1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 #define MEM_CGROUP_RECLAIM_RETRIES 5
78 /* Socket memory accounting disabled? */
79 static bool cgroup_memory_nosocket;
81 /* Kernel memory accounting disabled? */
82 static bool cgroup_memory_nokmem;
84 /* Whether the swap controller is active */
85 #ifdef CONFIG_MEMCG_SWAP
86 int do_swap_account __read_mostly;
88 #define do_swap_account 0
91 #ifdef CONFIG_CGROUP_WRITEBACK
92 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
95 /* Whether legacy memory+swap accounting is active */
96 static bool do_memsw_account(void)
98 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && do_swap_account;
101 static const char *const mem_cgroup_lru_names[] = {
109 #define THRESHOLDS_EVENTS_TARGET 128
110 #define SOFTLIMIT_EVENTS_TARGET 1024
113 * Cgroups above their limits are maintained in a RB-Tree, independent of
114 * their hierarchy representation
117 struct mem_cgroup_tree_per_node {
118 struct rb_root rb_root;
119 struct rb_node *rb_rightmost;
123 struct mem_cgroup_tree {
124 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
127 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
130 struct mem_cgroup_eventfd_list {
131 struct list_head list;
132 struct eventfd_ctx *eventfd;
136 * cgroup_event represents events which userspace want to receive.
138 struct mem_cgroup_event {
140 * memcg which the event belongs to.
142 struct mem_cgroup *memcg;
144 * eventfd to signal userspace about the event.
146 struct eventfd_ctx *eventfd;
148 * Each of these stored in a list by the cgroup.
150 struct list_head list;
152 * register_event() callback will be used to add new userspace
153 * waiter for changes related to this event. Use eventfd_signal()
154 * on eventfd to send notification to userspace.
156 int (*register_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd, const char *args);
159 * unregister_event() callback will be called when userspace closes
160 * the eventfd or on cgroup removing. This callback must be set,
161 * if you want provide notification functionality.
163 void (*unregister_event)(struct mem_cgroup *memcg,
164 struct eventfd_ctx *eventfd);
166 * All fields below needed to unregister event when
167 * userspace closes eventfd.
170 wait_queue_head_t *wqh;
171 wait_queue_entry_t wait;
172 struct work_struct remove;
175 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
176 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
178 /* Stuffs for move charges at task migration. */
180 * Types of charges to be moved.
182 #define MOVE_ANON 0x1U
183 #define MOVE_FILE 0x2U
184 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
186 /* "mc" and its members are protected by cgroup_mutex */
187 static struct move_charge_struct {
188 spinlock_t lock; /* for from, to */
189 struct mm_struct *mm;
190 struct mem_cgroup *from;
191 struct mem_cgroup *to;
193 unsigned long precharge;
194 unsigned long moved_charge;
195 unsigned long moved_swap;
196 struct task_struct *moving_task; /* a task moving charges */
197 wait_queue_head_t waitq; /* a waitq for other context */
199 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
200 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
204 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
205 * limit reclaim to prevent infinite loops, if they ever occur.
207 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
208 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
211 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
212 MEM_CGROUP_CHARGE_TYPE_ANON,
213 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
214 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
218 /* for encoding cft->private value on file */
227 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
228 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
229 #define MEMFILE_ATTR(val) ((val) & 0xffff)
230 /* Used for OOM nofiier */
231 #define OOM_CONTROL (0)
234 * Iteration constructs for visiting all cgroups (under a tree). If
235 * loops are exited prematurely (break), mem_cgroup_iter_break() must
236 * be used for reference counting.
238 #define for_each_mem_cgroup_tree(iter, root) \
239 for (iter = mem_cgroup_iter(root, NULL, NULL); \
241 iter = mem_cgroup_iter(root, iter, NULL))
243 #define for_each_mem_cgroup(iter) \
244 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
246 iter = mem_cgroup_iter(NULL, iter, NULL))
248 static inline bool should_force_charge(void)
250 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
251 (current->flags & PF_EXITING);
254 /* Some nice accessors for the vmpressure. */
255 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
258 memcg = root_mem_cgroup;
259 return &memcg->vmpressure;
262 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
264 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
267 #ifdef CONFIG_MEMCG_KMEM
269 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
270 * The main reason for not using cgroup id for this:
271 * this works better in sparse environments, where we have a lot of memcgs,
272 * but only a few kmem-limited. Or also, if we have, for instance, 200
273 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
274 * 200 entry array for that.
276 * The current size of the caches array is stored in memcg_nr_cache_ids. It
277 * will double each time we have to increase it.
279 static DEFINE_IDA(memcg_cache_ida);
280 int memcg_nr_cache_ids;
282 /* Protects memcg_nr_cache_ids */
283 static DECLARE_RWSEM(memcg_cache_ids_sem);
285 void memcg_get_cache_ids(void)
287 down_read(&memcg_cache_ids_sem);
290 void memcg_put_cache_ids(void)
292 up_read(&memcg_cache_ids_sem);
296 * MIN_SIZE is different than 1, because we would like to avoid going through
297 * the alloc/free process all the time. In a small machine, 4 kmem-limited
298 * cgroups is a reasonable guess. In the future, it could be a parameter or
299 * tunable, but that is strictly not necessary.
301 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
302 * this constant directly from cgroup, but it is understandable that this is
303 * better kept as an internal representation in cgroup.c. In any case, the
304 * cgrp_id space is not getting any smaller, and we don't have to necessarily
305 * increase ours as well if it increases.
307 #define MEMCG_CACHES_MIN_SIZE 4
308 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
311 * A lot of the calls to the cache allocation functions are expected to be
312 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
313 * conditional to this static branch, we'll have to allow modules that does
314 * kmem_cache_alloc and the such to see this symbol as well
316 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
317 EXPORT_SYMBOL(memcg_kmem_enabled_key);
319 struct workqueue_struct *memcg_kmem_cache_wq;
322 static int memcg_shrinker_map_size;
323 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
325 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
327 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
330 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
331 int size, int old_size)
333 struct memcg_shrinker_map *new, *old;
336 lockdep_assert_held(&memcg_shrinker_map_mutex);
339 old = rcu_dereference_protected(
340 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
341 /* Not yet online memcg */
345 new = kvmalloc(sizeof(*new) + size, GFP_KERNEL);
349 /* Set all old bits, clear all new bits */
350 memset(new->map, (int)0xff, old_size);
351 memset((void *)new->map + old_size, 0, size - old_size);
353 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
354 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
360 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
362 struct mem_cgroup_per_node *pn;
363 struct memcg_shrinker_map *map;
366 if (mem_cgroup_is_root(memcg))
370 pn = mem_cgroup_nodeinfo(memcg, nid);
371 map = rcu_dereference_protected(pn->shrinker_map, true);
374 rcu_assign_pointer(pn->shrinker_map, NULL);
378 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
380 struct memcg_shrinker_map *map;
381 int nid, size, ret = 0;
383 if (mem_cgroup_is_root(memcg))
386 mutex_lock(&memcg_shrinker_map_mutex);
387 size = memcg_shrinker_map_size;
389 map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
391 memcg_free_shrinker_maps(memcg);
395 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
397 mutex_unlock(&memcg_shrinker_map_mutex);
402 int memcg_expand_shrinker_maps(int new_id)
404 int size, old_size, ret = 0;
405 struct mem_cgroup *memcg;
407 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
408 old_size = memcg_shrinker_map_size;
409 if (size <= old_size)
412 mutex_lock(&memcg_shrinker_map_mutex);
413 if (!root_mem_cgroup)
416 for_each_mem_cgroup(memcg) {
417 if (mem_cgroup_is_root(memcg))
419 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
425 memcg_shrinker_map_size = size;
426 mutex_unlock(&memcg_shrinker_map_mutex);
430 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
432 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
433 struct memcg_shrinker_map *map;
436 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
437 /* Pairs with smp mb in shrink_slab() */
438 smp_mb__before_atomic();
439 set_bit(shrinker_id, map->map);
445 * mem_cgroup_css_from_page - css of the memcg associated with a page
446 * @page: page of interest
448 * If memcg is bound to the default hierarchy, css of the memcg associated
449 * with @page is returned. The returned css remains associated with @page
450 * until it is released.
452 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
455 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
457 struct mem_cgroup *memcg;
459 memcg = page->mem_cgroup;
461 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
462 memcg = root_mem_cgroup;
468 * page_cgroup_ino - return inode number of the memcg a page is charged to
471 * Look up the closest online ancestor of the memory cgroup @page is charged to
472 * and return its inode number or 0 if @page is not charged to any cgroup. It
473 * is safe to call this function without holding a reference to @page.
475 * Note, this function is inherently racy, because there is nothing to prevent
476 * the cgroup inode from getting torn down and potentially reallocated a moment
477 * after page_cgroup_ino() returns, so it only should be used by callers that
478 * do not care (such as procfs interfaces).
480 ino_t page_cgroup_ino(struct page *page)
482 struct mem_cgroup *memcg;
483 unsigned long ino = 0;
486 if (PageSlab(page) && !PageTail(page))
487 memcg = memcg_from_slab_page(page);
489 memcg = READ_ONCE(page->mem_cgroup);
490 while (memcg && !(memcg->css.flags & CSS_ONLINE))
491 memcg = parent_mem_cgroup(memcg);
493 ino = cgroup_ino(memcg->css.cgroup);
498 static struct mem_cgroup_per_node *
499 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
501 int nid = page_to_nid(page);
503 return memcg->nodeinfo[nid];
506 static struct mem_cgroup_tree_per_node *
507 soft_limit_tree_node(int nid)
509 return soft_limit_tree.rb_tree_per_node[nid];
512 static struct mem_cgroup_tree_per_node *
513 soft_limit_tree_from_page(struct page *page)
515 int nid = page_to_nid(page);
517 return soft_limit_tree.rb_tree_per_node[nid];
520 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
521 struct mem_cgroup_tree_per_node *mctz,
522 unsigned long new_usage_in_excess)
524 struct rb_node **p = &mctz->rb_root.rb_node;
525 struct rb_node *parent = NULL;
526 struct mem_cgroup_per_node *mz_node;
527 bool rightmost = true;
532 mz->usage_in_excess = new_usage_in_excess;
533 if (!mz->usage_in_excess)
537 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
539 if (mz->usage_in_excess < mz_node->usage_in_excess) {
545 * We can't avoid mem cgroups that are over their soft
546 * limit by the same amount
548 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
553 mctz->rb_rightmost = &mz->tree_node;
555 rb_link_node(&mz->tree_node, parent, p);
556 rb_insert_color(&mz->tree_node, &mctz->rb_root);
560 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
561 struct mem_cgroup_tree_per_node *mctz)
566 if (&mz->tree_node == mctz->rb_rightmost)
567 mctz->rb_rightmost = rb_prev(&mz->tree_node);
569 rb_erase(&mz->tree_node, &mctz->rb_root);
573 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
574 struct mem_cgroup_tree_per_node *mctz)
578 spin_lock_irqsave(&mctz->lock, flags);
579 __mem_cgroup_remove_exceeded(mz, mctz);
580 spin_unlock_irqrestore(&mctz->lock, flags);
583 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
585 unsigned long nr_pages = page_counter_read(&memcg->memory);
586 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
587 unsigned long excess = 0;
589 if (nr_pages > soft_limit)
590 excess = nr_pages - soft_limit;
595 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
597 unsigned long excess;
598 struct mem_cgroup_per_node *mz;
599 struct mem_cgroup_tree_per_node *mctz;
601 mctz = soft_limit_tree_from_page(page);
605 * Necessary to update all ancestors when hierarchy is used.
606 * because their event counter is not touched.
608 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
609 mz = mem_cgroup_page_nodeinfo(memcg, page);
610 excess = soft_limit_excess(memcg);
612 * We have to update the tree if mz is on RB-tree or
613 * mem is over its softlimit.
615 if (excess || mz->on_tree) {
618 spin_lock_irqsave(&mctz->lock, flags);
619 /* if on-tree, remove it */
621 __mem_cgroup_remove_exceeded(mz, mctz);
623 * Insert again. mz->usage_in_excess will be updated.
624 * If excess is 0, no tree ops.
626 __mem_cgroup_insert_exceeded(mz, mctz, excess);
627 spin_unlock_irqrestore(&mctz->lock, flags);
632 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
634 struct mem_cgroup_tree_per_node *mctz;
635 struct mem_cgroup_per_node *mz;
639 mz = mem_cgroup_nodeinfo(memcg, nid);
640 mctz = soft_limit_tree_node(nid);
642 mem_cgroup_remove_exceeded(mz, mctz);
646 static struct mem_cgroup_per_node *
647 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
649 struct mem_cgroup_per_node *mz;
653 if (!mctz->rb_rightmost)
654 goto done; /* Nothing to reclaim from */
656 mz = rb_entry(mctz->rb_rightmost,
657 struct mem_cgroup_per_node, tree_node);
659 * Remove the node now but someone else can add it back,
660 * we will to add it back at the end of reclaim to its correct
661 * position in the tree.
663 __mem_cgroup_remove_exceeded(mz, mctz);
664 if (!soft_limit_excess(mz->memcg) ||
665 !css_tryget_online(&mz->memcg->css))
671 static struct mem_cgroup_per_node *
672 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
674 struct mem_cgroup_per_node *mz;
676 spin_lock_irq(&mctz->lock);
677 mz = __mem_cgroup_largest_soft_limit_node(mctz);
678 spin_unlock_irq(&mctz->lock);
683 * __mod_memcg_state - update cgroup memory statistics
684 * @memcg: the memory cgroup
685 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
686 * @val: delta to add to the counter, can be negative
688 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
692 if (mem_cgroup_disabled())
695 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
696 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
697 struct mem_cgroup *mi;
700 * Batch local counters to keep them in sync with
701 * the hierarchical ones.
703 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
704 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
705 atomic_long_add(x, &mi->vmstats[idx]);
708 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
711 static struct mem_cgroup_per_node *
712 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
714 struct mem_cgroup *parent;
716 parent = parent_mem_cgroup(pn->memcg);
719 return mem_cgroup_nodeinfo(parent, nid);
723 * __mod_lruvec_state - update lruvec memory statistics
724 * @lruvec: the lruvec
725 * @idx: the stat item
726 * @val: delta to add to the counter, can be negative
728 * The lruvec is the intersection of the NUMA node and a cgroup. This
729 * function updates the all three counters that are affected by a
730 * change of state at this level: per-node, per-cgroup, per-lruvec.
732 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
735 pg_data_t *pgdat = lruvec_pgdat(lruvec);
736 struct mem_cgroup_per_node *pn;
737 struct mem_cgroup *memcg;
741 __mod_node_page_state(pgdat, idx, val);
743 if (mem_cgroup_disabled())
746 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
750 __mod_memcg_state(memcg, idx, val);
753 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
755 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
756 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
757 struct mem_cgroup_per_node *pi;
759 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
760 atomic_long_add(x, &pi->lruvec_stat[idx]);
763 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
766 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
768 struct page *page = virt_to_head_page(p);
769 pg_data_t *pgdat = page_pgdat(page);
770 struct mem_cgroup *memcg;
771 struct lruvec *lruvec;
774 memcg = memcg_from_slab_page(page);
776 /* Untracked pages have no memcg, no lruvec. Update only the node */
777 if (!memcg || memcg == root_mem_cgroup) {
778 __mod_node_page_state(pgdat, idx, val);
780 lruvec = mem_cgroup_lruvec(pgdat, memcg);
781 __mod_lruvec_state(lruvec, idx, val);
787 * __count_memcg_events - account VM events in a cgroup
788 * @memcg: the memory cgroup
789 * @idx: the event item
790 * @count: the number of events that occured
792 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
797 if (mem_cgroup_disabled())
800 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
801 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
802 struct mem_cgroup *mi;
805 * Batch local counters to keep them in sync with
806 * the hierarchical ones.
808 __this_cpu_add(memcg->vmstats_local->events[idx], x);
809 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
810 atomic_long_add(x, &mi->vmevents[idx]);
813 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
816 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
818 return atomic_long_read(&memcg->vmevents[event]);
821 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
826 for_each_possible_cpu(cpu)
827 x += per_cpu(memcg->vmstats_local->events[event], cpu);
831 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
833 bool compound, int nr_pages)
836 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
837 * counted as CACHE even if it's on ANON LRU.
840 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
842 __mod_memcg_state(memcg, MEMCG_CACHE, nr_pages);
843 if (PageSwapBacked(page))
844 __mod_memcg_state(memcg, NR_SHMEM, nr_pages);
848 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
849 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
852 /* pagein of a big page is an event. So, ignore page size */
854 __count_memcg_events(memcg, PGPGIN, 1);
856 __count_memcg_events(memcg, PGPGOUT, 1);
857 nr_pages = -nr_pages; /* for event */
860 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
863 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
864 enum mem_cgroup_events_target target)
866 unsigned long val, next;
868 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
869 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
870 /* from time_after() in jiffies.h */
871 if ((long)(next - val) < 0) {
873 case MEM_CGROUP_TARGET_THRESH:
874 next = val + THRESHOLDS_EVENTS_TARGET;
876 case MEM_CGROUP_TARGET_SOFTLIMIT:
877 next = val + SOFTLIMIT_EVENTS_TARGET;
882 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
889 * Check events in order.
892 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
894 /* threshold event is triggered in finer grain than soft limit */
895 if (unlikely(mem_cgroup_event_ratelimit(memcg,
896 MEM_CGROUP_TARGET_THRESH))) {
899 do_softlimit = mem_cgroup_event_ratelimit(memcg,
900 MEM_CGROUP_TARGET_SOFTLIMIT);
901 mem_cgroup_threshold(memcg);
902 if (unlikely(do_softlimit))
903 mem_cgroup_update_tree(memcg, page);
907 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
910 * mm_update_next_owner() may clear mm->owner to NULL
911 * if it races with swapoff, page migration, etc.
912 * So this can be called with p == NULL.
917 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
919 EXPORT_SYMBOL(mem_cgroup_from_task);
922 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
923 * @mm: mm from which memcg should be extracted. It can be NULL.
925 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
926 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
929 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
931 struct mem_cgroup *memcg;
933 if (mem_cgroup_disabled())
939 * Page cache insertions can happen withou an
940 * actual mm context, e.g. during disk probing
941 * on boot, loopback IO, acct() writes etc.
944 memcg = root_mem_cgroup;
946 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
947 if (unlikely(!memcg))
948 memcg = root_mem_cgroup;
950 } while (!css_tryget(&memcg->css));
954 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
957 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
958 * @page: page from which memcg should be extracted.
960 * Obtain a reference on page->memcg and returns it if successful. Otherwise
961 * root_mem_cgroup is returned.
963 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
965 struct mem_cgroup *memcg = page->mem_cgroup;
967 if (mem_cgroup_disabled())
971 if (!memcg || !css_tryget_online(&memcg->css))
972 memcg = root_mem_cgroup;
976 EXPORT_SYMBOL(get_mem_cgroup_from_page);
979 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
981 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
983 if (unlikely(current->active_memcg)) {
984 struct mem_cgroup *memcg = root_mem_cgroup;
987 if (css_tryget_online(¤t->active_memcg->css))
988 memcg = current->active_memcg;
992 return get_mem_cgroup_from_mm(current->mm);
996 * mem_cgroup_iter - iterate over memory cgroup hierarchy
997 * @root: hierarchy root
998 * @prev: previously returned memcg, NULL on first invocation
999 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1001 * Returns references to children of the hierarchy below @root, or
1002 * @root itself, or %NULL after a full round-trip.
1004 * Caller must pass the return value in @prev on subsequent
1005 * invocations for reference counting, or use mem_cgroup_iter_break()
1006 * to cancel a hierarchy walk before the round-trip is complete.
1008 * Reclaimers can specify a node and a priority level in @reclaim to
1009 * divide up the memcgs in the hierarchy among all concurrent
1010 * reclaimers operating on the same node and priority.
