1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 struct mem_cgroup *root_mem_cgroup __read_mostly;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char * const mem_cgroup_stat_names[] = {
111 static const char * const mem_cgroup_events_names[] = {
118 static const char * const mem_cgroup_lru_names[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_zone {
136 struct rb_root rb_root;
140 struct mem_cgroup_tree_per_node {
141 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
144 struct mem_cgroup_tree {
145 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
148 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
151 struct mem_cgroup_eventfd_list {
152 struct list_head list;
153 struct eventfd_ctx *eventfd;
157 * cgroup_event represents events which userspace want to receive.
159 struct mem_cgroup_event {
161 * memcg which the event belongs to.
163 struct mem_cgroup *memcg;
165 * eventfd to signal userspace about the event.
167 struct eventfd_ctx *eventfd;
169 * Each of these stored in a list by the cgroup.
171 struct list_head list;
173 * register_event() callback will be used to add new userspace
174 * waiter for changes related to this event. Use eventfd_signal()
175 * on eventfd to send notification to userspace.
177 int (*register_event)(struct mem_cgroup *memcg,
178 struct eventfd_ctx *eventfd, const char *args);
180 * unregister_event() callback will be called when userspace closes
181 * the eventfd or on cgroup removing. This callback must be set,
182 * if you want provide notification functionality.
184 void (*unregister_event)(struct mem_cgroup *memcg,
185 struct eventfd_ctx *eventfd);
187 * All fields below needed to unregister event when
188 * userspace closes eventfd.
191 wait_queue_head_t *wqh;
193 struct work_struct remove;
196 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
197 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
199 /* Stuffs for move charges at task migration. */
201 * Types of charges to be moved.
203 #define MOVE_ANON 0x1U
204 #define MOVE_FILE 0x2U
205 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
207 /* "mc" and its members are protected by cgroup_mutex */
208 static struct move_charge_struct {
209 spinlock_t lock; /* for from, to */
210 struct mem_cgroup *from;
211 struct mem_cgroup *to;
213 unsigned long precharge;
214 unsigned long moved_charge;
215 unsigned long moved_swap;
216 struct task_struct *moving_task; /* a task moving charges */
217 wait_queue_head_t waitq; /* a waitq for other context */
219 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
220 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
224 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
225 * limit reclaim to prevent infinite loops, if they ever occur.
227 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
228 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
231 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
232 MEM_CGROUP_CHARGE_TYPE_ANON,
233 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
234 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
238 /* for encoding cft->private value on file */
247 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
248 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
249 #define MEMFILE_ATTR(val) ((val) & 0xffff)
250 /* Used for OOM nofiier */
251 #define OOM_CONTROL (0)
253 /* Some nice accessors for the vmpressure. */
254 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
257 memcg = root_mem_cgroup;
258 return &memcg->vmpressure;
261 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
263 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
266 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
268 return (memcg == root_mem_cgroup);
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 */
325 static struct mem_cgroup_per_zone *
326 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
328 int nid = zone_to_nid(zone);
329 int zid = zone_idx(zone);
331 return &memcg->nodeinfo[nid]->zoneinfo[zid];
335 * mem_cgroup_css_from_page - css of the memcg associated with a page
336 * @page: page of interest
338 * If memcg is bound to the default hierarchy, css of the memcg associated
339 * with @page is returned. The returned css remains associated with @page
340 * until it is released.
342 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
345 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
347 struct mem_cgroup *memcg;
349 memcg = page->mem_cgroup;
351 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
352 memcg = root_mem_cgroup;
358 * page_cgroup_ino - return inode number of the memcg a page is charged to
361 * Look up the closest online ancestor of the memory cgroup @page is charged to
362 * and return its inode number or 0 if @page is not charged to any cgroup. It
363 * is safe to call this function without holding a reference to @page.
365 * Note, this function is inherently racy, because there is nothing to prevent
366 * the cgroup inode from getting torn down and potentially reallocated a moment
367 * after page_cgroup_ino() returns, so it only should be used by callers that
368 * do not care (such as procfs interfaces).
370 ino_t page_cgroup_ino(struct page *page)
372 struct mem_cgroup *memcg;
373 unsigned long ino = 0;
376 memcg = READ_ONCE(page->mem_cgroup);
377 while (memcg && !(memcg->css.flags & CSS_ONLINE))
378 memcg = parent_mem_cgroup(memcg);
380 ino = cgroup_ino(memcg->css.cgroup);
385 static struct mem_cgroup_per_zone *
386 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
388 int nid = page_to_nid(page);
389 int zid = page_zonenum(page);
391 return &memcg->nodeinfo[nid]->zoneinfo[zid];
394 static struct mem_cgroup_tree_per_zone *
395 soft_limit_tree_node_zone(int nid, int zid)
397 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
400 static struct mem_cgroup_tree_per_zone *
401 soft_limit_tree_from_page(struct page *page)
403 int nid = page_to_nid(page);
404 int zid = page_zonenum(page);
406 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
409 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
410 struct mem_cgroup_tree_per_zone *mctz,
411 unsigned long new_usage_in_excess)
413 struct rb_node **p = &mctz->rb_root.rb_node;
414 struct rb_node *parent = NULL;
415 struct mem_cgroup_per_zone *mz_node;
420 mz->usage_in_excess = new_usage_in_excess;
421 if (!mz->usage_in_excess)
425 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
427 if (mz->usage_in_excess < mz_node->usage_in_excess)
430 * We can't avoid mem cgroups that are over their soft
431 * limit by the same amount
433 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
436 rb_link_node(&mz->tree_node, parent, p);
437 rb_insert_color(&mz->tree_node, &mctz->rb_root);
441 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
442 struct mem_cgroup_tree_per_zone *mctz)
446 rb_erase(&mz->tree_node, &mctz->rb_root);
450 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
451 struct mem_cgroup_tree_per_zone *mctz)
455 spin_lock_irqsave(&mctz->lock, flags);
456 __mem_cgroup_remove_exceeded(mz, mctz);
457 spin_unlock_irqrestore(&mctz->lock, flags);
460 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
462 unsigned long nr_pages = page_counter_read(&memcg->memory);
463 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
464 unsigned long excess = 0;
466 if (nr_pages > soft_limit)
467 excess = nr_pages - soft_limit;
472 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
474 unsigned long excess;
475 struct mem_cgroup_per_zone *mz;
476 struct mem_cgroup_tree_per_zone *mctz;
478 mctz = soft_limit_tree_from_page(page);
480 * Necessary to update all ancestors when hierarchy is used.
481 * because their event counter is not touched.
483 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
484 mz = mem_cgroup_page_zoneinfo(memcg, page);
485 excess = soft_limit_excess(memcg);
487 * We have to update the tree if mz is on RB-tree or
488 * mem is over its softlimit.
490 if (excess || mz->on_tree) {
493 spin_lock_irqsave(&mctz->lock, flags);
494 /* if on-tree, remove it */
496 __mem_cgroup_remove_exceeded(mz, mctz);
498 * Insert again. mz->usage_in_excess will be updated.
499 * If excess is 0, no tree ops.
501 __mem_cgroup_insert_exceeded(mz, mctz, excess);
502 spin_unlock_irqrestore(&mctz->lock, flags);
507 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
509 struct mem_cgroup_tree_per_zone *mctz;
510 struct mem_cgroup_per_zone *mz;
514 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
515 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
516 mctz = soft_limit_tree_node_zone(nid, zid);
517 mem_cgroup_remove_exceeded(mz, mctz);
522 static struct mem_cgroup_per_zone *
523 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
525 struct rb_node *rightmost = NULL;
526 struct mem_cgroup_per_zone *mz;
530 rightmost = rb_last(&mctz->rb_root);
532 goto done; /* Nothing to reclaim from */
534 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
536 * Remove the node now but someone else can add it back,
537 * we will to add it back at the end of reclaim to its correct
538 * position in the tree.
540 __mem_cgroup_remove_exceeded(mz, mctz);
541 if (!soft_limit_excess(mz->memcg) ||
542 !css_tryget_online(&mz->memcg->css))
548 static struct mem_cgroup_per_zone *
549 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
551 struct mem_cgroup_per_zone *mz;
553 spin_lock_irq(&mctz->lock);
554 mz = __mem_cgroup_largest_soft_limit_node(mctz);
555 spin_unlock_irq(&mctz->lock);
560 * Return page count for single (non recursive) @memcg.
562 * Implementation Note: reading percpu statistics for memcg.
564 * Both of vmstat[] and percpu_counter has threshold and do periodic
565 * synchronization to implement "quick" read. There are trade-off between
566 * reading cost and precision of value. Then, we may have a chance to implement
567 * a periodic synchronization of counter in memcg's counter.
569 * But this _read() function is used for user interface now. The user accounts
570 * memory usage by memory cgroup and he _always_ requires exact value because
571 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
572 * have to visit all online cpus and make sum. So, for now, unnecessary
573 * synchronization is not implemented. (just implemented for cpu hotplug)
575 * If there are kernel internal actions which can make use of some not-exact
576 * value, and reading all cpu value can be performance bottleneck in some
577 * common workload, threshold and synchronization as vmstat[] should be
581 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
586 /* Per-cpu values can be negative, use a signed accumulator */
587 for_each_possible_cpu(cpu)
588 val += per_cpu(memcg->stat->count[idx], cpu);
590 * Summing races with updates, so val may be negative. Avoid exposing
591 * transient negative values.
598 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
599 enum mem_cgroup_events_index idx)
601 unsigned long val = 0;
604 for_each_possible_cpu(cpu)
605 val += per_cpu(memcg->stat->events[idx], cpu);
609 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
611 bool compound, int nr_pages)
614 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
615 * counted as CACHE even if it's on ANON LRU.
618 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
621 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
625 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
626 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
630 /* pagein of a big page is an event. So, ignore page size */
632 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
634 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
635 nr_pages = -nr_pages; /* for event */
638 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
641 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
642 int nid, unsigned int lru_mask)
644 unsigned long nr = 0;
647 VM_BUG_ON((unsigned)nid >= nr_node_ids);
649 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
650 struct mem_cgroup_per_zone *mz;
654 if (!(BIT(lru) & lru_mask))
656 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
657 nr += mz->lru_size[lru];
663 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
664 unsigned int lru_mask)
666 unsigned long nr = 0;
669 for_each_node_state(nid, N_MEMORY)
670 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
674 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
675 enum mem_cgroup_events_target target)
677 unsigned long val, next;
679 val = __this_cpu_read(memcg->stat->nr_page_events);
680 next = __this_cpu_read(memcg->stat->targets[target]);
681 /* from time_after() in jiffies.h */
682 if ((long)next - (long)val < 0) {
684 case MEM_CGROUP_TARGET_THRESH:
685 next = val + THRESHOLDS_EVENTS_TARGET;
687 case MEM_CGROUP_TARGET_SOFTLIMIT:
688 next = val + SOFTLIMIT_EVENTS_TARGET;
690 case MEM_CGROUP_TARGET_NUMAINFO:
691 next = val + NUMAINFO_EVENTS_TARGET;
696 __this_cpu_write(memcg->stat->targets[target], next);
703 * Check events in order.
706 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
708 /* threshold event is triggered in finer grain than soft limit */
709 if (unlikely(mem_cgroup_event_ratelimit(memcg,
710 MEM_CGROUP_TARGET_THRESH))) {
712 bool do_numainfo __maybe_unused;
714 do_softlimit = mem_cgroup_event_ratelimit(memcg,
715 MEM_CGROUP_TARGET_SOFTLIMIT);
717 do_numainfo = mem_cgroup_event_ratelimit(memcg,
718 MEM_CGROUP_TARGET_NUMAINFO);
720 mem_cgroup_threshold(memcg);
721 if (unlikely(do_softlimit))
722 mem_cgroup_update_tree(memcg, page);
724 if (unlikely(do_numainfo))
725 atomic_inc(&memcg->numainfo_events);
730 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
733 * mm_update_next_owner() may clear mm->owner to NULL
734 * if it races with swapoff, page migration, etc.
735 * So this can be called with p == NULL.
740 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
742 EXPORT_SYMBOL(mem_cgroup_from_task);
744 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
746 struct mem_cgroup *memcg = NULL;
751 * Page cache insertions can happen withou an
752 * actual mm context, e.g. during disk probing
753 * on boot, loopback IO, acct() writes etc.
756 memcg = root_mem_cgroup;
758 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
759 if (unlikely(!memcg))
760 memcg = root_mem_cgroup;
762 } while (!css_tryget_online(&memcg->css));
768 * mem_cgroup_iter - iterate over memory cgroup hierarchy
769 * @root: hierarchy root
770 * @prev: previously returned memcg, NULL on first invocation
771 * @reclaim: cookie for shared reclaim walks, NULL for full walks
773 * Returns references to children of the hierarchy below @root, or
774 * @root itself, or %NULL after a full round-trip.
776 * Caller must pass the return value in @prev on subsequent
777 * invocations for reference counting, or use mem_cgroup_iter_break()
778 * to cancel a hierarchy walk before the round-trip is complete.
780 * Reclaimers can specify a zone and a priority level in @reclaim to
781 * divide up the memcgs in the hierarchy among all concurrent
782 * reclaimers operating on the same zone and priority.
784 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
785 struct mem_cgroup *prev,
786 struct mem_cgroup_reclaim_cookie *reclaim)
788 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
789 struct cgroup_subsys_state *css = NULL;
790 struct mem_cgroup *memcg = NULL;
791 struct mem_cgroup *pos = NULL;
793 if (mem_cgroup_disabled())
797 root = root_mem_cgroup;
799 if (prev && !reclaim)
802 if (!root->use_hierarchy && root != root_mem_cgroup) {
811 struct mem_cgroup_per_zone *mz;
813 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
814 iter = &mz->iter[reclaim->priority];
816 if (prev && reclaim->generation != iter->generation)
820 pos = READ_ONCE(iter->position);
821 if (!pos || css_tryget(&pos->css))
824 * css reference reached zero, so iter->position will
825 * be cleared by ->css_released. However, we should not
826 * rely on this happening soon, because ->css_released
827 * is called from a work queue, and by busy-waiting we
828 * might block it. So we clear iter->position right
831 (void)cmpxchg(&iter->position, pos, NULL);
839 css = css_next_descendant_pre(css, &root->css);
842 * Reclaimers share the hierarchy walk, and a
843 * new one might jump in right at the end of
844 * the hierarchy - make sure they see at least
845 * one group and restart from the beginning.
853 * Verify the css and acquire a reference. The root
854 * is provided by the caller, so we know it's alive
855 * and kicking, and don't take an extra reference.
857 memcg = mem_cgroup_from_css(css);
859 if (css == &root->css)
870 * The position could have already been updated by a competing
871 * thread, so check that the value hasn't changed since we read
872 * it to avoid reclaiming from the same cgroup twice.
874 (void)cmpxchg(&iter->position, pos, memcg);
882 reclaim->generation = iter->generation;
888 if (prev && prev != root)
895 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
896 * @root: hierarchy root
897 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
899 void mem_cgroup_iter_break(struct mem_cgroup *root,
900 struct mem_cgroup *prev)
903 root = root_mem_cgroup;
904 if (prev && prev != root)
908 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
910 struct mem_cgroup *memcg = dead_memcg;
911 struct mem_cgroup_reclaim_iter *iter;
912 struct mem_cgroup_per_zone *mz;
916 while ((memcg = parent_mem_cgroup(memcg))) {
918 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
919 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
920 for (i = 0; i <= DEF_PRIORITY; i++) {
922 cmpxchg(&iter->position,
931 * Iteration constructs for visiting all cgroups (under a tree). If
932 * loops are exited prematurely (break), mem_cgroup_iter_break() must
933 * be used for reference counting.
