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>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 struct mem_cgroup *root_mem_cgroup __read_mostly;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char * const mem_cgroup_stat_names[] = {
111 static const char * const mem_cgroup_events_names[] = {
118 static const char * const mem_cgroup_lru_names[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_zone {
136 struct rb_root rb_root;
140 struct mem_cgroup_tree_per_node {
141 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
144 struct mem_cgroup_tree {
145 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
148 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
151 struct mem_cgroup_eventfd_list {
152 struct list_head list;
153 struct eventfd_ctx *eventfd;
157 * cgroup_event represents events which userspace want to receive.
159 struct mem_cgroup_event {
161 * memcg which the event belongs to.
163 struct mem_cgroup *memcg;
165 * eventfd to signal userspace about the event.
167 struct eventfd_ctx *eventfd;
169 * Each of these stored in a list by the cgroup.
171 struct list_head list;
173 * register_event() callback will be used to add new userspace
174 * waiter for changes related to this event. Use eventfd_signal()
175 * on eventfd to send notification to userspace.
177 int (*register_event)(struct mem_cgroup *memcg,
178 struct eventfd_ctx *eventfd, const char *args);
180 * unregister_event() callback will be called when userspace closes
181 * the eventfd or on cgroup removing. This callback must be set,
182 * if you want provide notification functionality.
184 void (*unregister_event)(struct mem_cgroup *memcg,
185 struct eventfd_ctx *eventfd);
187 * All fields below needed to unregister event when
188 * userspace closes eventfd.
191 wait_queue_head_t *wqh;
193 struct work_struct remove;
196 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
197 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
199 /* Stuffs for move charges at task migration. */
201 * Types of charges to be moved.
203 #define MOVE_ANON 0x1U
204 #define MOVE_FILE 0x2U
205 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
207 /* "mc" and its members are protected by cgroup_mutex */
208 static struct move_charge_struct {
209 spinlock_t lock; /* for from, to */
210 struct mem_cgroup *from;
211 struct mem_cgroup *to;
213 unsigned long precharge;
214 unsigned long moved_charge;
215 unsigned long moved_swap;
216 struct task_struct *moving_task; /* a task moving charges */
217 wait_queue_head_t waitq; /* a waitq for other context */
219 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
220 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
224 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
225 * limit reclaim to prevent infinite loops, if they ever occur.
227 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
228 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
231 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
232 MEM_CGROUP_CHARGE_TYPE_ANON,
233 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
234 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
238 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
253 /* Some nice accessors for the vmpressure. */
254 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
257 memcg = root_mem_cgroup;
258 return &memcg->vmpressure;
261 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
263 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
266 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
268 return (memcg == root_mem_cgroup);
272 * We restrict the id in the range of [1, 65535], so it can fit into
275 #define MEM_CGROUP_ID_MAX USHRT_MAX
277 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
279 return memcg->css.id;
283 * A helper function to get mem_cgroup from ID. must be called under
284 * rcu_read_lock(). The caller is responsible for calling
285 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
286 * refcnt from swap can be called against removed memcg.)
288 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
290 struct cgroup_subsys_state *css;
292 css = css_from_id(id, &memory_cgrp_subsys);
293 return mem_cgroup_from_css(css);
298 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
299 * The main reason for not using cgroup id for this:
300 * this works better in sparse environments, where we have a lot of memcgs,
301 * but only a few kmem-limited. Or also, if we have, for instance, 200
302 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
303 * 200 entry array for that.
305 * The current size of the caches array is stored in memcg_nr_cache_ids. It
306 * will double each time we have to increase it.
308 static DEFINE_IDA(memcg_cache_ida);
309 int memcg_nr_cache_ids;
311 /* Protects memcg_nr_cache_ids */
312 static DECLARE_RWSEM(memcg_cache_ids_sem);
314 void memcg_get_cache_ids(void)
316 down_read(&memcg_cache_ids_sem);
319 void memcg_put_cache_ids(void)
321 up_read(&memcg_cache_ids_sem);
325 * MIN_SIZE is different than 1, because we would like to avoid going through
326 * the alloc/free process all the time. In a small machine, 4 kmem-limited
327 * cgroups is a reasonable guess. In the future, it could be a parameter or
328 * tunable, but that is strictly not necessary.
330 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
331 * this constant directly from cgroup, but it is understandable that this is
332 * better kept as an internal representation in cgroup.c. In any case, the
333 * cgrp_id space is not getting any smaller, and we don't have to necessarily
334 * increase ours as well if it increases.
336 #define MEMCG_CACHES_MIN_SIZE 4
337 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
340 * A lot of the calls to the cache allocation functions are expected to be
341 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
342 * conditional to this static branch, we'll have to allow modules that does
343 * kmem_cache_alloc and the such to see this symbol as well
345 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
346 EXPORT_SYMBOL(memcg_kmem_enabled_key);
348 #endif /* !CONFIG_SLOB */
350 static struct mem_cgroup_per_zone *
351 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
353 int nid = zone_to_nid(zone);
354 int zid = zone_idx(zone);
356 return &memcg->nodeinfo[nid]->zoneinfo[zid];
360 * mem_cgroup_css_from_page - css of the memcg associated with a page
361 * @page: page of interest
363 * If memcg is bound to the default hierarchy, css of the memcg associated
364 * with @page is returned. The returned css remains associated with @page
365 * until it is released.
367 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
370 * XXX: The above description of behavior on the default hierarchy isn't
371 * strictly true yet as replace_page_cache_page() can modify the
372 * association before @page is released even on the default hierarchy;
373 * however, the current and planned usages don't mix the the two functions
374 * and replace_page_cache_page() will soon be updated to make the invariant
377 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
379 struct mem_cgroup *memcg;
381 memcg = page->mem_cgroup;
383 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
384 memcg = root_mem_cgroup;
390 * page_cgroup_ino - return inode number of the memcg a page is charged to
393 * Look up the closest online ancestor of the memory cgroup @page is charged to
394 * and return its inode number or 0 if @page is not charged to any cgroup. It
395 * is safe to call this function without holding a reference to @page.
397 * Note, this function is inherently racy, because there is nothing to prevent
398 * the cgroup inode from getting torn down and potentially reallocated a moment
399 * after page_cgroup_ino() returns, so it only should be used by callers that
400 * do not care (such as procfs interfaces).
402 ino_t page_cgroup_ino(struct page *page)
404 struct mem_cgroup *memcg;
405 unsigned long ino = 0;
408 memcg = READ_ONCE(page->mem_cgroup);
409 while (memcg && !(memcg->css.flags & CSS_ONLINE))
410 memcg = parent_mem_cgroup(memcg);
412 ino = cgroup_ino(memcg->css.cgroup);
417 static struct mem_cgroup_per_zone *
418 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
420 int nid = page_to_nid(page);
421 int zid = page_zonenum(page);
423 return &memcg->nodeinfo[nid]->zoneinfo[zid];
426 static struct mem_cgroup_tree_per_zone *
427 soft_limit_tree_node_zone(int nid, int zid)
429 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
432 static struct mem_cgroup_tree_per_zone *
433 soft_limit_tree_from_page(struct page *page)
435 int nid = page_to_nid(page);
436 int zid = page_zonenum(page);
438 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
441 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
442 struct mem_cgroup_tree_per_zone *mctz,
443 unsigned long new_usage_in_excess)
445 struct rb_node **p = &mctz->rb_root.rb_node;
446 struct rb_node *parent = NULL;
447 struct mem_cgroup_per_zone *mz_node;
452 mz->usage_in_excess = new_usage_in_excess;
453 if (!mz->usage_in_excess)
457 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
459 if (mz->usage_in_excess < mz_node->usage_in_excess)
462 * We can't avoid mem cgroups that are over their soft
463 * limit by the same amount
465 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
468 rb_link_node(&mz->tree_node, parent, p);
469 rb_insert_color(&mz->tree_node, &mctz->rb_root);
473 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
474 struct mem_cgroup_tree_per_zone *mctz)
478 rb_erase(&mz->tree_node, &mctz->rb_root);
482 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
483 struct mem_cgroup_tree_per_zone *mctz)
487 spin_lock_irqsave(&mctz->lock, flags);
488 __mem_cgroup_remove_exceeded(mz, mctz);
489 spin_unlock_irqrestore(&mctz->lock, flags);
492 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
494 unsigned long nr_pages = page_counter_read(&memcg->memory);
495 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
496 unsigned long excess = 0;
498 if (nr_pages > soft_limit)
499 excess = nr_pages - soft_limit;
504 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
506 unsigned long excess;
507 struct mem_cgroup_per_zone *mz;
508 struct mem_cgroup_tree_per_zone *mctz;
510 mctz = soft_limit_tree_from_page(page);
512 * Necessary to update all ancestors when hierarchy is used.
513 * because their event counter is not touched.
515 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
516 mz = mem_cgroup_page_zoneinfo(memcg, page);
517 excess = soft_limit_excess(memcg);
519 * We have to update the tree if mz is on RB-tree or
520 * mem is over its softlimit.
522 if (excess || mz->on_tree) {
525 spin_lock_irqsave(&mctz->lock, flags);
526 /* if on-tree, remove it */
528 __mem_cgroup_remove_exceeded(mz, mctz);
530 * Insert again. mz->usage_in_excess will be updated.
531 * If excess is 0, no tree ops.
533 __mem_cgroup_insert_exceeded(mz, mctz, excess);
534 spin_unlock_irqrestore(&mctz->lock, flags);
539 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
541 struct mem_cgroup_tree_per_zone *mctz;
542 struct mem_cgroup_per_zone *mz;
546 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
547 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
548 mctz = soft_limit_tree_node_zone(nid, zid);
549 mem_cgroup_remove_exceeded(mz, mctz);
554 static struct mem_cgroup_per_zone *
555 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
557 struct rb_node *rightmost = NULL;
558 struct mem_cgroup_per_zone *mz;
562 rightmost = rb_last(&mctz->rb_root);
564 goto done; /* Nothing to reclaim from */
566 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
568 * Remove the node now but someone else can add it back,
569 * we will to add it back at the end of reclaim to its correct
570 * position in the tree.
572 __mem_cgroup_remove_exceeded(mz, mctz);
573 if (!soft_limit_excess(mz->memcg) ||
574 !css_tryget_online(&mz->memcg->css))
580 static struct mem_cgroup_per_zone *
581 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
583 struct mem_cgroup_per_zone *mz;
585 spin_lock_irq(&mctz->lock);
586 mz = __mem_cgroup_largest_soft_limit_node(mctz);
587 spin_unlock_irq(&mctz->lock);
592 * Return page count for single (non recursive) @memcg.
594 * Implementation Note: reading percpu statistics for memcg.
596 * Both of vmstat[] and percpu_counter has threshold and do periodic
597 * synchronization to implement "quick" read. There are trade-off between
598 * reading cost and precision of value. Then, we may have a chance to implement
599 * a periodic synchronization of counter in memcg's counter.
601 * But this _read() function is used for user interface now. The user accounts
602 * memory usage by memory cgroup and he _always_ requires exact value because
603 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
604 * have to visit all online cpus and make sum. So, for now, unnecessary
605 * synchronization is not implemented. (just implemented for cpu hotplug)
607 * If there are kernel internal actions which can make use of some not-exact
608 * value, and reading all cpu value can be performance bottleneck in some
609 * common workload, threshold and synchronization as vmstat[] should be
613 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
618 /* Per-cpu values can be negative, use a signed accumulator */
619 for_each_possible_cpu(cpu)
620 val += per_cpu(memcg->stat->count[idx], cpu);
622 * Summing races with updates, so val may be negative. Avoid exposing
623 * transient negative values.
630 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
631 enum mem_cgroup_events_index idx)
633 unsigned long val = 0;
636 for_each_possible_cpu(cpu)
637 val += per_cpu(memcg->stat->events[idx], cpu);
641 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
643 bool compound, int nr_pages)
646 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
647 * counted as CACHE even if it's on ANON LRU.
650 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
653 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
657 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
658 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
662 /* pagein of a big page is an event. So, ignore page size */
664 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
666 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
667 nr_pages = -nr_pages; /* for event */
670 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
673 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
675 unsigned int lru_mask)
677 unsigned long nr = 0;
680 VM_BUG_ON((unsigned)nid >= nr_node_ids);
682 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
683 struct mem_cgroup_per_zone *mz;
687 if (!(BIT(lru) & lru_mask))
689 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
690 nr += mz->lru_size[lru];
696 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
697 unsigned int lru_mask)
699 unsigned long nr = 0;
702 for_each_node_state(nid, N_MEMORY)
703 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
707 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
708 enum mem_cgroup_events_target target)
710 unsigned long val, next;
712 val = __this_cpu_read(memcg->stat->nr_page_events);
713 next = __this_cpu_read(memcg->stat->targets[target]);
714 /* from time_after() in jiffies.h */
715 if ((long)next - (long)val < 0) {
717 case MEM_CGROUP_TARGET_THRESH:
718 next = val + THRESHOLDS_EVENTS_TARGET;
720 case MEM_CGROUP_TARGET_SOFTLIMIT:
721 next = val + SOFTLIMIT_EVENTS_TARGET;
723 case MEM_CGROUP_TARGET_NUMAINFO:
724 next = val + NUMAINFO_EVENTS_TARGET;
729 __this_cpu_write(memcg->stat->targets[target], next);
736 * Check events in order.
739 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
741 /* threshold event is triggered in finer grain than soft limit */
742 if (unlikely(mem_cgroup_event_ratelimit(memcg,
743 MEM_CGROUP_TARGET_THRESH))) {
745 bool do_numainfo __maybe_unused;
747 do_softlimit = mem_cgroup_event_ratelimit(memcg,
748 MEM_CGROUP_TARGET_SOFTLIMIT);
750 do_numainfo = mem_cgroup_event_ratelimit(memcg,
751 MEM_CGROUP_TARGET_NUMAINFO);
753 mem_cgroup_threshold(memcg);
754 if (unlikely(do_softlimit))
755 mem_cgroup_update_tree(memcg, page);
757 if (unlikely(do_numainfo))
758 atomic_inc(&memcg->numainfo_events);
763 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
766 * mm_update_next_owner() may clear mm->owner to NULL
767 * if it races with swapoff, page migration, etc.
768 * So this can be called with p == NULL.
773 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
775 EXPORT_SYMBOL(mem_cgroup_from_task);
777 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
779 struct mem_cgroup *memcg = NULL;
784 * Page cache insertions can happen withou an
785 * actual mm context, e.g. during disk probing
786 * on boot, loopback IO, acct() writes etc.
789 memcg = root_mem_cgroup;
791 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
792 if (unlikely(!memcg))
793 memcg = root_mem_cgroup;
795 } while (!css_tryget_online(&memcg->css));
801 * mem_cgroup_iter - iterate over memory cgroup hierarchy
802 * @root: hierarchy root
803 * @prev: previously returned memcg, NULL on first invocation
804 * @reclaim: cookie for shared reclaim walks, NULL for full walks
806 * Returns references to children of the hierarchy below @root, or
807 * @root itself, or %NULL after a full round-trip.
809 * Caller must pass the return value in @prev on subsequent
810 * invocations for reference counting, or use mem_cgroup_iter_break()
811 * to cancel a hierarchy walk before the round-trip is complete.
813 * Reclaimers can specify a zone and a priority level in @reclaim to
814 * divide up the memcgs in the hierarchy among all concurrent
815 * reclaimers operating on the same zone and priority.
