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;
125 struct mem_cgroup_tree {
126 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
129 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
132 struct mem_cgroup_eventfd_list {
133 struct list_head list;
134 struct eventfd_ctx *eventfd;
138 * cgroup_event represents events which userspace want to receive.
140 struct mem_cgroup_event {
142 * memcg which the event belongs to.
144 struct mem_cgroup *memcg;
146 * eventfd to signal userspace about the event.
148 struct eventfd_ctx *eventfd;
150 * Each of these stored in a list by the cgroup.
152 struct list_head list;
154 * register_event() callback will be used to add new userspace
155 * waiter for changes related to this event. Use eventfd_signal()
156 * on eventfd to send notification to userspace.
158 int (*register_event)(struct mem_cgroup *memcg,
159 struct eventfd_ctx *eventfd, const char *args);
161 * unregister_event() callback will be called when userspace closes
162 * the eventfd or on cgroup removing. This callback must be set,
163 * if you want provide notification functionality.
165 void (*unregister_event)(struct mem_cgroup *memcg,
166 struct eventfd_ctx *eventfd);
168 * All fields below needed to unregister event when
169 * userspace closes eventfd.
172 wait_queue_head_t *wqh;
173 wait_queue_entry_t wait;
174 struct work_struct remove;
177 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
178 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
180 /* Stuffs for move charges at task migration. */
182 * Types of charges to be moved.
184 #define MOVE_ANON 0x1U
185 #define MOVE_FILE 0x2U
186 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
188 /* "mc" and its members are protected by cgroup_mutex */
189 static struct move_charge_struct {
190 spinlock_t lock; /* for from, to */
191 struct mm_struct *mm;
192 struct mem_cgroup *from;
193 struct mem_cgroup *to;
195 unsigned long precharge;
196 unsigned long moved_charge;
197 unsigned long moved_swap;
198 struct task_struct *moving_task; /* a task moving charges */
199 wait_queue_head_t waitq; /* a waitq for other context */
201 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
202 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
206 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
207 * limit reclaim to prevent infinite loops, if they ever occur.
209 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
210 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
213 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
214 MEM_CGROUP_CHARGE_TYPE_ANON,
215 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
216 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
220 /* for encoding cft->private value on file */
229 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
230 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
231 #define MEMFILE_ATTR(val) ((val) & 0xffff)
232 /* Used for OOM nofiier */
233 #define OOM_CONTROL (0)
235 /* Some nice accessors for the vmpressure. */
236 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
239 memcg = root_mem_cgroup;
240 return &memcg->vmpressure;
243 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
245 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
248 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
250 return (memcg == root_mem_cgroup);
255 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
256 * The main reason for not using cgroup id for this:
257 * this works better in sparse environments, where we have a lot of memcgs,
258 * but only a few kmem-limited. Or also, if we have, for instance, 200
259 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
260 * 200 entry array for that.
262 * The current size of the caches array is stored in memcg_nr_cache_ids. It
263 * will double each time we have to increase it.
265 static DEFINE_IDA(memcg_cache_ida);
266 int memcg_nr_cache_ids;
268 /* Protects memcg_nr_cache_ids */
269 static DECLARE_RWSEM(memcg_cache_ids_sem);
271 void memcg_get_cache_ids(void)
273 down_read(&memcg_cache_ids_sem);
276 void memcg_put_cache_ids(void)
278 up_read(&memcg_cache_ids_sem);
282 * MIN_SIZE is different than 1, because we would like to avoid going through
283 * the alloc/free process all the time. In a small machine, 4 kmem-limited
284 * cgroups is a reasonable guess. In the future, it could be a parameter or
285 * tunable, but that is strictly not necessary.
287 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
288 * this constant directly from cgroup, but it is understandable that this is
289 * better kept as an internal representation in cgroup.c. In any case, the
290 * cgrp_id space is not getting any smaller, and we don't have to necessarily
291 * increase ours as well if it increases.
293 #define MEMCG_CACHES_MIN_SIZE 4
294 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
297 * A lot of the calls to the cache allocation functions are expected to be
298 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
299 * conditional to this static branch, we'll have to allow modules that does
300 * kmem_cache_alloc and the such to see this symbol as well
302 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
303 EXPORT_SYMBOL(memcg_kmem_enabled_key);
305 struct workqueue_struct *memcg_kmem_cache_wq;
307 #endif /* !CONFIG_SLOB */
310 * mem_cgroup_css_from_page - css of the memcg associated with a page
311 * @page: page of interest
313 * If memcg is bound to the default hierarchy, css of the memcg associated
314 * with @page is returned. The returned css remains associated with @page
315 * until it is released.
317 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
320 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
322 struct mem_cgroup *memcg;
324 memcg = page->mem_cgroup;
326 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
327 memcg = root_mem_cgroup;
333 * page_cgroup_ino - return inode number of the memcg a page is charged to
336 * Look up the closest online ancestor of the memory cgroup @page is charged to
337 * and return its inode number or 0 if @page is not charged to any cgroup. It
338 * is safe to call this function without holding a reference to @page.
340 * Note, this function is inherently racy, because there is nothing to prevent
341 * the cgroup inode from getting torn down and potentially reallocated a moment
342 * after page_cgroup_ino() returns, so it only should be used by callers that
343 * do not care (such as procfs interfaces).
345 ino_t page_cgroup_ino(struct page *page)
347 struct mem_cgroup *memcg;
348 unsigned long ino = 0;
351 memcg = READ_ONCE(page->mem_cgroup);
352 while (memcg && !(memcg->css.flags & CSS_ONLINE))
353 memcg = parent_mem_cgroup(memcg);
355 ino = cgroup_ino(memcg->css.cgroup);
360 static struct mem_cgroup_per_node *
361 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
363 int nid = page_to_nid(page);
365 return memcg->nodeinfo[nid];
368 static struct mem_cgroup_tree_per_node *
369 soft_limit_tree_node(int nid)
371 return soft_limit_tree.rb_tree_per_node[nid];
374 static struct mem_cgroup_tree_per_node *
375 soft_limit_tree_from_page(struct page *page)
377 int nid = page_to_nid(page);
379 return soft_limit_tree.rb_tree_per_node[nid];
382 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
383 struct mem_cgroup_tree_per_node *mctz,
384 unsigned long new_usage_in_excess)
386 struct rb_node **p = &mctz->rb_root.rb_node;
387 struct rb_node *parent = NULL;
388 struct mem_cgroup_per_node *mz_node;
393 mz->usage_in_excess = new_usage_in_excess;
394 if (!mz->usage_in_excess)
398 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
400 if (mz->usage_in_excess < mz_node->usage_in_excess)
403 * We can't avoid mem cgroups that are over their soft
404 * limit by the same amount
406 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
409 rb_link_node(&mz->tree_node, parent, p);
410 rb_insert_color(&mz->tree_node, &mctz->rb_root);
414 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
415 struct mem_cgroup_tree_per_node *mctz)
419 rb_erase(&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)
428 spin_lock_irqsave(&mctz->lock, flags);
429 __mem_cgroup_remove_exceeded(mz, mctz);
430 spin_unlock_irqrestore(&mctz->lock, flags);
433 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
435 unsigned long nr_pages = page_counter_read(&memcg->memory);
436 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
437 unsigned long excess = 0;
439 if (nr_pages > soft_limit)
440 excess = nr_pages - soft_limit;
445 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
447 unsigned long excess;
448 struct mem_cgroup_per_node *mz;
449 struct mem_cgroup_tree_per_node *mctz;
451 mctz = soft_limit_tree_from_page(page);
455 * Necessary to update all ancestors when hierarchy is used.
456 * because their event counter is not touched.
458 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
459 mz = mem_cgroup_page_nodeinfo(memcg, page);
460 excess = soft_limit_excess(memcg);
462 * We have to update the tree if mz is on RB-tree or
463 * mem is over its softlimit.
465 if (excess || mz->on_tree) {
468 spin_lock_irqsave(&mctz->lock, flags);
469 /* if on-tree, remove it */
471 __mem_cgroup_remove_exceeded(mz, mctz);
473 * Insert again. mz->usage_in_excess will be updated.
474 * If excess is 0, no tree ops.
476 __mem_cgroup_insert_exceeded(mz, mctz, excess);
477 spin_unlock_irqrestore(&mctz->lock, flags);
482 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
484 struct mem_cgroup_tree_per_node *mctz;
485 struct mem_cgroup_per_node *mz;
489 mz = mem_cgroup_nodeinfo(memcg, nid);
490 mctz = soft_limit_tree_node(nid);
492 mem_cgroup_remove_exceeded(mz, mctz);
496 static struct mem_cgroup_per_node *
497 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
499 struct rb_node *rightmost = NULL;
500 struct mem_cgroup_per_node *mz;
504 rightmost = rb_last(&mctz->rb_root);
506 goto done; /* Nothing to reclaim from */
508 mz = rb_entry(rightmost, struct mem_cgroup_per_node, tree_node);
510 * Remove the node now but someone else can add it back,
511 * we will to add it back at the end of reclaim to its correct
512 * position in the tree.
514 __mem_cgroup_remove_exceeded(mz, mctz);
515 if (!soft_limit_excess(mz->memcg) ||
516 !css_tryget_online(&mz->memcg->css))
522 static struct mem_cgroup_per_node *
523 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
525 struct mem_cgroup_per_node *mz;
527 spin_lock_irq(&mctz->lock);
528 mz = __mem_cgroup_largest_soft_limit_node(mctz);
529 spin_unlock_irq(&mctz->lock);
534 * Return page count for single (non recursive) @memcg.
536 * Implementation Note: reading percpu statistics for memcg.
538 * Both of vmstat[] and percpu_counter has threshold and do periodic
539 * synchronization to implement "quick" read. There are trade-off between
540 * reading cost and precision of value. Then, we may have a chance to implement
541 * a periodic synchronization of counter in memcg's counter.
543 * But this _read() function is used for user interface now. The user accounts
544 * memory usage by memory cgroup and he _always_ requires exact value because
545 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
546 * have to visit all online cpus and make sum. So, for now, unnecessary
547 * synchronization is not implemented. (just implemented for cpu hotplug)
549 * If there are kernel internal actions which can make use of some not-exact
550 * value, and reading all cpu value can be performance bottleneck in some
551 * common workload, threshold and synchronization as vmstat[] should be
554 * The parameter idx can be of type enum memcg_event_item or vm_event_item.
557 static unsigned long memcg_sum_events(struct mem_cgroup *memcg,
560 unsigned long val = 0;
563 for_each_possible_cpu(cpu)
564 val += per_cpu(memcg->stat->events[event], cpu);
568 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
570 bool compound, int nr_pages)
573 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
574 * counted as CACHE even if it's on ANON LRU.
577 __this_cpu_add(memcg->stat->count[MEMCG_RSS], nr_pages);
579 __this_cpu_add(memcg->stat->count[MEMCG_CACHE], nr_pages);
580 if (PageSwapBacked(page))
581 __this_cpu_add(memcg->stat->count[NR_SHMEM], nr_pages);
585 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
586 __this_cpu_add(memcg->stat->count[MEMCG_RSS_HUGE], nr_pages);
589 /* pagein of a big page is an event. So, ignore page size */
591 __this_cpu_inc(memcg->stat->events[PGPGIN]);
593 __this_cpu_inc(memcg->stat->events[PGPGOUT]);
594 nr_pages = -nr_pages; /* for event */
597 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
600 unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
601 int nid, unsigned int lru_mask)
603 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
604 unsigned long nr = 0;
607 VM_BUG_ON((unsigned)nid >= nr_node_ids);
610 if (!(BIT(lru) & lru_mask))
612 nr += mem_cgroup_get_lru_size(lruvec, lru);
617 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
618 unsigned int lru_mask)
620 unsigned long nr = 0;
623 for_each_node_state(nid, N_MEMORY)
624 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
628 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
629 enum mem_cgroup_events_target target)
631 unsigned long val, next;
633 val = __this_cpu_read(memcg->stat->nr_page_events);
634 next = __this_cpu_read(memcg->stat->targets[target]);
635 /* from time_after() in jiffies.h */
636 if ((long)(next - val) < 0) {
638 case MEM_CGROUP_TARGET_THRESH:
639 next = val + THRESHOLDS_EVENTS_TARGET;
641 case MEM_CGROUP_TARGET_SOFTLIMIT:
642 next = val + SOFTLIMIT_EVENTS_TARGET;
644 case MEM_CGROUP_TARGET_NUMAINFO:
645 next = val + NUMAINFO_EVENTS_TARGET;
650 __this_cpu_write(memcg->stat->targets[target], next);
657 * Check events in order.
660 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
662 /* threshold event is triggered in finer grain than soft limit */
663 if (unlikely(mem_cgroup_event_ratelimit(memcg,
664 MEM_CGROUP_TARGET_THRESH))) {
666 bool do_numainfo __maybe_unused;
668 do_softlimit = mem_cgroup_event_ratelimit(memcg,
669 MEM_CGROUP_TARGET_SOFTLIMIT);
671 do_numainfo = mem_cgroup_event_ratelimit(memcg,
672 MEM_CGROUP_TARGET_NUMAINFO);
674 mem_cgroup_threshold(memcg);
675 if (unlikely(do_softlimit))
676 mem_cgroup_update_tree(memcg, page);
678 if (unlikely(do_numainfo))
679 atomic_inc(&memcg->numainfo_events);
684 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
687 * mm_update_next_owner() may clear mm->owner to NULL
688 * if it races with swapoff, page migration, etc.
689 * So this can be called with p == NULL.
694 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
696 EXPORT_SYMBOL(mem_cgroup_from_task);
698 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
700 struct mem_cgroup *memcg = NULL;
705 * Page cache insertions can happen withou an
706 * actual mm context, e.g. during disk probing
707 * on boot, loopback IO, acct() writes etc.
710 memcg = root_mem_cgroup;
712 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
713 if (unlikely(!memcg))
714 memcg = root_mem_cgroup;
716 } while (!css_tryget_online(&memcg->css));
722 * mem_cgroup_iter - iterate over memory cgroup hierarchy
723 * @root: hierarchy root
724 * @prev: previously returned memcg, NULL on first invocation
725 * @reclaim: cookie for shared reclaim walks, NULL for full walks
727 * Returns references to children of the hierarchy below @root, or
728 * @root itself, or %NULL after a full round-trip.
730 * Caller must pass the return value in @prev on subsequent
731 * invocations for reference counting, or use mem_cgroup_iter_break()
732 * to cancel a hierarchy walk before the round-trip is complete.
734 * Reclaimers can specify a zone and a priority level in @reclaim to
735 * divide up the memcgs in the hierarchy among all concurrent
736 * reclaimers operating on the same zone and priority.
738 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
739 struct mem_cgroup *prev,
740 struct mem_cgroup_reclaim_cookie *reclaim)
742 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
743 struct cgroup_subsys_state *css = NULL;
744 struct mem_cgroup *memcg = NULL;
745 struct mem_cgroup *pos = NULL;
747 if (mem_cgroup_disabled())
751 root = root_mem_cgroup;
753 if (prev && !reclaim)
756 if (!root->use_hierarchy && root != root_mem_cgroup) {
765 struct mem_cgroup_per_node *mz;
767 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
768 iter = &mz->iter[reclaim->priority];
770 if (prev && reclaim->generation != iter->generation)
774 pos = READ_ONCE(iter->position);
775 if (!pos || css_tryget(&pos->css))
778 * css reference reached zero, so iter->position will
779 * be cleared by ->css_released. However, we should not
780 * rely on this happening soon, because ->css_released
781 * is called from a work queue, and by busy-waiting we
782 * might block it. So we clear iter->position right
785 (void)cmpxchg(&iter->position, pos, NULL);
793 css = css_next_descendant_pre(css, &root->css);
796 * Reclaimers share the hierarchy walk, and a
797 * new one might jump in right at the end of
798 * the hierarchy - make sure they see at least
799 * one group and restart from the beginning.
807 * Verify the css and acquire a reference. The root
808 * is provided by the caller, so we know it's alive
809 * and kicking, and don't take an extra reference.
811 memcg = mem_cgroup_from_css(css);
813 if (css == &root->css)
824 * The position could have already been updated by a competing
825 * thread, so check that the value hasn't changed since we read
826 * it to avoid reclaiming from the same cgroup twice.
828 (void)cmpxchg(&iter->position, pos, memcg);
836 reclaim->generation = iter->generation;
842 if (prev && prev != root)
849 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
850 * @root: hierarchy root
851 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
853 void mem_cgroup_iter_break(struct mem_cgroup *root,
854 struct mem_cgroup *prev)
857 root = root_mem_cgroup;
858 if (prev && prev != root)
862 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
864 struct mem_cgroup *memcg = dead_memcg;
865 struct mem_cgroup_reclaim_iter *iter;
866 struct mem_cgroup_per_node *mz;
870 while ((memcg = parent_mem_cgroup(memcg))) {
872 mz = mem_cgroup_nodeinfo(memcg, nid);
873 for (i = 0; i <= DEF_PRIORITY; i++) {
875 cmpxchg(&iter->position,
883 * Iteration constructs for visiting all cgroups (under a tree). If
884 * loops are exited prematurely (break), mem_cgroup_iter_break() must
885 * be used for reference counting.
887 #define for_each_mem_cgroup_tree(iter, root) \
888 for (iter = mem_cgroup_iter(root, NULL, NULL); \
890 iter = mem_cgroup_iter(root, iter, NULL))
892 #define for_each_mem_cgroup(iter) \
893 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
895 iter = mem_cgroup_iter(NULL, iter, NULL))
898 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
899 * @memcg: hierarchy root
900 * @fn: function to call for each task
901 * @arg: argument passed to @fn
903 * This function iterates over tasks attached to @memcg or to any of its
904 * descendants and calls @fn for each task. If @fn returns a non-zero
905 * value, the function breaks the iteration loop and returns the value.
