1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
69 #include <net/tcp_memcontrol.h>
70 #include <linux/locallock.h>
74 #include <asm/uaccess.h>
76 #include <trace/events/vmscan.h>
78 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
79 EXPORT_SYMBOL(memory_cgrp_subsys);
81 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 static struct mem_cgroup *root_mem_cgroup __read_mostly;
83 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 int do_swap_account __read_mostly;
89 #define do_swap_account 0
92 static DEFINE_LOCAL_IRQ_LOCK(event_lock);
93 static const char * const mem_cgroup_stat_names[] = {
103 static const char * const mem_cgroup_events_names[] = {
110 static const char * const mem_cgroup_lru_names[] = {
118 #define THRESHOLDS_EVENTS_TARGET 128
119 #define SOFTLIMIT_EVENTS_TARGET 1024
120 #define NUMAINFO_EVENTS_TARGET 1024
123 * Cgroups above their limits are maintained in a RB-Tree, independent of
124 * their hierarchy representation
127 struct mem_cgroup_tree_per_zone {
128 struct rb_root rb_root;
132 struct mem_cgroup_tree_per_node {
133 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
136 struct mem_cgroup_tree {
137 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
140 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
143 struct mem_cgroup_eventfd_list {
144 struct list_head list;
145 struct eventfd_ctx *eventfd;
149 * cgroup_event represents events which userspace want to receive.
151 struct mem_cgroup_event {
153 * memcg which the event belongs to.
155 struct mem_cgroup *memcg;
157 * eventfd to signal userspace about the event.
159 struct eventfd_ctx *eventfd;
161 * Each of these stored in a list by the cgroup.
163 struct list_head list;
165 * register_event() callback will be used to add new userspace
166 * waiter for changes related to this event. Use eventfd_signal()
167 * on eventfd to send notification to userspace.
169 int (*register_event)(struct mem_cgroup *memcg,
170 struct eventfd_ctx *eventfd, const char *args);
172 * unregister_event() callback will be called when userspace closes
173 * the eventfd or on cgroup removing. This callback must be set,
174 * if you want provide notification functionality.
176 void (*unregister_event)(struct mem_cgroup *memcg,
177 struct eventfd_ctx *eventfd);
179 * All fields below needed to unregister event when
180 * userspace closes eventfd.
183 wait_queue_head_t *wqh;
185 struct work_struct remove;
188 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
189 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
191 /* Stuffs for move charges at task migration. */
193 * Types of charges to be moved.
195 #define MOVE_ANON 0x1U
196 #define MOVE_FILE 0x2U
197 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
199 /* "mc" and its members are protected by cgroup_mutex */
200 static struct move_charge_struct {
201 spinlock_t lock; /* for from, to */
202 struct mem_cgroup *from;
203 struct mem_cgroup *to;
205 unsigned long precharge;
206 unsigned long moved_charge;
207 unsigned long moved_swap;
208 struct task_struct *moving_task; /* a task moving charges */
209 wait_queue_head_t waitq; /* a waitq for other context */
211 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
212 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
216 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
217 * limit reclaim to prevent infinite loops, if they ever occur.
219 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
220 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
223 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
224 MEM_CGROUP_CHARGE_TYPE_ANON,
225 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
226 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
230 /* for encoding cft->private value on file */
238 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
239 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
240 #define MEMFILE_ATTR(val) ((val) & 0xffff)
241 /* Used for OOM nofiier */
242 #define OOM_CONTROL (0)
245 * The memcg_create_mutex will be held whenever a new cgroup is created.
246 * As a consequence, any change that needs to protect against new child cgroups
247 * appearing has to hold it as well.
249 static DEFINE_MUTEX(memcg_create_mutex);
251 /* Some nice accessors for the vmpressure. */
252 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
255 memcg = root_mem_cgroup;
256 return &memcg->vmpressure;
259 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
261 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
264 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
266 return (memcg == root_mem_cgroup);
270 * We restrict the id in the range of [1, 65535], so it can fit into
273 #define MEM_CGROUP_ID_MAX USHRT_MAX
275 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
277 return memcg->css.id;
281 * A helper function to get mem_cgroup from ID. must be called under
282 * rcu_read_lock(). The caller is responsible for calling
283 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
284 * refcnt from swap can be called against removed memcg.)
286 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
288 struct cgroup_subsys_state *css;
290 css = css_from_id(id, &memory_cgrp_subsys);
291 return mem_cgroup_from_css(css);
294 /* Writing them here to avoid exposing memcg's inner layout */
295 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
297 void sock_update_memcg(struct sock *sk)
299 if (mem_cgroup_sockets_enabled) {
300 struct mem_cgroup *memcg;
301 struct cg_proto *cg_proto;
303 BUG_ON(!sk->sk_prot->proto_cgroup);
305 /* Socket cloning can throw us here with sk_cgrp already
306 * filled. It won't however, necessarily happen from
307 * process context. So the test for root memcg given
308 * the current task's memcg won't help us in this case.
310 * Respecting the original socket's memcg is a better
311 * decision in this case.
314 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
315 css_get(&sk->sk_cgrp->memcg->css);
320 memcg = mem_cgroup_from_task(current);
321 cg_proto = sk->sk_prot->proto_cgroup(memcg);
322 if (cg_proto && test_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags) &&
323 css_tryget_online(&memcg->css)) {
324 sk->sk_cgrp = cg_proto;
329 EXPORT_SYMBOL(sock_update_memcg);
331 void sock_release_memcg(struct sock *sk)
333 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
334 struct mem_cgroup *memcg;
335 WARN_ON(!sk->sk_cgrp->memcg);
336 memcg = sk->sk_cgrp->memcg;
337 css_put(&sk->sk_cgrp->memcg->css);
341 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
343 if (!memcg || mem_cgroup_is_root(memcg))
346 return &memcg->tcp_mem;
348 EXPORT_SYMBOL(tcp_proto_cgroup);
352 #ifdef CONFIG_MEMCG_KMEM
354 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
355 * The main reason for not using cgroup id for this:
356 * this works better in sparse environments, where we have a lot of memcgs,
357 * but only a few kmem-limited. Or also, if we have, for instance, 200
358 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
359 * 200 entry array for that.
361 * The current size of the caches array is stored in memcg_nr_cache_ids. It
362 * will double each time we have to increase it.
364 static DEFINE_IDA(memcg_cache_ida);
365 int memcg_nr_cache_ids;
367 /* Protects memcg_nr_cache_ids */
368 static DECLARE_RWSEM(memcg_cache_ids_sem);
370 void memcg_get_cache_ids(void)
372 down_read(&memcg_cache_ids_sem);
375 void memcg_put_cache_ids(void)
377 up_read(&memcg_cache_ids_sem);
381 * MIN_SIZE is different than 1, because we would like to avoid going through
382 * the alloc/free process all the time. In a small machine, 4 kmem-limited
383 * cgroups is a reasonable guess. In the future, it could be a parameter or
384 * tunable, but that is strictly not necessary.
386 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
387 * this constant directly from cgroup, but it is understandable that this is
388 * better kept as an internal representation in cgroup.c. In any case, the
389 * cgrp_id space is not getting any smaller, and we don't have to necessarily
390 * increase ours as well if it increases.
392 #define MEMCG_CACHES_MIN_SIZE 4
393 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
396 * A lot of the calls to the cache allocation functions are expected to be
397 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
398 * conditional to this static branch, we'll have to allow modules that does
399 * kmem_cache_alloc and the such to see this symbol as well
401 struct static_key memcg_kmem_enabled_key;
402 EXPORT_SYMBOL(memcg_kmem_enabled_key);
404 #endif /* CONFIG_MEMCG_KMEM */
406 static struct mem_cgroup_per_zone *
407 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
409 int nid = zone_to_nid(zone);
410 int zid = zone_idx(zone);
412 return &memcg->nodeinfo[nid]->zoneinfo[zid];
416 * mem_cgroup_css_from_page - css of the memcg associated with a page
417 * @page: page of interest
419 * If memcg is bound to the default hierarchy, css of the memcg associated
420 * with @page is returned. The returned css remains associated with @page
421 * until it is released.
423 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
426 * XXX: The above description of behavior on the default hierarchy isn't
427 * strictly true yet as replace_page_cache_page() can modify the
428 * association before @page is released even on the default hierarchy;
429 * however, the current and planned usages don't mix the the two functions
430 * and replace_page_cache_page() will soon be updated to make the invariant
433 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
435 struct mem_cgroup *memcg;
439 memcg = page->mem_cgroup;
441 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
442 memcg = root_mem_cgroup;
449 * page_cgroup_ino - return inode number of the memcg a page is charged to
452 * Look up the closest online ancestor of the memory cgroup @page is charged to
453 * and return its inode number or 0 if @page is not charged to any cgroup. It
454 * is safe to call this function without holding a reference to @page.
456 * Note, this function is inherently racy, because there is nothing to prevent
457 * the cgroup inode from getting torn down and potentially reallocated a moment
458 * after page_cgroup_ino() returns, so it only should be used by callers that
459 * do not care (such as procfs interfaces).
461 ino_t page_cgroup_ino(struct page *page)
463 struct mem_cgroup *memcg;
464 unsigned long ino = 0;
467 memcg = READ_ONCE(page->mem_cgroup);
468 while (memcg && !(memcg->css.flags & CSS_ONLINE))
469 memcg = parent_mem_cgroup(memcg);
471 ino = cgroup_ino(memcg->css.cgroup);
476 static struct mem_cgroup_per_zone *
477 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
479 int nid = page_to_nid(page);
480 int zid = page_zonenum(page);
482 return &memcg->nodeinfo[nid]->zoneinfo[zid];
485 static struct mem_cgroup_tree_per_zone *
486 soft_limit_tree_node_zone(int nid, int zid)
488 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
491 static struct mem_cgroup_tree_per_zone *
492 soft_limit_tree_from_page(struct page *page)
494 int nid = page_to_nid(page);
495 int zid = page_zonenum(page);
497 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
500 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
501 struct mem_cgroup_tree_per_zone *mctz,
502 unsigned long new_usage_in_excess)
504 struct rb_node **p = &mctz->rb_root.rb_node;
505 struct rb_node *parent = NULL;
506 struct mem_cgroup_per_zone *mz_node;
511 mz->usage_in_excess = new_usage_in_excess;
512 if (!mz->usage_in_excess)
516 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
518 if (mz->usage_in_excess < mz_node->usage_in_excess)
521 * We can't avoid mem cgroups that are over their soft
522 * limit by the same amount
524 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
527 rb_link_node(&mz->tree_node, parent, p);
528 rb_insert_color(&mz->tree_node, &mctz->rb_root);
532 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
533 struct mem_cgroup_tree_per_zone *mctz)
537 rb_erase(&mz->tree_node, &mctz->rb_root);
541 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
542 struct mem_cgroup_tree_per_zone *mctz)
546 spin_lock_irqsave(&mctz->lock, flags);
547 __mem_cgroup_remove_exceeded(mz, mctz);
548 spin_unlock_irqrestore(&mctz->lock, flags);
551 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
553 unsigned long nr_pages = page_counter_read(&memcg->memory);
554 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
555 unsigned long excess = 0;
557 if (nr_pages > soft_limit)
558 excess = nr_pages - soft_limit;
563 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
565 unsigned long excess;
566 struct mem_cgroup_per_zone *mz;
567 struct mem_cgroup_tree_per_zone *mctz;
569 mctz = soft_limit_tree_from_page(page);
571 * Necessary to update all ancestors when hierarchy is used.
572 * because their event counter is not touched.
574 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
575 mz = mem_cgroup_page_zoneinfo(memcg, page);
576 excess = soft_limit_excess(memcg);
578 * We have to update the tree if mz is on RB-tree or
579 * mem is over its softlimit.
581 if (excess || mz->on_tree) {
584 spin_lock_irqsave(&mctz->lock, flags);
585 /* if on-tree, remove it */
587 __mem_cgroup_remove_exceeded(mz, mctz);
589 * Insert again. mz->usage_in_excess will be updated.
590 * If excess is 0, no tree ops.
592 __mem_cgroup_insert_exceeded(mz, mctz, excess);
593 spin_unlock_irqrestore(&mctz->lock, flags);
598 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
600 struct mem_cgroup_tree_per_zone *mctz;
601 struct mem_cgroup_per_zone *mz;
605 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
606 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
607 mctz = soft_limit_tree_node_zone(nid, zid);
608 mem_cgroup_remove_exceeded(mz, mctz);
613 static struct mem_cgroup_per_zone *
614 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
616 struct rb_node *rightmost = NULL;
617 struct mem_cgroup_per_zone *mz;
621 rightmost = rb_last(&mctz->rb_root);
623 goto done; /* Nothing to reclaim from */
625 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
627 * Remove the node now but someone else can add it back,
628 * we will to add it back at the end of reclaim to its correct
629 * position in the tree.
631 __mem_cgroup_remove_exceeded(mz, mctz);
632 if (!soft_limit_excess(mz->memcg) ||
633 !css_tryget_online(&mz->memcg->css))
639 static struct mem_cgroup_per_zone *
640 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
642 struct mem_cgroup_per_zone *mz;
644 spin_lock_irq(&mctz->lock);
645 mz = __mem_cgroup_largest_soft_limit_node(mctz);
646 spin_unlock_irq(&mctz->lock);
651 * Return page count for single (non recursive) @memcg.
653 * Implementation Note: reading percpu statistics for memcg.
655 * Both of vmstat[] and percpu_counter has threshold and do periodic
656 * synchronization to implement "quick" read. There are trade-off between
657 * reading cost and precision of value. Then, we may have a chance to implement
658 * a periodic synchronization of counter in memcg's counter.
660 * But this _read() function is used for user interface now. The user accounts
661 * memory usage by memory cgroup and he _always_ requires exact value because
662 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
663 * have to visit all online cpus and make sum. So, for now, unnecessary
664 * synchronization is not implemented. (just implemented for cpu hotplug)
666 * If there are kernel internal actions which can make use of some not-exact
667 * value, and reading all cpu value can be performance bottleneck in some
668 * common workload, threshold and synchronization as vmstat[] should be
672 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
677 /* Per-cpu values can be negative, use a signed accumulator */
678 for_each_possible_cpu(cpu)
679 val += per_cpu(memcg->stat->count[idx], cpu);
681 * Summing races with updates, so val may be negative. Avoid exposing
682 * transient negative values.
689 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
690 enum mem_cgroup_events_index idx)
692 unsigned long val = 0;
695 for_each_possible_cpu(cpu)
696 val += per_cpu(memcg->stat->events[idx], cpu);
700 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
705 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
706 * counted as CACHE even if it's on ANON LRU.
709 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
712 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
715 if (PageTransHuge(page))
716 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
719 /* pagein of a big page is an event. So, ignore page size */
721 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
723 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
724 nr_pages = -nr_pages; /* for event */
727 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
730 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
732 unsigned int lru_mask)
734 unsigned long nr = 0;
737 VM_BUG_ON((unsigned)nid >= nr_node_ids);
739 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
740 struct mem_cgroup_per_zone *mz;
744 if (!(BIT(lru) & lru_mask))
746 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
747 nr += mz->lru_size[lru];
753 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
754 unsigned int lru_mask)
756 unsigned long nr = 0;
759 for_each_node_state(nid, N_MEMORY)
760 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
764 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
765 enum mem_cgroup_events_target target)
767 unsigned long val, next;
769 val = __this_cpu_read(memcg->stat->nr_page_events);
770 next = __this_cpu_read(memcg->stat->targets[target]);
771 /* from time_after() in jiffies.h */
772 if ((long)next - (long)val < 0) {
774 case MEM_CGROUP_TARGET_THRESH:
775 next = val + THRESHOLDS_EVENTS_TARGET;
777 case MEM_CGROUP_TARGET_SOFTLIMIT:
778 next = val + SOFTLIMIT_EVENTS_TARGET;
780 case MEM_CGROUP_TARGET_NUMAINFO:
781 next = val + NUMAINFO_EVENTS_TARGET;
786 __this_cpu_write(memcg->stat->targets[target], next);
793 * Check events in order.
