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 #define THRESHOLDS_EVENTS_TARGET 128
102 #define SOFTLIMIT_EVENTS_TARGET 1024
105 * Cgroups above their limits are maintained in a RB-Tree, independent of
106 * their hierarchy representation
109 struct mem_cgroup_tree_per_node {
110 struct rb_root rb_root;
111 struct rb_node *rb_rightmost;
115 struct mem_cgroup_tree {
116 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
119 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
122 struct mem_cgroup_eventfd_list {
123 struct list_head list;
124 struct eventfd_ctx *eventfd;
128 * cgroup_event represents events which userspace want to receive.
130 struct mem_cgroup_event {
132 * memcg which the event belongs to.
134 struct mem_cgroup *memcg;
136 * eventfd to signal userspace about the event.
138 struct eventfd_ctx *eventfd;
140 * Each of these stored in a list by the cgroup.
142 struct list_head list;
144 * register_event() callback will be used to add new userspace
145 * waiter for changes related to this event. Use eventfd_signal()
146 * on eventfd to send notification to userspace.
148 int (*register_event)(struct mem_cgroup *memcg,
149 struct eventfd_ctx *eventfd, const char *args);
151 * unregister_event() callback will be called when userspace closes
152 * the eventfd or on cgroup removing. This callback must be set,
153 * if you want provide notification functionality.
155 void (*unregister_event)(struct mem_cgroup *memcg,
156 struct eventfd_ctx *eventfd);
158 * All fields below needed to unregister event when
159 * userspace closes eventfd.
162 wait_queue_head_t *wqh;
163 wait_queue_entry_t wait;
164 struct work_struct remove;
167 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
170 /* Stuffs for move charges at task migration. */
172 * Types of charges to be moved.
174 #define MOVE_ANON 0x1U
175 #define MOVE_FILE 0x2U
176 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
178 /* "mc" and its members are protected by cgroup_mutex */
179 static struct move_charge_struct {
180 spinlock_t lock; /* for from, to */
181 struct mm_struct *mm;
182 struct mem_cgroup *from;
183 struct mem_cgroup *to;
185 unsigned long precharge;
186 unsigned long moved_charge;
187 unsigned long moved_swap;
188 struct task_struct *moving_task; /* a task moving charges */
189 wait_queue_head_t waitq; /* a waitq for other context */
191 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
192 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
196 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
197 * limit reclaim to prevent infinite loops, if they ever occur.
199 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
200 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
204 MEM_CGROUP_CHARGE_TYPE_ANON,
205 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
206 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
210 /* for encoding cft->private value on file */
219 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
220 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
221 #define MEMFILE_ATTR(val) ((val) & 0xffff)
222 /* Used for OOM nofiier */
223 #define OOM_CONTROL (0)
226 * Iteration constructs for visiting all cgroups (under a tree). If
227 * loops are exited prematurely (break), mem_cgroup_iter_break() must
228 * be used for reference counting.
230 #define for_each_mem_cgroup_tree(iter, root) \
231 for (iter = mem_cgroup_iter(root, NULL, NULL); \
233 iter = mem_cgroup_iter(root, iter, NULL))
235 #define for_each_mem_cgroup(iter) \
236 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
238 iter = mem_cgroup_iter(NULL, iter, NULL))
240 static inline bool should_force_charge(void)
242 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
243 (current->flags & PF_EXITING);
246 /* Some nice accessors for the vmpressure. */
247 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
250 memcg = root_mem_cgroup;
251 return &memcg->vmpressure;
254 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
256 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
259 #ifdef CONFIG_MEMCG_KMEM
261 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
262 * The main reason for not using cgroup id for this:
263 * this works better in sparse environments, where we have a lot of memcgs,
264 * but only a few kmem-limited. Or also, if we have, for instance, 200
265 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
266 * 200 entry array for that.
268 * The current size of the caches array is stored in memcg_nr_cache_ids. It
269 * will double each time we have to increase it.
271 static DEFINE_IDA(memcg_cache_ida);
272 int memcg_nr_cache_ids;
274 /* Protects memcg_nr_cache_ids */
275 static DECLARE_RWSEM(memcg_cache_ids_sem);
277 void memcg_get_cache_ids(void)
279 down_read(&memcg_cache_ids_sem);
282 void memcg_put_cache_ids(void)
284 up_read(&memcg_cache_ids_sem);
288 * MIN_SIZE is different than 1, because we would like to avoid going through
289 * the alloc/free process all the time. In a small machine, 4 kmem-limited
290 * cgroups is a reasonable guess. In the future, it could be a parameter or
291 * tunable, but that is strictly not necessary.
293 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
294 * this constant directly from cgroup, but it is understandable that this is
295 * better kept as an internal representation in cgroup.c. In any case, the
296 * cgrp_id space is not getting any smaller, and we don't have to necessarily
297 * increase ours as well if it increases.
299 #define MEMCG_CACHES_MIN_SIZE 4
300 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
303 * A lot of the calls to the cache allocation functions are expected to be
304 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
305 * conditional to this static branch, we'll have to allow modules that does
306 * kmem_cache_alloc and the such to see this symbol as well
308 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
309 EXPORT_SYMBOL(memcg_kmem_enabled_key);
311 struct workqueue_struct *memcg_kmem_cache_wq;
314 static int memcg_shrinker_map_size;
315 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
317 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
319 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
322 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
323 int size, int old_size)
325 struct memcg_shrinker_map *new, *old;
328 lockdep_assert_held(&memcg_shrinker_map_mutex);
331 old = rcu_dereference_protected(
332 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
333 /* Not yet online memcg */
337 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
341 /* Set all old bits, clear all new bits */
342 memset(new->map, (int)0xff, old_size);
343 memset((void *)new->map + old_size, 0, size - old_size);
345 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
346 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
352 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
354 struct mem_cgroup_per_node *pn;
355 struct memcg_shrinker_map *map;
358 if (mem_cgroup_is_root(memcg))
362 pn = mem_cgroup_nodeinfo(memcg, nid);
363 map = rcu_dereference_protected(pn->shrinker_map, true);
366 rcu_assign_pointer(pn->shrinker_map, NULL);
370 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
372 struct memcg_shrinker_map *map;
373 int nid, size, ret = 0;
375 if (mem_cgroup_is_root(memcg))
378 mutex_lock(&memcg_shrinker_map_mutex);
379 size = memcg_shrinker_map_size;
381 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
383 memcg_free_shrinker_maps(memcg);
387 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
389 mutex_unlock(&memcg_shrinker_map_mutex);
394 int memcg_expand_shrinker_maps(int new_id)
396 int size, old_size, ret = 0;
397 struct mem_cgroup *memcg;
399 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
400 old_size = memcg_shrinker_map_size;
401 if (size <= old_size)
404 mutex_lock(&memcg_shrinker_map_mutex);
405 if (!root_mem_cgroup)
408 for_each_mem_cgroup(memcg) {
409 if (mem_cgroup_is_root(memcg))
411 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
413 mem_cgroup_iter_break(NULL, memcg);
419 memcg_shrinker_map_size = size;
420 mutex_unlock(&memcg_shrinker_map_mutex);
424 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
426 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
427 struct memcg_shrinker_map *map;
430 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
431 /* Pairs with smp mb in shrink_slab() */
432 smp_mb__before_atomic();
433 set_bit(shrinker_id, map->map);
439 * mem_cgroup_css_from_page - css of the memcg associated with a page
440 * @page: page of interest
442 * If memcg is bound to the default hierarchy, css of the memcg associated
443 * with @page is returned. The returned css remains associated with @page
444 * until it is released.
446 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
449 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
451 struct mem_cgroup *memcg;
453 memcg = page->mem_cgroup;
455 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
456 memcg = root_mem_cgroup;
462 * page_cgroup_ino - return inode number of the memcg a page is charged to
465 * Look up the closest online ancestor of the memory cgroup @page is charged to
466 * and return its inode number or 0 if @page is not charged to any cgroup. It
467 * is safe to call this function without holding a reference to @page.
469 * Note, this function is inherently racy, because there is nothing to prevent
470 * the cgroup inode from getting torn down and potentially reallocated a moment
471 * after page_cgroup_ino() returns, so it only should be used by callers that
472 * do not care (such as procfs interfaces).
474 ino_t page_cgroup_ino(struct page *page)
476 struct mem_cgroup *memcg;
477 unsigned long ino = 0;
480 if (PageSlab(page) && !PageTail(page))
481 memcg = memcg_from_slab_page(page);
483 memcg = READ_ONCE(page->mem_cgroup);
484 while (memcg && !(memcg->css.flags & CSS_ONLINE))
485 memcg = parent_mem_cgroup(memcg);
487 ino = cgroup_ino(memcg->css.cgroup);
492 static struct mem_cgroup_per_node *
493 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
495 int nid = page_to_nid(page);
497 return memcg->nodeinfo[nid];
500 static struct mem_cgroup_tree_per_node *
501 soft_limit_tree_node(int nid)
503 return soft_limit_tree.rb_tree_per_node[nid];
506 static struct mem_cgroup_tree_per_node *
507 soft_limit_tree_from_page(struct page *page)
509 int nid = page_to_nid(page);
511 return soft_limit_tree.rb_tree_per_node[nid];
514 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
515 struct mem_cgroup_tree_per_node *mctz,
516 unsigned long new_usage_in_excess)
518 struct rb_node **p = &mctz->rb_root.rb_node;
519 struct rb_node *parent = NULL;
520 struct mem_cgroup_per_node *mz_node;
521 bool rightmost = true;
526 mz->usage_in_excess = new_usage_in_excess;
527 if (!mz->usage_in_excess)
531 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
533 if (mz->usage_in_excess < mz_node->usage_in_excess) {
539 * We can't avoid mem cgroups that are over their soft
540 * limit by the same amount
542 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
547 mctz->rb_rightmost = &mz->tree_node;
549 rb_link_node(&mz->tree_node, parent, p);
550 rb_insert_color(&mz->tree_node, &mctz->rb_root);
554 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
555 struct mem_cgroup_tree_per_node *mctz)
560 if (&mz->tree_node == mctz->rb_rightmost)
561 mctz->rb_rightmost = rb_prev(&mz->tree_node);
563 rb_erase(&mz->tree_node, &mctz->rb_root);
567 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
568 struct mem_cgroup_tree_per_node *mctz)
572 spin_lock_irqsave(&mctz->lock, flags);
573 __mem_cgroup_remove_exceeded(mz, mctz);
574 spin_unlock_irqrestore(&mctz->lock, flags);
577 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
579 unsigned long nr_pages = page_counter_read(&memcg->memory);
580 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
581 unsigned long excess = 0;
583 if (nr_pages > soft_limit)
584 excess = nr_pages - soft_limit;
589 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
591 unsigned long excess;
592 struct mem_cgroup_per_node *mz;
593 struct mem_cgroup_tree_per_node *mctz;
595 mctz = soft_limit_tree_from_page(page);
599 * Necessary to update all ancestors when hierarchy is used.
600 * because their event counter is not touched.
602 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
603 mz = mem_cgroup_page_nodeinfo(memcg, page);
604 excess = soft_limit_excess(memcg);
606 * We have to update the tree if mz is on RB-tree or
607 * mem is over its softlimit.
609 if (excess || mz->on_tree) {
612 spin_lock_irqsave(&mctz->lock, flags);
613 /* if on-tree, remove it */
615 __mem_cgroup_remove_exceeded(mz, mctz);
617 * Insert again. mz->usage_in_excess will be updated.
618 * If excess is 0, no tree ops.
620 __mem_cgroup_insert_exceeded(mz, mctz, excess);
621 spin_unlock_irqrestore(&mctz->lock, flags);
626 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
628 struct mem_cgroup_tree_per_node *mctz;
629 struct mem_cgroup_per_node *mz;
633 mz = mem_cgroup_nodeinfo(memcg, nid);
634 mctz = soft_limit_tree_node(nid);
636 mem_cgroup_remove_exceeded(mz, mctz);
640 static struct mem_cgroup_per_node *
641 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
643 struct mem_cgroup_per_node *mz;
647 if (!mctz->rb_rightmost)
648 goto done; /* Nothing to reclaim from */
650 mz = rb_entry(mctz->rb_rightmost,
651 struct mem_cgroup_per_node, tree_node);
653 * Remove the node now but someone else can add it back,
654 * we will to add it back at the end of reclaim to its correct
655 * position in the tree.
657 __mem_cgroup_remove_exceeded(mz, mctz);
658 if (!soft_limit_excess(mz->memcg) ||
659 !css_tryget(&mz->memcg->css))
665 static struct mem_cgroup_per_node *
666 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
668 struct mem_cgroup_per_node *mz;
670 spin_lock_irq(&mctz->lock);
671 mz = __mem_cgroup_largest_soft_limit_node(mctz);
672 spin_unlock_irq(&mctz->lock);
677 * __mod_memcg_state - update cgroup memory statistics
678 * @memcg: the memory cgroup
679 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
680 * @val: delta to add to the counter, can be negative
682 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
686 if (mem_cgroup_disabled())
689 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
690 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
691 struct mem_cgroup *mi;
694 * Batch local counters to keep them in sync with
695 * the hierarchical ones.
697 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
698 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
699 atomic_long_add(x, &mi->vmstats[idx]);
702 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
705 static struct mem_cgroup_per_node *
706 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
708 struct mem_cgroup *parent;
710 parent = parent_mem_cgroup(pn->memcg);
713 return mem_cgroup_nodeinfo(parent, nid);
717 * __mod_lruvec_state - update lruvec memory statistics
718 * @lruvec: the lruvec
719 * @idx: the stat item
720 * @val: delta to add to the counter, can be negative
722 * The lruvec is the intersection of the NUMA node and a cgroup. This
723 * function updates the all three counters that are affected by a
724 * change of state at this level: per-node, per-cgroup, per-lruvec.
726 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
729 pg_data_t *pgdat = lruvec_pgdat(lruvec);
730 struct mem_cgroup_per_node *pn;
731 struct mem_cgroup *memcg;
735 __mod_node_page_state(pgdat, idx, val);
737 if (mem_cgroup_disabled())
740 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
744 __mod_memcg_state(memcg, idx, val);
747 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
749 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
750 if (unlikely(abs(x) > MEMCG_CHARGE_BATCH)) {
751 struct mem_cgroup_per_node *pi;
753 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
754 atomic_long_add(x, &pi->lruvec_stat[idx]);
757 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
760 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
762 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
763 struct mem_cgroup *memcg;
764 struct lruvec *lruvec;
767 memcg = mem_cgroup_from_obj(p);
769 /* Untracked pages have no memcg, no lruvec. Update only the node */
770 if (!memcg || memcg == root_mem_cgroup) {
771 __mod_node_page_state(pgdat, idx, val);
773 lruvec = mem_cgroup_lruvec(memcg, pgdat);
774 __mod_lruvec_state(lruvec, idx, val);
779 void mod_memcg_obj_state(void *p, int idx, int val)
781 struct mem_cgroup *memcg;
784 memcg = mem_cgroup_from_obj(p);
786 mod_memcg_state(memcg, idx, val);
791 * __count_memcg_events - account VM events in a cgroup
792 * @memcg: the memory cgroup
793 * @idx: the event item
794 * @count: the number of events that occured
796 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
801 if (mem_cgroup_disabled())
804 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
805 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
806 struct mem_cgroup *mi;
809 * Batch local counters to keep them in sync with
810 * the hierarchical ones.
812 __this_cpu_add(memcg->vmstats_local->events[idx], x);
813 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
814 atomic_long_add(x, &mi->vmevents[idx]);
817 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
820 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
822 return atomic_long_read(&memcg->vmevents[event]);
825 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
830 for_each_possible_cpu(cpu)
831 x += per_cpu(memcg->vmstats_local->events[event], cpu);
835 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
840 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
841 * counted as CACHE even if it's on ANON LRU.
844 __mod_memcg_state(memcg, MEMCG_RSS, nr_pages);
846 if (abs(nr_pages) > 1) {
847 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
848 __mod_memcg_state(memcg, MEMCG_RSS_HUGE, nr_pages);
851 /* pagein of a big page is an event. So, ignore page size */
853 __count_memcg_events(memcg, PGPGIN, 1);
855 __count_memcg_events(memcg, PGPGOUT, 1);
856 nr_pages = -nr_pages; /* for event */
859 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
862 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
863 enum mem_cgroup_events_target target)
865 unsigned long val, next;
867 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
868 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
869 /* from time_after() in jiffies.h */
870 if ((long)(next - val) < 0) {
872 case MEM_CGROUP_TARGET_THRESH:
873 next = val + THRESHOLDS_EVENTS_TARGET;
875 case MEM_CGROUP_TARGET_SOFTLIMIT:
876 next = val + SOFTLIMIT_EVENTS_TARGET;
881 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
888 * Check events in order.
891 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
893 /* threshold event is triggered in finer grain than soft limit */
894 if (unlikely(mem_cgroup_event_ratelimit(memcg,
895 MEM_CGROUP_TARGET_THRESH))) {
898 do_softlimit = mem_cgroup_event_ratelimit(memcg,
899 MEM_CGROUP_TARGET_SOFTLIMIT);
900 mem_cgroup_threshold(memcg);
901 if (unlikely(do_softlimit))
902 mem_cgroup_update_tree(memcg, page);
906 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
909 * mm_update_next_owner() may clear mm->owner to NULL
910 * if it races with swapoff, page migration, etc.
911 * So this can be called with p == NULL.
916 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
918 EXPORT_SYMBOL(mem_cgroup_from_task);
921 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
922 * @mm: mm from which memcg should be extracted. It can be NULL.
924 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
925 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
928 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
930 struct mem_cgroup *memcg;
932 if (mem_cgroup_disabled())
938 * Page cache insertions can happen withou an
939 * actual mm context, e.g. during disk probing
940 * on boot, loopback IO, acct() writes etc.
943 memcg = root_mem_cgroup;
945 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
946 if (unlikely(!memcg))
947 memcg = root_mem_cgroup;
949 } while (!css_tryget(&memcg->css));
953 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
956 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
957 * @page: page from which memcg should be extracted.
959 * Obtain a reference on page->memcg and returns it if successful. Otherwise
960 * root_mem_cgroup is returned.
962 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
964 struct mem_cgroup *memcg = page->mem_cgroup;
966 if (mem_cgroup_disabled())
970 /* Page should not get uncharged and freed memcg under us. */
971 if (!memcg || WARN_ON_ONCE(!css_tryget(&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;
987 /* current->active_memcg must hold a ref. */
988 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
989 memcg = root_mem_cgroup;
991 memcg = current->active_memcg;
995 return get_mem_cgroup_from_mm(current->mm);
999 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1000 * @root: hierarchy root
1001 * @prev: previously returned memcg, NULL on first invocation
1002 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1004 * Returns references to children of the hierarchy below @root, or
1005 * @root itself, or %NULL after a full round-trip.
1007 * Caller must pass the return value in @prev on subsequent
1008 * invocations for reference counting, or use mem_cgroup_iter_break()
1009 * to cancel a hierarchy walk before the round-trip is complete.
1011 * Reclaimers can specify a node and a priority level in @reclaim to
1012 * divide up the memcgs in the hierarchy among all concurrent
1013 * reclaimers operating on the same node and priority.
1015 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1016 struct mem_cgroup *prev,
1017 struct mem_cgroup_reclaim_cookie *reclaim)
1019 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1020 struct cgroup_subsys_state *css = NULL;
1021 struct mem_cgroup *memcg = NULL;
1022 struct mem_cgroup *pos = NULL;
1024 if (mem_cgroup_disabled())
1028 root = root_mem_cgroup;
1030 if (prev && !reclaim)
1033 if (!root->use_hierarchy && root != root_mem_cgroup) {
1042 struct mem_cgroup_per_node *mz;
1044 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1047 if (prev && reclaim->generation != iter->generation)
1051 pos = READ_ONCE(iter->position);
1052 if (!pos || css_tryget(&pos->css))
1055 * css reference reached zero, so iter->position will
1056 * be cleared by ->css_released. However, we should not
1057 * rely on this happening soon, because ->css_released
1058 * is called from a work queue, and by busy-waiting we
1059 * might block it. So we clear iter->position right
1062 (void)cmpxchg(&iter->position, pos, NULL);
1070 css = css_next_descendant_pre(css, &root->css);
1073 * Reclaimers share the hierarchy walk, and a
1074 * new one might jump in right at the end of
1075 * the hierarchy - make sure they see at least
1076 * one group and restart from the beginning.
