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;
869 pg_data_t *pgdat = page_pgdat(page);
870 struct lruvec *lruvec;
873 memcg = page_memcg(head);
874 /* Untracked pages have no memcg, no lruvec. Update only the node */
877 __mod_node_page_state(pgdat, idx, val);
881 lruvec = mem_cgroup_lruvec(memcg, pgdat);
882 __mod_lruvec_state(lruvec, idx, val);
885 EXPORT_SYMBOL(__mod_lruvec_page_state);
887 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
889 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
890 struct mem_cgroup *memcg;
891 struct lruvec *lruvec;
894 memcg = mem_cgroup_from_obj(p);
897 * Untracked pages have no memcg, no lruvec. Update only the
898 * node. If we reparent the slab objects to the root memcg,
899 * when we free the slab object, we need to update the per-memcg
900 * vmstats to keep it correct for the root memcg.
903 __mod_node_page_state(pgdat, idx, val);
905 lruvec = mem_cgroup_lruvec(memcg, pgdat);
906 __mod_lruvec_state(lruvec, idx, val);
912 * __count_memcg_events - account VM events in a cgroup
913 * @memcg: the memory cgroup
914 * @idx: the event item
915 * @count: the number of events that occured
917 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
920 if (mem_cgroup_disabled())
923 __this_cpu_add(memcg->vmstats_percpu->events[idx], count);
924 cgroup_rstat_updated(memcg->css.cgroup, smp_processor_id());
927 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
929 return READ_ONCE(memcg->vmstats.events[event]);
932 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
937 for_each_possible_cpu(cpu)
938 x += per_cpu(memcg->vmstats_percpu->events[event], cpu);
942 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
946 /* pagein of a big page is an event. So, ignore page size */
948 __count_memcg_events(memcg, PGPGIN, 1);
950 __count_memcg_events(memcg, PGPGOUT, 1);
951 nr_pages = -nr_pages; /* for event */
954 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
957 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
958 enum mem_cgroup_events_target target)
960 unsigned long val, next;
962 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
963 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
964 /* from time_after() in jiffies.h */
965 if ((long)(next - val) < 0) {
967 case MEM_CGROUP_TARGET_THRESH:
968 next = val + THRESHOLDS_EVENTS_TARGET;
970 case MEM_CGROUP_TARGET_SOFTLIMIT:
971 next = val + SOFTLIMIT_EVENTS_TARGET;
976 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
983 * Check events in order.
986 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
988 /* threshold event is triggered in finer grain than soft limit */
989 if (unlikely(mem_cgroup_event_ratelimit(memcg,
990 MEM_CGROUP_TARGET_THRESH))) {
993 do_softlimit = mem_cgroup_event_ratelimit(memcg,
994 MEM_CGROUP_TARGET_SOFTLIMIT);
995 mem_cgroup_threshold(memcg);
996 if (unlikely(do_softlimit))
997 mem_cgroup_update_tree(memcg, page);
1001 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1004 * mm_update_next_owner() may clear mm->owner to NULL
1005 * if it races with swapoff, page migration, etc.
1006 * So this can be called with p == NULL.
1011 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1013 EXPORT_SYMBOL(mem_cgroup_from_task);
1016 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1017 * @mm: mm from which memcg should be extracted. It can be NULL.
1019 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1020 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1023 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1025 struct mem_cgroup *memcg;
1027 if (mem_cgroup_disabled())
1033 * Page cache insertions can happen withou an
1034 * actual mm context, e.g. during disk probing
1035 * on boot, loopback IO, acct() writes etc.
1038 memcg = root_mem_cgroup;
1040 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1041 if (unlikely(!memcg))
1042 memcg = root_mem_cgroup;
1044 } while (!css_tryget(&memcg->css));
1048 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1050 static __always_inline struct mem_cgroup *active_memcg(void)
1053 return this_cpu_read(int_active_memcg);
1055 return current->active_memcg;
1058 static __always_inline bool memcg_kmem_bypass(void)
1060 /* Allow remote memcg charging from any context. */
1061 if (unlikely(active_memcg()))
1064 /* Memcg to charge can't be determined. */
1065 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1072 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1073 * @root: hierarchy root
1074 * @prev: previously returned memcg, NULL on first invocation
1075 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1077 * Returns references to children of the hierarchy below @root, or
1078 * @root itself, or %NULL after a full round-trip.
1080 * Caller must pass the return value in @prev on subsequent
1081 * invocations for reference counting, or use mem_cgroup_iter_break()
1082 * to cancel a hierarchy walk before the round-trip is complete.
1084 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1085 * in the hierarchy among all concurrent reclaimers operating on the
1088 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1089 struct mem_cgroup *prev,
1090 struct mem_cgroup_reclaim_cookie *reclaim)
1092 struct mem_cgroup_reclaim_iter *iter;
1093 struct cgroup_subsys_state *css = NULL;
1094 struct mem_cgroup *memcg = NULL;
1095 struct mem_cgroup *pos = NULL;
1097 if (mem_cgroup_disabled())
1101 root = root_mem_cgroup;
1103 if (prev && !reclaim)
1109 struct mem_cgroup_per_node *mz;
1111 mz = root->nodeinfo[reclaim->pgdat->node_id];
1114 if (prev && reclaim->generation != iter->generation)
1118 pos = READ_ONCE(iter->position);
1119 if (!pos || css_tryget(&pos->css))
1122 * css reference reached zero, so iter->position will
1123 * be cleared by ->css_released. However, we should not
1124 * rely on this happening soon, because ->css_released
1125 * is called from a work queue, and by busy-waiting we
1126 * might block it. So we clear iter->position right
1129 (void)cmpxchg(&iter->position, pos, NULL);
1137 css = css_next_descendant_pre(css, &root->css);
1140 * Reclaimers share the hierarchy walk, and a
1141 * new one might jump in right at the end of
1142 * the hierarchy - make sure they see at least
1143 * one group and restart from the beginning.
