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_node {
136 struct rb_root rb_root;
140 struct mem_cgroup_tree {
141 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
144 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
147 struct mem_cgroup_eventfd_list {
148 struct list_head list;
149 struct eventfd_ctx *eventfd;
153 * cgroup_event represents events which userspace want to receive.
155 struct mem_cgroup_event {
157 * memcg which the event belongs to.
159 struct mem_cgroup *memcg;
161 * eventfd to signal userspace about the event.
163 struct eventfd_ctx *eventfd;
165 * Each of these stored in a list by the cgroup.
167 struct list_head list;
169 * register_event() callback will be used to add new userspace
170 * waiter for changes related to this event. Use eventfd_signal()
171 * on eventfd to send notification to userspace.
173 int (*register_event)(struct mem_cgroup *memcg,
174 struct eventfd_ctx *eventfd, const char *args);
176 * unregister_event() callback will be called when userspace closes
177 * the eventfd or on cgroup removing. This callback must be set,
178 * if you want provide notification functionality.
180 void (*unregister_event)(struct mem_cgroup *memcg,
181 struct eventfd_ctx *eventfd);
183 * All fields below needed to unregister event when
184 * userspace closes eventfd.
187 wait_queue_head_t *wqh;
189 struct work_struct remove;
192 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
195 /* Stuffs for move charges at task migration. */
197 * Types of charges to be moved.
199 #define MOVE_ANON 0x1U
200 #define MOVE_FILE 0x2U
201 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
203 /* "mc" and its members are protected by cgroup_mutex */
204 static struct move_charge_struct {
205 spinlock_t lock; /* for from, to */
206 struct mm_struct *mm;
207 struct mem_cgroup *from;
208 struct mem_cgroup *to;
210 unsigned long precharge;
211 unsigned long moved_charge;
212 unsigned long moved_swap;
213 struct task_struct *moving_task; /* a task moving charges */
214 wait_queue_head_t waitq; /* a waitq for other context */
216 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
217 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
221 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
222 * limit reclaim to prevent infinite loops, if they ever occur.
224 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
225 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
228 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
229 MEM_CGROUP_CHARGE_TYPE_ANON,
230 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
231 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
235 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
254 memcg = root_mem_cgroup;
255 return &memcg->vmpressure;
258 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
260 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
263 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
265 return (memcg == root_mem_cgroup);
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida);
281 int memcg_nr_cache_ids;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key);
320 #endif /* !CONFIG_SLOB */
323 * mem_cgroup_css_from_page - css of the memcg associated with a page
324 * @page: page of interest
326 * If memcg is bound to the default hierarchy, css of the memcg associated
327 * with @page is returned. The returned css remains associated with @page
328 * until it is released.
330 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
333 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
335 struct mem_cgroup *memcg;
337 memcg = page->mem_cgroup;
339 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
340 memcg = root_mem_cgroup;
346 * page_cgroup_ino - return inode number of the memcg a page is charged to
349 * Look up the closest online ancestor of the memory cgroup @page is charged to
350 * and return its inode number or 0 if @page is not charged to any cgroup. It
351 * is safe to call this function without holding a reference to @page.
353 * Note, this function is inherently racy, because there is nothing to prevent
354 * the cgroup inode from getting torn down and potentially reallocated a moment
355 * after page_cgroup_ino() returns, so it only should be used by callers that
356 * do not care (such as procfs interfaces).
358 ino_t page_cgroup_ino(struct page *page)
360 struct mem_cgroup *memcg;
361 unsigned long ino = 0;
364 memcg = READ_ONCE(page->mem_cgroup);
365 while (memcg && !(memcg->css.flags & CSS_ONLINE))
366 memcg = parent_mem_cgroup(memcg);
368 ino = cgroup_ino(memcg->css.cgroup);
373 static struct mem_cgroup_per_node *
374 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
376 int nid = page_to_nid(page);
378 return memcg->nodeinfo[nid];
381 static struct mem_cgroup_tree_per_node *
382 soft_limit_tree_node(int nid)
384 return soft_limit_tree.rb_tree_per_node[nid];
387 static struct mem_cgroup_tree_per_node *
388 soft_limit_tree_from_page(struct page *page)
390 int nid = page_to_nid(page);
392 return soft_limit_tree.rb_tree_per_node[nid];
395 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
396 struct mem_cgroup_tree_per_node *mctz,
397 unsigned long new_usage_in_excess)
399 struct rb_node **p = &mctz->rb_root.rb_node;
400 struct rb_node *parent = NULL;
401 struct mem_cgroup_per_node *mz_node;
406 mz->usage_in_excess = new_usage_in_excess;
407 if (!mz->usage_in_excess)
411 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
413 if (mz->usage_in_excess < mz_node->usage_in_excess)
416 * We can't avoid mem cgroups that are over their soft
417 * limit by the same amount
419 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
422 rb_link_node(&mz->tree_node, parent, p);
423 rb_insert_color(&mz->tree_node, &mctz->rb_root);
427 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
428 struct mem_cgroup_tree_per_node *mctz)
432 rb_erase(&mz->tree_node, &mctz->rb_root);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
437 struct mem_cgroup_tree_per_node *mctz)
441 spin_lock_irqsave(&mctz->lock, flags);
442 __mem_cgroup_remove_exceeded(mz, mctz);
443 spin_unlock_irqrestore(&mctz->lock, flags);
446 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
448 unsigned long nr_pages = page_counter_read(&memcg->memory);
449 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
450 unsigned long excess = 0;
452 if (nr_pages > soft_limit)
453 excess = nr_pages - soft_limit;
458 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
460 unsigned long excess;
461 struct mem_cgroup_per_node *mz;
462 struct mem_cgroup_tree_per_node *mctz;
464 mctz = soft_limit_tree_from_page(page);
468 * Necessary to update all ancestors when hierarchy is used.
469 * because their event counter is not touched.
471 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
472 mz = mem_cgroup_page_nodeinfo(memcg, page);
473 excess = soft_limit_excess(memcg);
475 * We have to update the tree if mz is on RB-tree or
476 * mem is over its softlimit.
478 if (excess || mz->on_tree) {
481 spin_lock_irqsave(&mctz->lock, flags);
482 /* if on-tree, remove it */
484 __mem_cgroup_remove_exceeded(mz, mctz);
486 * Insert again. mz->usage_in_excess will be updated.
487 * If excess is 0, no tree ops.
489 __mem_cgroup_insert_exceeded(mz, mctz, excess);
490 spin_unlock_irqrestore(&mctz->lock, flags);
495 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
497 struct mem_cgroup_tree_per_node *mctz;
498 struct mem_cgroup_per_node *mz;
502 mz = mem_cgroup_nodeinfo(memcg, nid);
503 mctz = soft_limit_tree_node(nid);
505 mem_cgroup_remove_exceeded(mz, mctz);
509 static struct mem_cgroup_per_node *
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
512 struct rb_node *rightmost = NULL;
513 struct mem_cgroup_per_node *mz;
517 rightmost = rb_last(&mctz->rb_root);
519 goto done; /* Nothing to reclaim from */
521 mz = rb_entry(rightmost, struct mem_cgroup_per_node, tree_node);
523 * Remove the node now but someone else can add it back,
524 * we will to add it back at the end of reclaim to its correct
525 * position in the tree.
527 __mem_cgroup_remove_exceeded(mz, mctz);
528 if (!soft_limit_excess(mz->memcg) ||
529 !css_tryget_online(&mz->memcg->css))
535 static struct mem_cgroup_per_node *
536 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
538 struct mem_cgroup_per_node *mz;
540 spin_lock_irq(&mctz->lock);
541 mz = __mem_cgroup_largest_soft_limit_node(mctz);
542 spin_unlock_irq(&mctz->lock);
547 * Return page count for single (non recursive) @memcg.
549 * Implementation Note: reading percpu statistics for memcg.
551 * Both of vmstat[] and percpu_counter has threshold and do periodic
552 * synchronization to implement "quick" read. There are trade-off between
553 * reading cost and precision of value. Then, we may have a chance to implement
554 * a periodic synchronization of counter in memcg's counter.
556 * But this _read() function is used for user interface now. The user accounts
557 * memory usage by memory cgroup and he _always_ requires exact value because
558 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
559 * have to visit all online cpus and make sum. So, for now, unnecessary
560 * synchronization is not implemented. (just implemented for cpu hotplug)
562 * If there are kernel internal actions which can make use of some not-exact
563 * value, and reading all cpu value can be performance bottleneck in some
564 * common workload, threshold and synchronization as vmstat[] should be
568 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
573 /* Per-cpu values can be negative, use a signed accumulator */
574 for_each_possible_cpu(cpu)
575 val += per_cpu(memcg->stat->count[idx], cpu);
577 * Summing races with updates, so val may be negative. Avoid exposing
578 * transient negative values.
585 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
586 enum mem_cgroup_events_index idx)
588 unsigned long val = 0;
591 for_each_possible_cpu(cpu)
592 val += per_cpu(memcg->stat->events[idx], cpu);
596 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
598 bool compound, int nr_pages)
601 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
602 * counted as CACHE even if it's on ANON LRU.
605 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
608 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
612 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
613 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
617 /* pagein of a big page is an event. So, ignore page size */
619 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
621 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
622 nr_pages = -nr_pages; /* for event */
625 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
628 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
629 int nid, unsigned int lru_mask)
631 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
632 unsigned long nr = 0;
635 VM_BUG_ON((unsigned)nid >= nr_node_ids);
638 if (!(BIT(lru) & lru_mask))
640 nr += mem_cgroup_get_lru_size(lruvec, lru);
645 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
646 unsigned int lru_mask)
648 unsigned long nr = 0;
651 for_each_node_state(nid, N_MEMORY)
652 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
656 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
657 enum mem_cgroup_events_target target)
659 unsigned long val, next;
661 val = __this_cpu_read(memcg->stat->nr_page_events);
662 next = __this_cpu_read(memcg->stat->targets[target]);
663 /* from time_after() in jiffies.h */
664 if ((long)next - (long)val < 0) {
666 case MEM_CGROUP_TARGET_THRESH:
667 next = val + THRESHOLDS_EVENTS_TARGET;
669 case MEM_CGROUP_TARGET_SOFTLIMIT:
670 next = val + SOFTLIMIT_EVENTS_TARGET;
672 case MEM_CGROUP_TARGET_NUMAINFO:
673 next = val + NUMAINFO_EVENTS_TARGET;
678 __this_cpu_write(memcg->stat->targets[target], next);
685 * Check events in order.
688 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
690 /* threshold event is triggered in finer grain than soft limit */
691 if (unlikely(mem_cgroup_event_ratelimit(memcg,
692 MEM_CGROUP_TARGET_THRESH))) {
694 bool do_numainfo __maybe_unused;
696 do_softlimit = mem_cgroup_event_ratelimit(memcg,
697 MEM_CGROUP_TARGET_SOFTLIMIT);
699 do_numainfo = mem_cgroup_event_ratelimit(memcg,
700 MEM_CGROUP_TARGET_NUMAINFO);
702 mem_cgroup_threshold(memcg);
703 if (unlikely(do_softlimit))
704 mem_cgroup_update_tree(memcg, page);
706 if (unlikely(do_numainfo))
707 atomic_inc(&memcg->numainfo_events);
712 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
715 * mm_update_next_owner() may clear mm->owner to NULL
716 * if it races with swapoff, page migration, etc.
717 * So this can be called with p == NULL.
722 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
724 EXPORT_SYMBOL(mem_cgroup_from_task);
726 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
728 struct mem_cgroup *memcg = NULL;
733 * Page cache insertions can happen withou an
734 * actual mm context, e.g. during disk probing
735 * on boot, loopback IO, acct() writes etc.
738 memcg = root_mem_cgroup;
740 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
741 if (unlikely(!memcg))
742 memcg = root_mem_cgroup;
744 } while (!css_tryget(&memcg->css));
750 * mem_cgroup_iter - iterate over memory cgroup hierarchy
751 * @root: hierarchy root
752 * @prev: previously returned memcg, NULL on first invocation
753 * @reclaim: cookie for shared reclaim walks, NULL for full walks
755 * Returns references to children of the hierarchy below @root, or
756 * @root itself, or %NULL after a full round-trip.
758 * Caller must pass the return value in @prev on subsequent
759 * invocations for reference counting, or use mem_cgroup_iter_break()
760 * to cancel a hierarchy walk before the round-trip is complete.
762 * Reclaimers can specify a zone and a priority level in @reclaim to
763 * divide up the memcgs in the hierarchy among all concurrent
764 * reclaimers operating on the same zone and priority.
766 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
767 struct mem_cgroup *prev,
768 struct mem_cgroup_reclaim_cookie *reclaim)
770 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
771 struct cgroup_subsys_state *css = NULL;
772 struct mem_cgroup *memcg = NULL;
773 struct mem_cgroup *pos = NULL;
775 if (mem_cgroup_disabled())
779 root = root_mem_cgroup;
781 if (prev && !reclaim)
784 if (!root->use_hierarchy && root != root_mem_cgroup) {
793 struct mem_cgroup_per_node *mz;
795 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
796 iter = &mz->iter[reclaim->priority];
798 if (prev && reclaim->generation != iter->generation)
802 pos = READ_ONCE(iter->position);
803 if (!pos || css_tryget(&pos->css))
806 * css reference reached zero, so iter->position will
807 * be cleared by ->css_released. However, we should not
808 * rely on this happening soon, because ->css_released
809 * is called from a work queue, and by busy-waiting we
810 * might block it. So we clear iter->position right
813 (void)cmpxchg(&iter->position, pos, NULL);
821 css = css_next_descendant_pre(css, &root->css);
824 * Reclaimers share the hierarchy walk, and a
825 * new one might jump in right at the end of
826 * the hierarchy - make sure they see at least
827 * one group and restart from the beginning.
835 * Verify the css and acquire a reference. The root
836 * is provided by the caller, so we know it's alive
837 * and kicking, and don't take an extra reference.
839 memcg = mem_cgroup_from_css(css);
841 if (css == &root->css)
852 * The position could have already been updated by a competing
853 * thread, so check that the value hasn't changed since we read
854 * it to avoid reclaiming from the same cgroup twice.
856 (void)cmpxchg(&iter->position, pos, memcg);
864 reclaim->generation = iter->generation;
870 if (prev && prev != root)
877 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
878 * @root: hierarchy root
879 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
881 void mem_cgroup_iter_break(struct mem_cgroup *root,
882 struct mem_cgroup *prev)
885 root = root_mem_cgroup;
886 if (prev && prev != root)
890 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
891 struct mem_cgroup *dead_memcg)
893 struct mem_cgroup_reclaim_iter *iter;
894 struct mem_cgroup_per_node *mz;
899 mz = mem_cgroup_nodeinfo(from, nid);
900 for (i = 0; i <= DEF_PRIORITY; i++) {
902 cmpxchg(&iter->position,
908 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
910 struct mem_cgroup *memcg = dead_memcg;
911 struct mem_cgroup *last;
914 __invalidate_reclaim_iterators(memcg, dead_memcg);
916 } while ((memcg = parent_mem_cgroup(memcg)));
919 * When cgruop1 non-hierarchy mode is used,
920 * parent_mem_cgroup() does not walk all the way up to the
921 * cgroup root (root_mem_cgroup). So we have to handle
922 * dead_memcg from cgroup root separately.
924 if (last != root_mem_cgroup)
925 __invalidate_reclaim_iterators(root_mem_cgroup,
930 * Iteration constructs for visiting all cgroups (under a tree). If
931 * loops are exited prematurely (break), mem_cgroup_iter_break() must
932 * be used for reference counting.
934 #define for_each_mem_cgroup_tree(iter, root) \
935 for (iter = mem_cgroup_iter(root, NULL, NULL); \
937 iter = mem_cgroup_iter(root, iter, NULL))
939 #define for_each_mem_cgroup(iter) \
940 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
942 iter = mem_cgroup_iter(NULL, iter, NULL))
945 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
946 * @memcg: hierarchy root
947 * @fn: function to call for each task
948 * @arg: argument passed to @fn
950 * This function iterates over tasks attached to @memcg or to any of its
951 * descendants and calls @fn for each task. If @fn returns a non-zero
952 * value, the function breaks the iteration loop and returns the value.
953 * Otherwise, it will iterate over all tasks and return 0.
