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 mm_struct *mm;
211 struct mem_cgroup *from;
212 struct mem_cgroup *to;
214 unsigned long precharge;
215 unsigned long moved_charge;
216 unsigned long moved_swap;
217 struct task_struct *moving_task; /* a task moving charges */
218 wait_queue_head_t waitq; /* a waitq for other context */
220 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
221 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
225 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
226 * limit reclaim to prevent infinite loops, if they ever occur.
228 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
229 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
232 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
233 MEM_CGROUP_CHARGE_TYPE_ANON,
234 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
235 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
239 /* for encoding cft->private value on file */
248 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
249 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
250 #define MEMFILE_ATTR(val) ((val) & 0xffff)
251 /* Used for OOM nofiier */
252 #define OOM_CONTROL (0)
254 /* Some nice accessors for the vmpressure. */
255 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
258 memcg = root_mem_cgroup;
259 return &memcg->vmpressure;
262 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
264 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
267 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
269 return (memcg == root_mem_cgroup);
274 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
275 * The main reason for not using cgroup id for this:
276 * this works better in sparse environments, where we have a lot of memcgs,
277 * but only a few kmem-limited. Or also, if we have, for instance, 200
278 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
279 * 200 entry array for that.
281 * The current size of the caches array is stored in memcg_nr_cache_ids. It
282 * will double each time we have to increase it.
284 static DEFINE_IDA(memcg_cache_ida);
285 int memcg_nr_cache_ids;
287 /* Protects memcg_nr_cache_ids */
288 static DECLARE_RWSEM(memcg_cache_ids_sem);
290 void memcg_get_cache_ids(void)
292 down_read(&memcg_cache_ids_sem);
295 void memcg_put_cache_ids(void)
297 up_read(&memcg_cache_ids_sem);
301 * MIN_SIZE is different than 1, because we would like to avoid going through
302 * the alloc/free process all the time. In a small machine, 4 kmem-limited
303 * cgroups is a reasonable guess. In the future, it could be a parameter or
304 * tunable, but that is strictly not necessary.
306 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
307 * this constant directly from cgroup, but it is understandable that this is
308 * better kept as an internal representation in cgroup.c. In any case, the
309 * cgrp_id space is not getting any smaller, and we don't have to necessarily
310 * increase ours as well if it increases.
312 #define MEMCG_CACHES_MIN_SIZE 4
313 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
316 * A lot of the calls to the cache allocation functions are expected to be
317 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
318 * conditional to this static branch, we'll have to allow modules that does
319 * kmem_cache_alloc and the such to see this symbol as well
321 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
322 EXPORT_SYMBOL(memcg_kmem_enabled_key);
324 #endif /* !CONFIG_SLOB */
326 static struct mem_cgroup_per_zone *
327 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
329 int nid = zone_to_nid(zone);
330 int zid = zone_idx(zone);
332 return &memcg->nodeinfo[nid]->zoneinfo[zid];
336 * mem_cgroup_css_from_page - css of the memcg associated with a page
337 * @page: page of interest
339 * If memcg is bound to the default hierarchy, css of the memcg associated
340 * with @page is returned. The returned css remains associated with @page
341 * until it is released.
343 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
346 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
348 struct mem_cgroup *memcg;
350 memcg = page->mem_cgroup;
352 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
353 memcg = root_mem_cgroup;
359 * page_cgroup_ino - return inode number of the memcg a page is charged to
362 * Look up the closest online ancestor of the memory cgroup @page is charged to
363 * and return its inode number or 0 if @page is not charged to any cgroup. It
364 * is safe to call this function without holding a reference to @page.
366 * Note, this function is inherently racy, because there is nothing to prevent
367 * the cgroup inode from getting torn down and potentially reallocated a moment
368 * after page_cgroup_ino() returns, so it only should be used by callers that
369 * do not care (such as procfs interfaces).
371 ino_t page_cgroup_ino(struct page *page)
373 struct mem_cgroup *memcg;
374 unsigned long ino = 0;
377 memcg = READ_ONCE(page->mem_cgroup);
378 while (memcg && !(memcg->css.flags & CSS_ONLINE))
379 memcg = parent_mem_cgroup(memcg);
381 ino = cgroup_ino(memcg->css.cgroup);
386 static struct mem_cgroup_per_zone *
387 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
389 int nid = page_to_nid(page);
390 int zid = page_zonenum(page);
392 return &memcg->nodeinfo[nid]->zoneinfo[zid];
395 static struct mem_cgroup_tree_per_zone *
396 soft_limit_tree_node_zone(int nid, int zid)
398 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
401 static struct mem_cgroup_tree_per_zone *
402 soft_limit_tree_from_page(struct page *page)
404 int nid = page_to_nid(page);
405 int zid = page_zonenum(page);
407 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
410 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
411 struct mem_cgroup_tree_per_zone *mctz,
412 unsigned long new_usage_in_excess)
414 struct rb_node **p = &mctz->rb_root.rb_node;
415 struct rb_node *parent = NULL;
416 struct mem_cgroup_per_zone *mz_node;
421 mz->usage_in_excess = new_usage_in_excess;
422 if (!mz->usage_in_excess)
426 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
428 if (mz->usage_in_excess < mz_node->usage_in_excess)
431 * We can't avoid mem cgroups that are over their soft
432 * limit by the same amount
434 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
437 rb_link_node(&mz->tree_node, parent, p);
438 rb_insert_color(&mz->tree_node, &mctz->rb_root);
442 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
443 struct mem_cgroup_tree_per_zone *mctz)
447 rb_erase(&mz->tree_node, &mctz->rb_root);
451 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
452 struct mem_cgroup_tree_per_zone *mctz)
456 spin_lock_irqsave(&mctz->lock, flags);
457 __mem_cgroup_remove_exceeded(mz, mctz);
458 spin_unlock_irqrestore(&mctz->lock, flags);
461 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
463 unsigned long nr_pages = page_counter_read(&memcg->memory);
464 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
465 unsigned long excess = 0;
467 if (nr_pages > soft_limit)
468 excess = nr_pages - soft_limit;
473 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
475 unsigned long excess;
476 struct mem_cgroup_per_zone *mz;
477 struct mem_cgroup_tree_per_zone *mctz;
479 mctz = soft_limit_tree_from_page(page);
481 * Necessary to update all ancestors when hierarchy is used.
482 * because their event counter is not touched.
484 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
485 mz = mem_cgroup_page_zoneinfo(memcg, page);
486 excess = soft_limit_excess(memcg);
488 * We have to update the tree if mz is on RB-tree or
489 * mem is over its softlimit.
491 if (excess || mz->on_tree) {
494 spin_lock_irqsave(&mctz->lock, flags);
495 /* if on-tree, remove it */
497 __mem_cgroup_remove_exceeded(mz, mctz);
499 * Insert again. mz->usage_in_excess will be updated.
500 * If excess is 0, no tree ops.
502 __mem_cgroup_insert_exceeded(mz, mctz, excess);
503 spin_unlock_irqrestore(&mctz->lock, flags);
508 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
510 struct mem_cgroup_tree_per_zone *mctz;
511 struct mem_cgroup_per_zone *mz;
515 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
516 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
517 mctz = soft_limit_tree_node_zone(nid, zid);
518 mem_cgroup_remove_exceeded(mz, mctz);
523 static struct mem_cgroup_per_zone *
524 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
526 struct rb_node *rightmost = NULL;
527 struct mem_cgroup_per_zone *mz;
531 rightmost = rb_last(&mctz->rb_root);
533 goto done; /* Nothing to reclaim from */
535 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
537 * Remove the node now but someone else can add it back,
538 * we will to add it back at the end of reclaim to its correct
539 * position in the tree.
541 __mem_cgroup_remove_exceeded(mz, mctz);
542 if (!soft_limit_excess(mz->memcg) ||
543 !css_tryget_online(&mz->memcg->css))
549 static struct mem_cgroup_per_zone *
550 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
552 struct mem_cgroup_per_zone *mz;
554 spin_lock_irq(&mctz->lock);
555 mz = __mem_cgroup_largest_soft_limit_node(mctz);
556 spin_unlock_irq(&mctz->lock);
561 * Return page count for single (non recursive) @memcg.
563 * Implementation Note: reading percpu statistics for memcg.
565 * Both of vmstat[] and percpu_counter has threshold and do periodic
566 * synchronization to implement "quick" read. There are trade-off between
567 * reading cost and precision of value. Then, we may have a chance to implement
568 * a periodic synchronization of counter in memcg's counter.
570 * But this _read() function is used for user interface now. The user accounts
571 * memory usage by memory cgroup and he _always_ requires exact value because
572 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
573 * have to visit all online cpus and make sum. So, for now, unnecessary
574 * synchronization is not implemented. (just implemented for cpu hotplug)
576 * If there are kernel internal actions which can make use of some not-exact
577 * value, and reading all cpu value can be performance bottleneck in some
578 * common workload, threshold and synchronization as vmstat[] should be
582 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
587 /* Per-cpu values can be negative, use a signed accumulator */
588 for_each_possible_cpu(cpu)
589 val += per_cpu(memcg->stat->count[idx], cpu);
591 * Summing races with updates, so val may be negative. Avoid exposing
592 * transient negative values.
599 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
600 enum mem_cgroup_events_index idx)
602 unsigned long val = 0;
605 for_each_possible_cpu(cpu)
606 val += per_cpu(memcg->stat->events[idx], cpu);
610 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
612 bool compound, int nr_pages)
615 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
616 * counted as CACHE even if it's on ANON LRU.
619 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
622 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
626 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
627 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
631 /* pagein of a big page is an event. So, ignore page size */
633 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
635 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
636 nr_pages = -nr_pages; /* for event */
639 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
642 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
643 int nid, unsigned int lru_mask)
645 unsigned long nr = 0;
648 VM_BUG_ON((unsigned)nid >= nr_node_ids);
650 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
651 struct mem_cgroup_per_zone *mz;
655 if (!(BIT(lru) & lru_mask))
657 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
658 nr += mz->lru_size[lru];
664 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
665 unsigned int lru_mask)
667 unsigned long nr = 0;
670 for_each_node_state(nid, N_MEMORY)
671 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
675 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
676 enum mem_cgroup_events_target target)
678 unsigned long val, next;
680 val = __this_cpu_read(memcg->stat->nr_page_events);
681 next = __this_cpu_read(memcg->stat->targets[target]);
682 /* from time_after() in jiffies.h */
683 if ((long)next - (long)val < 0) {
685 case MEM_CGROUP_TARGET_THRESH:
686 next = val + THRESHOLDS_EVENTS_TARGET;
688 case MEM_CGROUP_TARGET_SOFTLIMIT:
689 next = val + SOFTLIMIT_EVENTS_TARGET;
691 case MEM_CGROUP_TARGET_NUMAINFO:
692 next = val + NUMAINFO_EVENTS_TARGET;
697 __this_cpu_write(memcg->stat->targets[target], next);
704 * Check events in order.
707 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
709 /* threshold event is triggered in finer grain than soft limit */
710 if (unlikely(mem_cgroup_event_ratelimit(memcg,
711 MEM_CGROUP_TARGET_THRESH))) {
713 bool do_numainfo __maybe_unused;
715 do_softlimit = mem_cgroup_event_ratelimit(memcg,
716 MEM_CGROUP_TARGET_SOFTLIMIT);
718 do_numainfo = mem_cgroup_event_ratelimit(memcg,
719 MEM_CGROUP_TARGET_NUMAINFO);
721 mem_cgroup_threshold(memcg);
722 if (unlikely(do_softlimit))
723 mem_cgroup_update_tree(memcg, page);
725 if (unlikely(do_numainfo))
726 atomic_inc(&memcg->numainfo_events);
731 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
734 * mm_update_next_owner() may clear mm->owner to NULL
735 * if it races with swapoff, page migration, etc.
736 * So this can be called with p == NULL.
741 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
743 EXPORT_SYMBOL(mem_cgroup_from_task);
745 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
747 struct mem_cgroup *memcg = NULL;
752 * Page cache insertions can happen withou an
753 * actual mm context, e.g. during disk probing
754 * on boot, loopback IO, acct() writes etc.
757 memcg = root_mem_cgroup;
759 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
760 if (unlikely(!memcg))
761 memcg = root_mem_cgroup;
763 } while (!css_tryget_online(&memcg->css));
769 * mem_cgroup_iter - iterate over memory cgroup hierarchy
770 * @root: hierarchy root
771 * @prev: previously returned memcg, NULL on first invocation
772 * @reclaim: cookie for shared reclaim walks, NULL for full walks
774 * Returns references to children of the hierarchy below @root, or
775 * @root itself, or %NULL after a full round-trip.
777 * Caller must pass the return value in @prev on subsequent
778 * invocations for reference counting, or use mem_cgroup_iter_break()
779 * to cancel a hierarchy walk before the round-trip is complete.
781 * Reclaimers can specify a zone and a priority level in @reclaim to
782 * divide up the memcgs in the hierarchy among all concurrent
783 * reclaimers operating on the same zone and priority.
785 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
786 struct mem_cgroup *prev,
787 struct mem_cgroup_reclaim_cookie *reclaim)
789 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
790 struct cgroup_subsys_state *css = NULL;
791 struct mem_cgroup *memcg = NULL;
792 struct mem_cgroup *pos = NULL;
794 if (mem_cgroup_disabled())
798 root = root_mem_cgroup;
800 if (prev && !reclaim)
803 if (!root->use_hierarchy && root != root_mem_cgroup) {
812 struct mem_cgroup_per_zone *mz;
814 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
815 iter = &mz->iter[reclaim->priority];
817 if (prev && reclaim->generation != iter->generation)
821 pos = READ_ONCE(iter->position);
822 if (!pos || css_tryget(&pos->css))
825 * css reference reached zero, so iter->position will
826 * be cleared by ->css_released. However, we should not
827 * rely on this happening soon, because ->css_released
828 * is called from a work queue, and by busy-waiting we
829 * might block it. So we clear iter->position right
832 (void)cmpxchg(&iter->position, pos, NULL);
840 css = css_next_descendant_pre(css, &root->css);
843 * Reclaimers share the hierarchy walk, and a
844 * new one might jump in right at the end of
845 * the hierarchy - make sure they see at least
846 * one group and restart from the beginning.
854 * Verify the css and acquire a reference. The root
855 * is provided by the caller, so we know it's alive
856 * and kicking, and don't take an extra reference.
858 memcg = mem_cgroup_from_css(css);
860 if (css == &root->css)
871 * The position could have already been updated by a competing
872 * thread, so check that the value hasn't changed since we read
873 * it to avoid reclaiming from the same cgroup twice.
875 (void)cmpxchg(&iter->position, pos, memcg);
883 reclaim->generation = iter->generation;
889 if (prev && prev != root)
896 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
897 * @root: hierarchy root
898 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
900 void mem_cgroup_iter_break(struct mem_cgroup *root,
901 struct mem_cgroup *prev)
904 root = root_mem_cgroup;
905 if (prev && prev != root)
909 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
911 struct mem_cgroup *memcg = dead_memcg;
912 struct mem_cgroup_reclaim_iter *iter;
913 struct mem_cgroup_per_zone *mz;
917 while ((memcg = parent_mem_cgroup(memcg))) {
919 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
920 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
921 for (i = 0; i <= DEF_PRIORITY; i++) {
923 cmpxchg(&iter->position,
932 * Iteration constructs for visiting all cgroups (under a tree). If
933 * loops are exited prematurely (break), mem_cgroup_iter_break() must
934 * be used for reference counting.