1151 * Verify the css and acquire a reference. The root
1152 * is provided by the caller, so we know it's alive
1153 * and kicking, and don't take an extra reference.
1155 memcg = mem_cgroup_from_css(css);
1157 if (css == &root->css)
1160 if (css_tryget(css))
1168 * The position could have already been updated by a competing
1169 * thread, so check that the value hasn't changed since we read
1170 * it to avoid reclaiming from the same cgroup twice.
1172 (void)cmpxchg(&iter->position, pos, memcg);
1180 reclaim->generation = iter->generation;
1185 if (prev && prev != root)
1186 css_put(&prev->css);
1192 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1193 * @root: hierarchy root
1194 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1196 void mem_cgroup_iter_break(struct mem_cgroup *root,
1197 struct mem_cgroup *prev)
1200 root = root_mem_cgroup;
1201 if (prev && prev != root)
1202 css_put(&prev->css);
1205 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1206 struct mem_cgroup *dead_memcg)
1208 struct mem_cgroup_reclaim_iter *iter;
1209 struct mem_cgroup_per_node *mz;
1212 for_each_node(nid) {
1213 mz = from->nodeinfo[nid];
1215 cmpxchg(&iter->position, dead_memcg, NULL);
1219 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1221 struct mem_cgroup *memcg = dead_memcg;
1222 struct mem_cgroup *last;
1225 __invalidate_reclaim_iterators(memcg, dead_memcg);
1227 } while ((memcg = parent_mem_cgroup(memcg)));
1230 * When cgruop1 non-hierarchy mode is used,
1231 * parent_mem_cgroup() does not walk all the way up to the
1232 * cgroup root (root_mem_cgroup). So we have to handle
1233 * dead_memcg from cgroup root separately.
1235 if (last != root_mem_cgroup)
1236 __invalidate_reclaim_iterators(root_mem_cgroup,
1241 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1242 * @memcg: hierarchy root
1243 * @fn: function to call for each task
1244 * @arg: argument passed to @fn
1246 * This function iterates over tasks attached to @memcg or to any of its
1247 * descendants and calls @fn for each task. If @fn returns a non-zero
1248 * value, the function breaks the iteration loop and returns the value.
1249 * Otherwise, it will iterate over all tasks and return 0.
1251 * This function must not be called for the root memory cgroup.
1253 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1254 int (*fn)(struct task_struct *, void *), void *arg)
1256 struct mem_cgroup *iter;
1259 BUG_ON(memcg == root_mem_cgroup);
1261 for_each_mem_cgroup_tree(iter, memcg) {
1262 struct css_task_iter it;
1263 struct task_struct *task;
1265 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1266 while (!ret && (task = css_task_iter_next(&it)))
1267 ret = fn(task, arg);
1268 css_task_iter_end(&it);
1270 mem_cgroup_iter_break(memcg, iter);
1277 #ifdef CONFIG_DEBUG_VM
1278 void lruvec_memcg_debug(struct lruvec *lruvec, struct page *page)
1280 struct mem_cgroup *memcg;
1282 if (mem_cgroup_disabled())
1285 memcg = page_memcg(page);
1288 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != root_mem_cgroup, page);
1290 VM_BUG_ON_PAGE(lruvec_memcg(lruvec) != memcg, page);
1295 * lock_page_lruvec - lock and return lruvec for a given page.
1298 * These functions are safe to use under any of the following conditions:
1301 * - lock_page_memcg()
1302 * - page->_refcount is zero
1304 struct lruvec *lock_page_lruvec(struct page *page)
1306 struct lruvec *lruvec;
1307 struct pglist_data *pgdat = page_pgdat(page);
1309 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1310 spin_lock(&lruvec->lru_lock);
1312 lruvec_memcg_debug(lruvec, page);
1317 struct lruvec *lock_page_lruvec_irq(struct page *page)
1319 struct lruvec *lruvec;
1320 struct pglist_data *pgdat = page_pgdat(page);
1322 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1323 spin_lock_irq(&lruvec->lru_lock);
1325 lruvec_memcg_debug(lruvec, page);
1330 struct lruvec *lock_page_lruvec_irqsave(struct page *page, unsigned long *flags)
1332 struct lruvec *lruvec;
1333 struct pglist_data *pgdat = page_pgdat(page);
1335 lruvec = mem_cgroup_page_lruvec(page, pgdat);
1336 spin_lock_irqsave(&lruvec->lru_lock, *flags);
1338 lruvec_memcg_debug(lruvec, page);
1344 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1345 * @lruvec: mem_cgroup per zone lru vector
1346 * @lru: index of lru list the page is sitting on
1347 * @zid: zone id of the accounted pages
1348 * @nr_pages: positive when adding or negative when removing
1350 * This function must be called under lru_lock, just before a page is added
1351 * to or just after a page is removed from an lru list (that ordering being
1352 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1354 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1355 int zid, int nr_pages)
1357 struct mem_cgroup_per_node *mz;
1358 unsigned long *lru_size;
1361 if (mem_cgroup_disabled())
1364 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1365 lru_size = &mz->lru_zone_size[zid][lru];
1368 *lru_size += nr_pages;
1371 if (WARN_ONCE(size < 0,
1372 "%s(%p, %d, %d): lru_size %ld\n",
1373 __func__, lruvec, lru, nr_pages, size)) {
1379 *lru_size += nr_pages;
1383 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1384 * @memcg: the memory cgroup
1386 * Returns the maximum amount of memory @mem can be charged with, in
1389 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1391 unsigned long margin = 0;
1392 unsigned long count;
1393 unsigned long limit;
1395 count = page_counter_read(&memcg->memory);
1396 limit = READ_ONCE(memcg->memory.max);
1398 margin = limit - count;
1400 if (do_memsw_account()) {
1401 count = page_counter_read(&memcg->memsw);
1402 limit = READ_ONCE(memcg->memsw.max);
1404 margin = min(margin, limit - count);
1413 * A routine for checking "mem" is under move_account() or not.