1012 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1013 struct mem_cgroup *prev,
1014 struct mem_cgroup_reclaim_cookie *reclaim)
1016 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1017 struct cgroup_subsys_state *css = NULL;
1018 struct mem_cgroup *memcg = NULL;
1019 struct mem_cgroup *pos = NULL;
1021 if (mem_cgroup_disabled())
1025 root = root_mem_cgroup;
1027 if (prev && !reclaim)
1030 if (!root->use_hierarchy && root != root_mem_cgroup) {
1039 struct mem_cgroup_per_node *mz;
1041 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1044 if (prev && reclaim->generation != iter->generation)
1048 pos = READ_ONCE(iter->position);
1049 if (!pos || css_tryget(&pos->css))
1052 * css reference reached zero, so iter->position will
1053 * be cleared by ->css_released. However, we should not
1054 * rely on this happening soon, because ->css_released
1055 * is called from a work queue, and by busy-waiting we
1056 * might block it. So we clear iter->position right
1059 (void)cmpxchg(&iter->position, pos, NULL);
1067 css = css_next_descendant_pre(css, &root->css);
1070 * Reclaimers share the hierarchy walk, and a
1071 * new one might jump in right at the end of
1072 * the hierarchy - make sure they see at least
1073 * one group and restart from the beginning.
1081 * Verify the css and acquire a reference. The root
1082 * is provided by the caller, so we know it's alive
1083 * and kicking, and don't take an extra reference.
1085 memcg = mem_cgroup_from_css(css);
1087 if (css == &root->css)
1090 if (css_tryget(css))
1098 * The position could have already been updated by a competing
1099 * thread, so check that the value hasn't changed since we read
1100 * it to avoid reclaiming from the same cgroup twice.
1102 (void)cmpxchg(&iter->position, pos, memcg);
1110 reclaim->generation = iter->generation;
1116 if (prev && prev != root)
1117 css_put(&prev->css);
1123 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1124 * @root: hierarchy root
1125 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1127 void mem_cgroup_iter_break(struct mem_cgroup *root,
1128 struct mem_cgroup *prev)
1131 root = root_mem_cgroup;
1132 if (prev && prev != root)
1133 css_put(&prev->css);
1136 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1137 struct mem_cgroup *dead_memcg)
1139 struct mem_cgroup_reclaim_iter *iter;
1140 struct mem_cgroup_per_node *mz;
1143 for_each_node(nid) {
1144 mz = mem_cgroup_nodeinfo(from, nid);
1146 cmpxchg(&iter->position, dead_memcg, NULL);
1150 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1152 struct mem_cgroup *memcg = dead_memcg;
1153 struct mem_cgroup *last;
1156 __invalidate_reclaim_iterators(memcg, dead_memcg);
1158 } while ((memcg = parent_mem_cgroup(memcg)));
1161 * When cgruop1 non-hierarchy mode is used,
1162 * parent_mem_cgroup() does not walk all the way up to the
1163 * cgroup root (root_mem_cgroup). So we have to handle
1164 * dead_memcg from cgroup root separately.
1166 if (last != root_mem_cgroup)
1167 __invalidate_reclaim_iterators(root_mem_cgroup,
1172 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1173 * @memcg: hierarchy root
1174 * @fn: function to call for each task
1175 * @arg: argument passed to @fn
1177 * This function iterates over tasks attached to @memcg or to any of its
1178 * descendants and calls @fn for each task. If @fn returns a non-zero
1179 * value, the function breaks the iteration loop and returns the value.
1180 * Otherwise, it will iterate over all tasks and return 0.
1182 * This function must not be called for the root memory cgroup.
1184 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1185 int (*fn)(struct task_struct *, void *), void *arg)
1187 struct mem_cgroup *iter;
1190 BUG_ON(memcg == root_mem_cgroup);
1192 for_each_mem_cgroup_tree(iter, memcg) {
1193 struct css_task_iter it;
1194 struct task_struct *task;
1196 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1197 while (!ret && (task = css_task_iter_next(&it)))
1198 ret = fn(task, arg);
1199 css_task_iter_end(&it);
1201 mem_cgroup_iter_break(memcg, iter);
1209 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1211 * @pgdat: pgdat of the page
1213 * This function is only safe when following the LRU page isolation
1214 * and putback protocol: the LRU lock must be held, and the page must
1215 * either be PageLRU() or the caller must have isolated/allocated it.
1217 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1219 struct mem_cgroup_per_node *mz;
1220 struct mem_cgroup *memcg;
1221 struct lruvec *lruvec;
1223 if (mem_cgroup_disabled()) {
1224 lruvec = &pgdat->lruvec;
1228 memcg = page->mem_cgroup;
1230 * Swapcache readahead pages are added to the LRU - and
1231 * possibly migrated - before they are charged.
1234 memcg = root_mem_cgroup;
1236 mz = mem_cgroup_page_nodeinfo(memcg, page);
1237 lruvec = &mz->lruvec;
1240 * Since a node can be onlined after the mem_cgroup was created,
1241 * we have to be prepared to initialize lruvec->zone here;
1242 * and if offlined then reonlined, we need to reinitialize it.
1244 if (unlikely(lruvec->pgdat != pgdat))
1245 lruvec->pgdat = pgdat;
1250 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1251 * @lruvec: mem_cgroup per zone lru vector
1252 * @lru: index of lru list the page is sitting on
1253 * @zid: zone id of the accounted pages
1254 * @nr_pages: positive when adding or negative when removing
1256 * This function must be called under lru_lock, just before a page is added
1257 * to or just after a page is removed from an lru list (that ordering being
1258 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1260 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1261 int zid, int nr_pages)
1263 struct mem_cgroup_per_node *mz;
1264 unsigned long *lru_size;
1267 if (mem_cgroup_disabled())
1270 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1271 lru_size = &mz->lru_zone_size[zid][lru];
1274 *lru_size += nr_pages;
1277 if (WARN_ONCE(size < 0,
1278 "%s(%p, %d, %d): lru_size %ld\n",
1279 __func__, lruvec, lru, nr_pages, size)) {
1285 *lru_size += nr_pages;
1289 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1290 * @memcg: the memory cgroup
1292 * Returns the maximum amount of memory @mem can be charged with, in
1295 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1297 unsigned long margin = 0;
1298 unsigned long count;
1299 unsigned long limit;
1301 count = page_counter_read(&memcg->memory);
1302 limit = READ_ONCE(memcg->memory.max);
1304 margin = limit - count;
1306 if (do_memsw_account()) {
1307 count = page_counter_read(&memcg->memsw);
1308 limit = READ_ONCE(memcg->memsw.max);
1310 margin = min(margin, limit - count);
1319 * A routine for checking "mem" is under move_account() or not.
1321 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1322 * moving cgroups. This is for waiting at high-memory pressure
1325 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1327 struct mem_cgroup *from;
1328 struct mem_cgroup *to;
1331 * Unlike task_move routines, we access mc.to, mc.from not under
1332 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1334 spin_lock(&mc.lock);
1340 ret = mem_cgroup_is_descendant(from, memcg) ||
1341 mem_cgroup_is_descendant(to, memcg);
1343 spin_unlock(&mc.lock);
1347 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1349 if (mc.moving_task && current != mc.moving_task) {
1350 if (mem_cgroup_under_move(memcg)) {
1352 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1353 /* moving charge context might have finished. */
1356 finish_wait(&mc.waitq, &wait);
1363 static char *memory_stat_format(struct mem_cgroup *memcg)
1368 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1373 * Provide statistics on the state of the memory subsystem as
1374 * well as cumulative event counters that show past behavior.
1376 * This list is ordered following a combination of these gradients:
1377 * 1) generic big picture -> specifics and details
1378 * 2) reflecting userspace activity -> reflecting kernel heuristics
1380 * Current memory state:
1383 seq_buf_printf(&s, "anon %llu\n",
1384 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1386 seq_buf_printf(&s, "file %llu\n",
1387 (u64)memcg_page_state(memcg, MEMCG_CACHE) *
1389 seq_buf_printf(&s, "kernel_stack %llu\n",
1390 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1392 seq_buf_printf(&s, "slab %llu\n",
1393 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1394 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1396 seq_buf_printf(&s, "sock %llu\n",
1397 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1400 seq_buf_printf(&s, "shmem %llu\n",
1401 (u64)memcg_page_state(memcg, NR_SHMEM) *
1403 seq_buf_printf(&s, "file_mapped %llu\n",
1404 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1406 seq_buf_printf(&s, "file_dirty %llu\n",
1407 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1409 seq_buf_printf(&s, "file_writeback %llu\n",
1410 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1414 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1415 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1416 * arse because it requires migrating the work out of rmap to a place
1417 * where the page->mem_cgroup is set up and stable.
1419 seq_buf_printf(&s, "anon_thp %llu\n",
1420 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1423 for (i = 0; i < NR_LRU_LISTS; i++)
1424 seq_buf_printf(&s, "%s %llu\n", mem_cgroup_lru_names[i],
1425 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1428 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1429 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1431 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1432 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1435 /* Accumulated memory events */
1437 seq_buf_printf(&s, "pgfault %lu\n", memcg_events(memcg, PGFAULT));
1438 seq_buf_printf(&s, "pgmajfault %lu\n", memcg_events(memcg, PGMAJFAULT));
1440 seq_buf_printf(&s, "workingset_refault %lu\n",
1441 memcg_page_state(memcg, WORKINGSET_REFAULT));
1442 seq_buf_printf(&s, "workingset_activate %lu\n",
1443 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1444 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1445 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1447 seq_buf_printf(&s, "pgrefill %lu\n", memcg_events(memcg, PGREFILL));
1448 seq_buf_printf(&s, "pgscan %lu\n",
1449 memcg_events(memcg, PGSCAN_KSWAPD) +
1450 memcg_events(memcg, PGSCAN_DIRECT));
1451 seq_buf_printf(&s, "pgsteal %lu\n",
1452 memcg_events(memcg, PGSTEAL_KSWAPD) +
1453 memcg_events(memcg, PGSTEAL_DIRECT));
1454 seq_buf_printf(&s, "pgactivate %lu\n", memcg_events(memcg, PGACTIVATE));
1455 seq_buf_printf(&s, "pgdeactivate %lu\n", memcg_events(memcg, PGDEACTIVATE));
1456 seq_buf_printf(&s, "pglazyfree %lu\n", memcg_events(memcg, PGLAZYFREE));
1457 seq_buf_printf(&s, "pglazyfreed %lu\n", memcg_events(memcg, PGLAZYFREED));
1459 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1460 seq_buf_printf(&s, "thp_fault_alloc %lu\n",
1461 memcg_events(memcg, THP_FAULT_ALLOC));
1462 seq_buf_printf(&s, "thp_collapse_alloc %lu\n",
1463 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1464 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1466 /* The above should easily fit into one page */
1467 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1472 #define K(x) ((x) << (PAGE_SHIFT-10))
1474 * mem_cgroup_print_oom_context: Print OOM information relevant to
1475 * memory controller.
1476 * @memcg: The memory cgroup that went over limit
1477 * @p: Task that is going to be killed
1479 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1482 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1487 pr_cont(",oom_memcg=");
1488 pr_cont_cgroup_path(memcg->css.cgroup);
1490 pr_cont(",global_oom");
1492 pr_cont(",task_memcg=");
1493 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1499 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1500 * memory controller.
1501 * @memcg: The memory cgroup that went over limit
1503 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1507 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1508 K((u64)page_counter_read(&memcg->memory)),
1509 K((u64)memcg->memory.max), memcg->memory.failcnt);
1510 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1511 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1512 K((u64)page_counter_read(&memcg->swap)),
1513 K((u64)memcg->swap.max), memcg->swap.failcnt);
1515 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1516 K((u64)page_counter_read(&memcg->memsw)),
1517 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1518 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1519 K((u64)page_counter_read(&memcg->kmem)),
1520 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1523 pr_info("Memory cgroup stats for ");
1524 pr_cont_cgroup_path(memcg->css.cgroup);
1526 buf = memory_stat_format(memcg);
1534 * Return the memory (and swap, if configured) limit for a memcg.
1536 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1540 max = memcg->memory.max;
1541 if (mem_cgroup_swappiness(memcg)) {
1542 unsigned long memsw_max;
1543 unsigned long swap_max;
1545 memsw_max = memcg->memsw.max;
1546 swap_max = memcg->swap.max;
1547 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1548 max = min(max + swap_max, memsw_max);
1553 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1555 return page_counter_read(&memcg->memory);
1558 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1561 struct oom_control oc = {
1565 .gfp_mask = gfp_mask,
1570 if (mutex_lock_killable(&oom_lock))
1573 * A few threads which were not waiting at mutex_lock_killable() can
1574 * fail to bail out. Therefore, check again after holding oom_lock.
1576 ret = should_force_charge() || out_of_memory(&oc);
1577 mutex_unlock(&oom_lock);
1581 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1584 unsigned long *total_scanned)
1586 struct mem_cgroup *victim = NULL;
1589 unsigned long excess;
1590 unsigned long nr_scanned;
1591 struct mem_cgroup_reclaim_cookie reclaim = {
1595 excess = soft_limit_excess(root_memcg);
1598 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1603 * If we have not been able to reclaim
1604 * anything, it might because there are
1605 * no reclaimable pages under this hierarchy
1610 * We want to do more targeted reclaim.
1611 * excess >> 2 is not to excessive so as to
1612 * reclaim too much, nor too less that we keep
1613 * coming back to reclaim from this cgroup
1615 if (total >= (excess >> 2) ||
1616 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1621 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1622 pgdat, &nr_scanned);
1623 *total_scanned += nr_scanned;
1624 if (!soft_limit_excess(root_memcg))
1627 mem_cgroup_iter_break(root_memcg, victim);
1631 #ifdef CONFIG_LOCKDEP
1632 static struct lockdep_map memcg_oom_lock_dep_map = {
1633 .name = "memcg_oom_lock",
1637 static DEFINE_SPINLOCK(memcg_oom_lock);
1640 * Check OOM-Killer is already running under our hierarchy.
1641 * If someone is running, return false.
1643 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1645 struct mem_cgroup *iter, *failed = NULL;
1647 spin_lock(&memcg_oom_lock);
1649 for_each_mem_cgroup_tree(iter, memcg) {
1650 if (iter->oom_lock) {
1652 * this subtree of our hierarchy is already locked
1653 * so we cannot give a lock.
1656 mem_cgroup_iter_break(memcg, iter);
1659 iter->oom_lock = true;
1664 * OK, we failed to lock the whole subtree so we have
1665 * to clean up what we set up to the failing subtree
1667 for_each_mem_cgroup_tree(iter, memcg) {
1668 if (iter == failed) {
1669 mem_cgroup_iter_break(memcg, iter);
1672 iter->oom_lock = false;
1675 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1677 spin_unlock(&memcg_oom_lock);
1682 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1684 struct mem_cgroup *iter;
1686 spin_lock(&memcg_oom_lock);
1687 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1688 for_each_mem_cgroup_tree(iter, memcg)
1689 iter->oom_lock = false;
1690 spin_unlock(&memcg_oom_lock);
1693 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1695 struct mem_cgroup *iter;
1697 spin_lock(&memcg_oom_lock);
1698 for_each_mem_cgroup_tree(iter, memcg)
1700 spin_unlock(&memcg_oom_lock);
1703 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1705 struct mem_cgroup *iter;
1708 * When a new child is created while the hierarchy is under oom,
1709 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1711 spin_lock(&memcg_oom_lock);
1712 for_each_mem_cgroup_tree(iter, memcg)
1713 if (iter->under_oom > 0)
1715 spin_unlock(&memcg_oom_lock);
1718 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1720 struct oom_wait_info {
1721 struct mem_cgroup *memcg;
1722 wait_queue_entry_t wait;
1725 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1726 unsigned mode, int sync, void *arg)
1728 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1729 struct mem_cgroup *oom_wait_memcg;
1730 struct oom_wait_info *oom_wait_info;
1732 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1733 oom_wait_memcg = oom_wait_info->memcg;
1735 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1736 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1738 return autoremove_wake_function(wait, mode, sync, arg);
1741 static void memcg_oom_recover(struct mem_cgroup *memcg)
1744 * For the following lockless ->under_oom test, the only required
1745 * guarantee is that it must see the state asserted by an OOM when
1746 * this function is called as a result of userland actions
1747 * triggered by the notification of the OOM. This is trivially
1748 * achieved by invoking mem_cgroup_mark_under_oom() before
1749 * triggering notification.
1751 if (memcg && memcg->under_oom)
1752 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1762 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1764 enum oom_status ret;
1767 if (order > PAGE_ALLOC_COSTLY_ORDER)
1770 memcg_memory_event(memcg, MEMCG_OOM);
1773 * We are in the middle of the charge context here, so we
1774 * don't want to block when potentially sitting on a callstack
1775 * that holds all kinds of filesystem and mm locks.
1777 * cgroup1 allows disabling the OOM killer and waiting for outside
1778 * handling until the charge can succeed; remember the context and put
1779 * the task to sleep at the end of the page fault when all locks are
1782 * On the other hand, in-kernel OOM killer allows for an async victim
1783 * memory reclaim (oom_reaper) and that means that we are not solely
1784 * relying on the oom victim to make a forward progress and we can
1785 * invoke the oom killer here.
1787 * Please note that mem_cgroup_out_of_memory might fail to find a
1788 * victim and then we have to bail out from the charge path.
1790 if (memcg->oom_kill_disable) {
1791 if (!current->in_user_fault)
1793 css_get(&memcg->css);
1794 current->memcg_in_oom = memcg;
1795 current->memcg_oom_gfp_mask = mask;
1796 current->memcg_oom_order = order;
1801 mem_cgroup_mark_under_oom(memcg);
1803 locked = mem_cgroup_oom_trylock(memcg);
1806 mem_cgroup_oom_notify(memcg);
1808 mem_cgroup_unmark_under_oom(memcg);
1809 if (mem_cgroup_out_of_memory(memcg, mask, order))
1815 mem_cgroup_oom_unlock(memcg);
1821 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1822 * @handle: actually kill/wait or just clean up the OOM state
1824 * This has to be called at the end of a page fault if the memcg OOM
1825 * handler was enabled.
1827 * Memcg supports userspace OOM handling where failed allocations must
1828 * sleep on a waitqueue until the userspace task resolves the
1829 * situation. Sleeping directly in the charge context with all kinds
1830 * of locks held is not a good idea, instead we remember an OOM state
1831 * in the task and mem_cgroup_oom_synchronize() has to be called at
1832 * the end of the page fault to complete the OOM handling.
1834 * Returns %true if an ongoing memcg OOM situation was detected and
1835 * completed, %false otherwise.
1837 bool mem_cgroup_oom_synchronize(bool handle)
1839 struct mem_cgroup *memcg = current->memcg_in_oom;
1840 struct oom_wait_info owait;
1843 /* OOM is global, do not handle */
1850 owait.memcg = memcg;
1851 owait.wait.flags = 0;
1852 owait.wait.func = memcg_oom_wake_function;
1853 owait.wait.private = current;
1854 INIT_LIST_HEAD(&owait.wait.entry);
1856 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1857 mem_cgroup_mark_under_oom(memcg);
1859 locked = mem_cgroup_oom_trylock(memcg);
1862 mem_cgroup_oom_notify(memcg);
1864 if (locked && !memcg->oom_kill_disable) {
1865 mem_cgroup_unmark_under_oom(memcg);
1866 finish_wait(&memcg_oom_waitq, &owait.wait);
1867 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1868 current->memcg_oom_order);
1871 mem_cgroup_unmark_under_oom(memcg);
1872 finish_wait(&memcg_oom_waitq, &owait.wait);
1876 mem_cgroup_oom_unlock(memcg);
1878 * There is no guarantee that an OOM-lock contender
1879 * sees the wakeups triggered by the OOM kill
1880 * uncharges. Wake any sleepers explicitely.