817 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
818 struct mem_cgroup *prev,
819 struct mem_cgroup_reclaim_cookie *reclaim)
821 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
822 struct cgroup_subsys_state *css = NULL;
823 struct mem_cgroup *memcg = NULL;
824 struct mem_cgroup *pos = NULL;
826 if (mem_cgroup_disabled())
830 root = root_mem_cgroup;
832 if (prev && !reclaim)
835 if (!root->use_hierarchy && root != root_mem_cgroup) {
844 struct mem_cgroup_per_zone *mz;
846 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
847 iter = &mz->iter[reclaim->priority];
849 if (prev && reclaim->generation != iter->generation)
853 pos = READ_ONCE(iter->position);
854 if (!pos || css_tryget(&pos->css))
857 * css reference reached zero, so iter->position will
858 * be cleared by ->css_released. However, we should not
859 * rely on this happening soon, because ->css_released
860 * is called from a work queue, and by busy-waiting we
861 * might block it. So we clear iter->position right
864 (void)cmpxchg(&iter->position, pos, NULL);
872 css = css_next_descendant_pre(css, &root->css);
875 * Reclaimers share the hierarchy walk, and a
876 * new one might jump in right at the end of
877 * the hierarchy - make sure they see at least
878 * one group and restart from the beginning.
886 * Verify the css and acquire a reference. The root
887 * is provided by the caller, so we know it's alive
888 * and kicking, and don't take an extra reference.
890 memcg = mem_cgroup_from_css(css);
892 if (css == &root->css)
903 * The position could have already been updated by a competing
904 * thread, so check that the value hasn't changed since we read
905 * it to avoid reclaiming from the same cgroup twice.
907 (void)cmpxchg(&iter->position, pos, memcg);
915 reclaim->generation = iter->generation;
921 if (prev && prev != root)
928 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
929 * @root: hierarchy root
930 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
932 void mem_cgroup_iter_break(struct mem_cgroup *root,
933 struct mem_cgroup *prev)
936 root = root_mem_cgroup;
937 if (prev && prev != root)
941 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
943 struct mem_cgroup *memcg = dead_memcg;
944 struct mem_cgroup_reclaim_iter *iter;
945 struct mem_cgroup_per_zone *mz;
949 while ((memcg = parent_mem_cgroup(memcg))) {
951 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
952 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
953 for (i = 0; i <= DEF_PRIORITY; i++) {
955 cmpxchg(&iter->position,
964 * Iteration constructs for visiting all cgroups (under a tree). If
965 * loops are exited prematurely (break), mem_cgroup_iter_break() must
966 * be used for reference counting.
968 #define for_each_mem_cgroup_tree(iter, root) \
969 for (iter = mem_cgroup_iter(root, NULL, NULL); \
971 iter = mem_cgroup_iter(root, iter, NULL))
973 #define for_each_mem_cgroup(iter) \
974 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
976 iter = mem_cgroup_iter(NULL, iter, NULL))
979 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
980 * @zone: zone of the wanted lruvec
981 * @memcg: memcg of the wanted lruvec
983 * Returns the lru list vector holding pages for the given @zone and
984 * @mem. This can be the global zone lruvec, if the memory controller
987 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
988 struct mem_cgroup *memcg)
990 struct mem_cgroup_per_zone *mz;
991 struct lruvec *lruvec;
993 if (mem_cgroup_disabled()) {
994 lruvec = &zone->lruvec;
998 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
999 lruvec = &mz->lruvec;
1002 * Since a node can be onlined after the mem_cgroup was created,
1003 * we have to be prepared to initialize lruvec->zone here;
1004 * and if offlined then reonlined, we need to reinitialize it.
1006 if (unlikely(lruvec->zone != zone))
1007 lruvec->zone = zone;
1012 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1014 * @zone: zone of the page
1016 * This function is only safe when following the LRU page isolation
1017 * and putback protocol: the LRU lock must be held, and the page must
1018 * either be PageLRU() or the caller must have isolated/allocated it.
1020 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1022 struct mem_cgroup_per_zone *mz;
1023 struct mem_cgroup *memcg;
1024 struct lruvec *lruvec;
1026 if (mem_cgroup_disabled()) {
1027 lruvec = &zone->lruvec;
1031 memcg = page->mem_cgroup;
1033 * Swapcache readahead pages are added to the LRU - and
1034 * possibly migrated - before they are charged.
1037 memcg = root_mem_cgroup;
1039 mz = mem_cgroup_page_zoneinfo(memcg, page);
1040 lruvec = &mz->lruvec;
1043 * Since a node can be onlined after the mem_cgroup was created,
1044 * we have to be prepared to initialize lruvec->zone here;
1045 * and if offlined then reonlined, we need to reinitialize it.
1047 if (unlikely(lruvec->zone != zone))
1048 lruvec->zone = zone;
1053 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1054 * @lruvec: mem_cgroup per zone lru vector
1055 * @lru: index of lru list the page is sitting on
1056 * @nr_pages: positive when adding or negative when removing
1058 * This function must be called when a page is added to or removed from an
1061 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1064 struct mem_cgroup_per_zone *mz;
1065 unsigned long *lru_size;
1067 if (mem_cgroup_disabled())
1070 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1071 lru_size = mz->lru_size + lru;
1072 *lru_size += nr_pages;
1073 VM_BUG_ON((long)(*lru_size) < 0);
1076 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1078 struct mem_cgroup *task_memcg;
1079 struct task_struct *p;
1082 p = find_lock_task_mm(task);
1084 task_memcg = get_mem_cgroup_from_mm(p->mm);
1088 * All threads may have already detached their mm's, but the oom
1089 * killer still needs to detect if they have already been oom
1090 * killed to prevent needlessly killing additional tasks.
1093 task_memcg = mem_cgroup_from_task(task);
1094 css_get(&task_memcg->css);
1097 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1098 css_put(&task_memcg->css);
1103 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1104 * @memcg: the memory cgroup
1106 * Returns the maximum amount of memory @mem can be charged with, in
1109 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1111 unsigned long margin = 0;
1112 unsigned long count;
1113 unsigned long limit;
1115 count = page_counter_read(&memcg->memory);
1116 limit = READ_ONCE(memcg->memory.limit);
1118 margin = limit - count;
1120 if (do_memsw_account()) {
1121 count = page_counter_read(&memcg->memsw);
1122 limit = READ_ONCE(memcg->memsw.limit);
1124 margin = min(margin, limit - count);
1131 * A routine for checking "mem" is under move_account() or not.
1133 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1134 * moving cgroups. This is for waiting at high-memory pressure
1137 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1139 struct mem_cgroup *from;
1140 struct mem_cgroup *to;
1143 * Unlike task_move routines, we access mc.to, mc.from not under
1144 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1146 spin_lock(&mc.lock);
1152 ret = mem_cgroup_is_descendant(from, memcg) ||
1153 mem_cgroup_is_descendant(to, memcg);
1155 spin_unlock(&mc.lock);
1159 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1161 if (mc.moving_task && current != mc.moving_task) {
1162 if (mem_cgroup_under_move(memcg)) {
1164 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1165 /* moving charge context might have finished. */
1168 finish_wait(&mc.waitq, &wait);
1175 #define K(x) ((x) << (PAGE_SHIFT-10))
1177 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1178 * @memcg: The memory cgroup that went over limit
1179 * @p: Task that is going to be killed
1181 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1184 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1186 /* oom_info_lock ensures that parallel ooms do not interleave */
1187 static DEFINE_MUTEX(oom_info_lock);
1188 struct mem_cgroup *iter;
1191 mutex_lock(&oom_info_lock);
1195 pr_info("Task in ");
1196 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1197 pr_cont(" killed as a result of limit of ");
1199 pr_info("Memory limit reached of cgroup ");
1202 pr_cont_cgroup_path(memcg->css.cgroup);
1207 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1208 K((u64)page_counter_read(&memcg->memory)),
1209 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1210 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1211 K((u64)page_counter_read(&memcg->memsw)),
1212 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1213 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1214 K((u64)page_counter_read(&memcg->kmem)),
1215 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1217 for_each_mem_cgroup_tree(iter, memcg) {
1218 pr_info("Memory cgroup stats for ");
1219 pr_cont_cgroup_path(iter->css.cgroup);
1222 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1223 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1225 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1226 K(mem_cgroup_read_stat(iter, i)));
1229 for (i = 0; i < NR_LRU_LISTS; i++)
1230 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1231 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1235 mutex_unlock(&oom_info_lock);
1239 * This function returns the number of memcg under hierarchy tree. Returns
1240 * 1(self count) if no children.
1242 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1245 struct mem_cgroup *iter;
1247 for_each_mem_cgroup_tree(iter, memcg)
1253 * Return the memory (and swap, if configured) limit for a memcg.
1255 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1257 unsigned long limit;
1259 limit = memcg->memory.limit;
1260 if (mem_cgroup_swappiness(memcg)) {
1261 unsigned long memsw_limit;
1262 unsigned long swap_limit;
1264 memsw_limit = memcg->memsw.limit;
1265 swap_limit = memcg->swap.limit;
1266 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1267 limit = min(limit + swap_limit, memsw_limit);
1272 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1275 struct oom_control oc = {
1278 .gfp_mask = gfp_mask,
1281 struct mem_cgroup *iter;
1282 unsigned long chosen_points = 0;
1283 unsigned long totalpages;
1284 unsigned int points = 0;
1285 struct task_struct *chosen = NULL;
1287 mutex_lock(&oom_lock);
1290 * If current has a pending SIGKILL or is exiting, then automatically
1291 * select it. The goal is to allow it to allocate so that it may
1292 * quickly exit and free its memory.
1294 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1295 mark_oom_victim(current);
1299 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1300 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1301 for_each_mem_cgroup_tree(iter, memcg) {
1302 struct css_task_iter it;
1303 struct task_struct *task;
1305 css_task_iter_start(&iter->css, &it);
1306 while ((task = css_task_iter_next(&it))) {
1307 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1308 case OOM_SCAN_SELECT:
1310 put_task_struct(chosen);
1312 chosen_points = ULONG_MAX;
1313 get_task_struct(chosen);
1315 case OOM_SCAN_CONTINUE:
1317 case OOM_SCAN_ABORT:
1318 css_task_iter_end(&it);
1319 mem_cgroup_iter_break(memcg, iter);
1321 put_task_struct(chosen);
1326 points = oom_badness(task, memcg, NULL, totalpages);
1327 if (!points || points < chosen_points)
1329 /* Prefer thread group leaders for display purposes */
1330 if (points == chosen_points &&
1331 thread_group_leader(chosen))
1335 put_task_struct(chosen);
1337 chosen_points = points;
1338 get_task_struct(chosen);
1340 css_task_iter_end(&it);
1344 points = chosen_points * 1000 / totalpages;
1345 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1346 "Memory cgroup out of memory");
1349 mutex_unlock(&oom_lock);
1352 #if MAX_NUMNODES > 1
1355 * test_mem_cgroup_node_reclaimable
1356 * @memcg: the target memcg
1357 * @nid: the node ID to be checked.
1358 * @noswap : specify true here if the user wants flle only information.
1360 * This function returns whether the specified memcg contains any
1361 * reclaimable pages on a node. Returns true if there are any reclaimable
1362 * pages in the node.
1364 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1365 int nid, bool noswap)
1367 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1369 if (noswap || !total_swap_pages)
1371 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1378 * Always updating the nodemask is not very good - even if we have an empty
1379 * list or the wrong list here, we can start from some node and traverse all
1380 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1383 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1387 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1388 * pagein/pageout changes since the last update.
1390 if (!atomic_read(&memcg->numainfo_events))
1392 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1395 /* make a nodemask where this memcg uses memory from */
1396 memcg->scan_nodes = node_states[N_MEMORY];
1398 for_each_node_mask(nid, node_states[N_MEMORY]) {
1400 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1401 node_clear(nid, memcg->scan_nodes);
1404 atomic_set(&memcg->numainfo_events, 0);
1405 atomic_set(&memcg->numainfo_updating, 0);
1409 * Selecting a node where we start reclaim from. Because what we need is just
1410 * reducing usage counter, start from anywhere is O,K. Considering
1411 * memory reclaim from current node, there are pros. and cons.
1413 * Freeing memory from current node means freeing memory from a node which
1414 * we'll use or we've used. So, it may make LRU bad. And if several threads
1415 * hit limits, it will see a contention on a node. But freeing from remote
1416 * node means more costs for memory reclaim because of memory latency.
1418 * Now, we use round-robin. Better algorithm is welcomed.
1420 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1424 mem_cgroup_may_update_nodemask(memcg);
1425 node = memcg->last_scanned_node;
1427 node = next_node(node, memcg->scan_nodes);
1428 if (node == MAX_NUMNODES)
1429 node = first_node(memcg->scan_nodes);
1431 * We call this when we hit limit, not when pages are added to LRU.
1432 * No LRU may hold pages because all pages are UNEVICTABLE or
1433 * memcg is too small and all pages are not on LRU. In that case,
1434 * we use curret node.
1436 if (unlikely(node == MAX_NUMNODES))
1437 node = numa_node_id();
1439 memcg->last_scanned_node = node;
1443 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1449 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1452 unsigned long *total_scanned)
1454 struct mem_cgroup *victim = NULL;
1457 unsigned long excess;
1458 unsigned long nr_scanned;
1459 struct mem_cgroup_reclaim_cookie reclaim = {
1464 excess = soft_limit_excess(root_memcg);
1467 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1472 * If we have not been able to reclaim
1473 * anything, it might because there are
1474 * no reclaimable pages under this hierarchy
1479 * We want to do more targeted reclaim.
1480 * excess >> 2 is not to excessive so as to
1481 * reclaim too much, nor too less that we keep
1482 * coming back to reclaim from this cgroup
1484 if (total >= (excess >> 2) ||
1485 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1490 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1492 *total_scanned += nr_scanned;
1493 if (!soft_limit_excess(root_memcg))
1496 mem_cgroup_iter_break(root_memcg, victim);
1500 #ifdef CONFIG_LOCKDEP
1501 static struct lockdep_map memcg_oom_lock_dep_map = {
1502 .name = "memcg_oom_lock",
1506 static DEFINE_SPINLOCK(memcg_oom_lock);
1509 * Check OOM-Killer is already running under our hierarchy.
1510 * If someone is running, return false.
1512 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1514 struct mem_cgroup *iter, *failed = NULL;
1516 spin_lock(&memcg_oom_lock);
1518 for_each_mem_cgroup_tree(iter, memcg) {
1519 if (iter->oom_lock) {
1521 * this subtree of our hierarchy is already locked
1522 * so we cannot give a lock.
1525 mem_cgroup_iter_break(memcg, iter);
1528 iter->oom_lock = true;
1533 * OK, we failed to lock the whole subtree so we have
1534 * to clean up what we set up to the failing subtree
1536 for_each_mem_cgroup_tree(iter, memcg) {
1537 if (iter == failed) {
1538 mem_cgroup_iter_break(memcg, iter);
1541 iter->oom_lock = false;
1544 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1546 spin_unlock(&memcg_oom_lock);
1551 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1553 struct mem_cgroup *iter;
1555 spin_lock(&memcg_oom_lock);
1556 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1557 for_each_mem_cgroup_tree(iter, memcg)
1558 iter->oom_lock = false;
1559 spin_unlock(&memcg_oom_lock);
1562 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1564 struct mem_cgroup *iter;
1566 spin_lock(&memcg_oom_lock);
1567 for_each_mem_cgroup_tree(iter, memcg)
1569 spin_unlock(&memcg_oom_lock);
1572 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1574 struct mem_cgroup *iter;
1577 * When a new child is created while the hierarchy is under oom,
1578 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1580 spin_lock(&memcg_oom_lock);
1581 for_each_mem_cgroup_tree(iter, memcg)
1582 if (iter->under_oom > 0)
1584 spin_unlock(&memcg_oom_lock);
1587 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1589 struct oom_wait_info {
1590 struct mem_cgroup *memcg;
1594 static int memcg_oom_wake_function(wait_queue_t *wait,
1595 unsigned mode, int sync, void *arg)
1597 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1598 struct mem_cgroup *oom_wait_memcg;
1599 struct oom_wait_info *oom_wait_info;
1601 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1602 oom_wait_memcg = oom_wait_info->memcg;
1604 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1605 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1607 return autoremove_wake_function(wait, mode, sync, arg);
1610 static void memcg_oom_recover(struct mem_cgroup *memcg)
1613 * For the following lockless ->under_oom test, the only required
1614 * guarantee is that it must see the state asserted by an OOM when
1615 * this function is called as a result of userland actions
1616 * triggered by the notification of the OOM. This is trivially
1617 * achieved by invoking mem_cgroup_mark_under_oom() before
1618 * triggering notification.