1415 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1416 * moving cgroups. This is for waiting at high-memory pressure
1419 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1421 struct mem_cgroup *from;
1422 struct mem_cgroup *to;
1425 * Unlike task_move routines, we access mc.to, mc.from not under
1426 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1428 spin_lock(&mc.lock);
1434 ret = mem_cgroup_is_descendant(from, memcg) ||
1435 mem_cgroup_is_descendant(to, memcg);
1437 spin_unlock(&mc.lock);
1441 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1443 if (mc.moving_task && current != mc.moving_task) {
1444 if (mem_cgroup_under_move(memcg)) {
1446 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1447 /* moving charge context might have finished. */
1450 finish_wait(&mc.waitq, &wait);
1457 struct memory_stat {
1462 static const struct memory_stat memory_stats[] = {
1463 { "anon", NR_ANON_MAPPED },
1464 { "file", NR_FILE_PAGES },
1465 { "kernel_stack", NR_KERNEL_STACK_KB },
1466 { "pagetables", NR_PAGETABLE },
1467 { "percpu", MEMCG_PERCPU_B },
1468 { "sock", MEMCG_SOCK },
1469 { "shmem", NR_SHMEM },
1470 { "file_mapped", NR_FILE_MAPPED },
1471 { "file_dirty", NR_FILE_DIRTY },
1472 { "file_writeback", NR_WRITEBACK },
1474 { "swapcached", NR_SWAPCACHE },
1476 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1477 { "anon_thp", NR_ANON_THPS },
1478 { "file_thp", NR_FILE_THPS },
1479 { "shmem_thp", NR_SHMEM_THPS },
1481 { "inactive_anon", NR_INACTIVE_ANON },
1482 { "active_anon", NR_ACTIVE_ANON },
1483 { "inactive_file", NR_INACTIVE_FILE },
1484 { "active_file", NR_ACTIVE_FILE },
1485 { "unevictable", NR_UNEVICTABLE },
1486 { "slab_reclaimable", NR_SLAB_RECLAIMABLE_B },
1487 { "slab_unreclaimable", NR_SLAB_UNRECLAIMABLE_B },
1489 /* The memory events */
1490 { "workingset_refault_anon", WORKINGSET_REFAULT_ANON },
1491 { "workingset_refault_file", WORKINGSET_REFAULT_FILE },
1492 { "workingset_activate_anon", WORKINGSET_ACTIVATE_ANON },
1493 { "workingset_activate_file", WORKINGSET_ACTIVATE_FILE },
1494 { "workingset_restore_anon", WORKINGSET_RESTORE_ANON },
1495 { "workingset_restore_file", WORKINGSET_RESTORE_FILE },
1496 { "workingset_nodereclaim", WORKINGSET_NODERECLAIM },
1499 /* Translate stat items to the correct unit for memory.stat output */
1500 static int memcg_page_state_unit(int item)
1503 case MEMCG_PERCPU_B:
1504 case NR_SLAB_RECLAIMABLE_B:
1505 case NR_SLAB_UNRECLAIMABLE_B:
1506 case WORKINGSET_REFAULT_ANON:
1507 case WORKINGSET_REFAULT_FILE:
1508 case WORKINGSET_ACTIVATE_ANON:
1509 case WORKINGSET_ACTIVATE_FILE:
1510 case WORKINGSET_RESTORE_ANON:
1511 case WORKINGSET_RESTORE_FILE:
1512 case WORKINGSET_NODERECLAIM:
1514 case NR_KERNEL_STACK_KB:
1521 static inline unsigned long memcg_page_state_output(struct mem_cgroup *memcg,
1524 return memcg_page_state(memcg, item) * memcg_page_state_unit(item);
1527 static char *memory_stat_format(struct mem_cgroup *memcg)
1532 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1537 * Provide statistics on the state of the memory subsystem as
1538 * well as cumulative event counters that show past behavior.
1540 * This list is ordered following a combination of these gradients:
1541 * 1) generic big picture -> specifics and details
1542 * 2) reflecting userspace activity -> reflecting kernel heuristics
1544 * Current memory state:
1546 cgroup_rstat_flush(memcg->css.cgroup);
1548 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1551 size = memcg_page_state_output(memcg, memory_stats[i].idx);
1552 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1554 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1555 size += memcg_page_state_output(memcg,
1556 NR_SLAB_RECLAIMABLE_B);
1557 seq_buf_printf(&s, "slab %llu\n", size);
1561 /* Accumulated memory events */
1563 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1564 memcg_events(memcg, PGFAULT));
1565 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1566 memcg_events(memcg, PGMAJFAULT));
1567 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1568 memcg_events(memcg, PGREFILL));
1569 seq_buf_printf(&s, "pgscan %lu\n",
1570 memcg_events(memcg, PGSCAN_KSWAPD) +
1571 memcg_events(memcg, PGSCAN_DIRECT));
1572 seq_buf_printf(&s, "pgsteal %lu\n",
1573 memcg_events(memcg, PGSTEAL_KSWAPD) +
1574 memcg_events(memcg, PGSTEAL_DIRECT));
1575 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1576 memcg_events(memcg, PGACTIVATE));
1577 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1578 memcg_events(memcg, PGDEACTIVATE));
1579 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1580 memcg_events(memcg, PGLAZYFREE));
1581 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1582 memcg_events(memcg, PGLAZYFREED));
1584 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1585 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1586 memcg_events(memcg, THP_FAULT_ALLOC));
1587 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1588 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1589 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1591 /* The above should easily fit into one page */
1592 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1597 #define K(x) ((x) << (PAGE_SHIFT-10))
1599 * mem_cgroup_print_oom_context: Print OOM information relevant to
1600 * memory controller.