796 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
798 /* threshold event is triggered in finer grain than soft limit */
799 if (unlikely(mem_cgroup_event_ratelimit(memcg,
800 MEM_CGROUP_TARGET_THRESH))) {
802 bool do_numainfo __maybe_unused;
804 do_softlimit = mem_cgroup_event_ratelimit(memcg,
805 MEM_CGROUP_TARGET_SOFTLIMIT);
807 do_numainfo = mem_cgroup_event_ratelimit(memcg,
808 MEM_CGROUP_TARGET_NUMAINFO);
810 mem_cgroup_threshold(memcg);
811 if (unlikely(do_softlimit))
812 mem_cgroup_update_tree(memcg, page);
814 if (unlikely(do_numainfo))
815 atomic_inc(&memcg->numainfo_events);
820 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
823 * mm_update_next_owner() may clear mm->owner to NULL
824 * if it races with swapoff, page migration, etc.
825 * So this can be called with p == NULL.
830 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
832 EXPORT_SYMBOL(mem_cgroup_from_task);
834 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
836 struct mem_cgroup *memcg = NULL;
841 * Page cache insertions can happen withou an
842 * actual mm context, e.g. during disk probing
843 * on boot, loopback IO, acct() writes etc.
846 memcg = root_mem_cgroup;
848 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
849 if (unlikely(!memcg))
850 memcg = root_mem_cgroup;
852 } while (!css_tryget_online(&memcg->css));
858 * mem_cgroup_iter - iterate over memory cgroup hierarchy
859 * @root: hierarchy root
860 * @prev: previously returned memcg, NULL on first invocation
861 * @reclaim: cookie for shared reclaim walks, NULL for full walks
863 * Returns references to children of the hierarchy below @root, or
864 * @root itself, or %NULL after a full round-trip.
866 * Caller must pass the return value in @prev on subsequent
867 * invocations for reference counting, or use mem_cgroup_iter_break()
868 * to cancel a hierarchy walk before the round-trip is complete.
870 * Reclaimers can specify a zone and a priority level in @reclaim to
871 * divide up the memcgs in the hierarchy among all concurrent
872 * reclaimers operating on the same zone and priority.
874 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
875 struct mem_cgroup *prev,
876 struct mem_cgroup_reclaim_cookie *reclaim)
878 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
879 struct cgroup_subsys_state *css = NULL;
880 struct mem_cgroup *memcg = NULL;
881 struct mem_cgroup *pos = NULL;
883 if (mem_cgroup_disabled())
887 root = root_mem_cgroup;
889 if (prev && !reclaim)
892 if (!root->use_hierarchy && root != root_mem_cgroup) {
901 struct mem_cgroup_per_zone *mz;
903 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
904 iter = &mz->iter[reclaim->priority];
906 if (prev && reclaim->generation != iter->generation)
910 pos = READ_ONCE(iter->position);
911 if (!pos || css_tryget(&pos->css))
914 * css reference reached zero, so iter->position will
915 * be cleared by ->css_released. However, we should not
916 * rely on this happening soon, because ->css_released
917 * is called from a work queue, and by busy-waiting we
918 * might block it. So we clear iter->position right
921 (void)cmpxchg(&iter->position, pos, NULL);
929 css = css_next_descendant_pre(css, &root->css);
932 * Reclaimers share the hierarchy walk, and a
933 * new one might jump in right at the end of
934 * the hierarchy - make sure they see at least
935 * one group and restart from the beginning.
943 * Verify the css and acquire a reference. The root
944 * is provided by the caller, so we know it's alive
945 * and kicking, and don't take an extra reference.
947 memcg = mem_cgroup_from_css(css);
949 if (css == &root->css)
952 if (css_tryget(css)) {
954 * Make sure the memcg is initialized:
955 * mem_cgroup_css_online() orders the the
956 * initialization against setting the flag.
958 if (smp_load_acquire(&memcg->initialized))
969 * The position could have already been updated by a competing
970 * thread, so check that the value hasn't changed since we read
971 * it to avoid reclaiming from the same cgroup twice.
973 (void)cmpxchg(&iter->position, pos, memcg);
981 reclaim->generation = iter->generation;
987 if (prev && prev != root)
994 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
995 * @root: hierarchy root
996 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
998 void mem_cgroup_iter_break(struct mem_cgroup *root,
999 struct mem_cgroup *prev)
1002 root = root_mem_cgroup;
1003 if (prev && prev != root)
1004 css_put(&prev->css);
1007 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1009 struct mem_cgroup *memcg = dead_memcg;
1010 struct mem_cgroup_reclaim_iter *iter;
1011 struct mem_cgroup_per_zone *mz;
1015 while ((memcg = parent_mem_cgroup(memcg))) {
1016 for_each_node(nid) {
1017 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1018 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1019 for (i = 0; i <= DEF_PRIORITY; i++) {
1020 iter = &mz->iter[i];
1021 cmpxchg(&iter->position,
1030 * Iteration constructs for visiting all cgroups (under a tree). If
1031 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1032 * be used for reference counting.
1034 #define for_each_mem_cgroup_tree(iter, root) \
1035 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1037 iter = mem_cgroup_iter(root, iter, NULL))
1039 #define for_each_mem_cgroup(iter) \
1040 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1042 iter = mem_cgroup_iter(NULL, iter, NULL))
1045 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1046 * @zone: zone of the wanted lruvec
1047 * @memcg: memcg of the wanted lruvec
1049 * Returns the lru list vector holding pages for the given @zone and
1050 * @mem. This can be the global zone lruvec, if the memory controller
1053 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1054 struct mem_cgroup *memcg)
1056 struct mem_cgroup_per_zone *mz;
1057 struct lruvec *lruvec;
1059 if (mem_cgroup_disabled()) {
1060 lruvec = &zone->lruvec;
1064 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1065 lruvec = &mz->lruvec;
1068 * Since a node can be onlined after the mem_cgroup was created,
1069 * we have to be prepared to initialize lruvec->zone here;
1070 * and if offlined then reonlined, we need to reinitialize it.
1072 if (unlikely(lruvec->zone != zone))
1073 lruvec->zone = zone;
1078 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1080 * @zone: zone of the page
1082 * This function is only safe when following the LRU page isolation
1083 * and putback protocol: the LRU lock must be held, and the page must
1084 * either be PageLRU() or the caller must have isolated/allocated it.
1086 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1088 struct mem_cgroup_per_zone *mz;
1089 struct mem_cgroup *memcg;
1090 struct lruvec *lruvec;
1092 if (mem_cgroup_disabled()) {
1093 lruvec = &zone->lruvec;
1097 memcg = page->mem_cgroup;
1099 * Swapcache readahead pages are added to the LRU - and
1100 * possibly migrated - before they are charged.
1103 memcg = root_mem_cgroup;
1105 mz = mem_cgroup_page_zoneinfo(memcg, page);
1106 lruvec = &mz->lruvec;
1109 * Since a node can be onlined after the mem_cgroup was created,
1110 * we have to be prepared to initialize lruvec->zone here;
1111 * and if offlined then reonlined, we need to reinitialize it.
1113 if (unlikely(lruvec->zone != zone))
1114 lruvec->zone = zone;
1119 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1120 * @lruvec: mem_cgroup per zone lru vector
1121 * @lru: index of lru list the page is sitting on
1122 * @nr_pages: positive when adding or negative when removing
1124 * This function must be called when a page is added to or removed from an
1127 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1130 struct mem_cgroup_per_zone *mz;
1131 unsigned long *lru_size;
1133 if (mem_cgroup_disabled())
1136 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1137 lru_size = mz->lru_size + lru;
1138 *lru_size += nr_pages;
1139 VM_BUG_ON((long)(*lru_size) < 0);
1142 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1144 struct mem_cgroup *task_memcg;
1145 struct task_struct *p;
1148 p = find_lock_task_mm(task);
1150 task_memcg = get_mem_cgroup_from_mm(p->mm);
1154 * All threads may have already detached their mm's, but the oom
1155 * killer still needs to detect if they have already been oom
1156 * killed to prevent needlessly killing additional tasks.
1159 task_memcg = mem_cgroup_from_task(task);
1160 css_get(&task_memcg->css);
1163 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1164 css_put(&task_memcg->css);
1168 #define mem_cgroup_from_counter(counter, member) \
1169 container_of(counter, struct mem_cgroup, member)
1172 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1173 * @memcg: the memory cgroup
1175 * Returns the maximum amount of memory @mem can be charged with, in
1178 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1180 unsigned long margin = 0;
1181 unsigned long count;
1182 unsigned long limit;
1184 count = page_counter_read(&memcg->memory);
1185 limit = READ_ONCE(memcg->memory.limit);
1187 margin = limit - count;
1189 if (do_swap_account) {
1190 count = page_counter_read(&memcg->memsw);
1191 limit = READ_ONCE(memcg->memsw.limit);
1193 margin = min(margin, limit - count);
1200 * A routine for checking "mem" is under move_account() or not.
1202 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1203 * moving cgroups. This is for waiting at high-memory pressure
1206 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1208 struct mem_cgroup *from;
1209 struct mem_cgroup *to;
1212 * Unlike task_move routines, we access mc.to, mc.from not under
1213 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1215 spin_lock(&mc.lock);
1221 ret = mem_cgroup_is_descendant(from, memcg) ||
1222 mem_cgroup_is_descendant(to, memcg);
1224 spin_unlock(&mc.lock);
1228 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1230 if (mc.moving_task && current != mc.moving_task) {
1231 if (mem_cgroup_under_move(memcg)) {
1233 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1234 /* moving charge context might have finished. */
1237 finish_wait(&mc.waitq, &wait);
1244 #define K(x) ((x) << (PAGE_SHIFT-10))
1246 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1247 * @memcg: The memory cgroup that went over limit
1248 * @p: Task that is going to be killed
1250 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1253 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1255 /* oom_info_lock ensures that parallel ooms do not interleave */
1256 static DEFINE_MUTEX(oom_info_lock);
1257 struct mem_cgroup *iter;
1260 mutex_lock(&oom_info_lock);
1264 pr_info("Task in ");
1265 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1266 pr_cont(" killed as a result of limit of ");
1268 pr_info("Memory limit reached of cgroup ");
1271 pr_cont_cgroup_path(memcg->css.cgroup);
1276 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1277 K((u64)page_counter_read(&memcg->memory)),
1278 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1279 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1280 K((u64)page_counter_read(&memcg->memsw)),
1281 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1282 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1283 K((u64)page_counter_read(&memcg->kmem)),
1284 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1286 for_each_mem_cgroup_tree(iter, memcg) {
1287 pr_info("Memory cgroup stats for ");
1288 pr_cont_cgroup_path(iter->css.cgroup);
1291 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1292 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1294 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1295 K(mem_cgroup_read_stat(iter, i)));
1298 for (i = 0; i < NR_LRU_LISTS; i++)
1299 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1300 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1304 mutex_unlock(&oom_info_lock);
1308 * This function returns the number of memcg under hierarchy tree. Returns
1309 * 1(self count) if no children.
1311 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1314 struct mem_cgroup *iter;
1316 for_each_mem_cgroup_tree(iter, memcg)
1322 * Return the memory (and swap, if configured) limit for a memcg.
1324 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1326 unsigned long limit;
1328 limit = memcg->memory.limit;
1329 if (mem_cgroup_swappiness(memcg)) {
1330 unsigned long memsw_limit;
1332 memsw_limit = memcg->memsw.limit;
1333 limit = min(limit + total_swap_pages, memsw_limit);
1338 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1341 struct oom_control oc = {
1344 .gfp_mask = gfp_mask,
1347 struct mem_cgroup *iter;
1348 unsigned long chosen_points = 0;
1349 unsigned long totalpages;
1350 unsigned int points = 0;
1351 struct task_struct *chosen = NULL;
1353 mutex_lock(&oom_lock);
1356 * If current has a pending SIGKILL or is exiting, then automatically
1357 * select it. The goal is to allow it to allocate so that it may
1358 * quickly exit and free its memory.
1360 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1361 mark_oom_victim(current);
1365 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1366 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1367 for_each_mem_cgroup_tree(iter, memcg) {
1368 struct css_task_iter it;
1369 struct task_struct *task;
1371 css_task_iter_start(&iter->css, &it);
1372 while ((task = css_task_iter_next(&it))) {
1373 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1374 case OOM_SCAN_SELECT:
1376 put_task_struct(chosen);
1378 chosen_points = ULONG_MAX;
1379 get_task_struct(chosen);
1381 case OOM_SCAN_CONTINUE:
1383 case OOM_SCAN_ABORT:
1384 css_task_iter_end(&it);
1385 mem_cgroup_iter_break(memcg, iter);
1387 put_task_struct(chosen);
1392 points = oom_badness(task, memcg, NULL, totalpages);
1393 if (!points || points < chosen_points)
1395 /* Prefer thread group leaders for display purposes */
1396 if (points == chosen_points &&
1397 thread_group_leader(chosen))
1401 put_task_struct(chosen);
1403 chosen_points = points;
1404 get_task_struct(chosen);
1406 css_task_iter_end(&it);
1410 points = chosen_points * 1000 / totalpages;
1411 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1412 "Memory cgroup out of memory");
1415 mutex_unlock(&oom_lock);
1418 #if MAX_NUMNODES > 1
1421 * test_mem_cgroup_node_reclaimable
1422 * @memcg: the target memcg
1423 * @nid: the node ID to be checked.
1424 * @noswap : specify true here if the user wants flle only information.
1426 * This function returns whether the specified memcg contains any
1427 * reclaimable pages on a node. Returns true if there are any reclaimable
1428 * pages in the node.
1430 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1431 int nid, bool noswap)
1433 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1435 if (noswap || !total_swap_pages)
1437 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1444 * Always updating the nodemask is not very good - even if we have an empty
1445 * list or the wrong list here, we can start from some node and traverse all
1446 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1449 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1453 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1454 * pagein/pageout changes since the last update.
1456 if (!atomic_read(&memcg->numainfo_events))
1458 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1461 /* make a nodemask where this memcg uses memory from */
1462 memcg->scan_nodes = node_states[N_MEMORY];
1464 for_each_node_mask(nid, node_states[N_MEMORY]) {
1466 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1467 node_clear(nid, memcg->scan_nodes);
1470 atomic_set(&memcg->numainfo_events, 0);
1471 atomic_set(&memcg->numainfo_updating, 0);
1475 * Selecting a node where we start reclaim from. Because what we need is just
1476 * reducing usage counter, start from anywhere is O,K. Considering
1477 * memory reclaim from current node, there are pros. and cons.
1479 * Freeing memory from current node means freeing memory from a node which
1480 * we'll use or we've used. So, it may make LRU bad. And if several threads
1481 * hit limits, it will see a contention on a node. But freeing from remote
1482 * node means more costs for memory reclaim because of memory latency.
1484 * Now, we use round-robin. Better algorithm is welcomed.
1486 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1490 mem_cgroup_may_update_nodemask(memcg);
1491 node = memcg->last_scanned_node;
1493 node = next_node(node, memcg->scan_nodes);
1494 if (node == MAX_NUMNODES)
1495 node = first_node(memcg->scan_nodes);
1497 * We call this when we hit limit, not when pages are added to LRU.
