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/sched/mm.h>
39 #include <linux/shmem_fs.h>
40 #include <linux/hugetlb.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp.h>
43 #include <linux/page-flags.h>
44 #include <linux/backing-dev.h>
45 #include <linux/bit_spinlock.h>
46 #include <linux/rcupdate.h>
47 #include <linux/limits.h>
48 #include <linux/export.h>
49 #include <linux/mutex.h>
50 #include <linux/rbtree.h>
51 #include <linux/slab.h>
52 #include <linux/swap.h>
53 #include <linux/swapops.h>
54 #include <linux/spinlock.h>
55 #include <linux/eventfd.h>
56 #include <linux/poll.h>
57 #include <linux/sort.h>
59 #include <linux/seq_file.h>
60 #include <linux/vmpressure.h>
61 #include <linux/mm_inline.h>
62 #include <linux/swap_cgroup.h>
63 #include <linux/cpu.h>
64 #include <linux/oom.h>
65 #include <linux/lockdep.h>
66 #include <linux/file.h>
67 #include <linux/tracehook.h>
73 #include <linux/uaccess.h>
75 #include <trace/events/vmscan.h>
77 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
78 EXPORT_SYMBOL(memory_cgrp_subsys);
80 struct mem_cgroup *root_mem_cgroup __read_mostly;
82 #define MEM_CGROUP_RECLAIM_RETRIES 5
84 /* Socket memory accounting disabled? */
85 static bool cgroup_memory_nosocket;
87 /* Kernel memory accounting disabled? */
88 static bool cgroup_memory_nokmem;
90 /* Whether the swap controller is active */
91 #ifdef CONFIG_MEMCG_SWAP
92 int do_swap_account __read_mostly;
94 #define do_swap_account 0
97 /* Whether legacy memory+swap accounting is active */
98 static bool do_memsw_account(void)
100 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
103 static const char *const mem_cgroup_lru_names[] = {
111 #define THRESHOLDS_EVENTS_TARGET 128
112 #define SOFTLIMIT_EVENTS_TARGET 1024
113 #define NUMAINFO_EVENTS_TARGET 1024
116 * Cgroups above their limits are maintained in a RB-Tree, independent of
117 * their hierarchy representation
120 struct mem_cgroup_tree_per_node {
121 struct rb_root rb_root;
122 struct rb_node *rb_rightmost;
126 struct mem_cgroup_tree {
127 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
130 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
133 struct mem_cgroup_eventfd_list {
134 struct list_head list;
135 struct eventfd_ctx *eventfd;
139 * cgroup_event represents events which userspace want to receive.
141 struct mem_cgroup_event {
143 * memcg which the event belongs to.
145 struct mem_cgroup *memcg;
147 * eventfd to signal userspace about the event.
149 struct eventfd_ctx *eventfd;
151 * Each of these stored in a list by the cgroup.
153 struct list_head list;
155 * register_event() callback will be used to add new userspace
156 * waiter for changes related to this event. Use eventfd_signal()
157 * on eventfd to send notification to userspace.
159 int (*register_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd, const char *args);
162 * unregister_event() callback will be called when userspace closes
163 * the eventfd or on cgroup removing. This callback must be set,
164 * if you want provide notification functionality.
166 void (*unregister_event)(struct mem_cgroup *memcg,
167 struct eventfd_ctx *eventfd);
169 * All fields below needed to unregister event when
170 * userspace closes eventfd.
173 wait_queue_head_t *wqh;
174 wait_queue_entry_t wait;
175 struct work_struct remove;
178 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
179 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
181 /* Stuffs for move charges at task migration. */
183 * Types of charges to be moved.
185 #define MOVE_ANON 0x1U
186 #define MOVE_FILE 0x2U
187 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
189 /* "mc" and its members are protected by cgroup_mutex */
190 static struct move_charge_struct {
191 spinlock_t lock; /* for from, to */
192 struct mm_struct *mm;
193 struct mem_cgroup *from;
194 struct mem_cgroup *to;
196 unsigned long precharge;
197 unsigned long moved_charge;
198 unsigned long moved_swap;
199 struct task_struct *moving_task; /* a task moving charges */
200 wait_queue_head_t waitq; /* a waitq for other context */
202 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
203 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
207 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
208 * limit reclaim to prevent infinite loops, if they ever occur.
210 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
211 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
214 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
215 MEM_CGROUP_CHARGE_TYPE_ANON,
216 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
217 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
221 /* for encoding cft->private value on file */
230 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
231 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
232 #define MEMFILE_ATTR(val) ((val) & 0xffff)
233 /* Used for OOM nofiier */
234 #define OOM_CONTROL (0)
236 /* Some nice accessors for the vmpressure. */
237 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
240 memcg = root_mem_cgroup;
241 return &memcg->vmpressure;
244 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
246 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
249 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
251 return (memcg == root_mem_cgroup);
254 #ifdef CONFIG_MEMCG_KMEM
256 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
257 * The main reason for not using cgroup id for this:
258 * this works better in sparse environments, where we have a lot of memcgs,
259 * but only a few kmem-limited. Or also, if we have, for instance, 200
260 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
261 * 200 entry array for that.
263 * The current size of the caches array is stored in memcg_nr_cache_ids. It
264 * will double each time we have to increase it.
266 static DEFINE_IDA(memcg_cache_ida);
267 int memcg_nr_cache_ids;
269 /* Protects memcg_nr_cache_ids */
270 static DECLARE_RWSEM(memcg_cache_ids_sem);
272 void memcg_get_cache_ids(void)
274 down_read(&memcg_cache_ids_sem);
277 void memcg_put_cache_ids(void)
279 up_read(&memcg_cache_ids_sem);
283 * MIN_SIZE is different than 1, because we would like to avoid going through
284 * the alloc/free process all the time. In a small machine, 4 kmem-limited
285 * cgroups is a reasonable guess. In the future, it could be a parameter or
286 * tunable, but that is strictly not necessary.
288 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
289 * this constant directly from cgroup, but it is understandable that this is
290 * better kept as an internal representation in cgroup.c. In any case, the
291 * cgrp_id space is not getting any smaller, and we don't have to necessarily
292 * increase ours as well if it increases.
294 #define MEMCG_CACHES_MIN_SIZE 4
295 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
298 * A lot of the calls to the cache allocation functions are expected to be
299 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
300 * conditional to this static branch, we'll have to allow modules that does
301 * kmem_cache_alloc and the such to see this symbol as well
303 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
304 EXPORT_SYMBOL(memcg_kmem_enabled_key);
306 struct workqueue_struct *memcg_kmem_cache_wq;
308 #endif /* CONFIG_MEMCG_KMEM */
311 * mem_cgroup_css_from_page - css of the memcg associated with a page
312 * @page: page of interest
314 * If memcg is bound to the default hierarchy, css of the memcg associated
315 * with @page is returned. The returned css remains associated with @page
316 * until it is released.
318 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
321 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
323 struct mem_cgroup *memcg;
325 memcg = page->mem_cgroup;
327 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
328 memcg = root_mem_cgroup;
334 * page_cgroup_ino - return inode number of the memcg a page is charged to
337 * Look up the closest online ancestor of the memory cgroup @page is charged to
338 * and return its inode number or 0 if @page is not charged to any cgroup. It
339 * is safe to call this function without holding a reference to @page.
341 * Note, this function is inherently racy, because there is nothing to prevent
342 * the cgroup inode from getting torn down and potentially reallocated a moment
343 * after page_cgroup_ino() returns, so it only should be used by callers that
344 * do not care (such as procfs interfaces).
346 ino_t page_cgroup_ino(struct page *page)
348 struct mem_cgroup *memcg;
349 unsigned long ino = 0;
352 memcg = READ_ONCE(page->mem_cgroup);
353 while (memcg && !(memcg->css.flags & CSS_ONLINE))
354 memcg = parent_mem_cgroup(memcg);
356 ino = cgroup_ino(memcg->css.cgroup);
361 static struct mem_cgroup_per_node *
362 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
364 int nid = page_to_nid(page);
366 return memcg->nodeinfo[nid];
369 static struct mem_cgroup_tree_per_node *
370 soft_limit_tree_node(int nid)
372 return soft_limit_tree.rb_tree_per_node[nid];
375 static struct mem_cgroup_tree_per_node *
376 soft_limit_tree_from_page(struct page *page)
378 int nid = page_to_nid(page);
380 return soft_limit_tree.rb_tree_per_node[nid];
383 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
384 struct mem_cgroup_tree_per_node *mctz,
385 unsigned long new_usage_in_excess)
387 struct rb_node **p = &mctz->rb_root.rb_node;
388 struct rb_node *parent = NULL;
389 struct mem_cgroup_per_node *mz_node;
390 bool rightmost = true;
395 mz->usage_in_excess = new_usage_in_excess;
396 if (!mz->usage_in_excess)
400 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
402 if (mz->usage_in_excess < mz_node->usage_in_excess) {
408 * We can't avoid mem cgroups that are over their soft
409 * limit by the same amount
411 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
416 mctz->rb_rightmost = &mz->tree_node;
418 rb_link_node(&mz->tree_node, parent, p);
419 rb_insert_color(&mz->tree_node, &mctz->rb_root);
423 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
424 struct mem_cgroup_tree_per_node *mctz)
429 if (&mz->tree_node == mctz->rb_rightmost)
430 mctz->rb_rightmost = rb_prev(&mz->tree_node);
432 rb_erase(&mz->tree_node, &mctz->rb_root);
436 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
437 struct mem_cgroup_tree_per_node *mctz)
441 spin_lock_irqsave(&mctz->lock, flags);
442 __mem_cgroup_remove_exceeded(mz, mctz);
443 spin_unlock_irqrestore(&mctz->lock, flags);
446 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
448 unsigned long nr_pages = page_counter_read(&memcg->memory);
449 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
450 unsigned long excess = 0;
452 if (nr_pages > soft_limit)
453 excess = nr_pages - soft_limit;
458 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
460 unsigned long excess;
461 struct mem_cgroup_per_node *mz;
462 struct mem_cgroup_tree_per_node *mctz;
464 mctz = soft_limit_tree_from_page(page);
468 * Necessary to update all ancestors when hierarchy is used.
469 * because their event counter is not touched.
471 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
472 mz = mem_cgroup_page_nodeinfo(memcg, page);
473 excess = soft_limit_excess(memcg);
475 * We have to update the tree if mz is on RB-tree or
476 * mem is over its softlimit.
478 if (excess || mz->on_tree) {
481 spin_lock_irqsave(&mctz->lock, flags);
482 /* if on-tree, remove it */
484 __mem_cgroup_remove_exceeded(mz, mctz);
486 * Insert again. mz->usage_in_excess will be updated.
487 * If excess is 0, no tree ops.
489 __mem_cgroup_insert_exceeded(mz, mctz, excess);
490 spin_unlock_irqrestore(&mctz->lock, flags);
495 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
497 struct mem_cgroup_tree_per_node *mctz;
498 struct mem_cgroup_per_node *mz;
502 mz = mem_cgroup_nodeinfo(memcg, nid);
503 mctz = soft_limit_tree_node(nid);
505 mem_cgroup_remove_exceeded(mz, mctz);
509 static struct mem_cgroup_per_node *
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
512 struct mem_cgroup_per_node *mz;
516 if (!mctz->rb_rightmost)
517 goto done; /* Nothing to reclaim from */
519 mz = rb_entry(mctz->rb_rightmost,
520 struct mem_cgroup_per_node, tree_node);
522 * Remove the node now but someone else can add it back,
523 * we will to add it back at the end of reclaim to its correct
524 * position in the tree.
526 __mem_cgroup_remove_exceeded(mz, mctz);
527 if (!soft_limit_excess(mz->memcg) ||
528 !css_tryget_online(&mz->memcg->css))
534 static struct mem_cgroup_per_node *
535 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
537 struct mem_cgroup_per_node *mz;
539 spin_lock_irq(&mctz->lock);
540 mz = __mem_cgroup_largest_soft_limit_node(mctz);
541 spin_unlock_irq(&mctz->lock);
545 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
548 return atomic_long_read(&memcg->events[event]);
551 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
553 bool compound, int nr_pages)
556 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
557 * counted as CACHE even if it's on ANON LRU.
560 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
562 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
563 if (PageSwapBacked(page))
564 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
568 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
569 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
572 /* pagein of a big page is an event. So, ignore page size */
574 __count_memcg_events(memcg, PGPGIN, 1);
576 __count_memcg_events(memcg, PGPGOUT, 1);
577 nr_pages = -nr_pages; /* for event */
580 __this_cpu_add(memcg->stat_cpu->nr_page_events, nr_pages);
583 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
584 int nid, unsigned int lru_mask)
586 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
587 unsigned long nr = 0;
590 VM_BUG_ON((unsigned)nid >= nr_node_ids);
593 if (!(BIT(lru) & lru_mask))
595 nr += mem_cgroup_get_lru_size(lruvec, lru);
600 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
601 unsigned int lru_mask)
603 unsigned long nr = 0;
606 for_each_node_state(nid, N_MEMORY)
607 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
611 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
612 enum mem_cgroup_events_target target)
614 unsigned long val, next;
616 val = __this_cpu_read(memcg->stat_cpu->nr_page_events);
617 next = __this_cpu_read(memcg->stat_cpu->targets[target]);
618 /* from time_after() in jiffies.h */
619 if ((long)(next - val) < 0) {
621 case MEM_CGROUP_TARGET_THRESH:
622 next = val + THRESHOLDS_EVENTS_TARGET;
624 case MEM_CGROUP_TARGET_SOFTLIMIT:
625 next = val + SOFTLIMIT_EVENTS_TARGET;
627 case MEM_CGROUP_TARGET_NUMAINFO:
628 next = val + NUMAINFO_EVENTS_TARGET;
633 __this_cpu_write(memcg->stat_cpu->targets[target], next);
640 * Check events in order.
643 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
645 /* threshold event is triggered in finer grain than soft limit */
646 if (unlikely(mem_cgroup_event_ratelimit(memcg,
647 MEM_CGROUP_TARGET_THRESH))) {
649 bool do_numainfo __maybe_unused;
651 do_softlimit = mem_cgroup_event_ratelimit(memcg,
652 MEM_CGROUP_TARGET_SOFTLIMIT);
654 do_numainfo = mem_cgroup_event_ratelimit(memcg,
655 MEM_CGROUP_TARGET_NUMAINFO);
657 mem_cgroup_threshold(memcg);
658 if (unlikely(do_softlimit))
659 mem_cgroup_update_tree(memcg, page);
661 if (unlikely(do_numainfo))
662 atomic_inc(&memcg->numainfo_events);
667 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
670 * mm_update_next_owner() may clear mm->owner to NULL
671 * if it races with swapoff, page migration, etc.
672 * So this can be called with p == NULL.
677 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
679 EXPORT_SYMBOL(mem_cgroup_from_task);
682 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
683 * @mm: mm from which memcg should be extracted. It can be NULL.
685 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
686 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
689 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
691 struct mem_cgroup *memcg;
693 if (mem_cgroup_disabled())
699 * Page cache insertions can happen withou an
700 * actual mm context, e.g. during disk probing
701 * on boot, loopback IO, acct() writes etc.
704 memcg = root_mem_cgroup;
706 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
707 if (unlikely(!memcg))
708 memcg = root_mem_cgroup;
710 } while (!css_tryget_online(&memcg->css));
714 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
717 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
718 * @page: page from which memcg should be extracted.
720 * Obtain a reference on page->memcg and returns it if successful. Otherwise
721 * root_mem_cgroup is returned.
723 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
725 struct mem_cgroup *memcg = page->mem_cgroup;
727 if (mem_cgroup_disabled())
731 if (!memcg || !css_tryget_online(&memcg->css))
732 memcg = root_mem_cgroup;
736 EXPORT_SYMBOL(get_mem_cgroup_from_page);
739 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
741 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
743 if (unlikely(current->active_memcg)) {
744 struct mem_cgroup *memcg = root_mem_cgroup;
747 if (css_tryget_online(¤t->active_memcg->css))
748 memcg = current->active_memcg;
752 return get_mem_cgroup_from_mm(current->mm);
756 * mem_cgroup_iter - iterate over memory cgroup hierarchy
757 * @root: hierarchy root
758 * @prev: previously returned memcg, NULL on first invocation
759 * @reclaim: cookie for shared reclaim walks, NULL for full walks
761 * Returns references to children of the hierarchy below @root, or
762 * @root itself, or %NULL after a full round-trip.
764 * Caller must pass the return value in @prev on subsequent
765 * invocations for reference counting, or use mem_cgroup_iter_break()
766 * to cancel a hierarchy walk before the round-trip is complete.
768 * Reclaimers can specify a node and a priority level in @reclaim to
769 * divide up the memcgs in the hierarchy among all concurrent
770 * reclaimers operating on the same node and priority.
772 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
773 struct mem_cgroup *prev,
774 struct mem_cgroup_reclaim_cookie *reclaim)
776 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
777 struct cgroup_subsys_state *css = NULL;
778 struct mem_cgroup *memcg = NULL;
779 struct mem_cgroup *pos = NULL;
781 if (mem_cgroup_disabled())
785 root = root_mem_cgroup;
787 if (prev && !reclaim)
790 if (!root->use_hierarchy && root != root_mem_cgroup) {
799 struct mem_cgroup_per_node *mz;
801 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
802 iter = &mz->iter[reclaim->priority];
804 if (prev && reclaim->generation != iter->generation)
808 pos = READ_ONCE(iter->position);
809 if (!pos || css_tryget(&pos->css))
812 * css reference reached zero, so iter->position will
813 * be cleared by ->css_released. However, we should not
814 * rely on this happening soon, because ->css_released
815 * is called from a work queue, and by busy-waiting we
816 * might block it. So we clear iter->position right
819 (void)cmpxchg(&iter->position, pos, NULL);
827 css = css_next_descendant_pre(css, &root->css);
830 * Reclaimers share the hierarchy walk, and a
831 * new one might jump in right at the end of
832 * the hierarchy - make sure they see at least
833 * one group and restart from the beginning.