1601 * @memcg: The memory cgroup that went over limit
1602 * @p: Task that is going to be killed
1604 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1607 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1612 pr_cont(",oom_memcg=");
1613 pr_cont_cgroup_path(memcg->css.cgroup);
1615 pr_cont(",global_oom");
1617 pr_cont(",task_memcg=");
1618 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1624 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1625 * memory controller.
1626 * @memcg: The memory cgroup that went over limit
1628 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1632 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1633 K((u64)page_counter_read(&memcg->memory)),
1634 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1635 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1636 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1637 K((u64)page_counter_read(&memcg->swap)),
1638 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1640 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1641 K((u64)page_counter_read(&memcg->memsw)),
1642 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1643 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1644 K((u64)page_counter_read(&memcg->kmem)),
1645 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1648 pr_info("Memory cgroup stats for ");
1649 pr_cont_cgroup_path(memcg->css.cgroup);
1651 buf = memory_stat_format(memcg);
1659 * Return the memory (and swap, if configured) limit for a memcg.
1661 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1663 unsigned long max = READ_ONCE(memcg->memory.max);
1665 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1666 if (mem_cgroup_swappiness(memcg))
1667 max += min(READ_ONCE(memcg->swap.max),
1668 (unsigned long)total_swap_pages);
1670 if (mem_cgroup_swappiness(memcg)) {
1671 /* Calculate swap excess capacity from memsw limit */
1672 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1674 max += min(swap, (unsigned long)total_swap_pages);
1680 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1682 return page_counter_read(&memcg->memory);
1685 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1688 struct oom_control oc = {
1692 .gfp_mask = gfp_mask,
1697 if (mutex_lock_killable(&oom_lock))
1700 if (mem_cgroup_margin(memcg) >= (1 << order))
1704 * A few threads which were not waiting at mutex_lock_killable() can
1705 * fail to bail out. Therefore, check again after holding oom_lock.
1707 ret = should_force_charge() || out_of_memory(&oc);
1710 mutex_unlock(&oom_lock);
1714 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1717 unsigned long *total_scanned)
1719 struct mem_cgroup *victim = NULL;
1722 unsigned long excess;
1723 unsigned long nr_scanned;
1724 struct mem_cgroup_reclaim_cookie reclaim = {
1728 excess = soft_limit_excess(root_memcg);
1731 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1736 * If we have not been able to reclaim
1737 * anything, it might because there are
1738 * no reclaimable pages under this hierarchy
1743 * We want to do more targeted reclaim.
1744 * excess >> 2 is not to excessive so as to
1745 * reclaim too much, nor too less that we keep
1746 * coming back to reclaim from this cgroup
1748 if (total >= (excess >> 2) ||
1749 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1754 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1755 pgdat, &nr_scanned);
1756 *total_scanned += nr_scanned;
1757 if (!soft_limit_excess(root_memcg))
1760 mem_cgroup_iter_break(root_memcg, victim);
1764 #ifdef CONFIG_LOCKDEP
1765 static struct lockdep_map memcg_oom_lock_dep_map = {
1766 .name = "memcg_oom_lock",
1770 static DEFINE_SPINLOCK(memcg_oom_lock);
1773 * Check OOM-Killer is already running under our hierarchy.
1774 * If someone is running, return false.
1776 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter, *failed = NULL;
1780 spin_lock(&memcg_oom_lock);
1782 for_each_mem_cgroup_tree(iter, memcg) {
1783 if (iter->oom_lock) {
1785 * this subtree of our hierarchy is already locked
1786 * so we cannot give a lock.
1789 mem_cgroup_iter_break(memcg, iter);
1792 iter->oom_lock = true;
1797 * OK, we failed to lock the whole subtree so we have
1798 * to clean up what we set up to the failing subtree
1800 for_each_mem_cgroup_tree(iter, memcg) {
1801 if (iter == failed) {
1802 mem_cgroup_iter_break(memcg, iter);
1805 iter->oom_lock = false;
1808 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1810 spin_unlock(&memcg_oom_lock);
1815 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1817 struct mem_cgroup *iter;
1819 spin_lock(&memcg_oom_lock);
1820 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1821 for_each_mem_cgroup_tree(iter, memcg)
1822 iter->oom_lock = false;
1823 spin_unlock(&memcg_oom_lock);
1826 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1828 struct mem_cgroup *iter;
1830 spin_lock(&memcg_oom_lock);
1831 for_each_mem_cgroup_tree(iter, memcg)
1833 spin_unlock(&memcg_oom_lock);
1836 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1841 * Be careful about under_oom underflows becase a child memcg
1842 * could have been added after mem_cgroup_mark_under_oom.
