1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
69 #include <net/tcp_memcontrol.h>
72 #include <asm/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
77 EXPORT_SYMBOL(memory_cgrp_subsys);
79 #define MEM_CGROUP_RECLAIM_RETRIES 5
80 static struct mem_cgroup *root_mem_cgroup __read_mostly;
81 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
83 /* Whether the swap controller is active */
84 #ifdef CONFIG_MEMCG_SWAP
85 int do_swap_account __read_mostly;
87 #define do_swap_account 0
90 static const char * const mem_cgroup_stat_names[] = {
100 static const char * const mem_cgroup_events_names[] = {
107 static const char * const mem_cgroup_lru_names[] = {
115 #define THRESHOLDS_EVENTS_TARGET 128
116 #define SOFTLIMIT_EVENTS_TARGET 1024
117 #define NUMAINFO_EVENTS_TARGET 1024
120 * Cgroups above their limits are maintained in a RB-Tree, independent of
121 * their hierarchy representation
124 struct mem_cgroup_tree_per_zone {
125 struct rb_root rb_root;
129 struct mem_cgroup_tree_per_node {
130 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
133 struct mem_cgroup_tree {
134 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
137 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
140 struct mem_cgroup_eventfd_list {
141 struct list_head list;
142 struct eventfd_ctx *eventfd;
146 * cgroup_event represents events which userspace want to receive.
148 struct mem_cgroup_event {
150 * memcg which the event belongs to.
152 struct mem_cgroup *memcg;
154 * eventfd to signal userspace about the event.
156 struct eventfd_ctx *eventfd;
158 * Each of these stored in a list by the cgroup.
160 struct list_head list;
162 * register_event() callback will be used to add new userspace
163 * waiter for changes related to this event. Use eventfd_signal()
164 * on eventfd to send notification to userspace.
166 int (*register_event)(struct mem_cgroup *memcg,
167 struct eventfd_ctx *eventfd, const char *args);
169 * unregister_event() callback will be called when userspace closes
170 * the eventfd or on cgroup removing. This callback must be set,
171 * if you want provide notification functionality.
173 void (*unregister_event)(struct mem_cgroup *memcg,
174 struct eventfd_ctx *eventfd);
176 * All fields below needed to unregister event when
177 * userspace closes eventfd.
180 wait_queue_head_t *wqh;
182 struct work_struct remove;
185 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
186 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
188 /* Stuffs for move charges at task migration. */
190 * Types of charges to be moved.
192 #define MOVE_ANON 0x1U
193 #define MOVE_FILE 0x2U
194 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
196 /* "mc" and its members are protected by cgroup_mutex */
197 static struct move_charge_struct {
198 spinlock_t lock; /* for from, to */
199 struct mem_cgroup *from;
200 struct mem_cgroup *to;
202 unsigned long precharge;
203 unsigned long moved_charge;
204 unsigned long moved_swap;
205 struct task_struct *moving_task; /* a task moving charges */
206 wait_queue_head_t waitq; /* a waitq for other context */
208 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
209 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
213 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
214 * limit reclaim to prevent infinite loops, if they ever occur.
216 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
217 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
220 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
221 MEM_CGROUP_CHARGE_TYPE_ANON,
222 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
223 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
227 /* for encoding cft->private value on file */
235 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
236 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
237 #define MEMFILE_ATTR(val) ((val) & 0xffff)
238 /* Used for OOM nofiier */
239 #define OOM_CONTROL (0)
242 * The memcg_create_mutex will be held whenever a new cgroup is created.
243 * As a consequence, any change that needs to protect against new child cgroups
244 * appearing has to hold it as well.
246 static DEFINE_MUTEX(memcg_create_mutex);
248 /* Some nice accessors for the vmpressure. */
249 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
252 memcg = root_mem_cgroup;
253 return &memcg->vmpressure;
256 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
258 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
261 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
263 return (memcg == root_mem_cgroup);
267 * We restrict the id in the range of [1, 65535], so it can fit into
270 #define MEM_CGROUP_ID_MAX USHRT_MAX
272 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
274 return memcg->css.id;
278 * A helper function to get mem_cgroup from ID. must be called under
279 * rcu_read_lock(). The caller is responsible for calling
280 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
281 * refcnt from swap can be called against removed memcg.)
283 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
285 struct cgroup_subsys_state *css;
287 css = css_from_id(id, &memory_cgrp_subsys);
288 return mem_cgroup_from_css(css);
291 /* Writing them here to avoid exposing memcg's inner layout */
292 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
294 void sock_update_memcg(struct sock *sk)
296 if (mem_cgroup_sockets_enabled) {
297 struct mem_cgroup *memcg;
298 struct cg_proto *cg_proto;
300 BUG_ON(!sk->sk_prot->proto_cgroup);
302 /* Socket cloning can throw us here with sk_cgrp already
303 * filled. It won't however, necessarily happen from
304 * process context. So the test for root memcg given
305 * the current task's memcg won't help us in this case.
307 * Respecting the original socket's memcg is a better
308 * decision in this case.
311 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
312 css_get(&sk->sk_cgrp->memcg->css);
317 memcg = mem_cgroup_from_task(current);
318 cg_proto = sk->sk_prot->proto_cgroup(memcg);
319 if (cg_proto && test_bit(MEMCG_SOCK_ACTIVE, &cg_proto->flags) &&
320 css_tryget_online(&memcg->css)) {
321 sk->sk_cgrp = cg_proto;
326 EXPORT_SYMBOL(sock_update_memcg);
328 void sock_release_memcg(struct sock *sk)
330 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
331 struct mem_cgroup *memcg;
332 WARN_ON(!sk->sk_cgrp->memcg);
333 memcg = sk->sk_cgrp->memcg;
334 css_put(&sk->sk_cgrp->memcg->css);
338 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
340 if (!memcg || mem_cgroup_is_root(memcg))
343 return &memcg->tcp_mem;
345 EXPORT_SYMBOL(tcp_proto_cgroup);
349 #ifdef CONFIG_MEMCG_KMEM
351 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 struct static_key memcg_kmem_enabled_key;
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
401 #endif /* CONFIG_MEMCG_KMEM */
403 static struct mem_cgroup_per_zone *
404 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
406 int nid = zone_to_nid(zone);
407 int zid = zone_idx(zone);
409 return &memcg->nodeinfo[nid]->zoneinfo[zid];
413 * mem_cgroup_css_from_page - css of the memcg associated with a page
414 * @page: page of interest
416 * If memcg is bound to the default hierarchy, css of the memcg associated
417 * with @page is returned. The returned css remains associated with @page
418 * until it is released.
420 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
423 * XXX: The above description of behavior on the default hierarchy isn't
424 * strictly true yet as replace_page_cache_page() can modify the
425 * association before @page is released even on the default hierarchy;
426 * however, the current and planned usages don't mix the the two functions
427 * and replace_page_cache_page() will soon be updated to make the invariant
430 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
432 struct mem_cgroup *memcg;
436 memcg = page->mem_cgroup;
438 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
439 memcg = root_mem_cgroup;
446 * page_cgroup_ino - return inode number of the memcg a page is charged to
449 * Look up the closest online ancestor of the memory cgroup @page is charged to
450 * and return its inode number or 0 if @page is not charged to any cgroup. It
451 * is safe to call this function without holding a reference to @page.
453 * Note, this function is inherently racy, because there is nothing to prevent
454 * the cgroup inode from getting torn down and potentially reallocated a moment
455 * after page_cgroup_ino() returns, so it only should be used by callers that
456 * do not care (such as procfs interfaces).
458 ino_t page_cgroup_ino(struct page *page)
460 struct mem_cgroup *memcg;
461 unsigned long ino = 0;
464 memcg = READ_ONCE(page->mem_cgroup);
465 while (memcg && !(memcg->css.flags & CSS_ONLINE))
466 memcg = parent_mem_cgroup(memcg);
468 ino = cgroup_ino(memcg->css.cgroup);
473 static struct mem_cgroup_per_zone *
474 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
476 int nid = page_to_nid(page);
477 int zid = page_zonenum(page);
479 return &memcg->nodeinfo[nid]->zoneinfo[zid];
482 static struct mem_cgroup_tree_per_zone *
483 soft_limit_tree_node_zone(int nid, int zid)
485 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
488 static struct mem_cgroup_tree_per_zone *
489 soft_limit_tree_from_page(struct page *page)
491 int nid = page_to_nid(page);
492 int zid = page_zonenum(page);
494 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
497 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
498 struct mem_cgroup_tree_per_zone *mctz,
499 unsigned long new_usage_in_excess)
501 struct rb_node **p = &mctz->rb_root.rb_node;
502 struct rb_node *parent = NULL;
503 struct mem_cgroup_per_zone *mz_node;
508 mz->usage_in_excess = new_usage_in_excess;
509 if (!mz->usage_in_excess)
513 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
515 if (mz->usage_in_excess < mz_node->usage_in_excess)
518 * We can't avoid mem cgroups that are over their soft
519 * limit by the same amount
521 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
524 rb_link_node(&mz->tree_node, parent, p);
525 rb_insert_color(&mz->tree_node, &mctz->rb_root);
529 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
530 struct mem_cgroup_tree_per_zone *mctz)
534 rb_erase(&mz->tree_node, &mctz->rb_root);
538 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
539 struct mem_cgroup_tree_per_zone *mctz)
543 spin_lock_irqsave(&mctz->lock, flags);
544 __mem_cgroup_remove_exceeded(mz, mctz);
545 spin_unlock_irqrestore(&mctz->lock, flags);
548 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
550 unsigned long nr_pages = page_counter_read(&memcg->memory);
551 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
552 unsigned long excess = 0;
554 if (nr_pages > soft_limit)
555 excess = nr_pages - soft_limit;
560 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
562 unsigned long excess;
563 struct mem_cgroup_per_zone *mz;
564 struct mem_cgroup_tree_per_zone *mctz;
566 mctz = soft_limit_tree_from_page(page);
568 * Necessary to update all ancestors when hierarchy is used.
569 * because their event counter is not touched.
571 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
572 mz = mem_cgroup_page_zoneinfo(memcg, page);
573 excess = soft_limit_excess(memcg);
575 * We have to update the tree if mz is on RB-tree or
576 * mem is over its softlimit.
578 if (excess || mz->on_tree) {
581 spin_lock_irqsave(&mctz->lock, flags);
582 /* if on-tree, remove it */
584 __mem_cgroup_remove_exceeded(mz, mctz);
586 * Insert again. mz->usage_in_excess will be updated.
587 * If excess is 0, no tree ops.
589 __mem_cgroup_insert_exceeded(mz, mctz, excess);
590 spin_unlock_irqrestore(&mctz->lock, flags);
595 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
597 struct mem_cgroup_tree_per_zone *mctz;
598 struct mem_cgroup_per_zone *mz;
602 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
603 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
604 mctz = soft_limit_tree_node_zone(nid, zid);
605 mem_cgroup_remove_exceeded(mz, mctz);
610 static struct mem_cgroup_per_zone *
611 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
613 struct rb_node *rightmost = NULL;
614 struct mem_cgroup_per_zone *mz;
618 rightmost = rb_last(&mctz->rb_root);
620 goto done; /* Nothing to reclaim from */
622 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
624 * Remove the node now but someone else can add it back,
625 * we will to add it back at the end of reclaim to its correct
626 * position in the tree.
628 __mem_cgroup_remove_exceeded(mz, mctz);
629 if (!soft_limit_excess(mz->memcg) ||
630 !css_tryget_online(&mz->memcg->css))
636 static struct mem_cgroup_per_zone *
637 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
639 struct mem_cgroup_per_zone *mz;
641 spin_lock_irq(&mctz->lock);
642 mz = __mem_cgroup_largest_soft_limit_node(mctz);
643 spin_unlock_irq(&mctz->lock);
648 * Return page count for single (non recursive) @memcg.
650 * Implementation Note: reading percpu statistics for memcg.
652 * Both of vmstat[] and percpu_counter has threshold and do periodic
653 * synchronization to implement "quick" read. There are trade-off between
654 * reading cost and precision of value. Then, we may have a chance to implement
655 * a periodic synchronization of counter in memcg's counter.
657 * But this _read() function is used for user interface now. The user accounts
658 * memory usage by memory cgroup and he _always_ requires exact value because
659 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
660 * have to visit all online cpus and make sum. So, for now, unnecessary
661 * synchronization is not implemented. (just implemented for cpu hotplug)
663 * If there are kernel internal actions which can make use of some not-exact
664 * value, and reading all cpu value can be performance bottleneck in some
665 * common workload, threshold and synchronization as vmstat[] should be
669 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
674 /* Per-cpu values can be negative, use a signed accumulator */
675 for_each_possible_cpu(cpu)
676 val += per_cpu(memcg->stat->count[idx], cpu);
678 * Summing races with updates, so val may be negative. Avoid exposing
679 * transient negative values.
686 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
687 enum mem_cgroup_events_index idx)
689 unsigned long val = 0;
692 for_each_possible_cpu(cpu)
693 val += per_cpu(memcg->stat->events[idx], cpu);
697 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
702 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
703 * counted as CACHE even if it's on ANON LRU.
706 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
709 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
712 if (PageTransHuge(page))
713 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
716 /* pagein of a big page is an event. So, ignore page size */
718 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
720 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
721 nr_pages = -nr_pages; /* for event */
724 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
727 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
729 unsigned int lru_mask)
731 unsigned long nr = 0;
734 VM_BUG_ON((unsigned)nid >= nr_node_ids);
736 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
737 struct mem_cgroup_per_zone *mz;
741 if (!(BIT(lru) & lru_mask))
743 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
744 nr += mz->lru_size[lru];
750 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
751 unsigned int lru_mask)
753 unsigned long nr = 0;
756 for_each_node_state(nid, N_MEMORY)
757 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
761 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
762 enum mem_cgroup_events_target target)
764 unsigned long val, next;
766 val = __this_cpu_read(memcg->stat->nr_page_events);
767 next = __this_cpu_read(memcg->stat->targets[target]);
768 /* from time_after() in jiffies.h */
769 if ((long)next - (long)val < 0) {
771 case MEM_CGROUP_TARGET_THRESH:
772 next = val + THRESHOLDS_EVENTS_TARGET;
774 case MEM_CGROUP_TARGET_SOFTLIMIT:
775 next = val + SOFTLIMIT_EVENTS_TARGET;
777 case MEM_CGROUP_TARGET_NUMAINFO:
778 next = val + NUMAINFO_EVENTS_TARGET;
783 __this_cpu_write(memcg->stat->targets[target], next);
790 * Check events in order.
793 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
795 /* threshold event is triggered in finer grain than soft limit */
796 if (unlikely(mem_cgroup_event_ratelimit(memcg,
797 MEM_CGROUP_TARGET_THRESH))) {
799 bool do_numainfo __maybe_unused;
801 do_softlimit = mem_cgroup_event_ratelimit(memcg,
802 MEM_CGROUP_TARGET_SOFTLIMIT);
804 do_numainfo = mem_cgroup_event_ratelimit(memcg,
805 MEM_CGROUP_TARGET_NUMAINFO);
807 mem_cgroup_threshold(memcg);
808 if (unlikely(do_softlimit))
809 mem_cgroup_update_tree(memcg, page);
811 if (unlikely(do_numainfo))
812 atomic_inc(&memcg->numainfo_events);
817 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
820 * mm_update_next_owner() may clear mm->owner to NULL
821 * if it races with swapoff, page migration, etc.