906 * Otherwise, it will iterate over all tasks and return 0.
908 * This function must not be called for the root memory cgroup.
910 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
911 int (*fn)(struct task_struct *, void *), void *arg)
913 struct mem_cgroup *iter;
916 BUG_ON(memcg == root_mem_cgroup);
918 for_each_mem_cgroup_tree(iter, memcg) {
919 struct css_task_iter it;
920 struct task_struct *task;
922 css_task_iter_start(&iter->css, 0, &it);
923 while (!ret && (task = css_task_iter_next(&it)))
925 css_task_iter_end(&it);
927 mem_cgroup_iter_break(memcg, iter);
935 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
937 * @zone: zone of the page
939 * This function is only safe when following the LRU page isolation
940 * and putback protocol: the LRU lock must be held, and the page must
941 * either be PageLRU() or the caller must have isolated/allocated it.
943 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
945 struct mem_cgroup_per_node *mz;
946 struct mem_cgroup *memcg;
947 struct lruvec *lruvec;
949 if (mem_cgroup_disabled()) {
950 lruvec = &pgdat->lruvec;
954 memcg = page->mem_cgroup;
956 * Swapcache readahead pages are added to the LRU - and
957 * possibly migrated - before they are charged.
960 memcg = root_mem_cgroup;
962 mz = mem_cgroup_page_nodeinfo(memcg, page);
963 lruvec = &mz->lruvec;
966 * Since a node can be onlined after the mem_cgroup was created,
967 * we have to be prepared to initialize lruvec->zone here;
968 * and if offlined then reonlined, we need to reinitialize it.
970 if (unlikely(lruvec->pgdat != pgdat))
971 lruvec->pgdat = pgdat;
976 * mem_cgroup_update_lru_size - account for adding or removing an lru page
977 * @lruvec: mem_cgroup per zone lru vector
978 * @lru: index of lru list the page is sitting on
979 * @zid: zone id of the accounted pages
980 * @nr_pages: positive when adding or negative when removing
982 * This function must be called under lru_lock, just before a page is added
983 * to or just after a page is removed from an lru list (that ordering being
984 * so as to allow it to check that lru_size 0 is consistent with list_empty).
986 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
987 int zid, int nr_pages)
989 struct mem_cgroup_per_node *mz;
990 unsigned long *lru_size;
993 if (mem_cgroup_disabled())
996 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
997 lru_size = &mz->lru_zone_size[zid][lru];
1000 *lru_size += nr_pages;
1003 if (WARN_ONCE(size < 0,
1004 "%s(%p, %d, %d): lru_size %ld\n",
1005 __func__, lruvec, lru, nr_pages, size)) {
1011 *lru_size += nr_pages;
1014 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1016 struct mem_cgroup *task_memcg;
1017 struct task_struct *p;
1020 p = find_lock_task_mm(task);
1022 task_memcg = get_mem_cgroup_from_mm(p->mm);
1026 * All threads may have already detached their mm's, but the oom
1027 * killer still needs to detect if they have already been oom
1028 * killed to prevent needlessly killing additional tasks.
1031 task_memcg = mem_cgroup_from_task(task);
1032 css_get(&task_memcg->css);
1035 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1036 css_put(&task_memcg->css);
1041 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1042 * @memcg: the memory cgroup
1044 * Returns the maximum amount of memory @mem can be charged with, in
1047 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1049 unsigned long margin = 0;
1050 unsigned long count;
1051 unsigned long limit;
1053 count = page_counter_read(&memcg->memory);
1054 limit = READ_ONCE(memcg->memory.limit);
1056 margin = limit - count;
1058 if (do_memsw_account()) {
1059 count = page_counter_read(&memcg->memsw);
1060 limit = READ_ONCE(memcg->memsw.limit);
1062 margin = min(margin, limit - count);
1071 * A routine for checking "mem" is under move_account() or not.
1073 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1074 * moving cgroups. This is for waiting at high-memory pressure
1077 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1079 struct mem_cgroup *from;
1080 struct mem_cgroup *to;
1083 * Unlike task_move routines, we access mc.to, mc.from not under
1084 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1086 spin_lock(&mc.lock);
1092 ret = mem_cgroup_is_descendant(from, memcg) ||
1093 mem_cgroup_is_descendant(to, memcg);
1095 spin_unlock(&mc.lock);
1099 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1101 if (mc.moving_task && current != mc.moving_task) {
1102 if (mem_cgroup_under_move(memcg)) {
1104 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1105 /* moving charge context might have finished. */
1108 finish_wait(&mc.waitq, &wait);
1115 unsigned int memcg1_stats[] = {
1126 static const char *const memcg1_stat_names[] = {
1137 #define K(x) ((x) << (PAGE_SHIFT-10))
1139 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1140 * @memcg: The memory cgroup that went over limit
1141 * @p: Task that is going to be killed
1143 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1146 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1148 struct mem_cgroup *iter;
1154 pr_info("Task in ");
1155 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1156 pr_cont(" killed as a result of limit of ");
1158 pr_info("Memory limit reached of cgroup ");
1161 pr_cont_cgroup_path(memcg->css.cgroup);
1166 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1167 K((u64)page_counter_read(&memcg->memory)),
1168 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1169 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1170 K((u64)page_counter_read(&memcg->memsw)),
1171 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1172 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1173 K((u64)page_counter_read(&memcg->kmem)),
1174 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1176 for_each_mem_cgroup_tree(iter, memcg) {
1177 pr_info("Memory cgroup stats for ");
1178 pr_cont_cgroup_path(iter->css.cgroup);
1181 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
1182 if (memcg1_stats[i] == MEMCG_SWAP && !do_swap_account)
1184 pr_cont(" %s:%luKB", memcg1_stat_names[i],
1185 K(memcg_page_state(iter, memcg1_stats[i])));
1188 for (i = 0; i < NR_LRU_LISTS; i++)
1189 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1190 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1197 * This function returns the number of memcg under hierarchy tree. Returns
1198 * 1(self count) if no children.
1200 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1203 struct mem_cgroup *iter;
1205 for_each_mem_cgroup_tree(iter, memcg)
1211 * Return the memory (and swap, if configured) limit for a memcg.
1213 unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1215 unsigned long limit;
1217 limit = memcg->memory.limit;
1218 if (mem_cgroup_swappiness(memcg)) {
1219 unsigned long memsw_limit;
1220 unsigned long swap_limit;
1222 memsw_limit = memcg->memsw.limit;
1223 swap_limit = memcg->swap.limit;
1224 swap_limit = min(swap_limit, (unsigned long)total_swap_pages);
1225 limit = min(limit + swap_limit, memsw_limit);
1230 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1233 struct oom_control oc = {
1237 .gfp_mask = gfp_mask,
1242 mutex_lock(&oom_lock);
1243 ret = out_of_memory(&oc);
1244 mutex_unlock(&oom_lock);
1248 #if MAX_NUMNODES > 1
1251 * test_mem_cgroup_node_reclaimable
1252 * @memcg: the target memcg
1253 * @nid: the node ID to be checked.
1254 * @noswap : specify true here if the user wants flle only information.
1256 * This function returns whether the specified memcg contains any
1257 * reclaimable pages on a node. Returns true if there are any reclaimable
1258 * pages in the node.
1260 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1261 int nid, bool noswap)
1263 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1265 if (noswap || !total_swap_pages)
1267 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1274 * Always updating the nodemask is not very good - even if we have an empty
1275 * list or the wrong list here, we can start from some node and traverse all
1276 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1279 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1283 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1284 * pagein/pageout changes since the last update.
1286 if (!atomic_read(&memcg->numainfo_events))
1288 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1291 /* make a nodemask where this memcg uses memory from */
1292 memcg->scan_nodes = node_states[N_MEMORY];
1294 for_each_node_mask(nid, node_states[N_MEMORY]) {
1296 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1297 node_clear(nid, memcg->scan_nodes);
1300 atomic_set(&memcg->numainfo_events, 0);
1301 atomic_set(&memcg->numainfo_updating, 0);
1305 * Selecting a node where we start reclaim from. Because what we need is just
1306 * reducing usage counter, start from anywhere is O,K. Considering
1307 * memory reclaim from current node, there are pros. and cons.
1309 * Freeing memory from current node means freeing memory from a node which
1310 * we'll use or we've used. So, it may make LRU bad. And if several threads
1311 * hit limits, it will see a contention on a node. But freeing from remote
1312 * node means more costs for memory reclaim because of memory latency.
1314 * Now, we use round-robin. Better algorithm is welcomed.
1316 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1320 mem_cgroup_may_update_nodemask(memcg);
1321 node = memcg->last_scanned_node;
1323 node = next_node_in(node, memcg->scan_nodes);
1325 * mem_cgroup_may_update_nodemask might have seen no reclaimmable pages
1326 * last time it really checked all the LRUs due to rate limiting.
1327 * Fallback to the current node in that case for simplicity.
1329 if (unlikely(node == MAX_NUMNODES))
1330 node = numa_node_id();
1332 memcg->last_scanned_node = node;
1336 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1342 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1345 unsigned long *total_scanned)
1347 struct mem_cgroup *victim = NULL;
1350 unsigned long excess;
1351 unsigned long nr_scanned;
1352 struct mem_cgroup_reclaim_cookie reclaim = {
1357 excess = soft_limit_excess(root_memcg);
1360 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1365 * If we have not been able to reclaim
1366 * anything, it might because there are
1367 * no reclaimable pages under this hierarchy
1372 * We want to do more targeted reclaim.
1373 * excess >> 2 is not to excessive so as to
1374 * reclaim too much, nor too less that we keep
1375 * coming back to reclaim from this cgroup
1377 if (total >= (excess >> 2) ||
1378 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1383 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1384 pgdat, &nr_scanned);
1385 *total_scanned += nr_scanned;
1386 if (!soft_limit_excess(root_memcg))
1389 mem_cgroup_iter_break(root_memcg, victim);
1393 #ifdef CONFIG_LOCKDEP
1394 static struct lockdep_map memcg_oom_lock_dep_map = {
1395 .name = "memcg_oom_lock",
1399 static DEFINE_SPINLOCK(memcg_oom_lock);
1402 * Check OOM-Killer is already running under our hierarchy.
1403 * If someone is running, return false.
1405 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1407 struct mem_cgroup *iter, *failed = NULL;
1409 spin_lock(&memcg_oom_lock);
1411 for_each_mem_cgroup_tree(iter, memcg) {
1412 if (iter->oom_lock) {
1414 * this subtree of our hierarchy is already locked
1415 * so we cannot give a lock.
1418 mem_cgroup_iter_break(memcg, iter);
1421 iter->oom_lock = true;
1426 * OK, we failed to lock the whole subtree so we have
1427 * to clean up what we set up to the failing subtree
1429 for_each_mem_cgroup_tree(iter, memcg) {
1430 if (iter == failed) {
1431 mem_cgroup_iter_break(memcg, iter);
1434 iter->oom_lock = false;
1437 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1439 spin_unlock(&memcg_oom_lock);
1444 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1446 struct mem_cgroup *iter;
1448 spin_lock(&memcg_oom_lock);
1449 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1450 for_each_mem_cgroup_tree(iter, memcg)
1451 iter->oom_lock = false;
1452 spin_unlock(&memcg_oom_lock);
1455 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1457 struct mem_cgroup *iter;
1459 spin_lock(&memcg_oom_lock);
1460 for_each_mem_cgroup_tree(iter, memcg)
1462 spin_unlock(&memcg_oom_lock);
1465 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1467 struct mem_cgroup *iter;
1470 * When a new child is created while the hierarchy is under oom,
1471 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1473 spin_lock(&memcg_oom_lock);
1474 for_each_mem_cgroup_tree(iter, memcg)
1475 if (iter->under_oom > 0)
1477 spin_unlock(&memcg_oom_lock);
1480 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1482 struct oom_wait_info {
1483 struct mem_cgroup *memcg;
1484 wait_queue_entry_t wait;
1487 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1488 unsigned mode, int sync, void *arg)
1490 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1491 struct mem_cgroup *oom_wait_memcg;
1492 struct oom_wait_info *oom_wait_info;
1494 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1495 oom_wait_memcg = oom_wait_info->memcg;
1497 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1498 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1500 return autoremove_wake_function(wait, mode, sync, arg);
1503 static void memcg_oom_recover(struct mem_cgroup *memcg)
1506 * For the following lockless ->under_oom test, the only required
1507 * guarantee is that it must see the state asserted by an OOM when
1508 * this function is called as a result of userland actions
1509 * triggered by the notification of the OOM. This is trivially
1510 * achieved by invoking mem_cgroup_mark_under_oom() before
1511 * triggering notification.
1513 if (memcg && memcg->under_oom)
1514 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1517 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1519 if (!current->memcg_may_oom)
1522 * We are in the middle of the charge context here, so we
1523 * don't want to block when potentially sitting on a callstack
1524 * that holds all kinds of filesystem and mm locks.
1526 * Also, the caller may handle a failed allocation gracefully
1527 * (like optional page cache readahead) and so an OOM killer
1528 * invocation might not even be necessary.
1530 * That's why we don't do anything here except remember the
1531 * OOM context and then deal with it at the end of the page
1532 * fault when the stack is unwound, the locks are released,
1533 * and when we know whether the fault was overall successful.
1535 css_get(&memcg->css);
1536 current->memcg_in_oom = memcg;
1537 current->memcg_oom_gfp_mask = mask;
1538 current->memcg_oom_order = order;
1542 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1543 * @handle: actually kill/wait or just clean up the OOM state
1545 * This has to be called at the end of a page fault if the memcg OOM
1546 * handler was enabled.
1548 * Memcg supports userspace OOM handling where failed allocations must
1549 * sleep on a waitqueue until the userspace task resolves the
1550 * situation. Sleeping directly in the charge context with all kinds
1551 * of locks held is not a good idea, instead we remember an OOM state
1552 * in the task and mem_cgroup_oom_synchronize() has to be called at
1553 * the end of the page fault to complete the OOM handling.
1555 * Returns %true if an ongoing memcg OOM situation was detected and
1556 * completed, %false otherwise.
1558 bool mem_cgroup_oom_synchronize(bool handle)
1560 struct mem_cgroup *memcg = current->memcg_in_oom;
1561 struct oom_wait_info owait;
1564 /* OOM is global, do not handle */
1571 owait.memcg = memcg;
1572 owait.wait.flags = 0;
1573 owait.wait.func = memcg_oom_wake_function;
1574 owait.wait.private = current;
1575 INIT_LIST_HEAD(&owait.wait.entry);
1577 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1578 mem_cgroup_mark_under_oom(memcg);
1580 locked = mem_cgroup_oom_trylock(memcg);
1583 mem_cgroup_oom_notify(memcg);
1585 if (locked && !memcg->oom_kill_disable) {
1586 mem_cgroup_unmark_under_oom(memcg);
1587 finish_wait(&memcg_oom_waitq, &owait.wait);
1588 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1589 current->memcg_oom_order);
1592 mem_cgroup_unmark_under_oom(memcg);
1593 finish_wait(&memcg_oom_waitq, &owait.wait);
1597 mem_cgroup_oom_unlock(memcg);
1599 * There is no guarantee that an OOM-lock contender
1600 * sees the wakeups triggered by the OOM kill
1601 * uncharges. Wake any sleepers explicitely.
1603 memcg_oom_recover(memcg);
1606 current->memcg_in_oom = NULL;
1607 css_put(&memcg->css);
1612 * lock_page_memcg - lock a page->mem_cgroup binding
1615 * This function protects unlocked LRU pages from being moved to
1618 * It ensures lifetime of the returned memcg. Caller is responsible
1619 * for the lifetime of the page; __unlock_page_memcg() is available
1620 * when @page might get freed inside the locked section.
1622 struct mem_cgroup *lock_page_memcg(struct page *page)
1624 struct mem_cgroup *memcg;
1625 unsigned long flags;
1628 * The RCU lock is held throughout the transaction. The fast
1629 * path can get away without acquiring the memcg->move_lock
1630 * because page moving starts with an RCU grace period.
1632 * The RCU lock also protects the memcg from being freed when
1633 * the page state that is going to change is the only thing
1634 * preventing the page itself from being freed. E.g. writeback
1635 * doesn't hold a page reference and relies on PG_writeback to
1636 * keep off truncation, migration and so forth.
1640 if (mem_cgroup_disabled())
1643 memcg = page->mem_cgroup;
1644 if (unlikely(!memcg))
1647 if (atomic_read(&memcg->moving_account) <= 0)
1650 spin_lock_irqsave(&memcg->move_lock, flags);
1651 if (memcg != page->mem_cgroup) {
1652 spin_unlock_irqrestore(&memcg->move_lock, flags);
1657 * When charge migration first begins, we can have locked and
1658 * unlocked page stat updates happening concurrently. Track
1659 * the task who has the lock for unlock_page_memcg().
1661 memcg->move_lock_task = current;
1662 memcg->move_lock_flags = flags;
1666 EXPORT_SYMBOL(lock_page_memcg);
1669 * __unlock_page_memcg - unlock and unpin a memcg
1672 * Unlock and unpin a memcg returned by lock_page_memcg().
