4981 if (!isolate_lru_page(page)) {
4982 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4984 mc.precharge -= HPAGE_PMD_NR;
4985 mc.moved_charge += HPAGE_PMD_NR;
4987 putback_lru_page(page);
4995 if (pmd_trans_unstable(pmd))
4998 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4999 for (; addr != end; addr += PAGE_SIZE) {
5000 pte_t ptent = *(pte++);
5006 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5007 case MC_TARGET_PAGE:
5009 if (isolate_lru_page(page))
5011 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5013 /* we uncharge from mc.from later. */
5016 putback_lru_page(page);
5017 put: /* get_mctgt_type() gets the page */
5020 case MC_TARGET_SWAP:
5022 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5024 /* we fixup refcnts and charges later. */
5032 pte_unmap_unlock(pte - 1, ptl);
5037 * We have consumed all precharges we got in can_attach().
5038 * We try charge one by one, but don't do any additional
5039 * charges to mc.to if we have failed in charge once in attach()
5042 ret = mem_cgroup_do_precharge(1);
5050 static void mem_cgroup_move_charge(void)
5052 struct mm_walk mem_cgroup_move_charge_walk = {
5053 .pmd_entry = mem_cgroup_move_charge_pte_range,
5057 lru_add_drain_all();
5059 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5060 * move_lock while we're moving its pages to another memcg.
5061 * Then wait for already started RCU-only updates to finish.
5063 atomic_inc(&mc.from->moving_account);
5066 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5068 * Someone who are holding the mmap_sem might be waiting in
5069 * waitq. So we cancel all extra charges, wake up all waiters,
5070 * and retry. Because we cancel precharges, we might not be able
5071 * to move enough charges, but moving charge is a best-effort
5072 * feature anyway, so it wouldn't be a big problem.
5074 __mem_cgroup_clear_mc();
5079 * When we have consumed all precharges and failed in doing
5080 * additional charge, the page walk just aborts.
5082 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5083 up_read(&mc.mm->mmap_sem);
5084 atomic_dec(&mc.from->moving_account);
5087 static void mem_cgroup_move_task(void)
5090 mem_cgroup_move_charge();
5091 mem_cgroup_clear_mc();
5094 #else /* !CONFIG_MMU */
5095 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5099 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5102 static void mem_cgroup_move_task(void)
5108 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5109 * to verify whether we're attached to the default hierarchy on each mount
5112 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5115 * use_hierarchy is forced on the default hierarchy. cgroup core
5116 * guarantees that @root doesn't have any children, so turning it
5117 * on for the root memcg is enough.
5119 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5120 root_mem_cgroup->use_hierarchy = true;
5122 root_mem_cgroup->use_hierarchy = false;
5125 static u64 memory_current_read(struct cgroup_subsys_state *css,
5128 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5130 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5133 static int memory_low_show(struct seq_file *m, void *v)
5135 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5136 unsigned long low = READ_ONCE(memcg->low);
5138 if (low == PAGE_COUNTER_MAX)
5139 seq_puts(m, "max\n");
5141 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5146 static ssize_t memory_low_write(struct kernfs_open_file *of,
5147 char *buf, size_t nbytes, loff_t off)
5149 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5153 buf = strstrip(buf);
5154 err = page_counter_memparse(buf, "max", &low);
5163 static int memory_high_show(struct seq_file *m, void *v)
5165 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5166 unsigned long high = READ_ONCE(memcg->high);
5168 if (high == PAGE_COUNTER_MAX)
5169 seq_puts(m, "max\n");
5171 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5176 static ssize_t memory_high_write(struct kernfs_open_file *of,
5177 char *buf, size_t nbytes, loff_t off)