955 * This function must not be called for the root memory cgroup.
957 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
958 int (*fn)(struct task_struct *, void *), void *arg)
960 struct mem_cgroup *iter;
963 BUG_ON(memcg == root_mem_cgroup);
965 for_each_mem_cgroup_tree(iter, memcg) {
966 struct css_task_iter it;
967 struct task_struct *task;
969 css_task_iter_start(&iter->css, &it);
970 while (!ret && (task = css_task_iter_next(&it)))
972 css_task_iter_end(&it);
974 mem_cgroup_iter_break(memcg, iter);
982 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
984 * @zone: zone of the page
986 * This function is only safe when following the LRU page isolation
987 * and putback protocol: the LRU lock must be held, and the page must
988 * either be PageLRU() or the caller must have isolated/allocated it.
990 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
992 struct mem_cgroup_per_node *mz;
993 struct mem_cgroup *memcg;
994 struct lruvec *lruvec;
996 if (mem_cgroup_disabled()) {
997 lruvec = &pgdat->lruvec;
1001 memcg = page->mem_cgroup;
1003 * Swapcache readahead pages are added to the LRU - and
1004 * possibly migrated - before they are charged.
1007 memcg = root_mem_cgroup;
1009 mz = mem_cgroup_page_nodeinfo(memcg, page);
1010 lruvec = &mz->lruvec;
1013 * Since a node can be onlined after the mem_cgroup was created,
1014 * we have to be prepared to initialize lruvec->zone here;
1015 * and if offlined then reonlined, we need to reinitialize it.
1017 if (unlikely(lruvec->pgdat != pgdat))
1018 lruvec->pgdat = pgdat;
1023 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1024 * @lruvec: mem_cgroup per zone lru vector
1025 * @lru: index of lru list the page is sitting on
1026 * @zid: zone id of the accounted pages
1027 * @nr_pages: positive when adding or negative when removing
1029 * This function must be called under lru_lock, just before a page is added
1030 * to or just after a page is removed from an lru list (that ordering being
1031 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1033 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1034 int zid, int nr_pages)
1036 struct mem_cgroup_per_node *mz;
1037 unsigned long *lru_size;
1040 if (mem_cgroup_disabled())
1043 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1044 lru_size = &mz->lru_zone_size[zid][lru];
1047 *lru_size += nr_pages;
1050 if (WARN_ONCE(size < 0,
1051 "%s(%p, %d, %d): lru_size %ld\n",
1052 __func__, lruvec, lru, nr_pages, size)) {
1058 *lru_size += nr_pages;
1061 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1063 struct mem_cgroup *task_memcg;
1064 struct task_struct *p;
1067 p = find_lock_task_mm(task);
1069 task_memcg = get_mem_cgroup_from_mm(p->mm);
1073 * All threads may have already detached their mm's, but the oom
1074 * killer still needs to detect if they have already been oom
1075 * killed to prevent needlessly killing additional tasks.
1078 task_memcg = mem_cgroup_from_task(task);
1079 css_get(&task_memcg->css);
1082 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1083 css_put(&task_memcg->css);
1088 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1089 * @memcg: the memory cgroup
1091 * Returns the maximum amount of memory @mem can be charged with, in
1094 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1096 unsigned long margin = 0;
1097 unsigned long count;
1098 unsigned long limit;
1100 count = page_counter_read(&memcg->memory);
1101 limit = READ_ONCE(memcg->memory.limit);
1103 margin = limit - count;
1105 if (do_memsw_account()) {
1106 count = page_counter_read(&memcg->memsw);
1107 limit = READ_ONCE(memcg->memsw.limit);
1109 margin = min(margin, limit - count);
1118 * A routine for checking "mem" is under move_account() or not.
1120 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1121 * moving cgroups. This is for waiting at high-memory pressure
1124 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1126 struct mem_cgroup *from;
1127 struct mem_cgroup *to;
1130 * Unlike task_move routines, we access mc.to, mc.from not under
1131 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1133 spin_lock(&mc.lock);
1139 ret = mem_cgroup_is_descendant(from, memcg) ||
1140 mem_cgroup_is_descendant(to, memcg);
1142 spin_unlock(&mc.lock);
1146 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1148 if (mc.moving_task && current != mc.moving_task) {
1149 if (mem_cgroup_under_move(memcg)) {
1151 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1152 /* moving charge context might have finished. */
1155 finish_wait(&mc.waitq, &wait);
1162 #define K(x) ((x) << (PAGE_SHIFT-10))
1164 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1165 * @memcg: The memory cgroup that went over limit
1166 * @p: Task that is going to be killed
1168 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1171 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1173 struct mem_cgroup *iter;
1179 pr_info("Task in ");
1180 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1181 pr_cont(" killed as a result of limit of ");
1183 pr_info("Memory limit reached of cgroup ");
1186 pr_cont_cgroup_path(memcg->css.cgroup);
1191 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1192 K((u64)page_counter_read(&memcg->memory)),
1193 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1194 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1195 K((u64)page_counter_read(&memcg->memsw)),
1196 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1197 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1198 K((u64)page_counter_read(&memcg->kmem)),
1199 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1201 for_each_mem_cgroup_tree(iter, memcg) {
1202 pr_info("Memory cgroup stats for ");
1203 pr_cont_cgroup_path(iter->css.cgroup);
1206 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1207 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1209 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1210 K(mem_cgroup_read_stat(iter, i)));
1213 for (i = 0; i < NR_LRU_LISTS; i++)
1214 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1215 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1222 * This function returns the number of memcg under hierarchy tree. Returns
1223 * 1(self count) if no children.
1225 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1228 struct mem_cgroup *iter;
1230 for_each_mem_cgroup_tree(iter, memcg)
1236 * Return the memory (and swap, if configured) limit for a memcg.
1238 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1240 unsigned long limit;
1242 limit = memcg->memory.limit;
1243 if (mem_cgroup_swappiness(memcg)) {
1244 unsigned long memsw_limit;
1245 unsigned long swap_limit;
1247 memsw_limit = memcg->memsw.limit;
1248 swap_limit = memcg->swap.limit;
1249 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1250 limit = min(limit + swap_limit, memsw_limit);
1255 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1258 struct oom_control oc = {
1262 .gfp_mask = gfp_mask,
1267 mutex_lock(&oom_lock);
1268 ret = out_of_memory(&oc);
1269 mutex_unlock(&oom_lock);
1273 #if MAX_NUMNODES > 1
1276 * test_mem_cgroup_node_reclaimable
1277 * @memcg: the target memcg
1278 * @nid: the node ID to be checked.
1279 * @noswap : specify true here if the user wants flle only information.
1281 * This function returns whether the specified memcg contains any
1282 * reclaimable pages on a node. Returns true if there are any reclaimable
1283 * pages in the node.
1285 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1286 int nid, bool noswap)
1288 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1290 if (noswap || !total_swap_pages)
1292 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1299 * Always updating the nodemask is not very good - even if we have an empty
1300 * list or the wrong list here, we can start from some node and traverse all
1301 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1304 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1308 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1309 * pagein/pageout changes since the last update.
1311 if (!atomic_read(&memcg->numainfo_events))
1313 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1316 /* make a nodemask where this memcg uses memory from */
1317 memcg->scan_nodes = node_states[N_MEMORY];
1319 for_each_node_mask(nid, node_states[N_MEMORY]) {
1321 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1322 node_clear(nid, memcg->scan_nodes);
1325 atomic_set(&memcg->numainfo_events, 0);
1326 atomic_set(&memcg->numainfo_updating, 0);
1330 * Selecting a node where we start reclaim from. Because what we need is just
1331 * reducing usage counter, start from anywhere is O,K. Considering
1332 * memory reclaim from current node, there are pros. and cons.
1334 * Freeing memory from current node means freeing memory from a node which
1335 * we'll use or we've used. So, it may make LRU bad. And if several threads
1336 * hit limits, it will see a contention on a node. But freeing from remote
1337 * node means more costs for memory reclaim because of memory latency.
1339 * Now, we use round-robin. Better algorithm is welcomed.
1341 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1345 mem_cgroup_may_update_nodemask(memcg);
1346 node = memcg->last_scanned_node;
1348 node = next_node_in(node, memcg->scan_nodes);
1350 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1351 * last time it really checked all the LRUs due to rate limiting.
1352 * Fallback to the current node in that case for simplicity.
1354 if (unlikely(node == MAX_NUMNODES))
1355 node = numa_node_id();
1357 memcg->last_scanned_node = node;
1361 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1367 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1370 unsigned long *total_scanned)
1372 struct mem_cgroup *victim = NULL;
1375 unsigned long excess;
1376 unsigned long nr_scanned;
1377 struct mem_cgroup_reclaim_cookie reclaim = {
1382 excess = soft_limit_excess(root_memcg);
1385 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1390 * If we have not been able to reclaim
1391 * anything, it might because there are
1392 * no reclaimable pages under this hierarchy
1397 * We want to do more targeted reclaim.
1398 * excess >> 2 is not to excessive so as to
1399 * reclaim too much, nor too less that we keep
1400 * coming back to reclaim from this cgroup
1402 if (total >= (excess >> 2) ||
1403 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1408 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1409 pgdat, &nr_scanned);
1410 *total_scanned += nr_scanned;
1411 if (!soft_limit_excess(root_memcg))
1414 mem_cgroup_iter_break(root_memcg, victim);
1418 #ifdef CONFIG_LOCKDEP
1419 static struct lockdep_map memcg_oom_lock_dep_map = {
1420 .name = "memcg_oom_lock",
1424 static DEFINE_SPINLOCK(memcg_oom_lock);
1427 * Check OOM-Killer is already running under our hierarchy.
1428 * If someone is running, return false.
1430 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1432 struct mem_cgroup *iter, *failed = NULL;
1434 spin_lock(&memcg_oom_lock);
1436 for_each_mem_cgroup_tree(iter, memcg) {
1437 if (iter->oom_lock) {
1439 * this subtree of our hierarchy is already locked
1440 * so we cannot give a lock.
1443 mem_cgroup_iter_break(memcg, iter);
1446 iter->oom_lock = true;
1451 * OK, we failed to lock the whole subtree so we have
1452 * to clean up what we set up to the failing subtree
1454 for_each_mem_cgroup_tree(iter, memcg) {
1455 if (iter == failed) {
1456 mem_cgroup_iter_break(memcg, iter);
1459 iter->oom_lock = false;
1462 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1464 spin_unlock(&memcg_oom_lock);
1469 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1471 struct mem_cgroup *iter;
1473 spin_lock(&memcg_oom_lock);
1474 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1475 for_each_mem_cgroup_tree(iter, memcg)
1476 iter->oom_lock = false;
1477 spin_unlock(&memcg_oom_lock);
1480 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1482 struct mem_cgroup *iter;
1484 spin_lock(&memcg_oom_lock);
1485 for_each_mem_cgroup_tree(iter, memcg)
1487 spin_unlock(&memcg_oom_lock);
1490 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1492 struct mem_cgroup *iter;
1495 * When a new child is created while the hierarchy is under oom,
1496 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1498 spin_lock(&memcg_oom_lock);
1499 for_each_mem_cgroup_tree(iter, memcg)
1500 if (iter->under_oom > 0)
1502 spin_unlock(&memcg_oom_lock);
1505 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1507 struct oom_wait_info {
1508 struct mem_cgroup *memcg;
1512 static int memcg_oom_wake_function(wait_queue_t *wait,
1513 unsigned mode, int sync, void *arg)
1515 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1516 struct mem_cgroup *oom_wait_memcg;
1517 struct oom_wait_info *oom_wait_info;
1519 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1520 oom_wait_memcg = oom_wait_info->memcg;
1522 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1523 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1525 return autoremove_wake_function(wait, mode, sync, arg);
1528 static void memcg_oom_recover(struct mem_cgroup *memcg)
1531 * For the following lockless ->under_oom test, the only required
1532 * guarantee is that it must see the state asserted by an OOM when
1533 * this function is called as a result of userland actions
1534 * triggered by the notification of the OOM. This is trivially
1535 * achieved by invoking mem_cgroup_mark_under_oom() before
1536 * triggering notification.
1538 if (memcg && memcg->under_oom)
1539 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1542 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1544 if (!current->memcg_may_oom)
1547 * We are in the middle of the charge context here, so we
1548 * don't want to block when potentially sitting on a callstack
1549 * that holds all kinds of filesystem and mm locks.
1551 * Also, the caller may handle a failed allocation gracefully
1552 * (like optional page cache readahead) and so an OOM killer
1553 * invocation might not even be necessary.
1555 * That's why we don't do anything here except remember the
1556 * OOM context and then deal with it at the end of the page
1557 * fault when the stack is unwound, the locks are released,
1558 * and when we know whether the fault was overall successful.
1560 css_get(&memcg->css);
1561 current->memcg_in_oom = memcg;
1562 current->memcg_oom_gfp_mask = mask;
1563 current->memcg_oom_order = order;
1567 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1568 * @handle: actually kill/wait or just clean up the OOM state
1570 * This has to be called at the end of a page fault if the memcg OOM
1571 * handler was enabled.
1573 * Memcg supports userspace OOM handling where failed allocations must
1574 * sleep on a waitqueue until the userspace task resolves the
1575 * situation. Sleeping directly in the charge context with all kinds
1576 * of locks held is not a good idea, instead we remember an OOM state
1577 * in the task and mem_cgroup_oom_synchronize() has to be called at
1578 * the end of the page fault to complete the OOM handling.
1580 * Returns %true if an ongoing memcg OOM situation was detected and
1581 * completed, %false otherwise.
1583 bool mem_cgroup_oom_synchronize(bool handle)
1585 struct mem_cgroup *memcg = current->memcg_in_oom;
1586 struct oom_wait_info owait;
1589 /* OOM is global, do not handle */
1596 owait.memcg = memcg;
1597 owait.wait.flags = 0;
1598 owait.wait.func = memcg_oom_wake_function;
1599 owait.wait.private = current;
1600 INIT_LIST_HEAD(&owait.wait.task_list);
1602 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1603 mem_cgroup_mark_under_oom(memcg);
1605 locked = mem_cgroup_oom_trylock(memcg);
1608 mem_cgroup_oom_notify(memcg);
1610 if (locked && !memcg->oom_kill_disable) {
1611 mem_cgroup_unmark_under_oom(memcg);
1612 finish_wait(&memcg_oom_waitq, &owait.wait);
1613 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1614 current->memcg_oom_order);
1617 mem_cgroup_unmark_under_oom(memcg);
1618 finish_wait(&memcg_oom_waitq, &owait.wait);
1622 mem_cgroup_oom_unlock(memcg);
1624 * There is no guarantee that an OOM-lock contender
1625 * sees the wakeups triggered by the OOM kill
1626 * uncharges. Wake any sleepers explicitely.
1628 memcg_oom_recover(memcg);
1631 current->memcg_in_oom = NULL;
1632 css_put(&memcg->css);
1637 * lock_page_memcg - lock a page->mem_cgroup binding
1640 * This function protects unlocked LRU pages from being moved to
1643 * It ensures lifetime of the returned memcg. Caller is responsible
1644 * for the lifetime of the page; __unlock_page_memcg() is available
1645 * when @page might get freed inside the locked section.
1647 struct mem_cgroup *lock_page_memcg(struct page *page)
1649 struct mem_cgroup *memcg;
1650 unsigned long flags;
1653 * The RCU lock is held throughout the transaction. The fast
1654 * path can get away without acquiring the memcg->move_lock
1655 * because page moving starts with an RCU grace period.
1657 * The RCU lock also protects the memcg from being freed when
1658 * the page state that is going to change is the only thing
1659 * preventing the page itself from being freed. E.g. writeback
1660 * doesn't hold a page reference and relies on PG_writeback to
1661 * keep off truncation, migration and so forth.
1665 if (mem_cgroup_disabled())
1668 memcg = page->mem_cgroup;
1669 if (unlikely(!memcg))
1672 if (atomic_read(&memcg->moving_account) <= 0)
1675 spin_lock_irqsave(&memcg->move_lock, flags);
1676 if (memcg != page->mem_cgroup) {
1677 spin_unlock_irqrestore(&memcg->move_lock, flags);
1682 * When charge migration first begins, we can have locked and
1683 * unlocked page stat updates happening concurrently. Track
1684 * the task who has the lock for unlock_page_memcg().