1674 void __unlock_page_memcg(struct mem_cgroup *memcg)
1676 if (memcg && memcg->move_lock_task == current) {
1677 unsigned long flags = memcg->move_lock_flags;
1679 memcg->move_lock_task = NULL;
1680 memcg->move_lock_flags = 0;
1682 spin_unlock_irqrestore(&memcg->move_lock, flags);
1689 * unlock_page_memcg - unlock a page->mem_cgroup binding
1692 void unlock_page_memcg(struct page *page)
1694 __unlock_page_memcg(page->mem_cgroup);
1696 EXPORT_SYMBOL(unlock_page_memcg);
1699 * size of first charge trial. "32" comes from vmscan.c's magic value.
1700 * TODO: maybe necessary to use big numbers in big irons.
1702 #define CHARGE_BATCH 32U
1703 struct memcg_stock_pcp {
1704 struct mem_cgroup *cached; /* this never be root cgroup */
1705 unsigned int nr_pages;
1706 struct work_struct work;
1707 unsigned long flags;
1708 #define FLUSHING_CACHED_CHARGE 0
1710 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1711 static DEFINE_MUTEX(percpu_charge_mutex);
1714 * consume_stock: Try to consume stocked charge on this cpu.
1715 * @memcg: memcg to consume from.
1716 * @nr_pages: how many pages to charge.
1718 * The charges will only happen if @memcg matches the current cpu's memcg
1719 * stock, and at least @nr_pages are available in that stock. Failure to
1720 * service an allocation will refill the stock.
1722 * returns true if successful, false otherwise.
1724 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1726 struct memcg_stock_pcp *stock;
1727 unsigned long flags;
1730 if (nr_pages > CHARGE_BATCH)
1733 local_irq_save(flags);
1735 stock = this_cpu_ptr(&memcg_stock);
1736 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
1737 stock->nr_pages -= nr_pages;
1741 local_irq_restore(flags);
1747 * Returns stocks cached in percpu and reset cached information.
1749 static void drain_stock(struct memcg_stock_pcp *stock)
1751 struct mem_cgroup *old = stock->cached;
1753 if (stock->nr_pages) {
1754 page_counter_uncharge(&old->memory, stock->nr_pages);
1755 if (do_memsw_account())
1756 page_counter_uncharge(&old->memsw, stock->nr_pages);
1757 css_put_many(&old->css, stock->nr_pages);
1758 stock->nr_pages = 0;
1760 stock->cached = NULL;
1763 static void drain_local_stock(struct work_struct *dummy)
1765 struct memcg_stock_pcp *stock;
1766 unsigned long flags;
1768 local_irq_save(flags);
1770 stock = this_cpu_ptr(&memcg_stock);
1772 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1774 local_irq_restore(flags);
1778 * Cache charges(val) to local per_cpu area.
1779 * This will be consumed by consume_stock() function, later.
1781 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
1783 struct memcg_stock_pcp *stock;
1784 unsigned long flags;
1786 local_irq_save(flags);
1788 stock = this_cpu_ptr(&memcg_stock);
1789 if (stock->cached != memcg) { /* reset if necessary */
1791 stock->cached = memcg;
1793 stock->nr_pages += nr_pages;
1795 local_irq_restore(flags);
1799 * Drains all per-CPU charge caches for given root_memcg resp. subtree
1800 * of the hierarchy under it.
1802 static void drain_all_stock(struct mem_cgroup *root_memcg)
1806 /* If someone's already draining, avoid adding running more workers. */
1807 if (!mutex_trylock(&percpu_charge_mutex))
1809 /* Notify other cpus that system-wide "drain" is running */
1812 for_each_online_cpu(cpu) {
1813 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1814 struct mem_cgroup *memcg;
1816 memcg = stock->cached;
1817 if (!memcg || !stock->nr_pages)
1819 if (!mem_cgroup_is_descendant(memcg, root_memcg))
1821 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
1823 drain_local_stock(&stock->work);
1825 schedule_work_on(cpu, &stock->work);
1830 mutex_unlock(&percpu_charge_mutex);
1833 static int memcg_hotplug_cpu_dead(unsigned int cpu)
1835 struct memcg_stock_pcp *stock;
1837 stock = &per_cpu(memcg_stock, cpu);
1842 static void reclaim_high(struct mem_cgroup *memcg,
1843 unsigned int nr_pages,
1847 if (page_counter_read(&memcg->memory) <= memcg->high)
1849 mem_cgroup_event(memcg, MEMCG_HIGH);
1850 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
1851 } while ((memcg = parent_mem_cgroup(memcg)));
1854 static void high_work_func(struct work_struct *work)
1856 struct mem_cgroup *memcg;
1858 memcg = container_of(work, struct mem_cgroup, high_work);
1859 reclaim_high(memcg, CHARGE_BATCH, GFP_KERNEL);
1863 * Scheduled by try_charge() to be executed from the userland return path
1864 * and reclaims memory over the high limit.
1866 void mem_cgroup_handle_over_high(void)
1868 unsigned int nr_pages = current->memcg_nr_pages_over_high;
1869 struct mem_cgroup *memcg;
1871 if (likely(!nr_pages))
1874 memcg = get_mem_cgroup_from_mm(current->mm);
1875 reclaim_high(memcg, nr_pages, GFP_KERNEL);
1876 css_put(&memcg->css);
1877 current->memcg_nr_pages_over_high = 0;
1880 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
1881 unsigned int nr_pages)
1883 unsigned int batch = max(CHARGE_BATCH, nr_pages);
1884 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1885 struct mem_cgroup *mem_over_limit;
1886 struct page_counter *counter;
1887 unsigned long nr_reclaimed;
1888 bool may_swap = true;
1889 bool drained = false;
1891 if (mem_cgroup_is_root(memcg))
1894 if (consume_stock(memcg, nr_pages))
1897 if (!do_memsw_account() ||
1898 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
1899 if (page_counter_try_charge(&memcg->memory, batch, &counter))
1901 if (do_memsw_account())
1902 page_counter_uncharge(&memcg->memsw, batch);
1903 mem_over_limit = mem_cgroup_from_counter(counter, memory);
1905 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
1909 if (batch > nr_pages) {
1915 * Unlike in global OOM situations, memcg is not in a physical
1916 * memory shortage. Allow dying and OOM-killed tasks to
1917 * bypass the last charges so that they can exit quickly and
1918 * free their memory.
1920 if (unlikely(tsk_is_oom_victim(current) ||
1921 fatal_signal_pending(current) ||
1922 current->flags & PF_EXITING))
1926 * Prevent unbounded recursion when reclaim operations need to
1927 * allocate memory. This might exceed the limits temporarily,
1928 * but we prefer facilitating memory reclaim and getting back
1929 * under the limit over triggering OOM kills in these cases.
1931 if (unlikely(current->flags & PF_MEMALLOC))
1934 if (unlikely(task_in_memcg_oom(current)))
1937 if (!gfpflags_allow_blocking(gfp_mask))
1940 mem_cgroup_event(mem_over_limit, MEMCG_MAX);
1942 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
1943 gfp_mask, may_swap);
1945 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
1949 drain_all_stock(mem_over_limit);
1954 if (gfp_mask & __GFP_NORETRY)
1957 * Even though the limit is exceeded at this point, reclaim
1958 * may have been able to free some pages. Retry the charge
1959 * before killing the task.
1961 * Only for regular pages, though: huge pages are rather
1962 * unlikely to succeed so close to the limit, and we fall back
1963 * to regular pages anyway in case of failure.
1965 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
1968 * At task move, charge accounts can be doubly counted. So, it's
1969 * better to wait until the end of task_move if something is going on.
1971 if (mem_cgroup_wait_acct_move(mem_over_limit))
1977 if (gfp_mask & __GFP_NOFAIL)
1980 if (fatal_signal_pending(current))
1983 mem_cgroup_event(mem_over_limit, MEMCG_OOM);
1985 mem_cgroup_oom(mem_over_limit, gfp_mask,
1986 get_order(nr_pages * PAGE_SIZE));
1988 if (!(gfp_mask & __GFP_NOFAIL))
1992 * The allocation either can't fail or will lead to more memory
1993 * being freed very soon. Allow memory usage go over the limit
1994 * temporarily by force charging it.
1996 page_counter_charge(&memcg->memory, nr_pages);
1997 if (do_memsw_account())
1998 page_counter_charge(&memcg->memsw, nr_pages);
1999 css_get_many(&memcg->css, nr_pages);
2004 css_get_many(&memcg->css, batch);
2005 if (batch > nr_pages)
2006 refill_stock(memcg, batch - nr_pages);
2009 * If the hierarchy is above the normal consumption range, schedule
2010 * reclaim on returning to userland. We can perform reclaim here
2011 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2012 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2013 * not recorded as it most likely matches current's and won't
2014 * change in the meantime. As high limit is checked again before
2015 * reclaim, the cost of mismatch is negligible.
2018 if (page_counter_read(&memcg->memory) > memcg->high) {
2019 /* Don't bother a random interrupted task */
2020 if (in_interrupt()) {
2021 schedule_work(&memcg->high_work);
2024 current->memcg_nr_pages_over_high += batch;
2025 set_notify_resume(current);
2028 } while ((memcg = parent_mem_cgroup(memcg)));
2033 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2035 if (mem_cgroup_is_root(memcg))
2038 page_counter_uncharge(&memcg->memory, nr_pages);
2039 if (do_memsw_account())
2040 page_counter_uncharge(&memcg->memsw, nr_pages);
2042 css_put_many(&memcg->css, nr_pages);
2045 static void lock_page_lru(struct page *page, int *isolated)
2047 struct zone *zone = page_zone(page);
2049 spin_lock_irq(zone_lru_lock(zone));
2050 if (PageLRU(page)) {
2051 struct lruvec *lruvec;
2053 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2055 del_page_from_lru_list(page, lruvec, page_lru(page));
2061 static void unlock_page_lru(struct page *page, int isolated)
2063 struct zone *zone = page_zone(page);
2066 struct lruvec *lruvec;
2068 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
2069 VM_BUG_ON_PAGE(PageLRU(page), page);
2071 add_page_to_lru_list(page, lruvec, page_lru(page));
2073 spin_unlock_irq(zone_lru_lock(zone));
2076 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2081 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2084 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2085 * may already be on some other mem_cgroup's LRU. Take care of it.
2088 lock_page_lru(page, &isolated);
2091 * Nobody should be changing or seriously looking at
2092 * page->mem_cgroup at this point:
2094 * - the page is uncharged
2096 * - the page is off-LRU
2098 * - an anonymous fault has exclusive page access, except for
2099 * a locked page table
2101 * - a page cache insertion, a swapin fault, or a migration
2102 * have the page locked
2104 page->mem_cgroup = memcg;
2107 unlock_page_lru(page, isolated);
2111 static int memcg_alloc_cache_id(void)
2116 id = ida_simple_get(&memcg_cache_ida,
2117 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2121 if (id < memcg_nr_cache_ids)
2125 * There's no space for the new id in memcg_caches arrays,
2126 * so we have to grow them.
2128 down_write(&memcg_cache_ids_sem);
2130 size = 2 * (id + 1);
2131 if (size < MEMCG_CACHES_MIN_SIZE)
2132 size = MEMCG_CACHES_MIN_SIZE;
2133 else if (size > MEMCG_CACHES_MAX_SIZE)
2134 size = MEMCG_CACHES_MAX_SIZE;
2136 err = memcg_update_all_caches(size);
2138 err = memcg_update_all_list_lrus(size);
2140 memcg_nr_cache_ids = size;
2142 up_write(&memcg_cache_ids_sem);
2145 ida_simple_remove(&memcg_cache_ida, id);
2151 static void memcg_free_cache_id(int id)
2153 ida_simple_remove(&memcg_cache_ida, id);
2156 struct memcg_kmem_cache_create_work {
2157 struct mem_cgroup *memcg;
2158 struct kmem_cache *cachep;
2159 struct work_struct work;
2162 static void memcg_kmem_cache_create_func(struct work_struct *w)
2164 struct memcg_kmem_cache_create_work *cw =
2165 container_of(w, struct memcg_kmem_cache_create_work, work);
2166 struct mem_cgroup *memcg = cw->memcg;
2167 struct kmem_cache *cachep = cw->cachep;
2169 memcg_create_kmem_cache(memcg, cachep);
2171 css_put(&memcg->css);
2176 * Enqueue the creation of a per-memcg kmem_cache.
2178 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2179 struct kmem_cache *cachep)
2181 struct memcg_kmem_cache_create_work *cw;
2183 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2187 css_get(&memcg->css);
2190 cw->cachep = cachep;
2191 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2193 queue_work(memcg_kmem_cache_wq, &cw->work);
2196 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2197 struct kmem_cache *cachep)
2200 * We need to stop accounting when we kmalloc, because if the
2201 * corresponding kmalloc cache is not yet created, the first allocation
2202 * in __memcg_schedule_kmem_cache_create will recurse.
2204 * However, it is better to enclose the whole function. Depending on
2205 * the debugging options enabled, INIT_WORK(), for instance, can
2206 * trigger an allocation. This too, will make us recurse. Because at
2207 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2208 * the safest choice is to do it like this, wrapping the whole function.
2210 current->memcg_kmem_skip_account = 1;
2211 __memcg_schedule_kmem_cache_create(memcg, cachep);
2212 current->memcg_kmem_skip_account = 0;
2215 static inline bool memcg_kmem_bypass(void)
2217 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2223 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2224 * @cachep: the original global kmem cache
2226 * Return the kmem_cache we're supposed to use for a slab allocation.
2227 * We try to use the current memcg's version of the cache.
2229 * If the cache does not exist yet, if we are the first user of it, we
2230 * create it asynchronously in a workqueue and let the current allocation
2231 * go through with the original cache.
2233 * This function takes a reference to the cache it returns to assure it
2234 * won't get destroyed while we are working with it. Once the caller is
2235 * done with it, memcg_kmem_put_cache() must be called to release the
2238 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2240 struct mem_cgroup *memcg;
2241 struct kmem_cache *memcg_cachep;
2244 VM_BUG_ON(!is_root_cache(cachep));
2246 if (memcg_kmem_bypass())
2249 if (current->memcg_kmem_skip_account)
2252 memcg = get_mem_cgroup_from_mm(current->mm);
2253 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2257 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2258 if (likely(memcg_cachep))
2259 return memcg_cachep;
2262 * If we are in a safe context (can wait, and not in interrupt
2263 * context), we could be be predictable and return right away.
2264 * This would guarantee that the allocation being performed
2265 * already belongs in the new cache.
2267 * However, there are some clashes that can arrive from locking.
2268 * For instance, because we acquire the slab_mutex while doing
2269 * memcg_create_kmem_cache, this means no further allocation
2270 * could happen with the slab_mutex held. So it's better to
2273 memcg_schedule_kmem_cache_create(memcg, cachep);
2275 css_put(&memcg->css);
2280 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2281 * @cachep: the cache returned by memcg_kmem_get_cache
2283 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2285 if (!is_root_cache(cachep))
2286 css_put(&cachep->memcg_params.memcg->css);
2290 * memcg_kmem_charge: charge a kmem page
2291 * @page: page to charge
2292 * @gfp: reclaim mode
2293 * @order: allocation order
2294 * @memcg: memory cgroup to charge
2296 * Returns 0 on success, an error code on failure.
2298 int memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2299 struct mem_cgroup *memcg)
2301 unsigned int nr_pages = 1 << order;
2302 struct page_counter *counter;
2305 ret = try_charge(memcg, gfp, nr_pages);
2309 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2310 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2311 cancel_charge(memcg, nr_pages);
2315 page->mem_cgroup = memcg;
2321 * memcg_kmem_charge: charge a kmem page to the current memory cgroup
2322 * @page: page to charge
2323 * @gfp: reclaim mode
2324 * @order: allocation order
2326 * Returns 0 on success, an error code on failure.
2328 int memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2330 struct mem_cgroup *memcg;
2333 if (memcg_kmem_bypass())
2336 memcg = get_mem_cgroup_from_mm(current->mm);
2337 if (!mem_cgroup_is_root(memcg)) {
2338 ret = memcg_kmem_charge_memcg(page, gfp, order, memcg);
2340 __SetPageKmemcg(page);
2342 css_put(&memcg->css);
2346 * memcg_kmem_uncharge: uncharge a kmem page
2347 * @page: page to uncharge
2348 * @order: allocation order
2350 void memcg_kmem_uncharge(struct page *page, int order)
2352 struct mem_cgroup *memcg = page->mem_cgroup;
2353 unsigned int nr_pages = 1 << order;
2358 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2360 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2361 page_counter_uncharge(&memcg->kmem, nr_pages);
2363 page_counter_uncharge(&memcg->memory, nr_pages);
2364 if (do_memsw_account())
2365 page_counter_uncharge(&memcg->memsw, nr_pages);
2367 page->mem_cgroup = NULL;
2369 /* slab pages do not have PageKmemcg flag set */
2370 if (PageKmemcg(page))
2371 __ClearPageKmemcg(page);
2373 css_put_many(&memcg->css, nr_pages);
2375 #endif /* !CONFIG_SLOB */
2377 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2380 * Because tail pages are not marked as "used", set it. We're under
2381 * zone_lru_lock and migration entries setup in all page mappings.
2383 void mem_cgroup_split_huge_fixup(struct page *head)
2387 if (mem_cgroup_disabled())
2390 for (i = 1; i < HPAGE_PMD_NR; i++)
2391 head[i].mem_cgroup = head->mem_cgroup;
2393 __this_cpu_sub(head->mem_cgroup->stat->count[MEMCG_RSS_HUGE],
2396 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2398 #ifdef CONFIG_MEMCG_SWAP
2399 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2402 this_cpu_add(memcg->stat->count[MEMCG_SWAP], nr_entries);
2406 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2407 * @entry: swap entry to be moved
2408 * @from: mem_cgroup which the entry is moved from
2409 * @to: mem_cgroup which the entry is moved to
2411 * It succeeds only when the swap_cgroup's record for this entry is the same
2412 * as the mem_cgroup's id of @from.