1498 * No LRU may hold pages because all pages are UNEVICTABLE or
1499 * memcg is too small and all pages are not on LRU. In that case,
1500 * we use curret node.
1502 if (unlikely(node == MAX_NUMNODES))
1503 node = numa_node_id();
1505 memcg->last_scanned_node = node;
1509 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1515 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1518 unsigned long *total_scanned)
1520 struct mem_cgroup *victim = NULL;
1523 unsigned long excess;
1524 unsigned long nr_scanned;
1525 struct mem_cgroup_reclaim_cookie reclaim = {
1530 excess = soft_limit_excess(root_memcg);
1533 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1538 * If we have not been able to reclaim
1539 * anything, it might because there are
1540 * no reclaimable pages under this hierarchy
1545 * We want to do more targeted reclaim.
1546 * excess >> 2 is not to excessive so as to
1547 * reclaim too much, nor too less that we keep
1548 * coming back to reclaim from this cgroup
1550 if (total >= (excess >> 2) ||
1551 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1556 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1558 *total_scanned += nr_scanned;
1559 if (!soft_limit_excess(root_memcg))
1562 mem_cgroup_iter_break(root_memcg, victim);
1566 #ifdef CONFIG_LOCKDEP
1567 static struct lockdep_map memcg_oom_lock_dep_map = {
1568 .name = "memcg_oom_lock",
1572 static DEFINE_SPINLOCK(memcg_oom_lock);
1575 * Check OOM-Killer is already running under our hierarchy.
1576 * If someone is running, return false.
1578 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1580 struct mem_cgroup *iter, *failed = NULL;
1582 spin_lock(&memcg_oom_lock);
1584 for_each_mem_cgroup_tree(iter, memcg) {
1585 if (iter->oom_lock) {
1587 * this subtree of our hierarchy is already locked
1588 * so we cannot give a lock.
1591 mem_cgroup_iter_break(memcg, iter);
1594 iter->oom_lock = true;
1599 * OK, we failed to lock the whole subtree so we have
1600 * to clean up what we set up to the failing subtree
1602 for_each_mem_cgroup_tree(iter, memcg) {
1603 if (iter == failed) {
1604 mem_cgroup_iter_break(memcg, iter);
1607 iter->oom_lock = false;
1610 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1612 spin_unlock(&memcg_oom_lock);
1617 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1619 struct mem_cgroup *iter;
1621 spin_lock(&memcg_oom_lock);
1622 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1623 for_each_mem_cgroup_tree(iter, memcg)
1624 iter->oom_lock = false;
1625 spin_unlock(&memcg_oom_lock);
1628 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1630 struct mem_cgroup *iter;
1632 spin_lock(&memcg_oom_lock);
1633 for_each_mem_cgroup_tree(iter, memcg)
1635 spin_unlock(&memcg_oom_lock);
1638 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1640 struct mem_cgroup *iter;
1643 * When a new child is created while the hierarchy is under oom,
1644 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1646 spin_lock(&memcg_oom_lock);
1647 for_each_mem_cgroup_tree(iter, memcg)
1648 if (iter->under_oom > 0)
1650 spin_unlock(&memcg_oom_lock);
1653 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1655 struct oom_wait_info {
1656 struct mem_cgroup *memcg;
1660 static int memcg_oom_wake_function(wait_queue_t *wait,
1661 unsigned mode, int sync, void *arg)
1663 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1664 struct mem_cgroup *oom_wait_memcg;
1665 struct oom_wait_info *oom_wait_info;
1667 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1668 oom_wait_memcg = oom_wait_info->memcg;
1670 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1671 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1673 return autoremove_wake_function(wait, mode, sync, arg);
1676 static void memcg_oom_recover(struct mem_cgroup *memcg)
1679 * For the following lockless ->under_oom test, the only required
1680 * guarantee is that it must see the state asserted by an OOM when
1681 * this function is called as a result of userland actions
1682 * triggered by the notification of the OOM. This is trivially
1683 * achieved by invoking mem_cgroup_mark_under_oom() before
1684 * triggering notification.
1686 if (memcg && memcg->under_oom)
1687 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1690 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1692 if (!current->memcg_may_oom)
1695 * We are in the middle of the charge context here, so we
1696 * don't want to block when potentially sitting on a callstack
1697 * that holds all kinds of filesystem and mm locks.
1699 * Also, the caller may handle a failed allocation gracefully
1700 * (like optional page cache readahead) and so an OOM killer
1701 * invocation might not even be necessary.
1703 * That's why we don't do anything here except remember the
1704 * OOM context and then deal with it at the end of the page
1705 * fault when the stack is unwound, the locks are released,
1706 * and when we know whether the fault was overall successful.
1708 css_get(&memcg->css);
1709 current->memcg_in_oom = memcg;
1710 current->memcg_oom_gfp_mask = mask;
1711 current->memcg_oom_order = order;
1715 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1716 * @handle: actually kill/wait or just clean up the OOM state
1718 * This has to be called at the end of a page fault if the memcg OOM
1719 * handler was enabled.
1721 * Memcg supports userspace OOM handling where failed allocations must
1722 * sleep on a waitqueue until the userspace task resolves the
1723 * situation. Sleeping directly in the charge context with all kinds
1724 * of locks held is not a good idea, instead we remember an OOM state
1725 * in the task and mem_cgroup_oom_synchronize() has to be called at
1726 * the end of the page fault to complete the OOM handling.
1728 * Returns %true if an ongoing memcg OOM situation was detected and
1729 * completed, %false otherwise.
1731 bool mem_cgroup_oom_synchronize(bool handle)
1733 struct mem_cgroup *memcg = current->memcg_in_oom;
1734 struct oom_wait_info owait;
1737 /* OOM is global, do not handle */
1741 if (!handle || oom_killer_disabled)
1744 owait.memcg = memcg;
1745 owait.wait.flags = 0;
1746 owait.wait.func = memcg_oom_wake_function;
1747 owait.wait.private = current;
1748 INIT_LIST_HEAD(&owait.wait.task_list);
1750 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1751 mem_cgroup_mark_under_oom(memcg);
1753 locked = mem_cgroup_oom_trylock(memcg);
1756 mem_cgroup_oom_notify(memcg);
1758 if (locked && !memcg->oom_kill_disable) {
1759 mem_cgroup_unmark_under_oom(memcg);
1760 finish_wait(&memcg_oom_waitq, &owait.wait);
1761 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1762 current->memcg_oom_order);
1765 mem_cgroup_unmark_under_oom(memcg);
1766 finish_wait(&memcg_oom_waitq, &owait.wait);
1770 mem_cgroup_oom_unlock(memcg);
1772 * There is no guarantee that an OOM-lock contender
1773 * sees the wakeups triggered by the OOM kill
1774 * uncharges. Wake any sleepers explicitely.
1776 memcg_oom_recover(memcg);
1779 current->memcg_in_oom = NULL;
1780 css_put(&memcg->css);
1785 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1786 * @page: page that is going to change accounted state
1788 * This function must mark the beginning of an accounted page state
1789 * change to prevent double accounting when the page is concurrently
1790 * being moved to another memcg:
1792 * memcg = mem_cgroup_begin_page_stat(page);
1793 * if (TestClearPageState(page))
1794 * mem_cgroup_update_page_stat(memcg, state, -1);
1795 * mem_cgroup_end_page_stat(memcg);
1797 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1799 struct mem_cgroup *memcg;
1800 unsigned long flags;
1803 * The RCU lock is held throughout the transaction. The fast
1804 * path can get away without acquiring the memcg->move_lock
1805 * because page moving starts with an RCU grace period.
1807 * The RCU lock also protects the memcg from being freed when
1808 * the page state that is going to change is the only thing
1809 * preventing the page from being uncharged.
1810 * E.g. end-writeback clearing PageWriteback(), which allows
1811 * migration to go ahead and uncharge the page before the
1812 * account transaction might be complete.
1816 if (mem_cgroup_disabled())
1819 memcg = page->mem_cgroup;
1820 if (unlikely(!memcg))
1823 if (atomic_read(&memcg->moving_account) <= 0)
1826 spin_lock_irqsave(&memcg->move_lock, flags);
1827 if (memcg != page->mem_cgroup) {
1828 spin_unlock_irqrestore(&memcg->move_lock, flags);
1833 * When charge migration first begins, we can have locked and
1834 * unlocked page stat updates happening concurrently. Track
1835 * the task who has the lock for mem_cgroup_end_page_stat().
1837 memcg->move_lock_task = current;
1838 memcg->move_lock_flags = flags;
1842 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1845 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1846 * @memcg: the memcg that was accounted against
1848 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1850 if (memcg && memcg->move_lock_task == current) {
1851 unsigned long flags = memcg->move_lock_flags;
1853 memcg->move_lock_task = NULL;
1854 memcg->move_lock_flags = 0;
1856 spin_unlock_irqrestore(&memcg->move_lock, flags);
1861 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1864 * size of first charge trial. "32" comes from vmscan.c's magic value.
1865 * TODO: maybe necessary to use big numbers in big irons.
1867 #define CHARGE_BATCH 32U
1868 struct memcg_stock_pcp {
1869 struct mem_cgroup *cached; /* this never be root cgroup */
1870 unsigned int nr_pages;
1871 struct work_struct work;
1872 unsigned long flags;
1873 #define FLUSHING_CACHED_CHARGE 0
1875 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1876 static DEFINE_MUTEX(percpu_charge_mutex);
1879 * consume_stock: Try to consume stocked charge on this cpu.
1880 * @memcg: memcg to consume from.
1881 * @nr_pages: how many pages to charge.
1883 * The charges will only happen if @memcg matches the current cpu's memcg
1884 * stock, and at least @nr_pages are available in that stock. Failure to
1885 * service an allocation will refill the stock.
1887 * returns true if successful, false otherwise.
1889 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1891 struct memcg_stock_pcp *stock;
1894 if (nr_pages > CHARGE_BATCH)
1897 stock = &get_cpu_var(memcg_stock);
1898 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1899 stock->nr_pages -= nr_pages;
1902 put_cpu_var(memcg_stock);
1907 * Returns stocks cached in percpu and reset cached information.
1909 static void drain_stock(struct memcg_stock_pcp *stock)
1911 struct mem_cgroup *old = stock->cached;
1913 if (stock->nr_pages) {
1914 page_counter_uncharge(&old->memory, stock->nr_pages);
1915 if (do_swap_account)
1916 page_counter_uncharge(&old->memsw, stock->nr_pages);
1917 css_put_many(&old->css, stock->nr_pages);
1918 stock->nr_pages = 0;
1920 stock->cached = NULL;
1924 * This must be called under preempt disabled or must be called by
1925 * a thread which is pinned to local cpu.
1927 static void drain_local_stock(struct work_struct *dummy)
1929 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1931 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1935 * Cache charges(val) to local per_cpu area.
1936 * This will be consumed by consume_stock() function, later.
1938 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1940 struct memcg_stock_pcp *stock;
1941 int cpu = get_cpu_light();
1943 stock = &per_cpu(memcg_stock, cpu);
1945 if (stock->cached != memcg) { /* reset if necessary */
1947 stock->cached = memcg;
1949 stock->nr_pages += nr_pages;
1954 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1955 * of the hierarchy under it.
1957 static void drain_all_stock(struct mem_cgroup *root_memcg)
1961 /* If someone's already draining, avoid adding running more workers. */
1962 if (!mutex_trylock(&percpu_charge_mutex))
1964 /* Notify other cpus that system-wide "drain" is running */
1966 curcpu = get_cpu_light();
1967 for_each_online_cpu(cpu) {
1968 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1969 struct mem_cgroup *memcg;
1971 memcg = stock->cached;
1972 if (!memcg || !stock->nr_pages)
1974 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1976 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1978 drain_local_stock(&stock->work);
1980 schedule_work_on(cpu, &stock->work);
1985 mutex_unlock(&percpu_charge_mutex);
1988 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1989 unsigned long action,
1992 int cpu = (unsigned long)hcpu;
1993 struct memcg_stock_pcp *stock;
1995 if (action == CPU_ONLINE)
1998 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2001 stock = &per_cpu(memcg_stock, cpu);
2007 * Scheduled by try_charge() to be executed from the userland return path
2008 * and reclaims memory over the high limit.
2010 void mem_cgroup_handle_over_high(void)
2012 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2013 struct mem_cgroup *memcg, *pos;
2015 if (likely(!nr_pages))
2018 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2021 if (page_counter_read(&pos->memory) <= pos->high)
2023 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2024 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2025 } while ((pos = parent_mem_cgroup(pos)));
2027 css_put(&memcg->css);
2028 current->memcg_nr_pages_over_high = 0;
2031 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2032 unsigned int nr_pages)
2034 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2035 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2036 struct mem_cgroup *mem_over_limit;
2037 struct page_counter *counter;
2038 unsigned long nr_reclaimed;
2039 bool may_swap = true;
2040 bool drained = false;
2042 if (mem_cgroup_is_root(memcg))
2045 if (consume_stock(memcg, nr_pages))
2048 if (!do_swap_account ||
2049 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2050 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2052 if (do_swap_account)
2053 page_counter_uncharge(&memcg->memsw, batch);
2054 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2056 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2060 if (batch > nr_pages) {
2066 * Unlike in global OOM situations, memcg is not in a physical
2067 * memory shortage. Allow dying and OOM-killed tasks to
2068 * bypass the last charges so that they can exit quickly and
2069 * free their memory.
2071 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2072 fatal_signal_pending(current) ||
2073 current->flags & PF_EXITING))
2076 if (unlikely(task_in_memcg_oom(current)))
2079 if (!gfpflags_allow_blocking(gfp_mask))
2082 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2084 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2085 gfp_mask, may_swap);
2087 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2091 drain_all_stock(mem_over_limit);
2096 if (gfp_mask & __GFP_NORETRY)
2099 * Even though the limit is exceeded at this point, reclaim
2100 * may have been able to free some pages. Retry the charge
2101 * before killing the task.
2103 * Only for regular pages, though: huge pages are rather
2104 * unlikely to succeed so close to the limit, and we fall back
2105 * to regular pages anyway in case of failure.
2107 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2110 * At task move, charge accounts can be doubly counted. So, it's
2111 * better to wait until the end of task_move if something is going on.
2113 if (mem_cgroup_wait_acct_move(mem_over_limit))
2119 if (gfp_mask & __GFP_NOFAIL)
2122 if (fatal_signal_pending(current))
2125 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2127 mem_cgroup_oom(mem_over_limit, gfp_mask,
2128 get_order(nr_pages * PAGE_SIZE));
2130 if (!(gfp_mask & __GFP_NOFAIL))
2134 * The allocation either can't fail or will lead to more memory
2135 * being freed very soon. Allow memory usage go over the limit
2136 * temporarily by force charging it.
2138 page_counter_charge(&memcg->memory, nr_pages);
2139 if (do_swap_account)
2140 page_counter_charge(&memcg->memsw, nr_pages);
2141 css_get_many(&memcg->css, nr_pages);
2146 css_get_many(&memcg->css, batch);
2147 if (batch > nr_pages)
2148 refill_stock(memcg, batch - nr_pages);
2151 * If the hierarchy is above the normal consumption range, schedule
2152 * reclaim on returning to userland. We can perform reclaim here
2153 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2154 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2155 * not recorded as it most likely matches current's and won't
2156 * change in the meantime. As high limit is checked again before
2157 * reclaim, the cost of mismatch is negligible.