1620 if (memcg && memcg->under_oom)
1621 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1624 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1626 if (!current->memcg_may_oom)
1629 * We are in the middle of the charge context here, so we
1630 * don't want to block when potentially sitting on a callstack
1631 * that holds all kinds of filesystem and mm locks.
1633 * Also, the caller may handle a failed allocation gracefully
1634 * (like optional page cache readahead) and so an OOM killer
1635 * invocation might not even be necessary.
1637 * That's why we don't do anything here except remember the
1638 * OOM context and then deal with it at the end of the page
1639 * fault when the stack is unwound, the locks are released,
1640 * and when we know whether the fault was overall successful.
1642 css_get(&memcg->css);
1643 current->memcg_in_oom = memcg;
1644 current->memcg_oom_gfp_mask = mask;
1645 current->memcg_oom_order = order;
1649 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1650 * @handle: actually kill/wait or just clean up the OOM state
1652 * This has to be called at the end of a page fault if the memcg OOM
1653 * handler was enabled.
1655 * Memcg supports userspace OOM handling where failed allocations must
1656 * sleep on a waitqueue until the userspace task resolves the
1657 * situation. Sleeping directly in the charge context with all kinds
1658 * of locks held is not a good idea, instead we remember an OOM state
1659 * in the task and mem_cgroup_oom_synchronize() has to be called at
1660 * the end of the page fault to complete the OOM handling.
1662 * Returns %true if an ongoing memcg OOM situation was detected and
1663 * completed, %false otherwise.
1665 bool mem_cgroup_oom_synchronize(bool handle)
1667 struct mem_cgroup *memcg = current->memcg_in_oom;
1668 struct oom_wait_info owait;
1671 /* OOM is global, do not handle */
1675 if (!handle || oom_killer_disabled)
1678 owait.memcg = memcg;
1679 owait.wait.flags = 0;
1680 owait.wait.func = memcg_oom_wake_function;
1681 owait.wait.private = current;
1682 INIT_LIST_HEAD(&owait.wait.task_list);
1684 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1685 mem_cgroup_mark_under_oom(memcg);
1687 locked = mem_cgroup_oom_trylock(memcg);
1690 mem_cgroup_oom_notify(memcg);
1692 if (locked && !memcg->oom_kill_disable) {
1693 mem_cgroup_unmark_under_oom(memcg);
1694 finish_wait(&memcg_oom_waitq, &owait.wait);
1695 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1696 current->memcg_oom_order);
1699 mem_cgroup_unmark_under_oom(memcg);
1700 finish_wait(&memcg_oom_waitq, &owait.wait);
1704 mem_cgroup_oom_unlock(memcg);
1706 * There is no guarantee that an OOM-lock contender
1707 * sees the wakeups triggered by the OOM kill
1708 * uncharges. Wake any sleepers explicitely.
1710 memcg_oom_recover(memcg);
1713 current->memcg_in_oom = NULL;
1714 css_put(&memcg->css);
1719 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1720 * @page: page that is going to change accounted state
1722 * This function must mark the beginning of an accounted page state
1723 * change to prevent double accounting when the page is concurrently
1724 * being moved to another memcg:
1726 * memcg = mem_cgroup_begin_page_stat(page);
1727 * if (TestClearPageState(page))
1728 * mem_cgroup_update_page_stat(memcg, state, -1);
1729 * mem_cgroup_end_page_stat(memcg);
1731 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1733 struct mem_cgroup *memcg;
1734 unsigned long flags;
1737 * The RCU lock is held throughout the transaction. The fast
1738 * path can get away without acquiring the memcg->move_lock
1739 * because page moving starts with an RCU grace period.
1741 * The RCU lock also protects the memcg from being freed when
1742 * the page state that is going to change is the only thing
1743 * preventing the page from being uncharged.
1744 * E.g. end-writeback clearing PageWriteback(), which allows
1745 * migration to go ahead and uncharge the page before the
1746 * account transaction might be complete.
1750 if (mem_cgroup_disabled())
1753 memcg = page->mem_cgroup;
1754 if (unlikely(!memcg))
1757 if (atomic_read(&memcg->moving_account) <= 0)
1760 spin_lock_irqsave(&memcg->move_lock, flags);
1761 if (memcg != page->mem_cgroup) {
1762 spin_unlock_irqrestore(&memcg->move_lock, flags);
1767 * When charge migration first begins, we can have locked and
1768 * unlocked page stat updates happening concurrently. Track
1769 * the task who has the lock for mem_cgroup_end_page_stat().
1771 memcg->move_lock_task = current;
1772 memcg->move_lock_flags = flags;
1776 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
1779 * mem_cgroup_end_page_stat - finish a page state statistics transaction
1780 * @memcg: the memcg that was accounted against
1782 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
1784 if (memcg && memcg->move_lock_task == current) {
1785 unsigned long flags = memcg->move_lock_flags;
1787 memcg->move_lock_task = NULL;
1788 memcg->move_lock_flags = 0;
1790 spin_unlock_irqrestore(&memcg->move_lock, flags);
1795 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
1798 * size of first charge trial. "32" comes from vmscan.c's magic value.
1799 * TODO: maybe necessary to use big numbers in big irons.
1801 #define CHARGE_BATCH 32U
1802 struct memcg_stock_pcp {
1803 struct mem_cgroup *cached; /* this never be root cgroup */
1804 unsigned int nr_pages;
1805 struct work_struct work;
1806 unsigned long flags;
1807 #define FLUSHING_CACHED_CHARGE 0
1809 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1810 static DEFINE_MUTEX(percpu_charge_mutex);
1813 * consume_stock: Try to consume stocked charge on this cpu.
1814 * @memcg: memcg to consume from.
1815 * @nr_pages: how many pages to charge.
1817 * The charges will only happen if @memcg matches the current cpu's memcg
1818 * stock, and at least @nr_pages are available in that stock. Failure to
1819 * service an allocation will refill the stock.
1821 * returns true if successful, false otherwise.
1823 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1825 struct memcg_stock_pcp *stock;
1828 if (nr_pages > CHARGE_BATCH)
1831 stock = &get_cpu_var(memcg_stock);
1832 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1833 stock->nr_pages -= nr_pages;
1836 put_cpu_var(memcg_stock);
1841 * Returns stocks cached in percpu and reset cached information.
1843 static void drain_stock(struct memcg_stock_pcp *stock)
1845 struct mem_cgroup *old = stock->cached;
1847 if (stock->nr_pages) {
1848 page_counter_uncharge(&old->memory, stock->nr_pages);
1849 if (do_memsw_account())
1850 page_counter_uncharge(&old->memsw, stock->nr_pages);
1851 css_put_many(&old->css, stock->nr_pages);
1852 stock->nr_pages = 0;
1854 stock->cached = NULL;
1858 * This must be called under preempt disabled or must be called by
1859 * a thread which is pinned to local cpu.
1861 static void drain_local_stock(struct work_struct *dummy)
1863 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1865 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1869 * Cache charges(val) to local per_cpu area.
1870 * This will be consumed by consume_stock() function, later.
1872 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1874 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1876 if (stock->cached != memcg) { /* reset if necessary */
1878 stock->cached = memcg;
1880 stock->nr_pages += nr_pages;
1881 put_cpu_var(memcg_stock);
1885 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1886 * of the hierarchy under it.
1888 static void drain_all_stock(struct mem_cgroup *root_memcg)
1892 /* If someone's already draining, avoid adding running more workers. */
1893 if (!mutex_trylock(&percpu_charge_mutex))
1895 /* Notify other cpus that system-wide "drain" is running */
1898 for_each_online_cpu(cpu) {
1899 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1900 struct mem_cgroup *memcg;
1902 memcg = stock->cached;
1903 if (!memcg || !stock->nr_pages)
1905 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1907 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1909 drain_local_stock(&stock->work);
1911 schedule_work_on(cpu, &stock->work);
1916 mutex_unlock(&percpu_charge_mutex);
1919 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1920 unsigned long action,
1923 int cpu = (unsigned long)hcpu;
1924 struct memcg_stock_pcp *stock;
1926 if (action == CPU_ONLINE)
1929 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1932 stock = &per_cpu(memcg_stock, cpu);
1937 static void reclaim_high(struct mem_cgroup *memcg,
1938 unsigned int nr_pages,
1942 if (page_counter_read(&memcg->memory) <= memcg->high)
1944 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1945 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1946 } while ((memcg = parent_mem_cgroup(memcg)));
1949 static void high_work_func(struct work_struct *work)
1951 struct mem_cgroup *memcg;
1953 memcg = container_of(work, struct mem_cgroup, high_work);
1954 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1958 * Scheduled by try_charge() to be executed from the userland return path
1959 * and reclaims memory over the high limit.
1961 void mem_cgroup_handle_over_high(void)
1963 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1964 struct mem_cgroup *memcg;
1966 if (likely(!nr_pages))
1969 memcg = get_mem_cgroup_from_mm(current->mm);
1970 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1971 css_put(&memcg->css);
1972 current->memcg_nr_pages_over_high = 0;
1975 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1976 unsigned int nr_pages)
1978 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1979 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1980 struct mem_cgroup *mem_over_limit;
1981 struct page_counter *counter;
1982 unsigned long nr_reclaimed;
1983 bool may_swap = true;
1984 bool drained = false;
1986 if (mem_cgroup_is_root(memcg))
1989 if (consume_stock(memcg, nr_pages))
1992 if (!do_memsw_account() ||
1993 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1994 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1996 if (do_memsw_account())
1997 page_counter_uncharge(&memcg->memsw, batch);
1998 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2000 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2004 if (batch > nr_pages) {
2010 * Unlike in global OOM situations, memcg is not in a physical
2011 * memory shortage. Allow dying and OOM-killed tasks to
2012 * bypass the last charges so that they can exit quickly and
2013 * free their memory.
2015 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2016 fatal_signal_pending(current) ||
2017 current->flags & PF_EXITING))
2020 if (unlikely(task_in_memcg_oom(current)))
2023 if (!gfpflags_allow_blocking(gfp_mask))
2026 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2028 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2029 gfp_mask, may_swap);
2031 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2035 drain_all_stock(mem_over_limit);
2040 if (gfp_mask & __GFP_NORETRY)
2043 * Even though the limit is exceeded at this point, reclaim
2044 * may have been able to free some pages. Retry the charge
2045 * before killing the task.
2047 * Only for regular pages, though: huge pages are rather
2048 * unlikely to succeed so close to the limit, and we fall back
2049 * to regular pages anyway in case of failure.
2051 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2054 * At task move, charge accounts can be doubly counted. So, it's
2055 * better to wait until the end of task_move if something is going on.
2057 if (mem_cgroup_wait_acct_move(mem_over_limit))
2063 if (gfp_mask & __GFP_NOFAIL)
2066 if (fatal_signal_pending(current))
2069 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2071 mem_cgroup_oom(mem_over_limit, gfp_mask,
2072 get_order(nr_pages * PAGE_SIZE));
2074 if (!(gfp_mask & __GFP_NOFAIL))
2078 * The allocation either can't fail or will lead to more memory
2079 * being freed very soon. Allow memory usage go over the limit
2080 * temporarily by force charging it.
2082 page_counter_charge(&memcg->memory, nr_pages);
2083 if (do_memsw_account())
2084 page_counter_charge(&memcg->memsw, nr_pages);
2085 css_get_many(&memcg->css, nr_pages);
2090 css_get_many(&memcg->css, batch);
2091 if (batch > nr_pages)
2092 refill_stock(memcg, batch - nr_pages);
2095 * If the hierarchy is above the normal consumption range, schedule
2096 * reclaim on returning to userland. We can perform reclaim here
2097 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2098 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2099 * not recorded as it most likely matches current's and won't
2100 * change in the meantime. As high limit is checked again before
2101 * reclaim, the cost of mismatch is negligible.
2104 if (page_counter_read(&memcg->memory) > memcg->high) {
2105 /* Don't bother a random interrupted task */
2106 if (in_interrupt()) {
2107 schedule_work(&memcg->high_work);
2110 current->memcg_nr_pages_over_high += batch;
2111 set_notify_resume(current);
2114 } while ((memcg = parent_mem_cgroup(memcg)));
2119 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2121 if (mem_cgroup_is_root(memcg))
2124 page_counter_uncharge(&memcg->memory, nr_pages);
2125 if (do_memsw_account())
2126 page_counter_uncharge(&memcg->memsw, nr_pages);
2128 css_put_many(&memcg->css, nr_pages);
2131 static void lock_page_lru(struct page *page, int *isolated)
2133 struct zone *zone = page_zone(page);
2135 spin_lock_irq(&zone->lru_lock);
2136 if (PageLRU(page)) {
2137 struct lruvec *lruvec;
2139 lruvec = mem_cgroup_page_lruvec(page, zone);
2141 del_page_from_lru_list(page, lruvec, page_lru(page));
2147 static void unlock_page_lru(struct page *page, int isolated)
2149 struct zone *zone = page_zone(page);
2152 struct lruvec *lruvec;
2154 lruvec = mem_cgroup_page_lruvec(page, zone);
2155 VM_BUG_ON_PAGE(PageLRU(page), page);
2157 add_page_to_lru_list(page, lruvec, page_lru(page));
2159 spin_unlock_irq(&zone->lru_lock);
2162 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2167 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2170 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2171 * may already be on some other mem_cgroup's LRU. Take care of it.
2174 lock_page_lru(page, &isolated);
2177 * Nobody should be changing or seriously looking at
2178 * page->mem_cgroup at this point:
2180 * - the page is uncharged
2182 * - the page is off-LRU
2184 * - an anonymous fault has exclusive page access, except for
2185 * a locked page table
2187 * - a page cache insertion, a swapin fault, or a migration
2188 * have the page locked
2190 page->mem_cgroup = memcg;
2193 unlock_page_lru(page, isolated);
2197 static int memcg_alloc_cache_id(void)
2202 id = ida_simple_get(&memcg_cache_ida,
2203 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2207 if (id < memcg_nr_cache_ids)
2211 * There's no space for the new id in memcg_caches arrays,
2212 * so we have to grow them.
