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>
70 #include <linux/locallock.h>
72 #include <asm/uaccess.h>
74 #include <trace/events/vmscan.h>
76 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
77 EXPORT_SYMBOL(memory_cgrp_subsys);
79 struct mem_cgroup *root_mem_cgroup __read_mostly;
81 #define MEM_CGROUP_RECLAIM_RETRIES 5
83 /* Socket memory accounting disabled? */
84 static bool cgroup_memory_nosocket;
86 /* Kernel memory accounting disabled? */
87 static bool cgroup_memory_nokmem;
89 /* Whether the swap controller is active */
90 #ifdef CONFIG_MEMCG_SWAP
91 int do_swap_account __read_mostly;
93 #define do_swap_account 0
96 static DEFINE_LOCAL_IRQ_LOCK(event_lock);
98 /* Whether legacy memory+swap accounting is active */
99 static bool do_memsw_account(void)
101 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
104 static const char * const mem_cgroup_stat_names[] = {
114 static const char * const mem_cgroup_events_names[] = {
121 static const char * const mem_cgroup_lru_names[] = {
129 #define THRESHOLDS_EVENTS_TARGET 128
130 #define SOFTLIMIT_EVENTS_TARGET 1024
131 #define NUMAINFO_EVENTS_TARGET 1024
134 * Cgroups above their limits are maintained in a RB-Tree, independent of
135 * their hierarchy representation
138 struct mem_cgroup_tree_per_node {
139 struct rb_root rb_root;
143 struct mem_cgroup_tree {
144 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
147 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
150 struct mem_cgroup_eventfd_list {
151 struct list_head list;
152 struct eventfd_ctx *eventfd;
156 * cgroup_event represents events which userspace want to receive.
158 struct mem_cgroup_event {
160 * memcg which the event belongs to.
162 struct mem_cgroup *memcg;
164 * eventfd to signal userspace about the event.
166 struct eventfd_ctx *eventfd;
168 * Each of these stored in a list by the cgroup.
170 struct list_head list;
172 * register_event() callback will be used to add new userspace
173 * waiter for changes related to this event. Use eventfd_signal()
174 * on eventfd to send notification to userspace.
176 int (*register_event)(struct mem_cgroup *memcg,
177 struct eventfd_ctx *eventfd, const char *args);
179 * unregister_event() callback will be called when userspace closes
180 * the eventfd or on cgroup removing. This callback must be set,
181 * if you want provide notification functionality.
183 void (*unregister_event)(struct mem_cgroup *memcg,
184 struct eventfd_ctx *eventfd);
186 * All fields below needed to unregister event when
187 * userspace closes eventfd.
190 wait_queue_head_t *wqh;
192 struct work_struct remove;
195 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
196 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
198 /* Stuffs for move charges at task migration. */
200 * Types of charges to be moved.
202 #define MOVE_ANON 0x1U
203 #define MOVE_FILE 0x2U
204 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
206 /* "mc" and its members are protected by cgroup_mutex */
207 static struct move_charge_struct {
208 spinlock_t lock; /* for from, to */
209 struct mm_struct *mm;
210 struct mem_cgroup *from;
211 struct mem_cgroup *to;
213 unsigned long precharge;
214 unsigned long moved_charge;
215 unsigned long moved_swap;
216 struct task_struct *moving_task; /* a task moving charges */
217 wait_queue_head_t waitq; /* a waitq for other context */
219 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
220 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
224 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
225 * limit reclaim to prevent infinite loops, if they ever occur.
227 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
228 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
231 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
232 MEM_CGROUP_CHARGE_TYPE_ANON,
233 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
234 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
238 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
253 /* Some nice accessors for the vmpressure. */
254 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
257 memcg = root_mem_cgroup;
258 return &memcg->vmpressure;
261 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
263 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
266 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
268 return (memcg == root_mem_cgroup);
273 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
274 * The main reason for not using cgroup id for this:
275 * this works better in sparse environments, where we have a lot of memcgs,
276 * but only a few kmem-limited. Or also, if we have, for instance, 200
277 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
278 * 200 entry array for that.
280 * The current size of the caches array is stored in memcg_nr_cache_ids. It
281 * will double each time we have to increase it.
283 static DEFINE_IDA(memcg_cache_ida);
284 int memcg_nr_cache_ids;
286 /* Protects memcg_nr_cache_ids */
287 static DECLARE_RWSEM(memcg_cache_ids_sem);
289 void memcg_get_cache_ids(void)
291 down_read(&memcg_cache_ids_sem);
294 void memcg_put_cache_ids(void)
296 up_read(&memcg_cache_ids_sem);
300 * MIN_SIZE is different than 1, because we would like to avoid going through
301 * the alloc/free process all the time. In a small machine, 4 kmem-limited
302 * cgroups is a reasonable guess. In the future, it could be a parameter or
303 * tunable, but that is strictly not necessary.
305 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
306 * this constant directly from cgroup, but it is understandable that this is
307 * better kept as an internal representation in cgroup.c. In any case, the
308 * cgrp_id space is not getting any smaller, and we don't have to necessarily
309 * increase ours as well if it increases.
311 #define MEMCG_CACHES_MIN_SIZE 4
312 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
315 * A lot of the calls to the cache allocation functions are expected to be
316 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
317 * conditional to this static branch, we'll have to allow modules that does
318 * kmem_cache_alloc and the such to see this symbol as well
320 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
321 EXPORT_SYMBOL(memcg_kmem_enabled_key);
323 #endif /* !CONFIG_SLOB */
326 * mem_cgroup_css_from_page - css of the memcg associated with a page
327 * @page: page of interest
329 * If memcg is bound to the default hierarchy, css of the memcg associated
330 * with @page is returned. The returned css remains associated with @page
331 * until it is released.
333 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
336 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
338 struct mem_cgroup *memcg;
340 memcg = page->mem_cgroup;
342 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
343 memcg = root_mem_cgroup;
349 * page_cgroup_ino - return inode number of the memcg a page is charged to
352 * Look up the closest online ancestor of the memory cgroup @page is charged to
353 * and return its inode number or 0 if @page is not charged to any cgroup. It
354 * is safe to call this function without holding a reference to @page.
356 * Note, this function is inherently racy, because there is nothing to prevent
357 * the cgroup inode from getting torn down and potentially reallocated a moment
358 * after page_cgroup_ino() returns, so it only should be used by callers that
359 * do not care (such as procfs interfaces).
361 ino_t page_cgroup_ino(struct page *page)
363 struct mem_cgroup *memcg;
364 unsigned long ino = 0;
367 memcg = READ_ONCE(page->mem_cgroup);
368 while (memcg && !(memcg->css.flags & CSS_ONLINE))
369 memcg = parent_mem_cgroup(memcg);
371 ino = cgroup_ino(memcg->css.cgroup);
376 static struct mem_cgroup_per_node *
377 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
379 int nid = page_to_nid(page);
381 return memcg->nodeinfo[nid];
384 static struct mem_cgroup_tree_per_node *
385 soft_limit_tree_node(int nid)
387 return soft_limit_tree.rb_tree_per_node[nid];
390 static struct mem_cgroup_tree_per_node *
391 soft_limit_tree_from_page(struct page *page)
393 int nid = page_to_nid(page);
395 return soft_limit_tree.rb_tree_per_node[nid];
398 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
399 struct mem_cgroup_tree_per_node *mctz,
400 unsigned long new_usage_in_excess)
402 struct rb_node **p = &mctz->rb_root.rb_node;
403 struct rb_node *parent = NULL;
404 struct mem_cgroup_per_node *mz_node;
409 mz->usage_in_excess = new_usage_in_excess;
410 if (!mz->usage_in_excess)
414 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
416 if (mz->usage_in_excess < mz_node->usage_in_excess)
419 * We can't avoid mem cgroups that are over their soft
420 * limit by the same amount
422 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
425 rb_link_node(&mz->tree_node, parent, p);
426 rb_insert_color(&mz->tree_node, &mctz->rb_root);
430 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
431 struct mem_cgroup_tree_per_node *mctz)
435 rb_erase(&mz->tree_node, &mctz->rb_root);
439 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
440 struct mem_cgroup_tree_per_node *mctz)
444 spin_lock_irqsave(&mctz->lock, flags);
445 __mem_cgroup_remove_exceeded(mz, mctz);
446 spin_unlock_irqrestore(&mctz->lock, flags);
449 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
451 unsigned long nr_pages = page_counter_read(&memcg->memory);
452 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
453 unsigned long excess = 0;
455 if (nr_pages > soft_limit)
456 excess = nr_pages - soft_limit;
461 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
463 unsigned long excess;
464 struct mem_cgroup_per_node *mz;
465 struct mem_cgroup_tree_per_node *mctz;
467 mctz = soft_limit_tree_from_page(page);
469 * Necessary to update all ancestors when hierarchy is used.
470 * because their event counter is not touched.
472 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
473 mz = mem_cgroup_page_nodeinfo(memcg, page);
474 excess = soft_limit_excess(memcg);
476 * We have to update the tree if mz is on RB-tree or
477 * mem is over its softlimit.
479 if (excess || mz->on_tree) {
482 spin_lock_irqsave(&mctz->lock, flags);
483 /* if on-tree, remove it */
485 __mem_cgroup_remove_exceeded(mz, mctz);
487 * Insert again. mz->usage_in_excess will be updated.
488 * If excess is 0, no tree ops.
490 __mem_cgroup_insert_exceeded(mz, mctz, excess);
491 spin_unlock_irqrestore(&mctz->lock, flags);
496 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
498 struct mem_cgroup_tree_per_node *mctz;
499 struct mem_cgroup_per_node *mz;
503 mz = mem_cgroup_nodeinfo(memcg, nid);
504 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 unsigned long nr = 0;
632 struct mem_cgroup_per_node *mz;
635 VM_BUG_ON((unsigned)nid >= nr_node_ids);
638 if (!(BIT(lru) & lru_mask))
640 mz = mem_cgroup_nodeinfo(memcg, nid);
641 nr += mz->lru_size[lru];
646 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
647 unsigned int lru_mask)
649 unsigned long nr = 0;
652 for_each_node_state(nid, N_MEMORY)
653 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
657 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
658 enum mem_cgroup_events_target target)
660 unsigned long val, next;
662 val = __this_cpu_read(memcg->stat->nr_page_events);
663 next = __this_cpu_read(memcg->stat->targets[target]);
664 /* from time_after() in jiffies.h */
665 if ((long)next - (long)val < 0) {
667 case MEM_CGROUP_TARGET_THRESH:
668 next = val + THRESHOLDS_EVENTS_TARGET;
670 case MEM_CGROUP_TARGET_SOFTLIMIT:
671 next = val + SOFTLIMIT_EVENTS_TARGET;
673 case MEM_CGROUP_TARGET_NUMAINFO:
674 next = val + NUMAINFO_EVENTS_TARGET;
679 __this_cpu_write(memcg->stat->targets[target], next);
686 * Check events in order.
689 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
691 /* threshold event is triggered in finer grain than soft limit */
692 if (unlikely(mem_cgroup_event_ratelimit(memcg,
693 MEM_CGROUP_TARGET_THRESH))) {
695 bool do_numainfo __maybe_unused;
697 do_softlimit = mem_cgroup_event_ratelimit(memcg,
698 MEM_CGROUP_TARGET_SOFTLIMIT);
700 do_numainfo = mem_cgroup_event_ratelimit(memcg,
701 MEM_CGROUP_TARGET_NUMAINFO);
703 mem_cgroup_threshold(memcg);
704 if (unlikely(do_softlimit))
705 mem_cgroup_update_tree(memcg, page);
707 if (unlikely(do_numainfo))
708 atomic_inc(&memcg->numainfo_events);
713 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
716 * mm_update_next_owner() may clear mm->owner to NULL
717 * if it races with swapoff, page migration, etc.
718 * So this can be called with p == NULL.
723 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
725 EXPORT_SYMBOL(mem_cgroup_from_task);
727 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
729 struct mem_cgroup *memcg = NULL;
734 * Page cache insertions can happen withou an
735 * actual mm context, e.g. during disk probing
736 * on boot, loopback IO, acct() writes etc.
739 memcg = root_mem_cgroup;
741 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
742 if (unlikely(!memcg))
743 memcg = root_mem_cgroup;
745 } while (!css_tryget_online(&memcg->css));
751 * mem_cgroup_iter - iterate over memory cgroup hierarchy
752 * @root: hierarchy root
753 * @prev: previously returned memcg, NULL on first invocation
754 * @reclaim: cookie for shared reclaim walks, NULL for full walks
756 * Returns references to children of the hierarchy below @root, or
757 * @root itself, or %NULL after a full round-trip.
759 * Caller must pass the return value in @prev on subsequent
760 * invocations for reference counting, or use mem_cgroup_iter_break()
761 * to cancel a hierarchy walk before the round-trip is complete.
763 * Reclaimers can specify a zone and a priority level in @reclaim to
764 * divide up the memcgs in the hierarchy among all concurrent
765 * reclaimers operating on the same zone and priority.
767 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
768 struct mem_cgroup *prev,
769 struct mem_cgroup_reclaim_cookie *reclaim)
771 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
772 struct cgroup_subsys_state *css = NULL;
773 struct mem_cgroup *memcg = NULL;
774 struct mem_cgroup *pos = NULL;
776 if (mem_cgroup_disabled())
780 root = root_mem_cgroup;
782 if (prev && !reclaim)
785 if (!root->use_hierarchy && root != root_mem_cgroup) {
794 struct mem_cgroup_per_node *mz;
796 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
797 iter = &mz->iter[reclaim->priority];
799 if (prev && reclaim->generation != iter->generation)
803 pos = READ_ONCE(iter->position);
804 if (!pos || css_tryget(&pos->css))
807 * css reference reached zero, so iter->position will
808 * be cleared by ->css_released. However, we should not
809 * rely on this happening soon, because ->css_released
810 * is called from a work queue, and by busy-waiting we
811 * might block it. So we clear iter->position right
814 (void)cmpxchg(&iter->position, pos, NULL);
822 css = css_next_descendant_pre(css, &root->css);
825 * Reclaimers share the hierarchy walk, and a
826 * new one might jump in right at the end of
827 * the hierarchy - make sure they see at least
828 * one group and restart from the beginning.
836 * Verify the css and acquire a reference. The root
837 * is provided by the caller, so we know it's alive
838 * and kicking, and don't take an extra reference.
840 memcg = mem_cgroup_from_css(css);
842 if (css == &root->css)
853 * The position could have already been updated by a competing
854 * thread, so check that the value hasn't changed since we read
855 * it to avoid reclaiming from the same cgroup twice.
857 (void)cmpxchg(&iter->position, pos, memcg);
865 reclaim->generation = iter->generation;
871 if (prev && prev != root)
878 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
879 * @root: hierarchy root
880 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
882 void mem_cgroup_iter_break(struct mem_cgroup *root,
883 struct mem_cgroup *prev)
886 root = root_mem_cgroup;
887 if (prev && prev != root)
891 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
893 struct mem_cgroup *memcg = dead_memcg;
894 struct mem_cgroup_reclaim_iter *iter;
895 struct mem_cgroup_per_node *mz;
899 while ((memcg = parent_mem_cgroup(memcg))) {
901 mz = mem_cgroup_nodeinfo(memcg, nid);
902 for (i = 0; i <= DEF_PRIORITY; i++) {
904 cmpxchg(&iter->position,
912 * Iteration constructs for visiting all cgroups (under a tree). If
913 * loops are exited prematurely (break), mem_cgroup_iter_break() must
914 * be used for reference counting.
916 #define for_each_mem_cgroup_tree(iter, root) \
917 for (iter = mem_cgroup_iter(root, NULL, NULL); \
919 iter = mem_cgroup_iter(root, iter, NULL))
921 #define for_each_mem_cgroup(iter) \
922 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
924 iter = mem_cgroup_iter(NULL, iter, NULL))
927 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
928 * @memcg: hierarchy root
929 * @fn: function to call for each task
930 * @arg: argument passed to @fn
932 * This function iterates over tasks attached to @memcg or to any of its
933 * descendants and calls @fn for each task. If @fn returns a non-zero
934 * value, the function breaks the iteration loop and returns the value.