2414 * Returns 0 on success, -EINVAL on failure.
2416 * The caller must have charged to @to, IOW, called page_counter_charge() about
2417 * both res and memsw, and called css_get().
2419 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2420 struct mem_cgroup *from, struct mem_cgroup *to)
2422 unsigned short old_id, new_id;
2424 old_id = mem_cgroup_id(from);
2425 new_id = mem_cgroup_id(to);
2427 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2428 mem_cgroup_swap_statistics(from, -1);
2429 mem_cgroup_swap_statistics(to, 1);
2435 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2436 struct mem_cgroup *from, struct mem_cgroup *to)
2442 static DEFINE_MUTEX(memcg_limit_mutex);
2444 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2445 unsigned long limit)
2447 unsigned long curusage;
2448 unsigned long oldusage;
2449 bool enlarge = false;
2454 * For keeping hierarchical_reclaim simple, how long we should retry
2455 * is depends on callers. We set our retry-count to be function
2456 * of # of children which we should visit in this loop.
2458 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2459 mem_cgroup_count_children(memcg);
2461 oldusage = page_counter_read(&memcg->memory);
2464 if (signal_pending(current)) {
2469 mutex_lock(&memcg_limit_mutex);
2470 if (limit > memcg->memsw.limit) {
2471 mutex_unlock(&memcg_limit_mutex);
2475 if (limit > memcg->memory.limit)
2477 ret = page_counter_limit(&memcg->memory, limit);
2478 mutex_unlock(&memcg_limit_mutex);
2483 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2485 curusage = page_counter_read(&memcg->memory);
2486 /* Usage is reduced ? */
2487 if (curusage >= oldusage)
2490 oldusage = curusage;
2491 } while (retry_count);
2493 if (!ret && enlarge)
2494 memcg_oom_recover(memcg);
2499 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2500 unsigned long limit)
2502 unsigned long curusage;
2503 unsigned long oldusage;
2504 bool enlarge = false;
2508 /* see mem_cgroup_resize_res_limit */
2509 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2510 mem_cgroup_count_children(memcg);
2512 oldusage = page_counter_read(&memcg->memsw);
2515 if (signal_pending(current)) {
2520 mutex_lock(&memcg_limit_mutex);
2521 if (limit < memcg->memory.limit) {
2522 mutex_unlock(&memcg_limit_mutex);
2526 if (limit > memcg->memsw.limit)
2528 ret = page_counter_limit(&memcg->memsw, limit);
2529 mutex_unlock(&memcg_limit_mutex);
2534 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2536 curusage = page_counter_read(&memcg->memsw);
2537 /* Usage is reduced ? */
2538 if (curusage >= oldusage)
2541 oldusage = curusage;
2542 } while (retry_count);
2544 if (!ret && enlarge)
2545 memcg_oom_recover(memcg);
2550 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
2552 unsigned long *total_scanned)
2554 unsigned long nr_reclaimed = 0;
2555 struct mem_cgroup_per_node *mz, *next_mz = NULL;
2556 unsigned long reclaimed;
2558 struct mem_cgroup_tree_per_node *mctz;
2559 unsigned long excess;
2560 unsigned long nr_scanned;
2565 mctz = soft_limit_tree_node(pgdat->node_id);
2568 * Do not even bother to check the largest node if the root
2569 * is empty. Do it lockless to prevent lock bouncing. Races
2570 * are acceptable as soft limit is best effort anyway.
2572 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
2576 * This loop can run a while, specially if mem_cgroup's continuously
2577 * keep exceeding their soft limit and putting the system under
2584 mz = mem_cgroup_largest_soft_limit_node(mctz);
2589 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
2590 gfp_mask, &nr_scanned);
2591 nr_reclaimed += reclaimed;
2592 *total_scanned += nr_scanned;
2593 spin_lock_irq(&mctz->lock);
2594 __mem_cgroup_remove_exceeded(mz, mctz);
2597 * If we failed to reclaim anything from this memory cgroup
2598 * it is time to move on to the next cgroup
2602 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2604 excess = soft_limit_excess(mz->memcg);
2606 * One school of thought says that we should not add
2607 * back the node to the tree if reclaim returns 0.
2608 * But our reclaim could return 0, simply because due
2609 * to priority we are exposing a smaller subset of
2610 * memory to reclaim from. Consider this as a longer
2613 /* If excess == 0, no tree ops */
2614 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2615 spin_unlock_irq(&mctz->lock);
2616 css_put(&mz->memcg->css);
2619 * Could not reclaim anything and there are no more
2620 * mem cgroups to try or we seem to be looping without
2621 * reclaiming anything.
2623 if (!nr_reclaimed &&
2625 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2627 } while (!nr_reclaimed);
2629 css_put(&next_mz->memcg->css);
2630 return nr_reclaimed;
2634 * Test whether @memcg has children, dead or alive. Note that this
2635 * function doesn't care whether @memcg has use_hierarchy enabled and
2636 * returns %true if there are child csses according to the cgroup
2637 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
2639 static inline bool memcg_has_children(struct mem_cgroup *memcg)
2644 ret = css_next_child(NULL, &memcg->css);
2650 * Reclaims as many pages from the given memcg as possible.
2652 * Caller is responsible for holding css reference for memcg.
2654 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
2656 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2658 /* we call try-to-free pages for make this cgroup empty */
2659 lru_add_drain_all();
2660 /* try to free all pages in this cgroup */
2661 while (nr_retries && page_counter_read(&memcg->memory)) {
2664 if (signal_pending(current))
2667 progress = try_to_free_mem_cgroup_pages(memcg, 1,
2671 /* maybe some writeback is necessary */
2672 congestion_wait(BLK_RW_ASYNC, HZ/10);
2680 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
2681 char *buf, size_t nbytes,
2684 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2686 if (mem_cgroup_is_root(memcg))
2688 return mem_cgroup_force_empty(memcg) ?: nbytes;
2691 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
2694 return mem_cgroup_from_css(css)->use_hierarchy;
2697 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
2698 struct cftype *cft, u64 val)
2701 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2702 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
2704 if (memcg->use_hierarchy == val)
2708 * If parent's use_hierarchy is set, we can't make any modifications
2709 * in the child subtrees. If it is unset, then the change can
2710 * occur, provided the current cgroup has no children.
2712 * For the root cgroup, parent_mem is NULL, we allow value to be
2713 * set if there are no children.
2715 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
2716 (val == 1 || val == 0)) {
2717 if (!memcg_has_children(memcg))
2718 memcg->use_hierarchy = val;
2727 static void tree_stat(struct mem_cgroup *memcg, unsigned long *stat)
2729 struct mem_cgroup *iter;
2732 memset(stat, 0, sizeof(*stat) * MEMCG_NR_STAT);
2734 for_each_mem_cgroup_tree(iter, memcg) {
2735 for (i = 0; i < MEMCG_NR_STAT; i++)
2736 stat[i] += memcg_page_state(iter, i);
2740 static void tree_events(struct mem_cgroup *memcg, unsigned long *events)
2742 struct mem_cgroup *iter;
2745 memset(events, 0, sizeof(*events) * MEMCG_NR_EVENTS);
2747 for_each_mem_cgroup_tree(iter, memcg) {
2748 for (i = 0; i < MEMCG_NR_EVENTS; i++)
2749 events[i] += memcg_sum_events(iter, i);
2753 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
2755 unsigned long val = 0;
2757 if (mem_cgroup_is_root(memcg)) {
2758 struct mem_cgroup *iter;
2760 for_each_mem_cgroup_tree(iter, memcg) {
2761 val += memcg_page_state(iter, MEMCG_CACHE);
2762 val += memcg_page_state(iter, MEMCG_RSS);
2764 val += memcg_page_state(iter, MEMCG_SWAP);
2768 val = page_counter_read(&memcg->memory);
2770 val = page_counter_read(&memcg->memsw);
2783 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
2786 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2787 struct page_counter *counter;
2789 switch (MEMFILE_TYPE(cft->private)) {
2791 counter = &memcg->memory;
2794 counter = &memcg->memsw;
2797 counter = &memcg->kmem;
2800 counter = &memcg->tcpmem;
2806 switch (MEMFILE_ATTR(cft->private)) {
2808 if (counter == &memcg->memory)
2809 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
2810 if (counter == &memcg->memsw)
2811 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
2812 return (u64)page_counter_read(counter) * PAGE_SIZE;
2814 return (u64)counter->limit * PAGE_SIZE;
2816 return (u64)counter->watermark * PAGE_SIZE;
2818 return counter->failcnt;
2819 case RES_SOFT_LIMIT:
2820 return (u64)memcg->soft_limit * PAGE_SIZE;
2827 static int memcg_online_kmem(struct mem_cgroup *memcg)
2831 if (cgroup_memory_nokmem)
2834 BUG_ON(memcg->kmemcg_id >= 0);
2835 BUG_ON(memcg->kmem_state);
2837 memcg_id = memcg_alloc_cache_id();
2841 static_branch_inc(&memcg_kmem_enabled_key);
2843 * A memory cgroup is considered kmem-online as soon as it gets
2844 * kmemcg_id. Setting the id after enabling static branching will
2845 * guarantee no one starts accounting before all call sites are
2848 memcg->kmemcg_id = memcg_id;
2849 memcg->kmem_state = KMEM_ONLINE;
2850 INIT_LIST_HEAD(&memcg->kmem_caches);
2855 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2857 struct cgroup_subsys_state *css;
2858 struct mem_cgroup *parent, *child;
2861 if (memcg->kmem_state != KMEM_ONLINE)
2864 * Clear the online state before clearing memcg_caches array
2865 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
2866 * guarantees that no cache will be created for this cgroup
2867 * after we are done (see memcg_create_kmem_cache()).
2869 memcg->kmem_state = KMEM_ALLOCATED;
2871 memcg_deactivate_kmem_caches(memcg);
2873 kmemcg_id = memcg->kmemcg_id;
2874 BUG_ON(kmemcg_id < 0);
2876 parent = parent_mem_cgroup(memcg);
2878 parent = root_mem_cgroup;
2881 * Change kmemcg_id of this cgroup and all its descendants to the
2882 * parent's id, and then move all entries from this cgroup's list_lrus
2883 * to ones of the parent. After we have finished, all list_lrus
2884 * corresponding to this cgroup are guaranteed to remain empty. The
2885 * ordering is imposed by list_lru_node->lock taken by
2886 * memcg_drain_all_list_lrus().
2888 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
2889 css_for_each_descendant_pre(css, &memcg->css) {
2890 child = mem_cgroup_from_css(css);
2891 BUG_ON(child->kmemcg_id != kmemcg_id);
2892 child->kmemcg_id = parent->kmemcg_id;
2893 if (!memcg->use_hierarchy)
2898 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
2900 memcg_free_cache_id(kmemcg_id);
2903 static void memcg_free_kmem(struct mem_cgroup *memcg)
2905 /* css_alloc() failed, offlining didn't happen */
2906 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
2907 memcg_offline_kmem(memcg);
2909 if (memcg->kmem_state == KMEM_ALLOCATED) {
2910 memcg_destroy_kmem_caches(memcg);
2911 static_branch_dec(&memcg_kmem_enabled_key);
2912 WARN_ON(page_counter_read(&memcg->kmem));
2916 static int memcg_online_kmem(struct mem_cgroup *memcg)
2920 static void memcg_offline_kmem(struct mem_cgroup *memcg)
2923 static void memcg_free_kmem(struct mem_cgroup *memcg)
2926 #endif /* !CONFIG_SLOB */
2928 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
2929 unsigned long limit)
2933 mutex_lock(&memcg_limit_mutex);
2934 ret = page_counter_limit(&memcg->kmem, limit);
2935 mutex_unlock(&memcg_limit_mutex);
2939 static int memcg_update_tcp_limit(struct mem_cgroup *memcg, unsigned long limit)
2943 mutex_lock(&memcg_limit_mutex);
2945 ret = page_counter_limit(&memcg->tcpmem, limit);
2949 if (!memcg->tcpmem_active) {
2951 * The active flag needs to be written after the static_key
2952 * update. This is what guarantees that the socket activation
2953 * function is the last one to run. See mem_cgroup_sk_alloc()
2954 * for details, and note that we don't mark any socket as
2955 * belonging to this memcg until that flag is up.
2957 * We need to do this, because static_keys will span multiple
2958 * sites, but we can't control their order. If we mark a socket
2959 * as accounted, but the accounting functions are not patched in
2960 * yet, we'll lose accounting.
2962 * We never race with the readers in mem_cgroup_sk_alloc(),
2963 * because when this value change, the code to process it is not
2966 static_branch_inc(&memcg_sockets_enabled_key);
2967 memcg->tcpmem_active = true;
2970 mutex_unlock(&memcg_limit_mutex);
2975 * The user of this function is...
2978 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
2979 char *buf, size_t nbytes, loff_t off)
2981 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
2982 unsigned long nr_pages;
2985 buf = strstrip(buf);
2986 ret = page_counter_memparse(buf, "-1", &nr_pages);
2990 switch (MEMFILE_ATTR(of_cft(of)->private)) {
2992 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
2996 switch (MEMFILE_TYPE(of_cft(of)->private)) {
2998 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3001 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3004 ret = memcg_update_kmem_limit(memcg, nr_pages);
3007 ret = memcg_update_tcp_limit(memcg, nr_pages);
3011 case RES_SOFT_LIMIT:
3012 memcg->soft_limit = nr_pages;
3016 return ret ?: nbytes;
3019 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3020 size_t nbytes, loff_t off)
3022 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3023 struct page_counter *counter;
3025 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3027 counter = &memcg->memory;
3030 counter = &memcg->memsw;
3033 counter = &memcg->kmem;
3036 counter = &memcg->tcpmem;
3042 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3044 page_counter_reset_watermark(counter);
3047 counter->failcnt = 0;
3056 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3059 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3063 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3064 struct cftype *cft, u64 val)
3066 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3068 if (val & ~MOVE_MASK)
3072 * No kind of locking is needed in here, because ->can_attach() will
3073 * check this value once in the beginning of the process, and then carry
3074 * on with stale data. This means that changes to this value will only
3075 * affect task migrations starting after the change.
3077 memcg->move_charge_at_immigrate = val;
3081 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3082 struct cftype *cft, u64 val)
3089 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3093 unsigned int lru_mask;
3096 static const struct numa_stat stats[] = {
3097 { "total", LRU_ALL },
3098 { "file", LRU_ALL_FILE },
3099 { "anon", LRU_ALL_ANON },
3100 { "unevictable", BIT(LRU_UNEVICTABLE) },
3102 const struct numa_stat *stat;
3105 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3107 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3108 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3109 seq_printf(m, "%s=%lu", stat->name, nr);
3110 for_each_node_state(nid, N_MEMORY) {
3111 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3113 seq_printf(m, " N%d=%lu", nid, nr);
3118 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3119 struct mem_cgroup *iter;
3122 for_each_mem_cgroup_tree(iter, memcg)
3123 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3124 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3125 for_each_node_state(nid, N_MEMORY) {
3127 for_each_mem_cgroup_tree(iter, memcg)
3128 nr += mem_cgroup_node_nr_lru_pages(
3129 iter, nid, stat->lru_mask);
3130 seq_printf(m, " N%d=%lu", nid, nr);
3137 #endif /* CONFIG_NUMA */
3139 /* Universal VM events cgroup1 shows, original sort order */
3140 unsigned int memcg1_events[] = {
3147 static const char *const memcg1_event_names[] = {
3154 static int memcg_stat_show(struct seq_file *m, void *v)
3156 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3157 unsigned long memory, memsw;
3158 struct mem_cgroup *mi;
3161 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3162 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3164 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3165 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3167 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3168 memcg_page_state(memcg, memcg1_stats[i]) *
3172 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3173 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3174 memcg_sum_events(memcg, memcg1_events[i]));
3176 for (i = 0; i < NR_LRU_LISTS; i++)
3177 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3178 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3180 /* Hierarchical information */
3181 memory = memsw = PAGE_COUNTER_MAX;
3182 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3183 memory = min(memory, mi->memory.limit);
3184 memsw = min(memsw, mi->memsw.limit);
3186 seq_printf(m, "hierarchical_memory_limit %llu\n",
3187 (u64)memory * PAGE_SIZE);
3188 if (do_memsw_account())
3189 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3190 (u64)memsw * PAGE_SIZE);
3192 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3193 unsigned long long val = 0;
3195 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3197 for_each_mem_cgroup_tree(mi, memcg)
3198 val += memcg_page_state(mi, memcg1_stats[i]) *
3200 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i], val);
3203 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++) {
3204 unsigned long long val = 0;
3206 for_each_mem_cgroup_tree(mi, memcg)
3207 val += memcg_sum_events(mi, memcg1_events[i]);
3208 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i], val);
3211 for (i = 0; i < NR_LRU_LISTS; i++) {
3212 unsigned long long val = 0;
3214 for_each_mem_cgroup_tree(mi, memcg)
3215 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3216 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3219 #ifdef CONFIG_DEBUG_VM
3222 struct mem_cgroup_per_node *mz;
3223 struct zone_reclaim_stat *rstat;
3224 unsigned long recent_rotated[2] = {0, 0};
3225 unsigned long recent_scanned[2] = {0, 0};
3227 for_each_online_pgdat(pgdat) {
3228 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3229 rstat = &mz->lruvec.reclaim_stat;
3231 recent_rotated[0] += rstat->recent_rotated[0];
3232 recent_rotated[1] += rstat->recent_rotated[1];
3233 recent_scanned[0] += rstat->recent_scanned[0];
3234 recent_scanned[1] += rstat->recent_scanned[1];
3236 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3237 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3238 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3239 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3246 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3249 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3251 return mem_cgroup_swappiness(memcg);
3254 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3255 struct cftype *cft, u64 val)
3257 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3263 memcg->swappiness = val;
3265 vm_swappiness = val;
3270 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3272 struct mem_cgroup_threshold_ary *t;
3273 unsigned long usage;
3278 t = rcu_dereference(memcg->thresholds.primary);
3280 t = rcu_dereference(memcg->memsw_thresholds.primary);
3285 usage = mem_cgroup_usage(memcg, swap);
3288 * current_threshold points to threshold just below or equal to usage.