935 #define for_each_mem_cgroup_tree(iter, root) \
936 for (iter = mem_cgroup_iter(root, NULL, NULL); \
938 iter = mem_cgroup_iter(root, iter, NULL))
940 #define for_each_mem_cgroup(iter) \
941 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
943 iter = mem_cgroup_iter(NULL, iter, NULL))
946 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
947 * @zone: zone of the wanted lruvec
948 * @memcg: memcg of the wanted lruvec
950 * Returns the lru list vector holding pages for the given @zone and
951 * @mem. This can be the global zone lruvec, if the memory controller
954 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
955 struct mem_cgroup *memcg)
957 struct mem_cgroup_per_zone *mz;
958 struct lruvec *lruvec;
960 if (mem_cgroup_disabled()) {
961 lruvec = &zone->lruvec;
965 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
966 lruvec = &mz->lruvec;
969 * Since a node can be onlined after the mem_cgroup was created,
970 * we have to be prepared to initialize lruvec->zone here;
971 * and if offlined then reonlined, we need to reinitialize it.
973 if (unlikely(lruvec->zone != zone))
979 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
981 * @zone: zone of the page
983 * This function is only safe when following the LRU page isolation
984 * and putback protocol: the LRU lock must be held, and the page must
985 * either be PageLRU() or the caller must have isolated/allocated it.
987 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
989 struct mem_cgroup_per_zone *mz;
990 struct mem_cgroup *memcg;
991 struct lruvec *lruvec;
993 if (mem_cgroup_disabled()) {
994 lruvec = &zone->lruvec;
998 memcg = page->mem_cgroup;
1000 * Swapcache readahead pages are added to the LRU - and
1001 * possibly migrated - before they are charged.
1004 memcg = root_mem_cgroup;
1006 mz = mem_cgroup_page_zoneinfo(memcg, page);
1007 lruvec = &mz->lruvec;
1010 * Since a node can be onlined after the mem_cgroup was created,
1011 * we have to be prepared to initialize lruvec->zone here;
1012 * and if offlined then reonlined, we need to reinitialize it.
1014 if (unlikely(lruvec->zone != zone))
1015 lruvec->zone = zone;
1020 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1021 * @lruvec: mem_cgroup per zone lru vector
1022 * @lru: index of lru list the page is sitting on
1023 * @nr_pages: positive when adding or negative when removing
1025 * This function must be called when a page is added to or removed from an
1028 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1031 struct mem_cgroup_per_zone *mz;
1032 unsigned long *lru_size;
1034 if (mem_cgroup_disabled())
1037 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1038 lru_size = mz->lru_size + lru;
1039 *lru_size += nr_pages;
1040 VM_BUG_ON((long)(*lru_size) < 0);
1043 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1045 struct mem_cgroup *task_memcg;
1046 struct task_struct *p;
1049 p = find_lock_task_mm(task);
1051 task_memcg = get_mem_cgroup_from_mm(p->mm);
1055 * All threads may have already detached their mm's, but the oom
1056 * killer still needs to detect if they have already been oom
1057 * killed to prevent needlessly killing additional tasks.
1060 task_memcg = mem_cgroup_from_task(task);
1061 css_get(&task_memcg->css);
1064 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1065 css_put(&task_memcg->css);
1070 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1071 * @memcg: the memory cgroup
1073 * Returns the maximum amount of memory @mem can be charged with, in
1076 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1078 unsigned long margin = 0;
1079 unsigned long count;
1080 unsigned long limit;
1082 count = page_counter_read(&memcg->memory);
1083 limit = READ_ONCE(memcg->memory.limit);
1085 margin = limit - count;
1087 if (do_memsw_account()) {
1088 count = page_counter_read(&memcg->memsw);
1089 limit = READ_ONCE(memcg->memsw.limit);
1091 margin = min(margin, limit - count);
1098 * A routine for checking "mem" is under move_account() or not.
1100 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1101 * moving cgroups. This is for waiting at high-memory pressure
1104 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1106 struct mem_cgroup *from;
1107 struct mem_cgroup *to;
1110 * Unlike task_move routines, we access mc.to, mc.from not under
1111 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1113 spin_lock(&mc.lock);
1119 ret = mem_cgroup_is_descendant(from, memcg) ||
1120 mem_cgroup_is_descendant(to, memcg);
1122 spin_unlock(&mc.lock);
1126 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1128 if (mc.moving_task && current != mc.moving_task) {
1129 if (mem_cgroup_under_move(memcg)) {
1131 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1132 /* moving charge context might have finished. */
1135 finish_wait(&mc.waitq, &wait);
1142 #define K(x) ((x) << (PAGE_SHIFT-10))
1144 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1145 * @memcg: The memory cgroup that went over limit
1146 * @p: Task that is going to be killed
1148 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1151 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1153 /* oom_info_lock ensures that parallel ooms do not interleave */
1154 static DEFINE_MUTEX(oom_info_lock);
1155 struct mem_cgroup *iter;
1158 mutex_lock(&oom_info_lock);
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))));
1202 mutex_unlock(&oom_info_lock);
1206 * This function returns the number of memcg under hierarchy tree. Returns
1207 * 1(self count) if no children.
1209 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1212 struct mem_cgroup *iter;
1214 for_each_mem_cgroup_tree(iter, memcg)
1220 * Return the memory (and swap, if configured) limit for a memcg.
1222 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1224 unsigned long limit;
1226 limit = memcg->memory.limit;
1227 if (mem_cgroup_swappiness(memcg)) {
1228 unsigned long memsw_limit;
1229 unsigned long swap_limit;
1231 memsw_limit = memcg->memsw.limit;
1232 swap_limit = memcg->swap.limit;
1233 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1234 limit = min(limit + swap_limit, memsw_limit);
1239 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1242 struct oom_control oc = {
1245 .gfp_mask = gfp_mask,
1248 struct mem_cgroup *iter;
1249 unsigned long chosen_points = 0;
1250 unsigned long totalpages;
1251 unsigned int points = 0;
1252 struct task_struct *chosen = NULL;
1254 mutex_lock(&oom_lock);
1257 * If current has a pending SIGKILL or is exiting, then automatically
1258 * select it. The goal is to allow it to allocate so that it may
1259 * quickly exit and free its memory.
1261 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1262 mark_oom_victim(current);
1266 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1267 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1268 for_each_mem_cgroup_tree(iter, memcg) {
1269 struct css_task_iter it;
1270 struct task_struct *task;
1272 css_task_iter_start(&iter->css, &it);
1273 while ((task = css_task_iter_next(&it))) {
1274 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1275 case OOM_SCAN_SELECT:
1277 put_task_struct(chosen);
1279 chosen_points = ULONG_MAX;
1280 get_task_struct(chosen);
1282 case OOM_SCAN_CONTINUE:
1284 case OOM_SCAN_ABORT:
1285 css_task_iter_end(&it);
1286 mem_cgroup_iter_break(memcg, iter);
1288 put_task_struct(chosen);
1293 points = oom_badness(task, memcg, NULL, totalpages);
1294 if (!points || points < chosen_points)
1296 /* Prefer thread group leaders for display purposes */
1297 if (points == chosen_points &&
1298 thread_group_leader(chosen))
1302 put_task_struct(chosen);
1304 chosen_points = points;
1305 get_task_struct(chosen);
1307 css_task_iter_end(&it);
1311 points = chosen_points * 1000 / totalpages;
1312 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1313 "Memory cgroup out of memory");
1316 mutex_unlock(&oom_lock);
1320 #if MAX_NUMNODES > 1
1323 * test_mem_cgroup_node_reclaimable
1324 * @memcg: the target memcg
1325 * @nid: the node ID to be checked.
1326 * @noswap : specify true here if the user wants flle only information.
1328 * This function returns whether the specified memcg contains any
1329 * reclaimable pages on a node. Returns true if there are any reclaimable
1330 * pages in the node.
1332 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1333 int nid, bool noswap)
1335 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1337 if (noswap || !total_swap_pages)
1339 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1346 * Always updating the nodemask is not very good - even if we have an empty
1347 * list or the wrong list here, we can start from some node and traverse all
1348 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1351 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1355 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1356 * pagein/pageout changes since the last update.
1358 if (!atomic_read(&memcg->numainfo_events))
1360 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1363 /* make a nodemask where this memcg uses memory from */
1364 memcg->scan_nodes = node_states[N_MEMORY];
1366 for_each_node_mask(nid, node_states[N_MEMORY]) {
1368 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1369 node_clear(nid, memcg->scan_nodes);
1372 atomic_set(&memcg->numainfo_events, 0);
1373 atomic_set(&memcg->numainfo_updating, 0);
1377 * Selecting a node where we start reclaim from. Because what we need is just
1378 * reducing usage counter, start from anywhere is O,K. Considering
1379 * memory reclaim from current node, there are pros. and cons.
1381 * Freeing memory from current node means freeing memory from a node which
1382 * we'll use or we've used. So, it may make LRU bad. And if several threads
1383 * hit limits, it will see a contention on a node. But freeing from remote
1384 * node means more costs for memory reclaim because of memory latency.
1386 * Now, we use round-robin. Better algorithm is welcomed.
1388 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1392 mem_cgroup_may_update_nodemask(memcg);
1393 node = memcg->last_scanned_node;
1395 node = next_node(node, memcg->scan_nodes);
1396 if (node == MAX_NUMNODES)
1397 node = first_node(memcg->scan_nodes);
1399 * We call this when we hit limit, not when pages are added to LRU.
1400 * No LRU may hold pages because all pages are UNEVICTABLE or
1401 * memcg is too small and all pages are not on LRU. In that case,
1402 * we use curret node.
1404 if (unlikely(node == MAX_NUMNODES))
1405 node = numa_node_id();
1407 memcg->last_scanned_node = node;
1411 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1417 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1420 unsigned long *total_scanned)
1422 struct mem_cgroup *victim = NULL;
1425 unsigned long excess;
1426 unsigned long nr_scanned;
1427 struct mem_cgroup_reclaim_cookie reclaim = {
1432 excess = soft_limit_excess(root_memcg);
1435 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1440 * If we have not been able to reclaim
1441 * anything, it might because there are
1442 * no reclaimable pages under this hierarchy
1447 * We want to do more targeted reclaim.
1448 * excess >> 2 is not to excessive so as to
1449 * reclaim too much, nor too less that we keep
1450 * coming back to reclaim from this cgroup
1452 if (total >= (excess >> 2) ||
1453 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1458 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1460 *total_scanned += nr_scanned;
1461 if (!soft_limit_excess(root_memcg))
1464 mem_cgroup_iter_break(root_memcg, victim);
1468 #ifdef CONFIG_LOCKDEP
1469 static struct lockdep_map memcg_oom_lock_dep_map = {
1470 .name = "memcg_oom_lock",
1474 static DEFINE_SPINLOCK(memcg_oom_lock);
1477 * Check OOM-Killer is already running under our hierarchy.
1478 * If someone is running, return false.
1480 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1482 struct mem_cgroup *iter, *failed = NULL;
1484 spin_lock(&memcg_oom_lock);
1486 for_each_mem_cgroup_tree(iter, memcg) {
1487 if (iter->oom_lock) {
1489 * this subtree of our hierarchy is already locked
1490 * so we cannot give a lock.
1493 mem_cgroup_iter_break(memcg, iter);
1496 iter->oom_lock = true;
1501 * OK, we failed to lock the whole subtree so we have
1502 * to clean up what we set up to the failing subtree
1504 for_each_mem_cgroup_tree(iter, memcg) {
1505 if (iter == failed) {
1506 mem_cgroup_iter_break(memcg, iter);
1509 iter->oom_lock = false;
1512 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1514 spin_unlock(&memcg_oom_lock);
1519 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1521 struct mem_cgroup *iter;
1523 spin_lock(&memcg_oom_lock);
1524 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1525 for_each_mem_cgroup_tree(iter, memcg)
1526 iter->oom_lock = false;
1527 spin_unlock(&memcg_oom_lock);
1530 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1532 struct mem_cgroup *iter;
1534 spin_lock(&memcg_oom_lock);
1535 for_each_mem_cgroup_tree(iter, memcg)
1537 spin_unlock(&memcg_oom_lock);
1540 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1542 struct mem_cgroup *iter;
1545 * When a new child is created while the hierarchy is under oom,
1546 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1548 spin_lock(&memcg_oom_lock);
1549 for_each_mem_cgroup_tree(iter, memcg)
1550 if (iter->under_oom > 0)
1552 spin_unlock(&memcg_oom_lock);
1555 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1557 struct oom_wait_info {
1558 struct mem_cgroup *memcg;
1562 static int memcg_oom_wake_function(wait_queue_t *wait,
1563 unsigned mode, int sync, void *arg)
1565 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1566 struct mem_cgroup *oom_wait_memcg;
1567 struct oom_wait_info *oom_wait_info;
1569 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1570 oom_wait_memcg = oom_wait_info->memcg;
1572 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1573 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1575 return autoremove_wake_function(wait, mode, sync, arg);
1578 static void memcg_oom_recover(struct mem_cgroup *memcg)
1581 * For the following lockless ->under_oom test, the only required
1582 * guarantee is that it must see the state asserted by an OOM when
1583 * this function is called as a result of userland actions
1584 * triggered by the notification of the OOM. This is trivially
1585 * achieved by invoking mem_cgroup_mark_under_oom() before
1586 * triggering notification.
1588 if (memcg && memcg->under_oom)
1589 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1592 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1594 if (!current->memcg_may_oom)
1597 * We are in the middle of the charge context here, so we
1598 * don't want to block when potentially sitting on a callstack
1599 * that holds all kinds of filesystem and mm locks.
1601 * Also, the caller may handle a failed allocation gracefully
1602 * (like optional page cache readahead) and so an OOM killer
1603 * invocation might not even be necessary.
1605 * That's why we don't do anything here except remember the
1606 * OOM context and then deal with it at the end of the page
1607 * fault when the stack is unwound, the locks are released,
1608 * and when we know whether the fault was overall successful.
1610 css_get(&memcg->css);
1611 current->memcg_in_oom = memcg;
1612 current->memcg_oom_gfp_mask = mask;
1613 current->memcg_oom_order = order;
1617 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1618 * @handle: actually kill/wait or just clean up the OOM state
1620 * This has to be called at the end of a page fault if the memcg OOM
1621 * handler was enabled.
1623 * Memcg supports userspace OOM handling where failed allocations must
1624 * sleep on a waitqueue until the userspace task resolves the
1625 * situation. Sleeping directly in the charge context with all kinds
1626 * of locks held is not a good idea, instead we remember an OOM state
1627 * in the task and mem_cgroup_oom_synchronize() has to be called at
1628 * the end of the page fault to complete the OOM handling.
1630 * Returns %true if an ongoing memcg OOM situation was detected and
1631 * completed, %false otherwise.
1633 bool mem_cgroup_oom_synchronize(bool handle)
1635 struct mem_cgroup *memcg = current->memcg_in_oom;
1636 struct oom_wait_info owait;
1639 /* OOM is global, do not handle */
1643 if (!handle || oom_killer_disabled)
1646 owait.memcg = memcg;
1647 owait.wait.flags = 0;
1648 owait.wait.func = memcg_oom_wake_function;
1649 owait.wait.private = current;
1650 INIT_LIST_HEAD(&owait.wait.task_list);
1652 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1653 mem_cgroup_mark_under_oom(memcg);
1655 locked = mem_cgroup_oom_trylock(memcg);
1658 mem_cgroup_oom_notify(memcg);
1660 if (locked && !memcg->oom_kill_disable) {
1661 mem_cgroup_unmark_under_oom(memcg);
1662 finish_wait(&memcg_oom_waitq, &owait.wait);
1663 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1664 current->memcg_oom_order);
1667 mem_cgroup_unmark_under_oom(memcg);
1668 finish_wait(&memcg_oom_waitq, &owait.wait);
1672 mem_cgroup_oom_unlock(memcg);
1674 * There is no guarantee that an OOM-lock contender
1675 * sees the wakeups triggered by the OOM kill
1676 * uncharges. Wake any sleepers explicitely.