2160 if (page_counter_read(&memcg->memory) > memcg->high) {
2161 current->memcg_nr_pages_over_high += batch;
2162 set_notify_resume(current);
2165 } while ((memcg = parent_mem_cgroup(memcg)));
2170 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2172 if (mem_cgroup_is_root(memcg))
2175 page_counter_uncharge(&memcg->memory, nr_pages);
2176 if (do_swap_account)
2177 page_counter_uncharge(&memcg->memsw, nr_pages);
2179 css_put_many(&memcg->css, nr_pages);
2182 static void lock_page_lru(struct page *page, int *isolated)
2184 struct zone *zone = page_zone(page);
2186 spin_lock_irq(&zone->lru_lock);
2187 if (PageLRU(page)) {
2188 struct lruvec *lruvec;
2190 lruvec = mem_cgroup_page_lruvec(page, zone);
2192 del_page_from_lru_list(page, lruvec, page_lru(page));
2198 static void unlock_page_lru(struct page *page, int isolated)
2200 struct zone *zone = page_zone(page);
2203 struct lruvec *lruvec;
2205 lruvec = mem_cgroup_page_lruvec(page, zone);
2206 VM_BUG_ON_PAGE(PageLRU(page), page);
2208 add_page_to_lru_list(page, lruvec, page_lru(page));
2210 spin_unlock_irq(&zone->lru_lock);
2213 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2218 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2221 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2222 * may already be on some other mem_cgroup's LRU. Take care of it.
2225 lock_page_lru(page, &isolated);
2228 * Nobody should be changing or seriously looking at
2229 * page->mem_cgroup at this point:
2231 * - the page is uncharged
2233 * - the page is off-LRU
2235 * - an anonymous fault has exclusive page access, except for
2236 * a locked page table
2238 * - a page cache insertion, a swapin fault, or a migration
2239 * have the page locked
2241 page->mem_cgroup = memcg;
2244 unlock_page_lru(page, isolated);
2247 #ifdef CONFIG_MEMCG_KMEM
2248 static int memcg_alloc_cache_id(void)
2253 id = ida_simple_get(&memcg_cache_ida,
2254 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2258 if (id < memcg_nr_cache_ids)
2262 * There's no space for the new id in memcg_caches arrays,
2263 * so we have to grow them.
2265 down_write(&memcg_cache_ids_sem);
2267 size = 2 * (id + 1);
2268 if (size < MEMCG_CACHES_MIN_SIZE)
2269 size = MEMCG_CACHES_MIN_SIZE;
2270 else if (size > MEMCG_CACHES_MAX_SIZE)
2271 size = MEMCG_CACHES_MAX_SIZE;
2273 err = memcg_update_all_caches(size);
2275 err = memcg_update_all_list_lrus(size);
2277 memcg_nr_cache_ids = size;
2279 up_write(&memcg_cache_ids_sem);
2282 ida_simple_remove(&memcg_cache_ida, id);
2288 static void memcg_free_cache_id(int id)
2290 ida_simple_remove(&memcg_cache_ida, id);
2293 struct memcg_kmem_cache_create_work {
2294 struct mem_cgroup *memcg;
2295 struct kmem_cache *cachep;
2296 struct work_struct work;
2299 static void memcg_kmem_cache_create_func(struct work_struct *w)
2301 struct memcg_kmem_cache_create_work *cw =
2302 container_of(w, struct memcg_kmem_cache_create_work, work);
2303 struct mem_cgroup *memcg = cw->memcg;
2304 struct kmem_cache *cachep = cw->cachep;
2306 memcg_create_kmem_cache(memcg, cachep);
2308 css_put(&memcg->css);
2313 * Enqueue the creation of a per-memcg kmem_cache.
2315 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2316 struct kmem_cache *cachep)
2318 struct memcg_kmem_cache_create_work *cw;
2320 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2324 css_get(&memcg->css);
2327 cw->cachep = cachep;
2328 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2330 schedule_work(&cw->work);
2333 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2334 struct kmem_cache *cachep)
2337 * We need to stop accounting when we kmalloc, because if the
2338 * corresponding kmalloc cache is not yet created, the first allocation
2339 * in __memcg_schedule_kmem_cache_create will recurse.
2341 * However, it is better to enclose the whole function. Depending on
2342 * the debugging options enabled, INIT_WORK(), for instance, can
2343 * trigger an allocation. This too, will make us recurse. Because at
2344 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2345 * the safest choice is to do it like this, wrapping the whole function.
2347 current->memcg_kmem_skip_account = 1;
2348 __memcg_schedule_kmem_cache_create(memcg, cachep);
2349 current->memcg_kmem_skip_account = 0;
2353 * Return the kmem_cache we're supposed to use for a slab allocation.
2354 * We try to use the current memcg's version of the cache.
2356 * If the cache does not exist yet, if we are the first user of it,
2357 * we either create it immediately, if possible, or create it asynchronously
2359 * In the latter case, we will let the current allocation go through with
2360 * the original cache.
2362 * Can't be called in interrupt context or from kernel threads.
2363 * This function needs to be called with rcu_read_lock() held.
2365 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2367 struct mem_cgroup *memcg;
2368 struct kmem_cache *memcg_cachep;
2371 VM_BUG_ON(!is_root_cache(cachep));
2373 if (current->memcg_kmem_skip_account)
2376 memcg = get_mem_cgroup_from_mm(current->mm);
2377 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2381 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2382 if (likely(memcg_cachep))
2383 return memcg_cachep;
2386 * If we are in a safe context (can wait, and not in interrupt
2387 * context), we could be be predictable and return right away.
2388 * This would guarantee that the allocation being performed
2389 * already belongs in the new cache.
2391 * However, there are some clashes that can arrive from locking.
2392 * For instance, because we acquire the slab_mutex while doing
2393 * memcg_create_kmem_cache, this means no further allocation
2394 * could happen with the slab_mutex held. So it's better to
2397 memcg_schedule_kmem_cache_create(memcg, cachep);
2399 css_put(&memcg->css);
2403 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2405 if (!is_root_cache(cachep))
2406 css_put(&cachep->memcg_params.memcg->css);
2409 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2410 struct mem_cgroup *memcg)
2412 unsigned int nr_pages = 1 << order;
2413 struct page_counter *counter;
2416 if (!memcg_kmem_is_active(memcg))
2419 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2422 ret = try_charge(memcg, gfp, nr_pages);
2424 page_counter_uncharge(&memcg->kmem, nr_pages);
2428 page->mem_cgroup = memcg;
2433 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2435 struct mem_cgroup *memcg;
2438 memcg = get_mem_cgroup_from_mm(current->mm);
2439 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2440 css_put(&memcg->css);
2444 void __memcg_kmem_uncharge(struct page *page, int order)
2446 struct mem_cgroup *memcg = page->mem_cgroup;
2447 unsigned int nr_pages = 1 << order;
2452 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2454 page_counter_uncharge(&memcg->kmem, nr_pages);
2455 page_counter_uncharge(&memcg->memory, nr_pages);
2456 if (do_swap_account)
2457 page_counter_uncharge(&memcg->memsw, nr_pages);
2459 page->mem_cgroup = NULL;
2460 css_put_many(&memcg->css, nr_pages);
2462 #endif /* CONFIG_MEMCG_KMEM */
2464 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2467 * Because tail pages are not marked as "used", set it. We're under
2468 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2469 * charge/uncharge will be never happen and move_account() is done under
2470 * compound_lock(), so we don't have to take care of races.
2472 void mem_cgroup_split_huge_fixup(struct page *head)
2476 if (mem_cgroup_disabled())
2479 for (i = 1; i < HPAGE_PMD_NR; i++)
2480 head[i].mem_cgroup = head->mem_cgroup;
2482 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2485 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2487 #ifdef CONFIG_MEMCG_SWAP
2488 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2491 int val = (charge) ? 1 : -1;
2492 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2496 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2497 * @entry: swap entry to be moved
2498 * @from: mem_cgroup which the entry is moved from
2499 * @to: mem_cgroup which the entry is moved to
2501 * It succeeds only when the swap_cgroup's record for this entry is the same
2502 * as the mem_cgroup's id of @from.
2504 * Returns 0 on success, -EINVAL on failure.
2506 * The caller must have charged to @to, IOW, called page_counter_charge() about
2507 * both res and memsw, and called css_get().
2509 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2510 struct mem_cgroup *from, struct mem_cgroup *to)
2512 unsigned short old_id, new_id;
2514 old_id = mem_cgroup_id(from);
2515 new_id = mem_cgroup_id(to);
2517 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2518 mem_cgroup_swap_statistics(from, false);
2519 mem_cgroup_swap_statistics(to, true);
2525 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2526 struct mem_cgroup *from, struct mem_cgroup *to)
2532 static DEFINE_MUTEX(memcg_limit_mutex);
2534 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2535 unsigned long limit)
2537 unsigned long curusage;
2538 unsigned long oldusage;
2539 bool enlarge = false;
2544 * For keeping hierarchical_reclaim simple, how long we should retry
2545 * is depends on callers. We set our retry-count to be function
2546 * of # of children which we should visit in this loop.
2548 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2549 mem_cgroup_count_children(memcg);
2551 oldusage = page_counter_read(&memcg->memory);
2554 if (signal_pending(current)) {
2559 mutex_lock(&memcg_limit_mutex);
2560 if (limit > memcg->memsw.limit) {
2561 mutex_unlock(&memcg_limit_mutex);
2565 if (limit > memcg->memory.limit)
2567 ret = page_counter_limit(&memcg->memory, limit);
2568 mutex_unlock(&memcg_limit_mutex);
2573 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2575 curusage = page_counter_read(&memcg->memory);
2576 /* Usage is reduced ? */
2577 if (curusage >= oldusage)
2580 oldusage = curusage;
2581 } while (retry_count);
2583 if (!ret && enlarge)
2584 memcg_oom_recover(memcg);
2589 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2590 unsigned long limit)
2592 unsigned long curusage;
2593 unsigned long oldusage;
2594 bool enlarge = false;
2598 /* see mem_cgroup_resize_res_limit */
2599 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2600 mem_cgroup_count_children(memcg);
2602 oldusage = page_counter_read(&memcg->memsw);
2605 if (signal_pending(current)) {
2610 mutex_lock(&memcg_limit_mutex);
2611 if (limit < memcg->memory.limit) {
2612 mutex_unlock(&memcg_limit_mutex);
2616 if (limit > memcg->memsw.limit)
2618 ret = page_counter_limit(&memcg->memsw, limit);
2619 mutex_unlock(&memcg_limit_mutex);
2624 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2626 curusage = page_counter_read(&memcg->memsw);
2627 /* Usage is reduced ? */
2628 if (curusage >= oldusage)
2631 oldusage = curusage;
2632 } while (retry_count);
2634 if (!ret && enlarge)
2635 memcg_oom_recover(memcg);
2640 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2642 unsigned long *total_scanned)
2644 unsigned long nr_reclaimed = 0;
2645 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2646 unsigned long reclaimed;
2648 struct mem_cgroup_tree_per_zone *mctz;
2649 unsigned long excess;
2650 unsigned long nr_scanned;
2655 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2657 * This loop can run a while, specially if mem_cgroup's continuously
2658 * keep exceeding their soft limit and putting the system under
2665 mz = mem_cgroup_largest_soft_limit_node(mctz);
2670 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2671 gfp_mask, &nr_scanned);
2672 nr_reclaimed += reclaimed;
2673 *total_scanned += nr_scanned;
2674 spin_lock_irq(&mctz->lock);
2675 __mem_cgroup_remove_exceeded(mz, mctz);
2678 * If we failed to reclaim anything from this memory cgroup
2679 * it is time to move on to the next cgroup
2683 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2685 excess = soft_limit_excess(mz->memcg);
2687 * One school of thought says that we should not add
2688 * back the node to the tree if reclaim returns 0.
2689 * But our reclaim could return 0, simply because due
2690 * to priority we are exposing a smaller subset of
2691 * memory to reclaim from. Consider this as a longer
2694 /* If excess == 0, no tree ops */
2695 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2696 spin_unlock_irq(&mctz->lock);
2697 css_put(&mz->memcg->css);
2700 * Could not reclaim anything and there are no more
2701 * mem cgroups to try or we seem to be looping without
2702 * reclaiming anything.
2704 if (!nr_reclaimed &&
2706 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2708 } while (!nr_reclaimed);
2710 css_put(&next_mz->memcg->css);
2711 return nr_reclaimed;
2715 * Test whether @memcg has children, dead or alive. Note that this
2716 * function doesn't care whether @memcg has use_hierarchy enabled and
2717 * returns %true if there are child csses according to the cgroup
2718 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2720 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2725 * The lock does not prevent addition or deletion of children, but
2726 * it prevents a new child from being initialized based on this
2727 * parent in css_online(), so it's enough to decide whether
2728 * hierarchically inherited attributes can still be changed or not.
2730 lockdep_assert_held(&memcg_create_mutex);
2733 ret = css_next_child(NULL, &memcg->css);
2739 * Reclaims as many pages from the given memcg as possible and moves
2740 * the rest to the parent.
2742 * Caller is responsible for holding css reference for memcg.
2744 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2746 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2748 /* we call try-to-free pages for make this cgroup empty */
2749 lru_add_drain_all();
2750 /* try to free all pages in this cgroup */
2751 while (nr_retries && page_counter_read(&memcg->memory)) {
2754 if (signal_pending(current))
2757 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2761 /* maybe some writeback is necessary */
2762 congestion_wait(BLK_RW_ASYNC, HZ/10);
2770 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2771 char *buf, size_t nbytes,
2774 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2776 if (mem_cgroup_is_root(memcg))
2778 return mem_cgroup_force_empty(memcg) ?: nbytes;
2781 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2784 return mem_cgroup_from_css(css)->use_hierarchy;
2787 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2788 struct cftype *cft, u64 val)
2791 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2792 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2794 mutex_lock(&memcg_create_mutex);
2796 if (memcg->use_hierarchy == val)
2800 * If parent's use_hierarchy is set, we can't make any modifications
2801 * in the child subtrees. If it is unset, then the change can
2802 * occur, provided the current cgroup has no children.
2804 * For the root cgroup, parent_mem is NULL, we allow value to be
2805 * set if there are no children.
2807 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2808 (val == 1 || val == 0)) {
2809 if (!memcg_has_children(memcg))
2810 memcg->use_hierarchy = val;
2817 mutex_unlock(&memcg_create_mutex);
2822 static unsigned long tree_stat(struct mem_cgroup *memcg,
2823 enum mem_cgroup_stat_index idx)
2825 struct mem_cgroup *iter;
2826 unsigned long val = 0;
2828 for_each_mem_cgroup_tree(iter, memcg)
2829 val += mem_cgroup_read_stat(iter, idx);
2834 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2838 if (mem_cgroup_is_root(memcg)) {
2839 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2840 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2842 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2845 val = page_counter_read(&memcg->memory);
2847 val = page_counter_read(&memcg->memsw);
2860 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2863 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2864 struct page_counter *counter;
2866 switch (MEMFILE_TYPE(cft->private)) {
2868 counter = &memcg->memory;
2871 counter = &memcg->memsw;
2874 counter = &memcg->kmem;
2880 switch (MEMFILE_ATTR(cft->private)) {
2882 if (counter == &memcg->memory)
2883 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2884 if (counter == &memcg->memsw)
2885 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2886 return (u64)page_counter_read(counter) * PAGE_SIZE;
2888 return (u64)counter->limit * PAGE_SIZE;
2890 return (u64)counter->watermark * PAGE_SIZE;
2892 return counter->failcnt;
2893 case RES_SOFT_LIMIT:
2894 return (u64)memcg->soft_limit * PAGE_SIZE;
2900 #ifdef CONFIG_MEMCG_KMEM
2901 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2902 unsigned long nr_pages)
2907 BUG_ON(memcg->kmemcg_id >= 0);
2908 BUG_ON(memcg->kmem_acct_activated);
2909 BUG_ON(memcg->kmem_acct_active);
2912 * For simplicity, we won't allow this to be disabled. It also can't
2913 * be changed if the cgroup has children already, or if tasks had
2916 * If tasks join before we set the limit, a person looking at
2917 * kmem.usage_in_bytes will have no way to determine when it took
2918 * place, which makes the value quite meaningless.