841 * Verify the css and acquire a reference. The root
842 * is provided by the caller, so we know it's alive
843 * and kicking, and don't take an extra reference.
845 memcg = mem_cgroup_from_css(css);
847 if (css == &root->css)
858 * The position could have already been updated by a competing
859 * thread, so check that the value hasn't changed since we read
860 * it to avoid reclaiming from the same cgroup twice.
862 (void)cmpxchg(&iter->position, pos, memcg);
870 reclaim->generation = iter->generation;
876 if (prev && prev != root)
883 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
884 * @root: hierarchy root
885 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
887 void mem_cgroup_iter_break(struct mem_cgroup *root,
888 struct mem_cgroup *prev)
891 root = root_mem_cgroup;
892 if (prev && prev != root)
896 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
898 struct mem_cgroup *memcg = dead_memcg;
899 struct mem_cgroup_reclaim_iter *iter;
900 struct mem_cgroup_per_node *mz;
904 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
906 mz = mem_cgroup_nodeinfo(memcg, nid);
907 for (i = 0; i <= DEF_PRIORITY; i++) {
909 cmpxchg(&iter->position,
917 * Iteration constructs for visiting all cgroups (under a tree). If
918 * loops are exited prematurely (break), mem_cgroup_iter_break() must
919 * be used for reference counting.
921 #define for_each_mem_cgroup_tree(iter, root) \
922 for (iter = mem_cgroup_iter(root, NULL, NULL); \
924 iter = mem_cgroup_iter(root, iter, NULL))
926 #define for_each_mem_cgroup(iter) \
927 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
929 iter = mem_cgroup_iter(NULL, iter, NULL))
932 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
933 * @memcg: hierarchy root
934 * @fn: function to call for each task
935 * @arg: argument passed to @fn
937 * This function iterates over tasks attached to @memcg or to any of its
938 * descendants and calls @fn for each task. If @fn returns a non-zero
939 * value, the function breaks the iteration loop and returns the value.
940 * Otherwise, it will iterate over all tasks and return 0.
942 * This function must not be called for the root memory cgroup.
944 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
945 int (*fn)(struct task_struct *, void *), void *arg)
947 struct mem_cgroup *iter;
950 BUG_ON(memcg == root_mem_cgroup);
952 for_each_mem_cgroup_tree(iter, memcg) {
953 struct css_task_iter it;
954 struct task_struct *task;
956 css_task_iter_start(&iter->css, 0, &it);
957 while (!ret && (task = css_task_iter_next(&it)))
959 css_task_iter_end(&it);
961 mem_cgroup_iter_break(memcg, iter);
969 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
971 * @pgdat: pgdat of the page
973 * This function is only safe when following the LRU page isolation
974 * and putback protocol: the LRU lock must be held, and the page must
975 * either be PageLRU() or the caller must have isolated/allocated it.
977 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
979 struct mem_cgroup_per_node *mz;
980 struct mem_cgroup *memcg;
981 struct lruvec *lruvec;
983 if (mem_cgroup_disabled()) {
984 lruvec = &pgdat->lruvec;
988 memcg = page->mem_cgroup;
990 * Swapcache readahead pages are added to the LRU - and
991 * possibly migrated - before they are charged.
994 memcg = root_mem_cgroup;
996 mz = mem_cgroup_page_nodeinfo(memcg, page);
997 lruvec = &mz->lruvec;
1000 * Since a node can be onlined after the mem_cgroup was created,
1001 * we have to be prepared to initialize lruvec->zone here;
1002 * and if offlined then reonlined, we need to reinitialize it.
1004 if (unlikely(lruvec->pgdat != pgdat))
1005 lruvec->pgdat = pgdat;
1010 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1011 * @lruvec: mem_cgroup per zone lru vector
1012 * @lru: index of lru list the page is sitting on
1013 * @zid: zone id of the accounted pages
1014 * @nr_pages: positive when adding or negative when removing
1016 * This function must be called under lru_lock, just before a page is added
1017 * to or just after a page is removed from an lru list (that ordering being
1018 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1020 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1021 int zid, int nr_pages)
1023 struct mem_cgroup_per_node *mz;
1024 unsigned long *lru_size;
1027 if (mem_cgroup_disabled())
1030 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1031 lru_size = &mz->lru_zone_size[zid][lru];
1034 *lru_size += nr_pages;
1037 if (WARN_ONCE(size < 0,
1038 "%s(%p, %d, %d): lru_size %ld\n",
1039 __func__, lruvec, lru, nr_pages, size)) {
1045 *lru_size += nr_pages;
1048 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1050 struct mem_cgroup *task_memcg;
1051 struct task_struct *p;
1054 p = find_lock_task_mm(task);
1056 task_memcg = get_mem_cgroup_from_mm(p->mm);
1060 * All threads may have already detached their mm's, but the oom
1061 * killer still needs to detect if they have already been oom
1062 * killed to prevent needlessly killing additional tasks.
1065 task_memcg = mem_cgroup_from_task(task);
1066 css_get(&task_memcg->css);
1069 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1070 css_put(&task_memcg->css);
1075 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1076 * @memcg: the memory cgroup
1078 * Returns the maximum amount of memory @mem can be charged with, in
1081 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1083 unsigned long margin = 0;
1084 unsigned long count;
1085 unsigned long limit;
1087 count = page_counter_read(&memcg->memory);
1088 limit = READ_ONCE(memcg->memory.max);
1090 margin = limit - count;
1092 if (do_memsw_account()) {
1093 count = page_counter_read(&memcg->memsw);
1094 limit = READ_ONCE(memcg->memsw.max);
1096 margin = min(margin, limit - count);
1105 * A routine for checking "mem" is under move_account() or not.
1107 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1108 * moving cgroups. This is for waiting at high-memory pressure
1111 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1113 struct mem_cgroup *from;
1114 struct mem_cgroup *to;
1117 * Unlike task_move routines, we access mc.to, mc.from not under
1118 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1120 spin_lock(&mc.lock);
1126 ret = mem_cgroup_is_descendant(from, memcg) ||
1127 mem_cgroup_is_descendant(to, memcg);
1129 spin_unlock(&mc.lock);
1133 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1135 if (mc.moving_task && current != mc.moving_task) {
1136 if (mem_cgroup_under_move(memcg)) {
1138 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1139 /* moving charge context might have finished. */
1142 finish_wait(&mc.waitq, &wait);
1149 static const unsigned int memcg1_stats[] = {
1160 static const char *const memcg1_stat_names[] = {
1171 #define K(x) ((x) << (PAGE_SHIFT-10))
1173 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1174 * @memcg: The memory cgroup that went over limit
1175 * @p: Task that is going to be killed
1177 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1180 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1182 struct mem_cgroup *iter;
1188 pr_info("Task in ");
1189 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1190 pr_cont(" killed as a result of limit of ");
1192 pr_info("Memory limit reached of cgroup ");
1195 pr_cont_cgroup_path(memcg->css.cgroup);
1200 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1201 K((u64)page_counter_read(&memcg->memory)),
1202 K((u64)memcg->memory.max), memcg->memory.failcnt);
1203 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1204 K((u64)page_counter_read(&memcg->memsw)),
1205 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1206 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1207 K((u64)page_counter_read(&memcg->kmem)),
1208 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1210 for_each_mem_cgroup_tree(iter, memcg) {
1211 pr_info("Memory cgroup stats for ");
1212 pr_cont_cgroup_path(iter->css.cgroup);
1215 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1216 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1218 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1219 K(memcg_page_state(iter, memcg1_stats[i])));
1222 for (i = 0; i < NR_LRU_LISTS; i++)
1223 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1224 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1231 * Return the memory (and swap, if configured) limit for a memcg.
1233 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1237 max = memcg->memory.max;
1238 if (mem_cgroup_swappiness(memcg)) {
1239 unsigned long memsw_max;
1240 unsigned long swap_max;
1242 memsw_max = memcg->memsw.max;
1243 swap_max = memcg->swap.max;
1244 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1245 max = min(max + swap_max, memsw_max);
1250 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1253 struct oom_control oc = {
1257 .gfp_mask = gfp_mask,
1262 mutex_lock(&oom_lock);
1263 ret = out_of_memory(&oc);
1264 mutex_unlock(&oom_lock);
1268 #if MAX_NUMNODES > 1
1271 * test_mem_cgroup_node_reclaimable
1272 * @memcg: the target memcg
1273 * @nid: the node ID to be checked.
1274 * @noswap : specify true here if the user wants flle only information.
1276 * This function returns whether the specified memcg contains any
1277 * reclaimable pages on a node. Returns true if there are any reclaimable
1278 * pages in the node.
1280 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1281 int nid, bool noswap)
1283 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1285 if (noswap || !total_swap_pages)
1287 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1294 * Always updating the nodemask is not very good - even if we have an empty
1295 * list or the wrong list here, we can start from some node and traverse all
1296 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1299 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1303 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1304 * pagein/pageout changes since the last update.
1306 if (!atomic_read(&memcg->numainfo_events))
1308 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1311 /* make a nodemask where this memcg uses memory from */
1312 memcg->scan_nodes = node_states[N_MEMORY];
1314 for_each_node_mask(nid, node_states[N_MEMORY]) {
1316 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1317 node_clear(nid, memcg->scan_nodes);
1320 atomic_set(&memcg->numainfo_events, 0);
1321 atomic_set(&memcg->numainfo_updating, 0);
1325 * Selecting a node where we start reclaim from. Because what we need is just
1326 * reducing usage counter, start from anywhere is O,K. Considering
1327 * memory reclaim from current node, there are pros. and cons.
1329 * Freeing memory from current node means freeing memory from a node which
1330 * we'll use or we've used. So, it may make LRU bad. And if several threads
1331 * hit limits, it will see a contention on a node. But freeing from remote
1332 * node means more costs for memory reclaim because of memory latency.
1334 * Now, we use round-robin. Better algorithm is welcomed.
1336 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1340 mem_cgroup_may_update_nodemask(memcg);
1341 node = memcg->last_scanned_node;
1343 node = next_node_in(node, memcg->scan_nodes);
1345 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1346 * last time it really checked all the LRUs due to rate limiting.
1347 * Fallback to the current node in that case for simplicity.
1349 if (unlikely(node == MAX_NUMNODES))
1350 node = numa_node_id();
1352 memcg->last_scanned_node = node;
1356 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1362 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1365 unsigned long *total_scanned)
1367 struct mem_cgroup *victim = NULL;
1370 unsigned long excess;
1371 unsigned long nr_scanned;
1372 struct mem_cgroup_reclaim_cookie reclaim = {
1377 excess = soft_limit_excess(root_memcg);
1380 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1385 * If we have not been able to reclaim
1386 * anything, it might because there are
1387 * no reclaimable pages under this hierarchy
1392 * We want to do more targeted reclaim.
1393 * excess >> 2 is not to excessive so as to
1394 * reclaim too much, nor too less that we keep
1395 * coming back to reclaim from this cgroup
1397 if (total >= (excess >> 2) ||
1398 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1403 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1404 pgdat, &nr_scanned);
1405 *total_scanned += nr_scanned;
1406 if (!soft_limit_excess(root_memcg))
1409 mem_cgroup_iter_break(root_memcg, victim);
1413 #ifdef CONFIG_LOCKDEP
1414 static struct lockdep_map memcg_oom_lock_dep_map = {
1415 .name = "memcg_oom_lock",
1419 static DEFINE_SPINLOCK(memcg_oom_lock);
1422 * Check OOM-Killer is already running under our hierarchy.
1423 * If someone is running, return false.
1425 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1427 struct mem_cgroup *iter, *failed = NULL;
1429 spin_lock(&memcg_oom_lock);
1431 for_each_mem_cgroup_tree(iter, memcg) {
1432 if (iter->oom_lock) {
1434 * this subtree of our hierarchy is already locked
1435 * so we cannot give a lock.
1438 mem_cgroup_iter_break(memcg, iter);
1441 iter->oom_lock = true;
1446 * OK, we failed to lock the whole subtree so we have
1447 * to clean up what we set up to the failing subtree
1449 for_each_mem_cgroup_tree(iter, memcg) {
1450 if (iter == failed) {
1451 mem_cgroup_iter_break(memcg, iter);
1454 iter->oom_lock = false;
1457 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1459 spin_unlock(&memcg_oom_lock);
1464 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1466 struct mem_cgroup *iter;
1468 spin_lock(&memcg_oom_lock);
1469 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1470 for_each_mem_cgroup_tree(iter, memcg)
1471 iter->oom_lock = false;
1472 spin_unlock(&memcg_oom_lock);
1475 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1477 struct mem_cgroup *iter;
1479 spin_lock(&memcg_oom_lock);
1480 for_each_mem_cgroup_tree(iter, memcg)
1482 spin_unlock(&memcg_oom_lock);
1485 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1487 struct mem_cgroup *iter;
1490 * When a new child is created while the hierarchy is under oom,
1491 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1493 spin_lock(&memcg_oom_lock);
1494 for_each_mem_cgroup_tree(iter, memcg)
1495 if (iter->under_oom > 0)
1497 spin_unlock(&memcg_oom_lock);
1500 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1502 struct oom_wait_info {
1503 struct mem_cgroup *memcg;
1504 wait_queue_entry_t wait;
1507 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1508 unsigned mode, int sync, void *arg)
1510 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1511 struct mem_cgroup *oom_wait_memcg;
1512 struct oom_wait_info *oom_wait_info;
1514 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1515 oom_wait_memcg = oom_wait_info->memcg;
1517 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1518 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1520 return autoremove_wake_function(wait, mode, sync, arg);
1523 static void memcg_oom_recover(struct mem_cgroup *memcg)
1526 * For the following lockless ->under_oom test, the only required
1527 * guarantee is that it must see the state asserted by an OOM when
1528 * this function is called as a result of userland actions
1529 * triggered by the notification of the OOM. This is trivially
1530 * achieved by invoking mem_cgroup_mark_under_oom() before
1531 * triggering notification.
1533 if (memcg && memcg->under_oom)
1534 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1544 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1546 if (order > PAGE_ALLOC_COSTLY_ORDER)
1550 * We are in the middle of the charge context here, so we
1551 * don't want to block when potentially sitting on a callstack
1552 * that holds all kinds of filesystem and mm locks.
1554 * cgroup1 allows disabling the OOM killer and waiting for outside
1555 * handling until the charge can succeed; remember the context and put
1556 * the task to sleep at the end of the page fault when all locks are
1559 * On the other hand, in-kernel OOM killer allows for an async victim
1560 * memory reclaim (oom_reaper) and that means that we are not solely
1561 * relying on the oom victim to make a forward progress and we can
1562 * invoke the oom killer here.
1564 * Please note that mem_cgroup_out_of_memory might fail to find a
1565 * victim and then we have to bail out from the charge path.
1567 if (memcg->oom_kill_disable) {
1568 if (!current->in_user_fault)
1570 css_get(&memcg->css);
1571 current->memcg_in_oom = memcg;
1572 current->memcg_oom_gfp_mask = mask;
1573 current->memcg_oom_order = order;
1578 if (mem_cgroup_out_of_memory(memcg, mask, order))
1581 WARN(1,"Memory cgroup charge failed because of no reclaimable memory! "
1582 "This looks like a misconfiguration or a kernel bug.");
1587 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1588 * @handle: actually kill/wait or just clean up the OOM state
1590 * This has to be called at the end of a page fault if the memcg OOM
1591 * handler was enabled.
1593 * Memcg supports userspace OOM handling where failed allocations must
1594 * sleep on a waitqueue until the userspace task resolves the
1595 * situation. Sleeping directly in the charge context with all kinds
1596 * of locks held is not a good idea, instead we remember an OOM state
1597 * in the task and mem_cgroup_oom_synchronize() has to be called at
1598 * the end of the page fault to complete the OOM handling.
1600 * Returns %true if an ongoing memcg OOM situation was detected and
1601 * completed, %false otherwise.
1603 bool mem_cgroup_oom_synchronize(bool handle)
1605 struct mem_cgroup *memcg = current->memcg_in_oom;
1606 struct oom_wait_info owait;
1609 /* OOM is global, do not handle */
1616 owait.memcg = memcg;
1617 owait.wait.flags = 0;
1618 owait.wait.func = memcg_oom_wake_function;
1619 owait.wait.private = current;
1620 INIT_LIST_HEAD(&owait.wait.entry);
1622 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1623 mem_cgroup_mark_under_oom(memcg);
1625 locked = mem_cgroup_oom_trylock(memcg);
1628 mem_cgroup_oom_notify(memcg);
1630 if (locked && !memcg->oom_kill_disable) {
1631 mem_cgroup_unmark_under_oom(memcg);
1632 finish_wait(&memcg_oom_waitq, &owait.wait);
1633 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1634 current->memcg_oom_order);
1637 mem_cgroup_unmark_under_oom(memcg);
1638 finish_wait(&memcg_oom_waitq, &owait.wait);
1642 mem_cgroup_oom_unlock(memcg);
1644 * There is no guarantee that an OOM-lock contender
1645 * sees the wakeups triggered by the OOM kill
1646 * uncharges. Wake any sleepers explicitely.
1648 memcg_oom_recover(memcg);
1651 current->memcg_in_oom = NULL;
1652 css_put(&memcg->css);
1657 * lock_page_memcg - lock a page->mem_cgroup binding
1660 * This function protects unlocked LRU pages from being moved to
1663 * It ensures lifetime of the returned memcg. Caller is responsible
1664 * for the lifetime of the page; __unlock_page_memcg() is available
1665 * when @page might get freed inside the locked section.
1667 struct mem_cgroup *lock_page_memcg(struct page *page)
1669 struct mem_cgroup *memcg;
1670 unsigned long flags;
1673 * The RCU lock is held throughout the transaction. The fast
1674 * path can get away without acquiring the memcg->move_lock
1675 * because page moving starts with an RCU grace period.
1677 * The RCU lock also protects the memcg from being freed when
1678 * the page state that is going to change is the only thing
1679 * preventing the page itself from being freed. E.g. writeback
1680 * doesn't hold a page reference and relies on PG_writeback to
1681 * keep off truncation, migration and so forth.
