1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
65 #include <linux/tracehook.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 struct mem_cgroup *root_mem_cgroup __read_mostly;
80 #define MEM_CGROUP_RECLAIM_RETRIES 5
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 int do_swap_account __read_mostly;
92 #define do_swap_account 0
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char * const mem_cgroup_stat_names[] = {
111 static const char * const mem_cgroup_events_names[] = {
118 static const char * const mem_cgroup_lru_names[] = {
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
131 * Cgroups above their limits are maintained in a RB-Tree, independent of
132 * their hierarchy representation
135 struct mem_cgroup_tree_per_node {
136 struct rb_root rb_root;
140 struct mem_cgroup_tree {
141 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
144 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
147 struct mem_cgroup_eventfd_list {
148 struct list_head list;
149 struct eventfd_ctx *eventfd;
153 * cgroup_event represents events which userspace want to receive.
155 struct mem_cgroup_event {
157 * memcg which the event belongs to.
159 struct mem_cgroup *memcg;
161 * eventfd to signal userspace about the event.
163 struct eventfd_ctx *eventfd;
165 * Each of these stored in a list by the cgroup.
167 struct list_head list;
169 * register_event() callback will be used to add new userspace
170 * waiter for changes related to this event. Use eventfd_signal()
171 * on eventfd to send notification to userspace.
173 int (*register_event)(struct mem_cgroup *memcg,
174 struct eventfd_ctx *eventfd, const char *args);
176 * unregister_event() callback will be called when userspace closes
177 * the eventfd or on cgroup removing. This callback must be set,
178 * if you want provide notification functionality.
180 void (*unregister_event)(struct mem_cgroup *memcg,
181 struct eventfd_ctx *eventfd);
183 * All fields below needed to unregister event when
184 * userspace closes eventfd.
187 wait_queue_head_t *wqh;
189 struct work_struct remove;
192 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
195 /* Stuffs for move charges at task migration. */
197 * Types of charges to be moved.
199 #define MOVE_ANON 0x1U
200 #define MOVE_FILE 0x2U
201 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
203 /* "mc" and its members are protected by cgroup_mutex */
204 static struct move_charge_struct {
205 spinlock_t lock; /* for from, to */
206 struct mm_struct *mm;
207 struct mem_cgroup *from;
208 struct mem_cgroup *to;
210 unsigned long precharge;
211 unsigned long moved_charge;
212 unsigned long moved_swap;
213 struct task_struct *moving_task; /* a task moving charges */
214 wait_queue_head_t waitq; /* a waitq for other context */
216 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
217 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
221 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
222 * limit reclaim to prevent infinite loops, if they ever occur.
224 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
225 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
228 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
229 MEM_CGROUP_CHARGE_TYPE_ANON,
230 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
231 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
235 /* for encoding cft->private value on file */
244 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
245 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
246 #define MEMFILE_ATTR(val) ((val) & 0xffff)
247 /* Used for OOM nofiier */
248 #define OOM_CONTROL (0)
250 /* Some nice accessors for the vmpressure. */
251 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
254 memcg = root_mem_cgroup;
255 return &memcg->vmpressure;
258 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
260 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
263 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
265 return (memcg == root_mem_cgroup);
270 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
271 * The main reason for not using cgroup id for this:
272 * this works better in sparse environments, where we have a lot of memcgs,
273 * but only a few kmem-limited. Or also, if we have, for instance, 200
274 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
275 * 200 entry array for that.
277 * The current size of the caches array is stored in memcg_nr_cache_ids. It
278 * will double each time we have to increase it.
280 static DEFINE_IDA(memcg_cache_ida);
281 int memcg_nr_cache_ids;
283 /* Protects memcg_nr_cache_ids */
284 static DECLARE_RWSEM(memcg_cache_ids_sem);
286 void memcg_get_cache_ids(void)
288 down_read(&memcg_cache_ids_sem);
291 void memcg_put_cache_ids(void)
293 up_read(&memcg_cache_ids_sem);
297 * MIN_SIZE is different than 1, because we would like to avoid going through
298 * the alloc/free process all the time. In a small machine, 4 kmem-limited
299 * cgroups is a reasonable guess. In the future, it could be a parameter or
300 * tunable, but that is strictly not necessary.
302 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
303 * this constant directly from cgroup, but it is understandable that this is
304 * better kept as an internal representation in cgroup.c. In any case, the
305 * cgrp_id space is not getting any smaller, and we don't have to necessarily
306 * increase ours as well if it increases.
308 #define MEMCG_CACHES_MIN_SIZE 4
309 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
312 * A lot of the calls to the cache allocation functions are expected to be
313 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
314 * conditional to this static branch, we'll have to allow modules that does
315 * kmem_cache_alloc and the such to see this symbol as well
317 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
318 EXPORT_SYMBOL(memcg_kmem_enabled_key);
320 #endif /* !CONFIG_SLOB */
323 * mem_cgroup_css_from_page - css of the memcg associated with a page
324 * @page: page of interest
326 * If memcg is bound to the default hierarchy, css of the memcg associated
327 * with @page is returned. The returned css remains associated with @page
328 * until it is released.
330 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
333 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
335 struct mem_cgroup *memcg;
337 memcg = page->mem_cgroup;
339 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
340 memcg = root_mem_cgroup;
346 * page_cgroup_ino - return inode number of the memcg a page is charged to
349 * Look up the closest online ancestor of the memory cgroup @page is charged to
350 * and return its inode number or 0 if @page is not charged to any cgroup. It
351 * is safe to call this function without holding a reference to @page.
353 * Note, this function is inherently racy, because there is nothing to prevent
354 * the cgroup inode from getting torn down and potentially reallocated a moment
355 * after page_cgroup_ino() returns, so it only should be used by callers that
356 * do not care (such as procfs interfaces).
358 ino_t page_cgroup_ino(struct page *page)
360 struct mem_cgroup *memcg;
361 unsigned long ino = 0;
364 memcg = READ_ONCE(page->mem_cgroup);
365 while (memcg && !(memcg->css.flags & CSS_ONLINE))
366 memcg = parent_mem_cgroup(memcg);
368 ino = cgroup_ino(memcg->css.cgroup);
373 static struct mem_cgroup_per_node *
374 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
376 int nid = page_to_nid(page);
378 return memcg->nodeinfo[nid];
381 static struct mem_cgroup_tree_per_node *
382 soft_limit_tree_node(int nid)
384 return soft_limit_tree.rb_tree_per_node[nid];
387 static struct mem_cgroup_tree_per_node *
388 soft_limit_tree_from_page(struct page *page)
390 int nid = page_to_nid(page);
392 return soft_limit_tree.rb_tree_per_node[nid];
395 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
396 struct mem_cgroup_tree_per_node *mctz,
397 unsigned long new_usage_in_excess)
399 struct rb_node **p = &mctz->rb_root.rb_node;
400 struct rb_node *parent = NULL;
401 struct mem_cgroup_per_node *mz_node;
406 mz->usage_in_excess = new_usage_in_excess;
407 if (!mz->usage_in_excess)
411 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
413 if (mz->usage_in_excess < mz_node->usage_in_excess)
416 * We can't avoid mem cgroups that are over their soft
417 * limit by the same amount
419 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
422 rb_link_node(&mz->tree_node, parent, p);
423 rb_insert_color(&mz->tree_node, &mctz->rb_root);
427 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
428 struct mem_cgroup_tree_per_node *mctz)
432 rb_erase(&mz->tree_node, &mctz->rb_root);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
437 struct mem_cgroup_tree_per_node *mctz)
441 spin_lock_irqsave(&mctz->lock, flags);
442 __mem_cgroup_remove_exceeded(mz, mctz);
443 spin_unlock_irqrestore(&mctz->lock, flags);
446 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
448 unsigned long nr_pages = page_counter_read(&memcg->memory);
449 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
450 unsigned long excess = 0;
452 if (nr_pages > soft_limit)
453 excess = nr_pages - soft_limit;
458 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
460 unsigned long excess;
461 struct mem_cgroup_per_node *mz;
462 struct mem_cgroup_tree_per_node *mctz;
464 mctz = soft_limit_tree_from_page(page);
466 * Necessary to update all ancestors when hierarchy is used.
467 * because their event counter is not touched.
469 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
470 mz = mem_cgroup_page_nodeinfo(memcg, page);
471 excess = soft_limit_excess(memcg);
473 * We have to update the tree if mz is on RB-tree or
474 * mem is over its softlimit.
476 if (excess || mz->on_tree) {
479 spin_lock_irqsave(&mctz->lock, flags);
480 /* if on-tree, remove it */
482 __mem_cgroup_remove_exceeded(mz, mctz);
484 * Insert again. mz->usage_in_excess will be updated.
485 * If excess is 0, no tree ops.
487 __mem_cgroup_insert_exceeded(mz, mctz, excess);
488 spin_unlock_irqrestore(&mctz->lock, flags);
493 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
495 struct mem_cgroup_tree_per_node *mctz;
496 struct mem_cgroup_per_node *mz;
500 mz = mem_cgroup_nodeinfo(memcg, nid);
501 mctz = soft_limit_tree_node(nid);
502 mem_cgroup_remove_exceeded(mz, mctz);
506 static struct mem_cgroup_per_node *
507 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
509 struct rb_node *rightmost = NULL;
510 struct mem_cgroup_per_node *mz;
514 rightmost = rb_last(&mctz->rb_root);
516 goto done; /* Nothing to reclaim from */
518 mz = rb_entry(rightmost, struct mem_cgroup_per_node, tree_node);
520 * Remove the node now but someone else can add it back,
521 * we will to add it back at the end of reclaim to its correct
522 * position in the tree.
524 __mem_cgroup_remove_exceeded(mz, mctz);
525 if (!soft_limit_excess(mz->memcg) ||
526 !css_tryget_online(&mz->memcg->css))
532 static struct mem_cgroup_per_node *
533 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
535 struct mem_cgroup_per_node *mz;
537 spin_lock_irq(&mctz->lock);
538 mz = __mem_cgroup_largest_soft_limit_node(mctz);
539 spin_unlock_irq(&mctz->lock);
544 * Return page count for single (non recursive) @memcg.
546 * Implementation Note: reading percpu statistics for memcg.
548 * Both of vmstat[] and percpu_counter has threshold and do periodic
549 * synchronization to implement "quick" read. There are trade-off between
550 * reading cost and precision of value. Then, we may have a chance to implement
551 * a periodic synchronization of counter in memcg's counter.
553 * But this _read() function is used for user interface now. The user accounts
554 * memory usage by memory cgroup and he _always_ requires exact value because
555 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
556 * have to visit all online cpus and make sum. So, for now, unnecessary
557 * synchronization is not implemented. (just implemented for cpu hotplug)
559 * If there are kernel internal actions which can make use of some not-exact
560 * value, and reading all cpu value can be performance bottleneck in some
561 * common workload, threshold and synchronization as vmstat[] should be
565 mem_cgroup_read_stat(struct mem_cgroup *memcg, enum mem_cgroup_stat_index idx)
570 /* Per-cpu values can be negative, use a signed accumulator */
571 for_each_possible_cpu(cpu)
572 val += per_cpu(memcg->stat->count[idx], cpu);
574 * Summing races with updates, so val may be negative. Avoid exposing
575 * transient negative values.
582 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
583 enum mem_cgroup_events_index idx)
585 unsigned long val = 0;
588 for_each_possible_cpu(cpu)
589 val += per_cpu(memcg->stat->events[idx], cpu);
593 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
595 bool compound, int nr_pages)
598 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
599 * counted as CACHE even if it's on ANON LRU.
602 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
605 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
609 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
610 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
614 /* pagein of a big page is an event. So, ignore page size */
616 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
618 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
619 nr_pages = -nr_pages; /* for event */
622 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
625 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
626 int nid, unsigned int lru_mask)
628 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
629 unsigned long nr = 0;
632 VM_BUG_ON((unsigned)nid >= nr_node_ids);
635 if (!(BIT(lru) & lru_mask))
637 nr += mem_cgroup_get_lru_size(lruvec, lru);
642 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
643 unsigned int lru_mask)
645 unsigned long nr = 0;
648 for_each_node_state(nid, N_MEMORY)
649 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
653 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
654 enum mem_cgroup_events_target target)
656 unsigned long val, next;
658 val = __this_cpu_read(memcg->stat->nr_page_events);
659 next = __this_cpu_read(memcg->stat->targets[target]);
660 /* from time_after() in jiffies.h */
661 if ((long)next - (long)val < 0) {
663 case MEM_CGROUP_TARGET_THRESH:
664 next = val + THRESHOLDS_EVENTS_TARGET;
666 case MEM_CGROUP_TARGET_SOFTLIMIT:
667 next = val + SOFTLIMIT_EVENTS_TARGET;
669 case MEM_CGROUP_TARGET_NUMAINFO:
670 next = val + NUMAINFO_EVENTS_TARGET;
675 __this_cpu_write(memcg->stat->targets[target], next);
682 * Check events in order.
685 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
687 /* threshold event is triggered in finer grain than soft limit */
688 if (unlikely(mem_cgroup_event_ratelimit(memcg,
689 MEM_CGROUP_TARGET_THRESH))) {
691 bool do_numainfo __maybe_unused;
693 do_softlimit = mem_cgroup_event_ratelimit(memcg,
694 MEM_CGROUP_TARGET_SOFTLIMIT);
696 do_numainfo = mem_cgroup_event_ratelimit(memcg,
697 MEM_CGROUP_TARGET_NUMAINFO);
699 mem_cgroup_threshold(memcg);
700 if (unlikely(do_softlimit))
701 mem_cgroup_update_tree(memcg, page);
703 if (unlikely(do_numainfo))
704 atomic_inc(&memcg->numainfo_events);
709 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
712 * mm_update_next_owner() may clear mm->owner to NULL
713 * if it races with swapoff, page migration, etc.
714 * So this can be called with p == NULL.
719 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
721 EXPORT_SYMBOL(mem_cgroup_from_task);
723 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
725 struct mem_cgroup *memcg = NULL;
730 * Page cache insertions can happen withou an
731 * actual mm context, e.g. during disk probing
732 * on boot, loopback IO, acct() writes etc.
735 memcg = root_mem_cgroup;
737 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
738 if (unlikely(!memcg))
739 memcg = root_mem_cgroup;
741 } while (!css_tryget_online(&memcg->css));
747 * mem_cgroup_iter - iterate over memory cgroup hierarchy
748 * @root: hierarchy root
749 * @prev: previously returned memcg, NULL on first invocation
750 * @reclaim: cookie for shared reclaim walks, NULL for full walks
752 * Returns references to children of the hierarchy below @root, or
753 * @root itself, or %NULL after a full round-trip.
755 * Caller must pass the return value in @prev on subsequent
756 * invocations for reference counting, or use mem_cgroup_iter_break()
757 * to cancel a hierarchy walk before the round-trip is complete.
759 * Reclaimers can specify a zone and a priority level in @reclaim to
760 * divide up the memcgs in the hierarchy among all concurrent
761 * reclaimers operating on the same zone and priority.
763 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
764 struct mem_cgroup *prev,
765 struct mem_cgroup_reclaim_cookie *reclaim)
767 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
768 struct cgroup_subsys_state *css = NULL;
769 struct mem_cgroup *memcg = NULL;
770 struct mem_cgroup *pos = NULL;
772 if (mem_cgroup_disabled())
776 root = root_mem_cgroup;
778 if (prev && !reclaim)
781 if (!root->use_hierarchy && root != root_mem_cgroup) {
790 struct mem_cgroup_per_node *mz;
792 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
793 iter = &mz->iter[reclaim->priority];
795 if (prev && reclaim->generation != iter->generation)
799 pos = READ_ONCE(iter->position);
800 if (!pos || css_tryget(&pos->css))
803 * css reference reached zero, so iter->position will
804 * be cleared by ->css_released. However, we should not
805 * rely on this happening soon, because ->css_released
806 * is called from a work queue, and by busy-waiting we
807 * might block it. So we clear iter->position right
810 (void)cmpxchg(&iter->position, pos, NULL);
818 css = css_next_descendant_pre(css, &root->css);
821 * Reclaimers share the hierarchy walk, and a
822 * new one might jump in right at the end of
823 * the hierarchy - make sure they see at least
824 * one group and restart from the beginning.
832 * Verify the css and acquire a reference. The root
833 * is provided by the caller, so we know it's alive
834 * and kicking, and don't take an extra reference.