1678 memcg_oom_recover(memcg);
1681 current->memcg_in_oom = NULL;
1682 css_put(&memcg->css);
1687 * lock_page_memcg - lock a page->mem_cgroup binding
1690 * This function protects unlocked LRU pages from being moved to
1691 * another cgroup and stabilizes their page->mem_cgroup binding.
1693 void lock_page_memcg(struct page *page)
1695 struct mem_cgroup *memcg;
1696 unsigned long flags;
1699 * The RCU lock is held throughout the transaction. The fast
1700 * path can get away without acquiring the memcg->move_lock
1701 * because page moving starts with an RCU grace period.
1705 if (mem_cgroup_disabled())
1708 memcg = page->mem_cgroup;
1709 if (unlikely(!memcg))
1712 if (atomic_read(&memcg->moving_account) <= 0)
1715 spin_lock_irqsave(&memcg->move_lock, flags);
1716 if (memcg != page->mem_cgroup) {
1717 spin_unlock_irqrestore(&memcg->move_lock, flags);
1722 * When charge migration first begins, we can have locked and
1723 * unlocked page stat updates happening concurrently. Track
1724 * the task who has the lock for unlock_page_memcg().
1726 memcg->move_lock_task = current;
1727 memcg->move_lock_flags = flags;
1731 EXPORT_SYMBOL(lock_page_memcg);
1734 * unlock_page_memcg - unlock a page->mem_cgroup binding
1737 void unlock_page_memcg(struct page *page)
1739 struct mem_cgroup *memcg = page->mem_cgroup;
1741 if (memcg && memcg->move_lock_task == current) {
1742 unsigned long flags = memcg->move_lock_flags;
1744 memcg->move_lock_task = NULL;
1745 memcg->move_lock_flags = 0;
1747 spin_unlock_irqrestore(&memcg->move_lock, flags);
1752 EXPORT_SYMBOL(unlock_page_memcg);
1755 * size of first charge trial. "32" comes from vmscan.c's magic value.
1756 * TODO: maybe necessary to use big numbers in big irons.
1758 #define CHARGE_BATCH 32U
1759 struct memcg_stock_pcp {
1760 struct mem_cgroup *cached; /* this never be root cgroup */
1761 unsigned int nr_pages;
1762 struct work_struct work;
1763 unsigned long flags;
1764 #define FLUSHING_CACHED_CHARGE 0
1766 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1767 static DEFINE_MUTEX(percpu_charge_mutex);
1770 * consume_stock: Try to consume stocked charge on this cpu.
1771 * @memcg: memcg to consume from.
1772 * @nr_pages: how many pages to charge.
1774 * The charges will only happen if @memcg matches the current cpu's memcg
1775 * stock, and at least @nr_pages are available in that stock. Failure to
1776 * service an allocation will refill the stock.
1778 * returns true if successful, false otherwise.
1780 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1782 struct memcg_stock_pcp *stock;
1785 if (nr_pages > CHARGE_BATCH)
1788 stock = &get_cpu_var(memcg_stock);
1789 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1790 stock->nr_pages -= nr_pages;
1793 put_cpu_var(memcg_stock);
1798 * Returns stocks cached in percpu and reset cached information.
1800 static void drain_stock(struct memcg_stock_pcp *stock)
1802 struct mem_cgroup *old = stock->cached;
1804 if (stock->nr_pages) {
1805 page_counter_uncharge(&old->memory, stock->nr_pages);
1806 if (do_memsw_account())
1807 page_counter_uncharge(&old->memsw, stock->nr_pages);
1808 css_put_many(&old->css, stock->nr_pages);
1809 stock->nr_pages = 0;
1811 stock->cached = NULL;
1815 * This must be called under preempt disabled or must be called by
1816 * a thread which is pinned to local cpu.
1818 static void drain_local_stock(struct work_struct *dummy)
1820 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1822 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1826 * Cache charges(val) to local per_cpu area.
1827 * This will be consumed by consume_stock() function, later.
1829 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1831 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1833 if (stock->cached != memcg) { /* reset if necessary */
1835 stock->cached = memcg;
1837 stock->nr_pages += nr_pages;
1838 put_cpu_var(memcg_stock);
1842 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1843 * of the hierarchy under it.
1845 static void drain_all_stock(struct mem_cgroup *root_memcg)
1849 /* If someone's already draining, avoid adding running more workers. */
1850 if (!mutex_trylock(&percpu_charge_mutex))
1852 /* Notify other cpus that system-wide "drain" is running */
1855 for_each_online_cpu(cpu) {
1856 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1857 struct mem_cgroup *memcg;
1859 memcg = stock->cached;
1860 if (!memcg || !stock->nr_pages)
1862 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1864 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1866 drain_local_stock(&stock->work);
1868 schedule_work_on(cpu, &stock->work);
1873 mutex_unlock(&percpu_charge_mutex);
1876 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1877 unsigned long action,
1880 int cpu = (unsigned long)hcpu;
1881 struct memcg_stock_pcp *stock;
1883 if (action == CPU_ONLINE)
1886 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1889 stock = &per_cpu(memcg_stock, cpu);
1894 static void reclaim_high(struct mem_cgroup *memcg,
1895 unsigned int nr_pages,
1899 if (page_counter_read(&memcg->memory) <= memcg->high)
1901 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1902 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1903 } while ((memcg = parent_mem_cgroup(memcg)));
1906 static void high_work_func(struct work_struct *work)
1908 struct mem_cgroup *memcg;
1910 memcg = container_of(work, struct mem_cgroup, high_work);
1911 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1915 * Scheduled by try_charge() to be executed from the userland return path
1916 * and reclaims memory over the high limit.
1918 void mem_cgroup_handle_over_high(void)
1920 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1921 struct mem_cgroup *memcg;
1923 if (likely(!nr_pages))
1926 memcg = get_mem_cgroup_from_mm(current->mm);
1927 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1928 css_put(&memcg->css);
1929 current->memcg_nr_pages_over_high = 0;
1932 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1933 unsigned int nr_pages)
1935 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1936 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1937 struct mem_cgroup *mem_over_limit;
1938 struct page_counter *counter;
1939 unsigned long nr_reclaimed;
1940 bool may_swap = true;
1941 bool drained = false;
1943 if (mem_cgroup_is_root(memcg))
1946 if (consume_stock(memcg, nr_pages))
1949 if (!do_memsw_account() ||
1950 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1951 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1953 if (do_memsw_account())
1954 page_counter_uncharge(&memcg->memsw, batch);
1955 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1957 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1961 if (batch > nr_pages) {
1967 * Unlike in global OOM situations, memcg is not in a physical
1968 * memory shortage. Allow dying and OOM-killed tasks to
1969 * bypass the last charges so that they can exit quickly and
1970 * free their memory.
1972 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1973 fatal_signal_pending(current) ||
1974 current->flags & PF_EXITING))
1977 if (unlikely(task_in_memcg_oom(current)))
1980 if (!gfpflags_allow_blocking(gfp_mask))
1983 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1985 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1986 gfp_mask, may_swap);
1988 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1992 drain_all_stock(mem_over_limit);
1997 if (gfp_mask & __GFP_NORETRY)
2000 * Even though the limit is exceeded at this point, reclaim
2001 * may have been able to free some pages. Retry the charge
2002 * before killing the task.
2004 * Only for regular pages, though: huge pages are rather
2005 * unlikely to succeed so close to the limit, and we fall back
2006 * to regular pages anyway in case of failure.
2008 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2011 * At task move, charge accounts can be doubly counted. So, it's
2012 * better to wait until the end of task_move if something is going on.
2014 if (mem_cgroup_wait_acct_move(mem_over_limit))
2020 if (gfp_mask & __GFP_NOFAIL)
2023 if (fatal_signal_pending(current))
2026 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2028 mem_cgroup_oom(mem_over_limit, gfp_mask,
2029 get_order(nr_pages * PAGE_SIZE));
2031 if (!(gfp_mask & __GFP_NOFAIL))
2035 * The allocation either can't fail or will lead to more memory
2036 * being freed very soon. Allow memory usage go over the limit
2037 * temporarily by force charging it.
2039 page_counter_charge(&memcg->memory, nr_pages);
2040 if (do_memsw_account())
2041 page_counter_charge(&memcg->memsw, nr_pages);
2042 css_get_many(&memcg->css, nr_pages);
2047 css_get_many(&memcg->css, batch);
2048 if (batch > nr_pages)
2049 refill_stock(memcg, batch - nr_pages);
2052 * If the hierarchy is above the normal consumption range, schedule
2053 * reclaim on returning to userland. We can perform reclaim here
2054 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2055 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2056 * not recorded as it most likely matches current's and won't
2057 * change in the meantime. As high limit is checked again before
2058 * reclaim, the cost of mismatch is negligible.
2061 if (page_counter_read(&memcg->memory) > memcg->high) {
2062 /* Don't bother a random interrupted task */
2063 if (in_interrupt()) {
2064 schedule_work(&memcg->high_work);
2067 current->memcg_nr_pages_over_high += batch;
2068 set_notify_resume(current);
2071 } while ((memcg = parent_mem_cgroup(memcg)));
2076 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2078 if (mem_cgroup_is_root(memcg))
2081 page_counter_uncharge(&memcg->memory, nr_pages);
2082 if (do_memsw_account())
2083 page_counter_uncharge(&memcg->memsw, nr_pages);
2085 css_put_many(&memcg->css, nr_pages);
2088 static void lock_page_lru(struct page *page, int *isolated)
2090 struct zone *zone = page_zone(page);
2092 spin_lock_irq(&zone->lru_lock);
2093 if (PageLRU(page)) {
2094 struct lruvec *lruvec;
2096 lruvec = mem_cgroup_page_lruvec(page, zone);
2098 del_page_from_lru_list(page, lruvec, page_lru(page));
2104 static void unlock_page_lru(struct page *page, int isolated)
2106 struct zone *zone = page_zone(page);
2109 struct lruvec *lruvec;
2111 lruvec = mem_cgroup_page_lruvec(page, zone);
2112 VM_BUG_ON_PAGE(PageLRU(page), page);
2114 add_page_to_lru_list(page, lruvec, page_lru(page));
2116 spin_unlock_irq(&zone->lru_lock);
2119 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2124 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2127 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2128 * may already be on some other mem_cgroup's LRU. Take care of it.
2131 lock_page_lru(page, &isolated);
2134 * Nobody should be changing or seriously looking at
2135 * page->mem_cgroup at this point:
2137 * - the page is uncharged
2139 * - the page is off-LRU
2141 * - an anonymous fault has exclusive page access, except for
2142 * a locked page table
2144 * - a page cache insertion, a swapin fault, or a migration
2145 * have the page locked
2147 page->mem_cgroup = memcg;
2150 unlock_page_lru(page, isolated);
2154 static int memcg_alloc_cache_id(void)
2159 id = ida_simple_get(&memcg_cache_ida,
2160 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2164 if (id < memcg_nr_cache_ids)
2168 * There's no space for the new id in memcg_caches arrays,
2169 * so we have to grow them.
2171 down_write(&memcg_cache_ids_sem);
2173 size = 2 * (id + 1);
2174 if (size < MEMCG_CACHES_MIN_SIZE)
2175 size = MEMCG_CACHES_MIN_SIZE;
2176 else if (size > MEMCG_CACHES_MAX_SIZE)
2177 size = MEMCG_CACHES_MAX_SIZE;
2179 err = memcg_update_all_caches(size);
2181 err = memcg_update_all_list_lrus(size);
2183 memcg_nr_cache_ids = size;
2185 up_write(&memcg_cache_ids_sem);
2188 ida_simple_remove(&memcg_cache_ida, id);
2194 static void memcg_free_cache_id(int id)
2196 ida_simple_remove(&memcg_cache_ida, id);
2199 struct memcg_kmem_cache_create_work {
2200 struct mem_cgroup *memcg;
2201 struct kmem_cache *cachep;
2202 struct work_struct work;
2205 static void memcg_kmem_cache_create_func(struct work_struct *w)
2207 struct memcg_kmem_cache_create_work *cw =
2208 container_of(w, struct memcg_kmem_cache_create_work, work);
2209 struct mem_cgroup *memcg = cw->memcg;
2210 struct kmem_cache *cachep = cw->cachep;
2212 memcg_create_kmem_cache(memcg, cachep);
2214 css_put(&memcg->css);
2219 * Enqueue the creation of a per-memcg kmem_cache.
2221 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2222 struct kmem_cache *cachep)
2224 struct memcg_kmem_cache_create_work *cw;
2226 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2230 css_get(&memcg->css);
2233 cw->cachep = cachep;
2234 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2236 schedule_work(&cw->work);
2239 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2240 struct kmem_cache *cachep)
2243 * We need to stop accounting when we kmalloc, because if the
2244 * corresponding kmalloc cache is not yet created, the first allocation
2245 * in __memcg_schedule_kmem_cache_create will recurse.
2247 * However, it is better to enclose the whole function. Depending on
2248 * the debugging options enabled, INIT_WORK(), for instance, can
2249 * trigger an allocation. This too, will make us recurse. Because at
2250 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2251 * the safest choice is to do it like this, wrapping the whole function.
2253 current->memcg_kmem_skip_account = 1;
2254 __memcg_schedule_kmem_cache_create(memcg, cachep);
2255 current->memcg_kmem_skip_account = 0;
2259 * Return the kmem_cache we're supposed to use for a slab allocation.
2260 * We try to use the current memcg's version of the cache.
2262 * If the cache does not exist yet, if we are the first user of it,
2263 * we either create it immediately, if possible, or create it asynchronously
2265 * In the latter case, we will let the current allocation go through with
2266 * the original cache.
2268 * Can't be called in interrupt context or from kernel threads.
2269 * This function needs to be called with rcu_read_lock() held.
2271 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2273 struct mem_cgroup *memcg;
2274 struct kmem_cache *memcg_cachep;
2277 VM_BUG_ON(!is_root_cache(cachep));
2279 if (cachep->flags & SLAB_ACCOUNT)
2280 gfp |= __GFP_ACCOUNT;
2282 if (!(gfp & __GFP_ACCOUNT))
2285 if (current->memcg_kmem_skip_account)
2288 memcg = get_mem_cgroup_from_mm(current->mm);
2289 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2293 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2294 if (likely(memcg_cachep))
2295 return memcg_cachep;
2298 * If we are in a safe context (can wait, and not in interrupt
2299 * context), we could be be predictable and return right away.
2300 * This would guarantee that the allocation being performed
2301 * already belongs in the new cache.
2303 * However, there are some clashes that can arrive from locking.