936 #define for_each_mem_cgroup_tree(iter, root) \
937 for (iter = mem_cgroup_iter(root, NULL, NULL); \
939 iter = mem_cgroup_iter(root, iter, NULL))
941 #define for_each_mem_cgroup(iter) \
942 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
944 iter = mem_cgroup_iter(NULL, iter, NULL))
947 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
948 * @zone: zone of the wanted lruvec
949 * @memcg: memcg of the wanted lruvec
951 * Returns the lru list vector holding pages for the given @zone and
952 * @mem. This can be the global zone lruvec, if the memory controller
955 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
956 struct mem_cgroup *memcg)
958 struct mem_cgroup_per_zone *mz;
959 struct lruvec *lruvec;
961 if (mem_cgroup_disabled()) {
962 lruvec = &zone->lruvec;
966 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
967 lruvec = &mz->lruvec;
970 * Since a node can be onlined after the mem_cgroup was created,
971 * we have to be prepared to initialize lruvec->zone here;
972 * and if offlined then reonlined, we need to reinitialize it.
974 if (unlikely(lruvec->zone != zone))
980 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
982 * @zone: zone of the page
984 * This function is only safe when following the LRU page isolation
985 * and putback protocol: the LRU lock must be held, and the page must
986 * either be PageLRU() or the caller must have isolated/allocated it.
988 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
990 struct mem_cgroup_per_zone *mz;
991 struct mem_cgroup *memcg;
992 struct lruvec *lruvec;
994 if (mem_cgroup_disabled()) {
995 lruvec = &zone->lruvec;
999 memcg = page->mem_cgroup;
1001 * Swapcache readahead pages are added to the LRU - and
1002 * possibly migrated - before they are charged.
1005 memcg = root_mem_cgroup;
1007 mz = mem_cgroup_page_zoneinfo(memcg, page);
1008 lruvec = &mz->lruvec;
1011 * Since a node can be onlined after the mem_cgroup was created,
1012 * we have to be prepared to initialize lruvec->zone here;
1013 * and if offlined then reonlined, we need to reinitialize it.
1015 if (unlikely(lruvec->zone != zone))
1016 lruvec->zone = zone;
1021 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1022 * @lruvec: mem_cgroup per zone lru vector
1023 * @lru: index of lru list the page is sitting on
1024 * @nr_pages: positive when adding or negative when removing
1026 * This function must be called under lru_lock, just before a page is added
1027 * to or just after a page is removed from an lru list (that ordering being
1028 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1030 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1033 struct mem_cgroup_per_zone *mz;
1034 unsigned long *lru_size;
1038 __update_lru_size(lruvec, lru, nr_pages);
1040 if (mem_cgroup_disabled())
1043 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1044 lru_size = mz->lru_size + lru;
1045 empty = list_empty(lruvec->lists + lru);
1048 *lru_size += nr_pages;
1051 if (WARN_ONCE(size < 0 || empty != !size,
1052 "%s(%p, %d, %d): lru_size %ld but %sempty\n",
1053 __func__, lruvec, lru, nr_pages, size, empty ? "" : "not ")) {
1059 *lru_size += nr_pages;
1062 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1064 struct mem_cgroup *task_memcg;
1065 struct task_struct *p;
1068 p = find_lock_task_mm(task);
1070 task_memcg = get_mem_cgroup_from_mm(p->mm);
1074 * All threads may have already detached their mm's, but the oom
1075 * killer still needs to detect if they have already been oom
1076 * killed to prevent needlessly killing additional tasks.
1079 task_memcg = mem_cgroup_from_task(task);
1080 css_get(&task_memcg->css);
1083 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1084 css_put(&task_memcg->css);
1089 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1090 * @memcg: the memory cgroup
1092 * Returns the maximum amount of memory @mem can be charged with, in
1095 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1097 unsigned long margin = 0;
1098 unsigned long count;
1099 unsigned long limit;
1101 count = page_counter_read(&memcg->memory);
1102 limit = READ_ONCE(memcg->memory.limit);
1104 margin = limit - count;
1106 if (do_memsw_account()) {
1107 count = page_counter_read(&memcg->memsw);
1108 limit = READ_ONCE(memcg->memsw.limit);
1110 margin = min(margin, limit - count);
1117 * A routine for checking "mem" is under move_account() or not.
1119 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1120 * moving cgroups. This is for waiting at high-memory pressure
1123 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1125 struct mem_cgroup *from;
1126 struct mem_cgroup *to;
1129 * Unlike task_move routines, we access mc.to, mc.from not under
1130 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1132 spin_lock(&mc.lock);
1138 ret = mem_cgroup_is_descendant(from, memcg) ||
1139 mem_cgroup_is_descendant(to, memcg);
1141 spin_unlock(&mc.lock);
1145 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1147 if (mc.moving_task && current != mc.moving_task) {
1148 if (mem_cgroup_under_move(memcg)) {
1150 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1151 /* moving charge context might have finished. */
1154 finish_wait(&mc.waitq, &wait);
1161 #define K(x) ((x) << (PAGE_SHIFT-10))
1163 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1164 * @memcg: The memory cgroup that went over limit
1165 * @p: Task that is going to be killed
1167 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1170 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1172 struct mem_cgroup *iter;
1178 pr_info("Task in ");
1179 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1180 pr_cont(" killed as a result of limit of ");
1182 pr_info("Memory limit reached of cgroup ");
1185 pr_cont_cgroup_path(memcg->css.cgroup);
1190 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1191 K((u64)page_counter_read(&memcg->memory)),
1192 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1193 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1194 K((u64)page_counter_read(&memcg->memsw)),
1195 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1196 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1197 K((u64)page_counter_read(&memcg->kmem)),
1198 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1200 for_each_mem_cgroup_tree(iter, memcg) {
1201 pr_info("Memory cgroup stats for ");
1202 pr_cont_cgroup_path(iter->css.cgroup);
1205 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1206 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1208 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1209 K(mem_cgroup_read_stat(iter, i)));
1212 for (i = 0; i < NR_LRU_LISTS; i++)
1213 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1214 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1221 * This function returns the number of memcg under hierarchy tree. Returns
1222 * 1(self count) if no children.
1224 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1227 struct mem_cgroup *iter;
1229 for_each_mem_cgroup_tree(iter, memcg)
1235 * Return the memory (and swap, if configured) limit for a memcg.
1237 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1239 unsigned long limit;
1241 limit = memcg->memory.limit;
1242 if (mem_cgroup_swappiness(memcg)) {
1243 unsigned long memsw_limit;
1244 unsigned long swap_limit;
1246 memsw_limit = memcg->memsw.limit;
1247 swap_limit = memcg->swap.limit;
1248 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1249 limit = min(limit + swap_limit, memsw_limit);
1254 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1257 struct oom_control oc = {
1260 .gfp_mask = gfp_mask,
1263 struct mem_cgroup *iter;
1264 unsigned long chosen_points = 0;
1265 unsigned long totalpages;
1266 unsigned int points = 0;
1267 struct task_struct *chosen = NULL;
1269 mutex_lock(&oom_lock);
1272 * If current has a pending SIGKILL or is exiting, then automatically
1273 * select it. The goal is to allow it to allocate so that it may
1274 * quickly exit and free its memory.
1276 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1277 mark_oom_victim(current);
1278 try_oom_reaper(current);
1282 check_panic_on_oom(&oc, CONSTRAINT_MEMCG, memcg);
1283 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1284 for_each_mem_cgroup_tree(iter, memcg) {
1285 struct css_task_iter it;
1286 struct task_struct *task;
1288 css_task_iter_start(&iter->css, &it);
1289 while ((task = css_task_iter_next(&it))) {
1290 switch (oom_scan_process_thread(&oc, task, totalpages)) {
1291 case OOM_SCAN_SELECT:
1293 put_task_struct(chosen);
1295 chosen_points = ULONG_MAX;
1296 get_task_struct(chosen);
1298 case OOM_SCAN_CONTINUE:
1300 case OOM_SCAN_ABORT:
1301 css_task_iter_end(&it);
1302 mem_cgroup_iter_break(memcg, iter);
1304 put_task_struct(chosen);
1305 /* Set a dummy value to return "true". */
1306 chosen = (void *) 1;
1311 points = oom_badness(task, memcg, NULL, totalpages);
1312 if (!points || points < chosen_points)
1314 /* Prefer thread group leaders for display purposes */
1315 if (points == chosen_points &&
1316 thread_group_leader(chosen))
1320 put_task_struct(chosen);
1322 chosen_points = points;
1323 get_task_struct(chosen);
1325 css_task_iter_end(&it);
1329 points = chosen_points * 1000 / totalpages;
1330 oom_kill_process(&oc, chosen, points, totalpages, memcg,
1331 "Memory cgroup out of memory");
1334 mutex_unlock(&oom_lock);
1338 #if MAX_NUMNODES > 1
1341 * test_mem_cgroup_node_reclaimable
1342 * @memcg: the target memcg
1343 * @nid: the node ID to be checked.
1344 * @noswap : specify true here if the user wants flle only information.
1346 * This function returns whether the specified memcg contains any
1347 * reclaimable pages on a node. Returns true if there are any reclaimable
1348 * pages in the node.
1350 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1351 int nid, bool noswap)
1353 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1355 if (noswap || !total_swap_pages)
1357 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1364 * Always updating the nodemask is not very good - even if we have an empty
1365 * list or the wrong list here, we can start from some node and traverse all
1366 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1369 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1373 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1374 * pagein/pageout changes since the last update.
1376 if (!atomic_read(&memcg->numainfo_events))
1378 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1381 /* make a nodemask where this memcg uses memory from */
1382 memcg->scan_nodes = node_states[N_MEMORY];
1384 for_each_node_mask(nid, node_states[N_MEMORY]) {
1386 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1387 node_clear(nid, memcg->scan_nodes);
1390 atomic_set(&memcg->numainfo_events, 0);
1391 atomic_set(&memcg->numainfo_updating, 0);
1395 * Selecting a node where we start reclaim from. Because what we need is just
1396 * reducing usage counter, start from anywhere is O,K. Considering
1397 * memory reclaim from current node, there are pros. and cons.
1399 * Freeing memory from current node means freeing memory from a node which
1400 * we'll use or we've used. So, it may make LRU bad. And if several threads
1401 * hit limits, it will see a contention on a node. But freeing from remote
1402 * node means more costs for memory reclaim because of memory latency.
1404 * Now, we use round-robin. Better algorithm is welcomed.
1406 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1410 mem_cgroup_may_update_nodemask(memcg);
1411 node = memcg->last_scanned_node;
1413 node = next_node_in(node, memcg->scan_nodes);
1415 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1416 * last time it really checked all the LRUs due to rate limiting.
1417 * Fallback to the current node in that case for simplicity.
1419 if (unlikely(node == MAX_NUMNODES))
1420 node = numa_node_id();
1422 memcg->last_scanned_node = node;
1426 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1432 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1435 unsigned long *total_scanned)
1437 struct mem_cgroup *victim = NULL;
1440 unsigned long excess;
1441 unsigned long nr_scanned;
1442 struct mem_cgroup_reclaim_cookie reclaim = {
1447 excess = soft_limit_excess(root_memcg);
1450 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1455 * If we have not been able to reclaim
1456 * anything, it might because there are
1457 * no reclaimable pages under this hierarchy
1462 * We want to do more targeted reclaim.
1463 * excess >> 2 is not to excessive so as to
1464 * reclaim too much, nor too less that we keep
1465 * coming back to reclaim from this cgroup
1467 if (total >= (excess >> 2) ||
1468 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1473 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1475 *total_scanned += nr_scanned;
1476 if (!soft_limit_excess(root_memcg))
1479 mem_cgroup_iter_break(root_memcg, victim);
1483 #ifdef CONFIG_LOCKDEP
1484 static struct lockdep_map memcg_oom_lock_dep_map = {
1485 .name = "memcg_oom_lock",
1489 static DEFINE_SPINLOCK(memcg_oom_lock);
1492 * Check OOM-Killer is already running under our hierarchy.
1493 * If someone is running, return false.
1495 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1497 struct mem_cgroup *iter, *failed = NULL;
1499 spin_lock(&memcg_oom_lock);
1501 for_each_mem_cgroup_tree(iter, memcg) {
1502 if (iter->oom_lock) {
1504 * this subtree of our hierarchy is already locked
1505 * so we cannot give a lock.
1508 mem_cgroup_iter_break(memcg, iter);
1511 iter->oom_lock = true;
1516 * OK, we failed to lock the whole subtree so we have
1517 * to clean up what we set up to the failing subtree
1519 for_each_mem_cgroup_tree(iter, memcg) {
1520 if (iter == failed) {
1521 mem_cgroup_iter_break(memcg, iter);
1524 iter->oom_lock = false;
1527 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1529 spin_unlock(&memcg_oom_lock);
1534 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1536 struct mem_cgroup *iter;
1538 spin_lock(&memcg_oom_lock);
1539 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1540 for_each_mem_cgroup_tree(iter, memcg)
1541 iter->oom_lock = false;
1542 spin_unlock(&memcg_oom_lock);
1545 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1547 struct mem_cgroup *iter;
1549 spin_lock(&memcg_oom_lock);
1550 for_each_mem_cgroup_tree(iter, memcg)
1552 spin_unlock(&memcg_oom_lock);
1555 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1557 struct mem_cgroup *iter;
1560 * When a new child is created while the hierarchy is under oom,
1561 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1563 spin_lock(&memcg_oom_lock);
1564 for_each_mem_cgroup_tree(iter, memcg)
1565 if (iter->under_oom > 0)
1567 spin_unlock(&memcg_oom_lock);
1570 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1572 struct oom_wait_info {
1573 struct mem_cgroup *memcg;
1577 static int memcg_oom_wake_function(wait_queue_t *wait,
1578 unsigned mode, int sync, void *arg)
1580 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1581 struct mem_cgroup *oom_wait_memcg;
1582 struct oom_wait_info *oom_wait_info;
1584 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1585 oom_wait_memcg = oom_wait_info->memcg;
1587 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1588 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1590 return autoremove_wake_function(wait, mode, sync, arg);
1593 static void memcg_oom_recover(struct mem_cgroup *memcg)
1596 * For the following lockless ->under_oom test, the only required
1597 * guarantee is that it must see the state asserted by an OOM when
1598 * this function is called as a result of userland actions
1599 * triggered by the notification of the OOM. This is trivially
1600 * achieved by invoking mem_cgroup_mark_under_oom() before
1601 * triggering notification.
1603 if (memcg && memcg->under_oom)
1604 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1607 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1609 if (!current->memcg_may_oom || current->memcg_in_oom)
1612 * We are in the middle of the charge context here, so we
1613 * don't want to block when potentially sitting on a callstack
1614 * that holds all kinds of filesystem and mm locks.
1616 * Also, the caller may handle a failed allocation gracefully
1617 * (like optional page cache readahead) and so an OOM killer
1618 * invocation might not even be necessary.
1620 * That's why we don't do anything here except remember the
1621 * OOM context and then deal with it at the end of the page
1622 * fault when the stack is unwound, the locks are released,
1623 * and when we know whether the fault was overall successful.
1625 css_get(&memcg->css);
1626 current->memcg_in_oom = memcg;
1627 current->memcg_oom_gfp_mask = mask;
1628 current->memcg_oom_order = order;
1632 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1633 * @handle: actually kill/wait or just clean up the OOM state
1635 * This has to be called at the end of a page fault if the memcg OOM
1636 * handler was enabled.
1638 * Memcg supports userspace OOM handling where failed allocations must
1639 * sleep on a waitqueue until the userspace task resolves the
1640 * situation. Sleeping directly in the charge context with all kinds
1641 * of locks held is not a good idea, instead we remember an OOM state
1642 * in the task and mem_cgroup_oom_synchronize() has to be called at
1643 * the end of the page fault to complete the OOM handling.
