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
24 * Per memcg lru locking
25 * Copyright (C) 2020 Alibaba, Inc, Alex Shi
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
31 #include <linux/pagewalk.h>
32 #include <linux/sched/mm.h>
33 #include <linux/shmem_fs.h>
34 #include <linux/hugetlb.h>
35 #include <linux/pagemap.h>
36 #include <linux/vm_event_item.h>
37 #include <linux/smp.h>
38 #include <linux/page-flags.h>
39 #include <linux/backing-dev.h>
40 #include <linux/bit_spinlock.h>
41 #include <linux/rcupdate.h>
42 #include <linux/limits.h>
43 #include <linux/export.h>
44 #include <linux/mutex.h>
45 #include <linux/rbtree.h>
46 #include <linux/slab.h>
47 #include <linux/swap.h>
48 #include <linux/swapops.h>
49 #include <linux/spinlock.h>
50 #include <linux/eventfd.h>
51 #include <linux/poll.h>
52 #include <linux/sort.h>
54 #include <linux/seq_file.h>
55 #include <linux/vmpressure.h>
56 #include <linux/mm_inline.h>
57 #include <linux/swap_cgroup.h>
58 #include <linux/cpu.h>
59 #include <linux/oom.h>
60 #include <linux/lockdep.h>
61 #include <linux/file.h>
62 #include <linux/tracehook.h>
63 #include <linux/psi.h>
64 #include <linux/seq_buf.h>
70 #include <linux/uaccess.h>
72 #include <trace/events/vmscan.h>
74 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
75 EXPORT_SYMBOL(memory_cgrp_subsys);
77 struct mem_cgroup *root_mem_cgroup __read_mostly;
79 /* Active memory cgroup to use from an interrupt context */
80 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
82 /* Socket memory accounting disabled? */
83 static bool cgroup_memory_nosocket;
85 /* Kernel memory accounting disabled? */
86 static bool cgroup_memory_nokmem;
88 /* Whether the swap controller is active */
89 #ifdef CONFIG_MEMCG_SWAP
90 bool cgroup_memory_noswap __read_mostly;
92 #define cgroup_memory_noswap 1
95 #ifdef CONFIG_CGROUP_WRITEBACK
96 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
99 /* Whether legacy memory+swap accounting is active */
100 static bool do_memsw_account(void)
102 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
105 #define THRESHOLDS_EVENTS_TARGET 128
106 #define SOFTLIMIT_EVENTS_TARGET 1024
109 * Cgroups above their limits are maintained in a RB-Tree, independent of
110 * their hierarchy representation
113 struct mem_cgroup_tree_per_node {
114 struct rb_root rb_root;
115 struct rb_node *rb_rightmost;
119 struct mem_cgroup_tree {
120 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
123 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
126 struct mem_cgroup_eventfd_list {
127 struct list_head list;
128 struct eventfd_ctx *eventfd;
132 * cgroup_event represents events which userspace want to receive.
134 struct mem_cgroup_event {
136 * memcg which the event belongs to.
138 struct mem_cgroup *memcg;
140 * eventfd to signal userspace about the event.
142 struct eventfd_ctx *eventfd;
144 * Each of these stored in a list by the cgroup.
146 struct list_head list;
148 * register_event() callback will be used to add new userspace
149 * waiter for changes related to this event. Use eventfd_signal()
150 * on eventfd to send notification to userspace.
152 int (*register_event)(struct mem_cgroup *memcg,
153 struct eventfd_ctx *eventfd, const char *args);
155 * unregister_event() callback will be called when userspace closes
156 * the eventfd or on cgroup removing. This callback must be set,
157 * if you want provide notification functionality.
159 void (*unregister_event)(struct mem_cgroup *memcg,
160 struct eventfd_ctx *eventfd);
162 * All fields below needed to unregister event when
163 * userspace closes eventfd.
166 wait_queue_head_t *wqh;
167 wait_queue_entry_t wait;
168 struct work_struct remove;
171 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
172 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
174 /* Stuffs for move charges at task migration. */
176 * Types of charges to be moved.
178 #define MOVE_ANON 0x1U
179 #define MOVE_FILE 0x2U
180 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
182 /* "mc" and its members are protected by cgroup_mutex */
183 static struct move_charge_struct {
184 spinlock_t lock; /* for from, to */
185 struct mm_struct *mm;
186 struct mem_cgroup *from;
187 struct mem_cgroup *to;
189 unsigned long precharge;
190 unsigned long moved_charge;
191 unsigned long moved_swap;
192 struct task_struct *moving_task; /* a task moving charges */
193 wait_queue_head_t waitq; /* a waitq for other context */
195 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
196 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
200 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
201 * limit reclaim to prevent infinite loops, if they ever occur.
203 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
204 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
206 /* for encoding cft->private value on file */
215 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
216 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
217 #define MEMFILE_ATTR(val) ((val) & 0xffff)
218 /* Used for OOM nofiier */
219 #define OOM_CONTROL (0)
222 * Iteration constructs for visiting all cgroups (under a tree). If
223 * loops are exited prematurely (break), mem_cgroup_iter_break() must
224 * be used for reference counting.
226 #define for_each_mem_cgroup_tree(iter, root) \
227 for (iter = mem_cgroup_iter(root, NULL, NULL); \
229 iter = mem_cgroup_iter(root, iter, NULL))
231 #define for_each_mem_cgroup(iter) \
232 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
234 iter = mem_cgroup_iter(NULL, iter, NULL))
236 static inline bool should_force_charge(void)
238 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
239 (current->flags & PF_EXITING);
242 /* Some nice accessors for the vmpressure. */
243 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
246 memcg = root_mem_cgroup;
247 return &memcg->vmpressure;
250 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
252 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
255 #ifdef CONFIG_MEMCG_KMEM
256 extern spinlock_t css_set_lock;
258 static int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
259 unsigned int nr_pages);
260 static void __memcg_kmem_uncharge(struct mem_cgroup *memcg,
261 unsigned int nr_pages);
263 static void obj_cgroup_release(struct percpu_ref *ref)
265 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
266 struct mem_cgroup *memcg;
267 unsigned int nr_bytes;
268 unsigned int nr_pages;
272 * At this point all allocated objects are freed, and
273 * objcg->nr_charged_bytes can't have an arbitrary byte value.
274 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
276 * The following sequence can lead to it:
277 * 1) CPU0: objcg == stock->cached_objcg
278 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
279 * PAGE_SIZE bytes are charged
280 * 3) CPU1: a process from another memcg is allocating something,
281 * the stock if flushed,
282 * objcg->nr_charged_bytes = PAGE_SIZE - 92
283 * 5) CPU0: we do release this object,
284 * 92 bytes are added to stock->nr_bytes
285 * 6) CPU0: stock is flushed,
286 * 92 bytes are added to objcg->nr_charged_bytes
288 * In the result, nr_charged_bytes == PAGE_SIZE.
289 * This page will be uncharged in obj_cgroup_release().
291 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
292 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
293 nr_pages = nr_bytes >> PAGE_SHIFT;
295 spin_lock_irqsave(&css_set_lock, flags);
296 memcg = obj_cgroup_memcg(objcg);
298 __memcg_kmem_uncharge(memcg, nr_pages);
299 list_del(&objcg->list);
300 mem_cgroup_put(memcg);
301 spin_unlock_irqrestore(&css_set_lock, flags);
303 percpu_ref_exit(ref);
304 kfree_rcu(objcg, rcu);
307 static struct obj_cgroup *obj_cgroup_alloc(void)
309 struct obj_cgroup *objcg;
312 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
316 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
322 INIT_LIST_HEAD(&objcg->list);
326 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
327 struct mem_cgroup *parent)
329 struct obj_cgroup *objcg, *iter;
331 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
333 spin_lock_irq(&css_set_lock);
335 /* Move active objcg to the parent's list */
336 xchg(&objcg->memcg, parent);
337 css_get(&parent->css);
338 list_add(&objcg->list, &parent->objcg_list);
340 /* Move already reparented objcgs to the parent's list */
341 list_for_each_entry(iter, &memcg->objcg_list, list) {
342 css_get(&parent->css);
343 xchg(&iter->memcg, parent);
344 css_put(&memcg->css);
346 list_splice(&memcg->objcg_list, &parent->objcg_list);
348 spin_unlock_irq(&css_set_lock);
350 percpu_ref_kill(&objcg->refcnt);
354 * This will be used as a shrinker list's index.
355 * The main reason for not using cgroup id for this:
356 * this works better in sparse environments, where we have a lot of memcgs,
357 * but only a few kmem-limited. Or also, if we have, for instance, 200
358 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
359 * 200 entry array for that.
361 * The current size of the caches array is stored in memcg_nr_cache_ids. It
362 * will double each time we have to increase it.
364 static DEFINE_IDA(memcg_cache_ida);
365 int memcg_nr_cache_ids;
367 /* Protects memcg_nr_cache_ids */
368 static DECLARE_RWSEM(memcg_cache_ids_sem);
370 void memcg_get_cache_ids(void)
372 down_read(&memcg_cache_ids_sem);
375 void memcg_put_cache_ids(void)
377 up_read(&memcg_cache_ids_sem);
381 * MIN_SIZE is different than 1, because we would like to avoid going through
382 * the alloc/free process all the time. In a small machine, 4 kmem-limited
383 * cgroups is a reasonable guess. In the future, it could be a parameter or
384 * tunable, but that is strictly not necessary.
386 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
387 * this constant directly from cgroup, but it is understandable that this is
388 * better kept as an internal representation in cgroup.c. In any case, the
389 * cgrp_id space is not getting any smaller, and we don't have to necessarily
390 * increase ours as well if it increases.
392 #define MEMCG_CACHES_MIN_SIZE 4
393 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
396 * A lot of the calls to the cache allocation functions are expected to be
397 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
398 * conditional to this static branch, we'll have to allow modules that does
399 * kmem_cache_alloc and the such to see this symbol as well
401 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
402 EXPORT_SYMBOL(memcg_kmem_enabled_key);
405 static int memcg_shrinker_map_size;
406 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
408 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
410 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
413 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
414 int size, int old_size)
416 struct memcg_shrinker_map *new, *old;
417 struct mem_cgroup_per_node *pn;
420 lockdep_assert_held(&memcg_shrinker_map_mutex);
423 pn = memcg->nodeinfo[nid];
424 old = rcu_dereference_protected(pn->shrinker_map, true);
425 /* Not yet online memcg */
429 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
433 /* Set all old bits, clear all new bits */
434 memset(new->map, (int)0xff, old_size);
435 memset((void *)new->map + old_size, 0, size - old_size);
437 rcu_assign_pointer(pn->shrinker_map, new);
438 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
444 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
446 struct mem_cgroup_per_node *pn;
447 struct memcg_shrinker_map *map;
450 if (mem_cgroup_is_root(memcg))
454 pn = memcg->nodeinfo[nid];
455 map = rcu_dereference_protected(pn->shrinker_map, true);
457 rcu_assign_pointer(pn->shrinker_map, NULL);
461 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
463 struct memcg_shrinker_map *map;
464 int nid, size, ret = 0;
466 if (mem_cgroup_is_root(memcg))
469 mutex_lock(&memcg_shrinker_map_mutex);
470 size = memcg_shrinker_map_size;
472 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
474 memcg_free_shrinker_maps(memcg);
478 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
480 mutex_unlock(&memcg_shrinker_map_mutex);
485 int memcg_expand_shrinker_maps(int new_id)
487 int size, old_size, ret = 0;
488 struct mem_cgroup *memcg;
490 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
491 old_size = memcg_shrinker_map_size;
492 if (size <= old_size)
495 mutex_lock(&memcg_shrinker_map_mutex);
496 if (!root_mem_cgroup)
499 for_each_mem_cgroup(memcg) {
500 if (mem_cgroup_is_root(memcg))
502 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
504 mem_cgroup_iter_break(NULL, memcg);
510 memcg_shrinker_map_size = size;
511 mutex_unlock(&memcg_shrinker_map_mutex);
515 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
517 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
518 struct memcg_shrinker_map *map;
521 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
522 /* Pairs with smp mb in shrink_slab() */
523 smp_mb__before_atomic();
524 set_bit(shrinker_id, map->map);
530 * mem_cgroup_css_from_page - css of the memcg associated with a page
531 * @page: page of interest
533 * If memcg is bound to the default hierarchy, css of the memcg associated
534 * with @page is returned. The returned css remains associated with @page
535 * until it is released.
537 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
540 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
542 struct mem_cgroup *memcg;
544 memcg = page_memcg(page);
546 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
547 memcg = root_mem_cgroup;
553 * page_cgroup_ino - return inode number of the memcg a page is charged to
556 * Look up the closest online ancestor of the memory cgroup @page is charged to
557 * and return its inode number or 0 if @page is not charged to any cgroup. It
558 * is safe to call this function without holding a reference to @page.
560 * Note, this function is inherently racy, because there is nothing to prevent
561 * the cgroup inode from getting torn down and potentially reallocated a moment
562 * after page_cgroup_ino() returns, so it only should be used by callers that
563 * do not care (such as procfs interfaces).
565 ino_t page_cgroup_ino(struct page *page)
567 struct mem_cgroup *memcg;
568 unsigned long ino = 0;
571 memcg = page_memcg_check(page);
573 while (memcg && !(memcg->css.flags & CSS_ONLINE))
574 memcg = parent_mem_cgroup(memcg);
576 ino = cgroup_ino(memcg->css.cgroup);
581 static struct mem_cgroup_per_node *
582 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
584 int nid = page_to_nid(page);
586 return memcg->nodeinfo[nid];
589 static struct mem_cgroup_tree_per_node *
590 soft_limit_tree_node(int nid)
592 return soft_limit_tree.rb_tree_per_node[nid];
595 static struct mem_cgroup_tree_per_node *
596 soft_limit_tree_from_page(struct page *page)
598 int nid = page_to_nid(page);
600 return soft_limit_tree.rb_tree_per_node[nid];
603 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
604 struct mem_cgroup_tree_per_node *mctz,
605 unsigned long new_usage_in_excess)
607 struct rb_node **p = &mctz->rb_root.rb_node;
608 struct rb_node *parent = NULL;
609 struct mem_cgroup_per_node *mz_node;
610 bool rightmost = true;
615 mz->usage_in_excess = new_usage_in_excess;
616 if (!mz->usage_in_excess)
620 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
622 if (mz->usage_in_excess < mz_node->usage_in_excess) {
631 mctz->rb_rightmost = &mz->tree_node;
633 rb_link_node(&mz->tree_node, parent, p);
634 rb_insert_color(&mz->tree_node, &mctz->rb_root);
638 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
639 struct mem_cgroup_tree_per_node *mctz)
644 if (&mz->tree_node == mctz->rb_rightmost)
645 mctz->rb_rightmost = rb_prev(&mz->tree_node);
647 rb_erase(&mz->tree_node, &mctz->rb_root);
651 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
652 struct mem_cgroup_tree_per_node *mctz)
656 spin_lock_irqsave(&mctz->lock, flags);
657 __mem_cgroup_remove_exceeded(mz, mctz);
658 spin_unlock_irqrestore(&mctz->lock, flags);
661 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
663 unsigned long nr_pages = page_counter_read(&memcg->memory);
664 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
665 unsigned long excess = 0;
667 if (nr_pages > soft_limit)
668 excess = nr_pages - soft_limit;
673 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
675 unsigned long excess;
676 struct mem_cgroup_per_node *mz;
677 struct mem_cgroup_tree_per_node *mctz;
679 mctz = soft_limit_tree_from_page(page);
683 * Necessary to update all ancestors when hierarchy is used.
684 * because their event counter is not touched.
686 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
687 mz = mem_cgroup_page_nodeinfo(memcg, page);
688 excess = soft_limit_excess(memcg);
690 * We have to update the tree if mz is on RB-tree or
691 * mem is over its softlimit.
693 if (excess || mz->on_tree) {
696 spin_lock_irqsave(&mctz->lock, flags);
697 /* if on-tree, remove it */
699 __mem_cgroup_remove_exceeded(mz, mctz);
701 * Insert again. mz->usage_in_excess will be updated.
702 * If excess is 0, no tree ops.
704 __mem_cgroup_insert_exceeded(mz, mctz, excess);
705 spin_unlock_irqrestore(&mctz->lock, flags);
710 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
712 struct mem_cgroup_tree_per_node *mctz;
713 struct mem_cgroup_per_node *mz;
717 mz = memcg->nodeinfo[nid];
718 mctz = soft_limit_tree_node(nid);
720 mem_cgroup_remove_exceeded(mz, mctz);
724 static struct mem_cgroup_per_node *
725 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
727 struct mem_cgroup_per_node *mz;
731 if (!mctz->rb_rightmost)
732 goto done; /* Nothing to reclaim from */
734 mz = rb_entry(mctz->rb_rightmost,
735 struct mem_cgroup_per_node, tree_node);
737 * Remove the node now but someone else can add it back,
738 * we will to add it back at the end of reclaim to its correct
739 * position in the tree.
741 __mem_cgroup_remove_exceeded(mz, mctz);
742 if (!soft_limit_excess(mz->memcg) ||
743 !css_tryget(&mz->memcg->css))
749 static struct mem_cgroup_per_node *
750 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
752 struct mem_cgroup_per_node *mz;
754 spin_lock_irq(&mctz->lock);
755 mz = __mem_cgroup_largest_soft_limit_node(mctz);
756 spin_unlock_irq(&mctz->lock);
761 * __mod_memcg_state - update cgroup memory statistics
762 * @memcg: the memory cgroup
763 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
764 * @val: delta to add to the counter, can be negative
766 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
768 if (mem_cgroup_disabled())
771 __this_cpu_add(memcg->vmstats_percpu->state[idx], val);
772 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
775 /* idx can be of type enum memcg_stat_item or node_stat_item. */
776 static unsigned long memcg_page_state(struct mem_cgroup *memcg, int idx)
778 long x = READ_ONCE(memcg->vmstats.state[idx]);
786 /* idx can be of type enum memcg_stat_item or node_stat_item. */
787 static unsigned long memcg_page_state_local(struct mem_cgroup *memcg, int idx)
792 for_each_possible_cpu(cpu)
793 x += per_cpu(memcg->vmstats_percpu->state[idx], cpu);
801 static struct mem_cgroup_per_node *
802 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
804 struct mem_cgroup *parent;
806 parent = parent_mem_cgroup(pn->memcg);
809 return parent->nodeinfo[nid];
812 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
815 struct mem_cgroup_per_node *pn;
816 struct mem_cgroup *memcg;
817 long x, threshold = MEMCG_CHARGE_BATCH;
819 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
823 __mod_memcg_state(memcg, idx, val);
826 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
828 if (vmstat_item_in_bytes(idx))
829 threshold <<= PAGE_SHIFT;
831 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
832 if (unlikely(abs(x) > threshold)) {
833 pg_data_t *pgdat = lruvec_pgdat(lruvec);
834 struct mem_cgroup_per_node *pi;
836 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
837 atomic_long_add(x, &pi->lruvec_stat[idx]);
840 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
844 * __mod_lruvec_state - update lruvec memory statistics
845 * @lruvec: the lruvec
846 * @idx: the stat item
847 * @val: delta to add to the counter, can be negative
849 * The lruvec is the intersection of the NUMA node and a cgroup. This
850 * function updates the all three counters that are affected by a
851 * change of state at this level: per-node, per-cgroup, per-lruvec.
853 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
857 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
859 /* Update memcg and lruvec */
860 if (!mem_cgroup_disabled())
861 __mod_memcg_lruvec_state(lruvec, idx, val);
864 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
867 struct page *head = compound_head(page); /* rmap on tail pages */
868 struct mem_cgroup *memcg = page_memcg(head);
869 pg_data_t *pgdat = page_pgdat(page);
870 struct lruvec *lruvec;
872 /* Untracked pages have no memcg, no lruvec. Update only the node */
874 __mod_node_page_state(pgdat, idx, val);
878 lruvec = mem_cgroup_lruvec(memcg, pgdat);
879 __mod_lruvec_state(lruvec, idx, val);
881 EXPORT_SYMBOL(__mod_lruvec_page_state);
883 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
885 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
886 struct mem_cgroup *memcg;
887 struct lruvec *lruvec;
890 memcg = mem_cgroup_from_obj(p);
893 * Untracked pages have no memcg, no lruvec. Update only the
894 * node. If we reparent the slab objects to the root memcg,
895 * when we free the slab object, we need to update the per-memcg
896 * vmstats to keep it correct for the root memcg.