1685 if (mem_cgroup_disabled())
1688 memcg = page->mem_cgroup;
1689 if (unlikely(!memcg))
1692 if (atomic_read(&memcg->moving_account) <= 0)
1695 spin_lock_irqsave(&memcg->move_lock, flags);
1696 if (memcg != page->mem_cgroup) {
1697 spin_unlock_irqrestore(&memcg->move_lock, flags);
1702 * When charge migration first begins, we can have locked and
1703 * unlocked page stat updates happening concurrently. Track
1704 * the task who has the lock for unlock_page_memcg().
1706 memcg->move_lock_task = current;
1707 memcg->move_lock_flags = flags;
1711 EXPORT_SYMBOL(lock_page_memcg);
1714 * __unlock_page_memcg - unlock and unpin a memcg
1717 * Unlock and unpin a memcg returned by lock_page_memcg().
1719 void __unlock_page_memcg(struct mem_cgroup *memcg)
1721 if (memcg && memcg->move_lock_task == current) {
1722 unsigned long flags = memcg->move_lock_flags;
1724 memcg->move_lock_task = NULL;
1725 memcg->move_lock_flags = 0;
1727 spin_unlock_irqrestore(&memcg->move_lock, flags);
1734 * unlock_page_memcg - unlock a page->mem_cgroup binding
1737 void unlock_page_memcg(struct page *page)
1739 __unlock_page_memcg(page->mem_cgroup);
1741 EXPORT_SYMBOL(unlock_page_memcg);
1743 struct memcg_stock_pcp {
1744 struct mem_cgroup *cached; /* this never be root cgroup */
1745 unsigned int nr_pages;
1746 struct work_struct work;
1747 unsigned long flags;
1748 #define FLUSHING_CACHED_CHARGE 0
1750 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1751 static DEFINE_MUTEX(percpu_charge_mutex);
1754 * consume_stock: Try to consume stocked charge on this cpu.
1755 * @memcg: memcg to consume from.
1756 * @nr_pages: how many pages to charge.
1758 * The charges will only happen if @memcg matches the current cpu's memcg
1759 * stock, and at least @nr_pages are available in that stock. Failure to
1760 * service an allocation will refill the stock.
1762 * returns true if successful, false otherwise.
1764 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1766 struct memcg_stock_pcp *stock;
1767 unsigned long flags;
1770 if (nr_pages > MEMCG_CHARGE_BATCH)
1773 local_irq_save(flags);
1775 stock = this_cpu_ptr(&memcg_stock);
1776 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1777 stock->nr_pages -= nr_pages;
1781 local_irq_restore(flags);
1787 * Returns stocks cached in percpu and reset cached information.
1789 static void drain_stock(struct memcg_stock_pcp *stock)
1791 struct mem_cgroup *old = stock->cached;
1793 if (stock->nr_pages) {
1794 page_counter_uncharge(&old->memory, stock->nr_pages);
1795 if (do_memsw_account())
1796 page_counter_uncharge(&old->memsw, stock->nr_pages);
1797 css_put_many(&old->css, stock->nr_pages);
1798 stock->nr_pages = 0;
1800 stock->cached = NULL;
1803 static void drain_local_stock(struct work_struct *dummy)
1805 struct memcg_stock_pcp *stock;
1806 unsigned long flags;
1809 * The only protection from memory hotplug vs. drain_stock races is
1810 * that we always operate on local CPU stock here with IRQ disabled
1812 local_irq_save(flags);
1814 stock = this_cpu_ptr(&memcg_stock);
1816 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1818 local_irq_restore(flags);
1822 * Cache charges(val) to local per_cpu area.
1823 * This will be consumed by consume_stock() function, later.
1825 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1827 struct memcg_stock_pcp *stock;
1828 unsigned long flags;
1830 local_irq_save(flags);
1832 stock = this_cpu_ptr(&memcg_stock);
1833 if (stock->cached != memcg) { /* reset if necessary */
1835 stock->cached = memcg;
1837 stock->nr_pages += nr_pages;
1839 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
1842 local_irq_restore(flags);
1846 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1847 * of the hierarchy under it.
1849 static void drain_all_stock(struct mem_cgroup *root_memcg)
1853 /* If someone's already draining, avoid adding running more workers. */
1854 if (!mutex_trylock(&percpu_charge_mutex))
1857 * Notify other cpus that system-wide "drain" is running
1858 * We do not care about races with the cpu hotplug because cpu down
1859 * as well as workers from this path always operate on the local
1860 * per-cpu data. CPU up doesn't touch memcg_stock at all.
1863 for_each_online_cpu(cpu) {
1864 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1865 struct mem_cgroup *memcg;
1867 memcg = stock->cached;
1868 if (!memcg || !stock->nr_pages || !css_tryget(&memcg->css))
1870 if (!mem_cgroup_is_descendant(memcg, root_memcg)) {
1871 css_put(&memcg->css);
1874 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1876 drain_local_stock(&stock->work);
1878 schedule_work_on(cpu, &stock->work);
1880 css_put(&memcg->css);
1883 mutex_unlock(&percpu_charge_mutex);
1886 static int memcg_hotplug_cpu_dead(unsigned int cpu)
1888 struct memcg_stock_pcp *stock;
1889 struct mem_cgroup *memcg;
1891 stock = &per_cpu(memcg_stock, cpu);
1894 for_each_mem_cgroup(memcg) {
1897 for (i = 0; i < MEMCG_NR_STAT; i++) {
1901 x = this_cpu_xchg(memcg->stat_cpu->count[i], 0);
1903 atomic_long_add(x, &memcg->stat[i]);
1905 if (i >= NR_VM_NODE_STAT_ITEMS)
1908 for_each_node(nid) {
1909 struct mem_cgroup_per_node *pn;
1911 pn = mem_cgroup_nodeinfo(memcg, nid);
1912 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
1914 atomic_long_add(x, &pn->lruvec_stat[i]);
1918 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
1921 x = this_cpu_xchg(memcg->stat_cpu->events[i], 0);
1923 atomic_long_add(x, &memcg->events[i]);
1930 static void reclaim_high(struct mem_cgroup *memcg,
1931 unsigned int nr_pages,
1935 if (page_counter_read(&memcg->memory) <= memcg->high)
1937 memcg_memory_event(memcg, MEMCG_HIGH);
1938 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1939 } while ((memcg = parent_mem_cgroup(memcg)));
1942 static void high_work_func(struct work_struct *work)
1944 struct mem_cgroup *memcg;
1946 memcg = container_of(work, struct mem_cgroup, high_work);
1947 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
1951 * Scheduled by try_charge() to be executed from the userland return path
1952 * and reclaims memory over the high limit.
1954 void mem_cgroup_handle_over_high(void)
1956 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1957 struct mem_cgroup *memcg;
1959 if (likely(!nr_pages))
1962 memcg = get_mem_cgroup_from_mm(current->mm);
1963 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1964 css_put(&memcg->css);
1965 current->memcg_nr_pages_over_high = 0;
1968 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1969 unsigned int nr_pages)
1971 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
1972 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1973 struct mem_cgroup *mem_over_limit;
1974 struct page_counter *counter;
1975 unsigned long nr_reclaimed;
1976 bool may_swap = true;
1977 bool drained = false;
1979 enum oom_status oom_status;
1981 if (mem_cgroup_is_root(memcg))
1984 if (consume_stock(memcg, nr_pages))
1987 if (!do_memsw_account() ||
1988 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1989 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1991 if (do_memsw_account())
1992 page_counter_uncharge(&memcg->memsw, batch);
1993 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1995 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1999 if (batch > nr_pages) {
2005 * Unlike in global OOM situations, memcg is not in a physical
2006 * memory shortage. Allow dying and OOM-killed tasks to
2007 * bypass the last charges so that they can exit quickly and
2008 * free their memory.
2010 if (unlikely(tsk_is_oom_victim(current) ||
2011 fatal_signal_pending(current) ||
2012 current->flags & PF_EXITING))
2016 * Prevent unbounded recursion when reclaim operations need to
2017 * allocate memory. This might exceed the limits temporarily,
2018 * but we prefer facilitating memory reclaim and getting back
2019 * under the limit over triggering OOM kills in these cases.
2021 if (unlikely(current->flags & PF_MEMALLOC))
2024 if (unlikely(task_in_memcg_oom(current)))
2027 if (!gfpflags_allow_blocking(gfp_mask))
2030 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2032 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2033 gfp_mask, may_swap);
2035 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2039 drain_all_stock(mem_over_limit);
2044 if (gfp_mask & __GFP_NORETRY)
2047 * Even though the limit is exceeded at this point, reclaim
2048 * may have been able to free some pages. Retry the charge
2049 * before killing the task.
2051 * Only for regular pages, though: huge pages are rather
2052 * unlikely to succeed so close to the limit, and we fall back
2053 * to regular pages anyway in case of failure.
2055 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2058 * At task move, charge accounts can be doubly counted. So, it's
2059 * better to wait until the end of task_move if something is going on.
2061 if (mem_cgroup_wait_acct_move(mem_over_limit))
2067 if (gfp_mask & __GFP_RETRY_MAYFAIL && oomed)
2070 if (gfp_mask & __GFP_NOFAIL)
2073 if (fatal_signal_pending(current))
2076 memcg_memory_event(mem_over_limit, MEMCG_OOM);
2079 * keep retrying as long as the memcg oom killer is able to make
2080 * a forward progress or bypass the charge if the oom killer
2081 * couldn't make any progress.
2083 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2084 get_order(nr_pages * PAGE_SIZE));
2085 switch (oom_status) {
2087 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2096 if (!(gfp_mask & __GFP_NOFAIL))
2100 * The allocation either can't fail or will lead to more memory
2101 * being freed very soon. Allow memory usage go over the limit
2102 * temporarily by force charging it.
2104 page_counter_charge(&memcg->memory, nr_pages);
2105 if (do_memsw_account())
2106 page_counter_charge(&memcg->memsw, nr_pages);
2107 css_get_many(&memcg->css, nr_pages);
2112 css_get_many(&memcg->css, batch);
2113 if (batch > nr_pages)
2114 refill_stock(memcg, batch - nr_pages);
2117 * If the hierarchy is above the normal consumption range, schedule
2118 * reclaim on returning to userland. We can perform reclaim here
2119 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2120 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2121 * not recorded as it most likely matches current's and won't
2122 * change in the meantime. As high limit is checked again before
2123 * reclaim, the cost of mismatch is negligible.
2126 if (page_counter_read(&memcg->memory) > memcg->high) {
2127 /* Don't bother a random interrupted task */
2128 if (in_interrupt()) {
2129 schedule_work(&memcg->high_work);
2132 current->memcg_nr_pages_over_high += batch;
2133 set_notify_resume(current);
2136 } while ((memcg = parent_mem_cgroup(memcg)));
2141 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2143 if (mem_cgroup_is_root(memcg))
2146 page_counter_uncharge(&memcg->memory, nr_pages);
2147 if (do_memsw_account())
2148 page_counter_uncharge(&memcg->memsw, nr_pages);
2150 css_put_many(&memcg->css, nr_pages);
2153 static void lock_page_lru(struct page *page, int *isolated)
2155 struct zone *zone = page_zone(page);
2157 spin_lock_irq(zone_lru_lock(zone));
2158 if (PageLRU(page)) {
2159 struct lruvec *lruvec;
2161 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2163 del_page_from_lru_list(page, lruvec, page_lru(page));
2169 static void unlock_page_lru(struct page *page, int isolated)
2171 struct zone *zone = page_zone(page);
2174 struct lruvec *lruvec;
2176 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2177 VM_BUG_ON_PAGE(PageLRU(page), page);
2179 add_page_to_lru_list(page, lruvec, page_lru(page));
2181 spin_unlock_irq(zone_lru_lock(zone));
2184 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2189 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2192 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2193 * may already be on some other mem_cgroup's LRU. Take care of it.
2196 lock_page_lru(page, &isolated);
2199 * Nobody should be changing or seriously looking at
2200 * page->mem_cgroup at this point:
2202 * - the page is uncharged
2204 * - the page is off-LRU
2206 * - an anonymous fault has exclusive page access, except for
2207 * a locked page table
2209 * - a page cache insertion, a swapin fault, or a migration
2210 * have the page locked
2212 page->mem_cgroup = memcg;
2215 unlock_page_lru(page, isolated);
2218 #ifdef CONFIG_MEMCG_KMEM
2219 static int memcg_alloc_cache_id(void)
2224 id = ida_simple_get(&memcg_cache_ida,
2225 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2229 if (id < memcg_nr_cache_ids)
2233 * There's no space for the new id in memcg_caches arrays,
2234 * so we have to grow them.
2236 down_write(&memcg_cache_ids_sem);
2238 size = 2 * (id + 1);
2239 if (size < MEMCG_CACHES_MIN_SIZE)
2240 size = MEMCG_CACHES_MIN_SIZE;
2241 else if (size > MEMCG_CACHES_MAX_SIZE)
2242 size = MEMCG_CACHES_MAX_SIZE;
2244 err = memcg_update_all_caches(size);
2246 err = memcg_update_all_list_lrus(size);
2248 memcg_nr_cache_ids = size;
2250 up_write(&memcg_cache_ids_sem);
2253 ida_simple_remove(&memcg_cache_ida, id);
2259 static void memcg_free_cache_id(int id)
2261 ida_simple_remove(&memcg_cache_ida, id);
2264 struct memcg_kmem_cache_create_work {
2265 struct mem_cgroup *memcg;
2266 struct kmem_cache *cachep;
2267 struct work_struct work;
2270 static void memcg_kmem_cache_create_func(struct work_struct *w)
2272 struct memcg_kmem_cache_create_work *cw =
2273 container_of(w, struct memcg_kmem_cache_create_work, work);
2274 struct mem_cgroup *memcg = cw->memcg;
2275 struct kmem_cache *cachep = cw->cachep;
2277 memcg_create_kmem_cache(memcg, cachep);
2279 css_put(&memcg->css);
2284 * Enqueue the creation of a per-memcg kmem_cache.
2286 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2287 struct kmem_cache *cachep)
2289 struct memcg_kmem_cache_create_work *cw;
2291 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2295 css_get(&memcg->css);
2298 cw->cachep = cachep;
2299 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2301 queue_work(memcg_kmem_cache_wq, &cw->work);
2304 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2305 struct kmem_cache *cachep)
2308 * We need to stop accounting when we kmalloc, because if the
2309 * corresponding kmalloc cache is not yet created, the first allocation
2310 * in __memcg_schedule_kmem_cache_create will recurse.
2312 * However, it is better to enclose the whole function. Depending on
2313 * the debugging options enabled, INIT_WORK(), for instance, can
2314 * trigger an allocation. This too, will make us recurse. Because at
2315 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2316 * the safest choice is to do it like this, wrapping the whole function.
2318 current->memcg_kmem_skip_account = 1;
2319 __memcg_schedule_kmem_cache_create(memcg, cachep);
2320 current->memcg_kmem_skip_account = 0;
2323 static inline bool memcg_kmem_bypass(void)
2325 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2331 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2332 * @cachep: the original global kmem cache
2334 * Return the kmem_cache we're supposed to use for a slab allocation.
2335 * We try to use the current memcg's version of the cache.
2337 * If the cache does not exist yet, if we are the first user of it, we
2338 * create it asynchronously in a workqueue and let the current allocation
2339 * go through with the original cache.
2341 * This function takes a reference to the cache it returns to assure it
2342 * won't get destroyed while we are working with it. Once the caller is
2343 * done with it, memcg_kmem_put_cache() must be called to release the
2346 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2348 struct mem_cgroup *memcg;
2349 struct kmem_cache *memcg_cachep;
2352 VM_BUG_ON(!is_root_cache(cachep));
2354 if (memcg_kmem_bypass())
2357 if (current->memcg_kmem_skip_account)
2360 memcg = get_mem_cgroup_from_current();
2361 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2365 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2366 if (likely(memcg_cachep))
2367 return memcg_cachep;
2370 * If we are in a safe context (can wait, and not in interrupt
2371 * context), we could be be predictable and return right away.
2372 * This would guarantee that the allocation being performed
2373 * already belongs in the new cache.
2375 * However, there are some clashes that can arrive from locking.
2376 * For instance, because we acquire the slab_mutex while doing
2377 * memcg_create_kmem_cache, this means no further allocation
2378 * could happen with the slab_mutex held. So it's better to
2381 memcg_schedule_kmem_cache_create(memcg, cachep);
2383 css_put(&memcg->css);
2388 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2389 * @cachep: the cache returned by memcg_kmem_get_cache
2391 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2393 if (!is_root_cache(cachep))
2394 css_put(&cachep->memcg_params.memcg->css);
2398 * memcg_kmem_charge_memcg: charge a kmem page
2399 * @page: page to charge
2400 * @gfp: reclaim mode
2401 * @order: allocation order
2402 * @memcg: memory cgroup to charge
2404 * Returns 0 on success, an error code on failure.
2406 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2407 struct mem_cgroup *memcg)
2409 unsigned int nr_pages = 1 << order;
2410 struct page_counter *counter;
2413 ret = try_charge(memcg, gfp, nr_pages);
2417 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2418 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2419 cancel_charge(memcg, nr_pages);
2423 page->mem_cgroup = memcg;
2429 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2430 * @page: page to charge
2431 * @gfp: reclaim mode
2432 * @order: allocation order
2434 * Returns 0 on success, an error code on failure.
2436 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2438 struct mem_cgroup *memcg;
2441 if (memcg_kmem_bypass())
2444 memcg = get_mem_cgroup_from_current();
2445 if (!mem_cgroup_is_root(memcg)) {
2446 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2448 __SetPageKmemcg(page);
2450 css_put(&memcg->css);
2454 * memcg_kmem_uncharge: uncharge a kmem page
2455 * @page: page to uncharge
2456 * @order: allocation order
2458 void memcg_kmem_uncharge(struct page *page, int order)
2460 struct mem_cgroup *memcg = page->mem_cgroup;
2461 unsigned int nr_pages = 1 << order;
2466 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2468 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2469 page_counter_uncharge(&memcg->kmem, nr_pages);
2471 page_counter_uncharge(&memcg->memory, nr_pages);
2472 if (do_memsw_account())
2473 page_counter_uncharge(&memcg->memsw, nr_pages);
2475 page->mem_cgroup = NULL;
2477 /* slab pages do not have PageKmemcg flag set */
2478 if (PageKmemcg(page))
2479 __ClearPageKmemcg(page);
2481 css_put_many(&memcg->css, nr_pages);
2483 #endif /* CONFIG_MEMCG_KMEM */
2485 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2488 * Because tail pages are not marked as "used", set it. We're under
2489 * zone_lru_lock and migration entries setup in all page mappings.