3289 * If it's not true, a threshold was crossed after last
3290 * call of __mem_cgroup_threshold().
3292 i = t->current_threshold;
3295 * Iterate backward over array of thresholds starting from
3296 * current_threshold and check if a threshold is crossed.
3297 * If none of thresholds below usage is crossed, we read
3298 * only one element of the array here.
3300 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3301 eventfd_signal(t->entries[i].eventfd, 1);
3303 /* i = current_threshold + 1 */
3307 * Iterate forward over array of thresholds starting from
3308 * current_threshold+1 and check if a threshold is crossed.
3309 * If none of thresholds above usage is crossed, we read
3310 * only one element of the array here.
3312 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3313 eventfd_signal(t->entries[i].eventfd, 1);
3315 /* Update current_threshold */
3316 t->current_threshold = i - 1;
3321 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3324 __mem_cgroup_threshold(memcg, false);
3325 if (do_memsw_account())
3326 __mem_cgroup_threshold(memcg, true);
3328 memcg = parent_mem_cgroup(memcg);
3332 static int compare_thresholds(const void *a, const void *b)
3334 const struct mem_cgroup_threshold *_a = a;
3335 const struct mem_cgroup_threshold *_b = b;
3337 if (_a->threshold > _b->threshold)
3340 if (_a->threshold < _b->threshold)
3346 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3348 struct mem_cgroup_eventfd_list *ev;
3350 spin_lock(&memcg_oom_lock);
3352 list_for_each_entry(ev, &memcg->oom_notify, list)
3353 eventfd_signal(ev->eventfd, 1);
3355 spin_unlock(&memcg_oom_lock);
3359 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3361 struct mem_cgroup *iter;
3363 for_each_mem_cgroup_tree(iter, memcg)
3364 mem_cgroup_oom_notify_cb(iter);
3367 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3368 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3370 struct mem_cgroup_thresholds *thresholds;
3371 struct mem_cgroup_threshold_ary *new;
3372 unsigned long threshold;
3373 unsigned long usage;
3376 ret = page_counter_memparse(args, "-1", &threshold);
3380 mutex_lock(&memcg->thresholds_lock);
3383 thresholds = &memcg->thresholds;
3384 usage = mem_cgroup_usage(memcg, false);
3385 } else if (type == _MEMSWAP) {
3386 thresholds = &memcg->memsw_thresholds;
3387 usage = mem_cgroup_usage(memcg, true);
3391 /* Check if a threshold crossed before adding a new one */
3392 if (thresholds->primary)
3393 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3395 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3397 /* Allocate memory for new array of thresholds */
3398 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3406 /* Copy thresholds (if any) to new array */
3407 if (thresholds->primary) {
3408 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3409 sizeof(struct mem_cgroup_threshold));
3412 /* Add new threshold */
3413 new->entries[size - 1].eventfd = eventfd;
3414 new->entries[size - 1].threshold = threshold;
3416 /* Sort thresholds. Registering of new threshold isn't time-critical */
3417 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3418 compare_thresholds, NULL);
3420 /* Find current threshold */
3421 new->current_threshold = -1;
3422 for (i = 0; i < size; i++) {
3423 if (new->entries[i].threshold <= usage) {
3425 * new->current_threshold will not be used until
3426 * rcu_assign_pointer(), so it's safe to increment
3429 ++new->current_threshold;
3434 /* Free old spare buffer and save old primary buffer as spare */
3435 kfree(thresholds->spare);
3436 thresholds->spare = thresholds->primary;
3438 rcu_assign_pointer(thresholds->primary, new);
3440 /* To be sure that nobody uses thresholds */
3444 mutex_unlock(&memcg->thresholds_lock);
3449 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3450 struct eventfd_ctx *eventfd, const char *args)
3452 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3455 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3456 struct eventfd_ctx *eventfd, const char *args)
3458 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3461 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3462 struct eventfd_ctx *eventfd, enum res_type type)
3464 struct mem_cgroup_thresholds *thresholds;
3465 struct mem_cgroup_threshold_ary *new;
3466 unsigned long usage;
3469 mutex_lock(&memcg->thresholds_lock);
3472 thresholds = &memcg->thresholds;
3473 usage = mem_cgroup_usage(memcg, false);
3474 } else if (type == _MEMSWAP) {
3475 thresholds = &memcg->memsw_thresholds;
3476 usage = mem_cgroup_usage(memcg, true);
3480 if (!thresholds->primary)
3483 /* Check if a threshold crossed before removing */
3484 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3486 /* Calculate new number of threshold */
3488 for (i = 0; i < thresholds->primary->size; i++) {
3489 if (thresholds->primary->entries[i].eventfd != eventfd)
3493 new = thresholds->spare;
3495 /* Set thresholds array to NULL if we don't have thresholds */
3504 /* Copy thresholds and find current threshold */
3505 new->current_threshold = -1;
3506 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3507 if (thresholds->primary->entries[i].eventfd == eventfd)
3510 new->entries[j] = thresholds->primary->entries[i];
3511 if (new->entries[j].threshold <= usage) {
3513 * new->current_threshold will not be used
3514 * until rcu_assign_pointer(), so it's safe to increment
3517 ++new->current_threshold;
3523 /* Swap primary and spare array */
3524 thresholds->spare = thresholds->primary;
3526 rcu_assign_pointer(thresholds->primary, new);
3528 /* To be sure that nobody uses thresholds */
3531 /* If all events are unregistered, free the spare array */
3533 kfree(thresholds->spare);
3534 thresholds->spare = NULL;
3537 mutex_unlock(&memcg->thresholds_lock);
3540 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3541 struct eventfd_ctx *eventfd)
3543 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3546 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3547 struct eventfd_ctx *eventfd)
3549 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3552 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3553 struct eventfd_ctx *eventfd, const char *args)
3555 struct mem_cgroup_eventfd_list *event;
3557 event = kmalloc(sizeof(*event), GFP_KERNEL);
3561 spin_lock(&memcg_oom_lock);
3563 event->eventfd = eventfd;
3564 list_add(&event->list, &memcg->oom_notify);
3566 /* already in OOM ? */
3567 if (memcg->under_oom)
3568 eventfd_signal(eventfd, 1);
3569 spin_unlock(&memcg_oom_lock);
3574 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3575 struct eventfd_ctx *eventfd)
3577 struct mem_cgroup_eventfd_list *ev, *tmp;
3579 spin_lock(&memcg_oom_lock);
3581 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3582 if (ev->eventfd == eventfd) {
3583 list_del(&ev->list);
3588 spin_unlock(&memcg_oom_lock);
3591 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3593 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3595 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3596 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
3597 seq_printf(sf, "oom_kill %lu\n", memcg_sum_events(memcg, OOM_KILL));
3601 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3602 struct cftype *cft, u64 val)
3604 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3606 /* cannot set to root cgroup and only 0 and 1 are allowed */
3607 if (!css->parent || !((val == 0) || (val == 1)))
3610 memcg->oom_kill_disable = val;
3612 memcg_oom_recover(memcg);
3617 #ifdef CONFIG_CGROUP_WRITEBACK
3619 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3621 return &memcg->cgwb_list;
3624 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3626 return wb_domain_init(&memcg->cgwb_domain, gfp);
3629 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3631 wb_domain_exit(&memcg->cgwb_domain);
3634 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3636 wb_domain_size_changed(&memcg->cgwb_domain);
3639 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
3641 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3643 if (!memcg->css.parent)
3646 return &memcg->cgwb_domain;
3650 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
3651 * @wb: bdi_writeback in question
3652 * @pfilepages: out parameter for number of file pages
3653 * @pheadroom: out parameter for number of allocatable pages according to memcg
3654 * @pdirty: out parameter for number of dirty pages
3655 * @pwriteback: out parameter for number of pages under writeback
3657 * Determine the numbers of file, headroom, dirty, and writeback pages in
3658 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
3659 * is a bit more involved.
3661 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
3662 * headroom is calculated as the lowest headroom of itself and the
3663 * ancestors. Note that this doesn't consider the actual amount of
3664 * available memory in the system. The caller should further cap
3665 * *@pheadroom accordingly.
3667 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
3668 unsigned long *pheadroom, unsigned long *pdirty,
3669 unsigned long *pwriteback)
3671 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
3672 struct mem_cgroup *parent;
3674 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
3676 /* this should eventually include NR_UNSTABLE_NFS */
3677 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
3678 *pfilepages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
3679 (1 << LRU_ACTIVE_FILE));
3680 *pheadroom = PAGE_COUNTER_MAX;
3682 while ((parent = parent_mem_cgroup(memcg))) {
3683 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
3684 unsigned long used = page_counter_read(&memcg->memory);
3686 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
3691 #else /* CONFIG_CGROUP_WRITEBACK */
3693 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
3698 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
3702 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
3706 #endif /* CONFIG_CGROUP_WRITEBACK */
3709 * DO NOT USE IN NEW FILES.
3711 * "cgroup.event_control" implementation.
3713 * This is way over-engineered. It tries to support fully configurable
3714 * events for each user. Such level of flexibility is completely
3715 * unnecessary especially in the light of the planned unified hierarchy.
3717 * Please deprecate this and replace with something simpler if at all
3722 * Unregister event and free resources.
3724 * Gets called from workqueue.
3726 static void memcg_event_remove(struct work_struct *work)
3728 struct mem_cgroup_event *event =
3729 container_of(work, struct mem_cgroup_event, remove);
3730 struct mem_cgroup *memcg = event->memcg;
3732 remove_wait_queue(event->wqh, &event->wait);
3734 event->unregister_event(memcg, event->eventfd);
3736 /* Notify userspace the event is going away. */
3737 eventfd_signal(event->eventfd, 1);
3739 eventfd_ctx_put(event->eventfd);
3741 css_put(&memcg->css);
3745 * Gets called on POLLHUP on eventfd when user closes it.
3747 * Called with wqh->lock held and interrupts disabled.
3749 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
3750 int sync, void *key)
3752 struct mem_cgroup_event *event =
3753 container_of(wait, struct mem_cgroup_event, wait);
3754 struct mem_cgroup *memcg = event->memcg;
3755 unsigned long flags = (unsigned long)key;
3757 if (flags & POLLHUP) {
3759 * If the event has been detached at cgroup removal, we
3760 * can simply return knowing the other side will cleanup
3763 * We can't race against event freeing since the other
3764 * side will require wqh->lock via remove_wait_queue(),
3767 spin_lock(&memcg->event_list_lock);
3768 if (!list_empty(&event->list)) {
3769 list_del_init(&event->list);
3771 * We are in atomic context, but cgroup_event_remove()
3772 * may sleep, so we have to call it in workqueue.
3774 schedule_work(&event->remove);
3776 spin_unlock(&memcg->event_list_lock);
3782 static void memcg_event_ptable_queue_proc(struct file *file,
3783 wait_queue_head_t *wqh, poll_table *pt)
3785 struct mem_cgroup_event *event =
3786 container_of(pt, struct mem_cgroup_event, pt);
3789 add_wait_queue(wqh, &event->wait);
3793 * DO NOT USE IN NEW FILES.
3795 * Parse input and register new cgroup event handler.
3797 * Input must be in format '<event_fd> <control_fd> <args>'.
3798 * Interpretation of args is defined by control file implementation.
3800 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
3801 char *buf, size_t nbytes, loff_t off)
3803 struct cgroup_subsys_state *css = of_css(of);
3804 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3805 struct mem_cgroup_event *event;
3806 struct cgroup_subsys_state *cfile_css;
3807 unsigned int efd, cfd;
3814 buf = strstrip(buf);
3816 efd = simple_strtoul(buf, &endp, 10);
3821 cfd = simple_strtoul(buf, &endp, 10);
3822 if ((*endp != ' ') && (*endp != '\0'))
3826 event = kzalloc(sizeof(*event), GFP_KERNEL);
3830 event->memcg = memcg;
3831 INIT_LIST_HEAD(&event->list);
3832 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
3833 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
3834 INIT_WORK(&event->remove, memcg_event_remove);
3842 event->eventfd = eventfd_ctx_fileget(efile.file);
3843 if (IS_ERR(event->eventfd)) {
3844 ret = PTR_ERR(event->eventfd);
3851 goto out_put_eventfd;
3854 /* the process need read permission on control file */
3855 /* AV: shouldn't we check that it's been opened for read instead? */
3856 ret = inode_permission(file_inode(cfile.file), MAY_READ);
3861 * Determine the event callbacks and set them in @event. This used
3862 * to be done via struct cftype but cgroup core no longer knows
3863 * about these events. The following is crude but the whole thing
3864 * is for compatibility anyway.
3866 * DO NOT ADD NEW FILES.
3868 name = cfile.file->f_path.dentry->d_name.name;
3870 if (!strcmp(name, "memory.usage_in_bytes")) {
3871 event->register_event = mem_cgroup_usage_register_event;
3872 event->unregister_event = mem_cgroup_usage_unregister_event;
3873 } else if (!strcmp(name, "memory.oom_control")) {
3874 event->register_event = mem_cgroup_oom_register_event;
3875 event->unregister_event = mem_cgroup_oom_unregister_event;
3876 } else if (!strcmp(name, "memory.pressure_level")) {
3877 event->register_event = vmpressure_register_event;
3878 event->unregister_event = vmpressure_unregister_event;
3879 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
3880 event->register_event = memsw_cgroup_usage_register_event;
3881 event->unregister_event = memsw_cgroup_usage_unregister_event;
3888 * Verify @cfile should belong to @css. Also, remaining events are
3889 * automatically removed on cgroup destruction but the removal is
3890 * asynchronous, so take an extra ref on @css.
3892 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
3893 &memory_cgrp_subsys);
3895 if (IS_ERR(cfile_css))
3897 if (cfile_css != css) {
3902 ret = event->register_event(memcg, event->eventfd, buf);
3906 efile.file->f_op->poll(efile.file, &event->pt);
3908 spin_lock(&memcg->event_list_lock);
3909 list_add(&event->list, &memcg->event_list);
3910 spin_unlock(&memcg->event_list_lock);
3922 eventfd_ctx_put(event->eventfd);
3931 static struct cftype mem_cgroup_legacy_files[] = {
3933 .name = "usage_in_bytes",
3934 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3935 .read_u64 = mem_cgroup_read_u64,
3938 .name = "max_usage_in_bytes",
3939 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3940 .write = mem_cgroup_reset,
3941 .read_u64 = mem_cgroup_read_u64,
3944 .name = "limit_in_bytes",
3945 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3946 .write = mem_cgroup_write,
3947 .read_u64 = mem_cgroup_read_u64,
3950 .name = "soft_limit_in_bytes",
3951 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3952 .write = mem_cgroup_write,
3953 .read_u64 = mem_cgroup_read_u64,
3957 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3958 .write = mem_cgroup_reset,
3959 .read_u64 = mem_cgroup_read_u64,
3963 .seq_show = memcg_stat_show,
3966 .name = "force_empty",
3967 .write = mem_cgroup_force_empty_write,
3970 .name = "use_hierarchy",
3971 .write_u64 = mem_cgroup_hierarchy_write,
3972 .read_u64 = mem_cgroup_hierarchy_read,
3975 .name = "cgroup.event_control", /* XXX: for compat */
3976 .write = memcg_write_event_control,
3977 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
3980 .name = "swappiness",
3981 .read_u64 = mem_cgroup_swappiness_read,
3982 .write_u64 = mem_cgroup_swappiness_write,
3985 .name = "move_charge_at_immigrate",
3986 .read_u64 = mem_cgroup_move_charge_read,
3987 .write_u64 = mem_cgroup_move_charge_write,
3990 .name = "oom_control",
3991 .seq_show = mem_cgroup_oom_control_read,
3992 .write_u64 = mem_cgroup_oom_control_write,
3993 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3996 .name = "pressure_level",
4000 .name = "numa_stat",
4001 .seq_show = memcg_numa_stat_show,
4005 .name = "kmem.limit_in_bytes",
4006 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4007 .write = mem_cgroup_write,
4008 .read_u64 = mem_cgroup_read_u64,
4011 .name = "kmem.usage_in_bytes",
4012 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4013 .read_u64 = mem_cgroup_read_u64,
4016 .name = "kmem.failcnt",
4017 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4018 .write = mem_cgroup_reset,
4019 .read_u64 = mem_cgroup_read_u64,
4022 .name = "kmem.max_usage_in_bytes",
4023 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4024 .write = mem_cgroup_reset,
4025 .read_u64 = mem_cgroup_read_u64,
4027 #ifdef CONFIG_SLABINFO
4029 .name = "kmem.slabinfo",
4030 .seq_start = memcg_slab_start,
4031 .seq_next = memcg_slab_next,
4032 .seq_stop = memcg_slab_stop,
4033 .seq_show = memcg_slab_show,
4037 .name = "kmem.tcp.limit_in_bytes",
4038 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4039 .write = mem_cgroup_write,
4040 .read_u64 = mem_cgroup_read_u64,
4043 .name = "kmem.tcp.usage_in_bytes",
4044 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4045 .read_u64 = mem_cgroup_read_u64,
4048 .name = "kmem.tcp.failcnt",
4049 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4050 .write = mem_cgroup_reset,
4051 .read_u64 = mem_cgroup_read_u64,
4054 .name = "kmem.tcp.max_usage_in_bytes",
4055 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4056 .write = mem_cgroup_reset,
4057 .read_u64 = mem_cgroup_read_u64,
4059 { }, /* terminate */
4063 * Private memory cgroup IDR
4065 * Swap-out records and page cache shadow entries need to store memcg
4066 * references in constrained space, so we maintain an ID space that is
4067 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4068 * memory-controlled cgroups to 64k.