1686 memcg->move_lock_task = current;
1687 memcg->move_lock_flags = flags;
1691 EXPORT_SYMBOL(lock_page_memcg);
1694 * __unlock_page_memcg - unlock and unpin a memcg
1697 * Unlock and unpin a memcg returned by lock_page_memcg().
1699 void __unlock_page_memcg(struct mem_cgroup *memcg)
1701 if (memcg && memcg->move_lock_task == current) {
1702 unsigned long flags = memcg->move_lock_flags;
1704 memcg->move_lock_task = NULL;
1705 memcg->move_lock_flags = 0;
1707 spin_unlock_irqrestore(&memcg->move_lock, flags);
1714 * unlock_page_memcg - unlock a page->mem_cgroup binding
1717 void unlock_page_memcg(struct page *page)
1719 __unlock_page_memcg(page->mem_cgroup);
1721 EXPORT_SYMBOL(unlock_page_memcg);
1724 * size of first charge trial. "32" comes from vmscan.c's magic value.
1725 * TODO: maybe necessary to use big numbers in big irons.
1727 #define CHARGE_BATCH 32U
1728 struct memcg_stock_pcp {
1729 struct mem_cgroup *cached; /* this never be root cgroup */
1730 unsigned int nr_pages;
1731 struct work_struct work;
1732 unsigned long flags;
1733 #define FLUSHING_CACHED_CHARGE 0
1735 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1736 static DEFINE_MUTEX(percpu_charge_mutex);
1739 * consume_stock: Try to consume stocked charge on this cpu.
1740 * @memcg: memcg to consume from.
1741 * @nr_pages: how many pages to charge.
1743 * The charges will only happen if @memcg matches the current cpu's memcg
1744 * stock, and at least @nr_pages are available in that stock. Failure to
1745 * service an allocation will refill the stock.
1747 * returns true if successful, false otherwise.
1749 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1751 struct memcg_stock_pcp *stock;
1752 unsigned long flags;
1755 if (nr_pages > CHARGE_BATCH)
1758 local_irq_save(flags);
1760 stock = this_cpu_ptr(&memcg_stock);
1761 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1762 stock->nr_pages -= nr_pages;
1766 local_irq_restore(flags);
1772 * Returns stocks cached in percpu and reset cached information.
1774 static void drain_stock(struct memcg_stock_pcp *stock)
1776 struct mem_cgroup *old = stock->cached;
1778 if (stock->nr_pages) {
1779 page_counter_uncharge(&old->memory, stock->nr_pages);
1780 if (do_memsw_account())
1781 page_counter_uncharge(&old->memsw, stock->nr_pages);
1782 css_put_many(&old->css, stock->nr_pages);
1783 stock->nr_pages = 0;
1785 stock->cached = NULL;
1788 static void drain_local_stock(struct work_struct *dummy)
1790 struct memcg_stock_pcp *stock;
1791 unsigned long flags;
1793 local_irq_save(flags);
1795 stock = this_cpu_ptr(&memcg_stock);
1797 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1799 local_irq_restore(flags);
1803 * Cache charges(val) to local per_cpu area.
1804 * This will be consumed by consume_stock() function, later.
1806 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1808 struct memcg_stock_pcp *stock;
1809 unsigned long flags;
1811 local_irq_save(flags);
1813 stock = this_cpu_ptr(&memcg_stock);
1814 if (stock->cached != memcg) { /* reset if necessary */
1816 stock->cached = memcg;
1818 stock->nr_pages += nr_pages;
1820 local_irq_restore(flags);
1824 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1825 * of the hierarchy under it.
1827 static void drain_all_stock(struct mem_cgroup *root_memcg)
1831 /* If someone's already draining, avoid adding running more workers. */
1832 if (!mutex_trylock(&percpu_charge_mutex))
1834 /* Notify other cpus that system-wide "drain" is running */
1837 for_each_online_cpu(cpu) {
1838 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1839 struct mem_cgroup *memcg;
1841 memcg = stock->cached;
1842 if (!memcg || !stock->nr_pages)
1844 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1846 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1848 drain_local_stock(&stock->work);
1850 schedule_work_on(cpu, &stock->work);
1855 mutex_unlock(&percpu_charge_mutex);
1858 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1859 unsigned long action,
1862 int cpu = (unsigned long)hcpu;
1863 struct memcg_stock_pcp *stock;
1865 if (action == CPU_ONLINE)
1868 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1871 stock = &per_cpu(memcg_stock, cpu);
1876 static void reclaim_high(struct mem_cgroup *memcg,
1877 unsigned int nr_pages,
1881 if (page_counter_read(&memcg->memory) <= memcg->high)
1883 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1884 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1885 } while ((memcg = parent_mem_cgroup(memcg)));
1888 static void high_work_func(struct work_struct *work)
1890 struct mem_cgroup *memcg;
1892 memcg = container_of(work, struct mem_cgroup, high_work);
1893 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1897 * Scheduled by try_charge() to be executed from the userland return path
1898 * and reclaims memory over the high limit.
1900 void mem_cgroup_handle_over_high(void)
1902 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1903 struct mem_cgroup *memcg;
1905 if (likely(!nr_pages))
1908 memcg = get_mem_cgroup_from_mm(current->mm);
1909 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1910 css_put(&memcg->css);
1911 current->memcg_nr_pages_over_high = 0;
1914 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1915 unsigned int nr_pages)
1917 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1918 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1919 struct mem_cgroup *mem_over_limit;
1920 struct page_counter *counter;
1921 unsigned long nr_reclaimed;
1922 bool may_swap = true;
1923 bool drained = false;
1925 if (mem_cgroup_is_root(memcg))
1928 if (consume_stock(memcg, nr_pages))
1931 if (!do_memsw_account() ||
1932 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1933 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1935 if (do_memsw_account())
1936 page_counter_uncharge(&memcg->memsw, batch);
1937 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1939 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1943 if (batch > nr_pages) {
1949 * Unlike in global OOM situations, memcg is not in a physical
1950 * memory shortage. Allow dying and OOM-killed tasks to
1951 * bypass the last charges so that they can exit quickly and
1952 * free their memory.
1954 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1955 fatal_signal_pending(current) ||
1956 current->flags & PF_EXITING))
1960 * Prevent unbounded recursion when reclaim operations need to
1961 * allocate memory. This might exceed the limits temporarily,
1962 * but we prefer facilitating memory reclaim and getting back
1963 * under the limit over triggering OOM kills in these cases.
1965 if (unlikely(current->flags & PF_MEMALLOC))
1968 if (unlikely(task_in_memcg_oom(current)))
1971 if (!gfpflags_allow_blocking(gfp_mask))
1974 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1976 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1977 gfp_mask, may_swap);
1979 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1983 drain_all_stock(mem_over_limit);
1988 if (gfp_mask & __GFP_NORETRY)
1991 * Even though the limit is exceeded at this point, reclaim
1992 * may have been able to free some pages. Retry the charge
1993 * before killing the task.
1995 * Only for regular pages, though: huge pages are rather
1996 * unlikely to succeed so close to the limit, and we fall back
1997 * to regular pages anyway in case of failure.
1999 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2002 * At task move, charge accounts can be doubly counted. So, it's
2003 * better to wait until the end of task_move if something is going on.
2005 if (mem_cgroup_wait_acct_move(mem_over_limit))
2011 if (gfp_mask & __GFP_NOFAIL)
2014 if (fatal_signal_pending(current))
2017 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2019 mem_cgroup_oom(mem_over_limit, gfp_mask,
2020 get_order(nr_pages * PAGE_SIZE));
2022 if (!(gfp_mask & __GFP_NOFAIL))
2026 * The allocation either can't fail or will lead to more memory
2027 * being freed very soon. Allow memory usage go over the limit
2028 * temporarily by force charging it.
2030 page_counter_charge(&memcg->memory, nr_pages);
2031 if (do_memsw_account())
2032 page_counter_charge(&memcg->memsw, nr_pages);
2033 css_get_many(&memcg->css, nr_pages);
2038 css_get_many(&memcg->css, batch);
2039 if (batch > nr_pages)
2040 refill_stock(memcg, batch - nr_pages);
2043 * If the hierarchy is above the normal consumption range, schedule
2044 * reclaim on returning to userland. We can perform reclaim here
2045 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2046 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2047 * not recorded as it most likely matches current's and won't
2048 * change in the meantime. As high limit is checked again before
2049 * reclaim, the cost of mismatch is negligible.
2052 if (page_counter_read(&memcg->memory) > memcg->high) {
2053 /* Don't bother a random interrupted task */
2054 if (in_interrupt()) {
2055 schedule_work(&memcg->high_work);
2058 current->memcg_nr_pages_over_high += batch;
2059 set_notify_resume(current);
2062 } while ((memcg = parent_mem_cgroup(memcg)));
2067 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2069 if (mem_cgroup_is_root(memcg))
2072 page_counter_uncharge(&memcg->memory, nr_pages);
2073 if (do_memsw_account())
2074 page_counter_uncharge(&memcg->memsw, nr_pages);
2076 css_put_many(&memcg->css, nr_pages);
2079 static void lock_page_lru(struct page *page, int *isolated)
2081 struct zone *zone = page_zone(page);
2083 spin_lock_irq(zone_lru_lock(zone));
2084 if (PageLRU(page)) {
2085 struct lruvec *lruvec;
2087 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2089 del_page_from_lru_list(page, lruvec, page_lru(page));
2095 static void unlock_page_lru(struct page *page, int isolated)
2097 struct zone *zone = page_zone(page);
2100 struct lruvec *lruvec;
2102 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2103 VM_BUG_ON_PAGE(PageLRU(page), page);
2105 add_page_to_lru_list(page, lruvec, page_lru(page));
2107 spin_unlock_irq(zone_lru_lock(zone));
2110 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2115 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2118 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2119 * may already be on some other mem_cgroup's LRU. Take care of it.
2122 lock_page_lru(page, &isolated);
2125 * Nobody should be changing or seriously looking at
2126 * page->mem_cgroup at this point:
2128 * - the page is uncharged
2130 * - the page is off-LRU
2132 * - an anonymous fault has exclusive page access, except for
2133 * a locked page table
2135 * - a page cache insertion, a swapin fault, or a migration
2136 * have the page locked
2138 page->mem_cgroup = memcg;
2141 unlock_page_lru(page, isolated);
2145 static int memcg_alloc_cache_id(void)
2150 id = ida_simple_get(&memcg_cache_ida,
2151 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2155 if (id < memcg_nr_cache_ids)
2159 * There's no space for the new id in memcg_caches arrays,
2160 * so we have to grow them.
2162 down_write(&memcg_cache_ids_sem);
2164 size = 2 * (id + 1);
2165 if (size < MEMCG_CACHES_MIN_SIZE)
2166 size = MEMCG_CACHES_MIN_SIZE;
2167 else if (size > MEMCG_CACHES_MAX_SIZE)
2168 size = MEMCG_CACHES_MAX_SIZE;
2170 err = memcg_update_all_caches(size);
2172 err = memcg_update_all_list_lrus(size);
2174 memcg_nr_cache_ids = size;
2176 up_write(&memcg_cache_ids_sem);
2179 ida_simple_remove(&memcg_cache_ida, id);
2185 static void memcg_free_cache_id(int id)
2187 ida_simple_remove(&memcg_cache_ida, id);
2190 struct memcg_kmem_cache_create_work {
2191 struct mem_cgroup *memcg;
2192 struct kmem_cache *cachep;
2193 struct work_struct work;
2196 static struct workqueue_struct *memcg_kmem_cache_create_wq;
2198 static void memcg_kmem_cache_create_func(struct work_struct *w)
2200 struct memcg_kmem_cache_create_work *cw =
2201 container_of(w, struct memcg_kmem_cache_create_work, work);
2202 struct mem_cgroup *memcg = cw->memcg;
2203 struct kmem_cache *cachep = cw->cachep;
2205 memcg_create_kmem_cache(memcg, cachep);
2207 css_put(&memcg->css);
2212 * Enqueue the creation of a per-memcg kmem_cache.
2214 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2215 struct kmem_cache *cachep)
2217 struct memcg_kmem_cache_create_work *cw;
2219 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2223 css_get(&memcg->css);
2226 cw->cachep = cachep;
2227 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2229 queue_work(memcg_kmem_cache_create_wq, &cw->work);
2232 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2233 struct kmem_cache *cachep)
2236 * We need to stop accounting when we kmalloc, because if the
2237 * corresponding kmalloc cache is not yet created, the first allocation
2238 * in __memcg_schedule_kmem_cache_create will recurse.
2240 * However, it is better to enclose the whole function. Depending on
2241 * the debugging options enabled, INIT_WORK(), for instance, can
2242 * trigger an allocation. This too, will make us recurse. Because at
2243 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2244 * the safest choice is to do it like this, wrapping the whole function.
2246 current->memcg_kmem_skip_account = 1;
2247 __memcg_schedule_kmem_cache_create(memcg, cachep);
2248 current->memcg_kmem_skip_account = 0;
2251 static inline bool memcg_kmem_bypass(void)
2253 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2259 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2260 * @cachep: the original global kmem cache
2262 * Return the kmem_cache we're supposed to use for a slab allocation.
2263 * We try to use the current memcg's version of the cache.
2265 * If the cache does not exist yet, if we are the first user of it, we
2266 * create it asynchronously in a workqueue and let the current allocation
2267 * go through with the original cache.
2269 * This function takes a reference to the cache it returns to assure it
2270 * won't get destroyed while we are working with it. Once the caller is
2271 * done with it, memcg_kmem_put_cache() must be called to release the
2274 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2276 struct mem_cgroup *memcg;
2277 struct kmem_cache *memcg_cachep;
2280 VM_BUG_ON(!is_root_cache(cachep));
2282 if (memcg_kmem_bypass())
2285 if (current->memcg_kmem_skip_account)
2288 memcg = get_mem_cgroup_from_mm(current->mm);
2289 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2293 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2294 if (likely(memcg_cachep))
2295 return memcg_cachep;
2298 * If we are in a safe context (can wait, and not in interrupt
2299 * context), we could be be predictable and return right away.
2300 * This would guarantee that the allocation being performed
2301 * already belongs in the new cache.
2303 * However, there are some clashes that can arrive from locking.
2304 * For instance, because we acquire the slab_mutex while doing
2305 * memcg_create_kmem_cache, this means no further allocation
2306 * could happen with the slab_mutex held. So it's better to
2309 memcg_schedule_kmem_cache_create(memcg, cachep);
2311 css_put(&memcg->css);
2316 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2317 * @cachep: the cache returned by memcg_kmem_get_cache
2319 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2321 if (!is_root_cache(cachep))
2322 css_put(&cachep->memcg_params.memcg->css);
2326 * memcg_kmem_charge: charge a kmem page
2327 * @page: page to charge
2328 * @gfp: reclaim mode
2329 * @order: allocation order
2330 * @memcg: memory cgroup to charge
2332 * Returns 0 on success, an error code on failure.
2334 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2335 struct mem_cgroup *memcg)
2337 unsigned int nr_pages = 1 << order;
2338 struct page_counter *counter;
2341 ret = try_charge(memcg, gfp, nr_pages);
2345 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2346 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2349 * Enforce __GFP_NOFAIL allocation because callers are not
2350 * prepared to see failures and likely do not have any failure
2353 if (gfp & __GFP_NOFAIL) {
2354 page_counter_charge(&memcg->kmem, nr_pages);
2357 cancel_charge(memcg, nr_pages);
2361 page->mem_cgroup = memcg;
2367 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2368 * @page: page to charge
2369 * @gfp: reclaim mode
2370 * @order: allocation order
2372 * Returns 0 on success, an error code on failure.
2374 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2376 struct mem_cgroup *memcg;
2379 if (memcg_kmem_bypass())
2382 memcg = get_mem_cgroup_from_mm(current->mm);
2383 if (!mem_cgroup_is_root(memcg)) {
2384 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2386 __SetPageKmemcg(page);
2388 css_put(&memcg->css);
2392 * memcg_kmem_uncharge: uncharge a kmem page
2393 * @page: page to uncharge
2394 * @order: allocation order
2396 void memcg_kmem_uncharge(struct page *page, int order)
2398 struct mem_cgroup *memcg = page->mem_cgroup;
2399 unsigned int nr_pages = 1 << order;
2404 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2406 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2407 page_counter_uncharge(&memcg->kmem, nr_pages);
2409 page_counter_uncharge(&memcg->memory, nr_pages);
2410 if (do_memsw_account())
2411 page_counter_uncharge(&memcg->memsw, nr_pages);
2413 page->mem_cgroup = NULL;
2415 /* slab pages do not have PageKmemcg flag set */
2416 if (PageKmemcg(page))
2417 __ClearPageKmemcg(page);
2419 css_put_many(&memcg->css, nr_pages);
2421 #endif /* !CONFIG_SLOB */
2423 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2426 * Because tail pages are not marked as "used", set it. We're under
2427 * zone_lru_lock and migration entries setup in all page mappings.