899 __mod_node_page_state(pgdat, idx, val);
901 lruvec = mem_cgroup_lruvec(memcg, pgdat);
902 __mod_lruvec_state(lruvec, idx, val);
908 * __count_memcg_events - account VM events in a cgroup
909 * @memcg: the memory cgroup
910 * @idx: the event item
911 * @count: the number of events that occured
913 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
916 if (mem_cgroup_disabled())
919 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
920 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
923 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
925 return READ_ONCE(memcg->vmstats.events[event]);
928 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
933 for_each_possible_cpu(cpu)
934 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
938 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
942 /* pagein of a big page is an event. So, ignore page size */
944 __count_memcg_events(memcg, PGPGIN, 1);
946 __count_memcg_events(memcg, PGPGOUT, 1);
947 nr_pages = -nr_pages; /* for event */
950 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
953 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
954 enum mem_cgroup_events_target target)
956 unsigned long val, next;
958 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
959 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
960 /* from time_after() in jiffies.h */
961 if ((long)(next - val) < 0) {
963 case MEM_CGROUP_TARGET_THRESH:
964 next = val + THRESHOLDS_EVENTS_TARGET;
966 case MEM_CGROUP_TARGET_SOFTLIMIT:
967 next = val + SOFTLIMIT_EVENTS_TARGET;
972 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
979 * Check events in order.
982 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
984 /* threshold event is triggered in finer grain than soft limit */
985 if (unlikely(mem_cgroup_event_ratelimit(memcg,
986 MEM_CGROUP_TARGET_THRESH))) {
989 do_softlimit = mem_cgroup_event_ratelimit(memcg,
990 MEM_CGROUP_TARGET_SOFTLIMIT);
991 mem_cgroup_threshold(memcg);
992 if (unlikely(do_softlimit))
993 mem_cgroup_update_tree(memcg, page);
997 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1000 * mm_update_next_owner() may clear mm->owner to NULL
1001 * if it races with swapoff, page migration, etc.
1002 * So this can be called with p == NULL.
1007 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1009 EXPORT_SYMBOL(mem_cgroup_from_task);
1012 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1013 * @mm: mm from which memcg should be extracted. It can be NULL.
1015 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1016 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1019 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1021 struct mem_cgroup *memcg;
1023 if (mem_cgroup_disabled())
1029 * Page cache insertions can happen withou an
1030 * actual mm context, e.g. during disk probing
1031 * on boot, loopback IO, acct() writes etc.
1034 memcg = root_mem_cgroup;
1036 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1037 if (unlikely(!memcg))
1038 memcg = root_mem_cgroup;
1040 } while (!css_tryget(&memcg->css));
1044 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1046 static __always_inline struct mem_cgroup *active_memcg(void)
1049 return this_cpu_read(int_active_memcg);
1051 return current->active_memcg;
1054 static __always_inline struct mem_cgroup *get_active_memcg(void)
1056 struct mem_cgroup *memcg;
1059 memcg = active_memcg();
1060 /* remote memcg must hold a ref. */
1061 if (memcg && WARN_ON_ONCE(!css_tryget(&memcg->css)))
1062 memcg = root_mem_cgroup;
1068 static __always_inline bool memcg_kmem_bypass(void)
1070 /* Allow remote memcg charging from any context. */
1071 if (unlikely(active_memcg()))
1074 /* Memcg to charge can't be determined. */
1075 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1082 * If active memcg is set, do not fallback to current->mm->memcg.
1084 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1086 if (memcg_kmem_bypass())
1089 if (unlikely(active_memcg()))
1090 return get_active_memcg();
1092 return get_mem_cgroup_from_mm(current->mm);
1096 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1097 * @root: hierarchy root
1098 * @prev: previously returned memcg, NULL on first invocation
1099 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1101 * Returns references to children of the hierarchy below @root, or
1102 * @root itself, or %NULL after a full round-trip.
1104 * Caller must pass the return value in @prev on subsequent
1105 * invocations for reference counting, or use mem_cgroup_iter_break()
1106 * to cancel a hierarchy walk before the round-trip is complete.
1108 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1109 * in the hierarchy among all concurrent reclaimers operating on the
1112 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1113 struct mem_cgroup *prev,
1114 struct mem_cgroup_reclaim_cookie *reclaim)
1116 struct mem_cgroup_reclaim_iter *iter;
1117 struct cgroup_subsys_state *css = NULL;
1118 struct mem_cgroup *memcg = NULL;
1119 struct mem_cgroup *pos = NULL;
1121 if (mem_cgroup_disabled())
1125 root = root_mem_cgroup;
1127 if (prev && !reclaim)
1133 struct mem_cgroup_per_node *mz;
1135 mz = root->nodeinfo[reclaim->pgdat->node_id];
1138 if (prev && reclaim->generation != iter->generation)
1142 pos = READ_ONCE(iter->position);
1143 if (!pos || css_tryget(&pos->css))
1146 * css reference reached zero, so iter->position will
1147 * be cleared by ->css_released. However, we should not
1148 * rely on this happening soon, because ->css_released
1149 * is called from a work queue, and by busy-waiting we
1150 * might block it. So we clear iter->position right
1153 (void)cmpxchg(&iter->position, pos, NULL);
1161 css = css_next_descendant_pre(css, &root->css);
1164 * Reclaimers share the hierarchy walk, and a
1165 * new one might jump in right at the end of
1166 * the hierarchy - make sure they see at least
1167 * one group and restart from the beginning.
1175 * Verify the css and acquire a reference. The root
1176 * is provided by the caller, so we know it's alive
1177 * and kicking, and don't take an extra reference.
1179 memcg = mem_cgroup_from_css(css);
1181 if (css == &root->css)
1184 if (css_tryget(css))
1192 * The position could have already been updated by a competing
1193 * thread, so check that the value hasn't changed since we read
1194 * it to avoid reclaiming from the same cgroup twice.
1196 (void)cmpxchg(&iter->position, pos, memcg);
1204 reclaim->generation = iter->generation;
1209 if (prev && prev != root)
1210 css_put(&prev->css);
1216 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1217 * @root: hierarchy root
1218 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1220 void mem_cgroup_iter_break(struct mem_cgroup *root,
1221 struct mem_cgroup *prev)
1224 root = root_mem_cgroup;
1225 if (prev && prev != root)
1226 css_put(&prev->css);
1229 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1230 struct mem_cgroup *dead_memcg)
1232 struct mem_cgroup_reclaim_iter *iter;
1233 struct mem_cgroup_per_node *mz;
1236 for_each_node(nid) {
1237 mz = from->nodeinfo[nid];
1239 cmpxchg(&iter->position, dead_memcg, NULL);
1243 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1245 struct mem_cgroup *memcg = dead_memcg;
1246 struct mem_cgroup *last;
1249 __invalidate_reclaim_iterators(memcg, dead_memcg);
1251 } while ((memcg = parent_mem_cgroup(memcg)));
1254 * When cgruop1 non-hierarchy mode is used,
1255 * parent_mem_cgroup() does not walk all the way up to the
1256 * cgroup root (root_mem_cgroup). So we have to handle
1257 * dead_memcg from cgroup root separately.
1259 if (last != root_mem_cgroup)
1260 __invalidate_reclaim_iterators(root_mem_cgroup,
1265 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1266 * @memcg: hierarchy root
1267 * @fn: function to call for each task
1268 * @arg: argument passed to @fn
1270 * This function iterates over tasks attached to @memcg or to any of its
1271 * descendants and calls @fn for each task. If @fn returns a non-zero
1272 * value, the function breaks the iteration loop and returns the value.
1273 * Otherwise, it will iterate over all tasks and return 0.
1275 * This function must not be called for the root memory cgroup.
1277 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1278 int (*fn)(struct task_struct *, void *), void *arg)
1280 struct mem_cgroup *iter;
1283 BUG_ON(memcg == root_mem_cgroup);
1285 for_each_mem_cgroup_tree(iter, memcg) {
1286 struct css_task_iter it;
1287 struct task_struct *task;
1289 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1290 while (!ret && (task = css_task_iter_next(&it)))
1291 ret = fn(task, arg);
1292 css_task_iter_end(&it);
1294 mem_cgroup_iter_break(memcg, iter);
1301 #ifdef CONFIG_DEBUG_VM
1302 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1304 struct mem_cgroup *memcg;
1306 if (mem_cgroup_disabled())
1309 memcg = page_memcg(page);
1312 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1314 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1319 * lock_page_lruvec - lock and return lruvec for a given page.
1322 * These functions are safe to use under any of the following conditions:
1325 * - lock_page_memcg()
1326 * - page->_refcount is zero
1328 struct lruvec *lock_page_lruvec(struct page *page)
1330 struct lruvec *lruvec;
1331 struct pglist_data *pgdat = page_pgdat(page);
1333 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1334 spin_lock(&lruvec->lru_lock);
1336 lruvec_memcg_debug(lruvec, page);
1341 struct lruvec *lock_page_lruvec_irq(struct page *page)
1343 struct lruvec *lruvec;
1344 struct pglist_data *pgdat = page_pgdat(page);
1346 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1347 spin_lock_irq(&lruvec->lru_lock);
1349 lruvec_memcg_debug(lruvec, page);
1354 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1356 struct lruvec *lruvec;
1357 struct pglist_data *pgdat = page_pgdat(page);
1359 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1360 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1362 lruvec_memcg_debug(lruvec, page);
1368 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1369 * @lruvec: mem_cgroup per zone lru vector
1370 * @lru: index of lru list the page is sitting on
1371 * @zid: zone id of the accounted pages
1372 * @nr_pages: positive when adding or negative when removing
1374 * This function must be called under lru_lock, just before a page is added
1375 * to or just after a page is removed from an lru list (that ordering being
1376 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1378 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1379 int zid, int nr_pages)
1381 struct mem_cgroup_per_node *mz;
1382 unsigned long *lru_size;
1385 if (mem_cgroup_disabled())
1388 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1389 lru_size = &mz->lru_zone_size[zid][lru];
1392 *lru_size += nr_pages;
1395 if (WARN_ONCE(size < 0,
1396 "%s(%p, %d, %d): lru_size %ld\n",
1397 __func__, lruvec, lru, nr_pages, size)) {
1403 *lru_size += nr_pages;
1407 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1408 * @memcg: the memory cgroup
1410 * Returns the maximum amount of memory @mem can be charged with, in
1413 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1415 unsigned long margin = 0;
1416 unsigned long count;
1417 unsigned long limit;
1419 count = page_counter_read(&memcg->memory);
1420 limit = READ_ONCE(memcg->memory.max);
1422 margin = limit - count;
1424 if (do_memsw_account()) {
1425 count = page_counter_read(&memcg->memsw);
1426 limit = READ_ONCE(memcg->memsw.max);
1428 margin = min(margin, limit - count);
1437 * A routine for checking "mem" is under move_account() or not.
1439 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1440 * moving cgroups. This is for waiting at high-memory pressure
1443 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1445 struct mem_cgroup *from;
1446 struct mem_cgroup *to;
1449 * Unlike task_move routines, we access mc.to, mc.from not under
1450 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1452 spin_lock(&mc.lock);
1458 ret = mem_cgroup_is_descendant(from, memcg) ||
1459 mem_cgroup_is_descendant(to, memcg);
1461 spin_unlock(&mc.lock);
1465 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1467 if (mc.moving_task && current != mc.moving_task) {
1468 if (mem_cgroup_under_move(memcg)) {
1470 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1471 /* moving charge context might have finished. */
1474 finish_wait(&mc.waitq, &wait);
1481 struct memory_stat {
1486 static const struct memory_stat memory_stats[] = {
1487 { "anon", NR_ANON_MAPPED },
1488 { "file", NR_FILE_PAGES },
1489 { "kernel_stack", NR_KERNEL_STACK_KB },
1490 { "pagetables", NR_PAGETABLE },
1491 { "percpu", MEMCG_PERCPU_B },
1492 { "sock", MEMCG_SOCK },
1493 { "shmem", NR_SHMEM },
1494 { "file_mapped", NR_FILE_MAPPED },
1495 { "file_dirty", NR_FILE_DIRTY },
1496 { "file_writeback", NR_WRITEBACK },
1498 { "swapcached", NR_SWAPCACHE },
1500 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1501 { "anon_thp", NR_ANON_THPS },
1502 { "file_thp", NR_FILE_THPS },
1503 { "shmem_thp", NR_SHMEM_THPS },
1505 { "inactive_anon", NR_INACTIVE_ANON },
1506 { "active_anon", NR_ACTIVE_ANON },
1507 { "inactive_file", NR_INACTIVE_FILE },
1508 { "active_file", NR_ACTIVE_FILE },
1509 { "unevictable", NR_UNEVICTABLE },
1510 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1511 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1513 /* The memory events */
1514 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1515 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1516 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1517 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1518 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1519 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1520 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1523 /* Translate stat items to the correct unit for memory.stat output */
1524 static int memcg_page_state_unit(int item)
1527 case MEMCG_PERCPU_B:
1528 case NR_SLAB_RECLAIMABLE_B:
1529 case NR_SLAB_UNRECLAIMABLE_B:
1530 case WORKINGSET_REFAULT_ANON:
1531 case WORKINGSET_REFAULT_FILE:
1532 case WORKINGSET_ACTIVATE_ANON:
1533 case WORKINGSET_ACTIVATE_FILE:
1534 case WORKINGSET_RESTORE_ANON:
1535 case WORKINGSET_RESTORE_FILE:
1536 case WORKINGSET_NODERECLAIM:
1538 case NR_KERNEL_STACK_KB:
1545 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1548 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1551 static char *memory_stat_format(struct mem_cgroup *memcg)
1556 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1561 * Provide statistics on the state of the memory subsystem as
1562 * well as cumulative event counters that show past behavior.
1564 * This list is ordered following a combination of these gradients:
1565 * 1) generic big picture -> specifics and details
1566 * 2) reflecting userspace activity -> reflecting kernel heuristics
1568 * Current memory state:
1570 cgroup_rstat_flush(memcg->css.cgroup);
1572 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1575 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1576 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1578 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1579 size += memcg_page_state_output(memcg,
1580 NR_SLAB_RECLAIMABLE_B);
1581 seq_buf_printf(&s, "slab %llu\n", size);
1585 /* Accumulated memory events */
1587 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1588 memcg_events(memcg, PGFAULT));
1589 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1590 memcg_events(memcg, PGMAJFAULT));
1591 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1592 memcg_events(memcg, PGREFILL));
1593 seq_buf_printf(&s, "pgscan %lu\n",
1594 memcg_events(memcg, PGSCAN_KSWAPD) +
1595 memcg_events(memcg, PGSCAN_DIRECT));
1596 seq_buf_printf(&s, "pgsteal %lu\n",
1597 memcg_events(memcg, PGSTEAL_KSWAPD) +
1598 memcg_events(memcg, PGSTEAL_DIRECT));
1599 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1600 memcg_events(memcg, PGACTIVATE));
1601 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1602 memcg_events(memcg, PGDEACTIVATE));
1603 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1604 memcg_events(memcg, PGLAZYFREE));
1605 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1606 memcg_events(memcg, PGLAZYFREED));
1608 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1609 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1610 memcg_events(memcg, THP_FAULT_ALLOC));
1611 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1612 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1613 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1615 /* The above should easily fit into one page */
1616 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1621 #define K(x) ((x) << (PAGE_SHIFT-10))
1623 * mem_cgroup_print_oom_context: Print OOM information relevant to
1624 * memory controller.
1625 * @memcg: The memory cgroup that went over limit
1626 * @p: Task that is going to be killed
1628 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1631 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1636 pr_cont(",oom_memcg=");
1637 pr_cont_cgroup_path(memcg->css.cgroup);
1639 pr_cont(",global_oom");
1641 pr_cont(",task_memcg=");
1642 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1648 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1649 * memory controller.
1650 * @memcg: The memory cgroup that went over limit
1652 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1656 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1657 K((u64)page_counter_read(&memcg->memory)),
1658 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1659 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1660 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1661 K((u64)page_counter_read(&memcg->swap)),
1662 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1664 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1665 K((u64)page_counter_read(&memcg->memsw)),
1666 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1667 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1668 K((u64)page_counter_read(&memcg->kmem)),
1669 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1672 pr_info("Memory cgroup stats for ");
1673 pr_cont_cgroup_path(memcg->css.cgroup);
1675 buf = memory_stat_format(memcg);
1683 * Return the memory (and swap, if configured) limit for a memcg.
1685 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1687 unsigned long max = READ_ONCE(memcg->memory.max);
1689 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1690 if (mem_cgroup_swappiness(memcg))
1691 max += min(READ_ONCE(memcg->swap.max),
1692 (unsigned long)total_swap_pages);
1694 if (mem_cgroup_swappiness(memcg)) {
1695 /* Calculate swap excess capacity from memsw limit */
1696 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1698 max += min(swap, (unsigned long)total_swap_pages);
1704 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1706 return page_counter_read(&memcg->memory);
1709 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1712 struct oom_control oc = {
1716 .gfp_mask = gfp_mask,
1721 if (mutex_lock_killable(&oom_lock))
1724 if (mem_cgroup_margin(memcg) >= (1 << order))
1728 * A few threads which were not waiting at mutex_lock_killable() can
1729 * fail to bail out. Therefore, check again after holding oom_lock.
1731 ret = should_force_charge() || out_of_memory(&oc);
1734 mutex_unlock(&oom_lock);
1738 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1741 unsigned long *total_scanned)
1743 struct mem_cgroup *victim = NULL;
1746 unsigned long excess;
1747 unsigned long nr_scanned;
1748 struct mem_cgroup_reclaim_cookie reclaim = {
1752 excess = soft_limit_excess(root_memcg);
1755 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1760 * If we have not been able to reclaim
1761 * anything, it might because there are
1762 * no reclaimable pages under this hierarchy
1767 * We want to do more targeted reclaim.
1768 * excess >> 2 is not to excessive so as to
1769 * reclaim too much, nor too less that we keep
1770 * coming back to reclaim from this cgroup
1772 if (total >= (excess >> 2) ||
1773 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1778 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1779 pgdat, &nr_scanned);
1780 *total_scanned += nr_scanned;
1781 if (!soft_limit_excess(root_memcg))
1784 mem_cgroup_iter_break(root_memcg, victim);
1788 #ifdef CONFIG_LOCKDEP
1789 static struct lockdep_map memcg_oom_lock_dep_map = {
1790 .name = "memcg_oom_lock",
1794 static DEFINE_SPINLOCK(memcg_oom_lock);
1797 * Check OOM-Killer is already running under our hierarchy.
1798 * If someone is running, return false.
1800 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1802 struct mem_cgroup *iter, *failed = NULL;
1804 spin_lock(&memcg_oom_lock);
1806 for_each_mem_cgroup_tree(iter, memcg) {
1807 if (iter->oom_lock) {
1809 * this subtree of our hierarchy is already locked
1810 * so we cannot give a lock.
1813 mem_cgroup_iter_break(memcg, iter);
1816 iter->oom_lock = true;
1821 * OK, we failed to lock the whole subtree so we have
1822 * to clean up what we set up to the failing subtree
1824 for_each_mem_cgroup_tree(iter, memcg) {
1825 if (iter == failed) {
1826 mem_cgroup_iter_break(memcg, iter);
1829 iter->oom_lock = false;
1832 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1834 spin_unlock(&memcg_oom_lock);
1839 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1841 struct mem_cgroup *iter;
1843 spin_lock(&memcg_oom_lock);
1844 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1845 for_each_mem_cgroup_tree(iter, memcg)
1846 iter->oom_lock = false;
1847 spin_unlock(&memcg_oom_lock);
1850 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1852 struct mem_cgroup *iter;
1854 spin_lock(&memcg_oom_lock);
1855 for_each_mem_cgroup_tree(iter, memcg)
1857 spin_unlock(&memcg_oom_lock);
1860 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1862 struct mem_cgroup *iter;
1865 * Be careful about under_oom underflows becase a child memcg
1866 * could have been added after mem_cgroup_mark_under_oom.