2214 down_write(&memcg_cache_ids_sem);
2216 size = 2 * (id + 1);
2217 if (size < MEMCG_CACHES_MIN_SIZE)
2218 size = MEMCG_CACHES_MIN_SIZE;
2219 else if (size > MEMCG_CACHES_MAX_SIZE)
2220 size = MEMCG_CACHES_MAX_SIZE;
2222 err = memcg_update_all_caches(size);
2224 err = memcg_update_all_list_lrus(size);
2226 memcg_nr_cache_ids = size;
2228 up_write(&memcg_cache_ids_sem);
2231 ida_simple_remove(&memcg_cache_ida, id);
2237 static void memcg_free_cache_id(int id)
2239 ida_simple_remove(&memcg_cache_ida, id);
2242 struct memcg_kmem_cache_create_work {
2243 struct mem_cgroup *memcg;
2244 struct kmem_cache *cachep;
2245 struct work_struct work;
2248 static void memcg_kmem_cache_create_func(struct work_struct *w)
2250 struct memcg_kmem_cache_create_work *cw =
2251 container_of(w, struct memcg_kmem_cache_create_work, work);
2252 struct mem_cgroup *memcg = cw->memcg;
2253 struct kmem_cache *cachep = cw->cachep;
2255 memcg_create_kmem_cache(memcg, cachep);
2257 css_put(&memcg->css);
2262 * Enqueue the creation of a per-memcg kmem_cache.
2264 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2265 struct kmem_cache *cachep)
2267 struct memcg_kmem_cache_create_work *cw;
2269 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2273 css_get(&memcg->css);
2276 cw->cachep = cachep;
2277 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2279 schedule_work(&cw->work);
2282 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2283 struct kmem_cache *cachep)
2286 * We need to stop accounting when we kmalloc, because if the
2287 * corresponding kmalloc cache is not yet created, the first allocation
2288 * in __memcg_schedule_kmem_cache_create will recurse.
2290 * However, it is better to enclose the whole function. Depending on
2291 * the debugging options enabled, INIT_WORK(), for instance, can
2292 * trigger an allocation. This too, will make us recurse. Because at
2293 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2294 * the safest choice is to do it like this, wrapping the whole function.
2296 current->memcg_kmem_skip_account = 1;
2297 __memcg_schedule_kmem_cache_create(memcg, cachep);
2298 current->memcg_kmem_skip_account = 0;
2302 * Return the kmem_cache we're supposed to use for a slab allocation.
2303 * We try to use the current memcg's version of the cache.
2305 * If the cache does not exist yet, if we are the first user of it,
2306 * we either create it immediately, if possible, or create it asynchronously
2308 * In the latter case, we will let the current allocation go through with
2309 * the original cache.
2311 * Can't be called in interrupt context or from kernel threads.
2312 * This function needs to be called with rcu_read_lock() held.
2314 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2316 struct mem_cgroup *memcg;
2317 struct kmem_cache *memcg_cachep;
2320 VM_BUG_ON(!is_root_cache(cachep));
2322 if (cachep->flags & SLAB_ACCOUNT)
2323 gfp |= __GFP_ACCOUNT;
2325 if (!(gfp & __GFP_ACCOUNT))
2328 if (current->memcg_kmem_skip_account)
2331 memcg = get_mem_cgroup_from_mm(current->mm);
2332 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2336 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2337 if (likely(memcg_cachep))
2338 return memcg_cachep;
2341 * If we are in a safe context (can wait, and not in interrupt
2342 * context), we could be be predictable and return right away.
2343 * This would guarantee that the allocation being performed
2344 * already belongs in the new cache.
2346 * However, there are some clashes that can arrive from locking.
2347 * For instance, because we acquire the slab_mutex while doing
2348 * memcg_create_kmem_cache, this means no further allocation
2349 * could happen with the slab_mutex held. So it's better to
2352 memcg_schedule_kmem_cache_create(memcg, cachep);
2354 css_put(&memcg->css);
2358 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2360 if (!is_root_cache(cachep))
2361 css_put(&cachep->memcg_params.memcg->css);
2364 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2365 struct mem_cgroup *memcg)
2367 unsigned int nr_pages = 1 << order;
2368 struct page_counter *counter;
2371 if (!memcg_kmem_online(memcg))
2374 ret = try_charge(memcg, gfp, nr_pages);
2378 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2379 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2380 cancel_charge(memcg, nr_pages);
2384 page->mem_cgroup = memcg;
2389 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2391 struct mem_cgroup *memcg;
2394 memcg = get_mem_cgroup_from_mm(current->mm);
2395 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2396 css_put(&memcg->css);
2400 void __memcg_kmem_uncharge(struct page *page, int order)
2402 struct mem_cgroup *memcg = page->mem_cgroup;
2403 unsigned int nr_pages = 1 << order;
2408 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2410 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2411 page_counter_uncharge(&memcg->kmem, nr_pages);
2413 page_counter_uncharge(&memcg->memory, nr_pages);
2414 if (do_memsw_account())
2415 page_counter_uncharge(&memcg->memsw, nr_pages);
2417 page->mem_cgroup = NULL;
2418 css_put_many(&memcg->css, nr_pages);
2420 #endif /* !CONFIG_SLOB */
2422 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2425 * Because tail pages are not marked as "used", set it. We're under
2426 * zone->lru_lock and migration entries setup in all page mappings.
2428 void mem_cgroup_split_huge_fixup(struct page *head)
2432 if (mem_cgroup_disabled())
2435 for (i = 1; i < HPAGE_PMD_NR; i++)
2436 head[i].mem_cgroup = head->mem_cgroup;
2438 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2441 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2443 #ifdef CONFIG_MEMCG_SWAP
2444 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2447 int val = (charge) ? 1 : -1;
2448 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2452 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2453 * @entry: swap entry to be moved
2454 * @from: mem_cgroup which the entry is moved from
2455 * @to: mem_cgroup which the entry is moved to
2457 * It succeeds only when the swap_cgroup's record for this entry is the same
2458 * as the mem_cgroup's id of @from.
2460 * Returns 0 on success, -EINVAL on failure.
2462 * The caller must have charged to @to, IOW, called page_counter_charge() about
2463 * both res and memsw, and called css_get().
2465 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2466 struct mem_cgroup *from, struct mem_cgroup *to)
2468 unsigned short old_id, new_id;
2470 old_id = mem_cgroup_id(from);
2471 new_id = mem_cgroup_id(to);
2473 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2474 mem_cgroup_swap_statistics(from, false);
2475 mem_cgroup_swap_statistics(to, true);
2481 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2482 struct mem_cgroup *from, struct mem_cgroup *to)
2488 static DEFINE_MUTEX(memcg_limit_mutex);
2490 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2491 unsigned long limit)
2493 unsigned long curusage;
2494 unsigned long oldusage;
2495 bool enlarge = false;
2500 * For keeping hierarchical_reclaim simple, how long we should retry
2501 * is depends on callers. We set our retry-count to be function
2502 * of # of children which we should visit in this loop.
2504 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2505 mem_cgroup_count_children(memcg);
2507 oldusage = page_counter_read(&memcg->memory);
2510 if (signal_pending(current)) {
2515 mutex_lock(&memcg_limit_mutex);
2516 if (limit > memcg->memsw.limit) {
2517 mutex_unlock(&memcg_limit_mutex);
2521 if (limit > memcg->memory.limit)
2523 ret = page_counter_limit(&memcg->memory, limit);
2524 mutex_unlock(&memcg_limit_mutex);
2529 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2531 curusage = page_counter_read(&memcg->memory);
2532 /* Usage is reduced ? */
2533 if (curusage >= oldusage)
2536 oldusage = curusage;
2537 } while (retry_count);
2539 if (!ret && enlarge)
2540 memcg_oom_recover(memcg);
2545 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2546 unsigned long limit)
2548 unsigned long curusage;
2549 unsigned long oldusage;
2550 bool enlarge = false;
2554 /* see mem_cgroup_resize_res_limit */
2555 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2556 mem_cgroup_count_children(memcg);
2558 oldusage = page_counter_read(&memcg->memsw);
2561 if (signal_pending(current)) {
2566 mutex_lock(&memcg_limit_mutex);
2567 if (limit < memcg->memory.limit) {
2568 mutex_unlock(&memcg_limit_mutex);
2572 if (limit > memcg->memsw.limit)
2574 ret = page_counter_limit(&memcg->memsw, limit);
2575 mutex_unlock(&memcg_limit_mutex);
2580 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2582 curusage = page_counter_read(&memcg->memsw);
2583 /* Usage is reduced ? */
2584 if (curusage >= oldusage)
2587 oldusage = curusage;
2588 } while (retry_count);
2590 if (!ret && enlarge)
2591 memcg_oom_recover(memcg);
2596 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2598 unsigned long *total_scanned)
2600 unsigned long nr_reclaimed = 0;
2601 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2602 unsigned long reclaimed;
2604 struct mem_cgroup_tree_per_zone *mctz;
2605 unsigned long excess;
2606 unsigned long nr_scanned;
2611 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2613 * This loop can run a while, specially if mem_cgroup's continuously
2614 * keep exceeding their soft limit and putting the system under
2621 mz = mem_cgroup_largest_soft_limit_node(mctz);
2626 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2627 gfp_mask, &nr_scanned);
2628 nr_reclaimed += reclaimed;
2629 *total_scanned += nr_scanned;
2630 spin_lock_irq(&mctz->lock);
2631 __mem_cgroup_remove_exceeded(mz, mctz);
2634 * If we failed to reclaim anything from this memory cgroup
2635 * it is time to move on to the next cgroup
2639 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2641 excess = soft_limit_excess(mz->memcg);
2643 * One school of thought says that we should not add
2644 * back the node to the tree if reclaim returns 0.
2645 * But our reclaim could return 0, simply because due
2646 * to priority we are exposing a smaller subset of
2647 * memory to reclaim from. Consider this as a longer
2650 /* If excess == 0, no tree ops */
2651 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2652 spin_unlock_irq(&mctz->lock);
2653 css_put(&mz->memcg->css);
2656 * Could not reclaim anything and there are no more
2657 * mem cgroups to try or we seem to be looping without
2658 * reclaiming anything.
2660 if (!nr_reclaimed &&
2662 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2664 } while (!nr_reclaimed);
2666 css_put(&next_mz->memcg->css);
2667 return nr_reclaimed;
2671 * Test whether @memcg has children, dead or alive. Note that this
2672 * function doesn't care whether @memcg has use_hierarchy enabled and
2673 * returns %true if there are child csses according to the cgroup
2674 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2676 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2681 ret = css_next_child(NULL, &memcg->css);
2687 * Reclaims as many pages from the given memcg as possible and moves
2688 * the rest to the parent.
2690 * Caller is responsible for holding css reference for memcg.
2692 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2694 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2696 /* we call try-to-free pages for make this cgroup empty */
2697 lru_add_drain_all();
2698 /* try to free all pages in this cgroup */
2699 while (nr_retries && page_counter_read(&memcg->memory)) {
2702 if (signal_pending(current))
2705 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2709 /* maybe some writeback is necessary */
2710 congestion_wait(BLK_RW_ASYNC, HZ/10);
2718 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2719 char *buf, size_t nbytes,
2722 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2724 if (mem_cgroup_is_root(memcg))
2726 return mem_cgroup_force_empty(memcg) ?: nbytes;
2729 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2732 return mem_cgroup_from_css(css)->use_hierarchy;
2735 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2736 struct cftype *cft, u64 val)
2739 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2740 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2742 if (memcg->use_hierarchy == val)
2746 * If parent's use_hierarchy is set, we can't make any modifications
2747 * in the child subtrees. If it is unset, then the change can
2748 * occur, provided the current cgroup has no children.
2750 * For the root cgroup, parent_mem is NULL, we allow value to be
2751 * set if there are no children.
2753 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2754 (val == 1 || val == 0)) {
2755 if (!memcg_has_children(memcg))
2756 memcg->use_hierarchy = val;
2765 static unsigned long tree_stat(struct mem_cgroup *memcg,
2766 enum mem_cgroup_stat_index idx)
2768 struct mem_cgroup *iter;
2769 unsigned long val = 0;
2771 for_each_mem_cgroup_tree(iter, memcg)
2772 val += mem_cgroup_read_stat(iter, idx);
2777 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2781 if (mem_cgroup_is_root(memcg)) {
2782 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
2783 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
2785 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
2788 val = page_counter_read(&memcg->memory);
2790 val = page_counter_read(&memcg->memsw);
2803 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2806 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2807 struct page_counter *counter;
2809 switch (MEMFILE_TYPE(cft->private)) {
2811 counter = &memcg->memory;
2814 counter = &memcg->memsw;
2817 counter = &memcg->kmem;
2820 counter = &memcg->tcpmem;
2826 switch (MEMFILE_ATTR(cft->private)) {
2828 if (counter == &memcg->memory)
2829 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2830 if (counter == &memcg->memsw)
2831 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2832 return (u64)page_counter_read(counter) * PAGE_SIZE;
2834 return (u64)counter->limit * PAGE_SIZE;
2836 return (u64)counter->watermark * PAGE_SIZE;
2838 return counter->failcnt;
2839 case RES_SOFT_LIMIT:
2840 return (u64)memcg->soft_limit * PAGE_SIZE;
2847 static int memcg_online_kmem(struct mem_cgroup *memcg)
2851 BUG_ON(memcg->kmemcg_id >= 0);
2852 BUG_ON(memcg->kmem_state);
2854 memcg_id = memcg_alloc_cache_id();
2858 static_branch_inc(&memcg_kmem_enabled_key);
2860 * A memory cgroup is considered kmem-online as soon as it gets
2861 * kmemcg_id. Setting the id after enabling static branching will
2862 * guarantee no one starts accounting before all call sites are
2865 memcg->kmemcg_id = memcg_id;
2866 memcg->kmem_state = KMEM_ONLINE;
2871 static int memcg_propagate_kmem(struct mem_cgroup *parent,
2872 struct mem_cgroup *memcg)
2876 mutex_lock(&memcg_limit_mutex);
2878 * If the parent cgroup is not kmem-online now, it cannot be
2879 * onlined after this point, because it has at least one child
2882 if (memcg_kmem_online(parent) ||
2883 (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nokmem))
2884 ret = memcg_online_kmem(memcg);
2885 mutex_unlock(&memcg_limit_mutex);
2889 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2891 struct cgroup_subsys_state *css;
2892 struct mem_cgroup *parent, *child;
2895 if (memcg->kmem_state != KMEM_ONLINE)
2898 * Clear the online state before clearing memcg_caches array
2899 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2900 * guarantees that no cache will be created for this cgroup
2901 * after we are done (see memcg_create_kmem_cache()).
2903 memcg->kmem_state = KMEM_ALLOCATED;
2905 memcg_deactivate_kmem_caches(memcg);
2907 kmemcg_id = memcg->kmemcg_id;
2908 BUG_ON(kmemcg_id < 0);
2910 parent = parent_mem_cgroup(memcg);
2912 parent = root_mem_cgroup;
2915 * Change kmemcg_id of this cgroup and all its descendants to the
2916 * parent's id, and then move all entries from this cgroup's list_lrus
2917 * to ones of the parent. After we have finished, all list_lrus
2918 * corresponding to this cgroup are guaranteed to remain empty. The
2919 * ordering is imposed by list_lru_node->lock taken by
2920 * memcg_drain_all_list_lrus().
