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_online(&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 *dead_memcg)
892 struct mem_cgroup *memcg = dead_memcg;
893 struct mem_cgroup_reclaim_iter *iter;
894 struct mem_cgroup_per_node *mz;
898 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
900 mz = mem_cgroup_nodeinfo(memcg, nid);
901 for (i = 0; i <= DEF_PRIORITY; i++) {
903 cmpxchg(&iter->position,
911 * Iteration constructs for visiting all cgroups (under a tree). If
912 * loops are exited prematurely (break), mem_cgroup_iter_break() must
913 * be used for reference counting.
915 #define for_each_mem_cgroup_tree(iter, root) \
916 for (iter = mem_cgroup_iter(root, NULL, NULL); \
918 iter = mem_cgroup_iter(root, iter, NULL))
920 #define for_each_mem_cgroup(iter) \
921 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
923 iter = mem_cgroup_iter(NULL, iter, NULL))
926 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
927 * @memcg: hierarchy root
928 * @fn: function to call for each task
929 * @arg: argument passed to @fn
931 * This function iterates over tasks attached to @memcg or to any of its
932 * descendants and calls @fn for each task. If @fn returns a non-zero
933 * value, the function breaks the iteration loop and returns the value.
934 * Otherwise, it will iterate over all tasks and return 0.
936 * This function must not be called for the root memory cgroup.
938 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
939 int (*fn)(struct task_struct *, void *), void *arg)
941 struct mem_cgroup *iter;
944 BUG_ON(memcg == root_mem_cgroup);
946 for_each_mem_cgroup_tree(iter, memcg) {
947 struct css_task_iter it;
948 struct task_struct *task;
950 css_task_iter_start(&iter->css, &it);
951 while (!ret && (task = css_task_iter_next(&it)))
953 css_task_iter_end(&it);
955 mem_cgroup_iter_break(memcg, iter);
963 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
965 * @zone: zone of the page
967 * This function is only safe when following the LRU page isolation
968 * and putback protocol: the LRU lock must be held, and the page must
969 * either be PageLRU() or the caller must have isolated/allocated it.
971 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
973 struct mem_cgroup_per_node *mz;
974 struct mem_cgroup *memcg;
975 struct lruvec *lruvec;
977 if (mem_cgroup_disabled()) {
978 lruvec = &pgdat->lruvec;
982 memcg = page->mem_cgroup;
984 * Swapcache readahead pages are added to the LRU - and
985 * possibly migrated - before they are charged.
988 memcg = root_mem_cgroup;
990 mz = mem_cgroup_page_nodeinfo(memcg, page);
991 lruvec = &mz->lruvec;
994 * Since a node can be onlined after the mem_cgroup was created,
995 * we have to be prepared to initialize lruvec->zone here;
996 * and if offlined then reonlined, we need to reinitialize it.
998 if (unlikely(lruvec->pgdat != pgdat))
999 lruvec->pgdat = pgdat;
1004 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1005 * @lruvec: mem_cgroup per zone lru vector
1006 * @lru: index of lru list the page is sitting on
1007 * @zid: zone id of the accounted pages
1008 * @nr_pages: positive when adding or negative when removing
1010 * This function must be called under lru_lock, just before a page is added
1011 * to or just after a page is removed from an lru list (that ordering being
1012 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1014 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1015 int zid, int nr_pages)
1017 struct mem_cgroup_per_node *mz;
1018 unsigned long *lru_size;
1021 if (mem_cgroup_disabled())
1024 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1025 lru_size = &mz->lru_zone_size[zid][lru];
1028 *lru_size += nr_pages;
1031 if (WARN_ONCE(size < 0,
1032 "%s(%p, %d, %d): lru_size %ld\n",
1033 __func__, lruvec, lru, nr_pages, size)) {
1039 *lru_size += nr_pages;
1042 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1044 struct mem_cgroup *task_memcg;
1045 struct task_struct *p;
1048 p = find_lock_task_mm(task);
1050 task_memcg = get_mem_cgroup_from_mm(p->mm);
1054 * All threads may have already detached their mm's, but the oom
1055 * killer still needs to detect if they have already been oom
1056 * killed to prevent needlessly killing additional tasks.
1059 task_memcg = mem_cgroup_from_task(task);
1060 css_get(&task_memcg->css);
1063 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1064 css_put(&task_memcg->css);
1069 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1070 * @memcg: the memory cgroup
1072 * Returns the maximum amount of memory @mem can be charged with, in
1075 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1077 unsigned long margin = 0;
1078 unsigned long count;
1079 unsigned long limit;
1081 count = page_counter_read(&memcg->memory);
1082 limit = READ_ONCE(memcg->memory.limit);
1084 margin = limit - count;
1086 if (do_memsw_account()) {
1087 count = page_counter_read(&memcg->memsw);
1088 limit = READ_ONCE(memcg->memsw.limit);
1090 margin = min(margin, limit - count);
1099 * A routine for checking "mem" is under move_account() or not.
1101 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1102 * moving cgroups. This is for waiting at high-memory pressure
1105 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1107 struct mem_cgroup *from;
1108 struct mem_cgroup *to;
1111 * Unlike task_move routines, we access mc.to, mc.from not under
1112 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1114 spin_lock(&mc.lock);
1120 ret = mem_cgroup_is_descendant(from, memcg) ||
1121 mem_cgroup_is_descendant(to, memcg);
1123 spin_unlock(&mc.lock);
1127 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1129 if (mc.moving_task && current != mc.moving_task) {
1130 if (mem_cgroup_under_move(memcg)) {
1132 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1133 /* moving charge context might have finished. */
1136 finish_wait(&mc.waitq, &wait);
1143 #define K(x) ((x) << (PAGE_SHIFT-10))
1145 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1146 * @memcg: The memory cgroup that went over limit
1147 * @p: Task that is going to be killed
1149 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1152 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1154 struct mem_cgroup *iter;
1160 pr_info("Task in ");
1161 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1162 pr_cont(" killed as a result of limit of ");
1164 pr_info("Memory limit reached of cgroup ");
1167 pr_cont_cgroup_path(memcg->css.cgroup);
1172 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1173 K((u64)page_counter_read(&memcg->memory)),
1174 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1175 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1176 K((u64)page_counter_read(&memcg->memsw)),
1177 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1178 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1179 K((u64)page_counter_read(&memcg->kmem)),
1180 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1182 for_each_mem_cgroup_tree(iter, memcg) {
1183 pr_info("Memory cgroup stats for ");
1184 pr_cont_cgroup_path(iter->css.cgroup);
1187 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1188 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1190 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1191 K(mem_cgroup_read_stat(iter, i)));
1194 for (i = 0; i < NR_LRU_LISTS; i++)
1195 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1196 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1203 * This function returns the number of memcg under hierarchy tree. Returns
1204 * 1(self count) if no children.
1206 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1209 struct mem_cgroup *iter;
1211 for_each_mem_cgroup_tree(iter, memcg)
1217 * Return the memory (and swap, if configured) limit for a memcg.
1219 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1221 unsigned long limit;
1223 limit = memcg->memory.limit;
1224 if (mem_cgroup_swappiness(memcg)) {
1225 unsigned long memsw_limit;
1226 unsigned long swap_limit;
1228 memsw_limit = memcg->memsw.limit;
1229 swap_limit = memcg->swap.limit;
1230 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1231 limit = min(limit + swap_limit, memsw_limit);
1236 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1239 struct oom_control oc = {
1243 .gfp_mask = gfp_mask,
1248 mutex_lock(&oom_lock);
1249 ret = out_of_memory(&oc);
1250 mutex_unlock(&oom_lock);
1254 #if MAX_NUMNODES > 1
1257 * test_mem_cgroup_node_reclaimable
1258 * @memcg: the target memcg
1259 * @nid: the node ID to be checked.
1260 * @noswap : specify true here if the user wants flle only information.
1262 * This function returns whether the specified memcg contains any
1263 * reclaimable pages on a node. Returns true if there are any reclaimable
1264 * pages in the node.
1266 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1267 int nid, bool noswap)
1269 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1271 if (noswap || !total_swap_pages)
1273 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1280 * Always updating the nodemask is not very good - even if we have an empty
1281 * list or the wrong list here, we can start from some node and traverse all
1282 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1285 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1289 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1290 * pagein/pageout changes since the last update.
1292 if (!atomic_read(&memcg->numainfo_events))
1294 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1297 /* make a nodemask where this memcg uses memory from */
1298 memcg->scan_nodes = node_states[N_MEMORY];
1300 for_each_node_mask(nid, node_states[N_MEMORY]) {
1302 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1303 node_clear(nid, memcg->scan_nodes);
1306 atomic_set(&memcg->numainfo_events, 0);
1307 atomic_set(&memcg->numainfo_updating, 0);
1311 * Selecting a node where we start reclaim from. Because what we need is just
1312 * reducing usage counter, start from anywhere is O,K. Considering
1313 * memory reclaim from current node, there are pros. and cons.
1315 * Freeing memory from current node means freeing memory from a node which
1316 * we'll use or we've used. So, it may make LRU bad. And if several threads
1317 * hit limits, it will see a contention on a node. But freeing from remote
1318 * node means more costs for memory reclaim because of memory latency.
1320 * Now, we use round-robin. Better algorithm is welcomed.
1322 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1326 mem_cgroup_may_update_nodemask(memcg);
1327 node = memcg->last_scanned_node;
1329 node = next_node_in(node, memcg->scan_nodes);
1331 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1332 * last time it really checked all the LRUs due to rate limiting.
1333 * Fallback to the current node in that case for simplicity.
1335 if (unlikely(node == MAX_NUMNODES))
1336 node = numa_node_id();
1338 memcg->last_scanned_node = node;
1342 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1348 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1351 unsigned long *total_scanned)
1353 struct mem_cgroup *victim = NULL;
1356 unsigned long excess;
1357 unsigned long nr_scanned;
1358 struct mem_cgroup_reclaim_cookie reclaim = {
1363 excess = soft_limit_excess(root_memcg);
1366 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1371 * If we have not been able to reclaim
1372 * anything, it might because there are
1373 * no reclaimable pages under this hierarchy
1378 * We want to do more targeted reclaim.
1379 * excess >> 2 is not to excessive so as to
1380 * reclaim too much, nor too less that we keep
1381 * coming back to reclaim from this cgroup
1383 if (total >= (excess >> 2) ||
1384 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1389 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1390 pgdat, &nr_scanned);
1391 *total_scanned += nr_scanned;
1392 if (!soft_limit_excess(root_memcg))
1395 mem_cgroup_iter_break(root_memcg, victim);
1399 #ifdef CONFIG_LOCKDEP
1400 static struct lockdep_map memcg_oom_lock_dep_map = {
1401 .name = "memcg_oom_lock",
1405 static DEFINE_SPINLOCK(memcg_oom_lock);
1408 * Check OOM-Killer is already running under our hierarchy.
1409 * If someone is running, return false.
1411 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1413 struct mem_cgroup *iter, *failed = NULL;
1415 spin_lock(&memcg_oom_lock);
1417 for_each_mem_cgroup_tree(iter, memcg) {
1418 if (iter->oom_lock) {
1420 * this subtree of our hierarchy is already locked
1421 * so we cannot give a lock.
1424 mem_cgroup_iter_break(memcg, iter);
1427 iter->oom_lock = true;
1432 * OK, we failed to lock the whole subtree so we have
1433 * to clean up what we set up to the failing subtree
1435 for_each_mem_cgroup_tree(iter, memcg) {
1436 if (iter == failed) {
1437 mem_cgroup_iter_break(memcg, iter);
1440 iter->oom_lock = false;
1443 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1445 spin_unlock(&memcg_oom_lock);
1450 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1452 struct mem_cgroup *iter;
1454 spin_lock(&memcg_oom_lock);
1455 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1456 for_each_mem_cgroup_tree(iter, memcg)
1457 iter->oom_lock = false;
1458 spin_unlock(&memcg_oom_lock);
1461 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1463 struct mem_cgroup *iter;
1465 spin_lock(&memcg_oom_lock);
1466 for_each_mem_cgroup_tree(iter, memcg)
1468 spin_unlock(&memcg_oom_lock);
1471 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1473 struct mem_cgroup *iter;
1476 * When a new child is created while the hierarchy is under oom,
1477 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1479 spin_lock(&memcg_oom_lock);
1480 for_each_mem_cgroup_tree(iter, memcg)
1481 if (iter->under_oom > 0)
1483 spin_unlock(&memcg_oom_lock);
1486 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1488 struct oom_wait_info {
1489 struct mem_cgroup *memcg;
1493 static int memcg_oom_wake_function(wait_queue_t *wait,
1494 unsigned mode, int sync, void *arg)
1496 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1497 struct mem_cgroup *oom_wait_memcg;
1498 struct oom_wait_info *oom_wait_info;
1500 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1501 oom_wait_memcg = oom_wait_info->memcg;
1503 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1504 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1506 return autoremove_wake_function(wait, mode, sync, arg);
1509 static void memcg_oom_recover(struct mem_cgroup *memcg)
1512 * For the following lockless ->under_oom test, the only required
1513 * guarantee is that it must see the state asserted by an OOM when
1514 * this function is called as a result of userland actions
1515 * triggered by the notification of the OOM. This is trivially
1516 * achieved by invoking mem_cgroup_mark_under_oom() before
1517 * triggering notification.
1519 if (memcg && memcg->under_oom)
1520 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1523 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1525 if (!current->memcg_may_oom)
1528 * We are in the middle of the charge context here, so we
1529 * don't want to block when potentially sitting on a callstack
1530 * that holds all kinds of filesystem and mm locks.
1532 * Also, the caller may handle a failed allocation gracefully
1533 * (like optional page cache readahead) and so an OOM killer
1534 * invocation might not even be necessary.
1536 * That's why we don't do anything here except remember the
1537 * OOM context and then deal with it at the end of the page
1538 * fault when the stack is unwound, the locks are released,
1539 * and when we know whether the fault was overall successful.
1541 css_get(&memcg->css);
1542 current->memcg_in_oom = memcg;
1543 current->memcg_oom_gfp_mask = mask;
1544 current->memcg_oom_order = order;
1548 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1549 * @handle: actually kill/wait or just clean up the OOM state
1551 * This has to be called at the end of a page fault if the memcg OOM
1552 * handler was enabled.