836 memcg = mem_cgroup_from_css(css);
838 if (css == &root->css)
849 * The position could have already been updated by a competing
850 * thread, so check that the value hasn't changed since we read
851 * it to avoid reclaiming from the same cgroup twice.
853 (void)cmpxchg(&iter->position, pos, memcg);
861 reclaim->generation = iter->generation;
867 if (prev && prev != root)
874 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
875 * @root: hierarchy root
876 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
878 void mem_cgroup_iter_break(struct mem_cgroup *root,
879 struct mem_cgroup *prev)
882 root = root_mem_cgroup;
883 if (prev && prev != root)
887 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
889 struct mem_cgroup *memcg = dead_memcg;
890 struct mem_cgroup_reclaim_iter *iter;
891 struct mem_cgroup_per_node *mz;
895 while ((memcg = parent_mem_cgroup(memcg))) {
897 mz = mem_cgroup_nodeinfo(memcg, nid);
898 for (i = 0; i <= DEF_PRIORITY; i++) {
900 cmpxchg(&iter->position,
908 * Iteration constructs for visiting all cgroups (under a tree). If
909 * loops are exited prematurely (break), mem_cgroup_iter_break() must
910 * be used for reference counting.
912 #define for_each_mem_cgroup_tree(iter, root) \
913 for (iter = mem_cgroup_iter(root, NULL, NULL); \
915 iter = mem_cgroup_iter(root, iter, NULL))
917 #define for_each_mem_cgroup(iter) \
918 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
920 iter = mem_cgroup_iter(NULL, iter, NULL))
923 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
924 * @memcg: hierarchy root
925 * @fn: function to call for each task
926 * @arg: argument passed to @fn
928 * This function iterates over tasks attached to @memcg or to any of its
929 * descendants and calls @fn for each task. If @fn returns a non-zero
930 * value, the function breaks the iteration loop and returns the value.
931 * Otherwise, it will iterate over all tasks and return 0.
933 * This function must not be called for the root memory cgroup.
935 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
936 int (*fn)(struct task_struct *, void *), void *arg)
938 struct mem_cgroup *iter;
941 BUG_ON(memcg == root_mem_cgroup);
943 for_each_mem_cgroup_tree(iter, memcg) {
944 struct css_task_iter it;
945 struct task_struct *task;
947 css_task_iter_start(&iter->css, &it);
948 while (!ret && (task = css_task_iter_next(&it)))
950 css_task_iter_end(&it);
952 mem_cgroup_iter_break(memcg, iter);
960 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
962 * @zone: zone of the page
964 * This function is only safe when following the LRU page isolation
965 * and putback protocol: the LRU lock must be held, and the page must
966 * either be PageLRU() or the caller must have isolated/allocated it.
968 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
970 struct mem_cgroup_per_node *mz;
971 struct mem_cgroup *memcg;
972 struct lruvec *lruvec;
974 if (mem_cgroup_disabled()) {
975 lruvec = &pgdat->lruvec;
979 memcg = page->mem_cgroup;
981 * Swapcache readahead pages are added to the LRU - and
982 * possibly migrated - before they are charged.
985 memcg = root_mem_cgroup;
987 mz = mem_cgroup_page_nodeinfo(memcg, page);
988 lruvec = &mz->lruvec;
991 * Since a node can be onlined after the mem_cgroup was created,
992 * we have to be prepared to initialize lruvec->zone here;
993 * and if offlined then reonlined, we need to reinitialize it.
995 if (unlikely(lruvec->pgdat != pgdat))
996 lruvec->pgdat = pgdat;
1001 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1002 * @lruvec: mem_cgroup per zone lru vector
1003 * @lru: index of lru list the page is sitting on
1004 * @zid: zone id of the accounted pages
1005 * @nr_pages: positive when adding or negative when removing
1007 * This function must be called under lru_lock, just before a page is added
1008 * to or just after a page is removed from an lru list (that ordering being
1009 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1011 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1012 int zid, int nr_pages)
1014 struct mem_cgroup_per_node *mz;
1015 unsigned long *lru_size;
1018 if (mem_cgroup_disabled())
1021 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1022 lru_size = &mz->lru_zone_size[zid][lru];
1025 *lru_size += nr_pages;
1028 if (WARN_ONCE(size < 0,
1029 "%s(%p, %d, %d): lru_size %ld\n",
1030 __func__, lruvec, lru, nr_pages, size)) {
1036 *lru_size += nr_pages;
1039 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1041 struct mem_cgroup *task_memcg;
1042 struct task_struct *p;
1045 p = find_lock_task_mm(task);
1047 task_memcg = get_mem_cgroup_from_mm(p->mm);
1051 * All threads may have already detached their mm's, but the oom
1052 * killer still needs to detect if they have already been oom
1053 * killed to prevent needlessly killing additional tasks.
1056 task_memcg = mem_cgroup_from_task(task);
1057 css_get(&task_memcg->css);
1060 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1061 css_put(&task_memcg->css);
1066 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1067 * @memcg: the memory cgroup
1069 * Returns the maximum amount of memory @mem can be charged with, in
1072 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1074 unsigned long margin = 0;
1075 unsigned long count;
1076 unsigned long limit;
1078 count = page_counter_read(&memcg->memory);
1079 limit = READ_ONCE(memcg->memory.limit);
1081 margin = limit - count;
1083 if (do_memsw_account()) {
1084 count = page_counter_read(&memcg->memsw);
1085 limit = READ_ONCE(memcg->memsw.limit);
1087 margin = min(margin, limit - count);
1096 * A routine for checking "mem" is under move_account() or not.
1098 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1099 * moving cgroups. This is for waiting at high-memory pressure
1102 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1104 struct mem_cgroup *from;
1105 struct mem_cgroup *to;
1108 * Unlike task_move routines, we access mc.to, mc.from not under
1109 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1111 spin_lock(&mc.lock);
1117 ret = mem_cgroup_is_descendant(from, memcg) ||
1118 mem_cgroup_is_descendant(to, memcg);
1120 spin_unlock(&mc.lock);
1124 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1126 if (mc.moving_task && current != mc.moving_task) {
1127 if (mem_cgroup_under_move(memcg)) {
1129 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1130 /* moving charge context might have finished. */
1133 finish_wait(&mc.waitq, &wait);
1140 #define K(x) ((x) << (PAGE_SHIFT-10))
1142 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1143 * @memcg: The memory cgroup that went over limit
1144 * @p: Task that is going to be killed
1146 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1149 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1151 struct mem_cgroup *iter;
1157 pr_info("Task in ");
1158 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1159 pr_cont(" killed as a result of limit of ");
1161 pr_info("Memory limit reached of cgroup ");
1164 pr_cont_cgroup_path(memcg->css.cgroup);
1169 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1170 K((u64)page_counter_read(&memcg->memory)),
1171 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1172 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1173 K((u64)page_counter_read(&memcg->memsw)),
1174 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1175 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1176 K((u64)page_counter_read(&memcg->kmem)),
1177 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1179 for_each_mem_cgroup_tree(iter, memcg) {
1180 pr_info("Memory cgroup stats for ");
1181 pr_cont_cgroup_path(iter->css.cgroup);
1184 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1185 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1187 pr_cont(" %s:%luKB", mem_cgroup_stat_names[i],
1188 K(mem_cgroup_read_stat(iter, i)));
1191 for (i = 0; i < NR_LRU_LISTS; i++)
1192 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1193 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1200 * This function returns the number of memcg under hierarchy tree. Returns
1201 * 1(self count) if no children.
1203 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1206 struct mem_cgroup *iter;
1208 for_each_mem_cgroup_tree(iter, memcg)
1214 * Return the memory (and swap, if configured) limit for a memcg.
1216 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1218 unsigned long limit;
1220 limit = memcg->memory.limit;
1221 if (mem_cgroup_swappiness(memcg)) {
1222 unsigned long memsw_limit;
1223 unsigned long swap_limit;
1225 memsw_limit = memcg->memsw.limit;
1226 swap_limit = memcg->swap.limit;
1227 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1228 limit = min(limit + swap_limit, memsw_limit);
1233 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1236 struct oom_control oc = {
1240 .gfp_mask = gfp_mask,
1245 mutex_lock(&oom_lock);
1246 ret = out_of_memory(&oc);
1247 mutex_unlock(&oom_lock);
1251 #if MAX_NUMNODES > 1
1254 * test_mem_cgroup_node_reclaimable
1255 * @memcg: the target memcg
1256 * @nid: the node ID to be checked.
1257 * @noswap : specify true here if the user wants flle only information.
1259 * This function returns whether the specified memcg contains any
1260 * reclaimable pages on a node. Returns true if there are any reclaimable
1261 * pages in the node.
1263 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1264 int nid, bool noswap)
1266 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1268 if (noswap || !total_swap_pages)
1270 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1277 * Always updating the nodemask is not very good - even if we have an empty
1278 * list or the wrong list here, we can start from some node and traverse all
1279 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1282 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1286 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1287 * pagein/pageout changes since the last update.
1289 if (!atomic_read(&memcg->numainfo_events))
1291 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1294 /* make a nodemask where this memcg uses memory from */
1295 memcg->scan_nodes = node_states[N_MEMORY];
1297 for_each_node_mask(nid, node_states[N_MEMORY]) {
1299 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1300 node_clear(nid, memcg->scan_nodes);
1303 atomic_set(&memcg->numainfo_events, 0);
1304 atomic_set(&memcg->numainfo_updating, 0);
1308 * Selecting a node where we start reclaim from. Because what we need is just
1309 * reducing usage counter, start from anywhere is O,K. Considering
1310 * memory reclaim from current node, there are pros. and cons.
1312 * Freeing memory from current node means freeing memory from a node which
1313 * we'll use or we've used. So, it may make LRU bad. And if several threads
1314 * hit limits, it will see a contention on a node. But freeing from remote
1315 * node means more costs for memory reclaim because of memory latency.
1317 * Now, we use round-robin. Better algorithm is welcomed.
1319 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1323 mem_cgroup_may_update_nodemask(memcg);
1324 node = memcg->last_scanned_node;
1326 node = next_node_in(node, memcg->scan_nodes);
1328 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1329 * last time it really checked all the LRUs due to rate limiting.
1330 * Fallback to the current node in that case for simplicity.
1332 if (unlikely(node == MAX_NUMNODES))
1333 node = numa_node_id();
1335 memcg->last_scanned_node = node;
1339 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1345 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1348 unsigned long *total_scanned)
1350 struct mem_cgroup *victim = NULL;
1353 unsigned long excess;
1354 unsigned long nr_scanned;
1355 struct mem_cgroup_reclaim_cookie reclaim = {
1360 excess = soft_limit_excess(root_memcg);
1363 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1368 * If we have not been able to reclaim
1369 * anything, it might because there are
1370 * no reclaimable pages under this hierarchy
1375 * We want to do more targeted reclaim.
1376 * excess >> 2 is not to excessive so as to
1377 * reclaim too much, nor too less that we keep
1378 * coming back to reclaim from this cgroup
1380 if (total >= (excess >> 2) ||
1381 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1386 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1387 pgdat, &nr_scanned);
1388 *total_scanned += nr_scanned;
1389 if (!soft_limit_excess(root_memcg))
1392 mem_cgroup_iter_break(root_memcg, victim);
1396 #ifdef CONFIG_LOCKDEP
1397 static struct lockdep_map memcg_oom_lock_dep_map = {
1398 .name = "memcg_oom_lock",
1402 static DEFINE_SPINLOCK(memcg_oom_lock);
1405 * Check OOM-Killer is already running under our hierarchy.
1406 * If someone is running, return false.
1408 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1410 struct mem_cgroup *iter, *failed = NULL;
1412 spin_lock(&memcg_oom_lock);
1414 for_each_mem_cgroup_tree(iter, memcg) {
1415 if (iter->oom_lock) {
1417 * this subtree of our hierarchy is already locked
1418 * so we cannot give a lock.
1421 mem_cgroup_iter_break(memcg, iter);
1424 iter->oom_lock = true;
1429 * OK, we failed to lock the whole subtree so we have
1430 * to clean up what we set up to the failing subtree
1432 for_each_mem_cgroup_tree(iter, memcg) {
1433 if (iter == failed) {
1434 mem_cgroup_iter_break(memcg, iter);
1437 iter->oom_lock = false;
1440 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1442 spin_unlock(&memcg_oom_lock);
1447 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1449 struct mem_cgroup *iter;
1451 spin_lock(&memcg_oom_lock);
1452 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1453 for_each_mem_cgroup_tree(iter, memcg)
1454 iter->oom_lock = false;
1455 spin_unlock(&memcg_oom_lock);
1458 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1460 struct mem_cgroup *iter;
1462 spin_lock(&memcg_oom_lock);
1463 for_each_mem_cgroup_tree(iter, memcg)
1465 spin_unlock(&memcg_oom_lock);
1468 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1470 struct mem_cgroup *iter;
1473 * When a new child is created while the hierarchy is under oom,
1474 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1476 spin_lock(&memcg_oom_lock);
1477 for_each_mem_cgroup_tree(iter, memcg)
1478 if (iter->under_oom > 0)
1480 spin_unlock(&memcg_oom_lock);
1483 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1485 struct oom_wait_info {
1486 struct mem_cgroup *memcg;
1490 static int memcg_oom_wake_function(wait_queue_t *wait,
1491 unsigned mode, int sync, void *arg)
1493 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1494 struct mem_cgroup *oom_wait_memcg;
1495 struct oom_wait_info *oom_wait_info;
1497 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1498 oom_wait_memcg = oom_wait_info->memcg;
1500 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1501 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1503 return autoremove_wake_function(wait, mode, sync, arg);
1506 static void memcg_oom_recover(struct mem_cgroup *memcg)
1509 * For the following lockless ->under_oom test, the only required
1510 * guarantee is that it must see the state asserted by an OOM when
1511 * this function is called as a result of userland actions
1512 * triggered by the notification of the OOM. This is trivially
1513 * achieved by invoking mem_cgroup_mark_under_oom() before
1514 * triggering notification.
1516 if (memcg && memcg->under_oom)
1517 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1520 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1522 if (!current->memcg_may_oom)
1525 * We are in the middle of the charge context here, so we
1526 * don't want to block when potentially sitting on a callstack
1527 * that holds all kinds of filesystem and mm locks.
1529 * Also, the caller may handle a failed allocation gracefully
1530 * (like optional page cache readahead) and so an OOM killer
1531 * invocation might not even be necessary.
1533 * That's why we don't do anything here except remember the
1534 * OOM context and then deal with it at the end of the page
1535 * fault when the stack is unwound, the locks are released,
1536 * and when we know whether the fault was overall successful.
1538 css_get(&memcg->css);
1539 current->memcg_in_oom = memcg;
1540 current->memcg_oom_gfp_mask = mask;
1541 current->memcg_oom_order = order;
1545 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1546 * @handle: actually kill/wait or just clean up the OOM state
1548 * This has to be called at the end of a page fault if the memcg OOM
1549 * handler was enabled.
1551 * Memcg supports userspace OOM handling where failed allocations must
1552 * sleep on a waitqueue until the userspace task resolves the
1553 * situation. Sleeping directly in the charge context with all kinds
1554 * of locks held is not a good idea, instead we remember an OOM state
1555 * in the task and mem_cgroup_oom_synchronize() has to be called at
1556 * the end of the page fault to complete the OOM handling.
1558 * Returns %true if an ongoing memcg OOM situation was detected and
1559 * completed, %false otherwise.
1561 bool mem_cgroup_oom_synchronize(bool handle)
1563 struct mem_cgroup *memcg = current->memcg_in_oom;
1564 struct oom_wait_info owait;
1567 /* OOM is global, do not handle */
1574 owait.memcg = memcg;
1575 owait.wait.flags = 0;
1576 owait.wait.func = memcg_oom_wake_function;
1577 owait.wait.private = current;
1578 INIT_LIST_HEAD(&owait.wait.task_list);
1580 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1581 mem_cgroup_mark_under_oom(memcg);
1583 locked = mem_cgroup_oom_trylock(memcg);
1586 mem_cgroup_oom_notify(memcg);
1588 if (locked && !memcg->oom_kill_disable) {
1589 mem_cgroup_unmark_under_oom(memcg);
1590 finish_wait(&memcg_oom_waitq, &owait.wait);
1591 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1592 current->memcg_oom_order);
1595 mem_cgroup_unmark_under_oom(memcg);
1596 finish_wait(&memcg_oom_waitq, &owait.wait);
1600 mem_cgroup_oom_unlock(memcg);
1602 * There is no guarantee that an OOM-lock contender
1603 * sees the wakeups triggered by the OOM kill
1604 * uncharges. Wake any sleepers explicitely.