2922 css_for_each_descendant_pre(css, &memcg->css) {
2923 child = mem_cgroup_from_css(css);
2924 BUG_ON(child->kmemcg_id != kmemcg_id);
2925 child->kmemcg_id = parent->kmemcg_id;
2926 if (!memcg->use_hierarchy)
2929 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2931 memcg_free_cache_id(kmemcg_id);
2934 static void memcg_free_kmem(struct mem_cgroup *memcg)
2936 /* css_alloc() failed, offlining didn't happen */
2937 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2938 memcg_offline_kmem(memcg);
2940 if (memcg->kmem_state == KMEM_ALLOCATED) {
2941 memcg_destroy_kmem_caches(memcg);
2942 static_branch_dec(&memcg_kmem_enabled_key);
2943 WARN_ON(page_counter_read(&memcg->kmem));
2947 static int memcg_propagate_kmem(struct mem_cgroup *parent, struct mem_cgroup *memcg)
2951 static int memcg_online_kmem(struct mem_cgroup *memcg)
2955 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2958 static void memcg_free_kmem(struct mem_cgroup *memcg)
2961 #endif /* !CONFIG_SLOB */
2963 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2964 unsigned long limit)
2968 mutex_lock(&memcg_limit_mutex);
2969 /* Top-level cgroup doesn't propagate from root */
2970 if (!memcg_kmem_online(memcg)) {
2971 if (cgroup_is_populated(memcg->css.cgroup) ||
2972 (memcg->use_hierarchy && memcg_has_children(memcg)))
2976 ret = memcg_online_kmem(memcg);
2980 ret = page_counter_limit(&memcg->kmem, limit);
2982 mutex_unlock(&memcg_limit_mutex);
2986 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2990 mutex_lock(&memcg_limit_mutex);
2992 ret = page_counter_limit(&memcg->tcpmem, limit);
2996 if (!memcg->tcpmem_active) {
2998 * The active flag needs to be written after the static_key
2999 * update. This is what guarantees that the socket activation
3000 * function is the last one to run. See sock_update_memcg() for
3001 * details, and note that we don't mark any socket as belonging
3002 * to this memcg until that flag is up.
3004 * We need to do this, because static_keys will span multiple
3005 * sites, but we can't control their order. If we mark a socket
3006 * as accounted, but the accounting functions are not patched in
3007 * yet, we'll lose accounting.
3009 * We never race with the readers in sock_update_memcg(),
3010 * because when this value change, the code to process it is not
3013 static_branch_inc(&memcg_sockets_enabled_key);
3014 memcg->tcpmem_active = true;
3017 mutex_unlock(&memcg_limit_mutex);
3022 * The user of this function is...
3025 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3026 char *buf, size_t nbytes, loff_t off)
3028 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3029 unsigned long nr_pages;
3032 buf = strstrip(buf);
3033 ret = page_counter_memparse(buf, "-1", &nr_pages);
3037 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3039 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3043 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3045 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3048 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3051 ret = memcg_update_kmem_limit(memcg, nr_pages);
3054 ret = memcg_update_tcp_limit(memcg, nr_pages);
3058 case RES_SOFT_LIMIT:
3059 memcg->soft_limit = nr_pages;
3063 return ret ?: nbytes;
3066 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3067 size_t nbytes, loff_t off)
3069 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3070 struct page_counter *counter;
3072 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3074 counter = &memcg->memory;
3077 counter = &memcg->memsw;
3080 counter = &memcg->kmem;
3083 counter = &memcg->tcpmem;
3089 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3091 page_counter_reset_watermark(counter);
3094 counter->failcnt = 0;
3103 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3106 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3110 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3111 struct cftype *cft, u64 val)
3113 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3115 if (val & ~MOVE_MASK)
3119 * No kind of locking is needed in here, because ->can_attach() will
3120 * check this value once in the beginning of the process, and then carry
3121 * on with stale data. This means that changes to this value will only
3122 * affect task migrations starting after the change.
3124 memcg->move_charge_at_immigrate = val;
3128 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3129 struct cftype *cft, u64 val)
3136 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3140 unsigned int lru_mask;
3143 static const struct numa_stat stats[] = {
3144 { "total", LRU_ALL },
3145 { "file", LRU_ALL_FILE },
3146 { "anon", LRU_ALL_ANON },
3147 { "unevictable", BIT(LRU_UNEVICTABLE) },
3149 const struct numa_stat *stat;
3152 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3154 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3155 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3156 seq_printf(m, "%s=%lu", stat->name, nr);
3157 for_each_node_state(nid, N_MEMORY) {
3158 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3160 seq_printf(m, " N%d=%lu", nid, nr);
3165 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3166 struct mem_cgroup *iter;
3169 for_each_mem_cgroup_tree(iter, memcg)
3170 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3171 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3172 for_each_node_state(nid, N_MEMORY) {
3174 for_each_mem_cgroup_tree(iter, memcg)
3175 nr += mem_cgroup_node_nr_lru_pages(
3176 iter, nid, stat->lru_mask);
3177 seq_printf(m, " N%d=%lu", nid, nr);
3184 #endif /* CONFIG_NUMA */
3186 static int memcg_stat_show(struct seq_file *m, void *v)
3188 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3189 unsigned long memory, memsw;
3190 struct mem_cgroup *mi;
3193 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3194 MEM_CGROUP_STAT_NSTATS);
3195 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3196 MEM_CGROUP_EVENTS_NSTATS);
3197 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3199 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3200 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3202 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3203 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3206 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3207 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3208 mem_cgroup_read_events(memcg, i));
3210 for (i = 0; i < NR_LRU_LISTS; i++)
3211 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3212 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3214 /* Hierarchical information */
3215 memory = memsw = PAGE_COUNTER_MAX;
3216 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3217 memory = min(memory, mi->memory.limit);
3218 memsw = min(memsw, mi->memsw.limit);
3220 seq_printf(m, "hierarchical_memory_limit %llu\n",
3221 (u64)memory * PAGE_SIZE);
3222 if (do_memsw_account())
3223 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3224 (u64)memsw * PAGE_SIZE);
3226 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3227 unsigned long long val = 0;
3229 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3231 for_each_mem_cgroup_tree(mi, memcg)
3232 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3233 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3236 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3237 unsigned long long val = 0;
3239 for_each_mem_cgroup_tree(mi, memcg)
3240 val += mem_cgroup_read_events(mi, i);
3241 seq_printf(m, "total_%s %llu\n",
3242 mem_cgroup_events_names[i], val);
3245 for (i = 0; i < NR_LRU_LISTS; i++) {
3246 unsigned long long val = 0;
3248 for_each_mem_cgroup_tree(mi, memcg)
3249 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3250 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3253 #ifdef CONFIG_DEBUG_VM
3256 struct mem_cgroup_per_zone *mz;
3257 struct zone_reclaim_stat *rstat;
3258 unsigned long recent_rotated[2] = {0, 0};
3259 unsigned long recent_scanned[2] = {0, 0};
3261 for_each_online_node(nid)
3262 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3263 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3264 rstat = &mz->lruvec.reclaim_stat;
3266 recent_rotated[0] += rstat->recent_rotated[0];
3267 recent_rotated[1] += rstat->recent_rotated[1];
3268 recent_scanned[0] += rstat->recent_scanned[0];
3269 recent_scanned[1] += rstat->recent_scanned[1];
3271 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3272 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3273 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3274 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3281 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3284 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3286 return mem_cgroup_swappiness(memcg);
3289 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3290 struct cftype *cft, u64 val)
3292 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3298 memcg->swappiness = val;
3300 vm_swappiness = val;
3305 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3307 struct mem_cgroup_threshold_ary *t;
3308 unsigned long usage;
3313 t = rcu_dereference(memcg->thresholds.primary);
3315 t = rcu_dereference(memcg->memsw_thresholds.primary);
3320 usage = mem_cgroup_usage(memcg, swap);
3323 * current_threshold points to threshold just below or equal to usage.
3324 * If it's not true, a threshold was crossed after last
3325 * call of __mem_cgroup_threshold().
3327 i = t->current_threshold;
3330 * Iterate backward over array of thresholds starting from
3331 * current_threshold and check if a threshold is crossed.
3332 * If none of thresholds below usage is crossed, we read
3333 * only one element of the array here.
3335 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3336 eventfd_signal(t->entries[i].eventfd, 1);
3338 /* i = current_threshold + 1 */
3342 * Iterate forward over array of thresholds starting from
3343 * current_threshold+1 and check if a threshold is crossed.
3344 * If none of thresholds above usage is crossed, we read
3345 * only one element of the array here.
3347 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3348 eventfd_signal(t->entries[i].eventfd, 1);
3350 /* Update current_threshold */
3351 t->current_threshold = i - 1;
3356 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3359 __mem_cgroup_threshold(memcg, false);
3360 if (do_memsw_account())
3361 __mem_cgroup_threshold(memcg, true);
3363 memcg = parent_mem_cgroup(memcg);
3367 static int compare_thresholds(const void *a, const void *b)
3369 const struct mem_cgroup_threshold *_a = a;
3370 const struct mem_cgroup_threshold *_b = b;
3372 if (_a->threshold > _b->threshold)
3375 if (_a->threshold < _b->threshold)
3381 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3383 struct mem_cgroup_eventfd_list *ev;
3385 spin_lock(&memcg_oom_lock);
3387 list_for_each_entry(ev, &memcg->oom_notify, list)
3388 eventfd_signal(ev->eventfd, 1);
3390 spin_unlock(&memcg_oom_lock);
3394 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3396 struct mem_cgroup *iter;
3398 for_each_mem_cgroup_tree(iter, memcg)
3399 mem_cgroup_oom_notify_cb(iter);
3402 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3403 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3405 struct mem_cgroup_thresholds *thresholds;
3406 struct mem_cgroup_threshold_ary *new;
3407 unsigned long threshold;
3408 unsigned long usage;
3411 ret = page_counter_memparse(args, "-1", &threshold);
3415 mutex_lock(&memcg->thresholds_lock);
3418 thresholds = &memcg->thresholds;
3419 usage = mem_cgroup_usage(memcg, false);
3420 } else if (type == _MEMSWAP) {
3421 thresholds = &memcg->memsw_thresholds;
3422 usage = mem_cgroup_usage(memcg, true);
3426 /* Check if a threshold crossed before adding a new one */
3427 if (thresholds->primary)
3428 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3430 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3432 /* Allocate memory for new array of thresholds */
3433 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3441 /* Copy thresholds (if any) to new array */
3442 if (thresholds->primary) {
3443 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3444 sizeof(struct mem_cgroup_threshold));
3447 /* Add new threshold */
3448 new->entries[size - 1].eventfd = eventfd;
3449 new->entries[size - 1].threshold = threshold;
3451 /* Sort thresholds. Registering of new threshold isn't time-critical */
3452 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3453 compare_thresholds, NULL);
3455 /* Find current threshold */
3456 new->current_threshold = -1;
3457 for (i = 0; i < size; i++) {
3458 if (new->entries[i].threshold <= usage) {
3460 * new->current_threshold will not be used until
3461 * rcu_assign_pointer(), so it's safe to increment
3464 ++new->current_threshold;
3469 /* Free old spare buffer and save old primary buffer as spare */
3470 kfree(thresholds->spare);
3471 thresholds->spare = thresholds->primary;
3473 rcu_assign_pointer(thresholds->primary, new);
3475 /* To be sure that nobody uses thresholds */
3479 mutex_unlock(&memcg->thresholds_lock);
3484 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3485 struct eventfd_ctx *eventfd, const char *args)
3487 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3490 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3491 struct eventfd_ctx *eventfd, const char *args)
3493 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3496 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3497 struct eventfd_ctx *eventfd, enum res_type type)
3499 struct mem_cgroup_thresholds *thresholds;
3500 struct mem_cgroup_threshold_ary *new;
3501 unsigned long usage;
3504 mutex_lock(&memcg->thresholds_lock);
3507 thresholds = &memcg->thresholds;
3508 usage = mem_cgroup_usage(memcg, false);
3509 } else if (type == _MEMSWAP) {
3510 thresholds = &memcg->memsw_thresholds;
3511 usage = mem_cgroup_usage(memcg, true);
3515 if (!thresholds->primary)
3518 /* Check if a threshold crossed before removing */
3519 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3521 /* Calculate new number of threshold */
3523 for (i = 0; i < thresholds->primary->size; i++) {
3524 if (thresholds->primary->entries[i].eventfd != eventfd)
3528 new = thresholds->spare;
3530 /* Set thresholds array to NULL if we don't have thresholds */
3539 /* Copy thresholds and find current threshold */
3540 new->current_threshold = -1;
3541 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3542 if (thresholds->primary->entries[i].eventfd == eventfd)
3545 new->entries[j] = thresholds->primary->entries[i];
3546 if (new->entries[j].threshold <= usage) {
3548 * new->current_threshold will not be used
3549 * until rcu_assign_pointer(), so it's safe to increment
3552 ++new->current_threshold;
3558 /* Swap primary and spare array */
3559 thresholds->spare = thresholds->primary;
3561 rcu_assign_pointer(thresholds->primary, new);
3563 /* To be sure that nobody uses thresholds */
3566 /* If all events are unregistered, free the spare array */
3568 kfree(thresholds->spare);
3569 thresholds->spare = NULL;
3572 mutex_unlock(&memcg->thresholds_lock);
3575 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3576 struct eventfd_ctx *eventfd)
3578 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3581 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3582 struct eventfd_ctx *eventfd)
3584 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3587 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3588 struct eventfd_ctx *eventfd, const char *args)
3590 struct mem_cgroup_eventfd_list *event;
3592 event = kmalloc(sizeof(*event), GFP_KERNEL);
3596 spin_lock(&memcg_oom_lock);
3598 event->eventfd = eventfd;
3599 list_add(&event->list, &memcg->oom_notify);
3601 /* already in OOM ? */
3602 if (memcg->under_oom)
3603 eventfd_signal(eventfd, 1);
3604 spin_unlock(&memcg_oom_lock);
3609 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3610 struct eventfd_ctx *eventfd)
3612 struct mem_cgroup_eventfd_list *ev, *tmp;
3614 spin_lock(&memcg_oom_lock);
3616 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3617 if (ev->eventfd == eventfd) {
3618 list_del(&ev->list);
3623 spin_unlock(&memcg_oom_lock);
3626 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3628 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3630 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3631 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3635 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3636 struct cftype *cft, u64 val)
3638 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3640 /* cannot set to root cgroup and only 0 and 1 are allowed */
3641 if (!css->parent || !((val == 0) || (val == 1)))
3644 memcg->oom_kill_disable = val;
3646 memcg_oom_recover(memcg);
3651 #ifdef CONFIG_CGROUP_WRITEBACK
3653 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3655 return &memcg->cgwb_list;
3658 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3660 return wb_domain_init(&memcg->cgwb_domain, gfp);
3663 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3665 wb_domain_exit(&memcg->cgwb_domain);
3668 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3670 wb_domain_size_changed(&memcg->cgwb_domain);
3673 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3675 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3677 if (!memcg->css.parent)
3680 return &memcg->cgwb_domain;
3684 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3685 * @wb: bdi_writeback in question
3686 * @pfilepages: out parameter for number of file pages
3687 * @pheadroom: out parameter for number of allocatable pages according to memcg
3688 * @pdirty: out parameter for number of dirty pages
3689 * @pwriteback: out parameter for number of pages under writeback
3691 * Determine the numbers of file, headroom, dirty, and writeback pages in
3692 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3693 * is a bit more involved.
3695 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3696 * headroom is calculated as the lowest headroom of itself and the
3697 * ancestors. Note that this doesn't consider the actual amount of
3698 * available memory in the system. The caller should further cap
3699 * *@pheadroom accordingly.
3701 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3702 unsigned long *pheadroom, unsigned long *pdirty,
3703 unsigned long *pwriteback)
3705 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3706 struct mem_cgroup *parent;
3708 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3710 /* this should eventually include NR_UNSTABLE_NFS */
3711 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3712 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3713 (1 << LRU_ACTIVE_FILE));
3714 *pheadroom = PAGE_COUNTER_MAX;
3716 while ((parent = parent_mem_cgroup(memcg))) {
3717 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3718 unsigned long used = page_counter_read(&memcg->memory);
3720 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3725 #else /* CONFIG_CGROUP_WRITEBACK */
3727 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3732 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3736 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3740 #endif /* CONFIG_CGROUP_WRITEBACK */
3743 * DO NOT USE IN NEW FILES.
3745 * "cgroup.event_control" implementation.