1645 * Returns %true if an ongoing memcg OOM situation was detected and
1646 * completed, %false otherwise.
1648 bool mem_cgroup_oom_synchronize(bool handle)
1650 struct mem_cgroup *memcg = current->memcg_in_oom;
1651 struct oom_wait_info owait;
1654 /* OOM is global, do not handle */
1658 if (!handle || oom_killer_disabled)
1661 owait.memcg = memcg;
1662 owait.wait.flags = 0;
1663 owait.wait.func = memcg_oom_wake_function;
1664 owait.wait.private = current;
1665 INIT_LIST_HEAD(&owait.wait.task_list);
1667 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1668 mem_cgroup_mark_under_oom(memcg);
1670 locked = mem_cgroup_oom_trylock(memcg);
1673 mem_cgroup_oom_notify(memcg);
1675 if (locked && !memcg->oom_kill_disable) {
1676 mem_cgroup_unmark_under_oom(memcg);
1677 finish_wait(&memcg_oom_waitq, &owait.wait);
1678 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1679 current->memcg_oom_order);
1682 mem_cgroup_unmark_under_oom(memcg);
1683 finish_wait(&memcg_oom_waitq, &owait.wait);
1687 mem_cgroup_oom_unlock(memcg);
1689 * There is no guarantee that an OOM-lock contender
1690 * sees the wakeups triggered by the OOM kill
1691 * uncharges. Wake any sleepers explicitely.
1693 memcg_oom_recover(memcg);
1696 current->memcg_in_oom = NULL;
1697 css_put(&memcg->css);
1702 * lock_page_memcg - lock a page->mem_cgroup binding
1705 * This function protects unlocked LRU pages from being moved to
1706 * another cgroup and stabilizes their page->mem_cgroup binding.
1708 void lock_page_memcg(struct page *page)
1710 struct mem_cgroup *memcg;
1711 unsigned long flags;
1714 * The RCU lock is held throughout the transaction. The fast
1715 * path can get away without acquiring the memcg->move_lock
1716 * because page moving starts with an RCU grace period.
1720 if (mem_cgroup_disabled())
1723 memcg = page->mem_cgroup;
1724 if (unlikely(!memcg))
1727 if (atomic_read(&memcg->moving_account) <= 0)
1730 spin_lock_irqsave(&memcg->move_lock, flags);
1731 if (memcg != page->mem_cgroup) {
1732 spin_unlock_irqrestore(&memcg->move_lock, flags);
1737 * When charge migration first begins, we can have locked and
1738 * unlocked page stat updates happening concurrently. Track
1739 * the task who has the lock for unlock_page_memcg().
1741 memcg->move_lock_task = current;
1742 memcg->move_lock_flags = flags;
1746 EXPORT_SYMBOL(lock_page_memcg);
1749 * unlock_page_memcg - unlock a page->mem_cgroup binding
1752 void unlock_page_memcg(struct page *page)
1754 struct mem_cgroup *memcg = page->mem_cgroup;
1756 if (memcg && memcg->move_lock_task == current) {
1757 unsigned long flags = memcg->move_lock_flags;
1759 memcg->move_lock_task = NULL;
1760 memcg->move_lock_flags = 0;
1762 spin_unlock_irqrestore(&memcg->move_lock, flags);
1767 EXPORT_SYMBOL(unlock_page_memcg);
1770 * size of first charge trial. "32" comes from vmscan.c's magic value.
1771 * TODO: maybe necessary to use big numbers in big irons.
1773 #define CHARGE_BATCH 32U
1774 struct memcg_stock_pcp {
1775 struct mem_cgroup *cached; /* this never be root cgroup */
1776 unsigned int nr_pages;
1777 struct work_struct work;
1778 unsigned long flags;
1779 #define FLUSHING_CACHED_CHARGE 0
1781 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1782 static DEFINE_MUTEX(percpu_charge_mutex);
1785 * consume_stock: Try to consume stocked charge on this cpu.
1786 * @memcg: memcg to consume from.
1787 * @nr_pages: how many pages to charge.
1789 * The charges will only happen if @memcg matches the current cpu's memcg
1790 * stock, and at least @nr_pages are available in that stock. Failure to
1791 * service an allocation will refill the stock.
1793 * returns true if successful, false otherwise.
1795 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1797 struct memcg_stock_pcp *stock;
1800 if (nr_pages > CHARGE_BATCH)
1803 stock = &get_cpu_var(memcg_stock);
1804 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1805 stock->nr_pages -= nr_pages;
1808 put_cpu_var(memcg_stock);
1813 * Returns stocks cached in percpu and reset cached information.
1815 static void drain_stock(struct memcg_stock_pcp *stock)
1817 struct mem_cgroup *old = stock->cached;
1819 if (stock->nr_pages) {
1820 page_counter_uncharge(&old->memory, stock->nr_pages);
1821 if (do_memsw_account())
1822 page_counter_uncharge(&old->memsw, stock->nr_pages);
1823 css_put_many(&old->css, stock->nr_pages);
1824 stock->nr_pages = 0;
1826 stock->cached = NULL;
1830 * This must be called under preempt disabled or must be called by
1831 * a thread which is pinned to local cpu.
1833 static void drain_local_stock(struct work_struct *dummy)
1835 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
1837 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1841 * Cache charges(val) to local per_cpu area.
1842 * This will be consumed by consume_stock() function, later.
1844 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1846 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1848 if (stock->cached != memcg) { /* reset if necessary */
1850 stock->cached = memcg;
1852 stock->nr_pages += nr_pages;
1853 put_cpu_var(memcg_stock);
1857 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1858 * of the hierarchy under it.
1860 static void drain_all_stock(struct mem_cgroup *root_memcg)
1864 /* If someone's already draining, avoid adding running more workers. */
1865 if (!mutex_trylock(&percpu_charge_mutex))
1867 /* Notify other cpus that system-wide "drain" is running */
1870 for_each_online_cpu(cpu) {
1871 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1872 struct mem_cgroup *memcg;
1874 memcg = stock->cached;
1875 if (!memcg || !stock->nr_pages)
1877 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1879 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1881 drain_local_stock(&stock->work);
1883 schedule_work_on(cpu, &stock->work);
1888 mutex_unlock(&percpu_charge_mutex);
1891 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1892 unsigned long action,
1895 int cpu = (unsigned long)hcpu;
1896 struct memcg_stock_pcp *stock;
1898 if (action == CPU_ONLINE)
1901 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1904 stock = &per_cpu(memcg_stock, cpu);
1909 static void reclaim_high(struct mem_cgroup *memcg,
1910 unsigned int nr_pages,
1914 if (page_counter_read(&memcg->memory) <= memcg->high)
1916 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1917 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1918 } while ((memcg = parent_mem_cgroup(memcg)));
1921 static void high_work_func(struct work_struct *work)
1923 struct mem_cgroup *memcg;
1925 memcg = container_of(work, struct mem_cgroup, high_work);
1926 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1930 * Scheduled by try_charge() to be executed from the userland return path
1931 * and reclaims memory over the high limit.
1933 void mem_cgroup_handle_over_high(void)
1935 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1936 struct mem_cgroup *memcg;
1938 if (likely(!nr_pages))
1941 memcg = get_mem_cgroup_from_mm(current->mm);
1942 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1943 css_put(&memcg->css);
1944 current->memcg_nr_pages_over_high = 0;
1947 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1948 unsigned int nr_pages)
1950 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1951 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1952 struct mem_cgroup *mem_over_limit;
1953 struct page_counter *counter;
1954 unsigned long nr_reclaimed;
1955 bool may_swap = true;
1956 bool drained = false;
1958 if (mem_cgroup_is_root(memcg))
1961 if (consume_stock(memcg, nr_pages))
1964 if (!do_memsw_account() ||
1965 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1966 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1968 if (do_memsw_account())
1969 page_counter_uncharge(&memcg->memsw, batch);
1970 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1972 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1976 if (batch > nr_pages) {
1982 * Unlike in global OOM situations, memcg is not in a physical
1983 * memory shortage. Allow dying and OOM-killed tasks to
1984 * bypass the last charges so that they can exit quickly and
1985 * free their memory.
1987 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1988 fatal_signal_pending(current) ||
1989 current->flags & PF_EXITING))
1992 if (unlikely(task_in_memcg_oom(current)))
1995 if (!gfpflags_allow_blocking(gfp_mask))
1998 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2000 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2001 gfp_mask, may_swap);
2003 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2007 drain_all_stock(mem_over_limit);
2012 if (gfp_mask & __GFP_NORETRY)
2015 * Even though the limit is exceeded at this point, reclaim
2016 * may have been able to free some pages. Retry the charge
2017 * before killing the task.
2019 * Only for regular pages, though: huge pages are rather
2020 * unlikely to succeed so close to the limit, and we fall back
2021 * to regular pages anyway in case of failure.
2023 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2026 * At task move, charge accounts can be doubly counted. So, it's
2027 * better to wait until the end of task_move if something is going on.
2029 if (mem_cgroup_wait_acct_move(mem_over_limit))
2035 if (gfp_mask & __GFP_NOFAIL)
2038 if (fatal_signal_pending(current))
2041 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2043 mem_cgroup_oom(mem_over_limit, gfp_mask,
2044 get_order(nr_pages * PAGE_SIZE));
2046 if (!(gfp_mask & __GFP_NOFAIL))
2050 * The allocation either can't fail or will lead to more memory
2051 * being freed very soon. Allow memory usage go over the limit
2052 * temporarily by force charging it.
2054 page_counter_charge(&memcg->memory, nr_pages);
2055 if (do_memsw_account())
2056 page_counter_charge(&memcg->memsw, nr_pages);
2057 css_get_many(&memcg->css, nr_pages);
2062 css_get_many(&memcg->css, batch);
2063 if (batch > nr_pages)
2064 refill_stock(memcg, batch - nr_pages);
2067 * If the hierarchy is above the normal consumption range, schedule
2068 * reclaim on returning to userland. We can perform reclaim here
2069 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2070 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2071 * not recorded as it most likely matches current's and won't
2072 * change in the meantime. As high limit is checked again before
2073 * reclaim, the cost of mismatch is negligible.
2076 if (page_counter_read(&memcg->memory) > memcg->high) {
2077 /* Don't bother a random interrupted task */
2078 if (in_interrupt()) {
2079 schedule_work(&memcg->high_work);
2082 current->memcg_nr_pages_over_high += batch;
2083 set_notify_resume(current);
2086 } while ((memcg = parent_mem_cgroup(memcg)));
2091 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2093 if (mem_cgroup_is_root(memcg))
2096 page_counter_uncharge(&memcg->memory, nr_pages);
2097 if (do_memsw_account())
2098 page_counter_uncharge(&memcg->memsw, nr_pages);
2100 css_put_many(&memcg->css, nr_pages);
2103 static void lock_page_lru(struct page *page, int *isolated)
2105 struct zone *zone = page_zone(page);
2107 spin_lock_irq(&zone->lru_lock);
2108 if (PageLRU(page)) {
2109 struct lruvec *lruvec;
2111 lruvec = mem_cgroup_page_lruvec(page, zone);
2113 del_page_from_lru_list(page, lruvec, page_lru(page));
2119 static void unlock_page_lru(struct page *page, int isolated)
2121 struct zone *zone = page_zone(page);
2124 struct lruvec *lruvec;
2126 lruvec = mem_cgroup_page_lruvec(page, zone);
2127 VM_BUG_ON_PAGE(PageLRU(page), page);
2129 add_page_to_lru_list(page, lruvec, page_lru(page));
2131 spin_unlock_irq(&zone->lru_lock);
2134 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2139 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2142 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2143 * may already be on some other mem_cgroup's LRU. Take care of it.
2146 lock_page_lru(page, &isolated);
2149 * Nobody should be changing or seriously looking at
2150 * page->mem_cgroup at this point:
2152 * - the page is uncharged
2154 * - the page is off-LRU
2156 * - an anonymous fault has exclusive page access, except for
2157 * a locked page table
2159 * - a page cache insertion, a swapin fault, or a migration
2160 * have the page locked
2162 page->mem_cgroup = memcg;
2165 unlock_page_lru(page, isolated);
2169 static int memcg_alloc_cache_id(void)
2174 id = ida_simple_get(&memcg_cache_ida,
2175 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2179 if (id < memcg_nr_cache_ids)
2183 * There's no space for the new id in memcg_caches arrays,
2184 * so we have to grow them.
2186 down_write(&memcg_cache_ids_sem);
2188 size = 2 * (id + 1);
2189 if (size < MEMCG_CACHES_MIN_SIZE)
2190 size = MEMCG_CACHES_MIN_SIZE;
2191 else if (size > MEMCG_CACHES_MAX_SIZE)
2192 size = MEMCG_CACHES_MAX_SIZE;
2194 err = memcg_update_all_caches(size);
2196 err = memcg_update_all_list_lrus(size);
2198 memcg_nr_cache_ids = size;
2200 up_write(&memcg_cache_ids_sem);
2203 ida_simple_remove(&memcg_cache_ida, id);
2209 static void memcg_free_cache_id(int id)
2211 ida_simple_remove(&memcg_cache_ida, id);
2214 struct memcg_kmem_cache_create_work {
2215 struct mem_cgroup *memcg;
2216 struct kmem_cache *cachep;
2217 struct work_struct work;
2220 static void memcg_kmem_cache_create_func(struct work_struct *w)
2222 struct memcg_kmem_cache_create_work *cw =
2223 container_of(w, struct memcg_kmem_cache_create_work, work);
2224 struct mem_cgroup *memcg = cw->memcg;
2225 struct kmem_cache *cachep = cw->cachep;
2227 memcg_create_kmem_cache(memcg, cachep);
2229 css_put(&memcg->css);
2234 * Enqueue the creation of a per-memcg kmem_cache.
2236 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2237 struct kmem_cache *cachep)
2239 struct memcg_kmem_cache_create_work *cw;
2241 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2245 css_get(&memcg->css);
2248 cw->cachep = cachep;
2249 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2251 schedule_work(&cw->work);
2254 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2255 struct kmem_cache *cachep)
2258 * We need to stop accounting when we kmalloc, because if the
2259 * corresponding kmalloc cache is not yet created, the first allocation
2260 * in __memcg_schedule_kmem_cache_create will recurse.
2262 * However, it is better to enclose the whole function. Depending on
2263 * the debugging options enabled, INIT_WORK(), for instance, can
2264 * trigger an allocation. This too, will make us recurse. Because at
2265 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2266 * the safest choice is to do it like this, wrapping the whole function.
2268 current->memcg_kmem_skip_account = 1;
2269 __memcg_schedule_kmem_cache_create(memcg, cachep);
2270 current->memcg_kmem_skip_account = 0;
2274 * Return the kmem_cache we're supposed to use for a slab allocation.
2275 * We try to use the current memcg's version of the cache.
2277 * If the cache does not exist yet, if we are the first user of it,
2278 * we either create it immediately, if possible, or create it asynchronously
2280 * In the latter case, we will let the current allocation go through with
2281 * the original cache.
2283 * Can't be called in interrupt context or from kernel threads.
2284 * This function needs to be called with rcu_read_lock() held.
2286 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep, gfp_t gfp)
2288 struct mem_cgroup *memcg;
2289 struct kmem_cache *memcg_cachep;
2292 VM_BUG_ON(!is_root_cache(cachep));
2294 if (cachep->flags & SLAB_ACCOUNT)
2295 gfp |= __GFP_ACCOUNT;
2297 if (!(gfp & __GFP_ACCOUNT))
2300 if (current->memcg_kmem_skip_account)
2303 memcg = get_mem_cgroup_from_mm(current->mm);
2304 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2308 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2309 if (likely(memcg_cachep))
2310 return memcg_cachep;
2313 * If we are in a safe context (can wait, and not in interrupt
2314 * context), we could be be predictable and return right away.