1084 * Verify the css and acquire a reference. The root
1085 * is provided by the caller, so we know it's alive
1086 * and kicking, and don't take an extra reference.
1088 memcg = mem_cgroup_from_css(css);
1090 if (css == &root->css)
1093 if (css_tryget(css))
1101 * The position could have already been updated by a competing
1102 * thread, so check that the value hasn't changed since we read
1103 * it to avoid reclaiming from the same cgroup twice.
1105 (void)cmpxchg(&iter->position, pos, memcg);
1113 reclaim->generation = iter->generation;
1119 if (prev && prev != root)
1120 css_put(&prev->css);
1126 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1127 * @root: hierarchy root
1128 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1130 void mem_cgroup_iter_break(struct mem_cgroup *root,
1131 struct mem_cgroup *prev)
1134 root = root_mem_cgroup;
1135 if (prev && prev != root)
1136 css_put(&prev->css);
1139 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1140 struct mem_cgroup *dead_memcg)
1142 struct mem_cgroup_reclaim_iter *iter;
1143 struct mem_cgroup_per_node *mz;
1146 for_each_node(nid) {
1147 mz = mem_cgroup_nodeinfo(from, nid);
1149 cmpxchg(&iter->position, dead_memcg, NULL);
1153 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1155 struct mem_cgroup *memcg = dead_memcg;
1156 struct mem_cgroup *last;
1159 __invalidate_reclaim_iterators(memcg, dead_memcg);
1161 } while ((memcg = parent_mem_cgroup(memcg)));
1164 * When cgruop1 non-hierarchy mode is used,
1165 * parent_mem_cgroup() does not walk all the way up to the
1166 * cgroup root (root_mem_cgroup). So we have to handle
1167 * dead_memcg from cgroup root separately.
1169 if (last != root_mem_cgroup)
1170 __invalidate_reclaim_iterators(root_mem_cgroup,
1175 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1176 * @memcg: hierarchy root
1177 * @fn: function to call for each task
1178 * @arg: argument passed to @fn
1180 * This function iterates over tasks attached to @memcg or to any of its
1181 * descendants and calls @fn for each task. If @fn returns a non-zero
1182 * value, the function breaks the iteration loop and returns the value.
1183 * Otherwise, it will iterate over all tasks and return 0.
1185 * This function must not be called for the root memory cgroup.
1187 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1188 int (*fn)(struct task_struct *, void *), void *arg)
1190 struct mem_cgroup *iter;
1193 BUG_ON(memcg == root_mem_cgroup);
1195 for_each_mem_cgroup_tree(iter, memcg) {
1196 struct css_task_iter it;
1197 struct task_struct *task;
1199 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1200 while (!ret && (task = css_task_iter_next(&it)))
1201 ret = fn(task, arg);
1202 css_task_iter_end(&it);
1204 mem_cgroup_iter_break(memcg, iter);
1212 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1214 * @pgdat: pgdat of the page
1216 * This function is only safe when following the LRU page isolation
1217 * and putback protocol: the LRU lock must be held, and the page must
1218 * either be PageLRU() or the caller must have isolated/allocated it.
1220 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1222 struct mem_cgroup_per_node *mz;
1223 struct mem_cgroup *memcg;
1224 struct lruvec *lruvec;
1226 if (mem_cgroup_disabled()) {
1227 lruvec = &pgdat->__lruvec;
1231 memcg = page->mem_cgroup;
1233 * Swapcache readahead pages are added to the LRU - and
1234 * possibly migrated - before they are charged.
1237 memcg = root_mem_cgroup;
1239 mz = mem_cgroup_page_nodeinfo(memcg, page);
1240 lruvec = &mz->lruvec;
1243 * Since a node can be onlined after the mem_cgroup was created,
1244 * we have to be prepared to initialize lruvec->zone here;
1245 * and if offlined then reonlined, we need to reinitialize it.
1247 if (unlikely(lruvec->pgdat != pgdat))
1248 lruvec->pgdat = pgdat;
1253 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1254 * @lruvec: mem_cgroup per zone lru vector
1255 * @lru: index of lru list the page is sitting on
1256 * @zid: zone id of the accounted pages
1257 * @nr_pages: positive when adding or negative when removing
1259 * This function must be called under lru_lock, just before a page is added
1260 * to or just after a page is removed from an lru list (that ordering being
1261 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1263 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1264 int zid, int nr_pages)
1266 struct mem_cgroup_per_node *mz;
1267 unsigned long *lru_size;
1270 if (mem_cgroup_disabled())
1273 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1274 lru_size = &mz->lru_zone_size[zid][lru];
1277 *lru_size += nr_pages;
1280 if (WARN_ONCE(size < 0,
1281 "%s(%p, %d, %d): lru_size %ld\n",
1282 __func__, lruvec, lru, nr_pages, size)) {
1288 *lru_size += nr_pages;
1292 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1293 * @memcg: the memory cgroup
1295 * Returns the maximum amount of memory @mem can be charged with, in
1298 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1300 unsigned long margin = 0;
1301 unsigned long count;
1302 unsigned long limit;
1304 count = page_counter_read(&memcg->memory);
1305 limit = READ_ONCE(memcg->memory.max);
1307 margin = limit - count;
1309 if (do_memsw_account()) {
1310 count = page_counter_read(&memcg->memsw);
1311 limit = READ_ONCE(memcg->memsw.max);
1313 margin = min(margin, limit - count);
1322 * A routine for checking "mem" is under move_account() or not.
1324 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1325 * moving cgroups. This is for waiting at high-memory pressure
1328 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1330 struct mem_cgroup *from;
1331 struct mem_cgroup *to;
1334 * Unlike task_move routines, we access mc.to, mc.from not under
1335 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1337 spin_lock(&mc.lock);
1343 ret = mem_cgroup_is_descendant(from, memcg) ||
1344 mem_cgroup_is_descendant(to, memcg);
1346 spin_unlock(&mc.lock);
1350 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1352 if (mc.moving_task && current != mc.moving_task) {
1353 if (mem_cgroup_under_move(memcg)) {
1355 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1356 /* moving charge context might have finished. */
1359 finish_wait(&mc.waitq, &wait);
1366 static char *memory_stat_format(struct mem_cgroup *memcg)
1371 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1376 * Provide statistics on the state of the memory subsystem as
1377 * well as cumulative event counters that show past behavior.
1379 * This list is ordered following a combination of these gradients:
1380 * 1) generic big picture -> specifics and details
1381 * 2) reflecting userspace activity -> reflecting kernel heuristics
1383 * Current memory state:
1386 seq_buf_printf(&s, "anon %llu\n",
1387 (u64)memcg_page_state(memcg, MEMCG_RSS) *
1389 seq_buf_printf(&s, "file %llu\n",
1390 (u64)memcg_page_state(memcg, NR_FILE_PAGES) *
1392 seq_buf_printf(&s, "kernel_stack %llu\n",
1393 (u64)memcg_page_state(memcg, MEMCG_KERNEL_STACK_KB) *
1395 seq_buf_printf(&s, "slab %llu\n",
1396 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) +
1397 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE)) *
1399 seq_buf_printf(&s, "sock %llu\n",
1400 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1403 seq_buf_printf(&s, "shmem %llu\n",
1404 (u64)memcg_page_state(memcg, NR_SHMEM) *
1406 seq_buf_printf(&s, "file_mapped %llu\n",
1407 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1409 seq_buf_printf(&s, "file_dirty %llu\n",
1410 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1412 seq_buf_printf(&s, "file_writeback %llu\n",
1413 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1417 * TODO: We should eventually replace our own MEMCG_RSS_HUGE counter
1418 * with the NR_ANON_THP vm counter, but right now it's a pain in the
1419 * arse because it requires migrating the work out of rmap to a place
1420 * where the page->mem_cgroup is set up and stable.
1422 seq_buf_printf(&s, "anon_thp %llu\n",
1423 (u64)memcg_page_state(memcg, MEMCG_RSS_HUGE) *
1426 for (i = 0; i < NR_LRU_LISTS; i++)
1427 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1428 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1431 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1432 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE) *
1434 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1435 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE) *
1438 /* Accumulated memory events */
1440 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1441 memcg_events(memcg, PGFAULT));
1442 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1443 memcg_events(memcg, PGMAJFAULT));
1445 seq_buf_printf(&s, "workingset_refault %lu\n",
1446 memcg_page_state(memcg, WORKINGSET_REFAULT));
1447 seq_buf_printf(&s, "workingset_activate %lu\n",
1448 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1449 seq_buf_printf(&s, "workingset_restore %lu\n",
1450 memcg_page_state(memcg, WORKINGSET_RESTORE));
1451 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1452 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1454 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1455 memcg_events(memcg, PGREFILL));
1456 seq_buf_printf(&s, "pgscan %lu\n",
1457 memcg_events(memcg, PGSCAN_KSWAPD) +
1458 memcg_events(memcg, PGSCAN_DIRECT));
1459 seq_buf_printf(&s, "pgsteal %lu\n",
1460 memcg_events(memcg, PGSTEAL_KSWAPD) +
1461 memcg_events(memcg, PGSTEAL_DIRECT));
1462 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1463 memcg_events(memcg, PGACTIVATE));
1464 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1465 memcg_events(memcg, PGDEACTIVATE));
1466 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1467 memcg_events(memcg, PGLAZYFREE));
1468 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1469 memcg_events(memcg, PGLAZYFREED));
1471 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1472 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1473 memcg_events(memcg, THP_FAULT_ALLOC));
1474 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1475 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1476 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1478 /* The above should easily fit into one page */
1479 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1484 #define K(x) ((x) << (PAGE_SHIFT-10))
1486 * mem_cgroup_print_oom_context: Print OOM information relevant to
1487 * memory controller.
1488 * @memcg: The memory cgroup that went over limit
1489 * @p: Task that is going to be killed
1491 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1494 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1499 pr_cont(",oom_memcg=");
1500 pr_cont_cgroup_path(memcg->css.cgroup);
1502 pr_cont(",global_oom");
1504 pr_cont(",task_memcg=");
1505 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1511 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1512 * memory controller.
1513 * @memcg: The memory cgroup that went over limit
1515 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1519 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1520 K((u64)page_counter_read(&memcg->memory)),
1521 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1522 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1523 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1524 K((u64)page_counter_read(&memcg->swap)),
1525 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1527 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1528 K((u64)page_counter_read(&memcg->memsw)),
1529 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1530 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1531 K((u64)page_counter_read(&memcg->kmem)),
1532 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1535 pr_info("Memory cgroup stats for ");
1536 pr_cont_cgroup_path(memcg->css.cgroup);
1538 buf = memory_stat_format(memcg);
1546 * Return the memory (and swap, if configured) limit for a memcg.
1548 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1552 max = READ_ONCE(memcg->memory.max);
1553 if (mem_cgroup_swappiness(memcg)) {
1554 unsigned long memsw_max;
1555 unsigned long swap_max;
1557 memsw_max = memcg->memsw.max;
1558 swap_max = READ_ONCE(memcg->swap.max);
1559 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1560 max = min(max + swap_max, memsw_max);
1565 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1567 return page_counter_read(&memcg->memory);
1570 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1573 struct oom_control oc = {
1577 .gfp_mask = gfp_mask,
1582 if (mutex_lock_killable(&oom_lock))
1585 * A few threads which were not waiting at mutex_lock_killable() can
1586 * fail to bail out. Therefore, check again after holding oom_lock.
1588 ret = should_force_charge() || out_of_memory(&oc);
1589 mutex_unlock(&oom_lock);
1593 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1596 unsigned long *total_scanned)
1598 struct mem_cgroup *victim = NULL;
1601 unsigned long excess;
1602 unsigned long nr_scanned;
1603 struct mem_cgroup_reclaim_cookie reclaim = {
1607 excess = soft_limit_excess(root_memcg);
1610 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1615 * If we have not been able to reclaim
1616 * anything, it might because there are
1617 * no reclaimable pages under this hierarchy
1622 * We want to do more targeted reclaim.
1623 * excess >> 2 is not to excessive so as to
1624 * reclaim too much, nor too less that we keep
1625 * coming back to reclaim from this cgroup
1627 if (total >= (excess >> 2) ||
1628 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1633 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1634 pgdat, &nr_scanned);
1635 *total_scanned += nr_scanned;
1636 if (!soft_limit_excess(root_memcg))
1639 mem_cgroup_iter_break(root_memcg, victim);
1643 #ifdef CONFIG_LOCKDEP
1644 static struct lockdep_map memcg_oom_lock_dep_map = {
1645 .name = "memcg_oom_lock",
1649 static DEFINE_SPINLOCK(memcg_oom_lock);
1652 * Check OOM-Killer is already running under our hierarchy.
1653 * If someone is running, return false.
1655 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1657 struct mem_cgroup *iter, *failed = NULL;
1659 spin_lock(&memcg_oom_lock);
1661 for_each_mem_cgroup_tree(iter, memcg) {
1662 if (iter->oom_lock) {
1664 * this subtree of our hierarchy is already locked
1665 * so we cannot give a lock.
1668 mem_cgroup_iter_break(memcg, iter);
1671 iter->oom_lock = true;
1676 * OK, we failed to lock the whole subtree so we have
1677 * to clean up what we set up to the failing subtree
1679 for_each_mem_cgroup_tree(iter, memcg) {
1680 if (iter == failed) {
1681 mem_cgroup_iter_break(memcg, iter);
1684 iter->oom_lock = false;
1687 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1689 spin_unlock(&memcg_oom_lock);
1694 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1696 struct mem_cgroup *iter;
1698 spin_lock(&memcg_oom_lock);
1699 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1700 for_each_mem_cgroup_tree(iter, memcg)
1701 iter->oom_lock = false;
1702 spin_unlock(&memcg_oom_lock);
1705 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1707 struct mem_cgroup *iter;
1709 spin_lock(&memcg_oom_lock);
1710 for_each_mem_cgroup_tree(iter, memcg)
1712 spin_unlock(&memcg_oom_lock);
1715 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1717 struct mem_cgroup *iter;
1720 * When a new child is created while the hierarchy is under oom,
1721 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1723 spin_lock(&memcg_oom_lock);
1724 for_each_mem_cgroup_tree(iter, memcg)
1725 if (iter->under_oom > 0)
1727 spin_unlock(&memcg_oom_lock);
1730 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1732 struct oom_wait_info {
1733 struct mem_cgroup *memcg;
1734 wait_queue_entry_t wait;
1737 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1738 unsigned mode, int sync, void *arg)
1740 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1741 struct mem_cgroup *oom_wait_memcg;
1742 struct oom_wait_info *oom_wait_info;
1744 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1745 oom_wait_memcg = oom_wait_info->memcg;
1747 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1748 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1750 return autoremove_wake_function(wait, mode, sync, arg);
1753 static void memcg_oom_recover(struct mem_cgroup *memcg)
1756 * For the following lockless ->under_oom test, the only required
1757 * guarantee is that it must see the state asserted by an OOM when
1758 * this function is called as a result of userland actions
1759 * triggered by the notification of the OOM. This is trivially
1760 * achieved by invoking mem_cgroup_mark_under_oom() before
1761 * triggering notification.
1763 if (memcg && memcg->under_oom)
1764 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1774 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1776 enum oom_status ret;
1779 if (order > PAGE_ALLOC_COSTLY_ORDER)
1782 memcg_memory_event(memcg, MEMCG_OOM);
1785 * We are in the middle of the charge context here, so we
1786 * don't want to block when potentially sitting on a callstack
1787 * that holds all kinds of filesystem and mm locks.
1789 * cgroup1 allows disabling the OOM killer and waiting for outside
1790 * handling until the charge can succeed; remember the context and put
1791 * the task to sleep at the end of the page fault when all locks are
1794 * On the other hand, in-kernel OOM killer allows for an async victim
1795 * memory reclaim (oom_reaper) and that means that we are not solely
1796 * relying on the oom victim to make a forward progress and we can
1797 * invoke the oom killer here.
1799 * Please note that mem_cgroup_out_of_memory might fail to find a
1800 * victim and then we have to bail out from the charge path.
1802 if (memcg->oom_kill_disable) {
1803 if (!current->in_user_fault)
1805 css_get(&memcg->css);
1806 current->memcg_in_oom = memcg;
1807 current->memcg_oom_gfp_mask = mask;
1808 current->memcg_oom_order = order;
1813 mem_cgroup_mark_under_oom(memcg);
1815 locked = mem_cgroup_oom_trylock(memcg);
1818 mem_cgroup_oom_notify(memcg);
1820 mem_cgroup_unmark_under_oom(memcg);
1821 if (mem_cgroup_out_of_memory(memcg, mask, order))
1827 mem_cgroup_oom_unlock(memcg);
1833 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1834 * @handle: actually kill/wait or just clean up the OOM state
1836 * This has to be called at the end of a page fault if the memcg OOM
1837 * handler was enabled.
1839 * Memcg supports userspace OOM handling where failed allocations must
1840 * sleep on a waitqueue until the userspace task resolves the
1841 * situation. Sleeping directly in the charge context with all kinds
1842 * of locks held is not a good idea, instead we remember an OOM state
1843 * in the task and mem_cgroup_oom_synchronize() has to be called at
1844 * the end of the page fault to complete the OOM handling.
1846 * Returns %true if an ongoing memcg OOM situation was detected and
1847 * completed, %false otherwise.
1849 bool mem_cgroup_oom_synchronize(bool handle)
1851 struct mem_cgroup *memcg = current->memcg_in_oom;
1852 struct oom_wait_info owait;
1855 /* OOM is global, do not handle */
1862 owait.memcg = memcg;
1863 owait.wait.flags = 0;
1864 owait.wait.func = memcg_oom_wake_function;
1865 owait.wait.private = current;
1866 INIT_LIST_HEAD(&owait.wait.entry);
1868 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1869 mem_cgroup_mark_under_oom(memcg);
1871 locked = mem_cgroup_oom_trylock(memcg);
1874 mem_cgroup_oom_notify(memcg);
1876 if (locked && !memcg->oom_kill_disable) {
1877 mem_cgroup_unmark_under_oom(memcg);
1878 finish_wait(&memcg_oom_waitq, &owait.wait);
1879 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1880 current->memcg_oom_order);
1883 mem_cgroup_unmark_under_oom(memcg);
1884 finish_wait(&memcg_oom_waitq, &owait.wait);
1888 mem_cgroup_oom_unlock(memcg);
1890 * There is no guarantee that an OOM-lock contender
1891 * sees the wakeups triggered by the OOM kill
1892 * uncharges. Wake any sleepers explicitely.
1894 memcg_oom_recover(memcg);
1897 current->memcg_in_oom = NULL;
1898 css_put(&memcg->css);
1903 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1904 * @victim: task to be killed by the OOM killer
1905 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1907 * Returns a pointer to a memory cgroup, which has to be cleaned up
1908 * by killing all belonging OOM-killable tasks.
1910 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
1912 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
1913 struct mem_cgroup *oom_domain)
1915 struct mem_cgroup *oom_group = NULL;
1916 struct mem_cgroup *memcg;
1918 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
1922 oom_domain = root_mem_cgroup;
1926 memcg = mem_cgroup_from_task(victim);
1927 if (memcg == root_mem_cgroup)
1931 * If the victim task has been asynchronously moved to a different
1932 * memory cgroup, we might end up killing tasks outside oom_domain.
1933 * In this case it's better to ignore memory.group.oom.
1935 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
1939 * Traverse the memory cgroup hierarchy from the victim task's
1940 * cgroup up to the OOMing cgroup (or root) to find the
1941 * highest-level memory cgroup with oom.group set.