3747 * This is way over-engineered. It tries to support fully configurable
3748 * events for each user. Such level of flexibility is completely
3749 * unnecessary especially in the light of the planned unified hierarchy.
3751 * Please deprecate this and replace with something simpler if at all
3756 * Unregister event and free resources.
3758 * Gets called from workqueue.
3760 static void memcg_event_remove(struct work_struct *work)
3762 struct mem_cgroup_event *event =
3763 container_of(work, struct mem_cgroup_event, remove);
3764 struct mem_cgroup *memcg = event->memcg;
3766 remove_wait_queue(event->wqh, &event->wait);
3768 event->unregister_event(memcg, event->eventfd);
3770 /* Notify userspace the event is going away. */
3771 eventfd_signal(event->eventfd, 1);
3773 eventfd_ctx_put(event->eventfd);
3775 css_put(&memcg->css);
3779 * Gets called on POLLHUP on eventfd when user closes it.
3781 * Called with wqh->lock held and interrupts disabled.
3783 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3784 int sync, void *key)
3786 struct mem_cgroup_event *event =
3787 container_of(wait, struct mem_cgroup_event, wait);
3788 struct mem_cgroup *memcg = event->memcg;
3789 unsigned long flags = (unsigned long)key;
3791 if (flags & POLLHUP) {
3793 * If the event has been detached at cgroup removal, we
3794 * can simply return knowing the other side will cleanup
3797 * We can't race against event freeing since the other
3798 * side will require wqh->lock via remove_wait_queue(),
3801 spin_lock(&memcg->event_list_lock);
3802 if (!list_empty(&event->list)) {
3803 list_del_init(&event->list);
3805 * We are in atomic context, but cgroup_event_remove()
3806 * may sleep, so we have to call it in workqueue.
3808 schedule_work(&event->remove);
3810 spin_unlock(&memcg->event_list_lock);
3816 static void memcg_event_ptable_queue_proc(struct file *file,
3817 wait_queue_head_t *wqh, poll_table *pt)
3819 struct mem_cgroup_event *event =
3820 container_of(pt, struct mem_cgroup_event, pt);
3823 add_wait_queue(wqh, &event->wait);
3827 * DO NOT USE IN NEW FILES.
3829 * Parse input and register new cgroup event handler.
3831 * Input must be in format '<event_fd> <control_fd> <args>'.
3832 * Interpretation of args is defined by control file implementation.
3834 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3835 char *buf, size_t nbytes, loff_t off)
3837 struct cgroup_subsys_state *css = of_css(of);
3838 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3839 struct mem_cgroup_event *event;
3840 struct cgroup_subsys_state *cfile_css;
3841 unsigned int efd, cfd;
3848 buf = strstrip(buf);
3850 efd = simple_strtoul(buf, &endp, 10);
3855 cfd = simple_strtoul(buf, &endp, 10);
3856 if ((*endp != ' ') && (*endp != '\0'))
3860 event = kzalloc(sizeof(*event), GFP_KERNEL);
3864 event->memcg = memcg;
3865 INIT_LIST_HEAD(&event->list);
3866 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3867 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3868 INIT_WORK(&event->remove, memcg_event_remove);
3876 event->eventfd = eventfd_ctx_fileget(efile.file);
3877 if (IS_ERR(event->eventfd)) {
3878 ret = PTR_ERR(event->eventfd);
3885 goto out_put_eventfd;
3888 /* the process need read permission on control file */
3889 /* AV: shouldn't we check that it's been opened for read instead? */
3890 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3895 * Determine the event callbacks and set them in @event. This used
3896 * to be done via struct cftype but cgroup core no longer knows
3897 * about these events. The following is crude but the whole thing
3898 * is for compatibility anyway.
3900 * DO NOT ADD NEW FILES.
3902 name = cfile.file->f_path.dentry->d_name.name;
3904 if (!strcmp(name, "memory.usage_in_bytes")) {
3905 event->register_event = mem_cgroup_usage_register_event;
3906 event->unregister_event = mem_cgroup_usage_unregister_event;
3907 } else if (!strcmp(name, "memory.oom_control")) {
3908 event->register_event = mem_cgroup_oom_register_event;
3909 event->unregister_event = mem_cgroup_oom_unregister_event;
3910 } else if (!strcmp(name, "memory.pressure_level")) {
3911 event->register_event = vmpressure_register_event;
3912 event->unregister_event = vmpressure_unregister_event;
3913 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3914 event->register_event = memsw_cgroup_usage_register_event;
3915 event->unregister_event = memsw_cgroup_usage_unregister_event;
3922 * Verify @cfile should belong to @css. Also, remaining events are
3923 * automatically removed on cgroup destruction but the removal is
3924 * asynchronous, so take an extra ref on @css.
3926 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3927 &memory_cgrp_subsys);
3929 if (IS_ERR(cfile_css))
3931 if (cfile_css != css) {
3936 ret = event->register_event(memcg, event->eventfd, buf);
3940 efile.file->f_op->poll(efile.file, &event->pt);
3942 spin_lock(&memcg->event_list_lock);
3943 list_add(&event->list, &memcg->event_list);
3944 spin_unlock(&memcg->event_list_lock);
3956 eventfd_ctx_put(event->eventfd);
3965 static struct cftype mem_cgroup_legacy_files[] = {
3967 .name = "usage_in_bytes",
3968 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3969 .read_u64 = mem_cgroup_read_u64,
3972 .name = "max_usage_in_bytes",
3973 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3974 .write = mem_cgroup_reset,
3975 .read_u64 = mem_cgroup_read_u64,
3978 .name = "limit_in_bytes",
3979 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3980 .write = mem_cgroup_write,
3981 .read_u64 = mem_cgroup_read_u64,
3984 .name = "soft_limit_in_bytes",
3985 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3986 .write = mem_cgroup_write,
3987 .read_u64 = mem_cgroup_read_u64,
3991 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3992 .write = mem_cgroup_reset,
3993 .read_u64 = mem_cgroup_read_u64,
3997 .seq_show = memcg_stat_show,
4000 .name = "force_empty",
4001 .write = mem_cgroup_force_empty_write,
4004 .name = "use_hierarchy",
4005 .write_u64 = mem_cgroup_hierarchy_write,
4006 .read_u64 = mem_cgroup_hierarchy_read,
4009 .name = "cgroup.event_control", /* XXX: for compat */
4010 .write = memcg_write_event_control,
4011 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4014 .name = "swappiness",
4015 .read_u64 = mem_cgroup_swappiness_read,
4016 .write_u64 = mem_cgroup_swappiness_write,
4019 .name = "move_charge_at_immigrate",
4020 .read_u64 = mem_cgroup_move_charge_read,
4021 .write_u64 = mem_cgroup_move_charge_write,
4024 .name = "oom_control",
4025 .seq_show = mem_cgroup_oom_control_read,
4026 .write_u64 = mem_cgroup_oom_control_write,
4027 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4030 .name = "pressure_level",
4034 .name = "numa_stat",
4035 .seq_show = memcg_numa_stat_show,
4039 .name = "kmem.limit_in_bytes",
4040 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4041 .write = mem_cgroup_write,
4042 .read_u64 = mem_cgroup_read_u64,
4045 .name = "kmem.usage_in_bytes",
4046 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4047 .read_u64 = mem_cgroup_read_u64,
4050 .name = "kmem.failcnt",
4051 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4052 .write = mem_cgroup_reset,
4053 .read_u64 = mem_cgroup_read_u64,
4056 .name = "kmem.max_usage_in_bytes",
4057 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4058 .write = mem_cgroup_reset,
4059 .read_u64 = mem_cgroup_read_u64,
4061 #ifdef CONFIG_SLABINFO
4063 .name = "kmem.slabinfo",
4064 .seq_start = slab_start,
4065 .seq_next = slab_next,
4066 .seq_stop = slab_stop,
4067 .seq_show = memcg_slab_show,
4071 .name = "kmem.tcp.limit_in_bytes",
4072 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4073 .write = mem_cgroup_write,
4074 .read_u64 = mem_cgroup_read_u64,
4077 .name = "kmem.tcp.usage_in_bytes",
4078 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4079 .read_u64 = mem_cgroup_read_u64,
4082 .name = "kmem.tcp.failcnt",
4083 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4084 .write = mem_cgroup_reset,
4085 .read_u64 = mem_cgroup_read_u64,
4088 .name = "kmem.tcp.max_usage_in_bytes",
4089 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4090 .write = mem_cgroup_reset,
4091 .read_u64 = mem_cgroup_read_u64,
4093 { }, /* terminate */
4096 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4098 struct mem_cgroup_per_node *pn;
4099 struct mem_cgroup_per_zone *mz;
4100 int zone, tmp = node;
4102 * This routine is called against possible nodes.
4103 * But it's BUG to call kmalloc() against offline node.
4105 * TODO: this routine can waste much memory for nodes which will
4106 * never be onlined. It's better to use memory hotplug callback
4109 if (!node_state(node, N_NORMAL_MEMORY))
4111 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4115 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4116 mz = &pn->zoneinfo[zone];
4117 lruvec_init(&mz->lruvec);
4118 mz->usage_in_excess = 0;
4119 mz->on_tree = false;
4122 memcg->nodeinfo[node] = pn;
4126 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4128 kfree(memcg->nodeinfo[node]);
4131 static void mem_cgroup_free(struct mem_cgroup *memcg)
4135 memcg_wb_domain_exit(memcg);
4137 free_mem_cgroup_per_zone_info(memcg, node);
4138 free_percpu(memcg->stat);
4142 static struct mem_cgroup *mem_cgroup_alloc(void)
4144 struct mem_cgroup *memcg;
4148 size = sizeof(struct mem_cgroup);
4149 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4151 memcg = kzalloc(size, GFP_KERNEL);
4155 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4160 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4163 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4166 INIT_WORK(&memcg->high_work, high_work_func);
4167 memcg->last_scanned_node = MAX_NUMNODES;
4168 INIT_LIST_HEAD(&memcg->oom_notify);
4169 mutex_init(&memcg->thresholds_lock);
4170 spin_lock_init(&memcg->move_lock);
4171 vmpressure_init(&memcg->vmpressure);
4172 INIT_LIST_HEAD(&memcg->event_list);
4173 spin_lock_init(&memcg->event_list_lock);
4174 memcg->socket_pressure = jiffies;
4176 memcg->kmemcg_id = -1;
4178 #ifdef CONFIG_CGROUP_WRITEBACK
4179 INIT_LIST_HEAD(&memcg->cgwb_list);
4183 mem_cgroup_free(memcg);
4187 static struct cgroup_subsys_state * __ref
4188 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4190 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4191 struct mem_cgroup *memcg;
4192 long error = -ENOMEM;
4194 memcg = mem_cgroup_alloc();
4196 return ERR_PTR(error);
4198 memcg->high = PAGE_COUNTER_MAX;
4199 memcg->soft_limit = PAGE_COUNTER_MAX;
4201 memcg->swappiness = mem_cgroup_swappiness(parent);
4202 memcg->oom_kill_disable = parent->oom_kill_disable;
4204 if (parent && parent->use_hierarchy) {
4205 memcg->use_hierarchy = true;
4206 page_counter_init(&memcg->memory, &parent->memory);
4207 page_counter_init(&memcg->swap, &parent->swap);
4208 page_counter_init(&memcg->memsw, &parent->memsw);
4209 page_counter_init(&memcg->kmem, &parent->kmem);
4210 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4212 page_counter_init(&memcg->memory, NULL);
4213 page_counter_init(&memcg->swap, NULL);
4214 page_counter_init(&memcg->memsw, NULL);
4215 page_counter_init(&memcg->kmem, NULL);
4216 page_counter_init(&memcg->tcpmem, NULL);
4218 * Deeper hierachy with use_hierarchy == false doesn't make
4219 * much sense so let cgroup subsystem know about this
4220 * unfortunate state in our controller.
4222 if (parent != root_mem_cgroup)
4223 memory_cgrp_subsys.broken_hierarchy = true;
4226 /* The following stuff does not apply to the root */
4228 root_mem_cgroup = memcg;
4232 error = memcg_propagate_kmem(parent, memcg);
4236 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4237 static_branch_inc(&memcg_sockets_enabled_key);
4241 mem_cgroup_free(memcg);
4246 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4248 if (css->id > MEM_CGROUP_ID_MAX)
4254 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4256 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4257 struct mem_cgroup_event *event, *tmp;
4260 * Unregister events and notify userspace.
4261 * Notify userspace about cgroup removing only after rmdir of cgroup
4262 * directory to avoid race between userspace and kernelspace.
4264 spin_lock(&memcg->event_list_lock);
4265 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4266 list_del_init(&event->list);
4267 schedule_work(&event->remove);
4269 spin_unlock(&memcg->event_list_lock);
4271 memcg_offline_kmem(memcg);
4272 wb_memcg_offline(memcg);
4275 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4277 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4279 invalidate_reclaim_iterators(memcg);
4282 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4284 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4286 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4287 static_branch_dec(&memcg_sockets_enabled_key);
4289 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4290 static_branch_dec(&memcg_sockets_enabled_key);
4292 vmpressure_cleanup(&memcg->vmpressure);
4293 cancel_work_sync(&memcg->high_work);
4294 mem_cgroup_remove_from_trees(memcg);
4295 memcg_free_kmem(memcg);
4296 mem_cgroup_free(memcg);
4300 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4301 * @css: the target css
4303 * Reset the states of the mem_cgroup associated with @css. This is
4304 * invoked when the userland requests disabling on the default hierarchy
4305 * but the memcg is pinned through dependency. The memcg should stop
4306 * applying policies and should revert to the vanilla state as it may be
4307 * made visible again.
4309 * The current implementation only resets the essential configurations.
4310 * This needs to be expanded to cover all the visible parts.
4312 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4314 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4316 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4317 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4318 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4320 memcg->high = PAGE_COUNTER_MAX;
4321 memcg->soft_limit = PAGE_COUNTER_MAX;
4322 memcg_wb_domain_size_changed(memcg);
4326 /* Handlers for move charge at task migration. */
4327 static int mem_cgroup_do_precharge(unsigned long count)
4331 /* Try a single bulk charge without reclaim first, kswapd may wake */
4332 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4334 mc.precharge += count;
4338 /* Try charges one by one with reclaim */
4340 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4350 * get_mctgt_type - get target type of moving charge
4351 * @vma: the vma the pte to be checked belongs
4352 * @addr: the address corresponding to the pte to be checked
4353 * @ptent: the pte to be checked
4354 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4357 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4358 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4359 * move charge. if @target is not NULL, the page is stored in target->page
4360 * with extra refcnt got(Callers should handle it).
4361 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4362 * target for charge migration. if @target is not NULL, the entry is stored
4365 * Called with pte lock held.
4372 enum mc_target_type {
4378 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4379 unsigned long addr, pte_t ptent)
4381 struct page *page = vm_normal_page(vma, addr, ptent);
4383 if (!page || !page_mapped(page))
4385 if (PageAnon(page)) {
4386 if (!(mc.flags & MOVE_ANON))
4389 if (!(mc.flags & MOVE_FILE))
4392 if (!get_page_unless_zero(page))
4399 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4400 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4402 struct page *page = NULL;
4403 swp_entry_t ent = pte_to_swp_entry(ptent);
4405 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4408 * Because lookup_swap_cache() updates some statistics counter,
4409 * we call find_get_page() with swapper_space directly.