2315 * This would guarantee that the allocation being performed
2316 * already belongs in the new cache.
2318 * However, there are some clashes that can arrive from locking.
2319 * For instance, because we acquire the slab_mutex while doing
2320 * memcg_create_kmem_cache, this means no further allocation
2321 * could happen with the slab_mutex held. So it's better to
2324 memcg_schedule_kmem_cache_create(memcg, cachep);
2326 css_put(&memcg->css);
2330 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2332 if (!is_root_cache(cachep))
2333 css_put(&cachep->memcg_params.memcg->css);
2336 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2337 struct mem_cgroup *memcg)
2339 unsigned int nr_pages = 1 << order;
2340 struct page_counter *counter;
2343 ret = try_charge(memcg, gfp, nr_pages);
2347 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2348 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2349 cancel_charge(memcg, nr_pages);
2353 page->mem_cgroup = memcg;
2358 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2360 struct mem_cgroup *memcg;
2363 memcg = get_mem_cgroup_from_mm(current->mm);
2364 if (!mem_cgroup_is_root(memcg))
2365 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2366 css_put(&memcg->css);
2370 void __memcg_kmem_uncharge(struct page *page, int order)
2372 struct mem_cgroup *memcg = page->mem_cgroup;
2373 unsigned int nr_pages = 1 << order;
2378 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2380 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2381 page_counter_uncharge(&memcg->kmem, nr_pages);
2383 page_counter_uncharge(&memcg->memory, nr_pages);
2384 if (do_memsw_account())
2385 page_counter_uncharge(&memcg->memsw, nr_pages);
2387 page->mem_cgroup = NULL;
2388 css_put_many(&memcg->css, nr_pages);
2390 #endif /* !CONFIG_SLOB */
2392 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2395 * Because tail pages are not marked as "used", set it. We're under
2396 * zone->lru_lock and migration entries setup in all page mappings.
2398 void mem_cgroup_split_huge_fixup(struct page *head)
2402 if (mem_cgroup_disabled())
2405 for (i = 1; i < HPAGE_PMD_NR; i++)
2406 head[i].mem_cgroup = head->mem_cgroup;
2408 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2411 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2413 #ifdef CONFIG_MEMCG_SWAP
2414 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2417 int val = (charge) ? 1 : -1;
2418 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2422 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2423 * @entry: swap entry to be moved
2424 * @from: mem_cgroup which the entry is moved from
2425 * @to: mem_cgroup which the entry is moved to
2427 * It succeeds only when the swap_cgroup's record for this entry is the same
2428 * as the mem_cgroup's id of @from.
2430 * Returns 0 on success, -EINVAL on failure.
2432 * The caller must have charged to @to, IOW, called page_counter_charge() about
2433 * both res and memsw, and called css_get().
2435 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2436 struct mem_cgroup *from, struct mem_cgroup *to)
2438 unsigned short old_id, new_id;
2440 old_id = mem_cgroup_id(from);
2441 new_id = mem_cgroup_id(to);
2443 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2444 mem_cgroup_swap_statistics(from, false);
2445 mem_cgroup_swap_statistics(to, true);
2451 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2452 struct mem_cgroup *from, struct mem_cgroup *to)
2458 static DEFINE_MUTEX(memcg_limit_mutex);
2460 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2461 unsigned long limit)
2463 unsigned long curusage;
2464 unsigned long oldusage;
2465 bool enlarge = false;
2470 * For keeping hierarchical_reclaim simple, how long we should retry
2471 * is depends on callers. We set our retry-count to be function
2472 * of # of children which we should visit in this loop.
2474 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2475 mem_cgroup_count_children(memcg);
2477 oldusage = page_counter_read(&memcg->memory);
2480 if (signal_pending(current)) {
2485 mutex_lock(&memcg_limit_mutex);
2486 if (limit > memcg->memsw.limit) {
2487 mutex_unlock(&memcg_limit_mutex);
2491 if (limit > memcg->memory.limit)
2493 ret = page_counter_limit(&memcg->memory, limit);
2494 mutex_unlock(&memcg_limit_mutex);
2499 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2501 curusage = page_counter_read(&memcg->memory);
2502 /* Usage is reduced ? */
2503 if (curusage >= oldusage)
2506 oldusage = curusage;
2507 } while (retry_count);
2509 if (!ret && enlarge)
2510 memcg_oom_recover(memcg);
2515 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2516 unsigned long limit)
2518 unsigned long curusage;
2519 unsigned long oldusage;
2520 bool enlarge = false;
2524 /* see mem_cgroup_resize_res_limit */
2525 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2526 mem_cgroup_count_children(memcg);
2528 oldusage = page_counter_read(&memcg->memsw);
2531 if (signal_pending(current)) {
2536 mutex_lock(&memcg_limit_mutex);
2537 if (limit < memcg->memory.limit) {
2538 mutex_unlock(&memcg_limit_mutex);
2542 if (limit > memcg->memsw.limit)
2544 ret = page_counter_limit(&memcg->memsw, limit);
2545 mutex_unlock(&memcg_limit_mutex);
2550 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2552 curusage = page_counter_read(&memcg->memsw);
2553 /* Usage is reduced ? */
2554 if (curusage >= oldusage)
2557 oldusage = curusage;
2558 } while (retry_count);
2560 if (!ret && enlarge)
2561 memcg_oom_recover(memcg);
2566 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2568 unsigned long *total_scanned)
2570 unsigned long nr_reclaimed = 0;
2571 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2572 unsigned long reclaimed;
2574 struct mem_cgroup_tree_per_zone *mctz;
2575 unsigned long excess;
2576 unsigned long nr_scanned;
2581 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2583 * This loop can run a while, specially if mem_cgroup's continuously
2584 * keep exceeding their soft limit and putting the system under
2591 mz = mem_cgroup_largest_soft_limit_node(mctz);
2596 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2597 gfp_mask, &nr_scanned);
2598 nr_reclaimed += reclaimed;
2599 *total_scanned += nr_scanned;
2600 spin_lock_irq(&mctz->lock);
2601 __mem_cgroup_remove_exceeded(mz, mctz);
2604 * If we failed to reclaim anything from this memory cgroup
2605 * it is time to move on to the next cgroup
2609 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2611 excess = soft_limit_excess(mz->memcg);
2613 * One school of thought says that we should not add
2614 * back the node to the tree if reclaim returns 0.
2615 * But our reclaim could return 0, simply because due
2616 * to priority we are exposing a smaller subset of
2617 * memory to reclaim from. Consider this as a longer
2620 /* If excess == 0, no tree ops */
2621 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2622 spin_unlock_irq(&mctz->lock);
2623 css_put(&mz->memcg->css);
2626 * Could not reclaim anything and there are no more
2627 * mem cgroups to try or we seem to be looping without
2628 * reclaiming anything.
2630 if (!nr_reclaimed &&
2632 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2634 } while (!nr_reclaimed);
2636 css_put(&next_mz->memcg->css);
2637 return nr_reclaimed;
2641 * Test whether @memcg has children, dead or alive. Note that this
2642 * function doesn't care whether @memcg has use_hierarchy enabled and
2643 * returns %true if there are child csses according to the cgroup
2644 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2646 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2651 ret = css_next_child(NULL, &memcg->css);
2657 * Reclaims as many pages from the given memcg as possible.
2659 * Caller is responsible for holding css reference for memcg.
2661 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2663 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2665 /* we call try-to-free pages for make this cgroup empty */
2666 lru_add_drain_all();
2667 /* try to free all pages in this cgroup */
2668 while (nr_retries && page_counter_read(&memcg->memory)) {
2671 if (signal_pending(current))
2674 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2678 /* maybe some writeback is necessary */
2679 congestion_wait(BLK_RW_ASYNC, HZ/10);
2687 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2688 char *buf, size_t nbytes,
2691 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2693 if (mem_cgroup_is_root(memcg))
2695 return mem_cgroup_force_empty(memcg) ?: nbytes;
2698 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2701 return mem_cgroup_from_css(css)->use_hierarchy;
2704 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2705 struct cftype *cft, u64 val)
2708 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2709 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2711 if (memcg->use_hierarchy == val)
2715 * If parent's use_hierarchy is set, we can't make any modifications
2716 * in the child subtrees. If it is unset, then the change can
2717 * occur, provided the current cgroup has no children.
2719 * For the root cgroup, parent_mem is NULL, we allow value to be
2720 * set if there are no children.
2722 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2723 (val == 1 || val == 0)) {
2724 if (!memcg_has_children(memcg))
2725 memcg->use_hierarchy = val;
2734 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2736 struct mem_cgroup *iter;
2739 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2741 for_each_mem_cgroup_tree(iter, memcg) {
2742 for (i = 0; i < MEMCG_NR_STAT; i++)
2743 stat[i] += mem_cgroup_read_stat(iter, i);
2747 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2749 struct mem_cgroup *iter;
2752 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2754 for_each_mem_cgroup_tree(iter, memcg) {
2755 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2756 events[i] += mem_cgroup_read_events(iter, i);
2760 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2762 unsigned long val = 0;
2764 if (mem_cgroup_is_root(memcg)) {
2765 struct mem_cgroup *iter;
2767 for_each_mem_cgroup_tree(iter, memcg) {
2768 val += mem_cgroup_read_stat(iter,
2769 MEM_CGROUP_STAT_CACHE);
2770 val += mem_cgroup_read_stat(iter,
2771 MEM_CGROUP_STAT_RSS);
2773 val += mem_cgroup_read_stat(iter,
2774 MEM_CGROUP_STAT_SWAP);
2778 val = page_counter_read(&memcg->memory);
2780 val = page_counter_read(&memcg->memsw);
2793 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2796 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2797 struct page_counter *counter;
2799 switch (MEMFILE_TYPE(cft->private)) {
2801 counter = &memcg->memory;
2804 counter = &memcg->memsw;
2807 counter = &memcg->kmem;
2810 counter = &memcg->tcpmem;
2816 switch (MEMFILE_ATTR(cft->private)) {
2818 if (counter == &memcg->memory)
2819 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2820 if (counter == &memcg->memsw)
2821 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2822 return (u64)page_counter_read(counter) * PAGE_SIZE;
2824 return (u64)counter->limit * PAGE_SIZE;
2826 return (u64)counter->watermark * PAGE_SIZE;
2828 return counter->failcnt;
2829 case RES_SOFT_LIMIT:
2830 return (u64)memcg->soft_limit * PAGE_SIZE;
2837 static int memcg_online_kmem(struct mem_cgroup *memcg)
2841 if (cgroup_memory_nokmem)
2844 BUG_ON(memcg->kmemcg_id >= 0);
2845 BUG_ON(memcg->kmem_state);
2847 memcg_id = memcg_alloc_cache_id();
2851 static_branch_inc(&memcg_kmem_enabled_key);
2853 * A memory cgroup is considered kmem-online as soon as it gets
2854 * kmemcg_id. Setting the id after enabling static branching will
2855 * guarantee no one starts accounting before all call sites are
2858 memcg->kmemcg_id = memcg_id;
2859 memcg->kmem_state = KMEM_ONLINE;
2864 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2866 struct cgroup_subsys_state *css;
2867 struct mem_cgroup *parent, *child;
2870 if (memcg->kmem_state != KMEM_ONLINE)
2873 * Clear the online state before clearing memcg_caches array
2874 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2875 * guarantees that no cache will be created for this cgroup
2876 * after we are done (see memcg_create_kmem_cache()).
2878 memcg->kmem_state = KMEM_ALLOCATED;
2880 memcg_deactivate_kmem_caches(memcg);
2882 kmemcg_id = memcg->kmemcg_id;
2883 BUG_ON(kmemcg_id < 0);
2885 parent = parent_mem_cgroup(memcg);
2887 parent = root_mem_cgroup;
2890 * Change kmemcg_id of this cgroup and all its descendants to the
2891 * parent's id, and then move all entries from this cgroup's list_lrus
2892 * to ones of the parent. After we have finished, all list_lrus
2893 * corresponding to this cgroup are guaranteed to remain empty. The
2894 * ordering is imposed by list_lru_node->lock taken by
2895 * memcg_drain_all_list_lrus().
2897 css_for_each_descendant_pre(css, &memcg->css) {
2898 child = mem_cgroup_from_css(css);
2899 BUG_ON(child->kmemcg_id != kmemcg_id);
2900 child->kmemcg_id = parent->kmemcg_id;
2901 if (!memcg->use_hierarchy)
2904 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2906 memcg_free_cache_id(kmemcg_id);
2909 static void memcg_free_kmem(struct mem_cgroup *memcg)
2911 /* css_alloc() failed, offlining didn't happen */
2912 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2913 memcg_offline_kmem(memcg);
2915 if (memcg->kmem_state == KMEM_ALLOCATED) {
2916 memcg_destroy_kmem_caches(memcg);
2917 static_branch_dec(&memcg_kmem_enabled_key);
2918 WARN_ON(page_counter_read(&memcg->kmem));
2922 static int memcg_online_kmem(struct mem_cgroup *memcg)
2926 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2929 static void memcg_free_kmem(struct mem_cgroup *memcg)
2932 #endif /* !CONFIG_SLOB */
2934 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2935 unsigned long limit)
2939 mutex_lock(&memcg_limit_mutex);
2940 ret = page_counter_limit(&memcg->kmem, limit);
2941 mutex_unlock(&memcg_limit_mutex);
2945 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2949 mutex_lock(&memcg_limit_mutex);
2951 ret = page_counter_limit(&memcg->tcpmem, limit);
2955 if (!memcg->tcpmem_active) {
2957 * The active flag needs to be written after the static_key
2958 * update. This is what guarantees that the socket activation
2959 * function is the last one to run. See sock_update_memcg() for
2960 * details, and note that we don't mark any socket as belonging
2961 * to this memcg until that flag is up.
2963 * We need to do this, because static_keys will span multiple
2964 * sites, but we can't control their order. If we mark a socket
2965 * as accounted, but the accounting functions are not patched in
2966 * yet, we'll lose accounting.
2968 * We never race with the readers in sock_update_memcg(),
2969 * because when this value change, the code to process it is not
2972 static_branch_inc(&memcg_sockets_enabled_key);
2973 memcg->tcpmem_active = true;
2976 mutex_unlock(&memcg_limit_mutex);
2981 * The user of this function is...
2984 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2985 char *buf, size_t nbytes, loff_t off)
2987 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2988 unsigned long nr_pages;
2991 buf = strstrip(buf);
2992 ret = page_counter_memparse(buf, "-1", &nr_pages);
2996 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2998 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3002 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3004 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3007 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3010 ret = memcg_update_kmem_limit(memcg, nr_pages);
3013 ret = memcg_update_tcp_limit(memcg, nr_pages);
3017 case RES_SOFT_LIMIT:
3018 memcg->soft_limit = nr_pages;
3022 return ret ?: nbytes;
3025 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3026 size_t nbytes, loff_t off)
3028 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3029 struct page_counter *counter;
3031 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3033 counter = &memcg->memory;
3036 counter = &memcg->memsw;
3039 counter = &memcg->kmem;
3042 counter = &memcg->tcpmem;
3048 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3050 page_counter_reset_watermark(counter);
3053 counter->failcnt = 0;
3062 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3065 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3069 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3070 struct cftype *cft, u64 val)
3072 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3074 if (val & ~MOVE_MASK)
3078 * No kind of locking is needed in here, because ->can_attach() will
3079 * check this value once in the beginning of the process, and then carry
3080 * on with stale data. This means that changes to this value will only
3081 * affect task migrations starting after the change.