1882 memcg_oom_recover(memcg);
1885 current->memcg_in_oom = NULL;
1886 css_put(&memcg->css);
1891 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1892 * @victim: task to be killed by the OOM killer
1893 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1895 * Returns a pointer to a memory cgroup, which has to be cleaned up
1896 * by killing all belonging OOM-killable tasks.
1898 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1900 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1901 struct mem_cgroup *oom_domain)
1903 struct mem_cgroup *oom_group = NULL;
1904 struct mem_cgroup *memcg;
1906 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1910 oom_domain = root_mem_cgroup;
1914 memcg = mem_cgroup_from_task(victim);
1915 if (memcg == root_mem_cgroup)
1919 * Traverse the memory cgroup hierarchy from the victim task's
1920 * cgroup up to the OOMing cgroup (or root) to find the
1921 * highest-level memory cgroup with oom.group set.
1923 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1924 if (memcg->oom_group)
1927 if (memcg == oom_domain)
1932 css_get(&oom_group->css);
1939 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1941 pr_info("Tasks in ");
1942 pr_cont_cgroup_path(memcg->css.cgroup);
1943 pr_cont(" are going to be killed due to memory.oom.group set\n");
1947 * lock_page_memcg - lock a page->mem_cgroup binding
1950 * This function protects unlocked LRU pages from being moved to
1953 * It ensures lifetime of the returned memcg. Caller is responsible
1954 * for the lifetime of the page; __unlock_page_memcg() is available
1955 * when @page might get freed inside the locked section.
1957 struct mem_cgroup *lock_page_memcg(struct page *page)
1959 struct mem_cgroup *memcg;
1960 unsigned long flags;
1963 * The RCU lock is held throughout the transaction. The fast
1964 * path can get away without acquiring the memcg->move_lock
1965 * because page moving starts with an RCU grace period.
1967 * The RCU lock also protects the memcg from being freed when
1968 * the page state that is going to change is the only thing
1969 * preventing the page itself from being freed. E.g. writeback
1970 * doesn't hold a page reference and relies on PG_writeback to
1971 * keep off truncation, migration and so forth.
1975 if (mem_cgroup_disabled())
1978 memcg = page->mem_cgroup;
1979 if (unlikely(!memcg))
1982 if (atomic_read(&memcg->moving_account) <= 0)
1985 spin_lock_irqsave(&memcg->move_lock, flags);
1986 if (memcg != page->mem_cgroup) {
1987 spin_unlock_irqrestore(&memcg->move_lock, flags);
1992 * When charge migration first begins, we can have locked and
1993 * unlocked page stat updates happening concurrently. Track
1994 * the task who has the lock for unlock_page_memcg().
1996 memcg->move_lock_task = current;
1997 memcg->move_lock_flags = flags;
2001 EXPORT_SYMBOL(lock_page_memcg);
2004 * __unlock_page_memcg - unlock and unpin a memcg
2007 * Unlock and unpin a memcg returned by lock_page_memcg().
2009 void __unlock_page_memcg(struct mem_cgroup *memcg)
2011 if (memcg && memcg->move_lock_task == current) {
2012 unsigned long flags = memcg->move_lock_flags;
2014 memcg->move_lock_task = NULL;
2015 memcg->move_lock_flags = 0;
2017 spin_unlock_irqrestore(&memcg->move_lock, flags);
2024 * unlock_page_memcg - unlock a page->mem_cgroup binding
2027 void unlock_page_memcg(struct page *page)
2029 __unlock_page_memcg(page->mem_cgroup);
2031 EXPORT_SYMBOL(unlock_page_memcg);
2033 struct memcg_stock_pcp {
2034 struct mem_cgroup *cached; /* this never be root cgroup */
2035 unsigned int nr_pages;
2036 struct work_struct work;
2037 unsigned long flags;
2038 #define FLUSHING_CACHED_CHARGE 0
2040 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2041 static DEFINE_MUTEX(percpu_charge_mutex);
2044 * consume_stock: Try to consume stocked charge on this cpu.
2045 * @memcg: memcg to consume from.
2046 * @nr_pages: how many pages to charge.
2048 * The charges will only happen if @memcg matches the current cpu's memcg
2049 * stock, and at least @nr_pages are available in that stock. Failure to
2050 * service an allocation will refill the stock.
2052 * returns true if successful, false otherwise.
2054 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2056 struct memcg_stock_pcp *stock;
2057 unsigned long flags;
2060 if (nr_pages > MEMCG_CHARGE_BATCH)
2063 local_irq_save(flags);
2065 stock = this_cpu_ptr(&memcg_stock);
2066 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2067 stock->nr_pages -= nr_pages;
2071 local_irq_restore(flags);
2077 * Returns stocks cached in percpu and reset cached information.
2079 static void drain_stock(struct memcg_stock_pcp *stock)
2081 struct mem_cgroup *old = stock->cached;
2083 if (stock->nr_pages) {
2084 page_counter_uncharge(&old->memory, stock->nr_pages);
2085 if (do_memsw_account())
2086 page_counter_uncharge(&old->memsw, stock->nr_pages);
2087 css_put_many(&old->css, stock->nr_pages);
2088 stock->nr_pages = 0;
2090 stock->cached = NULL;
2093 static void drain_local_stock(struct work_struct *dummy)
2095 struct memcg_stock_pcp *stock;
2096 unsigned long flags;
2099 * The only protection from memory hotplug vs. drain_stock races is
2100 * that we always operate on local CPU stock here with IRQ disabled
2102 local_irq_save(flags);
2104 stock = this_cpu_ptr(&memcg_stock);
2106 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2108 local_irq_restore(flags);
2112 * Cache charges(val) to local per_cpu area.
2113 * This will be consumed by consume_stock() function, later.
2115 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2117 struct memcg_stock_pcp *stock;
2118 unsigned long flags;
2120 local_irq_save(flags);
2122 stock = this_cpu_ptr(&memcg_stock);
2123 if (stock->cached != memcg) { /* reset if necessary */
2125 stock->cached = memcg;
2127 stock->nr_pages += nr_pages;
2129 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2132 local_irq_restore(flags);
2136 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2137 * of the hierarchy under it.
2139 static void drain_all_stock(struct mem_cgroup *root_memcg)
2143 /* If someone's already draining, avoid adding running more workers. */
2144 if (!mutex_trylock(&percpu_charge_mutex))
2147 * Notify other cpus that system-wide "drain" is running
2148 * We do not care about races with the cpu hotplug because cpu down
2149 * as well as workers from this path always operate on the local
2150 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2153 for_each_online_cpu(cpu) {
2154 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2155 struct mem_cgroup *memcg;
2159 memcg = stock->cached;
2160 if (memcg && stock->nr_pages &&
2161 mem_cgroup_is_descendant(memcg, root_memcg))
2166 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2168 drain_local_stock(&stock->work);
2170 schedule_work_on(cpu, &stock->work);
2174 mutex_unlock(&percpu_charge_mutex);
2177 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2179 struct memcg_stock_pcp *stock;
2180 struct mem_cgroup *memcg, *mi;
2182 stock = &per_cpu(memcg_stock, cpu);
2185 for_each_mem_cgroup(memcg) {
2188 for (i = 0; i < MEMCG_NR_STAT; i++) {
2192 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2194 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2195 atomic_long_add(x, &memcg->vmstats[i]);
2197 if (i >= NR_VM_NODE_STAT_ITEMS)
2200 for_each_node(nid) {
2201 struct mem_cgroup_per_node *pn;
2203 pn = mem_cgroup_nodeinfo(memcg, nid);
2204 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2207 atomic_long_add(x, &pn->lruvec_stat[i]);
2208 } while ((pn = parent_nodeinfo(pn, nid)));
2212 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2215 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2217 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2218 atomic_long_add(x, &memcg->vmevents[i]);
2225 static void reclaim_high(struct mem_cgroup *memcg,
2226 unsigned int nr_pages,
2230 if (page_counter_read(&memcg->memory) <= memcg->high)
2232 memcg_memory_event(memcg, MEMCG_HIGH);
2233 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2234 } while ((memcg = parent_mem_cgroup(memcg)));
2237 static void high_work_func(struct work_struct *work)
2239 struct mem_cgroup *memcg;
2241 memcg = container_of(work, struct mem_cgroup, high_work);
2242 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2246 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2247 * enough to still cause a significant slowdown in most cases, while still
2248 * allowing diagnostics and tracing to proceed without becoming stuck.
2250 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2253 * When calculating the delay, we use these either side of the exponentiation to
2254 * maintain precision and scale to a reasonable number of jiffies (see the table
2257 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2258 * overage ratio to a delay.
2259 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2260 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2261 * to produce a reasonable delay curve.
2263 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2264 * reasonable delay curve compared to precision-adjusted overage, not
2265 * penalising heavily at first, but still making sure that growth beyond the
2266 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2267 * example, with a high of 100 megabytes:
2269 * +-------+------------------------+
2270 * | usage | time to allocate in ms |
2271 * +-------+------------------------+
2293 * +-------+------------------------+
2295 #define MEMCG_DELAY_PRECISION_SHIFT 20
2296 #define MEMCG_DELAY_SCALING_SHIFT 14
2299 * Scheduled by try_charge() to be executed from the userland return path
2300 * and reclaims memory over the high limit.
2302 void mem_cgroup_handle_over_high(void)
2304 unsigned long usage, high, clamped_high;
2305 unsigned long pflags;
2306 unsigned long penalty_jiffies, overage;
2307 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2308 struct mem_cgroup *memcg;
2310 if (likely(!nr_pages))
2313 memcg = get_mem_cgroup_from_mm(current->mm);
2314 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2315 current->memcg_nr_pages_over_high = 0;
2318 * memory.high is breached and reclaim is unable to keep up. Throttle
2319 * allocators proactively to slow down excessive growth.
2321 * We use overage compared to memory.high to calculate the number of
2322 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2323 * fairly lenient on small overages, and increasingly harsh when the
2324 * memcg in question makes it clear that it has no intention of stopping
2325 * its crazy behaviour, so we exponentially increase the delay based on
2329 usage = page_counter_read(&memcg->memory);
2330 high = READ_ONCE(memcg->high);
2336 * Prevent division by 0 in overage calculation by acting as if it was a
2337 * threshold of 1 page
2339 clamped_high = max(high, 1UL);
2341 overage = div_u64((u64)(usage - high) << MEMCG_DELAY_PRECISION_SHIFT,
2344 penalty_jiffies = ((u64)overage * overage * HZ)
2345 >> (MEMCG_DELAY_PRECISION_SHIFT + MEMCG_DELAY_SCALING_SHIFT);
2348 * Factor in the task's own contribution to the overage, such that four
2349 * N-sized allocations are throttled approximately the same as one
2350 * 4N-sized allocation.
2352 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2353 * larger the current charge patch is than that.
2355 penalty_jiffies = penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2358 * Clamp the max delay per usermode return so as to still keep the
2359 * application moving forwards and also permit diagnostics, albeit
2362 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2365 * Don't sleep if the amount of jiffies this memcg owes us is so low
2366 * that it's not even worth doing, in an attempt to be nice to those who
2367 * go only a small amount over their memory.high value and maybe haven't
2368 * been aggressively reclaimed enough yet.
2370 if (penalty_jiffies <= HZ / 100)
2374 * If we exit early, we're guaranteed to die (since
2375 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2376 * need to account for any ill-begotten jiffies to pay them off later.
2378 psi_memstall_enter(&pflags);
2379 schedule_timeout_killable(penalty_jiffies);
2380 psi_memstall_leave(&pflags);
2383 css_put(&memcg->css);
2386 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2387 unsigned int nr_pages)
2389 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2390 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2391 struct mem_cgroup *mem_over_limit;
2392 struct page_counter *counter;
2393 unsigned long nr_reclaimed;
2394 bool may_swap = true;
2395 bool drained = false;
2396 enum oom_status oom_status;
2398 if (mem_cgroup_is_root(memcg))
2401 if (consume_stock(memcg, nr_pages))
2404 if (!do_memsw_account() ||
2405 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2406 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2408 if (do_memsw_account())
2409 page_counter_uncharge(&memcg->memsw, batch);
2410 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2412 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2416 if (batch > nr_pages) {
2422 * Memcg doesn't have a dedicated reserve for atomic
2423 * allocations. But like the global atomic pool, we need to
2424 * put the burden of reclaim on regular allocation requests
2425 * and let these go through as privileged allocations.
2427 if (gfp_mask & __GFP_ATOMIC)
2431 * Unlike in global OOM situations, memcg is not in a physical
2432 * memory shortage. Allow dying and OOM-killed tasks to
2433 * bypass the last charges so that they can exit quickly and
2434 * free their memory.
2436 if (unlikely(should_force_charge()))
2440 * Prevent unbounded recursion when reclaim operations need to
2441 * allocate memory. This might exceed the limits temporarily,
2442 * but we prefer facilitating memory reclaim and getting back
2443 * under the limit over triggering OOM kills in these cases.
2445 if (unlikely(current->flags & PF_MEMALLOC))
2448 if (unlikely(task_in_memcg_oom(current)))
2451 if (!gfpflags_allow_blocking(gfp_mask))
2454 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2456 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2457 gfp_mask, may_swap);
2459 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2463 drain_all_stock(mem_over_limit);
2468 if (gfp_mask & __GFP_NORETRY)
2471 * Even though the limit is exceeded at this point, reclaim
2472 * may have been able to free some pages. Retry the charge
2473 * before killing the task.
2475 * Only for regular pages, though: huge pages are rather
2476 * unlikely to succeed so close to the limit, and we fall back
2477 * to regular pages anyway in case of failure.
2479 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2482 * At task move, charge accounts can be doubly counted. So, it's
2483 * better to wait until the end of task_move if something is going on.
2485 if (mem_cgroup_wait_acct_move(mem_over_limit))
2491 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2494 if (gfp_mask & __GFP_NOFAIL)
2497 if (fatal_signal_pending(current))
2501 * keep retrying as long as the memcg oom killer is able to make
2502 * a forward progress or bypass the charge if the oom killer
2503 * couldn't make any progress.
2505 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2506 get_order(nr_pages * PAGE_SIZE));
2507 switch (oom_status) {
2509 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2517 if (!(gfp_mask & __GFP_NOFAIL))
2521 * The allocation either can't fail or will lead to more memory
2522 * being freed very soon. Allow memory usage go over the limit
2523 * temporarily by force charging it.
2525 page_counter_charge(&memcg->memory, nr_pages);
2526 if (do_memsw_account())
2527 page_counter_charge(&memcg->memsw, nr_pages);
2528 css_get_many(&memcg->css, nr_pages);
2533 css_get_many(&memcg->css, batch);
2534 if (batch > nr_pages)
2535 refill_stock(memcg, batch - nr_pages);
2538 * If the hierarchy is above the normal consumption range, schedule
2539 * reclaim on returning to userland. We can perform reclaim here
2540 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2541 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2542 * not recorded as it most likely matches current's and won't
2543 * change in the meantime. As high limit is checked again before
2544 * reclaim, the cost of mismatch is negligible.
2547 if (page_counter_read(&memcg->memory) > memcg->high) {
2548 /* Don't bother a random interrupted task */
2549 if (in_interrupt()) {
2550 schedule_work(&memcg->high_work);
2553 current->memcg_nr_pages_over_high += batch;
2554 set_notify_resume(current);
2557 } while ((memcg = parent_mem_cgroup(memcg)));
2562 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2564 if (mem_cgroup_is_root(memcg))
2567 page_counter_uncharge(&memcg->memory, nr_pages);
2568 if (do_memsw_account())
2569 page_counter_uncharge(&memcg->memsw, nr_pages);
2571 css_put_many(&memcg->css, nr_pages);
2574 static void lock_page_lru(struct page *page, int *isolated)
2576 pg_data_t *pgdat = page_pgdat(page);
2578 spin_lock_irq(&pgdat->lru_lock);
2579 if (PageLRU(page)) {
2580 struct lruvec *lruvec;
2582 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2584 del_page_from_lru_list(page, lruvec, page_lru(page));
2590 static void unlock_page_lru(struct page *page, int isolated)
2592 pg_data_t *pgdat = page_pgdat(page);
2595 struct lruvec *lruvec;
2597 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2598 VM_BUG_ON_PAGE(PageLRU(page), page);
2600 add_page_to_lru_list(page, lruvec, page_lru(page));
2602 spin_unlock_irq(&pgdat->lru_lock);
2605 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2610 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2613 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2614 * may already be on some other mem_cgroup's LRU. Take care of it.
2617 lock_page_lru(page, &isolated);
2620 * Nobody should be changing or seriously looking at
2621 * page->mem_cgroup at this point:
2623 * - the page is uncharged
2625 * - the page is off-LRU
2627 * - an anonymous fault has exclusive page access, except for
2628 * a locked page table
2630 * - a page cache insertion, a swapin fault, or a migration
2631 * have the page locked
2633 page->mem_cgroup = memcg;
2636 unlock_page_lru(page, isolated);
2639 #ifdef CONFIG_MEMCG_KMEM
2640 static int memcg_alloc_cache_id(void)
2645 id = ida_simple_get(&memcg_cache_ida,
2646 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2650 if (id < memcg_nr_cache_ids)
2654 * There's no space for the new id in memcg_caches arrays,
2655 * so we have to grow them.
2657 down_write(&memcg_cache_ids_sem);
2659 size = 2 * (id + 1);
2660 if (size < MEMCG_CACHES_MIN_SIZE)
2661 size = MEMCG_CACHES_MIN_SIZE;
2662 else if (size > MEMCG_CACHES_MAX_SIZE)
2663 size = MEMCG_CACHES_MAX_SIZE;
2665 err = memcg_update_all_caches(size);
2667 err = memcg_update_all_list_lrus(size);
2669 memcg_nr_cache_ids = size;
2671 up_write(&memcg_cache_ids_sem);
2674 ida_simple_remove(&memcg_cache_ida, id);
2680 static void memcg_free_cache_id(int id)
2682 ida_simple_remove(&memcg_cache_ida, id);
2685 struct memcg_kmem_cache_create_work {
2686 struct mem_cgroup *memcg;
2687 struct kmem_cache *cachep;
2688 struct work_struct work;
2691 static void memcg_kmem_cache_create_func(struct work_struct *w)
2693 struct memcg_kmem_cache_create_work *cw =
2694 container_of(w, struct memcg_kmem_cache_create_work, work);
2695 struct mem_cgroup *memcg = cw->memcg;
2696 struct kmem_cache *cachep = cw->cachep;
2698 memcg_create_kmem_cache(memcg, cachep);
2700 css_put(&memcg->css);
2705 * Enqueue the creation of a per-memcg kmem_cache.
2707 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2708 struct kmem_cache *cachep)
2710 struct memcg_kmem_cache_create_work *cw;
2712 if (!css_tryget_online(&memcg->css))
2715 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2720 cw->cachep = cachep;
2721 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2723 queue_work(memcg_kmem_cache_wq, &cw->work);
2726 static inline bool memcg_kmem_bypass(void)
2728 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
2734 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2735 * @cachep: the original global kmem cache
2737 * Return the kmem_cache we're supposed to use for a slab allocation.
2738 * We try to use the current memcg's version of the cache.
2740 * If the cache does not exist yet, if we are the first user of it, we
2741 * create it asynchronously in a workqueue and let the current allocation
2742 * go through with the original cache.