2304 * For instance, because we acquire the slab_mutex while doing
2305 * memcg_create_kmem_cache, this means no further allocation
2306 * could happen with the slab_mutex held. So it's better to
2309 memcg_schedule_kmem_cache_create(memcg, cachep);
2311 css_put(&memcg->css);
2315 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2317 if (!is_root_cache(cachep))
2318 css_put(&cachep->memcg_params.memcg->css);
2321 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2322 struct mem_cgroup *memcg)
2324 unsigned int nr_pages = 1 << order;
2325 struct page_counter *counter;
2328 ret = try_charge(memcg, gfp, nr_pages);
2332 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2333 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2334 cancel_charge(memcg, nr_pages);
2338 page->mem_cgroup = memcg;
2343 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2345 struct mem_cgroup *memcg;
2348 memcg = get_mem_cgroup_from_mm(current->mm);
2349 if (!mem_cgroup_is_root(memcg))
2350 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2351 css_put(&memcg->css);
2355 void __memcg_kmem_uncharge(struct page *page, int order)
2357 struct mem_cgroup *memcg = page->mem_cgroup;
2358 unsigned int nr_pages = 1 << order;
2363 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2365 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2366 page_counter_uncharge(&memcg->kmem, nr_pages);
2368 page_counter_uncharge(&memcg->memory, nr_pages);
2369 if (do_memsw_account())
2370 page_counter_uncharge(&memcg->memsw, nr_pages);
2372 page->mem_cgroup = NULL;
2373 css_put_many(&memcg->css, nr_pages);
2375 #endif /* !CONFIG_SLOB */
2377 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2380 * Because tail pages are not marked as "used", set it. We're under
2381 * zone->lru_lock and migration entries setup in all page mappings.
2383 void mem_cgroup_split_huge_fixup(struct page *head)
2387 if (mem_cgroup_disabled())
2390 for (i = 1; i < HPAGE_PMD_NR; i++)
2391 head[i].mem_cgroup = head->mem_cgroup;
2393 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2396 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2398 #ifdef CONFIG_MEMCG_SWAP
2399 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2402 int val = (charge) ? 1 : -1;
2403 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2407 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2408 * @entry: swap entry to be moved
2409 * @from: mem_cgroup which the entry is moved from
2410 * @to: mem_cgroup which the entry is moved to
2412 * It succeeds only when the swap_cgroup's record for this entry is the same
2413 * as the mem_cgroup's id of @from.
2415 * Returns 0 on success, -EINVAL on failure.
2417 * The caller must have charged to @to, IOW, called page_counter_charge() about
2418 * both res and memsw, and called css_get().
2420 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2421 struct mem_cgroup *from, struct mem_cgroup *to)
2423 unsigned short old_id, new_id;
2425 old_id = mem_cgroup_id(from);
2426 new_id = mem_cgroup_id(to);
2428 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2429 mem_cgroup_swap_statistics(from, false);
2430 mem_cgroup_swap_statistics(to, true);
2436 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2437 struct mem_cgroup *from, struct mem_cgroup *to)
2443 static DEFINE_MUTEX(memcg_limit_mutex);
2445 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2446 unsigned long limit)
2448 unsigned long curusage;
2449 unsigned long oldusage;
2450 bool enlarge = false;
2455 * For keeping hierarchical_reclaim simple, how long we should retry
2456 * is depends on callers. We set our retry-count to be function
2457 * of # of children which we should visit in this loop.
2459 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2460 mem_cgroup_count_children(memcg);
2462 oldusage = page_counter_read(&memcg->memory);
2465 if (signal_pending(current)) {
2470 mutex_lock(&memcg_limit_mutex);
2471 if (limit > memcg->memsw.limit) {
2472 mutex_unlock(&memcg_limit_mutex);
2476 if (limit > memcg->memory.limit)
2478 ret = page_counter_limit(&memcg->memory, limit);
2479 mutex_unlock(&memcg_limit_mutex);
2484 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2486 curusage = page_counter_read(&memcg->memory);
2487 /* Usage is reduced ? */
2488 if (curusage >= oldusage)
2491 oldusage = curusage;
2492 } while (retry_count);
2494 if (!ret && enlarge)
2495 memcg_oom_recover(memcg);
2500 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2501 unsigned long limit)
2503 unsigned long curusage;
2504 unsigned long oldusage;
2505 bool enlarge = false;
2509 /* see mem_cgroup_resize_res_limit */
2510 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2511 mem_cgroup_count_children(memcg);
2513 oldusage = page_counter_read(&memcg->memsw);
2516 if (signal_pending(current)) {
2521 mutex_lock(&memcg_limit_mutex);
2522 if (limit < memcg->memory.limit) {
2523 mutex_unlock(&memcg_limit_mutex);
2527 if (limit > memcg->memsw.limit)
2529 ret = page_counter_limit(&memcg->memsw, limit);
2530 mutex_unlock(&memcg_limit_mutex);
2535 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2537 curusage = page_counter_read(&memcg->memsw);
2538 /* Usage is reduced ? */
2539 if (curusage >= oldusage)
2542 oldusage = curusage;
2543 } while (retry_count);
2545 if (!ret && enlarge)
2546 memcg_oom_recover(memcg);
2551 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2553 unsigned long *total_scanned)
2555 unsigned long nr_reclaimed = 0;
2556 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2557 unsigned long reclaimed;
2559 struct mem_cgroup_tree_per_zone *mctz;
2560 unsigned long excess;
2561 unsigned long nr_scanned;
2566 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2568 * This loop can run a while, specially if mem_cgroup's continuously
2569 * keep exceeding their soft limit and putting the system under
2576 mz = mem_cgroup_largest_soft_limit_node(mctz);
2581 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2582 gfp_mask, &nr_scanned);
2583 nr_reclaimed += reclaimed;
2584 *total_scanned += nr_scanned;
2585 spin_lock_irq(&mctz->lock);
2586 __mem_cgroup_remove_exceeded(mz, mctz);
2589 * If we failed to reclaim anything from this memory cgroup
2590 * it is time to move on to the next cgroup
2594 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2596 excess = soft_limit_excess(mz->memcg);
2598 * One school of thought says that we should not add
2599 * back the node to the tree if reclaim returns 0.
2600 * But our reclaim could return 0, simply because due
2601 * to priority we are exposing a smaller subset of
2602 * memory to reclaim from. Consider this as a longer
2605 /* If excess == 0, no tree ops */
2606 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2607 spin_unlock_irq(&mctz->lock);
2608 css_put(&mz->memcg->css);
2611 * Could not reclaim anything and there are no more
2612 * mem cgroups to try or we seem to be looping without
2613 * reclaiming anything.
2615 if (!nr_reclaimed &&
2617 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2619 } while (!nr_reclaimed);
2621 css_put(&next_mz->memcg->css);
2622 return nr_reclaimed;
2626 * Test whether @memcg has children, dead or alive. Note that this
2627 * function doesn't care whether @memcg has use_hierarchy enabled and
2628 * returns %true if there are child csses according to the cgroup
2629 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2631 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2636 ret = css_next_child(NULL, &memcg->css);
2642 * Reclaims as many pages from the given memcg as possible and moves
2643 * the rest to the parent.
2645 * Caller is responsible for holding css reference for memcg.
2647 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2649 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2651 /* we call try-to-free pages for make this cgroup empty */
2652 lru_add_drain_all();
2653 /* try to free all pages in this cgroup */
2654 while (nr_retries && page_counter_read(&memcg->memory)) {
2657 if (signal_pending(current))
2660 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2664 /* maybe some writeback is necessary */
2665 congestion_wait(BLK_RW_ASYNC, HZ/10);
2673 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2674 char *buf, size_t nbytes,
2677 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2679 if (mem_cgroup_is_root(memcg))
2681 return mem_cgroup_force_empty(memcg) ?: nbytes;
2684 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2687 return mem_cgroup_from_css(css)->use_hierarchy;
2690 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2691 struct cftype *cft, u64 val)
2694 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2695 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2697 if (memcg->use_hierarchy == val)
2701 * If parent's use_hierarchy is set, we can't make any modifications
2702 * in the child subtrees. If it is unset, then the change can
2703 * occur, provided the current cgroup has no children.
2705 * For the root cgroup, parent_mem is NULL, we allow value to be
2706 * set if there are no children.
2708 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2709 (val == 1 || val == 0)) {
2710 if (!memcg_has_children(memcg))
2711 memcg->use_hierarchy = val;
2720 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2722 struct mem_cgroup *iter;
2725 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2727 for_each_mem_cgroup_tree(iter, memcg) {
2728 for (i = 0; i < MEMCG_NR_STAT; i++)
2729 stat[i] += mem_cgroup_read_stat(iter, i);
2733 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2735 struct mem_cgroup *iter;
2738 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2740 for_each_mem_cgroup_tree(iter, memcg) {
2741 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2742 events[i] += mem_cgroup_read_events(iter, i);
2746 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2748 unsigned long val = 0;
2750 if (mem_cgroup_is_root(memcg)) {
2751 struct mem_cgroup *iter;
2753 for_each_mem_cgroup_tree(iter, memcg) {
2754 val += mem_cgroup_read_stat(iter,
2755 MEM_CGROUP_STAT_CACHE);
2756 val += mem_cgroup_read_stat(iter,
2757 MEM_CGROUP_STAT_RSS);
2759 val += mem_cgroup_read_stat(iter,
2760 MEM_CGROUP_STAT_SWAP);
2764 val = page_counter_read(&memcg->memory);
2766 val = page_counter_read(&memcg->memsw);
2779 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2782 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2783 struct page_counter *counter;
2785 switch (MEMFILE_TYPE(cft->private)) {
2787 counter = &memcg->memory;
2790 counter = &memcg->memsw;
2793 counter = &memcg->kmem;
2796 counter = &memcg->tcpmem;
2802 switch (MEMFILE_ATTR(cft->private)) {
2804 if (counter == &memcg->memory)
2805 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2806 if (counter == &memcg->memsw)
2807 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2808 return (u64)page_counter_read(counter) * PAGE_SIZE;
2810 return (u64)counter->limit * PAGE_SIZE;
2812 return (u64)counter->watermark * PAGE_SIZE;
2814 return counter->failcnt;
2815 case RES_SOFT_LIMIT:
2816 return (u64)memcg->soft_limit * PAGE_SIZE;
2823 static int memcg_online_kmem(struct mem_cgroup *memcg)
2827 if (cgroup_memory_nokmem)
2830 BUG_ON(memcg->kmemcg_id >= 0);
2831 BUG_ON(memcg->kmem_state);
2833 memcg_id = memcg_alloc_cache_id();
2837 static_branch_inc(&memcg_kmem_enabled_key);
2839 * A memory cgroup is considered kmem-online as soon as it gets
2840 * kmemcg_id. Setting the id after enabling static branching will
2841 * guarantee no one starts accounting before all call sites are
2844 memcg->kmemcg_id = memcg_id;
2845 memcg->kmem_state = KMEM_ONLINE;
2850 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2852 struct cgroup_subsys_state *css;
2853 struct mem_cgroup *parent, *child;
2856 if (memcg->kmem_state != KMEM_ONLINE)
2859 * Clear the online state before clearing memcg_caches array
2860 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2861 * guarantees that no cache will be created for this cgroup
2862 * after we are done (see memcg_create_kmem_cache()).
2864 memcg->kmem_state = KMEM_ALLOCATED;
2866 memcg_deactivate_kmem_caches(memcg);
2868 kmemcg_id = memcg->kmemcg_id;
2869 BUG_ON(kmemcg_id < 0);
2871 parent = parent_mem_cgroup(memcg);
2873 parent = root_mem_cgroup;
2876 * Change kmemcg_id of this cgroup and all its descendants to the
2877 * parent's id, and then move all entries from this cgroup's list_lrus
2878 * to ones of the parent. After we have finished, all list_lrus
2879 * corresponding to this cgroup are guaranteed to remain empty. The
2880 * ordering is imposed by list_lru_node->lock taken by
2881 * memcg_drain_all_list_lrus().
2883 css_for_each_descendant_pre(css, &memcg->css) {
2884 child = mem_cgroup_from_css(css);
2885 BUG_ON(child->kmemcg_id != kmemcg_id);
2886 child->kmemcg_id = parent->kmemcg_id;
2887 if (!memcg->use_hierarchy)
2890 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2892 memcg_free_cache_id(kmemcg_id);
2895 static void memcg_free_kmem(struct mem_cgroup *memcg)
2897 /* css_alloc() failed, offlining didn't happen */
2898 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2899 memcg_offline_kmem(memcg);
2901 if (memcg->kmem_state == KMEM_ALLOCATED) {
2902 memcg_destroy_kmem_caches(memcg);
2903 static_branch_dec(&memcg_kmem_enabled_key);
2904 WARN_ON(page_counter_read(&memcg->kmem));
2908 static int memcg_online_kmem(struct mem_cgroup *memcg)
2912 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2915 static void memcg_free_kmem(struct mem_cgroup *memcg)
2918 #endif /* !CONFIG_SLOB */
2920 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2921 unsigned long limit)
2925 mutex_lock(&memcg_limit_mutex);
2926 ret = page_counter_limit(&memcg->kmem, limit);
2927 mutex_unlock(&memcg_limit_mutex);
2931 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2935 mutex_lock(&memcg_limit_mutex);
2937 ret = page_counter_limit(&memcg->tcpmem, limit);
2941 if (!memcg->tcpmem_active) {
2943 * The active flag needs to be written after the static_key
2944 * update. This is what guarantees that the socket activation
2945 * function is the last one to run. See sock_update_memcg() for
2946 * details, and note that we don't mark any socket as belonging
2947 * to this memcg until that flag is up.
2949 * We need to do this, because static_keys will span multiple
2950 * sites, but we can't control their order. If we mark a socket
2951 * as accounted, but the accounting functions are not patched in
2952 * yet, we'll lose accounting.
2954 * We never race with the readers in sock_update_memcg(),
2955 * because when this value change, the code to process it is not
2958 static_branch_inc(&memcg_sockets_enabled_key);
2959 memcg->tcpmem_active = true;
2962 mutex_unlock(&memcg_limit_mutex);
2967 * The user of this function is...
2970 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2971 char *buf, size_t nbytes, loff_t off)
2973 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2974 unsigned long nr_pages;
2977 buf = strstrip(buf);
2978 ret = page_counter_memparse(buf, "-1", &nr_pages);
2982 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2984 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2988 switch (MEMFILE_TYPE(of_cft(of)->private)) {
2990 ret = mem_cgroup_resize_limit(memcg, nr_pages);
2993 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
2996 ret = memcg_update_kmem_limit(memcg, nr_pages);
2999 ret = memcg_update_tcp_limit(memcg, nr_pages);
3003 case RES_SOFT_LIMIT:
3004 memcg->soft_limit = nr_pages;
3008 return ret ?: nbytes;
3011 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3012 size_t nbytes, loff_t off)
3014 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3015 struct page_counter *counter;
3017 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3019 counter = &memcg->memory;
3022 counter = &memcg->memsw;
3025 counter = &memcg->kmem;
3028 counter = &memcg->tcpmem;
3034 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3036 page_counter_reset_watermark(counter);
3039 counter->failcnt = 0;
3048 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3051 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3055 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3056 struct cftype *cft, u64 val)
3058 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3060 if (val & ~MOVE_MASK)
3064 * No kind of locking is needed in here, because ->can_attach() will
3065 * check this value once in the beginning of the process, and then carry
3066 * on with stale data. This means that changes to this value will only
3067 * affect task migrations starting after the change.