1554 * Memcg supports userspace OOM handling where failed allocations must
1555 * sleep on a waitqueue until the userspace task resolves the
1556 * situation. Sleeping directly in the charge context with all kinds
1557 * of locks held is not a good idea, instead we remember an OOM state
1558 * in the task and mem_cgroup_oom_synchronize() has to be called at
1559 * the end of the page fault to complete the OOM handling.
1561 * Returns %true if an ongoing memcg OOM situation was detected and
1562 * completed, %false otherwise.
1564 bool mem_cgroup_oom_synchronize(bool handle)
1566 struct mem_cgroup *memcg = current->memcg_in_oom;
1567 struct oom_wait_info owait;
1570 /* OOM is global, do not handle */
1577 owait.memcg = memcg;
1578 owait.wait.flags = 0;
1579 owait.wait.func = memcg_oom_wake_function;
1580 owait.wait.private = current;
1581 INIT_LIST_HEAD(&owait.wait.task_list);
1583 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1584 mem_cgroup_mark_under_oom(memcg);
1586 locked = mem_cgroup_oom_trylock(memcg);
1589 mem_cgroup_oom_notify(memcg);
1591 if (locked && !memcg->oom_kill_disable) {
1592 mem_cgroup_unmark_under_oom(memcg);
1593 finish_wait(&memcg_oom_waitq, &owait.wait);
1594 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1595 current->memcg_oom_order);
1598 mem_cgroup_unmark_under_oom(memcg);
1599 finish_wait(&memcg_oom_waitq, &owait.wait);
1603 mem_cgroup_oom_unlock(memcg);
1605 * There is no guarantee that an OOM-lock contender
1606 * sees the wakeups triggered by the OOM kill
1607 * uncharges. Wake any sleepers explicitely.
1609 memcg_oom_recover(memcg);
1612 current->memcg_in_oom = NULL;
1613 css_put(&memcg->css);
1618 * lock_page_memcg - lock a page->mem_cgroup binding
1621 * This function protects unlocked LRU pages from being moved to
1622 * another cgroup and stabilizes their page->mem_cgroup binding.
1624 void lock_page_memcg(struct page *page)
1626 struct mem_cgroup *memcg;
1627 unsigned long flags;
1630 * The RCU lock is held throughout the transaction. The fast
1631 * path can get away without acquiring the memcg->move_lock
1632 * because page moving starts with an RCU grace period.
1636 if (mem_cgroup_disabled())
1639 memcg = page->mem_cgroup;
1640 if (unlikely(!memcg))
1643 if (atomic_read(&memcg->moving_account) <= 0)
1646 spin_lock_irqsave(&memcg->move_lock, flags);
1647 if (memcg != page->mem_cgroup) {
1648 spin_unlock_irqrestore(&memcg->move_lock, flags);
1653 * When charge migration first begins, we can have locked and
1654 * unlocked page stat updates happening concurrently. Track
1655 * the task who has the lock for unlock_page_memcg().
1657 memcg->move_lock_task = current;
1658 memcg->move_lock_flags = flags;
1662 EXPORT_SYMBOL(lock_page_memcg);
1665 * unlock_page_memcg - unlock a page->mem_cgroup binding
1668 void unlock_page_memcg(struct page *page)
1670 struct mem_cgroup *memcg = page->mem_cgroup;
1672 if (memcg && memcg->move_lock_task == current) {
1673 unsigned long flags = memcg->move_lock_flags;
1675 memcg->move_lock_task = NULL;
1676 memcg->move_lock_flags = 0;
1678 spin_unlock_irqrestore(&memcg->move_lock, flags);
1683 EXPORT_SYMBOL(unlock_page_memcg);
1686 * size of first charge trial. "32" comes from vmscan.c's magic value.
1687 * TODO: maybe necessary to use big numbers in big irons.
1689 #define CHARGE_BATCH 32U
1690 struct memcg_stock_pcp {
1691 struct mem_cgroup *cached; /* this never be root cgroup */
1692 unsigned int nr_pages;
1693 struct work_struct work;
1694 unsigned long flags;
1695 #define FLUSHING_CACHED_CHARGE 0
1697 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1698 static DEFINE_MUTEX(percpu_charge_mutex);
1701 * consume_stock: Try to consume stocked charge on this cpu.
1702 * @memcg: memcg to consume from.
1703 * @nr_pages: how many pages to charge.
1705 * The charges will only happen if @memcg matches the current cpu's memcg
1706 * stock, and at least @nr_pages are available in that stock. Failure to
1707 * service an allocation will refill the stock.
1709 * returns true if successful, false otherwise.
1711 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1713 struct memcg_stock_pcp *stock;
1714 unsigned long flags;
1717 if (nr_pages > CHARGE_BATCH)
1720 local_irq_save(flags);
1722 stock = this_cpu_ptr(&memcg_stock);
1723 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1724 stock->nr_pages -= nr_pages;
1728 local_irq_restore(flags);
1734 * Returns stocks cached in percpu and reset cached information.
1736 static void drain_stock(struct memcg_stock_pcp *stock)
1738 struct mem_cgroup *old = stock->cached;
1740 if (stock->nr_pages) {
1741 page_counter_uncharge(&old->memory, stock->nr_pages);
1742 if (do_memsw_account())
1743 page_counter_uncharge(&old->memsw, stock->nr_pages);
1744 css_put_many(&old->css, stock->nr_pages);
1745 stock->nr_pages = 0;
1747 stock->cached = NULL;
1750 static void drain_local_stock(struct work_struct *dummy)
1752 struct memcg_stock_pcp *stock;
1753 unsigned long flags;
1755 local_irq_save(flags);
1757 stock = this_cpu_ptr(&memcg_stock);
1759 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1761 local_irq_restore(flags);
1765 * Cache charges(val) to local per_cpu area.
1766 * This will be consumed by consume_stock() function, later.
1768 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1770 struct memcg_stock_pcp *stock;
1771 unsigned long flags;
1773 local_irq_save(flags);
1775 stock = this_cpu_ptr(&memcg_stock);
1776 if (stock->cached != memcg) { /* reset if necessary */
1778 stock->cached = memcg;
1780 stock->nr_pages += nr_pages;
1782 local_irq_restore(flags);
1786 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1787 * of the hierarchy under it.
1789 static void drain_all_stock(struct mem_cgroup *root_memcg)
1793 /* If someone's already draining, avoid adding running more workers. */
1794 if (!mutex_trylock(&percpu_charge_mutex))
1796 /* Notify other cpus that system-wide "drain" is running */
1799 for_each_online_cpu(cpu) {
1800 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1801 struct mem_cgroup *memcg;
1803 memcg = stock->cached;
1804 if (!memcg || !stock->nr_pages)
1806 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1808 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1810 drain_local_stock(&stock->work);
1812 schedule_work_on(cpu, &stock->work);
1817 mutex_unlock(&percpu_charge_mutex);
1820 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1821 unsigned long action,
1824 int cpu = (unsigned long)hcpu;
1825 struct memcg_stock_pcp *stock;
1827 if (action == CPU_ONLINE)
1830 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1833 stock = &per_cpu(memcg_stock, cpu);
1838 static void reclaim_high(struct mem_cgroup *memcg,
1839 unsigned int nr_pages,
1843 if (page_counter_read(&memcg->memory) <= memcg->high)
1845 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1846 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1847 } while ((memcg = parent_mem_cgroup(memcg)));
1850 static void high_work_func(struct work_struct *work)
1852 struct mem_cgroup *memcg;
1854 memcg = container_of(work, struct mem_cgroup, high_work);
1855 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1859 * Scheduled by try_charge() to be executed from the userland return path
1860 * and reclaims memory over the high limit.
1862 void mem_cgroup_handle_over_high(void)
1864 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1865 struct mem_cgroup *memcg;
1867 if (likely(!nr_pages))
1870 memcg = get_mem_cgroup_from_mm(current->mm);
1871 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1872 css_put(&memcg->css);
1873 current->memcg_nr_pages_over_high = 0;
1876 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1877 unsigned int nr_pages)
1879 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1880 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1881 struct mem_cgroup *mem_over_limit;
1882 struct page_counter *counter;
1883 unsigned long nr_reclaimed;
1884 bool may_swap = true;
1885 bool drained = false;
1887 if (mem_cgroup_is_root(memcg))
1890 if (consume_stock(memcg, nr_pages))
1893 if (!do_memsw_account() ||
1894 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1895 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1897 if (do_memsw_account())
1898 page_counter_uncharge(&memcg->memsw, batch);
1899 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1901 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1905 if (batch > nr_pages) {
1911 * Unlike in global OOM situations, memcg is not in a physical
1912 * memory shortage. Allow dying and OOM-killed tasks to
1913 * bypass the last charges so that they can exit quickly and
1914 * free their memory.
1916 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1917 fatal_signal_pending(current) ||
1918 current->flags & PF_EXITING))
1922 * Prevent unbounded recursion when reclaim operations need to
1923 * allocate memory. This might exceed the limits temporarily,
1924 * but we prefer facilitating memory reclaim and getting back
1925 * under the limit over triggering OOM kills in these cases.
1927 if (unlikely(current->flags & PF_MEMALLOC))
1930 if (unlikely(task_in_memcg_oom(current)))
1933 if (!gfpflags_allow_blocking(gfp_mask))
1936 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1938 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1939 gfp_mask, may_swap);
1941 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1945 drain_all_stock(mem_over_limit);
1950 if (gfp_mask & __GFP_NORETRY)
1953 * Even though the limit is exceeded at this point, reclaim
1954 * may have been able to free some pages. Retry the charge
1955 * before killing the task.
1957 * Only for regular pages, though: huge pages are rather
1958 * unlikely to succeed so close to the limit, and we fall back
1959 * to regular pages anyway in case of failure.
1961 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1964 * At task move, charge accounts can be doubly counted. So, it's
1965 * better to wait until the end of task_move if something is going on.
1967 if (mem_cgroup_wait_acct_move(mem_over_limit))
1973 if (gfp_mask & __GFP_NOFAIL)
1976 if (fatal_signal_pending(current))
1979 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
1981 mem_cgroup_oom(mem_over_limit, gfp_mask,
1982 get_order(nr_pages * PAGE_SIZE));
1984 if (!(gfp_mask & __GFP_NOFAIL))
1988 * The allocation either can't fail or will lead to more memory
1989 * being freed very soon. Allow memory usage go over the limit
1990 * temporarily by force charging it.
1992 page_counter_charge(&memcg->memory, nr_pages);
1993 if (do_memsw_account())
1994 page_counter_charge(&memcg->memsw, nr_pages);
1995 css_get_many(&memcg->css, nr_pages);
2000 css_get_many(&memcg->css, batch);
2001 if (batch > nr_pages)
2002 refill_stock(memcg, batch - nr_pages);
2005 * If the hierarchy is above the normal consumption range, schedule
2006 * reclaim on returning to userland. We can perform reclaim here
2007 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2008 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2009 * not recorded as it most likely matches current's and won't
2010 * change in the meantime. As high limit is checked again before
2011 * reclaim, the cost of mismatch is negligible.
2014 if (page_counter_read(&memcg->memory) > memcg->high) {
2015 /* Don't bother a random interrupted task */
2016 if (in_interrupt()) {
2017 schedule_work(&memcg->high_work);
2020 current->memcg_nr_pages_over_high += batch;
2021 set_notify_resume(current);
2024 } while ((memcg = parent_mem_cgroup(memcg)));
2029 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2031 if (mem_cgroup_is_root(memcg))
2034 page_counter_uncharge(&memcg->memory, nr_pages);
2035 if (do_memsw_account())
2036 page_counter_uncharge(&memcg->memsw, nr_pages);
2038 css_put_many(&memcg->css, nr_pages);
2041 static void lock_page_lru(struct page *page, int *isolated)
2043 struct zone *zone = page_zone(page);
2045 spin_lock_irq(zone_lru_lock(zone));
2046 if (PageLRU(page)) {
2047 struct lruvec *lruvec;
2049 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2051 del_page_from_lru_list(page, lruvec, page_lru(page));
2057 static void unlock_page_lru(struct page *page, int isolated)
2059 struct zone *zone = page_zone(page);
2062 struct lruvec *lruvec;
2064 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2065 VM_BUG_ON_PAGE(PageLRU(page), page);
2067 add_page_to_lru_list(page, lruvec, page_lru(page));
2069 spin_unlock_irq(zone_lru_lock(zone));
2072 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2077 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2080 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2081 * may already be on some other mem_cgroup's LRU. Take care of it.
2084 lock_page_lru(page, &isolated);
2087 * Nobody should be changing or seriously looking at
2088 * page->mem_cgroup at this point:
2090 * - the page is uncharged
2092 * - the page is off-LRU
2094 * - an anonymous fault has exclusive page access, except for
2095 * a locked page table
2097 * - a page cache insertion, a swapin fault, or a migration
2098 * have the page locked
2100 page->mem_cgroup = memcg;
2103 unlock_page_lru(page, isolated);
2107 static int memcg_alloc_cache_id(void)
2112 id = ida_simple_get(&memcg_cache_ida,
2113 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2117 if (id < memcg_nr_cache_ids)
2121 * There's no space for the new id in memcg_caches arrays,
2122 * so we have to grow them.
2124 down_write(&memcg_cache_ids_sem);
2126 size = 2 * (id + 1);
2127 if (size < MEMCG_CACHES_MIN_SIZE)
2128 size = MEMCG_CACHES_MIN_SIZE;
2129 else if (size > MEMCG_CACHES_MAX_SIZE)
2130 size = MEMCG_CACHES_MAX_SIZE;
2132 err = memcg_update_all_caches(size);
2134 err = memcg_update_all_list_lrus(size);
2136 memcg_nr_cache_ids = size;
2138 up_write(&memcg_cache_ids_sem);
2141 ida_simple_remove(&memcg_cache_ida, id);
2147 static void memcg_free_cache_id(int id)
2149 ida_simple_remove(&memcg_cache_ida, id);
2152 struct memcg_kmem_cache_create_work {
2153 struct mem_cgroup *memcg;
2154 struct kmem_cache *cachep;
2155 struct work_struct work;
2158 static struct workqueue_struct *memcg_kmem_cache_create_wq;
2160 static void memcg_kmem_cache_create_func(struct work_struct *w)
2162 struct memcg_kmem_cache_create_work *cw =
2163 container_of(w, struct memcg_kmem_cache_create_work, work);
2164 struct mem_cgroup *memcg = cw->memcg;
2165 struct kmem_cache *cachep = cw->cachep;
2167 memcg_create_kmem_cache(memcg, cachep);
2169 css_put(&memcg->css);
2174 * Enqueue the creation of a per-memcg kmem_cache.