1868 spin_lock(&memcg_oom_lock);
1869 for_each_mem_cgroup_tree(iter, memcg)
1870 if (iter->under_oom > 0)
1872 spin_unlock(&memcg_oom_lock);
1875 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1877 struct oom_wait_info {
1878 struct mem_cgroup *memcg;
1879 wait_queue_entry_t wait;
1882 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1883 unsigned mode, int sync, void *arg)
1885 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1886 struct mem_cgroup *oom_wait_memcg;
1887 struct oom_wait_info *oom_wait_info;
1889 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1890 oom_wait_memcg = oom_wait_info->memcg;
1892 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1893 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1895 return autoremove_wake_function(wait, mode, sync, arg);
1898 static void memcg_oom_recover(struct mem_cgroup *memcg)
1901 * For the following lockless ->under_oom test, the only required
1902 * guarantee is that it must see the state asserted by an OOM when
1903 * this function is called as a result of userland actions
1904 * triggered by the notification of the OOM. This is trivially
1905 * achieved by invoking mem_cgroup_mark_under_oom() before
1906 * triggering notification.
1908 if (memcg && memcg->under_oom)
1909 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1919 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1921 enum oom_status ret;
1924 if (order > PAGE_ALLOC_COSTLY_ORDER)
1927 memcg_memory_event(memcg, MEMCG_OOM);
1930 * We are in the middle of the charge context here, so we
1931 * don't want to block when potentially sitting on a callstack
1932 * that holds all kinds of filesystem and mm locks.
1934 * cgroup1 allows disabling the OOM killer and waiting for outside
1935 * handling until the charge can succeed; remember the context and put
1936 * the task to sleep at the end of the page fault when all locks are
1939 * On the other hand, in-kernel OOM killer allows for an async victim
1940 * memory reclaim (oom_reaper) and that means that we are not solely
1941 * relying on the oom victim to make a forward progress and we can
1942 * invoke the oom killer here.
1944 * Please note that mem_cgroup_out_of_memory might fail to find a
1945 * victim and then we have to bail out from the charge path.
1947 if (memcg->oom_kill_disable) {
1948 if (!current->in_user_fault)
1950 css_get(&memcg->css);
1951 current->memcg_in_oom = memcg;
1952 current->memcg_oom_gfp_mask = mask;
1953 current->memcg_oom_order = order;
1958 mem_cgroup_mark_under_oom(memcg);
1960 locked = mem_cgroup_oom_trylock(memcg);
1963 mem_cgroup_oom_notify(memcg);
1965 mem_cgroup_unmark_under_oom(memcg);
1966 if (mem_cgroup_out_of_memory(memcg, mask, order))
1972 mem_cgroup_oom_unlock(memcg);
1978 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1979 * @handle: actually kill/wait or just clean up the OOM state
1981 * This has to be called at the end of a page fault if the memcg OOM
1982 * handler was enabled.
1984 * Memcg supports userspace OOM handling where failed allocations must
1985 * sleep on a waitqueue until the userspace task resolves the
1986 * situation. Sleeping directly in the charge context with all kinds
1987 * of locks held is not a good idea, instead we remember an OOM state
1988 * in the task and mem_cgroup_oom_synchronize() has to be called at
1989 * the end of the page fault to complete the OOM handling.
1991 * Returns %true if an ongoing memcg OOM situation was detected and
1992 * completed, %false otherwise.
1994 bool mem_cgroup_oom_synchronize(bool handle)
1996 struct mem_cgroup *memcg = current->memcg_in_oom;
1997 struct oom_wait_info owait;
2000 /* OOM is global, do not handle */
2007 owait.memcg = memcg;
2008 owait.wait.flags = 0;
2009 owait.wait.func = memcg_oom_wake_function;
2010 owait.wait.private = current;
2011 INIT_LIST_HEAD(&owait.wait.entry);
2013 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2014 mem_cgroup_mark_under_oom(memcg);
2016 locked = mem_cgroup_oom_trylock(memcg);
2019 mem_cgroup_oom_notify(memcg);
2021 if (locked && !memcg->oom_kill_disable) {
2022 mem_cgroup_unmark_under_oom(memcg);
2023 finish_wait(&memcg_oom_waitq, &owait.wait);
2024 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2025 current->memcg_oom_order);
2028 mem_cgroup_unmark_under_oom(memcg);
2029 finish_wait(&memcg_oom_waitq, &owait.wait);
2033 mem_cgroup_oom_unlock(memcg);
2035 * There is no guarantee that an OOM-lock contender
2036 * sees the wakeups triggered by the OOM kill
2037 * uncharges. Wake any sleepers explicitely.
2039 memcg_oom_recover(memcg);
2042 current->memcg_in_oom = NULL;
2043 css_put(&memcg->css);
2048 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2049 * @victim: task to be killed by the OOM killer
2050 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2052 * Returns a pointer to a memory cgroup, which has to be cleaned up
2053 * by killing all belonging OOM-killable tasks.
2055 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2057 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2058 struct mem_cgroup *oom_domain)
2060 struct mem_cgroup *oom_group = NULL;
2061 struct mem_cgroup *memcg;
2063 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2067 oom_domain = root_mem_cgroup;
2071 memcg = mem_cgroup_from_task(victim);
2072 if (memcg == root_mem_cgroup)
2076 * If the victim task has been asynchronously moved to a different
2077 * memory cgroup, we might end up killing tasks outside oom_domain.
2078 * In this case it's better to ignore memory.group.oom.
2080 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2084 * Traverse the memory cgroup hierarchy from the victim task's
2085 * cgroup up to the OOMing cgroup (or root) to find the
2086 * highest-level memory cgroup with oom.group set.
2088 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2089 if (memcg->oom_group)
2092 if (memcg == oom_domain)
2097 css_get(&oom_group->css);
2104 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2106 pr_info("Tasks in ");
2107 pr_cont_cgroup_path(memcg->css.cgroup);
2108 pr_cont(" are going to be killed due to memory.oom.group set\n");
2112 * lock_page_memcg - lock a page and memcg binding
2115 * This function protects unlocked LRU pages from being moved to
2118 * It ensures lifetime of the locked memcg. Caller is responsible
2119 * for the lifetime of the page.
2121 void lock_page_memcg(struct page *page)
2123 struct page *head = compound_head(page); /* rmap on tail pages */
2124 struct mem_cgroup *memcg;
2125 unsigned long flags;
2128 * The RCU lock is held throughout the transaction. The fast
2129 * path can get away without acquiring the memcg->move_lock
2130 * because page moving starts with an RCU grace period.
2134 if (mem_cgroup_disabled())
2137 memcg = page_memcg(head);
2138 if (unlikely(!memcg))
2141 #ifdef CONFIG_PROVE_LOCKING
2142 local_irq_save(flags);
2143 might_lock(&memcg->move_lock);
2144 local_irq_restore(flags);
2147 if (atomic_read(&memcg->moving_account) <= 0)
2150 spin_lock_irqsave(&memcg->move_lock, flags);
2151 if (memcg != page_memcg(head)) {
2152 spin_unlock_irqrestore(&memcg->move_lock, flags);
2157 * When charge migration first begins, we can have multiple
2158 * critical sections holding the fast-path RCU lock and one
2159 * holding the slowpath move_lock. Track the task who has the
2160 * move_lock for unlock_page_memcg().
2162 memcg->move_lock_task = current;
2163 memcg->move_lock_flags = flags;
2165 EXPORT_SYMBOL(lock_page_memcg);
2167 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2169 if (memcg && memcg->move_lock_task == current) {
2170 unsigned long flags = memcg->move_lock_flags;
2172 memcg->move_lock_task = NULL;
2173 memcg->move_lock_flags = 0;
2175 spin_unlock_irqrestore(&memcg->move_lock, flags);
2182 * unlock_page_memcg - unlock a page and memcg binding
2185 void unlock_page_memcg(struct page *page)
2187 struct page *head = compound_head(page);
2189 __unlock_page_memcg(page_memcg(head));
2191 EXPORT_SYMBOL(unlock_page_memcg);
2193 struct memcg_stock_pcp {
2194 struct mem_cgroup *cached; /* this never be root cgroup */
2195 unsigned int nr_pages;
2197 #ifdef CONFIG_MEMCG_KMEM
2198 struct obj_cgroup *cached_objcg;
2199 unsigned int nr_bytes;
2202 struct work_struct work;
2203 unsigned long flags;
2204 #define FLUSHING_CACHED_CHARGE 0
2206 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2207 static DEFINE_MUTEX(percpu_charge_mutex);
2209 #ifdef CONFIG_MEMCG_KMEM
2210 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2211 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2212 struct mem_cgroup *root_memcg);
2215 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2218 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2219 struct mem_cgroup *root_memcg)
2226 * consume_stock: Try to consume stocked charge on this cpu.
2227 * @memcg: memcg to consume from.
2228 * @nr_pages: how many pages to charge.
2230 * The charges will only happen if @memcg matches the current cpu's memcg
2231 * stock, and at least @nr_pages are available in that stock. Failure to
2232 * service an allocation will refill the stock.
2234 * returns true if successful, false otherwise.
2236 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2238 struct memcg_stock_pcp *stock;
2239 unsigned long flags;
2242 if (nr_pages > MEMCG_CHARGE_BATCH)
2245 local_irq_save(flags);
2247 stock = this_cpu_ptr(&memcg_stock);
2248 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2249 stock->nr_pages -= nr_pages;
2253 local_irq_restore(flags);
2259 * Returns stocks cached in percpu and reset cached information.
2261 static void drain_stock(struct memcg_stock_pcp *stock)
2263 struct mem_cgroup *old = stock->cached;
2268 if (stock->nr_pages) {
2269 page_counter_uncharge(&old->memory, stock->nr_pages);
2270 if (do_memsw_account())
2271 page_counter_uncharge(&old->memsw, stock->nr_pages);
2272 stock->nr_pages = 0;
2276 stock->cached = NULL;
2279 static void drain_local_stock(struct work_struct *dummy)
2281 struct memcg_stock_pcp *stock;
2282 unsigned long flags;
2285 * The only protection from memory hotplug vs. drain_stock races is
2286 * that we always operate on local CPU stock here with IRQ disabled
2288 local_irq_save(flags);
2290 stock = this_cpu_ptr(&memcg_stock);
2291 drain_obj_stock(stock);
2293 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2295 local_irq_restore(flags);
2299 * Cache charges(val) to local per_cpu area.
2300 * This will be consumed by consume_stock() function, later.
2302 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2304 struct memcg_stock_pcp *stock;
2305 unsigned long flags;
2307 local_irq_save(flags);
2309 stock = this_cpu_ptr(&memcg_stock);
2310 if (stock->cached != memcg) { /* reset if necessary */
2312 css_get(&memcg->css);
2313 stock->cached = memcg;
2315 stock->nr_pages += nr_pages;
2317 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2320 local_irq_restore(flags);
2324 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2325 * of the hierarchy under it.
2327 static void drain_all_stock(struct mem_cgroup *root_memcg)
2331 /* If someone's already draining, avoid adding running more workers. */
2332 if (!mutex_trylock(&percpu_charge_mutex))
2335 * Notify other cpus that system-wide "drain" is running
2336 * We do not care about races with the cpu hotplug because cpu down
2337 * as well as workers from this path always operate on the local
2338 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2341 for_each_online_cpu(cpu) {
2342 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2343 struct mem_cgroup *memcg;
2347 memcg = stock->cached;
2348 if (memcg && stock->nr_pages &&
2349 mem_cgroup_is_descendant(memcg, root_memcg))
2351 if (obj_stock_flush_required(stock, root_memcg))
2356 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2358 drain_local_stock(&stock->work);
2360 schedule_work_on(cpu, &stock->work);
2364 mutex_unlock(&percpu_charge_mutex);
2367 static void memcg_flush_lruvec_page_state(struct mem_cgroup *memcg, int cpu)
2371 for_each_node(nid) {
2372 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
2373 unsigned long stat[NR_VM_NODE_STAT_ITEMS];
2374 struct batched_lruvec_stat *lstatc;
2377 lstatc = per_cpu_ptr(pn->lruvec_stat_cpu, cpu);
2378 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
2379 stat[i] = lstatc->count[i];
2380 lstatc->count[i] = 0;
2384 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
2385 atomic_long_add(stat[i], &pn->lruvec_stat[i]);
2386 } while ((pn = parent_nodeinfo(pn, nid)));
2390 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2392 struct memcg_stock_pcp *stock;
2393 struct mem_cgroup *memcg;
2395 stock = &per_cpu(memcg_stock, cpu);
2398 for_each_mem_cgroup(memcg)
2399 memcg_flush_lruvec_page_state(memcg, cpu);
2404 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2405 unsigned int nr_pages,
2408 unsigned long nr_reclaimed = 0;
2411 unsigned long pflags;
2413 if (page_counter_read(&memcg->memory) <=
2414 READ_ONCE(memcg->memory.high))
2417 memcg_memory_event(memcg, MEMCG_HIGH);
2419 psi_memstall_enter(&pflags);
2420 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2422 psi_memstall_leave(&pflags);
2423 } while ((memcg = parent_mem_cgroup(memcg)) &&
2424 !mem_cgroup_is_root(memcg));
2426 return nr_reclaimed;
2429 static void high_work_func(struct work_struct *work)
2431 struct mem_cgroup *memcg;
2433 memcg = container_of(work, struct mem_cgroup, high_work);
2434 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2438 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2439 * enough to still cause a significant slowdown in most cases, while still
2440 * allowing diagnostics and tracing to proceed without becoming stuck.
2442 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2445 * When calculating the delay, we use these either side of the exponentiation to
2446 * maintain precision and scale to a reasonable number of jiffies (see the table
2449 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2450 * overage ratio to a delay.
2451 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2452 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2453 * to produce a reasonable delay curve.
2455 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2456 * reasonable delay curve compared to precision-adjusted overage, not
2457 * penalising heavily at first, but still making sure that growth beyond the
2458 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2459 * example, with a high of 100 megabytes:
2461 * +-------+------------------------+
2462 * | usage | time to allocate in ms |
2463 * +-------+------------------------+
2485 * +-------+------------------------+
2487 #define MEMCG_DELAY_PRECISION_SHIFT 20
2488 #define MEMCG_DELAY_SCALING_SHIFT 14
2490 static u64 calculate_overage(unsigned long usage, unsigned long high)
2498 * Prevent division by 0 in overage calculation by acting as if
2499 * it was a threshold of 1 page
2501 high = max(high, 1UL);
2503 overage = usage - high;
2504 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2505 return div64_u64(overage, high);
2508 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2510 u64 overage, max_overage = 0;
2513 overage = calculate_overage(page_counter_read(&memcg->memory),
2514 READ_ONCE(memcg->memory.high));
2515 max_overage = max(overage, max_overage);
2516 } while ((memcg = parent_mem_cgroup(memcg)) &&
2517 !mem_cgroup_is_root(memcg));
2522 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2524 u64 overage, max_overage = 0;
2527 overage = calculate_overage(page_counter_read(&memcg->swap),
2528 READ_ONCE(memcg->swap.high));
2530 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2531 max_overage = max(overage, max_overage);
2532 } while ((memcg = parent_mem_cgroup(memcg)) &&
2533 !mem_cgroup_is_root(memcg));
2539 * Get the number of jiffies that we should penalise a mischievous cgroup which
2540 * is exceeding its memory.high by checking both it and its ancestors.
2542 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2543 unsigned int nr_pages,
2546 unsigned long penalty_jiffies;
2552 * We use overage compared to memory.high to calculate the number of
2553 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2554 * fairly lenient on small overages, and increasingly harsh when the
2555 * memcg in question makes it clear that it has no intention of stopping
2556 * its crazy behaviour, so we exponentially increase the delay based on
2559 penalty_jiffies = max_overage * max_overage * HZ;
2560 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2561 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2564 * Factor in the task's own contribution to the overage, such that four
2565 * N-sized allocations are throttled approximately the same as one
2566 * 4N-sized allocation.
2568 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2569 * larger the current charge patch is than that.
2571 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2575 * Scheduled by try_charge() to be executed from the userland return path
2576 * and reclaims memory over the high limit.
2578 void mem_cgroup_handle_over_high(void)
2580 unsigned long penalty_jiffies;
2581 unsigned long pflags;
2582 unsigned long nr_reclaimed;
2583 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2584 int nr_retries = MAX_RECLAIM_RETRIES;
2585 struct mem_cgroup *memcg;
2586 bool in_retry = false;
2588 if (likely(!nr_pages))
2591 memcg = get_mem_cgroup_from_mm(current->mm);
2592 current->memcg_nr_pages_over_high = 0;
2596 * The allocating task should reclaim at least the batch size, but for
2597 * subsequent retries we only want to do what's necessary to prevent oom
2598 * or breaching resource isolation.
2600 * This is distinct from memory.max or page allocator behaviour because
2601 * memory.high is currently batched, whereas memory.max and the page
2602 * allocator run every time an allocation is made.
2604 nr_reclaimed = reclaim_high(memcg,
2605 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2609 * memory.high is breached and reclaim is unable to keep up. Throttle
2610 * allocators proactively to slow down excessive growth.
2612 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2613 mem_find_max_overage(memcg));
2615 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2616 swap_find_max_overage(memcg));
2619 * Clamp the max delay per usermode return so as to still keep the
2620 * application moving forwards and also permit diagnostics, albeit
2623 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2626 * Don't sleep if the amount of jiffies this memcg owes us is so low
2627 * that it's not even worth doing, in an attempt to be nice to those who
2628 * go only a small amount over their memory.high value and maybe haven't
2629 * been aggressively reclaimed enough yet.
2631 if (penalty_jiffies <= HZ / 100)
2635 * If reclaim is making forward progress but we're still over
2636 * memory.high, we want to encourage that rather than doing allocator
2639 if (nr_reclaimed || nr_retries--) {
2645 * If we exit early, we're guaranteed to die (since
2646 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2647 * need to account for any ill-begotten jiffies to pay them off later.
2649 psi_memstall_enter(&pflags);
2650 schedule_timeout_killable(penalty_jiffies);
2651 psi_memstall_leave(&pflags);
2654 css_put(&memcg->css);
2657 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2658 unsigned int nr_pages)
2660 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2661 int nr_retries = MAX_RECLAIM_RETRIES;
2662 struct mem_cgroup *mem_over_limit;
2663 struct page_counter *counter;
2664 enum oom_status oom_status;
2665 unsigned long nr_reclaimed;
2666 bool may_swap = true;
2667 bool drained = false;
2668 unsigned long pflags;
2670 if (mem_cgroup_is_root(memcg))
2673 if (consume_stock(memcg, nr_pages))
2676 if (!do_memsw_account() ||
2677 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2678 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2680 if (do_memsw_account())
2681 page_counter_uncharge(&memcg->memsw, batch);
2682 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2684 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2688 if (batch > nr_pages) {
2694 * Memcg doesn't have a dedicated reserve for atomic
2695 * allocations. But like the global atomic pool, we need to
2696 * put the burden of reclaim on regular allocation requests
2697 * and let these go through as privileged allocations.
2699 if (gfp_mask & __GFP_ATOMIC)
2703 * Unlike in global OOM situations, memcg is not in a physical
2704 * memory shortage. Allow dying and OOM-killed tasks to
2705 * bypass the last charges so that they can exit quickly and
2706 * free their memory.
2708 if (unlikely(should_force_charge()))
2712 * Prevent unbounded recursion when reclaim operations need to
2713 * allocate memory. This might exceed the limits temporarily,
2714 * but we prefer facilitating memory reclaim and getting back
2715 * under the limit over triggering OOM kills in these cases.