2429 void mem_cgroup_split_huge_fixup(struct page *head)
2433 if (mem_cgroup_disabled())
2436 for (i = 1; i < HPAGE_PMD_NR; i++)
2437 head[i].mem_cgroup = head->mem_cgroup;
2439 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2442 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2444 #ifdef CONFIG_MEMCG_SWAP
2445 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2448 int val = (charge) ? 1 : -1;
2449 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2453 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2454 * @entry: swap entry to be moved
2455 * @from: mem_cgroup which the entry is moved from
2456 * @to: mem_cgroup which the entry is moved to
2458 * It succeeds only when the swap_cgroup's record for this entry is the same
2459 * as the mem_cgroup's id of @from.
2461 * Returns 0 on success, -EINVAL on failure.
2463 * The caller must have charged to @to, IOW, called page_counter_charge() about
2464 * both res and memsw, and called css_get().
2466 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2467 struct mem_cgroup *from, struct mem_cgroup *to)
2469 unsigned short old_id, new_id;
2471 old_id = mem_cgroup_id(from);
2472 new_id = mem_cgroup_id(to);
2474 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2475 mem_cgroup_swap_statistics(from, false);
2476 mem_cgroup_swap_statistics(to, true);
2482 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2483 struct mem_cgroup *from, struct mem_cgroup *to)
2489 static DEFINE_MUTEX(memcg_limit_mutex);
2491 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2492 unsigned long limit)
2494 unsigned long curusage;
2495 unsigned long oldusage;
2496 bool enlarge = false;
2501 * For keeping hierarchical_reclaim simple, how long we should retry
2502 * is depends on callers. We set our retry-count to be function
2503 * of # of children which we should visit in this loop.
2505 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2506 mem_cgroup_count_children(memcg);
2508 oldusage = page_counter_read(&memcg->memory);
2511 if (signal_pending(current)) {
2516 mutex_lock(&memcg_limit_mutex);
2517 if (limit > memcg->memsw.limit) {
2518 mutex_unlock(&memcg_limit_mutex);
2522 if (limit > memcg->memory.limit)
2524 ret = page_counter_limit(&memcg->memory, limit);
2525 mutex_unlock(&memcg_limit_mutex);
2530 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2532 curusage = page_counter_read(&memcg->memory);
2533 /* Usage is reduced ? */
2534 if (curusage >= oldusage)
2537 oldusage = curusage;
2538 } while (retry_count);
2540 if (!ret && enlarge)
2541 memcg_oom_recover(memcg);
2546 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2547 unsigned long limit)
2549 unsigned long curusage;
2550 unsigned long oldusage;
2551 bool enlarge = false;
2555 /* see mem_cgroup_resize_res_limit */
2556 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2557 mem_cgroup_count_children(memcg);
2559 oldusage = page_counter_read(&memcg->memsw);
2562 if (signal_pending(current)) {
2567 mutex_lock(&memcg_limit_mutex);
2568 if (limit < memcg->memory.limit) {
2569 mutex_unlock(&memcg_limit_mutex);
2573 if (limit > memcg->memsw.limit)
2575 ret = page_counter_limit(&memcg->memsw, limit);
2576 mutex_unlock(&memcg_limit_mutex);
2581 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2583 curusage = page_counter_read(&memcg->memsw);
2584 /* Usage is reduced ? */
2585 if (curusage >= oldusage)
2588 oldusage = curusage;
2589 } while (retry_count);
2591 if (!ret && enlarge)
2592 memcg_oom_recover(memcg);
2597 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2599 unsigned long *total_scanned)
2601 unsigned long nr_reclaimed = 0;
2602 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2603 unsigned long reclaimed;
2605 struct mem_cgroup_tree_per_node *mctz;
2606 unsigned long excess;
2607 unsigned long nr_scanned;
2612 mctz = soft_limit_tree_node(pgdat->node_id);
2615 * Do not even bother to check the largest node if the root
2616 * is empty. Do it lockless to prevent lock bouncing. Races
2617 * are acceptable as soft limit is best effort anyway.
2619 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2623 * This loop can run a while, specially if mem_cgroup's continuously
2624 * keep exceeding their soft limit and putting the system under
2631 mz = mem_cgroup_largest_soft_limit_node(mctz);
2636 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2637 gfp_mask, &nr_scanned);
2638 nr_reclaimed += reclaimed;
2639 *total_scanned += nr_scanned;
2640 spin_lock_irq(&mctz->lock);
2641 __mem_cgroup_remove_exceeded(mz, mctz);
2644 * If we failed to reclaim anything from this memory cgroup
2645 * it is time to move on to the next cgroup
2649 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2651 excess = soft_limit_excess(mz->memcg);
2653 * One school of thought says that we should not add
2654 * back the node to the tree if reclaim returns 0.
2655 * But our reclaim could return 0, simply because due
2656 * to priority we are exposing a smaller subset of
2657 * memory to reclaim from. Consider this as a longer
2660 /* If excess == 0, no tree ops */
2661 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2662 spin_unlock_irq(&mctz->lock);
2663 css_put(&mz->memcg->css);
2666 * Could not reclaim anything and there are no more
2667 * mem cgroups to try or we seem to be looping without
2668 * reclaiming anything.
2670 if (!nr_reclaimed &&
2672 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2674 } while (!nr_reclaimed);
2676 css_put(&next_mz->memcg->css);
2677 return nr_reclaimed;
2681 * Test whether @memcg has children, dead or alive. Note that this
2682 * function doesn't care whether @memcg has use_hierarchy enabled and
2683 * returns %true if there are child csses according to the cgroup
2684 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2686 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2691 ret = css_next_child(NULL, &memcg->css);
2697 * Reclaims as many pages from the given memcg as possible.
2699 * Caller is responsible for holding css reference for memcg.
2701 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2703 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2705 /* we call try-to-free pages for make this cgroup empty */
2706 lru_add_drain_all();
2707 /* try to free all pages in this cgroup */
2708 while (nr_retries && page_counter_read(&memcg->memory)) {
2711 if (signal_pending(current))
2714 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2718 /* maybe some writeback is necessary */
2719 congestion_wait(BLK_RW_ASYNC, HZ/10);
2727 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2728 char *buf, size_t nbytes,
2731 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2733 if (mem_cgroup_is_root(memcg))
2735 return mem_cgroup_force_empty(memcg) ?: nbytes;
2738 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2741 return mem_cgroup_from_css(css)->use_hierarchy;
2744 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2745 struct cftype *cft, u64 val)
2748 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2749 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2751 if (memcg->use_hierarchy == val)
2755 * If parent's use_hierarchy is set, we can't make any modifications
2756 * in the child subtrees. If it is unset, then the change can
2757 * occur, provided the current cgroup has no children.
2759 * For the root cgroup, parent_mem is NULL, we allow value to be
2760 * set if there are no children.
2762 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2763 (val == 1 || val == 0)) {
2764 if (!memcg_has_children(memcg))
2765 memcg->use_hierarchy = val;
2774 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2776 struct mem_cgroup *iter;
2779 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2781 for_each_mem_cgroup_tree(iter, memcg) {
2782 for (i = 0; i < MEMCG_NR_STAT; i++)
2783 stat[i] += mem_cgroup_read_stat(iter, i);
2787 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2789 struct mem_cgroup *iter;
2792 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2794 for_each_mem_cgroup_tree(iter, memcg) {
2795 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2796 events[i] += mem_cgroup_read_events(iter, i);
2800 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2802 unsigned long val = 0;
2804 if (mem_cgroup_is_root(memcg)) {
2805 struct mem_cgroup *iter;
2807 for_each_mem_cgroup_tree(iter, memcg) {
2808 val += mem_cgroup_read_stat(iter,
2809 MEM_CGROUP_STAT_CACHE);
2810 val += mem_cgroup_read_stat(iter,
2811 MEM_CGROUP_STAT_RSS);
2813 val += mem_cgroup_read_stat(iter,
2814 MEM_CGROUP_STAT_SWAP);
2818 val = page_counter_read(&memcg->memory);
2820 val = page_counter_read(&memcg->memsw);
2833 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2836 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2837 struct page_counter *counter;
2839 switch (MEMFILE_TYPE(cft->private)) {
2841 counter = &memcg->memory;
2844 counter = &memcg->memsw;
2847 counter = &memcg->kmem;
2850 counter = &memcg->tcpmem;
2856 switch (MEMFILE_ATTR(cft->private)) {
2858 if (counter == &memcg->memory)
2859 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2860 if (counter == &memcg->memsw)
2861 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2862 return (u64)page_counter_read(counter) * PAGE_SIZE;
2864 return (u64)counter->limit * PAGE_SIZE;
2866 return (u64)counter->watermark * PAGE_SIZE;
2868 return counter->failcnt;
2869 case RES_SOFT_LIMIT:
2870 return (u64)memcg->soft_limit * PAGE_SIZE;
2877 static int memcg_online_kmem(struct mem_cgroup *memcg)
2881 if (cgroup_memory_nokmem)
2884 BUG_ON(memcg->kmemcg_id >= 0);
2885 BUG_ON(memcg->kmem_state);
2887 memcg_id = memcg_alloc_cache_id();
2891 static_branch_inc(&memcg_kmem_enabled_key);
2893 * A memory cgroup is considered kmem-online as soon as it gets
2894 * kmemcg_id. Setting the id after enabling static branching will
2895 * guarantee no one starts accounting before all call sites are
2898 memcg->kmemcg_id = memcg_id;
2899 memcg->kmem_state = KMEM_ONLINE;
2904 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2906 struct cgroup_subsys_state *css;
2907 struct mem_cgroup *parent, *child;
2910 if (memcg->kmem_state != KMEM_ONLINE)
2913 * Clear the online state before clearing memcg_caches array
2914 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2915 * guarantees that no cache will be created for this cgroup
2916 * after we are done (see memcg_create_kmem_cache()).
2918 memcg->kmem_state = KMEM_ALLOCATED;
2920 memcg_deactivate_kmem_caches(memcg);
2922 kmemcg_id = memcg->kmemcg_id;
2923 BUG_ON(kmemcg_id < 0);
2925 parent = parent_mem_cgroup(memcg);
2927 parent = root_mem_cgroup;
2930 * Change kmemcg_id of this cgroup and all its descendants to the
2931 * parent's id, and then move all entries from this cgroup's list_lrus
2932 * to ones of the parent. After we have finished, all list_lrus
2933 * corresponding to this cgroup are guaranteed to remain empty. The
2934 * ordering is imposed by list_lru_node->lock taken by
2935 * memcg_drain_all_list_lrus().
2937 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2938 css_for_each_descendant_pre(css, &memcg->css) {
2939 child = mem_cgroup_from_css(css);
2940 BUG_ON(child->kmemcg_id != kmemcg_id);
2941 child->kmemcg_id = parent->kmemcg_id;
2942 if (!memcg->use_hierarchy)
2947 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2949 memcg_free_cache_id(kmemcg_id);
2952 static void memcg_free_kmem(struct mem_cgroup *memcg)
2954 /* css_alloc() failed, offlining didn't happen */
2955 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2956 memcg_offline_kmem(memcg);
2958 if (memcg->kmem_state == KMEM_ALLOCATED) {
2959 memcg_destroy_kmem_caches(memcg);
2960 static_branch_dec(&memcg_kmem_enabled_key);
2961 WARN_ON(page_counter_read(&memcg->kmem));
2965 static int memcg_online_kmem(struct mem_cgroup *memcg)
2969 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2972 static void memcg_free_kmem(struct mem_cgroup *memcg)
2975 #endif /* !CONFIG_SLOB */
2977 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2978 unsigned long limit)
2982 mutex_lock(&memcg_limit_mutex);
2983 ret = page_counter_limit(&memcg->kmem, limit);
2984 mutex_unlock(&memcg_limit_mutex);
2988 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2992 mutex_lock(&memcg_limit_mutex);
2994 ret = page_counter_limit(&memcg->tcpmem, limit);
2998 if (!memcg->tcpmem_active) {
3000 * The active flag needs to be written after the static_key
3001 * update. This is what guarantees that the socket activation
3002 * function is the last one to run. See mem_cgroup_sk_alloc()
3003 * for details, and note that we don't mark any socket as
3004 * belonging to this memcg until that flag is up.
3006 * We need to do this, because static_keys will span multiple
3007 * sites, but we can't control their order. If we mark a socket
3008 * as accounted, but the accounting functions are not patched in
3009 * yet, we'll lose accounting.
3011 * We never race with the readers in mem_cgroup_sk_alloc(),
3012 * because when this value change, the code to process it is not
3015 static_branch_inc(&memcg_sockets_enabled_key);
3016 memcg->tcpmem_active = true;
3019 mutex_unlock(&memcg_limit_mutex);
3024 * The user of this function is...
3027 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3028 char *buf, size_t nbytes, loff_t off)
3030 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3031 unsigned long nr_pages;
3034 buf = strstrip(buf);
3035 ret = page_counter_memparse(buf, "-1", &nr_pages);
3039 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3041 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3045 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3047 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3050 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3053 ret = memcg_update_kmem_limit(memcg, nr_pages);
3056 ret = memcg_update_tcp_limit(memcg, nr_pages);
3060 case RES_SOFT_LIMIT:
3061 memcg->soft_limit = nr_pages;
3065 return ret ?: nbytes;
3068 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3069 size_t nbytes, loff_t off)
3071 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3072 struct page_counter *counter;
3074 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3076 counter = &memcg->memory;
3079 counter = &memcg->memsw;
3082 counter = &memcg->kmem;
3085 counter = &memcg->tcpmem;
3091 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3093 page_counter_reset_watermark(counter);
3096 counter->failcnt = 0;
3105 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3108 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3112 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3113 struct cftype *cft, u64 val)
3115 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3117 if (val & ~MOVE_MASK)
3121 * No kind of locking is needed in here, because ->can_attach() will
3122 * check this value once in the beginning of the process, and then carry
3123 * on with stale data. This means that changes to this value will only
3124 * affect task migrations starting after the change.