4411 page = find_get_page(swap_address_space(ent), ent.val);
4412 if (do_memsw_account())
4413 entry->val = ent.val;
4418 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4419 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4425 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4426 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4428 struct page *page = NULL;
4429 struct address_space *mapping;
4432 if (!vma->vm_file) /* anonymous vma */
4434 if (!(mc.flags & MOVE_FILE))
4437 mapping = vma->vm_file->f_mapping;
4438 pgoff = linear_page_index(vma, addr);
4440 /* page is moved even if it's not RSS of this task(page-faulted). */
4442 /* shmem/tmpfs may report page out on swap: account for that too. */
4443 if (shmem_mapping(mapping)) {
4444 page = find_get_entry(mapping, pgoff);
4445 if (radix_tree_exceptional_entry(page)) {
4446 swp_entry_t swp = radix_to_swp_entry(page);
4447 if (do_memsw_account())
4449 page = find_get_page(swap_address_space(swp), swp.val);
4452 page = find_get_page(mapping, pgoff);
4454 page = find_get_page(mapping, pgoff);
4460 * mem_cgroup_move_account - move account of the page
4462 * @nr_pages: number of regular pages (>1 for huge pages)
4463 * @from: mem_cgroup which the page is moved from.
4464 * @to: mem_cgroup which the page is moved to. @from != @to.
4466 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4468 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4471 static int mem_cgroup_move_account(struct page *page,
4473 struct mem_cgroup *from,
4474 struct mem_cgroup *to)
4476 unsigned long flags;
4477 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4481 VM_BUG_ON(from == to);
4482 VM_BUG_ON_PAGE(PageLRU(page), page);
4483 VM_BUG_ON(compound && !PageTransHuge(page));
4486 * Prevent mem_cgroup_replace_page() from looking at
4487 * page->mem_cgroup of its source page while we change it.
4490 if (!trylock_page(page))
4494 if (page->mem_cgroup != from)
4497 anon = PageAnon(page);
4499 spin_lock_irqsave(&from->move_lock, flags);
4501 if (!anon && page_mapped(page)) {
4502 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4504 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4509 * move_lock grabbed above and caller set from->moving_account, so
4510 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4511 * So mapping should be stable for dirty pages.
4513 if (!anon && PageDirty(page)) {
4514 struct address_space *mapping = page_mapping(page);
4516 if (mapping_cap_account_dirty(mapping)) {
4517 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4519 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4524 if (PageWriteback(page)) {
4525 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4527 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4532 * It is safe to change page->mem_cgroup here because the page
4533 * is referenced, charged, and isolated - we can't race with
4534 * uncharging, charging, migration, or LRU putback.
4537 /* caller should have done css_get */
4538 page->mem_cgroup = to;
4539 spin_unlock_irqrestore(&from->move_lock, flags);
4543 local_irq_disable();
4544 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4545 memcg_check_events(to, page);
4546 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4547 memcg_check_events(from, page);
4555 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4556 unsigned long addr, pte_t ptent, union mc_target *target)
4558 struct page *page = NULL;
4559 enum mc_target_type ret = MC_TARGET_NONE;
4560 swp_entry_t ent = { .val = 0 };
4562 if (pte_present(ptent))
4563 page = mc_handle_present_pte(vma, addr, ptent);
4564 else if (is_swap_pte(ptent))
4565 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4566 else if (pte_none(ptent))
4567 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4569 if (!page && !ent.val)
4573 * Do only loose check w/o serialization.
4574 * mem_cgroup_move_account() checks the page is valid or
4575 * not under LRU exclusion.
4577 if (page->mem_cgroup == mc.from) {
4578 ret = MC_TARGET_PAGE;
4580 target->page = page;
4582 if (!ret || !target)
4585 /* There is a swap entry and a page doesn't exist or isn't charged */
4586 if (ent.val && !ret &&
4587 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4588 ret = MC_TARGET_SWAP;
4595 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4597 * We don't consider swapping or file mapped pages because THP does not
4598 * support them for now.
4599 * Caller should make sure that pmd_trans_huge(pmd) is true.
4601 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4602 unsigned long addr, pmd_t pmd, union mc_target *target)
4604 struct page *page = NULL;
4605 enum mc_target_type ret = MC_TARGET_NONE;
4607 page = pmd_page(pmd);
4608 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4609 if (!(mc.flags & MOVE_ANON))
4611 if (page->mem_cgroup == mc.from) {
4612 ret = MC_TARGET_PAGE;
4615 target->page = page;
4621 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4622 unsigned long addr, pmd_t pmd, union mc_target *target)
4624 return MC_TARGET_NONE;
4628 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4629 unsigned long addr, unsigned long end,
4630 struct mm_walk *walk)
4632 struct vm_area_struct *vma = walk->vma;
4636 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4637 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4638 mc.precharge += HPAGE_PMD_NR;
4643 if (pmd_trans_unstable(pmd))
4645 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4646 for (; addr != end; pte++, addr += PAGE_SIZE)
4647 if (get_mctgt_type(vma, addr, *pte, NULL))
4648 mc.precharge++; /* increment precharge temporarily */
4649 pte_unmap_unlock(pte - 1, ptl);
4655 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4657 unsigned long precharge;
4659 struct mm_walk mem_cgroup_count_precharge_walk = {
4660 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4663 down_read(&mm->mmap_sem);
4664 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4665 up_read(&mm->mmap_sem);
4667 precharge = mc.precharge;
4673 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4675 unsigned long precharge = mem_cgroup_count_precharge(mm);
4677 VM_BUG_ON(mc.moving_task);
4678 mc.moving_task = current;
4679 return mem_cgroup_do_precharge(precharge);
4682 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4683 static void __mem_cgroup_clear_mc(void)
4685 struct mem_cgroup *from = mc.from;
4686 struct mem_cgroup *to = mc.to;
4688 /* we must uncharge all the leftover precharges from mc.to */
4690 cancel_charge(mc.to, mc.precharge);
4694 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4695 * we must uncharge here.
4697 if (mc.moved_charge) {
4698 cancel_charge(mc.from, mc.moved_charge);
4699 mc.moved_charge = 0;
4701 /* we must fixup refcnts and charges */
4702 if (mc.moved_swap) {
4703 /* uncharge swap account from the old cgroup */
4704 if (!mem_cgroup_is_root(mc.from))
4705 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4708 * we charged both to->memory and to->memsw, so we
4709 * should uncharge to->memory.
4711 if (!mem_cgroup_is_root(mc.to))
4712 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4714 css_put_many(&mc.from->css, mc.moved_swap);
4716 /* we've already done css_get(mc.to) */
4719 memcg_oom_recover(from);
4720 memcg_oom_recover(to);
4721 wake_up_all(&mc.waitq);
4724 static void mem_cgroup_clear_mc(void)
4727 * we must clear moving_task before waking up waiters at the end of
4730 mc.moving_task = NULL;
4731 __mem_cgroup_clear_mc();
4732 spin_lock(&mc.lock);
4735 spin_unlock(&mc.lock);
4738 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4740 struct cgroup_subsys_state *css;
4741 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4742 struct mem_cgroup *from;
4743 struct task_struct *leader, *p;
4744 struct mm_struct *mm;
4745 unsigned long move_flags;
4748 /* charge immigration isn't supported on the default hierarchy */
4749 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4753 * Multi-process migrations only happen on the default hierarchy
4754 * where charge immigration is not used. Perform charge
4755 * immigration if @tset contains a leader and whine if there are
4759 cgroup_taskset_for_each_leader(leader, css, tset) {
4762 memcg = mem_cgroup_from_css(css);
4768 * We are now commited to this value whatever it is. Changes in this
4769 * tunable will only affect upcoming migrations, not the current one.
4770 * So we need to save it, and keep it going.
4772 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4776 from = mem_cgroup_from_task(p);
4778 VM_BUG_ON(from == memcg);
4780 mm = get_task_mm(p);
4783 /* We move charges only when we move a owner of the mm */
4784 if (mm->owner == p) {
4787 VM_BUG_ON(mc.precharge);
4788 VM_BUG_ON(mc.moved_charge);
4789 VM_BUG_ON(mc.moved_swap);
4791 spin_lock(&mc.lock);
4794 mc.flags = move_flags;
4795 spin_unlock(&mc.lock);
4796 /* We set mc.moving_task later */
4798 ret = mem_cgroup_precharge_mc(mm);
4800 mem_cgroup_clear_mc();
4806 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4809 mem_cgroup_clear_mc();
4812 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4813 unsigned long addr, unsigned long end,
4814 struct mm_walk *walk)
4817 struct vm_area_struct *vma = walk->vma;
4820 enum mc_target_type target_type;
4821 union mc_target target;
4824 if (pmd_trans_huge_lock(pmd, vma, &ptl)) {
4825 if (mc.precharge < HPAGE_PMD_NR) {
4829 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4830 if (target_type == MC_TARGET_PAGE) {
4832 if (!isolate_lru_page(page)) {
4833 if (!mem_cgroup_move_account(page, true,
4835 mc.precharge -= HPAGE_PMD_NR;
4836 mc.moved_charge += HPAGE_PMD_NR;
4838 putback_lru_page(page);
4846 if (pmd_trans_unstable(pmd))
4849 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4850 for (; addr != end; addr += PAGE_SIZE) {
4851 pte_t ptent = *(pte++);
4857 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4858 case MC_TARGET_PAGE:
4861 * We can have a part of the split pmd here. Moving it
4862 * can be done but it would be too convoluted so simply
4863 * ignore such a partial THP and keep it in original
4864 * memcg. There should be somebody mapping the head.
4866 if (PageTransCompound(page))
4868 if (isolate_lru_page(page))
4870 if (!mem_cgroup_move_account(page, false,
4873 /* we uncharge from mc.from later. */
4876 putback_lru_page(page);
4877 put: /* get_mctgt_type() gets the page */
4880 case MC_TARGET_SWAP:
4882 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4884 /* we fixup refcnts and charges later. */
4892 pte_unmap_unlock(pte - 1, ptl);
4897 * We have consumed all precharges we got in can_attach().
4898 * We try charge one by one, but don't do any additional
4899 * charges to mc.to if we have failed in charge once in attach()
4902 ret = mem_cgroup_do_precharge(1);
4910 static void mem_cgroup_move_charge(struct mm_struct *mm)
4912 struct mm_walk mem_cgroup_move_charge_walk = {
4913 .pmd_entry = mem_cgroup_move_charge_pte_range,
4917 lru_add_drain_all();
4919 * Signal mem_cgroup_begin_page_stat() to take the memcg's
4920 * move_lock while we're moving its pages to another memcg.
4921 * Then wait for already started RCU-only updates to finish.
4923 atomic_inc(&mc.from->moving_account);
4926 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4928 * Someone who are holding the mmap_sem might be waiting in
4929 * waitq. So we cancel all extra charges, wake up all waiters,
4930 * and retry. Because we cancel precharges, we might not be able
4931 * to move enough charges, but moving charge is a best-effort
4932 * feature anyway, so it wouldn't be a big problem.
4934 __mem_cgroup_clear_mc();
4939 * When we have consumed all precharges and failed in doing
4940 * additional charge, the page walk just aborts.
4942 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4943 up_read(&mm->mmap_sem);
4944 atomic_dec(&mc.from->moving_account);
4947 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4949 struct cgroup_subsys_state *css;
4950 struct task_struct *p = cgroup_taskset_first(tset, &css);
4951 struct mm_struct *mm = get_task_mm(p);
4955 mem_cgroup_move_charge(mm);
4959 mem_cgroup_clear_mc();
4961 #else /* !CONFIG_MMU */
4962 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4966 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4969 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4975 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4976 * to verify whether we're attached to the default hierarchy on each mount
4979 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
4982 * use_hierarchy is forced on the default hierarchy. cgroup core
4983 * guarantees that @root doesn't have any children, so turning it
4984 * on for the root memcg is enough.
4986 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4987 root_mem_cgroup->use_hierarchy = true;
4989 root_mem_cgroup->use_hierarchy = false;
4992 static u64 memory_current_read(struct cgroup_subsys_state *css,
4995 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4997 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5000 static int memory_low_show(struct seq_file *m, void *v)
5002 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5003 unsigned long low = READ_ONCE(memcg->low);
5005 if (low == PAGE_COUNTER_MAX)
5006 seq_puts(m, "max\n");
5008 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5013 static ssize_t memory_low_write(struct kernfs_open_file *of,
5014 char *buf, size_t nbytes, loff_t off)
5016 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5020 buf = strstrip(buf);
5021 err = page_counter_memparse(buf, "max", &low);
5030 static int memory_high_show(struct seq_file *m, void *v)
5032 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5033 unsigned long high = READ_ONCE(memcg->high);
5035 if (high == PAGE_COUNTER_MAX)
5036 seq_puts(m, "max\n");
5038 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5043 static ssize_t memory_high_write(struct kernfs_open_file *of,
5044 char *buf, size_t nbytes, loff_t off)
5046 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5050 buf = strstrip(buf);
5051 err = page_counter_memparse(buf, "max", &high);
5057 memcg_wb_domain_size_changed(memcg);
5061 static int memory_max_show(struct seq_file *m, void *v)
5063 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5064 unsigned long max = READ_ONCE(memcg->memory.limit);
5066 if (max == PAGE_COUNTER_MAX)
5067 seq_puts(m, "max\n");
5069 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5074 static ssize_t memory_max_write(struct kernfs_open_file *of,
5075 char *buf, size_t nbytes, loff_t off)
5077 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5081 buf = strstrip(buf);
5082 err = page_counter_memparse(buf, "max", &max);
5086 err = mem_cgroup_resize_limit(memcg, max);
5090 memcg_wb_domain_size_changed(memcg);
5094 static int memory_events_show(struct seq_file *m, void *v)
5096 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5098 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5099 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5100 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5101 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5106 static struct cftype memory_files[] = {
5109 .flags = CFTYPE_NOT_ON_ROOT,
5110 .read_u64 = memory_current_read,
5114 .flags = CFTYPE_NOT_ON_ROOT,
5115 .seq_show = memory_low_show,
5116 .write = memory_low_write,
5120 .flags = CFTYPE_NOT_ON_ROOT,
5121 .seq_show = memory_high_show,
5122 .write = memory_high_write,
5126 .flags = CFTYPE_NOT_ON_ROOT,
5127 .seq_show = memory_max_show,
5128 .write = memory_max_write,
5132 .flags = CFTYPE_NOT_ON_ROOT,
5133 .file_offset = offsetof(struct mem_cgroup, events_file),
5134 .seq_show = memory_events_show,
5139 struct cgroup_subsys memory_cgrp_subsys = {
5140 .css_alloc = mem_cgroup_css_alloc,
5141 .css_online = mem_cgroup_css_online,
5142 .css_offline = mem_cgroup_css_offline,
5143 .css_released = mem_cgroup_css_released,
5144 .css_free = mem_cgroup_css_free,
5145 .css_reset = mem_cgroup_css_reset,
5146 .can_attach = mem_cgroup_can_attach,
5147 .cancel_attach = mem_cgroup_cancel_attach,
5148 .attach = mem_cgroup_move_task,
5149 .bind = mem_cgroup_bind,
5150 .dfl_cftypes = memory_files,
5151 .legacy_cftypes = mem_cgroup_legacy_files,
5156 * mem_cgroup_low - check if memory consumption is below the normal range
5157 * @root: the highest ancestor to consider
5158 * @memcg: the memory cgroup to check
5160 * Returns %true if memory consumption of @memcg, and that of all
5161 * configurable ancestors up to @root, is below the normal range.
5163 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5165 if (mem_cgroup_disabled())
5169 * The toplevel group doesn't have a configurable range, so
5170 * it's never low when looked at directly, and it is not
5171 * considered an ancestor when assessing the hierarchy.