822 * So this can be called with p == NULL.
827 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
829 EXPORT_SYMBOL(mem_cgroup_from_task);
831 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
833 struct mem_cgroup *memcg = NULL;
838 * Page cache insertions can happen withou an
839 * actual mm context, e.g. during disk probing
840 * on boot, loopback IO, acct() writes etc.
843 memcg = root_mem_cgroup;
845 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
846 if (unlikely(!memcg))
847 memcg = root_mem_cgroup;
849 } while (!css_tryget_online(&memcg->css));
855 * mem_cgroup_iter - iterate over memory cgroup hierarchy
856 * @root: hierarchy root
857 * @prev: previously returned memcg, NULL on first invocation
858 * @reclaim: cookie for shared reclaim walks, NULL for full walks
860 * Returns references to children of the hierarchy below @root, or
861 * @root itself, or %NULL after a full round-trip.
863 * Caller must pass the return value in @prev on subsequent
864 * invocations for reference counting, or use mem_cgroup_iter_break()
865 * to cancel a hierarchy walk before the round-trip is complete.
867 * Reclaimers can specify a zone and a priority level in @reclaim to
868 * divide up the memcgs in the hierarchy among all concurrent
869 * reclaimers operating on the same zone and priority.
871 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
872 struct mem_cgroup *prev,
873 struct mem_cgroup_reclaim_cookie *reclaim)
875 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
876 struct cgroup_subsys_state *css = NULL;
877 struct mem_cgroup *memcg = NULL;
878 struct mem_cgroup *pos = NULL;
880 if (mem_cgroup_disabled())
884 root = root_mem_cgroup;
886 if (prev && !reclaim)
889 if (!root->use_hierarchy && root != root_mem_cgroup) {
898 struct mem_cgroup_per_zone *mz;
900 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
901 iter = &mz->iter[reclaim->priority];
903 if (prev && reclaim->generation != iter->generation)
907 pos = READ_ONCE(iter->position);
908 if (!pos || css_tryget(&pos->css))
911 * css reference reached zero, so iter->position will
912 * be cleared by ->css_released. However, we should not
913 * rely on this happening soon, because ->css_released
914 * is called from a work queue, and by busy-waiting we
915 * might block it. So we clear iter->position right
918 (void)cmpxchg(&iter->position, pos, NULL);
926 css = css_next_descendant_pre(css, &root->css);
929 * Reclaimers share the hierarchy walk, and a
930 * new one might jump in right at the end of
931 * the hierarchy - make sure they see at least
932 * one group and restart from the beginning.
940 * Verify the css and acquire a reference. The root
941 * is provided by the caller, so we know it's alive
942 * and kicking, and don't take an extra reference.
944 memcg = mem_cgroup_from_css(css);
946 if (css == &root->css)
949 if (css_tryget(css)) {
951 * Make sure the memcg is initialized:
952 * mem_cgroup_css_online() orders the the
953 * initialization against setting the flag.
955 if (smp_load_acquire(&memcg->initialized))
966 * The position could have already been updated by a competing
967 * thread, so check that the value hasn't changed since we read
968 * it to avoid reclaiming from the same cgroup twice.
970 (void)cmpxchg(&iter->position, pos, memcg);
978 reclaim->generation = iter->generation;
984 if (prev && prev != root)
991 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
992 * @root: hierarchy root
993 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
995 void mem_cgroup_iter_break(struct mem_cgroup *root,
996 struct mem_cgroup *prev)
999 root = root_mem_cgroup;
1000 if (prev && prev != root)
1001 css_put(&prev->css);
1004 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1006 struct mem_cgroup *memcg = dead_memcg;
1007 struct mem_cgroup_reclaim_iter *iter;
1008 struct mem_cgroup_per_zone *mz;
1012 while ((memcg = parent_mem_cgroup(memcg))) {
1013 for_each_node(nid) {
1014 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1015 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
1016 for (i = 0; i <= DEF_PRIORITY; i++) {
1017 iter = &mz->iter[i];
1018 cmpxchg(&iter->position,
1027 * Iteration constructs for visiting all cgroups (under a tree). If
1028 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1029 * be used for reference counting.
1031 #define for_each_mem_cgroup_tree(iter, root) \
1032 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1034 iter = mem_cgroup_iter(root, iter, NULL))
1036 #define for_each_mem_cgroup(iter) \
1037 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1039 iter = mem_cgroup_iter(NULL, iter, NULL))
1042 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1043 * @zone: zone of the wanted lruvec
1044 * @memcg: memcg of the wanted lruvec
1046 * Returns the lru list vector holding pages for the given @zone and
1047 * @mem. This can be the global zone lruvec, if the memory controller
1050 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1051 struct mem_cgroup *memcg)
1053 struct mem_cgroup_per_zone *mz;
1054 struct lruvec *lruvec;
1056 if (mem_cgroup_disabled()) {
1057 lruvec = &zone->lruvec;
1061 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1062 lruvec = &mz->lruvec;
1065 * Since a node can be onlined after the mem_cgroup was created,
1066 * we have to be prepared to initialize lruvec->zone here;
1067 * and if offlined then reonlined, we need to reinitialize it.
1069 if (unlikely(lruvec->zone != zone))
1070 lruvec->zone = zone;
1075 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1077 * @zone: zone of the page
1079 * This function is only safe when following the LRU page isolation
1080 * and putback protocol: the LRU lock must be held, and the page must
1081 * either be PageLRU() or the caller must have isolated/allocated it.
1083 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1085 struct mem_cgroup_per_zone *mz;
1086 struct mem_cgroup *memcg;
1087 struct lruvec *lruvec;
1089 if (mem_cgroup_disabled()) {
1090 lruvec = &zone->lruvec;
1094 memcg = page->mem_cgroup;
1096 * Swapcache readahead pages are added to the LRU - and
1097 * possibly migrated - before they are charged.
1100 memcg = root_mem_cgroup;
1102 mz = mem_cgroup_page_zoneinfo(memcg, page);
1103 lruvec = &mz->lruvec;
1106 * Since a node can be onlined after the mem_cgroup was created,
1107 * we have to be prepared to initialize lruvec->zone here;
1108 * and if offlined then reonlined, we need to reinitialize it.
1110 if (unlikely(lruvec->zone != zone))
1111 lruvec->zone = zone;
1116 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1117 * @lruvec: mem_cgroup per zone lru vector
1118 * @lru: index of lru list the page is sitting on
1119 * @nr_pages: positive when adding or negative when removing
1121 * This function must be called when a page is added to or removed from an
1124 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1127 struct mem_cgroup_per_zone *mz;
1128 unsigned long *lru_size;
1130 if (mem_cgroup_disabled())
1133 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1134 lru_size = mz->lru_size + lru;
1135 *lru_size += nr_pages;
1136 VM_BUG_ON((long)(*lru_size) < 0);
1139 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1141 struct mem_cgroup *task_memcg;
1142 struct task_struct *p;
1145 p = find_lock_task_mm(task);
1147 task_memcg = get_mem_cgroup_from_mm(p->mm);
1151 * All threads may have already detached their mm's, but the oom
1152 * killer still needs to detect if they have already been oom
1153 * killed to prevent needlessly killing additional tasks.
1156 task_memcg = mem_cgroup_from_task(task);
1157 css_get(&task_memcg->css);
1160 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1161 css_put(&task_memcg->css);
1165 #define mem_cgroup_from_counter(counter, member) \
1166 container_of(counter, struct mem_cgroup, member)
1169 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1170 * @memcg: the memory cgroup
1172 * Returns the maximum amount of memory @mem can be charged with, in
1175 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1177 unsigned long margin = 0;
1178 unsigned long count;
1179 unsigned long limit;
1181 count = page_counter_read(&memcg->memory);
1182 limit = READ_ONCE(memcg->memory.limit);
1184 margin = limit - count;
1186 if (do_swap_account) {
1187 count = page_counter_read(&memcg->memsw);
1188 limit = READ_ONCE(memcg->memsw.limit);
1190 margin = min(margin, limit - count);
1197 * A routine for checking "mem" is under move_account() or not.
1199 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1200 * moving cgroups. This is for waiting at high-memory pressure
1203 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1205 struct mem_cgroup *from;
1206 struct mem_cgroup *to;
1209 * Unlike task_move routines, we access mc.to, mc.from not under
1210 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1212 spin_lock(&mc.lock);
1218 ret = mem_cgroup_is_descendant(from, memcg) ||
1219 mem_cgroup_is_descendant(to, memcg);
1221 spin_unlock(&mc.lock);
1225 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1227 if (mc.moving_task && current != mc.moving_task) {
1228 if (mem_cgroup_under_move(memcg)) {
1230 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1231 /* moving charge context might have finished. */
1234 finish_wait(&mc.waitq, &wait);
1241 #define K(x) ((x) << (PAGE_SHIFT-10))
1243 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1244 * @memcg: The memory cgroup that went over limit
1245 * @p: Task that is going to be killed
1247 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1250 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1252 /* oom_info_lock ensures that parallel ooms do not interleave */
1253 static DEFINE_MUTEX(oom_info_lock);
1254 struct mem_cgroup *iter;
1257 mutex_lock(&oom_info_lock);
1261 pr_info("Task in ");
1262 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1263 pr_cont(" killed as a result of limit of ");
1265 pr_info("Memory limit reached of cgroup ");
1268 pr_cont_cgroup_path(memcg->css.cgroup);
1273 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1274 K((u64)page_counter_read(&memcg->memory)),
1275 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1276 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1277 K((u64)page_counter_read(&memcg->memsw)),
1278 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1279 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1280 K((u64)page_counter_read(&memcg->kmem)),
1281 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1283 for_each_mem_cgroup_tree(iter, memcg) {
1284 pr_info("Memory cgroup stats for ");
1285 pr_cont_cgroup_path(iter->css.cgroup);
1288 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1289 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1291 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1292 K(mem_cgroup_read_stat(iter, i)));
1295 for (i = 0; i < NR_LRU_LISTS; i++)
1296 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1297 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1301 mutex_unlock(&oom_info_lock);
1305 * This function returns the number of memcg under hierarchy tree. Returns
1306 * 1(self count) if no children.
1308 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1311 struct mem_cgroup *iter;
1313 for_each_mem_cgroup_tree(iter, memcg)
1319 * Return the memory (and swap, if configured) limit for a memcg.
1321 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1323 unsigned long limit;
1325 limit = memcg->memory.limit;
1326 if (mem_cgroup_swappiness(memcg)) {
1327 unsigned long memsw_limit;
1329 memsw_limit = memcg->memsw.limit;
1330 limit = min(limit + total_swap_pages, memsw_limit);
1335 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1338 struct oom_control oc = {
1341 .gfp_mask = gfp_mask,
1344 struct mem_cgroup *iter;
1345 unsigned long chosen_points = 0;
1346 unsigned long totalpages;
1347 unsigned int points = 0;
1348 struct task_struct *chosen = NULL;
1350 mutex_lock(&oom_lock);
1353 * If current has a pending SIGKILL or is exiting, then automatically
1354 * select it. The goal is to allow it to allocate so that it may
1355 * quickly exit and free its memory.
1357 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1358 mark_oom_victim(current);
1362 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1363 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1364 for_each_mem_cgroup_tree(iter, memcg) {
1365 struct css_task_iter it;
1366 struct task_struct *task;
1368 css_task_iter_start(&iter->css, &it);
1369 while ((task = css_task_iter_next(&it))) {
1370 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1371 case OOM_SCAN_SELECT:
1373 put_task_struct(chosen);
1375 chosen_points = ULONG_MAX;
1376 get_task_struct(chosen);
1378 case OOM_SCAN_CONTINUE:
1380 case OOM_SCAN_ABORT:
1381 css_task_iter_end(&it);
1382 mem_cgroup_iter_break(memcg, iter);
1384 put_task_struct(chosen);
1389 points = oom_badness(task, memcg, NULL, totalpages);
1390 if (!points || points < chosen_points)
1392 /* Prefer thread group leaders for display purposes */
1393 if (points == chosen_points &&
1394 thread_group_leader(chosen))
1398 put_task_struct(chosen);
1400 chosen_points = points;
1401 get_task_struct(chosen);
1403 css_task_iter_end(&it);
1407 points = chosen_points * 1000 / totalpages;
1408 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1409 "Memory cgroup out of memory");
1412 mutex_unlock(&oom_lock);
1416 #if MAX_NUMNODES > 1
1419 * test_mem_cgroup_node_reclaimable
1420 * @memcg: the target memcg
1421 * @nid: the node ID to be checked.
1422 * @noswap : specify true here if the user wants flle only information.
1424 * This function returns whether the specified memcg contains any
1425 * reclaimable pages on a node. Returns true if there are any reclaimable
1426 * pages in the node.
1428 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1429 int nid, bool noswap)
1431 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1433 if (noswap || !total_swap_pages)
1435 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1442 * Always updating the nodemask is not very good - even if we have an empty
1443 * list or the wrong list here, we can start from some node and traverse all
1444 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1447 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1451 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1452 * pagein/pageout changes since the last update.
1454 if (!atomic_read(&memcg->numainfo_events))
1456 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1459 /* make a nodemask where this memcg uses memory from */
1460 memcg->scan_nodes = node_states[N_MEMORY];
1462 for_each_node_mask(nid, node_states[N_MEMORY]) {
1464 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1465 node_clear(nid, memcg->scan_nodes);
1468 atomic_set(&memcg->numainfo_events, 0);
1469 atomic_set(&memcg->numainfo_updating, 0);
1473 * Selecting a node where we start reclaim from. Because what we need is just
1474 * reducing usage counter, start from anywhere is O,K. Considering
1475 * memory reclaim from current node, there are pros. and cons.
1477 * Freeing memory from current node means freeing memory from a node which
1478 * we'll use or we've used. So, it may make LRU bad. And if several threads
1479 * hit limits, it will see a contention on a node. But freeing from remote
1480 * node means more costs for memory reclaim because of memory latency.
1482 * Now, we use round-robin. Better algorithm is welcomed.
1484 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1488 mem_cgroup_may_update_nodemask(memcg);
1489 node = memcg->last_scanned_node;
1491 node = next_node(node, memcg->scan_nodes);
1492 if (node == MAX_NUMNODES)
1493 node = first_node(memcg->scan_nodes);
1495 * We call this when we hit limit, not when pages are added to LRU.
1496 * No LRU may hold pages because all pages are UNEVICTABLE or
1497 * memcg is too small and all pages are not on LRU. In that case,
1498 * we use curret node.
1500 if (unlikely(node == MAX_NUMNODES))
1501 node = numa_node_id();
1503 memcg->last_scanned_node = node;
1507 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1513 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1516 unsigned long *total_scanned)
1518 struct mem_cgroup *victim = NULL;
1521 unsigned long excess;
1522 unsigned long nr_scanned;
1523 struct mem_cgroup_reclaim_cookie reclaim = {
1528 excess = soft_limit_excess(root_memcg);
1531 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1536 * If we have not been able to reclaim
1537 * anything, it might because there are
1538 * no reclaimable pages under this hierarchy
1543 * We want to do more targeted reclaim.