3069 memcg->move_charge_at_immigrate = val;
3073 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3074 struct cftype *cft, u64 val)
3081 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3085 unsigned int lru_mask;
3088 static const struct numa_stat stats[] = {
3089 { "total", LRU_ALL },
3090 { "file", LRU_ALL_FILE },
3091 { "anon", LRU_ALL_ANON },
3092 { "unevictable", BIT(LRU_UNEVICTABLE) },
3094 const struct numa_stat *stat;
3097 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3099 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3100 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3101 seq_printf(m, "%s=%lu", stat->name, nr);
3102 for_each_node_state(nid, N_MEMORY) {
3103 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3105 seq_printf(m, " N%d=%lu", nid, nr);
3110 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3111 struct mem_cgroup *iter;
3114 for_each_mem_cgroup_tree(iter, memcg)
3115 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3116 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3117 for_each_node_state(nid, N_MEMORY) {
3119 for_each_mem_cgroup_tree(iter, memcg)
3120 nr += mem_cgroup_node_nr_lru_pages(
3121 iter, nid, stat->lru_mask);
3122 seq_printf(m, " N%d=%lu", nid, nr);
3129 #endif /* CONFIG_NUMA */
3131 static int memcg_stat_show(struct seq_file *m, void *v)
3133 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3134 unsigned long memory, memsw;
3135 struct mem_cgroup *mi;
3138 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3139 MEM_CGROUP_STAT_NSTATS);
3140 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3141 MEM_CGROUP_EVENTS_NSTATS);
3142 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3144 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3145 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3147 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3148 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3151 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3152 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3153 mem_cgroup_read_events(memcg, i));
3155 for (i = 0; i < NR_LRU_LISTS; i++)
3156 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3157 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3159 /* Hierarchical information */
3160 memory = memsw = PAGE_COUNTER_MAX;
3161 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3162 memory = min(memory, mi->memory.limit);
3163 memsw = min(memsw, mi->memsw.limit);
3165 seq_printf(m, "hierarchical_memory_limit %llu\n",
3166 (u64)memory * PAGE_SIZE);
3167 if (do_memsw_account())
3168 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3169 (u64)memsw * PAGE_SIZE);
3171 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3172 unsigned long long val = 0;
3174 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3176 for_each_mem_cgroup_tree(mi, memcg)
3177 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3178 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3181 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3182 unsigned long long val = 0;
3184 for_each_mem_cgroup_tree(mi, memcg)
3185 val += mem_cgroup_read_events(mi, i);
3186 seq_printf(m, "total_%s %llu\n",
3187 mem_cgroup_events_names[i], val);
3190 for (i = 0; i < NR_LRU_LISTS; i++) {
3191 unsigned long long val = 0;
3193 for_each_mem_cgroup_tree(mi, memcg)
3194 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3195 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3198 #ifdef CONFIG_DEBUG_VM
3201 struct mem_cgroup_per_zone *mz;
3202 struct zone_reclaim_stat *rstat;
3203 unsigned long recent_rotated[2] = {0, 0};
3204 unsigned long recent_scanned[2] = {0, 0};
3206 for_each_online_node(nid)
3207 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3208 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3209 rstat = &mz->lruvec.reclaim_stat;
3211 recent_rotated[0] += rstat->recent_rotated[0];
3212 recent_rotated[1] += rstat->recent_rotated[1];
3213 recent_scanned[0] += rstat->recent_scanned[0];
3214 recent_scanned[1] += rstat->recent_scanned[1];
3216 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3217 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3218 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3219 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3226 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3229 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3231 return mem_cgroup_swappiness(memcg);
3234 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3235 struct cftype *cft, u64 val)
3237 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3243 memcg->swappiness = val;
3245 vm_swappiness = val;
3250 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3252 struct mem_cgroup_threshold_ary *t;
3253 unsigned long usage;
3258 t = rcu_dereference(memcg->thresholds.primary);
3260 t = rcu_dereference(memcg->memsw_thresholds.primary);
3265 usage = mem_cgroup_usage(memcg, swap);
3268 * current_threshold points to threshold just below or equal to usage.
3269 * If it's not true, a threshold was crossed after last
3270 * call of __mem_cgroup_threshold().
3272 i = t->current_threshold;
3275 * Iterate backward over array of thresholds starting from
3276 * current_threshold and check if a threshold is crossed.
3277 * If none of thresholds below usage is crossed, we read
3278 * only one element of the array here.
3280 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3281 eventfd_signal(t->entries[i].eventfd, 1);
3283 /* i = current_threshold + 1 */
3287 * Iterate forward over array of thresholds starting from
3288 * current_threshold+1 and check if a threshold is crossed.
3289 * If none of thresholds above usage is crossed, we read
3290 * only one element of the array here.
3292 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3293 eventfd_signal(t->entries[i].eventfd, 1);
3295 /* Update current_threshold */
3296 t->current_threshold = i - 1;
3301 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3304 __mem_cgroup_threshold(memcg, false);
3305 if (do_memsw_account())
3306 __mem_cgroup_threshold(memcg, true);
3308 memcg = parent_mem_cgroup(memcg);
3312 static int compare_thresholds(const void *a, const void *b)
3314 const struct mem_cgroup_threshold *_a = a;
3315 const struct mem_cgroup_threshold *_b = b;
3317 if (_a->threshold > _b->threshold)
3320 if (_a->threshold < _b->threshold)
3326 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3328 struct mem_cgroup_eventfd_list *ev;
3330 spin_lock(&memcg_oom_lock);
3332 list_for_each_entry(ev, &memcg->oom_notify, list)
3333 eventfd_signal(ev->eventfd, 1);
3335 spin_unlock(&memcg_oom_lock);
3339 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3341 struct mem_cgroup *iter;
3343 for_each_mem_cgroup_tree(iter, memcg)
3344 mem_cgroup_oom_notify_cb(iter);
3347 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3348 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3350 struct mem_cgroup_thresholds *thresholds;
3351 struct mem_cgroup_threshold_ary *new;
3352 unsigned long threshold;
3353 unsigned long usage;
3356 ret = page_counter_memparse(args, "-1", &threshold);
3360 mutex_lock(&memcg->thresholds_lock);
3363 thresholds = &memcg->thresholds;
3364 usage = mem_cgroup_usage(memcg, false);
3365 } else if (type == _MEMSWAP) {
3366 thresholds = &memcg->memsw_thresholds;
3367 usage = mem_cgroup_usage(memcg, true);
3371 /* Check if a threshold crossed before adding a new one */
3372 if (thresholds->primary)
3373 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3375 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3377 /* Allocate memory for new array of thresholds */
3378 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3386 /* Copy thresholds (if any) to new array */
3387 if (thresholds->primary) {
3388 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3389 sizeof(struct mem_cgroup_threshold));
3392 /* Add new threshold */
3393 new->entries[size - 1].eventfd = eventfd;
3394 new->entries[size - 1].threshold = threshold;
3396 /* Sort thresholds. Registering of new threshold isn't time-critical */
3397 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3398 compare_thresholds, NULL);
3400 /* Find current threshold */
3401 new->current_threshold = -1;
3402 for (i = 0; i < size; i++) {
3403 if (new->entries[i].threshold <= usage) {
3405 * new->current_threshold will not be used until
3406 * rcu_assign_pointer(), so it's safe to increment
3409 ++new->current_threshold;
3414 /* Free old spare buffer and save old primary buffer as spare */
3415 kfree(thresholds->spare);
3416 thresholds->spare = thresholds->primary;
3418 rcu_assign_pointer(thresholds->primary, new);
3420 /* To be sure that nobody uses thresholds */
3424 mutex_unlock(&memcg->thresholds_lock);
3429 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3430 struct eventfd_ctx *eventfd, const char *args)
3432 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3435 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3436 struct eventfd_ctx *eventfd, const char *args)
3438 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3441 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3442 struct eventfd_ctx *eventfd, enum res_type type)
3444 struct mem_cgroup_thresholds *thresholds;
3445 struct mem_cgroup_threshold_ary *new;
3446 unsigned long usage;
3449 mutex_lock(&memcg->thresholds_lock);
3452 thresholds = &memcg->thresholds;
3453 usage = mem_cgroup_usage(memcg, false);
3454 } else if (type == _MEMSWAP) {
3455 thresholds = &memcg->memsw_thresholds;
3456 usage = mem_cgroup_usage(memcg, true);
3460 if (!thresholds->primary)
3463 /* Check if a threshold crossed before removing */
3464 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3466 /* Calculate new number of threshold */
3468 for (i = 0; i < thresholds->primary->size; i++) {
3469 if (thresholds->primary->entries[i].eventfd != eventfd)
3473 new = thresholds->spare;
3475 /* Set thresholds array to NULL if we don't have thresholds */
3484 /* Copy thresholds and find current threshold */
3485 new->current_threshold = -1;
3486 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3487 if (thresholds->primary->entries[i].eventfd == eventfd)
3490 new->entries[j] = thresholds->primary->entries[i];
3491 if (new->entries[j].threshold <= usage) {
3493 * new->current_threshold will not be used
3494 * until rcu_assign_pointer(), so it's safe to increment
3497 ++new->current_threshold;
3503 /* Swap primary and spare array */
3504 thresholds->spare = thresholds->primary;
3506 rcu_assign_pointer(thresholds->primary, new);
3508 /* To be sure that nobody uses thresholds */
3511 /* If all events are unregistered, free the spare array */
3513 kfree(thresholds->spare);
3514 thresholds->spare = NULL;
3517 mutex_unlock(&memcg->thresholds_lock);
3520 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3521 struct eventfd_ctx *eventfd)
3523 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3526 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3527 struct eventfd_ctx *eventfd)
3529 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3532 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3533 struct eventfd_ctx *eventfd, const char *args)
3535 struct mem_cgroup_eventfd_list *event;
3537 event = kmalloc(sizeof(*event), GFP_KERNEL);
3541 spin_lock(&memcg_oom_lock);
3543 event->eventfd = eventfd;
3544 list_add(&event->list, &memcg->oom_notify);
3546 /* already in OOM ? */
3547 if (memcg->under_oom)
3548 eventfd_signal(eventfd, 1);
3549 spin_unlock(&memcg_oom_lock);
3554 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3555 struct eventfd_ctx *eventfd)
3557 struct mem_cgroup_eventfd_list *ev, *tmp;
3559 spin_lock(&memcg_oom_lock);
3561 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3562 if (ev->eventfd == eventfd) {
3563 list_del(&ev->list);
3568 spin_unlock(&memcg_oom_lock);
3571 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3573 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3575 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3576 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3580 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3581 struct cftype *cft, u64 val)
3583 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3585 /* cannot set to root cgroup and only 0 and 1 are allowed */
3586 if (!css->parent || !((val == 0) || (val == 1)))
3589 memcg->oom_kill_disable = val;
3591 memcg_oom_recover(memcg);
3596 #ifdef CONFIG_CGROUP_WRITEBACK
3598 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3600 return &memcg->cgwb_list;
3603 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3605 return wb_domain_init(&memcg->cgwb_domain, gfp);
3608 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3610 wb_domain_exit(&memcg->cgwb_domain);
3613 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3615 wb_domain_size_changed(&memcg->cgwb_domain);
3618 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3620 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3622 if (!memcg->css.parent)
3625 return &memcg->cgwb_domain;
3629 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3630 * @wb: bdi_writeback in question
3631 * @pfilepages: out parameter for number of file pages
3632 * @pheadroom: out parameter for number of allocatable pages according to memcg
3633 * @pdirty: out parameter for number of dirty pages
3634 * @pwriteback: out parameter for number of pages under writeback
3636 * Determine the numbers of file, headroom, dirty, and writeback pages in
3637 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3638 * is a bit more involved.
3640 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3641 * headroom is calculated as the lowest headroom of itself and the
3642 * ancestors. Note that this doesn't consider the actual amount of
3643 * available memory in the system. The caller should further cap
3644 * *@pheadroom accordingly.
3646 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3647 unsigned long *pheadroom, unsigned long *pdirty,
3648 unsigned long *pwriteback)
3650 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3651 struct mem_cgroup *parent;
3653 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3655 /* this should eventually include NR_UNSTABLE_NFS */
3656 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3657 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3658 (1 << LRU_ACTIVE_FILE));
3659 *pheadroom = PAGE_COUNTER_MAX;
3661 while ((parent = parent_mem_cgroup(memcg))) {
3662 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3663 unsigned long used = page_counter_read(&memcg->memory);
3665 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3670 #else /* CONFIG_CGROUP_WRITEBACK */
3672 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3677 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3681 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3685 #endif /* CONFIG_CGROUP_WRITEBACK */
3688 * DO NOT USE IN NEW FILES.
3690 * "cgroup.event_control" implementation.
3692 * This is way over-engineered. It tries to support fully configurable
3693 * events for each user. Such level of flexibility is completely
3694 * unnecessary especially in the light of the planned unified hierarchy.
3696 * Please deprecate this and replace with something simpler if at all
3701 * Unregister event and free resources.
3703 * Gets called from workqueue.
3705 static void memcg_event_remove(struct work_struct *work)
3707 struct mem_cgroup_event *event =
3708 container_of(work, struct mem_cgroup_event, remove);
3709 struct mem_cgroup *memcg = event->memcg;
3711 remove_wait_queue(event->wqh, &event->wait);
3713 event->unregister_event(memcg, event->eventfd);
3715 /* Notify userspace the event is going away. */
3716 eventfd_signal(event->eventfd, 1);
3718 eventfd_ctx_put(event->eventfd);
3720 css_put(&memcg->css);
3724 * Gets called on POLLHUP on eventfd when user closes it.
3726 * Called with wqh->lock held and interrupts disabled.
3728 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3729 int sync, void *key)
3731 struct mem_cgroup_event *event =
3732 container_of(wait, struct mem_cgroup_event, wait);
3733 struct mem_cgroup *memcg = event->memcg;
3734 unsigned long flags = (unsigned long)key;
3736 if (flags & POLLHUP) {
3738 * If the event has been detached at cgroup removal, we
3739 * can simply return knowing the other side will cleanup
3742 * We can't race against event freeing since the other
3743 * side will require wqh->lock via remove_wait_queue(),
3746 spin_lock(&memcg->event_list_lock);
3747 if (!list_empty(&event->list)) {
3748 list_del_init(&event->list);
3750 * We are in atomic context, but cgroup_event_remove()
3751 * may sleep, so we have to call it in workqueue.
3753 schedule_work(&event->remove);
3755 spin_unlock(&memcg->event_list_lock);
3761 static void memcg_event_ptable_queue_proc(struct file *file,
3762 wait_queue_head_t *wqh, poll_table *pt)
3764 struct mem_cgroup_event *event =
3765 container_of(pt, struct mem_cgroup_event, pt);
3768 add_wait_queue(wqh, &event->wait);
3772 * DO NOT USE IN NEW FILES.
3774 * Parse input and register new cgroup event handler.
3776 * Input must be in format '<event_fd> <control_fd> <args>'.
3777 * Interpretation of args is defined by control file implementation.
3779 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3780 char *buf, size_t nbytes, loff_t off)
3782 struct cgroup_subsys_state *css = of_css(of);
3783 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3784 struct mem_cgroup_event *event;
3785 struct cgroup_subsys_state *cfile_css;
3786 unsigned int efd, cfd;
3793 buf = strstrip(buf);
3795 efd = simple_strtoul(buf, &endp, 10);
3800 cfd = simple_strtoul(buf, &endp, 10);
3801 if ((*endp != ' ') && (*endp != '\0'))
3805 event = kzalloc(sizeof(*event), GFP_KERNEL);
3809 event->memcg = memcg;
3810 INIT_LIST_HEAD(&event->list);
3811 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3812 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3813 INIT_WORK(&event->remove, memcg_event_remove);
3821 event->eventfd = eventfd_ctx_fileget(efile.file);
3822 if (IS_ERR(event->eventfd)) {
3823 ret = PTR_ERR(event->eventfd);
3830 goto out_put_eventfd;
3833 /* the process need read permission on control file */
3834 /* AV: shouldn't we check that it's been opened for read instead? */
3835 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3840 * Determine the event callbacks and set them in @event. This used
3841 * to be done via struct cftype but cgroup core no longer knows
3842 * about these events. The following is crude but the whole thing
3843 * is for compatibility anyway.