3126 memcg->move_charge_at_immigrate = val;
3130 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3131 struct cftype *cft, u64 val)
3138 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3142 unsigned int lru_mask;
3145 static const struct numa_stat stats[] = {
3146 { "total", LRU_ALL },
3147 { "file", LRU_ALL_FILE },
3148 { "anon", LRU_ALL_ANON },
3149 { "unevictable", BIT(LRU_UNEVICTABLE) },
3151 const struct numa_stat *stat;
3154 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3156 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3157 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3158 seq_printf(m, "%s=%lu", stat->name, nr);
3159 for_each_node_state(nid, N_MEMORY) {
3160 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3162 seq_printf(m, " N%d=%lu", nid, nr);
3167 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3168 struct mem_cgroup *iter;
3171 for_each_mem_cgroup_tree(iter, memcg)
3172 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3173 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3174 for_each_node_state(nid, N_MEMORY) {
3176 for_each_mem_cgroup_tree(iter, memcg)
3177 nr += mem_cgroup_node_nr_lru_pages(
3178 iter, nid, stat->lru_mask);
3179 seq_printf(m, " N%d=%lu", nid, nr);
3186 #endif /* CONFIG_NUMA */
3188 static int memcg_stat_show(struct seq_file *m, void *v)
3190 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3191 unsigned long memory, memsw;
3192 struct mem_cgroup *mi;
3195 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3196 MEM_CGROUP_STAT_NSTATS);
3197 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3198 MEM_CGROUP_EVENTS_NSTATS);
3199 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3201 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3202 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3204 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3205 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3208 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3209 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3210 mem_cgroup_read_events(memcg, i));
3212 for (i = 0; i < NR_LRU_LISTS; i++)
3213 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3214 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3216 /* Hierarchical information */
3217 memory = memsw = PAGE_COUNTER_MAX;
3218 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3219 memory = min(memory, mi->memory.limit);
3220 memsw = min(memsw, mi->memsw.limit);
3222 seq_printf(m, "hierarchical_memory_limit %llu\n",
3223 (u64)memory * PAGE_SIZE);
3224 if (do_memsw_account())
3225 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3226 (u64)memsw * PAGE_SIZE);
3228 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3229 unsigned long long val = 0;
3231 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3233 for_each_mem_cgroup_tree(mi, memcg)
3234 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3235 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3238 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3239 unsigned long long val = 0;
3241 for_each_mem_cgroup_tree(mi, memcg)
3242 val += mem_cgroup_read_events(mi, i);
3243 seq_printf(m, "total_%s %llu\n",
3244 mem_cgroup_events_names[i], val);
3247 for (i = 0; i < NR_LRU_LISTS; i++) {
3248 unsigned long long val = 0;
3250 for_each_mem_cgroup_tree(mi, memcg)
3251 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3252 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3255 #ifdef CONFIG_DEBUG_VM
3258 struct mem_cgroup_per_node *mz;
3259 struct zone_reclaim_stat *rstat;
3260 unsigned long recent_rotated[2] = {0, 0};
3261 unsigned long recent_scanned[2] = {0, 0};
3263 for_each_online_pgdat(pgdat) {
3264 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3265 rstat = &mz->lruvec.reclaim_stat;
3267 recent_rotated[0] += rstat->recent_rotated[0];
3268 recent_rotated[1] += rstat->recent_rotated[1];
3269 recent_scanned[0] += rstat->recent_scanned[0];
3270 recent_scanned[1] += rstat->recent_scanned[1];
3272 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3273 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3274 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3275 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3282 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3285 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3287 return mem_cgroup_swappiness(memcg);
3290 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3291 struct cftype *cft, u64 val)
3293 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3299 memcg->swappiness = val;
3301 vm_swappiness = val;
3306 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3308 struct mem_cgroup_threshold_ary *t;
3309 unsigned long usage;
3314 t = rcu_dereference(memcg->thresholds.primary);
3316 t = rcu_dereference(memcg->memsw_thresholds.primary);
3321 usage = mem_cgroup_usage(memcg, swap);
3324 * current_threshold points to threshold just below or equal to usage.
3325 * If it's not true, a threshold was crossed after last
3326 * call of __mem_cgroup_threshold().
3328 i = t->current_threshold;
3331 * Iterate backward over array of thresholds starting from
3332 * current_threshold and check if a threshold is crossed.
3333 * If none of thresholds below usage is crossed, we read
3334 * only one element of the array here.
3336 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3337 eventfd_signal(t->entries[i].eventfd, 1);
3339 /* i = current_threshold + 1 */
3343 * Iterate forward over array of thresholds starting from
3344 * current_threshold+1 and check if a threshold is crossed.
3345 * If none of thresholds above usage is crossed, we read
3346 * only one element of the array here.
3348 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3349 eventfd_signal(t->entries[i].eventfd, 1);
3351 /* Update current_threshold */
3352 t->current_threshold = i - 1;
3357 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3360 __mem_cgroup_threshold(memcg, false);
3361 if (do_memsw_account())
3362 __mem_cgroup_threshold(memcg, true);
3364 memcg = parent_mem_cgroup(memcg);
3368 static int compare_thresholds(const void *a, const void *b)
3370 const struct mem_cgroup_threshold *_a = a;
3371 const struct mem_cgroup_threshold *_b = b;
3373 if (_a->threshold > _b->threshold)
3376 if (_a->threshold < _b->threshold)
3382 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3384 struct mem_cgroup_eventfd_list *ev;
3386 spin_lock(&memcg_oom_lock);
3388 list_for_each_entry(ev, &memcg->oom_notify, list)
3389 eventfd_signal(ev->eventfd, 1);
3391 spin_unlock(&memcg_oom_lock);
3395 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3397 struct mem_cgroup *iter;
3399 for_each_mem_cgroup_tree(iter, memcg)
3400 mem_cgroup_oom_notify_cb(iter);
3403 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3404 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3406 struct mem_cgroup_thresholds *thresholds;
3407 struct mem_cgroup_threshold_ary *new;
3408 unsigned long threshold;
3409 unsigned long usage;
3412 ret = page_counter_memparse(args, "-1", &threshold);
3416 mutex_lock(&memcg->thresholds_lock);
3419 thresholds = &memcg->thresholds;
3420 usage = mem_cgroup_usage(memcg, false);
3421 } else if (type == _MEMSWAP) {
3422 thresholds = &memcg->memsw_thresholds;
3423 usage = mem_cgroup_usage(memcg, true);
3427 /* Check if a threshold crossed before adding a new one */
3428 if (thresholds->primary)
3429 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3431 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3433 /* Allocate memory for new array of thresholds */
3434 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3442 /* Copy thresholds (if any) to new array */
3443 if (thresholds->primary) {
3444 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3445 sizeof(struct mem_cgroup_threshold));
3448 /* Add new threshold */
3449 new->entries[size - 1].eventfd = eventfd;
3450 new->entries[size - 1].threshold = threshold;
3452 /* Sort thresholds. Registering of new threshold isn't time-critical */
3453 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3454 compare_thresholds, NULL);
3456 /* Find current threshold */
3457 new->current_threshold = -1;
3458 for (i = 0; i < size; i++) {
3459 if (new->entries[i].threshold <= usage) {
3461 * new->current_threshold will not be used until
3462 * rcu_assign_pointer(), so it's safe to increment
3465 ++new->current_threshold;
3470 /* Free old spare buffer and save old primary buffer as spare */
3471 kfree(thresholds->spare);
3472 thresholds->spare = thresholds->primary;
3474 rcu_assign_pointer(thresholds->primary, new);
3476 /* To be sure that nobody uses thresholds */
3480 mutex_unlock(&memcg->thresholds_lock);
3485 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3486 struct eventfd_ctx *eventfd, const char *args)
3488 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3491 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3492 struct eventfd_ctx *eventfd, const char *args)
3494 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3497 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3498 struct eventfd_ctx *eventfd, enum res_type type)
3500 struct mem_cgroup_thresholds *thresholds;
3501 struct mem_cgroup_threshold_ary *new;
3502 unsigned long usage;
3503 int i, j, size, entries;
3505 mutex_lock(&memcg->thresholds_lock);
3508 thresholds = &memcg->thresholds;
3509 usage = mem_cgroup_usage(memcg, false);
3510 } else if (type == _MEMSWAP) {
3511 thresholds = &memcg->memsw_thresholds;
3512 usage = mem_cgroup_usage(memcg, true);
3516 if (!thresholds->primary)
3519 /* Check if a threshold crossed before removing */
3520 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3522 /* Calculate new number of threshold */
3524 for (i = 0; i < thresholds->primary->size; i++) {
3525 if (thresholds->primary->entries[i].eventfd != eventfd)
3531 new = thresholds->spare;
3533 /* If no items related to eventfd have been cleared, nothing to do */
3537 /* Set thresholds array to NULL if we don't have thresholds */
3546 /* Copy thresholds and find current threshold */
3547 new->current_threshold = -1;
3548 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3549 if (thresholds->primary->entries[i].eventfd == eventfd)
3552 new->entries[j] = thresholds->primary->entries[i];
3553 if (new->entries[j].threshold <= usage) {
3555 * new->current_threshold will not be used
3556 * until rcu_assign_pointer(), so it's safe to increment
3559 ++new->current_threshold;
3565 /* Swap primary and spare array */
3566 thresholds->spare = thresholds->primary;
3568 rcu_assign_pointer(thresholds->primary, new);
3570 /* To be sure that nobody uses thresholds */
3573 /* If all events are unregistered, free the spare array */
3575 kfree(thresholds->spare);
3576 thresholds->spare = NULL;
3579 mutex_unlock(&memcg->thresholds_lock);
3582 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3583 struct eventfd_ctx *eventfd)
3585 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3588 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3589 struct eventfd_ctx *eventfd)
3591 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3594 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3595 struct eventfd_ctx *eventfd, const char *args)
3597 struct mem_cgroup_eventfd_list *event;
3599 event = kmalloc(sizeof(*event), GFP_KERNEL);
3603 spin_lock(&memcg_oom_lock);
3605 event->eventfd = eventfd;
3606 list_add(&event->list, &memcg->oom_notify);
3608 /* already in OOM ? */
3609 if (memcg->under_oom)
3610 eventfd_signal(eventfd, 1);
3611 spin_unlock(&memcg_oom_lock);
3616 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3617 struct eventfd_ctx *eventfd)
3619 struct mem_cgroup_eventfd_list *ev, *tmp;
3621 spin_lock(&memcg_oom_lock);
3623 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3624 if (ev->eventfd == eventfd) {
3625 list_del(&ev->list);
3630 spin_unlock(&memcg_oom_lock);
3633 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3635 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3637 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3638 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3642 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3643 struct cftype *cft, u64 val)
3645 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3647 /* cannot set to root cgroup and only 0 and 1 are allowed */
3648 if (!css->parent || !((val == 0) || (val == 1)))
3651 memcg->oom_kill_disable = val;
3653 memcg_oom_recover(memcg);
3658 #ifdef CONFIG_CGROUP_WRITEBACK
3660 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3662 return &memcg->cgwb_list;
3665 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3667 return wb_domain_init(&memcg->cgwb_domain, gfp);
3670 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3672 wb_domain_exit(&memcg->cgwb_domain);
3675 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3677 wb_domain_size_changed(&memcg->cgwb_domain);
3680 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3682 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3684 if (!memcg->css.parent)
3687 return &memcg->cgwb_domain;
3691 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3692 * @wb: bdi_writeback in question
3693 * @pfilepages: out parameter for number of file pages
3694 * @pheadroom: out parameter for number of allocatable pages according to memcg
3695 * @pdirty: out parameter for number of dirty pages
3696 * @pwriteback: out parameter for number of pages under writeback
3698 * Determine the numbers of file, headroom, dirty, and writeback pages in
3699 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3700 * is a bit more involved.
3702 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3703 * headroom is calculated as the lowest headroom of itself and the
3704 * ancestors. Note that this doesn't consider the actual amount of
3705 * available memory in the system. The caller should further cap
3706 * *@pheadroom accordingly.
3708 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3709 unsigned long *pheadroom, unsigned long *pdirty,
3710 unsigned long *pwriteback)
3712 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3713 struct mem_cgroup *parent;
3715 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3717 /* this should eventually include NR_UNSTABLE_NFS */
3718 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3719 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3720 (1 << LRU_ACTIVE_FILE));
3721 *pheadroom = PAGE_COUNTER_MAX;
3723 while ((parent = parent_mem_cgroup(memcg))) {
3724 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3725 unsigned long used = page_counter_read(&memcg->memory);
3727 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3732 #else /* CONFIG_CGROUP_WRITEBACK */
3734 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3739 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3743 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3747 #endif /* CONFIG_CGROUP_WRITEBACK */
3750 * DO NOT USE IN NEW FILES.
3752 * "cgroup.event_control" implementation.
3754 * This is way over-engineered. It tries to support fully configurable
3755 * events for each user. Such level of flexibility is completely
3756 * unnecessary especially in the light of the planned unified hierarchy.
3758 * Please deprecate this and replace with something simpler if at all
3763 * Unregister event and free resources.
3765 * Gets called from workqueue.
3767 static void memcg_event_remove(struct work_struct *work)
3769 struct mem_cgroup_event *event =
3770 container_of(work, struct mem_cgroup_event, remove);
3771 struct mem_cgroup *memcg = event->memcg;
3773 remove_wait_queue(event->wqh, &event->wait);
3775 event->unregister_event(memcg, event->eventfd);
3777 /* Notify userspace the event is going away. */
3778 eventfd_signal(event->eventfd, 1);
3780 eventfd_ctx_put(event->eventfd);
3782 css_put(&memcg->css);
3786 * Gets called on POLLHUP on eventfd when user closes it.
3788 * Called with wqh->lock held and interrupts disabled.
3790 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3791 int sync, void *key)
3793 struct mem_cgroup_event *event =
3794 container_of(wait, struct mem_cgroup_event, wait);
3795 struct mem_cgroup *memcg = event->memcg;
3796 unsigned long flags = (unsigned long)key;
3798 if (flags & POLLHUP) {
3800 * If the event has been detached at cgroup removal, we
3801 * can simply return knowing the other side will cleanup
3804 * We can't race against event freeing since the other
3805 * side will require wqh->lock via remove_wait_queue(),
3808 spin_lock(&memcg->event_list_lock);
3809 if (!list_empty(&event->list)) {
3810 list_del_init(&event->list);
3812 * We are in atomic context, but cgroup_event_remove()
3813 * may sleep, so we have to call it in workqueue.
3815 schedule_work(&event->remove);
3817 spin_unlock(&memcg->event_list_lock);
3823 static void memcg_event_ptable_queue_proc(struct file *file,
3824 wait_queue_head_t *wqh, poll_table *pt)
3826 struct mem_cgroup_event *event =
3827 container_of(pt, struct mem_cgroup_event, pt);
3830 add_wait_queue(wqh, &event->wait);
3834 * DO NOT USE IN NEW FILES.
3836 * Parse input and register new cgroup event handler.
3838 * Input must be in format '<event_fd> <control_fd> <args>'.
3839 * Interpretation of args is defined by control file implementation.
3841 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3842 char *buf, size_t nbytes, loff_t off)
3844 struct cgroup_subsys_state *css = of_css(of);
3845 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3846 struct mem_cgroup_event *event;
3847 struct cgroup_subsys_state *cfile_css;
3848 unsigned int efd, cfd;
3855 buf = strstrip(buf);
3857 efd = simple_strtoul(buf, &endp, 10);
3862 cfd = simple_strtoul(buf, &endp, 10);
3863 if ((*endp != ' ') && (*endp != '\0'))
3867 event = kzalloc(sizeof(*event), GFP_KERNEL);
3871 event->memcg = memcg;
3872 INIT_LIST_HEAD(&event->list);
3873 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3874 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3875 INIT_WORK(&event->remove, memcg_event_remove);
3883 event->eventfd = eventfd_ctx_fileget(efile.file);
3884 if (IS_ERR(event->eventfd)) {
3885 ret = PTR_ERR(event->eventfd);
3892 goto out_put_eventfd;
3895 /* the process need read permission on control file */
3896 /* AV: shouldn't we check that it's been opened for read instead? */
3897 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3902 * Determine the event callbacks and set them in @event. This used
3903 * to be done via struct cftype but cgroup core no longer knows
3904 * about these events. The following is crude but the whole thing
3905 * is for compatibility anyway.
3907 * DO NOT ADD NEW FILES.
3909 name = cfile.file->f_path.dentry->d_name.name;
3911 if (!strcmp(name, "memory.usage_in_bytes")) {
3912 event->register_event = mem_cgroup_usage_register_event;
3913 event->unregister_event = mem_cgroup_usage_unregister_event;
3914 } else if (!strcmp(name, "memory.oom_control")) {
3915 event->register_event = mem_cgroup_oom_register_event;
3916 event->unregister_event = mem_cgroup_oom_unregister_event;
3917 } else if (!strcmp(name, "memory.pressure_level")) {
3918 event->register_event = vmpressure_register_event;
3919 event->unregister_event = vmpressure_unregister_event;
3920 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3921 event->register_event = memsw_cgroup_usage_register_event;
3922 event->unregister_event = memsw_cgroup_usage_unregister_event;
3929 * Verify @cfile should belong to @css. Also, remaining events are
3930 * automatically removed on cgroup destruction but the removal is
3931 * asynchronous, so take an extra ref on @css.