1844 spin_lock(&memcg_oom_lock);
1845 for_each_mem_cgroup_tree(iter, memcg)
1846 if (iter->under_oom > 0)
1848 spin_unlock(&memcg_oom_lock);
1851 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1853 struct oom_wait_info {
1854 struct mem_cgroup *memcg;
1855 wait_queue_entry_t wait;
1858 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1859 unsigned mode, int sync, void *arg)
1861 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1862 struct mem_cgroup *oom_wait_memcg;
1863 struct oom_wait_info *oom_wait_info;
1865 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1866 oom_wait_memcg = oom_wait_info->memcg;
1868 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1869 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1871 return autoremove_wake_function(wait, mode, sync, arg);
1874 static void memcg_oom_recover(struct mem_cgroup *memcg)
1877 * For the following lockless ->under_oom test, the only required
1878 * guarantee is that it must see the state asserted by an OOM when
1879 * this function is called as a result of userland actions
1880 * triggered by the notification of the OOM. This is trivially
1881 * achieved by invoking mem_cgroup_mark_under_oom() before
1882 * triggering notification.
1884 if (memcg && memcg->under_oom)
1885 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1895 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1897 enum oom_status ret;
1900 if (order > PAGE_ALLOC_COSTLY_ORDER)
1903 memcg_memory_event(memcg, MEMCG_OOM);
1906 * We are in the middle of the charge context here, so we
1907 * don't want to block when potentially sitting on a callstack
1908 * that holds all kinds of filesystem and mm locks.
1910 * cgroup1 allows disabling the OOM killer and waiting for outside
1911 * handling until the charge can succeed; remember the context and put
1912 * the task to sleep at the end of the page fault when all locks are
1915 * On the other hand, in-kernel OOM killer allows for an async victim
1916 * memory reclaim (oom_reaper) and that means that we are not solely
1917 * relying on the oom victim to make a forward progress and we can
1918 * invoke the oom killer here.
1920 * Please note that mem_cgroup_out_of_memory might fail to find a
1921 * victim and then we have to bail out from the charge path.
1923 if (memcg->oom_kill_disable) {
1924 if (!current->in_user_fault)
1926 css_get(&memcg->css);
1927 current->memcg_in_oom = memcg;
1928 current->memcg_oom_gfp_mask = mask;
1929 current->memcg_oom_order = order;
1934 mem_cgroup_mark_under_oom(memcg);
1936 locked = mem_cgroup_oom_trylock(memcg);
1939 mem_cgroup_oom_notify(memcg);
1941 mem_cgroup_unmark_under_oom(memcg);
1942 if (mem_cgroup_out_of_memory(memcg, mask, order))
1948 mem_cgroup_oom_unlock(memcg);
1954 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1955 * @handle: actually kill/wait or just clean up the OOM state
1957 * This has to be called at the end of a page fault if the memcg OOM
1958 * handler was enabled.
1960 * Memcg supports userspace OOM handling where failed allocations must
1961 * sleep on a waitqueue until the userspace task resolves the
1962 * situation. Sleeping directly in the charge context with all kinds
1963 * of locks held is not a good idea, instead we remember an OOM state
1964 * in the task and mem_cgroup_oom_synchronize() has to be called at
1965 * the end of the page fault to complete the OOM handling.
1967 * Returns %true if an ongoing memcg OOM situation was detected and
1968 * completed, %false otherwise.
1970 bool mem_cgroup_oom_synchronize(bool handle)
1972 struct mem_cgroup *memcg = current->memcg_in_oom;
1973 struct oom_wait_info owait;
1976 /* OOM is global, do not handle */
1983 owait.memcg = memcg;
1984 owait.wait.flags = 0;
1985 owait.wait.func = memcg_oom_wake_function;
1986 owait.wait.private = current;
1987 INIT_LIST_HEAD(&owait.wait.entry);
1989 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1990 mem_cgroup_mark_under_oom(memcg);
1992 locked = mem_cgroup_oom_trylock(memcg);
1995 mem_cgroup_oom_notify(memcg);
1997 if (locked && !memcg->oom_kill_disable) {
1998 mem_cgroup_unmark_under_oom(memcg);
1999 finish_wait(&memcg_oom_waitq, &owait.wait);
2000 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2001 current->memcg_oom_order);
2004 mem_cgroup_unmark_under_oom(memcg);
2005 finish_wait(&memcg_oom_waitq, &owait.wait);
2009 mem_cgroup_oom_unlock(memcg);
2011 * There is no guarantee that an OOM-lock contender
2012 * sees the wakeups triggered by the OOM kill
2013 * uncharges. Wake any sleepers explicitely.
2015 memcg_oom_recover(memcg);
2018 current->memcg_in_oom = NULL;
2019 css_put(&memcg->css);
2024 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2025 * @victim: task to be killed by the OOM killer
2026 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2028 * Returns a pointer to a memory cgroup, which has to be cleaned up
2029 * by killing all belonging OOM-killable tasks.
2031 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2033 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2034 struct mem_cgroup *oom_domain)
2036 struct mem_cgroup *oom_group = NULL;
2037 struct mem_cgroup *memcg;
2039 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2043 oom_domain = root_mem_cgroup;
2047 memcg = mem_cgroup_from_task(victim);
2048 if (memcg == root_mem_cgroup)
2052 * If the victim task has been asynchronously moved to a different
2053 * memory cgroup, we might end up killing tasks outside oom_domain.
2054 * In this case it's better to ignore memory.group.oom.
2056 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2060 * Traverse the memory cgroup hierarchy from the victim task's
2061 * cgroup up to the OOMing cgroup (or root) to find the
2062 * highest-level memory cgroup with oom.group set.