1544 * excess >> 2 is not to excessive so as to
1545 * reclaim too much, nor too less that we keep
1546 * coming back to reclaim from this cgroup
1548 if (total >= (excess >> 2) ||
1549 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1554 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1556 *total_scanned += nr_scanned;
1557 if (!soft_limit_excess(root_memcg))
1560 mem_cgroup_iter_break(root_memcg, victim);
1564 #ifdef CONFIG_LOCKDEP
1565 static struct lockdep_map memcg_oom_lock_dep_map = {
1566 .name = "memcg_oom_lock",
1570 static DEFINE_SPINLOCK(memcg_oom_lock);
1573 * Check OOM-Killer is already running under our hierarchy.
1574 * If someone is running, return false.
1576 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1578 struct mem_cgroup *iter, *failed = NULL;
1580 spin_lock(&memcg_oom_lock);
1582 for_each_mem_cgroup_tree(iter, memcg) {
1583 if (iter->oom_lock) {
1585 * this subtree of our hierarchy is already locked
1586 * so we cannot give a lock.
1589 mem_cgroup_iter_break(memcg, iter);
1592 iter->oom_lock = true;
1597 * OK, we failed to lock the whole subtree so we have
1598 * to clean up what we set up to the failing subtree
1600 for_each_mem_cgroup_tree(iter, memcg) {
1601 if (iter == failed) {
1602 mem_cgroup_iter_break(memcg, iter);
1605 iter->oom_lock = false;
1608 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1610 spin_unlock(&memcg_oom_lock);
1615 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1617 struct mem_cgroup *iter;
1619 spin_lock(&memcg_oom_lock);
1620 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1621 for_each_mem_cgroup_tree(iter, memcg)
1622 iter->oom_lock = false;
1623 spin_unlock(&memcg_oom_lock);
1626 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1628 struct mem_cgroup *iter;
1630 spin_lock(&memcg_oom_lock);
1631 for_each_mem_cgroup_tree(iter, memcg)
1633 spin_unlock(&memcg_oom_lock);
1636 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1638 struct mem_cgroup *iter;
1641 * When a new child is created while the hierarchy is under oom,
1642 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1644 spin_lock(&memcg_oom_lock);
1645 for_each_mem_cgroup_tree(iter, memcg)
1646 if (iter->under_oom > 0)
1648 spin_unlock(&memcg_oom_lock);
1651 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1653 struct oom_wait_info {
1654 struct mem_cgroup *memcg;
1658 static int memcg_oom_wake_function(wait_queue_t *wait,
1659 unsigned mode, int sync, void *arg)
1661 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1662 struct mem_cgroup *oom_wait_memcg;
1663 struct oom_wait_info *oom_wait_info;
1665 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1666 oom_wait_memcg = oom_wait_info->memcg;
1668 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1669 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1671 return autoremove_wake_function(wait, mode, sync, arg);
1674 static void memcg_oom_recover(struct mem_cgroup *memcg)
1677 * For the following lockless ->under_oom test, the only required
1678 * guarantee is that it must see the state asserted by an OOM when
1679 * this function is called as a result of userland actions
1680 * triggered by the notification of the OOM. This is trivially
1681 * achieved by invoking mem_cgroup_mark_under_oom() before
1682 * triggering notification.
1684 if (memcg && memcg->under_oom)
1685 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1688 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1690 if (!current->memcg_may_oom)
1693 * We are in the middle of the charge context here, so we
1694 * don't want to block when potentially sitting on a callstack
1695 * that holds all kinds of filesystem and mm locks.
1697 * Also, the caller may handle a failed allocation gracefully
1698 * (like optional page cache readahead) and so an OOM killer
1699 * invocation might not even be necessary.
1701 * That's why we don't do anything here except remember the
1702 * OOM context and then deal with it at the end of the page
1703 * fault when the stack is unwound, the locks are released,
1704 * and when we know whether the fault was overall successful.
1706 css_get(&memcg->css);
1707 current->memcg_in_oom = memcg;
1708 current->memcg_oom_gfp_mask = mask;
1709 current->memcg_oom_order = order;
1713 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1714 * @handle: actually kill/wait or just clean up the OOM state
1716 * This has to be called at the end of a page fault if the memcg OOM
1717 * handler was enabled.
1719 * Memcg supports userspace OOM handling where failed allocations must
1720 * sleep on a waitqueue until the userspace task resolves the
1721 * situation. Sleeping directly in the charge context with all kinds
1722 * of locks held is not a good idea, instead we remember an OOM state
1723 * in the task and mem_cgroup_oom_synchronize() has to be called at
1724 * the end of the page fault to complete the OOM handling.
1726 * Returns %true if an ongoing memcg OOM situation was detected and
1727 * completed, %false otherwise.
1729 bool mem_cgroup_oom_synchronize(bool handle)
1731 struct mem_cgroup *memcg = current->memcg_in_oom;
1732 struct oom_wait_info owait;
1735 /* OOM is global, do not handle */
1739 if (!handle || oom_killer_disabled)
1742 owait.memcg = memcg;
1743 owait.wait.flags = 0;
1744 owait.wait.func = memcg_oom_wake_function;
1745 owait.wait.private = current;
1746 INIT_LIST_HEAD(&owait.wait.task_list);
1748 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1749 mem_cgroup_mark_under_oom(memcg);
1751 locked = mem_cgroup_oom_trylock(memcg);
1754 mem_cgroup_oom_notify(memcg);
1756 if (locked && !memcg->oom_kill_disable) {
1757 mem_cgroup_unmark_under_oom(memcg);
1758 finish_wait(&memcg_oom_waitq, &owait.wait);
1759 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1760 current->memcg_oom_order);
1763 mem_cgroup_unmark_under_oom(memcg);
1764 finish_wait(&memcg_oom_waitq, &owait.wait);
1768 mem_cgroup_oom_unlock(memcg);
1770 * There is no guarantee that an OOM-lock contender
1771 * sees the wakeups triggered by the OOM kill
1772 * uncharges. Wake any sleepers explicitely.
1774 memcg_oom_recover(memcg);
1777 current->memcg_in_oom = NULL;
1778 css_put(&memcg->css);
1783 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1784 * @page: page that is going to change accounted state
1786 * This function must mark the beginning of an accounted page state
1787 * change to prevent double accounting when the page is concurrently
1788 * being moved to another memcg:
1790 * memcg = mem_cgroup_begin_page_stat(page);
1791 * if (TestClearPageState(page))
1792 * mem_cgroup_update_page_stat(memcg, state, -1);
1793 * mem_cgroup_end_page_stat(memcg);
1795 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1797 struct mem_cgroup *memcg;
1798 unsigned long flags;
1801 * The RCU lock is held throughout the transaction. The fast
1802 * path can get away without acquiring the memcg->move_lock
1803 * because page moving starts with an RCU grace period.
1805 * The RCU lock also protects the memcg from being freed when
1806 * the page state that is going to change is the only thing
1807 * preventing the page from being uncharged.
1808 * E.g. end-writeback clearing PageWriteback(), which allows
1809 * migration to go ahead and uncharge the page before the
1810 * account transaction might be complete.
1814 if (mem_cgroup_disabled())
1817 memcg = page->mem_cgroup;
1818 if (unlikely(!memcg))
1821 if (atomic_read(&memcg->moving_account) <= 0)
1824 spin_lock_irqsave(&memcg->move_lock, flags);
1825 if (memcg != page->mem_cgroup) {
1826 spin_unlock_irqrestore(&memcg->move_lock, flags);
1831 * When charge migration first begins, we can have locked and
1832 * unlocked page stat updates happening concurrently. Track
1833 * the task who has the lock for mem_cgroup_end_page_stat().
1835 memcg->move_lock_task = current;
1836 memcg->move_lock_flags = flags;
1840 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1843 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1844 * @memcg: the memcg that was accounted against
1846 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1848 if (memcg && memcg->move_lock_task == current) {
1849 unsigned long flags = memcg->move_lock_flags;
1851 memcg->move_lock_task = NULL;
1852 memcg->move_lock_flags = 0;
1854 spin_unlock_irqrestore(&memcg->move_lock, flags);
1859 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1862 * size of first charge trial. "32" comes from vmscan.c's magic value.
1863 * TODO: maybe necessary to use big numbers in big irons.
1865 #define CHARGE_BATCH 32U
1866 struct memcg_stock_pcp {
1867 struct mem_cgroup *cached; /* this never be root cgroup */
1868 unsigned int nr_pages;
1869 struct work_struct work;
1870 unsigned long flags;
1871 #define FLUSHING_CACHED_CHARGE 0
1873 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1874 static DEFINE_MUTEX(percpu_charge_mutex);
1877 * consume_stock: Try to consume stocked charge on this cpu.
1878 * @memcg: memcg to consume from.
1879 * @nr_pages: how many pages to charge.
1881 * The charges will only happen if @memcg matches the current cpu's memcg
1882 * stock, and at least @nr_pages are available in that stock. Failure to
1883 * service an allocation will refill the stock.
1885 * returns true if successful, false otherwise.
1887 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1889 struct memcg_stock_pcp *stock;
1892 if (nr_pages > CHARGE_BATCH)
1895 stock = &get_cpu_var(memcg_stock);
1896 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1897 stock->nr_pages -= nr_pages;
1900 put_cpu_var(memcg_stock);
1905 * Returns stocks cached in percpu and reset cached information.
1907 static void drain_stock(struct memcg_stock_pcp *stock)
1909 struct mem_cgroup *old = stock->cached;
1911 if (stock->nr_pages) {
1912 page_counter_uncharge(&old->memory, stock->nr_pages);
1913 if (do_swap_account)
1914 page_counter_uncharge(&old->memsw, stock->nr_pages);
1915 css_put_many(&old->css, stock->nr_pages);
1916 stock->nr_pages = 0;
1918 stock->cached = NULL;
1922 * This must be called under preempt disabled or must be called by
1923 * a thread which is pinned to local cpu.
1925 static void drain_local_stock(struct work_struct *dummy)
1927 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1929 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1933 * Cache charges(val) to local per_cpu area.
1934 * This will be consumed by consume_stock() function, later.
1936 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1938 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1940 if (stock->cached != memcg) { /* reset if necessary */
1942 stock->cached = memcg;
1944 stock->nr_pages += nr_pages;
1945 put_cpu_var(memcg_stock);
1949 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1950 * of the hierarchy under it.
1952 static void drain_all_stock(struct mem_cgroup *root_memcg)
1956 /* If someone's already draining, avoid adding running more workers. */
1957 if (!mutex_trylock(&percpu_charge_mutex))
1959 /* Notify other cpus that system-wide "drain" is running */
1962 for_each_online_cpu(cpu) {
1963 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1964 struct mem_cgroup *memcg;
1966 memcg = stock->cached;
1967 if (!memcg || !stock->nr_pages)
1969 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1971 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1973 drain_local_stock(&stock->work);
1975 schedule_work_on(cpu, &stock->work);
1980 mutex_unlock(&percpu_charge_mutex);
1983 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1984 unsigned long action,
1987 int cpu = (unsigned long)hcpu;
1988 struct memcg_stock_pcp *stock;
1990 if (action == CPU_ONLINE)
1993 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1996 stock = &per_cpu(memcg_stock, cpu);
2002 * Scheduled by try_charge() to be executed from the userland return path
2003 * and reclaims memory over the high limit.
2005 void mem_cgroup_handle_over_high(void)
2007 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2008 struct mem_cgroup *memcg, *pos;
2010 if (likely(!nr_pages))
2013 pos = memcg = get_mem_cgroup_from_mm(current->mm);
2016 if (page_counter_read(&pos->memory) <= pos->high)
2018 mem_cgroup_events(pos, MEMCG_HIGH, 1);
2019 try_to_free_mem_cgroup_pages(pos, nr_pages, GFP_KERNEL, true);
2020 } while ((pos = parent_mem_cgroup(pos)));
2022 css_put(&memcg->css);
2023 current->memcg_nr_pages_over_high = 0;
2026 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2027 unsigned int nr_pages)
2029 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2030 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2031 struct mem_cgroup *mem_over_limit;
2032 struct page_counter *counter;
2033 unsigned long nr_reclaimed;
2034 bool may_swap = true;
2035 bool drained = false;
2037 if (mem_cgroup_is_root(memcg))
2040 if (consume_stock(memcg, nr_pages))
2043 if (!do_swap_account ||
2044 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2045 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2047 if (do_swap_account)
2048 page_counter_uncharge(&memcg->memsw, batch);
2049 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2051 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2055 if (batch > nr_pages) {
2061 * Unlike in global OOM situations, memcg is not in a physical
2062 * memory shortage. Allow dying and OOM-killed tasks to
2063 * bypass the last charges so that they can exit quickly and
2064 * free their memory.
2066 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2067 fatal_signal_pending(current) ||
2068 current->flags & PF_EXITING))
2071 if (unlikely(task_in_memcg_oom(current)))
2074 if (!gfpflags_allow_blocking(gfp_mask))
2077 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2079 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2080 gfp_mask, may_swap);
2082 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2086 drain_all_stock(mem_over_limit);
2091 if (gfp_mask & __GFP_NORETRY)
2094 * Even though the limit is exceeded at this point, reclaim
2095 * may have been able to free some pages. Retry the charge
2096 * before killing the task.
2098 * Only for regular pages, though: huge pages are rather
2099 * unlikely to succeed so close to the limit, and we fall back
2100 * to regular pages anyway in case of failure.
2102 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2105 * At task move, charge accounts can be doubly counted. So, it's
2106 * better to wait until the end of task_move if something is going on.
2108 if (mem_cgroup_wait_acct_move(mem_over_limit))
2114 if (gfp_mask & __GFP_NOFAIL)
2117 if (fatal_signal_pending(current))
2120 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2122 mem_cgroup_oom(mem_over_limit, gfp_mask,
2123 get_order(nr_pages * PAGE_SIZE));
2125 if (!(gfp_mask & __GFP_NOFAIL))
2129 * The allocation either can't fail or will lead to more memory
2130 * being freed very soon. Allow memory usage go over the limit
2131 * temporarily by force charging it.
2133 page_counter_charge(&memcg->memory, nr_pages);
2134 if (do_swap_account)
2135 page_counter_charge(&memcg->memsw, nr_pages);
2136 css_get_many(&memcg->css, nr_pages);
2141 css_get_many(&memcg->css, batch);
2142 if (batch > nr_pages)
2143 refill_stock(memcg, batch - nr_pages);
2146 * If the hierarchy is above the normal consumption range, schedule
2147 * reclaim on returning to userland. We can perform reclaim here
2148 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2149 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2150 * not recorded as it most likely matches current's and won't
2151 * change in the meantime. As high limit is checked again before
2152 * reclaim, the cost of mismatch is negligible.
2155 if (page_counter_read(&memcg->memory) > memcg->high) {
2156 current->memcg_nr_pages_over_high += batch;
2157 set_notify_resume(current);
2160 } while ((memcg = parent_mem_cgroup(memcg)));
2165 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2167 if (mem_cgroup_is_root(memcg))
2170 page_counter_uncharge(&memcg->memory, nr_pages);
2171 if (do_swap_account)
2172 page_counter_uncharge(&memcg->memsw, nr_pages);
2174 css_put_many(&memcg->css, nr_pages);
2177 static void lock_page_lru(struct page *page, int *isolated)
2179 struct zone *zone = page_zone(page);
2181 spin_lock_irq(&zone->lru_lock);
2182 if (PageLRU(page)) {
2183 struct lruvec *lruvec;
2185 lruvec = mem_cgroup_page_lruvec(page, zone);
2187 del_page_from_lru_list(page, lruvec, page_lru(page));
2193 static void unlock_page_lru(struct page *page, int isolated)
2195 struct zone *zone = page_zone(page);
2198 struct lruvec *lruvec;
2200 lruvec = mem_cgroup_page_lruvec(page, zone);
2201 VM_BUG_ON_PAGE(PageLRU(page), page);
2203 add_page_to_lru_list(page, lruvec, page_lru(page));
2205 spin_unlock_irq(&zone->lru_lock);
2208 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2213 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2216 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2217 * may already be on some other mem_cgroup's LRU. Take care of it.