3933 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3934 &memory_cgrp_subsys);
3936 if (IS_ERR(cfile_css))
3938 if (cfile_css != css) {
3943 ret = event->register_event(memcg, event->eventfd, buf);
3947 efile.file->f_op->poll(efile.file, &event->pt);
3949 spin_lock(&memcg->event_list_lock);
3950 list_add(&event->list, &memcg->event_list);
3951 spin_unlock(&memcg->event_list_lock);
3963 eventfd_ctx_put(event->eventfd);
3972 static struct cftype mem_cgroup_legacy_files[] = {
3974 .name = "usage_in_bytes",
3975 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3976 .read_u64 = mem_cgroup_read_u64,
3979 .name = "max_usage_in_bytes",
3980 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3981 .write = mem_cgroup_reset,
3982 .read_u64 = mem_cgroup_read_u64,
3985 .name = "limit_in_bytes",
3986 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3987 .write = mem_cgroup_write,
3988 .read_u64 = mem_cgroup_read_u64,
3991 .name = "soft_limit_in_bytes",
3992 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3993 .write = mem_cgroup_write,
3994 .read_u64 = mem_cgroup_read_u64,
3998 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3999 .write = mem_cgroup_reset,
4000 .read_u64 = mem_cgroup_read_u64,
4004 .seq_show = memcg_stat_show,
4007 .name = "force_empty",
4008 .write = mem_cgroup_force_empty_write,
4011 .name = "use_hierarchy",
4012 .write_u64 = mem_cgroup_hierarchy_write,
4013 .read_u64 = mem_cgroup_hierarchy_read,
4016 .name = "cgroup.event_control", /* XXX: for compat */
4017 .write = memcg_write_event_control,
4018 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4021 .name = "swappiness",
4022 .read_u64 = mem_cgroup_swappiness_read,
4023 .write_u64 = mem_cgroup_swappiness_write,
4026 .name = "move_charge_at_immigrate",
4027 .read_u64 = mem_cgroup_move_charge_read,
4028 .write_u64 = mem_cgroup_move_charge_write,
4031 .name = "oom_control",
4032 .seq_show = mem_cgroup_oom_control_read,
4033 .write_u64 = mem_cgroup_oom_control_write,
4034 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4037 .name = "pressure_level",
4041 .name = "numa_stat",
4042 .seq_show = memcg_numa_stat_show,
4046 .name = "kmem.limit_in_bytes",
4047 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4048 .write = mem_cgroup_write,
4049 .read_u64 = mem_cgroup_read_u64,
4052 .name = "kmem.usage_in_bytes",
4053 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4054 .read_u64 = mem_cgroup_read_u64,
4057 .name = "kmem.failcnt",
4058 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4059 .write = mem_cgroup_reset,
4060 .read_u64 = mem_cgroup_read_u64,
4063 .name = "kmem.max_usage_in_bytes",
4064 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4065 .write = mem_cgroup_reset,
4066 .read_u64 = mem_cgroup_read_u64,
4068 #ifdef CONFIG_SLABINFO
4070 .name = "kmem.slabinfo",
4071 .seq_start = slab_start,
4072 .seq_next = slab_next,
4073 .seq_stop = slab_stop,
4074 .seq_show = memcg_slab_show,
4078 .name = "kmem.tcp.limit_in_bytes",
4079 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4080 .write = mem_cgroup_write,
4081 .read_u64 = mem_cgroup_read_u64,
4084 .name = "kmem.tcp.usage_in_bytes",
4085 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4086 .read_u64 = mem_cgroup_read_u64,
4089 .name = "kmem.tcp.failcnt",
4090 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4091 .write = mem_cgroup_reset,
4092 .read_u64 = mem_cgroup_read_u64,
4095 .name = "kmem.tcp.max_usage_in_bytes",
4096 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4097 .write = mem_cgroup_reset,
4098 .read_u64 = mem_cgroup_read_u64,
4100 { }, /* terminate */
4104 * Private memory cgroup IDR
4106 * Swap-out records and page cache shadow entries need to store memcg
4107 * references in constrained space, so we maintain an ID space that is
4108 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4109 * memory-controlled cgroups to 64k.
4111 * However, there usually are many references to the oflline CSS after
4112 * the cgroup has been destroyed, such as page cache or reclaimable
4113 * slab objects, that don't need to hang on to the ID. We want to keep
4114 * those dead CSS from occupying IDs, or we might quickly exhaust the
4115 * relatively small ID space and prevent the creation of new cgroups
4116 * even when there are much fewer than 64k cgroups - possibly none.
4118 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4119 * be freed and recycled when it's no longer needed, which is usually
4120 * when the CSS is offlined.
4122 * The only exception to that are records of swapped out tmpfs/shmem
4123 * pages that need to be attributed to live ancestors on swapin. But
4124 * those references are manageable from userspace.
4127 static DEFINE_IDR(mem_cgroup_idr);
4129 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4131 if (memcg->id.id > 0) {
4132 idr_remove(&mem_cgroup_idr, memcg->id.id);
4137 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4139 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4140 atomic_add(n, &memcg->id.ref);
4143 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4145 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4146 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4147 mem_cgroup_id_remove(memcg);
4149 /* Memcg ID pins CSS */
4150 css_put(&memcg->css);
4154 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4156 mem_cgroup_id_get_many(memcg, 1);
4159 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4161 mem_cgroup_id_put_many(memcg, 1);
4165 * mem_cgroup_from_id - look up a memcg from a memcg id
4166 * @id: the memcg id to look up
4168 * Caller must hold rcu_read_lock().
4170 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4172 WARN_ON_ONCE(!rcu_read_lock_held());
4173 return idr_find(&mem_cgroup_idr, id);
4176 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4178 struct mem_cgroup_per_node *pn;
4181 * This routine is called against possible nodes.
4182 * But it's BUG to call kmalloc() against offline node.
4184 * TODO: this routine can waste much memory for nodes which will
4185 * never be onlined. It's better to use memory hotplug callback
4188 if (!node_state(node, N_NORMAL_MEMORY))
4190 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4194 lruvec_init(&pn->lruvec);
4195 pn->usage_in_excess = 0;
4196 pn->on_tree = false;
4199 memcg->nodeinfo[node] = pn;
4203 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4205 kfree(memcg->nodeinfo[node]);
4208 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4213 free_mem_cgroup_per_node_info(memcg, node);
4214 free_percpu(memcg->stat);
4218 static void mem_cgroup_free(struct mem_cgroup *memcg)
4220 memcg_wb_domain_exit(memcg);
4221 __mem_cgroup_free(memcg);
4224 static struct mem_cgroup *mem_cgroup_alloc(void)
4226 struct mem_cgroup *memcg;
4230 size = sizeof(struct mem_cgroup);
4231 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4233 memcg = kzalloc(size, GFP_KERNEL);
4237 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4238 1, MEM_CGROUP_ID_MAX,
4240 if (memcg->id.id < 0)
4243 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4248 if (alloc_mem_cgroup_per_node_info(memcg, node))
4251 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4254 INIT_WORK(&memcg->high_work, high_work_func);
4255 memcg->last_scanned_node = MAX_NUMNODES;
4256 INIT_LIST_HEAD(&memcg->oom_notify);
4257 mutex_init(&memcg->thresholds_lock);
4258 spin_lock_init(&memcg->move_lock);
4259 vmpressure_init(&memcg->vmpressure);
4260 INIT_LIST_HEAD(&memcg->event_list);
4261 spin_lock_init(&memcg->event_list_lock);
4262 memcg->socket_pressure = jiffies;
4264 memcg->kmemcg_id = -1;
4266 #ifdef CONFIG_CGROUP_WRITEBACK
4267 INIT_LIST_HEAD(&memcg->cgwb_list);
4269 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4272 mem_cgroup_id_remove(memcg);
4273 __mem_cgroup_free(memcg);
4277 static struct cgroup_subsys_state * __ref
4278 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4280 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4281 struct mem_cgroup *memcg;
4282 long error = -ENOMEM;
4284 memcg = mem_cgroup_alloc();
4286 return ERR_PTR(error);
4288 memcg->high = PAGE_COUNTER_MAX;
4289 memcg->soft_limit = PAGE_COUNTER_MAX;
4291 memcg->swappiness = mem_cgroup_swappiness(parent);
4292 memcg->oom_kill_disable = parent->oom_kill_disable;
4294 if (parent && parent->use_hierarchy) {
4295 memcg->use_hierarchy = true;
4296 page_counter_init(&memcg->memory, &parent->memory);
4297 page_counter_init(&memcg->swap, &parent->swap);
4298 page_counter_init(&memcg->memsw, &parent->memsw);
4299 page_counter_init(&memcg->kmem, &parent->kmem);
4300 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4302 page_counter_init(&memcg->memory, NULL);
4303 page_counter_init(&memcg->swap, NULL);
4304 page_counter_init(&memcg->memsw, NULL);
4305 page_counter_init(&memcg->kmem, NULL);
4306 page_counter_init(&memcg->tcpmem, NULL);
4308 * Deeper hierachy with use_hierarchy == false doesn't make
4309 * much sense so let cgroup subsystem know about this
4310 * unfortunate state in our controller.
4312 if (parent != root_mem_cgroup)
4313 memory_cgrp_subsys.broken_hierarchy = true;
4316 /* The following stuff does not apply to the root */
4318 root_mem_cgroup = memcg;
4322 error = memcg_online_kmem(memcg);
4326 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4327 static_branch_inc(&memcg_sockets_enabled_key);
4331 mem_cgroup_id_remove(memcg);
4332 mem_cgroup_free(memcg);
4333 return ERR_PTR(-ENOMEM);
4336 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4338 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4340 /* Online state pins memcg ID, memcg ID pins CSS */
4341 atomic_set(&memcg->id.ref, 1);
4346 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4348 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4349 struct mem_cgroup_event *event, *tmp;
4352 * Unregister events and notify userspace.
4353 * Notify userspace about cgroup removing only after rmdir of cgroup
4354 * directory to avoid race between userspace and kernelspace.
4356 spin_lock(&memcg->event_list_lock);
4357 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4358 list_del_init(&event->list);
4359 schedule_work(&event->remove);
4361 spin_unlock(&memcg->event_list_lock);
4363 memcg_offline_kmem(memcg);
4364 wb_memcg_offline(memcg);
4366 mem_cgroup_id_put(memcg);
4369 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4373 invalidate_reclaim_iterators(memcg);
4376 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4378 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4380 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4381 static_branch_dec(&memcg_sockets_enabled_key);
4383 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4384 static_branch_dec(&memcg_sockets_enabled_key);
4386 vmpressure_cleanup(&memcg->vmpressure);
4387 cancel_work_sync(&memcg->high_work);
4388 mem_cgroup_remove_from_trees(memcg);
4389 memcg_free_kmem(memcg);
4390 mem_cgroup_free(memcg);
4394 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4395 * @css: the target css
4397 * Reset the states of the mem_cgroup associated with @css. This is
4398 * invoked when the userland requests disabling on the default hierarchy
4399 * but the memcg is pinned through dependency. The memcg should stop
4400 * applying policies and should revert to the vanilla state as it may be
4401 * made visible again.
4403 * The current implementation only resets the essential configurations.
4404 * This needs to be expanded to cover all the visible parts.
4406 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4408 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4410 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4411 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4412 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4413 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4414 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4416 memcg->high = PAGE_COUNTER_MAX;
4417 memcg->soft_limit = PAGE_COUNTER_MAX;
4418 memcg_wb_domain_size_changed(memcg);
4422 /* Handlers for move charge at task migration. */
4423 static int mem_cgroup_do_precharge(unsigned long count)
4427 /* Try a single bulk charge without reclaim first, kswapd may wake */
4428 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4430 mc.precharge += count;
4434 /* Try charges one by one with reclaim, but do not retry */
4436 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4450 enum mc_target_type {
4456 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4457 unsigned long addr, pte_t ptent)
4459 struct page *page = vm_normal_page(vma, addr, ptent);
4461 if (!page || !page_mapped(page))
4463 if (PageAnon(page)) {
4464 if (!(mc.flags & MOVE_ANON))
4467 if (!(mc.flags & MOVE_FILE))
4470 if (!get_page_unless_zero(page))
4477 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4478 pte_t ptent, swp_entry_t *entry)
4480 struct page *page = NULL;
4481 swp_entry_t ent = pte_to_swp_entry(ptent);
4483 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4486 * Because lookup_swap_cache() updates some statistics counter,
4487 * we call find_get_page() with swapper_space directly.
4489 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4490 if (do_memsw_account())
4491 entry->val = ent.val;
4496 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4497 pte_t ptent, swp_entry_t *entry)
4503 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4504 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4506 struct page *page = NULL;
4507 struct address_space *mapping;
4510 if (!vma->vm_file) /* anonymous vma */
4512 if (!(mc.flags & MOVE_FILE))
4515 mapping = vma->vm_file->f_mapping;
4516 pgoff = linear_page_index(vma, addr);
4518 /* page is moved even if it's not RSS of this task(page-faulted). */
4520 /* shmem/tmpfs may report page out on swap: account for that too. */
4521 if (shmem_mapping(mapping)) {
4522 page = find_get_entry(mapping, pgoff);
4523 if (radix_tree_exceptional_entry(page)) {
4524 swp_entry_t swp = radix_to_swp_entry(page);
4525 if (do_memsw_account())
4527 page = find_get_page(swap_address_space(swp),
4531 page = find_get_page(mapping, pgoff);
4533 page = find_get_page(mapping, pgoff);
4539 * mem_cgroup_move_account - move account of the page
4541 * @compound: charge the page as compound or small page
4542 * @from: mem_cgroup which the page is moved from.
4543 * @to: mem_cgroup which the page is moved to. @from != @to.
4545 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4547 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4550 static int mem_cgroup_move_account(struct page *page,
4552 struct mem_cgroup *from,
4553 struct mem_cgroup *to)
4555 unsigned long flags;
4556 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4560 VM_BUG_ON(from == to);
4561 VM_BUG_ON_PAGE(PageLRU(page), page);
4562 VM_BUG_ON(compound && !PageTransHuge(page));
4565 * Prevent mem_cgroup_migrate() from looking at
4566 * page->mem_cgroup of its source page while we change it.
4569 if (!trylock_page(page))
4573 if (page->mem_cgroup != from)
4576 anon = PageAnon(page);
4578 spin_lock_irqsave(&from->move_lock, flags);
4580 if (!anon && page_mapped(page)) {
4581 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4583 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4588 * move_lock grabbed above and caller set from->moving_account, so
4589 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4590 * So mapping should be stable for dirty pages.
4592 if (!anon && PageDirty(page)) {
4593 struct address_space *mapping = page_mapping(page);
4595 if (mapping_cap_account_dirty(mapping)) {
4596 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4598 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4603 if (PageWriteback(page)) {
4604 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4606 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4611 * It is safe to change page->mem_cgroup here because the page
4612 * is referenced, charged, and isolated - we can't race with
4613 * uncharging, charging, migration, or LRU putback.
4616 /* caller should have done css_get */
4617 page->mem_cgroup = to;
4618 spin_unlock_irqrestore(&from->move_lock, flags);
4622 local_irq_disable();
4623 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4624 memcg_check_events(to, page);
4625 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4626 memcg_check_events(from, page);
4635 * get_mctgt_type - get target type of moving charge
4636 * @vma: the vma the pte to be checked belongs
4637 * @addr: the address corresponding to the pte to be checked
4638 * @ptent: the pte to be checked
4639 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4642 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4643 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4644 * move charge. if @target is not NULL, the page is stored in target->page
4645 * with extra refcnt got(Callers should handle it).
4646 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4647 * target for charge migration. if @target is not NULL, the entry is stored
4650 * Called with pte lock held.
4653 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4654 unsigned long addr, pte_t ptent, union mc_target *target)
4656 struct page *page = NULL;
4657 enum mc_target_type ret = MC_TARGET_NONE;
4658 swp_entry_t ent = { .val = 0 };
4660 if (pte_present(ptent))
4661 page = mc_handle_present_pte(vma, addr, ptent);
4662 else if (is_swap_pte(ptent))
4663 page = mc_handle_swap_pte(vma, ptent, &ent);
4664 else if (pte_none(ptent))
4665 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4667 if (!page && !ent.val)
4671 * Do only loose check w/o serialization.
4672 * mem_cgroup_move_account() checks the page is valid or
4673 * not under LRU exclusion.
4675 if (page->mem_cgroup == mc.from) {
4676 ret = MC_TARGET_PAGE;
4678 target->page = page;
4680 if (!ret || !target)
4683 /* There is a swap entry and a page doesn't exist or isn't charged */
4684 if (ent.val && !ret &&
4685 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4686 ret = MC_TARGET_SWAP;
4693 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4695 * We don't consider swapping or file mapped pages because THP does not
4696 * support them for now.
4697 * Caller should make sure that pmd_trans_huge(pmd) is true.