2920 * After it first became limited, changes in the value of the limit are
2921 * of course permitted.
2923 mutex_lock(&memcg_create_mutex);
2924 if (cgroup_is_populated(memcg->css.cgroup) ||
2925 (memcg->use_hierarchy && memcg_has_children(memcg)))
2927 mutex_unlock(&memcg_create_mutex);
2931 memcg_id = memcg_alloc_cache_id();
2938 * We couldn't have accounted to this cgroup, because it hasn't got
2939 * activated yet, so this should succeed.
2941 err = page_counter_limit(&memcg->kmem, nr_pages);
2944 static_key_slow_inc(&memcg_kmem_enabled_key);
2946 * A memory cgroup is considered kmem-active as soon as it gets
2947 * kmemcg_id. Setting the id after enabling static branching will
2948 * guarantee no one starts accounting before all call sites are
2951 memcg->kmemcg_id = memcg_id;
2952 memcg->kmem_acct_activated = true;
2953 memcg->kmem_acct_active = true;
2958 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2959 unsigned long limit)
2963 mutex_lock(&memcg_limit_mutex);
2964 if (!memcg_kmem_is_active(memcg))
2965 ret = memcg_activate_kmem(memcg, limit);
2967 ret = page_counter_limit(&memcg->kmem, limit);
2968 mutex_unlock(&memcg_limit_mutex);
2972 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2975 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2980 mutex_lock(&memcg_limit_mutex);
2982 * If the parent cgroup is not kmem-active now, it cannot be activated
2983 * after this point, because it has at least one child already.
2985 if (memcg_kmem_is_active(parent))
2986 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2987 mutex_unlock(&memcg_limit_mutex);
2991 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2992 unsigned long limit)
2996 #endif /* CONFIG_MEMCG_KMEM */
2999 * The user of this function is...
3002 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3003 char *buf, size_t nbytes, loff_t off)
3005 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3006 unsigned long nr_pages;
3009 buf = strstrip(buf);
3010 ret = page_counter_memparse(buf, "-1", &nr_pages);
3014 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3016 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3020 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3022 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3025 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3028 ret = memcg_update_kmem_limit(memcg, nr_pages);
3032 case RES_SOFT_LIMIT:
3033 memcg->soft_limit = nr_pages;
3037 return ret ?: nbytes;
3040 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3041 size_t nbytes, loff_t off)
3043 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3044 struct page_counter *counter;
3046 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3048 counter = &memcg->memory;
3051 counter = &memcg->memsw;
3054 counter = &memcg->kmem;
3060 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3062 page_counter_reset_watermark(counter);
3065 counter->failcnt = 0;
3074 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3077 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3081 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3082 struct cftype *cft, u64 val)
3084 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3086 if (val & ~MOVE_MASK)
3090 * No kind of locking is needed in here, because ->can_attach() will
3091 * check this value once in the beginning of the process, and then carry
3092 * on with stale data. This means that changes to this value will only
3093 * affect task migrations starting after the change.
3095 memcg->move_charge_at_immigrate = val;
3099 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3100 struct cftype *cft, u64 val)
3107 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3111 unsigned int lru_mask;
3114 static const struct numa_stat stats[] = {
3115 { "total", LRU_ALL },
3116 { "file", LRU_ALL_FILE },
3117 { "anon", LRU_ALL_ANON },
3118 { "unevictable", BIT(LRU_UNEVICTABLE) },
3120 const struct numa_stat *stat;
3123 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3125 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3126 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3127 seq_printf(m, "%s=%lu", stat->name, nr);
3128 for_each_node_state(nid, N_MEMORY) {
3129 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3131 seq_printf(m, " N%d=%lu", nid, nr);
3136 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3137 struct mem_cgroup *iter;
3140 for_each_mem_cgroup_tree(iter, memcg)
3141 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3142 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3143 for_each_node_state(nid, N_MEMORY) {
3145 for_each_mem_cgroup_tree(iter, memcg)
3146 nr += mem_cgroup_node_nr_lru_pages(
3147 iter, nid, stat->lru_mask);
3148 seq_printf(m, " N%d=%lu", nid, nr);
3155 #endif /* CONFIG_NUMA */
3157 static int memcg_stat_show(struct seq_file *m, void *v)
3159 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3160 unsigned long memory, memsw;
3161 struct mem_cgroup *mi;
3164 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3165 MEM_CGROUP_STAT_NSTATS);
3166 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3167 MEM_CGROUP_EVENTS_NSTATS);
3168 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3170 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3171 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3173 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3174 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3177 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3178 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3179 mem_cgroup_read_events(memcg, i));
3181 for (i = 0; i < NR_LRU_LISTS; i++)
3182 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3183 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3185 /* Hierarchical information */
3186 memory = memsw = PAGE_COUNTER_MAX;
3187 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3188 memory = min(memory, mi->memory.limit);
3189 memsw = min(memsw, mi->memsw.limit);
3191 seq_printf(m, "hierarchical_memory_limit %llu\n",
3192 (u64)memory * PAGE_SIZE);
3193 if (do_swap_account)
3194 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3195 (u64)memsw * PAGE_SIZE);
3197 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3198 unsigned long long val = 0;
3200 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3202 for_each_mem_cgroup_tree(mi, memcg)
3203 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3204 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3207 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3208 unsigned long long val = 0;
3210 for_each_mem_cgroup_tree(mi, memcg)
3211 val += mem_cgroup_read_events(mi, i);
3212 seq_printf(m, "total_%s %llu\n",
3213 mem_cgroup_events_names[i], val);
3216 for (i = 0; i < NR_LRU_LISTS; i++) {
3217 unsigned long long val = 0;
3219 for_each_mem_cgroup_tree(mi, memcg)
3220 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3221 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3224 #ifdef CONFIG_DEBUG_VM
3227 struct mem_cgroup_per_zone *mz;
3228 struct zone_reclaim_stat *rstat;
3229 unsigned long recent_rotated[2] = {0, 0};
3230 unsigned long recent_scanned[2] = {0, 0};
3232 for_each_online_node(nid)
3233 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3234 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3235 rstat = &mz->lruvec.reclaim_stat;
3237 recent_rotated[0] += rstat->recent_rotated[0];
3238 recent_rotated[1] += rstat->recent_rotated[1];
3239 recent_scanned[0] += rstat->recent_scanned[0];
3240 recent_scanned[1] += rstat->recent_scanned[1];
3242 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3243 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3244 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3245 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3252 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3255 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3257 return mem_cgroup_swappiness(memcg);
3260 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3261 struct cftype *cft, u64 val)
3263 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3269 memcg->swappiness = val;
3271 vm_swappiness = val;
3276 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3278 struct mem_cgroup_threshold_ary *t;
3279 unsigned long usage;
3284 t = rcu_dereference(memcg->thresholds.primary);
3286 t = rcu_dereference(memcg->memsw_thresholds.primary);
3291 usage = mem_cgroup_usage(memcg, swap);
3294 * current_threshold points to threshold just below or equal to usage.
3295 * If it's not true, a threshold was crossed after last
3296 * call of __mem_cgroup_threshold().
3298 i = t->current_threshold;
3301 * Iterate backward over array of thresholds starting from
3302 * current_threshold and check if a threshold is crossed.
3303 * If none of thresholds below usage is crossed, we read
3304 * only one element of the array here.
3306 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3307 eventfd_signal(t->entries[i].eventfd, 1);
3309 /* i = current_threshold + 1 */
3313 * Iterate forward over array of thresholds starting from
3314 * current_threshold+1 and check if a threshold is crossed.
3315 * If none of thresholds above usage is crossed, we read
3316 * only one element of the array here.
3318 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3319 eventfd_signal(t->entries[i].eventfd, 1);
3321 /* Update current_threshold */
3322 t->current_threshold = i - 1;
3327 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3330 __mem_cgroup_threshold(memcg, false);
3331 if (do_swap_account)
3332 __mem_cgroup_threshold(memcg, true);
3334 memcg = parent_mem_cgroup(memcg);
3338 static int compare_thresholds(const void *a, const void *b)
3340 const struct mem_cgroup_threshold *_a = a;
3341 const struct mem_cgroup_threshold *_b = b;
3343 if (_a->threshold > _b->threshold)
3346 if (_a->threshold < _b->threshold)
3352 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3354 struct mem_cgroup_eventfd_list *ev;
3356 spin_lock(&memcg_oom_lock);
3358 list_for_each_entry(ev, &memcg->oom_notify, list)
3359 eventfd_signal(ev->eventfd, 1);
3361 spin_unlock(&memcg_oom_lock);
3365 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3367 struct mem_cgroup *iter;
3369 for_each_mem_cgroup_tree(iter, memcg)
3370 mem_cgroup_oom_notify_cb(iter);
3373 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3374 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3376 struct mem_cgroup_thresholds *thresholds;
3377 struct mem_cgroup_threshold_ary *new;
3378 unsigned long threshold;
3379 unsigned long usage;
3382 ret = page_counter_memparse(args, "-1", &threshold);
3386 mutex_lock(&memcg->thresholds_lock);
3389 thresholds = &memcg->thresholds;
3390 usage = mem_cgroup_usage(memcg, false);
3391 } else if (type == _MEMSWAP) {
3392 thresholds = &memcg->memsw_thresholds;
3393 usage = mem_cgroup_usage(memcg, true);
3397 /* Check if a threshold crossed before adding a new one */
3398 if (thresholds->primary)
3399 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3401 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3403 /* Allocate memory for new array of thresholds */
3404 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3412 /* Copy thresholds (if any) to new array */
3413 if (thresholds->primary) {
3414 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3415 sizeof(struct mem_cgroup_threshold));
3418 /* Add new threshold */
3419 new->entries[size - 1].eventfd = eventfd;
3420 new->entries[size - 1].threshold = threshold;
3422 /* Sort thresholds. Registering of new threshold isn't time-critical */
3423 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3424 compare_thresholds, NULL);
3426 /* Find current threshold */
3427 new->current_threshold = -1;
3428 for (i = 0; i < size; i++) {
3429 if (new->entries[i].threshold <= usage) {
3431 * new->current_threshold will not be used until
3432 * rcu_assign_pointer(), so it's safe to increment
3435 ++new->current_threshold;
3440 /* Free old spare buffer and save old primary buffer as spare */
3441 kfree(thresholds->spare);
3442 thresholds->spare = thresholds->primary;
3444 rcu_assign_pointer(thresholds->primary, new);
3446 /* To be sure that nobody uses thresholds */
3450 mutex_unlock(&memcg->thresholds_lock);
3455 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3456 struct eventfd_ctx *eventfd, const char *args)
3458 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3461 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3462 struct eventfd_ctx *eventfd, const char *args)
3464 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3467 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3468 struct eventfd_ctx *eventfd, enum res_type type)
3470 struct mem_cgroup_thresholds *thresholds;
3471 struct mem_cgroup_threshold_ary *new;
3472 unsigned long usage;
3475 mutex_lock(&memcg->thresholds_lock);
3478 thresholds = &memcg->thresholds;
3479 usage = mem_cgroup_usage(memcg, false);
3480 } else if (type == _MEMSWAP) {
3481 thresholds = &memcg->memsw_thresholds;
3482 usage = mem_cgroup_usage(memcg, true);
3486 if (!thresholds->primary)
3489 /* Check if a threshold crossed before removing */
3490 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3492 /* Calculate new number of threshold */
3494 for (i = 0; i < thresholds->primary->size; i++) {
3495 if (thresholds->primary->entries[i].eventfd != eventfd)
3499 new = thresholds->spare;
3501 /* Set thresholds array to NULL if we don't have thresholds */
3510 /* Copy thresholds and find current threshold */
3511 new->current_threshold = -1;
3512 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3513 if (thresholds->primary->entries[i].eventfd == eventfd)
3516 new->entries[j] = thresholds->primary->entries[i];
3517 if (new->entries[j].threshold <= usage) {
3519 * new->current_threshold will not be used
3520 * until rcu_assign_pointer(), so it's safe to increment
3523 ++new->current_threshold;
3529 /* Swap primary and spare array */
3530 thresholds->spare = thresholds->primary;
3532 rcu_assign_pointer(thresholds->primary, new);
3534 /* To be sure that nobody uses thresholds */
3537 /* If all events are unregistered, free the spare array */
3539 kfree(thresholds->spare);
3540 thresholds->spare = NULL;
3543 mutex_unlock(&memcg->thresholds_lock);
3546 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3547 struct eventfd_ctx *eventfd)
3549 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3552 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3553 struct eventfd_ctx *eventfd)
3555 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3558 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3559 struct eventfd_ctx *eventfd, const char *args)
3561 struct mem_cgroup_eventfd_list *event;
3563 event = kmalloc(sizeof(*event), GFP_KERNEL);
3567 spin_lock(&memcg_oom_lock);
3569 event->eventfd = eventfd;
3570 list_add(&event->list, &memcg->oom_notify);
3572 /* already in OOM ? */
3573 if (memcg->under_oom)
3574 eventfd_signal(eventfd, 1);
3575 spin_unlock(&memcg_oom_lock);
3580 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3581 struct eventfd_ctx *eventfd)
3583 struct mem_cgroup_eventfd_list *ev, *tmp;
3585 spin_lock(&memcg_oom_lock);
3587 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3588 if (ev->eventfd == eventfd) {
3589 list_del(&ev->list);
3594 spin_unlock(&memcg_oom_lock);
3597 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3599 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3601 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3602 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3606 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3607 struct cftype *cft, u64 val)
3609 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3611 /* cannot set to root cgroup and only 0 and 1 are allowed */
3612 if (!css->parent || !((val == 0) || (val == 1)))
3615 memcg->oom_kill_disable = val;
3617 memcg_oom_recover(memcg);
3622 #ifdef CONFIG_MEMCG_KMEM
3623 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3627 ret = memcg_propagate_kmem(memcg);
3631 return mem_cgroup_sockets_init(memcg, ss);
3634 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3636 struct cgroup_subsys_state *css;
3637 struct mem_cgroup *parent, *child;
3640 if (!memcg->kmem_acct_active)
3644 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3645 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3646 * guarantees no cache will be created for this cgroup after we are
3647 * done (see memcg_create_kmem_cache()).
3649 memcg->kmem_acct_active = false;
3651 memcg_deactivate_kmem_caches(memcg);
3653 kmemcg_id = memcg->kmemcg_id;
3654 BUG_ON(kmemcg_id < 0);
3656 parent = parent_mem_cgroup(memcg);
3658 parent = root_mem_cgroup;
3661 * Change kmemcg_id of this cgroup and all its descendants to the
3662 * parent's id, and then move all entries from this cgroup's list_lrus
3663 * to ones of the parent. After we have finished, all list_lrus
3664 * corresponding to this cgroup are guaranteed to remain empty. The
3665 * ordering is imposed by list_lru_node->lock taken by
3666 * memcg_drain_all_list_lrus().