2744 * This function takes a reference to the cache it returns to assure it
2745 * won't get destroyed while we are working with it. Once the caller is
2746 * done with it, memcg_kmem_put_cache() must be called to release the
2749 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2751 struct mem_cgroup *memcg;
2752 struct kmem_cache *memcg_cachep;
2753 struct memcg_cache_array *arr;
2756 VM_BUG_ON(!is_root_cache(cachep));
2758 if (memcg_kmem_bypass())
2763 if (unlikely(current->active_memcg))
2764 memcg = current->active_memcg;
2766 memcg = mem_cgroup_from_task(current);
2768 if (!memcg || memcg == root_mem_cgroup)
2771 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2775 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2778 * Make sure we will access the up-to-date value. The code updating
2779 * memcg_caches issues a write barrier to match the data dependency
2780 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2782 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2785 * If we are in a safe context (can wait, and not in interrupt
2786 * context), we could be be predictable and return right away.
2787 * This would guarantee that the allocation being performed
2788 * already belongs in the new cache.
2790 * However, there are some clashes that can arrive from locking.
2791 * For instance, because we acquire the slab_mutex while doing
2792 * memcg_create_kmem_cache, this means no further allocation
2793 * could happen with the slab_mutex held. So it's better to
2796 * If the memcg is dying or memcg_cache is about to be released,
2797 * don't bother creating new kmem_caches. Because memcg_cachep
2798 * is ZEROed as the fist step of kmem offlining, we don't need
2799 * percpu_ref_tryget_live() here. css_tryget_online() check in
2800 * memcg_schedule_kmem_cache_create() will prevent us from
2801 * creation of a new kmem_cache.
2803 if (unlikely(!memcg_cachep))
2804 memcg_schedule_kmem_cache_create(memcg, cachep);
2805 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2806 cachep = memcg_cachep;
2813 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2814 * @cachep: the cache returned by memcg_kmem_get_cache
2816 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2818 if (!is_root_cache(cachep))
2819 percpu_ref_put(&cachep->memcg_params.refcnt);
2823 * __memcg_kmem_charge_memcg: charge a kmem page
2824 * @page: page to charge
2825 * @gfp: reclaim mode
2826 * @order: allocation order
2827 * @memcg: memory cgroup to charge
2829 * Returns 0 on success, an error code on failure.
2831 int __memcg_kmem_charge_memcg(struct page *page, gfp_t gfp, int order,
2832 struct mem_cgroup *memcg)
2834 unsigned int nr_pages = 1 << order;
2835 struct page_counter *counter;
2838 ret = try_charge(memcg, gfp, nr_pages);
2842 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2843 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2846 * Enforce __GFP_NOFAIL allocation because callers are not
2847 * prepared to see failures and likely do not have any failure
2850 if (gfp & __GFP_NOFAIL) {
2851 page_counter_charge(&memcg->kmem, nr_pages);
2854 cancel_charge(memcg, nr_pages);
2861 * __memcg_kmem_charge: charge a kmem page to the current memory cgroup
2862 * @page: page to charge
2863 * @gfp: reclaim mode
2864 * @order: allocation order
2866 * Returns 0 on success, an error code on failure.
2868 int __memcg_kmem_charge(struct page *page, gfp_t gfp, int order)
2870 struct mem_cgroup *memcg;
2873 if (memcg_kmem_bypass())
2876 memcg = get_mem_cgroup_from_current();
2877 if (!mem_cgroup_is_root(memcg)) {
2878 ret = __memcg_kmem_charge_memcg(page, gfp, order, memcg);
2880 page->mem_cgroup = memcg;
2881 __SetPageKmemcg(page);
2884 css_put(&memcg->css);
2889 * __memcg_kmem_uncharge_memcg: uncharge a kmem page
2890 * @memcg: memcg to uncharge
2891 * @nr_pages: number of pages to uncharge
2893 void __memcg_kmem_uncharge_memcg(struct mem_cgroup *memcg,
2894 unsigned int nr_pages)
2896 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2897 page_counter_uncharge(&memcg->kmem, nr_pages);
2899 page_counter_uncharge(&memcg->memory, nr_pages);
2900 if (do_memsw_account())
2901 page_counter_uncharge(&memcg->memsw, nr_pages);
2904 * __memcg_kmem_uncharge: uncharge a kmem page
2905 * @page: page to uncharge
2906 * @order: allocation order
2908 void __memcg_kmem_uncharge(struct page *page, int order)
2910 struct mem_cgroup *memcg = page->mem_cgroup;
2911 unsigned int nr_pages = 1 << order;
2916 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2917 __memcg_kmem_uncharge_memcg(memcg, nr_pages);
2918 page->mem_cgroup = NULL;
2920 /* slab pages do not have PageKmemcg flag set */
2921 if (PageKmemcg(page))
2922 __ClearPageKmemcg(page);
2924 css_put_many(&memcg->css, nr_pages);
2926 #endif /* CONFIG_MEMCG_KMEM */
2928 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2931 * Because tail pages are not marked as "used", set it. We're under
2932 * pgdat->lru_lock and migration entries setup in all page mappings.
2934 void mem_cgroup_split_huge_fixup(struct page *head)
2938 if (mem_cgroup_disabled())
2941 for (i = 1; i < HPAGE_PMD_NR; i++)
2942 head[i].mem_cgroup = head->mem_cgroup;
2944 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
2946 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2948 #ifdef CONFIG_MEMCG_SWAP
2950 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2951 * @entry: swap entry to be moved
2952 * @from: mem_cgroup which the entry is moved from
2953 * @to: mem_cgroup which the entry is moved to
2955 * It succeeds only when the swap_cgroup's record for this entry is the same
2956 * as the mem_cgroup's id of @from.
2958 * Returns 0 on success, -EINVAL on failure.
2960 * The caller must have charged to @to, IOW, called page_counter_charge() about
2961 * both res and memsw, and called css_get().
2963 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2964 struct mem_cgroup *from, struct mem_cgroup *to)
2966 unsigned short old_id, new_id;
2968 old_id = mem_cgroup_id(from);
2969 new_id = mem_cgroup_id(to);
2971 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2972 mod_memcg_state(from, MEMCG_SWAP, -1);
2973 mod_memcg_state(to, MEMCG_SWAP, 1);
2979 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2980 struct mem_cgroup *from, struct mem_cgroup *to)
2986 static DEFINE_MUTEX(memcg_max_mutex);
2988 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
2989 unsigned long max, bool memsw)
2991 bool enlarge = false;
2992 bool drained = false;
2994 bool limits_invariant;
2995 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
2998 if (signal_pending(current)) {
3003 mutex_lock(&memcg_max_mutex);
3005 * Make sure that the new limit (memsw or memory limit) doesn't
3006 * break our basic invariant rule memory.max <= memsw.max.
3008 limits_invariant = memsw ? max >= memcg->memory.max :
3009 max <= memcg->memsw.max;
3010 if (!limits_invariant) {
3011 mutex_unlock(&memcg_max_mutex);
3015 if (max > counter->max)
3017 ret = page_counter_set_max(counter, max);
3018 mutex_unlock(&memcg_max_mutex);
3024 drain_all_stock(memcg);
3029 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3030 GFP_KERNEL, !memsw)) {
3036 if (!ret && enlarge)
3037 memcg_oom_recover(memcg);
3042 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3044 unsigned long *total_scanned)
3046 unsigned long nr_reclaimed = 0;
3047 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3048 unsigned long reclaimed;
3050 struct mem_cgroup_tree_per_node *mctz;
3051 unsigned long excess;
3052 unsigned long nr_scanned;
3057 mctz = soft_limit_tree_node(pgdat->node_id);
3060 * Do not even bother to check the largest node if the root
3061 * is empty. Do it lockless to prevent lock bouncing. Races
3062 * are acceptable as soft limit is best effort anyway.
3064 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3068 * This loop can run a while, specially if mem_cgroup's continuously
3069 * keep exceeding their soft limit and putting the system under
3076 mz = mem_cgroup_largest_soft_limit_node(mctz);
3081 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3082 gfp_mask, &nr_scanned);
3083 nr_reclaimed += reclaimed;
3084 *total_scanned += nr_scanned;
3085 spin_lock_irq(&mctz->lock);
3086 __mem_cgroup_remove_exceeded(mz, mctz);
3089 * If we failed to reclaim anything from this memory cgroup
3090 * it is time to move on to the next cgroup
3094 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3096 excess = soft_limit_excess(mz->memcg);
3098 * One school of thought says that we should not add
3099 * back the node to the tree if reclaim returns 0.
3100 * But our reclaim could return 0, simply because due
3101 * to priority we are exposing a smaller subset of
3102 * memory to reclaim from. Consider this as a longer
3105 /* If excess == 0, no tree ops */
3106 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3107 spin_unlock_irq(&mctz->lock);
3108 css_put(&mz->memcg->css);
3111 * Could not reclaim anything and there are no more
3112 * mem cgroups to try or we seem to be looping without
3113 * reclaiming anything.
3115 if (!nr_reclaimed &&
3117 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3119 } while (!nr_reclaimed);
3121 css_put(&next_mz->memcg->css);
3122 return nr_reclaimed;
3126 * Test whether @memcg has children, dead or alive. Note that this
3127 * function doesn't care whether @memcg has use_hierarchy enabled and
3128 * returns %true if there are child csses according to the cgroup
3129 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3131 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3136 ret = css_next_child(NULL, &memcg->css);
3142 * Reclaims as many pages from the given memcg as possible.
3144 * Caller is responsible for holding css reference for memcg.
3146 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3148 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3150 /* we call try-to-free pages for make this cgroup empty */
3151 lru_add_drain_all();
3153 drain_all_stock(memcg);
3155 /* try to free all pages in this cgroup */
3156 while (nr_retries && page_counter_read(&memcg->memory)) {
3159 if (signal_pending(current))
3162 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3166 /* maybe some writeback is necessary */
3167 congestion_wait(BLK_RW_ASYNC, HZ/10);
3175 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3176 char *buf, size_t nbytes,
3179 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3181 if (mem_cgroup_is_root(memcg))
3183 return mem_cgroup_force_empty(memcg) ?: nbytes;
3186 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3189 return mem_cgroup_from_css(css)->use_hierarchy;
3192 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3193 struct cftype *cft, u64 val)
3196 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3197 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3199 if (memcg->use_hierarchy == val)
3203 * If parent's use_hierarchy is set, we can't make any modifications
3204 * in the child subtrees. If it is unset, then the change can
3205 * occur, provided the current cgroup has no children.
3207 * For the root cgroup, parent_mem is NULL, we allow value to be
3208 * set if there are no children.
3210 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3211 (val == 1 || val == 0)) {
3212 if (!memcg_has_children(memcg))
3213 memcg->use_hierarchy = val;
3222 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3226 if (mem_cgroup_is_root(memcg)) {
3227 val = memcg_page_state(memcg, MEMCG_CACHE) +
3228 memcg_page_state(memcg, MEMCG_RSS);
3230 val += memcg_page_state(memcg, MEMCG_SWAP);
3233 val = page_counter_read(&memcg->memory);
3235 val = page_counter_read(&memcg->memsw);
3248 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3251 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3252 struct page_counter *counter;
3254 switch (MEMFILE_TYPE(cft->private)) {
3256 counter = &memcg->memory;
3259 counter = &memcg->memsw;
3262 counter = &memcg->kmem;
3265 counter = &memcg->tcpmem;
3271 switch (MEMFILE_ATTR(cft->private)) {
3273 if (counter == &memcg->memory)
3274 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3275 if (counter == &memcg->memsw)
3276 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3277 return (u64)page_counter_read(counter) * PAGE_SIZE;
3279 return (u64)counter->max * PAGE_SIZE;
3281 return (u64)counter->watermark * PAGE_SIZE;
3283 return counter->failcnt;
3284 case RES_SOFT_LIMIT:
3285 return (u64)memcg->soft_limit * PAGE_SIZE;
3291 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg, bool slab_only)
3293 unsigned long stat[MEMCG_NR_STAT];
3294 struct mem_cgroup *mi;
3296 int min_idx, max_idx;
3299 min_idx = NR_SLAB_RECLAIMABLE;
3300 max_idx = NR_SLAB_UNRECLAIMABLE;
3303 max_idx = MEMCG_NR_STAT;
3306 for (i = min_idx; i < max_idx; i++)
3309 for_each_online_cpu(cpu)
3310 for (i = min_idx; i < max_idx; i++)
3311 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3313 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3314 for (i = min_idx; i < max_idx; i++)
3315 atomic_long_add(stat[i], &mi->vmstats[i]);
3318 max_idx = NR_VM_NODE_STAT_ITEMS;
3320 for_each_node(node) {
3321 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3322 struct mem_cgroup_per_node *pi;
3324 for (i = min_idx; i < max_idx; i++)
3327 for_each_online_cpu(cpu)
3328 for (i = min_idx; i < max_idx; i++)
3330 pn->lruvec_stat_cpu->count[i], cpu);
3332 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3333 for (i = min_idx; i < max_idx; i++)
3334 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3338 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3340 unsigned long events[NR_VM_EVENT_ITEMS];
3341 struct mem_cgroup *mi;
3344 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3347 for_each_online_cpu(cpu)
3348 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3349 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3352 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3353 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3354 atomic_long_add(events[i], &mi->vmevents[i]);
3357 #ifdef CONFIG_MEMCG_KMEM
3358 static int memcg_online_kmem(struct mem_cgroup *memcg)
3362 if (cgroup_memory_nokmem)
3365 BUG_ON(memcg->kmemcg_id >= 0);
3366 BUG_ON(memcg->kmem_state);
3368 memcg_id = memcg_alloc_cache_id();
3372 static_branch_inc(&memcg_kmem_enabled_key);
3374 * A memory cgroup is considered kmem-online as soon as it gets
3375 * kmemcg_id. Setting the id after enabling static branching will
3376 * guarantee no one starts accounting before all call sites are
3379 memcg->kmemcg_id = memcg_id;
3380 memcg->kmem_state = KMEM_ONLINE;
3381 INIT_LIST_HEAD(&memcg->kmem_caches);
3386 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3388 struct cgroup_subsys_state *css;
3389 struct mem_cgroup *parent, *child;
3392 if (memcg->kmem_state != KMEM_ONLINE)
3395 * Clear the online state before clearing memcg_caches array
3396 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3397 * guarantees that no cache will be created for this cgroup
3398 * after we are done (see memcg_create_kmem_cache()).
3400 memcg->kmem_state = KMEM_ALLOCATED;
3402 parent = parent_mem_cgroup(memcg);
3404 parent = root_mem_cgroup;
3407 * Deactivate and reparent kmem_caches. Then flush percpu
3408 * slab statistics to have precise values at the parent and
3409 * all ancestor levels. It's required to keep slab stats
3410 * accurate after the reparenting of kmem_caches.
3412 memcg_deactivate_kmem_caches(memcg, parent);
3413 memcg_flush_percpu_vmstats(memcg, true);
3415 kmemcg_id = memcg->kmemcg_id;
3416 BUG_ON(kmemcg_id < 0);
3419 * Change kmemcg_id of this cgroup and all its descendants to the
3420 * parent's id, and then move all entries from this cgroup's list_lrus
3421 * to ones of the parent. After we have finished, all list_lrus
3422 * corresponding to this cgroup are guaranteed to remain empty. The
3423 * ordering is imposed by list_lru_node->lock taken by
3424 * memcg_drain_all_list_lrus().
3426 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3427 css_for_each_descendant_pre(css, &memcg->css) {
3428 child = mem_cgroup_from_css(css);
3429 BUG_ON(child->kmemcg_id != kmemcg_id);
3430 child->kmemcg_id = parent->kmemcg_id;
3431 if (!memcg->use_hierarchy)
3436 memcg_drain_all_list_lrus(kmemcg_id, parent);
3438 memcg_free_cache_id(kmemcg_id);
3441 static void memcg_free_kmem(struct mem_cgroup *memcg)
3443 /* css_alloc() failed, offlining didn't happen */
3444 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3445 memcg_offline_kmem(memcg);
3447 if (memcg->kmem_state == KMEM_ALLOCATED) {
3448 WARN_ON(!list_empty(&memcg->kmem_caches));
3449 static_branch_dec(&memcg_kmem_enabled_key);
3453 static int memcg_online_kmem(struct mem_cgroup *memcg)
3457 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3460 static void memcg_free_kmem(struct mem_cgroup *memcg)
3463 #endif /* CONFIG_MEMCG_KMEM */
3465 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3470 mutex_lock(&memcg_max_mutex);
3471 ret = page_counter_set_max(&memcg->kmem, max);
3472 mutex_unlock(&memcg_max_mutex);
3476 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3480 mutex_lock(&memcg_max_mutex);
3482 ret = page_counter_set_max(&memcg->tcpmem, max);
3486 if (!memcg->tcpmem_active) {
3488 * The active flag needs to be written after the static_key
3489 * update. This is what guarantees that the socket activation
3490 * function is the last one to run. See mem_cgroup_sk_alloc()
3491 * for details, and note that we don't mark any socket as
3492 * belonging to this memcg until that flag is up.
3494 * We need to do this, because static_keys will span multiple
3495 * sites, but we can't control their order. If we mark a socket
3496 * as accounted, but the accounting functions are not patched in
3497 * yet, we'll lose accounting.
3499 * We never race with the readers in mem_cgroup_sk_alloc(),
3500 * because when this value change, the code to process it is not
3503 static_branch_inc(&memcg_sockets_enabled_key);
3504 memcg->tcpmem_active = true;
3507 mutex_unlock(&memcg_max_mutex);
3512 * The user of this function is...
3515 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3516 char *buf, size_t nbytes, loff_t off)
3518 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3519 unsigned long nr_pages;
3522 buf = strstrip(buf);
3523 ret = page_counter_memparse(buf, "-1", &nr_pages);
3527 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3529 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3533 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3535 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3538 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3541 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3542 "Please report your usecase to linux-mm@kvack.org if you "
3543 "depend on this functionality.\n");
3544 ret = memcg_update_kmem_max(memcg, nr_pages);
3547 ret = memcg_update_tcp_max(memcg, nr_pages);
3551 case RES_SOFT_LIMIT:
3552 memcg->soft_limit = nr_pages;
3556 return ret ?: nbytes;
3559 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3560 size_t nbytes, loff_t off)
3562 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3563 struct page_counter *counter;
3565 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3567 counter = &memcg->memory;
3570 counter = &memcg->memsw;
3573 counter = &memcg->kmem;
3576 counter = &memcg->tcpmem;
3582 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3584 page_counter_reset_watermark(counter);
3587 counter->failcnt = 0;
3596 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3599 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3603 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3604 struct cftype *cft, u64 val)
3606 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3608 if (val & ~MOVE_MASK)
3612 * No kind of locking is needed in here, because ->can_attach() will
3613 * check this value once in the beginning of the process, and then carry
3614 * on with stale data. This means that changes to this value will only
3615 * affect task migrations starting after the change.