935 * Otherwise, it will iterate over all tasks and return 0.
937 * This function must not be called for the root memory cgroup.
939 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
940 int (*fn)(struct task_struct *, void *), void *arg)
942 struct mem_cgroup *iter;
945 BUG_ON(memcg == root_mem_cgroup);
947 for_each_mem_cgroup_tree(iter, memcg) {
948 struct css_task_iter it;
949 struct task_struct *task;
951 css_task_iter_start(&iter->css, &it);
952 while (!ret && (task = css_task_iter_next(&it)))
954 css_task_iter_end(&it);
956 mem_cgroup_iter_break(memcg, iter);
964 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
966 * @zone: zone of the page
968 * This function is only safe when following the LRU page isolation
969 * and putback protocol: the LRU lock must be held, and the page must
970 * either be PageLRU() or the caller must have isolated/allocated it.
972 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
974 struct mem_cgroup_per_node *mz;
975 struct mem_cgroup *memcg;
976 struct lruvec *lruvec;
978 if (mem_cgroup_disabled()) {
979 lruvec = &pgdat->lruvec;
983 memcg = page->mem_cgroup;
985 * Swapcache readahead pages are added to the LRU - and
986 * possibly migrated - before they are charged.
989 memcg = root_mem_cgroup;
991 mz = mem_cgroup_page_nodeinfo(memcg, page);
992 lruvec = &mz->lruvec;
995 * Since a node can be onlined after the mem_cgroup was created,
996 * we have to be prepared to initialize lruvec->zone here;
997 * and if offlined then reonlined, we need to reinitialize it.
999 if (unlikely(lruvec->pgdat != pgdat))
1000 lruvec->pgdat = pgdat;
1005 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1006 * @lruvec: mem_cgroup per zone lru vector
1007 * @lru: index of lru list the page is sitting on
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,
1017 struct mem_cgroup_per_node *mz;
1018 unsigned long *lru_size;
1022 if (mem_cgroup_disabled())
1025 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1026 lru_size = mz->lru_size + lru;
1027 empty = list_empty(lruvec->lists + lru);
1030 *lru_size += nr_pages;
1033 if (WARN_ONCE(size < 0 || empty != !size,
1034 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1035 __func__, lruvec, lru, nr_pages, size, empty ? "" : "not ")) {
1041 *lru_size += nr_pages;
1044 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1046 struct mem_cgroup *task_memcg;
1047 struct task_struct *p;
1050 p = find_lock_task_mm(task);
1052 task_memcg = get_mem_cgroup_from_mm(p->mm);
1056 * All threads may have already detached their mm's, but the oom
1057 * killer still needs to detect if they have already been oom
1058 * killed to prevent needlessly killing additional tasks.
1061 task_memcg = mem_cgroup_from_task(task);
1062 css_get(&task_memcg->css);
1065 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1066 css_put(&task_memcg->css);
1071 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1072 * @memcg: the memory cgroup
1074 * Returns the maximum amount of memory @mem can be charged with, in
1077 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1079 unsigned long margin = 0;
1080 unsigned long count;
1081 unsigned long limit;
1083 count = page_counter_read(&memcg->memory);
1084 limit = READ_ONCE(memcg->memory.limit);
1086 margin = limit - count;
1088 if (do_memsw_account()) {
1089 count = page_counter_read(&memcg->memsw);
1090 limit = READ_ONCE(memcg->memsw.limit);
1092 margin = min(margin, limit - count);
1101 * A routine for checking "mem" is under move_account() or not.
1103 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1104 * moving cgroups. This is for waiting at high-memory pressure
1107 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1109 struct mem_cgroup *from;
1110 struct mem_cgroup *to;
1113 * Unlike task_move routines, we access mc.to, mc.from not under
1114 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1116 spin_lock(&mc.lock);
1122 ret = mem_cgroup_is_descendant(from, memcg) ||
1123 mem_cgroup_is_descendant(to, memcg);
1125 spin_unlock(&mc.lock);
1129 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1131 if (mc.moving_task && current != mc.moving_task) {
1132 if (mem_cgroup_under_move(memcg)) {
1134 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1135 /* moving charge context might have finished. */
1138 finish_wait(&mc.waitq, &wait);
1145 #define K(x) ((x) << (PAGE_SHIFT-10))
1147 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1148 * @memcg: The memory cgroup that went over limit
1149 * @p: Task that is going to be killed
1151 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1154 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1156 struct mem_cgroup *iter;
1162 pr_info("Task in ");
1163 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1164 pr_cont(" killed as a result of limit of ");
1166 pr_info("Memory limit reached of cgroup ");
1169 pr_cont_cgroup_path(memcg->css.cgroup);
1174 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1175 K((u64)page_counter_read(&memcg->memory)),
1176 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1177 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1178 K((u64)page_counter_read(&memcg->memsw)),
1179 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1180 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1181 K((u64)page_counter_read(&memcg->kmem)),
1182 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1184 for_each_mem_cgroup_tree(iter, memcg) {
1185 pr_info("Memory cgroup stats for ");
1186 pr_cont_cgroup_path(iter->css.cgroup);
1189 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1190 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1192 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1193 K(mem_cgroup_read_stat(iter, i)));
1196 for (i = 0; i < NR_LRU_LISTS; i++)
1197 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1198 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1205 * This function returns the number of memcg under hierarchy tree. Returns
1206 * 1(self count) if no children.
1208 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1211 struct mem_cgroup *iter;
1213 for_each_mem_cgroup_tree(iter, memcg)
1219 * Return the memory (and swap, if configured) limit for a memcg.
1221 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1223 unsigned long limit;
1225 limit = memcg->memory.limit;
1226 if (mem_cgroup_swappiness(memcg)) {
1227 unsigned long memsw_limit;
1228 unsigned long swap_limit;
1230 memsw_limit = memcg->memsw.limit;
1231 swap_limit = memcg->swap.limit;
1232 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1233 limit = min(limit + swap_limit, memsw_limit);
1238 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1241 struct oom_control oc = {
1245 .gfp_mask = gfp_mask,
1250 mutex_lock(&oom_lock);
1251 ret = out_of_memory(&oc);
1252 mutex_unlock(&oom_lock);
1256 #if MAX_NUMNODES > 1
1259 * test_mem_cgroup_node_reclaimable
1260 * @memcg: the target memcg
1261 * @nid: the node ID to be checked.
1262 * @noswap : specify true here if the user wants flle only information.
1264 * This function returns whether the specified memcg contains any
1265 * reclaimable pages on a node. Returns true if there are any reclaimable
1266 * pages in the node.
1268 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1269 int nid, bool noswap)
1271 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1273 if (noswap || !total_swap_pages)
1275 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1282 * Always updating the nodemask is not very good - even if we have an empty
1283 * list or the wrong list here, we can start from some node and traverse all
1284 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1287 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1291 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1292 * pagein/pageout changes since the last update.
1294 if (!atomic_read(&memcg->numainfo_events))
1296 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1299 /* make a nodemask where this memcg uses memory from */
1300 memcg->scan_nodes = node_states[N_MEMORY];
1302 for_each_node_mask(nid, node_states[N_MEMORY]) {
1304 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1305 node_clear(nid, memcg->scan_nodes);
1308 atomic_set(&memcg->numainfo_events, 0);
1309 atomic_set(&memcg->numainfo_updating, 0);
1313 * Selecting a node where we start reclaim from. Because what we need is just
1314 * reducing usage counter, start from anywhere is O,K. Considering
1315 * memory reclaim from current node, there are pros. and cons.
1317 * Freeing memory from current node means freeing memory from a node which
1318 * we'll use or we've used. So, it may make LRU bad. And if several threads
1319 * hit limits, it will see a contention on a node. But freeing from remote
1320 * node means more costs for memory reclaim because of memory latency.
1322 * Now, we use round-robin. Better algorithm is welcomed.
1324 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1328 mem_cgroup_may_update_nodemask(memcg);
1329 node = memcg->last_scanned_node;
1331 node = next_node_in(node, memcg->scan_nodes);
1333 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1334 * last time it really checked all the LRUs due to rate limiting.
1335 * Fallback to the current node in that case for simplicity.
1337 if (unlikely(node == MAX_NUMNODES))
1338 node = numa_node_id();
1340 memcg->last_scanned_node = node;
1344 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1350 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1353 unsigned long *total_scanned)
1355 struct mem_cgroup *victim = NULL;
1358 unsigned long excess;
1359 unsigned long nr_scanned;
1360 struct mem_cgroup_reclaim_cookie reclaim = {
1365 excess = soft_limit_excess(root_memcg);
1368 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1373 * If we have not been able to reclaim
1374 * anything, it might because there are
1375 * no reclaimable pages under this hierarchy
1380 * We want to do more targeted reclaim.
1381 * excess >> 2 is not to excessive so as to
1382 * reclaim too much, nor too less that we keep
1383 * coming back to reclaim from this cgroup
1385 if (total >= (excess >> 2) ||
1386 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1391 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1392 pgdat, &nr_scanned);
1393 *total_scanned += nr_scanned;
1394 if (!soft_limit_excess(root_memcg))
1397 mem_cgroup_iter_break(root_memcg, victim);
1401 #ifdef CONFIG_LOCKDEP
1402 static struct lockdep_map memcg_oom_lock_dep_map = {
1403 .name = "memcg_oom_lock",
1407 static DEFINE_SPINLOCK(memcg_oom_lock);
1410 * Check OOM-Killer is already running under our hierarchy.
1411 * If someone is running, return false.
1413 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1415 struct mem_cgroup *iter, *failed = NULL;
1417 spin_lock(&memcg_oom_lock);
1419 for_each_mem_cgroup_tree(iter, memcg) {
1420 if (iter->oom_lock) {
1422 * this subtree of our hierarchy is already locked
1423 * so we cannot give a lock.
1426 mem_cgroup_iter_break(memcg, iter);
1429 iter->oom_lock = true;
1434 * OK, we failed to lock the whole subtree so we have
1435 * to clean up what we set up to the failing subtree
1437 for_each_mem_cgroup_tree(iter, memcg) {
1438 if (iter == failed) {
1439 mem_cgroup_iter_break(memcg, iter);
1442 iter->oom_lock = false;
1445 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1447 spin_unlock(&memcg_oom_lock);
1452 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1454 struct mem_cgroup *iter;
1456 spin_lock(&memcg_oom_lock);
1457 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1458 for_each_mem_cgroup_tree(iter, memcg)
1459 iter->oom_lock = false;
1460 spin_unlock(&memcg_oom_lock);
1463 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1465 struct mem_cgroup *iter;
1467 spin_lock(&memcg_oom_lock);
1468 for_each_mem_cgroup_tree(iter, memcg)
1470 spin_unlock(&memcg_oom_lock);
1473 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1475 struct mem_cgroup *iter;
1478 * When a new child is created while the hierarchy is under oom,
1479 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1481 spin_lock(&memcg_oom_lock);
1482 for_each_mem_cgroup_tree(iter, memcg)
1483 if (iter->under_oom > 0)
1485 spin_unlock(&memcg_oom_lock);
1488 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1490 struct oom_wait_info {
1491 struct mem_cgroup *memcg;
1495 static int memcg_oom_wake_function(wait_queue_t *wait,
1496 unsigned mode, int sync, void *arg)
1498 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1499 struct mem_cgroup *oom_wait_memcg;
1500 struct oom_wait_info *oom_wait_info;
1502 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1503 oom_wait_memcg = oom_wait_info->memcg;
1505 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1506 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1508 return autoremove_wake_function(wait, mode, sync, arg);
1511 static void memcg_oom_recover(struct mem_cgroup *memcg)
1514 * For the following lockless ->under_oom test, the only required
1515 * guarantee is that it must see the state asserted by an OOM when
1516 * this function is called as a result of userland actions
1517 * triggered by the notification of the OOM. This is trivially
1518 * achieved by invoking mem_cgroup_mark_under_oom() before
1519 * triggering notification.
1521 if (memcg && memcg->under_oom)
1522 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1525 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1527 if (!current->memcg_may_oom)
1530 * We are in the middle of the charge context here, so we
1531 * don't want to block when potentially sitting on a callstack
1532 * that holds all kinds of filesystem and mm locks.
1534 * Also, the caller may handle a failed allocation gracefully
1535 * (like optional page cache readahead) and so an OOM killer
1536 * invocation might not even be necessary.
1538 * That's why we don't do anything here except remember the
1539 * OOM context and then deal with it at the end of the page
1540 * fault when the stack is unwound, the locks are released,
1541 * and when we know whether the fault was overall successful.
1543 css_get(&memcg->css);
1544 current->memcg_in_oom = memcg;
1545 current->memcg_oom_gfp_mask = mask;
1546 current->memcg_oom_order = order;
1550 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1551 * @handle: actually kill/wait or just clean up the OOM state
1553 * This has to be called at the end of a page fault if the memcg OOM
1554 * handler was enabled.
1556 * Memcg supports userspace OOM handling where failed allocations must
1557 * sleep on a waitqueue until the userspace task resolves the
1558 * situation. Sleeping directly in the charge context with all kinds
1559 * of locks held is not a good idea, instead we remember an OOM state
1560 * in the task and mem_cgroup_oom_synchronize() has to be called at
1561 * the end of the page fault to complete the OOM handling.
1563 * Returns %true if an ongoing memcg OOM situation was detected and
1564 * completed, %false otherwise.
1566 bool mem_cgroup_oom_synchronize(bool handle)
1568 struct mem_cgroup *memcg = current->memcg_in_oom;
1569 struct oom_wait_info owait;
1572 /* OOM is global, do not handle */
1579 owait.memcg = memcg;
1580 owait.wait.flags = 0;
1581 owait.wait.func = memcg_oom_wake_function;
1582 owait.wait.private = current;
1583 INIT_LIST_HEAD(&owait.wait.task_list);
1585 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1586 mem_cgroup_mark_under_oom(memcg);
1588 locked = mem_cgroup_oom_trylock(memcg);
1591 mem_cgroup_oom_notify(memcg);
1593 if (locked && !memcg->oom_kill_disable) {
1594 mem_cgroup_unmark_under_oom(memcg);
1595 finish_wait(&memcg_oom_waitq, &owait.wait);
1596 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1597 current->memcg_oom_order);
1600 mem_cgroup_unmark_under_oom(memcg);
1601 finish_wait(&memcg_oom_waitq, &owait.wait);
1605 mem_cgroup_oom_unlock(memcg);
1607 * There is no guarantee that an OOM-lock contender
1608 * sees the wakeups triggered by the OOM kill
1609 * uncharges. Wake any sleepers explicitely.
1611 memcg_oom_recover(memcg);
1614 current->memcg_in_oom = NULL;
1615 css_put(&memcg->css);
1620 * lock_page_memcg - lock a page->mem_cgroup binding
1623 * This function protects unlocked LRU pages from being moved to
1624 * another cgroup and stabilizes their page->mem_cgroup binding.
1626 void lock_page_memcg(struct page *page)
1628 struct mem_cgroup *memcg;
1629 unsigned long flags;
1632 * The RCU lock is held throughout the transaction. The fast
1633 * path can get away without acquiring the memcg->move_lock
1634 * because page moving starts with an RCU grace period.
1638 if (mem_cgroup_disabled())
1641 memcg = page->mem_cgroup;
1642 if (unlikely(!memcg))
1645 if (atomic_read(&memcg->moving_account) <= 0)
1648 spin_lock_irqsave(&memcg->move_lock, flags);
1649 if (memcg != page->mem_cgroup) {
1650 spin_unlock_irqrestore(&memcg->move_lock, flags);
1655 * When charge migration first begins, we can have locked and
1656 * unlocked page stat updates happening concurrently. Track
1657 * the task who has the lock for unlock_page_memcg().
1659 memcg->move_lock_task = current;
1660 memcg->move_lock_flags = flags;
1664 EXPORT_SYMBOL(lock_page_memcg);
1667 * unlock_page_memcg - unlock a page->mem_cgroup binding
1670 void unlock_page_memcg(struct page *page)
1672 struct mem_cgroup *memcg = page->mem_cgroup;
1674 if (memcg && memcg->move_lock_task == current) {
1675 unsigned long flags = memcg->move_lock_flags;
1677 memcg->move_lock_task = NULL;
1678 memcg->move_lock_flags = 0;
1680 spin_unlock_irqrestore(&memcg->move_lock, flags);
1685 EXPORT_SYMBOL(unlock_page_memcg);
1688 * size of first charge trial. "32" comes from vmscan.c's magic value.