3083 memcg->move_charge_at_immigrate = val;
3087 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3088 struct cftype *cft, u64 val)
3095 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3099 unsigned int lru_mask;
3102 static const struct numa_stat stats[] = {
3103 { "total", LRU_ALL },
3104 { "file", LRU_ALL_FILE },
3105 { "anon", LRU_ALL_ANON },
3106 { "unevictable", BIT(LRU_UNEVICTABLE) },
3108 const struct numa_stat *stat;
3111 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3113 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3114 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3115 seq_printf(m, "%s=%lu", stat->name, nr);
3116 for_each_node_state(nid, N_MEMORY) {
3117 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3119 seq_printf(m, " N%d=%lu", nid, nr);
3124 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3125 struct mem_cgroup *iter;
3128 for_each_mem_cgroup_tree(iter, memcg)
3129 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3130 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3131 for_each_node_state(nid, N_MEMORY) {
3133 for_each_mem_cgroup_tree(iter, memcg)
3134 nr += mem_cgroup_node_nr_lru_pages(
3135 iter, nid, stat->lru_mask);
3136 seq_printf(m, " N%d=%lu", nid, nr);
3143 #endif /* CONFIG_NUMA */
3145 static int memcg_stat_show(struct seq_file *m, void *v)
3147 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3148 unsigned long memory, memsw;
3149 struct mem_cgroup *mi;
3152 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3153 MEM_CGROUP_STAT_NSTATS);
3154 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3155 MEM_CGROUP_EVENTS_NSTATS);
3156 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3158 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3159 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3161 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3162 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3165 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3166 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3167 mem_cgroup_read_events(memcg, i));
3169 for (i = 0; i < NR_LRU_LISTS; i++)
3170 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3171 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3173 /* Hierarchical information */
3174 memory = memsw = PAGE_COUNTER_MAX;
3175 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3176 memory = min(memory, mi->memory.limit);
3177 memsw = min(memsw, mi->memsw.limit);
3179 seq_printf(m, "hierarchical_memory_limit %llu\n",
3180 (u64)memory * PAGE_SIZE);
3181 if (do_memsw_account())
3182 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3183 (u64)memsw * PAGE_SIZE);
3185 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3186 unsigned long long val = 0;
3188 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3190 for_each_mem_cgroup_tree(mi, memcg)
3191 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3192 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3195 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3196 unsigned long long val = 0;
3198 for_each_mem_cgroup_tree(mi, memcg)
3199 val += mem_cgroup_read_events(mi, i);
3200 seq_printf(m, "total_%s %llu\n",
3201 mem_cgroup_events_names[i], val);
3204 for (i = 0; i < NR_LRU_LISTS; i++) {
3205 unsigned long long val = 0;
3207 for_each_mem_cgroup_tree(mi, memcg)
3208 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3209 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3212 #ifdef CONFIG_DEBUG_VM
3215 struct mem_cgroup_per_zone *mz;
3216 struct zone_reclaim_stat *rstat;
3217 unsigned long recent_rotated[2] = {0, 0};
3218 unsigned long recent_scanned[2] = {0, 0};
3220 for_each_online_node(nid)
3221 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3222 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3223 rstat = &mz->lruvec.reclaim_stat;
3225 recent_rotated[0] += rstat->recent_rotated[0];
3226 recent_rotated[1] += rstat->recent_rotated[1];
3227 recent_scanned[0] += rstat->recent_scanned[0];
3228 recent_scanned[1] += rstat->recent_scanned[1];
3230 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3231 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3232 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3233 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3240 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3243 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3245 return mem_cgroup_swappiness(memcg);
3248 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3249 struct cftype *cft, u64 val)
3251 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3257 memcg->swappiness = val;
3259 vm_swappiness = val;
3264 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3266 struct mem_cgroup_threshold_ary *t;
3267 unsigned long usage;
3272 t = rcu_dereference(memcg->thresholds.primary);
3274 t = rcu_dereference(memcg->memsw_thresholds.primary);
3279 usage = mem_cgroup_usage(memcg, swap);
3282 * current_threshold points to threshold just below or equal to usage.
3283 * If it's not true, a threshold was crossed after last
3284 * call of __mem_cgroup_threshold().
3286 i = t->current_threshold;
3289 * Iterate backward over array of thresholds starting from
3290 * current_threshold and check if a threshold is crossed.
3291 * If none of thresholds below usage is crossed, we read
3292 * only one element of the array here.
3294 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3295 eventfd_signal(t->entries[i].eventfd, 1);
3297 /* i = current_threshold + 1 */
3301 * Iterate forward over array of thresholds starting from
3302 * current_threshold+1 and check if a threshold is crossed.
3303 * If none of thresholds above usage is crossed, we read
3304 * only one element of the array here.
3306 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3307 eventfd_signal(t->entries[i].eventfd, 1);
3309 /* Update current_threshold */
3310 t->current_threshold = i - 1;
3315 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3318 __mem_cgroup_threshold(memcg, false);
3319 if (do_memsw_account())
3320 __mem_cgroup_threshold(memcg, true);
3322 memcg = parent_mem_cgroup(memcg);
3326 static int compare_thresholds(const void *a, const void *b)
3328 const struct mem_cgroup_threshold *_a = a;
3329 const struct mem_cgroup_threshold *_b = b;
3331 if (_a->threshold > _b->threshold)
3334 if (_a->threshold < _b->threshold)
3340 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3342 struct mem_cgroup_eventfd_list *ev;
3344 spin_lock(&memcg_oom_lock);
3346 list_for_each_entry(ev, &memcg->oom_notify, list)
3347 eventfd_signal(ev->eventfd, 1);
3349 spin_unlock(&memcg_oom_lock);
3353 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3355 struct mem_cgroup *iter;
3357 for_each_mem_cgroup_tree(iter, memcg)
3358 mem_cgroup_oom_notify_cb(iter);
3361 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3362 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3364 struct mem_cgroup_thresholds *thresholds;
3365 struct mem_cgroup_threshold_ary *new;
3366 unsigned long threshold;
3367 unsigned long usage;
3370 ret = page_counter_memparse(args, "-1", &threshold);
3374 mutex_lock(&memcg->thresholds_lock);
3377 thresholds = &memcg->thresholds;
3378 usage = mem_cgroup_usage(memcg, false);
3379 } else if (type == _MEMSWAP) {
3380 thresholds = &memcg->memsw_thresholds;
3381 usage = mem_cgroup_usage(memcg, true);
3385 /* Check if a threshold crossed before adding a new one */
3386 if (thresholds->primary)
3387 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3389 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3391 /* Allocate memory for new array of thresholds */
3392 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3400 /* Copy thresholds (if any) to new array */
3401 if (thresholds->primary) {
3402 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3403 sizeof(struct mem_cgroup_threshold));
3406 /* Add new threshold */
3407 new->entries[size - 1].eventfd = eventfd;
3408 new->entries[size - 1].threshold = threshold;
3410 /* Sort thresholds. Registering of new threshold isn't time-critical */
3411 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3412 compare_thresholds, NULL);
3414 /* Find current threshold */
3415 new->current_threshold = -1;
3416 for (i = 0; i < size; i++) {
3417 if (new->entries[i].threshold <= usage) {
3419 * new->current_threshold will not be used until
3420 * rcu_assign_pointer(), so it's safe to increment
3423 ++new->current_threshold;
3428 /* Free old spare buffer and save old primary buffer as spare */
3429 kfree(thresholds->spare);
3430 thresholds->spare = thresholds->primary;
3432 rcu_assign_pointer(thresholds->primary, new);
3434 /* To be sure that nobody uses thresholds */
3438 mutex_unlock(&memcg->thresholds_lock);
3443 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3444 struct eventfd_ctx *eventfd, const char *args)
3446 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3449 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3450 struct eventfd_ctx *eventfd, const char *args)
3452 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3455 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3456 struct eventfd_ctx *eventfd, enum res_type type)
3458 struct mem_cgroup_thresholds *thresholds;
3459 struct mem_cgroup_threshold_ary *new;
3460 unsigned long usage;
3463 mutex_lock(&memcg->thresholds_lock);
3466 thresholds = &memcg->thresholds;
3467 usage = mem_cgroup_usage(memcg, false);
3468 } else if (type == _MEMSWAP) {
3469 thresholds = &memcg->memsw_thresholds;
3470 usage = mem_cgroup_usage(memcg, true);
3474 if (!thresholds->primary)
3477 /* Check if a threshold crossed before removing */
3478 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3480 /* Calculate new number of threshold */
3482 for (i = 0; i < thresholds->primary->size; i++) {
3483 if (thresholds->primary->entries[i].eventfd != eventfd)
3487 new = thresholds->spare;
3489 /* Set thresholds array to NULL if we don't have thresholds */
3498 /* Copy thresholds and find current threshold */
3499 new->current_threshold = -1;
3500 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3501 if (thresholds->primary->entries[i].eventfd == eventfd)
3504 new->entries[j] = thresholds->primary->entries[i];
3505 if (new->entries[j].threshold <= usage) {
3507 * new->current_threshold will not be used
3508 * until rcu_assign_pointer(), so it's safe to increment
3511 ++new->current_threshold;
3517 /* Swap primary and spare array */
3518 thresholds->spare = thresholds->primary;
3520 rcu_assign_pointer(thresholds->primary, new);
3522 /* To be sure that nobody uses thresholds */
3525 /* If all events are unregistered, free the spare array */
3527 kfree(thresholds->spare);
3528 thresholds->spare = NULL;
3531 mutex_unlock(&memcg->thresholds_lock);
3534 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3535 struct eventfd_ctx *eventfd)
3537 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3540 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3541 struct eventfd_ctx *eventfd)
3543 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3546 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3547 struct eventfd_ctx *eventfd, const char *args)
3549 struct mem_cgroup_eventfd_list *event;
3551 event = kmalloc(sizeof(*event), GFP_KERNEL);
3555 spin_lock(&memcg_oom_lock);
3557 event->eventfd = eventfd;
3558 list_add(&event->list, &memcg->oom_notify);
3560 /* already in OOM ? */
3561 if (memcg->under_oom)
3562 eventfd_signal(eventfd, 1);
3563 spin_unlock(&memcg_oom_lock);
3568 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3569 struct eventfd_ctx *eventfd)
3571 struct mem_cgroup_eventfd_list *ev, *tmp;
3573 spin_lock(&memcg_oom_lock);
3575 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3576 if (ev->eventfd == eventfd) {
3577 list_del(&ev->list);
3582 spin_unlock(&memcg_oom_lock);
3585 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3587 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3589 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3590 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3594 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3595 struct cftype *cft, u64 val)
3597 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3599 /* cannot set to root cgroup and only 0 and 1 are allowed */
3600 if (!css->parent || !((val == 0) || (val == 1)))
3603 memcg->oom_kill_disable = val;
3605 memcg_oom_recover(memcg);
3610 #ifdef CONFIG_CGROUP_WRITEBACK
3612 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3614 return &memcg->cgwb_list;
3617 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3619 return wb_domain_init(&memcg->cgwb_domain, gfp);
3622 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3624 wb_domain_exit(&memcg->cgwb_domain);
3627 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3629 wb_domain_size_changed(&memcg->cgwb_domain);
3632 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3634 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3636 if (!memcg->css.parent)
3639 return &memcg->cgwb_domain;
3643 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3644 * @wb: bdi_writeback in question
3645 * @pfilepages: out parameter for number of file pages
3646 * @pheadroom: out parameter for number of allocatable pages according to memcg
3647 * @pdirty: out parameter for number of dirty pages
3648 * @pwriteback: out parameter for number of pages under writeback
3650 * Determine the numbers of file, headroom, dirty, and writeback pages in
3651 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3652 * is a bit more involved.
3654 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3655 * headroom is calculated as the lowest headroom of itself and the
3656 * ancestors. Note that this doesn't consider the actual amount of
3657 * available memory in the system. The caller should further cap
3658 * *@pheadroom accordingly.
3660 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3661 unsigned long *pheadroom, unsigned long *pdirty,
3662 unsigned long *pwriteback)
3664 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3665 struct mem_cgroup *parent;
3667 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3669 /* this should eventually include NR_UNSTABLE_NFS */
3670 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3671 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3672 (1 << LRU_ACTIVE_FILE));
3673 *pheadroom = PAGE_COUNTER_MAX;
3675 while ((parent = parent_mem_cgroup(memcg))) {
3676 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3677 unsigned long used = page_counter_read(&memcg->memory);
3679 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3684 #else /* CONFIG_CGROUP_WRITEBACK */
3686 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3691 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3695 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3699 #endif /* CONFIG_CGROUP_WRITEBACK */
3702 * DO NOT USE IN NEW FILES.
3704 * "cgroup.event_control" implementation.
3706 * This is way over-engineered. It tries to support fully configurable
3707 * events for each user. Such level of flexibility is completely
3708 * unnecessary especially in the light of the planned unified hierarchy.
3710 * Please deprecate this and replace with something simpler if at all
3715 * Unregister event and free resources.
3717 * Gets called from workqueue.
3719 static void memcg_event_remove(struct work_struct *work)
3721 struct mem_cgroup_event *event =
3722 container_of(work, struct mem_cgroup_event, remove);
3723 struct mem_cgroup *memcg = event->memcg;
3725 remove_wait_queue(event->wqh, &event->wait);
3727 event->unregister_event(memcg, event->eventfd);
3729 /* Notify userspace the event is going away. */
3730 eventfd_signal(event->eventfd, 1);
3732 eventfd_ctx_put(event->eventfd);
3734 css_put(&memcg->css);
3738 * Gets called on POLLHUP on eventfd when user closes it.
3740 * Called with wqh->lock held and interrupts disabled.
3742 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3743 int sync, void *key)
3745 struct mem_cgroup_event *event =
3746 container_of(wait, struct mem_cgroup_event, wait);
3747 struct mem_cgroup *memcg = event->memcg;
3748 unsigned long flags = (unsigned long)key;
3750 if (flags & POLLHUP) {
3752 * If the event has been detached at cgroup removal, we
3753 * can simply return knowing the other side will cleanup
3756 * We can't race against event freeing since the other
3757 * side will require wqh->lock via remove_wait_queue(),
3760 spin_lock(&memcg->event_list_lock);
3761 if (!list_empty(&event->list)) {
3762 list_del_init(&event->list);
3764 * We are in atomic context, but cgroup_event_remove()
3765 * may sleep, so we have to call it in workqueue.
3767 schedule_work(&event->remove);
3769 spin_unlock(&memcg->event_list_lock);
3775 static void memcg_event_ptable_queue_proc(struct file *file,
3776 wait_queue_head_t *wqh, poll_table *pt)
3778 struct mem_cgroup_event *event =
3779 container_of(pt, struct mem_cgroup_event, pt);
3782 add_wait_queue(wqh, &event->wait);
3786 * DO NOT USE IN NEW FILES.
3788 * Parse input and register new cgroup event handler.
3790 * Input must be in format '<event_fd> <control_fd> <args>'.
3791 * Interpretation of args is defined by control file implementation.
3793 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3794 char *buf, size_t nbytes, loff_t off)
3796 struct cgroup_subsys_state *css = of_css(of);
3797 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3798 struct mem_cgroup_event *event;
3799 struct cgroup_subsys_state *cfile_css;
3800 unsigned int efd, cfd;
3807 buf = strstrip(buf);
3809 efd = simple_strtoul(buf, &endp, 10);
3814 cfd = simple_strtoul(buf, &endp, 10);
3815 if ((*endp != ' ') && (*endp != '\0'))
3819 event = kzalloc(sizeof(*event), GFP_KERNEL);
3823 event->memcg = memcg;
3824 INIT_LIST_HEAD(&event->list);
3825 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3826 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3827 INIT_WORK(&event->remove, memcg_event_remove);
3835 event->eventfd = eventfd_ctx_fileget(efile.file);
3836 if (IS_ERR(event->eventfd)) {
3837 ret = PTR_ERR(event->eventfd);
3844 goto out_put_eventfd;
3847 /* the process need read permission on control file */
3848 /* AV: shouldn't we check that it's been opened for read instead? */
3849 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3854 * Determine the event callbacks and set them in @event. This used
3855 * to be done via struct cftype but cgroup core no longer knows
3856 * about these events. The following is crude but the whole thing
3857 * is for compatibility anyway.