3617 memcg->move_charge_at_immigrate = val;
3621 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3622 struct cftype *cft, u64 val)
3630 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3631 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3632 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3634 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3635 int nid, unsigned int lru_mask)
3637 struct lruvec *lruvec = mem_cgroup_lruvec(NODE_DATA(nid), memcg);
3638 unsigned long nr = 0;
3641 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3644 if (!(BIT(lru) & lru_mask))
3646 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3651 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3652 unsigned int lru_mask)
3654 unsigned long nr = 0;
3658 if (!(BIT(lru) & lru_mask))
3660 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3665 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3669 unsigned int lru_mask;
3672 static const struct numa_stat stats[] = {
3673 { "total", LRU_ALL },
3674 { "file", LRU_ALL_FILE },
3675 { "anon", LRU_ALL_ANON },
3676 { "unevictable", BIT(LRU_UNEVICTABLE) },
3678 const struct numa_stat *stat;
3681 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3683 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3684 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3685 seq_printf(m, "%s=%lu", stat->name, nr);
3686 for_each_node_state(nid, N_MEMORY) {
3687 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3689 seq_printf(m, " N%d=%lu", nid, nr);
3694 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3695 struct mem_cgroup *iter;
3698 for_each_mem_cgroup_tree(iter, memcg)
3699 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3700 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3701 for_each_node_state(nid, N_MEMORY) {
3703 for_each_mem_cgroup_tree(iter, memcg)
3704 nr += mem_cgroup_node_nr_lru_pages(
3705 iter, nid, stat->lru_mask);
3706 seq_printf(m, " N%d=%lu", nid, nr);
3713 #endif /* CONFIG_NUMA */
3715 static const unsigned int memcg1_stats[] = {
3726 static const char *const memcg1_stat_names[] = {
3737 /* Universal VM events cgroup1 shows, original sort order */
3738 static const unsigned int memcg1_events[] = {
3745 static const char *const memcg1_event_names[] = {
3752 static int memcg_stat_show(struct seq_file *m, void *v)
3754 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3755 unsigned long memory, memsw;
3756 struct mem_cgroup *mi;
3759 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3760 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3762 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3763 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3765 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3766 memcg_page_state_local(memcg, memcg1_stats[i]) *
3770 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3771 seq_printf(m, "%s %lu\n", memcg1_event_names[i],
3772 memcg_events_local(memcg, memcg1_events[i]));
3774 for (i = 0; i < NR_LRU_LISTS; i++)
3775 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3776 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3779 /* Hierarchical information */
3780 memory = memsw = PAGE_COUNTER_MAX;
3781 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3782 memory = min(memory, mi->memory.max);
3783 memsw = min(memsw, mi->memsw.max);
3785 seq_printf(m, "hierarchical_memory_limit %llu\n",
3786 (u64)memory * PAGE_SIZE);
3787 if (do_memsw_account())
3788 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3789 (u64)memsw * PAGE_SIZE);
3791 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3792 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3794 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3795 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3799 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3800 seq_printf(m, "total_%s %llu\n", memcg1_event_names[i],
3801 (u64)memcg_events(memcg, memcg1_events[i]));
3803 for (i = 0; i < NR_LRU_LISTS; i++)
3804 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i],
3805 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3808 #ifdef CONFIG_DEBUG_VM
3811 struct mem_cgroup_per_node *mz;
3812 struct zone_reclaim_stat *rstat;
3813 unsigned long recent_rotated[2] = {0, 0};
3814 unsigned long recent_scanned[2] = {0, 0};
3816 for_each_online_pgdat(pgdat) {
3817 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3818 rstat = &mz->lruvec.reclaim_stat;
3820 recent_rotated[0] += rstat->recent_rotated[0];
3821 recent_rotated[1] += rstat->recent_rotated[1];
3822 recent_scanned[0] += rstat->recent_scanned[0];
3823 recent_scanned[1] += rstat->recent_scanned[1];
3825 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3826 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3827 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3828 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3835 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3838 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3840 return mem_cgroup_swappiness(memcg);
3843 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3844 struct cftype *cft, u64 val)
3846 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3852 memcg->swappiness = val;
3854 vm_swappiness = val;
3859 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3861 struct mem_cgroup_threshold_ary *t;
3862 unsigned long usage;
3867 t = rcu_dereference(memcg->thresholds.primary);
3869 t = rcu_dereference(memcg->memsw_thresholds.primary);
3874 usage = mem_cgroup_usage(memcg, swap);
3877 * current_threshold points to threshold just below or equal to usage.
3878 * If it's not true, a threshold was crossed after last
3879 * call of __mem_cgroup_threshold().
3881 i = t->current_threshold;
3884 * Iterate backward over array of thresholds starting from
3885 * current_threshold and check if a threshold is crossed.
3886 * If none of thresholds below usage is crossed, we read
3887 * only one element of the array here.
3889 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3890 eventfd_signal(t->entries[i].eventfd, 1);
3892 /* i = current_threshold + 1 */
3896 * Iterate forward over array of thresholds starting from
3897 * current_threshold+1 and check if a threshold is crossed.
3898 * If none of thresholds above usage is crossed, we read
3899 * only one element of the array here.
3901 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3902 eventfd_signal(t->entries[i].eventfd, 1);
3904 /* Update current_threshold */
3905 t->current_threshold = i - 1;
3910 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3913 __mem_cgroup_threshold(memcg, false);
3914 if (do_memsw_account())
3915 __mem_cgroup_threshold(memcg, true);
3917 memcg = parent_mem_cgroup(memcg);
3921 static int compare_thresholds(const void *a, const void *b)
3923 const struct mem_cgroup_threshold *_a = a;
3924 const struct mem_cgroup_threshold *_b = b;
3926 if (_a->threshold > _b->threshold)
3929 if (_a->threshold < _b->threshold)
3935 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3937 struct mem_cgroup_eventfd_list *ev;
3939 spin_lock(&memcg_oom_lock);
3941 list_for_each_entry(ev, &memcg->oom_notify, list)
3942 eventfd_signal(ev->eventfd, 1);
3944 spin_unlock(&memcg_oom_lock);
3948 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3950 struct mem_cgroup *iter;
3952 for_each_mem_cgroup_tree(iter, memcg)
3953 mem_cgroup_oom_notify_cb(iter);
3956 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3957 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3959 struct mem_cgroup_thresholds *thresholds;
3960 struct mem_cgroup_threshold_ary *new;
3961 unsigned long threshold;
3962 unsigned long usage;
3965 ret = page_counter_memparse(args, "-1", &threshold);
3969 mutex_lock(&memcg->thresholds_lock);
3972 thresholds = &memcg->thresholds;
3973 usage = mem_cgroup_usage(memcg, false);
3974 } else if (type == _MEMSWAP) {
3975 thresholds = &memcg->memsw_thresholds;
3976 usage = mem_cgroup_usage(memcg, true);
3980 /* Check if a threshold crossed before adding a new one */
3981 if (thresholds->primary)
3982 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3984 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3986 /* Allocate memory for new array of thresholds */
3987 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
3994 /* Copy thresholds (if any) to new array */
3995 if (thresholds->primary) {
3996 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3997 sizeof(struct mem_cgroup_threshold));
4000 /* Add new threshold */
4001 new->entries[size - 1].eventfd = eventfd;
4002 new->entries[size - 1].threshold = threshold;
4004 /* Sort thresholds. Registering of new threshold isn't time-critical */
4005 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4006 compare_thresholds, NULL);
4008 /* Find current threshold */
4009 new->current_threshold = -1;
4010 for (i = 0; i < size; i++) {
4011 if (new->entries[i].threshold <= usage) {
4013 * new->current_threshold will not be used until
4014 * rcu_assign_pointer(), so it's safe to increment
4017 ++new->current_threshold;
4022 /* Free old spare buffer and save old primary buffer as spare */
4023 kfree(thresholds->spare);
4024 thresholds->spare = thresholds->primary;
4026 rcu_assign_pointer(thresholds->primary, new);
4028 /* To be sure that nobody uses thresholds */
4032 mutex_unlock(&memcg->thresholds_lock);
4037 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4038 struct eventfd_ctx *eventfd, const char *args)
4040 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4043 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4044 struct eventfd_ctx *eventfd, const char *args)
4046 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4049 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4050 struct eventfd_ctx *eventfd, enum res_type type)
4052 struct mem_cgroup_thresholds *thresholds;
4053 struct mem_cgroup_threshold_ary *new;
4054 unsigned long usage;
4057 mutex_lock(&memcg->thresholds_lock);
4060 thresholds = &memcg->thresholds;
4061 usage = mem_cgroup_usage(memcg, false);
4062 } else if (type == _MEMSWAP) {
4063 thresholds = &memcg->memsw_thresholds;
4064 usage = mem_cgroup_usage(memcg, true);
4068 if (!thresholds->primary)
4071 /* Check if a threshold crossed before removing */
4072 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4074 /* Calculate new number of threshold */
4076 for (i = 0; i < thresholds->primary->size; i++) {
4077 if (thresholds->primary->entries[i].eventfd != eventfd)
4081 new = thresholds->spare;
4083 /* Set thresholds array to NULL if we don't have thresholds */
4092 /* Copy thresholds and find current threshold */
4093 new->current_threshold = -1;
4094 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4095 if (thresholds->primary->entries[i].eventfd == eventfd)
4098 new->entries[j] = thresholds->primary->entries[i];
4099 if (new->entries[j].threshold <= usage) {
4101 * new->current_threshold will not be used
4102 * until rcu_assign_pointer(), so it's safe to increment
4105 ++new->current_threshold;
4111 /* Swap primary and spare array */
4112 thresholds->spare = thresholds->primary;
4114 rcu_assign_pointer(thresholds->primary, new);
4116 /* To be sure that nobody uses thresholds */
4119 /* If all events are unregistered, free the spare array */
4121 kfree(thresholds->spare);
4122 thresholds->spare = NULL;
4125 mutex_unlock(&memcg->thresholds_lock);
4128 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4129 struct eventfd_ctx *eventfd)
4131 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4134 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4135 struct eventfd_ctx *eventfd)
4137 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4140 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4141 struct eventfd_ctx *eventfd, const char *args)
4143 struct mem_cgroup_eventfd_list *event;
4145 event = kmalloc(sizeof(*event), GFP_KERNEL);
4149 spin_lock(&memcg_oom_lock);
4151 event->eventfd = eventfd;
4152 list_add(&event->list, &memcg->oom_notify);
4154 /* already in OOM ? */
4155 if (memcg->under_oom)
4156 eventfd_signal(eventfd, 1);
4157 spin_unlock(&memcg_oom_lock);
4162 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4163 struct eventfd_ctx *eventfd)
4165 struct mem_cgroup_eventfd_list *ev, *tmp;
4167 spin_lock(&memcg_oom_lock);
4169 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4170 if (ev->eventfd == eventfd) {
4171 list_del(&ev->list);
4176 spin_unlock(&memcg_oom_lock);
4179 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4181 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4183 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4184 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4185 seq_printf(sf, "oom_kill %lu\n",
4186 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4190 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4191 struct cftype *cft, u64 val)
4193 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4195 /* cannot set to root cgroup and only 0 and 1 are allowed */
4196 if (!css->parent || !((val == 0) || (val == 1)))
4199 memcg->oom_kill_disable = val;
4201 memcg_oom_recover(memcg);
4206 #ifdef CONFIG_CGROUP_WRITEBACK
4208 #include <trace/events/writeback.h>
4210 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4212 return wb_domain_init(&memcg->cgwb_domain, gfp);
4215 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4217 wb_domain_exit(&memcg->cgwb_domain);
4220 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4222 wb_domain_size_changed(&memcg->cgwb_domain);
4225 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4227 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4229 if (!memcg->css.parent)
4232 return &memcg->cgwb_domain;
4236 * idx can be of type enum memcg_stat_item or node_stat_item.
4237 * Keep in sync with memcg_exact_page().
4239 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4241 long x = atomic_long_read(&memcg->vmstats[idx]);
4244 for_each_online_cpu(cpu)
4245 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4252 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4253 * @wb: bdi_writeback in question
4254 * @pfilepages: out parameter for number of file pages
4255 * @pheadroom: out parameter for number of allocatable pages according to memcg
4256 * @pdirty: out parameter for number of dirty pages
4257 * @pwriteback: out parameter for number of pages under writeback
4259 * Determine the numbers of file, headroom, dirty, and writeback pages in
4260 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4261 * is a bit more involved.
4263 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4264 * headroom is calculated as the lowest headroom of itself and the
4265 * ancestors. Note that this doesn't consider the actual amount of
4266 * available memory in the system. The caller should further cap
4267 * *@pheadroom accordingly.
4269 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4270 unsigned long *pheadroom, unsigned long *pdirty,
4271 unsigned long *pwriteback)
4273 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4274 struct mem_cgroup *parent;
4276 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4278 /* this should eventually include NR_UNSTABLE_NFS */
4279 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4280 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4281 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4282 *pheadroom = PAGE_COUNTER_MAX;
4284 while ((parent = parent_mem_cgroup(memcg))) {
4285 unsigned long ceiling = min(memcg->memory.max, memcg->high);
4286 unsigned long used = page_counter_read(&memcg->memory);
4288 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4294 * Foreign dirty flushing
4296 * There's an inherent mismatch between memcg and writeback. The former
4297 * trackes ownership per-page while the latter per-inode. This was a
4298 * deliberate design decision because honoring per-page ownership in the
4299 * writeback path is complicated, may lead to higher CPU and IO overheads
4300 * and deemed unnecessary given that write-sharing an inode across
4301 * different cgroups isn't a common use-case.
4303 * Combined with inode majority-writer ownership switching, this works well
4304 * enough in most cases but there are some pathological cases. For
4305 * example, let's say there are two cgroups A and B which keep writing to
4306 * different but confined parts of the same inode. B owns the inode and
4307 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4308 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4309 * triggering background writeback. A will be slowed down without a way to
4310 * make writeback of the dirty pages happen.
4312 * Conditions like the above can lead to a cgroup getting repatedly and
4313 * severely throttled after making some progress after each
4314 * dirty_expire_interval while the underyling IO device is almost
4317 * Solving this problem completely requires matching the ownership tracking
4318 * granularities between memcg and writeback in either direction. However,
4319 * the more egregious behaviors can be avoided by simply remembering the
4320 * most recent foreign dirtying events and initiating remote flushes on
4321 * them when local writeback isn't enough to keep the memory clean enough.
4323 * The following two functions implement such mechanism. When a foreign
4324 * page - a page whose memcg and writeback ownerships don't match - is
4325 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4326 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4327 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4328 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4329 * foreign bdi_writebacks which haven't expired. Both the numbers of
4330 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4331 * limited to MEMCG_CGWB_FRN_CNT.
4333 * The mechanism only remembers IDs and doesn't hold any object references.
4334 * As being wrong occasionally doesn't matter, updates and accesses to the
4335 * records are lockless and racy.
4337 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4338 struct bdi_writeback *wb)
4340 struct mem_cgroup *memcg = page->mem_cgroup;
4341 struct memcg_cgwb_frn *frn;
4342 u64 now = get_jiffies_64();
4343 u64 oldest_at = now;
4347 trace_track_foreign_dirty(page, wb);
4350 * Pick the slot to use. If there is already a slot for @wb, keep
4351 * using it. If not replace the oldest one which isn't being
4354 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4355 frn = &memcg->cgwb_frn[i];
4356 if (frn->bdi_id == wb->bdi->id &&
4357 frn->memcg_id == wb->memcg_css->id)
4359 if (time_before64(frn->at, oldest_at) &&
4360 atomic_read(&frn->done.cnt) == 1) {
4362 oldest_at = frn->at;
4366 if (i < MEMCG_CGWB_FRN_CNT) {
4368 * Re-using an existing one. Update timestamp lazily to
4369 * avoid making the cacheline hot. We want them to be
4370 * reasonably up-to-date and significantly shorter than
4371 * dirty_expire_interval as that's what expires the record.
4372 * Use the shorter of 1s and dirty_expire_interval / 8.
4374 unsigned long update_intv =
4375 min_t(unsigned long, HZ,
4376 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4378 if (time_before64(frn->at, now - update_intv))
4380 } else if (oldest >= 0) {
4381 /* replace the oldest free one */
4382 frn = &memcg->cgwb_frn[oldest];
4383 frn->bdi_id = wb->bdi->id;
4384 frn->memcg_id = wb->memcg_css->id;
4389 /* issue foreign writeback flushes for recorded foreign dirtying events */
4390 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4392 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4393 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4394 u64 now = jiffies_64;
4397 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4398 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4401 * If the record is older than dirty_expire_interval,
4402 * writeback on it has already started. No need to kick it
4403 * off again. Also, don't start a new one if there's
4404 * already one in flight.
4406 if (time_after64(frn->at, now - intv) &&
4407 atomic_read(&frn->done.cnt) == 1) {
4409 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4410 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4411 WB_REASON_FOREIGN_FLUSH,
4417 #else /* CONFIG_CGROUP_WRITEBACK */
4419 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4424 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4428 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4432 #endif /* CONFIG_CGROUP_WRITEBACK */
4435 * DO NOT USE IN NEW FILES.
4437 * "cgroup.event_control" implementation.
4439 * This is way over-engineered. It tries to support fully configurable
4440 * events for each user. Such level of flexibility is completely
4441 * unnecessary especially in the light of the planned unified hierarchy.
4443 * Please deprecate this and replace with something simpler if at all
4448 * Unregister event and free resources.
4450 * Gets called from workqueue.
4452 static void memcg_event_remove(struct work_struct *work)
4454 struct mem_cgroup_event *event =
4455 container_of(work, struct mem_cgroup_event, remove);
4456 struct mem_cgroup *memcg = event->memcg;
4458 remove_wait_queue(event->wqh, &event->wait);
4460 event->unregister_event(memcg, event->eventfd);
4462 /* Notify userspace the event is going away. */
4463 eventfd_signal(event->eventfd, 1);
4465 eventfd_ctx_put(event->eventfd);
4467 css_put(&memcg->css);
4471 * Gets called on EPOLLHUP on eventfd when user closes it.
4473 * Called with wqh->lock held and interrupts disabled.
4475 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4476 int sync, void *key)
4478 struct mem_cgroup_event *event =
4479 container_of(wait, struct mem_cgroup_event, wait);
4480 struct mem_cgroup *memcg = event->memcg;
4481 __poll_t flags = key_to_poll(key);
4483 if (flags & EPOLLHUP) {
4485 * If the event has been detached at cgroup removal, we
4486 * can simply return knowing the other side will cleanup
4489 * We can't race against event freeing since the other
4490 * side will require wqh->lock via remove_wait_queue(),
4493 spin_lock(&memcg->event_list_lock);
4494 if (!list_empty(&event->list)) {
4495 list_del_init(&event->list);
4497 * We are in atomic context, but cgroup_event_remove()
4498 * may sleep, so we have to call it in workqueue.
4500 schedule_work(&event->remove);
4502 spin_unlock(&memcg->event_list_lock);
4508 static void memcg_event_ptable_queue_proc(struct file *file,
4509 wait_queue_head_t *wqh, poll_table *pt)
4511 struct mem_cgroup_event *event =
4512 container_of(pt, struct mem_cgroup_event, pt);
4515 add_wait_queue(wqh, &event->wait);
4519 * DO NOT USE IN NEW FILES.
4521 * Parse input and register new cgroup event handler.
4523 * Input must be in format '<event_fd> <control_fd> <args>'.
4524 * Interpretation of args is defined by control file implementation.
4526 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4527 char *buf, size_t nbytes, loff_t off)
4529 struct cgroup_subsys_state *css = of_css(of);
4530 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4531 struct mem_cgroup_event *event;
4532 struct cgroup_subsys_state *cfile_css;
4533 unsigned int efd, cfd;
4540 buf = strstrip(buf);
4542 efd = simple_strtoul(buf, &endp, 10);
4547 cfd = simple_strtoul(buf, &endp, 10);
4548 if ((*endp != ' ') && (*endp != '\0'))
4552 event = kzalloc(sizeof(*event), GFP_KERNEL);
4556 event->memcg = memcg;
4557 INIT_LIST_HEAD(&event->list);
4558 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4559 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4560 INIT_WORK(&event->remove, memcg_event_remove);
4568 event->eventfd = eventfd_ctx_fileget(efile.file);
4569 if (IS_ERR(event->eventfd)) {
4570 ret = PTR_ERR(event->eventfd);
4577 goto out_put_eventfd;
4580 /* the process need read permission on control file */
4581 /* AV: shouldn't we check that it's been opened for read instead? */
4582 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4587 * Determine the event callbacks and set them in @event. This used
4588 * to be done via struct cftype but cgroup core no longer knows
4589 * about these events. The following is crude but the whole thing
4590 * is for compatibility anyway.
4592 * DO NOT ADD NEW FILES.