5179 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5180 unsigned long nr_pages;
5184 buf = strstrip(buf);
5185 err = page_counter_memparse(buf, "max", &high);
5191 nr_pages = page_counter_read(&memcg->memory);
5192 if (nr_pages > high)
5193 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5196 memcg_wb_domain_size_changed(memcg);
5200 static int memory_max_show(struct seq_file *m, void *v)
5202 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5203 unsigned long max = READ_ONCE(memcg->memory.limit);
5205 if (max == PAGE_COUNTER_MAX)
5206 seq_puts(m, "max\n");
5208 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5213 static ssize_t memory_max_write(struct kernfs_open_file *of,
5214 char *buf, size_t nbytes, loff_t off)
5216 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5217 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5218 bool drained = false;
5222 buf = strstrip(buf);
5223 err = page_counter_memparse(buf, "max", &max);
5227 xchg(&memcg->memory.limit, max);
5230 unsigned long nr_pages = page_counter_read(&memcg->memory);
5232 if (nr_pages <= max)
5235 if (signal_pending(current)) {
5241 drain_all_stock(memcg);
5247 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5253 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5254 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5258 memcg_wb_domain_size_changed(memcg);
5262 static int memory_events_show(struct seq_file *m, void *v)
5264 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5266 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5267 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5268 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5269 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5274 static struct cftype memory_files[] = {
5277 .flags = CFTYPE_NOT_ON_ROOT,
5278 .read_u64 = memory_current_read,
5282 .flags = CFTYPE_NOT_ON_ROOT,
5283 .seq_show = memory_low_show,
5284 .write = memory_low_write,
5288 .flags = CFTYPE_NOT_ON_ROOT,
5289 .seq_show = memory_high_show,
5290 .write = memory_high_write,
5294 .flags = CFTYPE_NOT_ON_ROOT,
5295 .seq_show = memory_max_show,
5296 .write = memory_max_write,
5300 .flags = CFTYPE_NOT_ON_ROOT,
5301 .file_offset = offsetof(struct mem_cgroup, events_file),
5302 .seq_show = memory_events_show,
5307 struct cgroup_subsys memory_cgrp_subsys = {
5308 .css_alloc = mem_cgroup_css_alloc,
5309 .css_online = mem_cgroup_css_online,
5310 .css_offline = mem_cgroup_css_offline,
5311 .css_released = mem_cgroup_css_released,
5312 .css_free = mem_cgroup_css_free,
5313 .css_reset = mem_cgroup_css_reset,
5314 .can_attach = mem_cgroup_can_attach,
5315 .cancel_attach = mem_cgroup_cancel_attach,
5316 .post_attach = mem_cgroup_move_task,
5317 .bind = mem_cgroup_bind,
5318 .dfl_cftypes = memory_files,
5319 .legacy_cftypes = mem_cgroup_legacy_files,
5324 * mem_cgroup_low - check if memory consumption is below the normal range
5325 * @root: the highest ancestor to consider
5326 * @memcg: the memory cgroup to check
5328 * Returns %true if memory consumption of @memcg, and that of all
5329 * configurable ancestors up to @root, is below the normal range.
5331 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5333 if (mem_cgroup_disabled())
5337 * The toplevel group doesn't have a configurable range, so
5338 * it's never low when looked at directly, and it is not
5339 * considered an ancestor when assessing the hierarchy.
5342 if (memcg == root_mem_cgroup)
5345 if (page_counter_read(&memcg->memory) >= memcg->low)
5348 while (memcg != root) {
5349 memcg = parent_mem_cgroup(memcg);
5351 if (memcg == root_mem_cgroup)
5354 if (page_counter_read(&memcg->memory) >= memcg->low)
5361 * mem_cgroup_try_charge - try charging a page
5362 * @page: page to charge
5363 * @mm: mm context of the victim
5364 * @gfp_mask: reclaim mode
5365 * @memcgp: charged memcg return
5367 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5368 * pages according to @gfp_mask if necessary.
5370 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5371 * Otherwise, an error code is returned.