2491 void mem_cgroup_split_huge_fixup(struct page *head)
2495 if (mem_cgroup_disabled())
2498 for (i = 1; i < HPAGE_PMD_NR; i++)
2499 head[i].mem_cgroup = head->mem_cgroup;
2501 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2503 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2505 #ifdef CONFIG_MEMCG_SWAP
2507 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2508 * @entry: swap entry to be moved
2509 * @from: mem_cgroup which the entry is moved from
2510 * @to: mem_cgroup which the entry is moved to
2512 * It succeeds only when the swap_cgroup's record for this entry is the same
2513 * as the mem_cgroup's id of @from.
2515 * Returns 0 on success, -EINVAL on failure.
2517 * The caller must have charged to @to, IOW, called page_counter_charge() about
2518 * both res and memsw, and called css_get().
2520 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2521 struct mem_cgroup *from, struct mem_cgroup *to)
2523 unsigned short old_id, new_id;
2525 old_id = mem_cgroup_id(from);
2526 new_id = mem_cgroup_id(to);
2528 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2529 mod_memcg_state(from, MEMCG_SWAP, -1);
2530 mod_memcg_state(to, MEMCG_SWAP, 1);
2536 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2537 struct mem_cgroup *from, struct mem_cgroup *to)
2543 static DEFINE_MUTEX(memcg_max_mutex);
2545 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2546 unsigned long max, bool memsw)
2548 bool enlarge = false;
2549 bool drained = false;
2551 bool limits_invariant;
2552 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2555 if (signal_pending(current)) {
2560 mutex_lock(&memcg_max_mutex);
2562 * Make sure that the new limit (memsw or memory limit) doesn't
2563 * break our basic invariant rule memory.max <= memsw.max.
2565 limits_invariant = memsw ? max >= memcg->memory.max :
2566 max <= memcg->memsw.max;
2567 if (!limits_invariant) {
2568 mutex_unlock(&memcg_max_mutex);
2572 if (max > counter->max)
2574 ret = page_counter_set_max(counter, max);
2575 mutex_unlock(&memcg_max_mutex);
2581 drain_all_stock(memcg);
2586 if (!try_to_free_mem_cgroup_pages(memcg, 1,
2587 GFP_KERNEL, !memsw)) {
2593 if (!ret && enlarge)
2594 memcg_oom_recover(memcg);
2599 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2601 unsigned long *total_scanned)
2603 unsigned long nr_reclaimed = 0;
2604 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2605 unsigned long reclaimed;
2607 struct mem_cgroup_tree_per_node *mctz;
2608 unsigned long excess;
2609 unsigned long nr_scanned;
2614 mctz = soft_limit_tree_node(pgdat->node_id);
2617 * Do not even bother to check the largest node if the root
2618 * is empty. Do it lockless to prevent lock bouncing. Races
2619 * are acceptable as soft limit is best effort anyway.
2621 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2625 * This loop can run a while, specially if mem_cgroup's continuously
2626 * keep exceeding their soft limit and putting the system under
2633 mz = mem_cgroup_largest_soft_limit_node(mctz);
2638 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2639 gfp_mask, &nr_scanned);
2640 nr_reclaimed += reclaimed;
2641 *total_scanned += nr_scanned;
2642 spin_lock_irq(&mctz->lock);
2643 __mem_cgroup_remove_exceeded(mz, mctz);
2646 * If we failed to reclaim anything from this memory cgroup
2647 * it is time to move on to the next cgroup
2651 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2653 excess = soft_limit_excess(mz->memcg);
2655 * One school of thought says that we should not add
2656 * back the node to the tree if reclaim returns 0.
2657 * But our reclaim could return 0, simply because due
2658 * to priority we are exposing a smaller subset of
2659 * memory to reclaim from. Consider this as a longer
2662 /* If excess == 0, no tree ops */
2663 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2664 spin_unlock_irq(&mctz->lock);
2665 css_put(&mz->memcg->css);
2668 * Could not reclaim anything and there are no more
2669 * mem cgroups to try or we seem to be looping without
2670 * reclaiming anything.
2672 if (!nr_reclaimed &&
2674 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2676 } while (!nr_reclaimed);
2678 css_put(&next_mz->memcg->css);
2679 return nr_reclaimed;
2683 * Test whether @memcg has children, dead or alive. Note that this
2684 * function doesn't care whether @memcg has use_hierarchy enabled and
2685 * returns %true if there are child csses according to the cgroup
2686 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2688 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2693 ret = css_next_child(NULL, &memcg->css);
2699 * Reclaims as many pages from the given memcg as possible.
2701 * Caller is responsible for holding css reference for memcg.
2703 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2705 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2707 /* we call try-to-free pages for make this cgroup empty */
2708 lru_add_drain_all();
2710 drain_all_stock(memcg);
2712 /* try to free all pages in this cgroup */
2713 while (nr_retries && page_counter_read(&memcg->memory)) {
2716 if (signal_pending(current))
2719 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2723 /* maybe some writeback is necessary */
2724 congestion_wait(BLK_RW_ASYNC, HZ/10);
2732 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2733 char *buf, size_t nbytes,
2736 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2738 if (mem_cgroup_is_root(memcg))
2740 return mem_cgroup_force_empty(memcg) ?: nbytes;
2743 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2746 return mem_cgroup_from_css(css)->use_hierarchy;
2749 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2750 struct cftype *cft, u64 val)
2753 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2754 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2756 if (memcg->use_hierarchy == val)
2760 * If parent's use_hierarchy is set, we can't make any modifications
2761 * in the child subtrees. If it is unset, then the change can
2762 * occur, provided the current cgroup has no children.
2764 * For the root cgroup, parent_mem is NULL, we allow value to be
2765 * set if there are no children.
2767 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2768 (val == 1 || val == 0)) {
2769 if (!memcg_has_children(memcg))
2770 memcg->use_hierarchy = val;
2779 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2781 struct mem_cgroup *iter;
2784 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2786 for_each_mem_cgroup_tree(iter, memcg) {
2787 for (i = 0; i < MEMCG_NR_STAT; i++)
2788 stat[i] += memcg_page_state(iter, i);
2792 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2794 struct mem_cgroup *iter;
2797 memset(events, 0, sizeof(*events) * NR_VM_EVENT_ITEMS);
2799 for_each_mem_cgroup_tree(iter, memcg) {
2800 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
2801 events[i] += memcg_sum_events(iter, i);
2805 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2807 unsigned long val = 0;
2809 if (mem_cgroup_is_root(memcg)) {
2810 struct mem_cgroup *iter;
2812 for_each_mem_cgroup_tree(iter, memcg) {
2813 val += memcg_page_state(iter, MEMCG_CACHE);
2814 val += memcg_page_state(iter, MEMCG_RSS);
2816 val += memcg_page_state(iter, MEMCG_SWAP);
2820 val = page_counter_read(&memcg->memory);
2822 val = page_counter_read(&memcg->memsw);
2835 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2838 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2839 struct page_counter *counter;
2841 switch (MEMFILE_TYPE(cft->private)) {
2843 counter = &memcg->memory;
2846 counter = &memcg->memsw;
2849 counter = &memcg->kmem;
2852 counter = &memcg->tcpmem;
2858 switch (MEMFILE_ATTR(cft->private)) {
2860 if (counter == &memcg->memory)
2861 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2862 if (counter == &memcg->memsw)
2863 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2864 return (u64)page_counter_read(counter) * PAGE_SIZE;
2866 return (u64)counter->max * PAGE_SIZE;
2868 return (u64)counter->watermark * PAGE_SIZE;
2870 return counter->failcnt;
2871 case RES_SOFT_LIMIT:
2872 return (u64)memcg->soft_limit * PAGE_SIZE;
2878 #ifdef CONFIG_MEMCG_KMEM
2879 static int memcg_online_kmem(struct mem_cgroup *memcg)
2883 if (cgroup_memory_nokmem)
2886 BUG_ON(memcg->kmemcg_id >= 0);
2887 BUG_ON(memcg->kmem_state);
2889 memcg_id = memcg_alloc_cache_id();
2893 static_branch_inc(&memcg_kmem_enabled_key);
2895 * A memory cgroup is considered kmem-online as soon as it gets
2896 * kmemcg_id. Setting the id after enabling static branching will
2897 * guarantee no one starts accounting before all call sites are
2900 memcg->kmemcg_id = memcg_id;
2901 memcg->kmem_state = KMEM_ONLINE;
2902 INIT_LIST_HEAD(&memcg->kmem_caches);
2907 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2909 struct cgroup_subsys_state *css;
2910 struct mem_cgroup *parent, *child;
2913 if (memcg->kmem_state != KMEM_ONLINE)
2916 * Clear the online state before clearing memcg_caches array
2917 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2918 * guarantees that no cache will be created for this cgroup
2919 * after we are done (see memcg_create_kmem_cache()).
2921 memcg->kmem_state = KMEM_ALLOCATED;
2923 memcg_deactivate_kmem_caches(memcg);
2925 kmemcg_id = memcg->kmemcg_id;
2926 BUG_ON(kmemcg_id < 0);
2928 parent = parent_mem_cgroup(memcg);
2930 parent = root_mem_cgroup;
2933 * Change kmemcg_id of this cgroup and all its descendants to the
2934 * parent's id, and then move all entries from this cgroup's list_lrus
2935 * to ones of the parent. After we have finished, all list_lrus
2936 * corresponding to this cgroup are guaranteed to remain empty. The
2937 * ordering is imposed by list_lru_node->lock taken by
2938 * memcg_drain_all_list_lrus().
2940 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2941 css_for_each_descendant_pre(css, &memcg->css) {
2942 child = mem_cgroup_from_css(css);
2943 BUG_ON(child->kmemcg_id != kmemcg_id);
2944 child->kmemcg_id = parent->kmemcg_id;
2945 if (!memcg->use_hierarchy)
2950 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2952 memcg_free_cache_id(kmemcg_id);
2955 static void memcg_free_kmem(struct mem_cgroup *memcg)
2957 /* css_alloc() failed, offlining didn't happen */
2958 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2959 memcg_offline_kmem(memcg);
2961 if (memcg->kmem_state == KMEM_ALLOCATED) {
2962 memcg_destroy_kmem_caches(memcg);
2963 static_branch_dec(&memcg_kmem_enabled_key);
2964 WARN_ON(page_counter_read(&memcg->kmem));
2968 static int memcg_online_kmem(struct mem_cgroup *memcg)
2972 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2975 static void memcg_free_kmem(struct mem_cgroup *memcg)
2978 #endif /* CONFIG_MEMCG_KMEM */
2980 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
2985 mutex_lock(&memcg_max_mutex);
2986 ret = page_counter_set_max(&memcg->kmem, max);
2987 mutex_unlock(&memcg_max_mutex);
2991 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
2995 mutex_lock(&memcg_max_mutex);
2997 ret = page_counter_set_max(&memcg->tcpmem, max);
3001 if (!memcg->tcpmem_active) {
3003 * The active flag needs to be written after the static_key
3004 * update. This is what guarantees that the socket activation
3005 * function is the last one to run. See mem_cgroup_sk_alloc()
3006 * for details, and note that we don't mark any socket as
3007 * belonging to this memcg until that flag is up.
3009 * We need to do this, because static_keys will span multiple
3010 * sites, but we can't control their order. If we mark a socket
3011 * as accounted, but the accounting functions are not patched in
3012 * yet, we'll lose accounting.
3014 * We never race with the readers in mem_cgroup_sk_alloc(),
3015 * because when this value change, the code to process it is not
3018 static_branch_inc(&memcg_sockets_enabled_key);
3019 memcg->tcpmem_active = true;
3022 mutex_unlock(&memcg_max_mutex);
3027 * The user of this function is...
3030 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3031 char *buf, size_t nbytes, loff_t off)
3033 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3034 unsigned long nr_pages;
3037 buf = strstrip(buf);
3038 ret = page_counter_memparse(buf, "-1", &nr_pages);
3042 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3044 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3048 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3050 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3053 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3056 ret = memcg_update_kmem_max(memcg, nr_pages);
3059 ret = memcg_update_tcp_max(memcg, nr_pages);
3063 case RES_SOFT_LIMIT:
3064 memcg->soft_limit = nr_pages;
3068 return ret ?: nbytes;
3071 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3072 size_t nbytes, loff_t off)
3074 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3075 struct page_counter *counter;
3077 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3079 counter = &memcg->memory;
3082 counter = &memcg->memsw;
3085 counter = &memcg->kmem;
3088 counter = &memcg->tcpmem;
3094 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3096 page_counter_reset_watermark(counter);
3099 counter->failcnt = 0;
3108 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3111 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3115 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3116 struct cftype *cft, u64 val)
3118 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3120 if (val & ~MOVE_MASK)
3124 * No kind of locking is needed in here, because ->can_attach() will
3125 * check this value once in the beginning of the process, and then carry
3126 * on with stale data. This means that changes to this value will only
3127 * affect task migrations starting after the change.
3129 memcg->move_charge_at_immigrate = val;
3133 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3134 struct cftype *cft, u64 val)
3141 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3145 unsigned int lru_mask;
3148 static const struct numa_stat stats[] = {
3149 { "total", LRU_ALL },
3150 { "file", LRU_ALL_FILE },
3151 { "anon", LRU_ALL_ANON },
3152 { "unevictable", BIT(LRU_UNEVICTABLE) },
3154 const struct numa_stat *stat;
3157 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3159 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3160 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3161 seq_printf(m, "%s=%lu", stat->name, nr);
3162 for_each_node_state(nid, N_MEMORY) {
3163 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3165 seq_printf(m, " N%d=%lu", nid, nr);
3170 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3171 struct mem_cgroup *iter;
3174 for_each_mem_cgroup_tree(iter, memcg)
3175 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3176 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3177 for_each_node_state(nid, N_MEMORY) {
3179 for_each_mem_cgroup_tree(iter, memcg)
3180 nr += mem_cgroup_node_nr_lru_pages(
3181 iter, nid, stat->lru_mask);
3182 seq_printf(m, " N%d=%lu", nid, nr);
3189 #endif /* CONFIG_NUMA */
3191 /* Universal VM events cgroup1 shows, original sort order */
3192 static const unsigned int memcg1_events[] = {
3199 static const char *const memcg1_event_names[] = {
3206 static int memcg_stat_show(struct seq_file *m, void *v)
3208 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3209 unsigned long memory, memsw;
3210 struct mem_cgroup *mi;
3213 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3214 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3216 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3217 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3219 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3220 memcg_page_state(memcg, memcg1_stats[i]) *
3224 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3225 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3226 memcg_sum_events(memcg, memcg1_events[i]));
3228 for (i = 0; i < NR_LRU_LISTS; i++)
3229 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3230 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3232 /* Hierarchical information */
3233 memory = memsw = PAGE_COUNTER_MAX;
3234 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3235 memory = min(memory, mi->memory.max);
3236 memsw = min(memsw, mi->memsw.max);
3238 seq_printf(m, "hierarchical_memory_limit %llu\n",
3239 (u64)memory * PAGE_SIZE);
3240 if (do_memsw_account())
3241 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3242 (u64)memsw * PAGE_SIZE);
3244 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3245 unsigned long long val = 0;
3247 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3249 for_each_mem_cgroup_tree(mi, memcg)
3250 val += memcg_page_state(mi, memcg1_stats[i]) *
3252 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i], val);
3255 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++) {
3256 unsigned long long val = 0;
3258 for_each_mem_cgroup_tree(mi, memcg)
3259 val += memcg_sum_events(mi, memcg1_events[i]);
3260 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i], val);
3263 for (i = 0; i < NR_LRU_LISTS; i++) {
3264 unsigned long long val = 0;
3266 for_each_mem_cgroup_tree(mi, memcg)
3267 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3268 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3271 #ifdef CONFIG_DEBUG_VM
3274 struct mem_cgroup_per_node *mz;
3275 struct zone_reclaim_stat *rstat;
3276 unsigned long recent_rotated[2] = {0, 0};
3277 unsigned long recent_scanned[2] = {0, 0};
3279 for_each_online_pgdat(pgdat) {
3280 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3281 rstat = &mz->lruvec.reclaim_stat;
3283 recent_rotated[0] += rstat->recent_rotated[0];
3284 recent_rotated[1] += rstat->recent_rotated[1];
3285 recent_scanned[0] += rstat->recent_scanned[0];
3286 recent_scanned[1] += rstat->recent_scanned[1];
3288 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3289 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3290 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3291 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3298 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3301 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3303 return mem_cgroup_swappiness(memcg);
3306 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3307 struct cftype *cft, u64 val)
3309 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3315 memcg->swappiness = val;
3317 vm_swappiness = val;
3322 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3324 struct mem_cgroup_threshold_ary *t;
3325 unsigned long usage;
3330 t = rcu_dereference(memcg->thresholds.primary);
3332 t = rcu_dereference(memcg->memsw_thresholds.primary);
3337 usage = mem_cgroup_usage(memcg, swap);
3340 * current_threshold points to threshold just below or equal to usage.
3341 * If it's not true, a threshold was crossed after last
3342 * call of __mem_cgroup_threshold().
3344 i = t->current_threshold;
3347 * Iterate backward over array of thresholds starting from
3348 * current_threshold and check if a threshold is crossed.
3349 * If none of thresholds below usage is crossed, we read
3350 * only one element of the array here.
3352 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3353 eventfd_signal(t->entries[i].eventfd, 1);
3355 /* i = current_threshold + 1 */
3359 * Iterate forward over array of thresholds starting from
3360 * current_threshold+1 and check if a threshold is crossed.