3668 css_for_each_descendant_pre(css, &memcg->css) {
3669 child = mem_cgroup_from_css(css);
3670 BUG_ON(child->kmemcg_id != kmemcg_id);
3671 child->kmemcg_id = parent->kmemcg_id;
3672 if (!memcg->use_hierarchy)
3675 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3677 memcg_free_cache_id(kmemcg_id);
3680 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3682 if (memcg->kmem_acct_activated) {
3683 memcg_destroy_kmem_caches(memcg);
3684 static_key_slow_dec(&memcg_kmem_enabled_key);
3685 WARN_ON(page_counter_read(&memcg->kmem));
3687 mem_cgroup_sockets_destroy(memcg);
3690 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3695 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3699 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3704 #ifdef CONFIG_CGROUP_WRITEBACK
3706 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3708 return &memcg->cgwb_list;
3711 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3713 return wb_domain_init(&memcg->cgwb_domain, gfp);
3716 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3718 wb_domain_exit(&memcg->cgwb_domain);
3721 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3723 wb_domain_size_changed(&memcg->cgwb_domain);
3726 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3728 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3730 if (!memcg->css.parent)
3733 return &memcg->cgwb_domain;
3737 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3738 * @wb: bdi_writeback in question
3739 * @pfilepages: out parameter for number of file pages
3740 * @pheadroom: out parameter for number of allocatable pages according to memcg
3741 * @pdirty: out parameter for number of dirty pages
3742 * @pwriteback: out parameter for number of pages under writeback
3744 * Determine the numbers of file, headroom, dirty, and writeback pages in
3745 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3746 * is a bit more involved.
3748 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3749 * headroom is calculated as the lowest headroom of itself and the
3750 * ancestors. Note that this doesn't consider the actual amount of
3751 * available memory in the system. The caller should further cap
3752 * *@pheadroom accordingly.
3754 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3755 unsigned long *pheadroom, unsigned long *pdirty,
3756 unsigned long *pwriteback)
3758 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3759 struct mem_cgroup *parent;
3761 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3763 /* this should eventually include NR_UNSTABLE_NFS */
3764 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3765 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3766 (1 << LRU_ACTIVE_FILE));
3767 *pheadroom = PAGE_COUNTER_MAX;
3769 while ((parent = parent_mem_cgroup(memcg))) {
3770 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3771 unsigned long used = page_counter_read(&memcg->memory);
3773 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3778 #else /* CONFIG_CGROUP_WRITEBACK */
3780 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3785 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3789 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3793 #endif /* CONFIG_CGROUP_WRITEBACK */
3796 * DO NOT USE IN NEW FILES.
3798 * "cgroup.event_control" implementation.
3800 * This is way over-engineered. It tries to support fully configurable
3801 * events for each user. Such level of flexibility is completely
3802 * unnecessary especially in the light of the planned unified hierarchy.
3804 * Please deprecate this and replace with something simpler if at all
3809 * Unregister event and free resources.
3811 * Gets called from workqueue.
3813 static void memcg_event_remove(struct work_struct *work)
3815 struct mem_cgroup_event *event =
3816 container_of(work, struct mem_cgroup_event, remove);
3817 struct mem_cgroup *memcg = event->memcg;
3819 remove_wait_queue(event->wqh, &event->wait);
3821 event->unregister_event(memcg, event->eventfd);
3823 /* Notify userspace the event is going away. */
3824 eventfd_signal(event->eventfd, 1);
3826 eventfd_ctx_put(event->eventfd);
3828 css_put(&memcg->css);
3832 * Gets called on POLLHUP on eventfd when user closes it.
3834 * Called with wqh->lock held and interrupts disabled.
3836 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3837 int sync, void *key)
3839 struct mem_cgroup_event *event =
3840 container_of(wait, struct mem_cgroup_event, wait);
3841 struct mem_cgroup *memcg = event->memcg;
3842 unsigned long flags = (unsigned long)key;
3844 if (flags & POLLHUP) {
3846 * If the event has been detached at cgroup removal, we
3847 * can simply return knowing the other side will cleanup
3850 * We can't race against event freeing since the other
3851 * side will require wqh->lock via remove_wait_queue(),
3854 spin_lock(&memcg->event_list_lock);
3855 if (!list_empty(&event->list)) {
3856 list_del_init(&event->list);
3858 * We are in atomic context, but cgroup_event_remove()
3859 * may sleep, so we have to call it in workqueue.
3861 schedule_work(&event->remove);
3863 spin_unlock(&memcg->event_list_lock);
3869 static void memcg_event_ptable_queue_proc(struct file *file,
3870 wait_queue_head_t *wqh, poll_table *pt)
3872 struct mem_cgroup_event *event =
3873 container_of(pt, struct mem_cgroup_event, pt);
3876 add_wait_queue(wqh, &event->wait);
3880 * DO NOT USE IN NEW FILES.
3882 * Parse input and register new cgroup event handler.
3884 * Input must be in format '<event_fd> <control_fd> <args>'.
3885 * Interpretation of args is defined by control file implementation.
3887 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3888 char *buf, size_t nbytes, loff_t off)
3890 struct cgroup_subsys_state *css = of_css(of);
3891 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3892 struct mem_cgroup_event *event;
3893 struct cgroup_subsys_state *cfile_css;
3894 unsigned int efd, cfd;
3901 buf = strstrip(buf);
3903 efd = simple_strtoul(buf, &endp, 10);
3908 cfd = simple_strtoul(buf, &endp, 10);
3909 if ((*endp != ' ') && (*endp != '\0'))
3913 event = kzalloc(sizeof(*event), GFP_KERNEL);
3917 event->memcg = memcg;
3918 INIT_LIST_HEAD(&event->list);
3919 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3920 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3921 INIT_WORK(&event->remove, memcg_event_remove);
3929 event->eventfd = eventfd_ctx_fileget(efile.file);
3930 if (IS_ERR(event->eventfd)) {
3931 ret = PTR_ERR(event->eventfd);
3938 goto out_put_eventfd;
3941 /* the process need read permission on control file */
3942 /* AV: shouldn't we check that it's been opened for read instead? */
3943 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3948 * Determine the event callbacks and set them in @event. This used
3949 * to be done via struct cftype but cgroup core no longer knows
3950 * about these events. The following is crude but the whole thing
3951 * is for compatibility anyway.
3953 * DO NOT ADD NEW FILES.
3955 name = cfile.file->f_path.dentry->d_name.name;
3957 if (!strcmp(name, "memory.usage_in_bytes")) {
3958 event->register_event = mem_cgroup_usage_register_event;
3959 event->unregister_event = mem_cgroup_usage_unregister_event;
3960 } else if (!strcmp(name, "memory.oom_control")) {
3961 event->register_event = mem_cgroup_oom_register_event;
3962 event->unregister_event = mem_cgroup_oom_unregister_event;
3963 } else if (!strcmp(name, "memory.pressure_level")) {
3964 event->register_event = vmpressure_register_event;
3965 event->unregister_event = vmpressure_unregister_event;
3966 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3967 event->register_event = memsw_cgroup_usage_register_event;
3968 event->unregister_event = memsw_cgroup_usage_unregister_event;
3975 * Verify @cfile should belong to @css. Also, remaining events are
3976 * automatically removed on cgroup destruction but the removal is
3977 * asynchronous, so take an extra ref on @css.
3979 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3980 &memory_cgrp_subsys);
3982 if (IS_ERR(cfile_css))
3984 if (cfile_css != css) {
3989 ret = event->register_event(memcg, event->eventfd, buf);
3993 efile.file->f_op->poll(efile.file, &event->pt);
3995 spin_lock(&memcg->event_list_lock);
3996 list_add(&event->list, &memcg->event_list);
3997 spin_unlock(&memcg->event_list_lock);
4009 eventfd_ctx_put(event->eventfd);
4018 static struct cftype mem_cgroup_legacy_files[] = {
4020 .name = "usage_in_bytes",
4021 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4022 .read_u64 = mem_cgroup_read_u64,
4025 .name = "max_usage_in_bytes",
4026 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4027 .write = mem_cgroup_reset,
4028 .read_u64 = mem_cgroup_read_u64,
4031 .name = "limit_in_bytes",
4032 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4033 .write = mem_cgroup_write,
4034 .read_u64 = mem_cgroup_read_u64,
4037 .name = "soft_limit_in_bytes",
4038 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4039 .write = mem_cgroup_write,
4040 .read_u64 = mem_cgroup_read_u64,
4044 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4045 .write = mem_cgroup_reset,
4046 .read_u64 = mem_cgroup_read_u64,
4050 .seq_show = memcg_stat_show,
4053 .name = "force_empty",
4054 .write = mem_cgroup_force_empty_write,
4057 .name = "use_hierarchy",
4058 .write_u64 = mem_cgroup_hierarchy_write,
4059 .read_u64 = mem_cgroup_hierarchy_read,
4062 .name = "cgroup.event_control", /* XXX: for compat */
4063 .write = memcg_write_event_control,
4064 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4067 .name = "swappiness",
4068 .read_u64 = mem_cgroup_swappiness_read,
4069 .write_u64 = mem_cgroup_swappiness_write,
4072 .name = "move_charge_at_immigrate",
4073 .read_u64 = mem_cgroup_move_charge_read,
4074 .write_u64 = mem_cgroup_move_charge_write,
4077 .name = "oom_control",
4078 .seq_show = mem_cgroup_oom_control_read,
4079 .write_u64 = mem_cgroup_oom_control_write,
4080 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4083 .name = "pressure_level",
4087 .name = "numa_stat",
4088 .seq_show = memcg_numa_stat_show,
4091 #ifdef CONFIG_MEMCG_KMEM
4093 .name = "kmem.limit_in_bytes",
4094 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4095 .write = mem_cgroup_write,
4096 .read_u64 = mem_cgroup_read_u64,
4099 .name = "kmem.usage_in_bytes",
4100 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4101 .read_u64 = mem_cgroup_read_u64,
4104 .name = "kmem.failcnt",
4105 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4106 .write = mem_cgroup_reset,
4107 .read_u64 = mem_cgroup_read_u64,
4110 .name = "kmem.max_usage_in_bytes",
4111 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4112 .write = mem_cgroup_reset,
4113 .read_u64 = mem_cgroup_read_u64,
4115 #ifdef CONFIG_SLABINFO
4117 .name = "kmem.slabinfo",
4118 .seq_start = slab_start,
4119 .seq_next = slab_next,
4120 .seq_stop = slab_stop,
4121 .seq_show = memcg_slab_show,
4125 { }, /* terminate */
4128 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4130 struct mem_cgroup_per_node *pn;
4131 struct mem_cgroup_per_zone *mz;
4132 int zone, tmp = node;
4134 * This routine is called against possible nodes.
4135 * But it's BUG to call kmalloc() against offline node.
4137 * TODO: this routine can waste much memory for nodes which will
4138 * never be onlined. It's better to use memory hotplug callback
4141 if (!node_state(node, N_NORMAL_MEMORY))
4143 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4147 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4148 mz = &pn->zoneinfo[zone];
4149 lruvec_init(&mz->lruvec);
4150 mz->usage_in_excess = 0;
4151 mz->on_tree = false;
4154 memcg->nodeinfo[node] = pn;
4158 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4160 kfree(memcg->nodeinfo[node]);
4163 static struct mem_cgroup *mem_cgroup_alloc(void)
4165 struct mem_cgroup *memcg;
4168 size = sizeof(struct mem_cgroup);
4169 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4171 memcg = kzalloc(size, GFP_KERNEL);
4175 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4179 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4185 free_percpu(memcg->stat);
4192 * At destroying mem_cgroup, references from swap_cgroup can remain.
4193 * (scanning all at force_empty is too costly...)
4195 * Instead of clearing all references at force_empty, we remember
4196 * the number of reference from swap_cgroup and free mem_cgroup when
4197 * it goes down to 0.
4199 * Removal of cgroup itself succeeds regardless of refs from swap.
4202 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4206 mem_cgroup_remove_from_trees(memcg);
4209 free_mem_cgroup_per_zone_info(memcg, node);
4211 free_percpu(memcg->stat);
4212 memcg_wb_domain_exit(memcg);
4217 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4219 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4221 if (!memcg->memory.parent)
4223 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4225 EXPORT_SYMBOL(parent_mem_cgroup);
4227 static struct cgroup_subsys_state * __ref
4228 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4230 struct mem_cgroup *memcg;
4231 long error = -ENOMEM;
4234 memcg = mem_cgroup_alloc();
4236 return ERR_PTR(error);
4239 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4243 if (parent_css == NULL) {
4244 root_mem_cgroup = memcg;
4245 mem_cgroup_root_css = &memcg->css;
4246 page_counter_init(&memcg->memory, NULL);
4247 memcg->high = PAGE_COUNTER_MAX;
4248 memcg->soft_limit = PAGE_COUNTER_MAX;
4249 page_counter_init(&memcg->memsw, NULL);
4250 page_counter_init(&memcg->kmem, NULL);
4253 memcg->last_scanned_node = MAX_NUMNODES;
4254 INIT_LIST_HEAD(&memcg->oom_notify);
4255 memcg->move_charge_at_immigrate = 0;
4256 mutex_init(&memcg->thresholds_lock);
4257 spin_lock_init(&memcg->move_lock);
4258 vmpressure_init(&memcg->vmpressure);
4259 INIT_LIST_HEAD(&memcg->event_list);
4260 spin_lock_init(&memcg->event_list_lock);
4261 #ifdef CONFIG_MEMCG_KMEM
4262 memcg->kmemcg_id = -1;
4264 #ifdef CONFIG_CGROUP_WRITEBACK
4265 INIT_LIST_HEAD(&memcg->cgwb_list);
4270 __mem_cgroup_free(memcg);
4271 return ERR_PTR(error);
4275 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4277 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4278 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4281 if (css->id > MEM_CGROUP_ID_MAX)
4287 mutex_lock(&memcg_create_mutex);
4289 memcg->use_hierarchy = parent->use_hierarchy;
4290 memcg->oom_kill_disable = parent->oom_kill_disable;
4291 memcg->swappiness = mem_cgroup_swappiness(parent);
4293 if (parent->use_hierarchy) {
4294 page_counter_init(&memcg->memory, &parent->memory);
4295 memcg->high = PAGE_COUNTER_MAX;
4296 memcg->soft_limit = PAGE_COUNTER_MAX;
4297 page_counter_init(&memcg->memsw, &parent->memsw);
4298 page_counter_init(&memcg->kmem, &parent->kmem);
4301 * No need to take a reference to the parent because cgroup
4302 * core guarantees its existence.
4305 page_counter_init(&memcg->memory, NULL);
4306 memcg->high = PAGE_COUNTER_MAX;
4307 memcg->soft_limit = PAGE_COUNTER_MAX;
4308 page_counter_init(&memcg->memsw, NULL);
4309 page_counter_init(&memcg->kmem, NULL);
4311 * Deeper hierachy with use_hierarchy == false doesn't make
4312 * much sense so let cgroup subsystem know about this
4313 * unfortunate state in our controller.
4315 if (parent != root_mem_cgroup)
4316 memory_cgrp_subsys.broken_hierarchy = true;
4318 mutex_unlock(&memcg_create_mutex);
4320 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4325 * Make sure the memcg is initialized: mem_cgroup_iter()
4326 * orders reading memcg->initialized against its callers
4327 * reading the memcg members.
4329 smp_store_release(&memcg->initialized, 1);
4334 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4336 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4337 struct mem_cgroup_event *event, *tmp;
4340 * Unregister events and notify userspace.