2717 if (unlikely(current->flags & PF_MEMALLOC))
2720 if (unlikely(task_in_memcg_oom(current)))
2723 if (!gfpflags_allow_blocking(gfp_mask))
2726 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2728 psi_memstall_enter(&pflags);
2729 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2730 gfp_mask, may_swap);
2731 psi_memstall_leave(&pflags);
2733 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2737 drain_all_stock(mem_over_limit);
2742 if (gfp_mask & __GFP_NORETRY)
2745 * Even though the limit is exceeded at this point, reclaim
2746 * may have been able to free some pages. Retry the charge
2747 * before killing the task.
2749 * Only for regular pages, though: huge pages are rather
2750 * unlikely to succeed so close to the limit, and we fall back
2751 * to regular pages anyway in case of failure.
2753 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2756 * At task move, charge accounts can be doubly counted. So, it's
2757 * better to wait until the end of task_move if something is going on.
2759 if (mem_cgroup_wait_acct_move(mem_over_limit))
2765 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2768 if (fatal_signal_pending(current))
2772 * keep retrying as long as the memcg oom killer is able to make
2773 * a forward progress or bypass the charge if the oom killer
2774 * couldn't make any progress.
2776 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2777 get_order(nr_pages * PAGE_SIZE));
2778 switch (oom_status) {
2780 nr_retries = MAX_RECLAIM_RETRIES;
2788 if (!(gfp_mask & __GFP_NOFAIL))
2792 * The allocation either can't fail or will lead to more memory
2793 * being freed very soon. Allow memory usage go over the limit
2794 * temporarily by force charging it.
2796 page_counter_charge(&memcg->memory, nr_pages);
2797 if (do_memsw_account())
2798 page_counter_charge(&memcg->memsw, nr_pages);
2803 if (batch > nr_pages)
2804 refill_stock(memcg, batch - nr_pages);
2807 * If the hierarchy is above the normal consumption range, schedule
2808 * reclaim on returning to userland. We can perform reclaim here
2809 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2810 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2811 * not recorded as it most likely matches current's and won't
2812 * change in the meantime. As high limit is checked again before
2813 * reclaim, the cost of mismatch is negligible.
2816 bool mem_high, swap_high;
2818 mem_high = page_counter_read(&memcg->memory) >
2819 READ_ONCE(memcg->memory.high);
2820 swap_high = page_counter_read(&memcg->swap) >
2821 READ_ONCE(memcg->swap.high);
2823 /* Don't bother a random interrupted task */
2824 if (in_interrupt()) {
2826 schedule_work(&memcg->high_work);
2832 if (mem_high || swap_high) {
2834 * The allocating tasks in this cgroup will need to do
2835 * reclaim or be throttled to prevent further growth
2836 * of the memory or swap footprints.
2838 * Target some best-effort fairness between the tasks,
2839 * and distribute reclaim work and delay penalties
2840 * based on how much each task is actually allocating.
2842 current->memcg_nr_pages_over_high += batch;
2843 set_notify_resume(current);
2846 } while ((memcg = parent_mem_cgroup(memcg)));
2851 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2852 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2854 if (mem_cgroup_is_root(memcg))
2857 page_counter_uncharge(&memcg->memory, nr_pages);
2858 if (do_memsw_account())
2859 page_counter_uncharge(&memcg->memsw, nr_pages);
2863 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2865 VM_BUG_ON_PAGE(page_memcg(page), page);
2867 * Any of the following ensures page's memcg stability:
2871 * - lock_page_memcg()
2872 * - exclusive reference
2874 page->memcg_data = (unsigned long)memcg;
2877 #ifdef CONFIG_MEMCG_KMEM
2878 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2879 gfp_t gfp, bool new_page)
2881 unsigned int objects = objs_per_slab_page(s, page);
2882 unsigned long memcg_data;
2885 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2890 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2893 * If the slab page is brand new and nobody can yet access
2894 * it's memcg_data, no synchronization is required and
2895 * memcg_data can be simply assigned.
2897 page->memcg_data = memcg_data;
2898 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2900 * If the slab page is already in use, somebody can allocate
2901 * and assign obj_cgroups in parallel. In this case the existing
2902 * objcg vector should be reused.
2908 kmemleak_not_leak(vec);
2913 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2915 * A passed kernel object can be a slab object or a generic kernel page, so
2916 * different mechanisms for getting the memory cgroup pointer should be used.
2917 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2918 * can not know for sure how the kernel object is implemented.
2919 * mem_cgroup_from_obj() can be safely used in such cases.
2921 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2922 * cgroup_mutex, etc.
2924 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2928 if (mem_cgroup_disabled())
2931 page = virt_to_head_page(p);
2934 * Slab objects are accounted individually, not per-page.
2935 * Memcg membership data for each individual object is saved in
2936 * the page->obj_cgroups.
2938 if (page_objcgs_check(page)) {
2939 struct obj_cgroup *objcg;
2942 off = obj_to_index(page->slab_cache, page, p);
2943 objcg = page_objcgs(page)[off];
2945 return obj_cgroup_memcg(objcg);
2951 * page_memcg_check() is used here, because page_has_obj_cgroups()
2952 * check above could fail because the object cgroups vector wasn't set
2953 * at that moment, but it can be set concurrently.
2954 * page_memcg_check(page) will guarantee that a proper memory
2955 * cgroup pointer or NULL will be returned.
2957 return page_memcg_check(page);
2960 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2962 struct obj_cgroup *objcg = NULL;
2963 struct mem_cgroup *memcg;
2965 if (memcg_kmem_bypass())
2969 if (unlikely(active_memcg()))
2970 memcg = active_memcg();
2972 memcg = mem_cgroup_from_task(current);
2974 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2975 objcg = rcu_dereference(memcg->objcg);
2976 if (objcg && obj_cgroup_tryget(objcg))
2985 static int memcg_alloc_cache_id(void)
2990 id = ida_simple_get(&memcg_cache_ida,
2991 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2995 if (id < memcg_nr_cache_ids)
2999 * There's no space for the new id in memcg_caches arrays,
3000 * so we have to grow them.
3002 down_write(&memcg_cache_ids_sem);
3004 size = 2 * (id + 1);
3005 if (size < MEMCG_CACHES_MIN_SIZE)
3006 size = MEMCG_CACHES_MIN_SIZE;
3007 else if (size > MEMCG_CACHES_MAX_SIZE)
3008 size = MEMCG_CACHES_MAX_SIZE;
3010 err = memcg_update_all_list_lrus(size);
3012 memcg_nr_cache_ids = size;
3014 up_write(&memcg_cache_ids_sem);
3017 ida_simple_remove(&memcg_cache_ida, id);
3023 static void memcg_free_cache_id(int id)
3025 ida_simple_remove(&memcg_cache_ida, id);
3029 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3030 * @memcg: memory cgroup to charge
3031 * @gfp: reclaim mode
3032 * @nr_pages: number of pages to charge
3034 * Returns 0 on success, an error code on failure.
3036 static int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3037 unsigned int nr_pages)
3039 struct page_counter *counter;
3042 ret = try_charge(memcg, gfp, nr_pages);
3046 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3047 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3050 * Enforce __GFP_NOFAIL allocation because callers are not
3051 * prepared to see failures and likely do not have any failure
3054 if (gfp & __GFP_NOFAIL) {
3055 page_counter_charge(&memcg->kmem, nr_pages);
3058 cancel_charge(memcg, nr_pages);
3065 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3066 * @memcg: memcg to uncharge
3067 * @nr_pages: number of pages to uncharge
3069 static void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3071 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3072 page_counter_uncharge(&memcg->kmem, nr_pages);
3074 refill_stock(memcg, nr_pages);
3078 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3079 * @page: page to charge
3080 * @gfp: reclaim mode
3081 * @order: allocation order
3083 * Returns 0 on success, an error code on failure.
3085 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3087 struct mem_cgroup *memcg;
3090 memcg = get_mem_cgroup_from_current();
3091 if (memcg && !mem_cgroup_is_root(memcg)) {
3092 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3094 page->memcg_data = (unsigned long)memcg |
3098 css_put(&memcg->css);
3104 * __memcg_kmem_uncharge_page: uncharge a kmem page
3105 * @page: page to uncharge
3106 * @order: allocation order
3108 void __memcg_kmem_uncharge_page(struct page *page, int order)
3110 struct mem_cgroup *memcg = page_memcg(page);
3111 unsigned int nr_pages = 1 << order;
3116 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3117 __memcg_kmem_uncharge(memcg, nr_pages);
3118 page->memcg_data = 0;
3119 css_put(&memcg->css);
3122 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3124 struct memcg_stock_pcp *stock;
3125 unsigned long flags;
3128 local_irq_save(flags);
3130 stock = this_cpu_ptr(&memcg_stock);
3131 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3132 stock->nr_bytes -= nr_bytes;
3136 local_irq_restore(flags);
3141 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3143 struct obj_cgroup *old = stock->cached_objcg;
3148 if (stock->nr_bytes) {
3149 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3150 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3154 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3159 * The leftover is flushed to the centralized per-memcg value.
3160 * On the next attempt to refill obj stock it will be moved
3161 * to a per-cpu stock (probably, on an other CPU), see
3162 * refill_obj_stock().
3164 * How often it's flushed is a trade-off between the memory
3165 * limit enforcement accuracy and potential CPU contention,
3166 * so it might be changed in the future.
3168 atomic_add(nr_bytes, &old->nr_charged_bytes);
3169 stock->nr_bytes = 0;
3172 obj_cgroup_put(old);
3173 stock->cached_objcg = NULL;
3176 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3177 struct mem_cgroup *root_memcg)
3179 struct mem_cgroup *memcg;
3181 if (stock->cached_objcg) {
3182 memcg = obj_cgroup_memcg(stock->cached_objcg);
3183 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3190 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3192 struct memcg_stock_pcp *stock;
3193 unsigned long flags;
3195 local_irq_save(flags);
3197 stock = this_cpu_ptr(&memcg_stock);
3198 if (stock->cached_objcg != objcg) { /* reset if necessary */
3199 drain_obj_stock(stock);
3200 obj_cgroup_get(objcg);
3201 stock->cached_objcg = objcg;
3202 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3204 stock->nr_bytes += nr_bytes;
3206 if (stock->nr_bytes > PAGE_SIZE)
3207 drain_obj_stock(stock);
3209 local_irq_restore(flags);
3212 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3214 struct mem_cgroup *memcg;
3215 unsigned int nr_pages, nr_bytes;
3218 if (consume_obj_stock(objcg, size))
3222 * In theory, memcg->nr_charged_bytes can have enough
3223 * pre-charged bytes to satisfy the allocation. However,
3224 * flushing memcg->nr_charged_bytes requires two atomic
3225 * operations, and memcg->nr_charged_bytes can't be big,
3226 * so it's better to ignore it and try grab some new pages.
3227 * memcg->nr_charged_bytes will be flushed in
3228 * refill_obj_stock(), called from this function or
3229 * independently later.
3233 memcg = obj_cgroup_memcg(objcg);
3234 if (unlikely(!css_tryget(&memcg->css)))
3238 nr_pages = size >> PAGE_SHIFT;
3239 nr_bytes = size & (PAGE_SIZE - 1);
3244 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3245 if (!ret && nr_bytes)
3246 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3248 css_put(&memcg->css);
3252 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3254 refill_obj_stock(objcg, size);
3257 #endif /* CONFIG_MEMCG_KMEM */
3260 * Because page_memcg(head) is not set on tails, set it now.
3262 void split_page_memcg(struct page *head, unsigned int nr)
3264 struct mem_cgroup *memcg = page_memcg(head);
3267 if (mem_cgroup_disabled() || !memcg)
3270 for (i = 1; i < nr; i++)
3271 head[i].memcg_data = head->memcg_data;
3272 css_get_many(&memcg->css, nr - 1);
3275 #ifdef CONFIG_MEMCG_SWAP
3277 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3278 * @entry: swap entry to be moved
3279 * @from: mem_cgroup which the entry is moved from
3280 * @to: mem_cgroup which the entry is moved to
3282 * It succeeds only when the swap_cgroup's record for this entry is the same
3283 * as the mem_cgroup's id of @from.
3285 * Returns 0 on success, -EINVAL on failure.
3287 * The caller must have charged to @to, IOW, called page_counter_charge() about
3288 * both res and memsw, and called css_get().
3290 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3291 struct mem_cgroup *from, struct mem_cgroup *to)
3293 unsigned short old_id, new_id;
3295 old_id = mem_cgroup_id(from);
3296 new_id = mem_cgroup_id(to);
3298 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3299 mod_memcg_state(from, MEMCG_SWAP, -1);
3300 mod_memcg_state(to, MEMCG_SWAP, 1);
3306 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3307 struct mem_cgroup *from, struct mem_cgroup *to)
3313 static DEFINE_MUTEX(memcg_max_mutex);
3315 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3316 unsigned long max, bool memsw)
3318 bool enlarge = false;
3319 bool drained = false;
3321 bool limits_invariant;
3322 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3325 if (signal_pending(current)) {
3330 mutex_lock(&memcg_max_mutex);
3332 * Make sure that the new limit (memsw or memory limit) doesn't
3333 * break our basic invariant rule memory.max <= memsw.max.
3335 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3336 max <= memcg->memsw.max;
3337 if (!limits_invariant) {
3338 mutex_unlock(&memcg_max_mutex);
3342 if (max > counter->max)
3344 ret = page_counter_set_max(counter, max);
3345 mutex_unlock(&memcg_max_mutex);
3351 drain_all_stock(memcg);
3356 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3357 GFP_KERNEL, !memsw)) {
3363 if (!ret && enlarge)
3364 memcg_oom_recover(memcg);
3369 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3371 unsigned long *total_scanned)
3373 unsigned long nr_reclaimed = 0;
3374 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3375 unsigned long reclaimed;
3377 struct mem_cgroup_tree_per_node *mctz;
3378 unsigned long excess;
3379 unsigned long nr_scanned;
3384 mctz = soft_limit_tree_node(pgdat->node_id);
3387 * Do not even bother to check the largest node if the root
3388 * is empty. Do it lockless to prevent lock bouncing. Races
3389 * are acceptable as soft limit is best effort anyway.
3391 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3395 * This loop can run a while, specially if mem_cgroup's continuously
3396 * keep exceeding their soft limit and putting the system under
3403 mz = mem_cgroup_largest_soft_limit_node(mctz);
3408 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3409 gfp_mask, &nr_scanned);
3410 nr_reclaimed += reclaimed;
3411 *total_scanned += nr_scanned;
3412 spin_lock_irq(&mctz->lock);
3413 __mem_cgroup_remove_exceeded(mz, mctz);
3416 * If we failed to reclaim anything from this memory cgroup
3417 * it is time to move on to the next cgroup
3421 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3423 excess = soft_limit_excess(mz->memcg);
3425 * One school of thought says that we should not add
3426 * back the node to the tree if reclaim returns 0.
3427 * But our reclaim could return 0, simply because due
3428 * to priority we are exposing a smaller subset of
3429 * memory to reclaim from. Consider this as a longer
3432 /* If excess == 0, no tree ops */
3433 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3434 spin_unlock_irq(&mctz->lock);
3435 css_put(&mz->memcg->css);
3438 * Could not reclaim anything and there are no more
3439 * mem cgroups to try or we seem to be looping without
3440 * reclaiming anything.
3442 if (!nr_reclaimed &&
3444 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3446 } while (!nr_reclaimed);
3448 css_put(&next_mz->memcg->css);
3449 return nr_reclaimed;
3453 * Reclaims as many pages from the given memcg as possible.
3455 * Caller is responsible for holding css reference for memcg.
3457 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3459 int nr_retries = MAX_RECLAIM_RETRIES;
3461 /* we call try-to-free pages for make this cgroup empty */
3462 lru_add_drain_all();
3464 drain_all_stock(memcg);
3466 /* try to free all pages in this cgroup */
3467 while (nr_retries && page_counter_read(&memcg->memory)) {
3470 if (signal_pending(current))
3473 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3477 /* maybe some writeback is necessary */
3478 congestion_wait(BLK_RW_ASYNC, HZ/10);
3486 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3487 char *buf, size_t nbytes,
3490 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3492 if (mem_cgroup_is_root(memcg))
3494 return mem_cgroup_force_empty(memcg) ?: nbytes;
3497 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3503 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3504 struct cftype *cft, u64 val)
3509 pr_warn_once("Non-hierarchical mode is deprecated. "
3510 "Please report your usecase to linux-mm@kvack.org if you "
3511 "depend on this functionality.\n");
3516 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3520 if (mem_cgroup_is_root(memcg)) {
3521 cgroup_rstat_flush(memcg->css.cgroup);
3522 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3523 memcg_page_state(memcg, NR_ANON_MAPPED);
3525 val += memcg_page_state(memcg, MEMCG_SWAP);
3528 val = page_counter_read(&memcg->memory);
3530 val = page_counter_read(&memcg->memsw);
3543 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3546 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3547 struct page_counter *counter;
3549 switch (MEMFILE_TYPE(cft->private)) {
3551 counter = &memcg->memory;
3554 counter = &memcg->memsw;
3557 counter = &memcg->kmem;
3560 counter = &memcg->tcpmem;
3566 switch (MEMFILE_ATTR(cft->private)) {
3568 if (counter == &memcg->memory)
3569 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3570 if (counter == &memcg->memsw)
3571 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3572 return (u64)page_counter_read(counter) * PAGE_SIZE;
3574 return (u64)counter->max * PAGE_SIZE;
3576 return (u64)counter->watermark * PAGE_SIZE;
3578 return counter->failcnt;
3579 case RES_SOFT_LIMIT:
3580 return (u64)memcg->soft_limit * PAGE_SIZE;
3586 #ifdef CONFIG_MEMCG_KMEM
3587 static int memcg_online_kmem(struct mem_cgroup *memcg)
3589 struct obj_cgroup *objcg;
3592 if (cgroup_memory_nokmem)
3595 BUG_ON(memcg->kmemcg_id >= 0);
3596 BUG_ON(memcg->kmem_state);
3598 memcg_id = memcg_alloc_cache_id();
3602 objcg = obj_cgroup_alloc();
3604 memcg_free_cache_id(memcg_id);
3607 objcg->memcg = memcg;
3608 rcu_assign_pointer(memcg->objcg, objcg);
3610 static_branch_enable(&memcg_kmem_enabled_key);
3612 memcg->kmemcg_id = memcg_id;
3613 memcg->kmem_state = KMEM_ONLINE;
3618 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3620 struct cgroup_subsys_state *css;
3621 struct mem_cgroup *parent, *child;
3624 if (memcg->kmem_state != KMEM_ONLINE)
3627 memcg->kmem_state = KMEM_ALLOCATED;
3629 parent = parent_mem_cgroup(memcg);
3631 parent = root_mem_cgroup;
3633 memcg_reparent_objcgs(memcg, parent);
3635 kmemcg_id = memcg->kmemcg_id;
3636 BUG_ON(kmemcg_id < 0);
3639 * Change kmemcg_id of this cgroup and all its descendants to the
3640 * parent's id, and then move all entries from this cgroup's list_lrus
3641 * to ones of the parent. After we have finished, all list_lrus
3642 * corresponding to this cgroup are guaranteed to remain empty. The
3643 * ordering is imposed by list_lru_node->lock taken by
3644 * memcg_drain_all_list_lrus().
3646 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3647 css_for_each_descendant_pre(css, &memcg->css) {
3648 child = mem_cgroup_from_css(css);
3649 BUG_ON(child->kmemcg_id != kmemcg_id);
3650 child->kmemcg_id = parent->kmemcg_id;
3654 memcg_drain_all_list_lrus(kmemcg_id, parent);
3656 memcg_free_cache_id(kmemcg_id);
3659 static void memcg_free_kmem(struct mem_cgroup *memcg)
3661 /* css_alloc() failed, offlining didn't happen */
3662 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3663 memcg_offline_kmem(memcg);
3666 static int memcg_online_kmem(struct mem_cgroup *memcg)
3670 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3673 static void memcg_free_kmem(struct mem_cgroup *memcg)
3676 #endif /* CONFIG_MEMCG_KMEM */
3678 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3683 mutex_lock(&memcg_max_mutex);
3684 ret = page_counter_set_max(&memcg->kmem, max);
3685 mutex_unlock(&memcg_max_mutex);
3689 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3693 mutex_lock(&memcg_max_mutex);
3695 ret = page_counter_set_max(&memcg->tcpmem, max);
3699 if (!memcg->tcpmem_active) {
3701 * The active flag needs to be written after the static_key
3702 * update. This is what guarantees that the socket activation
3703 * function is the last one to run. See mem_cgroup_sk_alloc()
3704 * for details, and note that we don't mark any socket as
3705 * belonging to this memcg until that flag is up.