3361 * If none of thresholds above usage is crossed, we read
3362 * only one element of the array here.
3364 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3365 eventfd_signal(t->entries[i].eventfd, 1);
3367 /* Update current_threshold */
3368 t->current_threshold = i - 1;
3373 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3376 __mem_cgroup_threshold(memcg, false);
3377 if (do_memsw_account())
3378 __mem_cgroup_threshold(memcg, true);
3380 memcg = parent_mem_cgroup(memcg);
3384 static int compare_thresholds(const void *a, const void *b)
3386 const struct mem_cgroup_threshold *_a = a;
3387 const struct mem_cgroup_threshold *_b = b;
3389 if (_a->threshold > _b->threshold)
3392 if (_a->threshold < _b->threshold)
3398 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3400 struct mem_cgroup_eventfd_list *ev;
3402 spin_lock(&memcg_oom_lock);
3404 list_for_each_entry(ev, &memcg->oom_notify, list)
3405 eventfd_signal(ev->eventfd, 1);
3407 spin_unlock(&memcg_oom_lock);
3411 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3413 struct mem_cgroup *iter;
3415 for_each_mem_cgroup_tree(iter, memcg)
3416 mem_cgroup_oom_notify_cb(iter);
3419 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3420 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3422 struct mem_cgroup_thresholds *thresholds;
3423 struct mem_cgroup_threshold_ary *new;
3424 unsigned long threshold;
3425 unsigned long usage;
3428 ret = page_counter_memparse(args, "-1", &threshold);
3432 mutex_lock(&memcg->thresholds_lock);
3435 thresholds = &memcg->thresholds;
3436 usage = mem_cgroup_usage(memcg, false);
3437 } else if (type == _MEMSWAP) {
3438 thresholds = &memcg->memsw_thresholds;
3439 usage = mem_cgroup_usage(memcg, true);
3443 /* Check if a threshold crossed before adding a new one */
3444 if (thresholds->primary)
3445 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3447 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3449 /* Allocate memory for new array of thresholds */
3450 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3458 /* Copy thresholds (if any) to new array */
3459 if (thresholds->primary) {
3460 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3461 sizeof(struct mem_cgroup_threshold));
3464 /* Add new threshold */
3465 new->entries[size - 1].eventfd = eventfd;
3466 new->entries[size - 1].threshold = threshold;
3468 /* Sort thresholds. Registering of new threshold isn't time-critical */
3469 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3470 compare_thresholds, NULL);
3472 /* Find current threshold */
3473 new->current_threshold = -1;
3474 for (i = 0; i < size; i++) {
3475 if (new->entries[i].threshold <= usage) {
3477 * new->current_threshold will not be used until
3478 * rcu_assign_pointer(), so it's safe to increment
3481 ++new->current_threshold;
3486 /* Free old spare buffer and save old primary buffer as spare */
3487 kfree(thresholds->spare);
3488 thresholds->spare = thresholds->primary;
3490 rcu_assign_pointer(thresholds->primary, new);
3492 /* To be sure that nobody uses thresholds */
3496 mutex_unlock(&memcg->thresholds_lock);
3501 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3502 struct eventfd_ctx *eventfd, const char *args)
3504 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3507 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3508 struct eventfd_ctx *eventfd, const char *args)
3510 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3513 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3514 struct eventfd_ctx *eventfd, enum res_type type)
3516 struct mem_cgroup_thresholds *thresholds;
3517 struct mem_cgroup_threshold_ary *new;
3518 unsigned long usage;
3521 mutex_lock(&memcg->thresholds_lock);
3524 thresholds = &memcg->thresholds;
3525 usage = mem_cgroup_usage(memcg, false);
3526 } else if (type == _MEMSWAP) {
3527 thresholds = &memcg->memsw_thresholds;
3528 usage = mem_cgroup_usage(memcg, true);
3532 if (!thresholds->primary)
3535 /* Check if a threshold crossed before removing */
3536 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3538 /* Calculate new number of threshold */
3540 for (i = 0; i < thresholds->primary->size; i++) {
3541 if (thresholds->primary->entries[i].eventfd != eventfd)
3545 new = thresholds->spare;
3547 /* Set thresholds array to NULL if we don't have thresholds */
3556 /* Copy thresholds and find current threshold */
3557 new->current_threshold = -1;
3558 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3559 if (thresholds->primary->entries[i].eventfd == eventfd)
3562 new->entries[j] = thresholds->primary->entries[i];
3563 if (new->entries[j].threshold <= usage) {
3565 * new->current_threshold will not be used
3566 * until rcu_assign_pointer(), so it's safe to increment
3569 ++new->current_threshold;
3575 /* Swap primary and spare array */
3576 thresholds->spare = thresholds->primary;
3578 rcu_assign_pointer(thresholds->primary, new);
3580 /* To be sure that nobody uses thresholds */
3583 /* If all events are unregistered, free the spare array */
3585 kfree(thresholds->spare);
3586 thresholds->spare = NULL;
3589 mutex_unlock(&memcg->thresholds_lock);
3592 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3593 struct eventfd_ctx *eventfd)
3595 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3598 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3599 struct eventfd_ctx *eventfd)
3601 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3604 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3605 struct eventfd_ctx *eventfd, const char *args)
3607 struct mem_cgroup_eventfd_list *event;
3609 event = kmalloc(sizeof(*event), GFP_KERNEL);
3613 spin_lock(&memcg_oom_lock);
3615 event->eventfd = eventfd;
3616 list_add(&event->list, &memcg->oom_notify);
3618 /* already in OOM ? */
3619 if (memcg->under_oom)
3620 eventfd_signal(eventfd, 1);
3621 spin_unlock(&memcg_oom_lock);
3626 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3627 struct eventfd_ctx *eventfd)
3629 struct mem_cgroup_eventfd_list *ev, *tmp;
3631 spin_lock(&memcg_oom_lock);
3633 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3634 if (ev->eventfd == eventfd) {
3635 list_del(&ev->list);
3640 spin_unlock(&memcg_oom_lock);
3643 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3645 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3647 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3648 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3649 seq_printf(sf, "oom_kill %lu\n",
3650 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
3654 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3655 struct cftype *cft, u64 val)
3657 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3659 /* cannot set to root cgroup and only 0 and 1 are allowed */
3660 if (!css->parent || !((val == 0) || (val == 1)))
3663 memcg->oom_kill_disable = val;
3665 memcg_oom_recover(memcg);
3670 #ifdef CONFIG_CGROUP_WRITEBACK
3672 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3674 return wb_domain_init(&memcg->cgwb_domain, gfp);
3677 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3679 wb_domain_exit(&memcg->cgwb_domain);
3682 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3684 wb_domain_size_changed(&memcg->cgwb_domain);
3687 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3689 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3691 if (!memcg->css.parent)
3694 return &memcg->cgwb_domain;
3698 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3699 * @wb: bdi_writeback in question
3700 * @pfilepages: out parameter for number of file pages
3701 * @pheadroom: out parameter for number of allocatable pages according to memcg
3702 * @pdirty: out parameter for number of dirty pages
3703 * @pwriteback: out parameter for number of pages under writeback
3705 * Determine the numbers of file, headroom, dirty, and writeback pages in
3706 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3707 * is a bit more involved.
3709 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3710 * headroom is calculated as the lowest headroom of itself and the
3711 * ancestors. Note that this doesn't consider the actual amount of
3712 * available memory in the system. The caller should further cap
3713 * *@pheadroom accordingly.
3715 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3716 unsigned long *pheadroom, unsigned long *pdirty,
3717 unsigned long *pwriteback)
3719 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3720 struct mem_cgroup *parent;
3722 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
3724 /* this should eventually include NR_UNSTABLE_NFS */
3725 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
3726 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3727 (1 << LRU_ACTIVE_FILE));
3728 *pheadroom = PAGE_COUNTER_MAX;
3730 while ((parent = parent_mem_cgroup(memcg))) {
3731 unsigned long ceiling = min(memcg->memory.max, memcg->high);
3732 unsigned long used = page_counter_read(&memcg->memory);
3734 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3739 #else /* CONFIG_CGROUP_WRITEBACK */
3741 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3746 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3750 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3754 #endif /* CONFIG_CGROUP_WRITEBACK */
3757 * DO NOT USE IN NEW FILES.
3759 * "cgroup.event_control" implementation.
3761 * This is way over-engineered. It tries to support fully configurable
3762 * events for each user. Such level of flexibility is completely
3763 * unnecessary especially in the light of the planned unified hierarchy.
3765 * Please deprecate this and replace with something simpler if at all
3770 * Unregister event and free resources.
3772 * Gets called from workqueue.
3774 static void memcg_event_remove(struct work_struct *work)
3776 struct mem_cgroup_event *event =
3777 container_of(work, struct mem_cgroup_event, remove);
3778 struct mem_cgroup *memcg = event->memcg;
3780 remove_wait_queue(event->wqh, &event->wait);
3782 event->unregister_event(memcg, event->eventfd);
3784 /* Notify userspace the event is going away. */
3785 eventfd_signal(event->eventfd, 1);
3787 eventfd_ctx_put(event->eventfd);
3789 css_put(&memcg->css);
3793 * Gets called on EPOLLHUP on eventfd when user closes it.
3795 * Called with wqh->lock held and interrupts disabled.
3797 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
3798 int sync, void *key)
3800 struct mem_cgroup_event *event =
3801 container_of(wait, struct mem_cgroup_event, wait);
3802 struct mem_cgroup *memcg = event->memcg;
3803 __poll_t flags = key_to_poll(key);
3805 if (flags & EPOLLHUP) {
3807 * If the event has been detached at cgroup removal, we
3808 * can simply return knowing the other side will cleanup
3811 * We can't race against event freeing since the other
3812 * side will require wqh->lock via remove_wait_queue(),
3815 spin_lock(&memcg->event_list_lock);
3816 if (!list_empty(&event->list)) {
3817 list_del_init(&event->list);
3819 * We are in atomic context, but cgroup_event_remove()
3820 * may sleep, so we have to call it in workqueue.
3822 schedule_work(&event->remove);
3824 spin_unlock(&memcg->event_list_lock);
3830 static void memcg_event_ptable_queue_proc(struct file *file,
3831 wait_queue_head_t *wqh, poll_table *pt)
3833 struct mem_cgroup_event *event =
3834 container_of(pt, struct mem_cgroup_event, pt);
3837 add_wait_queue(wqh, &event->wait);
3841 * DO NOT USE IN NEW FILES.
3843 * Parse input and register new cgroup event handler.
3845 * Input must be in format '<event_fd> <control_fd> <args>'.
3846 * Interpretation of args is defined by control file implementation.
3848 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3849 char *buf, size_t nbytes, loff_t off)
3851 struct cgroup_subsys_state *css = of_css(of);
3852 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3853 struct mem_cgroup_event *event;
3854 struct cgroup_subsys_state *cfile_css;
3855 unsigned int efd, cfd;
3862 buf = strstrip(buf);
3864 efd = simple_strtoul(buf, &endp, 10);
3869 cfd = simple_strtoul(buf, &endp, 10);
3870 if ((*endp != ' ') && (*endp != '\0'))
3874 event = kzalloc(sizeof(*event), GFP_KERNEL);
3878 event->memcg = memcg;
3879 INIT_LIST_HEAD(&event->list);
3880 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3881 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3882 INIT_WORK(&event->remove, memcg_event_remove);
3890 event->eventfd = eventfd_ctx_fileget(efile.file);
3891 if (IS_ERR(event->eventfd)) {
3892 ret = PTR_ERR(event->eventfd);
3899 goto out_put_eventfd;
3902 /* the process need read permission on control file */
3903 /* AV: shouldn't we check that it's been opened for read instead? */
3904 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3909 * Determine the event callbacks and set them in @event. This used
3910 * to be done via struct cftype but cgroup core no longer knows
3911 * about these events. The following is crude but the whole thing
3912 * is for compatibility anyway.
3914 * DO NOT ADD NEW FILES.
3916 name = cfile.file->f_path.dentry->d_name.name;
3918 if (!strcmp(name, "memory.usage_in_bytes")) {
3919 event->register_event = mem_cgroup_usage_register_event;
3920 event->unregister_event = mem_cgroup_usage_unregister_event;
3921 } else if (!strcmp(name, "memory.oom_control")) {
3922 event->register_event = mem_cgroup_oom_register_event;
3923 event->unregister_event = mem_cgroup_oom_unregister_event;
3924 } else if (!strcmp(name, "memory.pressure_level")) {
3925 event->register_event = vmpressure_register_event;
3926 event->unregister_event = vmpressure_unregister_event;
3927 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3928 event->register_event = memsw_cgroup_usage_register_event;
3929 event->unregister_event = memsw_cgroup_usage_unregister_event;
3936 * Verify @cfile should belong to @css. Also, remaining events are
3937 * automatically removed on cgroup destruction but the removal is
3938 * asynchronous, so take an extra ref on @css.
3940 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3941 &memory_cgrp_subsys);
3943 if (IS_ERR(cfile_css))
3945 if (cfile_css != css) {
3950 ret = event->register_event(memcg, event->eventfd, buf);
3954 vfs_poll(efile.file, &event->pt);
3956 spin_lock(&memcg->event_list_lock);
3957 list_add(&event->list, &memcg->event_list);
3958 spin_unlock(&memcg->event_list_lock);
3970 eventfd_ctx_put(event->eventfd);
3979 static struct cftype mem_cgroup_legacy_files[] = {
3981 .name = "usage_in_bytes",
3982 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3983 .read_u64 = mem_cgroup_read_u64,
3986 .name = "max_usage_in_bytes",
3987 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3988 .write = mem_cgroup_reset,
3989 .read_u64 = mem_cgroup_read_u64,
3992 .name = "limit_in_bytes",
3993 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3994 .write = mem_cgroup_write,
3995 .read_u64 = mem_cgroup_read_u64,
3998 .name = "soft_limit_in_bytes",
3999 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4000 .write = mem_cgroup_write,
4001 .read_u64 = mem_cgroup_read_u64,
4005 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4006 .write = mem_cgroup_reset,
4007 .read_u64 = mem_cgroup_read_u64,
4011 .seq_show = memcg_stat_show,
4014 .name = "force_empty",
4015 .write = mem_cgroup_force_empty_write,
4018 .name = "use_hierarchy",
4019 .write_u64 = mem_cgroup_hierarchy_write,
4020 .read_u64 = mem_cgroup_hierarchy_read,
4023 .name = "cgroup.event_control", /* XXX: for compat */
4024 .write = memcg_write_event_control,
4025 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4028 .name = "swappiness",
4029 .read_u64 = mem_cgroup_swappiness_read,
4030 .write_u64 = mem_cgroup_swappiness_write,
4033 .name = "move_charge_at_immigrate",
4034 .read_u64 = mem_cgroup_move_charge_read,
4035 .write_u64 = mem_cgroup_move_charge_write,
4038 .name = "oom_control",
4039 .seq_show = mem_cgroup_oom_control_read,
4040 .write_u64 = mem_cgroup_oom_control_write,
4041 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4044 .name = "pressure_level",
4048 .name = "numa_stat",
4049 .seq_show = memcg_numa_stat_show,
4053 .name = "kmem.limit_in_bytes",
4054 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4055 .write = mem_cgroup_write,
4056 .read_u64 = mem_cgroup_read_u64,
4059 .name = "kmem.usage_in_bytes",
4060 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4061 .read_u64 = mem_cgroup_read_u64,
4064 .name = "kmem.failcnt",
4065 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4066 .write = mem_cgroup_reset,
4067 .read_u64 = mem_cgroup_read_u64,
4070 .name = "kmem.max_usage_in_bytes",
4071 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4072 .write = mem_cgroup_reset,
4073 .read_u64 = mem_cgroup_read_u64,
4075 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4077 .name = "kmem.slabinfo",
4078 .seq_start = memcg_slab_start,
4079 .seq_next = memcg_slab_next,
4080 .seq_stop = memcg_slab_stop,
4081 .seq_show = memcg_slab_show,
4085 .name = "kmem.tcp.limit_in_bytes",
4086 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4087 .write = mem_cgroup_write,
4088 .read_u64 = mem_cgroup_read_u64,
4091 .name = "kmem.tcp.usage_in_bytes",
4092 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4093 .read_u64 = mem_cgroup_read_u64,
4096 .name = "kmem.tcp.failcnt",
4097 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4098 .write = mem_cgroup_reset,
4099 .read_u64 = mem_cgroup_read_u64,
4102 .name = "kmem.tcp.max_usage_in_bytes",
4103 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4104 .write = mem_cgroup_reset,
4105 .read_u64 = mem_cgroup_read_u64,
4107 { }, /* terminate */
4111 * Private memory cgroup IDR
4113 * Swap-out records and page cache shadow entries need to store memcg
4114 * references in constrained space, so we maintain an ID space that is
4115 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4116 * memory-controlled cgroups to 64k.
4118 * However, there usually are many references to the oflline CSS after
4119 * the cgroup has been destroyed, such as page cache or reclaimable
4120 * slab objects, that don't need to hang on to the ID. We want to keep
4121 * those dead CSS from occupying IDs, or we might quickly exhaust the
4122 * relatively small ID space and prevent the creation of new cgroups
4123 * even when there are much fewer than 64k cgroups - possibly none.
4125 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4126 * be freed and recycled when it's no longer needed, which is usually
4127 * when the CSS is offlined.
4129 * The only exception to that are records of swapped out tmpfs/shmem
4130 * pages that need to be attributed to live ancestors on swapin. But
4131 * those references are manageable from userspace.
4134 static DEFINE_IDR(mem_cgroup_idr);
4136 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4138 if (memcg->id.id > 0) {
4139 idr_remove(&mem_cgroup_idr, memcg->id.id);
4144 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4146 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4147 atomic_add(n, &memcg->id.ref);
4150 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4152 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4153 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4154 mem_cgroup_id_remove(memcg);
4156 /* Memcg ID pins CSS */
4157 css_put(&memcg->css);
4161 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4163 mem_cgroup_id_get_many(memcg, 1);
4166 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4168 mem_cgroup_id_put_many(memcg, 1);
4172 * mem_cgroup_from_id - look up a memcg from a memcg id
4173 * @id: the memcg id to look up
4175 * Caller must hold rcu_read_lock().