4341 * Notify userspace about cgroup removing only after rmdir of cgroup
4342 * directory to avoid race between userspace and kernelspace.
4344 spin_lock(&memcg->event_list_lock);
4345 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4346 list_del_init(&event->list);
4347 schedule_work(&event->remove);
4349 spin_unlock(&memcg->event_list_lock);
4351 vmpressure_cleanup(&memcg->vmpressure);
4353 memcg_deactivate_kmem(memcg);
4355 wb_memcg_offline(memcg);
4358 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4360 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4362 invalidate_reclaim_iterators(memcg);
4365 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4367 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4369 memcg_destroy_kmem(memcg);
4370 __mem_cgroup_free(memcg);
4374 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4375 * @css: the target css
4377 * Reset the states of the mem_cgroup associated with @css. This is
4378 * invoked when the userland requests disabling on the default hierarchy
4379 * but the memcg is pinned through dependency. The memcg should stop
4380 * applying policies and should revert to the vanilla state as it may be
4381 * made visible again.
4383 * The current implementation only resets the essential configurations.
4384 * This needs to be expanded to cover all the visible parts.
4386 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4388 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4390 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4391 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4392 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4394 memcg->high = PAGE_COUNTER_MAX;
4395 memcg->soft_limit = PAGE_COUNTER_MAX;
4396 memcg_wb_domain_size_changed(memcg);
4400 /* Handlers for move charge at task migration. */
4401 static int mem_cgroup_do_precharge(unsigned long count)
4405 /* Try a single bulk charge without reclaim first, kswapd may wake */
4406 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4408 mc.precharge += count;
4412 /* Try charges one by one with reclaim */
4414 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4424 * get_mctgt_type - get target type of moving charge
4425 * @vma: the vma the pte to be checked belongs
4426 * @addr: the address corresponding to the pte to be checked
4427 * @ptent: the pte to be checked
4428 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4431 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4432 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4433 * move charge. if @target is not NULL, the page is stored in target->page
4434 * with extra refcnt got(Callers should handle it).
4435 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4436 * target for charge migration. if @target is not NULL, the entry is stored
4439 * Called with pte lock held.
4446 enum mc_target_type {
4452 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4453 unsigned long addr, pte_t ptent)
4455 struct page *page = vm_normal_page(vma, addr, ptent);
4457 if (!page || !page_mapped(page))
4459 if (PageAnon(page)) {
4460 if (!(mc.flags & MOVE_ANON))
4463 if (!(mc.flags & MOVE_FILE))
4466 if (!get_page_unless_zero(page))
4473 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4474 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4476 struct page *page = NULL;
4477 swp_entry_t ent = pte_to_swp_entry(ptent);
4479 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4482 * Because lookup_swap_cache() updates some statistics counter,
4483 * we call find_get_page() with swapper_space directly.
4485 page = find_get_page(swap_address_space(ent), ent.val);
4486 if (do_swap_account)
4487 entry->val = ent.val;
4492 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4493 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4499 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4500 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4502 struct page *page = NULL;
4503 struct address_space *mapping;
4506 if (!vma->vm_file) /* anonymous vma */
4508 if (!(mc.flags & MOVE_FILE))
4511 mapping = vma->vm_file->f_mapping;
4512 pgoff = linear_page_index(vma, addr);
4514 /* page is moved even if it's not RSS of this task(page-faulted). */
4516 /* shmem/tmpfs may report page out on swap: account for that too. */
4517 if (shmem_mapping(mapping)) {
4518 page = find_get_entry(mapping, pgoff);
4519 if (radix_tree_exceptional_entry(page)) {
4520 swp_entry_t swp = radix_to_swp_entry(page);
4521 if (do_swap_account)
4523 page = find_get_page(swap_address_space(swp), swp.val);
4526 page = find_get_page(mapping, pgoff);
4528 page = find_get_page(mapping, pgoff);
4534 * mem_cgroup_move_account - move account of the page
4536 * @nr_pages: number of regular pages (>1 for huge pages)
4537 * @from: mem_cgroup which the page is moved from.
4538 * @to: mem_cgroup which the page is moved to. @from != @to.
4540 * The caller must confirm following.
4541 * - page is not on LRU (isolate_page() is useful.)
4542 * - compound_lock is held when nr_pages > 1
4544 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4547 static int mem_cgroup_move_account(struct page *page,
4548 unsigned int nr_pages,
4549 struct mem_cgroup *from,
4550 struct mem_cgroup *to)
4552 unsigned long flags;
4556 VM_BUG_ON(from == to);
4557 VM_BUG_ON_PAGE(PageLRU(page), page);
4559 * The page is isolated from LRU. So, collapse function
4560 * will not handle this page. But page splitting can happen.
4561 * Do this check under compound_page_lock(). The caller should
4565 if (nr_pages > 1 && !PageTransHuge(page))
4569 * Prevent mem_cgroup_replace_page() from looking at
4570 * page->mem_cgroup of its source page while we change it.
4572 if (!trylock_page(page))
4576 if (page->mem_cgroup != from)
4579 anon = PageAnon(page);
4581 spin_lock_irqsave(&from->move_lock, flags);
4583 if (!anon && page_mapped(page)) {
4584 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4586 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4591 * move_lock grabbed above and caller set from->moving_account, so
4592 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4593 * So mapping should be stable for dirty pages.
4595 if (!anon && PageDirty(page)) {
4596 struct address_space *mapping = page_mapping(page);
4598 if (mapping_cap_account_dirty(mapping)) {
4599 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4601 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4606 if (PageWriteback(page)) {
4607 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4609 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4614 * It is safe to change page->mem_cgroup here because the page
4615 * is referenced, charged, and isolated - we can't race with
4616 * uncharging, charging, migration, or LRU putback.
4619 /* caller should have done css_get */
4620 page->mem_cgroup = to;
4621 spin_unlock_irqrestore(&from->move_lock, flags);
4625 local_lock_irq(event_lock);
4626 mem_cgroup_charge_statistics(to, page, nr_pages);
4627 memcg_check_events(to, page);
4628 mem_cgroup_charge_statistics(from, page, -nr_pages);
4629 memcg_check_events(from, page);
4630 local_unlock_irq(event_lock);
4637 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4638 unsigned long addr, pte_t ptent, union mc_target *target)
4640 struct page *page = NULL;
4641 enum mc_target_type ret = MC_TARGET_NONE;
4642 swp_entry_t ent = { .val = 0 };
4644 if (pte_present(ptent))
4645 page = mc_handle_present_pte(vma, addr, ptent);
4646 else if (is_swap_pte(ptent))
4647 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4648 else if (pte_none(ptent))
4649 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4651 if (!page && !ent.val)
4655 * Do only loose check w/o serialization.
4656 * mem_cgroup_move_account() checks the page is valid or
4657 * not under LRU exclusion.
4659 if (page->mem_cgroup == mc.from) {
4660 ret = MC_TARGET_PAGE;
4662 target->page = page;
4664 if (!ret || !target)
4667 /* There is a swap entry and a page doesn't exist or isn't charged */
4668 if (ent.val && !ret &&
4669 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4670 ret = MC_TARGET_SWAP;
4677 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4679 * We don't consider swapping or file mapped pages because THP does not
4680 * support them for now.
4681 * Caller should make sure that pmd_trans_huge(pmd) is true.
4683 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4684 unsigned long addr, pmd_t pmd, union mc_target *target)
4686 struct page *page = NULL;
4687 enum mc_target_type ret = MC_TARGET_NONE;
4689 page = pmd_page(pmd);
4690 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4691 if (!(mc.flags & MOVE_ANON))
4693 if (page->mem_cgroup == mc.from) {
4694 ret = MC_TARGET_PAGE;
4697 target->page = page;
4703 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4704 unsigned long addr, pmd_t pmd, union mc_target *target)
4706 return MC_TARGET_NONE;
4710 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4711 unsigned long addr, unsigned long end,
4712 struct mm_walk *walk)
4714 struct vm_area_struct *vma = walk->vma;
4718 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4719 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4720 mc.precharge += HPAGE_PMD_NR;
4725 if (pmd_trans_unstable(pmd))
4727 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4728 for (; addr != end; pte++, addr += PAGE_SIZE)
4729 if (get_mctgt_type(vma, addr, *pte, NULL))
4730 mc.precharge++; /* increment precharge temporarily */
4731 pte_unmap_unlock(pte - 1, ptl);
4737 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4739 unsigned long precharge;
4741 struct mm_walk mem_cgroup_count_precharge_walk = {
4742 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4745 down_read(&mm->mmap_sem);
4746 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4747 up_read(&mm->mmap_sem);
4749 precharge = mc.precharge;
4755 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4757 unsigned long precharge = mem_cgroup_count_precharge(mm);
4759 VM_BUG_ON(mc.moving_task);
4760 mc.moving_task = current;
4761 return mem_cgroup_do_precharge(precharge);
4764 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4765 static void __mem_cgroup_clear_mc(void)
4767 struct mem_cgroup *from = mc.from;
4768 struct mem_cgroup *to = mc.to;
4770 /* we must uncharge all the leftover precharges from mc.to */
4772 cancel_charge(mc.to, mc.precharge);
4776 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4777 * we must uncharge here.
4779 if (mc.moved_charge) {
4780 cancel_charge(mc.from, mc.moved_charge);
4781 mc.moved_charge = 0;
4783 /* we must fixup refcnts and charges */
4784 if (mc.moved_swap) {
4785 /* uncharge swap account from the old cgroup */
4786 if (!mem_cgroup_is_root(mc.from))
4787 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4790 * we charged both to->memory and to->memsw, so we
4791 * should uncharge to->memory.
4793 if (!mem_cgroup_is_root(mc.to))
4794 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4796 css_put_many(&mc.from->css, mc.moved_swap);
4798 /* we've already done css_get(mc.to) */
4801 memcg_oom_recover(from);
4802 memcg_oom_recover(to);
4803 wake_up_all(&mc.waitq);
4806 static void mem_cgroup_clear_mc(void)
4809 * we must clear moving_task before waking up waiters at the end of
4812 mc.moving_task = NULL;
4813 __mem_cgroup_clear_mc();
4814 spin_lock(&mc.lock);
4817 spin_unlock(&mc.lock);
4820 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4822 struct cgroup_subsys_state *css;
4823 struct mem_cgroup *memcg;
4824 struct mem_cgroup *from;
4825 struct task_struct *leader, *p;
4826 struct mm_struct *mm;
4827 unsigned long move_flags;
4830 /* charge immigration isn't supported on the default hierarchy */
4831 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4835 * Multi-process migrations only happen on the default hierarchy
4836 * where charge immigration is not used. Perform charge
4837 * immigration if @tset contains a leader and whine if there are
4841 cgroup_taskset_for_each_leader(leader, css, tset) {
4844 memcg = mem_cgroup_from_css(css);
4850 * We are now commited to this value whatever it is. Changes in this
4851 * tunable will only affect upcoming migrations, not the current one.
4852 * So we need to save it, and keep it going.
4854 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4858 from = mem_cgroup_from_task(p);
4860 VM_BUG_ON(from == memcg);
4862 mm = get_task_mm(p);
4865 /* We move charges only when we move a owner of the mm */
4866 if (mm->owner == p) {
4869 VM_BUG_ON(mc.precharge);
4870 VM_BUG_ON(mc.moved_charge);
4871 VM_BUG_ON(mc.moved_swap);
4873 spin_lock(&mc.lock);
4876 mc.flags = move_flags;
4877 spin_unlock(&mc.lock);
4878 /* We set mc.moving_task later */
4880 ret = mem_cgroup_precharge_mc(mm);
4882 mem_cgroup_clear_mc();
4888 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4891 mem_cgroup_clear_mc();
4894 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4895 unsigned long addr, unsigned long end,
4896 struct mm_walk *walk)
4899 struct vm_area_struct *vma = walk->vma;
4902 enum mc_target_type target_type;
4903 union mc_target target;
4907 * We don't take compound_lock() here but no race with splitting thp
4909 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4910 * under splitting, which means there's no concurrent thp split,
4911 * - if another thread runs into split_huge_page() just after we
4912 * entered this if-block, the thread must wait for page table lock
4913 * to be unlocked in __split_huge_page_splitting(), where the main
4914 * part of thp split is not executed yet.
4916 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4917 if (mc.precharge < HPAGE_PMD_NR) {
4921 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4922 if (target_type == MC_TARGET_PAGE) {
4924 if (!isolate_lru_page(page)) {
4925 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4927 mc.precharge -= HPAGE_PMD_NR;
4928 mc.moved_charge += HPAGE_PMD_NR;
4930 putback_lru_page(page);
4938 if (pmd_trans_unstable(pmd))
4941 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4942 for (; addr != end; addr += PAGE_SIZE) {
4943 pte_t ptent = *(pte++);
4949 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4950 case MC_TARGET_PAGE:
4952 if (isolate_lru_page(page))
4954 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4956 /* we uncharge from mc.from later. */
4959 putback_lru_page(page);
4960 put: /* get_mctgt_type() gets the page */
4963 case MC_TARGET_SWAP:
4965 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4967 /* we fixup refcnts and charges later. */
4975 pte_unmap_unlock(pte - 1, ptl);
4980 * We have consumed all precharges we got in can_attach().
4981 * We try charge one by one, but don't do any additional
4982 * charges to mc.to if we have failed in charge once in attach()
4985 ret = mem_cgroup_do_precharge(1);
4993 static void mem_cgroup_move_charge(struct mm_struct *mm)
4995 struct mm_walk mem_cgroup_move_charge_walk = {
4996 .pmd_entry = mem_cgroup_move_charge_pte_range,
5000 lru_add_drain_all();
5002 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5003 * move_lock while we're moving its pages to another memcg.
5004 * Then wait for already started RCU-only updates to finish.
5006 atomic_inc(&mc.from->moving_account);
5009 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5011 * Someone who are holding the mmap_sem might be waiting in
5012 * waitq. So we cancel all extra charges, wake up all waiters,
5013 * and retry. Because we cancel precharges, we might not be able
5014 * to move enough charges, but moving charge is a best-effort
5015 * feature anyway, so it wouldn't be a big problem.
5017 __mem_cgroup_clear_mc();
5022 * When we have consumed all precharges and failed in doing
5023 * additional charge, the page walk just aborts.
5025 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5026 up_read(&mm->mmap_sem);
5027 atomic_dec(&mc.from->moving_account);
5030 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5032 struct cgroup_subsys_state *css;
5033 struct task_struct *p = cgroup_taskset_first(tset, &css);
5034 struct mm_struct *mm = get_task_mm(p);
5038 mem_cgroup_move_charge(mm);
5042 mem_cgroup_clear_mc();
5044 #else /* !CONFIG_MMU */
5045 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5049 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5052 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5058 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5059 * to verify whether we're attached to the default hierarchy on each mount
5062 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5065 * use_hierarchy is forced on the default hierarchy. cgroup core
5066 * guarantees that @root doesn't have any children, so turning it
5067 * on for the root memcg is enough.