4070 * However, there usually are many references to the oflline CSS after
4071 * the cgroup has been destroyed, such as page cache or reclaimable
4072 * slab objects, that don't need to hang on to the ID. We want to keep
4073 * those dead CSS from occupying IDs, or we might quickly exhaust the
4074 * relatively small ID space and prevent the creation of new cgroups
4075 * even when there are much fewer than 64k cgroups - possibly none.
4077 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4078 * be freed and recycled when it's no longer needed, which is usually
4079 * when the CSS is offlined.
4081 * The only exception to that are records of swapped out tmpfs/shmem
4082 * pages that need to be attributed to live ancestors on swapin. But
4083 * those references are manageable from userspace.
4086 static DEFINE_IDR(mem_cgroup_idr);
4088 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4090 VM_BUG_ON(atomic_read(&memcg->id.ref) <= 0);
4091 atomic_add(n, &memcg->id.ref);
4094 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4096 VM_BUG_ON(atomic_read(&memcg->id.ref) < n);
4097 if (atomic_sub_and_test(n, &memcg->id.ref)) {
4098 idr_remove(&mem_cgroup_idr, memcg->id.id);
4101 /* Memcg ID pins CSS */
4102 css_put(&memcg->css);
4106 static inline void mem_cgroup_id_get(struct mem_cgroup *memcg)
4108 mem_cgroup_id_get_many(memcg, 1);
4111 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4113 mem_cgroup_id_put_many(memcg, 1);
4117 * mem_cgroup_from_id - look up a memcg from a memcg id
4118 * @id: the memcg id to look up
4120 * Caller must hold rcu_read_lock().
4122 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4124 WARN_ON_ONCE(!rcu_read_lock_held());
4125 return idr_find(&mem_cgroup_idr, id);
4128 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4130 struct mem_cgroup_per_node *pn;
4133 * This routine is called against possible nodes.
4134 * But it's BUG to call kmalloc() against offline node.
4136 * TODO: this routine can waste much memory for nodes which will
4137 * never be onlined. It's better to use memory hotplug callback
4140 if (!node_state(node, N_NORMAL_MEMORY))
4142 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4146 pn->lruvec_stat = alloc_percpu(struct lruvec_stat);
4147 if (!pn->lruvec_stat) {
4152 lruvec_init(&pn->lruvec);
4153 pn->usage_in_excess = 0;
4154 pn->on_tree = false;
4157 memcg->nodeinfo[node] = pn;
4161 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4163 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4165 free_percpu(pn->lruvec_stat);
4169 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4174 free_mem_cgroup_per_node_info(memcg, node);
4175 free_percpu(memcg->stat);
4179 static void mem_cgroup_free(struct mem_cgroup *memcg)
4181 memcg_wb_domain_exit(memcg);
4182 __mem_cgroup_free(memcg);
4185 static struct mem_cgroup *mem_cgroup_alloc(void)
4187 struct mem_cgroup *memcg;
4191 size = sizeof(struct mem_cgroup);
4192 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4194 memcg = kzalloc(size, GFP_KERNEL);
4198 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4199 1, MEM_CGROUP_ID_MAX,
4201 if (memcg->id.id < 0)
4204 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4209 if (alloc_mem_cgroup_per_node_info(memcg, node))
4212 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4215 INIT_WORK(&memcg->high_work, high_work_func);
4216 memcg->last_scanned_node = MAX_NUMNODES;
4217 INIT_LIST_HEAD(&memcg->oom_notify);
4218 mutex_init(&memcg->thresholds_lock);
4219 spin_lock_init(&memcg->move_lock);
4220 vmpressure_init(&memcg->vmpressure);
4221 INIT_LIST_HEAD(&memcg->event_list);
4222 spin_lock_init(&memcg->event_list_lock);
4223 memcg->socket_pressure = jiffies;
4225 memcg->kmemcg_id = -1;
4227 #ifdef CONFIG_CGROUP_WRITEBACK
4228 INIT_LIST_HEAD(&memcg->cgwb_list);
4230 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4233 if (memcg->id.id > 0)
4234 idr_remove(&mem_cgroup_idr, memcg->id.id);
4235 __mem_cgroup_free(memcg);
4239 static struct cgroup_subsys_state * __ref
4240 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4242 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4243 struct mem_cgroup *memcg;
4244 long error = -ENOMEM;
4246 memcg = mem_cgroup_alloc();
4248 return ERR_PTR(error);
4250 memcg->high = PAGE_COUNTER_MAX;
4251 memcg->soft_limit = PAGE_COUNTER_MAX;
4253 memcg->swappiness = mem_cgroup_swappiness(parent);
4254 memcg->oom_kill_disable = parent->oom_kill_disable;
4256 if (parent && parent->use_hierarchy) {
4257 memcg->use_hierarchy = true;
4258 page_counter_init(&memcg->memory, &parent->memory);
4259 page_counter_init(&memcg->swap, &parent->swap);
4260 page_counter_init(&memcg->memsw, &parent->memsw);
4261 page_counter_init(&memcg->kmem, &parent->kmem);
4262 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
4264 page_counter_init(&memcg->memory, NULL);
4265 page_counter_init(&memcg->swap, NULL);
4266 page_counter_init(&memcg->memsw, NULL);
4267 page_counter_init(&memcg->kmem, NULL);
4268 page_counter_init(&memcg->tcpmem, NULL);
4270 * Deeper hierachy with use_hierarchy == false doesn't make
4271 * much sense so let cgroup subsystem know about this
4272 * unfortunate state in our controller.
4274 if (parent != root_mem_cgroup)
4275 memory_cgrp_subsys.broken_hierarchy = true;
4278 /* The following stuff does not apply to the root */
4280 root_mem_cgroup = memcg;
4284 error = memcg_online_kmem(memcg);
4288 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4289 static_branch_inc(&memcg_sockets_enabled_key);
4293 mem_cgroup_free(memcg);
4294 return ERR_PTR(-ENOMEM);
4297 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
4299 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4301 /* Online state pins memcg ID, memcg ID pins CSS */
4302 atomic_set(&memcg->id.ref, 1);
4307 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4309 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4310 struct mem_cgroup_event *event, *tmp;
4313 * Unregister events and notify userspace.
4314 * Notify userspace about cgroup removing only after rmdir of cgroup
4315 * directory to avoid race between userspace and kernelspace.
4317 spin_lock(&memcg->event_list_lock);
4318 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4319 list_del_init(&event->list);
4320 schedule_work(&event->remove);
4322 spin_unlock(&memcg->event_list_lock);
4326 memcg_offline_kmem(memcg);
4327 wb_memcg_offline(memcg);
4329 mem_cgroup_id_put(memcg);
4332 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
4334 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4336 invalidate_reclaim_iterators(memcg);
4339 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4341 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4343 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
4344 static_branch_dec(&memcg_sockets_enabled_key);
4346 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
4347 static_branch_dec(&memcg_sockets_enabled_key);
4349 vmpressure_cleanup(&memcg->vmpressure);
4350 cancel_work_sync(&memcg->high_work);
4351 mem_cgroup_remove_from_trees(memcg);
4352 memcg_free_kmem(memcg);
4353 mem_cgroup_free(memcg);
4357 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4358 * @css: the target css
4360 * Reset the states of the mem_cgroup associated with @css. This is
4361 * invoked when the userland requests disabling on the default hierarchy
4362 * but the memcg is pinned through dependency. The memcg should stop
4363 * applying policies and should revert to the vanilla state as it may be
4364 * made visible again.
4366 * The current implementation only resets the essential configurations.
4367 * This needs to be expanded to cover all the visible parts.
4369 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4373 page_counter_limit(&memcg->memory, PAGE_COUNTER_MAX);
4374 page_counter_limit(&memcg->swap, PAGE_COUNTER_MAX);
4375 page_counter_limit(&memcg->memsw, PAGE_COUNTER_MAX);
4376 page_counter_limit(&memcg->kmem, PAGE_COUNTER_MAX);
4377 page_counter_limit(&memcg->tcpmem, PAGE_COUNTER_MAX);
4379 memcg->high = PAGE_COUNTER_MAX;
4380 memcg->soft_limit = PAGE_COUNTER_MAX;
4381 memcg_wb_domain_size_changed(memcg);
4385 /* Handlers for move charge at task migration. */
4386 static int mem_cgroup_do_precharge(unsigned long count)
4390 /* Try a single bulk charge without reclaim first, kswapd may wake */
4391 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
4393 mc.precharge += count;
4397 /* Try charges one by one with reclaim, but do not retry */
4399 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
4413 enum mc_target_type {
4419 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4420 unsigned long addr, pte_t ptent)
4422 struct page *page = vm_normal_page(vma, addr, ptent);
4424 if (!page || !page_mapped(page))
4426 if (PageAnon(page)) {
4427 if (!(mc.flags & MOVE_ANON))
4430 if (!(mc.flags & MOVE_FILE))
4433 if (!get_page_unless_zero(page))
4440 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4441 pte_t ptent, swp_entry_t *entry)
4443 struct page *page = NULL;
4444 swp_entry_t ent = pte_to_swp_entry(ptent);
4446 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4449 * Because lookup_swap_cache() updates some statistics counter,
4450 * we call find_get_page() with swapper_space directly.
4452 page = find_get_page(swap_address_space(ent), swp_offset(ent));
4453 if (do_memsw_account())
4454 entry->val = ent.val;
4459 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4460 pte_t ptent, swp_entry_t *entry)
4466 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4467 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4469 struct page *page = NULL;
4470 struct address_space *mapping;
4473 if (!vma->vm_file) /* anonymous vma */
4475 if (!(mc.flags & MOVE_FILE))
4478 mapping = vma->vm_file->f_mapping;
4479 pgoff = linear_page_index(vma, addr);
4481 /* page is moved even if it's not RSS of this task(page-faulted). */
4483 /* shmem/tmpfs may report page out on swap: account for that too. */
4484 if (shmem_mapping(mapping)) {
4485 page = find_get_entry(mapping, pgoff);
4486 if (radix_tree_exceptional_entry(page)) {
4487 swp_entry_t swp = radix_to_swp_entry(page);
4488 if (do_memsw_account())
4490 page = find_get_page(swap_address_space(swp),
4494 page = find_get_page(mapping, pgoff);
4496 page = find_get_page(mapping, pgoff);
4502 * mem_cgroup_move_account - move account of the page
4504 * @compound: charge the page as compound or small page
4505 * @from: mem_cgroup which the page is moved from.
4506 * @to: mem_cgroup which the page is moved to. @from != @to.
4508 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
4510 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4513 static int mem_cgroup_move_account(struct page *page,
4515 struct mem_cgroup *from,
4516 struct mem_cgroup *to)
4518 unsigned long flags;
4519 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
4523 VM_BUG_ON(from == to);
4524 VM_BUG_ON_PAGE(PageLRU(page), page);
4525 VM_BUG_ON(compound && !PageTransHuge(page));
4528 * Prevent mem_cgroup_migrate() from looking at
4529 * page->mem_cgroup of its source page while we change it.
4532 if (!trylock_page(page))
4536 if (page->mem_cgroup != from)
4539 anon = PageAnon(page);
4541 spin_lock_irqsave(&from->move_lock, flags);
4543 if (!anon && page_mapped(page)) {
4544 __this_cpu_sub(from->stat->count[NR_FILE_MAPPED], nr_pages);
4545 __this_cpu_add(to->stat->count[NR_FILE_MAPPED], nr_pages);
4549 * move_lock grabbed above and caller set from->moving_account, so
4550 * mod_memcg_page_state will serialize updates to PageDirty.
4551 * So mapping should be stable for dirty pages.
4553 if (!anon && PageDirty(page)) {
4554 struct address_space *mapping = page_mapping(page);
4556 if (mapping_cap_account_dirty(mapping)) {
4557 __this_cpu_sub(from->stat->count[NR_FILE_DIRTY],
4559 __this_cpu_add(to->stat->count[NR_FILE_DIRTY],
4564 if (PageWriteback(page)) {
4565 __this_cpu_sub(from->stat->count[NR_WRITEBACK], nr_pages);
4566 __this_cpu_add(to->stat->count[NR_WRITEBACK], nr_pages);
4570 * It is safe to change page->mem_cgroup here because the page
4571 * is referenced, charged, and isolated - we can't race with
4572 * uncharging, charging, migration, or LRU putback.
4575 /* caller should have done css_get */
4576 page->mem_cgroup = to;
4577 spin_unlock_irqrestore(&from->move_lock, flags);
4581 local_irq_disable();
4582 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
4583 memcg_check_events(to, page);
4584 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
4585 memcg_check_events(from, page);
4594 * get_mctgt_type - get target type of moving charge
4595 * @vma: the vma the pte to be checked belongs
4596 * @addr: the address corresponding to the pte to be checked
4597 * @ptent: the pte to be checked
4598 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4601 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4602 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4603 * move charge. if @target is not NULL, the page is stored in target->page
4604 * with extra refcnt got(Callers should handle it).
4605 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4606 * target for charge migration. if @target is not NULL, the entry is stored
4609 * Called with pte lock held.
4612 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4613 unsigned long addr, pte_t ptent, union mc_target *target)
4615 struct page *page = NULL;
4616 enum mc_target_type ret = MC_TARGET_NONE;
4617 swp_entry_t ent = { .val = 0 };
4619 if (pte_present(ptent))
4620 page = mc_handle_present_pte(vma, addr, ptent);
4621 else if (is_swap_pte(ptent))
4622 page = mc_handle_swap_pte(vma, ptent, &ent);
4623 else if (pte_none(ptent))
4624 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4626 if (!page && !ent.val)
4630 * Do only loose check w/o serialization.
4631 * mem_cgroup_move_account() checks the page is valid or
4632 * not under LRU exclusion.
4634 if (page->mem_cgroup == mc.from) {
4635 ret = MC_TARGET_PAGE;
4637 target->page = page;
4639 if (!ret || !target)
4643 * There is a swap entry and a page doesn't exist or isn't charged.
4644 * But we cannot move a tail-page in a THP.
4646 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
4647 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4648 ret = MC_TARGET_SWAP;
4655 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4657 * We don't consider PMD mapped swapping or file mapped pages because THP does
4658 * not support them for now.
4659 * Caller should make sure that pmd_trans_huge(pmd) is true.
4661 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4662 unsigned long addr, pmd_t pmd, union mc_target *target)
4664 struct page *page = NULL;
4665 enum mc_target_type ret = MC_TARGET_NONE;
4667 page = pmd_page(pmd);
4668 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4669 if (!(mc.flags & MOVE_ANON))
4671 if (page->mem_cgroup == mc.from) {
4672 ret = MC_TARGET_PAGE;
4675 target->page = page;
4681 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4682 unsigned long addr, pmd_t pmd, union mc_target *target)
4684 return MC_TARGET_NONE;
4688 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4689 unsigned long addr, unsigned long end,
4690 struct mm_walk *walk)
4692 struct vm_area_struct *vma = walk->vma;
4696 ptl = pmd_trans_huge_lock(pmd, vma);
4698 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4699 mc.precharge += HPAGE_PMD_NR;
4704 if (pmd_trans_unstable(pmd))
4706 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4707 for (; addr != end; pte++, addr += PAGE_SIZE)
4708 if (get_mctgt_type(vma, addr, *pte, NULL))
4709 mc.precharge++; /* increment precharge temporarily */
4710 pte_unmap_unlock(pte - 1, ptl);
4716 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4718 unsigned long precharge;
4720 struct mm_walk mem_cgroup_count_precharge_walk = {
4721 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4724 down_read(&mm->mmap_sem);
4725 walk_page_range(0, mm->highest_vm_end,
4726 &mem_cgroup_count_precharge_walk);
4727 up_read(&mm->mmap_sem);
4729 precharge = mc.precharge;
4735 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4737 unsigned long precharge = mem_cgroup_count_precharge(mm);
4739 VM_BUG_ON(mc.moving_task);
4740 mc.moving_task = current;
4741 return mem_cgroup_do_precharge(precharge);
4744 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4745 static void __mem_cgroup_clear_mc(void)
4747 struct mem_cgroup *from = mc.from;
4748 struct mem_cgroup *to = mc.to;
4750 /* we must uncharge all the leftover precharges from mc.to */
4752 cancel_charge(mc.to, mc.precharge);
4756 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4757 * we must uncharge here.
4759 if (mc.moved_charge) {
4760 cancel_charge(mc.from, mc.moved_charge);
4761 mc.moved_charge = 0;
4763 /* we must fixup refcnts and charges */
4764 if (mc.moved_swap) {
4765 /* uncharge swap account from the old cgroup */
4766 if (!mem_cgroup_is_root(mc.from))
4767 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4769 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
4772 * we charged both to->memory and to->memsw, so we
4773 * should uncharge to->memory.