5373 * After page->mapping has been set up, the caller must finalize the
5374 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5375 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5377 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5378 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5380 struct mem_cgroup *memcg = NULL;
5381 unsigned int nr_pages = 1;
5384 if (mem_cgroup_disabled())
5387 if (PageSwapCache(page)) {
5389 * Every swap fault against a single page tries to charge the
5390 * page, bail as early as possible. shmem_unuse() encounters
5391 * already charged pages, too. The USED bit is protected by
5392 * the page lock, which serializes swap cache removal, which
5393 * in turn serializes uncharging.
5395 VM_BUG_ON_PAGE(!PageLocked(page), page);
5396 if (page->mem_cgroup)
5399 if (do_swap_account) {
5400 swp_entry_t ent = { .val = page_private(page), };
5401 unsigned short id = lookup_swap_cgroup_id(ent);
5404 memcg = mem_cgroup_from_id(id);
5405 if (memcg && !css_tryget_online(&memcg->css))
5411 if (PageTransHuge(page)) {
5412 nr_pages <<= compound_order(page);
5413 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5417 memcg = get_mem_cgroup_from_mm(mm);
5419 ret = try_charge(memcg, gfp_mask, nr_pages);
5421 css_put(&memcg->css);
5428 * mem_cgroup_commit_charge - commit a page charge
5429 * @page: page to charge
5430 * @memcg: memcg to charge the page to
5431 * @lrucare: page might be on LRU already
5433 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5434 * after page->mapping has been set up. This must happen atomically
5435 * as part of the page instantiation, i.e. under the page table lock
5436 * for anonymous pages, under the page lock for page and swap cache.
5438 * In addition, the page must not be on the LRU during the commit, to
5439 * prevent racing with task migration. If it might be, use @lrucare.
5441 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5443 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5446 unsigned int nr_pages = 1;
5448 VM_BUG_ON_PAGE(!page->mapping, page);
5449 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5451 if (mem_cgroup_disabled())
5454 * Swap faults will attempt to charge the same page multiple
5455 * times. But reuse_swap_page() might have removed the page
5456 * from swapcache already, so we can't check PageSwapCache().
5461 commit_charge(page, memcg, lrucare);
5463 if (PageTransHuge(page)) {
5464 nr_pages <<= compound_order(page);
5465 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5468 local_irq_disable();
5469 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5470 memcg_check_events(memcg, page);
5473 if (do_swap_account && PageSwapCache(page)) {
5474 swp_entry_t entry = { .val = page_private(page) };
5476 * The swap entry might not get freed for a long time,
5477 * let's not wait for it. The page already received a
5478 * memory+swap charge, drop the swap entry duplicate.
5480 mem_cgroup_uncharge_swap(entry);
5485 * mem_cgroup_cancel_charge - cancel a page charge
5486 * @page: page to charge
5487 * @memcg: memcg to charge the page to
5489 * Cancel a charge transaction started by mem_cgroup_try_charge().
5491 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5493 unsigned int nr_pages = 1;
5495 if (mem_cgroup_disabled())
5498 * Swap faults will attempt to charge the same page multiple
5499 * times. But reuse_swap_page() might have removed the page
5500 * from swapcache already, so we can't check PageSwapCache().