4177 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4179 WARN_ON_ONCE(!rcu_read_lock_held());
4180 return idr_find(&mem_cgroup_idr, id);
4183 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4185 struct mem_cgroup_per_node *pn;
4188 * This routine is called against possible nodes.
4189 * But it's BUG to call kmalloc() against offline node.
4191 * TODO: this routine can waste much memory for nodes which will
4192 * never be onlined. It's better to use memory hotplug callback
4195 if (!node_state(node, N_NORMAL_MEMORY))
4197 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4201 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4202 if (!pn->lruvec_stat_cpu) {
4207 lruvec_init(&pn->lruvec);
4208 pn->usage_in_excess = 0;
4209 pn->on_tree = false;
4212 memcg->nodeinfo[node] = pn;
4216 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4218 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4223 free_percpu(pn->lruvec_stat_cpu);
4227 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4232 free_mem_cgroup_per_node_info(memcg, node);
4233 free_percpu(memcg->stat_cpu);
4237 static void mem_cgroup_free(struct mem_cgroup *memcg)
4239 memcg_wb_domain_exit(memcg);
4240 __mem_cgroup_free(memcg);
4243 static struct mem_cgroup *mem_cgroup_alloc(void)
4245 struct mem_cgroup *memcg;
4249 size = sizeof(struct mem_cgroup);
4250 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4252 memcg = kzalloc(size, GFP_KERNEL);
4256 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4257 1, MEM_CGROUP_ID_MAX,
4259 if (memcg->id.id < 0)
4262 memcg->stat_cpu = alloc_percpu(struct mem_cgroup_stat_cpu);
4263 if (!memcg->stat_cpu)
4267 if (alloc_mem_cgroup_per_node_info(memcg, node))
4270 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4273 INIT_WORK(&memcg->high_work, high_work_func);
4274 memcg->last_scanned_node = MAX_NUMNODES;
4275 INIT_LIST_HEAD(&memcg->oom_notify);
4276 mutex_init(&memcg->thresholds_lock);
4277 spin_lock_init(&memcg->move_lock);
4278 vmpressure_init(&memcg->vmpressure);
4279 INIT_LIST_HEAD(&memcg->event_list);
4280 spin_lock_init(&memcg->event_list_lock);
4281 memcg->socket_pressure = jiffies;
4282 #ifdef CONFIG_MEMCG_KMEM
4283 memcg->kmemcg_id = -1;
4285 #ifdef CONFIG_CGROUP_WRITEBACK
4286 INIT_LIST_HEAD(&memcg->cgwb_list);
4288 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4291 mem_cgroup_id_remove(memcg);
4292 __mem_cgroup_free(memcg);
4296 static struct cgroup_subsys_state * __ref
4297 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4299 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4300 struct mem_cgroup *memcg;
4301 long error = -ENOMEM;
4303 memcg = mem_cgroup_alloc();
4305 return ERR_PTR(error);
4307 memcg->high = PAGE_COUNTER_MAX;
4308 memcg->soft_limit = PAGE_COUNTER_MAX;
4310 memcg->swappiness = mem_cgroup_swappiness(parent);
4311 memcg->oom_kill_disable = parent->oom_kill_disable;
4313 if (parent && parent->use_hierarchy) {
4314 memcg->use_hierarchy = true;
4315 page_counter_init(&memcg->memory, &parent->memory);
4316 page_counter_init(&memcg->swap, &parent->swap);
4317 page_counter_init(&memcg->memsw, &parent->memsw);
4318 page_counter_init(&memcg->kmem, &parent->kmem);
4319 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4321 page_counter_init(&memcg->memory, NULL);
4322 page_counter_init(&memcg->swap, NULL);
4323 page_counter_init(&memcg->memsw, NULL);
4324 page_counter_init(&memcg->kmem, NULL);
4325 page_counter_init(&memcg->tcpmem, NULL);
4327 * Deeper hierachy with use_hierarchy == false doesn't make
4328 * much sense so let cgroup subsystem know about this
4329 * unfortunate state in our controller.
4331 if (parent != root_mem_cgroup)
4332 memory_cgrp_subsys.broken_hierarchy = true;
4335 /* The following stuff does not apply to the root */
4337 root_mem_cgroup = memcg;
4341 error = memcg_online_kmem(memcg);
4345 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4346 static_branch_inc(&memcg_sockets_enabled_key);
4350 mem_cgroup_id_remove(memcg);
4351 mem_cgroup_free(memcg);
4352 return ERR_PTR(-ENOMEM);
4355 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4357 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4359 /* Online state pins memcg ID, memcg ID pins CSS */
4360 atomic_set(&memcg->id.ref, 1);
4365 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4367 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4368 struct mem_cgroup_event *event, *tmp;
4371 * Unregister events and notify userspace.
4372 * Notify userspace about cgroup removing only after rmdir of cgroup
4373 * directory to avoid race between userspace and kernelspace.
4375 spin_lock(&memcg->event_list_lock);
4376 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4377 list_del_init(&event->list);
4378 schedule_work(&event->remove);
4380 spin_unlock(&memcg->event_list_lock);
4382 page_counter_set_min(&memcg->memory, 0);
4383 page_counter_set_low(&memcg->memory, 0);
4385 memcg_offline_kmem(memcg);
4386 wb_memcg_offline(memcg);
4388 mem_cgroup_id_put(memcg);
4391 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4393 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4395 invalidate_reclaim_iterators(memcg);
4398 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4400 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4402 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4403 static_branch_dec(&memcg_sockets_enabled_key);
4405 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4406 static_branch_dec(&memcg_sockets_enabled_key);
4408 vmpressure_cleanup(&memcg->vmpressure);
4409 cancel_work_sync(&memcg->high_work);
4410 mem_cgroup_remove_from_trees(memcg);
4411 memcg_free_kmem(memcg);
4412 mem_cgroup_free(memcg);
4416 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4417 * @css: the target css
4419 * Reset the states of the mem_cgroup associated with @css. This is
4420 * invoked when the userland requests disabling on the default hierarchy
4421 * but the memcg is pinned through dependency. The memcg should stop
4422 * applying policies and should revert to the vanilla state as it may be
4423 * made visible again.
4425 * The current implementation only resets the essential configurations.
4426 * This needs to be expanded to cover all the visible parts.
4428 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4430 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4432 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
4433 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
4434 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
4435 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
4436 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
4437 page_counter_set_min(&memcg->memory, 0);
4438 page_counter_set_low(&memcg->memory, 0);
4439 memcg->high = PAGE_COUNTER_MAX;
4440 memcg->soft_limit = PAGE_COUNTER_MAX;
4441 memcg_wb_domain_size_changed(memcg);
4445 /* Handlers for move charge at task migration. */
4446 static int mem_cgroup_do_precharge(unsigned long count)
4450 /* Try a single bulk charge without reclaim first, kswapd may wake */
4451 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4453 mc.precharge += count;
4457 /* Try charges one by one with reclaim, but do not retry */
4459 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4473 enum mc_target_type {
4480 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4481 unsigned long addr, pte_t ptent)
4483 struct page *page = _vm_normal_page(vma, addr, ptent, true);
4485 if (!page || !page_mapped(page))
4487 if (PageAnon(page)) {
4488 if (!(mc.flags & MOVE_ANON))
4491 if (!(mc.flags & MOVE_FILE))
4494 if (!get_page_unless_zero(page))
4500 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
4501 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4502 pte_t ptent, swp_entry_t *entry)
4504 struct page *page = NULL;
4505 swp_entry_t ent = pte_to_swp_entry(ptent);
4507 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4511 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
4512 * a device and because they are not accessible by CPU they are store
4513 * as special swap entry in the CPU page table.
4515 if (is_device_private_entry(ent)) {
4516 page = device_private_entry_to_page(ent);
4518 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
4519 * a refcount of 1 when free (unlike normal page)
4521 if (!page_ref_add_unless(page, 1, 1))
4527 * Because lookup_swap_cache() updates some statistics counter,
4528 * we call find_get_page() with swapper_space directly.
4530 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4531 if (do_memsw_account())
4532 entry->val = ent.val;
4537 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4538 pte_t ptent, swp_entry_t *entry)
4544 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4545 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4547 struct page *page = NULL;
4548 struct address_space *mapping;
4551 if (!vma->vm_file) /* anonymous vma */
4553 if (!(mc.flags & MOVE_FILE))
4556 mapping = vma->vm_file->f_mapping;
4557 pgoff = linear_page_index(vma, addr);
4559 /* page is moved even if it's not RSS of this task(page-faulted). */
4561 /* shmem/tmpfs may report page out on swap: account for that too. */
4562 if (shmem_mapping(mapping)) {
4563 page = find_get_entry(mapping, pgoff);
4564 if (radix_tree_exceptional_entry(page)) {
4565 swp_entry_t swp = radix_to_swp_entry(page);
4566 if (do_memsw_account())
4568 page = find_get_page(swap_address_space(swp),
4572 page = find_get_page(mapping, pgoff);
4574 page = find_get_page(mapping, pgoff);
4580 * mem_cgroup_move_account - move account of the page
4582 * @compound: charge the page as compound or small page
4583 * @from: mem_cgroup which the page is moved from.
4584 * @to: mem_cgroup which the page is moved to. @from != @to.
4586 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4588 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4591 static int mem_cgroup_move_account(struct page *page,
4593 struct mem_cgroup *from,
4594 struct mem_cgroup *to)
4596 unsigned long flags;
4597 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4601 VM_BUG_ON(from == to);
4602 VM_BUG_ON_PAGE(PageLRU(page), page);
4603 VM_BUG_ON(compound && !PageTransHuge(page));
4606 * Prevent mem_cgroup_migrate() from looking at
4607 * page->mem_cgroup of its source page while we change it.
4610 if (!trylock_page(page))
4614 if (page->mem_cgroup != from)
4617 anon = PageAnon(page);
4619 spin_lock_irqsave(&from->move_lock, flags);
4621 if (!anon && page_mapped(page)) {
4622 __mod_memcg_state(from, NR_FILE_MAPPED, -nr_pages);
4623 __mod_memcg_state(to, NR_FILE_MAPPED, nr_pages);
4627 * move_lock grabbed above and caller set from->moving_account, so
4628 * mod_memcg_page_state will serialize updates to PageDirty.
4629 * So mapping should be stable for dirty pages.
4631 if (!anon && PageDirty(page)) {
4632 struct address_space *mapping = page_mapping(page);
4634 if (mapping_cap_account_dirty(mapping)) {
4635 __mod_memcg_state(from, NR_FILE_DIRTY, -nr_pages);
4636 __mod_memcg_state(to, NR_FILE_DIRTY, nr_pages);
4640 if (PageWriteback(page)) {
4641 __mod_memcg_state(from, NR_WRITEBACK, -nr_pages);
4642 __mod_memcg_state(to, NR_WRITEBACK, nr_pages);
4646 * It is safe to change page->mem_cgroup here because the page
4647 * is referenced, charged, and isolated - we can't race with
4648 * uncharging, charging, migration, or LRU putback.
4651 /* caller should have done css_get */
4652 page->mem_cgroup = to;
4653 spin_unlock_irqrestore(&from->move_lock, flags);
4657 local_irq_disable();
4658 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4659 memcg_check_events(to, page);
4660 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4661 memcg_check_events(from, page);
4670 * get_mctgt_type - get target type of moving charge
4671 * @vma: the vma the pte to be checked belongs
4672 * @addr: the address corresponding to the pte to be checked
4673 * @ptent: the pte to be checked
4674 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4677 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4678 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4679 * move charge. if @target is not NULL, the page is stored in target->page
4680 * with extra refcnt got(Callers should handle it).
4681 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4682 * target for charge migration. if @target is not NULL, the entry is stored
4684 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PUBLIC
4685 * or MEMORY_DEVICE_PRIVATE (so ZONE_DEVICE page and thus not on the lru).
4686 * For now we such page is charge like a regular page would be as for all
4687 * intent and purposes it is just special memory taking the place of a
4690 * See Documentations/vm/hmm.txt and include/linux/hmm.h
4692 * Called with pte lock held.
4695 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4696 unsigned long addr, pte_t ptent, union mc_target *target)
4698 struct page *page = NULL;
4699 enum mc_target_type ret = MC_TARGET_NONE;
4700 swp_entry_t ent = { .val = 0 };
4702 if (pte_present(ptent))
4703 page = mc_handle_present_pte(vma, addr, ptent);
4704 else if (is_swap_pte(ptent))
4705 page = mc_handle_swap_pte(vma, ptent, &ent);
4706 else if (pte_none(ptent))
4707 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4709 if (!page && !ent.val)
4713 * Do only loose check w/o serialization.
4714 * mem_cgroup_move_account() checks the page is valid or
4715 * not under LRU exclusion.
4717 if (page->mem_cgroup == mc.from) {
4718 ret = MC_TARGET_PAGE;
4719 if (is_device_private_page(page) ||
4720 is_device_public_page(page))
4721 ret = MC_TARGET_DEVICE;
4723 target->page = page;
4725 if (!ret || !target)
4729 * There is a swap entry and a page doesn't exist or isn't charged.
4730 * But we cannot move a tail-page in a THP.
4732 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4733 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4734 ret = MC_TARGET_SWAP;
4741 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4743 * We don't consider PMD mapped swapping or file mapped pages because THP does
4744 * not support them for now.
4745 * Caller should make sure that pmd_trans_huge(pmd) is true.
4747 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4748 unsigned long addr, pmd_t pmd, union mc_target *target)
4750 struct page *page = NULL;
4751 enum mc_target_type ret = MC_TARGET_NONE;
4753 if (unlikely(is_swap_pmd(pmd))) {
4754 VM_BUG_ON(thp_migration_supported() &&
4755 !is_pmd_migration_entry(pmd));
4758 page = pmd_page(pmd);
4759 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4760 if (!(mc.flags & MOVE_ANON))
4762 if (page->mem_cgroup == mc.from) {
4763 ret = MC_TARGET_PAGE;
4766 target->page = page;
4772 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4773 unsigned long addr, pmd_t pmd, union mc_target *target)
4775 return MC_TARGET_NONE;
4779 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4780 unsigned long addr, unsigned long end,
4781 struct mm_walk *walk)
4783 struct vm_area_struct *vma = walk->vma;
4787 ptl = pmd_trans_huge_lock(pmd, vma);
4790 * Note their can not be MC_TARGET_DEVICE for now as we do not
4791 * support transparent huge page with MEMORY_DEVICE_PUBLIC or
4792 * MEMORY_DEVICE_PRIVATE but this might change.
4794 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4795 mc.precharge += HPAGE_PMD_NR;
4800 if (pmd_trans_unstable(pmd))
4802 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4803 for (; addr != end; pte++, addr += PAGE_SIZE)
4804 if (get_mctgt_type(vma, addr, *pte, NULL))
4805 mc.precharge++; /* increment precharge temporarily */
4806 pte_unmap_unlock(pte - 1, ptl);
4812 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4814 unsigned long precharge;
4816 struct mm_walk mem_cgroup_count_precharge_walk = {
4817 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4820 down_read(&mm->mmap_sem);
4821 walk_page_range(0, mm->highest_vm_end,
4822 &mem_cgroup_count_precharge_walk);
4823 up_read(&mm->mmap_sem);
4825 precharge = mc.precharge;
4831 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4833 unsigned long precharge = mem_cgroup_count_precharge(mm);
4835 VM_BUG_ON(mc.moving_task);
4836 mc.moving_task = current;
4837 return mem_cgroup_do_precharge(precharge);
4840 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4841 static void __mem_cgroup_clear_mc(void)
4843 struct mem_cgroup *from = mc.from;
4844 struct mem_cgroup *to = mc.to;
4846 /* we must uncharge all the leftover precharges from mc.to */
4848 cancel_charge(mc.to, mc.precharge);
4852 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4853 * we must uncharge here.
4855 if (mc.moved_charge) {
4856 cancel_charge(mc.from, mc.moved_charge);
4857 mc.moved_charge = 0;
4859 /* we must fixup refcnts and charges */
4860 if (mc.moved_swap) {
4861 /* uncharge swap account from the old cgroup */
4862 if (!mem_cgroup_is_root(mc.from))
4863 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4865 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4868 * we charged both to->memory and to->memsw, so we
4869 * should uncharge to->memory.
4871 if (!mem_cgroup_is_root(mc.to))
4872 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4874 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4875 css_put_many(&mc.to->css, mc.moved_swap);
4879 memcg_oom_recover(from);
4880 memcg_oom_recover(to);
4881 wake_up_all(&mc.waitq);
4884 static void mem_cgroup_clear_mc(void)
4886 struct mm_struct *mm = mc.mm;
4889 * we must clear moving_task before waking up waiters at the end of
4892 mc.moving_task = NULL;
4893 __mem_cgroup_clear_mc();
4894 spin_lock(&mc.lock);
4898 spin_unlock(&mc.lock);
4903 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4905 struct cgroup_subsys_state *css;
4906 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4907 struct mem_cgroup *from;
4908 struct task_struct *leader, *p;
4909 struct mm_struct *mm;
4910 unsigned long move_flags;
4913 /* charge immigration isn't supported on the default hierarchy */
4914 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4918 * Multi-process migrations only happen on the default hierarchy
4919 * where charge immigration is not used. Perform charge
4920 * immigration if @tset contains a leader and whine if there are
4924 cgroup_taskset_for_each_leader(leader, css, tset) {
4927 memcg = mem_cgroup_from_css(css);
4933 * We are now commited to this value whatever it is. Changes in this
4934 * tunable will only affect upcoming migrations, not the current one.
4935 * So we need to save it, and keep it going.