4775 if (!mem_cgroup_is_root(mc.to))
4776 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4778 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
4779 css_put_many(&mc.to->css, mc.moved_swap);
4783 memcg_oom_recover(from);
4784 memcg_oom_recover(to);
4785 wake_up_all(&mc.waitq);
4788 static void mem_cgroup_clear_mc(void)
4790 struct mm_struct *mm = mc.mm;
4793 * we must clear moving_task before waking up waiters at the end of
4796 mc.moving_task = NULL;
4797 __mem_cgroup_clear_mc();
4798 spin_lock(&mc.lock);
4802 spin_unlock(&mc.lock);
4807 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
4809 struct cgroup_subsys_state *css;
4810 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
4811 struct mem_cgroup *from;
4812 struct task_struct *leader, *p;
4813 struct mm_struct *mm;
4814 unsigned long move_flags;
4817 /* charge immigration isn't supported on the default hierarchy */
4818 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
4822 * Multi-process migrations only happen on the default hierarchy
4823 * where charge immigration is not used. Perform charge
4824 * immigration if @tset contains a leader and whine if there are
4828 cgroup_taskset_for_each_leader(leader, css, tset) {
4831 memcg = mem_cgroup_from_css(css);
4837 * We are now commited to this value whatever it is. Changes in this
4838 * tunable will only affect upcoming migrations, not the current one.
4839 * So we need to save it, and keep it going.
4841 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
4845 from = mem_cgroup_from_task(p);
4847 VM_BUG_ON(from == memcg);
4849 mm = get_task_mm(p);
4852 /* We move charges only when we move a owner of the mm */
4853 if (mm->owner == p) {
4856 VM_BUG_ON(mc.precharge);
4857 VM_BUG_ON(mc.moved_charge);
4858 VM_BUG_ON(mc.moved_swap);
4860 spin_lock(&mc.lock);
4864 mc.flags = move_flags;
4865 spin_unlock(&mc.lock);
4866 /* We set mc.moving_task later */
4868 ret = mem_cgroup_precharge_mc(mm);
4870 mem_cgroup_clear_mc();
4877 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
4880 mem_cgroup_clear_mc();
4883 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4884 unsigned long addr, unsigned long end,
4885 struct mm_walk *walk)
4888 struct vm_area_struct *vma = walk->vma;
4891 enum mc_target_type target_type;
4892 union mc_target target;
4895 ptl = pmd_trans_huge_lock(pmd, vma);
4897 if (mc.precharge < HPAGE_PMD_NR) {
4901 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
4902 if (target_type == MC_TARGET_PAGE) {
4904 if (!isolate_lru_page(page)) {
4905 if (!mem_cgroup_move_account(page, true,
4907 mc.precharge -= HPAGE_PMD_NR;
4908 mc.moved_charge += HPAGE_PMD_NR;
4910 putback_lru_page(page);
4918 if (pmd_trans_unstable(pmd))
4921 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4922 for (; addr != end; addr += PAGE_SIZE) {
4923 pte_t ptent = *(pte++);
4929 switch (get_mctgt_type(vma, addr, ptent, &target)) {
4930 case MC_TARGET_PAGE:
4933 * We can have a part of the split pmd here. Moving it
4934 * can be done but it would be too convoluted so simply
4935 * ignore such a partial THP and keep it in original
4936 * memcg. There should be somebody mapping the head.
4938 if (PageTransCompound(page))
4940 if (isolate_lru_page(page))
4942 if (!mem_cgroup_move_account(page, false,
4945 /* we uncharge from mc.from later. */
4948 putback_lru_page(page);
4949 put: /* get_mctgt_type() gets the page */
4952 case MC_TARGET_SWAP:
4954 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
4956 /* we fixup refcnts and charges later. */
4964 pte_unmap_unlock(pte - 1, ptl);
4969 * We have consumed all precharges we got in can_attach().
4970 * We try charge one by one, but don't do any additional
4971 * charges to mc.to if we have failed in charge once in attach()
4974 ret = mem_cgroup_do_precharge(1);
4982 static void mem_cgroup_move_charge(void)
4984 struct mm_walk mem_cgroup_move_charge_walk = {
4985 .pmd_entry = mem_cgroup_move_charge_pte_range,
4989 lru_add_drain_all();
4991 * Signal lock_page_memcg() to take the memcg's move_lock
4992 * while we're moving its pages to another memcg. Then wait
4993 * for already started RCU-only updates to finish.
4995 atomic_inc(&mc.from->moving_account);
4998 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5000 * Someone who are holding the mmap_sem might be waiting in
5001 * waitq. So we cancel all extra charges, wake up all waiters,
5002 * and retry. Because we cancel precharges, we might not be able
5003 * to move enough charges, but moving charge is a best-effort
5004 * feature anyway, so it wouldn't be a big problem.
5006 __mem_cgroup_clear_mc();
5011 * When we have consumed all precharges and failed in doing
5012 * additional charge, the page walk just aborts.
5014 walk_page_range(0, mc.mm->highest_vm_end, &mem_cgroup_move_charge_walk);
5016 up_read(&mc.mm->mmap_sem);
5017 atomic_dec(&mc.from->moving_account);
5020 static void mem_cgroup_move_task(void)
5023 mem_cgroup_move_charge();
5024 mem_cgroup_clear_mc();
5027 #else /* !CONFIG_MMU */
5028 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5032 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5035 static void mem_cgroup_move_task(void)
5041 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5042 * to verify whether we're attached to the default hierarchy on each mount
5045 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5048 * use_hierarchy is forced on the default hierarchy. cgroup core
5049 * guarantees that @root doesn't have any children, so turning it
5050 * on for the root memcg is enough.
5052 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5053 root_mem_cgroup->use_hierarchy = true;
5055 root_mem_cgroup->use_hierarchy = false;
5058 static u64 memory_current_read(struct cgroup_subsys_state *css,
5061 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5063 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5066 static int memory_low_show(struct seq_file *m, void *v)
5068 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5069 unsigned long low = READ_ONCE(memcg->low);
5071 if (low == PAGE_COUNTER_MAX)
5072 seq_puts(m, "max\n");
5074 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5079 static ssize_t memory_low_write(struct kernfs_open_file *of,
5080 char *buf, size_t nbytes, loff_t off)
5082 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5086 buf = strstrip(buf);
5087 err = page_counter_memparse(buf, "max", &low);
5096 static int memory_high_show(struct seq_file *m, void *v)
5098 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5099 unsigned long high = READ_ONCE(memcg->high);
5101 if (high == PAGE_COUNTER_MAX)
5102 seq_puts(m, "max\n");
5104 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5109 static ssize_t memory_high_write(struct kernfs_open_file *of,
5110 char *buf, size_t nbytes, loff_t off)
5112 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5113 unsigned long nr_pages;
5117 buf = strstrip(buf);
5118 err = page_counter_memparse(buf, "max", &high);
5124 nr_pages = page_counter_read(&memcg->memory);
5125 if (nr_pages > high)
5126 try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
5129 memcg_wb_domain_size_changed(memcg);
5133 static int memory_max_show(struct seq_file *m, void *v)
5135 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5136 unsigned long max = READ_ONCE(memcg->memory.limit);
5138 if (max == PAGE_COUNTER_MAX)
5139 seq_puts(m, "max\n");
5141 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5146 static ssize_t memory_max_write(struct kernfs_open_file *of,
5147 char *buf, size_t nbytes, loff_t off)
5149 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5150 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
5151 bool drained = false;
5155 buf = strstrip(buf);
5156 err = page_counter_memparse(buf, "max", &max);
5160 xchg(&memcg->memory.limit, max);
5163 unsigned long nr_pages = page_counter_read(&memcg->memory);
5165 if (nr_pages <= max)
5168 if (signal_pending(current)) {
5174 drain_all_stock(memcg);
5180 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
5186 mem_cgroup_event(memcg, MEMCG_OOM);
5187 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
5191 memcg_wb_domain_size_changed(memcg);
5195 static int memory_events_show(struct seq_file *m, void *v)
5197 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5199 seq_printf(m, "low %lu\n", memcg_sum_events(memcg, MEMCG_LOW));
5200 seq_printf(m, "high %lu\n", memcg_sum_events(memcg, MEMCG_HIGH));
5201 seq_printf(m, "max %lu\n", memcg_sum_events(memcg, MEMCG_MAX));
5202 seq_printf(m, "oom %lu\n", memcg_sum_events(memcg, MEMCG_OOM));
5203 seq_printf(m, "oom_kill %lu\n", memcg_sum_events(memcg, OOM_KILL));
5208 static int memory_stat_show(struct seq_file *m, void *v)
5210 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5211 unsigned long stat[MEMCG_NR_STAT];
5212 unsigned long events[MEMCG_NR_EVENTS];
5216 * Provide statistics on the state of the memory subsystem as
5217 * well as cumulative event counters that show past behavior.
5219 * This list is ordered following a combination of these gradients:
5220 * 1) generic big picture -> specifics and details
5221 * 2) reflecting userspace activity -> reflecting kernel heuristics
5223 * Current memory state:
5226 tree_stat(memcg, stat);
5227 tree_events(memcg, events);
5229 seq_printf(m, "anon %llu\n",
5230 (u64)stat[MEMCG_RSS] * PAGE_SIZE);
5231 seq_printf(m, "file %llu\n",
5232 (u64)stat[MEMCG_CACHE] * PAGE_SIZE);
5233 seq_printf(m, "kernel_stack %llu\n",
5234 (u64)stat[MEMCG_KERNEL_STACK_KB] * 1024);
5235 seq_printf(m, "slab %llu\n",
5236 (u64)(stat[NR_SLAB_RECLAIMABLE] +
5237 stat[NR_SLAB_UNRECLAIMABLE]) * PAGE_SIZE);
5238 seq_printf(m, "sock %llu\n",
5239 (u64)stat[MEMCG_SOCK] * PAGE_SIZE);
5241 seq_printf(m, "shmem %llu\n",
5242 (u64)stat[NR_SHMEM] * PAGE_SIZE);
5243 seq_printf(m, "file_mapped %llu\n",
5244 (u64)stat[NR_FILE_MAPPED] * PAGE_SIZE);
5245 seq_printf(m, "file_dirty %llu\n",
5246 (u64)stat[NR_FILE_DIRTY] * PAGE_SIZE);
5247 seq_printf(m, "file_writeback %llu\n",
5248 (u64)stat[NR_WRITEBACK] * PAGE_SIZE);
5250 for (i = 0; i < NR_LRU_LISTS; i++) {
5251 struct mem_cgroup *mi;
5252 unsigned long val = 0;
5254 for_each_mem_cgroup_tree(mi, memcg)
5255 val += mem_cgroup_nr_lru_pages(mi, BIT(i));
5256 seq_printf(m, "%s %llu\n",
5257 mem_cgroup_lru_names[i], (u64)val * PAGE_SIZE);
5260 seq_printf(m, "slab_reclaimable %llu\n",
5261 (u64)stat[NR_SLAB_RECLAIMABLE] * PAGE_SIZE);
5262 seq_printf(m, "slab_unreclaimable %llu\n",
5263 (u64)stat[NR_SLAB_UNRECLAIMABLE] * PAGE_SIZE);
5265 /* Accumulated memory events */
5267 seq_printf(m, "pgfault %lu\n", events[PGFAULT]);
5268 seq_printf(m, "pgmajfault %lu\n", events[PGMAJFAULT]);
5270 seq_printf(m, "pgrefill %lu\n", events[PGREFILL]);
5271 seq_printf(m, "pgscan %lu\n", events[PGSCAN_KSWAPD] +
5272 events[PGSCAN_DIRECT]);
5273 seq_printf(m, "pgsteal %lu\n", events[PGSTEAL_KSWAPD] +
5274 events[PGSTEAL_DIRECT]);
5275 seq_printf(m, "pgactivate %lu\n", events[PGACTIVATE]);
5276 seq_printf(m, "pgdeactivate %lu\n", events[PGDEACTIVATE]);
5277 seq_printf(m, "pglazyfree %lu\n", events[PGLAZYFREE]);
5278 seq_printf(m, "pglazyfreed %lu\n", events[PGLAZYFREED]);
5280 seq_printf(m, "workingset_refault %lu\n",
5281 stat[WORKINGSET_REFAULT]);
5282 seq_printf(m, "workingset_activate %lu\n",
5283 stat[WORKINGSET_ACTIVATE]);
5284 seq_printf(m, "workingset_nodereclaim %lu\n",
5285 stat[WORKINGSET_NODERECLAIM]);
5290 static struct cftype memory_files[] = {
5293 .flags = CFTYPE_NOT_ON_ROOT,
5294 .read_u64 = memory_current_read,
5298 .flags = CFTYPE_NOT_ON_ROOT,
5299 .seq_show = memory_low_show,
5300 .write = memory_low_write,
5304 .flags = CFTYPE_NOT_ON_ROOT,
5305 .seq_show = memory_high_show,
5306 .write = memory_high_write,
5310 .flags = CFTYPE_NOT_ON_ROOT,
5311 .seq_show = memory_max_show,
5312 .write = memory_max_write,
5316 .flags = CFTYPE_NOT_ON_ROOT,
5317 .file_offset = offsetof(struct mem_cgroup, events_file),
5318 .seq_show = memory_events_show,
5322 .flags = CFTYPE_NOT_ON_ROOT,
5323 .seq_show = memory_stat_show,
5328 struct cgroup_subsys memory_cgrp_subsys = {
5329 .css_alloc = mem_cgroup_css_alloc,
5330 .css_online = mem_cgroup_css_online,
5331 .css_offline = mem_cgroup_css_offline,
5332 .css_released = mem_cgroup_css_released,
5333 .css_free = mem_cgroup_css_free,
5334 .css_reset = mem_cgroup_css_reset,
5335 .can_attach = mem_cgroup_can_attach,
5336 .cancel_attach = mem_cgroup_cancel_attach,
5337 .post_attach = mem_cgroup_move_task,
5338 .bind = mem_cgroup_bind,
5339 .dfl_cftypes = memory_files,
5340 .legacy_cftypes = mem_cgroup_legacy_files,
5345 * mem_cgroup_low - check if memory consumption is below the normal range
5346 * @root: the top ancestor of the sub-tree being checked
5347 * @memcg: the memory cgroup to check
5349 * Returns %true if memory consumption of @memcg, and that of all
5350 * ancestors up to (but not including) @root, is below the normal range.
5352 * @root is exclusive; it is never low when looked at directly and isn't
5353 * checked when traversing the hierarchy.
5355 * Excluding @root enables using memory.low to prioritize memory usage
5356 * between cgroups within a subtree of the hierarchy that is limited by
5357 * memory.high or memory.max.
5359 * For example, given cgroup A with children B and C:
5367 * 1. A/memory.current > A/memory.high
5368 * 2. A/B/memory.current < A/B/memory.low
5369 * 3. A/C/memory.current >= A/C/memory.low
5371 * As 'A' is high, i.e. triggers reclaim from 'A', and 'B' is low, we
5372 * should reclaim from 'C' until 'A' is no longer high or until we can
5373 * no longer reclaim from 'C'. If 'A', i.e. @root, isn't excluded by
5374 * mem_cgroup_low when reclaming from 'A', then 'B' won't be considered
5375 * low and we will reclaim indiscriminately from both 'B' and 'C'.
5377 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5379 if (mem_cgroup_disabled())
5383 root = root_mem_cgroup;
5387 for (; memcg != root; memcg = parent_mem_cgroup(memcg)) {
5388 if (page_counter_read(&memcg->memory) >= memcg->low)
5396 * mem_cgroup_try_charge - try charging a page
5397 * @page: page to charge
5398 * @mm: mm context of the victim
5399 * @gfp_mask: reclaim mode
5400 * @memcgp: charged memcg return
5401 * @compound: charge the page as compound or small page
5403 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5404 * pages according to @gfp_mask if necessary.
5406 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5407 * Otherwise, an error code is returned.
5409 * After page->mapping has been set up, the caller must finalize the
5410 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5411 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5413 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5414 gfp_t gfp_mask, struct mem_cgroup **memcgp,
5417 struct mem_cgroup *memcg = NULL;
5418 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5421 if (mem_cgroup_disabled())
5424 if (PageSwapCache(page)) {
5426 * Every swap fault against a single page tries to charge the
5427 * page, bail as early as possible. shmem_unuse() encounters
5428 * already charged pages, too. The USED bit is protected by
5429 * the page lock, which serializes swap cache removal, which
5430 * in turn serializes uncharging.
5432 VM_BUG_ON_PAGE(!PageLocked(page), page);
5433 if (compound_head(page)->mem_cgroup)
5436 if (do_swap_account) {
5437 swp_entry_t ent = { .val = page_private(page), };
5438 unsigned short id = lookup_swap_cgroup_id(ent);
5441 memcg = mem_cgroup_from_id(id);
5442 if (memcg && !css_tryget_online(&memcg->css))
5449 memcg = get_mem_cgroup_from_mm(mm);
5451 ret = try_charge(memcg, gfp_mask, nr_pages);
5453 css_put(&memcg->css);
5460 * mem_cgroup_commit_charge - commit a page charge
5461 * @page: page to charge
5462 * @memcg: memcg to charge the page to
5463 * @lrucare: page might be on LRU already
5464 * @compound: charge the page as compound or small page
5466 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5467 * after page->mapping has been set up. This must happen atomically
5468 * as part of the page instantiation, i.e. under the page table lock
5469 * for anonymous pages, under the page lock for page and swap cache.
5471 * In addition, the page must not be on the LRU during the commit, to
5472 * prevent racing with task migration. If it might be, use @lrucare.