1689 * TODO: maybe necessary to use big numbers in big irons.
1691 #define CHARGE_BATCH 32U
1692 struct memcg_stock_pcp {
1693 struct mem_cgroup *cached; /* this never be root cgroup */
1694 unsigned int nr_pages;
1695 struct work_struct work;
1696 unsigned long flags;
1697 #define FLUSHING_CACHED_CHARGE 0
1699 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1700 static DEFINE_LOCAL_IRQ_LOCK(memcg_stock_ll);
1701 static DEFINE_MUTEX(percpu_charge_mutex);
1704 * consume_stock: Try to consume stocked charge on this cpu.
1705 * @memcg: memcg to consume from.
1706 * @nr_pages: how many pages to charge.
1708 * The charges will only happen if @memcg matches the current cpu's memcg
1709 * stock, and at least @nr_pages are available in that stock. Failure to
1710 * service an allocation will refill the stock.
1712 * returns true if successful, false otherwise.
1714 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1716 struct memcg_stock_pcp *stock;
1717 unsigned long flags;
1720 if (nr_pages > CHARGE_BATCH)
1723 local_lock_irqsave(memcg_stock_ll, flags);
1725 stock = this_cpu_ptr(&memcg_stock);
1726 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1727 stock->nr_pages -= nr_pages;
1731 local_unlock_irqrestore(memcg_stock_ll, flags);
1737 * Returns stocks cached in percpu and reset cached information.
1739 static void drain_stock(struct memcg_stock_pcp *stock)
1741 struct mem_cgroup *old = stock->cached;
1743 if (stock->nr_pages) {
1744 page_counter_uncharge(&old->memory, stock->nr_pages);
1745 if (do_memsw_account())
1746 page_counter_uncharge(&old->memsw, stock->nr_pages);
1747 css_put_many(&old->css, stock->nr_pages);
1748 stock->nr_pages = 0;
1750 stock->cached = NULL;
1753 static void drain_local_stock(struct work_struct *dummy)
1755 struct memcg_stock_pcp *stock;
1756 unsigned long flags;
1758 local_lock_irqsave(memcg_stock_ll, flags);
1760 stock = this_cpu_ptr(&memcg_stock);
1762 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1764 local_unlock_irqrestore(memcg_stock_ll, flags);
1768 * Cache charges(val) to local per_cpu area.
1769 * This will be consumed by consume_stock() function, later.
1771 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1773 struct memcg_stock_pcp *stock;
1774 unsigned long flags;
1776 local_lock_irqsave(memcg_stock_ll, flags);
1778 stock = this_cpu_ptr(&memcg_stock);
1779 if (stock->cached != memcg) { /* reset if necessary */
1781 stock->cached = memcg;
1783 stock->nr_pages += nr_pages;
1785 local_unlock_irqrestore(memcg_stock_ll, flags);
1789 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1790 * of the hierarchy under it.
1792 static void drain_all_stock(struct mem_cgroup *root_memcg)
1796 /* If someone's already draining, avoid adding running more workers. */
1797 if (!mutex_trylock(&percpu_charge_mutex))
1799 /* Notify other cpus that system-wide "drain" is running */
1801 curcpu = get_cpu_light();
1802 for_each_online_cpu(cpu) {
1803 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1804 struct mem_cgroup *memcg;
1806 memcg = stock->cached;
1807 if (!memcg || !stock->nr_pages)
1809 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1811 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1813 drain_local_stock(&stock->work);
1815 schedule_work_on(cpu, &stock->work);
1820 mutex_unlock(&percpu_charge_mutex);
1823 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1824 unsigned long action,
1827 int cpu = (unsigned long)hcpu;
1828 struct memcg_stock_pcp *stock;
1830 if (action == CPU_ONLINE)
1833 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1836 stock = &per_cpu(memcg_stock, cpu);
1841 static void reclaim_high(struct mem_cgroup *memcg,
1842 unsigned int nr_pages,
1846 if (page_counter_read(&memcg->memory) <= memcg->high)
1848 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1849 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1850 } while ((memcg = parent_mem_cgroup(memcg)));
1853 static void high_work_func(struct work_struct *work)
1855 struct mem_cgroup *memcg;
1857 memcg = container_of(work, struct mem_cgroup, high_work);
1858 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1862 * Scheduled by try_charge() to be executed from the userland return path
1863 * and reclaims memory over the high limit.
1865 void mem_cgroup_handle_over_high(void)
1867 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1868 struct mem_cgroup *memcg;
1870 if (likely(!nr_pages))
1873 memcg = get_mem_cgroup_from_mm(current->mm);
1874 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1875 css_put(&memcg->css);
1876 current->memcg_nr_pages_over_high = 0;
1879 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1880 unsigned int nr_pages)
1882 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1883 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1884 struct mem_cgroup *mem_over_limit;
1885 struct page_counter *counter;
1886 unsigned long nr_reclaimed;
1887 bool may_swap = true;
1888 bool drained = false;
1890 if (mem_cgroup_is_root(memcg))
1893 if (consume_stock(memcg, nr_pages))
1896 if (!do_memsw_account() ||
1897 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1898 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1900 if (do_memsw_account())
1901 page_counter_uncharge(&memcg->memsw, batch);
1902 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1904 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1908 if (batch > nr_pages) {
1914 * Unlike in global OOM situations, memcg is not in a physical
1915 * memory shortage. Allow dying and OOM-killed tasks to
1916 * bypass the last charges so that they can exit quickly and
1917 * free their memory.
1919 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1920 fatal_signal_pending(current) ||
1921 current->flags & PF_EXITING))
1925 * Prevent unbounded recursion when reclaim operations need to
1926 * allocate memory. This might exceed the limits temporarily,
1927 * but we prefer facilitating memory reclaim and getting back
1928 * under the limit over triggering OOM kills in these cases.
1930 if (unlikely(current->flags & PF_MEMALLOC))
1933 if (unlikely(task_in_memcg_oom(current)))
1936 if (!gfpflags_allow_blocking(gfp_mask))
1939 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1941 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1942 gfp_mask, may_swap);
1944 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1948 drain_all_stock(mem_over_limit);
1953 if (gfp_mask & __GFP_NORETRY)
1956 * Even though the limit is exceeded at this point, reclaim
1957 * may have been able to free some pages. Retry the charge
1958 * before killing the task.
1960 * Only for regular pages, though: huge pages are rather
1961 * unlikely to succeed so close to the limit, and we fall back
1962 * to regular pages anyway in case of failure.
1964 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1967 * At task move, charge accounts can be doubly counted. So, it's
1968 * better to wait until the end of task_move if something is going on.
1970 if (mem_cgroup_wait_acct_move(mem_over_limit))
1976 if (gfp_mask & __GFP_NOFAIL)
1979 if (fatal_signal_pending(current))
1982 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
1984 mem_cgroup_oom(mem_over_limit, gfp_mask,
1985 get_order(nr_pages * PAGE_SIZE));
1987 if (!(gfp_mask & __GFP_NOFAIL))
1991 * The allocation either can't fail or will lead to more memory
1992 * being freed very soon. Allow memory usage go over the limit
1993 * temporarily by force charging it.
1995 page_counter_charge(&memcg->memory, nr_pages);
1996 if (do_memsw_account())
1997 page_counter_charge(&memcg->memsw, nr_pages);
1998 css_get_many(&memcg->css, nr_pages);
2003 css_get_many(&memcg->css, batch);
2004 if (batch > nr_pages)
2005 refill_stock(memcg, batch - nr_pages);
2008 * If the hierarchy is above the normal consumption range, schedule
2009 * reclaim on returning to userland. We can perform reclaim here
2010 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2011 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2012 * not recorded as it most likely matches current's and won't
2013 * change in the meantime. As high limit is checked again before
2014 * reclaim, the cost of mismatch is negligible.
2017 if (page_counter_read(&memcg->memory) > memcg->high) {
2018 /* Don't bother a random interrupted task */
2019 if (in_interrupt()) {
2020 schedule_work(&memcg->high_work);
2023 current->memcg_nr_pages_over_high += batch;
2024 set_notify_resume(current);
2027 } while ((memcg = parent_mem_cgroup(memcg)));
2032 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2034 if (mem_cgroup_is_root(memcg))
2037 page_counter_uncharge(&memcg->memory, nr_pages);
2038 if (do_memsw_account())
2039 page_counter_uncharge(&memcg->memsw, nr_pages);
2041 css_put_many(&memcg->css, nr_pages);
2044 static void lock_page_lru(struct page *page, int *isolated)
2046 struct zone *zone = page_zone(page);
2048 spin_lock_irq(zone_lru_lock(zone));
2049 if (PageLRU(page)) {
2050 struct lruvec *lruvec;
2052 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2054 del_page_from_lru_list(page, lruvec, page_lru(page));
2060 static void unlock_page_lru(struct page *page, int isolated)
2062 struct zone *zone = page_zone(page);
2065 struct lruvec *lruvec;
2067 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2068 VM_BUG_ON_PAGE(PageLRU(page), page);
2070 add_page_to_lru_list(page, lruvec, page_lru(page));
2072 spin_unlock_irq(zone_lru_lock(zone));
2075 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2080 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2083 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2084 * may already be on some other mem_cgroup's LRU. Take care of it.
2087 lock_page_lru(page, &isolated);
2090 * Nobody should be changing or seriously looking at
2091 * page->mem_cgroup at this point:
2093 * - the page is uncharged
2095 * - the page is off-LRU
2097 * - an anonymous fault has exclusive page access, except for
2098 * a locked page table
2100 * - a page cache insertion, a swapin fault, or a migration
2101 * have the page locked
2103 page->mem_cgroup = memcg;
2106 unlock_page_lru(page, isolated);
2110 static int memcg_alloc_cache_id(void)
2115 id = ida_simple_get(&memcg_cache_ida,
2116 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2120 if (id < memcg_nr_cache_ids)
2124 * There's no space for the new id in memcg_caches arrays,
2125 * so we have to grow them.
2127 down_write(&memcg_cache_ids_sem);
2129 size = 2 * (id + 1);
2130 if (size < MEMCG_CACHES_MIN_SIZE)
2131 size = MEMCG_CACHES_MIN_SIZE;
2132 else if (size > MEMCG_CACHES_MAX_SIZE)
2133 size = MEMCG_CACHES_MAX_SIZE;
2135 err = memcg_update_all_caches(size);
2137 err = memcg_update_all_list_lrus(size);
2139 memcg_nr_cache_ids = size;
2141 up_write(&memcg_cache_ids_sem);
2144 ida_simple_remove(&memcg_cache_ida, id);
2150 static void memcg_free_cache_id(int id)
2152 ida_simple_remove(&memcg_cache_ida, id);
2155 struct memcg_kmem_cache_create_work {
2156 struct mem_cgroup *memcg;
2157 struct kmem_cache *cachep;
2158 struct work_struct work;
2161 static void memcg_kmem_cache_create_func(struct work_struct *w)
2163 struct memcg_kmem_cache_create_work *cw =
2164 container_of(w, struct memcg_kmem_cache_create_work, work);
2165 struct mem_cgroup *memcg = cw->memcg;
2166 struct kmem_cache *cachep = cw->cachep;
2168 memcg_create_kmem_cache(memcg, cachep);
2170 css_put(&memcg->css);
2175 * Enqueue the creation of a per-memcg kmem_cache.
2177 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2178 struct kmem_cache *cachep)
2180 struct memcg_kmem_cache_create_work *cw;
2182 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2186 css_get(&memcg->css);
2189 cw->cachep = cachep;
2190 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2192 schedule_work(&cw->work);
2195 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2196 struct kmem_cache *cachep)
2199 * We need to stop accounting when we kmalloc, because if the
2200 * corresponding kmalloc cache is not yet created, the first allocation
2201 * in __memcg_schedule_kmem_cache_create will recurse.
2203 * However, it is better to enclose the whole function. Depending on
2204 * the debugging options enabled, INIT_WORK(), for instance, can
2205 * trigger an allocation. This too, will make us recurse. Because at
2206 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2207 * the safest choice is to do it like this, wrapping the whole function.
2209 current->memcg_kmem_skip_account = 1;
2210 __memcg_schedule_kmem_cache_create(memcg, cachep);
2211 current->memcg_kmem_skip_account = 0;
2214 static inline bool memcg_kmem_bypass(void)
2216 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2222 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2223 * @cachep: the original global kmem cache
2225 * Return the kmem_cache we're supposed to use for a slab allocation.
2226 * We try to use the current memcg's version of the cache.
2228 * If the cache does not exist yet, if we are the first user of it, we
2229 * create it asynchronously in a workqueue and let the current allocation
2230 * go through with the original cache.
2232 * This function takes a reference to the cache it returns to assure it
2233 * won't get destroyed while we are working with it. Once the caller is
2234 * done with it, memcg_kmem_put_cache() must be called to release the
2237 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2239 struct mem_cgroup *memcg;
2240 struct kmem_cache *memcg_cachep;
2243 VM_BUG_ON(!is_root_cache(cachep));
2245 if (memcg_kmem_bypass())
2248 if (current->memcg_kmem_skip_account)
2251 memcg = get_mem_cgroup_from_mm(current->mm);
2252 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2256 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2257 if (likely(memcg_cachep))
2258 return memcg_cachep;
2261 * If we are in a safe context (can wait, and not in interrupt
2262 * context), we could be be predictable and return right away.
2263 * This would guarantee that the allocation being performed
2264 * already belongs in the new cache.
2266 * However, there are some clashes that can arrive from locking.
2267 * For instance, because we acquire the slab_mutex while doing
2268 * memcg_create_kmem_cache, this means no further allocation
2269 * could happen with the slab_mutex held. So it's better to
2272 memcg_schedule_kmem_cache_create(memcg, cachep);
2274 css_put(&memcg->css);
2279 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2280 * @cachep: the cache returned by memcg_kmem_get_cache
2282 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2284 if (!is_root_cache(cachep))
2285 css_put(&cachep->memcg_params.memcg->css);
2289 * memcg_kmem_charge: charge a kmem page
2290 * @page: page to charge
2291 * @gfp: reclaim mode
2292 * @order: allocation order
2293 * @memcg: memory cgroup to charge
2295 * Returns 0 on success, an error code on failure.
2297 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2298 struct mem_cgroup *memcg)
2300 unsigned int nr_pages = 1 << order;
2301 struct page_counter *counter;
2304 ret = try_charge(memcg, gfp, nr_pages);
2308 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2309 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2310 cancel_charge(memcg, nr_pages);
2314 page->mem_cgroup = memcg;
2320 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2321 * @page: page to charge
2322 * @gfp: reclaim mode
2323 * @order: allocation order
2325 * Returns 0 on success, an error code on failure.
2327 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2329 struct mem_cgroup *memcg;
2332 if (memcg_kmem_bypass())
2335 memcg = get_mem_cgroup_from_mm(current->mm);
2336 if (!mem_cgroup_is_root(memcg)) {
2337 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2339 __SetPageKmemcg(page);
2341 css_put(&memcg->css);
2345 * memcg_kmem_uncharge: uncharge a kmem page
2346 * @page: page to uncharge
2347 * @order: allocation order
2349 void memcg_kmem_uncharge(struct page *page, int order)
2351 struct mem_cgroup *memcg = page->mem_cgroup;
2352 unsigned int nr_pages = 1 << order;
2357 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2359 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2360 page_counter_uncharge(&memcg->kmem, nr_pages);
2362 page_counter_uncharge(&memcg->memory, nr_pages);
2363 if (do_memsw_account())
2364 page_counter_uncharge(&memcg->memsw, nr_pages);
2366 page->mem_cgroup = NULL;
2368 /* slab pages do not have PageKmemcg flag set */
2369 if (PageKmemcg(page))
2370 __ClearPageKmemcg(page);
2372 css_put_many(&memcg->css, nr_pages);
2374 #endif /* !CONFIG_SLOB */
2376 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2379 * Because tail pages are not marked as "used", set it. We're under
2380 * zone_lru_lock and migration entries setup in all page mappings.