1606 memcg_oom_recover(memcg);
1609 current->memcg_in_oom = NULL;
1610 css_put(&memcg->css);
1615 * lock_page_memcg - lock a page->mem_cgroup binding
1618 * This function protects unlocked LRU pages from being moved to
1619 * another cgroup and stabilizes their page->mem_cgroup binding.
1621 void lock_page_memcg(struct page *page)
1623 struct mem_cgroup *memcg;
1624 unsigned long flags;
1627 * The RCU lock is held throughout the transaction. The fast
1628 * path can get away without acquiring the memcg->move_lock
1629 * because page moving starts with an RCU grace period.
1633 if (mem_cgroup_disabled())
1636 memcg = page->mem_cgroup;
1637 if (unlikely(!memcg))
1640 if (atomic_read(&memcg->moving_account) <= 0)
1643 spin_lock_irqsave(&memcg->move_lock, flags);
1644 if (memcg != page->mem_cgroup) {
1645 spin_unlock_irqrestore(&memcg->move_lock, flags);
1650 * When charge migration first begins, we can have locked and
1651 * unlocked page stat updates happening concurrently. Track
1652 * the task who has the lock for unlock_page_memcg().
1654 memcg->move_lock_task = current;
1655 memcg->move_lock_flags = flags;
1659 EXPORT_SYMBOL(lock_page_memcg);
1662 * unlock_page_memcg - unlock a page->mem_cgroup binding
1665 void unlock_page_memcg(struct page *page)
1667 struct mem_cgroup *memcg = page->mem_cgroup;
1669 if (memcg && memcg->move_lock_task == current) {
1670 unsigned long flags = memcg->move_lock_flags;
1672 memcg->move_lock_task = NULL;
1673 memcg->move_lock_flags = 0;
1675 spin_unlock_irqrestore(&memcg->move_lock, flags);
1680 EXPORT_SYMBOL(unlock_page_memcg);
1683 * size of first charge trial. "32" comes from vmscan.c's magic value.
1684 * TODO: maybe necessary to use big numbers in big irons.
1686 #define CHARGE_BATCH 32U
1687 struct memcg_stock_pcp {
1688 struct mem_cgroup *cached; /* this never be root cgroup */
1689 unsigned int nr_pages;
1690 struct work_struct work;
1691 unsigned long flags;
1692 #define FLUSHING_CACHED_CHARGE 0
1694 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1695 static DEFINE_MUTEX(percpu_charge_mutex);
1698 * consume_stock: Try to consume stocked charge on this cpu.
1699 * @memcg: memcg to consume from.
1700 * @nr_pages: how many pages to charge.
1702 * The charges will only happen if @memcg matches the current cpu's memcg
1703 * stock, and at least @nr_pages are available in that stock. Failure to
1704 * service an allocation will refill the stock.
1706 * returns true if successful, false otherwise.
1708 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1710 struct memcg_stock_pcp *stock;
1711 unsigned long flags;
1714 if (nr_pages > CHARGE_BATCH)
1717 local_irq_save(flags);
1719 stock = this_cpu_ptr(&memcg_stock);
1720 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1721 stock->nr_pages -= nr_pages;
1725 local_irq_restore(flags);
1731 * Returns stocks cached in percpu and reset cached information.
1733 static void drain_stock(struct memcg_stock_pcp *stock)
1735 struct mem_cgroup *old = stock->cached;
1737 if (stock->nr_pages) {
1738 page_counter_uncharge(&old->memory, stock->nr_pages);
1739 if (do_memsw_account())
1740 page_counter_uncharge(&old->memsw, stock->nr_pages);
1741 css_put_many(&old->css, stock->nr_pages);
1742 stock->nr_pages = 0;
1744 stock->cached = NULL;
1747 static void drain_local_stock(struct work_struct *dummy)
1749 struct memcg_stock_pcp *stock;
1750 unsigned long flags;
1752 local_irq_save(flags);
1754 stock = this_cpu_ptr(&memcg_stock);
1756 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1758 local_irq_restore(flags);
1762 * Cache charges(val) to local per_cpu area.
1763 * This will be consumed by consume_stock() function, later.
1765 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1767 struct memcg_stock_pcp *stock;
1768 unsigned long flags;
1770 local_irq_save(flags);
1772 stock = this_cpu_ptr(&memcg_stock);
1773 if (stock->cached != memcg) { /* reset if necessary */
1775 stock->cached = memcg;
1777 stock->nr_pages += nr_pages;
1779 local_irq_restore(flags);
1783 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1784 * of the hierarchy under it.
1786 static void drain_all_stock(struct mem_cgroup *root_memcg)
1790 /* If someone's already draining, avoid adding running more workers. */
1791 if (!mutex_trylock(&percpu_charge_mutex))
1793 /* Notify other cpus that system-wide "drain" is running */
1796 for_each_online_cpu(cpu) {
1797 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1798 struct mem_cgroup *memcg;
1800 memcg = stock->cached;
1801 if (!memcg || !stock->nr_pages)
1803 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1805 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1807 drain_local_stock(&stock->work);
1809 schedule_work_on(cpu, &stock->work);
1814 mutex_unlock(&percpu_charge_mutex);
1817 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
1818 unsigned long action,
1821 int cpu = (unsigned long)hcpu;
1822 struct memcg_stock_pcp *stock;
1824 if (action == CPU_ONLINE)
1827 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
1830 stock = &per_cpu(memcg_stock, cpu);
1835 static void reclaim_high(struct mem_cgroup *memcg,
1836 unsigned int nr_pages,
1840 if (page_counter_read(&memcg->memory) <= memcg->high)
1842 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
1843 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1844 } while ((memcg = parent_mem_cgroup(memcg)));
1847 static void high_work_func(struct work_struct *work)
1849 struct mem_cgroup *memcg;
1851 memcg = container_of(work, struct mem_cgroup, high_work);
1852 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1856 * Scheduled by try_charge() to be executed from the userland return path
1857 * and reclaims memory over the high limit.
1859 void mem_cgroup_handle_over_high(void)
1861 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1862 struct mem_cgroup *memcg;
1864 if (likely(!nr_pages))
1867 memcg = get_mem_cgroup_from_mm(current->mm);
1868 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1869 css_put(&memcg->css);
1870 current->memcg_nr_pages_over_high = 0;
1873 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1874 unsigned int nr_pages)
1876 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1877 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1878 struct mem_cgroup *mem_over_limit;
1879 struct page_counter *counter;
1880 unsigned long nr_reclaimed;
1881 bool may_swap = true;
1882 bool drained = false;
1884 if (mem_cgroup_is_root(memcg))
1887 if (consume_stock(memcg, nr_pages))
1890 if (!do_memsw_account() ||
1891 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1892 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1894 if (do_memsw_account())
1895 page_counter_uncharge(&memcg->memsw, batch);
1896 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1898 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1902 if (batch > nr_pages) {
1908 * Unlike in global OOM situations, memcg is not in a physical
1909 * memory shortage. Allow dying and OOM-killed tasks to
1910 * bypass the last charges so that they can exit quickly and
1911 * free their memory.
1913 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
1914 fatal_signal_pending(current) ||
1915 current->flags & PF_EXITING))
1919 * Prevent unbounded recursion when reclaim operations need to
1920 * allocate memory. This might exceed the limits temporarily,
1921 * but we prefer facilitating memory reclaim and getting back
1922 * under the limit over triggering OOM kills in these cases.
1924 if (unlikely(current->flags & PF_MEMALLOC))
1927 if (unlikely(task_in_memcg_oom(current)))
1930 if (!gfpflags_allow_blocking(gfp_mask))
1933 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
1935 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1936 gfp_mask, may_swap);
1938 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1942 drain_all_stock(mem_over_limit);
1947 if (gfp_mask & __GFP_NORETRY)
1950 * Even though the limit is exceeded at this point, reclaim
1951 * may have been able to free some pages. Retry the charge
1952 * before killing the task.
1954 * Only for regular pages, though: huge pages are rather
1955 * unlikely to succeed so close to the limit, and we fall back
1956 * to regular pages anyway in case of failure.
1958 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1961 * At task move, charge accounts can be doubly counted. So, it's
1962 * better to wait until the end of task_move if something is going on.
1964 if (mem_cgroup_wait_acct_move(mem_over_limit))
1970 if (gfp_mask & __GFP_NOFAIL)
1973 if (fatal_signal_pending(current))
1976 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
1978 mem_cgroup_oom(mem_over_limit, gfp_mask,
1979 get_order(nr_pages * PAGE_SIZE));
1981 if (!(gfp_mask & __GFP_NOFAIL))
1985 * The allocation either can't fail or will lead to more memory
1986 * being freed very soon. Allow memory usage go over the limit
1987 * temporarily by force charging it.
1989 page_counter_charge(&memcg->memory, nr_pages);
1990 if (do_memsw_account())
1991 page_counter_charge(&memcg->memsw, nr_pages);
1992 css_get_many(&memcg->css, nr_pages);
1997 css_get_many(&memcg->css, batch);
1998 if (batch > nr_pages)
1999 refill_stock(memcg, batch - nr_pages);
2002 * If the hierarchy is above the normal consumption range, schedule
2003 * reclaim on returning to userland. We can perform reclaim here
2004 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2005 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2006 * not recorded as it most likely matches current's and won't
2007 * change in the meantime. As high limit is checked again before
2008 * reclaim, the cost of mismatch is negligible.
2011 if (page_counter_read(&memcg->memory) > memcg->high) {
2012 /* Don't bother a random interrupted task */
2013 if (in_interrupt()) {
2014 schedule_work(&memcg->high_work);
2017 current->memcg_nr_pages_over_high += batch;
2018 set_notify_resume(current);
2021 } while ((memcg = parent_mem_cgroup(memcg)));
2026 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2028 if (mem_cgroup_is_root(memcg))
2031 page_counter_uncharge(&memcg->memory, nr_pages);
2032 if (do_memsw_account())
2033 page_counter_uncharge(&memcg->memsw, nr_pages);
2035 css_put_many(&memcg->css, nr_pages);
2038 static void lock_page_lru(struct page *page, int *isolated)
2040 struct zone *zone = page_zone(page);
2042 spin_lock_irq(zone_lru_lock(zone));
2043 if (PageLRU(page)) {
2044 struct lruvec *lruvec;
2046 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2048 del_page_from_lru_list(page, lruvec, page_lru(page));
2054 static void unlock_page_lru(struct page *page, int isolated)
2056 struct zone *zone = page_zone(page);
2059 struct lruvec *lruvec;
2061 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2062 VM_BUG_ON_PAGE(PageLRU(page), page);
2064 add_page_to_lru_list(page, lruvec, page_lru(page));
2066 spin_unlock_irq(zone_lru_lock(zone));
2069 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2074 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2077 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2078 * may already be on some other mem_cgroup's LRU. Take care of it.
2081 lock_page_lru(page, &isolated);
2084 * Nobody should be changing or seriously looking at
2085 * page->mem_cgroup at this point:
2087 * - the page is uncharged
2089 * - the page is off-LRU
2091 * - an anonymous fault has exclusive page access, except for
2092 * a locked page table
2094 * - a page cache insertion, a swapin fault, or a migration
2095 * have the page locked
2097 page->mem_cgroup = memcg;
2100 unlock_page_lru(page, isolated);
2104 static int memcg_alloc_cache_id(void)
2109 id = ida_simple_get(&memcg_cache_ida,
2110 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2114 if (id < memcg_nr_cache_ids)
2118 * There's no space for the new id in memcg_caches arrays,
2119 * so we have to grow them.
2121 down_write(&memcg_cache_ids_sem);
2123 size = 2 * (id + 1);
2124 if (size < MEMCG_CACHES_MIN_SIZE)
2125 size = MEMCG_CACHES_MIN_SIZE;
2126 else if (size > MEMCG_CACHES_MAX_SIZE)
2127 size = MEMCG_CACHES_MAX_SIZE;
2129 err = memcg_update_all_caches(size);
2131 err = memcg_update_all_list_lrus(size);
2133 memcg_nr_cache_ids = size;
2135 up_write(&memcg_cache_ids_sem);
2138 ida_simple_remove(&memcg_cache_ida, id);
2144 static void memcg_free_cache_id(int id)
2146 ida_simple_remove(&memcg_cache_ida, id);
2149 struct memcg_kmem_cache_create_work {
2150 struct mem_cgroup *memcg;
2151 struct kmem_cache *cachep;
2152 struct work_struct work;
2155 static void memcg_kmem_cache_create_func(struct work_struct *w)
2157 struct memcg_kmem_cache_create_work *cw =
2158 container_of(w, struct memcg_kmem_cache_create_work, work);
2159 struct mem_cgroup *memcg = cw->memcg;
2160 struct kmem_cache *cachep = cw->cachep;
2162 memcg_create_kmem_cache(memcg, cachep);
2164 css_put(&memcg->css);
2169 * Enqueue the creation of a per-memcg kmem_cache.
2171 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2172 struct kmem_cache *cachep)
2174 struct memcg_kmem_cache_create_work *cw;
2176 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2180 css_get(&memcg->css);
2183 cw->cachep = cachep;
2184 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2186 schedule_work(&cw->work);
2189 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2190 struct kmem_cache *cachep)
2193 * We need to stop accounting when we kmalloc, because if the
2194 * corresponding kmalloc cache is not yet created, the first allocation
2195 * in __memcg_schedule_kmem_cache_create will recurse.
2197 * However, it is better to enclose the whole function. Depending on
2198 * the debugging options enabled, INIT_WORK(), for instance, can
2199 * trigger an allocation. This too, will make us recurse. Because at
2200 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2201 * the safest choice is to do it like this, wrapping the whole function.
2203 current->memcg_kmem_skip_account = 1;
2204 __memcg_schedule_kmem_cache_create(memcg, cachep);
2205 current->memcg_kmem_skip_account = 0;
2208 static inline bool memcg_kmem_bypass(void)
2210 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2216 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2217 * @cachep: the original global kmem cache
2219 * Return the kmem_cache we're supposed to use for a slab allocation.
2220 * We try to use the current memcg's version of the cache.
2222 * If the cache does not exist yet, if we are the first user of it, we
2223 * create it asynchronously in a workqueue and let the current allocation
2224 * go through with the original cache.
2226 * This function takes a reference to the cache it returns to assure it
2227 * won't get destroyed while we are working with it. Once the caller is
2228 * done with it, memcg_kmem_put_cache() must be called to release the
2231 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2233 struct mem_cgroup *memcg;
2234 struct kmem_cache *memcg_cachep;
2237 VM_BUG_ON(!is_root_cache(cachep));
2239 if (memcg_kmem_bypass())
2242 if (current->memcg_kmem_skip_account)
2245 memcg = get_mem_cgroup_from_mm(current->mm);
2246 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2250 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2251 if (likely(memcg_cachep))
2252 return memcg_cachep;
2255 * If we are in a safe context (can wait, and not in interrupt
2256 * context), we could be be predictable and return right away.
2257 * This would guarantee that the allocation being performed
2258 * already belongs in the new cache.
2260 * However, there are some clashes that can arrive from locking.
2261 * For instance, because we acquire the slab_mutex while doing
2262 * memcg_create_kmem_cache, this means no further allocation
2263 * could happen with the slab_mutex held. So it's better to
2266 memcg_schedule_kmem_cache_create(memcg, cachep);
2268 css_put(&memcg->css);
2273 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2274 * @cachep: the cache returned by memcg_kmem_get_cache
2276 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2278 if (!is_root_cache(cachep))
2279 css_put(&cachep->memcg_params.memcg->css);
2283 * memcg_kmem_charge: charge a kmem page
2284 * @page: page to charge
2285 * @gfp: reclaim mode
2286 * @order: allocation order
2287 * @memcg: memory cgroup to charge
2289 * Returns 0 on success, an error code on failure.
2291 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2292 struct mem_cgroup *memcg)
2294 unsigned int nr_pages = 1 << order;
2295 struct page_counter *counter;
2298 ret = try_charge(memcg, gfp, nr_pages);
2302 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2303 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2304 cancel_charge(memcg, nr_pages);
2308 page->mem_cgroup = memcg;
2314 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2315 * @page: page to charge
2316 * @gfp: reclaim mode
2317 * @order: allocation order
2319 * Returns 0 on success, an error code on failure.