5474 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5476 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5477 bool lrucare, bool compound)
5479 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5481 VM_BUG_ON_PAGE(!page->mapping, page);
5482 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5484 if (mem_cgroup_disabled())
5487 * Swap faults will attempt to charge the same page multiple
5488 * times. But reuse_swap_page() might have removed the page
5489 * from swapcache already, so we can't check PageSwapCache().
5494 commit_charge(page, memcg, lrucare);
5496 local_irq_disable();
5497 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
5498 memcg_check_events(memcg, page);
5501 if (do_memsw_account() && PageSwapCache(page)) {
5502 swp_entry_t entry = { .val = page_private(page) };
5504 * The swap entry might not get freed for a long time,
5505 * let's not wait for it. The page already received a
5506 * memory+swap charge, drop the swap entry duplicate.
5508 mem_cgroup_uncharge_swap(entry, nr_pages);
5513 * mem_cgroup_cancel_charge - cancel a page charge
5514 * @page: page to charge
5515 * @memcg: memcg to charge the page to
5516 * @compound: charge the page as compound or small page
5518 * Cancel a charge transaction started by mem_cgroup_try_charge().
5520 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
5523 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5525 if (mem_cgroup_disabled())
5528 * Swap faults will attempt to charge the same page multiple
5529 * times. But reuse_swap_page() might have removed the page
5530 * from swapcache already, so we can't check PageSwapCache().
5535 cancel_charge(memcg, nr_pages);
5538 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5539 unsigned long nr_anon, unsigned long nr_file,
5540 unsigned long nr_kmem, unsigned long nr_huge,
5541 unsigned long nr_shmem, struct page *dummy_page)
5543 unsigned long nr_pages = nr_anon + nr_file + nr_kmem;
5544 unsigned long flags;
5546 if (!mem_cgroup_is_root(memcg)) {
5547 page_counter_uncharge(&memcg->memory, nr_pages);
5548 if (do_memsw_account())
5549 page_counter_uncharge(&memcg->memsw, nr_pages);
5550 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && nr_kmem)
5551 page_counter_uncharge(&memcg->kmem, nr_kmem);
5552 memcg_oom_recover(memcg);
5555 local_irq_save(flags);
5556 __this_cpu_sub(memcg->stat->count[MEMCG_RSS], nr_anon);
5557 __this_cpu_sub(memcg->stat->count[MEMCG_CACHE], nr_file);
5558 __this_cpu_sub(memcg->stat->count[MEMCG_RSS_HUGE], nr_huge);
5559 __this_cpu_sub(memcg->stat->count[NR_SHMEM], nr_shmem);
5560 __this_cpu_add(memcg->stat->events[PGPGOUT], pgpgout);
5561 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5562 memcg_check_events(memcg, dummy_page);
5563 local_irq_restore(flags);
5565 if (!mem_cgroup_is_root(memcg))
5566 css_put_many(&memcg->css, nr_pages);
5569 static void uncharge_list(struct list_head *page_list)
5571 struct mem_cgroup *memcg = NULL;
5572 unsigned long nr_shmem = 0;
5573 unsigned long nr_anon = 0;
5574 unsigned long nr_file = 0;
5575 unsigned long nr_huge = 0;
5576 unsigned long nr_kmem = 0;
5577 unsigned long pgpgout = 0;
5578 struct list_head *next;
5582 * Note that the list can be a single page->lru; hence the
5583 * do-while loop instead of a simple list_for_each_entry().
5585 next = page_list->next;
5587 page = list_entry(next, struct page, lru);
5588 next = page->lru.next;
5590 VM_BUG_ON_PAGE(PageLRU(page), page);
5591 VM_BUG_ON_PAGE(!PageHWPoison(page) && page_count(page), page);
5593 if (!page->mem_cgroup)
5597 * Nobody should be changing or seriously looking at
5598 * page->mem_cgroup at this point, we have fully
5599 * exclusive access to the page.
5602 if (memcg != page->mem_cgroup) {
5604 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5605 nr_kmem, nr_huge, nr_shmem, page);
5606 pgpgout = nr_anon = nr_file = nr_kmem = 0;
5607 nr_huge = nr_shmem = 0;
5609 memcg = page->mem_cgroup;
5612 if (!PageKmemcg(page)) {
5613 unsigned int nr_pages = 1;
5615 if (PageTransHuge(page)) {
5616 nr_pages <<= compound_order(page);
5617 nr_huge += nr_pages;
5620 nr_anon += nr_pages;
5622 nr_file += nr_pages;
5623 if (PageSwapBacked(page))
5624 nr_shmem += nr_pages;
5628 nr_kmem += 1 << compound_order(page);
5629 __ClearPageKmemcg(page);
5632 page->mem_cgroup = NULL;
5633 } while (next != page_list);
5636 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5637 nr_kmem, nr_huge, nr_shmem, page);
5641 * mem_cgroup_uncharge - uncharge a page
5642 * @page: page to uncharge
5644 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5645 * mem_cgroup_commit_charge().
5647 void mem_cgroup_uncharge(struct page *page)
5649 if (mem_cgroup_disabled())
5652 /* Don't touch page->lru of any random page, pre-check: */
5653 if (!page->mem_cgroup)
5656 INIT_LIST_HEAD(&page->lru);
5657 uncharge_list(&page->lru);
5661 * mem_cgroup_uncharge_list - uncharge a list of page
5662 * @page_list: list of pages to uncharge
5664 * Uncharge a list of pages previously charged with
5665 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5667 void mem_cgroup_uncharge_list(struct list_head *page_list)
5669 if (mem_cgroup_disabled())
5672 if (!list_empty(page_list))
5673 uncharge_list(page_list);
5677 * mem_cgroup_migrate - charge a page's replacement
5678 * @oldpage: currently circulating page
5679 * @newpage: replacement page
5681 * Charge @newpage as a replacement page for @oldpage. @oldpage will
5682 * be uncharged upon free.
5684 * Both pages must be locked, @newpage->mapping must be set up.
5686 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
5688 struct mem_cgroup *memcg;
5689 unsigned int nr_pages;
5691 unsigned long flags;
5693 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5694 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5695 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5696 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5699 if (mem_cgroup_disabled())
5702 /* Page cache replacement: new page already charged? */
5703 if (newpage->mem_cgroup)
5706 /* Swapcache readahead pages can get replaced before being charged */
5707 memcg = oldpage->mem_cgroup;
5711 /* Force-charge the new page. The old one will be freed soon */
5712 compound = PageTransHuge(newpage);
5713 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
5715 page_counter_charge(&memcg->memory, nr_pages);
5716 if (do_memsw_account())
5717 page_counter_charge(&memcg->memsw, nr_pages);
5718 css_get_many(&memcg->css, nr_pages);
5720 commit_charge(newpage, memcg, false);
5722 local_irq_save(flags);
5723 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
5724 memcg_check_events(memcg, newpage);
5725 local_irq_restore(flags);
5728 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
5729 EXPORT_SYMBOL(memcg_sockets_enabled_key);
5731 void mem_cgroup_sk_alloc(struct sock *sk)
5733 struct mem_cgroup *memcg;
5735 if (!mem_cgroup_sockets_enabled)
5739 * Socket cloning can throw us here with sk_memcg already
5740 * filled. It won't however, necessarily happen from
5741 * process context. So the test for root memcg given
5742 * the current task's memcg won't help us in this case.
5744 * Respecting the original socket's memcg is a better
5745 * decision in this case.
5748 BUG_ON(mem_cgroup_is_root(sk->sk_memcg));
5749 css_get(&sk->sk_memcg->css);
5754 memcg = mem_cgroup_from_task(current);
5755 if (memcg == root_mem_cgroup)
5757 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
5759 if (css_tryget_online(&memcg->css))
5760 sk->sk_memcg = memcg;
5765 void mem_cgroup_sk_free(struct sock *sk)
5768 css_put(&sk->sk_memcg->css);
5772 * mem_cgroup_charge_skmem - charge socket memory
5773 * @memcg: memcg to charge
5774 * @nr_pages: number of pages to charge
5776 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
5777 * @memcg's configured limit, %false if the charge had to be forced.
5779 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5781 gfp_t gfp_mask = GFP_KERNEL;
5783 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5784 struct page_counter *fail;
5786 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
5787 memcg->tcpmem_pressure = 0;
5790 page_counter_charge(&memcg->tcpmem, nr_pages);
5791 memcg->tcpmem_pressure = 1;
5795 /* Don't block in the packet receive path */
5797 gfp_mask = GFP_NOWAIT;
5799 this_cpu_add(memcg->stat->count[MEMCG_SOCK], nr_pages);
5801 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
5804 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
5809 * mem_cgroup_uncharge_skmem - uncharge socket memory
5810 * @memcg - memcg to uncharge
5811 * @nr_pages - number of pages to uncharge
5813 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
5815 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
5816 page_counter_uncharge(&memcg->tcpmem, nr_pages);
5820 this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
5822 page_counter_uncharge(&memcg->memory, nr_pages);
5823 css_put_many(&memcg->css, nr_pages);
5826 static int __init cgroup_memory(char *s)
5830 while ((token = strsep(&s, ",")) != NULL) {
5833 if (!strcmp(token, "nosocket"))
5834 cgroup_memory_nosocket = true;
5835 if (!strcmp(token, "nokmem"))
5836 cgroup_memory_nokmem = true;
5840 __setup("cgroup.memory=", cgroup_memory);
5843 * subsys_initcall() for memory controller.
5845 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
5846 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
5847 * basically everything that doesn't depend on a specific mem_cgroup structure
5848 * should be initialized from here.
5850 static int __init mem_cgroup_init(void)
5856 * Kmem cache creation is mostly done with the slab_mutex held,
5857 * so use a workqueue with limited concurrency to avoid stalling
5858 * all worker threads in case lots of cgroups are created and
5859 * destroyed simultaneously.
5861 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
5862 BUG_ON(!memcg_kmem_cache_wq);
5865 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
5866 memcg_hotplug_cpu_dead);
5868 for_each_possible_cpu(cpu)
5869 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5872 for_each_node(node) {
5873 struct mem_cgroup_tree_per_node *rtpn;
5875 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5876 node_online(node) ? node : NUMA_NO_NODE);
5878 rtpn->rb_root = RB_ROOT;
5879 spin_lock_init(&rtpn->lock);
5880 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5885 subsys_initcall(mem_cgroup_init);
5887 #ifdef CONFIG_MEMCG_SWAP
5888 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
5890 while (!atomic_inc_not_zero(&memcg->id.ref)) {
5892 * The root cgroup cannot be destroyed, so it's refcount must
5895 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
5899 memcg = parent_mem_cgroup(memcg);
5901 memcg = root_mem_cgroup;
5907 * mem_cgroup_swapout - transfer a memsw charge to swap
5908 * @page: page whose memsw charge to transfer
5909 * @entry: swap entry to move the charge to
5911 * Transfer the memsw charge of @page to @entry.
5913 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5915 struct mem_cgroup *memcg, *swap_memcg;
5916 unsigned int nr_entries;
5917 unsigned short oldid;
5919 VM_BUG_ON_PAGE(PageLRU(page), page);
5920 VM_BUG_ON_PAGE(page_count(page), page);
5922 if (!do_memsw_account())
5925 memcg = page->mem_cgroup;
5927 /* Readahead page, never charged */
5932 * In case the memcg owning these pages has been offlined and doesn't
5933 * have an ID allocated to it anymore, charge the closest online
5934 * ancestor for the swap instead and transfer the memory+swap charge.
5936 swap_memcg = mem_cgroup_id_get_online(memcg);
5937 nr_entries = hpage_nr_pages(page);
5938 /* Get references for the tail pages, too */
5940 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
5941 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
5943 VM_BUG_ON_PAGE(oldid, page);
5944 mem_cgroup_swap_statistics(swap_memcg, nr_entries);
5946 page->mem_cgroup = NULL;
5948 if (!mem_cgroup_is_root(memcg))
5949 page_counter_uncharge(&memcg->memory, nr_entries);
5951 if (memcg != swap_memcg) {
5952 if (!mem_cgroup_is_root(swap_memcg))
5953 page_counter_charge(&swap_memcg->memsw, nr_entries);
5954 page_counter_uncharge(&memcg->memsw, nr_entries);
5958 * Interrupts should be disabled here because the caller holds the
5959 * mapping->tree_lock lock which is taken with interrupts-off. It is
5960 * important here to have the interrupts disabled because it is the
5961 * only synchronisation we have for udpating the per-CPU variables.
5963 VM_BUG_ON(!irqs_disabled());
5964 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
5966 memcg_check_events(memcg, page);
5968 if (!mem_cgroup_is_root(memcg))
5969 css_put(&memcg->css);
5973 * mem_cgroup_try_charge_swap - try charging swap space for a page
5974 * @page: page being added to swap
5975 * @entry: swap entry to charge
5977 * Try to charge @page's memcg for the swap space at @entry.
5979 * Returns 0 on success, -ENOMEM on failure.
5981 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
5983 unsigned int nr_pages = hpage_nr_pages(page);
5984 struct page_counter *counter;
5985 struct mem_cgroup *memcg;
5986 unsigned short oldid;
5988 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
5991 memcg = page->mem_cgroup;
5993 /* Readahead page, never charged */
5997 memcg = mem_cgroup_id_get_online(memcg);
5999 if (!mem_cgroup_is_root(memcg) &&
6000 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6001 mem_cgroup_id_put(memcg);
6005 /* Get references for the tail pages, too */
6007 mem_cgroup_id_get_many(memcg, nr_pages - 1);
6008 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
6009 VM_BUG_ON_PAGE(oldid, page);
6010 mem_cgroup_swap_statistics(memcg, nr_pages);
6016 * mem_cgroup_uncharge_swap - uncharge swap space
6017 * @entry: swap entry to uncharge
6018 * @nr_pages: the amount of swap space to uncharge
6020 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
6022 struct mem_cgroup *memcg;
6025 if (!do_swap_account)
6028 id = swap_cgroup_record(entry, 0, nr_pages);
6030 memcg = mem_cgroup_from_id(id);
6032 if (!mem_cgroup_is_root(memcg)) {
6033 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6034 page_counter_uncharge(&memcg->swap, nr_pages);
6036 page_counter_uncharge(&memcg->memsw, nr_pages);
6038 mem_cgroup_swap_statistics(memcg, -nr_pages);
6039 mem_cgroup_id_put_many(memcg, nr_pages);
6044 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
6046 long nr_swap_pages = get_nr_swap_pages();
6048 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6049 return nr_swap_pages;
6050 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6051 nr_swap_pages = min_t(long, nr_swap_pages,
6052 READ_ONCE(memcg->swap.limit) -
6053 page_counter_read(&memcg->swap));
6054 return nr_swap_pages;
6057 bool mem_cgroup_swap_full(struct page *page)
6059 struct mem_cgroup *memcg;
6061 VM_BUG_ON_PAGE(!PageLocked(page), page);
6065 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
6068 memcg = page->mem_cgroup;
6072 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
6073 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.limit)
6079 /* for remember boot option*/
6080 #ifdef CONFIG_MEMCG_SWAP_ENABLED
6081 static int really_do_swap_account __initdata = 1;
6083 static int really_do_swap_account __initdata;
6086 static int __init enable_swap_account(char *s)
6088 if (!strcmp(s, "1"))
6089 really_do_swap_account = 1;
6090 else if (!strcmp(s, "0"))
6091 really_do_swap_account = 0;
6094 __setup("swapaccount=", enable_swap_account);
6096 static u64 swap_current_read(struct cgroup_subsys_state *css,
6099 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6101 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
6104 static int swap_max_show(struct seq_file *m, void *v)
6106 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
6107 unsigned long max = READ_ONCE(memcg->swap.limit);
6109 if (max == PAGE_COUNTER_MAX)
6110 seq_puts(m, "max\n");
6112 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
6117 static ssize_t swap_max_write(struct kernfs_open_file *of,
6118 char *buf, size_t nbytes, loff_t off)
6120 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6124 buf = strstrip(buf);
6125 err = page_counter_memparse(buf, "max", &max);
6129 mutex_lock(&memcg_limit_mutex);
6130 err = page_counter_limit(&memcg->swap, max);
6131 mutex_unlock(&memcg_limit_mutex);
6138 static struct cftype swap_files[] = {
6140 .name = "swap.current",
6141 .flags = CFTYPE_NOT_ON_ROOT,
6142 .read_u64 = swap_current_read,
6146 .flags = CFTYPE_NOT_ON_ROOT,
6147 .seq_show = swap_max_show,
6148 .write = swap_max_write,
6153 static struct cftype memsw_cgroup_files[] = {
6155 .name = "memsw.usage_in_bytes",
6156 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6157 .read_u64 = mem_cgroup_read_u64,
6160 .name = "memsw.max_usage_in_bytes",
6161 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6162 .write = mem_cgroup_reset,
6163 .read_u64 = mem_cgroup_read_u64,
6166 .name = "memsw.limit_in_bytes",
6167 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6168 .write = mem_cgroup_write,
6169 .read_u64 = mem_cgroup_read_u64,
6172 .name = "memsw.failcnt",
6173 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6174 .write = mem_cgroup_reset,
6175 .read_u64 = mem_cgroup_read_u64,
6177 { }, /* terminate */
6180 static int __init mem_cgroup_swap_init(void)
6182 if (!mem_cgroup_disabled() && really_do_swap_account) {
6183 do_swap_account = 1;
6184 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
6186 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6187 memsw_cgroup_files));
6191 subsys_initcall(mem_cgroup_swap_init);
6193 #endif /* CONFIG_MEMCG_SWAP */