2064 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2065 if (memcg->oom_group)
2068 if (memcg == oom_domain)
2073 css_get(&oom_group->css);
2080 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2082 pr_info("Tasks in ");
2083 pr_cont_cgroup_path(memcg->css.cgroup);
2084 pr_cont(" are going to be killed due to memory.oom.group set\n");
2088 * lock_page_memcg - lock a page and memcg binding
2091 * This function protects unlocked LRU pages from being moved to
2094 * It ensures lifetime of the locked memcg. Caller is responsible
2095 * for the lifetime of the page.
2097 void lock_page_memcg(struct page *page)
2099 struct page *head = compound_head(page); /* rmap on tail pages */
2100 struct mem_cgroup *memcg;
2101 unsigned long flags;
2104 * The RCU lock is held throughout the transaction. The fast
2105 * path can get away without acquiring the memcg->move_lock
2106 * because page moving starts with an RCU grace period.
2110 if (mem_cgroup_disabled())
2113 memcg = page_memcg(head);
2114 if (unlikely(!memcg))
2117 #ifdef CONFIG_PROVE_LOCKING
2118 local_irq_save(flags);
2119 might_lock(&memcg->move_lock);
2120 local_irq_restore(flags);
2123 if (atomic_read(&memcg->moving_account) <= 0)
2126 spin_lock_irqsave(&memcg->move_lock, flags);
2127 if (memcg != page_memcg(head)) {
2128 spin_unlock_irqrestore(&memcg->move_lock, flags);
2133 * When charge migration first begins, we can have multiple
2134 * critical sections holding the fast-path RCU lock and one
2135 * holding the slowpath move_lock. Track the task who has the
2136 * move_lock for unlock_page_memcg().
2138 memcg->move_lock_task = current;
2139 memcg->move_lock_flags = flags;
2141 EXPORT_SYMBOL(lock_page_memcg);
2143 static void __unlock_page_memcg(struct mem_cgroup *memcg)
2145 if (memcg && memcg->move_lock_task == current) {
2146 unsigned long flags = memcg->move_lock_flags;
2148 memcg->move_lock_task = NULL;
2149 memcg->move_lock_flags = 0;
2151 spin_unlock_irqrestore(&memcg->move_lock, flags);
2158 * unlock_page_memcg - unlock a page and memcg binding
2161 void unlock_page_memcg(struct page *page)
2163 struct page *head = compound_head(page);
2165 __unlock_page_memcg(page_memcg(head));
2167 EXPORT_SYMBOL(unlock_page_memcg);
2169 struct memcg_stock_pcp {
2170 struct mem_cgroup *cached; /* this never be root cgroup */
2171 unsigned int nr_pages;
2173 #ifdef CONFIG_MEMCG_KMEM
2174 struct obj_cgroup *cached_objcg;
2175 unsigned int nr_bytes;
2178 struct work_struct work;
2179 unsigned long flags;
2180 #define FLUSHING_CACHED_CHARGE 0
2182 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2183 static DEFINE_MUTEX(percpu_charge_mutex);
2185 #ifdef CONFIG_MEMCG_KMEM
2186 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2187 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2188 struct mem_cgroup *root_memcg);
2191 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2194 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2195 struct mem_cgroup *root_memcg)
2202 * consume_stock: Try to consume stocked charge on this cpu.
2203 * @memcg: memcg to consume from.
2204 * @nr_pages: how many pages to charge.
2206 * The charges will only happen if @memcg matches the current cpu's memcg
2207 * stock, and at least @nr_pages are available in that stock. Failure to
2208 * service an allocation will refill the stock.
2210 * returns true if successful, false otherwise.
2212 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2214 struct memcg_stock_pcp *stock;
2215 unsigned long flags;
2218 if (nr_pages > MEMCG_CHARGE_BATCH)
2221 local_irq_save(flags);
2223 stock = this_cpu_ptr(&memcg_stock);
2224 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2225 stock->nr_pages -= nr_pages;
2229 local_irq_restore(flags);
2235 * Returns stocks cached in percpu and reset cached information.
2237 static void drain_stock(struct memcg_stock_pcp *stock)
2239 struct mem_cgroup *old = stock->cached;
2244 if (stock->nr_pages) {
2245 page_counter_uncharge(&old->memory, stock->nr_pages);
2246 if (do_memsw_account())
2247 page_counter_uncharge(&old->memsw, stock->nr_pages);
2248 stock->nr_pages = 0;
2252 stock->cached = NULL;
2255 static void drain_local_stock(struct work_struct *dummy)
2257 struct memcg_stock_pcp *stock;
2258 unsigned long flags;
2261 * The only protection from memory hotplug vs. drain_stock races is
2262 * that we always operate on local CPU stock here with IRQ disabled
2264 local_irq_save(flags);
2266 stock = this_cpu_ptr(&memcg_stock);
2267 drain_obj_stock(stock);
2269 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2271 local_irq_restore(flags);
2275 * Cache charges(val) to local per_cpu area.
2276 * This will be consumed by consume_stock() function, later.
2278 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2280 struct memcg_stock_pcp *stock;
2281 unsigned long flags;
2283 local_irq_save(flags);
2285 stock = this_cpu_ptr(&memcg_stock);
2286 if (stock->cached != memcg) { /* reset if necessary */
2288 css_get(&memcg->css);
2289 stock->cached = memcg;
2291 stock->nr_pages += nr_pages;
2293 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2296 local_irq_restore(flags);
2300 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2301 * of the hierarchy under it.