4699 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4700 unsigned long addr, pmd_t pmd, union mc_target *target)
4702 struct page *page = NULL;
4703 enum mc_target_type ret = MC_TARGET_NONE;
4705 page = pmd_page(pmd);
4706 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4707 if (!(mc.flags & MOVE_ANON))
4709 if (page->mem_cgroup == mc.from) {
4710 ret = MC_TARGET_PAGE;
4713 target->page = page;
4719 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4720 unsigned long addr, pmd_t pmd, union mc_target *target)
4722 return MC_TARGET_NONE;
4726 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4727 unsigned long addr, unsigned long end,
4728 struct mm_walk *walk)
4730 struct vm_area_struct *vma = walk->vma;
4734 ptl = pmd_trans_huge_lock(pmd, vma);
4736 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4737 mc.precharge += HPAGE_PMD_NR;
4742 if (pmd_trans_unstable(pmd))
4744 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4745 for (; addr != end; pte++, addr += PAGE_SIZE)
4746 if (get_mctgt_type(vma, addr, *pte, NULL))
4747 mc.precharge++; /* increment precharge temporarily */
4748 pte_unmap_unlock(pte - 1, ptl);
4754 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4756 unsigned long precharge;
4758 struct mm_walk mem_cgroup_count_precharge_walk = {
4759 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4762 down_read(&mm->mmap_sem);
4763 walk_page_range(0, mm->highest_vm_end,
4764 &mem_cgroup_count_precharge_walk);
4765 up_read(&mm->mmap_sem);
4767 precharge = mc.precharge;
4773 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4775 unsigned long precharge = mem_cgroup_count_precharge(mm);
4777 VM_BUG_ON(mc.moving_task);
4778 mc.moving_task = current;
4779 return mem_cgroup_do_precharge(precharge);
4782 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4783 static void __mem_cgroup_clear_mc(void)
4785 struct mem_cgroup *from = mc.from;
4786 struct mem_cgroup *to = mc.to;
4788 /* we must uncharge all the leftover precharges from mc.to */
4790 cancel_charge(mc.to, mc.precharge);
4794 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4795 * we must uncharge here.
4797 if (mc.moved_charge) {
4798 cancel_charge(mc.from, mc.moved_charge);
4799 mc.moved_charge = 0;
4801 /* we must fixup refcnts and charges */
4802 if (mc.moved_swap) {
4803 /* uncharge swap account from the old cgroup */
4804 if (!mem_cgroup_is_root(mc.from))
4805 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4807 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4810 * we charged both to->memory and to->memsw, so we
4811 * should uncharge to->memory.
4813 if (!mem_cgroup_is_root(mc.to))
4814 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4816 css_put_many(&mc.to->css, mc.moved_swap);
4820 memcg_oom_recover(from);
4821 memcg_oom_recover(to);
4822 wake_up_all(&mc.waitq);
4825 static void mem_cgroup_clear_mc(void)
4827 struct mm_struct *mm = mc.mm;
4830 * we must clear moving_task before waking up waiters at the end of
4833 mc.moving_task = NULL;
4834 __mem_cgroup_clear_mc();
4835 spin_lock(&mc.lock);
4839 spin_unlock(&mc.lock);
4844 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4846 struct cgroup_subsys_state *css;
4847 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4848 struct mem_cgroup *from;
4849 struct task_struct *leader, *p;
4850 struct mm_struct *mm;
4851 unsigned long move_flags;
4854 /* charge immigration isn't supported on the default hierarchy */
4855 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4859 * Multi-process migrations only happen on the default hierarchy
4860 * where charge immigration is not used. Perform charge
4861 * immigration if @tset contains a leader and whine if there are
4865 cgroup_taskset_for_each_leader(leader, css, tset) {
4868 memcg = mem_cgroup_from_css(css);
4874 * We are now commited to this value whatever it is. Changes in this
4875 * tunable will only affect upcoming migrations, not the current one.
4876 * So we need to save it, and keep it going.
4878 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4882 from = mem_cgroup_from_task(p);
4884 VM_BUG_ON(from == memcg);
4886 mm = get_task_mm(p);
4889 /* We move charges only when we move a owner of the mm */
4890 if (mm->owner == p) {
4893 VM_BUG_ON(mc.precharge);
4894 VM_BUG_ON(mc.moved_charge);
4895 VM_BUG_ON(mc.moved_swap);
4897 spin_lock(&mc.lock);
4901 mc.flags = move_flags;
4902 spin_unlock(&mc.lock);
4903 /* We set mc.moving_task later */
4905 ret = mem_cgroup_precharge_mc(mm);
4907 mem_cgroup_clear_mc();
4914 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4917 mem_cgroup_clear_mc();
4920 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4921 unsigned long addr, unsigned long end,
4922 struct mm_walk *walk)
4925 struct vm_area_struct *vma = walk->vma;
4928 enum mc_target_type target_type;
4929 union mc_target target;
4932 ptl = pmd_trans_huge_lock(pmd, vma);
4934 if (mc.precharge < HPAGE_PMD_NR) {
4938 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4939 if (target_type == MC_TARGET_PAGE) {
4941 if (!isolate_lru_page(page)) {
4942 if (!mem_cgroup_move_account(page, true,
4944 mc.precharge -= HPAGE_PMD_NR;
4945 mc.moved_charge += HPAGE_PMD_NR;
4947 putback_lru_page(page);
4955 if (pmd_trans_unstable(pmd))
4958 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4959 for (; addr != end; addr += PAGE_SIZE) {
4960 pte_t ptent = *(pte++);
4966 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4967 case MC_TARGET_PAGE:
4970 * We can have a part of the split pmd here. Moving it
4971 * can be done but it would be too convoluted so simply
4972 * ignore such a partial THP and keep it in original
4973 * memcg. There should be somebody mapping the head.
4975 if (PageTransCompound(page))
4977 if (isolate_lru_page(page))
4979 if (!mem_cgroup_move_account(page, false,
4982 /* we uncharge from mc.from later. */
4985 putback_lru_page(page);
4986 put: /* get_mctgt_type() gets the page */
4989 case MC_TARGET_SWAP:
4991 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4993 mem_cgroup_id_get_many(mc.to, 1);
4994 /* we fixup other refcnts and charges later. */
5002 pte_unmap_unlock(pte - 1, ptl);
5007 * We have consumed all precharges we got in can_attach().
5008 * We try charge one by one, but don't do any additional
5009 * charges to mc.to if we have failed in charge once in attach()
5012 ret = mem_cgroup_do_precharge(1);
5020 static void mem_cgroup_move_charge(void)
5022 struct mm_walk mem_cgroup_move_charge_walk = {
5023 .pmd_entry = mem_cgroup_move_charge_pte_range,
5027 lru_add_drain_all();
5029 * Signal lock_page_memcg() to take the memcg's move_lock
5030 * while we're moving its pages to another memcg. Then wait
5031 * for already started RCU-only updates to finish.
5033 atomic_inc(&mc.from->moving_account);
5036 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5038 * Someone who are holding the mmap_sem might be waiting in
5039 * waitq. So we cancel all extra charges, wake up all waiters,
5040 * and retry. Because we cancel precharges, we might not be able
5041 * to move enough charges, but moving charge is a best-effort
5042 * feature anyway, so it wouldn't be a big problem.
5044 __mem_cgroup_clear_mc();
5049 * When we have consumed all precharges and failed in doing
5050 * additional charge, the page walk just aborts.
5052 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5054 up_read(&mc.mm->mmap_sem);
5055 atomic_dec(&mc.from->moving_account);
5058 static void mem_cgroup_move_task(void)
5061 mem_cgroup_move_charge();
5062 mem_cgroup_clear_mc();
5065 #else /* !CONFIG_MMU */
5066 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5070 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5073 static void mem_cgroup_move_task(void)
5079 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5080 * to verify whether we're attached to the default hierarchy on each mount
5083 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5086 * use_hierarchy is forced on the default hierarchy. cgroup core
5087 * guarantees that @root doesn't have any children, so turning it
5088 * on for the root memcg is enough.
5090 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5091 root_mem_cgroup->use_hierarchy = true;
5093 root_mem_cgroup->use_hierarchy = false;
5096 static u64 memory_current_read(struct cgroup_subsys_state *css,
5099 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5101 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5104 static int memory_low_show(struct seq_file *m, void *v)
5106 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5107 unsigned long low = READ_ONCE(memcg->low);
5109 if (low == PAGE_COUNTER_MAX)
5110 seq_puts(m, "max\n");
5112 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5117 static ssize_t memory_low_write(struct kernfs_open_file *of,
5118 char *buf, size_t nbytes, loff_t off)
5120 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5124 buf = strstrip(buf);
5125 err = page_counter_memparse(buf, "max", &low);
5134 static int memory_high_show(struct seq_file *m, void *v)
5136 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5137 unsigned long high = READ_ONCE(memcg->high);
5139 if (high == PAGE_COUNTER_MAX)
5140 seq_puts(m, "max\n");
5142 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5147 static ssize_t memory_high_write(struct kernfs_open_file *of,
5148 char *buf, size_t nbytes, loff_t off)
5150 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5151 unsigned long nr_pages;
5155 buf = strstrip(buf);
5156 err = page_counter_memparse(buf, "max", &high);
5162 nr_pages = page_counter_read(&memcg->memory);
5163 if (nr_pages > high)
5164 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5167 memcg_wb_domain_size_changed(memcg);
5171 static int memory_max_show(struct seq_file *m, void *v)
5173 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5174 unsigned long max = READ_ONCE(memcg->memory.limit);
5176 if (max == PAGE_COUNTER_MAX)
5177 seq_puts(m, "max\n");
5179 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5184 static ssize_t memory_max_write(struct kernfs_open_file *of,
5185 char *buf, size_t nbytes, loff_t off)
5187 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5188 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5189 bool drained = false;
5193 buf = strstrip(buf);
5194 err = page_counter_memparse(buf, "max", &max);
5198 xchg(&memcg->memory.limit, max);
5201 unsigned long nr_pages = page_counter_read(&memcg->memory);
5203 if (nr_pages <= max)
5206 if (signal_pending(current)) {
5212 drain_all_stock(memcg);
5218 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5224 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5225 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5229 memcg_wb_domain_size_changed(memcg);
5233 static int memory_events_show(struct seq_file *m, void *v)
5235 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5237 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5238 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5239 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5240 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5245 static int memory_stat_show(struct seq_file *m, void *v)
5247 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5248 unsigned long stat[MEMCG_NR_STAT];
5249 unsigned long events[MEMCG_NR_EVENTS];
5253 * Provide statistics on the state of the memory subsystem as
5254 * well as cumulative event counters that show past behavior.
5256 * This list is ordered following a combination of these gradients:
5257 * 1) generic big picture -> specifics and details
5258 * 2) reflecting userspace activity -> reflecting kernel heuristics
5260 * Current memory state:
5263 tree_stat(memcg, stat);
5264 tree_events(memcg, events);
5266 seq_printf(m, "anon %llu\n",
5267 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5268 seq_printf(m, "file %llu\n",
5269 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5270 seq_printf(m, "kernel_stack %llu\n",
5271 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5272 seq_printf(m, "slab %llu\n",
5273 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5274 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5275 seq_printf(m, "sock %llu\n",
5276 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5278 seq_printf(m, "file_mapped %llu\n",
5279 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5280 seq_printf(m, "file_dirty %llu\n",
5281 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5282 seq_printf(m, "file_writeback %llu\n",
5283 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5285 for (i = 0; i < NR_LRU_LISTS; i++) {
5286 struct mem_cgroup *mi;
5287 unsigned long val = 0;
5289 for_each_mem_cgroup_tree(mi, memcg)
5290 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5291 seq_printf(m, "%s %llu\n",
5292 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5295 seq_printf(m, "slab_reclaimable %llu\n",
5296 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5297 seq_printf(m, "slab_unreclaimable %llu\n",
5298 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5300 /* Accumulated memory events */
5302 seq_printf(m, "pgfault %lu\n",
5303 events[MEM_CGROUP_EVENTS_PGFAULT]);
5304 seq_printf(m, "pgmajfault %lu\n",
5305 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5310 static struct cftype memory_files[] = {
5313 .flags = CFTYPE_NOT_ON_ROOT,
5314 .read_u64 = memory_current_read,
5318 .flags = CFTYPE_NOT_ON_ROOT,
5319 .seq_show = memory_low_show,
5320 .write = memory_low_write,
5324 .flags = CFTYPE_NOT_ON_ROOT,
5325 .seq_show = memory_high_show,
5326 .write = memory_high_write,
5330 .flags = CFTYPE_NOT_ON_ROOT,
5331 .seq_show = memory_max_show,
5332 .write = memory_max_write,
5336 .flags = CFTYPE_NOT_ON_ROOT,
5337 .file_offset = offsetof(struct mem_cgroup, events_file),
5338 .seq_show = memory_events_show,
5342 .flags = CFTYPE_NOT_ON_ROOT,
5343 .seq_show = memory_stat_show,
5348 struct cgroup_subsys memory_cgrp_subsys = {
5349 .css_alloc = mem_cgroup_css_alloc,
5350 .css_online = mem_cgroup_css_online,
5351 .css_offline = mem_cgroup_css_offline,
5352 .css_released = mem_cgroup_css_released,
5353 .css_free = mem_cgroup_css_free,
5354 .css_reset = mem_cgroup_css_reset,
5355 .can_attach = mem_cgroup_can_attach,
5356 .cancel_attach = mem_cgroup_cancel_attach,
5357 .post_attach = mem_cgroup_move_task,
5358 .bind = mem_cgroup_bind,
5359 .dfl_cftypes = memory_files,
5360 .legacy_cftypes = mem_cgroup_legacy_files,
5365 * mem_cgroup_low - check if memory consumption is below the normal range
5366 * @root: the highest ancestor to consider
5367 * @memcg: the memory cgroup to check
5369 * Returns %true if memory consumption of @memcg, and that of all
5370 * configurable ancestors up to @root, is below the normal range.
5372 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5374 if (mem_cgroup_disabled())
5378 * The toplevel group doesn't have a configurable range, so
5379 * it's never low when looked at directly, and it is not
5380 * considered an ancestor when assessing the hierarchy.
5383 if (memcg == root_mem_cgroup)
5386 if (page_counter_read(&memcg->memory) >= memcg->low)
5389 while (memcg != root) {
5390 memcg = parent_mem_cgroup(memcg);
5392 if (memcg == root_mem_cgroup)
5395 if (page_counter_read(&memcg->memory) >= memcg->low)
5402 * mem_cgroup_try_charge - try charging a page
5403 * @page: page to charge
5404 * @mm: mm context of the victim
5405 * @gfp_mask: reclaim mode
5406 * @memcgp: charged memcg return
5407 * @compound: charge the page as compound or small page
5409 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5410 * pages according to @gfp_mask if necessary.
5412 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5413 * Otherwise, an error code is returned.
5415 * After page->mapping has been set up, the caller must finalize the
5416 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5417 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5419 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5420 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5423 struct mem_cgroup *memcg = NULL;
5424 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5427 if (mem_cgroup_disabled())
5430 if (PageSwapCache(page)) {
5432 * Every swap fault against a single page tries to charge the
5433 * page, bail as early as possible. shmem_unuse() encounters
5434 * already charged pages, too. The USED bit is protected by
5435 * the page lock, which serializes swap cache removal, which
5436 * in turn serializes uncharging.
5438 VM_BUG_ON_PAGE(!PageLocked(page), page);
5439 if (page->mem_cgroup)
5442 if (do_swap_account) {
5443 swp_entry_t ent = { .val = page_private(page), };
5444 unsigned short id = lookup_swap_cgroup_id(ent);
5447 memcg = mem_cgroup_from_id(id);
5448 if (memcg && !css_tryget_online(&memcg->css))
5455 memcg = get_mem_cgroup_from_mm(mm);
5457 ret = try_charge(memcg, gfp_mask, nr_pages);
5459 css_put(&memcg->css);
5466 * mem_cgroup_commit_charge - commit a page charge
5467 * @page: page to charge
5468 * @memcg: memcg to charge the page to
5469 * @lrucare: page might be on LRU already
5470 * @compound: charge the page as compound or small page
5472 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5473 * after page->mapping has been set up. This must happen atomically
5474 * as part of the page instantiation, i.e. under the page table lock
5475 * for anonymous pages, under the page lock for page and swap cache.