3707 * We need to do this, because static_keys will span multiple
3708 * sites, but we can't control their order. If we mark a socket
3709 * as accounted, but the accounting functions are not patched in
3710 * yet, we'll lose accounting.
3712 * We never race with the readers in mem_cgroup_sk_alloc(),
3713 * because when this value change, the code to process it is not
3716 static_branch_inc(&memcg_sockets_enabled_key);
3717 memcg->tcpmem_active = true;
3720 mutex_unlock(&memcg_max_mutex);
3725 * The user of this function is...
3728 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3729 char *buf, size_t nbytes, loff_t off)
3731 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3732 unsigned long nr_pages;
3735 buf = strstrip(buf);
3736 ret = page_counter_memparse(buf, "-1", &nr_pages);
3740 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3742 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3746 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3748 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3751 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3754 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3755 "Please report your usecase to linux-mm@kvack.org if you "
3756 "depend on this functionality.\n");
3757 ret = memcg_update_kmem_max(memcg, nr_pages);
3760 ret = memcg_update_tcp_max(memcg, nr_pages);
3764 case RES_SOFT_LIMIT:
3765 memcg->soft_limit = nr_pages;
3769 return ret ?: nbytes;
3772 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3773 size_t nbytes, loff_t off)
3775 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3776 struct page_counter *counter;
3778 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3780 counter = &memcg->memory;
3783 counter = &memcg->memsw;
3786 counter = &memcg->kmem;
3789 counter = &memcg->tcpmem;
3795 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3797 page_counter_reset_watermark(counter);
3800 counter->failcnt = 0;
3809 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3812 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3816 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3817 struct cftype *cft, u64 val)
3819 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3821 if (val & ~MOVE_MASK)
3825 * No kind of locking is needed in here, because ->can_attach() will
3826 * check this value once in the beginning of the process, and then carry
3827 * on with stale data. This means that changes to this value will only
3828 * affect task migrations starting after the change.
3830 memcg->move_charge_at_immigrate = val;
3834 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3835 struct cftype *cft, u64 val)
3843 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3844 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3845 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3847 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3848 int nid, unsigned int lru_mask, bool tree)
3850 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3851 unsigned long nr = 0;
3854 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3857 if (!(BIT(lru) & lru_mask))
3860 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3862 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3867 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3868 unsigned int lru_mask,
3871 unsigned long nr = 0;
3875 if (!(BIT(lru) & lru_mask))
3878 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3880 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3885 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3889 unsigned int lru_mask;
3892 static const struct numa_stat stats[] = {
3893 { "total", LRU_ALL },
3894 { "file", LRU_ALL_FILE },
3895 { "anon", LRU_ALL_ANON },
3896 { "unevictable", BIT(LRU_UNEVICTABLE) },
3898 const struct numa_stat *stat;
3900 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3902 cgroup_rstat_flush(memcg->css.cgroup);
3904 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3905 seq_printf(m, "%s=%lu", stat->name,
3906 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3908 for_each_node_state(nid, N_MEMORY)
3909 seq_printf(m, " N%d=%lu", nid,
3910 mem_cgroup_node_nr_lru_pages(memcg, nid,
3911 stat->lru_mask, false));
3915 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3917 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3918 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3920 for_each_node_state(nid, N_MEMORY)
3921 seq_printf(m, " N%d=%lu", nid,
3922 mem_cgroup_node_nr_lru_pages(memcg, nid,
3923 stat->lru_mask, true));
3929 #endif /* CONFIG_NUMA */
3931 static const unsigned int memcg1_stats[] = {
3934 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3944 static const char *const memcg1_stat_names[] = {
3947 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3957 /* Universal VM events cgroup1 shows, original sort order */
3958 static const unsigned int memcg1_events[] = {
3965 static int memcg_stat_show(struct seq_file *m, void *v)
3967 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3968 unsigned long memory, memsw;
3969 struct mem_cgroup *mi;
3972 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3974 cgroup_rstat_flush(memcg->css.cgroup);
3976 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3979 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3981 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3982 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3985 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3986 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3987 memcg_events_local(memcg, memcg1_events[i]));
3989 for (i = 0; i < NR_LRU_LISTS; i++)
3990 seq_printf(m, "%s %lu\n", lru_list_name(i),
3991 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3994 /* Hierarchical information */
3995 memory = memsw = PAGE_COUNTER_MAX;
3996 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3997 memory = min(memory, READ_ONCE(mi->memory.max));
3998 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4000 seq_printf(m, "hierarchical_memory_limit %llu\n",
4001 (u64)memory * PAGE_SIZE);
4002 if (do_memsw_account())
4003 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4004 (u64)memsw * PAGE_SIZE);
4006 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4009 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4011 nr = memcg_page_state(memcg, memcg1_stats[i]);
4012 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4013 (u64)nr * PAGE_SIZE);
4016 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4017 seq_printf(m, "total_%s %llu\n",
4018 vm_event_name(memcg1_events[i]),
4019 (u64)memcg_events(memcg, memcg1_events[i]));
4021 for (i = 0; i < NR_LRU_LISTS; i++)
4022 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4023 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4026 #ifdef CONFIG_DEBUG_VM
4029 struct mem_cgroup_per_node *mz;
4030 unsigned long anon_cost = 0;
4031 unsigned long file_cost = 0;
4033 for_each_online_pgdat(pgdat) {
4034 mz = memcg->nodeinfo[pgdat->node_id];
4036 anon_cost += mz->lruvec.anon_cost;
4037 file_cost += mz->lruvec.file_cost;
4039 seq_printf(m, "anon_cost %lu\n", anon_cost);
4040 seq_printf(m, "file_cost %lu\n", file_cost);
4047 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4050 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4052 return mem_cgroup_swappiness(memcg);
4055 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4056 struct cftype *cft, u64 val)
4058 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4063 if (!mem_cgroup_is_root(memcg))
4064 memcg->swappiness = val;
4066 vm_swappiness = val;
4071 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4073 struct mem_cgroup_threshold_ary *t;
4074 unsigned long usage;
4079 t = rcu_dereference(memcg->thresholds.primary);
4081 t = rcu_dereference(memcg->memsw_thresholds.primary);
4086 usage = mem_cgroup_usage(memcg, swap);
4089 * current_threshold points to threshold just below or equal to usage.
4090 * If it's not true, a threshold was crossed after last
4091 * call of __mem_cgroup_threshold().
4093 i = t->current_threshold;
4096 * Iterate backward over array of thresholds starting from
4097 * current_threshold and check if a threshold is crossed.
4098 * If none of thresholds below usage is crossed, we read
4099 * only one element of the array here.
4101 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4102 eventfd_signal(t->entries[i].eventfd, 1);
4104 /* i = current_threshold + 1 */
4108 * Iterate forward over array of thresholds starting from
4109 * current_threshold+1 and check if a threshold is crossed.
4110 * If none of thresholds above usage is crossed, we read
4111 * only one element of the array here.
4113 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4114 eventfd_signal(t->entries[i].eventfd, 1);
4116 /* Update current_threshold */
4117 t->current_threshold = i - 1;
4122 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4125 __mem_cgroup_threshold(memcg, false);
4126 if (do_memsw_account())
4127 __mem_cgroup_threshold(memcg, true);
4129 memcg = parent_mem_cgroup(memcg);
4133 static int compare_thresholds(const void *a, const void *b)
4135 const struct mem_cgroup_threshold *_a = a;
4136 const struct mem_cgroup_threshold *_b = b;
4138 if (_a->threshold > _b->threshold)
4141 if (_a->threshold < _b->threshold)
4147 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4149 struct mem_cgroup_eventfd_list *ev;
4151 spin_lock(&memcg_oom_lock);
4153 list_for_each_entry(ev, &memcg->oom_notify, list)
4154 eventfd_signal(ev->eventfd, 1);
4156 spin_unlock(&memcg_oom_lock);
4160 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4162 struct mem_cgroup *iter;
4164 for_each_mem_cgroup_tree(iter, memcg)
4165 mem_cgroup_oom_notify_cb(iter);
4168 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4169 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4171 struct mem_cgroup_thresholds *thresholds;
4172 struct mem_cgroup_threshold_ary *new;
4173 unsigned long threshold;
4174 unsigned long usage;
4177 ret = page_counter_memparse(args, "-1", &threshold);
4181 mutex_lock(&memcg->thresholds_lock);
4184 thresholds = &memcg->thresholds;
4185 usage = mem_cgroup_usage(memcg, false);
4186 } else if (type == _MEMSWAP) {
4187 thresholds = &memcg->memsw_thresholds;
4188 usage = mem_cgroup_usage(memcg, true);
4192 /* Check if a threshold crossed before adding a new one */
4193 if (thresholds->primary)
4194 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4196 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4198 /* Allocate memory for new array of thresholds */
4199 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4206 /* Copy thresholds (if any) to new array */
4207 if (thresholds->primary)
4208 memcpy(new->entries, thresholds->primary->entries,
4209 flex_array_size(new, entries, size - 1));
4211 /* Add new threshold */
4212 new->entries[size - 1].eventfd = eventfd;
4213 new->entries[size - 1].threshold = threshold;
4215 /* Sort thresholds. Registering of new threshold isn't time-critical */
4216 sort(new->entries, size, sizeof(*new->entries),
4217 compare_thresholds, NULL);
4219 /* Find current threshold */
4220 new->current_threshold = -1;
4221 for (i = 0; i < size; i++) {
4222 if (new->entries[i].threshold <= usage) {
4224 * new->current_threshold will not be used until
4225 * rcu_assign_pointer(), so it's safe to increment
4228 ++new->current_threshold;
4233 /* Free old spare buffer and save old primary buffer as spare */
4234 kfree(thresholds->spare);
4235 thresholds->spare = thresholds->primary;
4237 rcu_assign_pointer(thresholds->primary, new);
4239 /* To be sure that nobody uses thresholds */
4243 mutex_unlock(&memcg->thresholds_lock);
4248 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4249 struct eventfd_ctx *eventfd, const char *args)
4251 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4254 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4255 struct eventfd_ctx *eventfd, const char *args)
4257 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4260 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4261 struct eventfd_ctx *eventfd, enum res_type type)
4263 struct mem_cgroup_thresholds *thresholds;
4264 struct mem_cgroup_threshold_ary *new;
4265 unsigned long usage;
4266 int i, j, size, entries;
4268 mutex_lock(&memcg->thresholds_lock);
4271 thresholds = &memcg->thresholds;
4272 usage = mem_cgroup_usage(memcg, false);
4273 } else if (type == _MEMSWAP) {
4274 thresholds = &memcg->memsw_thresholds;
4275 usage = mem_cgroup_usage(memcg, true);
4279 if (!thresholds->primary)
4282 /* Check if a threshold crossed before removing */
4283 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4285 /* Calculate new number of threshold */
4287 for (i = 0; i < thresholds->primary->size; i++) {
4288 if (thresholds->primary->entries[i].eventfd != eventfd)
4294 new = thresholds->spare;
4296 /* If no items related to eventfd have been cleared, nothing to do */
4300 /* Set thresholds array to NULL if we don't have thresholds */
4309 /* Copy thresholds and find current threshold */
4310 new->current_threshold = -1;
4311 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4312 if (thresholds->primary->entries[i].eventfd == eventfd)
4315 new->entries[j] = thresholds->primary->entries[i];
4316 if (new->entries[j].threshold <= usage) {
4318 * new->current_threshold will not be used
4319 * until rcu_assign_pointer(), so it's safe to increment
4322 ++new->current_threshold;
4328 /* Swap primary and spare array */
4329 thresholds->spare = thresholds->primary;
4331 rcu_assign_pointer(thresholds->primary, new);
4333 /* To be sure that nobody uses thresholds */
4336 /* If all events are unregistered, free the spare array */
4338 kfree(thresholds->spare);
4339 thresholds->spare = NULL;
4342 mutex_unlock(&memcg->thresholds_lock);
4345 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4346 struct eventfd_ctx *eventfd)
4348 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4351 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4352 struct eventfd_ctx *eventfd)
4354 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4357 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4358 struct eventfd_ctx *eventfd, const char *args)
4360 struct mem_cgroup_eventfd_list *event;
4362 event = kmalloc(sizeof(*event), GFP_KERNEL);
4366 spin_lock(&memcg_oom_lock);
4368 event->eventfd = eventfd;
4369 list_add(&event->list, &memcg->oom_notify);
4371 /* already in OOM ? */
4372 if (memcg->under_oom)
4373 eventfd_signal(eventfd, 1);
4374 spin_unlock(&memcg_oom_lock);
4379 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4380 struct eventfd_ctx *eventfd)
4382 struct mem_cgroup_eventfd_list *ev, *tmp;
4384 spin_lock(&memcg_oom_lock);
4386 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4387 if (ev->eventfd == eventfd) {
4388 list_del(&ev->list);
4393 spin_unlock(&memcg_oom_lock);
4396 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4398 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4400 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4401 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4402 seq_printf(sf, "oom_kill %lu\n",
4403 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4407 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4408 struct cftype *cft, u64 val)
4410 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4412 /* cannot set to root cgroup and only 0 and 1 are allowed */
4413 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4416 memcg->oom_kill_disable = val;
4418 memcg_oom_recover(memcg);
4423 #ifdef CONFIG_CGROUP_WRITEBACK
4425 #include <trace/events/writeback.h>
4427 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4429 return wb_domain_init(&memcg->cgwb_domain, gfp);
4432 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4434 wb_domain_exit(&memcg->cgwb_domain);
4437 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4439 wb_domain_size_changed(&memcg->cgwb_domain);
4442 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4444 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4446 if (!memcg->css.parent)
4449 return &memcg->cgwb_domain;
4453 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4454 * @wb: bdi_writeback in question
4455 * @pfilepages: out parameter for number of file pages
4456 * @pheadroom: out parameter for number of allocatable pages according to memcg
4457 * @pdirty: out parameter for number of dirty pages
4458 * @pwriteback: out parameter for number of pages under writeback
4460 * Determine the numbers of file, headroom, dirty, and writeback pages in
4461 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4462 * is a bit more involved.
4464 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4465 * headroom is calculated as the lowest headroom of itself and the
4466 * ancestors. Note that this doesn't consider the actual amount of
4467 * available memory in the system. The caller should further cap
4468 * *@pheadroom accordingly.
4470 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4471 unsigned long *pheadroom, unsigned long *pdirty,
4472 unsigned long *pwriteback)
4474 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4475 struct mem_cgroup *parent;
4477 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4479 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4480 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4481 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4482 memcg_page_state(memcg, NR_ACTIVE_FILE);
4484 *pheadroom = PAGE_COUNTER_MAX;
4485 while ((parent = parent_mem_cgroup(memcg))) {
4486 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4487 READ_ONCE(memcg->memory.high));
4488 unsigned long used = page_counter_read(&memcg->memory);
4490 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4496 * Foreign dirty flushing
4498 * There's an inherent mismatch between memcg and writeback. The former
4499 * trackes ownership per-page while the latter per-inode. This was a
4500 * deliberate design decision because honoring per-page ownership in the
4501 * writeback path is complicated, may lead to higher CPU and IO overheads
4502 * and deemed unnecessary given that write-sharing an inode across
4503 * different cgroups isn't a common use-case.
4505 * Combined with inode majority-writer ownership switching, this works well
4506 * enough in most cases but there are some pathological cases. For
4507 * example, let's say there are two cgroups A and B which keep writing to
4508 * different but confined parts of the same inode. B owns the inode and
4509 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4510 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4511 * triggering background writeback. A will be slowed down without a way to
4512 * make writeback of the dirty pages happen.
4514 * Conditions like the above can lead to a cgroup getting repatedly and
4515 * severely throttled after making some progress after each
4516 * dirty_expire_interval while the underyling IO device is almost
4519 * Solving this problem completely requires matching the ownership tracking
4520 * granularities between memcg and writeback in either direction. However,
4521 * the more egregious behaviors can be avoided by simply remembering the
4522 * most recent foreign dirtying events and initiating remote flushes on
4523 * them when local writeback isn't enough to keep the memory clean enough.
4525 * The following two functions implement such mechanism. When a foreign
4526 * page - a page whose memcg and writeback ownerships don't match - is
4527 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4528 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4529 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4530 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4531 * foreign bdi_writebacks which haven't expired. Both the numbers of
4532 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4533 * limited to MEMCG_CGWB_FRN_CNT.
4535 * The mechanism only remembers IDs and doesn't hold any object references.
4536 * As being wrong occasionally doesn't matter, updates and accesses to the
4537 * records are lockless and racy.
4539 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4540 struct bdi_writeback *wb)
4542 struct mem_cgroup *memcg = page_memcg(page);
4543 struct memcg_cgwb_frn *frn;
4544 u64 now = get_jiffies_64();
4545 u64 oldest_at = now;
4549 trace_track_foreign_dirty(page, wb);
4552 * Pick the slot to use. If there is already a slot for @wb, keep
4553 * using it. If not replace the oldest one which isn't being
4556 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4557 frn = &memcg->cgwb_frn[i];
4558 if (frn->bdi_id == wb->bdi->id &&
4559 frn->memcg_id == wb->memcg_css->id)
4561 if (time_before64(frn->at, oldest_at) &&
4562 atomic_read(&frn->done.cnt) == 1) {
4564 oldest_at = frn->at;
4568 if (i < MEMCG_CGWB_FRN_CNT) {
4570 * Re-using an existing one. Update timestamp lazily to
4571 * avoid making the cacheline hot. We want them to be
4572 * reasonably up-to-date and significantly shorter than
4573 * dirty_expire_interval as that's what expires the record.
4574 * Use the shorter of 1s and dirty_expire_interval / 8.
4576 unsigned long update_intv =
4577 min_t(unsigned long, HZ,
4578 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4580 if (time_before64(frn->at, now - update_intv))
4582 } else if (oldest >= 0) {
4583 /* replace the oldest free one */
4584 frn = &memcg->cgwb_frn[oldest];
4585 frn->bdi_id = wb->bdi->id;
4586 frn->memcg_id = wb->memcg_css->id;
4591 /* issue foreign writeback flushes for recorded foreign dirtying events */
4592 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4594 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4595 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4596 u64 now = jiffies_64;
4599 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4600 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4603 * If the record is older than dirty_expire_interval,
4604 * writeback on it has already started. No need to kick it
4605 * off again. Also, don't start a new one if there's
4606 * already one in flight.
4608 if (time_after64(frn->at, now - intv) &&
4609 atomic_read(&frn->done.cnt) == 1) {
4611 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4612 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4613 WB_REASON_FOREIGN_FLUSH,
4619 #else /* CONFIG_CGROUP_WRITEBACK */
4621 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4626 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4630 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4634 #endif /* CONFIG_CGROUP_WRITEBACK */
4637 * DO NOT USE IN NEW FILES.
4639 * "cgroup.event_control" implementation.
4641 * This is way over-engineered. It tries to support fully configurable
4642 * events for each user. Such level of flexibility is completely
4643 * unnecessary especially in the light of the planned unified hierarchy.
4645 * Please deprecate this and replace with something simpler if at all
4650 * Unregister event and free resources.
4652 * Gets called from workqueue.
4654 static void memcg_event_remove(struct work_struct *work)
4656 struct mem_cgroup_event *event =
4657 container_of(work, struct mem_cgroup_event, remove);
4658 struct mem_cgroup *memcg = event->memcg;
4660 remove_wait_queue(event->wqh, &event->wait);
4662 event->unregister_event(memcg, event->eventfd);
4664 /* Notify userspace the event is going away. */
4665 eventfd_signal(event->eventfd, 1);
4667 eventfd_ctx_put(event->eventfd);
4669 css_put(&memcg->css);
4673 * Gets called on EPOLLHUP on eventfd when user closes it.
4675 * Called with wqh->lock held and interrupts disabled.