2303 static void drain_all_stock(struct mem_cgroup *root_memcg)
2307 /* If someone's already draining, avoid adding running more workers. */
2308 if (!mutex_trylock(&percpu_charge_mutex))
2311 * Notify other cpus that system-wide "drain" is running
2312 * We do not care about races with the cpu hotplug because cpu down
2313 * as well as workers from this path always operate on the local
2314 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2317 for_each_online_cpu(cpu) {
2318 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2319 struct mem_cgroup *memcg;
2323 memcg = stock->cached;
2324 if (memcg && stock->nr_pages &&
2325 mem_cgroup_is_descendant(memcg, root_memcg))
2327 if (obj_stock_flush_required(stock, root_memcg))
2332 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2334 drain_local_stock(&stock->work);
2336 schedule_work_on(cpu, &stock->work);
2340 mutex_unlock(&percpu_charge_mutex);
2343 static void memcg_flush_lruvec_page_state(struct mem_cgroup *memcg, int cpu)
2347 for_each_node(nid) {
2348 struct mem_cgroup_per_node *pn = memcg->nodeinfo[nid];
2349 unsigned long stat[NR_VM_NODE_STAT_ITEMS];
2350 struct batched_lruvec_stat *lstatc;
2353 lstatc = per_cpu_ptr(pn->lruvec_stat_cpu, cpu);
2354 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++) {
2355 stat[i] = lstatc->count[i];
2356 lstatc->count[i] = 0;
2360 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
2361 atomic_long_add(stat[i], &pn->lruvec_stat[i]);
2362 } while ((pn = parent_nodeinfo(pn, nid)));
2366 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2368 struct memcg_stock_pcp *stock;
2369 struct mem_cgroup *memcg;
2371 stock = &per_cpu(memcg_stock, cpu);
2374 for_each_mem_cgroup(memcg)
2375 memcg_flush_lruvec_page_state(memcg, cpu);
2380 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2381 unsigned int nr_pages,
2384 unsigned long nr_reclaimed = 0;
2387 unsigned long pflags;
2389 if (page_counter_read(&memcg->memory) <=
2390 READ_ONCE(memcg->memory.high))
2393 memcg_memory_event(memcg, MEMCG_HIGH);
2395 psi_memstall_enter(&pflags);
2396 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2398 psi_memstall_leave(&pflags);
2399 } while ((memcg = parent_mem_cgroup(memcg)) &&
2400 !mem_cgroup_is_root(memcg));
2402 return nr_reclaimed;
2405 static void high_work_func(struct work_struct *work)
2407 struct mem_cgroup *memcg;
2409 memcg = container_of(work, struct mem_cgroup, high_work);
2410 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2414 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2415 * enough to still cause a significant slowdown in most cases, while still
2416 * allowing diagnostics and tracing to proceed without becoming stuck.
2418 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2421 * When calculating the delay, we use these either side of the exponentiation to
2422 * maintain precision and scale to a reasonable number of jiffies (see the table
2425 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2426 * overage ratio to a delay.
2427 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2428 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2429 * to produce a reasonable delay curve.
2431 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2432 * reasonable delay curve compared to precision-adjusted overage, not
2433 * penalising heavily at first, but still making sure that growth beyond the
2434 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2435 * example, with a high of 100 megabytes:
2437 * +-------+------------------------+
2438 * | usage | time to allocate in ms |
2439 * +-------+------------------------+
2461 * +-------+------------------------+
2463 #define MEMCG_DELAY_PRECISION_SHIFT 20
2464 #define MEMCG_DELAY_SCALING_SHIFT 14
2466 static u64 calculate_overage(unsigned long usage, unsigned long high)
2474 * Prevent division by 0 in overage calculation by acting as if
2475 * it was a threshold of 1 page
2477 high = max(high, 1UL);
2479 overage = usage - high;
2480 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2481 return div64_u64(overage, high);
2484 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2486 u64 overage, max_overage = 0;
2489 overage = calculate_overage(page_counter_read(&memcg->memory),
2490 READ_ONCE(memcg->memory.high));
2491 max_overage = max(overage, max_overage);
2492 } while ((memcg = parent_mem_cgroup(memcg)) &&
2493 !mem_cgroup_is_root(memcg));
2498 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2500 u64 overage, max_overage = 0;
2503 overage = calculate_overage(page_counter_read(&memcg->swap),
2504 READ_ONCE(memcg->swap.high));
2506 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2507 max_overage = max(overage, max_overage);
2508 } while ((memcg = parent_mem_cgroup(memcg)) &&
2509 !mem_cgroup_is_root(memcg));
2515 * Get the number of jiffies that we should penalise a mischievous cgroup which
2516 * is exceeding its memory.high by checking both it and its ancestors.
2518 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2519 unsigned int nr_pages,
2522 unsigned long penalty_jiffies;
2528 * We use overage compared to memory.high to calculate the number of
2529 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2530 * fairly lenient on small overages, and increasingly harsh when the
2531 * memcg in question makes it clear that it has no intention of stopping
2532 * its crazy behaviour, so we exponentially increase the delay based on
2535 penalty_jiffies = max_overage * max_overage * HZ;
2536 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2537 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2540 * Factor in the task's own contribution to the overage, such that four
2541 * N-sized allocations are throttled approximately the same as one
2542 * 4N-sized allocation.
2544 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2545 * larger the current charge patch is than that.
2547 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2551 * Scheduled by try_charge() to be executed from the userland return path
2552 * and reclaims memory over the high limit.