5069 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5070 root_mem_cgroup->use_hierarchy = true;
5072 root_mem_cgroup->use_hierarchy = false;
5075 static u64 memory_current_read(struct cgroup_subsys_state *css,
5078 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5080 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5083 static int memory_low_show(struct seq_file *m, void *v)
5085 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5086 unsigned long low = READ_ONCE(memcg->low);
5088 if (low == PAGE_COUNTER_MAX)
5089 seq_puts(m, "max\n");
5091 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5096 static ssize_t memory_low_write(struct kernfs_open_file *of,
5097 char *buf, size_t nbytes, loff_t off)
5099 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5103 buf = strstrip(buf);
5104 err = page_counter_memparse(buf, "max", &low);
5113 static int memory_high_show(struct seq_file *m, void *v)
5115 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5116 unsigned long high = READ_ONCE(memcg->high);
5118 if (high == PAGE_COUNTER_MAX)
5119 seq_puts(m, "max\n");
5121 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5126 static ssize_t memory_high_write(struct kernfs_open_file *of,
5127 char *buf, size_t nbytes, loff_t off)
5129 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5133 buf = strstrip(buf);
5134 err = page_counter_memparse(buf, "max", &high);
5140 memcg_wb_domain_size_changed(memcg);
5144 static int memory_max_show(struct seq_file *m, void *v)
5146 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5147 unsigned long max = READ_ONCE(memcg->memory.limit);
5149 if (max == PAGE_COUNTER_MAX)
5150 seq_puts(m, "max\n");
5152 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5157 static ssize_t memory_max_write(struct kernfs_open_file *of,
5158 char *buf, size_t nbytes, loff_t off)
5160 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5164 buf = strstrip(buf);
5165 err = page_counter_memparse(buf, "max", &max);
5169 err = mem_cgroup_resize_limit(memcg, max);
5173 memcg_wb_domain_size_changed(memcg);
5177 static int memory_events_show(struct seq_file *m, void *v)
5179 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5181 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5182 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5183 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5184 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5189 static struct cftype memory_files[] = {
5192 .flags = CFTYPE_NOT_ON_ROOT,
5193 .read_u64 = memory_current_read,
5197 .flags = CFTYPE_NOT_ON_ROOT,
5198 .seq_show = memory_low_show,
5199 .write = memory_low_write,
5203 .flags = CFTYPE_NOT_ON_ROOT,
5204 .seq_show = memory_high_show,
5205 .write = memory_high_write,
5209 .flags = CFTYPE_NOT_ON_ROOT,
5210 .seq_show = memory_max_show,
5211 .write = memory_max_write,
5215 .flags = CFTYPE_NOT_ON_ROOT,
5216 .file_offset = offsetof(struct mem_cgroup, events_file),
5217 .seq_show = memory_events_show,
5222 struct cgroup_subsys memory_cgrp_subsys = {
5223 .css_alloc = mem_cgroup_css_alloc,
5224 .css_online = mem_cgroup_css_online,
5225 .css_offline = mem_cgroup_css_offline,
5226 .css_released = mem_cgroup_css_released,
5227 .css_free = mem_cgroup_css_free,
5228 .css_reset = mem_cgroup_css_reset,
5229 .can_attach = mem_cgroup_can_attach,
5230 .cancel_attach = mem_cgroup_cancel_attach,
5231 .attach = mem_cgroup_move_task,
5232 .bind = mem_cgroup_bind,
5233 .dfl_cftypes = memory_files,
5234 .legacy_cftypes = mem_cgroup_legacy_files,
5239 * mem_cgroup_low - check if memory consumption is below the normal range
5240 * @root: the highest ancestor to consider
5241 * @memcg: the memory cgroup to check
5243 * Returns %true if memory consumption of @memcg, and that of all
5244 * configurable ancestors up to @root, is below the normal range.
5246 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5248 if (mem_cgroup_disabled())
5252 * The toplevel group doesn't have a configurable range, so
5253 * it's never low when looked at directly, and it is not
5254 * considered an ancestor when assessing the hierarchy.
5257 if (memcg == root_mem_cgroup)
5260 if (page_counter_read(&memcg->memory) >= memcg->low)
5263 while (memcg != root) {
5264 memcg = parent_mem_cgroup(memcg);
5266 if (memcg == root_mem_cgroup)
5269 if (page_counter_read(&memcg->memory) >= memcg->low)
5276 * mem_cgroup_try_charge - try charging a page
5277 * @page: page to charge
5278 * @mm: mm context of the victim
5279 * @gfp_mask: reclaim mode
5280 * @memcgp: charged memcg return
5282 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5283 * pages according to @gfp_mask if necessary.
5285 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5286 * Otherwise, an error code is returned.
5288 * After page->mapping has been set up, the caller must finalize the
5289 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5290 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5292 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5293 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5295 struct mem_cgroup *memcg = NULL;
5296 unsigned int nr_pages = 1;
5299 if (mem_cgroup_disabled())
5302 if (PageSwapCache(page)) {
5304 * Every swap fault against a single page tries to charge the
5305 * page, bail as early as possible. shmem_unuse() encounters
5306 * already charged pages, too. The USED bit is protected by
5307 * the page lock, which serializes swap cache removal, which
5308 * in turn serializes uncharging.
5310 VM_BUG_ON_PAGE(!PageLocked(page), page);
5311 if (page->mem_cgroup)
5314 if (do_swap_account) {
5315 swp_entry_t ent = { .val = page_private(page), };
5316 unsigned short id = lookup_swap_cgroup_id(ent);
5319 memcg = mem_cgroup_from_id(id);
5320 if (memcg && !css_tryget_online(&memcg->css))
5326 if (PageTransHuge(page)) {
5327 nr_pages <<= compound_order(page);
5328 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5332 memcg = get_mem_cgroup_from_mm(mm);
5334 ret = try_charge(memcg, gfp_mask, nr_pages);
5336 css_put(&memcg->css);
5343 * mem_cgroup_commit_charge - commit a page charge
5344 * @page: page to charge
5345 * @memcg: memcg to charge the page to
5346 * @lrucare: page might be on LRU already
5348 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5349 * after page->mapping has been set up. This must happen atomically
5350 * as part of the page instantiation, i.e. under the page table lock
5351 * for anonymous pages, under the page lock for page and swap cache.
5353 * In addition, the page must not be on the LRU during the commit, to
5354 * prevent racing with task migration. If it might be, use @lrucare.
5356 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5358 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5361 unsigned int nr_pages = 1;
5363 VM_BUG_ON_PAGE(!page->mapping, page);
5364 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5366 if (mem_cgroup_disabled())
5369 * Swap faults will attempt to charge the same page multiple
5370 * times. But reuse_swap_page() might have removed the page
5371 * from swapcache already, so we can't check PageSwapCache().
5376 commit_charge(page, memcg, lrucare);
5378 if (PageTransHuge(page)) {
5379 nr_pages <<= compound_order(page);
5380 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5383 local_lock_irq(event_lock);
5384 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5385 memcg_check_events(memcg, page);
5386 local_unlock_irq(event_lock);
5388 if (do_swap_account && PageSwapCache(page)) {
5389 swp_entry_t entry = { .val = page_private(page) };
5391 * The swap entry might not get freed for a long time,
5392 * let's not wait for it. The page already received a
5393 * memory+swap charge, drop the swap entry duplicate.
5395 mem_cgroup_uncharge_swap(entry);
5400 * mem_cgroup_cancel_charge - cancel a page charge
5401 * @page: page to charge
5402 * @memcg: memcg to charge the page to
5404 * Cancel a charge transaction started by mem_cgroup_try_charge().
5406 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5408 unsigned int nr_pages = 1;
5410 if (mem_cgroup_disabled())
5413 * Swap faults will attempt to charge the same page multiple
5414 * times. But reuse_swap_page() might have removed the page
5415 * from swapcache already, so we can't check PageSwapCache().
5420 if (PageTransHuge(page)) {
5421 nr_pages <<= compound_order(page);
5422 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5425 cancel_charge(memcg, nr_pages);
5428 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5429 unsigned long nr_anon, unsigned long nr_file,
5430 unsigned long nr_huge, struct page *dummy_page)
5432 unsigned long nr_pages = nr_anon + nr_file;
5433 unsigned long flags;
5435 if (!mem_cgroup_is_root(memcg)) {
5436 page_counter_uncharge(&memcg->memory, nr_pages);
5437 if (do_swap_account)
5438 page_counter_uncharge(&memcg->memsw, nr_pages);
5439 memcg_oom_recover(memcg);
5442 local_lock_irqsave(event_lock, flags);
5443 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5444 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5445 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5446 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5447 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5448 memcg_check_events(memcg, dummy_page);
5449 local_unlock_irqrestore(event_lock, flags);
5451 if (!mem_cgroup_is_root(memcg))
5452 css_put_many(&memcg->css, nr_pages);
5455 static void uncharge_list(struct list_head *page_list)
5457 struct mem_cgroup *memcg = NULL;
5458 unsigned long nr_anon = 0;
5459 unsigned long nr_file = 0;
5460 unsigned long nr_huge = 0;
5461 unsigned long pgpgout = 0;
5462 struct list_head *next;
5465 next = page_list->next;
5467 unsigned int nr_pages = 1;
5469 page = list_entry(next, struct page, lru);
5470 next = page->lru.next;
5472 VM_BUG_ON_PAGE(PageLRU(page), page);
5473 VM_BUG_ON_PAGE(page_count(page), page);
5475 if (!page->mem_cgroup)
5479 * Nobody should be changing or seriously looking at
5480 * page->mem_cgroup at this point, we have fully
5481 * exclusive access to the page.
5484 if (memcg != page->mem_cgroup) {
5486 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5488 pgpgout = nr_anon = nr_file = nr_huge = 0;
5490 memcg = page->mem_cgroup;
5493 if (PageTransHuge(page)) {
5494 nr_pages <<= compound_order(page);
5495 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5496 nr_huge += nr_pages;
5500 nr_anon += nr_pages;
5502 nr_file += nr_pages;
5504 page->mem_cgroup = NULL;
5507 } while (next != page_list);
5510 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5515 * mem_cgroup_uncharge - uncharge a page
5516 * @page: page to uncharge
5518 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5519 * mem_cgroup_commit_charge().
5521 void mem_cgroup_uncharge(struct page *page)
5523 if (mem_cgroup_disabled())
5526 /* Don't touch page->lru of any random page, pre-check: */
5527 if (!page->mem_cgroup)
5530 INIT_LIST_HEAD(&page->lru);
5531 uncharge_list(&page->lru);
5535 * mem_cgroup_uncharge_list - uncharge a list of page
5536 * @page_list: list of pages to uncharge
5538 * Uncharge a list of pages previously charged with
5539 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5541 void mem_cgroup_uncharge_list(struct list_head *page_list)
5543 if (mem_cgroup_disabled())
5546 if (!list_empty(page_list))
5547 uncharge_list(page_list);
5551 * mem_cgroup_replace_page - migrate a charge to another page
5552 * @oldpage: currently charged page
5553 * @newpage: page to transfer the charge to
5555 * Migrate the charge from @oldpage to @newpage.
5557 * Both pages must be locked, @newpage->mapping must be set up.
5558 * Either or both pages might be on the LRU already.
5560 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5562 struct mem_cgroup *memcg;
5565 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5566 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5567 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5568 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5571 if (mem_cgroup_disabled())
5574 /* Page cache replacement: new page already charged? */
5575 if (newpage->mem_cgroup)
5578 /* Swapcache readahead pages can get replaced before being charged */
5579 memcg = oldpage->mem_cgroup;
5583 lock_page_lru(oldpage, &isolated);
5584 oldpage->mem_cgroup = NULL;
5585 unlock_page_lru(oldpage, isolated);
5587 commit_charge(newpage, memcg, true);
5591 * subsys_initcall() for memory controller.
5593 * Some parts like hotcpu_notifier() have to be initialized from this context
5594 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5595 * everything that doesn't depend on a specific mem_cgroup structure should
5596 * be initialized from here.
5598 static int __init mem_cgroup_init(void)
5602 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5604 for_each_possible_cpu(cpu)
5605 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5608 for_each_node(node) {
5609 struct mem_cgroup_tree_per_node *rtpn;
5612 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5613 node_online(node) ? node : NUMA_NO_NODE);
5615 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5616 struct mem_cgroup_tree_per_zone *rtpz;
5618 rtpz = &rtpn->rb_tree_per_zone[zone];
5619 rtpz->rb_root = RB_ROOT;
5620 spin_lock_init(&rtpz->lock);
5622 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5627 subsys_initcall(mem_cgroup_init);
5629 #ifdef CONFIG_MEMCG_SWAP
5631 * mem_cgroup_swapout - transfer a memsw charge to swap
5632 * @page: page whose memsw charge to transfer
5633 * @entry: swap entry to move the charge to
5635 * Transfer the memsw charge of @page to @entry.
5637 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5639 struct mem_cgroup *memcg;
5640 unsigned short oldid;
5641 unsigned long flags;
5643 VM_BUG_ON_PAGE(PageLRU(page), page);
5644 VM_BUG_ON_PAGE(page_count(page), page);
5646 if (!do_swap_account)
5649 memcg = page->mem_cgroup;
5651 /* Readahead page, never charged */
5655 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5656 VM_BUG_ON_PAGE(oldid, page);
5657 mem_cgroup_swap_statistics(memcg, true);
5659 page->mem_cgroup = NULL;
5661 if (!mem_cgroup_is_root(memcg))
5662 page_counter_uncharge(&memcg->memory, 1);
5665 * Interrupts should be disabled here because the caller holds the
5666 * mapping->tree_lock lock which is taken with interrupts-off. It is
5667 * important here to have the interrupts disabled because it is the
5668 * only synchronisation we have for udpating the per-CPU variables.
5670 local_lock_irqsave(event_lock, flags);
5671 #ifndef CONFIG_PREEMPT_RT_BASE
5672 VM_BUG_ON(!irqs_disabled());
5674 mem_cgroup_charge_statistics(memcg, page, -1);
5675 memcg_check_events(memcg, page);
5676 local_unlock_irqrestore(event_lock, flags);
5680 * mem_cgroup_uncharge_swap - uncharge a swap entry
5681 * @entry: swap entry to uncharge
5683 * Drop the memsw charge associated with @entry.
5685 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5687 struct mem_cgroup *memcg;
5690 if (!do_swap_account)
5693 id = swap_cgroup_record(entry, 0);
5695 memcg = mem_cgroup_from_id(id);
5697 if (!mem_cgroup_is_root(memcg))
5698 page_counter_uncharge(&memcg->memsw, 1);
5699 mem_cgroup_swap_statistics(memcg, false);
5700 css_put(&memcg->css);
5705 /* for remember boot option*/
5706 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5707 static int really_do_swap_account __initdata = 1;
5709 static int really_do_swap_account __initdata;
5712 static int __init enable_swap_account(char *s)
5714 if (!strcmp(s, "1"))
5715 really_do_swap_account = 1;
5716 else if (!strcmp(s, "0"))
5717 really_do_swap_account = 0;
5720 __setup("swapaccount=", enable_swap_account);
5722 static struct cftype memsw_cgroup_files[] = {
5724 .name = "memsw.usage_in_bytes",
5725 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5726 .read_u64 = mem_cgroup_read_u64,
5729 .name = "memsw.max_usage_in_bytes",
5730 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5731 .write = mem_cgroup_reset,
5732 .read_u64 = mem_cgroup_read_u64,
5735 .name = "memsw.limit_in_bytes",
5736 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5737 .write = mem_cgroup_write,
5738 .read_u64 = mem_cgroup_read_u64,
5741 .name = "memsw.failcnt",
5742 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5743 .write = mem_cgroup_reset,
5744 .read_u64 = mem_cgroup_read_u64,
5746 { }, /* terminate */
5749 static int __init mem_cgroup_swap_init(void)
5751 if (!mem_cgroup_disabled() && really_do_swap_account) {
5752 do_swap_account = 1;
5753 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5754 memsw_cgroup_files));
5758 subsys_initcall(mem_cgroup_swap_init);
5760 #endif /* CONFIG_MEMCG_SWAP */