2382 void mem_cgroup_split_huge_fixup(struct page *head)
2386 if (mem_cgroup_disabled())
2389 for (i = 1; i < HPAGE_PMD_NR; i++)
2390 head[i].mem_cgroup = head->mem_cgroup;
2392 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2395 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2397 #ifdef CONFIG_MEMCG_SWAP
2398 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2401 int val = (charge) ? 1 : -1;
2402 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2406 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2407 * @entry: swap entry to be moved
2408 * @from: mem_cgroup which the entry is moved from
2409 * @to: mem_cgroup which the entry is moved to
2411 * It succeeds only when the swap_cgroup's record for this entry is the same
2412 * as the mem_cgroup's id of @from.
2414 * Returns 0 on success, -EINVAL on failure.
2416 * The caller must have charged to @to, IOW, called page_counter_charge() about
2417 * both res and memsw, and called css_get().
2419 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2420 struct mem_cgroup *from, struct mem_cgroup *to)
2422 unsigned short old_id, new_id;
2424 old_id = mem_cgroup_id(from);
2425 new_id = mem_cgroup_id(to);
2427 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2428 mem_cgroup_swap_statistics(from, false);
2429 mem_cgroup_swap_statistics(to, true);
2435 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2436 struct mem_cgroup *from, struct mem_cgroup *to)
2442 static DEFINE_MUTEX(memcg_limit_mutex);
2444 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2445 unsigned long limit)
2447 unsigned long curusage;
2448 unsigned long oldusage;
2449 bool enlarge = false;
2454 * For keeping hierarchical_reclaim simple, how long we should retry
2455 * is depends on callers. We set our retry-count to be function
2456 * of # of children which we should visit in this loop.
2458 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2459 mem_cgroup_count_children(memcg);
2461 oldusage = page_counter_read(&memcg->memory);
2464 if (signal_pending(current)) {
2469 mutex_lock(&memcg_limit_mutex);
2470 if (limit > memcg->memsw.limit) {
2471 mutex_unlock(&memcg_limit_mutex);
2475 if (limit > memcg->memory.limit)
2477 ret = page_counter_limit(&memcg->memory, limit);
2478 mutex_unlock(&memcg_limit_mutex);
2483 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2485 curusage = page_counter_read(&memcg->memory);
2486 /* Usage is reduced ? */
2487 if (curusage >= oldusage)
2490 oldusage = curusage;
2491 } while (retry_count);
2493 if (!ret && enlarge)
2494 memcg_oom_recover(memcg);
2499 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2500 unsigned long limit)
2502 unsigned long curusage;
2503 unsigned long oldusage;
2504 bool enlarge = false;
2508 /* see mem_cgroup_resize_res_limit */
2509 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2510 mem_cgroup_count_children(memcg);
2512 oldusage = page_counter_read(&memcg->memsw);
2515 if (signal_pending(current)) {
2520 mutex_lock(&memcg_limit_mutex);
2521 if (limit < memcg->memory.limit) {
2522 mutex_unlock(&memcg_limit_mutex);
2526 if (limit > memcg->memsw.limit)
2528 ret = page_counter_limit(&memcg->memsw, limit);
2529 mutex_unlock(&memcg_limit_mutex);
2534 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2536 curusage = page_counter_read(&memcg->memsw);
2537 /* Usage is reduced ? */
2538 if (curusage >= oldusage)
2541 oldusage = curusage;
2542 } while (retry_count);
2544 if (!ret && enlarge)
2545 memcg_oom_recover(memcg);
2550 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2552 unsigned long *total_scanned)
2554 unsigned long nr_reclaimed = 0;
2555 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2556 unsigned long reclaimed;
2558 struct mem_cgroup_tree_per_node *mctz;
2559 unsigned long excess;
2560 unsigned long nr_scanned;
2565 mctz = soft_limit_tree_node(pgdat->node_id);
2568 * Do not even bother to check the largest node if the root
2569 * is empty. Do it lockless to prevent lock bouncing. Races
2570 * are acceptable as soft limit is best effort anyway.
2572 if (RB_EMPTY_ROOT(&mctz->rb_root))
2576 * This loop can run a while, specially if mem_cgroup's continuously
2577 * keep exceeding their soft limit and putting the system under
2584 mz = mem_cgroup_largest_soft_limit_node(mctz);
2589 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2590 gfp_mask, &nr_scanned);
2591 nr_reclaimed += reclaimed;
2592 *total_scanned += nr_scanned;
2593 spin_lock_irq(&mctz->lock);
2594 __mem_cgroup_remove_exceeded(mz, mctz);
2597 * If we failed to reclaim anything from this memory cgroup
2598 * it is time to move on to the next cgroup
2602 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2604 excess = soft_limit_excess(mz->memcg);
2606 * One school of thought says that we should not add
2607 * back the node to the tree if reclaim returns 0.
2608 * But our reclaim could return 0, simply because due
2609 * to priority we are exposing a smaller subset of
2610 * memory to reclaim from. Consider this as a longer
2613 /* If excess == 0, no tree ops */
2614 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2615 spin_unlock_irq(&mctz->lock);
2616 css_put(&mz->memcg->css);
2619 * Could not reclaim anything and there are no more
2620 * mem cgroups to try or we seem to be looping without
2621 * reclaiming anything.
2623 if (!nr_reclaimed &&
2625 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2627 } while (!nr_reclaimed);
2629 css_put(&next_mz->memcg->css);
2630 return nr_reclaimed;
2634 * Test whether @memcg has children, dead or alive. Note that this
2635 * function doesn't care whether @memcg has use_hierarchy enabled and
2636 * returns %true if there are child csses according to the cgroup
2637 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2639 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2644 ret = css_next_child(NULL, &memcg->css);
2650 * Reclaims as many pages from the given memcg as possible.
2652 * Caller is responsible for holding css reference for memcg.
2654 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2656 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2658 /* we call try-to-free pages for make this cgroup empty */
2659 lru_add_drain_all();
2660 /* try to free all pages in this cgroup */
2661 while (nr_retries && page_counter_read(&memcg->memory)) {
2664 if (signal_pending(current))
2667 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2671 /* maybe some writeback is necessary */
2672 congestion_wait(BLK_RW_ASYNC, HZ/10);
2680 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2681 char *buf, size_t nbytes,
2684 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2686 if (mem_cgroup_is_root(memcg))
2688 return mem_cgroup_force_empty(memcg) ?: nbytes;
2691 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2694 return mem_cgroup_from_css(css)->use_hierarchy;
2697 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2698 struct cftype *cft, u64 val)
2701 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2702 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2704 if (memcg->use_hierarchy == val)
2708 * If parent's use_hierarchy is set, we can't make any modifications
2709 * in the child subtrees. If it is unset, then the change can
2710 * occur, provided the current cgroup has no children.
2712 * For the root cgroup, parent_mem is NULL, we allow value to be
2713 * set if there are no children.
2715 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2716 (val == 1 || val == 0)) {
2717 if (!memcg_has_children(memcg))
2718 memcg->use_hierarchy = val;
2727 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2729 struct mem_cgroup *iter;
2732 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2734 for_each_mem_cgroup_tree(iter, memcg) {
2735 for (i = 0; i < MEMCG_NR_STAT; i++)
2736 stat[i] += mem_cgroup_read_stat(iter, i);
2740 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2742 struct mem_cgroup *iter;
2745 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2747 for_each_mem_cgroup_tree(iter, memcg) {
2748 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2749 events[i] += mem_cgroup_read_events(iter, i);
2753 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2755 unsigned long val = 0;
2757 if (mem_cgroup_is_root(memcg)) {
2758 struct mem_cgroup *iter;
2760 for_each_mem_cgroup_tree(iter, memcg) {
2761 val += mem_cgroup_read_stat(iter,
2762 MEM_CGROUP_STAT_CACHE);
2763 val += mem_cgroup_read_stat(iter,
2764 MEM_CGROUP_STAT_RSS);
2766 val += mem_cgroup_read_stat(iter,
2767 MEM_CGROUP_STAT_SWAP);
2771 val = page_counter_read(&memcg->memory);
2773 val = page_counter_read(&memcg->memsw);
2786 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2789 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2790 struct page_counter *counter;
2792 switch (MEMFILE_TYPE(cft->private)) {
2794 counter = &memcg->memory;
2797 counter = &memcg->memsw;
2800 counter = &memcg->kmem;
2803 counter = &memcg->tcpmem;
2809 switch (MEMFILE_ATTR(cft->private)) {
2811 if (counter == &memcg->memory)
2812 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2813 if (counter == &memcg->memsw)
2814 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2815 return (u64)page_counter_read(counter) * PAGE_SIZE;
2817 return (u64)counter->limit * PAGE_SIZE;
2819 return (u64)counter->watermark * PAGE_SIZE;
2821 return counter->failcnt;
2822 case RES_SOFT_LIMIT:
2823 return (u64)memcg->soft_limit * PAGE_SIZE;
2830 static int memcg_online_kmem(struct mem_cgroup *memcg)
2834 if (cgroup_memory_nokmem)
2837 BUG_ON(memcg->kmemcg_id >= 0);
2838 BUG_ON(memcg->kmem_state);
2840 memcg_id = memcg_alloc_cache_id();
2844 static_branch_inc(&memcg_kmem_enabled_key);
2846 * A memory cgroup is considered kmem-online as soon as it gets
2847 * kmemcg_id. Setting the id after enabling static branching will
2848 * guarantee no one starts accounting before all call sites are
2851 memcg->kmemcg_id = memcg_id;
2852 memcg->kmem_state = KMEM_ONLINE;
2857 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2859 struct cgroup_subsys_state *css;
2860 struct mem_cgroup *parent, *child;
2863 if (memcg->kmem_state != KMEM_ONLINE)
2866 * Clear the online state before clearing memcg_caches array
2867 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2868 * guarantees that no cache will be created for this cgroup
2869 * after we are done (see memcg_create_kmem_cache()).
2871 memcg->kmem_state = KMEM_ALLOCATED;
2873 memcg_deactivate_kmem_caches(memcg);
2875 kmemcg_id = memcg->kmemcg_id;
2876 BUG_ON(kmemcg_id < 0);
2878 parent = parent_mem_cgroup(memcg);
2880 parent = root_mem_cgroup;
2883 * Change kmemcg_id of this cgroup and all its descendants to the
2884 * parent's id, and then move all entries from this cgroup's list_lrus
2885 * to ones of the parent. After we have finished, all list_lrus
2886 * corresponding to this cgroup are guaranteed to remain empty. The
2887 * ordering is imposed by list_lru_node->lock taken by
2888 * memcg_drain_all_list_lrus().
2890 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2891 css_for_each_descendant_pre(css, &memcg->css) {
2892 child = mem_cgroup_from_css(css);
2893 BUG_ON(child->kmemcg_id != kmemcg_id);
2894 child->kmemcg_id = parent->kmemcg_id;
2895 if (!memcg->use_hierarchy)
2900 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2902 memcg_free_cache_id(kmemcg_id);
2905 static void memcg_free_kmem(struct mem_cgroup *memcg)
2907 /* css_alloc() failed, offlining didn't happen */
2908 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2909 memcg_offline_kmem(memcg);
2911 if (memcg->kmem_state == KMEM_ALLOCATED) {
2912 memcg_destroy_kmem_caches(memcg);
2913 static_branch_dec(&memcg_kmem_enabled_key);
2914 WARN_ON(page_counter_read(&memcg->kmem));
2918 static int memcg_online_kmem(struct mem_cgroup *memcg)
2922 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2925 static void memcg_free_kmem(struct mem_cgroup *memcg)
2928 #endif /* !CONFIG_SLOB */
2930 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2931 unsigned long limit)
2935 mutex_lock(&memcg_limit_mutex);
2936 ret = page_counter_limit(&memcg->kmem, limit);
2937 mutex_unlock(&memcg_limit_mutex);
2941 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2945 mutex_lock(&memcg_limit_mutex);
2947 ret = page_counter_limit(&memcg->tcpmem, limit);
2951 if (!memcg->tcpmem_active) {
2953 * The active flag needs to be written after the static_key
2954 * update. This is what guarantees that the socket activation
2955 * function is the last one to run. See mem_cgroup_sk_alloc()
2956 * for details, and note that we don't mark any socket as
2957 * belonging to this memcg until that flag is up.
2959 * We need to do this, because static_keys will span multiple
2960 * sites, but we can't control their order. If we mark a socket
2961 * as accounted, but the accounting functions are not patched in
2962 * yet, we'll lose accounting.
2964 * We never race with the readers in mem_cgroup_sk_alloc(),
2965 * because when this value change, the code to process it is not
2968 static_branch_inc(&memcg_sockets_enabled_key);
2969 memcg->tcpmem_active = true;
2972 mutex_unlock(&memcg_limit_mutex);
2977 * The user of this function is...
2980 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2981 char *buf, size_t nbytes, loff_t off)
2983 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2984 unsigned long nr_pages;
2987 buf = strstrip(buf);
2988 ret = page_counter_memparse(buf, "-1", &nr_pages);
2992 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2994 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2998 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3000 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3003 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3006 ret = memcg_update_kmem_limit(memcg, nr_pages);
3009 ret = memcg_update_tcp_limit(memcg, nr_pages);
3013 case RES_SOFT_LIMIT:
3014 memcg->soft_limit = nr_pages;
3018 return ret ?: nbytes;
3021 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3022 size_t nbytes, loff_t off)
3024 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3025 struct page_counter *counter;
3027 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3029 counter = &memcg->memory;
3032 counter = &memcg->memsw;
3035 counter = &memcg->kmem;
3038 counter = &memcg->tcpmem;
3044 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3046 page_counter_reset_watermark(counter);
3049 counter->failcnt = 0;
3058 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3061 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3065 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3066 struct cftype *cft, u64 val)
3068 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3070 if (val & ~MOVE_MASK)
3074 * No kind of locking is needed in here, because ->can_attach() will
3075 * check this value once in the beginning of the process, and then carry
3076 * on with stale data. This means that changes to this value will only
3077 * affect task migrations starting after the change.