1943 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
1944 if (memcg->oom_group)
1947 if (memcg == oom_domain)
1952 css_get(&oom_group->css);
1959 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
1961 pr_info("Tasks in ");
1962 pr_cont_cgroup_path(memcg->css.cgroup);
1963 pr_cont(" are going to be killed due to memory.oom.group set\n");
1967 * lock_page_memcg - lock a page->mem_cgroup binding
1970 * This function protects unlocked LRU pages from being moved to
1973 * It ensures lifetime of the returned memcg. Caller is responsible
1974 * for the lifetime of the page; __unlock_page_memcg() is available
1975 * when @page might get freed inside the locked section.
1977 struct mem_cgroup *lock_page_memcg(struct page *page)
1979 struct page *head = compound_head(page); /* rmap on tail pages */
1980 struct mem_cgroup *memcg;
1981 unsigned long flags;
1984 * The RCU lock is held throughout the transaction. The fast
1985 * path can get away without acquiring the memcg->move_lock
1986 * because page moving starts with an RCU grace period.
1988 * The RCU lock also protects the memcg from being freed when
1989 * the page state that is going to change is the only thing
1990 * preventing the page itself from being freed. E.g. writeback
1991 * doesn't hold a page reference and relies on PG_writeback to
1992 * keep off truncation, migration and so forth.
1996 if (mem_cgroup_disabled())
1999 memcg = head->mem_cgroup;
2000 if (unlikely(!memcg))
2003 if (atomic_read(&memcg->moving_account) <= 0)
2006 spin_lock_irqsave(&memcg->move_lock, flags);
2007 if (memcg != head->mem_cgroup) {
2008 spin_unlock_irqrestore(&memcg->move_lock, flags);
2013 * When charge migration first begins, we can have locked and
2014 * unlocked page stat updates happening concurrently. Track
2015 * the task who has the lock for unlock_page_memcg().
2017 memcg->move_lock_task = current;
2018 memcg->move_lock_flags = flags;
2022 EXPORT_SYMBOL(lock_page_memcg);
2025 * __unlock_page_memcg - unlock and unpin a memcg
2028 * Unlock and unpin a memcg returned by lock_page_memcg().
2030 void __unlock_page_memcg(struct mem_cgroup *memcg)
2032 if (memcg && memcg->move_lock_task == current) {
2033 unsigned long flags = memcg->move_lock_flags;
2035 memcg->move_lock_task = NULL;
2036 memcg->move_lock_flags = 0;
2038 spin_unlock_irqrestore(&memcg->move_lock, flags);
2045 * unlock_page_memcg - unlock a page->mem_cgroup binding
2048 void unlock_page_memcg(struct page *page)
2050 struct page *head = compound_head(page);
2052 __unlock_page_memcg(head->mem_cgroup);
2054 EXPORT_SYMBOL(unlock_page_memcg);
2056 struct memcg_stock_pcp {
2057 struct mem_cgroup *cached; /* this never be root cgroup */
2058 unsigned int nr_pages;
2059 struct work_struct work;
2060 unsigned long flags;
2061 #define FLUSHING_CACHED_CHARGE 0
2063 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2064 static DEFINE_MUTEX(percpu_charge_mutex);
2067 * consume_stock: Try to consume stocked charge on this cpu.
2068 * @memcg: memcg to consume from.
2069 * @nr_pages: how many pages to charge.
2071 * The charges will only happen if @memcg matches the current cpu's memcg
2072 * stock, and at least @nr_pages are available in that stock. Failure to
2073 * service an allocation will refill the stock.
2075 * returns true if successful, false otherwise.
2077 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2079 struct memcg_stock_pcp *stock;
2080 unsigned long flags;
2083 if (nr_pages > MEMCG_CHARGE_BATCH)
2086 local_irq_save(flags);
2088 stock = this_cpu_ptr(&memcg_stock);
2089 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2090 stock->nr_pages -= nr_pages;
2094 local_irq_restore(flags);
2100 * Returns stocks cached in percpu and reset cached information.
2102 static void drain_stock(struct memcg_stock_pcp *stock)
2104 struct mem_cgroup *old = stock->cached;
2106 if (stock->nr_pages) {
2107 page_counter_uncharge(&old->memory, stock->nr_pages);
2108 if (do_memsw_account())
2109 page_counter_uncharge(&old->memsw, stock->nr_pages);
2110 css_put_many(&old->css, stock->nr_pages);
2111 stock->nr_pages = 0;
2113 stock->cached = NULL;
2116 static void drain_local_stock(struct work_struct *dummy)
2118 struct memcg_stock_pcp *stock;
2119 unsigned long flags;
2122 * The only protection from memory hotplug vs. drain_stock races is
2123 * that we always operate on local CPU stock here with IRQ disabled
2125 local_irq_save(flags);
2127 stock = this_cpu_ptr(&memcg_stock);
2129 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2131 local_irq_restore(flags);
2135 * Cache charges(val) to local per_cpu area.
2136 * This will be consumed by consume_stock() function, later.
2138 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2140 struct memcg_stock_pcp *stock;
2141 unsigned long flags;
2143 local_irq_save(flags);
2145 stock = this_cpu_ptr(&memcg_stock);
2146 if (stock->cached != memcg) { /* reset if necessary */
2148 stock->cached = memcg;
2150 stock->nr_pages += nr_pages;
2152 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2155 local_irq_restore(flags);
2159 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2160 * of the hierarchy under it.
2162 static void drain_all_stock(struct mem_cgroup *root_memcg)
2166 /* If someone's already draining, avoid adding running more workers. */
2167 if (!mutex_trylock(&percpu_charge_mutex))
2170 * Notify other cpus that system-wide "drain" is running
2171 * We do not care about races with the cpu hotplug because cpu down
2172 * as well as workers from this path always operate on the local
2173 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2176 for_each_online_cpu(cpu) {
2177 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2178 struct mem_cgroup *memcg;
2182 memcg = stock->cached;
2183 if (memcg && stock->nr_pages &&
2184 mem_cgroup_is_descendant(memcg, root_memcg))
2189 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2191 drain_local_stock(&stock->work);
2193 schedule_work_on(cpu, &stock->work);
2197 mutex_unlock(&percpu_charge_mutex);
2200 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2202 struct memcg_stock_pcp *stock;
2203 struct mem_cgroup *memcg, *mi;
2205 stock = &per_cpu(memcg_stock, cpu);
2208 for_each_mem_cgroup(memcg) {
2211 for (i = 0; i < MEMCG_NR_STAT; i++) {
2215 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2217 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2218 atomic_long_add(x, &memcg->vmstats[i]);
2220 if (i >= NR_VM_NODE_STAT_ITEMS)
2223 for_each_node(nid) {
2224 struct mem_cgroup_per_node *pn;
2226 pn = mem_cgroup_nodeinfo(memcg, nid);
2227 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2230 atomic_long_add(x, &pn->lruvec_stat[i]);
2231 } while ((pn = parent_nodeinfo(pn, nid)));
2235 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2238 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2240 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2241 atomic_long_add(x, &memcg->vmevents[i]);
2248 static void reclaim_high(struct mem_cgroup *memcg,
2249 unsigned int nr_pages,
2253 if (page_counter_read(&memcg->memory) <=
2254 READ_ONCE(memcg->memory.high))
2256 memcg_memory_event(memcg, MEMCG_HIGH);
2257 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2258 } while ((memcg = parent_mem_cgroup(memcg)) &&
2259 !mem_cgroup_is_root(memcg));
2262 static void high_work_func(struct work_struct *work)
2264 struct mem_cgroup *memcg;
2266 memcg = container_of(work, struct mem_cgroup, high_work);
2267 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2271 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2272 * enough to still cause a significant slowdown in most cases, while still
2273 * allowing diagnostics and tracing to proceed without becoming stuck.
2275 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2278 * When calculating the delay, we use these either side of the exponentiation to
2279 * maintain precision and scale to a reasonable number of jiffies (see the table
2282 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2283 * overage ratio to a delay.
2284 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2285 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2286 * to produce a reasonable delay curve.
2288 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2289 * reasonable delay curve compared to precision-adjusted overage, not
2290 * penalising heavily at first, but still making sure that growth beyond the
2291 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2292 * example, with a high of 100 megabytes:
2294 * +-------+------------------------+
2295 * | usage | time to allocate in ms |
2296 * +-------+------------------------+
2318 * +-------+------------------------+
2320 #define MEMCG_DELAY_PRECISION_SHIFT 20
2321 #define MEMCG_DELAY_SCALING_SHIFT 14
2323 static u64 calculate_overage(unsigned long usage, unsigned long high)
2331 * Prevent division by 0 in overage calculation by acting as if
2332 * it was a threshold of 1 page
2334 high = max(high, 1UL);
2336 overage = usage - high;
2337 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2338 return div64_u64(overage, high);
2341 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2343 u64 overage, max_overage = 0;
2346 overage = calculate_overage(page_counter_read(&memcg->memory),
2347 READ_ONCE(memcg->memory.high));
2348 max_overage = max(overage, max_overage);
2349 } while ((memcg = parent_mem_cgroup(memcg)) &&
2350 !mem_cgroup_is_root(memcg));
2355 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2357 u64 overage, max_overage = 0;
2360 overage = calculate_overage(page_counter_read(&memcg->swap),
2361 READ_ONCE(memcg->swap.high));
2363 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2364 max_overage = max(overage, max_overage);
2365 } while ((memcg = parent_mem_cgroup(memcg)) &&
2366 !mem_cgroup_is_root(memcg));
2372 * Get the number of jiffies that we should penalise a mischievous cgroup which
2373 * is exceeding its memory.high by checking both it and its ancestors.
2375 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2376 unsigned int nr_pages,
2379 unsigned long penalty_jiffies;
2385 * We use overage compared to memory.high to calculate the number of
2386 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2387 * fairly lenient on small overages, and increasingly harsh when the
2388 * memcg in question makes it clear that it has no intention of stopping
2389 * its crazy behaviour, so we exponentially increase the delay based on
2392 penalty_jiffies = max_overage * max_overage * HZ;
2393 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2394 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2397 * Factor in the task's own contribution to the overage, such that four
2398 * N-sized allocations are throttled approximately the same as one
2399 * 4N-sized allocation.
2401 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2402 * larger the current charge patch is than that.
2404 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2408 * Scheduled by try_charge() to be executed from the userland return path
2409 * and reclaims memory over the high limit.
2411 void mem_cgroup_handle_over_high(void)
2413 unsigned long penalty_jiffies;
2414 unsigned long pflags;
2415 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2416 struct mem_cgroup *memcg;
2418 if (likely(!nr_pages))
2421 memcg = get_mem_cgroup_from_mm(current->mm);
2422 reclaim_high(memcg, nr_pages, GFP_KERNEL);
2423 current->memcg_nr_pages_over_high = 0;
2426 * memory.high is breached and reclaim is unable to keep up. Throttle
2427 * allocators proactively to slow down excessive growth.
2429 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2430 mem_find_max_overage(memcg));
2432 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2433 swap_find_max_overage(memcg));
2436 * Clamp the max delay per usermode return so as to still keep the
2437 * application moving forwards and also permit diagnostics, albeit
2440 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2443 * Don't sleep if the amount of jiffies this memcg owes us is so low
2444 * that it's not even worth doing, in an attempt to be nice to those who
2445 * go only a small amount over their memory.high value and maybe haven't
2446 * been aggressively reclaimed enough yet.
2448 if (penalty_jiffies <= HZ / 100)
2452 * If we exit early, we're guaranteed to die (since
2453 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2454 * need to account for any ill-begotten jiffies to pay them off later.
2456 psi_memstall_enter(&pflags);
2457 schedule_timeout_killable(penalty_jiffies);
2458 psi_memstall_leave(&pflags);
2461 css_put(&memcg->css);
2464 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2465 unsigned int nr_pages)
2467 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2468 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2469 struct mem_cgroup *mem_over_limit;
2470 struct page_counter *counter;
2471 unsigned long nr_reclaimed;
2472 bool may_swap = true;
2473 bool drained = false;
2474 enum oom_status oom_status;
2476 if (mem_cgroup_is_root(memcg))
2479 if (consume_stock(memcg, nr_pages))
2482 if (!do_memsw_account() ||
2483 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2484 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2486 if (do_memsw_account())
2487 page_counter_uncharge(&memcg->memsw, batch);
2488 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2490 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2494 if (batch > nr_pages) {
2500 * Memcg doesn't have a dedicated reserve for atomic
2501 * allocations. But like the global atomic pool, we need to
2502 * put the burden of reclaim on regular allocation requests
2503 * and let these go through as privileged allocations.
2505 if (gfp_mask & __GFP_ATOMIC)
2509 * Unlike in global OOM situations, memcg is not in a physical
2510 * memory shortage. Allow dying and OOM-killed tasks to
2511 * bypass the last charges so that they can exit quickly and
2512 * free their memory.
2514 if (unlikely(should_force_charge()))
2518 * Prevent unbounded recursion when reclaim operations need to
2519 * allocate memory. This might exceed the limits temporarily,
2520 * but we prefer facilitating memory reclaim and getting back
2521 * under the limit over triggering OOM kills in these cases.
2523 if (unlikely(current->flags & PF_MEMALLOC))
2526 if (unlikely(task_in_memcg_oom(current)))
2529 if (!gfpflags_allow_blocking(gfp_mask))
2532 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2534 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2535 gfp_mask, may_swap);
2537 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2541 drain_all_stock(mem_over_limit);
2546 if (gfp_mask & __GFP_NORETRY)
2549 * Even though the limit is exceeded at this point, reclaim
2550 * may have been able to free some pages. Retry the charge
2551 * before killing the task.
2553 * Only for regular pages, though: huge pages are rather
2554 * unlikely to succeed so close to the limit, and we fall back
2555 * to regular pages anyway in case of failure.
2557 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2560 * At task move, charge accounts can be doubly counted. So, it's
2561 * better to wait until the end of task_move if something is going on.
2563 if (mem_cgroup_wait_acct_move(mem_over_limit))
2569 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2572 if (gfp_mask & __GFP_NOFAIL)
2575 if (fatal_signal_pending(current))
2579 * keep retrying as long as the memcg oom killer is able to make
2580 * a forward progress or bypass the charge if the oom killer
2581 * couldn't make any progress.
2583 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2584 get_order(nr_pages * PAGE_SIZE));
2585 switch (oom_status) {
2587 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2595 if (!(gfp_mask & __GFP_NOFAIL))
2599 * The allocation either can't fail or will lead to more memory
2600 * being freed very soon. Allow memory usage go over the limit
2601 * temporarily by force charging it.
2603 page_counter_charge(&memcg->memory, nr_pages);
2604 if (do_memsw_account())
2605 page_counter_charge(&memcg->memsw, nr_pages);
2606 css_get_many(&memcg->css, nr_pages);
2611 css_get_many(&memcg->css, batch);
2612 if (batch > nr_pages)
2613 refill_stock(memcg, batch - nr_pages);
2616 * If the hierarchy is above the normal consumption range, schedule
2617 * reclaim on returning to userland. We can perform reclaim here
2618 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2619 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2620 * not recorded as it most likely matches current's and won't
2621 * change in the meantime. As high limit is checked again before
2622 * reclaim, the cost of mismatch is negligible.
2625 bool mem_high, swap_high;
2627 mem_high = page_counter_read(&memcg->memory) >
2628 READ_ONCE(memcg->memory.high);
2629 swap_high = page_counter_read(&memcg->swap) >
2630 READ_ONCE(memcg->swap.high);
2632 /* Don't bother a random interrupted task */
2633 if (in_interrupt()) {
2635 schedule_work(&memcg->high_work);
2641 if (mem_high || swap_high) {
2643 * The allocating tasks in this cgroup will need to do
2644 * reclaim or be throttled to prevent further growth
2645 * of the memory or swap footprints.
2647 * Target some best-effort fairness between the tasks,
2648 * and distribute reclaim work and delay penalties
2649 * based on how much each task is actually allocating.
2651 current->memcg_nr_pages_over_high += batch;
2652 set_notify_resume(current);
2655 } while ((memcg = parent_mem_cgroup(memcg)));
2660 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2662 if (mem_cgroup_is_root(memcg))
2665 page_counter_uncharge(&memcg->memory, nr_pages);
2666 if (do_memsw_account())
2667 page_counter_uncharge(&memcg->memsw, nr_pages);
2669 css_put_many(&memcg->css, nr_pages);
2672 static void lock_page_lru(struct page *page, int *isolated)
2674 pg_data_t *pgdat = page_pgdat(page);
2676 spin_lock_irq(&pgdat->lru_lock);
2677 if (PageLRU(page)) {
2678 struct lruvec *lruvec;
2680 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2682 del_page_from_lru_list(page, lruvec, page_lru(page));
2688 static void unlock_page_lru(struct page *page, int isolated)
2690 pg_data_t *pgdat = page_pgdat(page);
2693 struct lruvec *lruvec;
2695 lruvec = mem_cgroup_page_lruvec(page, pgdat);
2696 VM_BUG_ON_PAGE(PageLRU(page), page);
2698 add_page_to_lru_list(page, lruvec, page_lru(page));
2700 spin_unlock_irq(&pgdat->lru_lock);
2703 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2708 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2711 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2712 * may already be on some other mem_cgroup's LRU. Take care of it.
2715 lock_page_lru(page, &isolated);
2718 * Nobody should be changing or seriously looking at
2719 * page->mem_cgroup at this point:
2721 * - the page is uncharged
2723 * - the page is off-LRU
2725 * - an anonymous fault has exclusive page access, except for
2726 * a locked page table
2728 * - a page cache insertion, a swapin fault, or a migration
2729 * have the page locked
2731 page->mem_cgroup = memcg;
2734 unlock_page_lru(page, isolated);
2737 #ifdef CONFIG_MEMCG_KMEM
2739 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2741 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2742 * cgroup_mutex, etc.
2744 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2748 if (mem_cgroup_disabled())
2751 page = virt_to_head_page(p);
2754 * Slab pages don't have page->mem_cgroup set because corresponding
2755 * kmem caches can be reparented during the lifetime. That's why
2756 * memcg_from_slab_page() should be used instead.
2759 return memcg_from_slab_page(page);
2761 /* All other pages use page->mem_cgroup */
2762 return page->mem_cgroup;
2765 static int memcg_alloc_cache_id(void)
2770 id = ida_simple_get(&memcg_cache_ida,
2771 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2775 if (id < memcg_nr_cache_ids)
2779 * There's no space for the new id in memcg_caches arrays,
2780 * so we have to grow them.
2782 down_write(&memcg_cache_ids_sem);
2784 size = 2 * (id + 1);
2785 if (size < MEMCG_CACHES_MIN_SIZE)
2786 size = MEMCG_CACHES_MIN_SIZE;
2787 else if (size > MEMCG_CACHES_MAX_SIZE)
2788 size = MEMCG_CACHES_MAX_SIZE;
2790 err = memcg_update_all_caches(size);
2792 err = memcg_update_all_list_lrus(size);
2794 memcg_nr_cache_ids = size;
2796 up_write(&memcg_cache_ids_sem);
2799 ida_simple_remove(&memcg_cache_ida, id);
2805 static void memcg_free_cache_id(int id)
2807 ida_simple_remove(&memcg_cache_ida, id);
2810 struct memcg_kmem_cache_create_work {
2811 struct mem_cgroup *memcg;
2812 struct kmem_cache *cachep;
2813 struct work_struct work;
2816 static void memcg_kmem_cache_create_func(struct work_struct *w)
2818 struct memcg_kmem_cache_create_work *cw =
2819 container_of(w, struct memcg_kmem_cache_create_work, work);
2820 struct mem_cgroup *memcg = cw->memcg;
2821 struct kmem_cache *cachep = cw->cachep;
2823 memcg_create_kmem_cache(memcg, cachep);
2825 css_put(&memcg->css);
2830 * Enqueue the creation of a per-memcg kmem_cache.