2220 lock_page_lru(page, &isolated);
2223 * Nobody should be changing or seriously looking at
2224 * page->mem_cgroup at this point:
2226 * - the page is uncharged
2228 * - the page is off-LRU
2230 * - an anonymous fault has exclusive page access, except for
2231 * a locked page table
2233 * - a page cache insertion, a swapin fault, or a migration
2234 * have the page locked
2236 page->mem_cgroup = memcg;
2239 unlock_page_lru(page, isolated);
2242 #ifdef CONFIG_MEMCG_KMEM
2243 static int memcg_alloc_cache_id(void)
2248 id = ida_simple_get(&memcg_cache_ida,
2249 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2253 if (id < memcg_nr_cache_ids)
2257 * There's no space for the new id in memcg_caches arrays,
2258 * so we have to grow them.
2260 down_write(&memcg_cache_ids_sem);
2262 size = 2 * (id + 1);
2263 if (size < MEMCG_CACHES_MIN_SIZE)
2264 size = MEMCG_CACHES_MIN_SIZE;
2265 else if (size > MEMCG_CACHES_MAX_SIZE)
2266 size = MEMCG_CACHES_MAX_SIZE;
2268 err = memcg_update_all_caches(size);
2270 err = memcg_update_all_list_lrus(size);
2272 memcg_nr_cache_ids = size;
2274 up_write(&memcg_cache_ids_sem);
2277 ida_simple_remove(&memcg_cache_ida, id);
2283 static void memcg_free_cache_id(int id)
2285 ida_simple_remove(&memcg_cache_ida, id);
2288 struct memcg_kmem_cache_create_work {
2289 struct mem_cgroup *memcg;
2290 struct kmem_cache *cachep;
2291 struct work_struct work;
2294 static void memcg_kmem_cache_create_func(struct work_struct *w)
2296 struct memcg_kmem_cache_create_work *cw =
2297 container_of(w, struct memcg_kmem_cache_create_work, work);
2298 struct mem_cgroup *memcg = cw->memcg;
2299 struct kmem_cache *cachep = cw->cachep;
2301 memcg_create_kmem_cache(memcg, cachep);
2303 css_put(&memcg->css);
2308 * Enqueue the creation of a per-memcg kmem_cache.
2310 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2311 struct kmem_cache *cachep)
2313 struct memcg_kmem_cache_create_work *cw;
2315 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2319 css_get(&memcg->css);
2322 cw->cachep = cachep;
2323 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2325 schedule_work(&cw->work);
2328 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2329 struct kmem_cache *cachep)
2332 * We need to stop accounting when we kmalloc, because if the
2333 * corresponding kmalloc cache is not yet created, the first allocation
2334 * in __memcg_schedule_kmem_cache_create will recurse.
2336 * However, it is better to enclose the whole function. Depending on
2337 * the debugging options enabled, INIT_WORK(), for instance, can
2338 * trigger an allocation. This too, will make us recurse. Because at
2339 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2340 * the safest choice is to do it like this, wrapping the whole function.
2342 current->memcg_kmem_skip_account = 1;
2343 __memcg_schedule_kmem_cache_create(memcg, cachep);
2344 current->memcg_kmem_skip_account = 0;
2348 * Return the kmem_cache we're supposed to use for a slab allocation.
2349 * We try to use the current memcg's version of the cache.
2351 * If the cache does not exist yet, if we are the first user of it,
2352 * we either create it immediately, if possible, or create it asynchronously
2354 * In the latter case, we will let the current allocation go through with
2355 * the original cache.
2357 * Can't be called in interrupt context or from kernel threads.
2358 * This function needs to be called with rcu_read_lock() held.
2360 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2362 struct mem_cgroup *memcg;
2363 struct kmem_cache *memcg_cachep;
2366 VM_BUG_ON(!is_root_cache(cachep));
2368 if (current->memcg_kmem_skip_account)
2371 memcg = get_mem_cgroup_from_mm(current->mm);
2372 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2376 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2377 if (likely(memcg_cachep))
2378 return memcg_cachep;
2381 * If we are in a safe context (can wait, and not in interrupt
2382 * context), we could be be predictable and return right away.
2383 * This would guarantee that the allocation being performed
2384 * already belongs in the new cache.
2386 * However, there are some clashes that can arrive from locking.
2387 * For instance, because we acquire the slab_mutex while doing
2388 * memcg_create_kmem_cache, this means no further allocation
2389 * could happen with the slab_mutex held. So it's better to
2392 memcg_schedule_kmem_cache_create(memcg, cachep);
2394 css_put(&memcg->css);
2398 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2400 if (!is_root_cache(cachep))
2401 css_put(&cachep->memcg_params.memcg->css);
2404 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2405 struct mem_cgroup *memcg)
2407 unsigned int nr_pages = 1 << order;
2408 struct page_counter *counter;
2411 if (!memcg_kmem_is_active(memcg))
2414 if (!page_counter_try_charge(&memcg->kmem, nr_pages, &counter))
2417 ret = try_charge(memcg, gfp, nr_pages);
2419 page_counter_uncharge(&memcg->kmem, nr_pages);
2423 page->mem_cgroup = memcg;
2428 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2430 struct mem_cgroup *memcg;
2433 memcg = get_mem_cgroup_from_mm(current->mm);
2434 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2435 css_put(&memcg->css);
2439 void __memcg_kmem_uncharge(struct page *page, int order)
2441 struct mem_cgroup *memcg = page->mem_cgroup;
2442 unsigned int nr_pages = 1 << order;
2447 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2449 page_counter_uncharge(&memcg->kmem, nr_pages);
2450 page_counter_uncharge(&memcg->memory, nr_pages);
2451 if (do_swap_account)
2452 page_counter_uncharge(&memcg->memsw, nr_pages);
2454 page->mem_cgroup = NULL;
2455 css_put_many(&memcg->css, nr_pages);
2457 #endif /* CONFIG_MEMCG_KMEM */
2459 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2462 * Because tail pages are not marked as "used", set it. We're under
2463 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2464 * charge/uncharge will be never happen and move_account() is done under
2465 * compound_lock(), so we don't have to take care of races.
2467 void mem_cgroup_split_huge_fixup(struct page *head)
2471 if (mem_cgroup_disabled())
2474 for (i = 1; i < HPAGE_PMD_NR; i++)
2475 head[i].mem_cgroup = head->mem_cgroup;
2477 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2480 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2482 #ifdef CONFIG_MEMCG_SWAP
2483 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2486 int val = (charge) ? 1 : -1;
2487 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2491 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2492 * @entry: swap entry to be moved
2493 * @from: mem_cgroup which the entry is moved from
2494 * @to: mem_cgroup which the entry is moved to
2496 * It succeeds only when the swap_cgroup's record for this entry is the same
2497 * as the mem_cgroup's id of @from.
2499 * Returns 0 on success, -EINVAL on failure.
2501 * The caller must have charged to @to, IOW, called page_counter_charge() about
2502 * both res and memsw, and called css_get().
2504 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2505 struct mem_cgroup *from, struct mem_cgroup *to)
2507 unsigned short old_id, new_id;
2509 old_id = mem_cgroup_id(from);
2510 new_id = mem_cgroup_id(to);
2512 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2513 mem_cgroup_swap_statistics(from, false);
2514 mem_cgroup_swap_statistics(to, true);
2520 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2521 struct mem_cgroup *from, struct mem_cgroup *to)
2527 static DEFINE_MUTEX(memcg_limit_mutex);
2529 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2530 unsigned long limit)
2532 unsigned long curusage;
2533 unsigned long oldusage;
2534 bool enlarge = false;
2539 * For keeping hierarchical_reclaim simple, how long we should retry
2540 * is depends on callers. We set our retry-count to be function
2541 * of # of children which we should visit in this loop.
2543 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2544 mem_cgroup_count_children(memcg);
2546 oldusage = page_counter_read(&memcg->memory);
2549 if (signal_pending(current)) {
2554 mutex_lock(&memcg_limit_mutex);
2555 if (limit > memcg->memsw.limit) {
2556 mutex_unlock(&memcg_limit_mutex);
2560 if (limit > memcg->memory.limit)
2562 ret = page_counter_limit(&memcg->memory, limit);
2563 mutex_unlock(&memcg_limit_mutex);
2568 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2570 curusage = page_counter_read(&memcg->memory);
2571 /* Usage is reduced ? */
2572 if (curusage >= oldusage)
2575 oldusage = curusage;
2576 } while (retry_count);
2578 if (!ret && enlarge)
2579 memcg_oom_recover(memcg);
2584 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2585 unsigned long limit)
2587 unsigned long curusage;
2588 unsigned long oldusage;
2589 bool enlarge = false;
2593 /* see mem_cgroup_resize_res_limit */
2594 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2595 mem_cgroup_count_children(memcg);
2597 oldusage = page_counter_read(&memcg->memsw);
2600 if (signal_pending(current)) {
2605 mutex_lock(&memcg_limit_mutex);
2606 if (limit < memcg->memory.limit) {
2607 mutex_unlock(&memcg_limit_mutex);
2611 if (limit > memcg->memsw.limit)
2613 ret = page_counter_limit(&memcg->memsw, limit);
2614 mutex_unlock(&memcg_limit_mutex);
2619 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2621 curusage = page_counter_read(&memcg->memsw);
2622 /* Usage is reduced ? */
2623 if (curusage >= oldusage)
2626 oldusage = curusage;
2627 } while (retry_count);
2629 if (!ret && enlarge)
2630 memcg_oom_recover(memcg);
2635 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2637 unsigned long *total_scanned)
2639 unsigned long nr_reclaimed = 0;
2640 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2641 unsigned long reclaimed;
2643 struct mem_cgroup_tree_per_zone *mctz;
2644 unsigned long excess;
2645 unsigned long nr_scanned;
2650 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2652 * This loop can run a while, specially if mem_cgroup's continuously
2653 * keep exceeding their soft limit and putting the system under
2660 mz = mem_cgroup_largest_soft_limit_node(mctz);
2665 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2666 gfp_mask, &nr_scanned);
2667 nr_reclaimed += reclaimed;
2668 *total_scanned += nr_scanned;
2669 spin_lock_irq(&mctz->lock);
2670 __mem_cgroup_remove_exceeded(mz, mctz);
2673 * If we failed to reclaim anything from this memory cgroup
2674 * it is time to move on to the next cgroup
2678 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2680 excess = soft_limit_excess(mz->memcg);
2682 * One school of thought says that we should not add
2683 * back the node to the tree if reclaim returns 0.
2684 * But our reclaim could return 0, simply because due
2685 * to priority we are exposing a smaller subset of
2686 * memory to reclaim from. Consider this as a longer
2689 /* If excess == 0, no tree ops */
2690 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2691 spin_unlock_irq(&mctz->lock);
2692 css_put(&mz->memcg->css);
2695 * Could not reclaim anything and there are no more
2696 * mem cgroups to try or we seem to be looping without
2697 * reclaiming anything.
2699 if (!nr_reclaimed &&
2701 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2703 } while (!nr_reclaimed);
2705 css_put(&next_mz->memcg->css);
2706 return nr_reclaimed;
2710 * Test whether @memcg has children, dead or alive. Note that this
2711 * function doesn't care whether @memcg has use_hierarchy enabled and
2712 * returns %true if there are child csses according to the cgroup
2713 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2715 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2720 * The lock does not prevent addition or deletion of children, but
2721 * it prevents a new child from being initialized based on this
2722 * parent in css_online(), so it's enough to decide whether
2723 * hierarchically inherited attributes can still be changed or not.
2725 lockdep_assert_held(&memcg_create_mutex);
2728 ret = css_next_child(NULL, &memcg->css);
2734 * Reclaims as many pages from the given memcg as possible and moves
2735 * the rest to the parent.
2737 * Caller is responsible for holding css reference for memcg.
2739 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2741 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2743 /* we call try-to-free pages for make this cgroup empty */
2744 lru_add_drain_all();
2745 /* try to free all pages in this cgroup */
2746 while (nr_retries && page_counter_read(&memcg->memory)) {
2749 if (signal_pending(current))
2752 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2756 /* maybe some writeback is necessary */
2757 congestion_wait(BLK_RW_ASYNC, HZ/10);
2765 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2766 char *buf, size_t nbytes,
2769 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2771 if (mem_cgroup_is_root(memcg))
2773 return mem_cgroup_force_empty(memcg) ?: nbytes;
2776 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2779 return mem_cgroup_from_css(css)->use_hierarchy;
2782 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2783 struct cftype *cft, u64 val)
2786 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2787 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2789 mutex_lock(&memcg_create_mutex);
2791 if (memcg->use_hierarchy == val)
2795 * If parent's use_hierarchy is set, we can't make any modifications
2796 * in the child subtrees. If it is unset, then the change can
2797 * occur, provided the current cgroup has no children.
2799 * For the root cgroup, parent_mem is NULL, we allow value to be
2800 * set if there are no children.
2802 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2803 (val == 1 || val == 0)) {
2804 if (!memcg_has_children(memcg))
2805 memcg->use_hierarchy = val;
2812 mutex_unlock(&memcg_create_mutex);
2817 static unsigned long tree_stat(struct mem_cgroup *memcg,
2818 enum mem_cgroup_stat_index idx)
2820 struct mem_cgroup *iter;
2821 unsigned long val = 0;
2823 for_each_mem_cgroup_tree(iter, memcg)
2824 val += mem_cgroup_read_stat(iter, idx);
2829 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2833 if (mem_cgroup_is_root(memcg)) {
2834 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2835 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2837 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2840 val = page_counter_read(&memcg->memory);
2842 val = page_counter_read(&memcg->memsw);
2855 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2858 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2859 struct page_counter *counter;
2861 switch (MEMFILE_TYPE(cft->private)) {
2863 counter = &memcg->memory;
2866 counter = &memcg->memsw;
2869 counter = &memcg->kmem;
2875 switch (MEMFILE_ATTR(cft->private)) {
2877 if (counter == &memcg->memory)
2878 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2879 if (counter == &memcg->memsw)
2880 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2881 return (u64)page_counter_read(counter) * PAGE_SIZE;
2883 return (u64)counter->limit * PAGE_SIZE;
2885 return (u64)counter->watermark * PAGE_SIZE;
2887 return counter->failcnt;
2888 case RES_SOFT_LIMIT:
2889 return (u64)memcg->soft_limit * PAGE_SIZE;
2895 #ifdef CONFIG_MEMCG_KMEM
2896 static int memcg_activate_kmem(struct mem_cgroup *memcg,
2897 unsigned long nr_pages)
2902 BUG_ON(memcg->kmemcg_id >= 0);
2903 BUG_ON(memcg->kmem_acct_activated);
2904 BUG_ON(memcg->kmem_acct_active);
2907 * For simplicity, we won't allow this to be disabled. It also can't
2908 * be changed if the cgroup has children already, or if tasks had
2911 * If tasks join before we set the limit, a person looking at
2912 * kmem.usage_in_bytes will have no way to determine when it took
2913 * place, which makes the value quite meaningless.