3079 memcg->move_charge_at_immigrate = val;
3083 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3084 struct cftype *cft, u64 val)
3091 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3095 unsigned int lru_mask;
3098 static const struct numa_stat stats[] = {
3099 { "total", LRU_ALL },
3100 { "file", LRU_ALL_FILE },
3101 { "anon", LRU_ALL_ANON },
3102 { "unevictable", BIT(LRU_UNEVICTABLE) },
3104 const struct numa_stat *stat;
3107 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3109 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3110 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3111 seq_printf(m, "%s=%lu", stat->name, nr);
3112 for_each_node_state(nid, N_MEMORY) {
3113 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3115 seq_printf(m, " N%d=%lu", nid, nr);
3120 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3121 struct mem_cgroup *iter;
3124 for_each_mem_cgroup_tree(iter, memcg)
3125 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3126 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3127 for_each_node_state(nid, N_MEMORY) {
3129 for_each_mem_cgroup_tree(iter, memcg)
3130 nr += mem_cgroup_node_nr_lru_pages(
3131 iter, nid, stat->lru_mask);
3132 seq_printf(m, " N%d=%lu", nid, nr);
3139 #endif /* CONFIG_NUMA */
3141 static int memcg_stat_show(struct seq_file *m, void *v)
3143 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3144 unsigned long memory, memsw;
3145 struct mem_cgroup *mi;
3148 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3149 MEM_CGROUP_STAT_NSTATS);
3150 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3151 MEM_CGROUP_EVENTS_NSTATS);
3152 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3154 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3155 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3157 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3158 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3161 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3162 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3163 mem_cgroup_read_events(memcg, i));
3165 for (i = 0; i < NR_LRU_LISTS; i++)
3166 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3167 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3169 /* Hierarchical information */
3170 memory = memsw = PAGE_COUNTER_MAX;
3171 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3172 memory = min(memory, mi->memory.limit);
3173 memsw = min(memsw, mi->memsw.limit);
3175 seq_printf(m, "hierarchical_memory_limit %llu\n",
3176 (u64)memory * PAGE_SIZE);
3177 if (do_memsw_account())
3178 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3179 (u64)memsw * PAGE_SIZE);
3181 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3182 unsigned long long val = 0;
3184 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3186 for_each_mem_cgroup_tree(mi, memcg)
3187 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3188 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3191 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3192 unsigned long long val = 0;
3194 for_each_mem_cgroup_tree(mi, memcg)
3195 val += mem_cgroup_read_events(mi, i);
3196 seq_printf(m, "total_%s %llu\n",
3197 mem_cgroup_events_names[i], val);
3200 for (i = 0; i < NR_LRU_LISTS; i++) {
3201 unsigned long long val = 0;
3203 for_each_mem_cgroup_tree(mi, memcg)
3204 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3205 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3208 #ifdef CONFIG_DEBUG_VM
3211 struct mem_cgroup_per_node *mz;
3212 struct zone_reclaim_stat *rstat;
3213 unsigned long recent_rotated[2] = {0, 0};
3214 unsigned long recent_scanned[2] = {0, 0};
3216 for_each_online_pgdat(pgdat) {
3217 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3218 rstat = &mz->lruvec.reclaim_stat;
3220 recent_rotated[0] += rstat->recent_rotated[0];
3221 recent_rotated[1] += rstat->recent_rotated[1];
3222 recent_scanned[0] += rstat->recent_scanned[0];
3223 recent_scanned[1] += rstat->recent_scanned[1];
3225 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3226 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3227 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3228 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3235 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3238 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3240 return mem_cgroup_swappiness(memcg);
3243 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3244 struct cftype *cft, u64 val)
3246 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3252 memcg->swappiness = val;
3254 vm_swappiness = val;
3259 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3261 struct mem_cgroup_threshold_ary *t;
3262 unsigned long usage;
3267 t = rcu_dereference(memcg->thresholds.primary);
3269 t = rcu_dereference(memcg->memsw_thresholds.primary);
3274 usage = mem_cgroup_usage(memcg, swap);
3277 * current_threshold points to threshold just below or equal to usage.
3278 * If it's not true, a threshold was crossed after last
3279 * call of __mem_cgroup_threshold().
3281 i = t->current_threshold;
3284 * Iterate backward over array of thresholds starting from
3285 * current_threshold and check if a threshold is crossed.
3286 * If none of thresholds below usage is crossed, we read
3287 * only one element of the array here.
3289 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3290 eventfd_signal(t->entries[i].eventfd, 1);
3292 /* i = current_threshold + 1 */
3296 * Iterate forward over array of thresholds starting from
3297 * current_threshold+1 and check if a threshold is crossed.
3298 * If none of thresholds above usage is crossed, we read
3299 * only one element of the array here.
3301 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3302 eventfd_signal(t->entries[i].eventfd, 1);
3304 /* Update current_threshold */
3305 t->current_threshold = i - 1;
3310 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3313 __mem_cgroup_threshold(memcg, false);
3314 if (do_memsw_account())
3315 __mem_cgroup_threshold(memcg, true);
3317 memcg = parent_mem_cgroup(memcg);
3321 static int compare_thresholds(const void *a, const void *b)
3323 const struct mem_cgroup_threshold *_a = a;
3324 const struct mem_cgroup_threshold *_b = b;
3326 if (_a->threshold > _b->threshold)
3329 if (_a->threshold < _b->threshold)
3335 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3337 struct mem_cgroup_eventfd_list *ev;
3339 spin_lock(&memcg_oom_lock);
3341 list_for_each_entry(ev, &memcg->oom_notify, list)
3342 eventfd_signal(ev->eventfd, 1);
3344 spin_unlock(&memcg_oom_lock);
3348 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3350 struct mem_cgroup *iter;
3352 for_each_mem_cgroup_tree(iter, memcg)
3353 mem_cgroup_oom_notify_cb(iter);
3356 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3357 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3359 struct mem_cgroup_thresholds *thresholds;
3360 struct mem_cgroup_threshold_ary *new;
3361 unsigned long threshold;
3362 unsigned long usage;
3365 ret = page_counter_memparse(args, "-1", &threshold);
3369 mutex_lock(&memcg->thresholds_lock);
3372 thresholds = &memcg->thresholds;
3373 usage = mem_cgroup_usage(memcg, false);
3374 } else if (type == _MEMSWAP) {
3375 thresholds = &memcg->memsw_thresholds;
3376 usage = mem_cgroup_usage(memcg, true);
3380 /* Check if a threshold crossed before adding a new one */
3381 if (thresholds->primary)
3382 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3384 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3386 /* Allocate memory for new array of thresholds */
3387 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3395 /* Copy thresholds (if any) to new array */
3396 if (thresholds->primary) {
3397 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3398 sizeof(struct mem_cgroup_threshold));
3401 /* Add new threshold */
3402 new->entries[size - 1].eventfd = eventfd;
3403 new->entries[size - 1].threshold = threshold;
3405 /* Sort thresholds. Registering of new threshold isn't time-critical */
3406 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3407 compare_thresholds, NULL);
3409 /* Find current threshold */
3410 new->current_threshold = -1;
3411 for (i = 0; i < size; i++) {
3412 if (new->entries[i].threshold <= usage) {
3414 * new->current_threshold will not be used until
3415 * rcu_assign_pointer(), so it's safe to increment
3418 ++new->current_threshold;
3423 /* Free old spare buffer and save old primary buffer as spare */
3424 kfree(thresholds->spare);
3425 thresholds->spare = thresholds->primary;
3427 rcu_assign_pointer(thresholds->primary, new);
3429 /* To be sure that nobody uses thresholds */
3433 mutex_unlock(&memcg->thresholds_lock);
3438 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3439 struct eventfd_ctx *eventfd, const char *args)
3441 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3444 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3445 struct eventfd_ctx *eventfd, const char *args)
3447 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3450 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3451 struct eventfd_ctx *eventfd, enum res_type type)
3453 struct mem_cgroup_thresholds *thresholds;
3454 struct mem_cgroup_threshold_ary *new;
3455 unsigned long usage;
3458 mutex_lock(&memcg->thresholds_lock);
3461 thresholds = &memcg->thresholds;
3462 usage = mem_cgroup_usage(memcg, false);
3463 } else if (type == _MEMSWAP) {
3464 thresholds = &memcg->memsw_thresholds;
3465 usage = mem_cgroup_usage(memcg, true);
3469 if (!thresholds->primary)
3472 /* Check if a threshold crossed before removing */
3473 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3475 /* Calculate new number of threshold */
3477 for (i = 0; i < thresholds->primary->size; i++) {
3478 if (thresholds->primary->entries[i].eventfd != eventfd)
3482 new = thresholds->spare;
3484 /* Set thresholds array to NULL if we don't have thresholds */
3493 /* Copy thresholds and find current threshold */
3494 new->current_threshold = -1;
3495 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3496 if (thresholds->primary->entries[i].eventfd == eventfd)
3499 new->entries[j] = thresholds->primary->entries[i];
3500 if (new->entries[j].threshold <= usage) {
3502 * new->current_threshold will not be used
3503 * until rcu_assign_pointer(), so it's safe to increment
3506 ++new->current_threshold;
3512 /* Swap primary and spare array */
3513 thresholds->spare = thresholds->primary;
3515 rcu_assign_pointer(thresholds->primary, new);
3517 /* To be sure that nobody uses thresholds */
3520 /* If all events are unregistered, free the spare array */
3522 kfree(thresholds->spare);
3523 thresholds->spare = NULL;
3526 mutex_unlock(&memcg->thresholds_lock);
3529 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3530 struct eventfd_ctx *eventfd)
3532 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3535 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3536 struct eventfd_ctx *eventfd)
3538 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3541 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3542 struct eventfd_ctx *eventfd, const char *args)
3544 struct mem_cgroup_eventfd_list *event;
3546 event = kmalloc(sizeof(*event), GFP_KERNEL);
3550 spin_lock(&memcg_oom_lock);
3552 event->eventfd = eventfd;
3553 list_add(&event->list, &memcg->oom_notify);
3555 /* already in OOM ? */
3556 if (memcg->under_oom)
3557 eventfd_signal(eventfd, 1);
3558 spin_unlock(&memcg_oom_lock);
3563 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3564 struct eventfd_ctx *eventfd)
3566 struct mem_cgroup_eventfd_list *ev, *tmp;
3568 spin_lock(&memcg_oom_lock);
3570 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3571 if (ev->eventfd == eventfd) {
3572 list_del(&ev->list);
3577 spin_unlock(&memcg_oom_lock);
3580 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3582 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3584 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3585 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3589 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3590 struct cftype *cft, u64 val)
3592 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3594 /* cannot set to root cgroup and only 0 and 1 are allowed */
3595 if (!css->parent || !((val == 0) || (val == 1)))
3598 memcg->oom_kill_disable = val;
3600 memcg_oom_recover(memcg);
3605 #ifdef CONFIG_CGROUP_WRITEBACK
3607 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3609 return &memcg->cgwb_list;
3612 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3614 return wb_domain_init(&memcg->cgwb_domain, gfp);
3617 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3619 wb_domain_exit(&memcg->cgwb_domain);
3622 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3624 wb_domain_size_changed(&memcg->cgwb_domain);
3627 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3629 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3631 if (!memcg->css.parent)
3634 return &memcg->cgwb_domain;
3638 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3639 * @wb: bdi_writeback in question
3640 * @pfilepages: out parameter for number of file pages
3641 * @pheadroom: out parameter for number of allocatable pages according to memcg
3642 * @pdirty: out parameter for number of dirty pages
3643 * @pwriteback: out parameter for number of pages under writeback
3645 * Determine the numbers of file, headroom, dirty, and writeback pages in
3646 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3647 * is a bit more involved.
3649 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3650 * headroom is calculated as the lowest headroom of itself and the
3651 * ancestors. Note that this doesn't consider the actual amount of
3652 * available memory in the system. The caller should further cap
3653 * *@pheadroom accordingly.
3655 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3656 unsigned long *pheadroom, unsigned long *pdirty,
3657 unsigned long *pwriteback)
3659 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3660 struct mem_cgroup *parent;
3662 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3664 /* this should eventually include NR_UNSTABLE_NFS */
3665 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3666 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3667 (1 << LRU_ACTIVE_FILE));
3668 *pheadroom = PAGE_COUNTER_MAX;
3670 while ((parent = parent_mem_cgroup(memcg))) {
3671 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3672 unsigned long used = page_counter_read(&memcg->memory);
3674 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3679 #else /* CONFIG_CGROUP_WRITEBACK */
3681 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3686 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3690 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3694 #endif /* CONFIG_CGROUP_WRITEBACK */
3697 * DO NOT USE IN NEW FILES.
3699 * "cgroup.event_control" implementation.
3701 * This is way over-engineered. It tries to support fully configurable
3702 * events for each user. Such level of flexibility is completely
3703 * unnecessary especially in the light of the planned unified hierarchy.
3705 * Please deprecate this and replace with something simpler if at all
3710 * Unregister event and free resources.
3712 * Gets called from workqueue.
3714 static void memcg_event_remove(struct work_struct *work)
3716 struct mem_cgroup_event *event =
3717 container_of(work, struct mem_cgroup_event, remove);
3718 struct mem_cgroup *memcg = event->memcg;
3720 remove_wait_queue(event->wqh, &event->wait);
3722 event->unregister_event(memcg, event->eventfd);
3724 /* Notify userspace the event is going away. */
3725 eventfd_signal(event->eventfd, 1);
3727 eventfd_ctx_put(event->eventfd);
3729 css_put(&memcg->css);
3733 * Gets called on POLLHUP on eventfd when user closes it.
3735 * Called with wqh->lock held and interrupts disabled.
3737 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3738 int sync, void *key)
3740 struct mem_cgroup_event *event =
3741 container_of(wait, struct mem_cgroup_event, wait);
3742 struct mem_cgroup *memcg = event->memcg;
3743 unsigned long flags = (unsigned long)key;
3745 if (flags & POLLHUP) {
3747 * If the event has been detached at cgroup removal, we
3748 * can simply return knowing the other side will cleanup
3751 * We can't race against event freeing since the other
3752 * side will require wqh->lock via remove_wait_queue(),
3755 spin_lock(&memcg->event_list_lock);
3756 if (!list_empty(&event->list)) {
3757 list_del_init(&event->list);
3759 * We are in atomic context, but cgroup_event_remove()
3760 * may sleep, so we have to call it in workqueue.
3762 schedule_work(&event->remove);
3764 spin_unlock(&memcg->event_list_lock);
3770 static void memcg_event_ptable_queue_proc(struct file *file,
3771 wait_queue_head_t *wqh, poll_table *pt)
3773 struct mem_cgroup_event *event =
3774 container_of(pt, struct mem_cgroup_event, pt);
3777 add_wait_queue(wqh, &event->wait);
3781 * DO NOT USE IN NEW FILES.
3783 * Parse input and register new cgroup event handler.
3785 * Input must be in format '<event_fd> <control_fd> <args>'.
3786 * Interpretation of args is defined by control file implementation.
3788 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3789 char *buf, size_t nbytes, loff_t off)
3791 struct cgroup_subsys_state *css = of_css(of);
3792 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3793 struct mem_cgroup_event *event;
3794 struct cgroup_subsys_state *cfile_css;
3795 unsigned int efd, cfd;
3802 buf = strstrip(buf);
3804 efd = simple_strtoul(buf, &endp, 10);
3809 cfd = simple_strtoul(buf, &endp, 10);
3810 if ((*endp != ' ') && (*endp != '\0'))
3814 event = kzalloc(sizeof(*event), GFP_KERNEL);
3818 event->memcg = memcg;
3819 INIT_LIST_HEAD(&event->list);
3820 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3821 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3822 INIT_WORK(&event->remove, memcg_event_remove);
3830 event->eventfd = eventfd_ctx_fileget(efile.file);
3831 if (IS_ERR(event->eventfd)) {
3832 ret = PTR_ERR(event->eventfd);
3839 goto out_put_eventfd;
3842 /* the process need read permission on control file */
3843 /* AV: shouldn't we check that it's been opened for read instead? */
3844 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3849 * Determine the event callbacks and set them in @event. This used
3850 * to be done via struct cftype but cgroup core no longer knows
3851 * about these events. The following is crude but the whole thing
3852 * is for compatibility anyway.
3854 * DO NOT ADD NEW FILES.
3856 name = cfile.file->f_path.dentry->d_name.name;
3858 if (!strcmp(name, "memory.usage_in_bytes")) {
3859 event->register_event = mem_cgroup_usage_register_event;
3860 event->unregister_event = mem_cgroup_usage_unregister_event;
3861 } else if (!strcmp(name, "memory.oom_control")) {
3862 event->register_event = mem_cgroup_oom_register_event;
3863 event->unregister_event = mem_cgroup_oom_unregister_event;
3864 } else if (!strcmp(name, "memory.pressure_level")) {
3865 event->register_event = vmpressure_register_event;
3866 event->unregister_event = vmpressure_unregister_event;
3867 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3868 event->register_event = memsw_cgroup_usage_register_event;
3869 event->unregister_event = memsw_cgroup_usage_unregister_event;
3876 * Verify @cfile should belong to @css. Also, remaining events are
3877 * automatically removed on cgroup destruction but the removal is
3878 * asynchronous, so take an extra ref on @css.