4677 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4678 int sync, void *key)
4680 struct mem_cgroup_event *event =
4681 container_of(wait, struct mem_cgroup_event, wait);
4682 struct mem_cgroup *memcg = event->memcg;
4683 __poll_t flags = key_to_poll(key);
4685 if (flags & EPOLLHUP) {
4687 * If the event has been detached at cgroup removal, we
4688 * can simply return knowing the other side will cleanup
4691 * We can't race against event freeing since the other
4692 * side will require wqh->lock via remove_wait_queue(),
4695 spin_lock(&memcg->event_list_lock);
4696 if (!list_empty(&event->list)) {
4697 list_del_init(&event->list);
4699 * We are in atomic context, but cgroup_event_remove()
4700 * may sleep, so we have to call it in workqueue.
4702 schedule_work(&event->remove);
4704 spin_unlock(&memcg->event_list_lock);
4710 static void memcg_event_ptable_queue_proc(struct file *file,
4711 wait_queue_head_t *wqh, poll_table *pt)
4713 struct mem_cgroup_event *event =
4714 container_of(pt, struct mem_cgroup_event, pt);
4717 add_wait_queue(wqh, &event->wait);
4721 * DO NOT USE IN NEW FILES.
4723 * Parse input and register new cgroup event handler.
4725 * Input must be in format '<event_fd> <control_fd> <args>'.
4726 * Interpretation of args is defined by control file implementation.
4728 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4729 char *buf, size_t nbytes, loff_t off)
4731 struct cgroup_subsys_state *css = of_css(of);
4732 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4733 struct mem_cgroup_event *event;
4734 struct cgroup_subsys_state *cfile_css;
4735 unsigned int efd, cfd;
4742 buf = strstrip(buf);
4744 efd = simple_strtoul(buf, &endp, 10);
4749 cfd = simple_strtoul(buf, &endp, 10);
4750 if ((*endp != ' ') && (*endp != '\0'))
4754 event = kzalloc(sizeof(*event), GFP_KERNEL);
4758 event->memcg = memcg;
4759 INIT_LIST_HEAD(&event->list);
4760 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4761 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4762 INIT_WORK(&event->remove, memcg_event_remove);
4770 event->eventfd = eventfd_ctx_fileget(efile.file);
4771 if (IS_ERR(event->eventfd)) {
4772 ret = PTR_ERR(event->eventfd);
4779 goto out_put_eventfd;
4782 /* the process need read permission on control file */
4783 /* AV: shouldn't we check that it's been opened for read instead? */
4784 ret = file_permission(cfile.file, MAY_READ);
4789 * Determine the event callbacks and set them in @event. This used
4790 * to be done via struct cftype but cgroup core no longer knows
4791 * about these events. The following is crude but the whole thing
4792 * is for compatibility anyway.
4794 * DO NOT ADD NEW FILES.
4796 name = cfile.file->f_path.dentry->d_name.name;
4798 if (!strcmp(name, "memory.usage_in_bytes")) {
4799 event->register_event = mem_cgroup_usage_register_event;
4800 event->unregister_event = mem_cgroup_usage_unregister_event;
4801 } else if (!strcmp(name, "memory.oom_control")) {
4802 event->register_event = mem_cgroup_oom_register_event;
4803 event->unregister_event = mem_cgroup_oom_unregister_event;
4804 } else if (!strcmp(name, "memory.pressure_level")) {
4805 event->register_event = vmpressure_register_event;
4806 event->unregister_event = vmpressure_unregister_event;
4807 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4808 event->register_event = memsw_cgroup_usage_register_event;
4809 event->unregister_event = memsw_cgroup_usage_unregister_event;
4816 * Verify @cfile should belong to @css. Also, remaining events are
4817 * automatically removed on cgroup destruction but the removal is
4818 * asynchronous, so take an extra ref on @css.
4820 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4821 &memory_cgrp_subsys);
4823 if (IS_ERR(cfile_css))
4825 if (cfile_css != css) {
4830 ret = event->register_event(memcg, event->eventfd, buf);
4834 vfs_poll(efile.file, &event->pt);
4836 spin_lock(&memcg->event_list_lock);
4837 list_add(&event->list, &memcg->event_list);
4838 spin_unlock(&memcg->event_list_lock);
4850 eventfd_ctx_put(event->eventfd);
4859 static struct cftype mem_cgroup_legacy_files[] = {
4861 .name = "usage_in_bytes",
4862 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4863 .read_u64 = mem_cgroup_read_u64,
4866 .name = "max_usage_in_bytes",
4867 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4868 .write = mem_cgroup_reset,
4869 .read_u64 = mem_cgroup_read_u64,
4872 .name = "limit_in_bytes",
4873 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4874 .write = mem_cgroup_write,
4875 .read_u64 = mem_cgroup_read_u64,
4878 .name = "soft_limit_in_bytes",
4879 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4880 .write = mem_cgroup_write,
4881 .read_u64 = mem_cgroup_read_u64,
4885 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4886 .write = mem_cgroup_reset,
4887 .read_u64 = mem_cgroup_read_u64,
4891 .seq_show = memcg_stat_show,
4894 .name = "force_empty",
4895 .write = mem_cgroup_force_empty_write,
4898 .name = "use_hierarchy",
4899 .write_u64 = mem_cgroup_hierarchy_write,
4900 .read_u64 = mem_cgroup_hierarchy_read,
4903 .name = "cgroup.event_control", /* XXX: for compat */
4904 .write = memcg_write_event_control,
4905 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4908 .name = "swappiness",
4909 .read_u64 = mem_cgroup_swappiness_read,
4910 .write_u64 = mem_cgroup_swappiness_write,
4913 .name = "move_charge_at_immigrate",
4914 .read_u64 = mem_cgroup_move_charge_read,
4915 .write_u64 = mem_cgroup_move_charge_write,
4918 .name = "oom_control",
4919 .seq_show = mem_cgroup_oom_control_read,
4920 .write_u64 = mem_cgroup_oom_control_write,
4921 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4924 .name = "pressure_level",
4928 .name = "numa_stat",
4929 .seq_show = memcg_numa_stat_show,
4933 .name = "kmem.limit_in_bytes",
4934 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4935 .write = mem_cgroup_write,
4936 .read_u64 = mem_cgroup_read_u64,
4939 .name = "kmem.usage_in_bytes",
4940 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4941 .read_u64 = mem_cgroup_read_u64,
4944 .name = "kmem.failcnt",
4945 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4946 .write = mem_cgroup_reset,
4947 .read_u64 = mem_cgroup_read_u64,
4950 .name = "kmem.max_usage_in_bytes",
4951 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4952 .write = mem_cgroup_reset,
4953 .read_u64 = mem_cgroup_read_u64,
4955 #if defined(CONFIG_MEMCG_KMEM) && \
4956 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4958 .name = "kmem.slabinfo",
4959 .seq_show = memcg_slab_show,
4963 .name = "kmem.tcp.limit_in_bytes",
4964 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4965 .write = mem_cgroup_write,
4966 .read_u64 = mem_cgroup_read_u64,
4969 .name = "kmem.tcp.usage_in_bytes",
4970 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4971 .read_u64 = mem_cgroup_read_u64,
4974 .name = "kmem.tcp.failcnt",
4975 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4976 .write = mem_cgroup_reset,
4977 .read_u64 = mem_cgroup_read_u64,
4980 .name = "kmem.tcp.max_usage_in_bytes",
4981 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4982 .write = mem_cgroup_reset,
4983 .read_u64 = mem_cgroup_read_u64,
4985 { }, /* terminate */
4989 * Private memory cgroup IDR
4991 * Swap-out records and page cache shadow entries need to store memcg
4992 * references in constrained space, so we maintain an ID space that is
4993 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4994 * memory-controlled cgroups to 64k.
4996 * However, there usually are many references to the offline CSS after
4997 * the cgroup has been destroyed, such as page cache or reclaimable
4998 * slab objects, that don't need to hang on to the ID. We want to keep
4999 * those dead CSS from occupying IDs, or we might quickly exhaust the
5000 * relatively small ID space and prevent the creation of new cgroups
5001 * even when there are much fewer than 64k cgroups - possibly none.
5003 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5004 * be freed and recycled when it's no longer needed, which is usually
5005 * when the CSS is offlined.
5007 * The only exception to that are records of swapped out tmpfs/shmem
5008 * pages that need to be attributed to live ancestors on swapin. But
5009 * those references are manageable from userspace.
5012 static DEFINE_IDR(mem_cgroup_idr);
5014 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5016 if (memcg->id.id > 0) {
5017 idr_remove(&mem_cgroup_idr, memcg->id.id);
5022 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5025 refcount_add(n, &memcg->id.ref);
5028 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5030 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5031 mem_cgroup_id_remove(memcg);
5033 /* Memcg ID pins CSS */
5034 css_put(&memcg->css);
5038 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5040 mem_cgroup_id_put_many(memcg, 1);
5044 * mem_cgroup_from_id - look up a memcg from a memcg id
5045 * @id: the memcg id to look up
5047 * Caller must hold rcu_read_lock().
5049 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5051 WARN_ON_ONCE(!rcu_read_lock_held());
5052 return idr_find(&mem_cgroup_idr, id);
5055 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5057 struct mem_cgroup_per_node *pn;
5060 * This routine is called against possible nodes.
5061 * But it's BUG to call kmalloc() against offline node.
5063 * TODO: this routine can waste much memory for nodes which will
5064 * never be onlined. It's better to use memory hotplug callback
5067 if (!node_state(node, N_NORMAL_MEMORY))
5069 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5073 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5074 GFP_KERNEL_ACCOUNT);
5075 if (!pn->lruvec_stat_local) {
5080 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5081 GFP_KERNEL_ACCOUNT);
5082 if (!pn->lruvec_stat_cpu) {
5083 free_percpu(pn->lruvec_stat_local);
5088 lruvec_init(&pn->lruvec);
5089 pn->usage_in_excess = 0;
5090 pn->on_tree = false;
5093 memcg->nodeinfo[node] = pn;
5097 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5099 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5104 free_percpu(pn->lruvec_stat_cpu);
5105 free_percpu(pn->lruvec_stat_local);
5109 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5114 free_mem_cgroup_per_node_info(memcg, node);
5115 free_percpu(memcg->vmstats_percpu);
5119 static void mem_cgroup_free(struct mem_cgroup *memcg)
5123 memcg_wb_domain_exit(memcg);
5125 * Flush percpu lruvec stats to guarantee the value
5126 * correctness on parent's and all ancestor levels.
5128 for_each_online_cpu(cpu)
5129 memcg_flush_lruvec_page_state(memcg, cpu);
5130 __mem_cgroup_free(memcg);
5133 static struct mem_cgroup *mem_cgroup_alloc(void)
5135 struct mem_cgroup *memcg;
5138 int __maybe_unused i;
5139 long error = -ENOMEM;
5141 size = sizeof(struct mem_cgroup);
5142 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5144 memcg = kzalloc(size, GFP_KERNEL);
5146 return ERR_PTR(error);
5148 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5149 1, MEM_CGROUP_ID_MAX,
5151 if (memcg->id.id < 0) {
5152 error = memcg->id.id;
5156 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5157 GFP_KERNEL_ACCOUNT);
5158 if (!memcg->vmstats_percpu)
5162 if (alloc_mem_cgroup_per_node_info(memcg, node))
5165 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5168 INIT_WORK(&memcg->high_work, high_work_func);
5169 INIT_LIST_HEAD(&memcg->oom_notify);
5170 mutex_init(&memcg->thresholds_lock);
5171 spin_lock_init(&memcg->move_lock);
5172 vmpressure_init(&memcg->vmpressure);
5173 INIT_LIST_HEAD(&memcg->event_list);
5174 spin_lock_init(&memcg->event_list_lock);
5175 memcg->socket_pressure = jiffies;
5176 #ifdef CONFIG_MEMCG_KMEM
5177 memcg->kmemcg_id = -1;
5178 INIT_LIST_HEAD(&memcg->objcg_list);
5180 #ifdef CONFIG_CGROUP_WRITEBACK
5181 INIT_LIST_HEAD(&memcg->cgwb_list);
5182 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5183 memcg->cgwb_frn[i].done =
5184 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5186 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5187 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5188 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5189 memcg->deferred_split_queue.split_queue_len = 0;
5191 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5194 mem_cgroup_id_remove(memcg);
5195 __mem_cgroup_free(memcg);
5196 return ERR_PTR(error);
5199 static struct cgroup_subsys_state * __ref
5200 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5202 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5203 struct mem_cgroup *memcg, *old_memcg;
5204 long error = -ENOMEM;
5206 old_memcg = set_active_memcg(parent);
5207 memcg = mem_cgroup_alloc();
5208 set_active_memcg(old_memcg);
5210 return ERR_CAST(memcg);
5212 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5213 memcg->soft_limit = PAGE_COUNTER_MAX;
5214 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5216 memcg->swappiness = mem_cgroup_swappiness(parent);
5217 memcg->oom_kill_disable = parent->oom_kill_disable;
5219 page_counter_init(&memcg->memory, &parent->memory);
5220 page_counter_init(&memcg->swap, &parent->swap);
5221 page_counter_init(&memcg->kmem, &parent->kmem);
5222 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5224 page_counter_init(&memcg->memory, NULL);
5225 page_counter_init(&memcg->swap, NULL);
5226 page_counter_init(&memcg->kmem, NULL);
5227 page_counter_init(&memcg->tcpmem, NULL);
5229 root_mem_cgroup = memcg;
5233 /* The following stuff does not apply to the root */
5234 error = memcg_online_kmem(memcg);
5238 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5239 static_branch_inc(&memcg_sockets_enabled_key);
5243 mem_cgroup_id_remove(memcg);
5244 mem_cgroup_free(memcg);
5245 return ERR_PTR(error);
5248 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5250 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5253 * A memcg must be visible for memcg_expand_shrinker_maps()
5254 * by the time the maps are allocated. So, we allocate maps
5255 * here, when for_each_mem_cgroup() can't skip it.
5257 if (memcg_alloc_shrinker_maps(memcg)) {
5258 mem_cgroup_id_remove(memcg);
5262 /* Online state pins memcg ID, memcg ID pins CSS */
5263 refcount_set(&memcg->id.ref, 1);
5268 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5270 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5271 struct mem_cgroup_event *event, *tmp;
5274 * Unregister events and notify userspace.
5275 * Notify userspace about cgroup removing only after rmdir of cgroup
5276 * directory to avoid race between userspace and kernelspace.
5278 spin_lock(&memcg->event_list_lock);
5279 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5280 list_del_init(&event->list);
5281 schedule_work(&event->remove);
5283 spin_unlock(&memcg->event_list_lock);
5285 page_counter_set_min(&memcg->memory, 0);
5286 page_counter_set_low(&memcg->memory, 0);
5288 memcg_offline_kmem(memcg);
5289 wb_memcg_offline(memcg);
5291 drain_all_stock(memcg);
5293 mem_cgroup_id_put(memcg);
5296 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5298 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5300 invalidate_reclaim_iterators(memcg);
5303 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5305 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5306 int __maybe_unused i;
5308 #ifdef CONFIG_CGROUP_WRITEBACK
5309 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5310 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5312 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5313 static_branch_dec(&memcg_sockets_enabled_key);
5315 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5316 static_branch_dec(&memcg_sockets_enabled_key);
5318 vmpressure_cleanup(&memcg->vmpressure);
5319 cancel_work_sync(&memcg->high_work);
5320 mem_cgroup_remove_from_trees(memcg);
5321 memcg_free_shrinker_maps(memcg);
5322 memcg_free_kmem(memcg);
5323 mem_cgroup_free(memcg);
5327 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5328 * @css: the target css
5330 * Reset the states of the mem_cgroup associated with @css. This is
5331 * invoked when the userland requests disabling on the default hierarchy
5332 * but the memcg is pinned through dependency. The memcg should stop
5333 * applying policies and should revert to the vanilla state as it may be
5334 * made visible again.
5336 * The current implementation only resets the essential configurations.
5337 * This needs to be expanded to cover all the visible parts.
5339 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5341 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5343 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5344 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5345 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5346 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5347 page_counter_set_min(&memcg->memory, 0);
5348 page_counter_set_low(&memcg->memory, 0);
5349 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5350 memcg->soft_limit = PAGE_COUNTER_MAX;
5351 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5352 memcg_wb_domain_size_changed(memcg);
5355 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5357 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5358 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5359 struct memcg_vmstats_percpu *statc;
5363 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5365 for (i = 0; i < MEMCG_NR_STAT; i++) {
5367 * Collect the aggregated propagation counts of groups
5368 * below us. We're in a per-cpu loop here and this is
5369 * a global counter, so the first cycle will get them.
5371 delta = memcg->vmstats.state_pending[i];
5373 memcg->vmstats.state_pending[i] = 0;
5375 /* Add CPU changes on this level since the last flush */
5376 v = READ_ONCE(statc->state[i]);
5377 if (v != statc->state_prev[i]) {
5378 delta += v - statc->state_prev[i];
5379 statc->state_prev[i] = v;
5385 /* Aggregate counts on this level and propagate upwards */
5386 memcg->vmstats.state[i] += delta;
5388 parent->vmstats.state_pending[i] += delta;
5391 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5392 delta = memcg->vmstats.events_pending[i];
5394 memcg->vmstats.events_pending[i] = 0;
5396 v = READ_ONCE(statc->events[i]);
5397 if (v != statc->events_prev[i]) {
5398 delta += v - statc->events_prev[i];
5399 statc->events_prev[i] = v;
5405 memcg->vmstats.events[i] += delta;
5407 parent->vmstats.events_pending[i] += delta;
5412 /* Handlers for move charge at task migration. */
5413 static int mem_cgroup_do_precharge(unsigned long count)
5417 /* Try a single bulk charge without reclaim first, kswapd may wake */
5418 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5420 mc.precharge += count;
5424 /* Try charges one by one with reclaim, but do not retry */
5426 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5440 enum mc_target_type {
5447 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5448 unsigned long addr, pte_t ptent)
5450 struct page *page = vm_normal_page(vma, addr, ptent);
5452 if (!page || !page_mapped(page))
5454 if (PageAnon(page)) {
5455 if (!(mc.flags & MOVE_ANON))
5458 if (!(mc.flags & MOVE_FILE))
5461 if (!get_page_unless_zero(page))
5467 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5468 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5469 pte_t ptent, swp_entry_t *entry)
5471 struct page *page = NULL;
5472 swp_entry_t ent = pte_to_swp_entry(ptent);
5474 if (!(mc.flags & MOVE_ANON))
5478 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5479 * a device and because they are not accessible by CPU they are store
5480 * as special swap entry in the CPU page table.
5482 if (is_device_private_entry(ent)) {
5483 page = device_private_entry_to_page(ent);
5485 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5486 * a refcount of 1 when free (unlike normal page)
5488 if (!page_ref_add_unless(page, 1, 1))
5493 if (non_swap_entry(ent))
5497 * Because lookup_swap_cache() updates some statistics counter,
5498 * we call find_get_page() with swapper_space directly.
5500 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5501 entry->val = ent.val;
5506 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5507 pte_t ptent, swp_entry_t *entry)
5513 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5514 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5516 if (!vma->vm_file) /* anonymous vma */
5518 if (!(mc.flags & MOVE_FILE))
5521 /* page is moved even if it's not RSS of this task(page-faulted). */
5522 /* shmem/tmpfs may report page out on swap: account for that too. */
5523 return find_get_incore_page(vma->vm_file->f_mapping,
5524 linear_page_index(vma, addr));
5528 * mem_cgroup_move_account - move account of the page
5530 * @compound: charge the page as compound or small page
5531 * @from: mem_cgroup which the page is moved from.
5532 * @to: mem_cgroup which the page is moved to. @from != @to.
5534 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5536 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5539 static int mem_cgroup_move_account(struct page *page,
5541 struct mem_cgroup *from,
5542 struct mem_cgroup *to)
5544 struct lruvec *from_vec, *to_vec;
5545 struct pglist_data *pgdat;
5546 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5549 VM_BUG_ON(from == to);
5550 VM_BUG_ON_PAGE(PageLRU(page), page);
5551 VM_BUG_ON(compound && !PageTransHuge(page));
5554 * Prevent mem_cgroup_migrate() from looking at
5555 * page's memory cgroup of its source page while we change it.