2176 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2177 struct kmem_cache *cachep)
2179 struct memcg_kmem_cache_create_work *cw;
2181 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2185 css_get(&memcg->css);
2188 cw->cachep = cachep;
2189 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2191 queue_work(memcg_kmem_cache_create_wq, &cw->work);
2194 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2195 struct kmem_cache *cachep)
2198 * We need to stop accounting when we kmalloc, because if the
2199 * corresponding kmalloc cache is not yet created, the first allocation
2200 * in __memcg_schedule_kmem_cache_create will recurse.
2202 * However, it is better to enclose the whole function. Depending on
2203 * the debugging options enabled, INIT_WORK(), for instance, can
2204 * trigger an allocation. This too, will make us recurse. Because at
2205 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2206 * the safest choice is to do it like this, wrapping the whole function.
2208 current->memcg_kmem_skip_account = 1;
2209 __memcg_schedule_kmem_cache_create(memcg, cachep);
2210 current->memcg_kmem_skip_account = 0;
2213 static inline bool memcg_kmem_bypass(void)
2215 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2221 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2222 * @cachep: the original global kmem cache
2224 * Return the kmem_cache we're supposed to use for a slab allocation.
2225 * We try to use the current memcg's version of the cache.
2227 * If the cache does not exist yet, if we are the first user of it, we
2228 * create it asynchronously in a workqueue and let the current allocation
2229 * go through with the original cache.
2231 * This function takes a reference to the cache it returns to assure it
2232 * won't get destroyed while we are working with it. Once the caller is
2233 * done with it, memcg_kmem_put_cache() must be called to release the
2236 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2238 struct mem_cgroup *memcg;
2239 struct kmem_cache *memcg_cachep;
2242 VM_BUG_ON(!is_root_cache(cachep));
2244 if (memcg_kmem_bypass())
2247 if (current->memcg_kmem_skip_account)
2250 memcg = get_mem_cgroup_from_mm(current->mm);
2251 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2255 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2256 if (likely(memcg_cachep))
2257 return memcg_cachep;
2260 * If we are in a safe context (can wait, and not in interrupt
2261 * context), we could be be predictable and return right away.
2262 * This would guarantee that the allocation being performed
2263 * already belongs in the new cache.
2265 * However, there are some clashes that can arrive from locking.
2266 * For instance, because we acquire the slab_mutex while doing
2267 * memcg_create_kmem_cache, this means no further allocation
2268 * could happen with the slab_mutex held. So it's better to
2271 memcg_schedule_kmem_cache_create(memcg, cachep);
2273 css_put(&memcg->css);
2278 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2279 * @cachep: the cache returned by memcg_kmem_get_cache
2281 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2283 if (!is_root_cache(cachep))
2284 css_put(&cachep->memcg_params.memcg->css);
2288 * memcg_kmem_charge: charge a kmem page
2289 * @page: page to charge
2290 * @gfp: reclaim mode
2291 * @order: allocation order
2292 * @memcg: memory cgroup to charge
2294 * Returns 0 on success, an error code on failure.
2296 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2297 struct mem_cgroup *memcg)
2299 unsigned int nr_pages = 1 << order;
2300 struct page_counter *counter;
2303 ret = try_charge(memcg, gfp, nr_pages);
2307 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2308 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2309 cancel_charge(memcg, nr_pages);
2313 page->mem_cgroup = memcg;
2319 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2320 * @page: page to charge
2321 * @gfp: reclaim mode
2322 * @order: allocation order
2324 * Returns 0 on success, an error code on failure.
2326 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2328 struct mem_cgroup *memcg;
2331 if (memcg_kmem_bypass())
2334 memcg = get_mem_cgroup_from_mm(current->mm);
2335 if (!mem_cgroup_is_root(memcg)) {
2336 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2338 __SetPageKmemcg(page);
2340 css_put(&memcg->css);
2344 * memcg_kmem_uncharge: uncharge a kmem page
2345 * @page: page to uncharge
2346 * @order: allocation order
2348 void memcg_kmem_uncharge(struct page *page, int order)
2350 struct mem_cgroup *memcg = page->mem_cgroup;
2351 unsigned int nr_pages = 1 << order;
2356 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2358 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2359 page_counter_uncharge(&memcg->kmem, nr_pages);
2361 page_counter_uncharge(&memcg->memory, nr_pages);
2362 if (do_memsw_account())
2363 page_counter_uncharge(&memcg->memsw, nr_pages);
2365 page->mem_cgroup = NULL;
2367 /* slab pages do not have PageKmemcg flag set */
2368 if (PageKmemcg(page))
2369 __ClearPageKmemcg(page);
2371 css_put_many(&memcg->css, nr_pages);
2373 #endif /* !CONFIG_SLOB */
2375 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2378 * Because tail pages are not marked as "used", set it. We're under
2379 * zone_lru_lock and migration entries setup in all page mappings.
2381 void mem_cgroup_split_huge_fixup(struct page *head)
2385 if (mem_cgroup_disabled())
2388 for (i = 1; i < HPAGE_PMD_NR; i++)
2389 head[i].mem_cgroup = head->mem_cgroup;
2391 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2394 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2396 #ifdef CONFIG_MEMCG_SWAP
2397 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2400 int val = (charge) ? 1 : -1;
2401 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2405 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2406 * @entry: swap entry to be moved
2407 * @from: mem_cgroup which the entry is moved from
2408 * @to: mem_cgroup which the entry is moved to
2410 * It succeeds only when the swap_cgroup's record for this entry is the same
2411 * as the mem_cgroup's id of @from.
2413 * Returns 0 on success, -EINVAL on failure.
2415 * The caller must have charged to @to, IOW, called page_counter_charge() about
2416 * both res and memsw, and called css_get().
2418 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2419 struct mem_cgroup *from, struct mem_cgroup *to)
2421 unsigned short old_id, new_id;
2423 old_id = mem_cgroup_id(from);
2424 new_id = mem_cgroup_id(to);
2426 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2427 mem_cgroup_swap_statistics(from, false);
2428 mem_cgroup_swap_statistics(to, true);
2434 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2435 struct mem_cgroup *from, struct mem_cgroup *to)
2441 static DEFINE_MUTEX(memcg_limit_mutex);
2443 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2444 unsigned long limit)
2446 unsigned long curusage;
2447 unsigned long oldusage;
2448 bool enlarge = false;
2453 * For keeping hierarchical_reclaim simple, how long we should retry
2454 * is depends on callers. We set our retry-count to be function
2455 * of # of children which we should visit in this loop.
2457 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2458 mem_cgroup_count_children(memcg);
2460 oldusage = page_counter_read(&memcg->memory);
2463 if (signal_pending(current)) {
2468 mutex_lock(&memcg_limit_mutex);
2469 if (limit > memcg->memsw.limit) {
2470 mutex_unlock(&memcg_limit_mutex);
2474 if (limit > memcg->memory.limit)
2476 ret = page_counter_limit(&memcg->memory, limit);
2477 mutex_unlock(&memcg_limit_mutex);
2482 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2484 curusage = page_counter_read(&memcg->memory);
2485 /* Usage is reduced ? */
2486 if (curusage >= oldusage)
2489 oldusage = curusage;
2490 } while (retry_count);
2492 if (!ret && enlarge)
2493 memcg_oom_recover(memcg);
2498 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2499 unsigned long limit)
2501 unsigned long curusage;
2502 unsigned long oldusage;
2503 bool enlarge = false;
2507 /* see mem_cgroup_resize_res_limit */
2508 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2509 mem_cgroup_count_children(memcg);
2511 oldusage = page_counter_read(&memcg->memsw);
2514 if (signal_pending(current)) {
2519 mutex_lock(&memcg_limit_mutex);
2520 if (limit < memcg->memory.limit) {
2521 mutex_unlock(&memcg_limit_mutex);
2525 if (limit > memcg->memsw.limit)
2527 ret = page_counter_limit(&memcg->memsw, limit);
2528 mutex_unlock(&memcg_limit_mutex);
2533 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2535 curusage = page_counter_read(&memcg->memsw);
2536 /* Usage is reduced ? */
2537 if (curusage >= oldusage)
2540 oldusage = curusage;
2541 } while (retry_count);
2543 if (!ret && enlarge)
2544 memcg_oom_recover(memcg);
2549 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2551 unsigned long *total_scanned)
2553 unsigned long nr_reclaimed = 0;
2554 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2555 unsigned long reclaimed;
2557 struct mem_cgroup_tree_per_node *mctz;
2558 unsigned long excess;
2559 unsigned long nr_scanned;
2564 mctz = soft_limit_tree_node(pgdat->node_id);
2567 * Do not even bother to check the largest node if the root
2568 * is empty. Do it lockless to prevent lock bouncing. Races
2569 * are acceptable as soft limit is best effort anyway.
2571 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2575 * This loop can run a while, specially if mem_cgroup's continuously
2576 * keep exceeding their soft limit and putting the system under
2583 mz = mem_cgroup_largest_soft_limit_node(mctz);
2588 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2589 gfp_mask, &nr_scanned);
2590 nr_reclaimed += reclaimed;
2591 *total_scanned += nr_scanned;
2592 spin_lock_irq(&mctz->lock);
2593 __mem_cgroup_remove_exceeded(mz, mctz);
2596 * If we failed to reclaim anything from this memory cgroup
2597 * it is time to move on to the next cgroup
2601 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2603 excess = soft_limit_excess(mz->memcg);
2605 * One school of thought says that we should not add
2606 * back the node to the tree if reclaim returns 0.
2607 * But our reclaim could return 0, simply because due
2608 * to priority we are exposing a smaller subset of
2609 * memory to reclaim from. Consider this as a longer
2612 /* If excess == 0, no tree ops */
2613 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2614 spin_unlock_irq(&mctz->lock);
2615 css_put(&mz->memcg->css);
2618 * Could not reclaim anything and there are no more
2619 * mem cgroups to try or we seem to be looping without
2620 * reclaiming anything.
2622 if (!nr_reclaimed &&
2624 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2626 } while (!nr_reclaimed);
2628 css_put(&next_mz->memcg->css);
2629 return nr_reclaimed;
2633 * Test whether @memcg has children, dead or alive. Note that this
2634 * function doesn't care whether @memcg has use_hierarchy enabled and
2635 * returns %true if there are child csses according to the cgroup
2636 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2638 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2643 ret = css_next_child(NULL, &memcg->css);
2649 * Reclaims as many pages from the given memcg as possible.
2651 * Caller is responsible for holding css reference for memcg.
2653 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2655 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2657 /* we call try-to-free pages for make this cgroup empty */
2658 lru_add_drain_all();
2659 /* try to free all pages in this cgroup */
2660 while (nr_retries && page_counter_read(&memcg->memory)) {
2663 if (signal_pending(current))
2666 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2670 /* maybe some writeback is necessary */
2671 congestion_wait(BLK_RW_ASYNC, HZ/10);
2679 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2680 char *buf, size_t nbytes,
2683 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2685 if (mem_cgroup_is_root(memcg))
2687 return mem_cgroup_force_empty(memcg) ?: nbytes;
2690 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2693 return mem_cgroup_from_css(css)->use_hierarchy;
2696 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2697 struct cftype *cft, u64 val)
2700 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2701 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2703 if (memcg->use_hierarchy == val)
2707 * If parent's use_hierarchy is set, we can't make any modifications
2708 * in the child subtrees. If it is unset, then the change can
2709 * occur, provided the current cgroup has no children.
2711 * For the root cgroup, parent_mem is NULL, we allow value to be
2712 * set if there are no children.
2714 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2715 (val == 1 || val == 0)) {
2716 if (!memcg_has_children(memcg))
2717 memcg->use_hierarchy = val;
2726 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2728 struct mem_cgroup *iter;
2731 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2733 for_each_mem_cgroup_tree(iter, memcg) {
2734 for (i = 0; i < MEMCG_NR_STAT; i++)
2735 stat[i] += mem_cgroup_read_stat(iter, i);
2739 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2741 struct mem_cgroup *iter;
2744 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2746 for_each_mem_cgroup_tree(iter, memcg) {
2747 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2748 events[i] += mem_cgroup_read_events(iter, i);
2752 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2754 unsigned long val = 0;
2756 if (mem_cgroup_is_root(memcg)) {
2757 struct mem_cgroup *iter;
2759 for_each_mem_cgroup_tree(iter, memcg) {
2760 val += mem_cgroup_read_stat(iter,
2761 MEM_CGROUP_STAT_CACHE);
2762 val += mem_cgroup_read_stat(iter,
2763 MEM_CGROUP_STAT_RSS);
2765 val += mem_cgroup_read_stat(iter,
2766 MEM_CGROUP_STAT_SWAP);
2770 val = page_counter_read(&memcg->memory);
2772 val = page_counter_read(&memcg->memsw);
2785 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2788 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2789 struct page_counter *counter;
2791 switch (MEMFILE_TYPE(cft->private)) {
2793 counter = &memcg->memory;
2796 counter = &memcg->memsw;
2799 counter = &memcg->kmem;
2802 counter = &memcg->tcpmem;
2808 switch (MEMFILE_ATTR(cft->private)) {
2810 if (counter == &memcg->memory)
2811 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2812 if (counter == &memcg->memsw)
2813 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2814 return (u64)page_counter_read(counter) * PAGE_SIZE;
2816 return (u64)counter->limit * PAGE_SIZE;
2818 return (u64)counter->watermark * PAGE_SIZE;
2820 return counter->failcnt;
2821 case RES_SOFT_LIMIT:
2822 return (u64)memcg->soft_limit * PAGE_SIZE;
2829 static int memcg_online_kmem(struct mem_cgroup *memcg)
2833 if (cgroup_memory_nokmem)
2836 BUG_ON(memcg->kmemcg_id >= 0);
2837 BUG_ON(memcg->kmem_state);
2839 memcg_id = memcg_alloc_cache_id();
2843 static_branch_inc(&memcg_kmem_enabled_key);
2845 * A memory cgroup is considered kmem-online as soon as it gets
2846 * kmemcg_id. Setting the id after enabling static branching will
2847 * guarantee no one starts accounting before all call sites are
2850 memcg->kmemcg_id = memcg_id;
2851 memcg->kmem_state = KMEM_ONLINE;
2856 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2858 struct cgroup_subsys_state *css;
2859 struct mem_cgroup *parent, *child;
2862 if (memcg->kmem_state != KMEM_ONLINE)
2865 * Clear the online state before clearing memcg_caches array
2866 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2867 * guarantees that no cache will be created for this cgroup
2868 * after we are done (see memcg_create_kmem_cache()).