3859 * DO NOT ADD NEW FILES.
3861 name = cfile.file->f_path.dentry->d_name.name;
3863 if (!strcmp(name, "memory.usage_in_bytes")) {
3864 event->register_event = mem_cgroup_usage_register_event;
3865 event->unregister_event = mem_cgroup_usage_unregister_event;
3866 } else if (!strcmp(name, "memory.oom_control")) {
3867 event->register_event = mem_cgroup_oom_register_event;
3868 event->unregister_event = mem_cgroup_oom_unregister_event;
3869 } else if (!strcmp(name, "memory.pressure_level")) {
3870 event->register_event = vmpressure_register_event;
3871 event->unregister_event = vmpressure_unregister_event;
3872 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3873 event->register_event = memsw_cgroup_usage_register_event;
3874 event->unregister_event = memsw_cgroup_usage_unregister_event;
3881 * Verify @cfile should belong to @css. Also, remaining events are
3882 * automatically removed on cgroup destruction but the removal is
3883 * asynchronous, so take an extra ref on @css.
3885 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3886 &memory_cgrp_subsys);
3888 if (IS_ERR(cfile_css))
3890 if (cfile_css != css) {
3895 ret = event->register_event(memcg, event->eventfd, buf);
3899 efile.file->f_op->poll(efile.file, &event->pt);
3901 spin_lock(&memcg->event_list_lock);
3902 list_add(&event->list, &memcg->event_list);
3903 spin_unlock(&memcg->event_list_lock);
3915 eventfd_ctx_put(event->eventfd);
3924 static struct cftype mem_cgroup_legacy_files[] = {
3926 .name = "usage_in_bytes",
3927 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3928 .read_u64 = mem_cgroup_read_u64,
3931 .name = "max_usage_in_bytes",
3932 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3933 .write = mem_cgroup_reset,
3934 .read_u64 = mem_cgroup_read_u64,
3937 .name = "limit_in_bytes",
3938 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3939 .write = mem_cgroup_write,
3940 .read_u64 = mem_cgroup_read_u64,
3943 .name = "soft_limit_in_bytes",
3944 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3945 .write = mem_cgroup_write,
3946 .read_u64 = mem_cgroup_read_u64,
3950 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3951 .write = mem_cgroup_reset,
3952 .read_u64 = mem_cgroup_read_u64,
3956 .seq_show = memcg_stat_show,
3959 .name = "force_empty",
3960 .write = mem_cgroup_force_empty_write,
3963 .name = "use_hierarchy",
3964 .write_u64 = mem_cgroup_hierarchy_write,
3965 .read_u64 = mem_cgroup_hierarchy_read,
3968 .name = "cgroup.event_control", /* XXX: for compat */
3969 .write = memcg_write_event_control,
3970 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3973 .name = "swappiness",
3974 .read_u64 = mem_cgroup_swappiness_read,
3975 .write_u64 = mem_cgroup_swappiness_write,
3978 .name = "move_charge_at_immigrate",
3979 .read_u64 = mem_cgroup_move_charge_read,
3980 .write_u64 = mem_cgroup_move_charge_write,
3983 .name = "oom_control",
3984 .seq_show = mem_cgroup_oom_control_read,
3985 .write_u64 = mem_cgroup_oom_control_write,
3986 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3989 .name = "pressure_level",
3993 .name = "numa_stat",
3994 .seq_show = memcg_numa_stat_show,
3998 .name = "kmem.limit_in_bytes",
3999 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4000 .write = mem_cgroup_write,
4001 .read_u64 = mem_cgroup_read_u64,
4004 .name = "kmem.usage_in_bytes",
4005 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4006 .read_u64 = mem_cgroup_read_u64,
4009 .name = "kmem.failcnt",
4010 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4011 .write = mem_cgroup_reset,
4012 .read_u64 = mem_cgroup_read_u64,
4015 .name = "kmem.max_usage_in_bytes",
4016 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4017 .write = mem_cgroup_reset,
4018 .read_u64 = mem_cgroup_read_u64,
4020 #ifdef CONFIG_SLABINFO
4022 .name = "kmem.slabinfo",
4023 .seq_start = slab_start,
4024 .seq_next = slab_next,
4025 .seq_stop = slab_stop,
4026 .seq_show = memcg_slab_show,
4030 .name = "kmem.tcp.limit_in_bytes",
4031 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4032 .write = mem_cgroup_write,
4033 .read_u64 = mem_cgroup_read_u64,
4036 .name = "kmem.tcp.usage_in_bytes",
4037 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4038 .read_u64 = mem_cgroup_read_u64,
4041 .name = "kmem.tcp.failcnt",
4042 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4043 .write = mem_cgroup_reset,
4044 .read_u64 = mem_cgroup_read_u64,
4047 .name = "kmem.tcp.max_usage_in_bytes",
4048 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4049 .write = mem_cgroup_reset,
4050 .read_u64 = mem_cgroup_read_u64,
4052 { }, /* terminate */
4055 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4057 struct mem_cgroup_per_node *pn;
4058 struct mem_cgroup_per_zone *mz;
4059 int zone, tmp = node;
4061 * This routine is called against possible nodes.
4062 * But it's BUG to call kmalloc() against offline node.
4064 * TODO: this routine can waste much memory for nodes which will
4065 * never be onlined. It's better to use memory hotplug callback
4068 if (!node_state(node, N_NORMAL_MEMORY))
4070 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4074 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4075 mz = &pn->zoneinfo[zone];
4076 lruvec_init(&mz->lruvec);
4077 mz->usage_in_excess = 0;
4078 mz->on_tree = false;
4081 memcg->nodeinfo[node] = pn;
4085 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4087 kfree(memcg->nodeinfo[node]);
4090 static void mem_cgroup_free(struct mem_cgroup *memcg)
4094 memcg_wb_domain_exit(memcg);
4096 free_mem_cgroup_per_zone_info(memcg, node);
4097 free_percpu(memcg->stat);
4101 static struct mem_cgroup *mem_cgroup_alloc(void)
4103 struct mem_cgroup *memcg;
4107 size = sizeof(struct mem_cgroup);
4108 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4110 memcg = kzalloc(size, GFP_KERNEL);
4114 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4119 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4122 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4125 INIT_WORK(&memcg->high_work, high_work_func);
4126 memcg->last_scanned_node = MAX_NUMNODES;
4127 INIT_LIST_HEAD(&memcg->oom_notify);
4128 mutex_init(&memcg->thresholds_lock);
4129 spin_lock_init(&memcg->move_lock);
4130 vmpressure_init(&memcg->vmpressure);
4131 INIT_LIST_HEAD(&memcg->event_list);
4132 spin_lock_init(&memcg->event_list_lock);
4133 memcg->socket_pressure = jiffies;
4135 memcg->kmemcg_id = -1;
4137 #ifdef CONFIG_CGROUP_WRITEBACK
4138 INIT_LIST_HEAD(&memcg->cgwb_list);
4142 mem_cgroup_free(memcg);
4146 static struct cgroup_subsys_state * __ref
4147 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4149 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4150 struct mem_cgroup *memcg;
4151 long error = -ENOMEM;
4153 memcg = mem_cgroup_alloc();
4155 return ERR_PTR(error);
4157 memcg->high = PAGE_COUNTER_MAX;
4158 memcg->soft_limit = PAGE_COUNTER_MAX;
4160 memcg->swappiness = mem_cgroup_swappiness(parent);
4161 memcg->oom_kill_disable = parent->oom_kill_disable;
4163 if (parent && parent->use_hierarchy) {
4164 memcg->use_hierarchy = true;
4165 page_counter_init(&memcg->memory, &parent->memory);
4166 page_counter_init(&memcg->swap, &parent->swap);
4167 page_counter_init(&memcg->memsw, &parent->memsw);
4168 page_counter_init(&memcg->kmem, &parent->kmem);
4169 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4171 page_counter_init(&memcg->memory, NULL);
4172 page_counter_init(&memcg->swap, NULL);
4173 page_counter_init(&memcg->memsw, NULL);
4174 page_counter_init(&memcg->kmem, NULL);
4175 page_counter_init(&memcg->tcpmem, NULL);
4177 * Deeper hierachy with use_hierarchy == false doesn't make
4178 * much sense so let cgroup subsystem know about this
4179 * unfortunate state in our controller.
4181 if (parent != root_mem_cgroup)
4182 memory_cgrp_subsys.broken_hierarchy = true;
4185 /* The following stuff does not apply to the root */
4187 root_mem_cgroup = memcg;
4191 error = memcg_online_kmem(memcg);
4195 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4196 static_branch_inc(&memcg_sockets_enabled_key);
4200 mem_cgroup_free(memcg);
4205 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4207 if (css->id > MEM_CGROUP_ID_MAX)
4213 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4215 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4216 struct mem_cgroup_event *event, *tmp;
4219 * Unregister events and notify userspace.
4220 * Notify userspace about cgroup removing only after rmdir of cgroup
4221 * directory to avoid race between userspace and kernelspace.
4223 spin_lock(&memcg->event_list_lock);
4224 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4225 list_del_init(&event->list);
4226 schedule_work(&event->remove);
4228 spin_unlock(&memcg->event_list_lock);
4230 memcg_offline_kmem(memcg);
4231 wb_memcg_offline(memcg);
4234 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4236 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4238 invalidate_reclaim_iterators(memcg);
4241 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4243 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4245 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4246 static_branch_dec(&memcg_sockets_enabled_key);
4248 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4249 static_branch_dec(&memcg_sockets_enabled_key);
4251 vmpressure_cleanup(&memcg->vmpressure);
4252 cancel_work_sync(&memcg->high_work);
4253 mem_cgroup_remove_from_trees(memcg);
4254 memcg_free_kmem(memcg);
4255 mem_cgroup_free(memcg);
4259 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4260 * @css: the target css
4262 * Reset the states of the mem_cgroup associated with @css. This is
4263 * invoked when the userland requests disabling on the default hierarchy
4264 * but the memcg is pinned through dependency. The memcg should stop
4265 * applying policies and should revert to the vanilla state as it may be
4266 * made visible again.
4268 * The current implementation only resets the essential configurations.
4269 * This needs to be expanded to cover all the visible parts.
4271 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4273 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4275 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4276 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4277 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4278 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4279 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4281 memcg->high = PAGE_COUNTER_MAX;
4282 memcg->soft_limit = PAGE_COUNTER_MAX;
4283 memcg_wb_domain_size_changed(memcg);
4287 /* Handlers for move charge at task migration. */
4288 static int mem_cgroup_do_precharge(unsigned long count)
4292 /* Try a single bulk charge without reclaim first, kswapd may wake */
4293 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4295 mc.precharge += count;
4299 /* Try charges one by one with reclaim */
4301 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4311 * get_mctgt_type - get target type of moving charge
4312 * @vma: the vma the pte to be checked belongs
4313 * @addr: the address corresponding to the pte to be checked
4314 * @ptent: the pte to be checked
4315 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4318 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4319 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4320 * move charge. if @target is not NULL, the page is stored in target->page
4321 * with extra refcnt got(Callers should handle it).
4322 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4323 * target for charge migration. if @target is not NULL, the entry is stored
4326 * Called with pte lock held.
4333 enum mc_target_type {
4339 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4340 unsigned long addr, pte_t ptent)
4342 struct page *page = vm_normal_page(vma, addr, ptent);
4344 if (!page || !page_mapped(page))
4346 if (PageAnon(page)) {
4347 if (!(mc.flags & MOVE_ANON))
4350 if (!(mc.flags & MOVE_FILE))
4353 if (!get_page_unless_zero(page))
4360 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4361 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4363 struct page *page = NULL;
4364 swp_entry_t ent = pte_to_swp_entry(ptent);
4366 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4369 * Because lookup_swap_cache() updates some statistics counter,
4370 * we call find_get_page() with swapper_space directly.
4372 page = find_get_page(swap_address_space(ent), ent.val);
4373 if (do_memsw_account())
4374 entry->val = ent.val;
4379 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4380 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4386 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4387 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4389 struct page *page = NULL;
4390 struct address_space *mapping;
4393 if (!vma->vm_file) /* anonymous vma */
4395 if (!(mc.flags & MOVE_FILE))
4398 mapping = vma->vm_file->f_mapping;
4399 pgoff = linear_page_index(vma, addr);
4401 /* page is moved even if it's not RSS of this task(page-faulted). */
4403 /* shmem/tmpfs may report page out on swap: account for that too. */
4404 if (shmem_mapping(mapping)) {
4405 page = find_get_entry(mapping, pgoff);
4406 if (radix_tree_exceptional_entry(page)) {
4407 swp_entry_t swp = radix_to_swp_entry(page);
4408 if (do_memsw_account())
4410 page = find_get_page(swap_address_space(swp), swp.val);
4413 page = find_get_page(mapping, pgoff);
4415 page = find_get_page(mapping, pgoff);
4421 * mem_cgroup_move_account - move account of the page
4423 * @nr_pages: number of regular pages (>1 for huge pages)
4424 * @from: mem_cgroup which the page is moved from.
4425 * @to: mem_cgroup which the page is moved to. @from != @to.
4427 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4429 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4432 static int mem_cgroup_move_account(struct page *page,
4434 struct mem_cgroup *from,
4435 struct mem_cgroup *to)
4437 unsigned long flags;
4438 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4442 VM_BUG_ON(from == to);
4443 VM_BUG_ON_PAGE(PageLRU(page), page);
4444 VM_BUG_ON(compound && !PageTransHuge(page));
4447 * Prevent mem_cgroup_migrate() from looking at
4448 * page->mem_cgroup of its source page while we change it.
4451 if (!trylock_page(page))
4455 if (page->mem_cgroup != from)
4458 anon = PageAnon(page);
4460 spin_lock_irqsave(&from->move_lock, flags);
4462 if (!anon && page_mapped(page)) {
4463 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4465 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4470 * move_lock grabbed above and caller set from->moving_account, so
4471 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4472 * So mapping should be stable for dirty pages.
4474 if (!anon && PageDirty(page)) {
4475 struct address_space *mapping = page_mapping(page);
4477 if (mapping_cap_account_dirty(mapping)) {
4478 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4480 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4485 if (PageWriteback(page)) {
4486 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4488 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4493 * It is safe to change page->mem_cgroup here because the page
4494 * is referenced, charged, and isolated - we can't race with
4495 * uncharging, charging, migration, or LRU putback.
4498 /* caller should have done css_get */
4499 page->mem_cgroup = to;
4500 spin_unlock_irqrestore(&from->move_lock, flags);
4504 local_irq_disable();
4505 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4506 memcg_check_events(to, page);
4507 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4508 memcg_check_events(from, page);
4516 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4517 unsigned long addr, pte_t ptent, union mc_target *target)
4519 struct page *page = NULL;
4520 enum mc_target_type ret = MC_TARGET_NONE;
4521 swp_entry_t ent = { .val = 0 };
4523 if (pte_present(ptent))
4524 page = mc_handle_present_pte(vma, addr, ptent);
4525 else if (is_swap_pte(ptent))
4526 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4527 else if (pte_none(ptent))
4528 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4530 if (!page && !ent.val)
4534 * Do only loose check w/o serialization.
4535 * mem_cgroup_move_account() checks the page is valid or
4536 * not under LRU exclusion.