5558 if (!trylock_page(page))
5562 if (page_memcg(page) != from)
5565 pgdat = page_pgdat(page);
5566 from_vec = mem_cgroup_lruvec(from, pgdat);
5567 to_vec = mem_cgroup_lruvec(to, pgdat);
5569 lock_page_memcg(page);
5571 if (PageAnon(page)) {
5572 if (page_mapped(page)) {
5573 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5574 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5575 if (PageTransHuge(page)) {
5576 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5578 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5583 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5584 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5586 if (PageSwapBacked(page)) {
5587 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5588 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5591 if (page_mapped(page)) {
5592 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5593 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5596 if (PageDirty(page)) {
5597 struct address_space *mapping = page_mapping(page);
5599 if (mapping_can_writeback(mapping)) {
5600 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5602 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5608 if (PageWriteback(page)) {
5609 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5610 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5614 * All state has been migrated, let's switch to the new memcg.
5616 * It is safe to change page's memcg here because the page
5617 * is referenced, charged, isolated, and locked: we can't race
5618 * with (un)charging, migration, LRU putback, or anything else
5619 * that would rely on a stable page's memory cgroup.
5621 * Note that lock_page_memcg is a memcg lock, not a page lock,
5622 * to save space. As soon as we switch page's memory cgroup to a
5623 * new memcg that isn't locked, the above state can change
5624 * concurrently again. Make sure we're truly done with it.
5629 css_put(&from->css);
5631 page->memcg_data = (unsigned long)to;
5633 __unlock_page_memcg(from);
5637 local_irq_disable();
5638 mem_cgroup_charge_statistics(to, page, nr_pages);
5639 memcg_check_events(to, page);
5640 mem_cgroup_charge_statistics(from, page, -nr_pages);
5641 memcg_check_events(from, page);
5650 * get_mctgt_type - get target type of moving charge
5651 * @vma: the vma the pte to be checked belongs
5652 * @addr: the address corresponding to the pte to be checked
5653 * @ptent: the pte to be checked
5654 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5657 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5658 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5659 * move charge. if @target is not NULL, the page is stored in target->page
5660 * with extra refcnt got(Callers should handle it).
5661 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5662 * target for charge migration. if @target is not NULL, the entry is stored
5664 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5665 * (so ZONE_DEVICE page and thus not on the lru).
5666 * For now we such page is charge like a regular page would be as for all
5667 * intent and purposes it is just special memory taking the place of a
5670 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5672 * Called with pte lock held.
5675 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5676 unsigned long addr, pte_t ptent, union mc_target *target)
5678 struct page *page = NULL;
5679 enum mc_target_type ret = MC_TARGET_NONE;
5680 swp_entry_t ent = { .val = 0 };
5682 if (pte_present(ptent))
5683 page = mc_handle_present_pte(vma, addr, ptent);
5684 else if (is_swap_pte(ptent))
5685 page = mc_handle_swap_pte(vma, ptent, &ent);
5686 else if (pte_none(ptent))
5687 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5689 if (!page && !ent.val)
5693 * Do only loose check w/o serialization.
5694 * mem_cgroup_move_account() checks the page is valid or
5695 * not under LRU exclusion.
5697 if (page_memcg(page) == mc.from) {
5698 ret = MC_TARGET_PAGE;
5699 if (is_device_private_page(page))
5700 ret = MC_TARGET_DEVICE;
5702 target->page = page;
5704 if (!ret || !target)
5708 * There is a swap entry and a page doesn't exist or isn't charged.
5709 * But we cannot move a tail-page in a THP.
5711 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5712 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5713 ret = MC_TARGET_SWAP;
5720 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5722 * We don't consider PMD mapped swapping or file mapped pages because THP does
5723 * not support them for now.
5724 * Caller should make sure that pmd_trans_huge(pmd) is true.
5726 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5727 unsigned long addr, pmd_t pmd, union mc_target *target)
5729 struct page *page = NULL;
5730 enum mc_target_type ret = MC_TARGET_NONE;
5732 if (unlikely(is_swap_pmd(pmd))) {
5733 VM_BUG_ON(thp_migration_supported() &&
5734 !is_pmd_migration_entry(pmd));
5737 page = pmd_page(pmd);
5738 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5739 if (!(mc.flags & MOVE_ANON))
5741 if (page_memcg(page) == mc.from) {
5742 ret = MC_TARGET_PAGE;
5745 target->page = page;
5751 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5752 unsigned long addr, pmd_t pmd, union mc_target *target)
5754 return MC_TARGET_NONE;
5758 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5759 unsigned long addr, unsigned long end,
5760 struct mm_walk *walk)
5762 struct vm_area_struct *vma = walk->vma;
5766 ptl = pmd_trans_huge_lock(pmd, vma);
5769 * Note their can not be MC_TARGET_DEVICE for now as we do not
5770 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5771 * this might change.
5773 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5774 mc.precharge += HPAGE_PMD_NR;
5779 if (pmd_trans_unstable(pmd))
5781 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5782 for (; addr != end; pte++, addr += PAGE_SIZE)
5783 if (get_mctgt_type(vma, addr, *pte, NULL))
5784 mc.precharge++; /* increment precharge temporarily */
5785 pte_unmap_unlock(pte - 1, ptl);
5791 static const struct mm_walk_ops precharge_walk_ops = {
5792 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5795 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5797 unsigned long precharge;
5800 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5801 mmap_read_unlock(mm);
5803 precharge = mc.precharge;
5809 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5811 unsigned long precharge = mem_cgroup_count_precharge(mm);
5813 VM_BUG_ON(mc.moving_task);
5814 mc.moving_task = current;
5815 return mem_cgroup_do_precharge(precharge);
5818 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5819 static void __mem_cgroup_clear_mc(void)
5821 struct mem_cgroup *from = mc.from;
5822 struct mem_cgroup *to = mc.to;
5824 /* we must uncharge all the leftover precharges from mc.to */
5826 cancel_charge(mc.to, mc.precharge);
5830 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5831 * we must uncharge here.
5833 if (mc.moved_charge) {
5834 cancel_charge(mc.from, mc.moved_charge);
5835 mc.moved_charge = 0;
5837 /* we must fixup refcnts and charges */
5838 if (mc.moved_swap) {
5839 /* uncharge swap account from the old cgroup */
5840 if (!mem_cgroup_is_root(mc.from))
5841 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5843 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5846 * we charged both to->memory and to->memsw, so we
5847 * should uncharge to->memory.
5849 if (!mem_cgroup_is_root(mc.to))
5850 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5854 memcg_oom_recover(from);
5855 memcg_oom_recover(to);
5856 wake_up_all(&mc.waitq);
5859 static void mem_cgroup_clear_mc(void)
5861 struct mm_struct *mm = mc.mm;
5864 * we must clear moving_task before waking up waiters at the end of
5867 mc.moving_task = NULL;
5868 __mem_cgroup_clear_mc();
5869 spin_lock(&mc.lock);
5873 spin_unlock(&mc.lock);
5878 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5880 struct cgroup_subsys_state *css;
5881 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5882 struct mem_cgroup *from;
5883 struct task_struct *leader, *p;
5884 struct mm_struct *mm;
5885 unsigned long move_flags;
5888 /* charge immigration isn't supported on the default hierarchy */
5889 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5893 * Multi-process migrations only happen on the default hierarchy
5894 * where charge immigration is not used. Perform charge
5895 * immigration if @tset contains a leader and whine if there are
5899 cgroup_taskset_for_each_leader(leader, css, tset) {
5902 memcg = mem_cgroup_from_css(css);
5908 * We are now commited to this value whatever it is. Changes in this
5909 * tunable will only affect upcoming migrations, not the current one.
5910 * So we need to save it, and keep it going.
5912 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5916 from = mem_cgroup_from_task(p);
5918 VM_BUG_ON(from == memcg);
5920 mm = get_task_mm(p);
5923 /* We move charges only when we move a owner of the mm */
5924 if (mm->owner == p) {
5927 VM_BUG_ON(mc.precharge);
5928 VM_BUG_ON(mc.moved_charge);
5929 VM_BUG_ON(mc.moved_swap);
5931 spin_lock(&mc.lock);
5935 mc.flags = move_flags;
5936 spin_unlock(&mc.lock);
5937 /* We set mc.moving_task later */
5939 ret = mem_cgroup_precharge_mc(mm);
5941 mem_cgroup_clear_mc();
5948 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5951 mem_cgroup_clear_mc();
5954 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5955 unsigned long addr, unsigned long end,
5956 struct mm_walk *walk)
5959 struct vm_area_struct *vma = walk->vma;
5962 enum mc_target_type target_type;
5963 union mc_target target;
5966 ptl = pmd_trans_huge_lock(pmd, vma);
5968 if (mc.precharge < HPAGE_PMD_NR) {
5972 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5973 if (target_type == MC_TARGET_PAGE) {
5975 if (!isolate_lru_page(page)) {
5976 if (!mem_cgroup_move_account(page, true,
5978 mc.precharge -= HPAGE_PMD_NR;
5979 mc.moved_charge += HPAGE_PMD_NR;
5981 putback_lru_page(page);
5984 } else if (target_type == MC_TARGET_DEVICE) {
5986 if (!mem_cgroup_move_account(page, true,
5988 mc.precharge -= HPAGE_PMD_NR;
5989 mc.moved_charge += HPAGE_PMD_NR;
5997 if (pmd_trans_unstable(pmd))
6000 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6001 for (; addr != end; addr += PAGE_SIZE) {
6002 pte_t ptent = *(pte++);
6003 bool device = false;
6009 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6010 case MC_TARGET_DEVICE:
6013 case MC_TARGET_PAGE:
6016 * We can have a part of the split pmd here. Moving it
6017 * can be done but it would be too convoluted so simply
6018 * ignore such a partial THP and keep it in original
6019 * memcg. There should be somebody mapping the head.
6021 if (PageTransCompound(page))
6023 if (!device && isolate_lru_page(page))
6025 if (!mem_cgroup_move_account(page, false,
6028 /* we uncharge from mc.from later. */
6032 putback_lru_page(page);
6033 put: /* get_mctgt_type() gets the page */
6036 case MC_TARGET_SWAP:
6038 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6040 mem_cgroup_id_get_many(mc.to, 1);
6041 /* we fixup other refcnts and charges later. */
6049 pte_unmap_unlock(pte - 1, ptl);
6054 * We have consumed all precharges we got in can_attach().
6055 * We try charge one by one, but don't do any additional
6056 * charges to mc.to if we have failed in charge once in attach()
6059 ret = mem_cgroup_do_precharge(1);
6067 static const struct mm_walk_ops charge_walk_ops = {
6068 .pmd_entry = mem_cgroup_move_charge_pte_range,
6071 static void mem_cgroup_move_charge(void)
6073 lru_add_drain_all();
6075 * Signal lock_page_memcg() to take the memcg's move_lock
6076 * while we're moving its pages to another memcg. Then wait
6077 * for already started RCU-only updates to finish.
6079 atomic_inc(&mc.from->moving_account);
6082 if (unlikely(!mmap_read_trylock(mc.mm))) {
6084 * Someone who are holding the mmap_lock might be waiting in
6085 * waitq. So we cancel all extra charges, wake up all waiters,
6086 * and retry. Because we cancel precharges, we might not be able
6087 * to move enough charges, but moving charge is a best-effort
6088 * feature anyway, so it wouldn't be a big problem.
6090 __mem_cgroup_clear_mc();
6095 * When we have consumed all precharges and failed in doing
6096 * additional charge, the page walk just aborts.
6098 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6101 mmap_read_unlock(mc.mm);
6102 atomic_dec(&mc.from->moving_account);
6105 static void mem_cgroup_move_task(void)
6108 mem_cgroup_move_charge();
6109 mem_cgroup_clear_mc();
6112 #else /* !CONFIG_MMU */
6113 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6117 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6120 static void mem_cgroup_move_task(void)
6125 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6127 if (value == PAGE_COUNTER_MAX)
6128 seq_puts(m, "max\n");
6130 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6135 static u64 memory_current_read(struct cgroup_subsys_state *css,
6138 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6140 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6143 static int memory_min_show(struct seq_file *m, void *v)
6145 return seq_puts_memcg_tunable(m,
6146 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6149 static ssize_t memory_min_write(struct kernfs_open_file *of,
6150 char *buf, size_t nbytes, loff_t off)
6152 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6156 buf = strstrip(buf);
6157 err = page_counter_memparse(buf, "max", &min);
6161 page_counter_set_min(&memcg->memory, min);
6166 static int memory_low_show(struct seq_file *m, void *v)
6168 return seq_puts_memcg_tunable(m,
6169 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6172 static ssize_t memory_low_write(struct kernfs_open_file *of,
6173 char *buf, size_t nbytes, loff_t off)
6175 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6179 buf = strstrip(buf);
6180 err = page_counter_memparse(buf, "max", &low);
6184 page_counter_set_low(&memcg->memory, low);
6189 static int memory_high_show(struct seq_file *m, void *v)
6191 return seq_puts_memcg_tunable(m,
6192 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6195 static ssize_t memory_high_write(struct kernfs_open_file *of,
6196 char *buf, size_t nbytes, loff_t off)
6198 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6199 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6200 bool drained = false;
6204 buf = strstrip(buf);
6205 err = page_counter_memparse(buf, "max", &high);
6209 page_counter_set_high(&memcg->memory, high);
6212 unsigned long nr_pages = page_counter_read(&memcg->memory);
6213 unsigned long reclaimed;
6215 if (nr_pages <= high)
6218 if (signal_pending(current))
6222 drain_all_stock(memcg);
6227 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6230 if (!reclaimed && !nr_retries--)
6234 memcg_wb_domain_size_changed(memcg);
6238 static int memory_max_show(struct seq_file *m, void *v)
6240 return seq_puts_memcg_tunable(m,
6241 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6244 static ssize_t memory_max_write(struct kernfs_open_file *of,
6245 char *buf, size_t nbytes, loff_t off)
6247 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6248 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6249 bool drained = false;
6253 buf = strstrip(buf);
6254 err = page_counter_memparse(buf, "max", &max);
6258 xchg(&memcg->memory.max, max);
6261 unsigned long nr_pages = page_counter_read(&memcg->memory);
6263 if (nr_pages <= max)
6266 if (signal_pending(current))
6270 drain_all_stock(memcg);
6276 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6282 memcg_memory_event(memcg, MEMCG_OOM);
6283 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6287 memcg_wb_domain_size_changed(memcg);
6291 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6293 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6294 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6295 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6296 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6297 seq_printf(m, "oom_kill %lu\n",
6298 atomic_long_read(&events[MEMCG_OOM_KILL]));
6301 static int memory_events_show(struct seq_file *m, void *v)
6303 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6305 __memory_events_show(m, memcg->memory_events);
6309 static int memory_events_local_show(struct seq_file *m, void *v)
6311 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6313 __memory_events_show(m, memcg->memory_events_local);
6317 static int memory_stat_show(struct seq_file *m, void *v)
6319 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6322 buf = memory_stat_format(memcg);
6331 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6334 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6337 static int memory_numa_stat_show(struct seq_file *m, void *v)
6340 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6342 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6345 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6348 seq_printf(m, "%s", memory_stats[i].name);
6349 for_each_node_state(nid, N_MEMORY) {
6351 struct lruvec *lruvec;
6353 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6354 size = lruvec_page_state_output(lruvec,
6355 memory_stats[i].idx);
6356 seq_printf(m, " N%d=%llu", nid, size);
6365 static int memory_oom_group_show(struct seq_file *m, void *v)
6367 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6369 seq_printf(m, "%d\n", memcg->oom_group);
6374 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6375 char *buf, size_t nbytes, loff_t off)
6377 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6380 buf = strstrip(buf);
6384 ret = kstrtoint(buf, 0, &oom_group);
6388 if (oom_group != 0 && oom_group != 1)
6391 memcg->oom_group = oom_group;
6396 static struct cftype memory_files[] = {
6399 .flags = CFTYPE_NOT_ON_ROOT,
6400 .read_u64 = memory_current_read,
6404 .flags = CFTYPE_NOT_ON_ROOT,
6405 .seq_show = memory_min_show,
6406 .write = memory_min_write,
6410 .flags = CFTYPE_NOT_ON_ROOT,
6411 .seq_show = memory_low_show,
6412 .write = memory_low_write,
6416 .flags = CFTYPE_NOT_ON_ROOT,
6417 .seq_show = memory_high_show,
6418 .write = memory_high_write,
6422 .flags = CFTYPE_NOT_ON_ROOT,
6423 .seq_show = memory_max_show,
6424 .write = memory_max_write,
6428 .flags = CFTYPE_NOT_ON_ROOT,
6429 .file_offset = offsetof(struct mem_cgroup, events_file),
6430 .seq_show = memory_events_show,
6433 .name = "events.local",
6434 .flags = CFTYPE_NOT_ON_ROOT,
6435 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6436 .seq_show = memory_events_local_show,
6440 .seq_show = memory_stat_show,
6444 .name = "numa_stat",
6445 .seq_show = memory_numa_stat_show,
6449 .name = "oom.group",
6450 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6451 .seq_show = memory_oom_group_show,
6452 .write = memory_oom_group_write,
6457 struct cgroup_subsys memory_cgrp_subsys = {
6458 .css_alloc = mem_cgroup_css_alloc,
6459 .css_online = mem_cgroup_css_online,
6460 .css_offline = mem_cgroup_css_offline,
6461 .css_released = mem_cgroup_css_released,
6462 .css_free = mem_cgroup_css_free,
6463 .css_reset = mem_cgroup_css_reset,
6464 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6465 .can_attach = mem_cgroup_can_attach,
6466 .cancel_attach = mem_cgroup_cancel_attach,
6467 .post_attach = mem_cgroup_move_task,
6468 .dfl_cftypes = memory_files,
6469 .legacy_cftypes = mem_cgroup_legacy_files,
6474 * This function calculates an individual cgroup's effective
6475 * protection which is derived from its own memory.min/low, its
6476 * parent's and siblings' settings, as well as the actual memory
6477 * distribution in the tree.
6479 * The following rules apply to the effective protection values:
6481 * 1. At the first level of reclaim, effective protection is equal to
6482 * the declared protection in memory.min and memory.low.
6484 * 2. To enable safe delegation of the protection configuration, at
6485 * subsequent levels the effective protection is capped to the
6486 * parent's effective protection.
6488 * 3. To make complex and dynamic subtrees easier to configure, the
6489 * user is allowed to overcommit the declared protection at a given
6490 * level. If that is the case, the parent's effective protection is
6491 * distributed to the children in proportion to how much protection
6492 * they have declared and how much of it they are utilizing.
6494 * This makes distribution proportional, but also work-conserving:
6495 * if one cgroup claims much more protection than it uses memory,
6496 * the unused remainder is available to its siblings.
6498 * 4. Conversely, when the declared protection is undercommitted at a
6499 * given level, the distribution of the larger parental protection
6500 * budget is NOT proportional. A cgroup's protection from a sibling
6501 * is capped to its own memory.min/low setting.
6503 * 5. However, to allow protecting recursive subtrees from each other
6504 * without having to declare each individual cgroup's fixed share
6505 * of the ancestor's claim to protection, any unutilized -
6506 * "floating" - protection from up the tree is distributed in
6507 * proportion to each cgroup's *usage*. This makes the protection
6508 * neutral wrt sibling cgroups and lets them compete freely over
6509 * the shared parental protection budget, but it protects the
6510 * subtree as a whole from neighboring subtrees.
6512 * Note that 4. and 5. are not in conflict: 4. is about protecting
6513 * against immediate siblings whereas 5. is about protecting against
6514 * neighboring subtrees.
6516 static unsigned long effective_protection(unsigned long usage,
6517 unsigned long parent_usage,
6518 unsigned long setting,
6519 unsigned long parent_effective,
6520 unsigned long siblings_protected)
6522 unsigned long protected;
6525 protected = min(usage, setting);
6527 * If all cgroups at this level combined claim and use more
6528 * protection then what the parent affords them, distribute
6529 * shares in proportion to utilization.