2870 memcg->kmem_state = KMEM_ALLOCATED;
2872 memcg_deactivate_kmem_caches(memcg);
2874 kmemcg_id = memcg->kmemcg_id;
2875 BUG_ON(kmemcg_id < 0);
2877 parent = parent_mem_cgroup(memcg);
2879 parent = root_mem_cgroup;
2882 * Change kmemcg_id of this cgroup and all its descendants to the
2883 * parent's id, and then move all entries from this cgroup's list_lrus
2884 * to ones of the parent. After we have finished, all list_lrus
2885 * corresponding to this cgroup are guaranteed to remain empty. The
2886 * ordering is imposed by list_lru_node->lock taken by
2887 * memcg_drain_all_list_lrus().
2889 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2890 css_for_each_descendant_pre(css, &memcg->css) {
2891 child = mem_cgroup_from_css(css);
2892 BUG_ON(child->kmemcg_id != kmemcg_id);
2893 child->kmemcg_id = parent->kmemcg_id;
2894 if (!memcg->use_hierarchy)
2899 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2901 memcg_free_cache_id(kmemcg_id);
2904 static void memcg_free_kmem(struct mem_cgroup *memcg)
2906 /* css_alloc() failed, offlining didn't happen */
2907 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2908 memcg_offline_kmem(memcg);
2910 if (memcg->kmem_state == KMEM_ALLOCATED) {
2911 memcg_destroy_kmem_caches(memcg);
2912 static_branch_dec(&memcg_kmem_enabled_key);
2913 WARN_ON(page_counter_read(&memcg->kmem));
2917 static int memcg_online_kmem(struct mem_cgroup *memcg)
2921 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2924 static void memcg_free_kmem(struct mem_cgroup *memcg)
2927 #endif /* !CONFIG_SLOB */
2929 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2930 unsigned long limit)
2934 mutex_lock(&memcg_limit_mutex);
2935 ret = page_counter_limit(&memcg->kmem, limit);
2936 mutex_unlock(&memcg_limit_mutex);
2940 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2944 mutex_lock(&memcg_limit_mutex);
2946 ret = page_counter_limit(&memcg->tcpmem, limit);
2950 if (!memcg->tcpmem_active) {
2952 * The active flag needs to be written after the static_key
2953 * update. This is what guarantees that the socket activation
2954 * function is the last one to run. See mem_cgroup_sk_alloc()
2955 * for details, and note that we don't mark any socket as
2956 * belonging to this memcg until that flag is up.
2958 * We need to do this, because static_keys will span multiple
2959 * sites, but we can't control their order. If we mark a socket
2960 * as accounted, but the accounting functions are not patched in
2961 * yet, we'll lose accounting.
2963 * We never race with the readers in mem_cgroup_sk_alloc(),
2964 * because when this value change, the code to process it is not
2967 static_branch_inc(&memcg_sockets_enabled_key);
2968 memcg->tcpmem_active = true;
2971 mutex_unlock(&memcg_limit_mutex);
2976 * The user of this function is...
2979 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2980 char *buf, size_t nbytes, loff_t off)
2982 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2983 unsigned long nr_pages;
2986 buf = strstrip(buf);
2987 ret = page_counter_memparse(buf, "-1", &nr_pages);
2991 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2993 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2997 switch (MEMFILE_TYPE(of_cft(of)->private)) {
2999 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3002 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3005 ret = memcg_update_kmem_limit(memcg, nr_pages);
3008 ret = memcg_update_tcp_limit(memcg, nr_pages);
3012 case RES_SOFT_LIMIT:
3013 memcg->soft_limit = nr_pages;
3017 return ret ?: nbytes;
3020 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3021 size_t nbytes, loff_t off)
3023 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3024 struct page_counter *counter;
3026 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3028 counter = &memcg->memory;
3031 counter = &memcg->memsw;
3034 counter = &memcg->kmem;
3037 counter = &memcg->tcpmem;
3043 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3045 page_counter_reset_watermark(counter);
3048 counter->failcnt = 0;
3057 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3060 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3064 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3065 struct cftype *cft, u64 val)
3067 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3069 if (val & ~MOVE_MASK)
3073 * No kind of locking is needed in here, because ->can_attach() will
3074 * check this value once in the beginning of the process, and then carry
3075 * on with stale data. This means that changes to this value will only
3076 * affect task migrations starting after the change.
3078 memcg->move_charge_at_immigrate = val;
3082 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3083 struct cftype *cft, u64 val)
3090 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3094 unsigned int lru_mask;
3097 static const struct numa_stat stats[] = {
3098 { "total", LRU_ALL },
3099 { "file", LRU_ALL_FILE },
3100 { "anon", LRU_ALL_ANON },
3101 { "unevictable", BIT(LRU_UNEVICTABLE) },
3103 const struct numa_stat *stat;
3106 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3108 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3109 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3110 seq_printf(m, "%s=%lu", stat->name, nr);
3111 for_each_node_state(nid, N_MEMORY) {
3112 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3114 seq_printf(m, " N%d=%lu", nid, nr);
3119 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3120 struct mem_cgroup *iter;
3123 for_each_mem_cgroup_tree(iter, memcg)
3124 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3125 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3126 for_each_node_state(nid, N_MEMORY) {
3128 for_each_mem_cgroup_tree(iter, memcg)
3129 nr += mem_cgroup_node_nr_lru_pages(
3130 iter, nid, stat->lru_mask);
3131 seq_printf(m, " N%d=%lu", nid, nr);
3138 #endif /* CONFIG_NUMA */
3140 static int memcg_stat_show(struct seq_file *m, void *v)
3142 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3143 unsigned long memory, memsw;
3144 struct mem_cgroup *mi;
3147 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3148 MEM_CGROUP_STAT_NSTATS);
3149 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3150 MEM_CGROUP_EVENTS_NSTATS);
3151 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3153 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3154 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3156 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3157 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3160 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3161 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3162 mem_cgroup_read_events(memcg, i));
3164 for (i = 0; i < NR_LRU_LISTS; i++)
3165 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3166 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3168 /* Hierarchical information */
3169 memory = memsw = PAGE_COUNTER_MAX;
3170 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3171 memory = min(memory, mi->memory.limit);
3172 memsw = min(memsw, mi->memsw.limit);
3174 seq_printf(m, "hierarchical_memory_limit %llu\n",
3175 (u64)memory * PAGE_SIZE);
3176 if (do_memsw_account())
3177 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3178 (u64)memsw * PAGE_SIZE);
3180 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3181 unsigned long long val = 0;
3183 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3185 for_each_mem_cgroup_tree(mi, memcg)
3186 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3187 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3190 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3191 unsigned long long val = 0;
3193 for_each_mem_cgroup_tree(mi, memcg)
3194 val += mem_cgroup_read_events(mi, i);
3195 seq_printf(m, "total_%s %llu\n",
3196 mem_cgroup_events_names[i], val);
3199 for (i = 0; i < NR_LRU_LISTS; i++) {
3200 unsigned long long val = 0;
3202 for_each_mem_cgroup_tree(mi, memcg)
3203 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3204 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3207 #ifdef CONFIG_DEBUG_VM
3210 struct mem_cgroup_per_node *mz;
3211 struct zone_reclaim_stat *rstat;
3212 unsigned long recent_rotated[2] = {0, 0};
3213 unsigned long recent_scanned[2] = {0, 0};
3215 for_each_online_pgdat(pgdat) {
3216 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3217 rstat = &mz->lruvec.reclaim_stat;
3219 recent_rotated[0] += rstat->recent_rotated[0];
3220 recent_rotated[1] += rstat->recent_rotated[1];
3221 recent_scanned[0] += rstat->recent_scanned[0];
3222 recent_scanned[1] += rstat->recent_scanned[1];
3224 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3225 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3226 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3227 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3234 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3237 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3239 return mem_cgroup_swappiness(memcg);
3242 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3243 struct cftype *cft, u64 val)
3245 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3251 memcg->swappiness = val;
3253 vm_swappiness = val;
3258 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3260 struct mem_cgroup_threshold_ary *t;
3261 unsigned long usage;
3266 t = rcu_dereference(memcg->thresholds.primary);
3268 t = rcu_dereference(memcg->memsw_thresholds.primary);
3273 usage = mem_cgroup_usage(memcg, swap);
3276 * current_threshold points to threshold just below or equal to usage.
3277 * If it's not true, a threshold was crossed after last
3278 * call of __mem_cgroup_threshold().
3280 i = t->current_threshold;
3283 * Iterate backward over array of thresholds starting from
3284 * current_threshold and check if a threshold is crossed.
3285 * If none of thresholds below usage is crossed, we read
3286 * only one element of the array here.
3288 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3289 eventfd_signal(t->entries[i].eventfd, 1);
3291 /* i = current_threshold + 1 */
3295 * Iterate forward over array of thresholds starting from
3296 * current_threshold+1 and check if a threshold is crossed.
3297 * If none of thresholds above usage is crossed, we read
3298 * only one element of the array here.
3300 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3301 eventfd_signal(t->entries[i].eventfd, 1);
3303 /* Update current_threshold */
3304 t->current_threshold = i - 1;
3309 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3312 __mem_cgroup_threshold(memcg, false);
3313 if (do_memsw_account())
3314 __mem_cgroup_threshold(memcg, true);
3316 memcg = parent_mem_cgroup(memcg);
3320 static int compare_thresholds(const void *a, const void *b)
3322 const struct mem_cgroup_threshold *_a = a;
3323 const struct mem_cgroup_threshold *_b = b;
3325 if (_a->threshold > _b->threshold)
3328 if (_a->threshold < _b->threshold)
3334 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3336 struct mem_cgroup_eventfd_list *ev;
3338 spin_lock(&memcg_oom_lock);
3340 list_for_each_entry(ev, &memcg->oom_notify, list)
3341 eventfd_signal(ev->eventfd, 1);
3343 spin_unlock(&memcg_oom_lock);
3347 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3349 struct mem_cgroup *iter;
3351 for_each_mem_cgroup_tree(iter, memcg)
3352 mem_cgroup_oom_notify_cb(iter);
3355 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3356 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3358 struct mem_cgroup_thresholds *thresholds;
3359 struct mem_cgroup_threshold_ary *new;
3360 unsigned long threshold;
3361 unsigned long usage;
3364 ret = page_counter_memparse(args, "-1", &threshold);
3368 mutex_lock(&memcg->thresholds_lock);
3371 thresholds = &memcg->thresholds;
3372 usage = mem_cgroup_usage(memcg, false);
3373 } else if (type == _MEMSWAP) {
3374 thresholds = &memcg->memsw_thresholds;
3375 usage = mem_cgroup_usage(memcg, true);
3379 /* Check if a threshold crossed before adding a new one */
3380 if (thresholds->primary)
3381 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3383 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3385 /* Allocate memory for new array of thresholds */
3386 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3394 /* Copy thresholds (if any) to new array */
3395 if (thresholds->primary) {
3396 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3397 sizeof(struct mem_cgroup_threshold));
3400 /* Add new threshold */
3401 new->entries[size - 1].eventfd = eventfd;
3402 new->entries[size - 1].threshold = threshold;
3404 /* Sort thresholds. Registering of new threshold isn't time-critical */
3405 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3406 compare_thresholds, NULL);
3408 /* Find current threshold */
3409 new->current_threshold = -1;
3410 for (i = 0; i < size; i++) {
3411 if (new->entries[i].threshold <= usage) {
3413 * new->current_threshold will not be used until
3414 * rcu_assign_pointer(), so it's safe to increment
3417 ++new->current_threshold;
3422 /* Free old spare buffer and save old primary buffer as spare */
3423 kfree(thresholds->spare);
3424 thresholds->spare = thresholds->primary;
3426 rcu_assign_pointer(thresholds->primary, new);
3428 /* To be sure that nobody uses thresholds */
3432 mutex_unlock(&memcg->thresholds_lock);
3437 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3438 struct eventfd_ctx *eventfd, const char *args)
3440 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3443 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3444 struct eventfd_ctx *eventfd, const char *args)
3446 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3449 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3450 struct eventfd_ctx *eventfd, enum res_type type)
3452 struct mem_cgroup_thresholds *thresholds;
3453 struct mem_cgroup_threshold_ary *new;
3454 unsigned long usage;
3457 mutex_lock(&memcg->thresholds_lock);
3460 thresholds = &memcg->thresholds;
3461 usage = mem_cgroup_usage(memcg, false);
3462 } else if (type == _MEMSWAP) {
3463 thresholds = &memcg->memsw_thresholds;
3464 usage = mem_cgroup_usage(memcg, true);
3468 if (!thresholds->primary)
3471 /* Check if a threshold crossed before removing */
3472 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3474 /* Calculate new number of threshold */
3476 for (i = 0; i < thresholds->primary->size; i++) {
3477 if (thresholds->primary->entries[i].eventfd != eventfd)
3481 new = thresholds->spare;
3483 /* Set thresholds array to NULL if we don't have thresholds */
3492 /* Copy thresholds and find current threshold */
3493 new->current_threshold = -1;
3494 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3495 if (thresholds->primary->entries[i].eventfd == eventfd)
3498 new->entries[j] = thresholds->primary->entries[i];
3499 if (new->entries[j].threshold <= usage) {
3501 * new->current_threshold will not be used
3502 * until rcu_assign_pointer(), so it's safe to increment
3505 ++new->current_threshold;
3511 /* Swap primary and spare array */
3512 thresholds->spare = thresholds->primary;
3514 rcu_assign_pointer(thresholds->primary, new);
3516 /* To be sure that nobody uses thresholds */
3519 /* If all events are unregistered, free the spare array */
3521 kfree(thresholds->spare);
3522 thresholds->spare = NULL;
3525 mutex_unlock(&memcg->thresholds_lock);
3528 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3529 struct eventfd_ctx *eventfd)
3531 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3534 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3535 struct eventfd_ctx *eventfd)
3537 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3540 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3541 struct eventfd_ctx *eventfd, const char *args)
3543 struct mem_cgroup_eventfd_list *event;
3545 event = kmalloc(sizeof(*event), GFP_KERNEL);
3549 spin_lock(&memcg_oom_lock);
3551 event->eventfd = eventfd;
3552 list_add(&event->list, &memcg->oom_notify);
3554 /* already in OOM ? */
3555 if (memcg->under_oom)
3556 eventfd_signal(eventfd, 1);
3557 spin_unlock(&memcg_oom_lock);
3562 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3563 struct eventfd_ctx *eventfd)
3565 struct mem_cgroup_eventfd_list *ev, *tmp;
3567 spin_lock(&memcg_oom_lock);
3569 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3570 if (ev->eventfd == eventfd) {
3571 list_del(&ev->list);
3576 spin_unlock(&memcg_oom_lock);
3579 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3581 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3583 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3584 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3588 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3589 struct cftype *cft, u64 val)
3591 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3593 /* cannot set to root cgroup and only 0 and 1 are allowed */
3594 if (!css->parent || !((val == 0) || (val == 1)))
3597 memcg->oom_kill_disable = val;
3599 memcg_oom_recover(memcg);
3604 #ifdef CONFIG_CGROUP_WRITEBACK
3606 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3608 return &memcg->cgwb_list;
3611 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3613 return wb_domain_init(&memcg->cgwb_domain, gfp);
3616 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3618 wb_domain_exit(&memcg->cgwb_domain);
3621 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3623 wb_domain_size_changed(&memcg->cgwb_domain);
3626 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3628 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3630 if (!memcg->css.parent)
3633 return &memcg->cgwb_domain;
3637 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3638 * @wb: bdi_writeback in question
3639 * @pfilepages: out parameter for number of file pages
3640 * @pheadroom: out parameter for number of allocatable pages according to memcg
3641 * @pdirty: out parameter for number of dirty pages
3642 * @pwriteback: out parameter for number of pages under writeback
3644 * Determine the numbers of file, headroom, dirty, and writeback pages in
3645 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3646 * is a bit more involved.