3880 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3881 &memory_cgrp_subsys);
3883 if (IS_ERR(cfile_css))
3885 if (cfile_css != css) {
3890 ret = event->register_event(memcg, event->eventfd, buf);
3894 efile.file->f_op->poll(efile.file, &event->pt);
3896 spin_lock(&memcg->event_list_lock);
3897 list_add(&event->list, &memcg->event_list);
3898 spin_unlock(&memcg->event_list_lock);
3910 eventfd_ctx_put(event->eventfd);
3919 static struct cftype mem_cgroup_legacy_files[] = {
3921 .name = "usage_in_bytes",
3922 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3923 .read_u64 = mem_cgroup_read_u64,
3926 .name = "max_usage_in_bytes",
3927 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3928 .write = mem_cgroup_reset,
3929 .read_u64 = mem_cgroup_read_u64,
3932 .name = "limit_in_bytes",
3933 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3934 .write = mem_cgroup_write,
3935 .read_u64 = mem_cgroup_read_u64,
3938 .name = "soft_limit_in_bytes",
3939 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3940 .write = mem_cgroup_write,
3941 .read_u64 = mem_cgroup_read_u64,
3945 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3946 .write = mem_cgroup_reset,
3947 .read_u64 = mem_cgroup_read_u64,
3951 .seq_show = memcg_stat_show,
3954 .name = "force_empty",
3955 .write = mem_cgroup_force_empty_write,
3958 .name = "use_hierarchy",
3959 .write_u64 = mem_cgroup_hierarchy_write,
3960 .read_u64 = mem_cgroup_hierarchy_read,
3963 .name = "cgroup.event_control", /* XXX: for compat */
3964 .write = memcg_write_event_control,
3965 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3968 .name = "swappiness",
3969 .read_u64 = mem_cgroup_swappiness_read,
3970 .write_u64 = mem_cgroup_swappiness_write,
3973 .name = "move_charge_at_immigrate",
3974 .read_u64 = mem_cgroup_move_charge_read,
3975 .write_u64 = mem_cgroup_move_charge_write,
3978 .name = "oom_control",
3979 .seq_show = mem_cgroup_oom_control_read,
3980 .write_u64 = mem_cgroup_oom_control_write,
3981 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3984 .name = "pressure_level",
3988 .name = "numa_stat",
3989 .seq_show = memcg_numa_stat_show,
3993 .name = "kmem.limit_in_bytes",
3994 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
3995 .write = mem_cgroup_write,
3996 .read_u64 = mem_cgroup_read_u64,
3999 .name = "kmem.usage_in_bytes",
4000 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4001 .read_u64 = mem_cgroup_read_u64,
4004 .name = "kmem.failcnt",
4005 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4006 .write = mem_cgroup_reset,
4007 .read_u64 = mem_cgroup_read_u64,
4010 .name = "kmem.max_usage_in_bytes",
4011 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4012 .write = mem_cgroup_reset,
4013 .read_u64 = mem_cgroup_read_u64,
4015 #ifdef CONFIG_SLABINFO
4017 .name = "kmem.slabinfo",
4018 .seq_start = slab_start,
4019 .seq_next = slab_next,
4020 .seq_stop = slab_stop,
4021 .seq_show = memcg_slab_show,
4025 .name = "kmem.tcp.limit_in_bytes",
4026 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4027 .write = mem_cgroup_write,
4028 .read_u64 = mem_cgroup_read_u64,
4031 .name = "kmem.tcp.usage_in_bytes",
4032 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4033 .read_u64 = mem_cgroup_read_u64,
4036 .name = "kmem.tcp.failcnt",
4037 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4038 .write = mem_cgroup_reset,
4039 .read_u64 = mem_cgroup_read_u64,
4042 .name = "kmem.tcp.max_usage_in_bytes",
4043 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4044 .write = mem_cgroup_reset,
4045 .read_u64 = mem_cgroup_read_u64,
4047 { }, /* terminate */
4051 * Private memory cgroup IDR
4053 * Swap-out records and page cache shadow entries need to store memcg
4054 * references in constrained space, so we maintain an ID space that is
4055 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4056 * memory-controlled cgroups to 64k.
4058 * However, there usually are many references to the oflline CSS after
4059 * the cgroup has been destroyed, such as page cache or reclaimable
4060 * slab objects, that don't need to hang on to the ID. We want to keep
4061 * those dead CSS from occupying IDs, or we might quickly exhaust the
4062 * relatively small ID space and prevent the creation of new cgroups
4063 * even when there are much fewer than 64k cgroups - possibly none.
4065 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4066 * be freed and recycled when it's no longer needed, which is usually
4067 * when the CSS is offlined.
4069 * The only exception to that are records of swapped out tmpfs/shmem
4070 * pages that need to be attributed to live ancestors on swapin. But
4071 * those references are manageable from userspace.
4074 static DEFINE_IDR(mem_cgroup_idr);
4076 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4078 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4079 atomic_add(n, &memcg->id.ref);
4082 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4084 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4085 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4086 idr_remove(&mem_cgroup_idr, memcg->id.id);
4089 /* Memcg ID pins CSS */
4090 css_put(&memcg->css);
4094 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4096 mem_cgroup_id_get_many(memcg, 1);
4099 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4101 mem_cgroup_id_put_many(memcg, 1);
4105 * mem_cgroup_from_id - look up a memcg from a memcg id
4106 * @id: the memcg id to look up
4108 * Caller must hold rcu_read_lock().
4110 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4112 WARN_ON_ONCE(!rcu_read_lock_held());
4113 return idr_find(&mem_cgroup_idr, id);
4116 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4118 struct mem_cgroup_per_node *pn;
4121 * This routine is called against possible nodes.
4122 * But it's BUG to call kmalloc() against offline node.
4124 * TODO: this routine can waste much memory for nodes which will
4125 * never be onlined. It's better to use memory hotplug callback
4128 if (!node_state(node, N_NORMAL_MEMORY))
4130 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4134 lruvec_init(&pn->lruvec);
4135 pn->usage_in_excess = 0;
4136 pn->on_tree = false;
4139 memcg->nodeinfo[node] = pn;
4143 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4145 kfree(memcg->nodeinfo[node]);
4148 static void mem_cgroup_free(struct mem_cgroup *memcg)
4152 memcg_wb_domain_exit(memcg);
4154 free_mem_cgroup_per_node_info(memcg, node);
4155 free_percpu(memcg->stat);
4159 static struct mem_cgroup *mem_cgroup_alloc(void)
4161 struct mem_cgroup *memcg;
4165 size = sizeof(struct mem_cgroup);
4166 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4168 memcg = kzalloc(size, GFP_KERNEL);
4172 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4173 1, MEM_CGROUP_ID_MAX,
4175 if (memcg->id.id < 0)
4178 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4183 if (alloc_mem_cgroup_per_node_info(memcg, node))
4186 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4189 INIT_WORK(&memcg->high_work, high_work_func);
4190 memcg->last_scanned_node = MAX_NUMNODES;
4191 INIT_LIST_HEAD(&memcg->oom_notify);
4192 mutex_init(&memcg->thresholds_lock);
4193 spin_lock_init(&memcg->move_lock);
4194 vmpressure_init(&memcg->vmpressure);
4195 INIT_LIST_HEAD(&memcg->event_list);
4196 spin_lock_init(&memcg->event_list_lock);
4197 memcg->socket_pressure = jiffies;
4199 memcg->kmemcg_id = -1;
4201 #ifdef CONFIG_CGROUP_WRITEBACK
4202 INIT_LIST_HEAD(&memcg->cgwb_list);
4204 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4207 if (memcg->id.id > 0)
4208 idr_remove(&mem_cgroup_idr, memcg->id.id);
4209 mem_cgroup_free(memcg);
4213 static struct cgroup_subsys_state * __ref
4214 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4216 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4217 struct mem_cgroup *memcg;
4218 long error = -ENOMEM;
4220 memcg = mem_cgroup_alloc();
4222 return ERR_PTR(error);
4224 memcg->high = PAGE_COUNTER_MAX;
4225 memcg->soft_limit = PAGE_COUNTER_MAX;
4227 memcg->swappiness = mem_cgroup_swappiness(parent);
4228 memcg->oom_kill_disable = parent->oom_kill_disable;
4230 if (parent && parent->use_hierarchy) {
4231 memcg->use_hierarchy = true;
4232 page_counter_init(&memcg->memory, &parent->memory);
4233 page_counter_init(&memcg->swap, &parent->swap);
4234 page_counter_init(&memcg->memsw, &parent->memsw);
4235 page_counter_init(&memcg->kmem, &parent->kmem);
4236 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4238 page_counter_init(&memcg->memory, NULL);
4239 page_counter_init(&memcg->swap, NULL);
4240 page_counter_init(&memcg->memsw, NULL);
4241 page_counter_init(&memcg->kmem, NULL);
4242 page_counter_init(&memcg->tcpmem, NULL);
4244 * Deeper hierachy with use_hierarchy == false doesn't make
4245 * much sense so let cgroup subsystem know about this
4246 * unfortunate state in our controller.
4248 if (parent != root_mem_cgroup)
4249 memory_cgrp_subsys.broken_hierarchy = true;
4252 /* The following stuff does not apply to the root */
4254 root_mem_cgroup = memcg;
4258 error = memcg_online_kmem(memcg);
4262 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4263 static_branch_inc(&memcg_sockets_enabled_key);
4267 mem_cgroup_free(memcg);
4268 return ERR_PTR(-ENOMEM);
4271 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4273 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4275 /* Online state pins memcg ID, memcg ID pins CSS */
4276 atomic_set(&memcg->id.ref, 1);
4281 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4283 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4284 struct mem_cgroup_event *event, *tmp;
4287 * Unregister events and notify userspace.
4288 * Notify userspace about cgroup removing only after rmdir of cgroup
4289 * directory to avoid race between userspace and kernelspace.
4291 spin_lock(&memcg->event_list_lock);
4292 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4293 list_del_init(&event->list);
4294 schedule_work(&event->remove);
4296 spin_unlock(&memcg->event_list_lock);
4298 memcg_offline_kmem(memcg);
4299 wb_memcg_offline(memcg);
4301 mem_cgroup_id_put(memcg);
4304 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4306 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4308 invalidate_reclaim_iterators(memcg);
4311 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4313 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4315 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4316 static_branch_dec(&memcg_sockets_enabled_key);
4318 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4319 static_branch_dec(&memcg_sockets_enabled_key);
4321 vmpressure_cleanup(&memcg->vmpressure);
4322 cancel_work_sync(&memcg->high_work);
4323 mem_cgroup_remove_from_trees(memcg);
4324 memcg_free_kmem(memcg);
4325 mem_cgroup_free(memcg);
4329 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4330 * @css: the target css
4332 * Reset the states of the mem_cgroup associated with @css. This is
4333 * invoked when the userland requests disabling on the default hierarchy
4334 * but the memcg is pinned through dependency. The memcg should stop
4335 * applying policies and should revert to the vanilla state as it may be
4336 * made visible again.
4338 * The current implementation only resets the essential configurations.
4339 * This needs to be expanded to cover all the visible parts.
4341 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4345 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4346 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4347 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4348 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4349 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4351 memcg->high = PAGE_COUNTER_MAX;
4352 memcg->soft_limit = PAGE_COUNTER_MAX;
4353 memcg_wb_domain_size_changed(memcg);
4357 /* Handlers for move charge at task migration. */
4358 static int mem_cgroup_do_precharge(unsigned long count)
4362 /* Try a single bulk charge without reclaim first, kswapd may wake */
4363 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4365 mc.precharge += count;
4369 /* Try charges one by one with reclaim */
4371 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4385 enum mc_target_type {
4391 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4392 unsigned long addr, pte_t ptent)
4394 struct page *page = vm_normal_page(vma, addr, ptent);
4396 if (!page || !page_mapped(page))
4398 if (PageAnon(page)) {
4399 if (!(mc.flags & MOVE_ANON))
4402 if (!(mc.flags & MOVE_FILE))
4405 if (!get_page_unless_zero(page))
4412 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4413 pte_t ptent, swp_entry_t *entry)
4415 struct page *page = NULL;
4416 swp_entry_t ent = pte_to_swp_entry(ptent);
4418 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4421 * Because lookup_swap_cache() updates some statistics counter,
4422 * we call find_get_page() with swapper_space directly.
4424 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4425 if (do_memsw_account())
4426 entry->val = ent.val;
4431 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4432 pte_t ptent, swp_entry_t *entry)
4438 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4439 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4441 struct page *page = NULL;
4442 struct address_space *mapping;
4445 if (!vma->vm_file) /* anonymous vma */
4447 if (!(mc.flags & MOVE_FILE))
4450 mapping = vma->vm_file->f_mapping;
4451 pgoff = linear_page_index(vma, addr);
4453 /* page is moved even if it's not RSS of this task(page-faulted). */
4455 /* shmem/tmpfs may report page out on swap: account for that too. */
4456 if (shmem_mapping(mapping)) {
4457 page = find_get_entry(mapping, pgoff);
4458 if (radix_tree_exceptional_entry(page)) {
4459 swp_entry_t swp = radix_to_swp_entry(page);
4460 if (do_memsw_account())
4462 page = find_get_page(swap_address_space(swp),
4466 page = find_get_page(mapping, pgoff);
4468 page = find_get_page(mapping, pgoff);
4474 * mem_cgroup_move_account - move account of the page
4476 * @compound: charge the page as compound or small page
4477 * @from: mem_cgroup which the page is moved from.
4478 * @to: mem_cgroup which the page is moved to. @from != @to.
4480 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4482 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4485 static int mem_cgroup_move_account(struct page *page,
4487 struct mem_cgroup *from,
4488 struct mem_cgroup *to)
4490 unsigned long flags;
4491 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4495 VM_BUG_ON(from == to);
4496 VM_BUG_ON_PAGE(PageLRU(page), page);
4497 VM_BUG_ON(compound && !PageTransHuge(page));
4500 * Prevent mem_cgroup_migrate() from looking at
4501 * page->mem_cgroup of its source page while we change it.
4504 if (!trylock_page(page))
4508 if (page->mem_cgroup != from)
4511 anon = PageAnon(page);
4513 spin_lock_irqsave(&from->move_lock, flags);
4515 if (!anon && page_mapped(page)) {
4516 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4518 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4523 * move_lock grabbed above and caller set from->moving_account, so
4524 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4525 * So mapping should be stable for dirty pages.
4527 if (!anon && PageDirty(page)) {
4528 struct address_space *mapping = page_mapping(page);
4530 if (mapping_cap_account_dirty(mapping)) {
4531 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4533 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4538 if (PageWriteback(page)) {
4539 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4541 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4546 * It is safe to change page->mem_cgroup here because the page
4547 * is referenced, charged, and isolated - we can't race with
4548 * uncharging, charging, migration, or LRU putback.
4551 /* caller should have done css_get */
4552 page->mem_cgroup = to;
4553 spin_unlock_irqrestore(&from->move_lock, flags);
4557 local_lock_irq(event_lock);
4558 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4559 memcg_check_events(to, page);
4560 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4561 memcg_check_events(from, page);
4562 local_unlock_irq(event_lock);
4570 * get_mctgt_type - get target type of moving charge
4571 * @vma: the vma the pte to be checked belongs
4572 * @addr: the address corresponding to the pte to be checked
4573 * @ptent: the pte to be checked
4574 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4577 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4578 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4579 * move charge. if @target is not NULL, the page is stored in target->page
4580 * with extra refcnt got(Callers should handle it).
4581 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4582 * target for charge migration. if @target is not NULL, the entry is stored
4585 * Called with pte lock held.
4588 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4589 unsigned long addr, pte_t ptent, union mc_target *target)
4591 struct page *page = NULL;
4592 enum mc_target_type ret = MC_TARGET_NONE;
4593 swp_entry_t ent = { .val = 0 };
4595 if (pte_present(ptent))
4596 page = mc_handle_present_pte(vma, addr, ptent);
4597 else if (is_swap_pte(ptent))
4598 page = mc_handle_swap_pte(vma, ptent, &ent);
4599 else if (pte_none(ptent))
4600 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4602 if (!page && !ent.val)
4606 * Do only loose check w/o serialization.
4607 * mem_cgroup_move_account() checks the page is valid or
4608 * not under LRU exclusion.
4610 if (page->mem_cgroup == mc.from) {
4611 ret = MC_TARGET_PAGE;
4613 target->page = page;
4615 if (!ret || !target)
4618 /* There is a swap entry and a page doesn't exist or isn't charged */
4619 if (ent.val && !ret &&
4620 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4621 ret = MC_TARGET_SWAP;
4628 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4630 * We don't consider swapping or file mapped pages because THP does not
4631 * support them for now.
4632 * Caller should make sure that pmd_trans_huge(pmd) is true.