2832 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2833 struct kmem_cache *cachep)
2835 struct memcg_kmem_cache_create_work *cw;
2837 if (!css_tryget_online(&memcg->css))
2840 cw = kmalloc(sizeof(*cw), GFP_NOWAIT | __GFP_NOWARN);
2845 cw->cachep = cachep;
2846 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2848 queue_work(memcg_kmem_cache_wq, &cw->work);
2851 static inline bool memcg_kmem_bypass(void)
2856 /* Allow remote memcg charging in kthread contexts. */
2857 if ((!current->mm || (current->flags & PF_KTHREAD)) &&
2858 !current->active_memcg)
2864 * memcg_kmem_get_cache: select the correct per-memcg cache for allocation
2865 * @cachep: the original global kmem cache
2867 * Return the kmem_cache we're supposed to use for a slab allocation.
2868 * We try to use the current memcg's version of the cache.
2870 * If the cache does not exist yet, if we are the first user of it, we
2871 * create it asynchronously in a workqueue and let the current allocation
2872 * go through with the original cache.
2874 * This function takes a reference to the cache it returns to assure it
2875 * won't get destroyed while we are working with it. Once the caller is
2876 * done with it, memcg_kmem_put_cache() must be called to release the
2879 struct kmem_cache *memcg_kmem_get_cache(struct kmem_cache *cachep)
2881 struct mem_cgroup *memcg;
2882 struct kmem_cache *memcg_cachep;
2883 struct memcg_cache_array *arr;
2886 VM_BUG_ON(!is_root_cache(cachep));
2888 if (memcg_kmem_bypass())
2893 if (unlikely(current->active_memcg))
2894 memcg = current->active_memcg;
2896 memcg = mem_cgroup_from_task(current);
2898 if (!memcg || memcg == root_mem_cgroup)
2901 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2905 arr = rcu_dereference(cachep->memcg_params.memcg_caches);
2908 * Make sure we will access the up-to-date value. The code updating
2909 * memcg_caches issues a write barrier to match the data dependency
2910 * barrier inside READ_ONCE() (see memcg_create_kmem_cache()).
2912 memcg_cachep = READ_ONCE(arr->entries[kmemcg_id]);
2915 * If we are in a safe context (can wait, and not in interrupt
2916 * context), we could be be predictable and return right away.
2917 * This would guarantee that the allocation being performed
2918 * already belongs in the new cache.
2920 * However, there are some clashes that can arrive from locking.
2921 * For instance, because we acquire the slab_mutex while doing
2922 * memcg_create_kmem_cache, this means no further allocation
2923 * could happen with the slab_mutex held. So it's better to
2926 * If the memcg is dying or memcg_cache is about to be released,
2927 * don't bother creating new kmem_caches. Because memcg_cachep
2928 * is ZEROed as the fist step of kmem offlining, we don't need
2929 * percpu_ref_tryget_live() here. css_tryget_online() check in
2930 * memcg_schedule_kmem_cache_create() will prevent us from
2931 * creation of a new kmem_cache.
2933 if (unlikely(!memcg_cachep))
2934 memcg_schedule_kmem_cache_create(memcg, cachep);
2935 else if (percpu_ref_tryget(&memcg_cachep->memcg_params.refcnt))
2936 cachep = memcg_cachep;
2943 * memcg_kmem_put_cache: drop reference taken by memcg_kmem_get_cache
2944 * @cachep: the cache returned by memcg_kmem_get_cache
2946 void memcg_kmem_put_cache(struct kmem_cache *cachep)
2948 if (!is_root_cache(cachep))
2949 percpu_ref_put(&cachep->memcg_params.refcnt);
2953 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2954 * @memcg: memory cgroup to charge
2955 * @gfp: reclaim mode
2956 * @nr_pages: number of pages to charge
2958 * Returns 0 on success, an error code on failure.
2960 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2961 unsigned int nr_pages)
2963 struct page_counter *counter;
2966 ret = try_charge(memcg, gfp, nr_pages);
2970 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2971 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2974 * Enforce __GFP_NOFAIL allocation because callers are not
2975 * prepared to see failures and likely do not have any failure
2978 if (gfp & __GFP_NOFAIL) {
2979 page_counter_charge(&memcg->kmem, nr_pages);
2982 cancel_charge(memcg, nr_pages);
2989 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
2990 * @memcg: memcg to uncharge
2991 * @nr_pages: number of pages to uncharge
2993 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
2995 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2996 page_counter_uncharge(&memcg->kmem, nr_pages);
2998 page_counter_uncharge(&memcg->memory, nr_pages);
2999 if (do_memsw_account())
3000 page_counter_uncharge(&memcg->memsw, nr_pages);
3004 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3005 * @page: page to charge
3006 * @gfp: reclaim mode
3007 * @order: allocation order
3009 * Returns 0 on success, an error code on failure.
3011 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3013 struct mem_cgroup *memcg;
3016 if (memcg_kmem_bypass())
3019 memcg = get_mem_cgroup_from_current();
3020 if (!mem_cgroup_is_root(memcg)) {
3021 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3023 page->mem_cgroup = memcg;
3024 __SetPageKmemcg(page);
3027 css_put(&memcg->css);
3032 * __memcg_kmem_uncharge_page: uncharge a kmem page
3033 * @page: page to uncharge
3034 * @order: allocation order
3036 void __memcg_kmem_uncharge_page(struct page *page, int order)
3038 struct mem_cgroup *memcg = page->mem_cgroup;
3039 unsigned int nr_pages = 1 << order;
3044 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3045 __memcg_kmem_uncharge(memcg, nr_pages);
3046 page->mem_cgroup = NULL;
3048 /* slab pages do not have PageKmemcg flag set */
3049 if (PageKmemcg(page))
3050 __ClearPageKmemcg(page);
3052 css_put_many(&memcg->css, nr_pages);
3054 #endif /* CONFIG_MEMCG_KMEM */
3056 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3059 * Because tail pages are not marked as "used", set it. We're under
3060 * pgdat->lru_lock and migration entries setup in all page mappings.
3062 void mem_cgroup_split_huge_fixup(struct page *head)
3066 if (mem_cgroup_disabled())
3069 for (i = 1; i < HPAGE_PMD_NR; i++)
3070 head[i].mem_cgroup = head->mem_cgroup;
3072 __mod_memcg_state(head->mem_cgroup, MEMCG_RSS_HUGE, -HPAGE_PMD_NR);
3074 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3076 #ifdef CONFIG_MEMCG_SWAP
3078 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3079 * @entry: swap entry to be moved
3080 * @from: mem_cgroup which the entry is moved from
3081 * @to: mem_cgroup which the entry is moved to
3083 * It succeeds only when the swap_cgroup's record for this entry is the same
3084 * as the mem_cgroup's id of @from.
3086 * Returns 0 on success, -EINVAL on failure.
3088 * The caller must have charged to @to, IOW, called page_counter_charge() about
3089 * both res and memsw, and called css_get().
3091 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3092 struct mem_cgroup *from, struct mem_cgroup *to)
3094 unsigned short old_id, new_id;
3096 old_id = mem_cgroup_id(from);
3097 new_id = mem_cgroup_id(to);
3099 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3100 mod_memcg_state(from, MEMCG_SWAP, -1);
3101 mod_memcg_state(to, MEMCG_SWAP, 1);
3107 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3108 struct mem_cgroup *from, struct mem_cgroup *to)
3114 static DEFINE_MUTEX(memcg_max_mutex);
3116 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3117 unsigned long max, bool memsw)
3119 bool enlarge = false;
3120 bool drained = false;
3122 bool limits_invariant;
3123 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3126 if (signal_pending(current)) {
3131 mutex_lock(&memcg_max_mutex);
3133 * Make sure that the new limit (memsw or memory limit) doesn't
3134 * break our basic invariant rule memory.max <= memsw.max.
3136 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3137 max <= memcg->memsw.max;
3138 if (!limits_invariant) {
3139 mutex_unlock(&memcg_max_mutex);
3143 if (max > counter->max)
3145 ret = page_counter_set_max(counter, max);
3146 mutex_unlock(&memcg_max_mutex);
3152 drain_all_stock(memcg);
3157 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3158 GFP_KERNEL, !memsw)) {
3164 if (!ret && enlarge)
3165 memcg_oom_recover(memcg);
3170 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3172 unsigned long *total_scanned)
3174 unsigned long nr_reclaimed = 0;
3175 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3176 unsigned long reclaimed;
3178 struct mem_cgroup_tree_per_node *mctz;
3179 unsigned long excess;
3180 unsigned long nr_scanned;
3185 mctz = soft_limit_tree_node(pgdat->node_id);
3188 * Do not even bother to check the largest node if the root
3189 * is empty. Do it lockless to prevent lock bouncing. Races
3190 * are acceptable as soft limit is best effort anyway.
3192 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3196 * This loop can run a while, specially if mem_cgroup's continuously
3197 * keep exceeding their soft limit and putting the system under
3204 mz = mem_cgroup_largest_soft_limit_node(mctz);
3209 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3210 gfp_mask, &nr_scanned);
3211 nr_reclaimed += reclaimed;
3212 *total_scanned += nr_scanned;
3213 spin_lock_irq(&mctz->lock);
3214 __mem_cgroup_remove_exceeded(mz, mctz);
3217 * If we failed to reclaim anything from this memory cgroup
3218 * it is time to move on to the next cgroup
3222 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3224 excess = soft_limit_excess(mz->memcg);
3226 * One school of thought says that we should not add
3227 * back the node to the tree if reclaim returns 0.
3228 * But our reclaim could return 0, simply because due
3229 * to priority we are exposing a smaller subset of
3230 * memory to reclaim from. Consider this as a longer
3233 /* If excess == 0, no tree ops */
3234 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3235 spin_unlock_irq(&mctz->lock);
3236 css_put(&mz->memcg->css);
3239 * Could not reclaim anything and there are no more
3240 * mem cgroups to try or we seem to be looping without
3241 * reclaiming anything.
3243 if (!nr_reclaimed &&
3245 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3247 } while (!nr_reclaimed);
3249 css_put(&next_mz->memcg->css);
3250 return nr_reclaimed;
3254 * Test whether @memcg has children, dead or alive. Note that this
3255 * function doesn't care whether @memcg has use_hierarchy enabled and
3256 * returns %true if there are child csses according to the cgroup
3257 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3259 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3264 ret = css_next_child(NULL, &memcg->css);
3270 * Reclaims as many pages from the given memcg as possible.
3272 * Caller is responsible for holding css reference for memcg.
3274 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3276 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3278 /* we call try-to-free pages for make this cgroup empty */
3279 lru_add_drain_all();
3281 drain_all_stock(memcg);
3283 /* try to free all pages in this cgroup */
3284 while (nr_retries && page_counter_read(&memcg->memory)) {
3287 if (signal_pending(current))
3290 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3294 /* maybe some writeback is necessary */
3295 congestion_wait(BLK_RW_ASYNC, HZ/10);
3303 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3304 char *buf, size_t nbytes,
3307 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3309 if (mem_cgroup_is_root(memcg))
3311 return mem_cgroup_force_empty(memcg) ?: nbytes;
3314 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3317 return mem_cgroup_from_css(css)->use_hierarchy;
3320 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3321 struct cftype *cft, u64 val)
3324 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3325 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3327 if (memcg->use_hierarchy == val)
3331 * If parent's use_hierarchy is set, we can't make any modifications
3332 * in the child subtrees. If it is unset, then the change can
3333 * occur, provided the current cgroup has no children.
3335 * For the root cgroup, parent_mem is NULL, we allow value to be
3336 * set if there are no children.
3338 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3339 (val == 1 || val == 0)) {
3340 if (!memcg_has_children(memcg))
3341 memcg->use_hierarchy = val;
3350 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3354 if (mem_cgroup_is_root(memcg)) {
3355 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3356 memcg_page_state(memcg, MEMCG_RSS);
3358 val += memcg_page_state(memcg, MEMCG_SWAP);
3361 val = page_counter_read(&memcg->memory);
3363 val = page_counter_read(&memcg->memsw);
3376 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3380 struct page_counter *counter;
3382 switch (MEMFILE_TYPE(cft->private)) {
3384 counter = &memcg->memory;
3387 counter = &memcg->memsw;
3390 counter = &memcg->kmem;
3393 counter = &memcg->tcpmem;
3399 switch (MEMFILE_ATTR(cft->private)) {
3401 if (counter == &memcg->memory)
3402 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3403 if (counter == &memcg->memsw)
3404 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3405 return (u64)page_counter_read(counter) * PAGE_SIZE;
3407 return (u64)counter->max * PAGE_SIZE;
3409 return (u64)counter->watermark * PAGE_SIZE;
3411 return counter->failcnt;
3412 case RES_SOFT_LIMIT:
3413 return (u64)memcg->soft_limit * PAGE_SIZE;
3419 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3421 unsigned long stat[MEMCG_NR_STAT] = {0};
3422 struct mem_cgroup *mi;
3425 for_each_online_cpu(cpu)
3426 for (i = 0; i < MEMCG_NR_STAT; i++)
3427 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3429 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3430 for (i = 0; i < MEMCG_NR_STAT; i++)
3431 atomic_long_add(stat[i], &mi->vmstats[i]);
3433 for_each_node(node) {
3434 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3435 struct mem_cgroup_per_node *pi;
3437 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3440 for_each_online_cpu(cpu)
3441 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3443 pn->lruvec_stat_cpu->count[i], cpu);
3445 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3446 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3447 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3451 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3453 unsigned long events[NR_VM_EVENT_ITEMS];
3454 struct mem_cgroup *mi;
3457 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3460 for_each_online_cpu(cpu)
3461 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3462 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3465 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3466 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3467 atomic_long_add(events[i], &mi->vmevents[i]);
3470 #ifdef CONFIG_MEMCG_KMEM
3471 static int memcg_online_kmem(struct mem_cgroup *memcg)
3475 if (cgroup_memory_nokmem)
3478 BUG_ON(memcg->kmemcg_id >= 0);
3479 BUG_ON(memcg->kmem_state);
3481 memcg_id = memcg_alloc_cache_id();
3485 static_branch_inc(&memcg_kmem_enabled_key);
3487 * A memory cgroup is considered kmem-online as soon as it gets
3488 * kmemcg_id. Setting the id after enabling static branching will
3489 * guarantee no one starts accounting before all call sites are
3492 memcg->kmemcg_id = memcg_id;
3493 memcg->kmem_state = KMEM_ONLINE;
3494 INIT_LIST_HEAD(&memcg->kmem_caches);
3499 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3501 struct cgroup_subsys_state *css;
3502 struct mem_cgroup *parent, *child;
3505 if (memcg->kmem_state != KMEM_ONLINE)
3508 * Clear the online state before clearing memcg_caches array
3509 * entries. The slab_mutex in memcg_deactivate_kmem_caches()
3510 * guarantees that no cache will be created for this cgroup
3511 * after we are done (see memcg_create_kmem_cache()).
3513 memcg->kmem_state = KMEM_ALLOCATED;
3515 parent = parent_mem_cgroup(memcg);
3517 parent = root_mem_cgroup;
3520 * Deactivate and reparent kmem_caches.
3522 memcg_deactivate_kmem_caches(memcg, parent);
3524 kmemcg_id = memcg->kmemcg_id;
3525 BUG_ON(kmemcg_id < 0);
3528 * Change kmemcg_id of this cgroup and all its descendants to the
3529 * parent's id, and then move all entries from this cgroup's list_lrus
3530 * to ones of the parent. After we have finished, all list_lrus
3531 * corresponding to this cgroup are guaranteed to remain empty. The
3532 * ordering is imposed by list_lru_node->lock taken by
3533 * memcg_drain_all_list_lrus().
3535 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3536 css_for_each_descendant_pre(css, &memcg->css) {
3537 child = mem_cgroup_from_css(css);
3538 BUG_ON(child->kmemcg_id != kmemcg_id);
3539 child->kmemcg_id = parent->kmemcg_id;
3540 if (!memcg->use_hierarchy)
3545 memcg_drain_all_list_lrus(kmemcg_id, parent);
3547 memcg_free_cache_id(kmemcg_id);
3550 static void memcg_free_kmem(struct mem_cgroup *memcg)
3552 /* css_alloc() failed, offlining didn't happen */
3553 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3554 memcg_offline_kmem(memcg);
3556 if (memcg->kmem_state == KMEM_ALLOCATED) {
3557 WARN_ON(!list_empty(&memcg->kmem_caches));
3558 static_branch_dec(&memcg_kmem_enabled_key);
3562 static int memcg_online_kmem(struct mem_cgroup *memcg)
3566 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3569 static void memcg_free_kmem(struct mem_cgroup *memcg)
3572 #endif /* CONFIG_MEMCG_KMEM */
3574 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3579 mutex_lock(&memcg_max_mutex);
3580 ret = page_counter_set_max(&memcg->kmem, max);
3581 mutex_unlock(&memcg_max_mutex);
3585 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3589 mutex_lock(&memcg_max_mutex);
3591 ret = page_counter_set_max(&memcg->tcpmem, max);
3595 if (!memcg->tcpmem_active) {
3597 * The active flag needs to be written after the static_key
3598 * update. This is what guarantees that the socket activation
3599 * function is the last one to run. See mem_cgroup_sk_alloc()
3600 * for details, and note that we don't mark any socket as
3601 * belonging to this memcg until that flag is up.
3603 * We need to do this, because static_keys will span multiple
3604 * sites, but we can't control their order. If we mark a socket
3605 * as accounted, but the accounting functions are not patched in
3606 * yet, we'll lose accounting.
3608 * We never race with the readers in mem_cgroup_sk_alloc(),
3609 * because when this value change, the code to process it is not
3612 static_branch_inc(&memcg_sockets_enabled_key);
3613 memcg->tcpmem_active = true;
3616 mutex_unlock(&memcg_max_mutex);
3621 * The user of this function is...
3624 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3625 char *buf, size_t nbytes, loff_t off)
3627 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3628 unsigned long nr_pages;
3631 buf = strstrip(buf);
3632 ret = page_counter_memparse(buf, "-1", &nr_pages);
3636 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3638 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3642 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3644 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3647 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3650 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3651 "Please report your usecase to linux-mm@kvack.org if you "
3652 "depend on this functionality.\n");
3653 ret = memcg_update_kmem_max(memcg, nr_pages);
3656 ret = memcg_update_tcp_max(memcg, nr_pages);
3660 case RES_SOFT_LIMIT:
3661 memcg->soft_limit = nr_pages;
3665 return ret ?: nbytes;
3668 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3669 size_t nbytes, loff_t off)
3671 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3672 struct page_counter *counter;
3674 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3676 counter = &memcg->memory;
3679 counter = &memcg->memsw;
3682 counter = &memcg->kmem;
3685 counter = &memcg->tcpmem;
3691 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3693 page_counter_reset_watermark(counter);
3696 counter->failcnt = 0;
3705 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3708 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3712 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3713 struct cftype *cft, u64 val)
3715 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3717 if (val & ~MOVE_MASK)
3721 * No kind of locking is needed in here, because ->can_attach() will
3722 * check this value once in the beginning of the process, and then carry
3723 * on with stale data. This means that changes to this value will only
3724 * affect task migrations starting after the change.