3648 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3649 * headroom is calculated as the lowest headroom of itself and the
3650 * ancestors. Note that this doesn't consider the actual amount of
3651 * available memory in the system. The caller should further cap
3652 * *@pheadroom accordingly.
3654 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3655 unsigned long *pheadroom, unsigned long *pdirty,
3656 unsigned long *pwriteback)
3658 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3659 struct mem_cgroup *parent;
3661 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3663 /* this should eventually include NR_UNSTABLE_NFS */
3664 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3665 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3666 (1 << LRU_ACTIVE_FILE));
3667 *pheadroom = PAGE_COUNTER_MAX;
3669 while ((parent = parent_mem_cgroup(memcg))) {
3670 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3671 unsigned long used = page_counter_read(&memcg->memory);
3673 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3678 #else /* CONFIG_CGROUP_WRITEBACK */
3680 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3685 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3689 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3693 #endif /* CONFIG_CGROUP_WRITEBACK */
3696 * DO NOT USE IN NEW FILES.
3698 * "cgroup.event_control" implementation.
3700 * This is way over-engineered. It tries to support fully configurable
3701 * events for each user. Such level of flexibility is completely
3702 * unnecessary especially in the light of the planned unified hierarchy.
3704 * Please deprecate this and replace with something simpler if at all
3709 * Unregister event and free resources.
3711 * Gets called from workqueue.
3713 static void memcg_event_remove(struct work_struct *work)
3715 struct mem_cgroup_event *event =
3716 container_of(work, struct mem_cgroup_event, remove);
3717 struct mem_cgroup *memcg = event->memcg;
3719 remove_wait_queue(event->wqh, &event->wait);
3721 event->unregister_event(memcg, event->eventfd);
3723 /* Notify userspace the event is going away. */
3724 eventfd_signal(event->eventfd, 1);
3726 eventfd_ctx_put(event->eventfd);
3728 css_put(&memcg->css);
3732 * Gets called on POLLHUP on eventfd when user closes it.
3734 * Called with wqh->lock held and interrupts disabled.
3736 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3737 int sync, void *key)
3739 struct mem_cgroup_event *event =
3740 container_of(wait, struct mem_cgroup_event, wait);
3741 struct mem_cgroup *memcg = event->memcg;
3742 unsigned long flags = (unsigned long)key;
3744 if (flags & POLLHUP) {
3746 * If the event has been detached at cgroup removal, we
3747 * can simply return knowing the other side will cleanup
3750 * We can't race against event freeing since the other
3751 * side will require wqh->lock via remove_wait_queue(),
3754 spin_lock(&memcg->event_list_lock);
3755 if (!list_empty(&event->list)) {
3756 list_del_init(&event->list);
3758 * We are in atomic context, but cgroup_event_remove()
3759 * may sleep, so we have to call it in workqueue.
3761 schedule_work(&event->remove);
3763 spin_unlock(&memcg->event_list_lock);
3769 static void memcg_event_ptable_queue_proc(struct file *file,
3770 wait_queue_head_t *wqh, poll_table *pt)
3772 struct mem_cgroup_event *event =
3773 container_of(pt, struct mem_cgroup_event, pt);
3776 add_wait_queue(wqh, &event->wait);
3780 * DO NOT USE IN NEW FILES.
3782 * Parse input and register new cgroup event handler.
3784 * Input must be in format '<event_fd> <control_fd> <args>'.
3785 * Interpretation of args is defined by control file implementation.
3787 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3788 char *buf, size_t nbytes, loff_t off)
3790 struct cgroup_subsys_state *css = of_css(of);
3791 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3792 struct mem_cgroup_event *event;
3793 struct cgroup_subsys_state *cfile_css;
3794 unsigned int efd, cfd;
3801 buf = strstrip(buf);
3803 efd = simple_strtoul(buf, &endp, 10);
3808 cfd = simple_strtoul(buf, &endp, 10);
3809 if ((*endp != ' ') && (*endp != '\0'))
3813 event = kzalloc(sizeof(*event), GFP_KERNEL);
3817 event->memcg = memcg;
3818 INIT_LIST_HEAD(&event->list);
3819 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3820 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3821 INIT_WORK(&event->remove, memcg_event_remove);
3829 event->eventfd = eventfd_ctx_fileget(efile.file);
3830 if (IS_ERR(event->eventfd)) {
3831 ret = PTR_ERR(event->eventfd);
3838 goto out_put_eventfd;
3841 /* the process need read permission on control file */
3842 /* AV: shouldn't we check that it's been opened for read instead? */
3843 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3848 * Determine the event callbacks and set them in @event. This used
3849 * to be done via struct cftype but cgroup core no longer knows
3850 * about these events. The following is crude but the whole thing
3851 * is for compatibility anyway.
3853 * DO NOT ADD NEW FILES.
3855 name = cfile.file->f_path.dentry->d_name.name;
3857 if (!strcmp(name, "memory.usage_in_bytes")) {
3858 event->register_event = mem_cgroup_usage_register_event;
3859 event->unregister_event = mem_cgroup_usage_unregister_event;
3860 } else if (!strcmp(name, "memory.oom_control")) {
3861 event->register_event = mem_cgroup_oom_register_event;
3862 event->unregister_event = mem_cgroup_oom_unregister_event;
3863 } else if (!strcmp(name, "memory.pressure_level")) {
3864 event->register_event = vmpressure_register_event;
3865 event->unregister_event = vmpressure_unregister_event;
3866 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3867 event->register_event = memsw_cgroup_usage_register_event;
3868 event->unregister_event = memsw_cgroup_usage_unregister_event;
3875 * Verify @cfile should belong to @css. Also, remaining events are
3876 * automatically removed on cgroup destruction but the removal is
3877 * asynchronous, so take an extra ref on @css.
3879 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3880 &memory_cgrp_subsys);
3882 if (IS_ERR(cfile_css))
3884 if (cfile_css != css) {
3889 ret = event->register_event(memcg, event->eventfd, buf);
3893 efile.file->f_op->poll(efile.file, &event->pt);
3895 spin_lock(&memcg->event_list_lock);
3896 list_add(&event->list, &memcg->event_list);
3897 spin_unlock(&memcg->event_list_lock);
3909 eventfd_ctx_put(event->eventfd);
3918 static struct cftype mem_cgroup_legacy_files[] = {
3920 .name = "usage_in_bytes",
3921 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3922 .read_u64 = mem_cgroup_read_u64,
3925 .name = "max_usage_in_bytes",
3926 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3927 .write = mem_cgroup_reset,
3928 .read_u64 = mem_cgroup_read_u64,
3931 .name = "limit_in_bytes",
3932 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3933 .write = mem_cgroup_write,
3934 .read_u64 = mem_cgroup_read_u64,
3937 .name = "soft_limit_in_bytes",
3938 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3939 .write = mem_cgroup_write,
3940 .read_u64 = mem_cgroup_read_u64,
3944 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3945 .write = mem_cgroup_reset,
3946 .read_u64 = mem_cgroup_read_u64,
3950 .seq_show = memcg_stat_show,
3953 .name = "force_empty",
3954 .write = mem_cgroup_force_empty_write,
3957 .name = "use_hierarchy",
3958 .write_u64 = mem_cgroup_hierarchy_write,
3959 .read_u64 = mem_cgroup_hierarchy_read,
3962 .name = "cgroup.event_control", /* XXX: for compat */
3963 .write = memcg_write_event_control,
3964 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3967 .name = "swappiness",
3968 .read_u64 = mem_cgroup_swappiness_read,
3969 .write_u64 = mem_cgroup_swappiness_write,
3972 .name = "move_charge_at_immigrate",
3973 .read_u64 = mem_cgroup_move_charge_read,
3974 .write_u64 = mem_cgroup_move_charge_write,
3977 .name = "oom_control",
3978 .seq_show = mem_cgroup_oom_control_read,
3979 .write_u64 = mem_cgroup_oom_control_write,
3980 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3983 .name = "pressure_level",
3987 .name = "numa_stat",
3988 .seq_show = memcg_numa_stat_show,
3992 .name = "kmem.limit_in_bytes",
3993 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
3994 .write = mem_cgroup_write,
3995 .read_u64 = mem_cgroup_read_u64,
3998 .name = "kmem.usage_in_bytes",
3999 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4000 .read_u64 = mem_cgroup_read_u64,
4003 .name = "kmem.failcnt",
4004 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4005 .write = mem_cgroup_reset,
4006 .read_u64 = mem_cgroup_read_u64,
4009 .name = "kmem.max_usage_in_bytes",
4010 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4011 .write = mem_cgroup_reset,
4012 .read_u64 = mem_cgroup_read_u64,
4014 #ifdef CONFIG_SLABINFO
4016 .name = "kmem.slabinfo",
4017 .seq_start = slab_start,
4018 .seq_next = slab_next,
4019 .seq_stop = slab_stop,
4020 .seq_show = memcg_slab_show,
4024 .name = "kmem.tcp.limit_in_bytes",
4025 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4026 .write = mem_cgroup_write,
4027 .read_u64 = mem_cgroup_read_u64,
4030 .name = "kmem.tcp.usage_in_bytes",
4031 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4032 .read_u64 = mem_cgroup_read_u64,
4035 .name = "kmem.tcp.failcnt",
4036 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4037 .write = mem_cgroup_reset,
4038 .read_u64 = mem_cgroup_read_u64,
4041 .name = "kmem.tcp.max_usage_in_bytes",
4042 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4043 .write = mem_cgroup_reset,
4044 .read_u64 = mem_cgroup_read_u64,
4046 { }, /* terminate */
4050 * Private memory cgroup IDR
4052 * Swap-out records and page cache shadow entries need to store memcg
4053 * references in constrained space, so we maintain an ID space that is
4054 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4055 * memory-controlled cgroups to 64k.
4057 * However, there usually are many references to the oflline CSS after
4058 * the cgroup has been destroyed, such as page cache or reclaimable
4059 * slab objects, that don't need to hang on to the ID. We want to keep
4060 * those dead CSS from occupying IDs, or we might quickly exhaust the
4061 * relatively small ID space and prevent the creation of new cgroups
4062 * even when there are much fewer than 64k cgroups - possibly none.
4064 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4065 * be freed and recycled when it's no longer needed, which is usually
4066 * when the CSS is offlined.
4068 * The only exception to that are records of swapped out tmpfs/shmem
4069 * pages that need to be attributed to live ancestors on swapin. But
4070 * those references are manageable from userspace.
4073 static DEFINE_IDR(mem_cgroup_idr);
4075 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4077 if (memcg->id.id > 0) {
4078 idr_remove(&mem_cgroup_idr, memcg->id.id);
4083 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4085 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4086 atomic_add(n, &memcg->id.ref);
4089 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4091 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4092 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4093 mem_cgroup_id_remove(memcg);
4095 /* Memcg ID pins CSS */
4096 css_put(&memcg->css);
4100 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4102 mem_cgroup_id_get_many(memcg, 1);
4105 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4107 mem_cgroup_id_put_many(memcg, 1);
4111 * mem_cgroup_from_id - look up a memcg from a memcg id
4112 * @id: the memcg id to look up
4114 * Caller must hold rcu_read_lock().
4116 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4118 WARN_ON_ONCE(!rcu_read_lock_held());
4119 return idr_find(&mem_cgroup_idr, id);
4122 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4124 struct mem_cgroup_per_node *pn;
4127 * This routine is called against possible nodes.
4128 * But it's BUG to call kmalloc() against offline node.