4937 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4941 from = mem_cgroup_from_task(p);
4943 VM_BUG_ON(from == memcg);
4945 mm = get_task_mm(p);
4948 /* We move charges only when we move a owner of the mm */
4949 if (mm->owner == p) {
4952 VM_BUG_ON(mc.precharge);
4953 VM_BUG_ON(mc.moved_charge);
4954 VM_BUG_ON(mc.moved_swap);
4956 spin_lock(&mc.lock);
4960 mc.flags = move_flags;
4961 spin_unlock(&mc.lock);
4962 /* We set mc.moving_task later */
4964 ret = mem_cgroup_precharge_mc(mm);
4966 mem_cgroup_clear_mc();
4973 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4976 mem_cgroup_clear_mc();
4979 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4980 unsigned long addr, unsigned long end,
4981 struct mm_walk *walk)
4984 struct vm_area_struct *vma = walk->vma;
4987 enum mc_target_type target_type;
4988 union mc_target target;
4991 ptl = pmd_trans_huge_lock(pmd, vma);
4993 if (mc.precharge < HPAGE_PMD_NR) {
4997 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4998 if (target_type == MC_TARGET_PAGE) {
5000 if (!isolate_lru_page(page)) {
5001 if (!mem_cgroup_move_account(page, true,
5003 mc.precharge -= HPAGE_PMD_NR;
5004 mc.moved_charge += HPAGE_PMD_NR;
5006 putback_lru_page(page);
5009 } else if (target_type == MC_TARGET_DEVICE) {
5011 if (!mem_cgroup_move_account(page, true,
5013 mc.precharge -= HPAGE_PMD_NR;
5014 mc.moved_charge += HPAGE_PMD_NR;
5022 if (pmd_trans_unstable(pmd))
5025 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5026 for (; addr != end; addr += PAGE_SIZE) {
5027 pte_t ptent = *(pte++);
5028 bool device = false;
5034 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5035 case MC_TARGET_DEVICE:
5038 case MC_TARGET_PAGE:
5041 * We can have a part of the split pmd here. Moving it
5042 * can be done but it would be too convoluted so simply
5043 * ignore such a partial THP and keep it in original
5044 * memcg. There should be somebody mapping the head.
5046 if (PageTransCompound(page))
5048 if (!device && isolate_lru_page(page))
5050 if (!mem_cgroup_move_account(page, false,
5053 /* we uncharge from mc.from later. */
5057 putback_lru_page(page);
5058 put: /* get_mctgt_type() gets the page */
5061 case MC_TARGET_SWAP:
5063 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5065 /* we fixup refcnts and charges later. */
5073 pte_unmap_unlock(pte - 1, ptl);
5078 * We have consumed all precharges we got in can_attach().
5079 * We try charge one by one, but don't do any additional
5080 * charges to mc.to if we have failed in charge once in attach()
5083 ret = mem_cgroup_do_precharge(1);
5091 static void mem_cgroup_move_charge(void)
5093 struct mm_walk mem_cgroup_move_charge_walk = {
5094 .pmd_entry = mem_cgroup_move_charge_pte_range,
5098 lru_add_drain_all();
5100 * Signal lock_page_memcg() to take the memcg's move_lock
5101 * while we're moving its pages to another memcg. Then wait
5102 * for already started RCU-only updates to finish.
5104 atomic_inc(&mc.from->moving_account);
5107 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5109 * Someone who are holding the mmap_sem might be waiting in
5110 * waitq. So we cancel all extra charges, wake up all waiters,
5111 * and retry. Because we cancel precharges, we might not be able
5112 * to move enough charges, but moving charge is a best-effort
5113 * feature anyway, so it wouldn't be a big problem.
5115 __mem_cgroup_clear_mc();
5120 * When we have consumed all precharges and failed in doing
5121 * additional charge, the page walk just aborts.
5123 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5125 up_read(&mc.mm->mmap_sem);
5126 atomic_dec(&mc.from->moving_account);
5129 static void mem_cgroup_move_task(void)
5132 mem_cgroup_move_charge();
5133 mem_cgroup_clear_mc();
5136 #else /* !CONFIG_MMU */
5137 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5141 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5144 static void mem_cgroup_move_task(void)
5150 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5151 * to verify whether we're attached to the default hierarchy on each mount
5154 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5157 * use_hierarchy is forced on the default hierarchy. cgroup core
5158 * guarantees that @root doesn't have any children, so turning it
5159 * on for the root memcg is enough.
5161 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5162 root_mem_cgroup->use_hierarchy = true;
5164 root_mem_cgroup->use_hierarchy = false;
5167 static u64 memory_current_read(struct cgroup_subsys_state *css,
5170 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5172 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5175 static int memory_min_show(struct seq_file *m, void *v)
5177 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5178 unsigned long min = READ_ONCE(memcg->memory.min);
5180 if (min == PAGE_COUNTER_MAX)
5181 seq_puts(m, "max\n");
5183 seq_printf(m, "%llu\n", (u64)min * PAGE_SIZE);
5188 static ssize_t memory_min_write(struct kernfs_open_file *of,
5189 char *buf, size_t nbytes, loff_t off)
5191 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5195 buf = strstrip(buf);
5196 err = page_counter_memparse(buf, "max", &min);
5200 page_counter_set_min(&memcg->memory, min);
5205 static int memory_low_show(struct seq_file *m, void *v)
5207 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5208 unsigned long low = READ_ONCE(memcg->memory.low);
5210 if (low == PAGE_COUNTER_MAX)
5211 seq_puts(m, "max\n");
5213 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5218 static ssize_t memory_low_write(struct kernfs_open_file *of,
5219 char *buf, size_t nbytes, loff_t off)
5221 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5225 buf = strstrip(buf);
5226 err = page_counter_memparse(buf, "max", &low);
5230 page_counter_set_low(&memcg->memory, low);
5235 static int memory_high_show(struct seq_file *m, void *v)
5237 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5238 unsigned long high = READ_ONCE(memcg->high);
5240 if (high == PAGE_COUNTER_MAX)
5241 seq_puts(m, "max\n");
5243 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5248 static ssize_t memory_high_write(struct kernfs_open_file *of,
5249 char *buf, size_t nbytes, loff_t off)
5251 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5252 unsigned long nr_pages;
5256 buf = strstrip(buf);
5257 err = page_counter_memparse(buf, "max", &high);
5263 nr_pages = page_counter_read(&memcg->memory);
5264 if (nr_pages > high)
5265 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5268 memcg_wb_domain_size_changed(memcg);
5272 static int memory_max_show(struct seq_file *m, void *v)
5274 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5275 unsigned long max = READ_ONCE(memcg->memory.max);
5277 if (max == PAGE_COUNTER_MAX)
5278 seq_puts(m, "max\n");
5280 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5285 static ssize_t memory_max_write(struct kernfs_open_file *of,
5286 char *buf, size_t nbytes, loff_t off)
5288 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5289 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5290 bool drained = false;
5294 buf = strstrip(buf);
5295 err = page_counter_memparse(buf, "max", &max);
5299 xchg(&memcg->memory.max, max);
5302 unsigned long nr_pages = page_counter_read(&memcg->memory);
5304 if (nr_pages <= max)
5307 if (signal_pending(current)) {
5313 drain_all_stock(memcg);
5319 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5325 memcg_memory_event(memcg, MEMCG_OOM);
5326 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5330 memcg_wb_domain_size_changed(memcg);
5334 static int memory_events_show(struct seq_file *m, void *v)
5336 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5338 seq_printf(m, "low %lu\n",
5339 atomic_long_read(&memcg->memory_events[MEMCG_LOW]));
5340 seq_printf(m, "high %lu\n",
5341 atomic_long_read(&memcg->memory_events[MEMCG_HIGH]));
5342 seq_printf(m, "max %lu\n",
5343 atomic_long_read(&memcg->memory_events[MEMCG_MAX]));
5344 seq_printf(m, "oom %lu\n",
5345 atomic_long_read(&memcg->memory_events[MEMCG_OOM]));
5346 seq_printf(m, "oom_kill %lu\n",
5347 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
5352 static int memory_stat_show(struct seq_file *m, void *v)
5354 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5355 unsigned long stat[MEMCG_NR_STAT];
5356 unsigned long events[NR_VM_EVENT_ITEMS];
5360 * Provide statistics on the state of the memory subsystem as
5361 * well as cumulative event counters that show past behavior.
5363 * This list is ordered following a combination of these gradients:
5364 * 1) generic big picture -> specifics and details
5365 * 2) reflecting userspace activity -> reflecting kernel heuristics
5367 * Current memory state:
5370 tree_stat(memcg, stat);
5371 tree_events(memcg, events);
5373 seq_printf(m, "anon %llu\n",
5374 (u64)stat[MEMCG_RSS] * PAGE_SIZE);
5375 seq_printf(m, "file %llu\n",
5376 (u64)stat[MEMCG_CACHE] * PAGE_SIZE);
5377 seq_printf(m, "kernel_stack %llu\n",
5378 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5379 seq_printf(m, "slab %llu\n",
5380 (u64)(stat[NR_SLAB_RECLAIMABLE] +
5381 stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5382 seq_printf(m, "sock %llu\n",
5383 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5385 seq_printf(m, "shmem %llu\n",
5386 (u64)stat[NR_SHMEM] * PAGE_SIZE);
5387 seq_printf(m, "file_mapped %llu\n",
5388 (u64)stat[NR_FILE_MAPPED] * PAGE_SIZE);
5389 seq_printf(m, "file_dirty %llu\n",
5390 (u64)stat[NR_FILE_DIRTY] * PAGE_SIZE);
5391 seq_printf(m, "file_writeback %llu\n",
5392 (u64)stat[NR_WRITEBACK] * PAGE_SIZE);
5394 for (i = 0; i < NR_LRU_LISTS; i++) {
5395 struct mem_cgroup *mi;
5396 unsigned long val = 0;
5398 for_each_mem_cgroup_tree(mi, memcg)
5399 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5400 seq_printf(m, "%s %llu\n",
5401 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5404 seq_printf(m, "slab_reclaimable %llu\n",
5405 (u64)stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5406 seq_printf(m, "slab_unreclaimable %llu\n",
5407 (u64)stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5409 /* Accumulated memory events */
5411 seq_printf(m, "pgfault %lu\n", events[PGFAULT]);
5412 seq_printf(m, "pgmajfault %lu\n", events[PGMAJFAULT]);
5414 seq_printf(m, "pgrefill %lu\n", events[PGREFILL]);
5415 seq_printf(m, "pgscan %lu\n", events[PGSCAN_KSWAPD] +
5416 events[PGSCAN_DIRECT]);
5417 seq_printf(m, "pgsteal %lu\n", events[PGSTEAL_KSWAPD] +
5418 events[PGSTEAL_DIRECT]);
5419 seq_printf(m, "pgactivate %lu\n", events[PGACTIVATE]);
5420 seq_printf(m, "pgdeactivate %lu\n", events[PGDEACTIVATE]);
5421 seq_printf(m, "pglazyfree %lu\n", events[PGLAZYFREE]);
5422 seq_printf(m, "pglazyfreed %lu\n", events[PGLAZYFREED]);
5424 seq_printf(m, "workingset_refault %lu\n",
5425 stat[WORKINGSET_REFAULT]);
5426 seq_printf(m, "workingset_activate %lu\n",
5427 stat[WORKINGSET_ACTIVATE]);
5428 seq_printf(m, "workingset_nodereclaim %lu\n",
5429 stat[WORKINGSET_NODERECLAIM]);
5434 static struct cftype memory_files[] = {
5437 .flags = CFTYPE_NOT_ON_ROOT,
5438 .read_u64 = memory_current_read,
5442 .flags = CFTYPE_NOT_ON_ROOT,
5443 .seq_show = memory_min_show,
5444 .write = memory_min_write,
5448 .flags = CFTYPE_NOT_ON_ROOT,
5449 .seq_show = memory_low_show,
5450 .write = memory_low_write,
5454 .flags = CFTYPE_NOT_ON_ROOT,
5455 .seq_show = memory_high_show,
5456 .write = memory_high_write,
5460 .flags = CFTYPE_NOT_ON_ROOT,
5461 .seq_show = memory_max_show,
5462 .write = memory_max_write,
5466 .flags = CFTYPE_NOT_ON_ROOT,
5467 .file_offset = offsetof(struct mem_cgroup, events_file),
5468 .seq_show = memory_events_show,
5472 .flags = CFTYPE_NOT_ON_ROOT,
5473 .seq_show = memory_stat_show,
5478 struct cgroup_subsys memory_cgrp_subsys = {
5479 .css_alloc = mem_cgroup_css_alloc,
5480 .css_online = mem_cgroup_css_online,
5481 .css_offline = mem_cgroup_css_offline,
5482 .css_released = mem_cgroup_css_released,
5483 .css_free = mem_cgroup_css_free,
5484 .css_reset = mem_cgroup_css_reset,
5485 .can_attach = mem_cgroup_can_attach,
5486 .cancel_attach = mem_cgroup_cancel_attach,
5487 .post_attach = mem_cgroup_move_task,
5488 .bind = mem_cgroup_bind,
5489 .dfl_cftypes = memory_files,
5490 .legacy_cftypes = mem_cgroup_legacy_files,
5495 * mem_cgroup_protected - check if memory consumption is in the normal range
5496 * @root: the top ancestor of the sub-tree being checked
5497 * @memcg: the memory cgroup to check
5499 * WARNING: This function is not stateless! It can only be used as part
5500 * of a top-down tree iteration, not for isolated queries.
5502 * Returns one of the following:
5503 * MEMCG_PROT_NONE: cgroup memory is not protected
5504 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
5505 * an unprotected supply of reclaimable memory from other cgroups.
5506 * MEMCG_PROT_MIN: cgroup memory is protected
5508 * @root is exclusive; it is never protected when looked at directly
5510 * To provide a proper hierarchical behavior, effective memory.min/low values
5511 * are used. Below is the description of how effective memory.low is calculated.
5512 * Effective memory.min values is calculated in the same way.
5514 * Effective memory.low is always equal or less than the original memory.low.
5515 * If there is no memory.low overcommittment (which is always true for
5516 * top-level memory cgroups), these two values are equal.
5517 * Otherwise, it's a part of parent's effective memory.low,
5518 * calculated as a cgroup's memory.low usage divided by sum of sibling's
5519 * memory.low usages, where memory.low usage is the size of actually
5523 * elow = min( memory.low, parent->elow * ------------------ ),
5524 * siblings_low_usage
5526 * | memory.current, if memory.current < memory.low
5531 * Such definition of the effective memory.low provides the expected
5532 * hierarchical behavior: parent's memory.low value is limiting
5533 * children, unprotected memory is reclaimed first and cgroups,
5534 * which are not using their guarantee do not affect actual memory
5537 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
5539 * A A/memory.low = 2G, A/memory.current = 6G
5541 * BC DE B/memory.low = 3G B/memory.current = 2G
5542 * C/memory.low = 1G C/memory.current = 2G
5543 * D/memory.low = 0 D/memory.current = 2G
5544 * E/memory.low = 10G E/memory.current = 0
5546 * and the memory pressure is applied, the following memory distribution
5547 * is expected (approximately):
5549 * A/memory.current = 2G
5551 * B/memory.current = 1.3G
5552 * C/memory.current = 0.6G
5553 * D/memory.current = 0
5554 * E/memory.current = 0
5556 * These calculations require constant tracking of the actual low usages
5557 * (see propagate_protected_usage()), as well as recursive calculation of
5558 * effective memory.low values. But as we do call mem_cgroup_protected()
5559 * path for each memory cgroup top-down from the reclaim,
5560 * it's possible to optimize this part, and save calculated elow
5561 * for next usage. This part is intentionally racy, but it's ok,
5562 * as memory.low is a best-effort mechanism.
5564 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
5565 struct mem_cgroup *memcg)
5567 struct mem_cgroup *parent;
5568 unsigned long emin, parent_emin;
5569 unsigned long elow, parent_elow;
5570 unsigned long usage;
5572 if (mem_cgroup_disabled())
5573 return MEMCG_PROT_NONE;
5576 root = root_mem_cgroup;
5578 return MEMCG_PROT_NONE;
5580 usage = page_counter_read(&memcg->memory);
5582 return MEMCG_PROT_NONE;
5584 emin = memcg->memory.min;
5585 elow = memcg->memory.low;
5587 parent = parent_mem_cgroup(memcg);
5588 /* No parent means a non-hierarchical mode on v1 memcg */
5590 return MEMCG_PROT_NONE;
5595 parent_emin = READ_ONCE(parent->memory.emin);
5596 emin = min(emin, parent_emin);
5597 if (emin && parent_emin) {
5598 unsigned long min_usage, siblings_min_usage;
5600 min_usage = min(usage, memcg->memory.min);
5601 siblings_min_usage = atomic_long_read(
5602 &parent->memory.children_min_usage);
5604 if (min_usage && siblings_min_usage)
5605 emin = min(emin, parent_emin * min_usage /
5606 siblings_min_usage);
5609 parent_elow = READ_ONCE(parent->memory.elow);
5610 elow = min(elow, parent_elow);
5611 if (elow && parent_elow) {
5612 unsigned long low_usage, siblings_low_usage;
5614 low_usage = min(usage, memcg->memory.low);
5615 siblings_low_usage = atomic_long_read(
5616 &parent->memory.children_low_usage);
5618 if (low_usage && siblings_low_usage)
5619 elow = min(elow, parent_elow * low_usage /
5620 siblings_low_usage);
5624 memcg->memory.emin = emin;
5625 memcg->memory.elow = elow;
5628 return MEMCG_PROT_MIN;
5629 else if (usage <= elow)
5630 return MEMCG_PROT_LOW;
5632 return MEMCG_PROT_NONE;
5636 * mem_cgroup_try_charge - try charging a page
5637 * @page: page to charge
5638 * @mm: mm context of the victim
5639 * @gfp_mask: reclaim mode
5640 * @memcgp: charged memcg return
5641 * @compound: charge the page as compound or small page
5643 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5644 * pages according to @gfp_mask if necessary.
5646 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5647 * Otherwise, an error code is returned.
5649 * After page->mapping has been set up, the caller must finalize the
5650 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5651 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5653 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5654 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5657 struct mem_cgroup *memcg = NULL;
5658 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5661 if (mem_cgroup_disabled())
5664 if (PageSwapCache(page)) {
5666 * Every swap fault against a single page tries to charge the
5667 * page, bail as early as possible. shmem_unuse() encounters
5668 * already charged pages, too. The USED bit is protected by
5669 * the page lock, which serializes swap cache removal, which
5670 * in turn serializes uncharging.