3845 * DO NOT ADD NEW FILES.
3847 name = cfile.file->f_path.dentry->d_name.name;
3849 if (!strcmp(name, "memory.usage_in_bytes")) {
3850 event->register_event = mem_cgroup_usage_register_event;
3851 event->unregister_event = mem_cgroup_usage_unregister_event;
3852 } else if (!strcmp(name, "memory.oom_control")) {
3853 event->register_event = mem_cgroup_oom_register_event;
3854 event->unregister_event = mem_cgroup_oom_unregister_event;
3855 } else if (!strcmp(name, "memory.pressure_level")) {
3856 event->register_event = vmpressure_register_event;
3857 event->unregister_event = vmpressure_unregister_event;
3858 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3859 event->register_event = memsw_cgroup_usage_register_event;
3860 event->unregister_event = memsw_cgroup_usage_unregister_event;
3867 * Verify @cfile should belong to @css. Also, remaining events are
3868 * automatically removed on cgroup destruction but the removal is
3869 * asynchronous, so take an extra ref on @css.
3871 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3872 &memory_cgrp_subsys);
3874 if (IS_ERR(cfile_css))
3876 if (cfile_css != css) {
3881 ret = event->register_event(memcg, event->eventfd, buf);
3885 efile.file->f_op->poll(efile.file, &event->pt);
3887 spin_lock(&memcg->event_list_lock);
3888 list_add(&event->list, &memcg->event_list);
3889 spin_unlock(&memcg->event_list_lock);
3901 eventfd_ctx_put(event->eventfd);
3910 static struct cftype mem_cgroup_legacy_files[] = {
3912 .name = "usage_in_bytes",
3913 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3914 .read_u64 = mem_cgroup_read_u64,
3917 .name = "max_usage_in_bytes",
3918 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3919 .write = mem_cgroup_reset,
3920 .read_u64 = mem_cgroup_read_u64,
3923 .name = "limit_in_bytes",
3924 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3925 .write = mem_cgroup_write,
3926 .read_u64 = mem_cgroup_read_u64,
3929 .name = "soft_limit_in_bytes",
3930 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3931 .write = mem_cgroup_write,
3932 .read_u64 = mem_cgroup_read_u64,
3936 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3937 .write = mem_cgroup_reset,
3938 .read_u64 = mem_cgroup_read_u64,
3942 .seq_show = memcg_stat_show,
3945 .name = "force_empty",
3946 .write = mem_cgroup_force_empty_write,
3949 .name = "use_hierarchy",
3950 .write_u64 = mem_cgroup_hierarchy_write,
3951 .read_u64 = mem_cgroup_hierarchy_read,
3954 .name = "cgroup.event_control", /* XXX: for compat */
3955 .write = memcg_write_event_control,
3956 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3959 .name = "swappiness",
3960 .read_u64 = mem_cgroup_swappiness_read,
3961 .write_u64 = mem_cgroup_swappiness_write,
3964 .name = "move_charge_at_immigrate",
3965 .read_u64 = mem_cgroup_move_charge_read,
3966 .write_u64 = mem_cgroup_move_charge_write,
3969 .name = "oom_control",
3970 .seq_show = mem_cgroup_oom_control_read,
3971 .write_u64 = mem_cgroup_oom_control_write,
3972 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3975 .name = "pressure_level",
3979 .name = "numa_stat",
3980 .seq_show = memcg_numa_stat_show,
3984 .name = "kmem.limit_in_bytes",
3985 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
3986 .write = mem_cgroup_write,
3987 .read_u64 = mem_cgroup_read_u64,
3990 .name = "kmem.usage_in_bytes",
3991 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
3992 .read_u64 = mem_cgroup_read_u64,
3995 .name = "kmem.failcnt",
3996 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
3997 .write = mem_cgroup_reset,
3998 .read_u64 = mem_cgroup_read_u64,
4001 .name = "kmem.max_usage_in_bytes",
4002 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4003 .write = mem_cgroup_reset,
4004 .read_u64 = mem_cgroup_read_u64,
4006 #ifdef CONFIG_SLABINFO
4008 .name = "kmem.slabinfo",
4009 .seq_start = slab_start,
4010 .seq_next = slab_next,
4011 .seq_stop = slab_stop,
4012 .seq_show = memcg_slab_show,
4016 .name = "kmem.tcp.limit_in_bytes",
4017 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4018 .write = mem_cgroup_write,
4019 .read_u64 = mem_cgroup_read_u64,
4022 .name = "kmem.tcp.usage_in_bytes",
4023 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4024 .read_u64 = mem_cgroup_read_u64,
4027 .name = "kmem.tcp.failcnt",
4028 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4029 .write = mem_cgroup_reset,
4030 .read_u64 = mem_cgroup_read_u64,
4033 .name = "kmem.tcp.max_usage_in_bytes",
4034 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4035 .write = mem_cgroup_reset,
4036 .read_u64 = mem_cgroup_read_u64,
4038 { }, /* terminate */
4041 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4043 struct mem_cgroup_per_node *pn;
4044 struct mem_cgroup_per_zone *mz;
4045 int zone, tmp = node;
4047 * This routine is called against possible nodes.
4048 * But it's BUG to call kmalloc() against offline node.
4050 * TODO: this routine can waste much memory for nodes which will
4051 * never be onlined. It's better to use memory hotplug callback
4054 if (!node_state(node, N_NORMAL_MEMORY))
4056 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4060 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4061 mz = &pn->zoneinfo[zone];
4062 lruvec_init(&mz->lruvec);
4063 mz->usage_in_excess = 0;
4064 mz->on_tree = false;
4067 memcg->nodeinfo[node] = pn;
4071 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4073 kfree(memcg->nodeinfo[node]);
4076 static void mem_cgroup_free(struct mem_cgroup *memcg)
4080 memcg_wb_domain_exit(memcg);
4082 free_mem_cgroup_per_zone_info(memcg, node);
4083 free_percpu(memcg->stat);
4087 static struct mem_cgroup *mem_cgroup_alloc(void)
4089 struct mem_cgroup *memcg;
4093 size = sizeof(struct mem_cgroup);
4094 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4096 memcg = kzalloc(size, GFP_KERNEL);
4100 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4105 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4108 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4111 INIT_WORK(&memcg->high_work, high_work_func);
4112 memcg->last_scanned_node = MAX_NUMNODES;
4113 INIT_LIST_HEAD(&memcg->oom_notify);
4114 mutex_init(&memcg->thresholds_lock);
4115 spin_lock_init(&memcg->move_lock);
4116 vmpressure_init(&memcg->vmpressure);
4117 INIT_LIST_HEAD(&memcg->event_list);
4118 spin_lock_init(&memcg->event_list_lock);
4119 memcg->socket_pressure = jiffies;
4121 memcg->kmemcg_id = -1;
4123 #ifdef CONFIG_CGROUP_WRITEBACK
4124 INIT_LIST_HEAD(&memcg->cgwb_list);
4128 mem_cgroup_free(memcg);
4132 static struct cgroup_subsys_state * __ref
4133 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4135 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4136 struct mem_cgroup *memcg;
4137 long error = -ENOMEM;
4139 memcg = mem_cgroup_alloc();
4141 return ERR_PTR(error);
4143 memcg->high = PAGE_COUNTER_MAX;
4144 memcg->soft_limit = PAGE_COUNTER_MAX;
4146 memcg->swappiness = mem_cgroup_swappiness(parent);
4147 memcg->oom_kill_disable = parent->oom_kill_disable;
4149 if (parent && parent->use_hierarchy) {
4150 memcg->use_hierarchy = true;
4151 page_counter_init(&memcg->memory, &parent->memory);
4152 page_counter_init(&memcg->swap, &parent->swap);
4153 page_counter_init(&memcg->memsw, &parent->memsw);
4154 page_counter_init(&memcg->kmem, &parent->kmem);
4155 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4157 page_counter_init(&memcg->memory, NULL);
4158 page_counter_init(&memcg->swap, NULL);
4159 page_counter_init(&memcg->memsw, NULL);
4160 page_counter_init(&memcg->kmem, NULL);
4161 page_counter_init(&memcg->tcpmem, NULL);
4163 * Deeper hierachy with use_hierarchy == false doesn't make
4164 * much sense so let cgroup subsystem know about this
4165 * unfortunate state in our controller.
4167 if (parent != root_mem_cgroup)
4168 memory_cgrp_subsys.broken_hierarchy = true;
4171 /* The following stuff does not apply to the root */
4173 root_mem_cgroup = memcg;
4177 error = memcg_online_kmem(memcg);
4181 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4182 static_branch_inc(&memcg_sockets_enabled_key);
4186 mem_cgroup_free(memcg);
4191 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4193 if (css->id > MEM_CGROUP_ID_MAX)
4199 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4201 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4202 struct mem_cgroup_event *event, *tmp;
4205 * Unregister events and notify userspace.
4206 * Notify userspace about cgroup removing only after rmdir of cgroup
4207 * directory to avoid race between userspace and kernelspace.
4209 spin_lock(&memcg->event_list_lock);
4210 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4211 list_del_init(&event->list);
4212 schedule_work(&event->remove);
4214 spin_unlock(&memcg->event_list_lock);
4216 memcg_offline_kmem(memcg);
4217 wb_memcg_offline(memcg);
4220 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4222 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4224 invalidate_reclaim_iterators(memcg);
4227 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4229 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4231 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4232 static_branch_dec(&memcg_sockets_enabled_key);
4234 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4235 static_branch_dec(&memcg_sockets_enabled_key);
4237 vmpressure_cleanup(&memcg->vmpressure);
4238 cancel_work_sync(&memcg->high_work);
4239 mem_cgroup_remove_from_trees(memcg);
4240 memcg_free_kmem(memcg);
4241 mem_cgroup_free(memcg);
4245 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4246 * @css: the target css
4248 * Reset the states of the mem_cgroup associated with @css. This is
4249 * invoked when the userland requests disabling on the default hierarchy
4250 * but the memcg is pinned through dependency. The memcg should stop
4251 * applying policies and should revert to the vanilla state as it may be
4252 * made visible again.
4254 * The current implementation only resets the essential configurations.
4255 * This needs to be expanded to cover all the visible parts.
4257 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4259 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4261 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4262 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4263 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4264 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4265 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4267 memcg->high = PAGE_COUNTER_MAX;
4268 memcg->soft_limit = PAGE_COUNTER_MAX;
4269 memcg_wb_domain_size_changed(memcg);
4273 /* Handlers for move charge at task migration. */
4274 static int mem_cgroup_do_precharge(unsigned long count)
4278 /* Try a single bulk charge without reclaim first, kswapd may wake */
4279 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4281 mc.precharge += count;
4285 /* Try charges one by one with reclaim */
4287 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4297 * get_mctgt_type - get target type of moving charge
4298 * @vma: the vma the pte to be checked belongs
4299 * @addr: the address corresponding to the pte to be checked
4300 * @ptent: the pte to be checked
4301 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4304 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4305 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4306 * move charge. if @target is not NULL, the page is stored in target->page
4307 * with extra refcnt got(Callers should handle it).
4308 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4309 * target for charge migration. if @target is not NULL, the entry is stored
4312 * Called with pte lock held.
4319 enum mc_target_type {
4325 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4326 unsigned long addr, pte_t ptent)
4328 struct page *page = vm_normal_page(vma, addr, ptent);
4330 if (!page || !page_mapped(page))
4332 if (PageAnon(page)) {
4333 if (!(mc.flags & MOVE_ANON))
4336 if (!(mc.flags & MOVE_FILE))
4339 if (!get_page_unless_zero(page))
4346 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4347 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4349 struct page *page = NULL;
4350 swp_entry_t ent = pte_to_swp_entry(ptent);
4352 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4355 * Because lookup_swap_cache() updates some statistics counter,
4356 * we call find_get_page() with swapper_space directly.
4358 page = find_get_page(swap_address_space(ent), ent.val);
4359 if (do_memsw_account())
4360 entry->val = ent.val;
4365 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4366 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4372 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4373 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4375 struct page *page = NULL;
4376 struct address_space *mapping;
4379 if (!vma->vm_file) /* anonymous vma */
4381 if (!(mc.flags & MOVE_FILE))
4384 mapping = vma->vm_file->f_mapping;
4385 pgoff = linear_page_index(vma, addr);
4387 /* page is moved even if it's not RSS of this task(page-faulted). */
4389 /* shmem/tmpfs may report page out on swap: account for that too. */
4390 if (shmem_mapping(mapping)) {
4391 page = find_get_entry(mapping, pgoff);
4392 if (radix_tree_exceptional_entry(page)) {
4393 swp_entry_t swp = radix_to_swp_entry(page);
4394 if (do_memsw_account())
4396 page = find_get_page(swap_address_space(swp), swp.val);
4399 page = find_get_page(mapping, pgoff);
4401 page = find_get_page(mapping, pgoff);
4407 * mem_cgroup_move_account - move account of the page
4409 * @nr_pages: number of regular pages (>1 for huge pages)
4410 * @from: mem_cgroup which the page is moved from.
4411 * @to: mem_cgroup which the page is moved to. @from != @to.
4413 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4415 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4418 static int mem_cgroup_move_account(struct page *page,
4420 struct mem_cgroup *from,
4421 struct mem_cgroup *to)
4423 unsigned long flags;
4424 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4428 VM_BUG_ON(from == to);
4429 VM_BUG_ON_PAGE(PageLRU(page), page);
4430 VM_BUG_ON(compound && !PageTransHuge(page));
4433 * Prevent mem_cgroup_migrate() from looking at
4434 * page->mem_cgroup of its source page while we change it.
4437 if (!trylock_page(page))
4441 if (page->mem_cgroup != from)
4444 anon = PageAnon(page);
4446 spin_lock_irqsave(&from->move_lock, flags);
4448 if (!anon && page_mapped(page)) {
4449 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4451 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4456 * move_lock grabbed above and caller set from->moving_account, so
4457 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4458 * So mapping should be stable for dirty pages.
4460 if (!anon && PageDirty(page)) {
4461 struct address_space *mapping = page_mapping(page);
4463 if (mapping_cap_account_dirty(mapping)) {
4464 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4466 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4471 if (PageWriteback(page)) {
4472 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4474 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4479 * It is safe to change page->mem_cgroup here because the page
4480 * is referenced, charged, and isolated - we can't race with
4481 * uncharging, charging, migration, or LRU putback.
4484 /* caller should have done css_get */
4485 page->mem_cgroup = to;
4486 spin_unlock_irqrestore(&from->move_lock, flags);
4490 local_irq_disable();
4491 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4492 memcg_check_events(to, page);
4493 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4494 memcg_check_events(from, page);
4502 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4503 unsigned long addr, pte_t ptent, union mc_target *target)
4505 struct page *page = NULL;
4506 enum mc_target_type ret = MC_TARGET_NONE;
4507 swp_entry_t ent = { .val = 0 };
4509 if (pte_present(ptent))
4510 page = mc_handle_present_pte(vma, addr, ptent);
4511 else if (is_swap_pte(ptent))
4512 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4513 else if (pte_none(ptent))
4514 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4516 if (!page && !ent.val)
4520 * Do only loose check w/o serialization.
4521 * mem_cgroup_move_account() checks the page is valid or
4522 * not under LRU exclusion.