4130 * TODO: this routine can waste much memory for nodes which will
4131 * never be onlined. It's better to use memory hotplug callback
4134 if (!node_state(node, N_NORMAL_MEMORY))
4136 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4140 lruvec_init(&pn->lruvec);
4141 pn->usage_in_excess = 0;
4142 pn->on_tree = false;
4145 memcg->nodeinfo[node] = pn;
4149 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4151 kfree(memcg->nodeinfo[node]);
4154 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4159 free_mem_cgroup_per_node_info(memcg, node);
4160 free_percpu(memcg->stat);
4164 static void mem_cgroup_free(struct mem_cgroup *memcg)
4166 memcg_wb_domain_exit(memcg);
4167 __mem_cgroup_free(memcg);
4170 static struct mem_cgroup *mem_cgroup_alloc(void)
4172 struct mem_cgroup *memcg;
4176 size = sizeof(struct mem_cgroup);
4177 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4179 memcg = kzalloc(size, GFP_KERNEL);
4183 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4184 1, MEM_CGROUP_ID_MAX,
4186 if (memcg->id.id < 0)
4189 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4194 if (alloc_mem_cgroup_per_node_info(memcg, node))
4197 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4200 INIT_WORK(&memcg->high_work, high_work_func);
4201 memcg->last_scanned_node = MAX_NUMNODES;
4202 INIT_LIST_HEAD(&memcg->oom_notify);
4203 mutex_init(&memcg->thresholds_lock);
4204 spin_lock_init(&memcg->move_lock);
4205 vmpressure_init(&memcg->vmpressure);
4206 INIT_LIST_HEAD(&memcg->event_list);
4207 spin_lock_init(&memcg->event_list_lock);
4208 memcg->socket_pressure = jiffies;
4210 memcg->kmemcg_id = -1;
4212 #ifdef CONFIG_CGROUP_WRITEBACK
4213 INIT_LIST_HEAD(&memcg->cgwb_list);
4215 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4218 mem_cgroup_id_remove(memcg);
4219 __mem_cgroup_free(memcg);
4223 static struct cgroup_subsys_state * __ref
4224 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4226 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4227 struct mem_cgroup *memcg;
4228 long error = -ENOMEM;
4230 memcg = mem_cgroup_alloc();
4232 return ERR_PTR(error);
4234 memcg->high = PAGE_COUNTER_MAX;
4235 memcg->soft_limit = PAGE_COUNTER_MAX;
4237 memcg->swappiness = mem_cgroup_swappiness(parent);
4238 memcg->oom_kill_disable = parent->oom_kill_disable;
4240 if (parent && parent->use_hierarchy) {
4241 memcg->use_hierarchy = true;
4242 page_counter_init(&memcg->memory, &parent->memory);
4243 page_counter_init(&memcg->swap, &parent->swap);
4244 page_counter_init(&memcg->memsw, &parent->memsw);
4245 page_counter_init(&memcg->kmem, &parent->kmem);
4246 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4248 page_counter_init(&memcg->memory, NULL);
4249 page_counter_init(&memcg->swap, NULL);
4250 page_counter_init(&memcg->memsw, NULL);
4251 page_counter_init(&memcg->kmem, NULL);
4252 page_counter_init(&memcg->tcpmem, NULL);
4254 * Deeper hierachy with use_hierarchy == false doesn't make
4255 * much sense so let cgroup subsystem know about this
4256 * unfortunate state in our controller.
4258 if (parent != root_mem_cgroup)
4259 memory_cgrp_subsys.broken_hierarchy = true;
4262 /* The following stuff does not apply to the root */
4264 root_mem_cgroup = memcg;
4268 error = memcg_online_kmem(memcg);
4272 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4273 static_branch_inc(&memcg_sockets_enabled_key);
4277 mem_cgroup_id_remove(memcg);
4278 mem_cgroup_free(memcg);
4279 return ERR_PTR(-ENOMEM);
4282 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4284 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4286 /* Online state pins memcg ID, memcg ID pins CSS */
4287 atomic_set(&memcg->id.ref, 1);
4292 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4294 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4295 struct mem_cgroup_event *event, *tmp;
4298 * Unregister events and notify userspace.
4299 * Notify userspace about cgroup removing only after rmdir of cgroup
4300 * directory to avoid race between userspace and kernelspace.
4302 spin_lock(&memcg->event_list_lock);
4303 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4304 list_del_init(&event->list);
4305 schedule_work(&event->remove);
4307 spin_unlock(&memcg->event_list_lock);
4309 memcg_offline_kmem(memcg);
4310 wb_memcg_offline(memcg);
4312 mem_cgroup_id_put(memcg);
4315 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4317 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4319 invalidate_reclaim_iterators(memcg);
4322 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4324 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4326 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4327 static_branch_dec(&memcg_sockets_enabled_key);
4329 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4330 static_branch_dec(&memcg_sockets_enabled_key);
4332 vmpressure_cleanup(&memcg->vmpressure);
4333 cancel_work_sync(&memcg->high_work);
4334 mem_cgroup_remove_from_trees(memcg);
4335 memcg_free_kmem(memcg);
4336 mem_cgroup_free(memcg);
4340 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4341 * @css: the target css
4343 * Reset the states of the mem_cgroup associated with @css. This is
4344 * invoked when the userland requests disabling on the default hierarchy
4345 * but the memcg is pinned through dependency. The memcg should stop
4346 * applying policies and should revert to the vanilla state as it may be
4347 * made visible again.
4349 * The current implementation only resets the essential configurations.
4350 * This needs to be expanded to cover all the visible parts.
4352 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4354 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4356 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4357 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4358 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4359 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4360 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4362 memcg->high = PAGE_COUNTER_MAX;
4363 memcg->soft_limit = PAGE_COUNTER_MAX;
4364 memcg_wb_domain_size_changed(memcg);
4368 /* Handlers for move charge at task migration. */
4369 static int mem_cgroup_do_precharge(unsigned long count)
4373 /* Try a single bulk charge without reclaim first, kswapd may wake */
4374 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4376 mc.precharge += count;
4380 /* Try charges one by one with reclaim, but do not retry */
4382 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4396 enum mc_target_type {
4402 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4403 unsigned long addr, pte_t ptent)
4405 struct page *page = vm_normal_page(vma, addr, ptent);
4407 if (!page || !page_mapped(page))
4409 if (PageAnon(page)) {
4410 if (!(mc.flags & MOVE_ANON))
4413 if (!(mc.flags & MOVE_FILE))
4416 if (!get_page_unless_zero(page))
4423 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4424 pte_t ptent, swp_entry_t *entry)
4426 struct page *page = NULL;
4427 swp_entry_t ent = pte_to_swp_entry(ptent);
4429 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4432 * Because lookup_swap_cache() updates some statistics counter,
4433 * we call find_get_page() with swapper_space directly.
4435 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4436 if (do_memsw_account())
4437 entry->val = ent.val;
4442 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4443 pte_t ptent, swp_entry_t *entry)
4449 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4450 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4452 struct page *page = NULL;
4453 struct address_space *mapping;
4456 if (!vma->vm_file) /* anonymous vma */
4458 if (!(mc.flags & MOVE_FILE))
4461 mapping = vma->vm_file->f_mapping;
4462 pgoff = linear_page_index(vma, addr);
4464 /* page is moved even if it's not RSS of this task(page-faulted). */
4466 /* shmem/tmpfs may report page out on swap: account for that too. */
4467 if (shmem_mapping(mapping)) {
4468 page = find_get_entry(mapping, pgoff);
4469 if (radix_tree_exceptional_entry(page)) {
4470 swp_entry_t swp = radix_to_swp_entry(page);
4471 if (do_memsw_account())
4473 page = find_get_page(swap_address_space(swp),
4477 page = find_get_page(mapping, pgoff);
4479 page = find_get_page(mapping, pgoff);
4485 * mem_cgroup_move_account - move account of the page
4487 * @compound: charge the page as compound or small page
4488 * @from: mem_cgroup which the page is moved from.
4489 * @to: mem_cgroup which the page is moved to. @from != @to.
4491 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4493 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4496 static int mem_cgroup_move_account(struct page *page,
4498 struct mem_cgroup *from,
4499 struct mem_cgroup *to)
4501 unsigned long flags;
4502 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4506 VM_BUG_ON(from == to);
4507 VM_BUG_ON_PAGE(PageLRU(page), page);
4508 VM_BUG_ON(compound && !PageTransHuge(page));
4511 * Prevent mem_cgroup_migrate() from looking at
4512 * page->mem_cgroup of its source page while we change it.
4515 if (!trylock_page(page))
4519 if (page->mem_cgroup != from)
4522 anon = PageAnon(page);
4524 spin_lock_irqsave(&from->move_lock, flags);
4526 if (!anon && page_mapped(page)) {
4527 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4529 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4534 * move_lock grabbed above and caller set from->moving_account, so
4535 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4536 * So mapping should be stable for dirty pages.
4538 if (!anon && PageDirty(page)) {
4539 struct address_space *mapping = page_mapping(page);
4541 if (mapping_cap_account_dirty(mapping)) {
4542 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4544 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4549 if (PageWriteback(page)) {
4550 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4552 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4557 * It is safe to change page->mem_cgroup here because the page
4558 * is referenced, charged, and isolated - we can't race with
4559 * uncharging, charging, migration, or LRU putback.
4562 /* caller should have done css_get */
4563 page->mem_cgroup = to;
4564 spin_unlock_irqrestore(&from->move_lock, flags);
4568 local_irq_disable();
4569 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4570 memcg_check_events(to, page);
4571 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4572 memcg_check_events(from, page);
4581 * get_mctgt_type - get target type of moving charge
4582 * @vma: the vma the pte to be checked belongs
4583 * @addr: the address corresponding to the pte to be checked
4584 * @ptent: the pte to be checked
4585 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4588 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4589 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4590 * move charge. if @target is not NULL, the page is stored in target->page
4591 * with extra refcnt got(Callers should handle it).
4592 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4593 * target for charge migration. if @target is not NULL, the entry is stored
4596 * Called with pte lock held.
4599 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4600 unsigned long addr, pte_t ptent, union mc_target *target)
4602 struct page *page = NULL;
4603 enum mc_target_type ret = MC_TARGET_NONE;
4604 swp_entry_t ent = { .val = 0 };
4606 if (pte_present(ptent))
4607 page = mc_handle_present_pte(vma, addr, ptent);
4608 else if (is_swap_pte(ptent))
4609 page = mc_handle_swap_pte(vma, ptent, &ent);
4610 else if (pte_none(ptent))
4611 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4613 if (!page && !ent.val)
4617 * Do only loose check w/o serialization.
4618 * mem_cgroup_move_account() checks the page is valid or
4619 * not under LRU exclusion.
4621 if (page->mem_cgroup == mc.from) {
4622 ret = MC_TARGET_PAGE;
4624 target->page = page;
4626 if (!ret || !target)
4629 /* There is a swap entry and a page doesn't exist or isn't charged */
4630 if (ent.val && !ret &&
4631 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4632 ret = MC_TARGET_SWAP;
4639 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4641 * We don't consider swapping or file mapped pages because THP does not
4642 * support them for now.
4643 * Caller should make sure that pmd_trans_huge(pmd) is true.
4645 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4646 unsigned long addr, pmd_t pmd, union mc_target *target)
4648 struct page *page = NULL;
4649 enum mc_target_type ret = MC_TARGET_NONE;
4651 page = pmd_page(pmd);
4652 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4653 if (!(mc.flags & MOVE_ANON))
4655 if (page->mem_cgroup == mc.from) {
4656 ret = MC_TARGET_PAGE;
4659 target->page = page;
4665 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4666 unsigned long addr, pmd_t pmd, union mc_target *target)
4668 return MC_TARGET_NONE;
4672 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4673 unsigned long addr, unsigned long end,
4674 struct mm_walk *walk)
4676 struct vm_area_struct *vma = walk->vma;
4680 ptl = pmd_trans_huge_lock(pmd, vma);
4682 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4683 mc.precharge += HPAGE_PMD_NR;
4688 if (pmd_trans_unstable(pmd))
4690 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4691 for (; addr != end; pte++, addr += PAGE_SIZE)
4692 if (get_mctgt_type(vma, addr, *pte, NULL))
4693 mc.precharge++; /* increment precharge temporarily */
4694 pte_unmap_unlock(pte - 1, ptl);
4700 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4702 unsigned long precharge;
4704 struct mm_walk mem_cgroup_count_precharge_walk = {
4705 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4708 down_read(&mm->mmap_sem);
4709 walk_page_range(0, mm->highest_vm_end,
4710 &mem_cgroup_count_precharge_walk);
4711 up_read(&mm->mmap_sem);
4713 precharge = mc.precharge;
4719 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4721 unsigned long precharge = mem_cgroup_count_precharge(mm);
4723 VM_BUG_ON(mc.moving_task);
4724 mc.moving_task = current;
4725 return mem_cgroup_do_precharge(precharge);
4728 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4729 static void __mem_cgroup_clear_mc(void)
4731 struct mem_cgroup *from = mc.from;
4732 struct mem_cgroup *to = mc.to;
4734 /* we must uncharge all the leftover precharges from mc.to */
4736 cancel_charge(mc.to, mc.precharge);
4740 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4741 * we must uncharge here.
4743 if (mc.moved_charge) {
4744 cancel_charge(mc.from, mc.moved_charge);
4745 mc.moved_charge = 0;
4747 /* we must fixup refcnts and charges */
4748 if (mc.moved_swap) {
4749 /* uncharge swap account from the old cgroup */
4750 if (!mem_cgroup_is_root(mc.from))
4751 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4753 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4756 * we charged both to->memory and to->memsw, so we
4757 * should uncharge to->memory.
4759 if (!mem_cgroup_is_root(mc.to))
4760 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4762 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4763 css_put_many(&mc.to->css, mc.moved_swap);
4767 memcg_oom_recover(from);
4768 memcg_oom_recover(to);
4769 wake_up_all(&mc.waitq);
4772 static void mem_cgroup_clear_mc(void)
4774 struct mm_struct *mm = mc.mm;
4777 * we must clear moving_task before waking up waiters at the end of
4780 mc.moving_task = NULL;
4781 __mem_cgroup_clear_mc();
4782 spin_lock(&mc.lock);
4786 spin_unlock(&mc.lock);
4791 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4793 struct cgroup_subsys_state *css;
4794 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4795 struct mem_cgroup *from;
4796 struct task_struct *leader, *p;
4797 struct mm_struct *mm;
4798 unsigned long move_flags;
4801 /* charge immigration isn't supported on the default hierarchy */
4802 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4806 * Multi-process migrations only happen on the default hierarchy
4807 * where charge immigration is not used. Perform charge
4808 * immigration if @tset contains a leader and whine if there are
4812 cgroup_taskset_for_each_leader(leader, css, tset) {
4815 memcg = mem_cgroup_from_css(css);
4821 * We are now commited to this value whatever it is. Changes in this
4822 * tunable will only affect upcoming migrations, not the current one.
4823 * So we need to save it, and keep it going.