4634 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4635 unsigned long addr, pmd_t pmd, union mc_target *target)
4637 struct page *page = NULL;
4638 enum mc_target_type ret = MC_TARGET_NONE;
4640 page = pmd_page(pmd);
4641 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4642 if (!(mc.flags & MOVE_ANON))
4644 if (page->mem_cgroup == mc.from) {
4645 ret = MC_TARGET_PAGE;
4648 target->page = page;
4654 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4655 unsigned long addr, pmd_t pmd, union mc_target *target)
4657 return MC_TARGET_NONE;
4661 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4662 unsigned long addr, unsigned long end,
4663 struct mm_walk *walk)
4665 struct vm_area_struct *vma = walk->vma;
4669 ptl = pmd_trans_huge_lock(pmd, vma);
4671 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4672 mc.precharge += HPAGE_PMD_NR;
4677 if (pmd_trans_unstable(pmd))
4679 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4680 for (; addr != end; pte++, addr += PAGE_SIZE)
4681 if (get_mctgt_type(vma, addr, *pte, NULL))
4682 mc.precharge++; /* increment precharge temporarily */
4683 pte_unmap_unlock(pte - 1, ptl);
4689 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4691 unsigned long precharge;
4693 struct mm_walk mem_cgroup_count_precharge_walk = {
4694 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4697 down_read(&mm->mmap_sem);
4698 walk_page_range(0, mm->highest_vm_end,
4699 &mem_cgroup_count_precharge_walk);
4700 up_read(&mm->mmap_sem);
4702 precharge = mc.precharge;
4708 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4710 unsigned long precharge = mem_cgroup_count_precharge(mm);
4712 VM_BUG_ON(mc.moving_task);
4713 mc.moving_task = current;
4714 return mem_cgroup_do_precharge(precharge);
4717 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4718 static void __mem_cgroup_clear_mc(void)
4720 struct mem_cgroup *from = mc.from;
4721 struct mem_cgroup *to = mc.to;
4723 /* we must uncharge all the leftover precharges from mc.to */
4725 cancel_charge(mc.to, mc.precharge);
4729 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4730 * we must uncharge here.
4732 if (mc.moved_charge) {
4733 cancel_charge(mc.from, mc.moved_charge);
4734 mc.moved_charge = 0;
4736 /* we must fixup refcnts and charges */
4737 if (mc.moved_swap) {
4738 /* uncharge swap account from the old cgroup */
4739 if (!mem_cgroup_is_root(mc.from))
4740 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4742 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4745 * we charged both to->memory and to->memsw, so we
4746 * should uncharge to->memory.
4748 if (!mem_cgroup_is_root(mc.to))
4749 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4751 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4752 css_put_many(&mc.to->css, mc.moved_swap);
4756 memcg_oom_recover(from);
4757 memcg_oom_recover(to);
4758 wake_up_all(&mc.waitq);
4761 static void mem_cgroup_clear_mc(void)
4763 struct mm_struct *mm = mc.mm;
4766 * we must clear moving_task before waking up waiters at the end of
4769 mc.moving_task = NULL;
4770 __mem_cgroup_clear_mc();
4771 spin_lock(&mc.lock);
4775 spin_unlock(&mc.lock);
4780 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4782 struct cgroup_subsys_state *css;
4783 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4784 struct mem_cgroup *from;
4785 struct task_struct *leader, *p;
4786 struct mm_struct *mm;
4787 unsigned long move_flags;
4790 /* charge immigration isn't supported on the default hierarchy */
4791 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4795 * Multi-process migrations only happen on the default hierarchy
4796 * where charge immigration is not used. Perform charge
4797 * immigration if @tset contains a leader and whine if there are
4801 cgroup_taskset_for_each_leader(leader, css, tset) {
4804 memcg = mem_cgroup_from_css(css);
4810 * We are now commited to this value whatever it is. Changes in this
4811 * tunable will only affect upcoming migrations, not the current one.
4812 * So we need to save it, and keep it going.
4814 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4818 from = mem_cgroup_from_task(p);
4820 VM_BUG_ON(from == memcg);
4822 mm = get_task_mm(p);
4825 /* We move charges only when we move a owner of the mm */
4826 if (mm->owner == p) {
4829 VM_BUG_ON(mc.precharge);
4830 VM_BUG_ON(mc.moved_charge);
4831 VM_BUG_ON(mc.moved_swap);
4833 spin_lock(&mc.lock);
4837 mc.flags = move_flags;
4838 spin_unlock(&mc.lock);
4839 /* We set mc.moving_task later */
4841 ret = mem_cgroup_precharge_mc(mm);
4843 mem_cgroup_clear_mc();
4850 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4853 mem_cgroup_clear_mc();
4856 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4857 unsigned long addr, unsigned long end,
4858 struct mm_walk *walk)
4861 struct vm_area_struct *vma = walk->vma;
4864 enum mc_target_type target_type;
4865 union mc_target target;
4868 ptl = pmd_trans_huge_lock(pmd, vma);
4870 if (mc.precharge < HPAGE_PMD_NR) {
4874 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4875 if (target_type == MC_TARGET_PAGE) {
4877 if (!isolate_lru_page(page)) {
4878 if (!mem_cgroup_move_account(page, true,
4880 mc.precharge -= HPAGE_PMD_NR;
4881 mc.moved_charge += HPAGE_PMD_NR;
4883 putback_lru_page(page);
4891 if (pmd_trans_unstable(pmd))
4894 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4895 for (; addr != end; addr += PAGE_SIZE) {
4896 pte_t ptent = *(pte++);
4902 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4903 case MC_TARGET_PAGE:
4906 * We can have a part of the split pmd here. Moving it
4907 * can be done but it would be too convoluted so simply
4908 * ignore such a partial THP and keep it in original
4909 * memcg. There should be somebody mapping the head.
4911 if (PageTransCompound(page))
4913 if (isolate_lru_page(page))
4915 if (!mem_cgroup_move_account(page, false,
4918 /* we uncharge from mc.from later. */
4921 putback_lru_page(page);
4922 put: /* get_mctgt_type() gets the page */
4925 case MC_TARGET_SWAP:
4927 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4929 /* we fixup refcnts and charges later. */
4937 pte_unmap_unlock(pte - 1, ptl);
4942 * We have consumed all precharges we got in can_attach().
4943 * We try charge one by one, but don't do any additional
4944 * charges to mc.to if we have failed in charge once in attach()
4947 ret = mem_cgroup_do_precharge(1);
4955 static void mem_cgroup_move_charge(void)
4957 struct mm_walk mem_cgroup_move_charge_walk = {
4958 .pmd_entry = mem_cgroup_move_charge_pte_range,
4962 lru_add_drain_all();
4964 * Signal lock_page_memcg() to take the memcg's move_lock
4965 * while we're moving its pages to another memcg. Then wait
4966 * for already started RCU-only updates to finish.
4968 atomic_inc(&mc.from->moving_account);
4971 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
4973 * Someone who are holding the mmap_sem might be waiting in
4974 * waitq. So we cancel all extra charges, wake up all waiters,
4975 * and retry. Because we cancel precharges, we might not be able
4976 * to move enough charges, but moving charge is a best-effort
4977 * feature anyway, so it wouldn't be a big problem.
4979 __mem_cgroup_clear_mc();
4984 * When we have consumed all precharges and failed in doing
4985 * additional charge, the page walk just aborts.
4987 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
4989 up_read(&mc.mm->mmap_sem);
4990 atomic_dec(&mc.from->moving_account);
4993 static void mem_cgroup_move_task(void)
4996 mem_cgroup_move_charge();
4997 mem_cgroup_clear_mc();
5000 #else /* !CONFIG_MMU */
5001 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5005 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5008 static void mem_cgroup_move_task(void)
5014 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5015 * to verify whether we're attached to the default hierarchy on each mount
5018 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5021 * use_hierarchy is forced on the default hierarchy. cgroup core
5022 * guarantees that @root doesn't have any children, so turning it
5023 * on for the root memcg is enough.
5025 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5026 root_mem_cgroup->use_hierarchy = true;
5028 root_mem_cgroup->use_hierarchy = false;
5031 static u64 memory_current_read(struct cgroup_subsys_state *css,
5034 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5036 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5039 static int memory_low_show(struct seq_file *m, void *v)
5041 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5042 unsigned long low = READ_ONCE(memcg->low);
5044 if (low == PAGE_COUNTER_MAX)
5045 seq_puts(m, "max\n");
5047 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5052 static ssize_t memory_low_write(struct kernfs_open_file *of,
5053 char *buf, size_t nbytes, loff_t off)
5055 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5059 buf = strstrip(buf);
5060 err = page_counter_memparse(buf, "max", &low);
5069 static int memory_high_show(struct seq_file *m, void *v)
5071 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5072 unsigned long high = READ_ONCE(memcg->high);
5074 if (high == PAGE_COUNTER_MAX)
5075 seq_puts(m, "max\n");
5077 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5082 static ssize_t memory_high_write(struct kernfs_open_file *of,
5083 char *buf, size_t nbytes, loff_t off)
5085 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5086 unsigned long nr_pages;
5090 buf = strstrip(buf);
5091 err = page_counter_memparse(buf, "max", &high);
5097 nr_pages = page_counter_read(&memcg->memory);
5098 if (nr_pages > high)
5099 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5102 memcg_wb_domain_size_changed(memcg);
5106 static int memory_max_show(struct seq_file *m, void *v)
5108 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5109 unsigned long max = READ_ONCE(memcg->memory.limit);
5111 if (max == PAGE_COUNTER_MAX)
5112 seq_puts(m, "max\n");
5114 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5119 static ssize_t memory_max_write(struct kernfs_open_file *of,
5120 char *buf, size_t nbytes, loff_t off)
5122 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5123 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5124 bool drained = false;
5128 buf = strstrip(buf);
5129 err = page_counter_memparse(buf, "max", &max);
5133 xchg(&memcg->memory.limit, max);
5136 unsigned long nr_pages = page_counter_read(&memcg->memory);
5138 if (nr_pages <= max)
5141 if (signal_pending(current)) {
5147 drain_all_stock(memcg);
5153 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5159 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5160 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5164 memcg_wb_domain_size_changed(memcg);
5168 static int memory_events_show(struct seq_file *m, void *v)
5170 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5172 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5173 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5174 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5175 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5180 static int memory_stat_show(struct seq_file *m, void *v)
5182 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5183 unsigned long stat[MEMCG_NR_STAT];
5184 unsigned long events[MEMCG_NR_EVENTS];
5188 * Provide statistics on the state of the memory subsystem as
5189 * well as cumulative event counters that show past behavior.
5191 * This list is ordered following a combination of these gradients:
5192 * 1) generic big picture -> specifics and details
5193 * 2) reflecting userspace activity -> reflecting kernel heuristics
5195 * Current memory state:
5198 tree_stat(memcg, stat);
5199 tree_events(memcg, events);
5201 seq_printf(m, "anon %llu\n",
5202 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5203 seq_printf(m, "file %llu\n",
5204 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5205 seq_printf(m, "kernel_stack %llu\n",
5206 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5207 seq_printf(m, "slab %llu\n",
5208 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5209 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5210 seq_printf(m, "sock %llu\n",
5211 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5213 seq_printf(m, "file_mapped %llu\n",
5214 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5215 seq_printf(m, "file_dirty %llu\n",
5216 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5217 seq_printf(m, "file_writeback %llu\n",
5218 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5220 for (i = 0; i < NR_LRU_LISTS; i++) {
5221 struct mem_cgroup *mi;
5222 unsigned long val = 0;
5224 for_each_mem_cgroup_tree(mi, memcg)
5225 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5226 seq_printf(m, "%s %llu\n",
5227 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5230 seq_printf(m, "slab_reclaimable %llu\n",
5231 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5232 seq_printf(m, "slab_unreclaimable %llu\n",
5233 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5235 /* Accumulated memory events */
5237 seq_printf(m, "pgfault %lu\n",
5238 events[MEM_CGROUP_EVENTS_PGFAULT]);
5239 seq_printf(m, "pgmajfault %lu\n",
5240 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5245 static struct cftype memory_files[] = {
5248 .flags = CFTYPE_NOT_ON_ROOT,
5249 .read_u64 = memory_current_read,
5253 .flags = CFTYPE_NOT_ON_ROOT,
5254 .seq_show = memory_low_show,
5255 .write = memory_low_write,
5259 .flags = CFTYPE_NOT_ON_ROOT,
5260 .seq_show = memory_high_show,
5261 .write = memory_high_write,
5265 .flags = CFTYPE_NOT_ON_ROOT,
5266 .seq_show = memory_max_show,
5267 .write = memory_max_write,
5271 .flags = CFTYPE_NOT_ON_ROOT,
5272 .file_offset = offsetof(struct mem_cgroup, events_file),
5273 .seq_show = memory_events_show,
5277 .flags = CFTYPE_NOT_ON_ROOT,
5278 .seq_show = memory_stat_show,
5283 struct cgroup_subsys memory_cgrp_subsys = {
5284 .css_alloc = mem_cgroup_css_alloc,
5285 .css_online = mem_cgroup_css_online,
5286 .css_offline = mem_cgroup_css_offline,
5287 .css_released = mem_cgroup_css_released,
5288 .css_free = mem_cgroup_css_free,
5289 .css_reset = mem_cgroup_css_reset,
5290 .can_attach = mem_cgroup_can_attach,
5291 .cancel_attach = mem_cgroup_cancel_attach,
5292 .post_attach = mem_cgroup_move_task,
5293 .bind = mem_cgroup_bind,
5294 .dfl_cftypes = memory_files,
5295 .legacy_cftypes = mem_cgroup_legacy_files,
5300 * mem_cgroup_low - check if memory consumption is below the normal range
5301 * @root: the highest ancestor to consider
5302 * @memcg: the memory cgroup to check
5304 * Returns %true if memory consumption of @memcg, and that of all
5305 * configurable ancestors up to @root, is below the normal range.
5307 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5309 if (mem_cgroup_disabled())
5313 * The toplevel group doesn't have a configurable range, so
5314 * it's never low when looked at directly, and it is not
5315 * considered an ancestor when assessing the hierarchy.
5318 if (memcg == root_mem_cgroup)
5321 if (page_counter_read(&memcg->memory) >= memcg->low)
5324 while (memcg != root) {
5325 memcg = parent_mem_cgroup(memcg);
5327 if (memcg == root_mem_cgroup)
5330 if (page_counter_read(&memcg->memory) >= memcg->low)
5337 * mem_cgroup_try_charge - try charging a page
5338 * @page: page to charge
5339 * @mm: mm context of the victim
5340 * @gfp_mask: reclaim mode
5341 * @memcgp: charged memcg return
5342 * @compound: charge the page as compound or small page
5344 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5345 * pages according to @gfp_mask if necessary.
5347 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5348 * Otherwise, an error code is returned.
5350 * After page->mapping has been set up, the caller must finalize the
5351 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5352 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5354 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5355 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5358 struct mem_cgroup *memcg = NULL;
5359 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5362 if (mem_cgroup_disabled())
5365 if (PageSwapCache(page)) {
5367 * Every swap fault against a single page tries to charge the
5368 * page, bail as early as possible. shmem_unuse() encounters
5369 * already charged pages, too. The USED bit is protected by
5370 * the page lock, which serializes swap cache removal, which
5371 * in turn serializes uncharging.
5373 VM_BUG_ON_PAGE(!PageLocked(page), page);
5374 if (page->mem_cgroup)
5377 if (do_swap_account) {
5378 swp_entry_t ent = { .val = page_private(page), };
5379 unsigned short id = lookup_swap_cgroup_id(ent);
5382 memcg = mem_cgroup_from_id(id);
5383 if (memcg && !css_tryget_online(&memcg->css))
5390 memcg = get_mem_cgroup_from_mm(mm);
5392 ret = try_charge(memcg, gfp_mask, nr_pages);
5394 css_put(&memcg->css);
5401 * mem_cgroup_commit_charge - commit a page charge
5402 * @page: page to charge
5403 * @memcg: memcg to charge the page to
5404 * @lrucare: page might be on LRU already
5405 * @compound: charge the page as compound or small page
5407 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5408 * after page->mapping has been set up. This must happen atomically
5409 * as part of the page instantiation, i.e. under the page table lock
5410 * for anonymous pages, under the page lock for page and swap cache.