2915 * After it first became limited, changes in the value of the limit are
2916 * of course permitted.
2918 mutex_lock(&memcg_create_mutex);
2919 if (cgroup_is_populated(memcg->css.cgroup) ||
2920 (memcg->use_hierarchy && memcg_has_children(memcg)))
2922 mutex_unlock(&memcg_create_mutex);
2926 memcg_id = memcg_alloc_cache_id();
2933 * We couldn't have accounted to this cgroup, because it hasn't got
2934 * activated yet, so this should succeed.
2936 err = page_counter_limit(&memcg->kmem, nr_pages);
2939 static_key_slow_inc(&memcg_kmem_enabled_key);
2941 * A memory cgroup is considered kmem-active as soon as it gets
2942 * kmemcg_id. Setting the id after enabling static branching will
2943 * guarantee no one starts accounting before all call sites are
2946 memcg->kmemcg_id = memcg_id;
2947 memcg->kmem_acct_activated = true;
2948 memcg->kmem_acct_active = true;
2953 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2954 unsigned long limit)
2958 mutex_lock(&memcg_limit_mutex);
2959 if (!memcg_kmem_is_active(memcg))
2960 ret = memcg_activate_kmem(memcg, limit);
2962 ret = page_counter_limit(&memcg->kmem, limit);
2963 mutex_unlock(&memcg_limit_mutex);
2967 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
2970 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
2975 mutex_lock(&memcg_limit_mutex);
2977 * If the parent cgroup is not kmem-active now, it cannot be activated
2978 * after this point, because it has at least one child already.
2980 if (memcg_kmem_is_active(parent))
2981 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
2982 mutex_unlock(&memcg_limit_mutex);
2986 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2987 unsigned long limit)
2991 #endif /* CONFIG_MEMCG_KMEM */
2994 * The user of this function is...
2997 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2998 char *buf, size_t nbytes, loff_t off)
3000 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3001 unsigned long nr_pages;
3004 buf = strstrip(buf);
3005 ret = page_counter_memparse(buf, "-1", &nr_pages);
3009 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3011 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3015 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3017 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3020 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3023 ret = memcg_update_kmem_limit(memcg, nr_pages);
3027 case RES_SOFT_LIMIT:
3028 memcg->soft_limit = nr_pages;
3032 return ret ?: nbytes;
3035 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3036 size_t nbytes, loff_t off)
3038 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3039 struct page_counter *counter;
3041 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3043 counter = &memcg->memory;
3046 counter = &memcg->memsw;
3049 counter = &memcg->kmem;
3055 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3057 page_counter_reset_watermark(counter);
3060 counter->failcnt = 0;
3069 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3072 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3076 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3077 struct cftype *cft, u64 val)
3079 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3081 if (val & ~MOVE_MASK)
3085 * No kind of locking is needed in here, because ->can_attach() will
3086 * check this value once in the beginning of the process, and then carry
3087 * on with stale data. This means that changes to this value will only
3088 * affect task migrations starting after the change.
3090 memcg->move_charge_at_immigrate = val;
3094 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3095 struct cftype *cft, u64 val)
3102 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3106 unsigned int lru_mask;
3109 static const struct numa_stat stats[] = {
3110 { "total", LRU_ALL },
3111 { "file", LRU_ALL_FILE },
3112 { "anon", LRU_ALL_ANON },
3113 { "unevictable", BIT(LRU_UNEVICTABLE) },
3115 const struct numa_stat *stat;
3118 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3120 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3121 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3122 seq_printf(m, "%s=%lu", stat->name, nr);
3123 for_each_node_state(nid, N_MEMORY) {
3124 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3126 seq_printf(m, " N%d=%lu", nid, nr);
3131 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3132 struct mem_cgroup *iter;
3135 for_each_mem_cgroup_tree(iter, memcg)
3136 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3137 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3138 for_each_node_state(nid, N_MEMORY) {
3140 for_each_mem_cgroup_tree(iter, memcg)
3141 nr += mem_cgroup_node_nr_lru_pages(
3142 iter, nid, stat->lru_mask);
3143 seq_printf(m, " N%d=%lu", nid, nr);
3150 #endif /* CONFIG_NUMA */
3152 static int memcg_stat_show(struct seq_file *m, void *v)
3154 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3155 unsigned long memory, memsw;
3156 struct mem_cgroup *mi;
3159 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3160 MEM_CGROUP_STAT_NSTATS);
3161 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3162 MEM_CGROUP_EVENTS_NSTATS);
3163 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3165 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3166 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3168 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3169 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3172 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3173 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3174 mem_cgroup_read_events(memcg, i));
3176 for (i = 0; i < NR_LRU_LISTS; i++)
3177 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3178 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3180 /* Hierarchical information */
3181 memory = memsw = PAGE_COUNTER_MAX;
3182 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3183 memory = min(memory, mi->memory.limit);
3184 memsw = min(memsw, mi->memsw.limit);
3186 seq_printf(m, "hierarchical_memory_limit %llu\n",
3187 (u64)memory * PAGE_SIZE);
3188 if (do_swap_account)
3189 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3190 (u64)memsw * PAGE_SIZE);
3192 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3193 unsigned long long val = 0;
3195 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3197 for_each_mem_cgroup_tree(mi, memcg)
3198 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3199 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3202 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3203 unsigned long long val = 0;
3205 for_each_mem_cgroup_tree(mi, memcg)
3206 val += mem_cgroup_read_events(mi, i);
3207 seq_printf(m, "total_%s %llu\n",
3208 mem_cgroup_events_names[i], val);
3211 for (i = 0; i < NR_LRU_LISTS; i++) {
3212 unsigned long long val = 0;
3214 for_each_mem_cgroup_tree(mi, memcg)
3215 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3216 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3219 #ifdef CONFIG_DEBUG_VM
3222 struct mem_cgroup_per_zone *mz;
3223 struct zone_reclaim_stat *rstat;
3224 unsigned long recent_rotated[2] = {0, 0};
3225 unsigned long recent_scanned[2] = {0, 0};
3227 for_each_online_node(nid)
3228 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3229 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3230 rstat = &mz->lruvec.reclaim_stat;
3232 recent_rotated[0] += rstat->recent_rotated[0];
3233 recent_rotated[1] += rstat->recent_rotated[1];
3234 recent_scanned[0] += rstat->recent_scanned[0];
3235 recent_scanned[1] += rstat->recent_scanned[1];
3237 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3238 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3239 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3240 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3247 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3250 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3252 return mem_cgroup_swappiness(memcg);
3255 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3256 struct cftype *cft, u64 val)
3258 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3264 memcg->swappiness = val;
3266 vm_swappiness = val;
3271 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3273 struct mem_cgroup_threshold_ary *t;
3274 unsigned long usage;
3279 t = rcu_dereference(memcg->thresholds.primary);
3281 t = rcu_dereference(memcg->memsw_thresholds.primary);
3286 usage = mem_cgroup_usage(memcg, swap);
3289 * current_threshold points to threshold just below or equal to usage.
3290 * If it's not true, a threshold was crossed after last
3291 * call of __mem_cgroup_threshold().
3293 i = t->current_threshold;
3296 * Iterate backward over array of thresholds starting from
3297 * current_threshold and check if a threshold is crossed.
3298 * If none of thresholds below usage is crossed, we read
3299 * only one element of the array here.
3301 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3302 eventfd_signal(t->entries[i].eventfd, 1);
3304 /* i = current_threshold + 1 */
3308 * Iterate forward over array of thresholds starting from
3309 * current_threshold+1 and check if a threshold is crossed.
3310 * If none of thresholds above usage is crossed, we read
3311 * only one element of the array here.
3313 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3314 eventfd_signal(t->entries[i].eventfd, 1);
3316 /* Update current_threshold */
3317 t->current_threshold = i - 1;
3322 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3325 __mem_cgroup_threshold(memcg, false);
3326 if (do_swap_account)
3327 __mem_cgroup_threshold(memcg, true);
3329 memcg = parent_mem_cgroup(memcg);
3333 static int compare_thresholds(const void *a, const void *b)
3335 const struct mem_cgroup_threshold *_a = a;
3336 const struct mem_cgroup_threshold *_b = b;
3338 if (_a->threshold > _b->threshold)
3341 if (_a->threshold < _b->threshold)
3347 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3349 struct mem_cgroup_eventfd_list *ev;
3351 spin_lock(&memcg_oom_lock);
3353 list_for_each_entry(ev, &memcg->oom_notify, list)
3354 eventfd_signal(ev->eventfd, 1);
3356 spin_unlock(&memcg_oom_lock);
3360 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3362 struct mem_cgroup *iter;
3364 for_each_mem_cgroup_tree(iter, memcg)
3365 mem_cgroup_oom_notify_cb(iter);
3368 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3369 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3371 struct mem_cgroup_thresholds *thresholds;
3372 struct mem_cgroup_threshold_ary *new;
3373 unsigned long threshold;
3374 unsigned long usage;
3377 ret = page_counter_memparse(args, "-1", &threshold);
3381 mutex_lock(&memcg->thresholds_lock);
3384 thresholds = &memcg->thresholds;
3385 usage = mem_cgroup_usage(memcg, false);
3386 } else if (type == _MEMSWAP) {
3387 thresholds = &memcg->memsw_thresholds;
3388 usage = mem_cgroup_usage(memcg, true);
3392 /* Check if a threshold crossed before adding a new one */
3393 if (thresholds->primary)
3394 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3396 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3398 /* Allocate memory for new array of thresholds */
3399 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3407 /* Copy thresholds (if any) to new array */
3408 if (thresholds->primary) {
3409 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3410 sizeof(struct mem_cgroup_threshold));
3413 /* Add new threshold */
3414 new->entries[size - 1].eventfd = eventfd;
3415 new->entries[size - 1].threshold = threshold;
3417 /* Sort thresholds. Registering of new threshold isn't time-critical */
3418 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3419 compare_thresholds, NULL);
3421 /* Find current threshold */
3422 new->current_threshold = -1;
3423 for (i = 0; i < size; i++) {
3424 if (new->entries[i].threshold <= usage) {
3426 * new->current_threshold will not be used until
3427 * rcu_assign_pointer(), so it's safe to increment
3430 ++new->current_threshold;
3435 /* Free old spare buffer and save old primary buffer as spare */
3436 kfree(thresholds->spare);
3437 thresholds->spare = thresholds->primary;
3439 rcu_assign_pointer(thresholds->primary, new);
3441 /* To be sure that nobody uses thresholds */
3445 mutex_unlock(&memcg->thresholds_lock);
3450 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3451 struct eventfd_ctx *eventfd, const char *args)
3453 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3456 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3457 struct eventfd_ctx *eventfd, const char *args)
3459 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3462 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3463 struct eventfd_ctx *eventfd, enum res_type type)
3465 struct mem_cgroup_thresholds *thresholds;
3466 struct mem_cgroup_threshold_ary *new;
3467 unsigned long usage;
3470 mutex_lock(&memcg->thresholds_lock);
3473 thresholds = &memcg->thresholds;
3474 usage = mem_cgroup_usage(memcg, false);
3475 } else if (type == _MEMSWAP) {
3476 thresholds = &memcg->memsw_thresholds;
3477 usage = mem_cgroup_usage(memcg, true);
3481 if (!thresholds->primary)
3484 /* Check if a threshold crossed before removing */
3485 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3487 /* Calculate new number of threshold */
3489 for (i = 0; i < thresholds->primary->size; i++) {
3490 if (thresholds->primary->entries[i].eventfd != eventfd)
3494 new = thresholds->spare;
3496 /* Set thresholds array to NULL if we don't have thresholds */
3505 /* Copy thresholds and find current threshold */
3506 new->current_threshold = -1;
3507 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3508 if (thresholds->primary->entries[i].eventfd == eventfd)
3511 new->entries[j] = thresholds->primary->entries[i];
3512 if (new->entries[j].threshold <= usage) {
3514 * new->current_threshold will not be used
3515 * until rcu_assign_pointer(), so it's safe to increment
3518 ++new->current_threshold;
3524 /* Swap primary and spare array */
3525 thresholds->spare = thresholds->primary;
3527 rcu_assign_pointer(thresholds->primary, new);
3529 /* To be sure that nobody uses thresholds */
3532 /* If all events are unregistered, free the spare array */
3534 kfree(thresholds->spare);
3535 thresholds->spare = NULL;
3538 mutex_unlock(&memcg->thresholds_lock);
3541 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3542 struct eventfd_ctx *eventfd)
3544 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3547 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3548 struct eventfd_ctx *eventfd)
3550 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3553 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3554 struct eventfd_ctx *eventfd, const char *args)
3556 struct mem_cgroup_eventfd_list *event;
3558 event = kmalloc(sizeof(*event), GFP_KERNEL);
3562 spin_lock(&memcg_oom_lock);
3564 event->eventfd = eventfd;
3565 list_add(&event->list, &memcg->oom_notify);
3567 /* already in OOM ? */
3568 if (memcg->under_oom)
3569 eventfd_signal(eventfd, 1);
3570 spin_unlock(&memcg_oom_lock);
3575 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3576 struct eventfd_ctx *eventfd)
3578 struct mem_cgroup_eventfd_list *ev, *tmp;
3580 spin_lock(&memcg_oom_lock);
3582 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3583 if (ev->eventfd == eventfd) {
3584 list_del(&ev->list);
3589 spin_unlock(&memcg_oom_lock);
3592 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3594 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3596 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3597 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3601 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3602 struct cftype *cft, u64 val)
3604 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3606 /* cannot set to root cgroup and only 0 and 1 are allowed */
3607 if (!css->parent || !((val == 0) || (val == 1)))
3610 memcg->oom_kill_disable = val;
3612 memcg_oom_recover(memcg);
3617 #ifdef CONFIG_MEMCG_KMEM
3618 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3622 ret = memcg_propagate_kmem(memcg);
3626 return mem_cgroup_sockets_init(memcg, ss);
3629 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3631 struct cgroup_subsys_state *css;
3632 struct mem_cgroup *parent, *child;
3635 if (!memcg->kmem_acct_active)
3639 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3640 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3641 * guarantees no cache will be created for this cgroup after we are
3642 * done (see memcg_create_kmem_cache()).
3644 memcg->kmem_acct_active = false;
3646 memcg_deactivate_kmem_caches(memcg);
3648 kmemcg_id = memcg->kmemcg_id;
3649 BUG_ON(kmemcg_id < 0);
3651 parent = parent_mem_cgroup(memcg);
3653 parent = root_mem_cgroup;
3656 * Change kmemcg_id of this cgroup and all its descendants to the
3657 * parent's id, and then move all entries from this cgroup's list_lrus
3658 * to ones of the parent. After we have finished, all list_lrus
3659 * corresponding to this cgroup are guaranteed to remain empty. The
3660 * ordering is imposed by list_lru_node->lock taken by
3661 * memcg_drain_all_list_lrus().