5505 if (PageTransHuge(page)) {
5506 nr_pages <<= compound_order(page);
5507 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5510 cancel_charge(memcg, nr_pages);
5513 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5514 unsigned long nr_anon, unsigned long nr_file,
5515 unsigned long nr_huge, struct page *dummy_page)
5517 unsigned long nr_pages = nr_anon + nr_file;
5518 unsigned long flags;
5520 if (!mem_cgroup_is_root(memcg)) {
5521 page_counter_uncharge(&memcg->memory, nr_pages);
5522 if (do_swap_account)
5523 page_counter_uncharge(&memcg->memsw, nr_pages);
5524 memcg_oom_recover(memcg);
5527 local_irq_save(flags);
5528 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5529 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5530 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5531 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5532 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5533 memcg_check_events(memcg, dummy_page);
5534 local_irq_restore(flags);
5536 if (!mem_cgroup_is_root(memcg))
5537 css_put_many(&memcg->css, nr_pages);
5540 static void uncharge_list(struct list_head *page_list)
5542 struct mem_cgroup *memcg = NULL;
5543 unsigned long nr_anon = 0;
5544 unsigned long nr_file = 0;
5545 unsigned long nr_huge = 0;
5546 unsigned long pgpgout = 0;
5547 struct list_head *next;
5550 next = page_list->next;
5552 unsigned int nr_pages = 1;
5554 page = list_entry(next, struct page, lru);
5555 next = page->lru.next;
5557 VM_BUG_ON_PAGE(PageLRU(page), page);
5558 VM_BUG_ON_PAGE(page_count(page), page);
5560 if (!page->mem_cgroup)
5564 * Nobody should be changing or seriously looking at
5565 * page->mem_cgroup at this point, we have fully
5566 * exclusive access to the page.
5569 if (memcg != page->mem_cgroup) {
5571 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5573 pgpgout = nr_anon = nr_file = nr_huge = 0;
5575 memcg = page->mem_cgroup;
5578 if (PageTransHuge(page)) {
5579 nr_pages <<= compound_order(page);
5580 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5581 nr_huge += nr_pages;
5585 nr_anon += nr_pages;
5587 nr_file += nr_pages;
5589 page->mem_cgroup = NULL;
5592 } while (next != page_list);
5595 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5600 * mem_cgroup_uncharge - uncharge a page
5601 * @page: page to uncharge
5603 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5604 * mem_cgroup_commit_charge().
5606 void mem_cgroup_uncharge(struct page *page)
5608 if (mem_cgroup_disabled())
5611 /* Don't touch page->lru of any random page, pre-check: */
5612 if (!page->mem_cgroup)
5615 INIT_LIST_HEAD(&page->lru);
5616 uncharge_list(&page->lru);
5620 * mem_cgroup_uncharge_list - uncharge a list of page
5621 * @page_list: list of pages to uncharge
5623 * Uncharge a list of pages previously charged with
5624 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5626 void mem_cgroup_uncharge_list(struct list_head *page_list)
5628 if (mem_cgroup_disabled())
5631 if (!list_empty(page_list))
5632 uncharge_list(page_list);
5636 * mem_cgroup_replace_page - migrate a charge to another page
5637 * @oldpage: currently charged page
5638 * @newpage: page to transfer the charge to
5640 * Migrate the charge from @oldpage to @newpage.
5642 * Both pages must be locked, @newpage->mapping must be set up.
5643 * Either or both pages might be on the LRU already.
5645 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5647 struct mem_cgroup *memcg;
5650 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5651 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5652 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5653 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5656 if (mem_cgroup_disabled())
5659 /* Page cache replacement: new page already charged? */
5660 if (newpage->mem_cgroup)
5663 /* Swapcache readahead pages can get replaced before being charged */
5664 memcg = oldpage->mem_cgroup;
5668 lock_page_lru(oldpage, &isolated);
5669 oldpage->mem_cgroup = NULL;
5670 unlock_page_lru(oldpage, isolated);
5672 commit_charge(newpage, memcg, true);
5676 * subsys_initcall() for memory controller.
5678 * Some parts like hotcpu_notifier() have to be initialized from this context
5679 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5680 * everything that doesn't depend on a specific mem_cgroup structure should
5681 * be initialized from here.