5174 if (memcg == root_mem_cgroup)
5177 if (page_counter_read(&memcg->memory) >= memcg->low)
5180 while (memcg != root) {
5181 memcg = parent_mem_cgroup(memcg);
5183 if (memcg == root_mem_cgroup)
5186 if (page_counter_read(&memcg->memory) >= memcg->low)
5193 * mem_cgroup_try_charge - try charging a page
5194 * @page: page to charge
5195 * @mm: mm context of the victim
5196 * @gfp_mask: reclaim mode
5197 * @memcgp: charged memcg return
5199 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5200 * pages according to @gfp_mask if necessary.
5202 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5203 * Otherwise, an error code is returned.
5205 * After page->mapping has been set up, the caller must finalize the
5206 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5207 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5209 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5210 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5213 struct mem_cgroup *memcg = NULL;
5214 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5217 if (mem_cgroup_disabled())
5220 if (PageSwapCache(page)) {
5222 * Every swap fault against a single page tries to charge the
5223 * page, bail as early as possible. shmem_unuse() encounters
5224 * already charged pages, too. The USED bit is protected by
5225 * the page lock, which serializes swap cache removal, which
5226 * in turn serializes uncharging.
5228 VM_BUG_ON_PAGE(!PageLocked(page), page);
5229 if (page->mem_cgroup)
5232 if (do_swap_account) {
5233 swp_entry_t ent = { .val = page_private(page), };
5234 unsigned short id = lookup_swap_cgroup_id(ent);
5237 memcg = mem_cgroup_from_id(id);
5238 if (memcg && !css_tryget_online(&memcg->css))
5245 memcg = get_mem_cgroup_from_mm(mm);
5247 ret = try_charge(memcg, gfp_mask, nr_pages);
5249 css_put(&memcg->css);
5256 * mem_cgroup_commit_charge - commit a page charge
5257 * @page: page to charge
5258 * @memcg: memcg to charge the page to
5259 * @lrucare: page might be on LRU already
5261 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5262 * after page->mapping has been set up. This must happen atomically
5263 * as part of the page instantiation, i.e. under the page table lock
5264 * for anonymous pages, under the page lock for page and swap cache.
5266 * In addition, the page must not be on the LRU during the commit, to
5267 * prevent racing with task migration. If it might be, use @lrucare.
5269 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5271 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5272 bool lrucare, bool compound)
5274 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5276 VM_BUG_ON_PAGE(!page->mapping, page);
5277 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5279 if (mem_cgroup_disabled())
5282 * Swap faults will attempt to charge the same page multiple
5283 * times. But reuse_swap_page() might have removed the page
5284 * from swapcache already, so we can't check PageSwapCache().
5289 commit_charge(page, memcg, lrucare);
5291 local_irq_disable();
5292 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5293 memcg_check_events(memcg, page);
5296 if (do_memsw_account() && PageSwapCache(page)) {
5297 swp_entry_t entry = { .val = page_private(page) };
5299 * The swap entry might not get freed for a long time,
5300 * let's not wait for it. The page already received a
5301 * memory+swap charge, drop the swap entry duplicate.
5303 mem_cgroup_uncharge_swap(entry);
5308 * mem_cgroup_cancel_charge - cancel a page charge
5309 * @page: page to charge
5310 * @memcg: memcg to charge the page to
5312 * Cancel a charge transaction started by mem_cgroup_try_charge().
5314 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5317 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5319 if (mem_cgroup_disabled())
5322 * Swap faults will attempt to charge the same page multiple
5323 * times. But reuse_swap_page() might have removed the page
5324 * from swapcache already, so we can't check PageSwapCache().
5329 cancel_charge(memcg, nr_pages);
5332 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5333 unsigned long nr_anon, unsigned long nr_file,
5334 unsigned long nr_huge, struct page *dummy_page)
5336 unsigned long nr_pages = nr_anon + nr_file;
5337 unsigned long flags;
5339 if (!mem_cgroup_is_root(memcg)) {
5340 page_counter_uncharge(&memcg->memory, nr_pages);
5341 if (do_memsw_account())
5342 page_counter_uncharge(&memcg->memsw, nr_pages);
5343 memcg_oom_recover(memcg);
5346 local_irq_save(flags);
5347 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5348 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5349 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5350 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5351 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5352 memcg_check_events(memcg, dummy_page);
5353 local_irq_restore(flags);
5355 if (!mem_cgroup_is_root(memcg))
5356 css_put_many(&memcg->css, nr_pages);
5359 static void uncharge_list(struct list_head *page_list)
5361 struct mem_cgroup *memcg = NULL;
5362 unsigned long nr_anon = 0;
5363 unsigned long nr_file = 0;
5364 unsigned long nr_huge = 0;
5365 unsigned long pgpgout = 0;
5366 struct list_head *next;
5369 next = page_list->next;
5371 unsigned int nr_pages = 1;
5373 page = list_entry(next, struct page, lru);
5374 next = page->lru.next;
5376 VM_BUG_ON_PAGE(PageLRU(page), page);
5377 VM_BUG_ON_PAGE(page_count(page), page);
5379 if (!page->mem_cgroup)
5383 * Nobody should be changing or seriously looking at
5384 * page->mem_cgroup at this point, we have fully
5385 * exclusive access to the page.
5388 if (memcg != page->mem_cgroup) {
5390 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5392 pgpgout = nr_anon = nr_file = nr_huge = 0;
5394 memcg = page->mem_cgroup;
5397 if (PageTransHuge(page)) {
5398 nr_pages <<= compound_order(page);
5399 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5400 nr_huge += nr_pages;
5404 nr_anon += nr_pages;
5406 nr_file += nr_pages;
5408 page->mem_cgroup = NULL;
5411 } while (next != page_list);
5414 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5419 * mem_cgroup_uncharge - uncharge a page
5420 * @page: page to uncharge
5422 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5423 * mem_cgroup_commit_charge().
5425 void mem_cgroup_uncharge(struct page *page)
5427 if (mem_cgroup_disabled())
5430 /* Don't touch page->lru of any random page, pre-check: */
5431 if (!page->mem_cgroup)
5434 INIT_LIST_HEAD(&page->lru);
5435 uncharge_list(&page->lru);
5439 * mem_cgroup_uncharge_list - uncharge a list of page
5440 * @page_list: list of pages to uncharge
5442 * Uncharge a list of pages previously charged with
5443 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5445 void mem_cgroup_uncharge_list(struct list_head *page_list)
5447 if (mem_cgroup_disabled())
5450 if (!list_empty(page_list))
5451 uncharge_list(page_list);
5455 * mem_cgroup_replace_page - migrate a charge to another page
5456 * @oldpage: currently charged page
5457 * @newpage: page to transfer the charge to
5459 * Migrate the charge from @oldpage to @newpage.
5461 * Both pages must be locked, @newpage->mapping must be set up.
5462 * Either or both pages might be on the LRU already.
5464 void mem_cgroup_replace_page(struct page *oldpage, struct page *newpage)
5466 struct mem_cgroup *memcg;
5469 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5470 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5471 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5472 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5475 if (mem_cgroup_disabled())
5478 /* Page cache replacement: new page already charged? */
5479 if (newpage->mem_cgroup)
5482 /* Swapcache readahead pages can get replaced before being charged */
5483 memcg = oldpage->mem_cgroup;
5487 lock_page_lru(oldpage, &isolated);
5488 oldpage->mem_cgroup = NULL;
5489 unlock_page_lru(oldpage, isolated);
5491 commit_charge(newpage, memcg, true);
5494 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5495 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5497 void sock_update_memcg(struct sock *sk)
5499 struct mem_cgroup *memcg;
5501 /* Socket cloning can throw us here with sk_cgrp already
5502 * filled. It won't however, necessarily happen from
5503 * process context. So the test for root memcg given
5504 * the current task's memcg won't help us in this case.
5506 * Respecting the original socket's memcg is a better
5507 * decision in this case.
5510 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5511 css_get(&sk->sk_memcg->css);
5516 memcg = mem_cgroup_from_task(current);
5517 if (memcg == root_mem_cgroup)
5519 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5521 if (css_tryget_online(&memcg->css))
5522 sk->sk_memcg = memcg;
5526 EXPORT_SYMBOL(sock_update_memcg);
5528 void sock_release_memcg(struct sock *sk)
5530 WARN_ON(!sk->sk_memcg);
5531 css_put(&sk->sk_memcg->css);
5535 * mem_cgroup_charge_skmem - charge socket memory
5536 * @memcg: memcg to charge
5537 * @nr_pages: number of pages to charge
5539 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5540 * @memcg's configured limit, %false if the charge had to be forced.
5542 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5544 gfp_t gfp_mask = GFP_KERNEL;
5546 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5547 struct page_counter *fail;
5549 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5550 memcg->tcpmem_pressure = 0;
5553 page_counter_charge(&memcg->tcpmem, nr_pages);
5554 memcg->tcpmem_pressure = 1;
5558 /* Don't block in the packet receive path */
5560 gfp_mask = GFP_NOWAIT;
5562 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5565 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5570 * mem_cgroup_uncharge_skmem - uncharge socket memory
5571 * @memcg - memcg to uncharge
5572 * @nr_pages - number of pages to uncharge
5574 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5576 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5577 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5581 page_counter_uncharge(&memcg->memory, nr_pages);
5582 css_put_many(&memcg->css, nr_pages);
5585 static int __init cgroup_memory(char *s)
5589 while ((token = strsep(&s, ",")) != NULL) {
5592 if (!strcmp(token, "nosocket"))
5593 cgroup_memory_nosocket = true;
5594 if (!strcmp(token, "nokmem"))
5595 cgroup_memory_nokmem = true;
5599 __setup("cgroup.memory=", cgroup_memory);
5602 * subsys_initcall() for memory controller.
5604 * Some parts like hotcpu_notifier() have to be initialized from this context
5605 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5606 * everything that doesn't depend on a specific mem_cgroup structure should
5607 * be initialized from here.
5609 static int __init mem_cgroup_init(void)
5613 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5615 for_each_possible_cpu(cpu)
5616 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5619 for_each_node(node) {
5620 struct mem_cgroup_tree_per_node *rtpn;
5623 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5624 node_online(node) ? node : NUMA_NO_NODE);
5626 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5627 struct mem_cgroup_tree_per_zone *rtpz;
5629 rtpz = &rtpn->rb_tree_per_zone[zone];
5630 rtpz->rb_root = RB_ROOT;
5631 spin_lock_init(&rtpz->lock);
5633 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5638 subsys_initcall(mem_cgroup_init);
5640 #ifdef CONFIG_MEMCG_SWAP
5642 * mem_cgroup_swapout - transfer a memsw charge to swap
5643 * @page: page whose memsw charge to transfer
5644 * @entry: swap entry to move the charge to
5646 * Transfer the memsw charge of @page to @entry.
5648 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5650 struct mem_cgroup *memcg;
5651 unsigned short oldid;
5653 VM_BUG_ON_PAGE(PageLRU(page), page);
5654 VM_BUG_ON_PAGE(page_count(page), page);
5656 if (!do_memsw_account())
5659 memcg = page->mem_cgroup;
5661 /* Readahead page, never charged */
5665 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5666 VM_BUG_ON_PAGE(oldid, page);
5667 mem_cgroup_swap_statistics(memcg, true);
5669 page->mem_cgroup = NULL;
5671 if (!mem_cgroup_is_root(memcg))
5672 page_counter_uncharge(&memcg->memory, 1);
5675 * Interrupts should be disabled here because the caller holds the
5676 * mapping->tree_lock lock which is taken with interrupts-off. It is
5677 * important here to have the interrupts disabled because it is the
5678 * only synchronisation we have for udpating the per-CPU variables.
5680 VM_BUG_ON(!irqs_disabled());
5681 mem_cgroup_charge_statistics(memcg, page, false, -1);
5682 memcg_check_events(memcg, page);
5686 * mem_cgroup_try_charge_swap - try charging a swap entry
5687 * @page: page being added to swap
5688 * @entry: swap entry to charge
5690 * Try to charge @entry to the memcg that @page belongs to.
5692 * Returns 0 on success, -ENOMEM on failure.
5694 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5696 struct mem_cgroup *memcg;
5697 struct page_counter *counter;
5698 unsigned short oldid;
5700 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5703 memcg = page->mem_cgroup;
5705 /* Readahead page, never charged */
5709 if (!mem_cgroup_is_root(memcg) &&
5710 !page_counter_try_charge(&memcg->swap, 1, &counter))
5713 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5714 VM_BUG_ON_PAGE(oldid, page);
5715 mem_cgroup_swap_statistics(memcg, true);
5717 css_get(&memcg->css);
5722 * mem_cgroup_uncharge_swap - uncharge a swap entry
5723 * @entry: swap entry to uncharge
5725 * Drop the swap charge associated with @entry.
5727 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5729 struct mem_cgroup *memcg;
5732 if (!do_swap_account)
5735 id = swap_cgroup_record(entry, 0);
5737 memcg = mem_cgroup_from_id(id);
5739 if (!mem_cgroup_is_root(memcg)) {
5740 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5741 page_counter_uncharge(&memcg->swap, 1);
5743 page_counter_uncharge(&memcg->memsw, 1);
5745 mem_cgroup_swap_statistics(memcg, false);
5746 css_put(&memcg->css);
5751 /* for remember boot option*/
5752 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5753 static int really_do_swap_account __initdata = 1;
5755 static int really_do_swap_account __initdata;
5758 static int __init enable_swap_account(char *s)
5760 if (!strcmp(s, "1"))
5761 really_do_swap_account = 1;
5762 else if (!strcmp(s, "0"))
5763 really_do_swap_account = 0;
5766 __setup("swapaccount=", enable_swap_account);
5768 static u64 swap_current_read(struct cgroup_subsys_state *css,
5771 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5773 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
5776 static int swap_max_show(struct seq_file *m, void *v)
5778 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5779 unsigned long max = READ_ONCE(memcg->swap.limit);
5781 if (max == PAGE_COUNTER_MAX)
5782 seq_puts(m, "max\n");
5784 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5789 static ssize_t swap_max_write(struct kernfs_open_file *of,
5790 char *buf, size_t nbytes, loff_t off)
5792 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5796 buf = strstrip(buf);
5797 err = page_counter_memparse(buf, "max", &max);
5801 mutex_lock(&memcg_limit_mutex);
5802 err = page_counter_limit(&memcg->swap, max);
5803 mutex_unlock(&memcg_limit_mutex);
5810 static struct cftype swap_files[] = {
5812 .name = "swap.current",
5813 .flags = CFTYPE_NOT_ON_ROOT,
5814 .read_u64 = swap_current_read,
5818 .flags = CFTYPE_NOT_ON_ROOT,
5819 .seq_show = swap_max_show,
5820 .write = swap_max_write,
5825 static struct cftype memsw_cgroup_files[] = {
5827 .name = "memsw.usage_in_bytes",
5828 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5829 .read_u64 = mem_cgroup_read_u64,
5832 .name = "memsw.max_usage_in_bytes",
5833 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5834 .write = mem_cgroup_reset,
5835 .read_u64 = mem_cgroup_read_u64,
5838 .name = "memsw.limit_in_bytes",
5839 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5840 .write = mem_cgroup_write,
5841 .read_u64 = mem_cgroup_read_u64,
5844 .name = "memsw.failcnt",
5845 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5846 .write = mem_cgroup_reset,
5847 .read_u64 = mem_cgroup_read_u64,
5849 { }, /* terminate */
5852 static int __init mem_cgroup_swap_init(void)
5854 if (!mem_cgroup_disabled() && really_do_swap_account) {
5855 do_swap_account = 1;
5856 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
5858 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5859 memsw_cgroup_files));
5863 subsys_initcall(mem_cgroup_swap_init);
5865 #endif /* CONFIG_MEMCG_SWAP */