4594 name = cfile.file->f_path.dentry->d_name.name;
4596 if (!strcmp(name, "memory.usage_in_bytes")) {
4597 event->register_event = mem_cgroup_usage_register_event;
4598 event->unregister_event = mem_cgroup_usage_unregister_event;
4599 } else if (!strcmp(name, "memory.oom_control")) {
4600 event->register_event = mem_cgroup_oom_register_event;
4601 event->unregister_event = mem_cgroup_oom_unregister_event;
4602 } else if (!strcmp(name, "memory.pressure_level")) {
4603 event->register_event = vmpressure_register_event;
4604 event->unregister_event = vmpressure_unregister_event;
4605 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4606 event->register_event = memsw_cgroup_usage_register_event;
4607 event->unregister_event = memsw_cgroup_usage_unregister_event;
4614 * Verify @cfile should belong to @css. Also, remaining events are
4615 * automatically removed on cgroup destruction but the removal is
4616 * asynchronous, so take an extra ref on @css.
4618 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4619 &memory_cgrp_subsys);
4621 if (IS_ERR(cfile_css))
4623 if (cfile_css != css) {
4628 ret = event->register_event(memcg, event->eventfd, buf);
4632 vfs_poll(efile.file, &event->pt);
4634 spin_lock(&memcg->event_list_lock);
4635 list_add(&event->list, &memcg->event_list);
4636 spin_unlock(&memcg->event_list_lock);
4648 eventfd_ctx_put(event->eventfd);
4657 static struct cftype mem_cgroup_legacy_files[] = {
4659 .name = "usage_in_bytes",
4660 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4661 .read_u64 = mem_cgroup_read_u64,
4664 .name = "max_usage_in_bytes",
4665 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4666 .write = mem_cgroup_reset,
4667 .read_u64 = mem_cgroup_read_u64,
4670 .name = "limit_in_bytes",
4671 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4672 .write = mem_cgroup_write,
4673 .read_u64 = mem_cgroup_read_u64,
4676 .name = "soft_limit_in_bytes",
4677 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4678 .write = mem_cgroup_write,
4679 .read_u64 = mem_cgroup_read_u64,
4683 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4684 .write = mem_cgroup_reset,
4685 .read_u64 = mem_cgroup_read_u64,
4689 .seq_show = memcg_stat_show,
4692 .name = "force_empty",
4693 .write = mem_cgroup_force_empty_write,
4696 .name = "use_hierarchy",
4697 .write_u64 = mem_cgroup_hierarchy_write,
4698 .read_u64 = mem_cgroup_hierarchy_read,
4701 .name = "cgroup.event_control", /* XXX: for compat */
4702 .write = memcg_write_event_control,
4703 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4706 .name = "swappiness",
4707 .read_u64 = mem_cgroup_swappiness_read,
4708 .write_u64 = mem_cgroup_swappiness_write,
4711 .name = "move_charge_at_immigrate",
4712 .read_u64 = mem_cgroup_move_charge_read,
4713 .write_u64 = mem_cgroup_move_charge_write,
4716 .name = "oom_control",
4717 .seq_show = mem_cgroup_oom_control_read,
4718 .write_u64 = mem_cgroup_oom_control_write,
4719 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4722 .name = "pressure_level",
4726 .name = "numa_stat",
4727 .seq_show = memcg_numa_stat_show,
4731 .name = "kmem.limit_in_bytes",
4732 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4733 .write = mem_cgroup_write,
4734 .read_u64 = mem_cgroup_read_u64,
4737 .name = "kmem.usage_in_bytes",
4738 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4739 .read_u64 = mem_cgroup_read_u64,
4742 .name = "kmem.failcnt",
4743 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4744 .write = mem_cgroup_reset,
4745 .read_u64 = mem_cgroup_read_u64,
4748 .name = "kmem.max_usage_in_bytes",
4749 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4750 .write = mem_cgroup_reset,
4751 .read_u64 = mem_cgroup_read_u64,
4753 #if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
4755 .name = "kmem.slabinfo",
4756 .seq_start = memcg_slab_start,
4757 .seq_next = memcg_slab_next,
4758 .seq_stop = memcg_slab_stop,
4759 .seq_show = memcg_slab_show,
4763 .name = "kmem.tcp.limit_in_bytes",
4764 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4765 .write = mem_cgroup_write,
4766 .read_u64 = mem_cgroup_read_u64,
4769 .name = "kmem.tcp.usage_in_bytes",
4770 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4771 .read_u64 = mem_cgroup_read_u64,
4774 .name = "kmem.tcp.failcnt",
4775 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4776 .write = mem_cgroup_reset,
4777 .read_u64 = mem_cgroup_read_u64,
4780 .name = "kmem.tcp.max_usage_in_bytes",
4781 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4782 .write = mem_cgroup_reset,
4783 .read_u64 = mem_cgroup_read_u64,
4785 { }, /* terminate */
4789 * Private memory cgroup IDR
4791 * Swap-out records and page cache shadow entries need to store memcg
4792 * references in constrained space, so we maintain an ID space that is
4793 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4794 * memory-controlled cgroups to 64k.
4796 * However, there usually are many references to the oflline CSS after
4797 * the cgroup has been destroyed, such as page cache or reclaimable
4798 * slab objects, that don't need to hang on to the ID. We want to keep
4799 * those dead CSS from occupying IDs, or we might quickly exhaust the
4800 * relatively small ID space and prevent the creation of new cgroups
4801 * even when there are much fewer than 64k cgroups - possibly none.
4803 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4804 * be freed and recycled when it's no longer needed, which is usually
4805 * when the CSS is offlined.
4807 * The only exception to that are records of swapped out tmpfs/shmem
4808 * pages that need to be attributed to live ancestors on swapin. But
4809 * those references are manageable from userspace.
4812 static DEFINE_IDR(mem_cgroup_idr);
4814 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4816 if (memcg->id.id > 0) {
4817 idr_remove(&mem_cgroup_idr, memcg->id.id);
4822 static void mem_cgroup_id_get_many(struct mem_cgroup *memcg, unsigned int n)
4824 refcount_add(n, &memcg->id.ref);
4827 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4829 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4830 mem_cgroup_id_remove(memcg);
4832 /* Memcg ID pins CSS */
4833 css_put(&memcg->css);
4837 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4839 mem_cgroup_id_put_many(memcg, 1);
4843 * mem_cgroup_from_id - look up a memcg from a memcg id
4844 * @id: the memcg id to look up
4846 * Caller must hold rcu_read_lock().
4848 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4850 WARN_ON_ONCE(!rcu_read_lock_held());
4851 return idr_find(&mem_cgroup_idr, id);
4854 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4856 struct mem_cgroup_per_node *pn;
4859 * This routine is called against possible nodes.
4860 * But it's BUG to call kmalloc() against offline node.
4862 * TODO: this routine can waste much memory for nodes which will
4863 * never be onlined. It's better to use memory hotplug callback
4866 if (!node_state(node, N_NORMAL_MEMORY))
4868 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4872 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4873 if (!pn->lruvec_stat_local) {
4878 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4879 if (!pn->lruvec_stat_cpu) {
4880 free_percpu(pn->lruvec_stat_local);
4885 lruvec_init(&pn->lruvec);
4886 pn->usage_in_excess = 0;
4887 pn->on_tree = false;
4890 memcg->nodeinfo[node] = pn;
4894 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4896 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
4901 free_percpu(pn->lruvec_stat_cpu);
4902 free_percpu(pn->lruvec_stat_local);
4906 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4911 free_mem_cgroup_per_node_info(memcg, node);
4912 free_percpu(memcg->vmstats_percpu);
4913 free_percpu(memcg->vmstats_local);
4917 static void mem_cgroup_free(struct mem_cgroup *memcg)
4919 memcg_wb_domain_exit(memcg);
4921 * Flush percpu vmstats and vmevents to guarantee the value correctness
4922 * on parent's and all ancestor levels.
4924 memcg_flush_percpu_vmstats(memcg, false);
4925 memcg_flush_percpu_vmevents(memcg);
4926 __mem_cgroup_free(memcg);
4929 static struct mem_cgroup *mem_cgroup_alloc(void)
4931 struct mem_cgroup *memcg;
4934 int __maybe_unused i;
4936 size = sizeof(struct mem_cgroup);
4937 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4939 memcg = kzalloc(size, GFP_KERNEL);
4943 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
4944 1, MEM_CGROUP_ID_MAX,
4946 if (memcg->id.id < 0)
4949 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
4950 if (!memcg->vmstats_local)
4953 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
4954 if (!memcg->vmstats_percpu)
4958 if (alloc_mem_cgroup_per_node_info(memcg, node))
4961 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4964 INIT_WORK(&memcg->high_work, high_work_func);
4965 INIT_LIST_HEAD(&memcg->oom_notify);
4966 mutex_init(&memcg->thresholds_lock);
4967 spin_lock_init(&memcg->move_lock);
4968 vmpressure_init(&memcg->vmpressure);
4969 INIT_LIST_HEAD(&memcg->event_list);
4970 spin_lock_init(&memcg->event_list_lock);
4971 memcg->socket_pressure = jiffies;
4972 #ifdef CONFIG_MEMCG_KMEM
4973 memcg->kmemcg_id = -1;
4975 #ifdef CONFIG_CGROUP_WRITEBACK
4976 INIT_LIST_HEAD(&memcg->cgwb_list);
4977 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
4978 memcg->cgwb_frn[i].done =
4979 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
4981 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4982 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
4983 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
4984 memcg->deferred_split_queue.split_queue_len = 0;
4986 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
4989 mem_cgroup_id_remove(memcg);
4990 __mem_cgroup_free(memcg);
4994 static struct cgroup_subsys_state * __ref
4995 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4997 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
4998 struct mem_cgroup *memcg;
4999 long error = -ENOMEM;
5001 memcg = mem_cgroup_alloc();
5003 return ERR_PTR(error);
5005 memcg->high = PAGE_COUNTER_MAX;
5006 memcg->soft_limit = PAGE_COUNTER_MAX;
5008 memcg->swappiness = mem_cgroup_swappiness(parent);
5009 memcg->oom_kill_disable = parent->oom_kill_disable;
5011 if (parent && parent->use_hierarchy) {
5012 memcg->use_hierarchy = true;
5013 page_counter_init(&memcg->memory, &parent->memory);
5014 page_counter_init(&memcg->swap, &parent->swap);
5015 page_counter_init(&memcg->memsw, &parent->memsw);
5016 page_counter_init(&memcg->kmem, &parent->kmem);
5017 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5019 page_counter_init(&memcg->memory, NULL);
5020 page_counter_init(&memcg->swap, NULL);
5021 page_counter_init(&memcg->memsw, NULL);
5022 page_counter_init(&memcg->kmem, NULL);
5023 page_counter_init(&memcg->tcpmem, NULL);
5025 * Deeper hierachy with use_hierarchy == false doesn't make
5026 * much sense so let cgroup subsystem know about this
5027 * unfortunate state in our controller.
5029 if (parent != root_mem_cgroup)
5030 memory_cgrp_subsys.broken_hierarchy = true;
5033 /* The following stuff does not apply to the root */
5035 #ifdef CONFIG_MEMCG_KMEM
5036 INIT_LIST_HEAD(&memcg->kmem_caches);
5038 root_mem_cgroup = memcg;
5042 error = memcg_online_kmem(memcg);
5046 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5047 static_branch_inc(&memcg_sockets_enabled_key);
5051 mem_cgroup_id_remove(memcg);
5052 mem_cgroup_free(memcg);
5053 return ERR_PTR(-ENOMEM);
5056 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5058 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5061 * A memcg must be visible for memcg_expand_shrinker_maps()
5062 * by the time the maps are allocated. So, we allocate maps
5063 * here, when for_each_mem_cgroup() can't skip it.
5065 if (memcg_alloc_shrinker_maps(memcg)) {
5066 mem_cgroup_id_remove(memcg);
5070 /* Online state pins memcg ID, memcg ID pins CSS */
5071 refcount_set(&memcg->id.ref, 1);
5076 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5078 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5079 struct mem_cgroup_event *event, *tmp;
5082 * Unregister events and notify userspace.
5083 * Notify userspace about cgroup removing only after rmdir of cgroup
5084 * directory to avoid race between userspace and kernelspace.
5086 spin_lock(&memcg->event_list_lock);
5087 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5088 list_del_init(&event->list);
5089 schedule_work(&event->remove);
5091 spin_unlock(&memcg->event_list_lock);
5093 page_counter_set_min(&memcg->memory, 0);
5094 page_counter_set_low(&memcg->memory, 0);
5096 memcg_offline_kmem(memcg);
5097 wb_memcg_offline(memcg);
5099 drain_all_stock(memcg);
5101 mem_cgroup_id_put(memcg);
5104 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5106 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5108 invalidate_reclaim_iterators(memcg);
5111 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5113 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5114 int __maybe_unused i;
5116 #ifdef CONFIG_CGROUP_WRITEBACK
5117 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5118 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5120 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5121 static_branch_dec(&memcg_sockets_enabled_key);
5123 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5124 static_branch_dec(&memcg_sockets_enabled_key);
5126 vmpressure_cleanup(&memcg->vmpressure);
5127 cancel_work_sync(&memcg->high_work);
5128 mem_cgroup_remove_from_trees(memcg);
5129 memcg_free_shrinker_maps(memcg);
5130 memcg_free_kmem(memcg);
5131 mem_cgroup_free(memcg);
5135 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5136 * @css: the target css
5138 * Reset the states of the mem_cgroup associated with @css. This is
5139 * invoked when the userland requests disabling on the default hierarchy
5140 * but the memcg is pinned through dependency. The memcg should stop
5141 * applying policies and should revert to the vanilla state as it may be
5142 * made visible again.
5144 * The current implementation only resets the essential configurations.
5145 * This needs to be expanded to cover all the visible parts.
5147 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5149 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5151 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5152 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5153 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5154 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5155 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5156 page_counter_set_min(&memcg->memory, 0);
5157 page_counter_set_low(&memcg->memory, 0);
5158 memcg->high = PAGE_COUNTER_MAX;
5159 memcg->soft_limit = PAGE_COUNTER_MAX;
5160 memcg_wb_domain_size_changed(memcg);
5164 /* Handlers for move charge at task migration. */
5165 static int mem_cgroup_do_precharge(unsigned long count)
5169 /* Try a single bulk charge without reclaim first, kswapd may wake */
5170 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5172 mc.precharge += count;
5176 /* Try charges one by one with reclaim, but do not retry */
5178 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5192 enum mc_target_type {
5199 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5200 unsigned long addr, pte_t ptent)
5202 struct page *page = vm_normal_page(vma, addr, ptent);
5204 if (!page || !page_mapped(page))
5206 if (PageAnon(page)) {
5207 if (!(mc.flags & MOVE_ANON))
5210 if (!(mc.flags & MOVE_FILE))
5213 if (!get_page_unless_zero(page))
5219 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5220 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5221 pte_t ptent, swp_entry_t *entry)
5223 struct page *page = NULL;
5224 swp_entry_t ent = pte_to_swp_entry(ptent);
5226 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5230 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5231 * a device and because they are not accessible by CPU they are store
5232 * as special swap entry in the CPU page table.
5234 if (is_device_private_entry(ent)) {
5235 page = device_private_entry_to_page(ent);
5237 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5238 * a refcount of 1 when free (unlike normal page)
5240 if (!page_ref_add_unless(page, 1, 1))
5246 * Because lookup_swap_cache() updates some statistics counter,
5247 * we call find_get_page() with swapper_space directly.
5249 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5250 if (do_memsw_account())
5251 entry->val = ent.val;
5256 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5257 pte_t ptent, swp_entry_t *entry)
5263 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5264 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5266 struct page *page = NULL;
5267 struct address_space *mapping;
5270 if (!vma->vm_file) /* anonymous vma */
5272 if (!(mc.flags & MOVE_FILE))
5275 mapping = vma->vm_file->f_mapping;
5276 pgoff = linear_page_index(vma, addr);
5278 /* page is moved even if it's not RSS of this task(page-faulted). */
5280 /* shmem/tmpfs may report page out on swap: account for that too. */
5281 if (shmem_mapping(mapping)) {
5282 page = find_get_entry(mapping, pgoff);
5283 if (xa_is_value(page)) {
5284 swp_entry_t swp = radix_to_swp_entry(page);
5285 if (do_memsw_account())
5287 page = find_get_page(swap_address_space(swp),
5291 page = find_get_page(mapping, pgoff);
5293 page = find_get_page(mapping, pgoff);
5299 * mem_cgroup_move_account - move account of the page
5301 * @compound: charge the page as compound or small page
5302 * @from: mem_cgroup which the page is moved from.
5303 * @to: mem_cgroup which the page is moved to. @from != @to.
5305 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5307 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5310 static int mem_cgroup_move_account(struct page *page,
5312 struct mem_cgroup *from,
5313 struct mem_cgroup *to)
5315 struct lruvec *from_vec, *to_vec;
5316 struct pglist_data *pgdat;
5317 unsigned long flags;
5318 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5322 VM_BUG_ON(from == to);
5323 VM_BUG_ON_PAGE(PageLRU(page), page);
5324 VM_BUG_ON(compound && !PageTransHuge(page));
5327 * Prevent mem_cgroup_migrate() from looking at
5328 * page->mem_cgroup of its source page while we change it.
5331 if (!trylock_page(page))
5335 if (page->mem_cgroup != from)
5338 anon = PageAnon(page);
5340 pgdat = page_pgdat(page);
5341 from_vec = mem_cgroup_lruvec(pgdat, from);
5342 to_vec = mem_cgroup_lruvec(pgdat, to);
5344 spin_lock_irqsave(&from->move_lock, flags);
5346 if (!anon && page_mapped(page)) {
5347 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5348 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5352 * move_lock grabbed above and caller set from->moving_account, so
5353 * mod_memcg_page_state will serialize updates to PageDirty.
5354 * So mapping should be stable for dirty pages.
5356 if (!anon && PageDirty(page)) {
5357 struct address_space *mapping = page_mapping(page);
5359 if (mapping_cap_account_dirty(mapping)) {
5360 __mod_lruvec_state(from_vec, NR_FILE_DIRTY, -nr_pages);
5361 __mod_lruvec_state(to_vec, NR_FILE_DIRTY, nr_pages);
5365 if (PageWriteback(page)) {
5366 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5367 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5370 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5371 if (compound && !list_empty(page_deferred_list(page))) {
5372 spin_lock(&from->deferred_split_queue.split_queue_lock);
5373 list_del_init(page_deferred_list(page));
5374 from->deferred_split_queue.split_queue_len--;
5375 spin_unlock(&from->deferred_split_queue.split_queue_lock);
5379 * It is safe to change page->mem_cgroup here because the page
5380 * is referenced, charged, and isolated - we can't race with
5381 * uncharging, charging, migration, or LRU putback.
5384 /* caller should have done css_get */
5385 page->mem_cgroup = to;
5387 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5388 if (compound && list_empty(page_deferred_list(page))) {
5389 spin_lock(&to->deferred_split_queue.split_queue_lock);
5390 list_add_tail(page_deferred_list(page),
5391 &to->deferred_split_queue.split_queue);
5392 to->deferred_split_queue.split_queue_len++;
5393 spin_unlock(&to->deferred_split_queue.split_queue_lock);
5397 spin_unlock_irqrestore(&from->move_lock, flags);
5401 local_irq_disable();
5402 mem_cgroup_charge_statistics(to, page, compound, nr_pages);
5403 memcg_check_events(to, page);
5404 mem_cgroup_charge_statistics(from, page, compound, -nr_pages);
5405 memcg_check_events(from, page);
5414 * get_mctgt_type - get target type of moving charge
5415 * @vma: the vma the pte to be checked belongs
5416 * @addr: the address corresponding to the pte to be checked
5417 * @ptent: the pte to be checked
5418 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5421 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5422 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5423 * move charge. if @target is not NULL, the page is stored in target->page
5424 * with extra refcnt got(Callers should handle it).
5425 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5426 * target for charge migration. if @target is not NULL, the entry is stored
5428 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5429 * (so ZONE_DEVICE page and thus not on the lru).