2321 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2323 struct mem_cgroup *memcg;
2326 if (memcg_kmem_bypass())
2329 memcg = get_mem_cgroup_from_mm(current->mm);
2330 if (!mem_cgroup_is_root(memcg)) {
2331 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2333 __SetPageKmemcg(page);
2335 css_put(&memcg->css);
2339 * memcg_kmem_uncharge: uncharge a kmem page
2340 * @page: page to uncharge
2341 * @order: allocation order
2343 void memcg_kmem_uncharge(struct page *page, int order)
2345 struct mem_cgroup *memcg = page->mem_cgroup;
2346 unsigned int nr_pages = 1 << order;
2351 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2353 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2354 page_counter_uncharge(&memcg->kmem, nr_pages);
2356 page_counter_uncharge(&memcg->memory, nr_pages);
2357 if (do_memsw_account())
2358 page_counter_uncharge(&memcg->memsw, nr_pages);
2360 page->mem_cgroup = NULL;
2362 /* slab pages do not have PageKmemcg flag set */
2363 if (PageKmemcg(page))
2364 __ClearPageKmemcg(page);
2366 css_put_many(&memcg->css, nr_pages);
2368 #endif /* !CONFIG_SLOB */
2370 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2373 * Because tail pages are not marked as "used", set it. We're under
2374 * zone_lru_lock and migration entries setup in all page mappings.
2376 void mem_cgroup_split_huge_fixup(struct page *head)
2380 if (mem_cgroup_disabled())
2383 for (i = 1; i < HPAGE_PMD_NR; i++)
2384 head[i].mem_cgroup = head->mem_cgroup;
2386 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2389 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2391 #ifdef CONFIG_MEMCG_SWAP
2392 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2395 int val = (charge) ? 1 : -1;
2396 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2400 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2401 * @entry: swap entry to be moved
2402 * @from: mem_cgroup which the entry is moved from
2403 * @to: mem_cgroup which the entry is moved to
2405 * It succeeds only when the swap_cgroup's record for this entry is the same
2406 * as the mem_cgroup's id of @from.
2408 * Returns 0 on success, -EINVAL on failure.
2410 * The caller must have charged to @to, IOW, called page_counter_charge() about
2411 * both res and memsw, and called css_get().
2413 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2414 struct mem_cgroup *from, struct mem_cgroup *to)
2416 unsigned short old_id, new_id;
2418 old_id = mem_cgroup_id(from);
2419 new_id = mem_cgroup_id(to);
2421 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2422 mem_cgroup_swap_statistics(from, false);
2423 mem_cgroup_swap_statistics(to, true);
2429 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2430 struct mem_cgroup *from, struct mem_cgroup *to)
2436 static DEFINE_MUTEX(memcg_limit_mutex);
2438 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2439 unsigned long limit)
2441 unsigned long curusage;
2442 unsigned long oldusage;
2443 bool enlarge = false;
2448 * For keeping hierarchical_reclaim simple, how long we should retry
2449 * is depends on callers. We set our retry-count to be function
2450 * of # of children which we should visit in this loop.
2452 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2453 mem_cgroup_count_children(memcg);
2455 oldusage = page_counter_read(&memcg->memory);
2458 if (signal_pending(current)) {
2463 mutex_lock(&memcg_limit_mutex);
2464 if (limit > memcg->memsw.limit) {
2465 mutex_unlock(&memcg_limit_mutex);
2469 if (limit > memcg->memory.limit)
2471 ret = page_counter_limit(&memcg->memory, limit);
2472 mutex_unlock(&memcg_limit_mutex);
2477 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2479 curusage = page_counter_read(&memcg->memory);
2480 /* Usage is reduced ? */
2481 if (curusage >= oldusage)
2484 oldusage = curusage;
2485 } while (retry_count);
2487 if (!ret && enlarge)
2488 memcg_oom_recover(memcg);
2493 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2494 unsigned long limit)
2496 unsigned long curusage;
2497 unsigned long oldusage;
2498 bool enlarge = false;
2502 /* see mem_cgroup_resize_res_limit */
2503 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2504 mem_cgroup_count_children(memcg);
2506 oldusage = page_counter_read(&memcg->memsw);
2509 if (signal_pending(current)) {
2514 mutex_lock(&memcg_limit_mutex);
2515 if (limit < memcg->memory.limit) {
2516 mutex_unlock(&memcg_limit_mutex);
2520 if (limit > memcg->memsw.limit)
2522 ret = page_counter_limit(&memcg->memsw, limit);
2523 mutex_unlock(&memcg_limit_mutex);
2528 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2530 curusage = page_counter_read(&memcg->memsw);
2531 /* Usage is reduced ? */
2532 if (curusage >= oldusage)
2535 oldusage = curusage;
2536 } while (retry_count);
2538 if (!ret && enlarge)
2539 memcg_oom_recover(memcg);
2544 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2546 unsigned long *total_scanned)
2548 unsigned long nr_reclaimed = 0;
2549 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2550 unsigned long reclaimed;
2552 struct mem_cgroup_tree_per_node *mctz;
2553 unsigned long excess;
2554 unsigned long nr_scanned;
2559 mctz = soft_limit_tree_node(pgdat->node_id);
2562 * Do not even bother to check the largest node if the root
2563 * is empty. Do it lockless to prevent lock bouncing. Races
2564 * are acceptable as soft limit is best effort anyway.
2566 if (RB_EMPTY_ROOT(&mctz->rb_root))
2570 * This loop can run a while, specially if mem_cgroup's continuously
2571 * keep exceeding their soft limit and putting the system under
2578 mz = mem_cgroup_largest_soft_limit_node(mctz);
2583 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2584 gfp_mask, &nr_scanned);
2585 nr_reclaimed += reclaimed;
2586 *total_scanned += nr_scanned;
2587 spin_lock_irq(&mctz->lock);
2588 __mem_cgroup_remove_exceeded(mz, mctz);
2591 * If we failed to reclaim anything from this memory cgroup
2592 * it is time to move on to the next cgroup
2596 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2598 excess = soft_limit_excess(mz->memcg);
2600 * One school of thought says that we should not add
2601 * back the node to the tree if reclaim returns 0.
2602 * But our reclaim could return 0, simply because due
2603 * to priority we are exposing a smaller subset of
2604 * memory to reclaim from. Consider this as a longer
2607 /* If excess == 0, no tree ops */
2608 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2609 spin_unlock_irq(&mctz->lock);
2610 css_put(&mz->memcg->css);
2613 * Could not reclaim anything and there are no more
2614 * mem cgroups to try or we seem to be looping without
2615 * reclaiming anything.
2617 if (!nr_reclaimed &&
2619 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2621 } while (!nr_reclaimed);
2623 css_put(&next_mz->memcg->css);
2624 return nr_reclaimed;
2628 * Test whether @memcg has children, dead or alive. Note that this
2629 * function doesn't care whether @memcg has use_hierarchy enabled and
2630 * returns %true if there are child csses according to the cgroup
2631 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2633 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2638 ret = css_next_child(NULL, &memcg->css);
2644 * Reclaims as many pages from the given memcg as possible.
2646 * Caller is responsible for holding css reference for memcg.
2648 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2650 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2652 /* we call try-to-free pages for make this cgroup empty */
2653 lru_add_drain_all();
2654 /* try to free all pages in this cgroup */
2655 while (nr_retries && page_counter_read(&memcg->memory)) {
2658 if (signal_pending(current))
2661 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2665 /* maybe some writeback is necessary */
2666 congestion_wait(BLK_RW_ASYNC, HZ/10);
2674 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2675 char *buf, size_t nbytes,
2678 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2680 if (mem_cgroup_is_root(memcg))
2682 return mem_cgroup_force_empty(memcg) ?: nbytes;
2685 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2688 return mem_cgroup_from_css(css)->use_hierarchy;
2691 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2692 struct cftype *cft, u64 val)
2695 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2696 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2698 if (memcg->use_hierarchy == val)
2702 * If parent's use_hierarchy is set, we can't make any modifications
2703 * in the child subtrees. If it is unset, then the change can
2704 * occur, provided the current cgroup has no children.
2706 * For the root cgroup, parent_mem is NULL, we allow value to be
2707 * set if there are no children.
2709 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2710 (val == 1 || val == 0)) {
2711 if (!memcg_has_children(memcg))
2712 memcg->use_hierarchy = val;
2721 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2723 struct mem_cgroup *iter;
2726 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2728 for_each_mem_cgroup_tree(iter, memcg) {
2729 for (i = 0; i < MEMCG_NR_STAT; i++)
2730 stat[i] += mem_cgroup_read_stat(iter, i);
2734 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2736 struct mem_cgroup *iter;
2739 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2741 for_each_mem_cgroup_tree(iter, memcg) {
2742 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2743 events[i] += mem_cgroup_read_events(iter, i);
2747 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2749 unsigned long val = 0;
2751 if (mem_cgroup_is_root(memcg)) {
2752 struct mem_cgroup *iter;
2754 for_each_mem_cgroup_tree(iter, memcg) {
2755 val += mem_cgroup_read_stat(iter,
2756 MEM_CGROUP_STAT_CACHE);
2757 val += mem_cgroup_read_stat(iter,
2758 MEM_CGROUP_STAT_RSS);
2760 val += mem_cgroup_read_stat(iter,
2761 MEM_CGROUP_STAT_SWAP);
2765 val = page_counter_read(&memcg->memory);
2767 val = page_counter_read(&memcg->memsw);
2780 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2783 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2784 struct page_counter *counter;
2786 switch (MEMFILE_TYPE(cft->private)) {
2788 counter = &memcg->memory;
2791 counter = &memcg->memsw;
2794 counter = &memcg->kmem;
2797 counter = &memcg->tcpmem;
2803 switch (MEMFILE_ATTR(cft->private)) {
2805 if (counter == &memcg->memory)
2806 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2807 if (counter == &memcg->memsw)
2808 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2809 return (u64)page_counter_read(counter) * PAGE_SIZE;
2811 return (u64)counter->limit * PAGE_SIZE;
2813 return (u64)counter->watermark * PAGE_SIZE;
2815 return counter->failcnt;
2816 case RES_SOFT_LIMIT:
2817 return (u64)memcg->soft_limit * PAGE_SIZE;
2824 static int memcg_online_kmem(struct mem_cgroup *memcg)
2828 if (cgroup_memory_nokmem)
2831 BUG_ON(memcg->kmemcg_id >= 0);
2832 BUG_ON(memcg->kmem_state);
2834 memcg_id = memcg_alloc_cache_id();
2838 static_branch_inc(&memcg_kmem_enabled_key);
2840 * A memory cgroup is considered kmem-online as soon as it gets
2841 * kmemcg_id. Setting the id after enabling static branching will
2842 * guarantee no one starts accounting before all call sites are
2845 memcg->kmemcg_id = memcg_id;
2846 memcg->kmem_state = KMEM_ONLINE;
2851 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2853 struct cgroup_subsys_state *css;
2854 struct mem_cgroup *parent, *child;
2857 if (memcg->kmem_state != KMEM_ONLINE)
2860 * Clear the online state before clearing memcg_caches array
2861 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2862 * guarantees that no cache will be created for this cgroup
2863 * after we are done (see memcg_create_kmem_cache()).
2865 memcg->kmem_state = KMEM_ALLOCATED;
2867 memcg_deactivate_kmem_caches(memcg);
2869 kmemcg_id = memcg->kmemcg_id;
2870 BUG_ON(kmemcg_id < 0);
2872 parent = parent_mem_cgroup(memcg);
2874 parent = root_mem_cgroup;
2877 * Change kmemcg_id of this cgroup and all its descendants to the
2878 * parent's id, and then move all entries from this cgroup's list_lrus
2879 * to ones of the parent. After we have finished, all list_lrus
2880 * corresponding to this cgroup are guaranteed to remain empty. The
2881 * ordering is imposed by list_lru_node->lock taken by
2882 * memcg_drain_all_list_lrus().
2884 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2885 css_for_each_descendant_pre(css, &memcg->css) {
2886 child = mem_cgroup_from_css(css);
2887 BUG_ON(child->kmemcg_id != kmemcg_id);
2888 child->kmemcg_id = parent->kmemcg_id;
2889 if (!memcg->use_hierarchy)
2894 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2896 memcg_free_cache_id(kmemcg_id);
2899 static void memcg_free_kmem(struct mem_cgroup *memcg)
2901 /* css_alloc() failed, offlining didn't happen */
2902 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2903 memcg_offline_kmem(memcg);
2905 if (memcg->kmem_state == KMEM_ALLOCATED) {
2906 memcg_destroy_kmem_caches(memcg);
2907 static_branch_dec(&memcg_kmem_enabled_key);
2908 WARN_ON(page_counter_read(&memcg->kmem));
2912 static int memcg_online_kmem(struct mem_cgroup *memcg)
2916 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2919 static void memcg_free_kmem(struct mem_cgroup *memcg)
2922 #endif /* !CONFIG_SLOB */
2924 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2925 unsigned long limit)
2929 mutex_lock(&memcg_limit_mutex);
2930 ret = page_counter_limit(&memcg->kmem, limit);
2931 mutex_unlock(&memcg_limit_mutex);
2935 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2939 mutex_lock(&memcg_limit_mutex);
2941 ret = page_counter_limit(&memcg->tcpmem, limit);
2945 if (!memcg->tcpmem_active) {
2947 * The active flag needs to be written after the static_key
2948 * update. This is what guarantees that the socket activation
2949 * function is the last one to run. See mem_cgroup_sk_alloc()
2950 * for details, and note that we don't mark any socket as
2951 * belonging to this memcg until that flag is up.
2953 * We need to do this, because static_keys will span multiple
2954 * sites, but we can't control their order. If we mark a socket
2955 * as accounted, but the accounting functions are not patched in
2956 * yet, we'll lose accounting.
2958 * We never race with the readers in mem_cgroup_sk_alloc(),
2959 * because when this value change, the code to process it is not
2962 static_branch_inc(&memcg_sockets_enabled_key);
2963 memcg->tcpmem_active = true;
2966 mutex_unlock(&memcg_limit_mutex);
2971 * The user of this function is...
2974 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2975 char *buf, size_t nbytes, loff_t off)
2977 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2978 unsigned long nr_pages;
2981 buf = strstrip(buf);
2982 ret = page_counter_memparse(buf, "-1", &nr_pages);
2986 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2988 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2992 switch (MEMFILE_TYPE(of_cft(of)->private)) {
2994 ret = mem_cgroup_resize_limit(memcg, nr_pages);
2997 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3000 ret = memcg_update_kmem_limit(memcg, nr_pages);
3003 ret = memcg_update_tcp_limit(memcg, nr_pages);
3007 case RES_SOFT_LIMIT:
3008 memcg->soft_limit = nr_pages;
3012 return ret ?: nbytes;
3015 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3016 size_t nbytes, loff_t off)
3018 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3019 struct page_counter *counter;
3021 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3023 counter = &memcg->memory;
3026 counter = &memcg->memsw;
3029 counter = &memcg->kmem;
3032 counter = &memcg->tcpmem;
3038 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3040 page_counter_reset_watermark(counter);
3043 counter->failcnt = 0;
3052 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3055 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3059 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3060 struct cftype *cft, u64 val)
3062 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3064 if (val & ~MOVE_MASK)
3068 * No kind of locking is needed in here, because ->can_attach() will
3069 * check this value once in the beginning of the process, and then carry
3070 * on with stale data. This means that changes to this value will only
3071 * affect task migrations starting after the change.