3663 css_for_each_descendant_pre(css, &memcg->css) {
3664 child = mem_cgroup_from_css(css);
3665 BUG_ON(child->kmemcg_id != kmemcg_id);
3666 child->kmemcg_id = parent->kmemcg_id;
3667 if (!memcg->use_hierarchy)
3670 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3672 memcg_free_cache_id(kmemcg_id);
3675 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3677 if (memcg->kmem_acct_activated) {
3678 memcg_destroy_kmem_caches(memcg);
3679 static_key_slow_dec(&memcg_kmem_enabled_key);
3680 WARN_ON(page_counter_read(&memcg->kmem));
3682 mem_cgroup_sockets_destroy(memcg);
3685 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3690 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3694 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3699 #ifdef CONFIG_CGROUP_WRITEBACK
3701 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3703 return &memcg->cgwb_list;
3706 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3708 return wb_domain_init(&memcg->cgwb_domain, gfp);
3711 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3713 wb_domain_exit(&memcg->cgwb_domain);
3716 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3718 wb_domain_size_changed(&memcg->cgwb_domain);
3721 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3723 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3725 if (!memcg->css.parent)
3728 return &memcg->cgwb_domain;
3732 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3733 * @wb: bdi_writeback in question
3734 * @pfilepages: out parameter for number of file pages
3735 * @pheadroom: out parameter for number of allocatable pages according to memcg
3736 * @pdirty: out parameter for number of dirty pages
3737 * @pwriteback: out parameter for number of pages under writeback
3739 * Determine the numbers of file, headroom, dirty, and writeback pages in
3740 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3741 * is a bit more involved.
3743 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3744 * headroom is calculated as the lowest headroom of itself and the
3745 * ancestors. Note that this doesn't consider the actual amount of
3746 * available memory in the system. The caller should further cap
3747 * *@pheadroom accordingly.
3749 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3750 unsigned long *pheadroom, unsigned long *pdirty,
3751 unsigned long *pwriteback)
3753 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3754 struct mem_cgroup *parent;
3756 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3758 /* this should eventually include NR_UNSTABLE_NFS */
3759 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3760 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3761 (1 << LRU_ACTIVE_FILE));
3762 *pheadroom = PAGE_COUNTER_MAX;
3764 while ((parent = parent_mem_cgroup(memcg))) {
3765 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3766 unsigned long used = page_counter_read(&memcg->memory);
3768 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3773 #else /* CONFIG_CGROUP_WRITEBACK */
3775 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3780 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3784 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3788 #endif /* CONFIG_CGROUP_WRITEBACK */
3791 * DO NOT USE IN NEW FILES.
3793 * "cgroup.event_control" implementation.
3795 * This is way over-engineered. It tries to support fully configurable
3796 * events for each user. Such level of flexibility is completely
3797 * unnecessary especially in the light of the planned unified hierarchy.
3799 * Please deprecate this and replace with something simpler if at all
3804 * Unregister event and free resources.
3806 * Gets called from workqueue.
3808 static void memcg_event_remove(struct work_struct *work)
3810 struct mem_cgroup_event *event =
3811 container_of(work, struct mem_cgroup_event, remove);
3812 struct mem_cgroup *memcg = event->memcg;
3814 remove_wait_queue(event->wqh, &event->wait);
3816 event->unregister_event(memcg, event->eventfd);
3818 /* Notify userspace the event is going away. */
3819 eventfd_signal(event->eventfd, 1);
3821 eventfd_ctx_put(event->eventfd);
3823 css_put(&memcg->css);
3827 * Gets called on POLLHUP on eventfd when user closes it.
3829 * Called with wqh->lock held and interrupts disabled.
3831 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3832 int sync, void *key)
3834 struct mem_cgroup_event *event =
3835 container_of(wait, struct mem_cgroup_event, wait);
3836 struct mem_cgroup *memcg = event->memcg;
3837 unsigned long flags = (unsigned long)key;
3839 if (flags & POLLHUP) {
3841 * If the event has been detached at cgroup removal, we
3842 * can simply return knowing the other side will cleanup
3845 * We can't race against event freeing since the other
3846 * side will require wqh->lock via remove_wait_queue(),
3849 spin_lock(&memcg->event_list_lock);
3850 if (!list_empty(&event->list)) {
3851 list_del_init(&event->list);
3853 * We are in atomic context, but cgroup_event_remove()
3854 * may sleep, so we have to call it in workqueue.
3856 schedule_work(&event->remove);
3858 spin_unlock(&memcg->event_list_lock);
3864 static void memcg_event_ptable_queue_proc(struct file *file,
3865 wait_queue_head_t *wqh, poll_table *pt)
3867 struct mem_cgroup_event *event =
3868 container_of(pt, struct mem_cgroup_event, pt);
3871 add_wait_queue(wqh, &event->wait);
3875 * DO NOT USE IN NEW FILES.
3877 * Parse input and register new cgroup event handler.
3879 * Input must be in format '<event_fd> <control_fd> <args>'.
3880 * Interpretation of args is defined by control file implementation.
3882 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3883 char *buf, size_t nbytes, loff_t off)
3885 struct cgroup_subsys_state *css = of_css(of);
3886 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3887 struct mem_cgroup_event *event;
3888 struct cgroup_subsys_state *cfile_css;
3889 unsigned int efd, cfd;
3896 buf = strstrip(buf);
3898 efd = simple_strtoul(buf, &endp, 10);
3903 cfd = simple_strtoul(buf, &endp, 10);
3904 if ((*endp != ' ') && (*endp != '\0'))
3908 event = kzalloc(sizeof(*event), GFP_KERNEL);
3912 event->memcg = memcg;
3913 INIT_LIST_HEAD(&event->list);
3914 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3915 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3916 INIT_WORK(&event->remove, memcg_event_remove);
3924 event->eventfd = eventfd_ctx_fileget(efile.file);
3925 if (IS_ERR(event->eventfd)) {
3926 ret = PTR_ERR(event->eventfd);
3933 goto out_put_eventfd;
3936 /* the process need read permission on control file */
3937 /* AV: shouldn't we check that it's been opened for read instead? */
3938 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3943 * Determine the event callbacks and set them in @event. This used
3944 * to be done via struct cftype but cgroup core no longer knows
3945 * about these events. The following is crude but the whole thing
3946 * is for compatibility anyway.
3948 * DO NOT ADD NEW FILES.
3950 name = cfile.file->f_path.dentry->d_name.name;
3952 if (!strcmp(name, "memory.usage_in_bytes")) {
3953 event->register_event = mem_cgroup_usage_register_event;
3954 event->unregister_event = mem_cgroup_usage_unregister_event;
3955 } else if (!strcmp(name, "memory.oom_control")) {
3956 event->register_event = mem_cgroup_oom_register_event;
3957 event->unregister_event = mem_cgroup_oom_unregister_event;
3958 } else if (!strcmp(name, "memory.pressure_level")) {
3959 event->register_event = vmpressure_register_event;
3960 event->unregister_event = vmpressure_unregister_event;
3961 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3962 event->register_event = memsw_cgroup_usage_register_event;
3963 event->unregister_event = memsw_cgroup_usage_unregister_event;
3970 * Verify @cfile should belong to @css. Also, remaining events are
3971 * automatically removed on cgroup destruction but the removal is
3972 * asynchronous, so take an extra ref on @css.
3974 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3975 &memory_cgrp_subsys);
3977 if (IS_ERR(cfile_css))
3979 if (cfile_css != css) {
3984 ret = event->register_event(memcg, event->eventfd, buf);
3988 efile.file->f_op->poll(efile.file, &event->pt);
3990 spin_lock(&memcg->event_list_lock);
3991 list_add(&event->list, &memcg->event_list);
3992 spin_unlock(&memcg->event_list_lock);
4004 eventfd_ctx_put(event->eventfd);
4013 static struct cftype mem_cgroup_legacy_files[] = {
4015 .name = "usage_in_bytes",
4016 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4017 .read_u64 = mem_cgroup_read_u64,
4020 .name = "max_usage_in_bytes",
4021 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4022 .write = mem_cgroup_reset,
4023 .read_u64 = mem_cgroup_read_u64,
4026 .name = "limit_in_bytes",
4027 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4028 .write = mem_cgroup_write,
4029 .read_u64 = mem_cgroup_read_u64,
4032 .name = "soft_limit_in_bytes",
4033 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4034 .write = mem_cgroup_write,
4035 .read_u64 = mem_cgroup_read_u64,
4039 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4040 .write = mem_cgroup_reset,
4041 .read_u64 = mem_cgroup_read_u64,
4045 .seq_show = memcg_stat_show,
4048 .name = "force_empty",
4049 .write = mem_cgroup_force_empty_write,
4052 .name = "use_hierarchy",
4053 .write_u64 = mem_cgroup_hierarchy_write,
4054 .read_u64 = mem_cgroup_hierarchy_read,
4057 .name = "cgroup.event_control", /* XXX: for compat */
4058 .write = memcg_write_event_control,
4059 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4062 .name = "swappiness",
4063 .read_u64 = mem_cgroup_swappiness_read,
4064 .write_u64 = mem_cgroup_swappiness_write,
4067 .name = "move_charge_at_immigrate",
4068 .read_u64 = mem_cgroup_move_charge_read,
4069 .write_u64 = mem_cgroup_move_charge_write,
4072 .name = "oom_control",
4073 .seq_show = mem_cgroup_oom_control_read,
4074 .write_u64 = mem_cgroup_oom_control_write,
4075 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4078 .name = "pressure_level",
4082 .name = "numa_stat",
4083 .seq_show = memcg_numa_stat_show,
4086 #ifdef CONFIG_MEMCG_KMEM
4088 .name = "kmem.limit_in_bytes",
4089 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4090 .write = mem_cgroup_write,
4091 .read_u64 = mem_cgroup_read_u64,
4094 .name = "kmem.usage_in_bytes",
4095 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4096 .read_u64 = mem_cgroup_read_u64,
4099 .name = "kmem.failcnt",
4100 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4101 .write = mem_cgroup_reset,
4102 .read_u64 = mem_cgroup_read_u64,
4105 .name = "kmem.max_usage_in_bytes",
4106 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4107 .write = mem_cgroup_reset,
4108 .read_u64 = mem_cgroup_read_u64,
4110 #ifdef CONFIG_SLABINFO
4112 .name = "kmem.slabinfo",
4113 .seq_start = slab_start,
4114 .seq_next = slab_next,
4115 .seq_stop = slab_stop,
4116 .seq_show = memcg_slab_show,
4120 { }, /* terminate */
4123 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4125 struct mem_cgroup_per_node *pn;
4126 struct mem_cgroup_per_zone *mz;
4127 int zone, tmp = node;
4129 * This routine is called against possible nodes.
4130 * But it's BUG to call kmalloc() against offline node.
4132 * TODO: this routine can waste much memory for nodes which will
4133 * never be onlined. It's better to use memory hotplug callback
4136 if (!node_state(node, N_NORMAL_MEMORY))
4138 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4142 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4143 mz = &pn->zoneinfo[zone];
4144 lruvec_init(&mz->lruvec);
4145 mz->usage_in_excess = 0;
4146 mz->on_tree = false;
4149 memcg->nodeinfo[node] = pn;
4153 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4155 kfree(memcg->nodeinfo[node]);
4158 static struct mem_cgroup *mem_cgroup_alloc(void)
4160 struct mem_cgroup *memcg;
4163 size = sizeof(struct mem_cgroup);
4164 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4166 memcg = kzalloc(size, GFP_KERNEL);
4170 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4174 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4180 free_percpu(memcg->stat);
4187 * At destroying mem_cgroup, references from swap_cgroup can remain.
4188 * (scanning all at force_empty is too costly...)
4190 * Instead of clearing all references at force_empty, we remember
4191 * the number of reference from swap_cgroup and free mem_cgroup when
4192 * it goes down to 0.
4194 * Removal of cgroup itself succeeds regardless of refs from swap.
4197 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4201 mem_cgroup_remove_from_trees(memcg);
4204 free_mem_cgroup_per_zone_info(memcg, node);
4206 free_percpu(memcg->stat);
4207 memcg_wb_domain_exit(memcg);
4212 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4214 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4216 if (!memcg->memory.parent)
4218 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4220 EXPORT_SYMBOL(parent_mem_cgroup);
4222 static struct cgroup_subsys_state * __ref
4223 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4225 struct mem_cgroup *memcg;
4226 long error = -ENOMEM;
4229 memcg = mem_cgroup_alloc();
4231 return ERR_PTR(error);
4234 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4238 if (parent_css == NULL) {
4239 root_mem_cgroup = memcg;
4240 mem_cgroup_root_css = &memcg->css;
4241 page_counter_init(&memcg->memory, NULL);
4242 memcg->high = PAGE_COUNTER_MAX;
4243 memcg->soft_limit = PAGE_COUNTER_MAX;
4244 page_counter_init(&memcg->memsw, NULL);
4245 page_counter_init(&memcg->kmem, NULL);
4248 memcg->last_scanned_node = MAX_NUMNODES;
4249 INIT_LIST_HEAD(&memcg->oom_notify);
4250 memcg->move_charge_at_immigrate = 0;
4251 mutex_init(&memcg->thresholds_lock);
4252 spin_lock_init(&memcg->move_lock);
4253 vmpressure_init(&memcg->vmpressure);
4254 INIT_LIST_HEAD(&memcg->event_list);
4255 spin_lock_init(&memcg->event_list_lock);
4256 #ifdef CONFIG_MEMCG_KMEM
4257 memcg->kmemcg_id = -1;
4259 #ifdef CONFIG_CGROUP_WRITEBACK
4260 INIT_LIST_HEAD(&memcg->cgwb_list);
4265 __mem_cgroup_free(memcg);
4266 return ERR_PTR(error);
4270 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4272 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4273 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4276 if (css->id > MEM_CGROUP_ID_MAX)
4282 mutex_lock(&memcg_create_mutex);
4284 memcg->use_hierarchy = parent->use_hierarchy;
4285 memcg->oom_kill_disable = parent->oom_kill_disable;
4286 memcg->swappiness = mem_cgroup_swappiness(parent);
4288 if (parent->use_hierarchy) {
4289 page_counter_init(&memcg->memory, &parent->memory);
4290 memcg->high = PAGE_COUNTER_MAX;
4291 memcg->soft_limit = PAGE_COUNTER_MAX;
4292 page_counter_init(&memcg->memsw, &parent->memsw);
4293 page_counter_init(&memcg->kmem, &parent->kmem);
4296 * No need to take a reference to the parent because cgroup
4297 * core guarantees its existence.
4300 page_counter_init(&memcg->memory, NULL);
4301 memcg->high = PAGE_COUNTER_MAX;
4302 memcg->soft_limit = PAGE_COUNTER_MAX;
4303 page_counter_init(&memcg->memsw, NULL);
4304 page_counter_init(&memcg->kmem, NULL);
4306 * Deeper hierachy with use_hierarchy == false doesn't make
4307 * much sense so let cgroup subsystem know about this
4308 * unfortunate state in our controller.
4310 if (parent != root_mem_cgroup)
4311 memory_cgrp_subsys.broken_hierarchy = true;
4313 mutex_unlock(&memcg_create_mutex);
4315 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4320 * Make sure the memcg is initialized: mem_cgroup_iter()
4321 * orders reading memcg->initialized against its callers
4322 * reading the memcg members.
4324 smp_store_release(&memcg->initialized, 1);
4329 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4331 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4332 struct mem_cgroup_event *event, *tmp;
4335 * Unregister events and notify userspace.
4336 * Notify userspace about cgroup removing only after rmdir of cgroup
4337 * directory to avoid race between userspace and kernelspace.