4825 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4829 from = mem_cgroup_from_task(p);
4831 VM_BUG_ON(from == memcg);
4833 mm = get_task_mm(p);
4836 /* We move charges only when we move a owner of the mm */
4837 if (mm->owner == p) {
4840 VM_BUG_ON(mc.precharge);
4841 VM_BUG_ON(mc.moved_charge);
4842 VM_BUG_ON(mc.moved_swap);
4844 spin_lock(&mc.lock);
4848 mc.flags = move_flags;
4849 spin_unlock(&mc.lock);
4850 /* We set mc.moving_task later */
4852 ret = mem_cgroup_precharge_mc(mm);
4854 mem_cgroup_clear_mc();
4861 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4864 mem_cgroup_clear_mc();
4867 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4868 unsigned long addr, unsigned long end,
4869 struct mm_walk *walk)
4872 struct vm_area_struct *vma = walk->vma;
4875 enum mc_target_type target_type;
4876 union mc_target target;
4879 ptl = pmd_trans_huge_lock(pmd, vma);
4881 if (mc.precharge < HPAGE_PMD_NR) {
4885 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4886 if (target_type == MC_TARGET_PAGE) {
4888 if (!isolate_lru_page(page)) {
4889 if (!mem_cgroup_move_account(page, true,
4891 mc.precharge -= HPAGE_PMD_NR;
4892 mc.moved_charge += HPAGE_PMD_NR;
4894 putback_lru_page(page);
4902 if (pmd_trans_unstable(pmd))
4905 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4906 for (; addr != end; addr += PAGE_SIZE) {
4907 pte_t ptent = *(pte++);
4913 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4914 case MC_TARGET_PAGE:
4917 * We can have a part of the split pmd here. Moving it
4918 * can be done but it would be too convoluted so simply
4919 * ignore such a partial THP and keep it in original
4920 * memcg. There should be somebody mapping the head.
4922 if (PageTransCompound(page))
4924 if (isolate_lru_page(page))
4926 if (!mem_cgroup_move_account(page, false,
4929 /* we uncharge from mc.from later. */
4932 putback_lru_page(page);
4933 put: /* get_mctgt_type() gets the page */
4936 case MC_TARGET_SWAP:
4938 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4940 /* we fixup refcnts and charges later. */
4948 pte_unmap_unlock(pte - 1, ptl);
4953 * We have consumed all precharges we got in can_attach().
4954 * We try charge one by one, but don't do any additional
4955 * charges to mc.to if we have failed in charge once in attach()
4958 ret = mem_cgroup_do_precharge(1);
4966 static void mem_cgroup_move_charge(void)
4968 struct mm_walk mem_cgroup_move_charge_walk = {
4969 .pmd_entry = mem_cgroup_move_charge_pte_range,
4973 lru_add_drain_all();
4975 * Signal lock_page_memcg() to take the memcg's move_lock
4976 * while we're moving its pages to another memcg. Then wait
4977 * for already started RCU-only updates to finish.
4979 atomic_inc(&mc.from->moving_account);
4982 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
4984 * Someone who are holding the mmap_sem might be waiting in
4985 * waitq. So we cancel all extra charges, wake up all waiters,
4986 * and retry. Because we cancel precharges, we might not be able
4987 * to move enough charges, but moving charge is a best-effort
4988 * feature anyway, so it wouldn't be a big problem.
4990 __mem_cgroup_clear_mc();
4995 * When we have consumed all precharges and failed in doing
4996 * additional charge, the page walk just aborts.
4998 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5000 up_read(&mc.mm->mmap_sem);
5001 atomic_dec(&mc.from->moving_account);
5004 static void mem_cgroup_move_task(void)
5007 mem_cgroup_move_charge();
5008 mem_cgroup_clear_mc();
5011 #else /* !CONFIG_MMU */
5012 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5016 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5019 static void mem_cgroup_move_task(void)
5025 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5026 * to verify whether we're attached to the default hierarchy on each mount
5029 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5032 * use_hierarchy is forced on the default hierarchy. cgroup core
5033 * guarantees that @root doesn't have any children, so turning it
5034 * on for the root memcg is enough.
5036 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5037 root_mem_cgroup->use_hierarchy = true;
5039 root_mem_cgroup->use_hierarchy = false;
5042 static u64 memory_current_read(struct cgroup_subsys_state *css,
5045 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5047 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5050 static int memory_low_show(struct seq_file *m, void *v)
5052 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5053 unsigned long low = READ_ONCE(memcg->low);
5055 if (low == PAGE_COUNTER_MAX)
5056 seq_puts(m, "max\n");
5058 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5063 static ssize_t memory_low_write(struct kernfs_open_file *of,
5064 char *buf, size_t nbytes, loff_t off)
5066 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5070 buf = strstrip(buf);
5071 err = page_counter_memparse(buf, "max", &low);
5080 static int memory_high_show(struct seq_file *m, void *v)
5082 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5083 unsigned long high = READ_ONCE(memcg->high);
5085 if (high == PAGE_COUNTER_MAX)
5086 seq_puts(m, "max\n");
5088 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5093 static ssize_t memory_high_write(struct kernfs_open_file *of,
5094 char *buf, size_t nbytes, loff_t off)
5096 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5097 unsigned long nr_pages;
5101 buf = strstrip(buf);
5102 err = page_counter_memparse(buf, "max", &high);
5108 nr_pages = page_counter_read(&memcg->memory);
5109 if (nr_pages > high)
5110 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5113 memcg_wb_domain_size_changed(memcg);
5117 static int memory_max_show(struct seq_file *m, void *v)
5119 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5120 unsigned long max = READ_ONCE(memcg->memory.limit);
5122 if (max == PAGE_COUNTER_MAX)
5123 seq_puts(m, "max\n");
5125 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5130 static ssize_t memory_max_write(struct kernfs_open_file *of,
5131 char *buf, size_t nbytes, loff_t off)
5133 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5134 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5135 bool drained = false;
5139 buf = strstrip(buf);
5140 err = page_counter_memparse(buf, "max", &max);
5144 xchg(&memcg->memory.limit, max);
5147 unsigned long nr_pages = page_counter_read(&memcg->memory);
5149 if (nr_pages <= max)
5152 if (signal_pending(current)) {
5158 drain_all_stock(memcg);
5164 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5170 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5171 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5175 memcg_wb_domain_size_changed(memcg);
5179 static int memory_events_show(struct seq_file *m, void *v)
5181 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5183 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5184 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5185 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5186 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5191 static int memory_stat_show(struct seq_file *m, void *v)
5193 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5194 unsigned long stat[MEMCG_NR_STAT];
5195 unsigned long events[MEMCG_NR_EVENTS];
5199 * Provide statistics on the state of the memory subsystem as
5200 * well as cumulative event counters that show past behavior.
5202 * This list is ordered following a combination of these gradients:
5203 * 1) generic big picture -> specifics and details
5204 * 2) reflecting userspace activity -> reflecting kernel heuristics
5206 * Current memory state:
5209 tree_stat(memcg, stat);
5210 tree_events(memcg, events);
5212 seq_printf(m, "anon %llu\n",
5213 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5214 seq_printf(m, "file %llu\n",
5215 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5216 seq_printf(m, "kernel_stack %llu\n",
5217 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5218 seq_printf(m, "slab %llu\n",
5219 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5220 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5221 seq_printf(m, "sock %llu\n",
5222 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5224 seq_printf(m, "file_mapped %llu\n",
5225 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5226 seq_printf(m, "file_dirty %llu\n",
5227 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5228 seq_printf(m, "file_writeback %llu\n",
5229 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5231 for (i = 0; i < NR_LRU_LISTS; i++) {
5232 struct mem_cgroup *mi;
5233 unsigned long val = 0;
5235 for_each_mem_cgroup_tree(mi, memcg)
5236 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5237 seq_printf(m, "%s %llu\n",
5238 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5241 seq_printf(m, "slab_reclaimable %llu\n",
5242 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5243 seq_printf(m, "slab_unreclaimable %llu\n",
5244 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5246 /* Accumulated memory events */
5248 seq_printf(m, "pgfault %lu\n",
5249 events[MEM_CGROUP_EVENTS_PGFAULT]);
5250 seq_printf(m, "pgmajfault %lu\n",
5251 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5256 static struct cftype memory_files[] = {
5259 .flags = CFTYPE_NOT_ON_ROOT,
5260 .read_u64 = memory_current_read,
5264 .flags = CFTYPE_NOT_ON_ROOT,
5265 .seq_show = memory_low_show,
5266 .write = memory_low_write,
5270 .flags = CFTYPE_NOT_ON_ROOT,
5271 .seq_show = memory_high_show,
5272 .write = memory_high_write,
5276 .flags = CFTYPE_NOT_ON_ROOT,
5277 .seq_show = memory_max_show,
5278 .write = memory_max_write,
5282 .flags = CFTYPE_NOT_ON_ROOT,
5283 .file_offset = offsetof(struct mem_cgroup, events_file),
5284 .seq_show = memory_events_show,
5288 .flags = CFTYPE_NOT_ON_ROOT,
5289 .seq_show = memory_stat_show,
5294 struct cgroup_subsys memory_cgrp_subsys = {
5295 .css_alloc = mem_cgroup_css_alloc,
5296 .css_online = mem_cgroup_css_online,
5297 .css_offline = mem_cgroup_css_offline,
5298 .css_released = mem_cgroup_css_released,
5299 .css_free = mem_cgroup_css_free,
5300 .css_reset = mem_cgroup_css_reset,
5301 .can_attach = mem_cgroup_can_attach,
5302 .cancel_attach = mem_cgroup_cancel_attach,
5303 .post_attach = mem_cgroup_move_task,
5304 .bind = mem_cgroup_bind,
5305 .dfl_cftypes = memory_files,
5306 .legacy_cftypes = mem_cgroup_legacy_files,
5311 * mem_cgroup_low - check if memory consumption is below the normal range
5312 * @root: the highest ancestor to consider
5313 * @memcg: the memory cgroup to check
5315 * Returns %true if memory consumption of @memcg, and that of all
5316 * configurable ancestors up to @root, is below the normal range.
5318 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5320 if (mem_cgroup_disabled())
5324 * The toplevel group doesn't have a configurable range, so
5325 * it's never low when looked at directly, and it is not
5326 * considered an ancestor when assessing the hierarchy.
5329 if (memcg == root_mem_cgroup)
5332 if (page_counter_read(&memcg->memory) >= memcg->low)
5335 while (memcg != root) {
5336 memcg = parent_mem_cgroup(memcg);
5338 if (memcg == root_mem_cgroup)
5341 if (page_counter_read(&memcg->memory) >= memcg->low)
5348 * mem_cgroup_try_charge - try charging a page
5349 * @page: page to charge
5350 * @mm: mm context of the victim
5351 * @gfp_mask: reclaim mode
5352 * @memcgp: charged memcg return
5353 * @compound: charge the page as compound or small page
5355 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5356 * pages according to @gfp_mask if necessary.
5358 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5359 * Otherwise, an error code is returned.
5361 * After page->mapping has been set up, the caller must finalize the
5362 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5363 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5365 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5366 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5369 struct mem_cgroup *memcg = NULL;
5370 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5373 if (mem_cgroup_disabled())
5376 if (PageSwapCache(page)) {
5378 * Every swap fault against a single page tries to charge the
5379 * page, bail as early as possible. shmem_unuse() encounters
5380 * already charged pages, too. The USED bit is protected by
5381 * the page lock, which serializes swap cache removal, which
5382 * in turn serializes uncharging.
5384 VM_BUG_ON_PAGE(!PageLocked(page), page);
5385 if (page->mem_cgroup)
5388 if (do_swap_account) {
5389 swp_entry_t ent = { .val = page_private(page), };
5390 unsigned short id = lookup_swap_cgroup_id(ent);
5393 memcg = mem_cgroup_from_id(id);
5394 if (memcg && !css_tryget_online(&memcg->css))
5401 memcg = get_mem_cgroup_from_mm(mm);
5403 ret = try_charge(memcg, gfp_mask, nr_pages);
5405 css_put(&memcg->css);
5412 * mem_cgroup_commit_charge - commit a page charge
5413 * @page: page to charge
5414 * @memcg: memcg to charge the page to
5415 * @lrucare: page might be on LRU already
5416 * @compound: charge the page as compound or small page
5418 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5419 * after page->mapping has been set up. This must happen atomically
5420 * as part of the page instantiation, i.e. under the page table lock
5421 * for anonymous pages, under the page lock for page and swap cache.
5423 * In addition, the page must not be on the LRU during the commit, to
5424 * prevent racing with task migration. If it might be, use @lrucare.
5426 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5428 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5429 bool lrucare, bool compound)
5431 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5433 VM_BUG_ON_PAGE(!page->mapping, page);
5434 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5436 if (mem_cgroup_disabled())
5439 * Swap faults will attempt to charge the same page multiple
5440 * times. But reuse_swap_page() might have removed the page
5441 * from swapcache already, so we can't check PageSwapCache().
5446 commit_charge(page, memcg, lrucare);
5448 local_irq_disable();
5449 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5450 memcg_check_events(memcg, page);
5453 if (do_memsw_account() && PageSwapCache(page)) {
5454 swp_entry_t entry = { .val = page_private(page) };
5456 * The swap entry might not get freed for a long time,
5457 * let's not wait for it. The page already received a
5458 * memory+swap charge, drop the swap entry duplicate.
5460 mem_cgroup_uncharge_swap(entry);
5465 * mem_cgroup_cancel_charge - cancel a page charge
5466 * @page: page to charge
5467 * @memcg: memcg to charge the page to
5468 * @compound: charge the page as compound or small page
5470 * Cancel a charge transaction started by mem_cgroup_try_charge().
5472 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5475 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5477 if (mem_cgroup_disabled())
5480 * Swap faults will attempt to charge the same page multiple
5481 * times. But reuse_swap_page() might have removed the page
5482 * from swapcache already, so we can't check PageSwapCache().