4538 if (page->mem_cgroup == mc.from) {
4539 ret = MC_TARGET_PAGE;
4541 target->page = page;
4543 if (!ret || !target)
4546 /* There is a swap entry and a page doesn't exist or isn't charged */
4547 if (ent.val && !ret &&
4548 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4549 ret = MC_TARGET_SWAP;
4556 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4558 * We don't consider swapping or file mapped pages because THP does not
4559 * support them for now.
4560 * Caller should make sure that pmd_trans_huge(pmd) is true.
4562 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4563 unsigned long addr, pmd_t pmd, union mc_target *target)
4565 struct page *page = NULL;
4566 enum mc_target_type ret = MC_TARGET_NONE;
4568 page = pmd_page(pmd);
4569 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4570 if (!(mc.flags & MOVE_ANON))
4572 if (page->mem_cgroup == mc.from) {
4573 ret = MC_TARGET_PAGE;
4576 target->page = page;
4582 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4583 unsigned long addr, pmd_t pmd, union mc_target *target)
4585 return MC_TARGET_NONE;
4589 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4590 unsigned long addr, unsigned long end,
4591 struct mm_walk *walk)
4593 struct vm_area_struct *vma = walk->vma;
4597 ptl = pmd_trans_huge_lock(pmd, vma);
4599 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4600 mc.precharge += HPAGE_PMD_NR;
4605 if (pmd_trans_unstable(pmd))
4607 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4608 for (; addr != end; pte++, addr += PAGE_SIZE)
4609 if (get_mctgt_type(vma, addr, *pte, NULL))
4610 mc.precharge++; /* increment precharge temporarily */
4611 pte_unmap_unlock(pte - 1, ptl);
4617 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4619 unsigned long precharge;
4621 struct mm_walk mem_cgroup_count_precharge_walk = {
4622 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4625 down_read(&mm->mmap_sem);
4626 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4627 up_read(&mm->mmap_sem);
4629 precharge = mc.precharge;
4635 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4637 unsigned long precharge = mem_cgroup_count_precharge(mm);
4639 VM_BUG_ON(mc.moving_task);
4640 mc.moving_task = current;
4641 return mem_cgroup_do_precharge(precharge);
4644 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4645 static void __mem_cgroup_clear_mc(void)
4647 struct mem_cgroup *from = mc.from;
4648 struct mem_cgroup *to = mc.to;
4650 /* we must uncharge all the leftover precharges from mc.to */
4652 cancel_charge(mc.to, mc.precharge);
4656 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4657 * we must uncharge here.
4659 if (mc.moved_charge) {
4660 cancel_charge(mc.from, mc.moved_charge);
4661 mc.moved_charge = 0;
4663 /* we must fixup refcnts and charges */
4664 if (mc.moved_swap) {
4665 /* uncharge swap account from the old cgroup */
4666 if (!mem_cgroup_is_root(mc.from))
4667 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4670 * we charged both to->memory and to->memsw, so we
4671 * should uncharge to->memory.
4673 if (!mem_cgroup_is_root(mc.to))
4674 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4676 css_put_many(&mc.from->css, mc.moved_swap);
4678 /* we've already done css_get(mc.to) */
4681 memcg_oom_recover(from);
4682 memcg_oom_recover(to);
4683 wake_up_all(&mc.waitq);
4686 static void mem_cgroup_clear_mc(void)
4688 struct mm_struct *mm = mc.mm;
4691 * we must clear moving_task before waking up waiters at the end of
4694 mc.moving_task = NULL;
4695 __mem_cgroup_clear_mc();
4696 spin_lock(&mc.lock);
4700 spin_unlock(&mc.lock);
4705 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4707 struct cgroup_subsys_state *css;
4708 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4709 struct mem_cgroup *from;
4710 struct task_struct *leader, *p;
4711 struct mm_struct *mm;
4712 unsigned long move_flags;
4715 /* charge immigration isn't supported on the default hierarchy */
4716 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4720 * Multi-process migrations only happen on the default hierarchy
4721 * where charge immigration is not used. Perform charge
4722 * immigration if @tset contains a leader and whine if there are
4726 cgroup_taskset_for_each_leader(leader, css, tset) {
4729 memcg = mem_cgroup_from_css(css);
4735 * We are now commited to this value whatever it is. Changes in this
4736 * tunable will only affect upcoming migrations, not the current one.
4737 * So we need to save it, and keep it going.
4739 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4743 from = mem_cgroup_from_task(p);
4745 VM_BUG_ON(from == memcg);
4747 mm = get_task_mm(p);
4750 /* We move charges only when we move a owner of the mm */
4751 if (mm->owner == p) {
4754 VM_BUG_ON(mc.precharge);
4755 VM_BUG_ON(mc.moved_charge);
4756 VM_BUG_ON(mc.moved_swap);
4758 spin_lock(&mc.lock);
4762 mc.flags = move_flags;
4763 spin_unlock(&mc.lock);
4764 /* We set mc.moving_task later */
4766 ret = mem_cgroup_precharge_mc(mm);
4768 mem_cgroup_clear_mc();
4775 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4778 mem_cgroup_clear_mc();
4781 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4782 unsigned long addr, unsigned long end,
4783 struct mm_walk *walk)
4786 struct vm_area_struct *vma = walk->vma;
4789 enum mc_target_type target_type;
4790 union mc_target target;
4793 ptl = pmd_trans_huge_lock(pmd, vma);
4795 if (mc.precharge < HPAGE_PMD_NR) {
4799 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4800 if (target_type == MC_TARGET_PAGE) {
4802 if (!isolate_lru_page(page)) {
4803 if (!mem_cgroup_move_account(page, true,
4805 mc.precharge -= HPAGE_PMD_NR;
4806 mc.moved_charge += HPAGE_PMD_NR;
4808 putback_lru_page(page);
4816 if (pmd_trans_unstable(pmd))
4819 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4820 for (; addr != end; addr += PAGE_SIZE) {
4821 pte_t ptent = *(pte++);
4827 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4828 case MC_TARGET_PAGE:
4831 * We can have a part of the split pmd here. Moving it
4832 * can be done but it would be too convoluted so simply
4833 * ignore such a partial THP and keep it in original
4834 * memcg. There should be somebody mapping the head.
4836 if (PageTransCompound(page))
4838 if (isolate_lru_page(page))
4840 if (!mem_cgroup_move_account(page, false,
4843 /* we uncharge from mc.from later. */
4846 putback_lru_page(page);
4847 put: /* get_mctgt_type() gets the page */
4850 case MC_TARGET_SWAP:
4852 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4854 /* we fixup refcnts and charges later. */
4862 pte_unmap_unlock(pte - 1, ptl);
4867 * We have consumed all precharges we got in can_attach().
4868 * We try charge one by one, but don't do any additional
4869 * charges to mc.to if we have failed in charge once in attach()
4872 ret = mem_cgroup_do_precharge(1);
4880 static void mem_cgroup_move_charge(void)
4882 struct mm_walk mem_cgroup_move_charge_walk = {
4883 .pmd_entry = mem_cgroup_move_charge_pte_range,
4887 lru_add_drain_all();
4889 * Signal lock_page_memcg() to take the memcg's move_lock
4890 * while we're moving its pages to another memcg. Then wait
4891 * for already started RCU-only updates to finish.
4893 atomic_inc(&mc.from->moving_account);
4896 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
4898 * Someone who are holding the mmap_sem might be waiting in
4899 * waitq. So we cancel all extra charges, wake up all waiters,
4900 * and retry. Because we cancel precharges, we might not be able
4901 * to move enough charges, but moving charge is a best-effort
4902 * feature anyway, so it wouldn't be a big problem.
4904 __mem_cgroup_clear_mc();
4909 * When we have consumed all precharges and failed in doing
4910 * additional charge, the page walk just aborts.
4912 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
4913 up_read(&mc.mm->mmap_sem);
4914 atomic_dec(&mc.from->moving_account);
4917 static void mem_cgroup_move_task(void)
4920 mem_cgroup_move_charge();
4921 mem_cgroup_clear_mc();
4924 #else /* !CONFIG_MMU */
4925 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4929 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4932 static void mem_cgroup_move_task(void)
4938 * Cgroup retains root cgroups across [un]mount cycles making it necessary
4939 * to verify whether we're attached to the default hierarchy on each mount
4942 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
4945 * use_hierarchy is forced on the default hierarchy. cgroup core
4946 * guarantees that @root doesn't have any children, so turning it
4947 * on for the root memcg is enough.
4949 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4950 root_mem_cgroup->use_hierarchy = true;
4952 root_mem_cgroup->use_hierarchy = false;
4955 static u64 memory_current_read(struct cgroup_subsys_state *css,
4958 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4960 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
4963 static int memory_low_show(struct seq_file *m, void *v)
4965 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4966 unsigned long low = READ_ONCE(memcg->low);
4968 if (low == PAGE_COUNTER_MAX)
4969 seq_puts(m, "max\n");
4971 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
4976 static ssize_t memory_low_write(struct kernfs_open_file *of,
4977 char *buf, size_t nbytes, loff_t off)
4979 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4983 buf = strstrip(buf);
4984 err = page_counter_memparse(buf, "max", &low);
4993 static int memory_high_show(struct seq_file *m, void *v)
4995 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4996 unsigned long high = READ_ONCE(memcg->high);
4998 if (high == PAGE_COUNTER_MAX)
4999 seq_puts(m, "max\n");
5001 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5006 static ssize_t memory_high_write(struct kernfs_open_file *of,
5007 char *buf, size_t nbytes, loff_t off)
5009 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5010 unsigned long nr_pages;
5014 buf = strstrip(buf);
5015 err = page_counter_memparse(buf, "max", &high);
5021 nr_pages = page_counter_read(&memcg->memory);
5022 if (nr_pages > high)
5023 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5026 memcg_wb_domain_size_changed(memcg);
5030 static int memory_max_show(struct seq_file *m, void *v)
5032 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5033 unsigned long max = READ_ONCE(memcg->memory.limit);
5035 if (max == PAGE_COUNTER_MAX)
5036 seq_puts(m, "max\n");
5038 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5043 static ssize_t memory_max_write(struct kernfs_open_file *of,
5044 char *buf, size_t nbytes, loff_t off)
5046 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5047 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5048 bool drained = false;
5052 buf = strstrip(buf);
5053 err = page_counter_memparse(buf, "max", &max);
5057 xchg(&memcg->memory.limit, max);
5060 unsigned long nr_pages = page_counter_read(&memcg->memory);
5062 if (nr_pages <= max)
5065 if (signal_pending(current)) {
5071 drain_all_stock(memcg);
5077 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5083 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5084 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5088 memcg_wb_domain_size_changed(memcg);
5092 static int memory_events_show(struct seq_file *m, void *v)
5094 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5096 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5097 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5098 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5099 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5104 static int memory_stat_show(struct seq_file *m, void *v)
5106 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5107 unsigned long stat[MEMCG_NR_STAT];
5108 unsigned long events[MEMCG_NR_EVENTS];
5112 * Provide statistics on the state of the memory subsystem as
5113 * well as cumulative event counters that show past behavior.
5115 * This list is ordered following a combination of these gradients:
5116 * 1) generic big picture -> specifics and details
5117 * 2) reflecting userspace activity -> reflecting kernel heuristics
5119 * Current memory state:
5122 tree_stat(memcg, stat);
5123 tree_events(memcg, events);
5125 seq_printf(m, "anon %llu\n",
5126 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5127 seq_printf(m, "file %llu\n",
5128 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5129 seq_printf(m, "kernel_stack %llu\n",
5130 (u64)stat[MEMCG_KERNEL_STACK] * PAGE_SIZE);
5131 seq_printf(m, "slab %llu\n",
5132 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5133 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5134 seq_printf(m, "sock %llu\n",
5135 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5137 seq_printf(m, "file_mapped %llu\n",
5138 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5139 seq_printf(m, "file_dirty %llu\n",
5140 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5141 seq_printf(m, "file_writeback %llu\n",
5142 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5144 for (i = 0; i < NR_LRU_LISTS; i++) {
5145 struct mem_cgroup *mi;
5146 unsigned long val = 0;
5148 for_each_mem_cgroup_tree(mi, memcg)
5149 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5150 seq_printf(m, "%s %llu\n",
5151 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5154 seq_printf(m, "slab_reclaimable %llu\n",
5155 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5156 seq_printf(m, "slab_unreclaimable %llu\n",
5157 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5159 /* Accumulated memory events */
5161 seq_printf(m, "pgfault %lu\n",
5162 events[MEM_CGROUP_EVENTS_PGFAULT]);
5163 seq_printf(m, "pgmajfault %lu\n",
5164 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5169 static struct cftype memory_files[] = {
5172 .flags = CFTYPE_NOT_ON_ROOT,
5173 .read_u64 = memory_current_read,
5177 .flags = CFTYPE_NOT_ON_ROOT,
5178 .seq_show = memory_low_show,
5179 .write = memory_low_write,
5183 .flags = CFTYPE_NOT_ON_ROOT,
5184 .seq_show = memory_high_show,
5185 .write = memory_high_write,
5189 .flags = CFTYPE_NOT_ON_ROOT,
5190 .seq_show = memory_max_show,
5191 .write = memory_max_write,
5195 .flags = CFTYPE_NOT_ON_ROOT,
5196 .file_offset = offsetof(struct mem_cgroup, events_file),
5197 .seq_show = memory_events_show,
5201 .flags = CFTYPE_NOT_ON_ROOT,
5202 .seq_show = memory_stat_show,
5207 struct cgroup_subsys memory_cgrp_subsys = {
5208 .css_alloc = mem_cgroup_css_alloc,
5209 .css_online = mem_cgroup_css_online,
5210 .css_offline = mem_cgroup_css_offline,
5211 .css_released = mem_cgroup_css_released,
5212 .css_free = mem_cgroup_css_free,
5213 .css_reset = mem_cgroup_css_reset,
5214 .can_attach = mem_cgroup_can_attach,
5215 .cancel_attach = mem_cgroup_cancel_attach,
5216 .post_attach = mem_cgroup_move_task,
5217 .bind = mem_cgroup_bind,
5218 .dfl_cftypes = memory_files,
5219 .legacy_cftypes = mem_cgroup_legacy_files,
5224 * mem_cgroup_low - check if memory consumption is below the normal range
5225 * @root: the highest ancestor to consider
5226 * @memcg: the memory cgroup to check
5228 * Returns %true if memory consumption of @memcg, and that of all
5229 * configurable ancestors up to @root, is below the normal range.
5231 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5233 if (mem_cgroup_disabled())
5237 * The toplevel group doesn't have a configurable range, so
5238 * it's never low when looked at directly, and it is not
5239 * considered an ancestor when assessing the hierarchy.
5242 if (memcg == root_mem_cgroup)
5245 if (page_counter_read(&memcg->memory) >= memcg->low)
5248 while (memcg != root) {
5249 memcg = parent_mem_cgroup(memcg);
5251 if (memcg == root_mem_cgroup)
5254 if (page_counter_read(&memcg->memory) >= memcg->low)
5261 * mem_cgroup_try_charge - try charging a page
5262 * @page: page to charge
5263 * @mm: mm context of the victim
5264 * @gfp_mask: reclaim mode
5265 * @memcgp: charged memcg return
5267 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5268 * pages according to @gfp_mask if necessary.
5270 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5271 * Otherwise, an error code is returned.
5273 * After page->mapping has been set up, the caller must finalize the
5274 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5275 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5277 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5278 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5281 struct mem_cgroup *memcg = NULL;
5282 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5285 if (mem_cgroup_disabled())
5288 if (PageSwapCache(page)) {
5290 * Every swap fault against a single page tries to charge the
5291 * page, bail as early as possible. shmem_unuse() encounters
5292 * already charged pages, too. The USED bit is protected by
5293 * the page lock, which serializes swap cache removal, which
5294 * in turn serializes uncharging.