4524 if (page->mem_cgroup == mc.from) {
4525 ret = MC_TARGET_PAGE;
4527 target->page = page;
4529 if (!ret || !target)
4532 /* There is a swap entry and a page doesn't exist or isn't charged */
4533 if (ent.val && !ret &&
4534 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4535 ret = MC_TARGET_SWAP;
4542 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4544 * We don't consider swapping or file mapped pages because THP does not
4545 * support them for now.
4546 * Caller should make sure that pmd_trans_huge(pmd) is true.
4548 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4549 unsigned long addr, pmd_t pmd, union mc_target *target)
4551 struct page *page = NULL;
4552 enum mc_target_type ret = MC_TARGET_NONE;
4554 page = pmd_page(pmd);
4555 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4556 if (!(mc.flags & MOVE_ANON))
4558 if (page->mem_cgroup == mc.from) {
4559 ret = MC_TARGET_PAGE;
4562 target->page = page;
4568 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4569 unsigned long addr, pmd_t pmd, union mc_target *target)
4571 return MC_TARGET_NONE;
4575 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4576 unsigned long addr, unsigned long end,
4577 struct mm_walk *walk)
4579 struct vm_area_struct *vma = walk->vma;
4583 ptl = pmd_trans_huge_lock(pmd, vma);
4585 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4586 mc.precharge += HPAGE_PMD_NR;
4591 if (pmd_trans_unstable(pmd))
4593 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4594 for (; addr != end; pte++, addr += PAGE_SIZE)
4595 if (get_mctgt_type(vma, addr, *pte, NULL))
4596 mc.precharge++; /* increment precharge temporarily */
4597 pte_unmap_unlock(pte - 1, ptl);
4603 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4605 unsigned long precharge;
4607 struct mm_walk mem_cgroup_count_precharge_walk = {
4608 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4611 down_read(&mm->mmap_sem);
4612 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4613 up_read(&mm->mmap_sem);
4615 precharge = mc.precharge;
4621 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4623 unsigned long precharge = mem_cgroup_count_precharge(mm);
4625 VM_BUG_ON(mc.moving_task);
4626 mc.moving_task = current;
4627 return mem_cgroup_do_precharge(precharge);
4630 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4631 static void __mem_cgroup_clear_mc(void)
4633 struct mem_cgroup *from = mc.from;
4634 struct mem_cgroup *to = mc.to;
4636 /* we must uncharge all the leftover precharges from mc.to */
4638 cancel_charge(mc.to, mc.precharge);
4642 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4643 * we must uncharge here.
4645 if (mc.moved_charge) {
4646 cancel_charge(mc.from, mc.moved_charge);
4647 mc.moved_charge = 0;
4649 /* we must fixup refcnts and charges */
4650 if (mc.moved_swap) {
4651 /* uncharge swap account from the old cgroup */
4652 if (!mem_cgroup_is_root(mc.from))
4653 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4656 * we charged both to->memory and to->memsw, so we
4657 * should uncharge to->memory.
4659 if (!mem_cgroup_is_root(mc.to))
4660 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4662 css_put_many(&mc.from->css, mc.moved_swap);
4664 /* we've already done css_get(mc.to) */
4667 memcg_oom_recover(from);
4668 memcg_oom_recover(to);
4669 wake_up_all(&mc.waitq);
4672 static void mem_cgroup_clear_mc(void)
4675 * we must clear moving_task before waking up waiters at the end of
4678 mc.moving_task = NULL;
4679 __mem_cgroup_clear_mc();
4680 spin_lock(&mc.lock);
4683 spin_unlock(&mc.lock);
4686 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4688 struct cgroup_subsys_state *css;
4689 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4690 struct mem_cgroup *from;
4691 struct task_struct *leader, *p;
4692 struct mm_struct *mm;
4693 unsigned long move_flags;
4696 /* charge immigration isn't supported on the default hierarchy */
4697 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4701 * Multi-process migrations only happen on the default hierarchy
4702 * where charge immigration is not used. Perform charge
4703 * immigration if @tset contains a leader and whine if there are
4707 cgroup_taskset_for_each_leader(leader, css, tset) {
4710 memcg = mem_cgroup_from_css(css);
4716 * We are now commited to this value whatever it is. Changes in this
4717 * tunable will only affect upcoming migrations, not the current one.
4718 * So we need to save it, and keep it going.
4720 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4724 from = mem_cgroup_from_task(p);
4726 VM_BUG_ON(from == memcg);
4728 mm = get_task_mm(p);
4731 /* We move charges only when we move a owner of the mm */
4732 if (mm->owner == p) {
4735 VM_BUG_ON(mc.precharge);
4736 VM_BUG_ON(mc.moved_charge);
4737 VM_BUG_ON(mc.moved_swap);
4739 spin_lock(&mc.lock);
4742 mc.flags = move_flags;
4743 spin_unlock(&mc.lock);
4744 /* We set mc.moving_task later */
4746 ret = mem_cgroup_precharge_mc(mm);
4748 mem_cgroup_clear_mc();
4754 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4757 mem_cgroup_clear_mc();
4760 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4761 unsigned long addr, unsigned long end,
4762 struct mm_walk *walk)
4765 struct vm_area_struct *vma = walk->vma;
4768 enum mc_target_type target_type;
4769 union mc_target target;
4772 ptl = pmd_trans_huge_lock(pmd, vma);
4774 if (mc.precharge < HPAGE_PMD_NR) {
4778 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4779 if (target_type == MC_TARGET_PAGE) {
4781 if (!isolate_lru_page(page)) {
4782 if (!mem_cgroup_move_account(page, true,
4784 mc.precharge -= HPAGE_PMD_NR;
4785 mc.moved_charge += HPAGE_PMD_NR;
4787 putback_lru_page(page);
4795 if (pmd_trans_unstable(pmd))
4798 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4799 for (; addr != end; addr += PAGE_SIZE) {
4800 pte_t ptent = *(pte++);
4806 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4807 case MC_TARGET_PAGE:
4810 * We can have a part of the split pmd here. Moving it
4811 * can be done but it would be too convoluted so simply
4812 * ignore such a partial THP and keep it in original
4813 * memcg. There should be somebody mapping the head.
4815 if (PageTransCompound(page))
4817 if (isolate_lru_page(page))
4819 if (!mem_cgroup_move_account(page, false,
4822 /* we uncharge from mc.from later. */
4825 putback_lru_page(page);
4826 put: /* get_mctgt_type() gets the page */
4829 case MC_TARGET_SWAP:
4831 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4833 /* we fixup refcnts and charges later. */
4841 pte_unmap_unlock(pte - 1, ptl);
4846 * We have consumed all precharges we got in can_attach().
4847 * We try charge one by one, but don't do any additional
4848 * charges to mc.to if we have failed in charge once in attach()
4851 ret = mem_cgroup_do_precharge(1);
4859 static void mem_cgroup_move_charge(struct mm_struct *mm)
4861 struct mm_walk mem_cgroup_move_charge_walk = {
4862 .pmd_entry = mem_cgroup_move_charge_pte_range,
4866 lru_add_drain_all();
4868 * Signal lock_page_memcg() to take the memcg's move_lock
4869 * while we're moving its pages to another memcg. Then wait
4870 * for already started RCU-only updates to finish.
4872 atomic_inc(&mc.from->moving_account);
4875 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
4877 * Someone who are holding the mmap_sem might be waiting in
4878 * waitq. So we cancel all extra charges, wake up all waiters,
4879 * and retry. Because we cancel precharges, we might not be able
4880 * to move enough charges, but moving charge is a best-effort
4881 * feature anyway, so it wouldn't be a big problem.
4883 __mem_cgroup_clear_mc();
4888 * When we have consumed all precharges and failed in doing
4889 * additional charge, the page walk just aborts.
4891 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4892 up_read(&mm->mmap_sem);
4893 atomic_dec(&mc.from->moving_account);
4896 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4898 struct cgroup_subsys_state *css;
4899 struct task_struct *p = cgroup_taskset_first(tset, &css);
4900 struct mm_struct *mm = get_task_mm(p);
4904 mem_cgroup_move_charge(mm);
4908 mem_cgroup_clear_mc();
4910 #else /* !CONFIG_MMU */
4911 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4915 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4918 static void mem_cgroup_move_task(struct cgroup_taskset *tset)
4924 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4925 * to verify whether we're attached to the default hierarchy on each mount
4928 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
4931 * use_hierarchy is forced on the default hierarchy. cgroup core
4932 * guarantees that @root doesn't have any children, so turning it
4933 * on for the root memcg is enough.
4935 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4936 root_mem_cgroup->use_hierarchy = true;
4938 root_mem_cgroup->use_hierarchy = false;
4941 static u64 memory_current_read(struct cgroup_subsys_state *css,
4944 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4946 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
4949 static int memory_low_show(struct seq_file *m, void *v)
4951 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4952 unsigned long low = READ_ONCE(memcg->low);
4954 if (low == PAGE_COUNTER_MAX)
4955 seq_puts(m, "max\n");
4957 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
4962 static ssize_t memory_low_write(struct kernfs_open_file *of,
4963 char *buf, size_t nbytes, loff_t off)
4965 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4969 buf = strstrip(buf);
4970 err = page_counter_memparse(buf, "max", &low);
4979 static int memory_high_show(struct seq_file *m, void *v)
4981 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4982 unsigned long high = READ_ONCE(memcg->high);
4984 if (high == PAGE_COUNTER_MAX)
4985 seq_puts(m, "max\n");
4987 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
4992 static ssize_t memory_high_write(struct kernfs_open_file *of,
4993 char *buf, size_t nbytes, loff_t off)
4995 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4996 unsigned long nr_pages;
5000 buf = strstrip(buf);
5001 err = page_counter_memparse(buf, "max", &high);
5007 nr_pages = page_counter_read(&memcg->memory);
5008 if (nr_pages > high)
5009 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5012 memcg_wb_domain_size_changed(memcg);
5016 static int memory_max_show(struct seq_file *m, void *v)
5018 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5019 unsigned long max = READ_ONCE(memcg->memory.limit);
5021 if (max == PAGE_COUNTER_MAX)
5022 seq_puts(m, "max\n");
5024 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5029 static ssize_t memory_max_write(struct kernfs_open_file *of,
5030 char *buf, size_t nbytes, loff_t off)
5032 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5033 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5034 bool drained = false;
5038 buf = strstrip(buf);
5039 err = page_counter_memparse(buf, "max", &max);
5043 xchg(&memcg->memory.limit, max);
5046 unsigned long nr_pages = page_counter_read(&memcg->memory);
5048 if (nr_pages <= max)
5051 if (signal_pending(current)) {
5057 drain_all_stock(memcg);
5063 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5069 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5070 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5074 memcg_wb_domain_size_changed(memcg);
5078 static int memory_events_show(struct seq_file *m, void *v)
5080 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5082 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5083 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5084 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5085 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5090 static int memory_stat_show(struct seq_file *m, void *v)
5092 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5093 unsigned long stat[MEMCG_NR_STAT];
5094 unsigned long events[MEMCG_NR_EVENTS];
5098 * Provide statistics on the state of the memory subsystem as
5099 * well as cumulative event counters that show past behavior.
5101 * This list is ordered following a combination of these gradients:
5102 * 1) generic big picture -> specifics and details
5103 * 2) reflecting userspace activity -> reflecting kernel heuristics
5105 * Current memory state:
5108 tree_stat(memcg, stat);
5109 tree_events(memcg, events);
5111 seq_printf(m, "anon %llu\n",
5112 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5113 seq_printf(m, "file %llu\n",
5114 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5115 seq_printf(m, "kernel_stack %llu\n",
5116 (u64)stat[MEMCG_KERNEL_STACK] * PAGE_SIZE);
5117 seq_printf(m, "slab %llu\n",
5118 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5119 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5120 seq_printf(m, "sock %llu\n",
5121 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5123 seq_printf(m, "file_mapped %llu\n",
5124 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5125 seq_printf(m, "file_dirty %llu\n",
5126 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5127 seq_printf(m, "file_writeback %llu\n",
5128 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5130 for (i = 0; i < NR_LRU_LISTS; i++) {
5131 struct mem_cgroup *mi;
5132 unsigned long val = 0;
5134 for_each_mem_cgroup_tree(mi, memcg)
5135 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5136 seq_printf(m, "%s %llu\n",
5137 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5140 seq_printf(m, "slab_reclaimable %llu\n",
5141 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5142 seq_printf(m, "slab_unreclaimable %llu\n",
5143 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5145 /* Accumulated memory events */
5147 seq_printf(m, "pgfault %lu\n",
5148 events[MEM_CGROUP_EVENTS_PGFAULT]);
5149 seq_printf(m, "pgmajfault %lu\n",
5150 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5155 static struct cftype memory_files[] = {
5158 .flags = CFTYPE_NOT_ON_ROOT,
5159 .read_u64 = memory_current_read,
5163 .flags = CFTYPE_NOT_ON_ROOT,
5164 .seq_show = memory_low_show,
5165 .write = memory_low_write,
5169 .flags = CFTYPE_NOT_ON_ROOT,
5170 .seq_show = memory_high_show,
5171 .write = memory_high_write,
5175 .flags = CFTYPE_NOT_ON_ROOT,
5176 .seq_show = memory_max_show,
5177 .write = memory_max_write,
5181 .flags = CFTYPE_NOT_ON_ROOT,
5182 .file_offset = offsetof(struct mem_cgroup, events_file),
5183 .seq_show = memory_events_show,
5187 .flags = CFTYPE_NOT_ON_ROOT,
5188 .seq_show = memory_stat_show,
5193 struct cgroup_subsys memory_cgrp_subsys = {
5194 .css_alloc = mem_cgroup_css_alloc,
5195 .css_online = mem_cgroup_css_online,
5196 .css_offline = mem_cgroup_css_offline,
5197 .css_released = mem_cgroup_css_released,
5198 .css_free = mem_cgroup_css_free,
5199 .css_reset = mem_cgroup_css_reset,
5200 .can_attach = mem_cgroup_can_attach,
5201 .cancel_attach = mem_cgroup_cancel_attach,
5202 .attach = mem_cgroup_move_task,
5203 .bind = mem_cgroup_bind,
5204 .dfl_cftypes = memory_files,
5205 .legacy_cftypes = mem_cgroup_legacy_files,
5210 * mem_cgroup_low - check if memory consumption is below the normal range
5211 * @root: the highest ancestor to consider
5212 * @memcg: the memory cgroup to check
5214 * Returns %true if memory consumption of @memcg, and that of all
5215 * configurable ancestors up to @root, is below the normal range.
5217 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5219 if (mem_cgroup_disabled())
5223 * The toplevel group doesn't have a configurable range, so
5224 * it's never low when looked at directly, and it is not
5225 * considered an ancestor when assessing the hierarchy.
5228 if (memcg == root_mem_cgroup)
5231 if (page_counter_read(&memcg->memory) >= memcg->low)
5234 while (memcg != root) {
5235 memcg = parent_mem_cgroup(memcg);
5237 if (memcg == root_mem_cgroup)
5240 if (page_counter_read(&memcg->memory) >= memcg->low)
5247 * mem_cgroup_try_charge - try charging a page
5248 * @page: page to charge
5249 * @mm: mm context of the victim
5250 * @gfp_mask: reclaim mode
5251 * @memcgp: charged memcg return
5253 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5254 * pages according to @gfp_mask if necessary.
5256 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5257 * Otherwise, an error code is returned.
5259 * After page->mapping has been set up, the caller must finalize the
5260 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5261 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5263 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5264 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5267 struct mem_cgroup *memcg = NULL;
5268 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5271 if (mem_cgroup_disabled())
5274 if (PageSwapCache(page)) {
5276 * Every swap fault against a single page tries to charge the
5277 * page, bail as early as possible. shmem_unuse() encounters
5278 * already charged pages, too. The USED bit is protected by
5279 * the page lock, which serializes swap cache removal, which
5280 * in turn serializes uncharging.