4339 spin_lock(&memcg->event_list_lock);
4340 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4341 list_del_init(&event->list);
4342 schedule_work(&event->remove);
4344 spin_unlock(&memcg->event_list_lock);
4346 vmpressure_cleanup(&memcg->vmpressure);
4348 memcg_deactivate_kmem(memcg);
4350 wb_memcg_offline(memcg);
4353 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4355 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4357 invalidate_reclaim_iterators(memcg);
4360 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4362 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4364 memcg_destroy_kmem(memcg);
4365 __mem_cgroup_free(memcg);
4369 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4370 * @css: the target css
4372 * Reset the states of the mem_cgroup associated with @css. This is
4373 * invoked when the userland requests disabling on the default hierarchy
4374 * but the memcg is pinned through dependency. The memcg should stop
4375 * applying policies and should revert to the vanilla state as it may be
4376 * made visible again.
4378 * The current implementation only resets the essential configurations.
4379 * This needs to be expanded to cover all the visible parts.
4381 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4383 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4385 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4386 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4387 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4389 memcg->high = PAGE_COUNTER_MAX;
4390 memcg->soft_limit = PAGE_COUNTER_MAX;
4391 memcg_wb_domain_size_changed(memcg);
4395 /* Handlers for move charge at task migration. */
4396 static int mem_cgroup_do_precharge(unsigned long count)
4400 /* Try a single bulk charge without reclaim first, kswapd may wake */
4401 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4403 mc.precharge += count;
4407 /* Try charges one by one with reclaim */
4409 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4419 * get_mctgt_type - get target type of moving charge
4420 * @vma: the vma the pte to be checked belongs
4421 * @addr: the address corresponding to the pte to be checked
4422 * @ptent: the pte to be checked
4423 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4426 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4427 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4428 * move charge. if @target is not NULL, the page is stored in target->page
4429 * with extra refcnt got(Callers should handle it).
4430 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4431 * target for charge migration. if @target is not NULL, the entry is stored
4434 * Called with pte lock held.
4441 enum mc_target_type {
4447 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4448 unsigned long addr, pte_t ptent)
4450 struct page *page = vm_normal_page(vma, addr, ptent);
4452 if (!page || !page_mapped(page))
4454 if (PageAnon(page)) {
4455 if (!(mc.flags & MOVE_ANON))
4458 if (!(mc.flags & MOVE_FILE))
4461 if (!get_page_unless_zero(page))
4468 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4469 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4471 struct page *page = NULL;
4472 swp_entry_t ent = pte_to_swp_entry(ptent);
4474 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4477 * Because lookup_swap_cache() updates some statistics counter,
4478 * we call find_get_page() with swapper_space directly.
4480 page = find_get_page(swap_address_space(ent), ent.val);
4481 if (do_swap_account)
4482 entry->val = ent.val;
4487 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4488 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4494 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4495 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4497 struct page *page = NULL;
4498 struct address_space *mapping;
4501 if (!vma->vm_file) /* anonymous vma */
4503 if (!(mc.flags & MOVE_FILE))
4506 mapping = vma->vm_file->f_mapping;
4507 pgoff = linear_page_index(vma, addr);
4509 /* page is moved even if it's not RSS of this task(page-faulted). */
4511 /* shmem/tmpfs may report page out on swap: account for that too. */
4512 if (shmem_mapping(mapping)) {
4513 page = find_get_entry(mapping, pgoff);
4514 if (radix_tree_exceptional_entry(page)) {
4515 swp_entry_t swp = radix_to_swp_entry(page);
4516 if (do_swap_account)
4518 page = find_get_page(swap_address_space(swp), swp.val);
4521 page = find_get_page(mapping, pgoff);
4523 page = find_get_page(mapping, pgoff);
4529 * mem_cgroup_move_account - move account of the page
4531 * @nr_pages: number of regular pages (>1 for huge pages)
4532 * @from: mem_cgroup which the page is moved from.
4533 * @to: mem_cgroup which the page is moved to. @from != @to.
4535 * The caller must confirm following.
4536 * - page is not on LRU (isolate_page() is useful.)
4537 * - compound_lock is held when nr_pages > 1
4539 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4542 static int mem_cgroup_move_account(struct page *page,
4543 unsigned int nr_pages,
4544 struct mem_cgroup *from,
4545 struct mem_cgroup *to)
4547 unsigned long flags;
4551 VM_BUG_ON(from == to);
4552 VM_BUG_ON_PAGE(PageLRU(page), page);
4554 * The page is isolated from LRU. So, collapse function
4555 * will not handle this page. But page splitting can happen.
4556 * Do this check under compound_page_lock(). The caller should
4560 if (nr_pages > 1 && !PageTransHuge(page))
4564 * Prevent mem_cgroup_replace_page() from looking at
4565 * page->mem_cgroup of its source page while we change it.
4567 if (!trylock_page(page))
4571 if (page->mem_cgroup != from)
4574 anon = PageAnon(page);
4576 spin_lock_irqsave(&from->move_lock, flags);
4578 if (!anon && page_mapped(page)) {
4579 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4581 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4586 * move_lock grabbed above and caller set from->moving_account, so
4587 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4588 * So mapping should be stable for dirty pages.
4590 if (!anon && PageDirty(page)) {
4591 struct address_space *mapping = page_mapping(page);
4593 if (mapping_cap_account_dirty(mapping)) {
4594 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4596 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4601 if (PageWriteback(page)) {
4602 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4604 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4609 * It is safe to change page->mem_cgroup here because the page
4610 * is referenced, charged, and isolated - we can't race with
4611 * uncharging, charging, migration, or LRU putback.
4614 /* caller should have done css_get */
4615 page->mem_cgroup = to;
4616 spin_unlock_irqrestore(&from->move_lock, flags);
4620 local_irq_disable();
4621 mem_cgroup_charge_statistics(to, page, nr_pages);
4622 memcg_check_events(to, page);
4623 mem_cgroup_charge_statistics(from, page, -nr_pages);
4624 memcg_check_events(from, page);
4632 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4633 unsigned long addr, pte_t ptent, union mc_target *target)
4635 struct page *page = NULL;
4636 enum mc_target_type ret = MC_TARGET_NONE;
4637 swp_entry_t ent = { .val = 0 };
4639 if (pte_present(ptent))
4640 page = mc_handle_present_pte(vma, addr, ptent);
4641 else if (is_swap_pte(ptent))
4642 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4643 else if (pte_none(ptent))
4644 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4646 if (!page && !ent.val)
4650 * Do only loose check w/o serialization.
4651 * mem_cgroup_move_account() checks the page is valid or
4652 * not under LRU exclusion.
4654 if (page->mem_cgroup == mc.from) {
4655 ret = MC_TARGET_PAGE;
4657 target->page = page;
4659 if (!ret || !target)
4662 /* There is a swap entry and a page doesn't exist or isn't charged */
4663 if (ent.val && !ret &&
4664 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4665 ret = MC_TARGET_SWAP;
4672 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4674 * We don't consider swapping or file mapped pages because THP does not
4675 * support them for now.
4676 * Caller should make sure that pmd_trans_huge(pmd) is true.
4678 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4679 unsigned long addr, pmd_t pmd, union mc_target *target)
4681 struct page *page = NULL;
4682 enum mc_target_type ret = MC_TARGET_NONE;
4684 page = pmd_page(pmd);
4685 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4686 if (!(mc.flags & MOVE_ANON))
4688 if (page->mem_cgroup == mc.from) {
4689 ret = MC_TARGET_PAGE;
4692 target->page = page;
4698 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4699 unsigned long addr, pmd_t pmd, union mc_target *target)
4701 return MC_TARGET_NONE;
4705 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4706 unsigned long addr, unsigned long end,
4707 struct mm_walk *walk)
4709 struct vm_area_struct *vma = walk->vma;
4713 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4714 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4715 mc.precharge += HPAGE_PMD_NR;
4720 if (pmd_trans_unstable(pmd))
4722 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4723 for (; addr != end; pte++, addr += PAGE_SIZE)
4724 if (get_mctgt_type(vma, addr, *pte, NULL))
4725 mc.precharge++; /* increment precharge temporarily */
4726 pte_unmap_unlock(pte - 1, ptl);
4732 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4734 unsigned long precharge;
4736 struct mm_walk mem_cgroup_count_precharge_walk = {
4737 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4740 down_read(&mm->mmap_sem);
4741 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4742 up_read(&mm->mmap_sem);
4744 precharge = mc.precharge;
4750 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4752 unsigned long precharge = mem_cgroup_count_precharge(mm);
4754 VM_BUG_ON(mc.moving_task);
4755 mc.moving_task = current;
4756 return mem_cgroup_do_precharge(precharge);
4759 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4760 static void __mem_cgroup_clear_mc(void)
4762 struct mem_cgroup *from = mc.from;
4763 struct mem_cgroup *to = mc.to;
4765 /* we must uncharge all the leftover precharges from mc.to */
4767 cancel_charge(mc.to, mc.precharge);
4771 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4772 * we must uncharge here.
4774 if (mc.moved_charge) {
4775 cancel_charge(mc.from, mc.moved_charge);
4776 mc.moved_charge = 0;
4778 /* we must fixup refcnts and charges */
4779 if (mc.moved_swap) {
4780 /* uncharge swap account from the old cgroup */
4781 if (!mem_cgroup_is_root(mc.from))
4782 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4785 * we charged both to->memory and to->memsw, so we
4786 * should uncharge to->memory.
4788 if (!mem_cgroup_is_root(mc.to))
4789 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4791 css_put_many(&mc.from->css, mc.moved_swap);
4793 /* we've already done css_get(mc.to) */
4796 memcg_oom_recover(from);
4797 memcg_oom_recover(to);
4798 wake_up_all(&mc.waitq);
4801 static void mem_cgroup_clear_mc(void)
4804 * we must clear moving_task before waking up waiters at the end of
4807 mc.moving_task = NULL;
4808 __mem_cgroup_clear_mc();
4809 spin_lock(&mc.lock);
4812 spin_unlock(&mc.lock);
4815 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4817 struct cgroup_subsys_state *css;
4818 struct mem_cgroup *memcg;
4819 struct mem_cgroup *from;
4820 struct task_struct *leader, *p;
4821 struct mm_struct *mm;
4822 unsigned long move_flags;
4825 /* charge immigration isn't supported on the default hierarchy */
4826 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4830 * Multi-process migrations only happen on the default hierarchy
4831 * where charge immigration is not used. Perform charge
4832 * immigration if @tset contains a leader and whine if there are
4836 cgroup_taskset_for_each_leader(leader, css, tset) {
4839 memcg = mem_cgroup_from_css(css);
4845 * We are now commited to this value whatever it is. Changes in this
4846 * tunable will only affect upcoming migrations, not the current one.
4847 * So we need to save it, and keep it going.
4849 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4853 from = mem_cgroup_from_task(p);
4855 VM_BUG_ON(from == memcg);
4857 mm = get_task_mm(p);
4860 /* We move charges only when we move a owner of the mm */
4861 if (mm->owner == p) {
4864 VM_BUG_ON(mc.precharge);
4865 VM_BUG_ON(mc.moved_charge);
4866 VM_BUG_ON(mc.moved_swap);
4868 spin_lock(&mc.lock);
4871 mc.flags = move_flags;
4872 spin_unlock(&mc.lock);
4873 /* We set mc.moving_task later */
4875 ret = mem_cgroup_precharge_mc(mm);
4877 mem_cgroup_clear_mc();
4883 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4886 mem_cgroup_clear_mc();
4889 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4890 unsigned long addr, unsigned long end,
4891 struct mm_walk *walk)
4894 struct vm_area_struct *vma = walk->vma;
4897 enum mc_target_type target_type;
4898 union mc_target target;
4902 * We don't take compound_lock() here but no race with splitting thp
4904 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
4905 * under splitting, which means there's no concurrent thp split,
4906 * - if another thread runs into split_huge_page() just after we
4907 * entered this if-block, the thread must wait for page table lock
4908 * to be unlocked in __split_huge_page_splitting(), where the main
4909 * part of thp split is not executed yet.
4911 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4912 if (mc.precharge < HPAGE_PMD_NR) {
4916 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4917 if (target_type == MC_TARGET_PAGE) {
4919 if (!isolate_lru_page(page)) {
4920 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
4922 mc.precharge -= HPAGE_PMD_NR;
4923 mc.moved_charge += HPAGE_PMD_NR;
4925 putback_lru_page(page);
4933 if (pmd_trans_unstable(pmd))
4936 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4937 for (; addr != end; addr += PAGE_SIZE) {
4938 pte_t ptent = *(pte++);
4944 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4945 case MC_TARGET_PAGE:
4947 if (isolate_lru_page(page))
4949 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
4951 /* we uncharge from mc.from later. */
4954 putback_lru_page(page);
4955 put: /* get_mctgt_type() gets the page */
4958 case MC_TARGET_SWAP:
4960 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4962 /* we fixup refcnts and charges later. */
4970 pte_unmap_unlock(pte - 1, ptl);
4975 * We have consumed all precharges we got in can_attach().
4976 * We try charge one by one, but don't do any additional
4977 * charges to mc.to if we have failed in charge once in attach()
4980 ret = mem_cgroup_do_precharge(1);
4988 static void mem_cgroup_move_charge(struct mm_struct *mm)
4990 struct mm_walk mem_cgroup_move_charge_walk = {
4991 .pmd_entry = mem_cgroup_move_charge_pte_range,
4995 lru_add_drain_all();
4997 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4998 * move_lock while we're moving its pages to another memcg.
4999 * Then wait for already started RCU-only updates to finish.
5001 atomic_inc(&mc.from->moving_account);
5004 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5006 * Someone who are holding the mmap_sem might be waiting in
5007 * waitq. So we cancel all extra charges, wake up all waiters,
5008 * and retry. Because we cancel precharges, we might not be able
5009 * to move enough charges, but moving charge is a best-effort
5010 * feature anyway, so it wouldn't be a big problem.
5012 __mem_cgroup_clear_mc();
5017 * When we have consumed all precharges and failed in doing
5018 * additional charge, the page walk just aborts.
5020 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5021 up_read(&mm->mmap_sem);
5022 atomic_dec(&mc.from->moving_account);
5025 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5027 struct cgroup_subsys_state *css;
5028 struct task_struct *p = cgroup_taskset_first(tset, &css);
5029 struct mm_struct *mm = get_task_mm(p);
5033 mem_cgroup_move_charge(mm);
5037 mem_cgroup_clear_mc();
5039 #else /* !CONFIG_MMU */
5040 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5044 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5047 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
5053 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5054 * to verify whether we're attached to the default hierarchy on each mount
5057 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5060 * use_hierarchy is forced on the default hierarchy. cgroup core
5061 * guarantees that @root doesn't have any children, so turning it
5062 * on for the root memcg is enough.