5477 * In addition, the page must not be on the LRU during the commit, to
5478 * prevent racing with task migration. If it might be, use @lrucare.
5480 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5482 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5483 bool lrucare, bool compound)
5485 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5487 VM_BUG_ON_PAGE(!page->mapping, page);
5488 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5490 if (mem_cgroup_disabled())
5493 * Swap faults will attempt to charge the same page multiple
5494 * times. But reuse_swap_page() might have removed the page
5495 * from swapcache already, so we can't check PageSwapCache().
5500 commit_charge(page, memcg, lrucare);
5502 local_irq_disable();
5503 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5504 memcg_check_events(memcg, page);
5507 if (do_memsw_account() && PageSwapCache(page)) {
5508 swp_entry_t entry = { .val = page_private(page) };
5510 * The swap entry might not get freed for a long time,
5511 * let's not wait for it. The page already received a
5512 * memory+swap charge, drop the swap entry duplicate.
5514 mem_cgroup_uncharge_swap(entry);
5519 * mem_cgroup_cancel_charge - cancel a page charge
5520 * @page: page to charge
5521 * @memcg: memcg to charge the page to
5522 * @compound: charge the page as compound or small page
5524 * Cancel a charge transaction started by mem_cgroup_try_charge().
5526 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5529 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5531 if (mem_cgroup_disabled())
5534 * Swap faults will attempt to charge the same page multiple
5535 * times. But reuse_swap_page() might have removed the page
5536 * from swapcache already, so we can't check PageSwapCache().
5541 cancel_charge(memcg, nr_pages);
5544 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5545 unsigned long nr_anon, unsigned long nr_file,
5546 unsigned long nr_huge, unsigned long nr_kmem,
5547 struct page *dummy_page)
5549 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5550 unsigned long flags;
5552 if (!mem_cgroup_is_root(memcg)) {
5553 page_counter_uncharge(&memcg->memory, nr_pages);
5554 if (do_memsw_account())
5555 page_counter_uncharge(&memcg->memsw, nr_pages);
5556 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5557 page_counter_uncharge(&memcg->kmem, nr_kmem);
5558 memcg_oom_recover(memcg);
5561 local_irq_save(flags);
5562 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5563 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5564 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5565 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5566 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5567 memcg_check_events(memcg, dummy_page);
5568 local_irq_restore(flags);
5570 if (!mem_cgroup_is_root(memcg))
5571 css_put_many(&memcg->css, nr_pages);
5574 static void uncharge_list(struct list_head *page_list)
5576 struct mem_cgroup *memcg = NULL;
5577 unsigned long nr_anon = 0;
5578 unsigned long nr_file = 0;
5579 unsigned long nr_huge = 0;
5580 unsigned long nr_kmem = 0;
5581 unsigned long pgpgout = 0;
5582 struct list_head *next;
5586 * Note that the list can be a single page->lru; hence the
5587 * do-while loop instead of a simple list_for_each_entry().
5589 next = page_list->next;
5591 page = list_entry(next, struct page, lru);
5592 next = page->lru.next;
5594 VM_BUG_ON_PAGE(PageLRU(page), page);
5595 VM_BUG_ON_PAGE(!PageHWPoison(page) && page_count(page), page);
5597 if (!page->mem_cgroup)
5601 * Nobody should be changing or seriously looking at
5602 * page->mem_cgroup at this point, we have fully
5603 * exclusive access to the page.
5606 if (memcg != page->mem_cgroup) {
5608 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5609 nr_huge, nr_kmem, page);
5610 pgpgout = nr_anon = nr_file =
5611 nr_huge = nr_kmem = 0;
5613 memcg = page->mem_cgroup;
5616 if (!PageKmemcg(page)) {
5617 unsigned int nr_pages = 1;
5619 if (PageTransHuge(page)) {
5620 nr_pages <<= compound_order(page);
5621 nr_huge += nr_pages;
5624 nr_anon += nr_pages;
5626 nr_file += nr_pages;
5629 nr_kmem += 1 << compound_order(page);
5630 __ClearPageKmemcg(page);
5633 page->mem_cgroup = NULL;
5634 } while (next != page_list);
5637 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5638 nr_huge, nr_kmem, page);
5642 * mem_cgroup_uncharge - uncharge a page
5643 * @page: page to uncharge
5645 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5646 * mem_cgroup_commit_charge().
5648 void mem_cgroup_uncharge(struct page *page)
5650 if (mem_cgroup_disabled())
5653 /* Don't touch page->lru of any random page, pre-check: */
5654 if (!page->mem_cgroup)
5657 INIT_LIST_HEAD(&page->lru);
5658 uncharge_list(&page->lru);
5662 * mem_cgroup_uncharge_list - uncharge a list of page
5663 * @page_list: list of pages to uncharge
5665 * Uncharge a list of pages previously charged with
5666 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5668 void mem_cgroup_uncharge_list(struct list_head *page_list)
5670 if (mem_cgroup_disabled())
5673 if (!list_empty(page_list))
5674 uncharge_list(page_list);
5678 * mem_cgroup_migrate - charge a page's replacement
5679 * @oldpage: currently circulating page
5680 * @newpage: replacement page
5682 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5683 * be uncharged upon free.
5685 * Both pages must be locked, @newpage->mapping must be set up.
5687 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5689 struct mem_cgroup *memcg;
5690 unsigned int nr_pages;
5692 unsigned long flags;
5694 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5695 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5696 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5697 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5700 if (mem_cgroup_disabled())
5703 /* Page cache replacement: new page already charged? */
5704 if (newpage->mem_cgroup)
5707 /* Swapcache readahead pages can get replaced before being charged */
5708 memcg = oldpage->mem_cgroup;
5712 /* Force-charge the new page. The old one will be freed soon */
5713 compound = PageTransHuge(newpage);
5714 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5716 page_counter_charge(&memcg->memory, nr_pages);
5717 if (do_memsw_account())
5718 page_counter_charge(&memcg->memsw, nr_pages);
5719 css_get_many(&memcg->css, nr_pages);
5721 commit_charge(newpage, memcg, false);
5723 local_irq_save(flags);
5724 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5725 memcg_check_events(memcg, newpage);
5726 local_irq_restore(flags);
5729 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5730 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5732 void mem_cgroup_sk_alloc(struct sock *sk)
5734 struct mem_cgroup *memcg;
5736 if (!mem_cgroup_sockets_enabled)
5740 * Socket cloning can throw us here with sk_memcg already
5741 * filled. It won't however, necessarily happen from
5742 * process context. So the test for root memcg given
5743 * the current task's memcg won't help us in this case.
5745 * Respecting the original socket's memcg is a better
5746 * decision in this case.
5749 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5750 css_get(&sk->sk_memcg->css);
5754 /* Do not associate the sock with unrelated interrupted task's memcg. */
5759 memcg = mem_cgroup_from_task(current);
5760 if (memcg == root_mem_cgroup)
5762 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5764 if (css_tryget_online(&memcg->css))
5765 sk->sk_memcg = memcg;
5770 void mem_cgroup_sk_free(struct sock *sk)
5773 css_put(&sk->sk_memcg->css);
5777 * mem_cgroup_charge_skmem - charge socket memory
5778 * @memcg: memcg to charge
5779 * @nr_pages: number of pages to charge
5781 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5782 * @memcg's configured limit, %false if the charge had to be forced.
5784 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5786 gfp_t gfp_mask = GFP_KERNEL;
5788 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5789 struct page_counter *fail;
5791 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5792 memcg->tcpmem_pressure = 0;
5795 page_counter_charge(&memcg->tcpmem, nr_pages);
5796 memcg->tcpmem_pressure = 1;
5800 /* Don't block in the packet receive path */
5802 gfp_mask = GFP_NOWAIT;
5804 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5806 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5809 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5814 * mem_cgroup_uncharge_skmem - uncharge socket memory
5815 * @memcg - memcg to uncharge
5816 * @nr_pages - number of pages to uncharge
5818 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5820 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5821 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5825 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5827 page_counter_uncharge(&memcg->memory, nr_pages);
5828 css_put_many(&memcg->css, nr_pages);
5831 static int __init cgroup_memory(char *s)
5835 while ((token = strsep(&s, ",")) != NULL) {
5838 if (!strcmp(token, "nosocket"))
5839 cgroup_memory_nosocket = true;
5840 if (!strcmp(token, "nokmem"))
5841 cgroup_memory_nokmem = true;
5845 __setup("cgroup.memory=", cgroup_memory);
5848 * subsys_initcall() for memory controller.
5850 * Some parts like hotcpu_notifier() have to be initialized from this context
5851 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5852 * everything that doesn't depend on a specific mem_cgroup structure should
5853 * be initialized from here.
5855 static int __init mem_cgroup_init(void)
5861 * Kmem cache creation is mostly done with the slab_mutex held,
5862 * so use a special workqueue to avoid stalling all worker
5863 * threads in case lots of cgroups are created simultaneously.
5865 memcg_kmem_cache_create_wq =
5866 alloc_ordered_workqueue("memcg_kmem_cache_create", 0);
5867 BUG_ON(!memcg_kmem_cache_create_wq);
5870 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5872 for_each_possible_cpu(cpu)
5873 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5876 for_each_node(node) {
5877 struct mem_cgroup_tree_per_node *rtpn;
5879 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5880 node_online(node) ? node : NUMA_NO_NODE);
5882 rtpn->rb_root = RB_ROOT;
5883 spin_lock_init(&rtpn->lock);
5884 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5889 subsys_initcall(mem_cgroup_init);
5891 #ifdef CONFIG_MEMCG_SWAP
5892 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5894 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5896 * The root cgroup cannot be destroyed, so it's refcount must
5899 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5903 memcg = parent_mem_cgroup(memcg);
5905 memcg = root_mem_cgroup;
5911 * mem_cgroup_swapout - transfer a memsw charge to swap
5912 * @page: page whose memsw charge to transfer
5913 * @entry: swap entry to move the charge to
5915 * Transfer the memsw charge of @page to @entry.
5917 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5919 struct mem_cgroup *memcg, *swap_memcg;
5920 unsigned short oldid;
5922 VM_BUG_ON_PAGE(PageLRU(page), page);
5923 VM_BUG_ON_PAGE(page_count(page), page);
5925 if (!do_memsw_account())
5928 memcg = page->mem_cgroup;
5930 /* Readahead page, never charged */
5935 * In case the memcg owning these pages has been offlined and doesn't
5936 * have an ID allocated to it anymore, charge the closest online
5937 * ancestor for the swap instead and transfer the memory+swap charge.
5939 swap_memcg = mem_cgroup_id_get_online(memcg);
5940 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5941 VM_BUG_ON_PAGE(oldid, page);
5942 mem_cgroup_swap_statistics(swap_memcg, true);
5944 page->mem_cgroup = NULL;
5946 if (!mem_cgroup_is_root(memcg))
5947 page_counter_uncharge(&memcg->memory, 1);
5949 if (memcg != swap_memcg) {
5950 if (!mem_cgroup_is_root(swap_memcg))
5951 page_counter_charge(&swap_memcg->memsw, 1);
5952 page_counter_uncharge(&memcg->memsw, 1);
5956 * Interrupts should be disabled here because the caller holds the
5957 * mapping->tree_lock lock which is taken with interrupts-off. It is
5958 * important here to have the interrupts disabled because it is the
5959 * only synchronisation we have for udpating the per-CPU variables.
5961 VM_BUG_ON(!irqs_disabled());
5962 mem_cgroup_charge_statistics(memcg, page, false, -1);
5963 memcg_check_events(memcg, page);
5965 if (!mem_cgroup_is_root(memcg))
5966 css_put(&memcg->css);
5970 * mem_cgroup_try_charge_swap - try charging a swap entry
5971 * @page: page being added to swap
5972 * @entry: swap entry to charge
5974 * Try to charge @entry to the memcg that @page belongs to.
5976 * Returns 0 on success, -ENOMEM on failure.
5978 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5980 struct mem_cgroup *memcg;
5981 struct page_counter *counter;
5982 unsigned short oldid;
5984 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5987 memcg = page->mem_cgroup;
5989 /* Readahead page, never charged */
5993 memcg = mem_cgroup_id_get_online(memcg);
5995 if (!mem_cgroup_is_root(memcg) &&
5996 !page_counter_try_charge(&memcg->swap, 1, &counter)) {
5997 mem_cgroup_id_put(memcg);
6001 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
6002 VM_BUG_ON_PAGE(oldid, page);
6003 mem_cgroup_swap_statistics(memcg, true);
6009 * mem_cgroup_uncharge_swap - uncharge a swap entry
6010 * @entry: swap entry to uncharge
6012 * Drop the swap charge associated with @entry.
6014 void mem_cgroup_uncharge_swap(swp_entry_t entry)
6016 struct mem_cgroup *memcg;
6019 if (!do_swap_account)
6022 id = swap_cgroup_record(entry, 0);
6024 memcg = mem_cgroup_from_id(id);
6026 if (!mem_cgroup_is_root(memcg)) {
6027 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6028 page_counter_uncharge(&memcg->swap, 1);
6030 page_counter_uncharge(&memcg->memsw, 1);
6032 mem_cgroup_swap_statistics(memcg, false);
6033 mem_cgroup_id_put(memcg);
6038 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6040 long nr_swap_pages = get_nr_swap_pages();
6042 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6043 return nr_swap_pages;
6044 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6045 nr_swap_pages = min_t(long, nr_swap_pages,
6046 READ_ONCE(memcg->swap.limit) -
6047 page_counter_read(&memcg->swap));
6048 return nr_swap_pages;
6051 bool mem_cgroup_swap_full(struct page *page)
6053 struct mem_cgroup *memcg;
6055 VM_BUG_ON_PAGE(!PageLocked(page), page);
6059 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6062 memcg = page->mem_cgroup;
6066 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6067 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
6073 /* for remember boot option*/
6074 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6075 static int really_do_swap_account __initdata = 1;
6077 static int really_do_swap_account __initdata;
6080 static int __init enable_swap_account(char *s)
6082 if (!strcmp(s, "1"))
6083 really_do_swap_account = 1;
6084 else if (!strcmp(s, "0"))
6085 really_do_swap_account = 0;
6088 __setup("swapaccount=", enable_swap_account);
6090 static u64 swap_current_read(struct cgroup_subsys_state *css,
6093 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6095 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6098 static int swap_max_show(struct seq_file *m, void *v)
6100 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6101 unsigned long max = READ_ONCE(memcg->swap.limit);
6103 if (max == PAGE_COUNTER_MAX)
6104 seq_puts(m, "max\n");
6106 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6111 static ssize_t swap_max_write(struct kernfs_open_file *of,
6112 char *buf, size_t nbytes, loff_t off)
6114 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6118 buf = strstrip(buf);
6119 err = page_counter_memparse(buf, "max", &max);
6123 mutex_lock(&memcg_limit_mutex);
6124 err = page_counter_limit(&memcg->swap, max);
6125 mutex_unlock(&memcg_limit_mutex);
6132 static struct cftype swap_files[] = {
6134 .name = "swap.current",
6135 .flags = CFTYPE_NOT_ON_ROOT,
6136 .read_u64 = swap_current_read,
6140 .flags = CFTYPE_NOT_ON_ROOT,
6141 .seq_show = swap_max_show,
6142 .write = swap_max_write,
6147 static struct cftype memsw_cgroup_files[] = {
6149 .name = "memsw.usage_in_bytes",
6150 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6151 .read_u64 = mem_cgroup_read_u64,
6154 .name = "memsw.max_usage_in_bytes",
6155 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6156 .write = mem_cgroup_reset,
6157 .read_u64 = mem_cgroup_read_u64,
6160 .name = "memsw.limit_in_bytes",
6161 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6162 .write = mem_cgroup_write,
6163 .read_u64 = mem_cgroup_read_u64,
6166 .name = "memsw.failcnt",
6167 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6168 .write = mem_cgroup_reset,
6169 .read_u64 = mem_cgroup_read_u64,
6171 { }, /* terminate */
6174 static int __init mem_cgroup_swap_init(void)
6176 if (!mem_cgroup_disabled() && really_do_swap_account) {
6177 do_swap_account = 1;
6178 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6180 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6181 memsw_cgroup_files));
6185 subsys_initcall(mem_cgroup_swap_init);
6187 #endif /* CONFIG_MEMCG_SWAP */