3073 memcg->move_charge_at_immigrate = val;
3077 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3078 struct cftype *cft, u64 val)
3085 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3089 unsigned int lru_mask;
3092 static const struct numa_stat stats[] = {
3093 { "total", LRU_ALL },
3094 { "file", LRU_ALL_FILE },
3095 { "anon", LRU_ALL_ANON },
3096 { "unevictable", BIT(LRU_UNEVICTABLE) },
3098 const struct numa_stat *stat;
3101 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3103 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3104 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3105 seq_printf(m, "%s=%lu", stat->name, nr);
3106 for_each_node_state(nid, N_MEMORY) {
3107 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3109 seq_printf(m, " N%d=%lu", nid, nr);
3114 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3115 struct mem_cgroup *iter;
3118 for_each_mem_cgroup_tree(iter, memcg)
3119 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3120 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3121 for_each_node_state(nid, N_MEMORY) {
3123 for_each_mem_cgroup_tree(iter, memcg)
3124 nr += mem_cgroup_node_nr_lru_pages(
3125 iter, nid, stat->lru_mask);
3126 seq_printf(m, " N%d=%lu", nid, nr);
3133 #endif /* CONFIG_NUMA */
3135 static int memcg_stat_show(struct seq_file *m, void *v)
3137 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3138 unsigned long memory, memsw;
3139 struct mem_cgroup *mi;
3142 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3143 MEM_CGROUP_STAT_NSTATS);
3144 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3145 MEM_CGROUP_EVENTS_NSTATS);
3146 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3148 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3149 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3151 seq_printf(m, "%s %lu\n", mem_cgroup_stat_names[i],
3152 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3155 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3156 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3157 mem_cgroup_read_events(memcg, i));
3159 for (i = 0; i < NR_LRU_LISTS; i++)
3160 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3161 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3163 /* Hierarchical information */
3164 memory = memsw = PAGE_COUNTER_MAX;
3165 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3166 memory = min(memory, mi->memory.limit);
3167 memsw = min(memsw, mi->memsw.limit);
3169 seq_printf(m, "hierarchical_memory_limit %llu\n",
3170 (u64)memory * PAGE_SIZE);
3171 if (do_memsw_account())
3172 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3173 (u64)memsw * PAGE_SIZE);
3175 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3176 unsigned long long val = 0;
3178 if (i == MEM_CGROUP_STAT_SWAP && !do_memsw_account())
3180 for_each_mem_cgroup_tree(mi, memcg)
3181 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3182 seq_printf(m, "total_%s %llu\n", mem_cgroup_stat_names[i], val);
3185 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3186 unsigned long long val = 0;
3188 for_each_mem_cgroup_tree(mi, memcg)
3189 val += mem_cgroup_read_events(mi, i);
3190 seq_printf(m, "total_%s %llu\n",
3191 mem_cgroup_events_names[i], val);
3194 for (i = 0; i < NR_LRU_LISTS; i++) {
3195 unsigned long long val = 0;
3197 for_each_mem_cgroup_tree(mi, memcg)
3198 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3199 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3202 #ifdef CONFIG_DEBUG_VM
3205 struct mem_cgroup_per_node *mz;
3206 struct zone_reclaim_stat *rstat;
3207 unsigned long recent_rotated[2] = {0, 0};
3208 unsigned long recent_scanned[2] = {0, 0};
3210 for_each_online_pgdat(pgdat) {
3211 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3212 rstat = &mz->lruvec.reclaim_stat;
3214 recent_rotated[0] += rstat->recent_rotated[0];
3215 recent_rotated[1] += rstat->recent_rotated[1];
3216 recent_scanned[0] += rstat->recent_scanned[0];
3217 recent_scanned[1] += rstat->recent_scanned[1];
3219 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3220 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3221 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3222 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3229 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3232 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3234 return mem_cgroup_swappiness(memcg);
3237 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3238 struct cftype *cft, u64 val)
3240 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3246 memcg->swappiness = val;
3248 vm_swappiness = val;
3253 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3255 struct mem_cgroup_threshold_ary *t;
3256 unsigned long usage;
3261 t = rcu_dereference(memcg->thresholds.primary);
3263 t = rcu_dereference(memcg->memsw_thresholds.primary);
3268 usage = mem_cgroup_usage(memcg, swap);
3271 * current_threshold points to threshold just below or equal to usage.
3272 * If it's not true, a threshold was crossed after last
3273 * call of __mem_cgroup_threshold().
3275 i = t->current_threshold;
3278 * Iterate backward over array of thresholds starting from
3279 * current_threshold and check if a threshold is crossed.
3280 * If none of thresholds below usage is crossed, we read
3281 * only one element of the array here.
3283 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3284 eventfd_signal(t->entries[i].eventfd, 1);
3286 /* i = current_threshold + 1 */
3290 * Iterate forward over array of thresholds starting from
3291 * current_threshold+1 and check if a threshold is crossed.
3292 * If none of thresholds above usage is crossed, we read
3293 * only one element of the array here.
3295 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3296 eventfd_signal(t->entries[i].eventfd, 1);
3298 /* Update current_threshold */
3299 t->current_threshold = i - 1;
3304 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3307 __mem_cgroup_threshold(memcg, false);
3308 if (do_memsw_account())
3309 __mem_cgroup_threshold(memcg, true);
3311 memcg = parent_mem_cgroup(memcg);
3315 static int compare_thresholds(const void *a, const void *b)
3317 const struct mem_cgroup_threshold *_a = a;
3318 const struct mem_cgroup_threshold *_b = b;
3320 if (_a->threshold > _b->threshold)
3323 if (_a->threshold < _b->threshold)
3329 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3331 struct mem_cgroup_eventfd_list *ev;
3333 spin_lock(&memcg_oom_lock);
3335 list_for_each_entry(ev, &memcg->oom_notify, list)
3336 eventfd_signal(ev->eventfd, 1);
3338 spin_unlock(&memcg_oom_lock);
3342 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3344 struct mem_cgroup *iter;
3346 for_each_mem_cgroup_tree(iter, memcg)
3347 mem_cgroup_oom_notify_cb(iter);
3350 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3351 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3353 struct mem_cgroup_thresholds *thresholds;
3354 struct mem_cgroup_threshold_ary *new;
3355 unsigned long threshold;
3356 unsigned long usage;
3359 ret = page_counter_memparse(args, "-1", &threshold);
3363 mutex_lock(&memcg->thresholds_lock);
3366 thresholds = &memcg->thresholds;
3367 usage = mem_cgroup_usage(memcg, false);
3368 } else if (type == _MEMSWAP) {
3369 thresholds = &memcg->memsw_thresholds;
3370 usage = mem_cgroup_usage(memcg, true);
3374 /* Check if a threshold crossed before adding a new one */
3375 if (thresholds->primary)
3376 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3378 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3380 /* Allocate memory for new array of thresholds */
3381 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3389 /* Copy thresholds (if any) to new array */
3390 if (thresholds->primary) {
3391 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3392 sizeof(struct mem_cgroup_threshold));
3395 /* Add new threshold */
3396 new->entries[size - 1].eventfd = eventfd;
3397 new->entries[size - 1].threshold = threshold;
3399 /* Sort thresholds. Registering of new threshold isn't time-critical */
3400 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3401 compare_thresholds, NULL);
3403 /* Find current threshold */
3404 new->current_threshold = -1;
3405 for (i = 0; i < size; i++) {
3406 if (new->entries[i].threshold <= usage) {
3408 * new->current_threshold will not be used until
3409 * rcu_assign_pointer(), so it's safe to increment
3412 ++new->current_threshold;
3417 /* Free old spare buffer and save old primary buffer as spare */
3418 kfree(thresholds->spare);
3419 thresholds->spare = thresholds->primary;
3421 rcu_assign_pointer(thresholds->primary, new);
3423 /* To be sure that nobody uses thresholds */
3427 mutex_unlock(&memcg->thresholds_lock);
3432 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3433 struct eventfd_ctx *eventfd, const char *args)
3435 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3438 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3439 struct eventfd_ctx *eventfd, const char *args)
3441 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3444 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3445 struct eventfd_ctx *eventfd, enum res_type type)
3447 struct mem_cgroup_thresholds *thresholds;
3448 struct mem_cgroup_threshold_ary *new;
3449 unsigned long usage;
3452 mutex_lock(&memcg->thresholds_lock);
3455 thresholds = &memcg->thresholds;
3456 usage = mem_cgroup_usage(memcg, false);
3457 } else if (type == _MEMSWAP) {
3458 thresholds = &memcg->memsw_thresholds;
3459 usage = mem_cgroup_usage(memcg, true);
3463 if (!thresholds->primary)
3466 /* Check if a threshold crossed before removing */
3467 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3469 /* Calculate new number of threshold */
3471 for (i = 0; i < thresholds->primary->size; i++) {
3472 if (thresholds->primary->entries[i].eventfd != eventfd)
3476 new = thresholds->spare;
3478 /* Set thresholds array to NULL if we don't have thresholds */
3487 /* Copy thresholds and find current threshold */
3488 new->current_threshold = -1;
3489 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3490 if (thresholds->primary->entries[i].eventfd == eventfd)
3493 new->entries[j] = thresholds->primary->entries[i];
3494 if (new->entries[j].threshold <= usage) {
3496 * new->current_threshold will not be used
3497 * until rcu_assign_pointer(), so it's safe to increment
3500 ++new->current_threshold;
3506 /* Swap primary and spare array */
3507 thresholds->spare = thresholds->primary;
3509 rcu_assign_pointer(thresholds->primary, new);
3511 /* To be sure that nobody uses thresholds */
3514 /* If all events are unregistered, free the spare array */
3516 kfree(thresholds->spare);
3517 thresholds->spare = NULL;
3520 mutex_unlock(&memcg->thresholds_lock);
3523 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3524 struct eventfd_ctx *eventfd)
3526 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3529 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3530 struct eventfd_ctx *eventfd)
3532 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3535 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3536 struct eventfd_ctx *eventfd, const char *args)
3538 struct mem_cgroup_eventfd_list *event;
3540 event = kmalloc(sizeof(*event), GFP_KERNEL);
3544 spin_lock(&memcg_oom_lock);
3546 event->eventfd = eventfd;
3547 list_add(&event->list, &memcg->oom_notify);
3549 /* already in OOM ? */
3550 if (memcg->under_oom)
3551 eventfd_signal(eventfd, 1);
3552 spin_unlock(&memcg_oom_lock);
3557 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3558 struct eventfd_ctx *eventfd)
3560 struct mem_cgroup_eventfd_list *ev, *tmp;
3562 spin_lock(&memcg_oom_lock);
3564 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3565 if (ev->eventfd == eventfd) {
3566 list_del(&ev->list);
3571 spin_unlock(&memcg_oom_lock);
3574 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3576 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3578 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3579 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3583 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3584 struct cftype *cft, u64 val)
3586 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3588 /* cannot set to root cgroup and only 0 and 1 are allowed */
3589 if (!css->parent || !((val == 0) || (val == 1)))
3592 memcg->oom_kill_disable = val;
3594 memcg_oom_recover(memcg);
3599 #ifdef CONFIG_CGROUP_WRITEBACK
3601 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3603 return &memcg->cgwb_list;
3606 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3608 return wb_domain_init(&memcg->cgwb_domain, gfp);
3611 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3613 wb_domain_exit(&memcg->cgwb_domain);
3616 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3618 wb_domain_size_changed(&memcg->cgwb_domain);
3621 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3623 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3625 if (!memcg->css.parent)
3628 return &memcg->cgwb_domain;
3632 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3633 * @wb: bdi_writeback in question
3634 * @pfilepages: out parameter for number of file pages
3635 * @pheadroom: out parameter for number of allocatable pages according to memcg
3636 * @pdirty: out parameter for number of dirty pages
3637 * @pwriteback: out parameter for number of pages under writeback
3639 * Determine the numbers of file, headroom, dirty, and writeback pages in
3640 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3641 * is a bit more involved.
3643 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3644 * headroom is calculated as the lowest headroom of itself and the
3645 * ancestors. Note that this doesn't consider the actual amount of
3646 * available memory in the system. The caller should further cap
3647 * *@pheadroom accordingly.
3649 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3650 unsigned long *pheadroom, unsigned long *pdirty,
3651 unsigned long *pwriteback)
3653 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3654 struct mem_cgroup *parent;
3656 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
3658 /* this should eventually include NR_UNSTABLE_NFS */
3659 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
3660 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3661 (1 << LRU_ACTIVE_FILE));
3662 *pheadroom = PAGE_COUNTER_MAX;
3664 while ((parent = parent_mem_cgroup(memcg))) {
3665 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3666 unsigned long used = page_counter_read(&memcg->memory);
3668 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3673 #else /* CONFIG_CGROUP_WRITEBACK */
3675 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3680 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3684 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3688 #endif /* CONFIG_CGROUP_WRITEBACK */
3691 * DO NOT USE IN NEW FILES.
3693 * "cgroup.event_control" implementation.
3695 * This is way over-engineered. It tries to support fully configurable
3696 * events for each user. Such level of flexibility is completely
3697 * unnecessary especially in the light of the planned unified hierarchy.
3699 * Please deprecate this and replace with something simpler if at all
3704 * Unregister event and free resources.
3706 * Gets called from workqueue.
3708 static void memcg_event_remove(struct work_struct *work)
3710 struct mem_cgroup_event *event =
3711 container_of(work, struct mem_cgroup_event, remove);
3712 struct mem_cgroup *memcg = event->memcg;
3714 remove_wait_queue(event->wqh, &event->wait);
3716 event->unregister_event(memcg, event->eventfd);
3718 /* Notify userspace the event is going away. */
3719 eventfd_signal(event->eventfd, 1);
3721 eventfd_ctx_put(event->eventfd);
3723 css_put(&memcg->css);
3727 * Gets called on POLLHUP on eventfd when user closes it.
3729 * Called with wqh->lock held and interrupts disabled.
3731 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
3732 int sync, void *key)
3734 struct mem_cgroup_event *event =
3735 container_of(wait, struct mem_cgroup_event, wait);
3736 struct mem_cgroup *memcg = event->memcg;
3737 unsigned long flags = (unsigned long)key;
3739 if (flags & POLLHUP) {
3741 * If the event has been detached at cgroup removal, we
3742 * can simply return knowing the other side will cleanup
3745 * We can't race against event freeing since the other
3746 * side will require wqh->lock via remove_wait_queue(),
3749 spin_lock(&memcg->event_list_lock);
3750 if (!list_empty(&event->list)) {
3751 list_del_init(&event->list);
3753 * We are in atomic context, but cgroup_event_remove()
3754 * may sleep, so we have to call it in workqueue.
3756 schedule_work(&event->remove);
3758 spin_unlock(&memcg->event_list_lock);
3764 static void memcg_event_ptable_queue_proc(struct file *file,
3765 wait_queue_head_t *wqh, poll_table *pt)
3767 struct mem_cgroup_event *event =
3768 container_of(pt, struct mem_cgroup_event, pt);
3771 add_wait_queue(wqh, &event->wait);
3775 * DO NOT USE IN NEW FILES.
3777 * Parse input and register new cgroup event handler.
3779 * Input must be in format '<event_fd> <control_fd> <args>'.
3780 * Interpretation of args is defined by control file implementation.
3782 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3783 char *buf, size_t nbytes, loff_t off)
3785 struct cgroup_subsys_state *css = of_css(of);
3786 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3787 struct mem_cgroup_event *event;
3788 struct cgroup_subsys_state *cfile_css;
3789 unsigned int efd, cfd;
3796 buf = strstrip(buf);
3798 efd = simple_strtoul(buf, &endp, 10);
3803 cfd = simple_strtoul(buf, &endp, 10);
3804 if ((*endp != ' ') && (*endp != '\0'))
3808 event = kzalloc(sizeof(*event), GFP_KERNEL);
3812 event->memcg = memcg;
3813 INIT_LIST_HEAD(&event->list);
3814 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3815 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3816 INIT_WORK(&event->remove, memcg_event_remove);
3824 event->eventfd = eventfd_ctx_fileget(efile.file);
3825 if (IS_ERR(event->eventfd)) {
3826 ret = PTR_ERR(event->eventfd);
3833 goto out_put_eventfd;
3836 /* the process need read permission on control file */
3837 /* AV: shouldn't we check that it's been opened for read instead? */
3838 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3843 * Determine the event callbacks and set them in @event. This used
3844 * to be done via struct cftype but cgroup core no longer knows
3845 * about these events. The following is crude but the whole thing
3846 * is for compatibility anyway.
3848 * DO NOT ADD NEW FILES.
3850 name = cfile.file->f_path.dentry->d_name.name;
3852 if (!strcmp(name, "memory.usage_in_bytes")) {
3853 event->register_event = mem_cgroup_usage_register_event;
3854 event->unregister_event = mem_cgroup_usage_unregister_event;
3855 } else if (!strcmp(name, "memory.oom_control")) {
3856 event->register_event = mem_cgroup_oom_register_event;
3857 event->unregister_event = mem_cgroup_oom_unregister_event;
3858 } else if (!strcmp(name, "memory.pressure_level")) {
3859 event->register_event = vmpressure_register_event;
3860 event->unregister_event = vmpressure_unregister_event;
3861 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3862 event->register_event = memsw_cgroup_usage_register_event;
3863 event->unregister_event = memsw_cgroup_usage_unregister_event;
3870 * Verify @cfile should belong to @css. Also, remaining events are
3871 * automatically removed on cgroup destruction but the removal is
3872 * asynchronous, so take an extra ref on @css.