5683 static int __init mem_cgroup_init(void)
5687 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5689 for_each_possible_cpu(cpu)
5690 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5693 for_each_node(node) {
5694 struct mem_cgroup_tree_per_node *rtpn;
5697 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5698 node_online(node) ? node : NUMA_NO_NODE);
5700 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5701 struct mem_cgroup_tree_per_zone *rtpz;
5703 rtpz = &rtpn->rb_tree_per_zone[zone];
5704 rtpz->rb_root = RB_ROOT;
5705 spin_lock_init(&rtpz->lock);
5707 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5712 subsys_initcall(mem_cgroup_init);
5714 #ifdef CONFIG_MEMCG_SWAP
5716 * mem_cgroup_swapout - transfer a memsw charge to swap
5717 * @page: page whose memsw charge to transfer
5718 * @entry: swap entry to move the charge to
5720 * Transfer the memsw charge of @page to @entry.
5722 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5724 struct mem_cgroup *memcg;
5725 unsigned short oldid;
5727 VM_BUG_ON_PAGE(PageLRU(page), page);
5728 VM_BUG_ON_PAGE(page_count(page), page);
5730 if (!do_swap_account)
5733 memcg = page->mem_cgroup;
5735 /* Readahead page, never charged */
5739 mem_cgroup_id_get(memcg);
5740 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5741 VM_BUG_ON_PAGE(oldid, page);
5742 mem_cgroup_swap_statistics(memcg, true);
5744 page->mem_cgroup = NULL;
5746 if (!mem_cgroup_is_root(memcg))
5747 page_counter_uncharge(&memcg->memory, 1);
5750 * Interrupts should be disabled here because the caller holds the
5751 * mapping->tree_lock lock which is taken with interrupts-off. It is
5752 * important here to have the interrupts disabled because it is the
5753 * only synchronisation we have for udpating the per-CPU variables.
5755 VM_BUG_ON(!irqs_disabled());
5756 mem_cgroup_charge_statistics(memcg, page, -1);
5757 memcg_check_events(memcg, page);
5759 if (!mem_cgroup_is_root(memcg))
5760 css_put(&memcg->css);
5764 * mem_cgroup_uncharge_swap - uncharge a swap entry
5765 * @entry: swap entry to uncharge
5767 * Drop the memsw charge associated with @entry.
5769 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5771 struct mem_cgroup *memcg;
5774 if (!do_swap_account)
5777 id = swap_cgroup_record(entry, 0);
5779 memcg = mem_cgroup_from_id(id);
5781 if (!mem_cgroup_is_root(memcg))
5782 page_counter_uncharge(&memcg->memsw, 1);
5783 mem_cgroup_swap_statistics(memcg, false);
5784 mem_cgroup_id_put(memcg);
5789 /* for remember boot option*/
5790 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5791 static int really_do_swap_account __initdata = 1;
5793 static int really_do_swap_account __initdata;
5796 static int __init enable_swap_account(char *s)
5798 if (!strcmp(s, "1"))
5799 really_do_swap_account = 1;
5800 else if (!strcmp(s, "0"))
5801 really_do_swap_account = 0;
5804 __setup("swapaccount=", enable_swap_account);
5806 static struct cftype memsw_cgroup_files[] = {
5808 .name = "memsw.usage_in_bytes",
5809 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5810 .read_u64 = mem_cgroup_read_u64,
5813 .name = "memsw.max_usage_in_bytes",
5814 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5815 .write = mem_cgroup_reset,
5816 .read_u64 = mem_cgroup_read_u64,
5819 .name = "memsw.limit_in_bytes",
5820 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5821 .write = mem_cgroup_write,
5822 .read_u64 = mem_cgroup_read_u64,
5825 .name = "memsw.failcnt",
5826 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5827 .write = mem_cgroup_reset,
5828 .read_u64 = mem_cgroup_read_u64,
5830 { }, /* terminate */
5833 static int __init mem_cgroup_swap_init(void)
5835 if (!mem_cgroup_disabled() && really_do_swap_account) {
5836 do_swap_account = 1;
5837 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5838 memsw_cgroup_files));
5842 subsys_initcall(mem_cgroup_swap_init);
5844 #endif /* CONFIG_MEMCG_SWAP */