3726 memcg->move_charge_at_immigrate = val;
3730 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3731 struct cftype *cft, u64 val)
3739 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3740 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3741 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3743 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3744 int nid, unsigned int lru_mask, bool tree)
3746 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3747 unsigned long nr = 0;
3750 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3753 if (!(BIT(lru) & lru_mask))
3756 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3758 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3763 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3764 unsigned int lru_mask,
3767 unsigned long nr = 0;
3771 if (!(BIT(lru) & lru_mask))
3774 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3776 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3781 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3785 unsigned int lru_mask;
3788 static const struct numa_stat stats[] = {
3789 { "total", LRU_ALL },
3790 { "file", LRU_ALL_FILE },
3791 { "anon", LRU_ALL_ANON },
3792 { "unevictable", BIT(LRU_UNEVICTABLE) },
3794 const struct numa_stat *stat;
3796 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3798 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3799 seq_printf(m, "%s=%lu", stat->name,
3800 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3802 for_each_node_state(nid, N_MEMORY)
3803 seq_printf(m, " N%d=%lu", nid,
3804 mem_cgroup_node_nr_lru_pages(memcg, nid,
3805 stat->lru_mask, false));
3809 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3811 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3812 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3814 for_each_node_state(nid, N_MEMORY)
3815 seq_printf(m, " N%d=%lu", nid,
3816 mem_cgroup_node_nr_lru_pages(memcg, nid,
3817 stat->lru_mask, true));
3823 #endif /* CONFIG_NUMA */
3825 static const unsigned int memcg1_stats[] = {
3836 static const char *const memcg1_stat_names[] = {
3847 /* Universal VM events cgroup1 shows, original sort order */
3848 static const unsigned int memcg1_events[] = {
3855 static int memcg_stat_show(struct seq_file *m, void *v)
3857 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3858 unsigned long memory, memsw;
3859 struct mem_cgroup *mi;
3862 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3864 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3865 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3867 seq_printf(m, "%s %lu\n", memcg1_stat_names[i],
3868 memcg_page_state_local(memcg, memcg1_stats[i]) *
3872 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3873 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3874 memcg_events_local(memcg, memcg1_events[i]));
3876 for (i = 0; i < NR_LRU_LISTS; i++)
3877 seq_printf(m, "%s %lu\n", lru_list_name(i),
3878 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3881 /* Hierarchical information */
3882 memory = memsw = PAGE_COUNTER_MAX;
3883 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3884 memory = min(memory, READ_ONCE(mi->memory.max));
3885 memsw = min(memsw, READ_ONCE(mi->memsw.max));
3887 seq_printf(m, "hierarchical_memory_limit %llu\n",
3888 (u64)memory * PAGE_SIZE);
3889 if (do_memsw_account())
3890 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3891 (u64)memsw * PAGE_SIZE);
3893 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3894 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3896 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
3897 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
3901 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3902 seq_printf(m, "total_%s %llu\n",
3903 vm_event_name(memcg1_events[i]),
3904 (u64)memcg_events(memcg, memcg1_events[i]));
3906 for (i = 0; i < NR_LRU_LISTS; i++)
3907 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
3908 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
3911 #ifdef CONFIG_DEBUG_VM
3914 struct mem_cgroup_per_node *mz;
3915 struct zone_reclaim_stat *rstat;
3916 unsigned long recent_rotated[2] = {0, 0};
3917 unsigned long recent_scanned[2] = {0, 0};
3919 for_each_online_pgdat(pgdat) {
3920 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
3921 rstat = &mz->lruvec.reclaim_stat;
3923 recent_rotated[0] += rstat->recent_rotated[0];
3924 recent_rotated[1] += rstat->recent_rotated[1];
3925 recent_scanned[0] += rstat->recent_scanned[0];
3926 recent_scanned[1] += rstat->recent_scanned[1];
3928 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3929 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3930 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3931 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3938 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3941 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3943 return mem_cgroup_swappiness(memcg);
3946 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3947 struct cftype *cft, u64 val)
3949 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3955 memcg->swappiness = val;
3957 vm_swappiness = val;
3962 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3964 struct mem_cgroup_threshold_ary *t;
3965 unsigned long usage;
3970 t = rcu_dereference(memcg->thresholds.primary);
3972 t = rcu_dereference(memcg->memsw_thresholds.primary);
3977 usage = mem_cgroup_usage(memcg, swap);
3980 * current_threshold points to threshold just below or equal to usage.
3981 * If it's not true, a threshold was crossed after last
3982 * call of __mem_cgroup_threshold().
3984 i = t->current_threshold;
3987 * Iterate backward over array of thresholds starting from
3988 * current_threshold and check if a threshold is crossed.
3989 * If none of thresholds below usage is crossed, we read
3990 * only one element of the array here.
3992 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3993 eventfd_signal(t->entries[i].eventfd, 1);
3995 /* i = current_threshold + 1 */
3999 * Iterate forward over array of thresholds starting from
4000 * current_threshold+1 and check if a threshold is crossed.
4001 * If none of thresholds above usage is crossed, we read
4002 * only one element of the array here.
4004 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4005 eventfd_signal(t->entries[i].eventfd, 1);
4007 /* Update current_threshold */
4008 t->current_threshold = i - 1;
4013 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4016 __mem_cgroup_threshold(memcg, false);
4017 if (do_memsw_account())
4018 __mem_cgroup_threshold(memcg, true);
4020 memcg = parent_mem_cgroup(memcg);
4024 static int compare_thresholds(const void *a, const void *b)
4026 const struct mem_cgroup_threshold *_a = a;
4027 const struct mem_cgroup_threshold *_b = b;
4029 if (_a->threshold > _b->threshold)
4032 if (_a->threshold < _b->threshold)
4038 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4040 struct mem_cgroup_eventfd_list *ev;
4042 spin_lock(&memcg_oom_lock);
4044 list_for_each_entry(ev, &memcg->oom_notify, list)
4045 eventfd_signal(ev->eventfd, 1);
4047 spin_unlock(&memcg_oom_lock);
4051 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4053 struct mem_cgroup *iter;
4055 for_each_mem_cgroup_tree(iter, memcg)
4056 mem_cgroup_oom_notify_cb(iter);
4059 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4060 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4062 struct mem_cgroup_thresholds *thresholds;
4063 struct mem_cgroup_threshold_ary *new;
4064 unsigned long threshold;
4065 unsigned long usage;
4068 ret = page_counter_memparse(args, "-1", &threshold);
4072 mutex_lock(&memcg->thresholds_lock);
4075 thresholds = &memcg->thresholds;
4076 usage = mem_cgroup_usage(memcg, false);
4077 } else if (type == _MEMSWAP) {
4078 thresholds = &memcg->memsw_thresholds;
4079 usage = mem_cgroup_usage(memcg, true);
4083 /* Check if a threshold crossed before adding a new one */
4084 if (thresholds->primary)
4085 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4087 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4089 /* Allocate memory for new array of thresholds */
4090 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4097 /* Copy thresholds (if any) to new array */
4098 if (thresholds->primary) {
4099 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4100 sizeof(struct mem_cgroup_threshold));
4103 /* Add new threshold */
4104 new->entries[size - 1].eventfd = eventfd;
4105 new->entries[size - 1].threshold = threshold;
4107 /* Sort thresholds. Registering of new threshold isn't time-critical */
4108 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4109 compare_thresholds, NULL);
4111 /* Find current threshold */
4112 new->current_threshold = -1;
4113 for (i = 0; i < size; i++) {
4114 if (new->entries[i].threshold <= usage) {
4116 * new->current_threshold will not be used until
4117 * rcu_assign_pointer(), so it's safe to increment
4120 ++new->current_threshold;
4125 /* Free old spare buffer and save old primary buffer as spare */
4126 kfree(thresholds->spare);
4127 thresholds->spare = thresholds->primary;
4129 rcu_assign_pointer(thresholds->primary, new);
4131 /* To be sure that nobody uses thresholds */
4135 mutex_unlock(&memcg->thresholds_lock);
4140 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4141 struct eventfd_ctx *eventfd, const char *args)
4143 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4146 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4147 struct eventfd_ctx *eventfd, const char *args)
4149 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4152 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4153 struct eventfd_ctx *eventfd, enum res_type type)
4155 struct mem_cgroup_thresholds *thresholds;
4156 struct mem_cgroup_threshold_ary *new;
4157 unsigned long usage;
4158 int i, j, size, entries;
4160 mutex_lock(&memcg->thresholds_lock);
4163 thresholds = &memcg->thresholds;
4164 usage = mem_cgroup_usage(memcg, false);
4165 } else if (type == _MEMSWAP) {
4166 thresholds = &memcg->memsw_thresholds;
4167 usage = mem_cgroup_usage(memcg, true);
4171 if (!thresholds->primary)
4174 /* Check if a threshold crossed before removing */
4175 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4177 /* Calculate new number of threshold */
4179 for (i = 0; i < thresholds->primary->size; i++) {
4180 if (thresholds->primary->entries[i].eventfd != eventfd)
4186 new = thresholds->spare;
4188 /* If no items related to eventfd have been cleared, nothing to do */
4192 /* Set thresholds array to NULL if we don't have thresholds */
4201 /* Copy thresholds and find current threshold */
4202 new->current_threshold = -1;
4203 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4204 if (thresholds->primary->entries[i].eventfd == eventfd)
4207 new->entries[j] = thresholds->primary->entries[i];
4208 if (new->entries[j].threshold <= usage) {
4210 * new->current_threshold will not be used
4211 * until rcu_assign_pointer(), so it's safe to increment
4214 ++new->current_threshold;
4220 /* Swap primary and spare array */
4221 thresholds->spare = thresholds->primary;
4223 rcu_assign_pointer(thresholds->primary, new);
4225 /* To be sure that nobody uses thresholds */
4228 /* If all events are unregistered, free the spare array */
4230 kfree(thresholds->spare);
4231 thresholds->spare = NULL;
4234 mutex_unlock(&memcg->thresholds_lock);
4237 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4238 struct eventfd_ctx *eventfd)
4240 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4243 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4244 struct eventfd_ctx *eventfd)
4246 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4249 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4250 struct eventfd_ctx *eventfd, const char *args)
4252 struct mem_cgroup_eventfd_list *event;
4254 event = kmalloc(sizeof(*event), GFP_KERNEL);
4258 spin_lock(&memcg_oom_lock);
4260 event->eventfd = eventfd;
4261 list_add(&event->list, &memcg->oom_notify);
4263 /* already in OOM ? */
4264 if (memcg->under_oom)
4265 eventfd_signal(eventfd, 1);
4266 spin_unlock(&memcg_oom_lock);
4271 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4272 struct eventfd_ctx *eventfd)
4274 struct mem_cgroup_eventfd_list *ev, *tmp;
4276 spin_lock(&memcg_oom_lock);
4278 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4279 if (ev->eventfd == eventfd) {
4280 list_del(&ev->list);
4285 spin_unlock(&memcg_oom_lock);
4288 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4290 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4292 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4293 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4294 seq_printf(sf, "oom_kill %lu\n",
4295 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4299 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4300 struct cftype *cft, u64 val)
4302 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4304 /* cannot set to root cgroup and only 0 and 1 are allowed */
4305 if (!css->parent || !((val == 0) || (val == 1)))
4308 memcg->oom_kill_disable = val;
4310 memcg_oom_recover(memcg);
4315 #ifdef CONFIG_CGROUP_WRITEBACK
4317 #include <trace/events/writeback.h>
4319 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4321 return wb_domain_init(&memcg->cgwb_domain, gfp);
4324 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4326 wb_domain_exit(&memcg->cgwb_domain);
4329 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4331 wb_domain_size_changed(&memcg->cgwb_domain);
4334 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4336 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4338 if (!memcg->css.parent)
4341 return &memcg->cgwb_domain;
4345 * idx can be of type enum memcg_stat_item or node_stat_item.
4346 * Keep in sync with memcg_exact_page().
4348 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4350 long x = atomic_long_read(&memcg->vmstats[idx]);
4353 for_each_online_cpu(cpu)
4354 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4361 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4362 * @wb: bdi_writeback in question
4363 * @pfilepages: out parameter for number of file pages
4364 * @pheadroom: out parameter for number of allocatable pages according to memcg
4365 * @pdirty: out parameter for number of dirty pages
4366 * @pwriteback: out parameter for number of pages under writeback
4368 * Determine the numbers of file, headroom, dirty, and writeback pages in
4369 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4370 * is a bit more involved.
4372 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4373 * headroom is calculated as the lowest headroom of itself and the
4374 * ancestors. Note that this doesn't consider the actual amount of
4375 * available memory in the system. The caller should further cap
4376 * *@pheadroom accordingly.
4378 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4379 unsigned long *pheadroom, unsigned long *pdirty,
4380 unsigned long *pwriteback)
4382 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4383 struct mem_cgroup *parent;
4385 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4387 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4388 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4389 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4390 *pheadroom = PAGE_COUNTER_MAX;
4392 while ((parent = parent_mem_cgroup(memcg))) {
4393 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4394 READ_ONCE(memcg->memory.high));
4395 unsigned long used = page_counter_read(&memcg->memory);
4397 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4403 * Foreign dirty flushing
4405 * There's an inherent mismatch between memcg and writeback. The former
4406 * trackes ownership per-page while the latter per-inode. This was a
4407 * deliberate design decision because honoring per-page ownership in the
4408 * writeback path is complicated, may lead to higher CPU and IO overheads
4409 * and deemed unnecessary given that write-sharing an inode across
4410 * different cgroups isn't a common use-case.
4412 * Combined with inode majority-writer ownership switching, this works well
4413 * enough in most cases but there are some pathological cases. For
4414 * example, let's say there are two cgroups A and B which keep writing to
4415 * different but confined parts of the same inode. B owns the inode and
4416 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4417 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4418 * triggering background writeback. A will be slowed down without a way to
4419 * make writeback of the dirty pages happen.
4421 * Conditions like the above can lead to a cgroup getting repatedly and
4422 * severely throttled after making some progress after each
4423 * dirty_expire_interval while the underyling IO device is almost
4426 * Solving this problem completely requires matching the ownership tracking
4427 * granularities between memcg and writeback in either direction. However,
4428 * the more egregious behaviors can be avoided by simply remembering the
4429 * most recent foreign dirtying events and initiating remote flushes on
4430 * them when local writeback isn't enough to keep the memory clean enough.
4432 * The following two functions implement such mechanism. When a foreign
4433 * page - a page whose memcg and writeback ownerships don't match - is
4434 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4435 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4436 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4437 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4438 * foreign bdi_writebacks which haven't expired. Both the numbers of
4439 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4440 * limited to MEMCG_CGWB_FRN_CNT.
4442 * The mechanism only remembers IDs and doesn't hold any object references.
4443 * As being wrong occasionally doesn't matter, updates and accesses to the
4444 * records are lockless and racy.
4446 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4447 struct bdi_writeback *wb)
4449 struct mem_cgroup *memcg = page->mem_cgroup;
4450 struct memcg_cgwb_frn *frn;
4451 u64 now = get_jiffies_64();
4452 u64 oldest_at = now;
4456 trace_track_foreign_dirty(page, wb);
4459 * Pick the slot to use. If there is already a slot for @wb, keep
4460 * using it. If not replace the oldest one which isn't being
4463 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4464 frn = &memcg->cgwb_frn[i];
4465 if (frn->bdi_id == wb->bdi->id &&
4466 frn->memcg_id == wb->memcg_css->id)
4468 if (time_before64(frn->at, oldest_at) &&
4469 atomic_read(&frn->done.cnt) == 1) {
4471 oldest_at = frn->at;
4475 if (i < MEMCG_CGWB_FRN_CNT) {
4477 * Re-using an existing one. Update timestamp lazily to
4478 * avoid making the cacheline hot. We want them to be
4479 * reasonably up-to-date and significantly shorter than
4480 * dirty_expire_interval as that's what expires the record.
4481 * Use the shorter of 1s and dirty_expire_interval / 8.
4483 unsigned long update_intv =
4484 min_t(unsigned long, HZ,
4485 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4487 if (time_before64(frn->at, now - update_intv))
4489 } else if (oldest >= 0) {
4490 /* replace the oldest free one */
4491 frn = &memcg->cgwb_frn[oldest];
4492 frn->bdi_id = wb->bdi->id;
4493 frn->memcg_id = wb->memcg_css->id;
4498 /* issue foreign writeback flushes for recorded foreign dirtying events */
4499 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4501 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4502 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4503 u64 now = jiffies_64;
4506 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4507 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4510 * If the record is older than dirty_expire_interval,
4511 * writeback on it has already started. No need to kick it
4512 * off again. Also, don't start a new one if there's
4513 * already one in flight.
4515 if (time_after64(frn->at, now - intv) &&
4516 atomic_read(&frn->done.cnt) == 1) {
4518 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4519 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4520 WB_REASON_FOREIGN_FLUSH,
4526 #else /* CONFIG_CGROUP_WRITEBACK */
4528 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4533 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4537 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4541 #endif /* CONFIG_CGROUP_WRITEBACK */
4544 * DO NOT USE IN NEW FILES.
4546 * "cgroup.event_control" implementation.
4548 * This is way over-engineered. It tries to support fully configurable
4549 * events for each user. Such level of flexibility is completely
4550 * unnecessary especially in the light of the planned unified hierarchy.
4552 * Please deprecate this and replace with something simpler if at all
4557 * Unregister event and free resources.
4559 * Gets called from workqueue.
4561 static void memcg_event_remove(struct work_struct *work)
4563 struct mem_cgroup_event *event =
4564 container_of(work, struct mem_cgroup_event, remove);
4565 struct mem_cgroup *memcg = event->memcg;
4567 remove_wait_queue(event->wqh, &event->wait);
4569 event->unregister_event(memcg, event->eventfd);
4571 /* Notify userspace the event is going away. */
4572 eventfd_signal(event->eventfd, 1);
4574 eventfd_ctx_put(event->eventfd);
4576 css_put(&memcg->css);
4580 * Gets called on EPOLLHUP on eventfd when user closes it.
4582 * Called with wqh->lock held and interrupts disabled.
4584 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4585 int sync, void *key)
4587 struct mem_cgroup_event *event =
4588 container_of(wait, struct mem_cgroup_event, wait);
4589 struct mem_cgroup *memcg = event->memcg;
4590 __poll_t flags = key_to_poll(key);
4592 if (flags & EPOLLHUP) {
4594 * If the event has been detached at cgroup removal, we
4595 * can simply return knowing the other side will cleanup
4598 * We can't race against event freeing since the other
4599 * side will require wqh->lock via remove_wait_queue(),
4602 spin_lock(&memcg->event_list_lock);
4603 if (!list_empty(&event->list)) {
4604 list_del_init(&event->list);
4606 * We are in atomic context, but cgroup_event_remove()
4607 * may sleep, so we have to call it in workqueue.
4609 schedule_work(&event->remove);
4611 spin_unlock(&memcg->event_list_lock);
4617 static void memcg_event_ptable_queue_proc(struct file *file,
4618 wait_queue_head_t *wqh, poll_table *pt)
4620 struct mem_cgroup_event *event =
4621 container_of(pt, struct mem_cgroup_event, pt);
4624 add_wait_queue(wqh, &event->wait);
4628 * DO NOT USE IN NEW FILES.
4630 * Parse input and register new cgroup event handler.
4632 * Input must be in format '<event_fd> <control_fd> <args>'.
4633 * Interpretation of args is defined by control file implementation.
4635 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4636 char *buf, size_t nbytes, loff_t off)
4638 struct cgroup_subsys_state *css = of_css(of);
4639 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4640 struct mem_cgroup_event *event;
4641 struct cgroup_subsys_state *cfile_css;
4642 unsigned int efd, cfd;
4649 buf = strstrip(buf);
4651 efd = simple_strtoul(buf, &endp, 10);
4656 cfd = simple_strtoul(buf, &endp, 10);
4657 if ((*endp != ' ') && (*endp != '\0'))
4661 event = kzalloc(sizeof(*event), GFP_KERNEL);
4665 event->memcg = memcg;
4666 INIT_LIST_HEAD(&event->list);
4667 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4668 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4669 INIT_WORK(&event->remove, memcg_event_remove);
4677 event->eventfd = eventfd_ctx_fileget(efile.file);
4678 if (IS_ERR(event->eventfd)) {
4679 ret = PTR_ERR(event->eventfd);
4686 goto out_put_eventfd;
4689 /* the process need read permission on control file */
4690 /* AV: shouldn't we check that it's been opened for read instead? */
4691 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4696 * Determine the event callbacks and set them in @event. This used
4697 * to be done via struct cftype but cgroup core no longer knows
4698 * about these events. The following is crude but the whole thing
4699 * is for compatibility anyway.
4701 * DO NOT ADD NEW FILES.
4703 name = cfile.file->f_path.dentry->d_name.name;
4705 if (!strcmp(name, "memory.usage_in_bytes")) {
4706 event->register_event = mem_cgroup_usage_register_event;
4707 event->unregister_event = mem_cgroup_usage_unregister_event;
4708 } else if (!strcmp(name, "memory.oom_control")) {
4709 event->register_event = mem_cgroup_oom_register_event;
4710 event->unregister_event = mem_cgroup_oom_unregister_event;
4711 } else if (!strcmp(name, "memory.pressure_level")) {
4712 event->register_event = vmpressure_register_event;
4713 event->unregister_event = vmpressure_unregister_event;
4714 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4715 event->register_event = memsw_cgroup_usage_register_event;
4716 event->unregister_event = memsw_cgroup_usage_unregister_event;
4723 * Verify @cfile should belong to @css. Also, remaining events are
4724 * automatically removed on cgroup destruction but the removal is
4725 * asynchronous, so take an extra ref on @css.