5282 VM_BUG_ON_PAGE(!PageLocked(page), page);
5283 if (page->mem_cgroup)
5286 if (do_swap_account) {
5287 swp_entry_t ent = { .val = page_private(page), };
5288 unsigned short id = lookup_swap_cgroup_id(ent);
5291 memcg = mem_cgroup_from_id(id);
5292 if (memcg && !css_tryget_online(&memcg->css))
5299 memcg = get_mem_cgroup_from_mm(mm);
5301 ret = try_charge(memcg, gfp_mask, nr_pages);
5303 css_put(&memcg->css);
5310 * mem_cgroup_commit_charge - commit a page charge
5311 * @page: page to charge
5312 * @memcg: memcg to charge the page to
5313 * @lrucare: page might be on LRU already
5315 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5316 * after page->mapping has been set up. This must happen atomically
5317 * as part of the page instantiation, i.e. under the page table lock
5318 * for anonymous pages, under the page lock for page and swap cache.
5320 * In addition, the page must not be on the LRU during the commit, to
5321 * prevent racing with task migration. If it might be, use @lrucare.
5323 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5325 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5326 bool lrucare, bool compound)
5328 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5330 VM_BUG_ON_PAGE(!page->mapping, page);
5331 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5333 if (mem_cgroup_disabled())
5336 * Swap faults will attempt to charge the same page multiple
5337 * times. But reuse_swap_page() might have removed the page
5338 * from swapcache already, so we can't check PageSwapCache().
5343 commit_charge(page, memcg, lrucare);
5345 local_irq_disable();
5346 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5347 memcg_check_events(memcg, page);
5350 if (do_memsw_account() && PageSwapCache(page)) {
5351 swp_entry_t entry = { .val = page_private(page) };
5353 * The swap entry might not get freed for a long time,
5354 * let's not wait for it. The page already received a
5355 * memory+swap charge, drop the swap entry duplicate.
5357 mem_cgroup_uncharge_swap(entry);
5362 * mem_cgroup_cancel_charge - cancel a page charge
5363 * @page: page to charge
5364 * @memcg: memcg to charge the page to
5366 * Cancel a charge transaction started by mem_cgroup_try_charge().
5368 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5371 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5373 if (mem_cgroup_disabled())
5376 * Swap faults will attempt to charge the same page multiple
5377 * times. But reuse_swap_page() might have removed the page
5378 * from swapcache already, so we can't check PageSwapCache().
5383 cancel_charge(memcg, nr_pages);
5386 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5387 unsigned long nr_anon, unsigned long nr_file,
5388 unsigned long nr_huge, struct page *dummy_page)
5390 unsigned long nr_pages = nr_anon + nr_file;
5391 unsigned long flags;
5393 if (!mem_cgroup_is_root(memcg)) {
5394 page_counter_uncharge(&memcg->memory, nr_pages);
5395 if (do_memsw_account())
5396 page_counter_uncharge(&memcg->memsw, nr_pages);
5397 memcg_oom_recover(memcg);
5400 local_irq_save(flags);
5401 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5402 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5403 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5404 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5405 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5406 memcg_check_events(memcg, dummy_page);
5407 local_irq_restore(flags);
5409 if (!mem_cgroup_is_root(memcg))
5410 css_put_many(&memcg->css, nr_pages);
5413 static void uncharge_list(struct list_head *page_list)
5415 struct mem_cgroup *memcg = NULL;
5416 unsigned long nr_anon = 0;
5417 unsigned long nr_file = 0;
5418 unsigned long nr_huge = 0;
5419 unsigned long pgpgout = 0;
5420 struct list_head *next;
5423 next = page_list->next;
5425 unsigned int nr_pages = 1;
5427 page = list_entry(next, struct page, lru);
5428 next = page->lru.next;
5430 VM_BUG_ON_PAGE(PageLRU(page), page);
5431 VM_BUG_ON_PAGE(page_count(page), page);
5433 if (!page->mem_cgroup)
5437 * Nobody should be changing or seriously looking at
5438 * page->mem_cgroup at this point, we have fully
5439 * exclusive access to the page.
5442 if (memcg != page->mem_cgroup) {
5444 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5446 pgpgout = nr_anon = nr_file = nr_huge = 0;
5448 memcg = page->mem_cgroup;
5451 if (PageTransHuge(page)) {
5452 nr_pages <<= compound_order(page);
5453 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5454 nr_huge += nr_pages;
5458 nr_anon += nr_pages;
5460 nr_file += nr_pages;
5462 page->mem_cgroup = NULL;
5465 } while (next != page_list);
5468 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5473 * mem_cgroup_uncharge - uncharge a page
5474 * @page: page to uncharge
5476 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5477 * mem_cgroup_commit_charge().
5479 void mem_cgroup_uncharge(struct page *page)
5481 if (mem_cgroup_disabled())
5484 /* Don't touch page->lru of any random page, pre-check: */
5485 if (!page->mem_cgroup)
5488 INIT_LIST_HEAD(&page->lru);
5489 uncharge_list(&page->lru);
5493 * mem_cgroup_uncharge_list - uncharge a list of page
5494 * @page_list: list of pages to uncharge
5496 * Uncharge a list of pages previously charged with
5497 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5499 void mem_cgroup_uncharge_list(struct list_head *page_list)
5501 if (mem_cgroup_disabled())
5504 if (!list_empty(page_list))
5505 uncharge_list(page_list);
5509 * mem_cgroup_migrate - charge a page's replacement
5510 * @oldpage: currently circulating page
5511 * @newpage: replacement page
5513 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5514 * be uncharged upon free.
5516 * Both pages must be locked, @newpage->mapping must be set up.
5518 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5520 struct mem_cgroup *memcg;
5521 unsigned int nr_pages;
5524 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5525 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5526 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5527 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5530 if (mem_cgroup_disabled())
5533 /* Page cache replacement: new page already charged? */
5534 if (newpage->mem_cgroup)
5537 /* Swapcache readahead pages can get replaced before being charged */
5538 memcg = oldpage->mem_cgroup;
5542 /* Force-charge the new page. The old one will be freed soon */
5543 compound = PageTransHuge(newpage);
5544 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5546 page_counter_charge(&memcg->memory, nr_pages);
5547 if (do_memsw_account())
5548 page_counter_charge(&memcg->memsw, nr_pages);
5549 css_get_many(&memcg->css, nr_pages);
5551 commit_charge(newpage, memcg, false);
5553 local_irq_disable();
5554 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5555 memcg_check_events(memcg, newpage);
5559 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5560 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5562 void sock_update_memcg(struct sock *sk)
5564 struct mem_cgroup *memcg;
5566 /* Socket cloning can throw us here with sk_cgrp already
5567 * filled. It won't however, necessarily happen from
5568 * process context. So the test for root memcg given
5569 * the current task's memcg won't help us in this case.
5571 * Respecting the original socket's memcg is a better
5572 * decision in this case.
5575 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5576 css_get(&sk->sk_memcg->css);
5581 memcg = mem_cgroup_from_task(current);
5582 if (memcg == root_mem_cgroup)
5584 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5586 if (css_tryget_online(&memcg->css))
5587 sk->sk_memcg = memcg;
5591 EXPORT_SYMBOL(sock_update_memcg);
5593 void sock_release_memcg(struct sock *sk)
5595 WARN_ON(!sk->sk_memcg);
5596 css_put(&sk->sk_memcg->css);
5600 * mem_cgroup_charge_skmem - charge socket memory
5601 * @memcg: memcg to charge
5602 * @nr_pages: number of pages to charge
5604 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5605 * @memcg's configured limit, %false if the charge had to be forced.
5607 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5609 gfp_t gfp_mask = GFP_KERNEL;
5611 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5612 struct page_counter *fail;
5614 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5615 memcg->tcpmem_pressure = 0;
5618 page_counter_charge(&memcg->tcpmem, nr_pages);
5619 memcg->tcpmem_pressure = 1;
5623 /* Don't block in the packet receive path */
5625 gfp_mask = GFP_NOWAIT;
5627 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5629 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5632 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5637 * mem_cgroup_uncharge_skmem - uncharge socket memory
5638 * @memcg - memcg to uncharge
5639 * @nr_pages - number of pages to uncharge
5641 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5643 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5644 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5648 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5650 page_counter_uncharge(&memcg->memory, nr_pages);
5651 css_put_many(&memcg->css, nr_pages);
5654 static int __init cgroup_memory(char *s)
5658 while ((token = strsep(&s, ",")) != NULL) {
5661 if (!strcmp(token, "nosocket"))
5662 cgroup_memory_nosocket = true;
5663 if (!strcmp(token, "nokmem"))
5664 cgroup_memory_nokmem = true;
5668 __setup("cgroup.memory=", cgroup_memory);
5671 * subsys_initcall() for memory controller.
5673 * Some parts like hotcpu_notifier() have to be initialized from this context
5674 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5675 * everything that doesn't depend on a specific mem_cgroup structure should
5676 * be initialized from here.
5678 static int __init mem_cgroup_init(void)
5682 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5684 for_each_possible_cpu(cpu)
5685 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5688 for_each_node(node) {
5689 struct mem_cgroup_tree_per_node *rtpn;
5692 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5693 node_online(node) ? node : NUMA_NO_NODE);
5695 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5696 struct mem_cgroup_tree_per_zone *rtpz;
5698 rtpz = &rtpn->rb_tree_per_zone[zone];
5699 rtpz->rb_root = RB_ROOT;
5700 spin_lock_init(&rtpz->lock);
5702 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5707 subsys_initcall(mem_cgroup_init);
5709 #ifdef CONFIG_MEMCG_SWAP
5711 * mem_cgroup_swapout - transfer a memsw charge to swap
5712 * @page: page whose memsw charge to transfer
5713 * @entry: swap entry to move the charge to
5715 * Transfer the memsw charge of @page to @entry.
5717 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5719 struct mem_cgroup *memcg;
5720 unsigned short oldid;
5722 VM_BUG_ON_PAGE(PageLRU(page), page);
5723 VM_BUG_ON_PAGE(page_count(page), page);
5725 if (!do_memsw_account())
5728 memcg = page->mem_cgroup;
5730 /* Readahead page, never charged */
5734 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5735 VM_BUG_ON_PAGE(oldid, page);
5736 mem_cgroup_swap_statistics(memcg, true);
5738 page->mem_cgroup = NULL;
5740 if (!mem_cgroup_is_root(memcg))
5741 page_counter_uncharge(&memcg->memory, 1);
5744 * Interrupts should be disabled here because the caller holds the
5745 * mapping->tree_lock lock which is taken with interrupts-off. It is
5746 * important here to have the interrupts disabled because it is the
5747 * only synchronisation we have for udpating the per-CPU variables.
5749 VM_BUG_ON(!irqs_disabled());
5750 mem_cgroup_charge_statistics(memcg, page, false, -1);
5751 memcg_check_events(memcg, page);
5755 * mem_cgroup_try_charge_swap - try charging a swap entry
5756 * @page: page being added to swap
5757 * @entry: swap entry to charge
5759 * Try to charge @entry to the memcg that @page belongs to.
5761 * Returns 0 on success, -ENOMEM on failure.
5763 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5765 struct mem_cgroup *memcg;
5766 struct page_counter *counter;
5767 unsigned short oldid;
5769 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5772 memcg = page->mem_cgroup;
5774 /* Readahead page, never charged */
5778 if (!mem_cgroup_is_root(memcg) &&
5779 !page_counter_try_charge(&memcg->swap, 1, &counter))
5782 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5783 VM_BUG_ON_PAGE(oldid, page);
5784 mem_cgroup_swap_statistics(memcg, true);
5786 css_get(&memcg->css);
5791 * mem_cgroup_uncharge_swap - uncharge a swap entry
5792 * @entry: swap entry to uncharge
5794 * Drop the swap charge associated with @entry.
5796 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5798 struct mem_cgroup *memcg;
5801 if (!do_swap_account)
5804 id = swap_cgroup_record(entry, 0);
5806 memcg = mem_cgroup_from_id(id);
5808 if (!mem_cgroup_is_root(memcg)) {
5809 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5810 page_counter_uncharge(&memcg->swap, 1);
5812 page_counter_uncharge(&memcg->memsw, 1);
5814 mem_cgroup_swap_statistics(memcg, false);
5815 css_put(&memcg->css);
5820 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5822 long nr_swap_pages = get_nr_swap_pages();
5824 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5825 return nr_swap_pages;
5826 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5827 nr_swap_pages = min_t(long, nr_swap_pages,
5828 READ_ONCE(memcg->swap.limit) -
5829 page_counter_read(&memcg->swap));
5830 return nr_swap_pages;
5833 bool mem_cgroup_swap_full(struct page *page)
5835 struct mem_cgroup *memcg;
5837 VM_BUG_ON_PAGE(!PageLocked(page), page);
5841 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5844 memcg = page->mem_cgroup;
5848 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5849 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
5855 /* for remember boot option*/
5856 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5857 static int really_do_swap_account __initdata = 1;
5859 static int really_do_swap_account __initdata;
5862 static int __init enable_swap_account(char *s)
5864 if (!strcmp(s, "1"))
5865 really_do_swap_account = 1;
5866 else if (!strcmp(s, "0"))
5867 really_do_swap_account = 0;
5870 __setup("swapaccount=", enable_swap_account);
5872 static u64 swap_current_read(struct cgroup_subsys_state *css,
5875 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5877 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
5880 static int swap_max_show(struct seq_file *m, void *v)
5882 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5883 unsigned long max = READ_ONCE(memcg->swap.limit);
5885 if (max == PAGE_COUNTER_MAX)
5886 seq_puts(m, "max\n");
5888 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5893 static ssize_t swap_max_write(struct kernfs_open_file *of,
5894 char *buf, size_t nbytes, loff_t off)
5896 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5900 buf = strstrip(buf);
5901 err = page_counter_memparse(buf, "max", &max);
5905 mutex_lock(&memcg_limit_mutex);
5906 err = page_counter_limit(&memcg->swap, max);
5907 mutex_unlock(&memcg_limit_mutex);
5914 static struct cftype swap_files[] = {
5916 .name = "swap.current",
5917 .flags = CFTYPE_NOT_ON_ROOT,
5918 .read_u64 = swap_current_read,
5922 .flags = CFTYPE_NOT_ON_ROOT,
5923 .seq_show = swap_max_show,
5924 .write = swap_max_write,
5929 static struct cftype memsw_cgroup_files[] = {
5931 .name = "memsw.usage_in_bytes",
5932 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5933 .read_u64 = mem_cgroup_read_u64,
5936 .name = "memsw.max_usage_in_bytes",
5937 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5938 .write = mem_cgroup_reset,
5939 .read_u64 = mem_cgroup_read_u64,
5942 .name = "memsw.limit_in_bytes",
5943 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5944 .write = mem_cgroup_write,
5945 .read_u64 = mem_cgroup_read_u64,
5948 .name = "memsw.failcnt",
5949 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5950 .write = mem_cgroup_reset,
5951 .read_u64 = mem_cgroup_read_u64,
5953 { }, /* terminate */
5956 static int __init mem_cgroup_swap_init(void)
5958 if (!mem_cgroup_disabled() && really_do_swap_account) {
5959 do_swap_account = 1;
5960 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
5962 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5963 memsw_cgroup_files));
5967 subsys_initcall(mem_cgroup_swap_init);
5969 #endif /* CONFIG_MEMCG_SWAP */