5296 VM_BUG_ON_PAGE(!PageLocked(page), page);
5297 if (page->mem_cgroup)
5300 if (do_swap_account) {
5301 swp_entry_t ent = { .val = page_private(page), };
5302 unsigned short id = lookup_swap_cgroup_id(ent);
5305 memcg = mem_cgroup_from_id(id);
5306 if (memcg && !css_tryget_online(&memcg->css))
5313 memcg = get_mem_cgroup_from_mm(mm);
5315 ret = try_charge(memcg, gfp_mask, nr_pages);
5317 css_put(&memcg->css);
5324 * mem_cgroup_commit_charge - commit a page charge
5325 * @page: page to charge
5326 * @memcg: memcg to charge the page to
5327 * @lrucare: page might be on LRU already
5329 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5330 * after page->mapping has been set up. This must happen atomically
5331 * as part of the page instantiation, i.e. under the page table lock
5332 * for anonymous pages, under the page lock for page and swap cache.
5334 * In addition, the page must not be on the LRU during the commit, to
5335 * prevent racing with task migration. If it might be, use @lrucare.
5337 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5339 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5340 bool lrucare, bool compound)
5342 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5344 VM_BUG_ON_PAGE(!page->mapping, page);
5345 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5347 if (mem_cgroup_disabled())
5350 * Swap faults will attempt to charge the same page multiple
5351 * times. But reuse_swap_page() might have removed the page
5352 * from swapcache already, so we can't check PageSwapCache().
5357 commit_charge(page, memcg, lrucare);
5359 local_irq_disable();
5360 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5361 memcg_check_events(memcg, page);
5364 if (do_memsw_account() && PageSwapCache(page)) {
5365 swp_entry_t entry = { .val = page_private(page) };
5367 * The swap entry might not get freed for a long time,
5368 * let's not wait for it. The page already received a
5369 * memory+swap charge, drop the swap entry duplicate.
5371 mem_cgroup_uncharge_swap(entry);
5376 * mem_cgroup_cancel_charge - cancel a page charge
5377 * @page: page to charge
5378 * @memcg: memcg to charge the page to
5380 * Cancel a charge transaction started by mem_cgroup_try_charge().
5382 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5385 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5387 if (mem_cgroup_disabled())
5390 * Swap faults will attempt to charge the same page multiple
5391 * times. But reuse_swap_page() might have removed the page
5392 * from swapcache already, so we can't check PageSwapCache().
5397 cancel_charge(memcg, nr_pages);
5400 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5401 unsigned long nr_anon, unsigned long nr_file,
5402 unsigned long nr_huge, struct page *dummy_page)
5404 unsigned long nr_pages = nr_anon + nr_file;
5405 unsigned long flags;
5407 if (!mem_cgroup_is_root(memcg)) {
5408 page_counter_uncharge(&memcg->memory, nr_pages);
5409 if (do_memsw_account())
5410 page_counter_uncharge(&memcg->memsw, nr_pages);
5411 memcg_oom_recover(memcg);
5414 local_irq_save(flags);
5415 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5416 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5417 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5418 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5419 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5420 memcg_check_events(memcg, dummy_page);
5421 local_irq_restore(flags);
5423 if (!mem_cgroup_is_root(memcg))
5424 css_put_many(&memcg->css, nr_pages);
5427 static void uncharge_list(struct list_head *page_list)
5429 struct mem_cgroup *memcg = NULL;
5430 unsigned long nr_anon = 0;
5431 unsigned long nr_file = 0;
5432 unsigned long nr_huge = 0;
5433 unsigned long pgpgout = 0;
5434 struct list_head *next;
5438 * Note that the list can be a single page->lru; hence the
5439 * do-while loop instead of a simple list_for_each_entry().
5441 next = page_list->next;
5443 unsigned int nr_pages = 1;
5445 page = list_entry(next, struct page, lru);
5446 next = page->lru.next;
5448 VM_BUG_ON_PAGE(PageLRU(page), page);
5449 VM_BUG_ON_PAGE(page_count(page), page);
5451 if (!page->mem_cgroup)
5455 * Nobody should be changing or seriously looking at
5456 * page->mem_cgroup at this point, we have fully
5457 * exclusive access to the page.
5460 if (memcg != page->mem_cgroup) {
5462 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5464 pgpgout = nr_anon = nr_file = nr_huge = 0;
5466 memcg = page->mem_cgroup;
5469 if (PageTransHuge(page)) {
5470 nr_pages <<= compound_order(page);
5471 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5472 nr_huge += nr_pages;
5476 nr_anon += nr_pages;
5478 nr_file += nr_pages;
5480 page->mem_cgroup = NULL;
5483 } while (next != page_list);
5486 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5491 * mem_cgroup_uncharge - uncharge a page
5492 * @page: page to uncharge
5494 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5495 * mem_cgroup_commit_charge().
5497 void mem_cgroup_uncharge(struct page *page)
5499 if (mem_cgroup_disabled())
5502 /* Don't touch page->lru of any random page, pre-check: */
5503 if (!page->mem_cgroup)
5506 INIT_LIST_HEAD(&page->lru);
5507 uncharge_list(&page->lru);
5511 * mem_cgroup_uncharge_list - uncharge a list of page
5512 * @page_list: list of pages to uncharge
5514 * Uncharge a list of pages previously charged with
5515 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5517 void mem_cgroup_uncharge_list(struct list_head *page_list)
5519 if (mem_cgroup_disabled())
5522 if (!list_empty(page_list))
5523 uncharge_list(page_list);
5527 * mem_cgroup_migrate - charge a page's replacement
5528 * @oldpage: currently circulating page
5529 * @newpage: replacement page
5531 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5532 * be uncharged upon free.
5534 * Both pages must be locked, @newpage->mapping must be set up.
5536 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5538 struct mem_cgroup *memcg;
5539 unsigned int nr_pages;
5542 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5543 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5544 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5545 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5548 if (mem_cgroup_disabled())
5551 /* Page cache replacement: new page already charged? */
5552 if (newpage->mem_cgroup)
5555 /* Swapcache readahead pages can get replaced before being charged */
5556 memcg = oldpage->mem_cgroup;
5560 /* Force-charge the new page. The old one will be freed soon */
5561 compound = PageTransHuge(newpage);
5562 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5564 page_counter_charge(&memcg->memory, nr_pages);
5565 if (do_memsw_account())
5566 page_counter_charge(&memcg->memsw, nr_pages);
5567 css_get_many(&memcg->css, nr_pages);
5569 commit_charge(newpage, memcg, false);
5571 local_irq_disable();
5572 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5573 memcg_check_events(memcg, newpage);
5577 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5578 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5580 void sock_update_memcg(struct sock *sk)
5582 struct mem_cgroup *memcg;
5584 /* Socket cloning can throw us here with sk_cgrp already
5585 * filled. It won't however, necessarily happen from
5586 * process context. So the test for root memcg given
5587 * the current task's memcg won't help us in this case.
5589 * Respecting the original socket's memcg is a better
5590 * decision in this case.
5593 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5594 css_get(&sk->sk_memcg->css);
5599 memcg = mem_cgroup_from_task(current);
5600 if (memcg == root_mem_cgroup)
5602 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5604 if (css_tryget_online(&memcg->css))
5605 sk->sk_memcg = memcg;
5609 EXPORT_SYMBOL(sock_update_memcg);
5611 void sock_release_memcg(struct sock *sk)
5613 WARN_ON(!sk->sk_memcg);
5614 css_put(&sk->sk_memcg->css);
5618 * mem_cgroup_charge_skmem - charge socket memory
5619 * @memcg: memcg to charge
5620 * @nr_pages: number of pages to charge
5622 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5623 * @memcg's configured limit, %false if the charge had to be forced.
5625 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5627 gfp_t gfp_mask = GFP_KERNEL;
5629 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5630 struct page_counter *fail;
5632 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5633 memcg->tcpmem_pressure = 0;
5636 page_counter_charge(&memcg->tcpmem, nr_pages);
5637 memcg->tcpmem_pressure = 1;
5641 /* Don't block in the packet receive path */
5643 gfp_mask = GFP_NOWAIT;
5645 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5647 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5650 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5655 * mem_cgroup_uncharge_skmem - uncharge socket memory
5656 * @memcg - memcg to uncharge
5657 * @nr_pages - number of pages to uncharge
5659 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5661 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5662 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5666 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5668 page_counter_uncharge(&memcg->memory, nr_pages);
5669 css_put_many(&memcg->css, nr_pages);
5672 static int __init cgroup_memory(char *s)
5676 while ((token = strsep(&s, ",")) != NULL) {
5679 if (!strcmp(token, "nosocket"))
5680 cgroup_memory_nosocket = true;
5681 if (!strcmp(token, "nokmem"))
5682 cgroup_memory_nokmem = true;
5686 __setup("cgroup.memory=", cgroup_memory);
5689 * subsys_initcall() for memory controller.
5691 * Some parts like hotcpu_notifier() have to be initialized from this context
5692 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5693 * everything that doesn't depend on a specific mem_cgroup structure should
5694 * be initialized from here.
5696 static int __init mem_cgroup_init(void)
5700 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5702 for_each_possible_cpu(cpu)
5703 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5706 for_each_node(node) {
5707 struct mem_cgroup_tree_per_node *rtpn;
5710 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5711 node_online(node) ? node : NUMA_NO_NODE);
5713 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5714 struct mem_cgroup_tree_per_zone *rtpz;
5716 rtpz = &rtpn->rb_tree_per_zone[zone];
5717 rtpz->rb_root = RB_ROOT;
5718 spin_lock_init(&rtpz->lock);
5720 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5725 subsys_initcall(mem_cgroup_init);
5727 #ifdef CONFIG_MEMCG_SWAP
5729 * mem_cgroup_swapout - transfer a memsw charge to swap
5730 * @page: page whose memsw charge to transfer
5731 * @entry: swap entry to move the charge to
5733 * Transfer the memsw charge of @page to @entry.
5735 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5737 struct mem_cgroup *memcg;
5738 unsigned short oldid;
5740 VM_BUG_ON_PAGE(PageLRU(page), page);
5741 VM_BUG_ON_PAGE(page_count(page), page);
5743 if (!do_memsw_account())
5746 memcg = page->mem_cgroup;
5748 /* Readahead page, never charged */
5752 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5753 VM_BUG_ON_PAGE(oldid, page);
5754 mem_cgroup_swap_statistics(memcg, true);
5756 page->mem_cgroup = NULL;
5758 if (!mem_cgroup_is_root(memcg))
5759 page_counter_uncharge(&memcg->memory, 1);
5762 * Interrupts should be disabled here because the caller holds the
5763 * mapping->tree_lock lock which is taken with interrupts-off. It is
5764 * important here to have the interrupts disabled because it is the
5765 * only synchronisation we have for udpating the per-CPU variables.
5767 VM_BUG_ON(!irqs_disabled());
5768 mem_cgroup_charge_statistics(memcg, page, false, -1);
5769 memcg_check_events(memcg, page);
5773 * mem_cgroup_try_charge_swap - try charging a swap entry
5774 * @page: page being added to swap
5775 * @entry: swap entry to charge
5777 * Try to charge @entry to the memcg that @page belongs to.
5779 * Returns 0 on success, -ENOMEM on failure.
5781 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5783 struct mem_cgroup *memcg;
5784 struct page_counter *counter;
5785 unsigned short oldid;
5787 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5790 memcg = page->mem_cgroup;
5792 /* Readahead page, never charged */
5796 if (!mem_cgroup_is_root(memcg) &&
5797 !page_counter_try_charge(&memcg->swap, 1, &counter))
5800 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5801 VM_BUG_ON_PAGE(oldid, page);
5802 mem_cgroup_swap_statistics(memcg, true);
5804 css_get(&memcg->css);
5809 * mem_cgroup_uncharge_swap - uncharge a swap entry
5810 * @entry: swap entry to uncharge
5812 * Drop the swap charge associated with @entry.
5814 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5816 struct mem_cgroup *memcg;
5819 if (!do_swap_account)
5822 id = swap_cgroup_record(entry, 0);
5824 memcg = mem_cgroup_from_id(id);
5826 if (!mem_cgroup_is_root(memcg)) {
5827 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5828 page_counter_uncharge(&memcg->swap, 1);
5830 page_counter_uncharge(&memcg->memsw, 1);
5832 mem_cgroup_swap_statistics(memcg, false);
5833 css_put(&memcg->css);
5838 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5840 long nr_swap_pages = get_nr_swap_pages();
5842 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5843 return nr_swap_pages;
5844 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5845 nr_swap_pages = min_t(long, nr_swap_pages,
5846 READ_ONCE(memcg->swap.limit) -
5847 page_counter_read(&memcg->swap));
5848 return nr_swap_pages;
5851 bool mem_cgroup_swap_full(struct page *page)
5853 struct mem_cgroup *memcg;
5855 VM_BUG_ON_PAGE(!PageLocked(page), page);
5859 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5862 memcg = page->mem_cgroup;
5866 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5867 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
5873 /* for remember boot option*/
5874 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5875 static int really_do_swap_account __initdata = 1;
5877 static int really_do_swap_account __initdata;
5880 static int __init enable_swap_account(char *s)
5882 if (!strcmp(s, "1"))
5883 really_do_swap_account = 1;
5884 else if (!strcmp(s, "0"))
5885 really_do_swap_account = 0;
5888 __setup("swapaccount=", enable_swap_account);
5890 static u64 swap_current_read(struct cgroup_subsys_state *css,
5893 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5895 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
5898 static int swap_max_show(struct seq_file *m, void *v)
5900 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5901 unsigned long max = READ_ONCE(memcg->swap.limit);
5903 if (max == PAGE_COUNTER_MAX)
5904 seq_puts(m, "max\n");
5906 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5911 static ssize_t swap_max_write(struct kernfs_open_file *of,
5912 char *buf, size_t nbytes, loff_t off)
5914 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5918 buf = strstrip(buf);
5919 err = page_counter_memparse(buf, "max", &max);
5923 mutex_lock(&memcg_limit_mutex);
5924 err = page_counter_limit(&memcg->swap, max);
5925 mutex_unlock(&memcg_limit_mutex);
5932 static struct cftype swap_files[] = {
5934 .name = "swap.current",
5935 .flags = CFTYPE_NOT_ON_ROOT,
5936 .read_u64 = swap_current_read,
5940 .flags = CFTYPE_NOT_ON_ROOT,
5941 .seq_show = swap_max_show,
5942 .write = swap_max_write,
5947 static struct cftype memsw_cgroup_files[] = {
5949 .name = "memsw.usage_in_bytes",
5950 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5951 .read_u64 = mem_cgroup_read_u64,
5954 .name = "memsw.max_usage_in_bytes",
5955 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5956 .write = mem_cgroup_reset,
5957 .read_u64 = mem_cgroup_read_u64,
5960 .name = "memsw.limit_in_bytes",
5961 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5962 .write = mem_cgroup_write,
5963 .read_u64 = mem_cgroup_read_u64,
5966 .name = "memsw.failcnt",
5967 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5968 .write = mem_cgroup_reset,
5969 .read_u64 = mem_cgroup_read_u64,
5971 { }, /* terminate */
5974 static int __init mem_cgroup_swap_init(void)
5976 if (!mem_cgroup_disabled() && really_do_swap_account) {
5977 do_swap_account = 1;
5978 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
5980 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5981 memsw_cgroup_files));
5985 subsys_initcall(mem_cgroup_swap_init);
5987 #endif /* CONFIG_MEMCG_SWAP */