4727 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4728 &memory_cgrp_subsys);
4730 if (IS_ERR(cfile_css))
4732 if (cfile_css != css) {
4737 ret = event->register_event(memcg, event->eventfd, buf);
4741 vfs_poll(efile.file, &event->pt);
4743 spin_lock(&memcg->event_list_lock);
4744 list_add(&event->list, &memcg->event_list);
4745 spin_unlock(&memcg->event_list_lock);
4757 eventfd_ctx_put(event->eventfd);
4766 static struct cftype mem_cgroup_legacy_files[] = {
4768 .name = "usage_in_bytes",
4769 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4770 .read_u64 = mem_cgroup_read_u64,
4773 .name = "max_usage_in_bytes",
4774 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4775 .write = mem_cgroup_reset,
4776 .read_u64 = mem_cgroup_read_u64,
4779 .name = "limit_in_bytes",
4780 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4781 .write = mem_cgroup_write,
4782 .read_u64 = mem_cgroup_read_u64,
4785 .name = "soft_limit_in_bytes",
4786 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4787 .write = mem_cgroup_write,
4788 .read_u64 = mem_cgroup_read_u64,
4792 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4793 .write = mem_cgroup_reset,
4794 .read_u64 = mem_cgroup_read_u64,
4798 .seq_show = memcg_stat_show,
4801 .name = "force_empty",
4802 .write = mem_cgroup_force_empty_write,
4805 .name = "use_hierarchy",
4806 .write_u64 = mem_cgroup_hierarchy_write,
4807 .read_u64 = mem_cgroup_hierarchy_read,
4810 .name = "cgroup.event_control", /* XXX: for compat */
4811 .write = memcg_write_event_control,
4812 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4815 .name = "swappiness",
4816 .read_u64 = mem_cgroup_swappiness_read,
4817 .write_u64 = mem_cgroup_swappiness_write,
4820 .name = "move_charge_at_immigrate",
4821 .read_u64 = mem_cgroup_move_charge_read,
4822 .write_u64 = mem_cgroup_move_charge_write,
4825 .name = "oom_control",
4826 .seq_show = mem_cgroup_oom_control_read,
4827 .write_u64 = mem_cgroup_oom_control_write,
4828 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4831 .name = "pressure_level",
4835 .name = "numa_stat",
4836 .seq_show = memcg_numa_stat_show,
4840 .name = "kmem.limit_in_bytes",
4841 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4842 .write = mem_cgroup_write,
4843 .read_u64 = mem_cgroup_read_u64,
4846 .name = "kmem.usage_in_bytes",
4847 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4848 .read_u64 = mem_cgroup_read_u64,
4851 .name = "kmem.failcnt",
4852 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4853 .write = mem_cgroup_reset,
4854 .read_u64 = mem_cgroup_read_u64,
4857 .name = "kmem.max_usage_in_bytes",
4858 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4859 .write = mem_cgroup_reset,
4860 .read_u64 = mem_cgroup_read_u64,
4862 #if defined(CONFIG_MEMCG_KMEM) && \
4863 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4865 .name = "kmem.slabinfo",
4866 .seq_start = memcg_slab_start,
4867 .seq_next = memcg_slab_next,
4868 .seq_stop = memcg_slab_stop,
4869 .seq_show = memcg_slab_show,
4873 .name = "kmem.tcp.limit_in_bytes",
4874 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4875 .write = mem_cgroup_write,
4876 .read_u64 = mem_cgroup_read_u64,
4879 .name = "kmem.tcp.usage_in_bytes",
4880 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4881 .read_u64 = mem_cgroup_read_u64,
4884 .name = "kmem.tcp.failcnt",
4885 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4886 .write = mem_cgroup_reset,
4887 .read_u64 = mem_cgroup_read_u64,
4890 .name = "kmem.tcp.max_usage_in_bytes",
4891 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4892 .write = mem_cgroup_reset,
4893 .read_u64 = mem_cgroup_read_u64,
4895 { }, /* terminate */
4899 * Private memory cgroup IDR
4901 * Swap-out records and page cache shadow entries need to store memcg
4902 * references in constrained space, so we maintain an ID space that is
4903 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4904 * memory-controlled cgroups to 64k.
4906 * However, there usually are many references to the oflline CSS after
4907 * the cgroup has been destroyed, such as page cache or reclaimable
4908 * slab objects, that don't need to hang on to the ID. We want to keep
4909 * those dead CSS from occupying IDs, or we might quickly exhaust the
4910 * relatively small ID space and prevent the creation of new cgroups
4911 * even when there are much fewer than 64k cgroups - possibly none.
4913 * Maintain a private 16-bit ID space for memcg, and allow the ID to
4914 * be freed and recycled when it's no longer needed, which is usually
4915 * when the CSS is offlined.
4917 * The only exception to that are records of swapped out tmpfs/shmem
4918 * pages that need to be attributed to live ancestors on swapin. But
4919 * those references are manageable from userspace.
4922 static DEFINE_IDR(mem_cgroup_idr);
4924 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
4926 if (memcg->id.id > 0) {
4927 idr_remove(&mem_cgroup_idr, memcg->id.id);
4932 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
4935 refcount_add(n, &memcg->id.ref);
4938 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
4940 if (refcount_sub_and_test(n, &memcg->id.ref)) {
4941 mem_cgroup_id_remove(memcg);
4943 /* Memcg ID pins CSS */
4944 css_put(&memcg->css);
4948 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
4950 mem_cgroup_id_put_many(memcg, 1);
4954 * mem_cgroup_from_id - look up a memcg from a memcg id
4955 * @id: the memcg id to look up
4957 * Caller must hold rcu_read_lock().
4959 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
4961 WARN_ON_ONCE(!rcu_read_lock_held());
4962 return idr_find(&mem_cgroup_idr, id);
4965 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
4967 struct mem_cgroup_per_node *pn;
4970 * This routine is called against possible nodes.
4971 * But it's BUG to call kmalloc() against offline node.
4973 * TODO: this routine can waste much memory for nodes which will
4974 * never be onlined. It's better to use memory hotplug callback
4977 if (!node_state(node, N_NORMAL_MEMORY))
4979 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4983 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
4984 if (!pn->lruvec_stat_local) {
4989 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
4990 if (!pn->lruvec_stat_cpu) {
4991 free_percpu(pn->lruvec_stat_local);
4996 lruvec_init(&pn->lruvec);
4997 pn->usage_in_excess = 0;
4998 pn->on_tree = false;
5001 memcg->nodeinfo[node] = pn;
5005 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5007 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5012 free_percpu(pn->lruvec_stat_cpu);
5013 free_percpu(pn->lruvec_stat_local);
5017 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5022 free_mem_cgroup_per_node_info(memcg, node);
5023 free_percpu(memcg->vmstats_percpu);
5024 free_percpu(memcg->vmstats_local);
5028 static void mem_cgroup_free(struct mem_cgroup *memcg)
5030 memcg_wb_domain_exit(memcg);
5032 * Flush percpu vmstats and vmevents to guarantee the value correctness
5033 * on parent's and all ancestor levels.
5035 memcg_flush_percpu_vmstats(memcg);
5036 memcg_flush_percpu_vmevents(memcg);
5037 __mem_cgroup_free(memcg);
5040 static struct mem_cgroup *mem_cgroup_alloc(void)
5042 struct mem_cgroup *memcg;
5045 int __maybe_unused i;
5046 long error = -ENOMEM;
5048 size = sizeof(struct mem_cgroup);
5049 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5051 memcg = kzalloc(size, GFP_KERNEL);
5053 return ERR_PTR(error);
5055 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5056 1, MEM_CGROUP_ID_MAX,
5058 if (memcg->id.id < 0) {
5059 error = memcg->id.id;
5063 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5064 if (!memcg->vmstats_local)
5067 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5068 if (!memcg->vmstats_percpu)
5072 if (alloc_mem_cgroup_per_node_info(memcg, node))
5075 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5078 INIT_WORK(&memcg->high_work, high_work_func);
5079 INIT_LIST_HEAD(&memcg->oom_notify);
5080 mutex_init(&memcg->thresholds_lock);
5081 spin_lock_init(&memcg->move_lock);
5082 vmpressure_init(&memcg->vmpressure);
5083 INIT_LIST_HEAD(&memcg->event_list);
5084 spin_lock_init(&memcg->event_list_lock);
5085 memcg->socket_pressure = jiffies;
5086 #ifdef CONFIG_MEMCG_KMEM
5087 memcg->kmemcg_id = -1;
5089 #ifdef CONFIG_CGROUP_WRITEBACK
5090 INIT_LIST_HEAD(&memcg->cgwb_list);
5091 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5092 memcg->cgwb_frn[i].done =
5093 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5095 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5096 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5097 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5098 memcg->deferred_split_queue.split_queue_len = 0;
5100 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5103 mem_cgroup_id_remove(memcg);
5104 __mem_cgroup_free(memcg);
5105 return ERR_PTR(error);
5108 static struct cgroup_subsys_state * __ref
5109 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5111 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5112 struct mem_cgroup *memcg;
5113 long error = -ENOMEM;
5115 memcg = mem_cgroup_alloc();
5117 return ERR_CAST(memcg);
5119 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5120 memcg->soft_limit = PAGE_COUNTER_MAX;
5121 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5123 memcg->swappiness = mem_cgroup_swappiness(parent);
5124 memcg->oom_kill_disable = parent->oom_kill_disable;
5126 if (parent && parent->use_hierarchy) {
5127 memcg->use_hierarchy = true;
5128 page_counter_init(&memcg->memory, &parent->memory);
5129 page_counter_init(&memcg->swap, &parent->swap);
5130 page_counter_init(&memcg->memsw, &parent->memsw);
5131 page_counter_init(&memcg->kmem, &parent->kmem);
5132 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5134 page_counter_init(&memcg->memory, NULL);
5135 page_counter_init(&memcg->swap, NULL);
5136 page_counter_init(&memcg->memsw, NULL);
5137 page_counter_init(&memcg->kmem, NULL);
5138 page_counter_init(&memcg->tcpmem, NULL);
5140 * Deeper hierachy with use_hierarchy == false doesn't make
5141 * much sense so let cgroup subsystem know about this
5142 * unfortunate state in our controller.
5144 if (parent != root_mem_cgroup)
5145 memory_cgrp_subsys.broken_hierarchy = true;
5148 /* The following stuff does not apply to the root */
5150 #ifdef CONFIG_MEMCG_KMEM
5151 INIT_LIST_HEAD(&memcg->kmem_caches);
5153 root_mem_cgroup = memcg;
5157 error = memcg_online_kmem(memcg);
5161 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5162 static_branch_inc(&memcg_sockets_enabled_key);
5166 mem_cgroup_id_remove(memcg);
5167 mem_cgroup_free(memcg);
5168 return ERR_PTR(error);
5171 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5173 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5176 * A memcg must be visible for memcg_expand_shrinker_maps()
5177 * by the time the maps are allocated. So, we allocate maps
5178 * here, when for_each_mem_cgroup() can't skip it.
5180 if (memcg_alloc_shrinker_maps(memcg)) {
5181 mem_cgroup_id_remove(memcg);
5185 /* Online state pins memcg ID, memcg ID pins CSS */
5186 refcount_set(&memcg->id.ref, 1);
5191 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5193 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5194 struct mem_cgroup_event *event, *tmp;
5197 * Unregister events and notify userspace.
5198 * Notify userspace about cgroup removing only after rmdir of cgroup
5199 * directory to avoid race between userspace and kernelspace.
5201 spin_lock(&memcg->event_list_lock);
5202 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5203 list_del_init(&event->list);
5204 schedule_work(&event->remove);
5206 spin_unlock(&memcg->event_list_lock);
5208 page_counter_set_min(&memcg->memory, 0);
5209 page_counter_set_low(&memcg->memory, 0);
5211 memcg_offline_kmem(memcg);
5212 wb_memcg_offline(memcg);
5214 drain_all_stock(memcg);
5216 mem_cgroup_id_put(memcg);
5219 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5221 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5223 invalidate_reclaim_iterators(memcg);
5226 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5228 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5229 int __maybe_unused i;
5231 #ifdef CONFIG_CGROUP_WRITEBACK
5232 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5233 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5235 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5236 static_branch_dec(&memcg_sockets_enabled_key);
5238 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5239 static_branch_dec(&memcg_sockets_enabled_key);
5241 vmpressure_cleanup(&memcg->vmpressure);
5242 cancel_work_sync(&memcg->high_work);
5243 mem_cgroup_remove_from_trees(memcg);
5244 memcg_free_shrinker_maps(memcg);
5245 memcg_free_kmem(memcg);
5246 mem_cgroup_free(memcg);
5250 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5251 * @css: the target css
5253 * Reset the states of the mem_cgroup associated with @css. This is
5254 * invoked when the userland requests disabling on the default hierarchy
5255 * but the memcg is pinned through dependency. The memcg should stop
5256 * applying policies and should revert to the vanilla state as it may be
5257 * made visible again.
5259 * The current implementation only resets the essential configurations.
5260 * This needs to be expanded to cover all the visible parts.
5262 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5264 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5266 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5267 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5268 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5269 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5270 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5271 page_counter_set_min(&memcg->memory, 0);
5272 page_counter_set_low(&memcg->memory, 0);
5273 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5274 memcg->soft_limit = PAGE_COUNTER_MAX;
5275 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5276 memcg_wb_domain_size_changed(memcg);
5280 /* Handlers for move charge at task migration. */
5281 static int mem_cgroup_do_precharge(unsigned long count)
5285 /* Try a single bulk charge without reclaim first, kswapd may wake */
5286 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5288 mc.precharge += count;
5292 /* Try charges one by one with reclaim, but do not retry */
5294 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5308 enum mc_target_type {
5315 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5316 unsigned long addr, pte_t ptent)
5318 struct page *page = vm_normal_page(vma, addr, ptent);
5320 if (!page || !page_mapped(page))
5322 if (PageAnon(page)) {
5323 if (!(mc.flags & MOVE_ANON))
5326 if (!(mc.flags & MOVE_FILE))
5329 if (!get_page_unless_zero(page))
5335 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5336 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5337 pte_t ptent, swp_entry_t *entry)
5339 struct page *page = NULL;
5340 swp_entry_t ent = pte_to_swp_entry(ptent);
5342 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5346 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5347 * a device and because they are not accessible by CPU they are store
5348 * as special swap entry in the CPU page table.
5350 if (is_device_private_entry(ent)) {
5351 page = device_private_entry_to_page(ent);
5353 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5354 * a refcount of 1 when free (unlike normal page)
5356 if (!page_ref_add_unless(page, 1, 1))
5362 * Because lookup_swap_cache() updates some statistics counter,
5363 * we call find_get_page() with swapper_space directly.
5365 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5366 if (do_memsw_account())
5367 entry->val = ent.val;
5372 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5373 pte_t ptent, swp_entry_t *entry)
5379 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5380 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5382 struct page *page = NULL;
5383 struct address_space *mapping;
5386 if (!vma->vm_file) /* anonymous vma */
5388 if (!(mc.flags & MOVE_FILE))
5391 mapping = vma->vm_file->f_mapping;
5392 pgoff = linear_page_index(vma, addr);
5394 /* page is moved even if it's not RSS of this task(page-faulted). */
5396 /* shmem/tmpfs may report page out on swap: account for that too. */
5397 if (shmem_mapping(mapping)) {
5398 page = find_get_entry(mapping, pgoff);
5399 if (xa_is_value(page)) {
5400 swp_entry_t swp = radix_to_swp_entry(page);
5401 if (do_memsw_account())
5403 page = find_get_page(swap_address_space(swp),
5407 page = find_get_page(mapping, pgoff);
5409 page = find_get_page(mapping, pgoff);
5415 * mem_cgroup_move_account - move account of the page
5417 * @compound: charge the page as compound or small page
5418 * @from: mem_cgroup which the page is moved from.
5419 * @to: mem_cgroup which the page is moved to. @from != @to.
5421 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5423 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5426 static int mem_cgroup_move_account(struct page *page,
5428 struct mem_cgroup *from,
5429 struct mem_cgroup *to)
5431 struct lruvec *from_vec, *to_vec;
5432 struct pglist_data *pgdat;
5433 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5436 VM_BUG_ON(from == to);
5437 VM_BUG_ON_PAGE(PageLRU(page), page);
5438 VM_BUG_ON(compound && !PageTransHuge(page));
5441 * Prevent mem_cgroup_migrate() from looking at
5442 * page->mem_cgroup of its source page while we change it.
5445 if (!trylock_page(page))
5449 if (page->mem_cgroup != from)
5452 pgdat = page_pgdat(page);
5453 from_vec = mem_cgroup_lruvec(from, pgdat);
5454 to_vec = mem_cgroup_lruvec(to, pgdat);
5456 lock_page_memcg(page);
5458 if (!PageAnon(page)) {
5459 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5460 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5462 if (PageSwapBacked(page)) {
5463 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5464 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5467 if (page_mapped(page)) {
5468 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5469 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5472 if (PageDirty(page)) {
5473 struct address_space *mapping = page_mapping(page);
5475 if (mapping_cap_account_dirty(mapping)) {
5476 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5478 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5484 if (PageWriteback(page)) {
5485 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5486 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5490 * All state has been migrated, let's switch to the new memcg.
5492 * It is safe to change page->mem_cgroup here because the page
5493 * is referenced, charged, isolated, and locked: we can't race
5494 * with (un)charging, migration, LRU putback, or anything else
5495 * that would rely on a stable page->mem_cgroup.
5497 * Note that lock_page_memcg is a memcg lock, not a page lock,
5498 * to save space. As soon as we switch page->mem_cgroup to a
5499 * new memcg that isn't locked, the above state can change
5500 * concurrently again. Make sure we're truly done with it.
5504 page->mem_cgroup = to; /* caller should have done css_get */
5506 __unlock_page_memcg(from);
5510 local_irq_disable();
5511 mem_cgroup_charge_statistics(to, page, nr_pages);
5512 memcg_check_events(to, page);
5513 mem_cgroup_charge_statistics(from, page, -nr_pages);
5514 memcg_check_events(from, page);
5523 * get_mctgt_type - get target type of moving charge
5524 * @vma: the vma the pte to be checked belongs
5525 * @addr: the address corresponding to the pte to be checked
5526 * @ptent: the pte to be checked
5527 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5530 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5531 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5532 * move charge. if @target is not NULL, the page is stored in target->page
5533 * with extra refcnt got(Callers should handle it).
5534 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5535 * target for charge migration. if @target is not NULL, the entry is stored
5537 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5538 * (so ZONE_DEVICE page and thus not on the lru).
5539 * For now we such page is charge like a regular page would be as for all
5540 * intent and purposes it is just special memory taking the place of a
5543 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5545 * Called with pte lock held.
5548 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5549 unsigned long addr, pte_t ptent, union mc_target *target)
5551 struct page *page = NULL;
5552 enum mc_target_type ret = MC_TARGET_NONE;
5553 swp_entry_t ent = { .val = 0 };
5555 if (pte_present(ptent))
5556 page = mc_handle_present_pte(vma, addr, ptent);
5557 else if (is_swap_pte(ptent))
5558 page = mc_handle_swap_pte(vma, ptent, &ent);
5559 else if (pte_none(ptent))
5560 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5562 if (!page && !ent.val)
5566 * Do only loose check w/o serialization.
5567 * mem_cgroup_move_account() checks the page is valid or
5568 * not under LRU exclusion.
5570 if (page->mem_cgroup == mc.from) {
5571 ret = MC_TARGET_PAGE;
5572 if (is_device_private_page(page))
5573 ret = MC_TARGET_DEVICE;
5575 target->page = page;
5577 if (!ret || !target)
5581 * There is a swap entry and a page doesn't exist or isn't charged.