5430 * For now we such page is charge like a regular page would be as for all
5431 * intent and purposes it is just special memory taking the place of a
5434 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5436 * Called with pte lock held.
5439 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5440 unsigned long addr, pte_t ptent, union mc_target *target)
5442 struct page *page = NULL;
5443 enum mc_target_type ret = MC_TARGET_NONE;
5444 swp_entry_t ent = { .val = 0 };
5446 if (pte_present(ptent))
5447 page = mc_handle_present_pte(vma, addr, ptent);
5448 else if (is_swap_pte(ptent))
5449 page = mc_handle_swap_pte(vma, ptent, &ent);
5450 else if (pte_none(ptent))
5451 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5453 if (!page && !ent.val)
5457 * Do only loose check w/o serialization.
5458 * mem_cgroup_move_account() checks the page is valid or
5459 * not under LRU exclusion.
5461 if (page->mem_cgroup == mc.from) {
5462 ret = MC_TARGET_PAGE;
5463 if (is_device_private_page(page))
5464 ret = MC_TARGET_DEVICE;
5466 target->page = page;
5468 if (!ret || !target)
5472 * There is a swap entry and a page doesn't exist or isn't charged.
5473 * But we cannot move a tail-page in a THP.
5475 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5476 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5477 ret = MC_TARGET_SWAP;
5484 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5486 * We don't consider PMD mapped swapping or file mapped pages because THP does
5487 * not support them for now.
5488 * Caller should make sure that pmd_trans_huge(pmd) is true.
5490 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5491 unsigned long addr, pmd_t pmd, union mc_target *target)
5493 struct page *page = NULL;
5494 enum mc_target_type ret = MC_TARGET_NONE;
5496 if (unlikely(is_swap_pmd(pmd))) {
5497 VM_BUG_ON(thp_migration_supported() &&
5498 !is_pmd_migration_entry(pmd));
5501 page = pmd_page(pmd);
5502 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5503 if (!(mc.flags & MOVE_ANON))
5505 if (page->mem_cgroup == mc.from) {
5506 ret = MC_TARGET_PAGE;
5509 target->page = page;
5515 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5516 unsigned long addr, pmd_t pmd, union mc_target *target)
5518 return MC_TARGET_NONE;
5522 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5523 unsigned long addr, unsigned long end,
5524 struct mm_walk *walk)
5526 struct vm_area_struct *vma = walk->vma;
5530 ptl = pmd_trans_huge_lock(pmd, vma);
5533 * Note their can not be MC_TARGET_DEVICE for now as we do not
5534 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5535 * this might change.
5537 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5538 mc.precharge += HPAGE_PMD_NR;
5543 if (pmd_trans_unstable(pmd))
5545 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5546 for (; addr != end; pte++, addr += PAGE_SIZE)
5547 if (get_mctgt_type(vma, addr, *pte, NULL))
5548 mc.precharge++; /* increment precharge temporarily */
5549 pte_unmap_unlock(pte - 1, ptl);
5555 static const struct mm_walk_ops precharge_walk_ops = {
5556 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5559 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5561 unsigned long precharge;
5563 down_read(&mm->mmap_sem);
5564 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5565 up_read(&mm->mmap_sem);
5567 precharge = mc.precharge;
5573 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5575 unsigned long precharge = mem_cgroup_count_precharge(mm);
5577 VM_BUG_ON(mc.moving_task);
5578 mc.moving_task = current;
5579 return mem_cgroup_do_precharge(precharge);
5582 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5583 static void __mem_cgroup_clear_mc(void)
5585 struct mem_cgroup *from = mc.from;
5586 struct mem_cgroup *to = mc.to;
5588 /* we must uncharge all the leftover precharges from mc.to */
5590 cancel_charge(mc.to, mc.precharge);
5594 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5595 * we must uncharge here.
5597 if (mc.moved_charge) {
5598 cancel_charge(mc.from, mc.moved_charge);
5599 mc.moved_charge = 0;
5601 /* we must fixup refcnts and charges */
5602 if (mc.moved_swap) {
5603 /* uncharge swap account from the old cgroup */
5604 if (!mem_cgroup_is_root(mc.from))
5605 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5607 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5610 * we charged both to->memory and to->memsw, so we
5611 * should uncharge to->memory.
5613 if (!mem_cgroup_is_root(mc.to))
5614 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5616 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5617 css_put_many(&mc.to->css, mc.moved_swap);
5621 memcg_oom_recover(from);
5622 memcg_oom_recover(to);
5623 wake_up_all(&mc.waitq);
5626 static void mem_cgroup_clear_mc(void)
5628 struct mm_struct *mm = mc.mm;
5631 * we must clear moving_task before waking up waiters at the end of
5634 mc.moving_task = NULL;
5635 __mem_cgroup_clear_mc();
5636 spin_lock(&mc.lock);
5640 spin_unlock(&mc.lock);
5645 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5647 struct cgroup_subsys_state *css;
5648 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5649 struct mem_cgroup *from;
5650 struct task_struct *leader, *p;
5651 struct mm_struct *mm;
5652 unsigned long move_flags;
5655 /* charge immigration isn't supported on the default hierarchy */
5656 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5660 * Multi-process migrations only happen on the default hierarchy
5661 * where charge immigration is not used. Perform charge
5662 * immigration if @tset contains a leader and whine if there are
5666 cgroup_taskset_for_each_leader(leader, css, tset) {
5669 memcg = mem_cgroup_from_css(css);
5675 * We are now commited to this value whatever it is. Changes in this
5676 * tunable will only affect upcoming migrations, not the current one.
5677 * So we need to save it, and keep it going.
5679 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5683 from = mem_cgroup_from_task(p);
5685 VM_BUG_ON(from == memcg);
5687 mm = get_task_mm(p);
5690 /* We move charges only when we move a owner of the mm */
5691 if (mm->owner == p) {
5694 VM_BUG_ON(mc.precharge);
5695 VM_BUG_ON(mc.moved_charge);
5696 VM_BUG_ON(mc.moved_swap);
5698 spin_lock(&mc.lock);
5702 mc.flags = move_flags;
5703 spin_unlock(&mc.lock);
5704 /* We set mc.moving_task later */
5706 ret = mem_cgroup_precharge_mc(mm);
5708 mem_cgroup_clear_mc();
5715 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5718 mem_cgroup_clear_mc();
5721 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5722 unsigned long addr, unsigned long end,
5723 struct mm_walk *walk)
5726 struct vm_area_struct *vma = walk->vma;
5729 enum mc_target_type target_type;
5730 union mc_target target;
5733 ptl = pmd_trans_huge_lock(pmd, vma);
5735 if (mc.precharge < HPAGE_PMD_NR) {
5739 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5740 if (target_type == MC_TARGET_PAGE) {
5742 if (!isolate_lru_page(page)) {
5743 if (!mem_cgroup_move_account(page, true,
5745 mc.precharge -= HPAGE_PMD_NR;
5746 mc.moved_charge += HPAGE_PMD_NR;
5748 putback_lru_page(page);
5751 } else if (target_type == MC_TARGET_DEVICE) {
5753 if (!mem_cgroup_move_account(page, true,
5755 mc.precharge -= HPAGE_PMD_NR;
5756 mc.moved_charge += HPAGE_PMD_NR;
5764 if (pmd_trans_unstable(pmd))
5767 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5768 for (; addr != end; addr += PAGE_SIZE) {
5769 pte_t ptent = *(pte++);
5770 bool device = false;
5776 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5777 case MC_TARGET_DEVICE:
5780 case MC_TARGET_PAGE:
5783 * We can have a part of the split pmd here. Moving it
5784 * can be done but it would be too convoluted so simply
5785 * ignore such a partial THP and keep it in original
5786 * memcg. There should be somebody mapping the head.
5788 if (PageTransCompound(page))
5790 if (!device && isolate_lru_page(page))
5792 if (!mem_cgroup_move_account(page, false,
5795 /* we uncharge from mc.from later. */
5799 putback_lru_page(page);
5800 put: /* get_mctgt_type() gets the page */
5803 case MC_TARGET_SWAP:
5805 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5807 /* we fixup refcnts and charges later. */
5815 pte_unmap_unlock(pte - 1, ptl);
5820 * We have consumed all precharges we got in can_attach().
5821 * We try charge one by one, but don't do any additional
5822 * charges to mc.to if we have failed in charge once in attach()
5825 ret = mem_cgroup_do_precharge(1);
5833 static const struct mm_walk_ops charge_walk_ops = {
5834 .pmd_entry = mem_cgroup_move_charge_pte_range,
5837 static void mem_cgroup_move_charge(void)
5839 lru_add_drain_all();
5841 * Signal lock_page_memcg() to take the memcg's move_lock
5842 * while we're moving its pages to another memcg. Then wait
5843 * for already started RCU-only updates to finish.
5845 atomic_inc(&mc.from->moving_account);
5848 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5850 * Someone who are holding the mmap_sem might be waiting in
5851 * waitq. So we cancel all extra charges, wake up all waiters,
5852 * and retry. Because we cancel precharges, we might not be able
5853 * to move enough charges, but moving charge is a best-effort
5854 * feature anyway, so it wouldn't be a big problem.
5856 __mem_cgroup_clear_mc();
5861 * When we have consumed all precharges and failed in doing
5862 * additional charge, the page walk just aborts.
5864 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5867 up_read(&mc.mm->mmap_sem);
5868 atomic_dec(&mc.from->moving_account);
5871 static void mem_cgroup_move_task(void)
5874 mem_cgroup_move_charge();
5875 mem_cgroup_clear_mc();
5878 #else /* !CONFIG_MMU */
5879 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5883 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5886 static void mem_cgroup_move_task(void)
5892 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5893 * to verify whether we're attached to the default hierarchy on each mount
5896 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5899 * use_hierarchy is forced on the default hierarchy. cgroup core
5900 * guarantees that @root doesn't have any children, so turning it
5901 * on for the root memcg is enough.
5903 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5904 root_mem_cgroup->use_hierarchy = true;
5906 root_mem_cgroup->use_hierarchy = false;
5909 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
5911 if (value == PAGE_COUNTER_MAX)
5912 seq_puts(m, "max\n");
5914 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
5919 static u64 memory_current_read(struct cgroup_subsys_state *css,
5922 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5924 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
5927 static int memory_min_show(struct seq_file *m, void *v)
5929 return seq_puts_memcg_tunable(m,
5930 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
5933 static ssize_t memory_min_write(struct kernfs_open_file *of,
5934 char *buf, size_t nbytes, loff_t off)
5936 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5940 buf = strstrip(buf);
5941 err = page_counter_memparse(buf, "max", &min);
5945 page_counter_set_min(&memcg->memory, min);
5950 static int memory_low_show(struct seq_file *m, void *v)
5952 return seq_puts_memcg_tunable(m,
5953 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
5956 static ssize_t memory_low_write(struct kernfs_open_file *of,
5957 char *buf, size_t nbytes, loff_t off)
5959 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5963 buf = strstrip(buf);
5964 err = page_counter_memparse(buf, "max", &low);
5968 page_counter_set_low(&memcg->memory, low);
5973 static int memory_high_show(struct seq_file *m, void *v)
5975 return seq_puts_memcg_tunable(m, READ_ONCE(mem_cgroup_from_seq(m)->high));
5978 static ssize_t memory_high_write(struct kernfs_open_file *of,
5979 char *buf, size_t nbytes, loff_t off)
5981 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5982 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5983 bool drained = false;
5987 buf = strstrip(buf);
5988 err = page_counter_memparse(buf, "max", &high);
5995 unsigned long nr_pages = page_counter_read(&memcg->memory);
5996 unsigned long reclaimed;
5998 if (nr_pages <= high)
6001 if (signal_pending(current))
6005 drain_all_stock(memcg);
6010 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6013 if (!reclaimed && !nr_retries--)
6020 static int memory_max_show(struct seq_file *m, void *v)
6022 return seq_puts_memcg_tunable(m,
6023 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6026 static ssize_t memory_max_write(struct kernfs_open_file *of,
6027 char *buf, size_t nbytes, loff_t off)
6029 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6030 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6031 bool drained = false;
6035 buf = strstrip(buf);
6036 err = page_counter_memparse(buf, "max", &max);
6040 xchg(&memcg->memory.max, max);
6043 unsigned long nr_pages = page_counter_read(&memcg->memory);
6045 if (nr_pages <= max)
6048 if (signal_pending(current))
6052 drain_all_stock(memcg);
6058 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6064 memcg_memory_event(memcg, MEMCG_OOM);
6065 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6069 memcg_wb_domain_size_changed(memcg);
6073 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6075 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6076 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6077 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6078 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6079 seq_printf(m, "oom_kill %lu\n",
6080 atomic_long_read(&events[MEMCG_OOM_KILL]));
6083 static int memory_events_show(struct seq_file *m, void *v)
6085 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6087 __memory_events_show(m, memcg->memory_events);
6091 static int memory_events_local_show(struct seq_file *m, void *v)
6093 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6095 __memory_events_show(m, memcg->memory_events_local);
6099 static int memory_stat_show(struct seq_file *m, void *v)
6101 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6104 buf = memory_stat_format(memcg);
6112 static int memory_oom_group_show(struct seq_file *m, void *v)
6114 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6116 seq_printf(m, "%d\n", memcg->oom_group);
6121 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6122 char *buf, size_t nbytes, loff_t off)
6124 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6127 buf = strstrip(buf);
6131 ret = kstrtoint(buf, 0, &oom_group);
6135 if (oom_group != 0 && oom_group != 1)
6138 memcg->oom_group = oom_group;
6143 static struct cftype memory_files[] = {
6146 .flags = CFTYPE_NOT_ON_ROOT,
6147 .read_u64 = memory_current_read,
6151 .flags = CFTYPE_NOT_ON_ROOT,
6152 .seq_show = memory_min_show,
6153 .write = memory_min_write,
6157 .flags = CFTYPE_NOT_ON_ROOT,
6158 .seq_show = memory_low_show,
6159 .write = memory_low_write,
6163 .flags = CFTYPE_NOT_ON_ROOT,
6164 .seq_show = memory_high_show,
6165 .write = memory_high_write,
6169 .flags = CFTYPE_NOT_ON_ROOT,
6170 .seq_show = memory_max_show,
6171 .write = memory_max_write,
6175 .flags = CFTYPE_NOT_ON_ROOT,
6176 .file_offset = offsetof(struct mem_cgroup, events_file),
6177 .seq_show = memory_events_show,
6180 .name = "events.local",
6181 .flags = CFTYPE_NOT_ON_ROOT,
6182 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6183 .seq_show = memory_events_local_show,
6187 .flags = CFTYPE_NOT_ON_ROOT,
6188 .seq_show = memory_stat_show,
6191 .name = "oom.group",
6192 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6193 .seq_show = memory_oom_group_show,
6194 .write = memory_oom_group_write,
6199 struct cgroup_subsys memory_cgrp_subsys = {
6200 .css_alloc = mem_cgroup_css_alloc,
6201 .css_online = mem_cgroup_css_online,
6202 .css_offline = mem_cgroup_css_offline,
6203 .css_released = mem_cgroup_css_released,
6204 .css_free = mem_cgroup_css_free,
6205 .css_reset = mem_cgroup_css_reset,
6206 .can_attach = mem_cgroup_can_attach,
6207 .cancel_attach = mem_cgroup_cancel_attach,
6208 .post_attach = mem_cgroup_move_task,
6209 .bind = mem_cgroup_bind,
6210 .dfl_cftypes = memory_files,
6211 .legacy_cftypes = mem_cgroup_legacy_files,
6216 * mem_cgroup_protected - check if memory consumption is in the normal range
6217 * @root: the top ancestor of the sub-tree being checked
6218 * @memcg: the memory cgroup to check
6220 * WARNING: This function is not stateless! It can only be used as part
6221 * of a top-down tree iteration, not for isolated queries.
6223 * Returns one of the following:
6224 * MEMCG_PROT_NONE: cgroup memory is not protected
6225 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6226 * an unprotected supply of reclaimable memory from other cgroups.
6227 * MEMCG_PROT_MIN: cgroup memory is protected
6229 * @root is exclusive; it is never protected when looked at directly
6231 * To provide a proper hierarchical behavior, effective memory.min/low values
6232 * are used. Below is the description of how effective memory.low is calculated.
6233 * Effective memory.min values is calculated in the same way.
6235 * Effective memory.low is always equal or less than the original memory.low.
6236 * If there is no memory.low overcommittment (which is always true for
6237 * top-level memory cgroups), these two values are equal.
6238 * Otherwise, it's a part of parent's effective memory.low,
6239 * calculated as a cgroup's memory.low usage divided by sum of sibling's
6240 * memory.low usages, where memory.low usage is the size of actually
6244 * elow = min( memory.low, parent->elow * ------------------ ),
6245 * siblings_low_usage
6247 * | memory.current, if memory.current < memory.low
6252 * Such definition of the effective memory.low provides the expected
6253 * hierarchical behavior: parent's memory.low value is limiting
6254 * children, unprotected memory is reclaimed first and cgroups,
6255 * which are not using their guarantee do not affect actual memory
6258 * For example, if there are memcgs A, A/B, A/C, A/D and A/E:
6260 * A A/memory.low = 2G, A/memory.current = 6G
6262 * BC DE B/memory.low = 3G B/memory.current = 2G
6263 * C/memory.low = 1G C/memory.current = 2G
6264 * D/memory.low = 0 D/memory.current = 2G
6265 * E/memory.low = 10G E/memory.current = 0
6267 * and the memory pressure is applied, the following memory distribution
6268 * is expected (approximately):
6270 * A/memory.current = 2G
6272 * B/memory.current = 1.3G
6273 * C/memory.current = 0.6G
6274 * D/memory.current = 0
6275 * E/memory.current = 0
6277 * These calculations require constant tracking of the actual low usages
6278 * (see propagate_protected_usage()), as well as recursive calculation of
6279 * effective memory.low values. But as we do call mem_cgroup_protected()
6280 * path for each memory cgroup top-down from the reclaim,
6281 * it's possible to optimize this part, and save calculated elow
6282 * for next usage. This part is intentionally racy, but it's ok,
6283 * as memory.low is a best-effort mechanism.
6285 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6286 struct mem_cgroup *memcg)
6288 struct mem_cgroup *parent;
6289 unsigned long emin, parent_emin;
6290 unsigned long elow, parent_elow;
6291 unsigned long usage;
6293 if (mem_cgroup_disabled())
6294 return MEMCG_PROT_NONE;
6297 root = root_mem_cgroup;
6299 return MEMCG_PROT_NONE;
6301 usage = page_counter_read(&memcg->memory);
6303 return MEMCG_PROT_NONE;
6305 emin = memcg->memory.min;
6306 elow = memcg->memory.low;
6308 parent = parent_mem_cgroup(memcg);
6309 /* No parent means a non-hierarchical mode on v1 memcg */
6311 return MEMCG_PROT_NONE;
6316 parent_emin = READ_ONCE(parent->memory.emin);
6317 emin = min(emin, parent_emin);
6318 if (emin && parent_emin) {
6319 unsigned long min_usage, siblings_min_usage;
6321 min_usage = min(usage, memcg->memory.min);
6322 siblings_min_usage = atomic_long_read(
6323 &parent->memory.children_min_usage);
6325 if (min_usage && siblings_min_usage)
6326 emin = min(emin, parent_emin * min_usage /
6327 siblings_min_usage);
6330 parent_elow = READ_ONCE(parent->memory.elow);
6331 elow = min(elow, parent_elow);
6332 if (elow && parent_elow) {
6333 unsigned long low_usage, siblings_low_usage;
6335 low_usage = min(usage, memcg->memory.low);
6336 siblings_low_usage = atomic_long_read(
6337 &parent->memory.children_low_usage);
6339 if (low_usage && siblings_low_usage)
6340 elow = min(elow, parent_elow * low_usage /
6341 siblings_low_usage);
6345 memcg->memory.emin = emin;
6346 memcg->memory.elow = elow;
6349 return MEMCG_PROT_MIN;
6350 else if (usage <= elow)
6351 return MEMCG_PROT_LOW;
6353 return MEMCG_PROT_NONE;
6357 * mem_cgroup_try_charge - try charging a page
6358 * @page: page to charge
6359 * @mm: mm context of the victim
6360 * @gfp_mask: reclaim mode
6361 * @memcgp: charged memcg return
6362 * @compound: charge the page as compound or small page
6364 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6365 * pages according to @gfp_mask if necessary.