5672 VM_BUG_ON_PAGE(!PageLocked(page), page);
5673 if (compound_head(page)->mem_cgroup)
5676 if (do_swap_account) {
5677 swp_entry_t ent = { .val = page_private(page), };
5678 unsigned short id = lookup_swap_cgroup_id(ent);
5681 memcg = mem_cgroup_from_id(id);
5682 if (memcg && !css_tryget_online(&memcg->css))
5689 memcg = get_mem_cgroup_from_mm(mm);
5691 ret = try_charge(memcg, gfp_mask, nr_pages);
5693 css_put(&memcg->css);
5699 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
5700 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5703 struct mem_cgroup *memcg;
5706 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
5708 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
5713 * mem_cgroup_commit_charge - commit a page charge
5714 * @page: page to charge
5715 * @memcg: memcg to charge the page to
5716 * @lrucare: page might be on LRU already
5717 * @compound: charge the page as compound or small page
5719 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5720 * after page->mapping has been set up. This must happen atomically
5721 * as part of the page instantiation, i.e. under the page table lock
5722 * for anonymous pages, under the page lock for page and swap cache.
5724 * In addition, the page must not be on the LRU during the commit, to
5725 * prevent racing with task migration. If it might be, use @lrucare.
5727 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5729 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5730 bool lrucare, bool compound)
5732 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5734 VM_BUG_ON_PAGE(!page->mapping, page);
5735 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5737 if (mem_cgroup_disabled())
5740 * Swap faults will attempt to charge the same page multiple
5741 * times. But reuse_swap_page() might have removed the page
5742 * from swapcache already, so we can't check PageSwapCache().
5747 commit_charge(page, memcg, lrucare);
5749 local_irq_disable();
5750 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5751 memcg_check_events(memcg, page);
5754 if (do_memsw_account() && PageSwapCache(page)) {
5755 swp_entry_t entry = { .val = page_private(page) };
5757 * The swap entry might not get freed for a long time,
5758 * let's not wait for it. The page already received a
5759 * memory+swap charge, drop the swap entry duplicate.
5761 mem_cgroup_uncharge_swap(entry, nr_pages);
5766 * mem_cgroup_cancel_charge - cancel a page charge
5767 * @page: page to charge
5768 * @memcg: memcg to charge the page to
5769 * @compound: charge the page as compound or small page
5771 * Cancel a charge transaction started by mem_cgroup_try_charge().
5773 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5776 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5778 if (mem_cgroup_disabled())
5781 * Swap faults will attempt to charge the same page multiple
5782 * times. But reuse_swap_page() might have removed the page
5783 * from swapcache already, so we can't check PageSwapCache().
5788 cancel_charge(memcg, nr_pages);
5791 struct uncharge_gather {
5792 struct mem_cgroup *memcg;
5793 unsigned long pgpgout;
5794 unsigned long nr_anon;
5795 unsigned long nr_file;
5796 unsigned long nr_kmem;
5797 unsigned long nr_huge;
5798 unsigned long nr_shmem;
5799 struct page *dummy_page;
5802 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
5804 memset(ug, 0, sizeof(*ug));
5807 static void uncharge_batch(const struct uncharge_gather *ug)
5809 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
5810 unsigned long flags;
5812 if (!mem_cgroup_is_root(ug->memcg)) {
5813 page_counter_uncharge(&ug->memcg->memory, nr_pages);
5814 if (do_memsw_account())
5815 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
5816 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
5817 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
5818 memcg_oom_recover(ug->memcg);
5821 local_irq_save(flags);
5822 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
5823 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
5824 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
5825 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
5826 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
5827 __this_cpu_add(ug->memcg->stat_cpu->nr_page_events, nr_pages);
5828 memcg_check_events(ug->memcg, ug->dummy_page);
5829 local_irq_restore(flags);
5831 if (!mem_cgroup_is_root(ug->memcg))
5832 css_put_many(&ug->memcg->css, nr_pages);
5835 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
5837 VM_BUG_ON_PAGE(PageLRU(page), page);
5838 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
5839 !PageHWPoison(page) , page);
5841 if (!page->mem_cgroup)
5845 * Nobody should be changing or seriously looking at
5846 * page->mem_cgroup at this point, we have fully
5847 * exclusive access to the page.
5850 if (ug->memcg != page->mem_cgroup) {
5853 uncharge_gather_clear(ug);
5855 ug->memcg = page->mem_cgroup;
5858 if (!PageKmemcg(page)) {
5859 unsigned int nr_pages = 1;
5861 if (PageTransHuge(page)) {
5862 nr_pages <<= compound_order(page);
5863 ug->nr_huge += nr_pages;
5866 ug->nr_anon += nr_pages;
5868 ug->nr_file += nr_pages;
5869 if (PageSwapBacked(page))
5870 ug->nr_shmem += nr_pages;
5874 ug->nr_kmem += 1 << compound_order(page);
5875 __ClearPageKmemcg(page);
5878 ug->dummy_page = page;
5879 page->mem_cgroup = NULL;
5882 static void uncharge_list(struct list_head *page_list)
5884 struct uncharge_gather ug;
5885 struct list_head *next;
5887 uncharge_gather_clear(&ug);
5890 * Note that the list can be a single page->lru; hence the
5891 * do-while loop instead of a simple list_for_each_entry().
5893 next = page_list->next;
5897 page = list_entry(next, struct page, lru);
5898 next = page->lru.next;
5900 uncharge_page(page, &ug);
5901 } while (next != page_list);
5904 uncharge_batch(&ug);
5908 * mem_cgroup_uncharge - uncharge a page
5909 * @page: page to uncharge
5911 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5912 * mem_cgroup_commit_charge().
5914 void mem_cgroup_uncharge(struct page *page)
5916 struct uncharge_gather ug;
5918 if (mem_cgroup_disabled())
5921 /* Don't touch page->lru of any random page, pre-check: */
5922 if (!page->mem_cgroup)
5925 uncharge_gather_clear(&ug);
5926 uncharge_page(page, &ug);
5927 uncharge_batch(&ug);
5931 * mem_cgroup_uncharge_list - uncharge a list of page
5932 * @page_list: list of pages to uncharge
5934 * Uncharge a list of pages previously charged with
5935 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5937 void mem_cgroup_uncharge_list(struct list_head *page_list)
5939 if (mem_cgroup_disabled())
5942 if (!list_empty(page_list))
5943 uncharge_list(page_list);
5947 * mem_cgroup_migrate - charge a page's replacement
5948 * @oldpage: currently circulating page
5949 * @newpage: replacement page
5951 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5952 * be uncharged upon free.
5954 * Both pages must be locked, @newpage->mapping must be set up.
5956 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5958 struct mem_cgroup *memcg;
5959 unsigned int nr_pages;
5961 unsigned long flags;
5963 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5964 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5965 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5966 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5969 if (mem_cgroup_disabled())
5972 /* Page cache replacement: new page already charged? */
5973 if (newpage->mem_cgroup)
5976 /* Swapcache readahead pages can get replaced before being charged */
5977 memcg = oldpage->mem_cgroup;
5981 /* Force-charge the new page. The old one will be freed soon */
5982 compound = PageTransHuge(newpage);
5983 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5985 page_counter_charge(&memcg->memory, nr_pages);
5986 if (do_memsw_account())
5987 page_counter_charge(&memcg->memsw, nr_pages);
5988 css_get_many(&memcg->css, nr_pages);
5990 commit_charge(newpage, memcg, false);
5992 local_irq_save(flags);
5993 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5994 memcg_check_events(memcg, newpage);
5995 local_irq_restore(flags);
5998 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5999 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6001 void mem_cgroup_sk_alloc(struct sock *sk)
6003 struct mem_cgroup *memcg;
6005 if (!mem_cgroup_sockets_enabled)
6009 * Socket cloning can throw us here with sk_memcg already
6010 * filled. It won't however, necessarily happen from
6011 * process context. So the test for root memcg given
6012 * the current task's memcg won't help us in this case.
6014 * Respecting the original socket's memcg is a better
6015 * decision in this case.
6018 css_get(&sk->sk_memcg->css);
6023 memcg = mem_cgroup_from_task(current);
6024 if (memcg == root_mem_cgroup)
6026 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6028 if (css_tryget_online(&memcg->css))
6029 sk->sk_memcg = memcg;
6034 void mem_cgroup_sk_free(struct sock *sk)
6037 css_put(&sk->sk_memcg->css);
6041 * mem_cgroup_charge_skmem - charge socket memory
6042 * @memcg: memcg to charge
6043 * @nr_pages: number of pages to charge
6045 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6046 * @memcg's configured limit, %false if the charge had to be forced.
6048 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6050 gfp_t gfp_mask = GFP_KERNEL;
6052 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6053 struct page_counter *fail;
6055 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6056 memcg->tcpmem_pressure = 0;
6059 page_counter_charge(&memcg->tcpmem, nr_pages);
6060 memcg->tcpmem_pressure = 1;
6064 /* Don't block in the packet receive path */
6066 gfp_mask = GFP_NOWAIT;
6068 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6070 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6073 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6078 * mem_cgroup_uncharge_skmem - uncharge socket memory
6079 * @memcg: memcg to uncharge
6080 * @nr_pages: number of pages to uncharge
6082 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6084 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6085 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6089 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6091 refill_stock(memcg, nr_pages);
6094 static int __init cgroup_memory(char *s)
6098 while ((token = strsep(&s, ",")) != NULL) {
6101 if (!strcmp(token, "nosocket"))
6102 cgroup_memory_nosocket = true;
6103 if (!strcmp(token, "nokmem"))
6104 cgroup_memory_nokmem = true;
6108 __setup("cgroup.memory=", cgroup_memory);
6111 * subsys_initcall() for memory controller.
6113 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6114 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6115 * basically everything that doesn't depend on a specific mem_cgroup structure
6116 * should be initialized from here.
6118 static int __init mem_cgroup_init(void)
6122 #ifdef CONFIG_MEMCG_KMEM
6124 * Kmem cache creation is mostly done with the slab_mutex held,
6125 * so use a workqueue with limited concurrency to avoid stalling
6126 * all worker threads in case lots of cgroups are created and
6127 * destroyed simultaneously.
6129 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6130 BUG_ON(!memcg_kmem_cache_wq);
6133 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6134 memcg_hotplug_cpu_dead);
6136 for_each_possible_cpu(cpu)
6137 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6140 for_each_node(node) {
6141 struct mem_cgroup_tree_per_node *rtpn;
6143 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6144 node_online(node) ? node : NUMA_NO_NODE);
6146 rtpn->rb_root = RB_ROOT;
6147 rtpn->rb_rightmost = NULL;
6148 spin_lock_init(&rtpn->lock);
6149 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6154 subsys_initcall(mem_cgroup_init);
6156 #ifdef CONFIG_MEMCG_SWAP
6157 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6159 while (!atomic_inc_not_zero(&memcg->id.ref)) {
6161 * The root cgroup cannot be destroyed, so it's refcount must
6164 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6168 memcg = parent_mem_cgroup(memcg);
6170 memcg = root_mem_cgroup;
6176 * mem_cgroup_swapout - transfer a memsw charge to swap
6177 * @page: page whose memsw charge to transfer
6178 * @entry: swap entry to move the charge to
6180 * Transfer the memsw charge of @page to @entry.
6182 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6184 struct mem_cgroup *memcg, *swap_memcg;
6185 unsigned int nr_entries;
6186 unsigned short oldid;
6188 VM_BUG_ON_PAGE(PageLRU(page), page);
6189 VM_BUG_ON_PAGE(page_count(page), page);
6191 if (!do_memsw_account())
6194 memcg = page->mem_cgroup;
6196 /* Readahead page, never charged */
6201 * In case the memcg owning these pages has been offlined and doesn't
6202 * have an ID allocated to it anymore, charge the closest online
6203 * ancestor for the swap instead and transfer the memory+swap charge.
6205 swap_memcg = mem_cgroup_id_get_online(memcg);
6206 nr_entries = hpage_nr_pages(page);
6207 /* Get references for the tail pages, too */
6209 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6210 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6212 VM_BUG_ON_PAGE(oldid, page);
6213 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6215 page->mem_cgroup = NULL;
6217 if (!mem_cgroup_is_root(memcg))
6218 page_counter_uncharge(&memcg->memory, nr_entries);
6220 if (memcg != swap_memcg) {
6221 if (!mem_cgroup_is_root(swap_memcg))
6222 page_counter_charge(&swap_memcg->memsw, nr_entries);
6223 page_counter_uncharge(&memcg->memsw, nr_entries);
6227 * Interrupts should be disabled here because the caller holds the
6228 * i_pages lock which is taken with interrupts-off. It is
6229 * important here to have the interrupts disabled because it is the
6230 * only synchronisation we have for updating the per-CPU variables.
6232 VM_BUG_ON(!irqs_disabled());
6233 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6235 memcg_check_events(memcg, page);
6237 if (!mem_cgroup_is_root(memcg))
6238 css_put_many(&memcg->css, nr_entries);
6242 * mem_cgroup_try_charge_swap - try charging swap space for a page
6243 * @page: page being added to swap
6244 * @entry: swap entry to charge
6246 * Try to charge @page's memcg for the swap space at @entry.
6248 * Returns 0 on success, -ENOMEM on failure.
6250 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6252 unsigned int nr_pages = hpage_nr_pages(page);
6253 struct page_counter *counter;
6254 struct mem_cgroup *memcg;
6255 unsigned short oldid;
6257 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6260 memcg = page->mem_cgroup;
6262 /* Readahead page, never charged */
6267 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6271 memcg = mem_cgroup_id_get_online(memcg);
6273 if (!mem_cgroup_is_root(memcg) &&
6274 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6275 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6276 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6277 mem_cgroup_id_put(memcg);
6281 /* Get references for the tail pages, too */
6283 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6284 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6285 VM_BUG_ON_PAGE(oldid, page);
6286 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
6292 * mem_cgroup_uncharge_swap - uncharge swap space
6293 * @entry: swap entry to uncharge
6294 * @nr_pages: the amount of swap space to uncharge
6296 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6298 struct mem_cgroup *memcg;
6301 if (!do_swap_account)
6304 id = swap_cgroup_record(entry, 0, nr_pages);
6306 memcg = mem_cgroup_from_id(id);
6308 if (!mem_cgroup_is_root(memcg)) {
6309 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6310 page_counter_uncharge(&memcg->swap, nr_pages);
6312 page_counter_uncharge(&memcg->memsw, nr_pages);
6314 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
6315 mem_cgroup_id_put_many(memcg, nr_pages);
6320 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6322 long nr_swap_pages = get_nr_swap_pages();
6324 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6325 return nr_swap_pages;
6326 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6327 nr_swap_pages = min_t(long, nr_swap_pages,
6328 READ_ONCE(memcg->swap.max) -
6329 page_counter_read(&memcg->swap));
6330 return nr_swap_pages;
6333 bool mem_cgroup_swap_full(struct page *page)
6335 struct mem_cgroup *memcg;
6337 VM_BUG_ON_PAGE(!PageLocked(page), page);
6341 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6344 memcg = page->mem_cgroup;
6348 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6349 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
6355 /* for remember boot option*/
6356 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6357 static int really_do_swap_account __initdata = 1;
6359 static int really_do_swap_account __initdata;
6362 static int __init enable_swap_account(char *s)
6364 if (!strcmp(s, "1"))
6365 really_do_swap_account = 1;
6366 else if (!strcmp(s, "0"))
6367 really_do_swap_account = 0;
6370 __setup("swapaccount=", enable_swap_account);
6372 static u64 swap_current_read(struct cgroup_subsys_state *css,
6375 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6377 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6380 static int swap_max_show(struct seq_file *m, void *v)
6382 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6383 unsigned long max = READ_ONCE(memcg->swap.max);
6385 if (max == PAGE_COUNTER_MAX)
6386 seq_puts(m, "max\n");
6388 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6393 static ssize_t swap_max_write(struct kernfs_open_file *of,
6394 char *buf, size_t nbytes, loff_t off)
6396 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6400 buf = strstrip(buf);
6401 err = page_counter_memparse(buf, "max", &max);
6405 xchg(&memcg->swap.max, max);
6410 static int swap_events_show(struct seq_file *m, void *v)
6412 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6414 seq_printf(m, "max %lu\n",
6415 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
6416 seq_printf(m, "fail %lu\n",
6417 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
6422 static struct cftype swap_files[] = {
6424 .name = "swap.current",
6425 .flags = CFTYPE_NOT_ON_ROOT,
6426 .read_u64 = swap_current_read,
6430 .flags = CFTYPE_NOT_ON_ROOT,
6431 .seq_show = swap_max_show,
6432 .write = swap_max_write,
6435 .name = "swap.events",
6436 .flags = CFTYPE_NOT_ON_ROOT,
6437 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
6438 .seq_show = swap_events_show,
6443 static struct cftype memsw_cgroup_files[] = {
6445 .name = "memsw.usage_in_bytes",
6446 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6447 .read_u64 = mem_cgroup_read_u64,
6450 .name = "memsw.max_usage_in_bytes",
6451 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6452 .write = mem_cgroup_reset,
6453 .read_u64 = mem_cgroup_read_u64,
6456 .name = "memsw.limit_in_bytes",
6457 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6458 .write = mem_cgroup_write,
6459 .read_u64 = mem_cgroup_read_u64,
6462 .name = "memsw.failcnt",
6463 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6464 .write = mem_cgroup_reset,
6465 .read_u64 = mem_cgroup_read_u64,
6467 { }, /* terminate */
6470 static int __init mem_cgroup_swap_init(void)
6472 if (!mem_cgroup_disabled() && really_do_swap_account) {
6473 do_swap_account = 1;
6474 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6476 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6477 memsw_cgroup_files));
6481 subsys_initcall(mem_cgroup_swap_init);
6483 #endif /* CONFIG_MEMCG_SWAP */