5582 * But we cannot move a tail-page in a THP.
5584 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5585 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5586 ret = MC_TARGET_SWAP;
5593 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5595 * We don't consider PMD mapped swapping or file mapped pages because THP does
5596 * not support them for now.
5597 * Caller should make sure that pmd_trans_huge(pmd) is true.
5599 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5600 unsigned long addr, pmd_t pmd, union mc_target *target)
5602 struct page *page = NULL;
5603 enum mc_target_type ret = MC_TARGET_NONE;
5605 if (unlikely(is_swap_pmd(pmd))) {
5606 VM_BUG_ON(thp_migration_supported() &&
5607 !is_pmd_migration_entry(pmd));
5610 page = pmd_page(pmd);
5611 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5612 if (!(mc.flags & MOVE_ANON))
5614 if (page->mem_cgroup == mc.from) {
5615 ret = MC_TARGET_PAGE;
5618 target->page = page;
5624 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5625 unsigned long addr, pmd_t pmd, union mc_target *target)
5627 return MC_TARGET_NONE;
5631 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5632 unsigned long addr, unsigned long end,
5633 struct mm_walk *walk)
5635 struct vm_area_struct *vma = walk->vma;
5639 ptl = pmd_trans_huge_lock(pmd, vma);
5642 * Note their can not be MC_TARGET_DEVICE for now as we do not
5643 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5644 * this might change.
5646 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5647 mc.precharge += HPAGE_PMD_NR;
5652 if (pmd_trans_unstable(pmd))
5654 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5655 for (; addr != end; pte++, addr += PAGE_SIZE)
5656 if (get_mctgt_type(vma, addr, *pte, NULL))
5657 mc.precharge++; /* increment precharge temporarily */
5658 pte_unmap_unlock(pte - 1, ptl);
5664 static const struct mm_walk_ops precharge_walk_ops = {
5665 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5668 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5670 unsigned long precharge;
5672 down_read(&mm->mmap_sem);
5673 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5674 up_read(&mm->mmap_sem);
5676 precharge = mc.precharge;
5682 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5684 unsigned long precharge = mem_cgroup_count_precharge(mm);
5686 VM_BUG_ON(mc.moving_task);
5687 mc.moving_task = current;
5688 return mem_cgroup_do_precharge(precharge);
5691 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5692 static void __mem_cgroup_clear_mc(void)
5694 struct mem_cgroup *from = mc.from;
5695 struct mem_cgroup *to = mc.to;
5697 /* we must uncharge all the leftover precharges from mc.to */
5699 cancel_charge(mc.to, mc.precharge);
5703 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5704 * we must uncharge here.
5706 if (mc.moved_charge) {
5707 cancel_charge(mc.from, mc.moved_charge);
5708 mc.moved_charge = 0;
5710 /* we must fixup refcnts and charges */
5711 if (mc.moved_swap) {
5712 /* uncharge swap account from the old cgroup */
5713 if (!mem_cgroup_is_root(mc.from))
5714 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5716 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5719 * we charged both to->memory and to->memsw, so we
5720 * should uncharge to->memory.
5722 if (!mem_cgroup_is_root(mc.to))
5723 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5725 mem_cgroup_id_get_many(mc.to, mc.moved_swap);
5726 css_put_many(&mc.to->css, mc.moved_swap);
5730 memcg_oom_recover(from);
5731 memcg_oom_recover(to);
5732 wake_up_all(&mc.waitq);
5735 static void mem_cgroup_clear_mc(void)
5737 struct mm_struct *mm = mc.mm;
5740 * we must clear moving_task before waking up waiters at the end of
5743 mc.moving_task = NULL;
5744 __mem_cgroup_clear_mc();
5745 spin_lock(&mc.lock);
5749 spin_unlock(&mc.lock);
5754 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5756 struct cgroup_subsys_state *css;
5757 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5758 struct mem_cgroup *from;
5759 struct task_struct *leader, *p;
5760 struct mm_struct *mm;
5761 unsigned long move_flags;
5764 /* charge immigration isn't supported on the default hierarchy */
5765 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5769 * Multi-process migrations only happen on the default hierarchy
5770 * where charge immigration is not used. Perform charge
5771 * immigration if @tset contains a leader and whine if there are
5775 cgroup_taskset_for_each_leader(leader, css, tset) {
5778 memcg = mem_cgroup_from_css(css);
5784 * We are now commited to this value whatever it is. Changes in this
5785 * tunable will only affect upcoming migrations, not the current one.
5786 * So we need to save it, and keep it going.
5788 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5792 from = mem_cgroup_from_task(p);
5794 VM_BUG_ON(from == memcg);
5796 mm = get_task_mm(p);
5799 /* We move charges only when we move a owner of the mm */
5800 if (mm->owner == p) {
5803 VM_BUG_ON(mc.precharge);
5804 VM_BUG_ON(mc.moved_charge);
5805 VM_BUG_ON(mc.moved_swap);
5807 spin_lock(&mc.lock);
5811 mc.flags = move_flags;
5812 spin_unlock(&mc.lock);
5813 /* We set mc.moving_task later */
5815 ret = mem_cgroup_precharge_mc(mm);
5817 mem_cgroup_clear_mc();
5824 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5827 mem_cgroup_clear_mc();
5830 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5831 unsigned long addr, unsigned long end,
5832 struct mm_walk *walk)
5835 struct vm_area_struct *vma = walk->vma;
5838 enum mc_target_type target_type;
5839 union mc_target target;
5842 ptl = pmd_trans_huge_lock(pmd, vma);
5844 if (mc.precharge < HPAGE_PMD_NR) {
5848 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5849 if (target_type == MC_TARGET_PAGE) {
5851 if (!isolate_lru_page(page)) {
5852 if (!mem_cgroup_move_account(page, true,
5854 mc.precharge -= HPAGE_PMD_NR;
5855 mc.moved_charge += HPAGE_PMD_NR;
5857 putback_lru_page(page);
5860 } else if (target_type == MC_TARGET_DEVICE) {
5862 if (!mem_cgroup_move_account(page, true,
5864 mc.precharge -= HPAGE_PMD_NR;
5865 mc.moved_charge += HPAGE_PMD_NR;
5873 if (pmd_trans_unstable(pmd))
5876 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5877 for (; addr != end; addr += PAGE_SIZE) {
5878 pte_t ptent = *(pte++);
5879 bool device = false;
5885 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5886 case MC_TARGET_DEVICE:
5889 case MC_TARGET_PAGE:
5892 * We can have a part of the split pmd here. Moving it
5893 * can be done but it would be too convoluted so simply
5894 * ignore such a partial THP and keep it in original
5895 * memcg. There should be somebody mapping the head.
5897 if (PageTransCompound(page))
5899 if (!device && isolate_lru_page(page))
5901 if (!mem_cgroup_move_account(page, false,
5904 /* we uncharge from mc.from later. */
5908 putback_lru_page(page);
5909 put: /* get_mctgt_type() gets the page */
5912 case MC_TARGET_SWAP:
5914 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5916 /* we fixup refcnts and charges later. */
5924 pte_unmap_unlock(pte - 1, ptl);
5929 * We have consumed all precharges we got in can_attach().
5930 * We try charge one by one, but don't do any additional
5931 * charges to mc.to if we have failed in charge once in attach()
5934 ret = mem_cgroup_do_precharge(1);
5942 static const struct mm_walk_ops charge_walk_ops = {
5943 .pmd_entry = mem_cgroup_move_charge_pte_range,
5946 static void mem_cgroup_move_charge(void)
5948 lru_add_drain_all();
5950 * Signal lock_page_memcg() to take the memcg's move_lock
5951 * while we're moving its pages to another memcg. Then wait
5952 * for already started RCU-only updates to finish.
5954 atomic_inc(&mc.from->moving_account);
5957 if (unlikely(!down_read_trylock(&mc.mm->mmap_sem))) {
5959 * Someone who are holding the mmap_sem might be waiting in
5960 * waitq. So we cancel all extra charges, wake up all waiters,
5961 * and retry. Because we cancel precharges, we might not be able
5962 * to move enough charges, but moving charge is a best-effort
5963 * feature anyway, so it wouldn't be a big problem.
5965 __mem_cgroup_clear_mc();
5970 * When we have consumed all precharges and failed in doing
5971 * additional charge, the page walk just aborts.
5973 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
5976 up_read(&mc.mm->mmap_sem);
5977 atomic_dec(&mc.from->moving_account);
5980 static void mem_cgroup_move_task(void)
5983 mem_cgroup_move_charge();
5984 mem_cgroup_clear_mc();
5987 #else /* !CONFIG_MMU */
5988 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5992 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5995 static void mem_cgroup_move_task(void)
6001 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6002 * to verify whether we're attached to the default hierarchy on each mount
6005 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6008 * use_hierarchy is forced on the default hierarchy. cgroup core
6009 * guarantees that @root doesn't have any children, so turning it
6010 * on for the root memcg is enough.
6012 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6013 root_mem_cgroup->use_hierarchy = true;
6015 root_mem_cgroup->use_hierarchy = false;
6018 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6020 if (value == PAGE_COUNTER_MAX)
6021 seq_puts(m, "max\n");
6023 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6028 static u64 memory_current_read(struct cgroup_subsys_state *css,
6031 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6033 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6036 static int memory_min_show(struct seq_file *m, void *v)
6038 return seq_puts_memcg_tunable(m,
6039 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6042 static ssize_t memory_min_write(struct kernfs_open_file *of,
6043 char *buf, size_t nbytes, loff_t off)
6045 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6049 buf = strstrip(buf);
6050 err = page_counter_memparse(buf, "max", &min);
6054 page_counter_set_min(&memcg->memory, min);
6059 static int memory_low_show(struct seq_file *m, void *v)
6061 return seq_puts_memcg_tunable(m,
6062 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6065 static ssize_t memory_low_write(struct kernfs_open_file *of,
6066 char *buf, size_t nbytes, loff_t off)
6068 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6072 buf = strstrip(buf);
6073 err = page_counter_memparse(buf, "max", &low);
6077 page_counter_set_low(&memcg->memory, low);
6082 static int memory_high_show(struct seq_file *m, void *v)
6084 return seq_puts_memcg_tunable(m,
6085 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6088 static ssize_t memory_high_write(struct kernfs_open_file *of,
6089 char *buf, size_t nbytes, loff_t off)
6091 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6092 unsigned int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
6093 bool drained = false;
6097 buf = strstrip(buf);
6098 err = page_counter_memparse(buf, "max", &high);
6102 page_counter_set_high(&memcg->memory, high);
6105 unsigned long nr_pages = page_counter_read(&memcg->memory);
6106 unsigned long reclaimed;
6108 if (nr_pages <= high)
6111 if (signal_pending(current))
6115 drain_all_stock(memcg);
6120 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6123 if (!reclaimed && !nr_retries--)
6130 static int memory_max_show(struct seq_file *m, void *v)
6132 return seq_puts_memcg_tunable(m,
6133 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6136 static ssize_t memory_max_write(struct kernfs_open_file *of,
6137 char *buf, size_t nbytes, loff_t off)
6139 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6140 unsigned int nr_reclaims = MEM_CGROUP_RECLAIM_RETRIES;
6141 bool drained = false;
6145 buf = strstrip(buf);
6146 err = page_counter_memparse(buf, "max", &max);
6150 xchg(&memcg->memory.max, max);
6153 unsigned long nr_pages = page_counter_read(&memcg->memory);
6155 if (nr_pages <= max)
6158 if (signal_pending(current))
6162 drain_all_stock(memcg);
6168 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6174 memcg_memory_event(memcg, MEMCG_OOM);
6175 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6179 memcg_wb_domain_size_changed(memcg);
6183 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6185 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6186 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6187 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6188 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6189 seq_printf(m, "oom_kill %lu\n",
6190 atomic_long_read(&events[MEMCG_OOM_KILL]));
6193 static int memory_events_show(struct seq_file *m, void *v)
6195 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6197 __memory_events_show(m, memcg->memory_events);
6201 static int memory_events_local_show(struct seq_file *m, void *v)
6203 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6205 __memory_events_show(m, memcg->memory_events_local);
6209 static int memory_stat_show(struct seq_file *m, void *v)
6211 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6214 buf = memory_stat_format(memcg);
6222 static int memory_oom_group_show(struct seq_file *m, void *v)
6224 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6226 seq_printf(m, "%d\n", memcg->oom_group);
6231 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6232 char *buf, size_t nbytes, loff_t off)
6234 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6237 buf = strstrip(buf);
6241 ret = kstrtoint(buf, 0, &oom_group);
6245 if (oom_group != 0 && oom_group != 1)
6248 memcg->oom_group = oom_group;
6253 static struct cftype memory_files[] = {
6256 .flags = CFTYPE_NOT_ON_ROOT,
6257 .read_u64 = memory_current_read,
6261 .flags = CFTYPE_NOT_ON_ROOT,
6262 .seq_show = memory_min_show,
6263 .write = memory_min_write,
6267 .flags = CFTYPE_NOT_ON_ROOT,
6268 .seq_show = memory_low_show,
6269 .write = memory_low_write,
6273 .flags = CFTYPE_NOT_ON_ROOT,
6274 .seq_show = memory_high_show,
6275 .write = memory_high_write,
6279 .flags = CFTYPE_NOT_ON_ROOT,
6280 .seq_show = memory_max_show,
6281 .write = memory_max_write,
6285 .flags = CFTYPE_NOT_ON_ROOT,
6286 .file_offset = offsetof(struct mem_cgroup, events_file),
6287 .seq_show = memory_events_show,
6290 .name = "events.local",
6291 .flags = CFTYPE_NOT_ON_ROOT,
6292 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6293 .seq_show = memory_events_local_show,
6297 .seq_show = memory_stat_show,
6300 .name = "oom.group",
6301 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6302 .seq_show = memory_oom_group_show,
6303 .write = memory_oom_group_write,
6308 struct cgroup_subsys memory_cgrp_subsys = {
6309 .css_alloc = mem_cgroup_css_alloc,
6310 .css_online = mem_cgroup_css_online,
6311 .css_offline = mem_cgroup_css_offline,
6312 .css_released = mem_cgroup_css_released,
6313 .css_free = mem_cgroup_css_free,
6314 .css_reset = mem_cgroup_css_reset,
6315 .can_attach = mem_cgroup_can_attach,
6316 .cancel_attach = mem_cgroup_cancel_attach,
6317 .post_attach = mem_cgroup_move_task,
6318 .bind = mem_cgroup_bind,
6319 .dfl_cftypes = memory_files,
6320 .legacy_cftypes = mem_cgroup_legacy_files,
6325 * This function calculates an individual cgroup's effective
6326 * protection which is derived from its own memory.min/low, its
6327 * parent's and siblings' settings, as well as the actual memory
6328 * distribution in the tree.
6330 * The following rules apply to the effective protection values:
6332 * 1. At the first level of reclaim, effective protection is equal to
6333 * the declared protection in memory.min and memory.low.
6335 * 2. To enable safe delegation of the protection configuration, at
6336 * subsequent levels the effective protection is capped to the
6337 * parent's effective protection.
6339 * 3. To make complex and dynamic subtrees easier to configure, the
6340 * user is allowed to overcommit the declared protection at a given
6341 * level. If that is the case, the parent's effective protection is
6342 * distributed to the children in proportion to how much protection
6343 * they have declared and how much of it they are utilizing.
6345 * This makes distribution proportional, but also work-conserving:
6346 * if one cgroup claims much more protection than it uses memory,
6347 * the unused remainder is available to its siblings.
6349 * 4. Conversely, when the declared protection is undercommitted at a
6350 * given level, the distribution of the larger parental protection
6351 * budget is NOT proportional. A cgroup's protection from a sibling
6352 * is capped to its own memory.min/low setting.
6354 * 5. However, to allow protecting recursive subtrees from each other
6355 * without having to declare each individual cgroup's fixed share
6356 * of the ancestor's claim to protection, any unutilized -
6357 * "floating" - protection from up the tree is distributed in
6358 * proportion to each cgroup's *usage*. This makes the protection
6359 * neutral wrt sibling cgroups and lets them compete freely over
6360 * the shared parental protection budget, but it protects the
6361 * subtree as a whole from neighboring subtrees.
6363 * Note that 4. and 5. are not in conflict: 4. is about protecting
6364 * against immediate siblings whereas 5. is about protecting against
6365 * neighboring subtrees.
6367 static unsigned long effective_protection(unsigned long usage,
6368 unsigned long parent_usage,
6369 unsigned long setting,
6370 unsigned long parent_effective,
6371 unsigned long siblings_protected)
6373 unsigned long protected;
6376 protected = min(usage, setting);
6378 * If all cgroups at this level combined claim and use more
6379 * protection then what the parent affords them, distribute
6380 * shares in proportion to utilization.
6382 * We are using actual utilization rather than the statically
6383 * claimed protection in order to be work-conserving: claimed
6384 * but unused protection is available to siblings that would
6385 * otherwise get a smaller chunk than what they claimed.
6387 if (siblings_protected > parent_effective)
6388 return protected * parent_effective / siblings_protected;
6391 * Ok, utilized protection of all children is within what the
6392 * parent affords them, so we know whatever this child claims
6393 * and utilizes is effectively protected.
6395 * If there is unprotected usage beyond this value, reclaim
6396 * will apply pressure in proportion to that amount.
6398 * If there is unutilized protection, the cgroup will be fully
6399 * shielded from reclaim, but we do return a smaller value for
6400 * protection than what the group could enjoy in theory. This
6401 * is okay. With the overcommit distribution above, effective
6402 * protection is always dependent on how memory is actually
6403 * consumed among the siblings anyway.
6408 * If the children aren't claiming (all of) the protection
6409 * afforded to them by the parent, distribute the remainder in
6410 * proportion to the (unprotected) memory of each cgroup. That
6411 * way, cgroups that aren't explicitly prioritized wrt each
6412 * other compete freely over the allowance, but they are
6413 * collectively protected from neighboring trees.
6415 * We're using unprotected memory for the weight so that if
6416 * some cgroups DO claim explicit protection, we don't protect
6417 * the same bytes twice.
6419 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6422 if (parent_effective > siblings_protected && usage > protected) {
6423 unsigned long unclaimed;
6425 unclaimed = parent_effective - siblings_protected;
6426 unclaimed *= usage - protected;
6427 unclaimed /= parent_usage - siblings_protected;
6436 * mem_cgroup_protected - check if memory consumption is in the normal range
6437 * @root: the top ancestor of the sub-tree being checked
6438 * @memcg: the memory cgroup to check
6440 * WARNING: This function is not stateless! It can only be used as part
6441 * of a top-down tree iteration, not for isolated queries.
6443 * Returns one of the following:
6444 * MEMCG_PROT_NONE: cgroup memory is not protected
6445 * MEMCG_PROT_LOW: cgroup memory is protected as long there is
6446 * an unprotected supply of reclaimable memory from other cgroups.
6447 * MEMCG_PROT_MIN: cgroup memory is protected
6449 enum mem_cgroup_protection mem_cgroup_protected(struct mem_cgroup *root,
6450 struct mem_cgroup *memcg)
6452 unsigned long usage, parent_usage;
6453 struct mem_cgroup *parent;
6455 if (mem_cgroup_disabled())
6456 return MEMCG_PROT_NONE;
6459 root = root_mem_cgroup;
6461 return MEMCG_PROT_NONE;
6463 usage = page_counter_read(&memcg->memory);
6465 return MEMCG_PROT_NONE;
6467 parent = parent_mem_cgroup(memcg);
6468 /* No parent means a non-hierarchical mode on v1 memcg */
6470 return MEMCG_PROT_NONE;
6472 if (parent == root) {
6473 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6474 memcg->memory.elow = memcg->memory.low;
6478 parent_usage = page_counter_read(&parent->memory);
6480 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6481 READ_ONCE(memcg->memory.min),
6482 READ_ONCE(parent->memory.emin),
6483 atomic_long_read(&parent->memory.children_min_usage)));
6485 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6486 memcg->memory.low, READ_ONCE(parent->memory.elow),
6487 atomic_long_read(&parent->memory.children_low_usage)));
6490 if (usage <= memcg->memory.emin)
6491 return MEMCG_PROT_MIN;
6492 else if (usage <= memcg->memory.elow)
6493 return MEMCG_PROT_LOW;
6495 return MEMCG_PROT_NONE;
6499 * mem_cgroup_try_charge - try charging a page
6500 * @page: page to charge
6501 * @mm: mm context of the victim
6502 * @gfp_mask: reclaim mode
6503 * @memcgp: charged memcg return
6505 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6506 * pages according to @gfp_mask if necessary.