6531 * We are using actual utilization rather than the statically
6532 * claimed protection in order to be work-conserving: claimed
6533 * but unused protection is available to siblings that would
6534 * otherwise get a smaller chunk than what they claimed.
6536 if (siblings_protected > parent_effective)
6537 return protected * parent_effective / siblings_protected;
6540 * Ok, utilized protection of all children is within what the
6541 * parent affords them, so we know whatever this child claims
6542 * and utilizes is effectively protected.
6544 * If there is unprotected usage beyond this value, reclaim
6545 * will apply pressure in proportion to that amount.
6547 * If there is unutilized protection, the cgroup will be fully
6548 * shielded from reclaim, but we do return a smaller value for
6549 * protection than what the group could enjoy in theory. This
6550 * is okay. With the overcommit distribution above, effective
6551 * protection is always dependent on how memory is actually
6552 * consumed among the siblings anyway.
6557 * If the children aren't claiming (all of) the protection
6558 * afforded to them by the parent, distribute the remainder in
6559 * proportion to the (unprotected) memory of each cgroup. That
6560 * way, cgroups that aren't explicitly prioritized wrt each
6561 * other compete freely over the allowance, but they are
6562 * collectively protected from neighboring trees.
6564 * We're using unprotected memory for the weight so that if
6565 * some cgroups DO claim explicit protection, we don't protect
6566 * the same bytes twice.
6568 * Check both usage and parent_usage against the respective
6569 * protected values. One should imply the other, but they
6570 * aren't read atomically - make sure the division is sane.
6572 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6574 if (parent_effective > siblings_protected &&
6575 parent_usage > siblings_protected &&
6576 usage > protected) {
6577 unsigned long unclaimed;
6579 unclaimed = parent_effective - siblings_protected;
6580 unclaimed *= usage - protected;
6581 unclaimed /= parent_usage - siblings_protected;
6590 * mem_cgroup_protected - check if memory consumption is in the normal range
6591 * @root: the top ancestor of the sub-tree being checked
6592 * @memcg: the memory cgroup to check
6594 * WARNING: This function is not stateless! It can only be used as part
6595 * of a top-down tree iteration, not for isolated queries.
6597 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6598 struct mem_cgroup *memcg)
6600 unsigned long usage, parent_usage;
6601 struct mem_cgroup *parent;
6603 if (mem_cgroup_disabled())
6607 root = root_mem_cgroup;
6610 * Effective values of the reclaim targets are ignored so they
6611 * can be stale. Have a look at mem_cgroup_protection for more
6613 * TODO: calculation should be more robust so that we do not need
6614 * that special casing.
6619 usage = page_counter_read(&memcg->memory);
6623 parent = parent_mem_cgroup(memcg);
6624 /* No parent means a non-hierarchical mode on v1 memcg */
6628 if (parent == root) {
6629 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6630 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6634 parent_usage = page_counter_read(&parent->memory);
6636 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6637 READ_ONCE(memcg->memory.min),
6638 READ_ONCE(parent->memory.emin),
6639 atomic_long_read(&parent->memory.children_min_usage)));
6641 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6642 READ_ONCE(memcg->memory.low),
6643 READ_ONCE(parent->memory.elow),
6644 atomic_long_read(&parent->memory.children_low_usage)));
6647 static int __mem_cgroup_charge(struct page *page, struct mem_cgroup *memcg,
6650 unsigned int nr_pages = thp_nr_pages(page);
6653 ret = try_charge(memcg, gfp, nr_pages);
6657 css_get(&memcg->css);
6658 commit_charge(page, memcg);
6660 local_irq_disable();
6661 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6662 memcg_check_events(memcg, page);
6669 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6670 * @page: page to charge
6671 * @mm: mm context of the victim
6672 * @gfp_mask: reclaim mode
6674 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6675 * pages according to @gfp_mask if necessary.
6677 * Do not use this for pages allocated for swapin.
6679 * Returns 0 on success. Otherwise, an error code is returned.
6681 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6683 struct mem_cgroup *memcg;
6686 if (mem_cgroup_disabled())
6689 memcg = get_mem_cgroup_from_mm(mm);
6690 ret = __mem_cgroup_charge(page, memcg, gfp_mask);
6691 css_put(&memcg->css);
6697 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6698 * @page: page to charge
6699 * @mm: mm context of the victim
6700 * @gfp: reclaim mode
6701 * @entry: swap entry for which the page is allocated
6703 * This function charges a page allocated for swapin. Please call this before
6704 * adding the page to the swapcache.
6706 * Returns 0 on success. Otherwise, an error code is returned.
6708 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6709 gfp_t gfp, swp_entry_t entry)
6711 struct mem_cgroup *memcg;
6715 if (mem_cgroup_disabled())
6718 id = lookup_swap_cgroup_id(entry);
6720 memcg = mem_cgroup_from_id(id);
6721 if (!memcg || !css_tryget_online(&memcg->css))
6722 memcg = get_mem_cgroup_from_mm(mm);
6725 ret = __mem_cgroup_charge(page, memcg, gfp);
6727 css_put(&memcg->css);
6732 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6733 * @entry: swap entry for which the page is charged
6735 * Call this function after successfully adding the charged page to swapcache.
6737 * Note: This function assumes the page for which swap slot is being uncharged
6740 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6743 * Cgroup1's unified memory+swap counter has been charged with the
6744 * new swapcache page, finish the transfer by uncharging the swap
6745 * slot. The swap slot would also get uncharged when it dies, but
6746 * it can stick around indefinitely and we'd count the page twice
6749 * Cgroup2 has separate resource counters for memory and swap,
6750 * so this is a non-issue here. Memory and swap charge lifetimes
6751 * correspond 1:1 to page and swap slot lifetimes: we charge the
6752 * page to memory here, and uncharge swap when the slot is freed.
6754 if (!mem_cgroup_disabled() && do_memsw_account()) {
6756 * The swap entry might not get freed for a long time,
6757 * let's not wait for it. The page already received a
6758 * memory+swap charge, drop the swap entry duplicate.
6760 mem_cgroup_uncharge_swap(entry, 1);
6764 struct uncharge_gather {
6765 struct mem_cgroup *memcg;
6766 unsigned long nr_pages;
6767 unsigned long pgpgout;
6768 unsigned long nr_kmem;
6769 struct page *dummy_page;
6772 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6774 memset(ug, 0, sizeof(*ug));
6777 static void uncharge_batch(const struct uncharge_gather *ug)
6779 unsigned long flags;
6781 if (!mem_cgroup_is_root(ug->memcg)) {
6782 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6783 if (do_memsw_account())
6784 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6785 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6786 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6787 memcg_oom_recover(ug->memcg);
6790 local_irq_save(flags);
6791 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6792 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6793 memcg_check_events(ug->memcg, ug->dummy_page);
6794 local_irq_restore(flags);
6796 /* drop reference from uncharge_page */
6797 css_put(&ug->memcg->css);
6800 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6802 unsigned long nr_pages;
6804 VM_BUG_ON_PAGE(PageLRU(page), page);
6806 if (!page_memcg(page))
6810 * Nobody should be changing or seriously looking at
6811 * page_memcg(page) at this point, we have fully
6812 * exclusive access to the page.
6815 if (ug->memcg != page_memcg(page)) {
6818 uncharge_gather_clear(ug);
6820 ug->memcg = page_memcg(page);
6822 /* pairs with css_put in uncharge_batch */
6823 css_get(&ug->memcg->css);
6826 nr_pages = compound_nr(page);
6827 ug->nr_pages += nr_pages;
6829 if (PageMemcgKmem(page))
6830 ug->nr_kmem += nr_pages;
6834 ug->dummy_page = page;
6835 page->memcg_data = 0;
6836 css_put(&ug->memcg->css);
6840 * mem_cgroup_uncharge - uncharge a page
6841 * @page: page to uncharge
6843 * Uncharge a page previously charged with mem_cgroup_charge().
6845 void mem_cgroup_uncharge(struct page *page)
6847 struct uncharge_gather ug;
6849 if (mem_cgroup_disabled())
6852 /* Don't touch page->lru of any random page, pre-check: */
6853 if (!page_memcg(page))
6856 uncharge_gather_clear(&ug);
6857 uncharge_page(page, &ug);
6858 uncharge_batch(&ug);
6862 * mem_cgroup_uncharge_list - uncharge a list of page
6863 * @page_list: list of pages to uncharge
6865 * Uncharge a list of pages previously charged with
6866 * mem_cgroup_charge().
6868 void mem_cgroup_uncharge_list(struct list_head *page_list)
6870 struct uncharge_gather ug;
6873 if (mem_cgroup_disabled())
6876 uncharge_gather_clear(&ug);
6877 list_for_each_entry(page, page_list, lru)
6878 uncharge_page(page, &ug);
6880 uncharge_batch(&ug);
6884 * mem_cgroup_migrate - charge a page's replacement
6885 * @oldpage: currently circulating page
6886 * @newpage: replacement page
6888 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6889 * be uncharged upon free.
6891 * Both pages must be locked, @newpage->mapping must be set up.
6893 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6895 struct mem_cgroup *memcg;
6896 unsigned int nr_pages;
6897 unsigned long flags;
6899 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6900 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6901 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6902 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6905 if (mem_cgroup_disabled())
6908 /* Page cache replacement: new page already charged? */
6909 if (page_memcg(newpage))
6912 memcg = page_memcg(oldpage);
6913 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6917 /* Force-charge the new page. The old one will be freed soon */
6918 nr_pages = thp_nr_pages(newpage);
6920 page_counter_charge(&memcg->memory, nr_pages);
6921 if (do_memsw_account())
6922 page_counter_charge(&memcg->memsw, nr_pages);
6924 css_get(&memcg->css);
6925 commit_charge(newpage, memcg);
6927 local_irq_save(flags);
6928 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6929 memcg_check_events(memcg, newpage);
6930 local_irq_restore(flags);
6933 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6934 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6936 void mem_cgroup_sk_alloc(struct sock *sk)
6938 struct mem_cgroup *memcg;
6940 if (!mem_cgroup_sockets_enabled)
6943 /* Do not associate the sock with unrelated interrupted task's memcg. */
6948 memcg = mem_cgroup_from_task(current);
6949 if (memcg == root_mem_cgroup)
6951 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6953 if (css_tryget(&memcg->css))
6954 sk->sk_memcg = memcg;
6959 void mem_cgroup_sk_free(struct sock *sk)
6962 css_put(&sk->sk_memcg->css);
6966 * mem_cgroup_charge_skmem - charge socket memory
6967 * @memcg: memcg to charge
6968 * @nr_pages: number of pages to charge
6970 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6971 * @memcg's configured limit, %false if the charge had to be forced.
6973 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6975 gfp_t gfp_mask = GFP_KERNEL;
6977 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6978 struct page_counter *fail;
6980 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6981 memcg->tcpmem_pressure = 0;
6984 page_counter_charge(&memcg->tcpmem, nr_pages);
6985 memcg->tcpmem_pressure = 1;
6989 /* Don't block in the packet receive path */
6991 gfp_mask = GFP_NOWAIT;
6993 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6995 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6998 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7003 * mem_cgroup_uncharge_skmem - uncharge socket memory
7004 * @memcg: memcg to uncharge
7005 * @nr_pages: number of pages to uncharge
7007 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7009 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7010 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7014 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7016 refill_stock(memcg, nr_pages);
7019 static int __init cgroup_memory(char *s)
7023 while ((token = strsep(&s, ",")) != NULL) {
7026 if (!strcmp(token, "nosocket"))
7027 cgroup_memory_nosocket = true;
7028 if (!strcmp(token, "nokmem"))
7029 cgroup_memory_nokmem = true;
7033 __setup("cgroup.memory=", cgroup_memory);
7036 * subsys_initcall() for memory controller.
7038 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7039 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7040 * basically everything that doesn't depend on a specific mem_cgroup structure
7041 * should be initialized from here.
7043 static int __init mem_cgroup_init(void)
7048 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7049 * used for per-memcg-per-cpu caching of per-node statistics. In order
7050 * to work fine, we should make sure that the overfill threshold can't
7051 * exceed S32_MAX / PAGE_SIZE.
7053 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7055 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7056 memcg_hotplug_cpu_dead);
7058 for_each_possible_cpu(cpu)
7059 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7062 for_each_node(node) {
7063 struct mem_cgroup_tree_per_node *rtpn;
7065 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7066 node_online(node) ? node : NUMA_NO_NODE);
7068 rtpn->rb_root = RB_ROOT;
7069 rtpn->rb_rightmost = NULL;
7070 spin_lock_init(&rtpn->lock);
7071 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7076 subsys_initcall(mem_cgroup_init);
7078 #ifdef CONFIG_MEMCG_SWAP
7079 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7081 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7083 * The root cgroup cannot be destroyed, so it's refcount must
7086 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7090 memcg = parent_mem_cgroup(memcg);
7092 memcg = root_mem_cgroup;
7098 * mem_cgroup_swapout - transfer a memsw charge to swap
7099 * @page: page whose memsw charge to transfer
7100 * @entry: swap entry to move the charge to
7102 * Transfer the memsw charge of @page to @entry.
7104 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7106 struct mem_cgroup *memcg, *swap_memcg;
7107 unsigned int nr_entries;
7108 unsigned short oldid;
7110 VM_BUG_ON_PAGE(PageLRU(page), page);
7111 VM_BUG_ON_PAGE(page_count(page), page);
7113 if (mem_cgroup_disabled())
7116 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7119 memcg = page_memcg(page);
7121 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7126 * In case the memcg owning these pages has been offlined and doesn't
7127 * have an ID allocated to it anymore, charge the closest online
7128 * ancestor for the swap instead and transfer the memory+swap charge.
7130 swap_memcg = mem_cgroup_id_get_online(memcg);
7131 nr_entries = thp_nr_pages(page);
7132 /* Get references for the tail pages, too */
7134 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7135 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7137 VM_BUG_ON_PAGE(oldid, page);
7138 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7140 page->memcg_data = 0;
7142 if (!mem_cgroup_is_root(memcg))
7143 page_counter_uncharge(&memcg->memory, nr_entries);
7145 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7146 if (!mem_cgroup_is_root(swap_memcg))
7147 page_counter_charge(&swap_memcg->memsw, nr_entries);
7148 page_counter_uncharge(&memcg->memsw, nr_entries);
7152 * Interrupts should be disabled here because the caller holds the
7153 * i_pages lock which is taken with interrupts-off. It is
7154 * important here to have the interrupts disabled because it is the
7155 * only synchronisation we have for updating the per-CPU variables.
7157 VM_BUG_ON(!irqs_disabled());
7158 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7159 memcg_check_events(memcg, page);
7161 css_put(&memcg->css);
7165 * mem_cgroup_try_charge_swap - try charging swap space for a page
7166 * @page: page being added to swap
7167 * @entry: swap entry to charge
7169 * Try to charge @page's memcg for the swap space at @entry.
7171 * Returns 0 on success, -ENOMEM on failure.
7173 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7175 unsigned int nr_pages = thp_nr_pages(page);
7176 struct page_counter *counter;
7177 struct mem_cgroup *memcg;
7178 unsigned short oldid;
7180 if (mem_cgroup_disabled())
7183 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7186 memcg = page_memcg(page);
7188 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7193 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7197 memcg = mem_cgroup_id_get_online(memcg);
7199 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7200 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7201 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7202 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7203 mem_cgroup_id_put(memcg);
7207 /* Get references for the tail pages, too */
7209 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7210 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7211 VM_BUG_ON_PAGE(oldid, page);
7212 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7218 * mem_cgroup_uncharge_swap - uncharge swap space
7219 * @entry: swap entry to uncharge
7220 * @nr_pages: the amount of swap space to uncharge
7222 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7224 struct mem_cgroup *memcg;
7227 id = swap_cgroup_record(entry, 0, nr_pages);
7229 memcg = mem_cgroup_from_id(id);
7231 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7232 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7233 page_counter_uncharge(&memcg->swap, nr_pages);
7235 page_counter_uncharge(&memcg->memsw, nr_pages);
7237 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7238 mem_cgroup_id_put_many(memcg, nr_pages);
7243 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7245 long nr_swap_pages = get_nr_swap_pages();
7247 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7248 return nr_swap_pages;
7249 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7250 nr_swap_pages = min_t(long, nr_swap_pages,
7251 READ_ONCE(memcg->swap.max) -
7252 page_counter_read(&memcg->swap));
7253 return nr_swap_pages;
7256 bool mem_cgroup_swap_full(struct page *page)
7258 struct mem_cgroup *memcg;
7260 VM_BUG_ON_PAGE(!PageLocked(page), page);
7264 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7267 memcg = page_memcg(page);
7271 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7272 unsigned long usage = page_counter_read(&memcg->swap);
7274 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7275 usage * 2 >= READ_ONCE(memcg->swap.max))
7282 static int __init setup_swap_account(char *s)
7284 if (!strcmp(s, "1"))
7285 cgroup_memory_noswap = false;
7286 else if (!strcmp(s, "0"))
7287 cgroup_memory_noswap = true;
7290 __setup("swapaccount=", setup_swap_account);
7292 static u64 swap_current_read(struct cgroup_subsys_state *css,
7295 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7297 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7300 static int swap_high_show(struct seq_file *m, void *v)
7302 return seq_puts_memcg_tunable(m,
7303 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7306 static ssize_t swap_high_write(struct kernfs_open_file *of,
7307 char *buf, size_t nbytes, loff_t off)
7309 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7313 buf = strstrip(buf);
7314 err = page_counter_memparse(buf, "max", &high);
7318 page_counter_set_high(&memcg->swap, high);
7323 static int swap_max_show(struct seq_file *m, void *v)
7325 return seq_puts_memcg_tunable(m,
7326 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7329 static ssize_t swap_max_write(struct kernfs_open_file *of,
7330 char *buf, size_t nbytes, loff_t off)
7332 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7336 buf = strstrip(buf);
7337 err = page_counter_memparse(buf, "max", &max);
7341 xchg(&memcg->swap.max, max);
7346 static int swap_events_show(struct seq_file *m, void *v)
7348 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7350 seq_printf(m, "high %lu\n",
7351 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7352 seq_printf(m, "max %lu\n",
7353 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7354 seq_printf(m, "fail %lu\n",
7355 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7360 static struct cftype swap_files[] = {
7362 .name = "swap.current",
7363 .flags = CFTYPE_NOT_ON_ROOT,
7364 .read_u64 = swap_current_read,
7367 .name = "swap.high",
7368 .flags = CFTYPE_NOT_ON_ROOT,
7369 .seq_show = swap_high_show,
7370 .write = swap_high_write,
7374 .flags = CFTYPE_NOT_ON_ROOT,
7375 .seq_show = swap_max_show,
7376 .write = swap_max_write,
7379 .name = "swap.events",
7380 .flags = CFTYPE_NOT_ON_ROOT,
7381 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7382 .seq_show = swap_events_show,
7387 static struct cftype memsw_files[] = {
7389 .name = "memsw.usage_in_bytes",
7390 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7391 .read_u64 = mem_cgroup_read_u64,
7394 .name = "memsw.max_usage_in_bytes",
7395 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7396 .write = mem_cgroup_reset,
7397 .read_u64 = mem_cgroup_read_u64,
7400 .name = "memsw.limit_in_bytes",
7401 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7402 .write = mem_cgroup_write,
7403 .read_u64 = mem_cgroup_read_u64,
7406 .name = "memsw.failcnt",
7407 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7408 .write = mem_cgroup_reset,
7409 .read_u64 = mem_cgroup_read_u64,
7411 { }, /* terminate */
7415 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7416 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7417 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7418 * boot parameter. This may result in premature OOPS inside
7419 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7421 static int __init mem_cgroup_swap_init(void)
7423 /* No memory control -> no swap control */
7424 if (mem_cgroup_disabled())
7425 cgroup_memory_noswap = true;
7427 if (cgroup_memory_noswap)
7430 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7431 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7435 core_initcall(mem_cgroup_swap_init);
7437 #endif /* CONFIG_MEMCG_SWAP */