5487 cancel_charge(memcg, nr_pages);
5490 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5491 unsigned long nr_anon, unsigned long nr_file,
5492 unsigned long nr_huge, unsigned long nr_kmem,
5493 struct page *dummy_page)
5495 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5496 unsigned long flags;
5498 if (!mem_cgroup_is_root(memcg)) {
5499 page_counter_uncharge(&memcg->memory, nr_pages);
5500 if (do_memsw_account())
5501 page_counter_uncharge(&memcg->memsw, nr_pages);
5502 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5503 page_counter_uncharge(&memcg->kmem, nr_kmem);
5504 memcg_oom_recover(memcg);
5507 local_irq_save(flags);
5508 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5509 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5510 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5511 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5512 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5513 memcg_check_events(memcg, dummy_page);
5514 local_irq_restore(flags);
5516 if (!mem_cgroup_is_root(memcg))
5517 css_put_many(&memcg->css, nr_pages);
5520 static void uncharge_list(struct list_head *page_list)
5522 struct mem_cgroup *memcg = NULL;
5523 unsigned long nr_anon = 0;
5524 unsigned long nr_file = 0;
5525 unsigned long nr_huge = 0;
5526 unsigned long nr_kmem = 0;
5527 unsigned long pgpgout = 0;
5528 struct list_head *next;
5532 * Note that the list can be a single page->lru; hence the
5533 * do-while loop instead of a simple list_for_each_entry().
5535 next = page_list->next;
5537 page = list_entry(next, struct page, lru);
5538 next = page->lru.next;
5540 VM_BUG_ON_PAGE(PageLRU(page), page);
5541 VM_BUG_ON_PAGE(!PageHWPoison(page) && page_count(page), page);
5543 if (!page->mem_cgroup)
5547 * Nobody should be changing or seriously looking at
5548 * page->mem_cgroup at this point, we have fully
5549 * exclusive access to the page.
5552 if (memcg != page->mem_cgroup) {
5554 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5555 nr_huge, nr_kmem, page);
5556 pgpgout = nr_anon = nr_file =
5557 nr_huge = nr_kmem = 0;
5559 memcg = page->mem_cgroup;
5562 if (!PageKmemcg(page)) {
5563 unsigned int nr_pages = 1;
5565 if (PageTransHuge(page)) {
5566 nr_pages <<= compound_order(page);
5567 nr_huge += nr_pages;
5570 nr_anon += nr_pages;
5572 nr_file += nr_pages;
5575 nr_kmem += 1 << compound_order(page);
5576 __ClearPageKmemcg(page);
5579 page->mem_cgroup = NULL;
5580 } while (next != page_list);
5583 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5584 nr_huge, nr_kmem, page);
5588 * mem_cgroup_uncharge - uncharge a page
5589 * @page: page to uncharge
5591 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5592 * mem_cgroup_commit_charge().
5594 void mem_cgroup_uncharge(struct page *page)
5596 if (mem_cgroup_disabled())
5599 /* Don't touch page->lru of any random page, pre-check: */
5600 if (!page->mem_cgroup)
5603 INIT_LIST_HEAD(&page->lru);
5604 uncharge_list(&page->lru);
5608 * mem_cgroup_uncharge_list - uncharge a list of page
5609 * @page_list: list of pages to uncharge
5611 * Uncharge a list of pages previously charged with
5612 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5614 void mem_cgroup_uncharge_list(struct list_head *page_list)
5616 if (mem_cgroup_disabled())
5619 if (!list_empty(page_list))
5620 uncharge_list(page_list);
5624 * mem_cgroup_migrate - charge a page's replacement
5625 * @oldpage: currently circulating page
5626 * @newpage: replacement page
5628 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5629 * be uncharged upon free.
5631 * Both pages must be locked, @newpage->mapping must be set up.
5633 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5635 struct mem_cgroup *memcg;
5636 unsigned int nr_pages;
5638 unsigned long flags;
5640 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5641 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5642 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5643 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5646 if (mem_cgroup_disabled())
5649 /* Page cache replacement: new page already charged? */
5650 if (newpage->mem_cgroup)
5653 /* Swapcache readahead pages can get replaced before being charged */
5654 memcg = oldpage->mem_cgroup;
5658 /* Force-charge the new page. The old one will be freed soon */
5659 compound = PageTransHuge(newpage);
5660 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5662 page_counter_charge(&memcg->memory, nr_pages);
5663 if (do_memsw_account())
5664 page_counter_charge(&memcg->memsw, nr_pages);
5665 css_get_many(&memcg->css, nr_pages);
5667 commit_charge(newpage, memcg, false);
5669 local_irq_save(flags);
5670 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5671 memcg_check_events(memcg, newpage);
5672 local_irq_restore(flags);
5675 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5676 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5678 void mem_cgroup_sk_alloc(struct sock *sk)
5680 struct mem_cgroup *memcg;
5682 if (!mem_cgroup_sockets_enabled)
5686 * Socket cloning can throw us here with sk_memcg already
5687 * filled. It won't however, necessarily happen from
5688 * process context. So the test for root memcg given
5689 * the current task's memcg won't help us in this case.
5691 * Respecting the original socket's memcg is a better
5692 * decision in this case.
5695 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5696 css_get(&sk->sk_memcg->css);
5701 memcg = mem_cgroup_from_task(current);
5702 if (memcg == root_mem_cgroup)
5704 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5706 if (css_tryget_online(&memcg->css))
5707 sk->sk_memcg = memcg;
5712 void mem_cgroup_sk_free(struct sock *sk)
5715 css_put(&sk->sk_memcg->css);
5719 * mem_cgroup_charge_skmem - charge socket memory
5720 * @memcg: memcg to charge
5721 * @nr_pages: number of pages to charge
5723 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5724 * @memcg's configured limit, %false if the charge had to be forced.
5726 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5728 gfp_t gfp_mask = GFP_KERNEL;
5730 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5731 struct page_counter *fail;
5733 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5734 memcg->tcpmem_pressure = 0;
5737 page_counter_charge(&memcg->tcpmem, nr_pages);
5738 memcg->tcpmem_pressure = 1;
5742 /* Don't block in the packet receive path */
5744 gfp_mask = GFP_NOWAIT;
5746 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5748 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5751 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5756 * mem_cgroup_uncharge_skmem - uncharge socket memory
5757 * @memcg - memcg to uncharge
5758 * @nr_pages - number of pages to uncharge
5760 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5762 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5763 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5767 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5769 page_counter_uncharge(&memcg->memory, nr_pages);
5770 css_put_many(&memcg->css, nr_pages);
5773 static int __init cgroup_memory(char *s)
5777 while ((token = strsep(&s, ",")) != NULL) {
5780 if (!strcmp(token, "nosocket"))
5781 cgroup_memory_nosocket = true;
5782 if (!strcmp(token, "nokmem"))
5783 cgroup_memory_nokmem = true;
5787 __setup("cgroup.memory=", cgroup_memory);
5790 * subsys_initcall() for memory controller.
5792 * Some parts like hotcpu_notifier() have to be initialized from this context
5793 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5794 * everything that doesn't depend on a specific mem_cgroup structure should
5795 * be initialized from here.
5797 static int __init mem_cgroup_init(void)
5803 * Kmem cache creation is mostly done with the slab_mutex held,
5804 * so use a special workqueue to avoid stalling all worker
5805 * threads in case lots of cgroups are created simultaneously.
5807 memcg_kmem_cache_create_wq =
5808 alloc_ordered_workqueue("memcg_kmem_cache_create", 0);
5809 BUG_ON(!memcg_kmem_cache_create_wq);
5812 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5814 for_each_possible_cpu(cpu)
5815 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5818 for_each_node(node) {
5819 struct mem_cgroup_tree_per_node *rtpn;
5821 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5822 node_online(node) ? node : NUMA_NO_NODE);
5824 rtpn->rb_root = RB_ROOT;
5825 spin_lock_init(&rtpn->lock);
5826 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5831 subsys_initcall(mem_cgroup_init);
5833 #ifdef CONFIG_MEMCG_SWAP
5834 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5836 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5838 * The root cgroup cannot be destroyed, so it's refcount must
5841 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5845 memcg = parent_mem_cgroup(memcg);
5847 memcg = root_mem_cgroup;
5853 * mem_cgroup_swapout - transfer a memsw charge to swap
5854 * @page: page whose memsw charge to transfer
5855 * @entry: swap entry to move the charge to
5857 * Transfer the memsw charge of @page to @entry.
5859 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5861 struct mem_cgroup *memcg, *swap_memcg;
5862 unsigned short oldid;
5864 VM_BUG_ON_PAGE(PageLRU(page), page);
5865 VM_BUG_ON_PAGE(page_count(page), page);
5867 if (!do_memsw_account())
5870 memcg = page->mem_cgroup;
5872 /* Readahead page, never charged */
5877 * In case the memcg owning these pages has been offlined and doesn't
5878 * have an ID allocated to it anymore, charge the closest online
5879 * ancestor for the swap instead and transfer the memory+swap charge.
5881 swap_memcg = mem_cgroup_id_get_online(memcg);
5882 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5883 VM_BUG_ON_PAGE(oldid, page);
5884 mem_cgroup_swap_statistics(swap_memcg, true);
5886 page->mem_cgroup = NULL;
5888 if (!mem_cgroup_is_root(memcg))
5889 page_counter_uncharge(&memcg->memory, 1);
5891 if (memcg != swap_memcg) {
5892 if (!mem_cgroup_is_root(swap_memcg))
5893 page_counter_charge(&swap_memcg->memsw, 1);
5894 page_counter_uncharge(&memcg->memsw, 1);
5898 * Interrupts should be disabled here because the caller holds the
5899 * mapping->tree_lock lock which is taken with interrupts-off. It is
5900 * important here to have the interrupts disabled because it is the
5901 * only synchronisation we have for udpating the per-CPU variables.
5903 VM_BUG_ON(!irqs_disabled());
5904 mem_cgroup_charge_statistics(memcg, page, false, -1);
5905 memcg_check_events(memcg, page);
5907 if (!mem_cgroup_is_root(memcg))
5908 css_put(&memcg->css);
5912 * mem_cgroup_try_charge_swap - try charging a swap entry
5913 * @page: page being added to swap
5914 * @entry: swap entry to charge
5916 * Try to charge @entry to the memcg that @page belongs to.
5918 * Returns 0 on success, -ENOMEM on failure.
5920 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5922 struct mem_cgroup *memcg;
5923 struct page_counter *counter;
5924 unsigned short oldid;
5926 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5929 memcg = page->mem_cgroup;
5931 /* Readahead page, never charged */
5935 memcg = mem_cgroup_id_get_online(memcg);
5937 if (!mem_cgroup_is_root(memcg) &&
5938 !page_counter_try_charge(&memcg->swap, 1, &counter)) {
5939 mem_cgroup_id_put(memcg);
5943 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5944 VM_BUG_ON_PAGE(oldid, page);
5945 mem_cgroup_swap_statistics(memcg, true);
5951 * mem_cgroup_uncharge_swap - uncharge a swap entry
5952 * @entry: swap entry to uncharge
5954 * Drop the swap charge associated with @entry.
5956 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5958 struct mem_cgroup *memcg;
5961 if (!do_swap_account)
5964 id = swap_cgroup_record(entry, 0);
5966 memcg = mem_cgroup_from_id(id);
5968 if (!mem_cgroup_is_root(memcg)) {
5969 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5970 page_counter_uncharge(&memcg->swap, 1);
5972 page_counter_uncharge(&memcg->memsw, 1);
5974 mem_cgroup_swap_statistics(memcg, false);
5975 mem_cgroup_id_put(memcg);
5980 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5982 long nr_swap_pages = get_nr_swap_pages();
5984 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5985 return nr_swap_pages;
5986 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5987 nr_swap_pages = min_t(long, nr_swap_pages,
5988 READ_ONCE(memcg->swap.limit) -
5989 page_counter_read(&memcg->swap));
5990 return nr_swap_pages;
5993 bool mem_cgroup_swap_full(struct page *page)
5995 struct mem_cgroup *memcg;
5997 VM_BUG_ON_PAGE(!PageLocked(page), page);
6001 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6004 memcg = page->mem_cgroup;
6008 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6009 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
6015 /* for remember boot option*/
6016 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6017 static int really_do_swap_account __initdata = 1;
6019 static int really_do_swap_account __initdata;
6022 static int __init enable_swap_account(char *s)
6024 if (!strcmp(s, "1"))
6025 really_do_swap_account = 1;
6026 else if (!strcmp(s, "0"))
6027 really_do_swap_account = 0;
6030 __setup("swapaccount=", enable_swap_account);
6032 static u64 swap_current_read(struct cgroup_subsys_state *css,
6035 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6037 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6040 static int swap_max_show(struct seq_file *m, void *v)
6042 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6043 unsigned long max = READ_ONCE(memcg->swap.limit);
6045 if (max == PAGE_COUNTER_MAX)
6046 seq_puts(m, "max\n");
6048 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6053 static ssize_t swap_max_write(struct kernfs_open_file *of,
6054 char *buf, size_t nbytes, loff_t off)
6056 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6060 buf = strstrip(buf);
6061 err = page_counter_memparse(buf, "max", &max);
6065 mutex_lock(&memcg_limit_mutex);
6066 err = page_counter_limit(&memcg->swap, max);
6067 mutex_unlock(&memcg_limit_mutex);
6074 static struct cftype swap_files[] = {
6076 .name = "swap.current",
6077 .flags = CFTYPE_NOT_ON_ROOT,
6078 .read_u64 = swap_current_read,
6082 .flags = CFTYPE_NOT_ON_ROOT,
6083 .seq_show = swap_max_show,
6084 .write = swap_max_write,
6089 static struct cftype memsw_cgroup_files[] = {
6091 .name = "memsw.usage_in_bytes",
6092 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6093 .read_u64 = mem_cgroup_read_u64,
6096 .name = "memsw.max_usage_in_bytes",
6097 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6098 .write = mem_cgroup_reset,
6099 .read_u64 = mem_cgroup_read_u64,
6102 .name = "memsw.limit_in_bytes",
6103 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6104 .write = mem_cgroup_write,
6105 .read_u64 = mem_cgroup_read_u64,
6108 .name = "memsw.failcnt",
6109 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6110 .write = mem_cgroup_reset,
6111 .read_u64 = mem_cgroup_read_u64,
6113 { }, /* terminate */
6116 static int __init mem_cgroup_swap_init(void)
6118 if (!mem_cgroup_disabled() && really_do_swap_account) {
6119 do_swap_account = 1;
6120 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6122 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6123 memsw_cgroup_files));
6127 subsys_initcall(mem_cgroup_swap_init);
6129 #endif /* CONFIG_MEMCG_SWAP */