3874 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3875 &memory_cgrp_subsys);
3877 if (IS_ERR(cfile_css))
3879 if (cfile_css != css) {
3884 ret = event->register_event(memcg, event->eventfd, buf);
3888 efile.file->f_op->poll(efile.file, &event->pt);
3890 spin_lock(&memcg->event_list_lock);
3891 list_add(&event->list, &memcg->event_list);
3892 spin_unlock(&memcg->event_list_lock);
3904 eventfd_ctx_put(event->eventfd);
3913 static struct cftype mem_cgroup_legacy_files[] = {
3915 .name = "usage_in_bytes",
3916 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3917 .read_u64 = mem_cgroup_read_u64,
3920 .name = "max_usage_in_bytes",
3921 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3922 .write = mem_cgroup_reset,
3923 .read_u64 = mem_cgroup_read_u64,
3926 .name = "limit_in_bytes",
3927 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3928 .write = mem_cgroup_write,
3929 .read_u64 = mem_cgroup_read_u64,
3932 .name = "soft_limit_in_bytes",
3933 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3934 .write = mem_cgroup_write,
3935 .read_u64 = mem_cgroup_read_u64,
3939 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3940 .write = mem_cgroup_reset,
3941 .read_u64 = mem_cgroup_read_u64,
3945 .seq_show = memcg_stat_show,
3948 .name = "force_empty",
3949 .write = mem_cgroup_force_empty_write,
3952 .name = "use_hierarchy",
3953 .write_u64 = mem_cgroup_hierarchy_write,
3954 .read_u64 = mem_cgroup_hierarchy_read,
3957 .name = "cgroup.event_control", /* XXX: for compat */
3958 .write = memcg_write_event_control,
3959 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3962 .name = "swappiness",
3963 .read_u64 = mem_cgroup_swappiness_read,
3964 .write_u64 = mem_cgroup_swappiness_write,
3967 .name = "move_charge_at_immigrate",
3968 .read_u64 = mem_cgroup_move_charge_read,
3969 .write_u64 = mem_cgroup_move_charge_write,
3972 .name = "oom_control",
3973 .seq_show = mem_cgroup_oom_control_read,
3974 .write_u64 = mem_cgroup_oom_control_write,
3975 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3978 .name = "pressure_level",
3982 .name = "numa_stat",
3983 .seq_show = memcg_numa_stat_show,
3987 .name = "kmem.limit_in_bytes",
3988 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
3989 .write = mem_cgroup_write,
3990 .read_u64 = mem_cgroup_read_u64,
3993 .name = "kmem.usage_in_bytes",
3994 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
3995 .read_u64 = mem_cgroup_read_u64,
3998 .name = "kmem.failcnt",
3999 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4000 .write = mem_cgroup_reset,
4001 .read_u64 = mem_cgroup_read_u64,
4004 .name = "kmem.max_usage_in_bytes",
4005 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4006 .write = mem_cgroup_reset,
4007 .read_u64 = mem_cgroup_read_u64,
4009 #ifdef CONFIG_SLABINFO
4011 .name = "kmem.slabinfo",
4012 .seq_start = slab_start,
4013 .seq_next = slab_next,
4014 .seq_stop = slab_stop,
4015 .seq_show = memcg_slab_show,
4019 .name = "kmem.tcp.limit_in_bytes",
4020 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4021 .write = mem_cgroup_write,
4022 .read_u64 = mem_cgroup_read_u64,
4025 .name = "kmem.tcp.usage_in_bytes",
4026 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4027 .read_u64 = mem_cgroup_read_u64,
4030 .name = "kmem.tcp.failcnt",
4031 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4032 .write = mem_cgroup_reset,
4033 .read_u64 = mem_cgroup_read_u64,
4036 .name = "kmem.tcp.max_usage_in_bytes",
4037 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4038 .write = mem_cgroup_reset,
4039 .read_u64 = mem_cgroup_read_u64,
4041 { }, /* terminate */
4045 * Private memory cgroup IDR
4047 * Swap-out records and page cache shadow entries need to store memcg
4048 * references in constrained space, so we maintain an ID space that is
4049 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4050 * memory-controlled cgroups to 64k.
4052 * However, there usually are many references to the oflline CSS after
4053 * the cgroup has been destroyed, such as page cache or reclaimable
4054 * slab objects, that don't need to hang on to the ID. We want to keep
4055 * those dead CSS from occupying IDs, or we might quickly exhaust the
4056 * relatively small ID space and prevent the creation of new cgroups
4057 * even when there are much fewer than 64k cgroups - possibly none.
4059 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4060 * be freed and recycled when it's no longer needed, which is usually
4061 * when the CSS is offlined.
4063 * The only exception to that are records of swapped out tmpfs/shmem
4064 * pages that need to be attributed to live ancestors on swapin. But
4065 * those references are manageable from userspace.
4068 static DEFINE_IDR(mem_cgroup_idr);
4070 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4072 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4073 atomic_add(n, &memcg->id.ref);
4076 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4078 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4079 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4080 idr_remove(&mem_cgroup_idr, memcg->id.id);
4083 /* Memcg ID pins CSS */
4084 css_put(&memcg->css);
4088 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4090 mem_cgroup_id_get_many(memcg, 1);
4093 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4095 mem_cgroup_id_put_many(memcg, 1);
4099 * mem_cgroup_from_id - look up a memcg from a memcg id
4100 * @id: the memcg id to look up
4102 * Caller must hold rcu_read_lock().
4104 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4106 WARN_ON_ONCE(!rcu_read_lock_held());
4107 return idr_find(&mem_cgroup_idr, id);
4110 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4112 struct mem_cgroup_per_node *pn;
4115 * This routine is called against possible nodes.
4116 * But it's BUG to call kmalloc() against offline node.
4118 * TODO: this routine can waste much memory for nodes which will
4119 * never be onlined. It's better to use memory hotplug callback
4122 if (!node_state(node, N_NORMAL_MEMORY))
4124 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4128 lruvec_init(&pn->lruvec);
4129 pn->usage_in_excess = 0;
4130 pn->on_tree = false;
4133 memcg->nodeinfo[node] = pn;
4137 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4139 kfree(memcg->nodeinfo[node]);
4142 static void mem_cgroup_free(struct mem_cgroup *memcg)
4146 memcg_wb_domain_exit(memcg);
4148 free_mem_cgroup_per_node_info(memcg, node);
4149 free_percpu(memcg->stat);
4153 static struct mem_cgroup *mem_cgroup_alloc(void)
4155 struct mem_cgroup *memcg;
4159 size = sizeof(struct mem_cgroup);
4160 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4162 memcg = kzalloc(size, GFP_KERNEL);
4166 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4167 1, MEM_CGROUP_ID_MAX,
4169 if (memcg->id.id < 0)
4172 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4177 if (alloc_mem_cgroup_per_node_info(memcg, node))
4180 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4183 INIT_WORK(&memcg->high_work, high_work_func);
4184 memcg->last_scanned_node = MAX_NUMNODES;
4185 INIT_LIST_HEAD(&memcg->oom_notify);
4186 mutex_init(&memcg->thresholds_lock);
4187 spin_lock_init(&memcg->move_lock);
4188 vmpressure_init(&memcg->vmpressure);
4189 INIT_LIST_HEAD(&memcg->event_list);
4190 spin_lock_init(&memcg->event_list_lock);
4191 memcg->socket_pressure = jiffies;
4193 memcg->kmemcg_id = -1;
4195 #ifdef CONFIG_CGROUP_WRITEBACK
4196 INIT_LIST_HEAD(&memcg->cgwb_list);
4198 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4201 if (memcg->id.id > 0)
4202 idr_remove(&mem_cgroup_idr, memcg->id.id);
4203 mem_cgroup_free(memcg);
4207 static struct cgroup_subsys_state * __ref
4208 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4210 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4211 struct mem_cgroup *memcg;
4212 long error = -ENOMEM;
4214 memcg = mem_cgroup_alloc();
4216 return ERR_PTR(error);
4218 memcg->high = PAGE_COUNTER_MAX;
4219 memcg->soft_limit = PAGE_COUNTER_MAX;
4221 memcg->swappiness = mem_cgroup_swappiness(parent);
4222 memcg->oom_kill_disable = parent->oom_kill_disable;
4224 if (parent && parent->use_hierarchy) {
4225 memcg->use_hierarchy = true;
4226 page_counter_init(&memcg->memory, &parent->memory);
4227 page_counter_init(&memcg->swap, &parent->swap);
4228 page_counter_init(&memcg->memsw, &parent->memsw);
4229 page_counter_init(&memcg->kmem, &parent->kmem);
4230 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4232 page_counter_init(&memcg->memory, NULL);
4233 page_counter_init(&memcg->swap, NULL);
4234 page_counter_init(&memcg->memsw, NULL);
4235 page_counter_init(&memcg->kmem, NULL);
4236 page_counter_init(&memcg->tcpmem, NULL);
4238 * Deeper hierachy with use_hierarchy == false doesn't make
4239 * much sense so let cgroup subsystem know about this
4240 * unfortunate state in our controller.
4242 if (parent != root_mem_cgroup)
4243 memory_cgrp_subsys.broken_hierarchy = true;
4246 /* The following stuff does not apply to the root */
4248 root_mem_cgroup = memcg;
4252 error = memcg_online_kmem(memcg);
4256 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4257 static_branch_inc(&memcg_sockets_enabled_key);
4261 mem_cgroup_free(memcg);
4262 return ERR_PTR(-ENOMEM);
4265 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4267 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4269 /* Online state pins memcg ID, memcg ID pins CSS */
4270 atomic_set(&memcg->id.ref, 1);
4275 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4277 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4278 struct mem_cgroup_event *event, *tmp;
4281 * Unregister events and notify userspace.
4282 * Notify userspace about cgroup removing only after rmdir of cgroup
4283 * directory to avoid race between userspace and kernelspace.
4285 spin_lock(&memcg->event_list_lock);
4286 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4287 list_del_init(&event->list);
4288 schedule_work(&event->remove);
4290 spin_unlock(&memcg->event_list_lock);
4292 memcg_offline_kmem(memcg);
4293 wb_memcg_offline(memcg);
4295 mem_cgroup_id_put(memcg);
4298 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4300 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4302 invalidate_reclaim_iterators(memcg);
4305 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4307 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4309 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4310 static_branch_dec(&memcg_sockets_enabled_key);
4312 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4313 static_branch_dec(&memcg_sockets_enabled_key);
4315 vmpressure_cleanup(&memcg->vmpressure);
4316 cancel_work_sync(&memcg->high_work);
4317 mem_cgroup_remove_from_trees(memcg);
4318 memcg_free_kmem(memcg);
4319 mem_cgroup_free(memcg);
4323 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4324 * @css: the target css
4326 * Reset the states of the mem_cgroup associated with @css. This is
4327 * invoked when the userland requests disabling on the default hierarchy
4328 * but the memcg is pinned through dependency. The memcg should stop
4329 * applying policies and should revert to the vanilla state as it may be
4330 * made visible again.
4332 * The current implementation only resets the essential configurations.
4333 * This needs to be expanded to cover all the visible parts.
4335 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4337 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4339 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4340 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4341 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4342 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4343 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4345 memcg->high = PAGE_COUNTER_MAX;
4346 memcg->soft_limit = PAGE_COUNTER_MAX;
4347 memcg_wb_domain_size_changed(memcg);
4351 /* Handlers for move charge at task migration. */
4352 static int mem_cgroup_do_precharge(unsigned long count)
4356 /* Try a single bulk charge without reclaim first, kswapd may wake */
4357 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4359 mc.precharge += count;
4363 /* Try charges one by one with reclaim, but do not retry */
4365 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4379 enum mc_target_type {
4385 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4386 unsigned long addr, pte_t ptent)
4388 struct page *page = vm_normal_page(vma, addr, ptent);
4390 if (!page || !page_mapped(page))
4392 if (PageAnon(page)) {
4393 if (!(mc.flags & MOVE_ANON))
4396 if (!(mc.flags & MOVE_FILE))
4399 if (!get_page_unless_zero(page))
4406 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4407 pte_t ptent, swp_entry_t *entry)
4409 struct page *page = NULL;
4410 swp_entry_t ent = pte_to_swp_entry(ptent);
4412 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4415 * Because lookup_swap_cache() updates some statistics counter,
4416 * we call find_get_page() with swapper_space directly.
4418 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4419 if (do_memsw_account())
4420 entry->val = ent.val;
4425 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4426 pte_t ptent, swp_entry_t *entry)
4432 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4433 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4435 struct page *page = NULL;
4436 struct address_space *mapping;
4439 if (!vma->vm_file) /* anonymous vma */
4441 if (!(mc.flags & MOVE_FILE))
4444 mapping = vma->vm_file->f_mapping;
4445 pgoff = linear_page_index(vma, addr);
4447 /* page is moved even if it's not RSS of this task(page-faulted). */
4449 /* shmem/tmpfs may report page out on swap: account for that too. */
4450 if (shmem_mapping(mapping)) {
4451 page = find_get_entry(mapping, pgoff);
4452 if (radix_tree_exceptional_entry(page)) {
4453 swp_entry_t swp = radix_to_swp_entry(page);
4454 if (do_memsw_account())
4456 page = find_get_page(swap_address_space(swp),
4460 page = find_get_page(mapping, pgoff);
4462 page = find_get_page(mapping, pgoff);
4468 * mem_cgroup_move_account - move account of the page
4470 * @compound: charge the page as compound or small page
4471 * @from: mem_cgroup which the page is moved from.
4472 * @to: mem_cgroup which the page is moved to. @from != @to.
4474 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4476 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4479 static int mem_cgroup_move_account(struct page *page,
4481 struct mem_cgroup *from,
4482 struct mem_cgroup *to)
4484 unsigned long flags;
4485 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4489 VM_BUG_ON(from == to);
4490 VM_BUG_ON_PAGE(PageLRU(page), page);
4491 VM_BUG_ON(compound && !PageTransHuge(page));
4494 * Prevent mem_cgroup_migrate() from looking at
4495 * page->mem_cgroup of its source page while we change it.
4498 if (!trylock_page(page))
4502 if (page->mem_cgroup != from)
4505 anon = PageAnon(page);
4507 spin_lock_irqsave(&from->move_lock, flags);
4509 if (!anon && page_mapped(page)) {
4510 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4512 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4517 * move_lock grabbed above and caller set from->moving_account, so
4518 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4519 * So mapping should be stable for dirty pages.
4521 if (!anon && PageDirty(page)) {
4522 struct address_space *mapping = page_mapping(page);
4524 if (mapping_cap_account_dirty(mapping)) {
4525 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4527 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4532 if (PageWriteback(page)) {
4533 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4535 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4540 * It is safe to change page->mem_cgroup here because the page
4541 * is referenced, charged, and isolated - we can't race with
4542 * uncharging, charging, migration, or LRU putback.
4545 /* caller should have done css_get */
4546 page->mem_cgroup = to;
4547 spin_unlock_irqrestore(&from->move_lock, flags);
4551 local_irq_disable();
4552 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4553 memcg_check_events(to, page);
4554 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4555 memcg_check_events(from, page);
4564 * get_mctgt_type - get target type of moving charge
4565 * @vma: the vma the pte to be checked belongs
4566 * @addr: the address corresponding to the pte to be checked
4567 * @ptent: the pte to be checked
4568 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4571 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4572 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4573 * move charge. if @target is not NULL, the page is stored in target->page
4574 * with extra refcnt got(Callers should handle it).
4575 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4576 * target for charge migration. if @target is not NULL, the entry is stored
4579 * Called with pte lock held.
4582 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4583 unsigned long addr, pte_t ptent, union mc_target *target)
4585 struct page *page = NULL;
4586 enum mc_target_type ret = MC_TARGET_NONE;
4587 swp_entry_t ent = { .val = 0 };
4589 if (pte_present(ptent))
4590 page = mc_handle_present_pte(vma, addr, ptent);
4591 else if (is_swap_pte(ptent))
4592 page = mc_handle_swap_pte(vma, ptent, &ent);
4593 else if (pte_none(ptent))
4594 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4596 if (!page && !ent.val)
4600 * Do only loose check w/o serialization.
4601 * mem_cgroup_move_account() checks the page is valid or
4602 * not under LRU exclusion.
4604 if (page->mem_cgroup == mc.from) {
4605 ret = MC_TARGET_PAGE;
4607 target->page = page;
4609 if (!ret || !target)
4612 /* There is a swap entry and a page doesn't exist or isn't charged */
4613 if (ent.val && !ret &&
4614 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4615 ret = MC_TARGET_SWAP;
4622 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4624 * We don't consider swapping or file mapped pages because THP does not
4625 * support them for now.