6508 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6509 * Otherwise, an error code is returned.
6511 * After page->mapping has been set up, the caller must finalize the
6512 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6513 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6515 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6516 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6518 unsigned int nr_pages = hpage_nr_pages(page);
6519 struct mem_cgroup *memcg = NULL;
6522 if (mem_cgroup_disabled())
6525 if (PageSwapCache(page)) {
6527 * Every swap fault against a single page tries to charge the
6528 * page, bail as early as possible. shmem_unuse() encounters
6529 * already charged pages, too. The USED bit is protected by
6530 * the page lock, which serializes swap cache removal, which
6531 * in turn serializes uncharging.
6533 VM_BUG_ON_PAGE(!PageLocked(page), page);
6534 if (compound_head(page)->mem_cgroup)
6537 if (do_swap_account) {
6538 swp_entry_t ent = { .val = page_private(page), };
6539 unsigned short id = lookup_swap_cgroup_id(ent);
6542 memcg = mem_cgroup_from_id(id);
6543 if (memcg && !css_tryget_online(&memcg->css))
6550 memcg = get_mem_cgroup_from_mm(mm);
6552 ret = try_charge(memcg, gfp_mask, nr_pages);
6554 css_put(&memcg->css);
6560 int mem_cgroup_try_charge_delay(struct page *page, struct mm_struct *mm,
6561 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6565 ret = mem_cgroup_try_charge(page, mm, gfp_mask, memcgp);
6567 cgroup_throttle_swaprate(page, gfp_mask);
6572 * mem_cgroup_commit_charge - commit a page charge
6573 * @page: page to charge
6574 * @memcg: memcg to charge the page to
6575 * @lrucare: page might be on LRU already
6577 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6578 * after page->mapping has been set up. This must happen atomically
6579 * as part of the page instantiation, i.e. under the page table lock
6580 * for anonymous pages, under the page lock for page and swap cache.
6582 * In addition, the page must not be on the LRU during the commit, to
6583 * prevent racing with task migration. If it might be, use @lrucare.
6585 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6587 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6590 unsigned int nr_pages = hpage_nr_pages(page);
6592 VM_BUG_ON_PAGE(!page->mapping, page);
6593 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6595 if (mem_cgroup_disabled())
6598 * Swap faults will attempt to charge the same page multiple
6599 * times. But reuse_swap_page() might have removed the page
6600 * from swapcache already, so we can't check PageSwapCache().
6605 commit_charge(page, memcg, lrucare);
6607 local_irq_disable();
6608 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6609 memcg_check_events(memcg, page);
6612 if (do_memsw_account() && PageSwapCache(page)) {
6613 swp_entry_t entry = { .val = page_private(page) };
6615 * The swap entry might not get freed for a long time,
6616 * let's not wait for it. The page already received a
6617 * memory+swap charge, drop the swap entry duplicate.
6619 mem_cgroup_uncharge_swap(entry, nr_pages);
6624 * mem_cgroup_cancel_charge - cancel a page charge
6625 * @page: page to charge
6626 * @memcg: memcg to charge the page to
6628 * Cancel a charge transaction started by mem_cgroup_try_charge().
6630 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
6632 unsigned int nr_pages = hpage_nr_pages(page);
6634 if (mem_cgroup_disabled())
6637 * Swap faults will attempt to charge the same page multiple
6638 * times. But reuse_swap_page() might have removed the page
6639 * from swapcache already, so we can't check PageSwapCache().
6644 cancel_charge(memcg, nr_pages);
6648 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6649 * @page: page to charge
6650 * @mm: mm context of the victim
6651 * @gfp_mask: reclaim mode
6652 * @lrucare: page might be on the LRU already
6654 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6655 * pages according to @gfp_mask if necessary.
6657 * Returns 0 on success. Otherwise, an error code is returned.
6659 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask,
6662 struct mem_cgroup *memcg;
6665 VM_BUG_ON_PAGE(!page->mapping, page);
6667 ret = mem_cgroup_try_charge(page, mm, gfp_mask, &memcg);
6670 mem_cgroup_commit_charge(page, memcg, lrucare);
6674 struct uncharge_gather {
6675 struct mem_cgroup *memcg;
6676 unsigned long nr_pages;
6677 unsigned long pgpgout;
6678 unsigned long nr_anon;
6679 unsigned long nr_kmem;
6680 unsigned long nr_huge;
6681 struct page *dummy_page;
6684 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6686 memset(ug, 0, sizeof(*ug));
6689 static void uncharge_batch(const struct uncharge_gather *ug)
6691 unsigned long flags;
6693 if (!mem_cgroup_is_root(ug->memcg)) {
6694 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6695 if (do_memsw_account())
6696 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6697 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6698 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6699 memcg_oom_recover(ug->memcg);
6702 local_irq_save(flags);
6703 __mod_memcg_state(ug->memcg, MEMCG_RSS, -ug->nr_anon);
6704 __mod_memcg_state(ug->memcg, MEMCG_RSS_HUGE, -ug->nr_huge);
6705 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6706 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6707 memcg_check_events(ug->memcg, ug->dummy_page);
6708 local_irq_restore(flags);
6710 if (!mem_cgroup_is_root(ug->memcg))
6711 css_put_many(&ug->memcg->css, ug->nr_pages);
6714 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6716 unsigned long nr_pages;
6718 VM_BUG_ON_PAGE(PageLRU(page), page);
6720 if (!page->mem_cgroup)
6724 * Nobody should be changing or seriously looking at
6725 * page->mem_cgroup at this point, we have fully
6726 * exclusive access to the page.
6729 if (ug->memcg != page->mem_cgroup) {
6732 uncharge_gather_clear(ug);
6734 ug->memcg = page->mem_cgroup;
6737 nr_pages = compound_nr(page);
6738 ug->nr_pages += nr_pages;
6740 if (!PageKmemcg(page)) {
6741 if (PageTransHuge(page))
6742 ug->nr_huge += nr_pages;
6744 ug->nr_anon += nr_pages;
6747 ug->nr_kmem += nr_pages;
6748 __ClearPageKmemcg(page);
6751 ug->dummy_page = page;
6752 page->mem_cgroup = NULL;
6755 static void uncharge_list(struct list_head *page_list)
6757 struct uncharge_gather ug;
6758 struct list_head *next;
6760 uncharge_gather_clear(&ug);
6763 * Note that the list can be a single page->lru; hence the
6764 * do-while loop instead of a simple list_for_each_entry().
6766 next = page_list->next;
6770 page = list_entry(next, struct page, lru);
6771 next = page->lru.next;
6773 uncharge_page(page, &ug);
6774 } while (next != page_list);
6777 uncharge_batch(&ug);
6781 * mem_cgroup_uncharge - uncharge a page
6782 * @page: page to uncharge
6784 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6785 * mem_cgroup_commit_charge().
6787 void mem_cgroup_uncharge(struct page *page)
6789 struct uncharge_gather ug;
6791 if (mem_cgroup_disabled())
6794 /* Don't touch page->lru of any random page, pre-check: */
6795 if (!page->mem_cgroup)
6798 uncharge_gather_clear(&ug);
6799 uncharge_page(page, &ug);
6800 uncharge_batch(&ug);
6804 * mem_cgroup_uncharge_list - uncharge a list of page
6805 * @page_list: list of pages to uncharge
6807 * Uncharge a list of pages previously charged with
6808 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6810 void mem_cgroup_uncharge_list(struct list_head *page_list)
6812 if (mem_cgroup_disabled())
6815 if (!list_empty(page_list))
6816 uncharge_list(page_list);
6820 * mem_cgroup_migrate - charge a page's replacement
6821 * @oldpage: currently circulating page
6822 * @newpage: replacement page
6824 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6825 * be uncharged upon free.
6827 * Both pages must be locked, @newpage->mapping must be set up.
6829 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6831 struct mem_cgroup *memcg;
6832 unsigned int nr_pages;
6833 unsigned long flags;
6835 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6836 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6837 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6838 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6841 if (mem_cgroup_disabled())
6844 /* Page cache replacement: new page already charged? */
6845 if (newpage->mem_cgroup)
6848 /* Swapcache readahead pages can get replaced before being charged */
6849 memcg = oldpage->mem_cgroup;
6853 /* Force-charge the new page. The old one will be freed soon */
6854 nr_pages = hpage_nr_pages(newpage);
6856 page_counter_charge(&memcg->memory, nr_pages);
6857 if (do_memsw_account())
6858 page_counter_charge(&memcg->memsw, nr_pages);
6859 css_get_many(&memcg->css, nr_pages);
6861 commit_charge(newpage, memcg, false);
6863 local_irq_save(flags);
6864 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6865 memcg_check_events(memcg, newpage);
6866 local_irq_restore(flags);
6869 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6870 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6872 void mem_cgroup_sk_alloc(struct sock *sk)
6874 struct mem_cgroup *memcg;
6876 if (!mem_cgroup_sockets_enabled)
6879 /* Do not associate the sock with unrelated interrupted task's memcg. */
6884 memcg = mem_cgroup_from_task(current);
6885 if (memcg == root_mem_cgroup)
6887 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6889 if (css_tryget(&memcg->css))
6890 sk->sk_memcg = memcg;
6895 void mem_cgroup_sk_free(struct sock *sk)
6898 css_put(&sk->sk_memcg->css);
6902 * mem_cgroup_charge_skmem - charge socket memory
6903 * @memcg: memcg to charge
6904 * @nr_pages: number of pages to charge
6906 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6907 * @memcg's configured limit, %false if the charge had to be forced.
6909 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6911 gfp_t gfp_mask = GFP_KERNEL;
6913 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6914 struct page_counter *fail;
6916 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6917 memcg->tcpmem_pressure = 0;
6920 page_counter_charge(&memcg->tcpmem, nr_pages);
6921 memcg->tcpmem_pressure = 1;
6925 /* Don't block in the packet receive path */
6927 gfp_mask = GFP_NOWAIT;
6929 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6931 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6934 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6939 * mem_cgroup_uncharge_skmem - uncharge socket memory
6940 * @memcg: memcg to uncharge
6941 * @nr_pages: number of pages to uncharge
6943 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6945 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6946 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6950 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6952 refill_stock(memcg, nr_pages);
6955 static int __init cgroup_memory(char *s)
6959 while ((token = strsep(&s, ",")) != NULL) {
6962 if (!strcmp(token, "nosocket"))
6963 cgroup_memory_nosocket = true;
6964 if (!strcmp(token, "nokmem"))
6965 cgroup_memory_nokmem = true;
6969 __setup("cgroup.memory=", cgroup_memory);
6972 * subsys_initcall() for memory controller.
6974 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
6975 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
6976 * basically everything that doesn't depend on a specific mem_cgroup structure
6977 * should be initialized from here.
6979 static int __init mem_cgroup_init(void)
6983 #ifdef CONFIG_MEMCG_KMEM
6985 * Kmem cache creation is mostly done with the slab_mutex held,
6986 * so use a workqueue with limited concurrency to avoid stalling
6987 * all worker threads in case lots of cgroups are created and
6988 * destroyed simultaneously.
6990 memcg_kmem_cache_wq = alloc_workqueue("memcg_kmem_cache", 0, 1);
6991 BUG_ON(!memcg_kmem_cache_wq);
6994 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
6995 memcg_hotplug_cpu_dead);
6997 for_each_possible_cpu(cpu)
6998 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7001 for_each_node(node) {
7002 struct mem_cgroup_tree_per_node *rtpn;
7004 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7005 node_online(node) ? node : NUMA_NO_NODE);
7007 rtpn->rb_root = RB_ROOT;
7008 rtpn->rb_rightmost = NULL;
7009 spin_lock_init(&rtpn->lock);
7010 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7015 subsys_initcall(mem_cgroup_init);
7017 #ifdef CONFIG_MEMCG_SWAP
7018 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7020 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7022 * The root cgroup cannot be destroyed, so it's refcount must
7025 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7029 memcg = parent_mem_cgroup(memcg);
7031 memcg = root_mem_cgroup;
7037 * mem_cgroup_swapout - transfer a memsw charge to swap
7038 * @page: page whose memsw charge to transfer
7039 * @entry: swap entry to move the charge to
7041 * Transfer the memsw charge of @page to @entry.
7043 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7045 struct mem_cgroup *memcg, *swap_memcg;
7046 unsigned int nr_entries;
7047 unsigned short oldid;
7049 VM_BUG_ON_PAGE(PageLRU(page), page);
7050 VM_BUG_ON_PAGE(page_count(page), page);
7052 if (!do_memsw_account())
7055 memcg = page->mem_cgroup;
7057 /* Readahead page, never charged */
7062 * In case the memcg owning these pages has been offlined and doesn't
7063 * have an ID allocated to it anymore, charge the closest online
7064 * ancestor for the swap instead and transfer the memory+swap charge.
7066 swap_memcg = mem_cgroup_id_get_online(memcg);
7067 nr_entries = hpage_nr_pages(page);
7068 /* Get references for the tail pages, too */
7070 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7071 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7073 VM_BUG_ON_PAGE(oldid, page);
7074 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7076 page->mem_cgroup = NULL;
7078 if (!mem_cgroup_is_root(memcg))
7079 page_counter_uncharge(&memcg->memory, nr_entries);
7081 if (memcg != swap_memcg) {
7082 if (!mem_cgroup_is_root(swap_memcg))
7083 page_counter_charge(&swap_memcg->memsw, nr_entries);
7084 page_counter_uncharge(&memcg->memsw, nr_entries);
7088 * Interrupts should be disabled here because the caller holds the
7089 * i_pages lock which is taken with interrupts-off. It is
7090 * important here to have the interrupts disabled because it is the
7091 * only synchronisation we have for updating the per-CPU variables.
7093 VM_BUG_ON(!irqs_disabled());
7094 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7095 memcg_check_events(memcg, page);
7097 if (!mem_cgroup_is_root(memcg))
7098 css_put_many(&memcg->css, nr_entries);
7102 * mem_cgroup_try_charge_swap - try charging swap space for a page
7103 * @page: page being added to swap
7104 * @entry: swap entry to charge
7106 * Try to charge @page's memcg for the swap space at @entry.
7108 * Returns 0 on success, -ENOMEM on failure.
7110 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7112 unsigned int nr_pages = hpage_nr_pages(page);
7113 struct page_counter *counter;
7114 struct mem_cgroup *memcg;
7115 unsigned short oldid;
7117 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) || !do_swap_account)
7120 memcg = page->mem_cgroup;
7122 /* Readahead page, never charged */
7127 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7131 memcg = mem_cgroup_id_get_online(memcg);
7133 if (!mem_cgroup_is_root(memcg) &&
7134 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7135 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7136 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7137 mem_cgroup_id_put(memcg);
7141 /* Get references for the tail pages, too */
7143 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7144 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7145 VM_BUG_ON_PAGE(oldid, page);
7146 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7152 * mem_cgroup_uncharge_swap - uncharge swap space
7153 * @entry: swap entry to uncharge
7154 * @nr_pages: the amount of swap space to uncharge
7156 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7158 struct mem_cgroup *memcg;
7161 if (!do_swap_account)
7164 id = swap_cgroup_record(entry, 0, nr_pages);
7166 memcg = mem_cgroup_from_id(id);
7168 if (!mem_cgroup_is_root(memcg)) {
7169 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7170 page_counter_uncharge(&memcg->swap, nr_pages);
7172 page_counter_uncharge(&memcg->memsw, nr_pages);
7174 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7175 mem_cgroup_id_put_many(memcg, nr_pages);
7180 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7182 long nr_swap_pages = get_nr_swap_pages();
7184 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7185 return nr_swap_pages;
7186 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7187 nr_swap_pages = min_t(long, nr_swap_pages,
7188 READ_ONCE(memcg->swap.max) -
7189 page_counter_read(&memcg->swap));
7190 return nr_swap_pages;
7193 bool mem_cgroup_swap_full(struct page *page)
7195 struct mem_cgroup *memcg;
7197 VM_BUG_ON_PAGE(!PageLocked(page), page);
7201 if (!do_swap_account || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7204 memcg = page->mem_cgroup;
7208 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7209 unsigned long usage = page_counter_read(&memcg->swap);
7211 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7212 usage * 2 >= READ_ONCE(memcg->swap.max))
7219 /* for remember boot option*/
7220 #ifdef CONFIG_MEMCG_SWAP_ENABLED
7221 static int really_do_swap_account __initdata = 1;
7223 static int really_do_swap_account __initdata;
7226 static int __init enable_swap_account(char *s)
7228 if (!strcmp(s, "1"))
7229 really_do_swap_account = 1;
7230 else if (!strcmp(s, "0"))
7231 really_do_swap_account = 0;
7234 __setup("swapaccount=", enable_swap_account);
7236 static u64 swap_current_read(struct cgroup_subsys_state *css,
7239 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7241 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7244 static int swap_high_show(struct seq_file *m, void *v)
7246 return seq_puts_memcg_tunable(m,
7247 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7250 static ssize_t swap_high_write(struct kernfs_open_file *of,
7251 char *buf, size_t nbytes, loff_t off)
7253 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7257 buf = strstrip(buf);
7258 err = page_counter_memparse(buf, "max", &high);
7262 page_counter_set_high(&memcg->swap, high);
7267 static int swap_max_show(struct seq_file *m, void *v)
7269 return seq_puts_memcg_tunable(m,
7270 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7273 static ssize_t swap_max_write(struct kernfs_open_file *of,
7274 char *buf, size_t nbytes, loff_t off)
7276 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7280 buf = strstrip(buf);
7281 err = page_counter_memparse(buf, "max", &max);
7285 xchg(&memcg->swap.max, max);
7290 static int swap_events_show(struct seq_file *m, void *v)
7292 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7294 seq_printf(m, "high %lu\n",
7295 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7296 seq_printf(m, "max %lu\n",
7297 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7298 seq_printf(m, "fail %lu\n",
7299 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7304 static struct cftype swap_files[] = {
7306 .name = "swap.current",
7307 .flags = CFTYPE_NOT_ON_ROOT,
7308 .read_u64 = swap_current_read,
7311 .name = "swap.high",
7312 .flags = CFTYPE_NOT_ON_ROOT,
7313 .seq_show = swap_high_show,
7314 .write = swap_high_write,
7318 .flags = CFTYPE_NOT_ON_ROOT,
7319 .seq_show = swap_max_show,
7320 .write = swap_max_write,
7323 .name = "swap.events",
7324 .flags = CFTYPE_NOT_ON_ROOT,
7325 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7326 .seq_show = swap_events_show,
7331 static struct cftype memsw_cgroup_files[] = {
7333 .name = "memsw.usage_in_bytes",
7334 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7335 .read_u64 = mem_cgroup_read_u64,
7338 .name = "memsw.max_usage_in_bytes",
7339 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7340 .write = mem_cgroup_reset,
7341 .read_u64 = mem_cgroup_read_u64,
7344 .name = "memsw.limit_in_bytes",
7345 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7346 .write = mem_cgroup_write,
7347 .read_u64 = mem_cgroup_read_u64,
7350 .name = "memsw.failcnt",
7351 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7352 .write = mem_cgroup_reset,
7353 .read_u64 = mem_cgroup_read_u64,
7355 { }, /* terminate */
7358 static int __init mem_cgroup_swap_init(void)
7360 if (!mem_cgroup_disabled() && really_do_swap_account) {
7361 do_swap_account = 1;
7362 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys,
7364 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7365 memsw_cgroup_files));
7369 subsys_initcall(mem_cgroup_swap_init);
7371 #endif /* CONFIG_MEMCG_SWAP */