6367 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6368 * Otherwise, an error code is returned.
6370 * After page->mapping has been set up, the caller must finalize the
6371 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6372 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6374 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6375 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6378 struct mem_cgroup *memcg = NULL;
6379 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6382 if (mem_cgroup_disabled())
6385 if (PageSwapCache(page)) {
6387 * Every swap fault against a single page tries to charge the
6388 * page, bail as early as possible. shmem_unuse() encounters
6389 * already charged pages, too. The USED bit is protected by
6390 * the page lock, which serializes swap cache removal, which
6391 * in turn serializes uncharging.
6393 VM_BUG_ON_PAGE(!PageLocked(page), page);
6394 if (compound_head(page)->mem_cgroup)
6397 if (do_swap_account) {
6398 swp_entry_t ent = { .val = page_private(page), };
6399 unsigned short id = lookup_swap_cgroup_id(ent);
6402 memcg = mem_cgroup_from_id(id);
6403 if (memcg && !css_tryget_online(&memcg->css))
6410 memcg = get_mem_cgroup_from_mm(mm);
6412 ret = try_charge(memcg, gfp_mask, nr_pages);
6414 css_put(&memcg->css);
6420 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6421 gfp_t gfp_mask, struct mem_cgroup **memcgp,
6424 struct mem_cgroup *memcg;
6427 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp, compound);
6429 mem_cgroup_throttle_swaprate(memcg, page_to_nid(page), gfp_mask);
6434 * mem_cgroup_commit_charge - commit a page charge
6435 * @page: page to charge
6436 * @memcg: memcg to charge the page to
6437 * @lrucare: page might be on LRU already
6438 * @compound: charge the page as compound or small page
6440 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6441 * after page->mapping has been set up. This must happen atomically
6442 * as part of the page instantiation, i.e. under the page table lock
6443 * for anonymous pages, under the page lock for page and swap cache.
6445 * In addition, the page must not be on the LRU during the commit, to
6446 * prevent racing with task migration. If it might be, use @lrucare.
6448 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6450 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6451 bool lrucare, bool compound)
6453 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6455 VM_BUG_ON_PAGE(!page->mapping, page);
6456 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6458 if (mem_cgroup_disabled())
6461 * Swap faults will attempt to charge the same page multiple
6462 * times. But reuse_swap_page() might have removed the page
6463 * from swapcache already, so we can't check PageSwapCache().
6468 commit_charge(page, memcg, lrucare);
6470 local_irq_disable();
6471 mem_cgroup_charge_statistics(memcg, page, compound, nr_pages);
6472 memcg_check_events(memcg, page);
6475 if (do_memsw_account() && PageSwapCache(page)) {
6476 swp_entry_t entry = { .val = page_private(page) };
6478 * The swap entry might not get freed for a long time,
6479 * let's not wait for it. The page already received a
6480 * memory+swap charge, drop the swap entry duplicate.
6482 mem_cgroup_uncharge_swap(entry, nr_pages);
6487 * mem_cgroup_cancel_charge - cancel a page charge
6488 * @page: page to charge
6489 * @memcg: memcg to charge the page to
6490 * @compound: charge the page as compound or small page
6492 * Cancel a charge transaction started by mem_cgroup_try_charge().
6494 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg,
6497 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
6499 if (mem_cgroup_disabled())
6502 * Swap faults will attempt to charge the same page multiple
6503 * times. But reuse_swap_page() might have removed the page
6504 * from swapcache already, so we can't check PageSwapCache().
6509 cancel_charge(memcg, nr_pages);
6512 struct uncharge_gather {
6513 struct mem_cgroup *memcg;
6514 unsigned long pgpgout;
6515 unsigned long nr_anon;
6516 unsigned long nr_file;
6517 unsigned long nr_kmem;
6518 unsigned long nr_huge;
6519 unsigned long nr_shmem;
6520 struct page *dummy_page;
6523 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6525 memset(ug, 0, sizeof(*ug));
6528 static void uncharge_batch(const struct uncharge_gather *ug)
6530 unsigned long nr_pages = ug->nr_anon + ug->nr_file + ug->nr_kmem;
6531 unsigned long flags;
6533 if (!mem_cgroup_is_root(ug->memcg)) {
6534 page_counter_uncharge(&ug->memcg->memory, nr_pages);
6535 if (do_memsw_account())
6536 page_counter_uncharge(&ug->memcg->memsw, nr_pages);
6537 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6538 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6539 memcg_oom_recover(ug->memcg);
6542 local_irq_save(flags);
6543 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6544 __mod_memcg_state(ug->memcg, MEMCG_CACHE, -ug->nr_file);
6545 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6546 __mod_memcg_state(ug->memcg, NR_SHMEM, -ug->nr_shmem);
6547 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6548 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, nr_pages);
6549 memcg_check_events(ug->memcg, ug->dummy_page);
6550 local_irq_restore(flags);
6552 if (!mem_cgroup_is_root(ug->memcg))
6553 css_put_many(&ug->memcg->css, nr_pages);
6556 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6558 VM_BUG_ON_PAGE(PageLRU(page), page);
6559 VM_BUG_ON_PAGE(page_count(page) && !is_zone_device_page(page) &&
6560 !PageHWPoison(page) , page);
6562 if (!page->mem_cgroup)
6566 * Nobody should be changing or seriously looking at
6567 * page->mem_cgroup at this point, we have fully
6568 * exclusive access to the page.
6571 if (ug->memcg != page->mem_cgroup) {
6574 uncharge_gather_clear(ug);
6576 ug->memcg = page->mem_cgroup;
6579 if (!PageKmemcg(page)) {
6580 unsigned int nr_pages = 1;
6582 if (PageTransHuge(page)) {
6583 nr_pages = compound_nr(page);
6584 ug->nr_huge += nr_pages;
6587 ug->nr_anon += nr_pages;
6589 ug->nr_file += nr_pages;
6590 if (PageSwapBacked(page))
6591 ug->nr_shmem += nr_pages;
6595 ug->nr_kmem += compound_nr(page);
6596 __ClearPageKmemcg(page);
6599 ug->dummy_page = page;
6600 page->mem_cgroup = NULL;
6603 static void uncharge_list(struct list_head *page_list)
6605 struct uncharge_gather ug;
6606 struct list_head *next;
6608 uncharge_gather_clear(&ug);
6611 * Note that the list can be a single page->lru; hence the
6612 * do-while loop instead of a simple list_for_each_entry().
6614 next = page_list->next;
6618 page = list_entry(next, struct page, lru);
6619 next = page->lru.next;
6621 uncharge_page(page, &ug);
6622 } while (next != page_list);
6625 uncharge_batch(&ug);
6629 * mem_cgroup_uncharge - uncharge a page
6630 * @page: page to uncharge
6632 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6633 * mem_cgroup_commit_charge().
6635 void mem_cgroup_uncharge(struct page *page)
6637 struct uncharge_gather ug;
6639 if (mem_cgroup_disabled())
6642 /* Don't touch page->lru of any random page, pre-check: */
6643 if (!page->mem_cgroup)
6646 uncharge_gather_clear(&ug);
6647 uncharge_page(page, &ug);
6648 uncharge_batch(&ug);
6652 * mem_cgroup_uncharge_list - uncharge a list of page
6653 * @page_list: list of pages to uncharge
6655 * Uncharge a list of pages previously charged with
6656 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6658 void mem_cgroup_uncharge_list(struct list_head *page_list)
6660 if (mem_cgroup_disabled())
6663 if (!list_empty(page_list))
6664 uncharge_list(page_list);
6668 * mem_cgroup_migrate - charge a page's replacement
6669 * @oldpage: currently circulating page
6670 * @newpage: replacement page
6672 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6673 * be uncharged upon free.
6675 * Both pages must be locked, @newpage->mapping must be set up.
6677 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6679 struct mem_cgroup *memcg;
6680 unsigned int nr_pages;
6682 unsigned long flags;
6684 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6685 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6686 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6687 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6690 if (mem_cgroup_disabled())
6693 /* Page cache replacement: new page already charged? */
6694 if (newpage->mem_cgroup)
6697 /* Swapcache readahead pages can get replaced before being charged */
6698 memcg = oldpage->mem_cgroup;
6702 /* Force-charge the new page. The old one will be freed soon */
6703 compound = PageTransHuge(newpage);
6704 nr_pages = compound ? hpage_nr_pages(newpage) : 1;
6706 page_counter_charge(&memcg->memory, nr_pages);
6707 if (do_memsw_account())
6708 page_counter_charge(&memcg->memsw, nr_pages);
6709 css_get_many(&memcg->css, nr_pages);
6711 commit_charge(newpage, memcg, false);
6713 local_irq_save(flags);
6714 mem_cgroup_charge_statistics(memcg, newpage, compound, nr_pages);
6715 memcg_check_events(memcg, newpage);
6716 local_irq_restore(flags);
6719 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6720 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6722 void mem_cgroup_sk_alloc(struct sock *sk)
6724 struct mem_cgroup *memcg;
6726 if (!mem_cgroup_sockets_enabled)
6730 * Socket cloning can throw us here with sk_memcg already
6731 * filled. It won't however, necessarily happen from
6732 * process context. So the test for root memcg given
6733 * the current task's memcg won't help us in this case.
6735 * Respecting the original socket's memcg is a better
6736 * decision in this case.
6739 css_get(&sk->sk_memcg->css);
6744 memcg = mem_cgroup_from_task(current);
6745 if (memcg == root_mem_cgroup)
6747 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6749 if (css_tryget_online(&memcg->css))
6750 sk->sk_memcg = memcg;
6755 void mem_cgroup_sk_free(struct sock *sk)
6758 css_put(&sk->sk_memcg->css);
6762 * mem_cgroup_charge_skmem - charge socket memory
6763 * @memcg: memcg to charge
6764 * @nr_pages: number of pages to charge
6766 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6767 * @memcg's configured limit, %false if the charge had to be forced.
6769 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6771 gfp_t gfp_mask = GFP_KERNEL;
6773 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6774 struct page_counter *fail;
6776 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6777 memcg->tcpmem_pressure = 0;
6780 page_counter_charge(&memcg->tcpmem, nr_pages);
6781 memcg->tcpmem_pressure = 1;
6785 /* Don't block in the packet receive path */
6787 gfp_mask = GFP_NOWAIT;
6789 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6791 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6794 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6799 * mem_cgroup_uncharge_skmem - uncharge socket memory
6800 * @memcg: memcg to uncharge
6801 * @nr_pages: number of pages to uncharge
6803 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6805 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6806 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6810 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6812 refill_stock(memcg, nr_pages);
6815 static int __init cgroup_memory(char *s)
6819 while ((token = strsep(&s, ",")) != NULL) {
6822 if (!strcmp(token, "nosocket"))
6823 cgroup_memory_nosocket = true;
6824 if (!strcmp(token, "nokmem"))
6825 cgroup_memory_nokmem = true;
6829 __setup("cgroup.memory=", cgroup_memory);
6832 * subsys_initcall() for memory controller.
6834 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6835 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6836 * basically everything that doesn't depend on a specific mem_cgroup structure
6837 * should be initialized from here.
6839 static int __init mem_cgroup_init(void)
6843 #ifdef CONFIG_MEMCG_KMEM
6845 * Kmem cache creation is mostly done with the slab_mutex held,
6846 * so use a workqueue with limited concurrency to avoid stalling
6847 * all worker threads in case lots of cgroups are created and
6848 * destroyed simultaneously.
6850 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6851 BUG_ON(!memcg_kmem_cache_wq);
6854 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6855 memcg_hotplug_cpu_dead);
6857 for_each_possible_cpu(cpu)
6858 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
6861 for_each_node(node) {
6862 struct mem_cgroup_tree_per_node *rtpn;
6864 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
6865 node_online(node) ? node : NUMA_NO_NODE);
6867 rtpn->rb_root = RB_ROOT;
6868 rtpn->rb_rightmost = NULL;
6869 spin_lock_init(&rtpn->lock);
6870 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6875 subsys_initcall(mem_cgroup_init);
6877 #ifdef CONFIG_MEMCG_SWAP
6878 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
6880 while (!refcount_inc_not_zero(&memcg->id.ref)) {
6882 * The root cgroup cannot be destroyed, so it's refcount must
6885 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
6889 memcg = parent_mem_cgroup(memcg);
6891 memcg = root_mem_cgroup;
6897 * mem_cgroup_swapout - transfer a memsw charge to swap
6898 * @page: page whose memsw charge to transfer
6899 * @entry: swap entry to move the charge to
6901 * Transfer the memsw charge of @page to @entry.
6903 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6905 struct mem_cgroup *memcg, *swap_memcg;
6906 unsigned int nr_entries;
6907 unsigned short oldid;
6909 VM_BUG_ON_PAGE(PageLRU(page), page);
6910 VM_BUG_ON_PAGE(page_count(page), page);
6912 if (!do_memsw_account())
6915 memcg = page->mem_cgroup;
6917 /* Readahead page, never charged */
6922 * In case the memcg owning these pages has been offlined and doesn't
6923 * have an ID allocated to it anymore, charge the closest online
6924 * ancestor for the swap instead and transfer the memory+swap charge.
6926 swap_memcg = mem_cgroup_id_get_online(memcg);
6927 nr_entries = hpage_nr_pages(page);
6928 /* Get references for the tail pages, too */
6930 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
6931 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
6933 VM_BUG_ON_PAGE(oldid, page);
6934 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
6936 page->mem_cgroup = NULL;
6938 if (!mem_cgroup_is_root(memcg))
6939 page_counter_uncharge(&memcg->memory, nr_entries);
6941 if (memcg != swap_memcg) {
6942 if (!mem_cgroup_is_root(swap_memcg))
6943 page_counter_charge(&swap_memcg->memsw, nr_entries);
6944 page_counter_uncharge(&memcg->memsw, nr_entries);
6948 * Interrupts should be disabled here because the caller holds the
6949 * i_pages lock which is taken with interrupts-off. It is
6950 * important here to have the interrupts disabled because it is the
6951 * only synchronisation we have for updating the per-CPU variables.
6953 VM_BUG_ON(!irqs_disabled());
6954 mem_cgroup_charge_statistics(memcg, page, PageTransHuge(page),
6956 memcg_check_events(memcg, page);
6958 if (!mem_cgroup_is_root(memcg))
6959 css_put_many(&memcg->css, nr_entries);
6963 * mem_cgroup_try_charge_swap - try charging swap space for a page
6964 * @page: page being added to swap
6965 * @entry: swap entry to charge
6967 * Try to charge @page's memcg for the swap space at @entry.
6969 * Returns 0 on success, -ENOMEM on failure.
6971 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
6973 unsigned int nr_pages = hpage_nr_pages(page);
6974 struct page_counter *counter;
6975 struct mem_cgroup *memcg;
6976 unsigned short oldid;
6978 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
6981 memcg = page->mem_cgroup;
6983 /* Readahead page, never charged */
6988 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6992 memcg = mem_cgroup_id_get_online(memcg);
6994 if (!mem_cgroup_is_root(memcg) &&
6995 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
6996 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
6997 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
6998 mem_cgroup_id_put(memcg);
7002 /* Get references for the tail pages, too */
7004 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7005 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7006 VM_BUG_ON_PAGE(oldid, page);
7007 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7013 * mem_cgroup_uncharge_swap - uncharge swap space
7014 * @entry: swap entry to uncharge
7015 * @nr_pages: the amount of swap space to uncharge
7017 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7019 struct mem_cgroup *memcg;
7022 if (!do_swap_account)
7025 id = swap_cgroup_record(entry, 0, nr_pages);
7027 memcg = mem_cgroup_from_id(id);
7029 if (!mem_cgroup_is_root(memcg)) {
7030 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7031 page_counter_uncharge(&memcg->swap, nr_pages);
7033 page_counter_uncharge(&memcg->memsw, nr_pages);
7035 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7036 mem_cgroup_id_put_many(memcg, nr_pages);
7041 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7043 long nr_swap_pages = get_nr_swap_pages();
7045 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7046 return nr_swap_pages;
7047 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7048 nr_swap_pages = min_t(long, nr_swap_pages,
7049 READ_ONCE(memcg->swap.max) -
7050 page_counter_read(&memcg->swap));
7051 return nr_swap_pages;
7054 bool mem_cgroup_swap_full(struct page *page)
7056 struct mem_cgroup *memcg;
7058 VM_BUG_ON_PAGE(!PageLocked(page), page);
7062 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7065 memcg = page->mem_cgroup;
7069 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7070 if (page_counter_read(&memcg->swap) * 2 >= memcg->swap.max)
7076 /* for remember boot option*/
7077 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7078 static int really_do_swap_account __initdata = 1;
7080 static int really_do_swap_account __initdata;
7083 static int __init enable_swap_account(char *s)
7085 if (!strcmp(s, "1"))
7086 really_do_swap_account = 1;
7087 else if (!strcmp(s, "0"))
7088 really_do_swap_account = 0;
7091 __setup("swapaccount=", enable_swap_account);
7093 static u64 swap_current_read(struct cgroup_subsys_state *css,
7096 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7098 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7101 static int swap_max_show(struct seq_file *m, void *v)
7103 return seq_puts_memcg_tunable(m,
7104 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7107 static ssize_t swap_max_write(struct kernfs_open_file *of,
7108 char *buf, size_t nbytes, loff_t off)
7110 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7114 buf = strstrip(buf);
7115 err = page_counter_memparse(buf, "max", &max);
7119 xchg(&memcg->swap.max, max);
7124 static int swap_events_show(struct seq_file *m, void *v)
7126 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7128 seq_printf(m, "max %lu\n",
7129 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7130 seq_printf(m, "fail %lu\n",
7131 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7136 static struct cftype swap_files[] = {
7138 .name = "swap.current",
7139 .flags = CFTYPE_NOT_ON_ROOT,
7140 .read_u64 = swap_current_read,
7144 .flags = CFTYPE_NOT_ON_ROOT,
7145 .seq_show = swap_max_show,
7146 .write = swap_max_write,
7149 .name = "swap.events",
7150 .flags = CFTYPE_NOT_ON_ROOT,
7151 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7152 .seq_show = swap_events_show,
7157 static struct cftype memsw_cgroup_files[] = {
7159 .name = "memsw.usage_in_bytes",
7160 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7161 .read_u64 = mem_cgroup_read_u64,
7164 .name = "memsw.max_usage_in_bytes",
7165 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7166 .write = mem_cgroup_reset,
7167 .read_u64 = mem_cgroup_read_u64,
7170 .name = "memsw.limit_in_bytes",
7171 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7172 .write = mem_cgroup_write,
7173 .read_u64 = mem_cgroup_read_u64,
7176 .name = "memsw.failcnt",
7177 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7178 .write = mem_cgroup_reset,
7179 .read_u64 = mem_cgroup_read_u64,
7181 { }, /* terminate */
7184 static int __init mem_cgroup_swap_init(void)
7186 if (!mem_cgroup_disabled() && really_do_swap_account) {
7187 do_swap_account = 1;
7188 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7190 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7191 memsw_cgroup_files));
7195 subsys_initcall(mem_cgroup_swap_init);
7197 #endif /* CONFIG_MEMCG_SWAP */