4626 * Caller should make sure that pmd_trans_huge(pmd) is true.
4628 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4629 unsigned long addr, pmd_t pmd, union mc_target *target)
4631 struct page *page = NULL;
4632 enum mc_target_type ret = MC_TARGET_NONE;
4634 page = pmd_page(pmd);
4635 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4636 if (!(mc.flags & MOVE_ANON))
4638 if (page->mem_cgroup == mc.from) {
4639 ret = MC_TARGET_PAGE;
4642 target->page = page;
4648 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4649 unsigned long addr, pmd_t pmd, union mc_target *target)
4651 return MC_TARGET_NONE;
4655 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4656 unsigned long addr, unsigned long end,
4657 struct mm_walk *walk)
4659 struct vm_area_struct *vma = walk->vma;
4663 ptl = pmd_trans_huge_lock(pmd, vma);
4665 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4666 mc.precharge += HPAGE_PMD_NR;
4671 if (pmd_trans_unstable(pmd))
4673 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4674 for (; addr != end; pte++, addr += PAGE_SIZE)
4675 if (get_mctgt_type(vma, addr, *pte, NULL))
4676 mc.precharge++; /* increment precharge temporarily */
4677 pte_unmap_unlock(pte - 1, ptl);
4683 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4685 unsigned long precharge;
4687 struct mm_walk mem_cgroup_count_precharge_walk = {
4688 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4691 down_read(&mm->mmap_sem);
4692 walk_page_range(0, mm->highest_vm_end,
4693 &mem_cgroup_count_precharge_walk);
4694 up_read(&mm->mmap_sem);
4696 precharge = mc.precharge;
4702 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4704 unsigned long precharge = mem_cgroup_count_precharge(mm);
4706 VM_BUG_ON(mc.moving_task);
4707 mc.moving_task = current;
4708 return mem_cgroup_do_precharge(precharge);
4711 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4712 static void __mem_cgroup_clear_mc(void)
4714 struct mem_cgroup *from = mc.from;
4715 struct mem_cgroup *to = mc.to;
4717 /* we must uncharge all the leftover precharges from mc.to */
4719 cancel_charge(mc.to, mc.precharge);
4723 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4724 * we must uncharge here.
4726 if (mc.moved_charge) {
4727 cancel_charge(mc.from, mc.moved_charge);
4728 mc.moved_charge = 0;
4730 /* we must fixup refcnts and charges */
4731 if (mc.moved_swap) {
4732 /* uncharge swap account from the old cgroup */
4733 if (!mem_cgroup_is_root(mc.from))
4734 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4736 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4739 * we charged both to->memory and to->memsw, so we
4740 * should uncharge to->memory.
4742 if (!mem_cgroup_is_root(mc.to))
4743 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4745 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4746 css_put_many(&mc.to->css, mc.moved_swap);
4750 memcg_oom_recover(from);
4751 memcg_oom_recover(to);
4752 wake_up_all(&mc.waitq);
4755 static void mem_cgroup_clear_mc(void)
4757 struct mm_struct *mm = mc.mm;
4760 * we must clear moving_task before waking up waiters at the end of
4763 mc.moving_task = NULL;
4764 __mem_cgroup_clear_mc();
4765 spin_lock(&mc.lock);
4769 spin_unlock(&mc.lock);
4774 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4776 struct cgroup_subsys_state *css;
4777 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4778 struct mem_cgroup *from;
4779 struct task_struct *leader, *p;
4780 struct mm_struct *mm;
4781 unsigned long move_flags;
4784 /* charge immigration isn't supported on the default hierarchy */
4785 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4789 * Multi-process migrations only happen on the default hierarchy
4790 * where charge immigration is not used. Perform charge
4791 * immigration if @tset contains a leader and whine if there are
4795 cgroup_taskset_for_each_leader(leader, css, tset) {
4798 memcg = mem_cgroup_from_css(css);
4804 * We are now commited to this value whatever it is. Changes in this
4805 * tunable will only affect upcoming migrations, not the current one.
4806 * So we need to save it, and keep it going.
4808 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4812 from = mem_cgroup_from_task(p);
4814 VM_BUG_ON(from == memcg);
4816 mm = get_task_mm(p);
4819 /* We move charges only when we move a owner of the mm */
4820 if (mm->owner == p) {
4823 VM_BUG_ON(mc.precharge);
4824 VM_BUG_ON(mc.moved_charge);
4825 VM_BUG_ON(mc.moved_swap);
4827 spin_lock(&mc.lock);
4831 mc.flags = move_flags;
4832 spin_unlock(&mc.lock);
4833 /* We set mc.moving_task later */
4835 ret = mem_cgroup_precharge_mc(mm);
4837 mem_cgroup_clear_mc();
4844 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4847 mem_cgroup_clear_mc();
4850 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4851 unsigned long addr, unsigned long end,
4852 struct mm_walk *walk)
4855 struct vm_area_struct *vma = walk->vma;
4858 enum mc_target_type target_type;
4859 union mc_target target;
4862 ptl = pmd_trans_huge_lock(pmd, vma);
4864 if (mc.precharge < HPAGE_PMD_NR) {
4868 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4869 if (target_type == MC_TARGET_PAGE) {
4871 if (!isolate_lru_page(page)) {
4872 if (!mem_cgroup_move_account(page, true,
4874 mc.precharge -= HPAGE_PMD_NR;
4875 mc.moved_charge += HPAGE_PMD_NR;
4877 putback_lru_page(page);
4885 if (pmd_trans_unstable(pmd))
4888 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4889 for (; addr != end; addr += PAGE_SIZE) {
4890 pte_t ptent = *(pte++);
4896 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4897 case MC_TARGET_PAGE:
4900 * We can have a part of the split pmd here. Moving it
4901 * can be done but it would be too convoluted so simply
4902 * ignore such a partial THP and keep it in original
4903 * memcg. There should be somebody mapping the head.
4905 if (PageTransCompound(page))
4907 if (isolate_lru_page(page))
4909 if (!mem_cgroup_move_account(page, false,
4912 /* we uncharge from mc.from later. */
4915 putback_lru_page(page);
4916 put: /* get_mctgt_type() gets the page */
4919 case MC_TARGET_SWAP:
4921 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4923 /* we fixup refcnts and charges later. */
4931 pte_unmap_unlock(pte - 1, ptl);
4936 * We have consumed all precharges we got in can_attach().
4937 * We try charge one by one, but don't do any additional
4938 * charges to mc.to if we have failed in charge once in attach()
4941 ret = mem_cgroup_do_precharge(1);
4949 static void mem_cgroup_move_charge(void)
4951 struct mm_walk mem_cgroup_move_charge_walk = {
4952 .pmd_entry = mem_cgroup_move_charge_pte_range,
4956 lru_add_drain_all();
4958 * Signal lock_page_memcg() to take the memcg's move_lock
4959 * while we're moving its pages to another memcg. Then wait
4960 * for already started RCU-only updates to finish.
4962 atomic_inc(&mc.from->moving_account);
4965 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
4967 * Someone who are holding the mmap_sem might be waiting in
4968 * waitq. So we cancel all extra charges, wake up all waiters,
4969 * and retry. Because we cancel precharges, we might not be able
4970 * to move enough charges, but moving charge is a best-effort
4971 * feature anyway, so it wouldn't be a big problem.
4973 __mem_cgroup_clear_mc();
4978 * When we have consumed all precharges and failed in doing
4979 * additional charge, the page walk just aborts.
4981 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
4983 up_read(&mc.mm->mmap_sem);
4984 atomic_dec(&mc.from->moving_account);
4987 static void mem_cgroup_move_task(void)
4990 mem_cgroup_move_charge();
4991 mem_cgroup_clear_mc();
4994 #else /* !CONFIG_MMU */
4995 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4999 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5002 static void mem_cgroup_move_task(void)
5008 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5009 * to verify whether we're attached to the default hierarchy on each mount
5012 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5015 * use_hierarchy is forced on the default hierarchy. cgroup core
5016 * guarantees that @root doesn't have any children, so turning it
5017 * on for the root memcg is enough.
5019 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5020 root_mem_cgroup->use_hierarchy = true;
5022 root_mem_cgroup->use_hierarchy = false;
5025 static u64 memory_current_read(struct cgroup_subsys_state *css,
5028 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5030 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5033 static int memory_low_show(struct seq_file *m, void *v)
5035 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5036 unsigned long low = READ_ONCE(memcg->low);
5038 if (low == PAGE_COUNTER_MAX)
5039 seq_puts(m, "max\n");
5041 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5046 static ssize_t memory_low_write(struct kernfs_open_file *of,
5047 char *buf, size_t nbytes, loff_t off)
5049 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5053 buf = strstrip(buf);
5054 err = page_counter_memparse(buf, "max", &low);
5063 static int memory_high_show(struct seq_file *m, void *v)
5065 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5066 unsigned long high = READ_ONCE(memcg->high);
5068 if (high == PAGE_COUNTER_MAX)
5069 seq_puts(m, "max\n");
5071 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5076 static ssize_t memory_high_write(struct kernfs_open_file *of,
5077 char *buf, size_t nbytes, loff_t off)
5079 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5080 unsigned long nr_pages;
5084 buf = strstrip(buf);
5085 err = page_counter_memparse(buf, "max", &high);
5091 nr_pages = page_counter_read(&memcg->memory);
5092 if (nr_pages > high)
5093 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5096 memcg_wb_domain_size_changed(memcg);
5100 static int memory_max_show(struct seq_file *m, void *v)
5102 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5103 unsigned long max = READ_ONCE(memcg->memory.limit);
5105 if (max == PAGE_COUNTER_MAX)
5106 seq_puts(m, "max\n");
5108 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5113 static ssize_t memory_max_write(struct kernfs_open_file *of,
5114 char *buf, size_t nbytes, loff_t off)
5116 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5117 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5118 bool drained = false;
5122 buf = strstrip(buf);
5123 err = page_counter_memparse(buf, "max", &max);
5127 xchg(&memcg->memory.limit, max);
5130 unsigned long nr_pages = page_counter_read(&memcg->memory);
5132 if (nr_pages <= max)
5135 if (signal_pending(current)) {
5141 drain_all_stock(memcg);
5147 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5153 mem_cgroup_events(memcg, MEMCG_OOM, 1);
5154 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5158 memcg_wb_domain_size_changed(memcg);
5162 static int memory_events_show(struct seq_file *m, void *v)
5164 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5166 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5167 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5168 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5169 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5174 static int memory_stat_show(struct seq_file *m, void *v)
5176 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5177 unsigned long stat[MEMCG_NR_STAT];
5178 unsigned long events[MEMCG_NR_EVENTS];
5182 * Provide statistics on the state of the memory subsystem as
5183 * well as cumulative event counters that show past behavior.
5185 * This list is ordered following a combination of these gradients:
5186 * 1) generic big picture -> specifics and details
5187 * 2) reflecting userspace activity -> reflecting kernel heuristics
5189 * Current memory state:
5192 tree_stat(memcg, stat);
5193 tree_events(memcg, events);
5195 seq_printf(m, "anon %llu\n",
5196 (u64)stat[MEM_CGROUP_STAT_RSS] * PAGE_SIZE);
5197 seq_printf(m, "file %llu\n",
5198 (u64)stat[MEM_CGROUP_STAT_CACHE] * PAGE_SIZE);
5199 seq_printf(m, "kernel_stack %llu\n",
5200 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5201 seq_printf(m, "slab %llu\n",
5202 (u64)(stat[MEMCG_SLAB_RECLAIMABLE] +
5203 stat[MEMCG_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5204 seq_printf(m, "sock %llu\n",
5205 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5207 seq_printf(m, "file_mapped %llu\n",
5208 (u64)stat[MEM_CGROUP_STAT_FILE_MAPPED] * PAGE_SIZE);
5209 seq_printf(m, "file_dirty %llu\n",
5210 (u64)stat[MEM_CGROUP_STAT_DIRTY] * PAGE_SIZE);
5211 seq_printf(m, "file_writeback %llu\n",
5212 (u64)stat[MEM_CGROUP_STAT_WRITEBACK] * PAGE_SIZE);
5214 for (i = 0; i < NR_LRU_LISTS; i++) {
5215 struct mem_cgroup *mi;
5216 unsigned long val = 0;
5218 for_each_mem_cgroup_tree(mi, memcg)
5219 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5220 seq_printf(m, "%s %llu\n",
5221 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5224 seq_printf(m, "slab_reclaimable %llu\n",
5225 (u64)stat[MEMCG_SLAB_RECLAIMABLE] * PAGE_SIZE);
5226 seq_printf(m, "slab_unreclaimable %llu\n",
5227 (u64)stat[MEMCG_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5229 /* Accumulated memory events */
5231 seq_printf(m, "pgfault %lu\n",
5232 events[MEM_CGROUP_EVENTS_PGFAULT]);
5233 seq_printf(m, "pgmajfault %lu\n",
5234 events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
5239 static struct cftype memory_files[] = {
5242 .flags = CFTYPE_NOT_ON_ROOT,
5243 .read_u64 = memory_current_read,
5247 .flags = CFTYPE_NOT_ON_ROOT,
5248 .seq_show = memory_low_show,
5249 .write = memory_low_write,
5253 .flags = CFTYPE_NOT_ON_ROOT,
5254 .seq_show = memory_high_show,
5255 .write = memory_high_write,
5259 .flags = CFTYPE_NOT_ON_ROOT,
5260 .seq_show = memory_max_show,
5261 .write = memory_max_write,
5265 .flags = CFTYPE_NOT_ON_ROOT,
5266 .file_offset = offsetof(struct mem_cgroup, events_file),
5267 .seq_show = memory_events_show,
5271 .flags = CFTYPE_NOT_ON_ROOT,
5272 .seq_show = memory_stat_show,
5277 struct cgroup_subsys memory_cgrp_subsys = {
5278 .css_alloc = mem_cgroup_css_alloc,
5279 .css_online = mem_cgroup_css_online,
5280 .css_offline = mem_cgroup_css_offline,
5281 .css_released = mem_cgroup_css_released,
5282 .css_free = mem_cgroup_css_free,
5283 .css_reset = mem_cgroup_css_reset,
5284 .can_attach = mem_cgroup_can_attach,
5285 .cancel_attach = mem_cgroup_cancel_attach,
5286 .post_attach = mem_cgroup_move_task,
5287 .bind = mem_cgroup_bind,
5288 .dfl_cftypes = memory_files,
5289 .legacy_cftypes = mem_cgroup_legacy_files,
5294 * mem_cgroup_low - check if memory consumption is below the normal range
5295 * @root: the highest ancestor to consider
5296 * @memcg: the memory cgroup to check
5298 * Returns %true if memory consumption of @memcg, and that of all
5299 * configurable ancestors up to @root, is below the normal range.
5301 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5303 if (mem_cgroup_disabled())
5307 * The toplevel group doesn't have a configurable range, so
5308 * it's never low when looked at directly, and it is not
5309 * considered an ancestor when assessing the hierarchy.
5312 if (memcg == root_mem_cgroup)
5315 if (page_counter_read(&memcg->memory) >= memcg->low)
5318 while (memcg != root) {
5319 memcg = parent_mem_cgroup(memcg);
5321 if (memcg == root_mem_cgroup)
5324 if (page_counter_read(&memcg->memory) >= memcg->low)
5331 * mem_cgroup_try_charge - try charging a page
5332 * @page: page to charge
5333 * @mm: mm context of the victim
5334 * @gfp_mask: reclaim mode
5335 * @memcgp: charged memcg return
5336 * @compound: charge the page as compound or small page
5338 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5339 * pages according to @gfp_mask if necessary.
5341 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5342 * Otherwise, an error code is returned.
5344 * After page->mapping has been set up, the caller must finalize the
5345 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5346 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5348 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5349 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5352 struct mem_cgroup *memcg = NULL;
5353 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5356 if (mem_cgroup_disabled())
5359 if (PageSwapCache(page)) {
5361 * Every swap fault against a single page tries to charge the
5362 * page, bail as early as possible. shmem_unuse() encounters
5363 * already charged pages, too. The USED bit is protected by
5364 * the page lock, which serializes swap cache removal, which
5365 * in turn serializes uncharging.