5064 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5065 root_mem_cgroup->use_hierarchy = true;
5067 root_mem_cgroup->use_hierarchy = false;
5070 static u64 memory_current_read(struct cgroup_subsys_state *css,
5073 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5075 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5078 static int memory_low_show(struct seq_file *m, void *v)
5080 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5081 unsigned long low = READ_ONCE(memcg->low);
5083 if (low == PAGE_COUNTER_MAX)
5084 seq_puts(m, "max\n");
5086 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5091 static ssize_t memory_low_write(struct kernfs_open_file *of,
5092 char *buf, size_t nbytes, loff_t off)
5094 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5098 buf = strstrip(buf);
5099 err = page_counter_memparse(buf, "max", &low);
5108 static int memory_high_show(struct seq_file *m, void *v)
5110 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5111 unsigned long high = READ_ONCE(memcg->high);
5113 if (high == PAGE_COUNTER_MAX)
5114 seq_puts(m, "max\n");
5116 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5121 static ssize_t memory_high_write(struct kernfs_open_file *of,
5122 char *buf, size_t nbytes, loff_t off)
5124 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5125 unsigned long nr_pages;
5129 buf = strstrip(buf);
5130 err = page_counter_memparse(buf, "max", &high);
5136 nr_pages = page_counter_read(&memcg->memory);
5137 if (nr_pages > high)
5138 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5141 memcg_wb_domain_size_changed(memcg);
5145 static int memory_max_show(struct seq_file *m, void *v)
5147 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5148 unsigned long max = READ_ONCE(memcg->memory.limit);
5150 if (max == PAGE_COUNTER_MAX)
5151 seq_puts(m, "max\n");
5153 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5158 static ssize_t memory_max_write(struct kernfs_open_file *of,
5159 char *buf, size_t nbytes, loff_t off)
5161 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5162 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5163 bool drained = false;
5167 buf = strstrip(buf);
5168 err = page_counter_memparse(buf, "max", &max);
5172 xchg(&memcg->memory.limit, max);
5175 unsigned long nr_pages = page_counter_read(&memcg->memory);
5177 if (nr_pages <= max)
5180 if (signal_pending(current)) {
5186 drain_all_stock(memcg);
5192 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5198 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5199 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5203 memcg_wb_domain_size_changed(memcg);
5207 static int memory_events_show(struct seq_file *m, void *v)
5209 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5211 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5212 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5213 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5214 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5219 static struct cftype memory_files[] = {
5222 .flags = CFTYPE_NOT_ON_ROOT,
5223 .read_u64 = memory_current_read,
5227 .flags = CFTYPE_NOT_ON_ROOT,
5228 .seq_show = memory_low_show,
5229 .write = memory_low_write,
5233 .flags = CFTYPE_NOT_ON_ROOT,
5234 .seq_show = memory_high_show,
5235 .write = memory_high_write,
5239 .flags = CFTYPE_NOT_ON_ROOT,
5240 .seq_show = memory_max_show,
5241 .write = memory_max_write,
5245 .flags = CFTYPE_NOT_ON_ROOT,
5246 .file_offset = offsetof(struct mem_cgroup, events_file),
5247 .seq_show = memory_events_show,
5252 struct cgroup_subsys memory_cgrp_subsys = {
5253 .css_alloc = mem_cgroup_css_alloc,
5254 .css_online = mem_cgroup_css_online,
5255 .css_offline = mem_cgroup_css_offline,
5256 .css_released = mem_cgroup_css_released,
5257 .css_free = mem_cgroup_css_free,
5258 .css_reset = mem_cgroup_css_reset,
5259 .can_attach = mem_cgroup_can_attach,
5260 .cancel_attach = mem_cgroup_cancel_attach,
5261 .attach = mem_cgroup_move_task,
5262 .bind = mem_cgroup_bind,
5263 .dfl_cftypes = memory_files,
5264 .legacy_cftypes = mem_cgroup_legacy_files,
5269 * mem_cgroup_low - check if memory consumption is below the normal range
5270 * @root: the highest ancestor to consider
5271 * @memcg: the memory cgroup to check
5273 * Returns %true if memory consumption of @memcg, and that of all
5274 * configurable ancestors up to @root, is below the normal range.
5276 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5278 if (mem_cgroup_disabled())
5282 * The toplevel group doesn't have a configurable range, so
5283 * it's never low when looked at directly, and it is not
5284 * considered an ancestor when assessing the hierarchy.
5287 if (memcg == root_mem_cgroup)
5290 if (page_counter_read(&memcg->memory) >= memcg->low)
5293 while (memcg != root) {
5294 memcg = parent_mem_cgroup(memcg);
5296 if (memcg == root_mem_cgroup)
5299 if (page_counter_read(&memcg->memory) >= memcg->low)
5306 * mem_cgroup_try_charge - try charging a page
5307 * @page: page to charge
5308 * @mm: mm context of the victim
5309 * @gfp_mask: reclaim mode
5310 * @memcgp: charged memcg return
5312 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5313 * pages according to @gfp_mask if necessary.
5315 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5316 * Otherwise, an error code is returned.
5318 * After page->mapping has been set up, the caller must finalize the
5319 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5320 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5322 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5323 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5325 struct mem_cgroup *memcg = NULL;
5326 unsigned int nr_pages = 1;
5329 if (mem_cgroup_disabled())
5332 if (PageSwapCache(page)) {
5334 * Every swap fault against a single page tries to charge the
5335 * page, bail as early as possible. shmem_unuse() encounters
5336 * already charged pages, too. The USED bit is protected by
5337 * the page lock, which serializes swap cache removal, which
5338 * in turn serializes uncharging.
5340 VM_BUG_ON_PAGE(!PageLocked(page), page);
5341 if (page->mem_cgroup)
5344 if (do_swap_account) {
5345 swp_entry_t ent = { .val = page_private(page), };
5346 unsigned short id = lookup_swap_cgroup_id(ent);
5349 memcg = mem_cgroup_from_id(id);
5350 if (memcg && !css_tryget_online(&memcg->css))
5356 if (PageTransHuge(page)) {
5357 nr_pages <<= compound_order(page);
5358 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5362 memcg = get_mem_cgroup_from_mm(mm);
5364 ret = try_charge(memcg, gfp_mask, nr_pages);
5366 css_put(&memcg->css);
5373 * mem_cgroup_commit_charge - commit a page charge
5374 * @page: page to charge
5375 * @memcg: memcg to charge the page to
5376 * @lrucare: page might be on LRU already
5378 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5379 * after page->mapping has been set up. This must happen atomically
5380 * as part of the page instantiation, i.e. under the page table lock
5381 * for anonymous pages, under the page lock for page and swap cache.
5383 * In addition, the page must not be on the LRU during the commit, to
5384 * prevent racing with task migration. If it might be, use @lrucare.
5386 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5388 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5391 unsigned int nr_pages = 1;
5393 VM_BUG_ON_PAGE(!page->mapping, page);
5394 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5396 if (mem_cgroup_disabled())
5399 * Swap faults will attempt to charge the same page multiple
5400 * times. But reuse_swap_page() might have removed the page
5401 * from swapcache already, so we can't check PageSwapCache().
5406 commit_charge(page, memcg, lrucare);
5408 if (PageTransHuge(page)) {
5409 nr_pages <<= compound_order(page);
5410 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5413 local_irq_disable();
5414 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5415 memcg_check_events(memcg, page);
5418 if (do_swap_account && PageSwapCache(page)) {
5419 swp_entry_t entry = { .val = page_private(page) };
5421 * The swap entry might not get freed for a long time,
5422 * let's not wait for it. The page already received a
5423 * memory+swap charge, drop the swap entry duplicate.
5425 mem_cgroup_uncharge_swap(entry);
5430 * mem_cgroup_cancel_charge - cancel a page charge
5431 * @page: page to charge
5432 * @memcg: memcg to charge the page to
5434 * Cancel a charge transaction started by mem_cgroup_try_charge().
5436 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5438 unsigned int nr_pages = 1;
5440 if (mem_cgroup_disabled())
5443 * Swap faults will attempt to charge the same page multiple
5444 * times. But reuse_swap_page() might have removed the page
5445 * from swapcache already, so we can't check PageSwapCache().
5450 if (PageTransHuge(page)) {
5451 nr_pages <<= compound_order(page);
5452 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5455 cancel_charge(memcg, nr_pages);
5458 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5459 unsigned long nr_anon, unsigned long nr_file,
5460 unsigned long nr_huge, struct page *dummy_page)
5462 unsigned long nr_pages = nr_anon + nr_file;
5463 unsigned long flags;
5465 if (!mem_cgroup_is_root(memcg)) {
5466 page_counter_uncharge(&memcg->memory, nr_pages);
5467 if (do_swap_account)
5468 page_counter_uncharge(&memcg->memsw, nr_pages);
5469 memcg_oom_recover(memcg);
5472 local_irq_save(flags);
5473 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5474 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5475 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5476 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5477 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5478 memcg_check_events(memcg, dummy_page);
5479 local_irq_restore(flags);
5481 if (!mem_cgroup_is_root(memcg))
5482 css_put_many(&memcg->css, nr_pages);
5485 static void uncharge_list(struct list_head *page_list)
5487 struct mem_cgroup *memcg = NULL;
5488 unsigned long nr_anon = 0;
5489 unsigned long nr_file = 0;
5490 unsigned long nr_huge = 0;
5491 unsigned long pgpgout = 0;
5492 struct list_head *next;
5495 next = page_list->next;
5497 unsigned int nr_pages = 1;
5499 page = list_entry(next, struct page, lru);
5500 next = page->lru.next;
5502 VM_BUG_ON_PAGE(PageLRU(page), page);
5503 VM_BUG_ON_PAGE(page_count(page), page);
5505 if (!page->mem_cgroup)
5509 * Nobody should be changing or seriously looking at
5510 * page->mem_cgroup at this point, we have fully
5511 * exclusive access to the page.
5514 if (memcg != page->mem_cgroup) {
5516 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5518 pgpgout = nr_anon = nr_file = nr_huge = 0;
5520 memcg = page->mem_cgroup;
5523 if (PageTransHuge(page)) {
5524 nr_pages <<= compound_order(page);
5525 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5526 nr_huge += nr_pages;
5530 nr_anon += nr_pages;
5532 nr_file += nr_pages;
5534 page->mem_cgroup = NULL;
5537 } while (next != page_list);
5540 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5545 * mem_cgroup_uncharge - uncharge a page
5546 * @page: page to uncharge
5548 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5549 * mem_cgroup_commit_charge().
5551 void mem_cgroup_uncharge(struct page *page)
5553 if (mem_cgroup_disabled())
5556 /* Don't touch page->lru of any random page, pre-check: */
5557 if (!page->mem_cgroup)
5560 INIT_LIST_HEAD(&page->lru);
5561 uncharge_list(&page->lru);
5565 * mem_cgroup_uncharge_list - uncharge a list of page
5566 * @page_list: list of pages to uncharge
5568 * Uncharge a list of pages previously charged with
5569 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5571 void mem_cgroup_uncharge_list(struct list_head *page_list)
5573 if (mem_cgroup_disabled())
5576 if (!list_empty(page_list))
5577 uncharge_list(page_list);
5581 * mem_cgroup_replace_page - migrate a charge to another page
5582 * @oldpage: currently charged page
5583 * @newpage: page to transfer the charge to
5585 * Migrate the charge from @oldpage to @newpage.
5587 * Both pages must be locked, @newpage->mapping must be set up.
5588 * Either or both pages might be on the LRU already.
5590 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5592 struct mem_cgroup *memcg;
5595 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5596 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5597 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5598 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5601 if (mem_cgroup_disabled())
5604 /* Page cache replacement: new page already charged? */
5605 if (newpage->mem_cgroup)
5608 /* Swapcache readahead pages can get replaced before being charged */
5609 memcg = oldpage->mem_cgroup;
5613 lock_page_lru(oldpage, &isolated);
5614 oldpage->mem_cgroup = NULL;
5615 unlock_page_lru(oldpage, isolated);
5617 commit_charge(newpage, memcg, true);
5621 * subsys_initcall() for memory controller.
5623 * Some parts like hotcpu_notifier() have to be initialized from this context
5624 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5625 * everything that doesn't depend on a specific mem_cgroup structure should
5626 * be initialized from here.
5628 static int __init mem_cgroup_init(void)
5632 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5634 for_each_possible_cpu(cpu)
5635 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5638 for_each_node(node) {
5639 struct mem_cgroup_tree_per_node *rtpn;
5642 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5643 node_online(node) ? node : NUMA_NO_NODE);
5645 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5646 struct mem_cgroup_tree_per_zone *rtpz;
5648 rtpz = &rtpn->rb_tree_per_zone[zone];
5649 rtpz->rb_root = RB_ROOT;
5650 spin_lock_init(&rtpz->lock);
5652 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5657 subsys_initcall(mem_cgroup_init);
5659 #ifdef CONFIG_MEMCG_SWAP
5661 * mem_cgroup_swapout - transfer a memsw charge to swap
5662 * @page: page whose memsw charge to transfer
5663 * @entry: swap entry to move the charge to
5665 * Transfer the memsw charge of @page to @entry.
5667 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5669 struct mem_cgroup *memcg;
5670 unsigned short oldid;
5672 VM_BUG_ON_PAGE(PageLRU(page), page);
5673 VM_BUG_ON_PAGE(page_count(page), page);
5675 if (!do_swap_account)
5678 memcg = page->mem_cgroup;
5680 /* Readahead page, never charged */
5684 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5685 VM_BUG_ON_PAGE(oldid, page);
5686 mem_cgroup_swap_statistics(memcg, true);
5688 page->mem_cgroup = NULL;
5690 if (!mem_cgroup_is_root(memcg))
5691 page_counter_uncharge(&memcg->memory, 1);
5694 * Interrupts should be disabled here because the caller holds the
5695 * mapping->tree_lock lock which is taken with interrupts-off. It is
5696 * important here to have the interrupts disabled because it is the
5697 * only synchronisation we have for udpating the per-CPU variables.
5699 VM_BUG_ON(!irqs_disabled());
5700 mem_cgroup_charge_statistics(memcg, page, -1);
5701 memcg_check_events(memcg, page);
5705 * mem_cgroup_uncharge_swap - uncharge a swap entry
5706 * @entry: swap entry to uncharge
5708 * Drop the memsw charge associated with @entry.
5710 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5712 struct mem_cgroup *memcg;
5715 if (!do_swap_account)
5718 id = swap_cgroup_record(entry, 0);
5720 memcg = mem_cgroup_from_id(id);
5722 if (!mem_cgroup_is_root(memcg))
5723 page_counter_uncharge(&memcg->memsw, 1);
5724 mem_cgroup_swap_statistics(memcg, false);
5725 css_put(&memcg->css);
5730 /* for remember boot option*/
5731 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5732 static int really_do_swap_account __initdata = 1;
5734 static int really_do_swap_account __initdata;
5737 static int __init enable_swap_account(char *s)
5739 if (!strcmp(s, "1"))
5740 really_do_swap_account = 1;
5741 else if (!strcmp(s, "0"))
5742 really_do_swap_account = 0;
5745 __setup("swapaccount=", enable_swap_account);
5747 static struct cftype memsw_cgroup_files[] = {
5749 .name = "memsw.usage_in_bytes",
5750 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5751 .read_u64 = mem_cgroup_read_u64,
5754 .name = "memsw.max_usage_in_bytes",
5755 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5756 .write = mem_cgroup_reset,
5757 .read_u64 = mem_cgroup_read_u64,
5760 .name = "memsw.limit_in_bytes",
5761 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5762 .write = mem_cgroup_write,
5763 .read_u64 = mem_cgroup_read_u64,
5766 .name = "memsw.failcnt",
5767 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5768 .write = mem_cgroup_reset,
5769 .read_u64 = mem_cgroup_read_u64,
5771 { }, /* terminate */
5774 static int __init mem_cgroup_swap_init(void)
5776 if (!mem_cgroup_disabled() && really_do_swap_account) {
5777 do_swap_account = 1;
5778 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5779 memsw_cgroup_files));
5783 subsys_initcall(mem_cgroup_swap_init);
5785 #endif /* CONFIG_MEMCG_SWAP */