5412 * In addition, the page must not be on the LRU during the commit, to
5413 * prevent racing with task migration. If it might be, use @lrucare.
5415 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5417 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5418 bool lrucare, bool compound)
5420 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5422 VM_BUG_ON_PAGE(!page->mapping, page);
5423 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5425 if (mem_cgroup_disabled())
5428 * Swap faults will attempt to charge the same page multiple
5429 * times. But reuse_swap_page() might have removed the page
5430 * from swapcache already, so we can't check PageSwapCache().
5435 commit_charge(page, memcg, lrucare);
5437 local_lock_irq(event_lock);
5438 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5439 memcg_check_events(memcg, page);
5440 local_unlock_irq(event_lock);
5442 if (do_memsw_account() && PageSwapCache(page)) {
5443 swp_entry_t entry = { .val = page_private(page) };
5445 * The swap entry might not get freed for a long time,
5446 * let's not wait for it. The page already received a
5447 * memory+swap charge, drop the swap entry duplicate.
5449 mem_cgroup_uncharge_swap(entry);
5454 * mem_cgroup_cancel_charge - cancel a page charge
5455 * @page: page to charge
5456 * @memcg: memcg to charge the page to
5457 * @compound: charge the page as compound or small page
5459 * Cancel a charge transaction started by mem_cgroup_try_charge().
5461 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5464 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5466 if (mem_cgroup_disabled())
5469 * Swap faults will attempt to charge the same page multiple
5470 * times. But reuse_swap_page() might have removed the page
5471 * from swapcache already, so we can't check PageSwapCache().
5476 cancel_charge(memcg, nr_pages);
5479 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5480 unsigned long nr_anon, unsigned long nr_file,
5481 unsigned long nr_huge, unsigned long nr_kmem,
5482 struct page *dummy_page)
5484 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5485 unsigned long flags;
5487 if (!mem_cgroup_is_root(memcg)) {
5488 page_counter_uncharge(&memcg->memory, nr_pages);
5489 if (do_memsw_account())
5490 page_counter_uncharge(&memcg->memsw, nr_pages);
5491 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5492 page_counter_uncharge(&memcg->kmem, nr_kmem);
5493 memcg_oom_recover(memcg);
5496 local_lock_irqsave(event_lock, flags);
5497 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5498 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5499 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5500 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5501 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5502 memcg_check_events(memcg, dummy_page);
5503 local_unlock_irqrestore(event_lock, flags);
5505 if (!mem_cgroup_is_root(memcg))
5506 css_put_many(&memcg->css, nr_pages);
5509 static void uncharge_list(struct list_head *page_list)
5511 struct mem_cgroup *memcg = NULL;
5512 unsigned long nr_anon = 0;
5513 unsigned long nr_file = 0;
5514 unsigned long nr_huge = 0;
5515 unsigned long nr_kmem = 0;
5516 unsigned long pgpgout = 0;
5517 struct list_head *next;
5521 * Note that the list can be a single page->lru; hence the
5522 * do-while loop instead of a simple list_for_each_entry().
5524 next = page_list->next;
5526 page = list_entry(next, struct page, lru);
5527 next = page->lru.next;
5529 VM_BUG_ON_PAGE(PageLRU(page), page);
5530 VM_BUG_ON_PAGE(page_count(page), page);
5532 if (!page->mem_cgroup)
5536 * Nobody should be changing or seriously looking at
5537 * page->mem_cgroup at this point, we have fully
5538 * exclusive access to the page.
5541 if (memcg != page->mem_cgroup) {
5543 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5544 nr_huge, nr_kmem, page);
5545 pgpgout = nr_anon = nr_file =
5546 nr_huge = nr_kmem = 0;
5548 memcg = page->mem_cgroup;
5551 if (!PageKmemcg(page)) {
5552 unsigned int nr_pages = 1;
5554 if (PageTransHuge(page)) {
5555 nr_pages <<= compound_order(page);
5556 nr_huge += nr_pages;
5559 nr_anon += nr_pages;
5561 nr_file += nr_pages;
5564 nr_kmem += 1 << compound_order(page);
5565 __ClearPageKmemcg(page);
5568 page->mem_cgroup = NULL;
5569 } while (next != page_list);
5572 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5573 nr_huge, nr_kmem, page);
5577 * mem_cgroup_uncharge - uncharge a page
5578 * @page: page to uncharge
5580 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5581 * mem_cgroup_commit_charge().
5583 void mem_cgroup_uncharge(struct page *page)
5585 if (mem_cgroup_disabled())
5588 /* Don't touch page->lru of any random page, pre-check: */
5589 if (!page->mem_cgroup)
5592 INIT_LIST_HEAD(&page->lru);
5593 uncharge_list(&page->lru);
5597 * mem_cgroup_uncharge_list - uncharge a list of page
5598 * @page_list: list of pages to uncharge
5600 * Uncharge a list of pages previously charged with
5601 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5603 void mem_cgroup_uncharge_list(struct list_head *page_list)
5605 if (mem_cgroup_disabled())
5608 if (!list_empty(page_list))
5609 uncharge_list(page_list);
5613 * mem_cgroup_migrate - charge a page's replacement
5614 * @oldpage: currently circulating page
5615 * @newpage: replacement page
5617 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5618 * be uncharged upon free.
5620 * Both pages must be locked, @newpage->mapping must be set up.
5622 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5624 struct mem_cgroup *memcg;
5625 unsigned int nr_pages;
5627 unsigned long flags;
5629 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5630 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5631 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5632 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5635 if (mem_cgroup_disabled())
5638 /* Page cache replacement: new page already charged? */
5639 if (newpage->mem_cgroup)
5642 /* Swapcache readahead pages can get replaced before being charged */
5643 memcg = oldpage->mem_cgroup;
5647 /* Force-charge the new page. The old one will be freed soon */
5648 compound = PageTransHuge(newpage);
5649 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5651 page_counter_charge(&memcg->memory, nr_pages);
5652 if (do_memsw_account())
5653 page_counter_charge(&memcg->memsw, nr_pages);
5654 css_get_many(&memcg->css, nr_pages);
5656 commit_charge(newpage, memcg, false);
5658 local_lock_irqsave(event_lock, flags);
5659 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5660 memcg_check_events(memcg, newpage);
5661 local_unlock_irqrestore(event_lock, flags);
5664 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5665 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5667 void mem_cgroup_sk_alloc(struct sock *sk)
5669 struct mem_cgroup *memcg;
5671 if (!mem_cgroup_sockets_enabled)
5675 * Socket cloning can throw us here with sk_memcg already
5676 * filled. It won't however, necessarily happen from
5677 * process context. So the test for root memcg given
5678 * the current task's memcg won't help us in this case.
5680 * Respecting the original socket's memcg is a better
5681 * decision in this case.
5684 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5685 css_get(&sk->sk_memcg->css);
5690 memcg = mem_cgroup_from_task(current);
5691 if (memcg == root_mem_cgroup)
5693 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5695 if (css_tryget_online(&memcg->css))
5696 sk->sk_memcg = memcg;
5701 void mem_cgroup_sk_free(struct sock *sk)
5704 css_put(&sk->sk_memcg->css);
5708 * mem_cgroup_charge_skmem - charge socket memory
5709 * @memcg: memcg to charge
5710 * @nr_pages: number of pages to charge
5712 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5713 * @memcg's configured limit, %false if the charge had to be forced.
5715 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5717 gfp_t gfp_mask = GFP_KERNEL;
5719 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5720 struct page_counter *fail;
5722 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5723 memcg->tcpmem_pressure = 0;
5726 page_counter_charge(&memcg->tcpmem, nr_pages);
5727 memcg->tcpmem_pressure = 1;
5731 /* Don't block in the packet receive path */
5733 gfp_mask = GFP_NOWAIT;
5735 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5737 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5740 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5745 * mem_cgroup_uncharge_skmem - uncharge socket memory
5746 * @memcg - memcg to uncharge
5747 * @nr_pages - number of pages to uncharge
5749 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5751 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5752 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5756 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5758 page_counter_uncharge(&memcg->memory, nr_pages);
5759 css_put_many(&memcg->css, nr_pages);
5762 static int __init cgroup_memory(char *s)
5766 while ((token = strsep(&s, ",")) != NULL) {
5769 if (!strcmp(token, "nosocket"))
5770 cgroup_memory_nosocket = true;
5771 if (!strcmp(token, "nokmem"))
5772 cgroup_memory_nokmem = true;
5776 __setup("cgroup.memory=", cgroup_memory);
5779 * subsys_initcall() for memory controller.
5781 * Some parts like hotcpu_notifier() have to be initialized from this context
5782 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5783 * everything that doesn't depend on a specific mem_cgroup structure should
5784 * be initialized from here.
5786 static int __init mem_cgroup_init(void)
5790 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5792 for_each_possible_cpu(cpu)
5793 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5796 for_each_node(node) {
5797 struct mem_cgroup_tree_per_node *rtpn;
5799 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5800 node_online(node) ? node : NUMA_NO_NODE);
5802 rtpn->rb_root = RB_ROOT;
5803 spin_lock_init(&rtpn->lock);
5804 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5809 subsys_initcall(mem_cgroup_init);
5811 #ifdef CONFIG_MEMCG_SWAP
5812 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5814 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5816 * The root cgroup cannot be destroyed, so it's refcount must
5819 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5823 memcg = parent_mem_cgroup(memcg);
5825 memcg = root_mem_cgroup;
5831 * mem_cgroup_swapout - transfer a memsw charge to swap
5832 * @page: page whose memsw charge to transfer
5833 * @entry: swap entry to move the charge to
5835 * Transfer the memsw charge of @page to @entry.
5837 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5839 struct mem_cgroup *memcg, *swap_memcg;
5840 unsigned short oldid;
5841 unsigned long flags;
5843 VM_BUG_ON_PAGE(PageLRU(page), page);
5844 VM_BUG_ON_PAGE(page_count(page), page);
5846 if (!do_memsw_account())
5849 memcg = page->mem_cgroup;
5851 /* Readahead page, never charged */
5856 * In case the memcg owning these pages has been offlined and doesn't
5857 * have an ID allocated to it anymore, charge the closest online
5858 * ancestor for the swap instead and transfer the memory+swap charge.
5860 swap_memcg = mem_cgroup_id_get_online(memcg);
5861 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5862 VM_BUG_ON_PAGE(oldid, page);
5863 mem_cgroup_swap_statistics(swap_memcg, true);
5865 page->mem_cgroup = NULL;
5867 if (!mem_cgroup_is_root(memcg))
5868 page_counter_uncharge(&memcg->memory, 1);
5870 if (memcg != swap_memcg) {
5871 if (!mem_cgroup_is_root(swap_memcg))
5872 page_counter_charge(&swap_memcg->memsw, 1);
5873 page_counter_uncharge(&memcg->memsw, 1);
5877 * Interrupts should be disabled here because the caller holds the
5878 * mapping->tree_lock lock which is taken with interrupts-off. It is
5879 * important here to have the interrupts disabled because it is the
5880 * only synchronisation we have for udpating the per-CPU variables.
5882 local_lock_irqsave(event_lock, flags);
5883 #ifndef CONFIG_PREEMPT_RT_BASE
5884 VM_BUG_ON(!irqs_disabled());
5886 mem_cgroup_charge_statistics(memcg, page, false, -1);
5887 memcg_check_events(memcg, page);
5889 if (!mem_cgroup_is_root(memcg))
5890 css_put(&memcg->css);
5891 local_unlock_irqrestore(event_lock, flags);
5895 * mem_cgroup_try_charge_swap - try charging a swap entry
5896 * @page: page being added to swap
5897 * @entry: swap entry to charge
5899 * Try to charge @entry to the memcg that @page belongs to.
5901 * Returns 0 on success, -ENOMEM on failure.
5903 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5905 struct mem_cgroup *memcg;
5906 struct page_counter *counter;
5907 unsigned short oldid;
5909 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5912 memcg = page->mem_cgroup;
5914 /* Readahead page, never charged */
5918 memcg = mem_cgroup_id_get_online(memcg);
5920 if (!mem_cgroup_is_root(memcg) &&
5921 !page_counter_try_charge(&memcg->swap, 1, &counter)) {
5922 mem_cgroup_id_put(memcg);
5926 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5927 VM_BUG_ON_PAGE(oldid, page);
5928 mem_cgroup_swap_statistics(memcg, true);
5934 * mem_cgroup_uncharge_swap - uncharge a swap entry
5935 * @entry: swap entry to uncharge
5937 * Drop the swap charge associated with @entry.
5939 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5941 struct mem_cgroup *memcg;
5944 if (!do_swap_account)
5947 id = swap_cgroup_record(entry, 0);
5949 memcg = mem_cgroup_from_id(id);
5951 if (!mem_cgroup_is_root(memcg)) {
5952 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5953 page_counter_uncharge(&memcg->swap, 1);
5955 page_counter_uncharge(&memcg->memsw, 1);
5957 mem_cgroup_swap_statistics(memcg, false);
5958 mem_cgroup_id_put(memcg);
5963 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5965 long nr_swap_pages = get_nr_swap_pages();
5967 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5968 return nr_swap_pages;
5969 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5970 nr_swap_pages = min_t(long, nr_swap_pages,
5971 READ_ONCE(memcg->swap.limit) -
5972 page_counter_read(&memcg->swap));
5973 return nr_swap_pages;
5976 bool mem_cgroup_swap_full(struct page *page)
5978 struct mem_cgroup *memcg;
5980 VM_BUG_ON_PAGE(!PageLocked(page), page);
5984 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5987 memcg = page->mem_cgroup;
5991 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5992 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
5998 /* for remember boot option*/
5999 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6000 static int really_do_swap_account __initdata = 1;
6002 static int really_do_swap_account __initdata;
6005 static int __init enable_swap_account(char *s)
6007 if (!strcmp(s, "1"))
6008 really_do_swap_account = 1;
6009 else if (!strcmp(s, "0"))
6010 really_do_swap_account = 0;
6013 __setup("swapaccount=", enable_swap_account);
6015 static u64 swap_current_read(struct cgroup_subsys_state *css,
6018 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6020 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6023 static int swap_max_show(struct seq_file *m, void *v)
6025 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6026 unsigned long max = READ_ONCE(memcg->swap.limit);
6028 if (max == PAGE_COUNTER_MAX)
6029 seq_puts(m, "max\n");
6031 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6036 static ssize_t swap_max_write(struct kernfs_open_file *of,
6037 char *buf, size_t nbytes, loff_t off)
6039 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6043 buf = strstrip(buf);
6044 err = page_counter_memparse(buf, "max", &max);
6048 mutex_lock(&memcg_limit_mutex);
6049 err = page_counter_limit(&memcg->swap, max);
6050 mutex_unlock(&memcg_limit_mutex);
6057 static struct cftype swap_files[] = {
6059 .name = "swap.current",
6060 .flags = CFTYPE_NOT_ON_ROOT,
6061 .read_u64 = swap_current_read,
6065 .flags = CFTYPE_NOT_ON_ROOT,
6066 .seq_show = swap_max_show,
6067 .write = swap_max_write,
6072 static struct cftype memsw_cgroup_files[] = {
6074 .name = "memsw.usage_in_bytes",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6076 .read_u64 = mem_cgroup_read_u64,
6079 .name = "memsw.max_usage_in_bytes",
6080 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6081 .write = mem_cgroup_reset,
6082 .read_u64 = mem_cgroup_read_u64,
6085 .name = "memsw.limit_in_bytes",
6086 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6087 .write = mem_cgroup_write,
6088 .read_u64 = mem_cgroup_read_u64,
6091 .name = "memsw.failcnt",
6092 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6093 .write = mem_cgroup_reset,
6094 .read_u64 = mem_cgroup_read_u64,
6096 { }, /* terminate */
6099 static int __init mem_cgroup_swap_init(void)
6101 if (!mem_cgroup_disabled() && really_do_swap_account) {
6102 do_swap_account = 1;
6103 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6105 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6106 memsw_cgroup_files));
6110 subsys_initcall(mem_cgroup_swap_init);
6112 #endif /* CONFIG_MEMCG_SWAP */