5367 VM_BUG_ON_PAGE(!PageLocked(page), page);
5368 if (page->mem_cgroup)
5371 if (do_swap_account) {
5372 swp_entry_t ent = { .val = page_private(page), };
5373 unsigned short id = lookup_swap_cgroup_id(ent);
5376 memcg = mem_cgroup_from_id(id);
5377 if (memcg && !css_tryget_online(&memcg->css))
5384 memcg = get_mem_cgroup_from_mm(mm);
5386 ret = try_charge(memcg, gfp_mask, nr_pages);
5388 css_put(&memcg->css);
5395 * mem_cgroup_commit_charge - commit a page charge
5396 * @page: page to charge
5397 * @memcg: memcg to charge the page to
5398 * @lrucare: page might be on LRU already
5399 * @compound: charge the page as compound or small page
5401 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5402 * after page->mapping has been set up. This must happen atomically
5403 * as part of the page instantiation, i.e. under the page table lock
5404 * for anonymous pages, under the page lock for page and swap cache.
5406 * In addition, the page must not be on the LRU during the commit, to
5407 * prevent racing with task migration. If it might be, use @lrucare.
5409 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5411 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5412 bool lrucare, bool compound)
5414 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5416 VM_BUG_ON_PAGE(!page->mapping, page);
5417 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5419 if (mem_cgroup_disabled())
5422 * Swap faults will attempt to charge the same page multiple
5423 * times. But reuse_swap_page() might have removed the page
5424 * from swapcache already, so we can't check PageSwapCache().
5429 commit_charge(page, memcg, lrucare);
5431 local_irq_disable();
5432 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5433 memcg_check_events(memcg, page);
5436 if (do_memsw_account() && PageSwapCache(page)) {
5437 swp_entry_t entry = { .val = page_private(page) };
5439 * The swap entry might not get freed for a long time,
5440 * let's not wait for it. The page already received a
5441 * memory+swap charge, drop the swap entry duplicate.
5443 mem_cgroup_uncharge_swap(entry);
5448 * mem_cgroup_cancel_charge - cancel a page charge
5449 * @page: page to charge
5450 * @memcg: memcg to charge the page to
5451 * @compound: charge the page as compound or small page
5453 * Cancel a charge transaction started by mem_cgroup_try_charge().
5455 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5458 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5460 if (mem_cgroup_disabled())
5463 * Swap faults will attempt to charge the same page multiple
5464 * times. But reuse_swap_page() might have removed the page
5465 * from swapcache already, so we can't check PageSwapCache().
5470 cancel_charge(memcg, nr_pages);
5473 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5474 unsigned long nr_anon, unsigned long nr_file,
5475 unsigned long nr_huge, unsigned long nr_kmem,
5476 struct page *dummy_page)
5478 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5479 unsigned long flags;
5481 if (!mem_cgroup_is_root(memcg)) {
5482 page_counter_uncharge(&memcg->memory, nr_pages);
5483 if (do_memsw_account())
5484 page_counter_uncharge(&memcg->memsw, nr_pages);
5485 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5486 page_counter_uncharge(&memcg->kmem, nr_kmem);
5487 memcg_oom_recover(memcg);
5490 local_irq_save(flags);
5491 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5492 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5493 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5494 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5495 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5496 memcg_check_events(memcg, dummy_page);
5497 local_irq_restore(flags);
5499 if (!mem_cgroup_is_root(memcg))
5500 css_put_many(&memcg->css, nr_pages);
5503 static void uncharge_list(struct list_head *page_list)
5505 struct mem_cgroup *memcg = NULL;
5506 unsigned long nr_anon = 0;
5507 unsigned long nr_file = 0;
5508 unsigned long nr_huge = 0;
5509 unsigned long nr_kmem = 0;
5510 unsigned long pgpgout = 0;
5511 struct list_head *next;
5515 * Note that the list can be a single page->lru; hence the
5516 * do-while loop instead of a simple list_for_each_entry().
5518 next = page_list->next;
5520 page = list_entry(next, struct page, lru);
5521 next = page->lru.next;
5523 VM_BUG_ON_PAGE(PageLRU(page), page);
5524 VM_BUG_ON_PAGE(page_count(page), page);
5526 if (!page->mem_cgroup)
5530 * Nobody should be changing or seriously looking at
5531 * page->mem_cgroup at this point, we have fully
5532 * exclusive access to the page.
5535 if (memcg != page->mem_cgroup) {
5537 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5538 nr_huge, nr_kmem, page);
5539 pgpgout = nr_anon = nr_file =
5540 nr_huge = nr_kmem = 0;
5542 memcg = page->mem_cgroup;
5545 if (!PageKmemcg(page)) {
5546 unsigned int nr_pages = 1;
5548 if (PageTransHuge(page)) {
5549 nr_pages <<= compound_order(page);
5550 nr_huge += nr_pages;
5553 nr_anon += nr_pages;
5555 nr_file += nr_pages;
5558 nr_kmem += 1 << compound_order(page);
5559 __ClearPageKmemcg(page);
5562 page->mem_cgroup = NULL;
5563 } while (next != page_list);
5566 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5567 nr_huge, nr_kmem, page);
5571 * mem_cgroup_uncharge - uncharge a page
5572 * @page: page to uncharge
5574 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5575 * mem_cgroup_commit_charge().
5577 void mem_cgroup_uncharge(struct page *page)
5579 if (mem_cgroup_disabled())
5582 /* Don't touch page->lru of any random page, pre-check: */
5583 if (!page->mem_cgroup)
5586 INIT_LIST_HEAD(&page->lru);
5587 uncharge_list(&page->lru);
5591 * mem_cgroup_uncharge_list - uncharge a list of page
5592 * @page_list: list of pages to uncharge
5594 * Uncharge a list of pages previously charged with
5595 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5597 void mem_cgroup_uncharge_list(struct list_head *page_list)
5599 if (mem_cgroup_disabled())
5602 if (!list_empty(page_list))
5603 uncharge_list(page_list);
5607 * mem_cgroup_migrate - charge a page's replacement
5608 * @oldpage: currently circulating page
5609 * @newpage: replacement page
5611 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5612 * be uncharged upon free.
5614 * Both pages must be locked, @newpage->mapping must be set up.
5616 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5618 struct mem_cgroup *memcg;
5619 unsigned int nr_pages;
5621 unsigned long flags;
5623 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5624 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5625 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5626 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5629 if (mem_cgroup_disabled())
5632 /* Page cache replacement: new page already charged? */
5633 if (newpage->mem_cgroup)
5636 /* Swapcache readahead pages can get replaced before being charged */
5637 memcg = oldpage->mem_cgroup;
5641 /* Force-charge the new page. The old one will be freed soon */
5642 compound = PageTransHuge(newpage);
5643 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5645 page_counter_charge(&memcg->memory, nr_pages);
5646 if (do_memsw_account())
5647 page_counter_charge(&memcg->memsw, nr_pages);
5648 css_get_many(&memcg->css, nr_pages);
5650 commit_charge(newpage, memcg, false);
5652 local_irq_save(flags);
5653 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5654 memcg_check_events(memcg, newpage);
5655 local_irq_restore(flags);
5658 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5659 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5661 void mem_cgroup_sk_alloc(struct sock *sk)
5663 struct mem_cgroup *memcg;
5665 if (!mem_cgroup_sockets_enabled)
5669 * Socket cloning can throw us here with sk_memcg already
5670 * filled. It won't however, necessarily happen from
5671 * process context. So the test for root memcg given
5672 * the current task's memcg won't help us in this case.
5674 * Respecting the original socket's memcg is a better
5675 * decision in this case.
5678 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5679 css_get(&sk->sk_memcg->css);
5684 memcg = mem_cgroup_from_task(current);
5685 if (memcg == root_mem_cgroup)
5687 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5689 if (css_tryget_online(&memcg->css))
5690 sk->sk_memcg = memcg;
5695 void mem_cgroup_sk_free(struct sock *sk)
5698 css_put(&sk->sk_memcg->css);
5702 * mem_cgroup_charge_skmem - charge socket memory
5703 * @memcg: memcg to charge
5704 * @nr_pages: number of pages to charge
5706 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5707 * @memcg's configured limit, %false if the charge had to be forced.
5709 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5711 gfp_t gfp_mask = GFP_KERNEL;
5713 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5714 struct page_counter *fail;
5716 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5717 memcg->tcpmem_pressure = 0;
5720 page_counter_charge(&memcg->tcpmem, nr_pages);
5721 memcg->tcpmem_pressure = 1;
5725 /* Don't block in the packet receive path */
5727 gfp_mask = GFP_NOWAIT;
5729 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5731 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5734 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5739 * mem_cgroup_uncharge_skmem - uncharge socket memory
5740 * @memcg - memcg to uncharge
5741 * @nr_pages - number of pages to uncharge
5743 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5745 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5746 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5750 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5752 page_counter_uncharge(&memcg->memory, nr_pages);
5753 css_put_many(&memcg->css, nr_pages);
5756 static int __init cgroup_memory(char *s)
5760 while ((token = strsep(&s, ",")) != NULL) {
5763 if (!strcmp(token, "nosocket"))
5764 cgroup_memory_nosocket = true;
5765 if (!strcmp(token, "nokmem"))
5766 cgroup_memory_nokmem = true;
5770 __setup("cgroup.memory=", cgroup_memory);
5773 * subsys_initcall() for memory controller.
5775 * Some parts like hotcpu_notifier() have to be initialized from this context
5776 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5777 * everything that doesn't depend on a specific mem_cgroup structure should
5778 * be initialized from here.
5780 static int __init mem_cgroup_init(void)
5784 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5786 for_each_possible_cpu(cpu)
5787 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5790 for_each_node(node) {
5791 struct mem_cgroup_tree_per_node *rtpn;
5793 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5794 node_online(node) ? node : NUMA_NO_NODE);
5796 rtpn->rb_root = RB_ROOT;
5797 spin_lock_init(&rtpn->lock);
5798 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5803 subsys_initcall(mem_cgroup_init);
5805 #ifdef CONFIG_MEMCG_SWAP
5806 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5808 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5810 * The root cgroup cannot be destroyed, so it's refcount must
5813 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5817 memcg = parent_mem_cgroup(memcg);
5819 memcg = root_mem_cgroup;
5825 * mem_cgroup_swapout - transfer a memsw charge to swap
5826 * @page: page whose memsw charge to transfer
5827 * @entry: swap entry to move the charge to
5829 * Transfer the memsw charge of @page to @entry.
5831 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5833 struct mem_cgroup *memcg, *swap_memcg;
5834 unsigned short oldid;
5836 VM_BUG_ON_PAGE(PageLRU(page), page);
5837 VM_BUG_ON_PAGE(page_count(page), page);
5839 if (!do_memsw_account())
5842 memcg = page->mem_cgroup;
5844 /* Readahead page, never charged */
5849 * In case the memcg owning these pages has been offlined and doesn't
5850 * have an ID allocated to it anymore, charge the closest online
5851 * ancestor for the swap instead and transfer the memory+swap charge.
5853 swap_memcg = mem_cgroup_id_get_online(memcg);
5854 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg));
5855 VM_BUG_ON_PAGE(oldid, page);
5856 mem_cgroup_swap_statistics(swap_memcg, true);
5858 page->mem_cgroup = NULL;
5860 if (!mem_cgroup_is_root(memcg))
5861 page_counter_uncharge(&memcg->memory, 1);
5863 if (memcg != swap_memcg) {
5864 if (!mem_cgroup_is_root(swap_memcg))
5865 page_counter_charge(&swap_memcg->memsw, 1);
5866 page_counter_uncharge(&memcg->memsw, 1);
5870 * Interrupts should be disabled here because the caller holds the
5871 * mapping->tree_lock lock which is taken with interrupts-off. It is
5872 * important here to have the interrupts disabled because it is the
5873 * only synchronisation we have for udpating the per-CPU variables.
5875 VM_BUG_ON(!irqs_disabled());
5876 mem_cgroup_charge_statistics(memcg, page, false, -1);
5877 memcg_check_events(memcg, page);
5879 if (!mem_cgroup_is_root(memcg))
5880 css_put(&memcg->css);
5884 * mem_cgroup_try_charge_swap - try charging a swap entry
5885 * @page: page being added to swap
5886 * @entry: swap entry to charge
5888 * Try to charge @entry to the memcg that @page belongs to.
5890 * Returns 0 on success, -ENOMEM on failure.
5892 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5894 struct mem_cgroup *memcg;
5895 struct page_counter *counter;
5896 unsigned short oldid;
5898 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5901 memcg = page->mem_cgroup;
5903 /* Readahead page, never charged */
5907 memcg = mem_cgroup_id_get_online(memcg);
5909 if (!mem_cgroup_is_root(memcg) &&
5910 !page_counter_try_charge(&memcg->swap, 1, &counter)) {
5911 mem_cgroup_id_put(memcg);
5915 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5916 VM_BUG_ON_PAGE(oldid, page);
5917 mem_cgroup_swap_statistics(memcg, true);
5923 * mem_cgroup_uncharge_swap - uncharge a swap entry
5924 * @entry: swap entry to uncharge
5926 * Drop the swap charge associated with @entry.
5928 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5930 struct mem_cgroup *memcg;
5933 if (!do_swap_account)
5936 id = swap_cgroup_record(entry, 0);
5938 memcg = mem_cgroup_from_id(id);
5940 if (!mem_cgroup_is_root(memcg)) {
5941 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5942 page_counter_uncharge(&memcg->swap, 1);
5944 page_counter_uncharge(&memcg->memsw, 1);
5946 mem_cgroup_swap_statistics(memcg, false);
5947 mem_cgroup_id_put(memcg);
5952 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
5954 long nr_swap_pages = get_nr_swap_pages();
5956 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5957 return nr_swap_pages;
5958 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5959 nr_swap_pages = min_t(long, nr_swap_pages,
5960 READ_ONCE(memcg->swap.limit) -
5961 page_counter_read(&memcg->swap));
5962 return nr_swap_pages;
5965 bool mem_cgroup_swap_full(struct page *page)
5967 struct mem_cgroup *memcg;
5969 VM_BUG_ON_PAGE(!PageLocked(page), page);
5973 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
5976 memcg = page->mem_cgroup;
5980 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
5981 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
5987 /* for remember boot option*/
5988 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5989 static int really_do_swap_account __initdata = 1;
5991 static int really_do_swap_account __initdata;
5994 static int __init enable_swap_account(char *s)
5996 if (!strcmp(s, "1"))
5997 really_do_swap_account = 1;
5998 else if (!strcmp(s, "0"))
5999 really_do_swap_account = 0;
6002 __setup("swapaccount=", enable_swap_account);
6004 static u64 swap_current_read(struct cgroup_subsys_state *css,
6007 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6009 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6012 static int swap_max_show(struct seq_file *m, void *v)
6014 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6015 unsigned long max = READ_ONCE(memcg->swap.limit);
6017 if (max == PAGE_COUNTER_MAX)
6018 seq_puts(m, "max\n");
6020 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6025 static ssize_t swap_max_write(struct kernfs_open_file *of,
6026 char *buf, size_t nbytes, loff_t off)
6028 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6032 buf = strstrip(buf);
6033 err = page_counter_memparse(buf, "max", &max);
6037 mutex_lock(&memcg_limit_mutex);
6038 err = page_counter_limit(&memcg->swap, max);
6039 mutex_unlock(&memcg_limit_mutex);
6046 static struct cftype swap_files[] = {
6048 .name = "swap.current",
6049 .flags = CFTYPE_NOT_ON_ROOT,
6050 .read_u64 = swap_current_read,
6054 .flags = CFTYPE_NOT_ON_ROOT,
6055 .seq_show = swap_max_show,
6056 .write = swap_max_write,
6061 static struct cftype memsw_cgroup_files[] = {
6063 .name = "memsw.usage_in_bytes",
6064 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6065 .read_u64 = mem_cgroup_read_u64,
6068 .name = "memsw.max_usage_in_bytes",
6069 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6070 .write = mem_cgroup_reset,
6071 .read_u64 = mem_cgroup_read_u64,
6074 .name = "memsw.limit_in_bytes",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6076 .write = mem_cgroup_write,
6077 .read_u64 = mem_cgroup_read_u64,
6080 .name = "memsw.failcnt",
6081 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6082 .write = mem_cgroup_reset,
6083 .read_u64 = mem_cgroup_read_u64,
6085 { }, /* terminate */
6088 static int __init mem_cgroup_swap_init(void)
6090 if (!mem_cgroup_disabled() && really_do_swap_account) {
6091 do_swap_account = 1;
6092 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6094 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6095 memsw_cgroup_files));
6099 subsys_initcall(mem_cgroup_swap_init);
6101 #endif /* CONFIG_MEMCG_SWAP */