2554 void mem_cgroup_handle_over_high(void)
2556 unsigned long penalty_jiffies;
2557 unsigned long pflags;
2558 unsigned long nr_reclaimed;
2559 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2560 int nr_retries = MAX_RECLAIM_RETRIES;
2561 struct mem_cgroup *memcg;
2562 bool in_retry = false;
2564 if (likely(!nr_pages))
2567 memcg = get_mem_cgroup_from_mm(current->mm);
2568 current->memcg_nr_pages_over_high = 0;
2572 * The allocating task should reclaim at least the batch size, but for
2573 * subsequent retries we only want to do what's necessary to prevent oom
2574 * or breaching resource isolation.
2576 * This is distinct from memory.max or page allocator behaviour because
2577 * memory.high is currently batched, whereas memory.max and the page
2578 * allocator run every time an allocation is made.
2580 nr_reclaimed = reclaim_high(memcg,
2581 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2585 * memory.high is breached and reclaim is unable to keep up. Throttle
2586 * allocators proactively to slow down excessive growth.
2588 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2589 mem_find_max_overage(memcg));
2591 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2592 swap_find_max_overage(memcg));
2595 * Clamp the max delay per usermode return so as to still keep the
2596 * application moving forwards and also permit diagnostics, albeit
2599 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2602 * Don't sleep if the amount of jiffies this memcg owes us is so low
2603 * that it's not even worth doing, in an attempt to be nice to those who
2604 * go only a small amount over their memory.high value and maybe haven't
2605 * been aggressively reclaimed enough yet.
2607 if (penalty_jiffies <= HZ / 100)
2611 * If reclaim is making forward progress but we're still over
2612 * memory.high, we want to encourage that rather than doing allocator
2615 if (nr_reclaimed || nr_retries--) {
2621 * If we exit early, we're guaranteed to die (since
2622 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2623 * need to account for any ill-begotten jiffies to pay them off later.
2625 psi_memstall_enter(&pflags);
2626 schedule_timeout_killable(penalty_jiffies);
2627 psi_memstall_leave(&pflags);
2630 css_put(&memcg->css);
2633 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2634 unsigned int nr_pages)
2636 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2637 int nr_retries = MAX_RECLAIM_RETRIES;
2638 struct mem_cgroup *mem_over_limit;
2639 struct page_counter *counter;
2640 enum oom_status oom_status;
2641 unsigned long nr_reclaimed;
2642 bool may_swap = true;
2643 bool drained = false;
2644 unsigned long pflags;
2646 if (mem_cgroup_is_root(memcg))
2649 if (consume_stock(memcg, nr_pages))
2652 if (!do_memsw_account() ||
2653 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2654 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2656 if (do_memsw_account())
2657 page_counter_uncharge(&memcg->memsw, batch);
2658 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2660 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2664 if (batch > nr_pages) {
2670 * Memcg doesn't have a dedicated reserve for atomic
2671 * allocations. But like the global atomic pool, we need to
2672 * put the burden of reclaim on regular allocation requests
2673 * and let these go through as privileged allocations.
2675 if (gfp_mask & __GFP_ATOMIC)
2679 * Unlike in global OOM situations, memcg is not in a physical
2680 * memory shortage. Allow dying and OOM-killed tasks to
2681 * bypass the last charges so that they can exit quickly and
2682 * free their memory.
2684 if (unlikely(should_force_charge()))
2688 * Prevent unbounded recursion when reclaim operations need to
2689 * allocate memory. This might exceed the limits temporarily,
2690 * but we prefer facilitating memory reclaim and getting back
2691 * under the limit over triggering OOM kills in these cases.
2693 if (unlikely(current->flags & PF_MEMALLOC))
2696 if (unlikely(task_in_memcg_oom(current)))
2699 if (!gfpflags_allow_blocking(gfp_mask))
2702 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2704 psi_memstall_enter(&pflags);
2705 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2706 gfp_mask, may_swap);
2707 psi_memstall_leave(&pflags);
2709 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2713 drain_all_stock(mem_over_limit);
2718 if (gfp_mask & __GFP_NORETRY)
2721 * Even though the limit is exceeded at this point, reclaim
2722 * may have been able to free some pages. Retry the charge
2723 * before killing the task.
2725 * Only for regular pages, though: huge pages are rather
2726 * unlikely to succeed so close to the limit, and we fall back
2727 * to regular pages anyway in case of failure.
2729 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2732 * At task move, charge accounts can be doubly counted. So, it's
2733 * better to wait until the end of task_move if something is going on.
2735 if (mem_cgroup_wait_acct_move(mem_over_limit))
2741 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2744 if (fatal_signal_pending(current))
2748 * keep retrying as long as the memcg oom killer is able to make
2749 * a forward progress or bypass the charge if the oom killer
2750 * couldn't make any progress.
2752 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2753 get_order(nr_pages * PAGE_SIZE));
2754 switch (oom_status) {
2756 nr_retries = MAX_RECLAIM_RETRIES;
2764 if (!(gfp_mask & __GFP_NOFAIL))
2768 * The allocation either can't fail or will lead to more memory
2769 * being freed very soon. Allow memory usage go over the limit
2770 * temporarily by force charging it.
2772 page_counter_charge(&memcg->memory, nr_pages);
2773 if (do_memsw_account())
2774 page_counter_charge(&memcg->memsw, nr_pages);
2779 if (batch > nr_pages)
2780 refill_stock(memcg, batch - nr_pages);
2783 * If the hierarchy is above the normal consumption range, schedule
2784 * reclaim on returning to userland. We can perform reclaim here
2785 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2786 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2787 * not recorded as it most likely matches current's and won't
2788 * change in the meantime. As high limit is checked again before
2789 * reclaim, the cost of mismatch is negligible.
2792 bool mem_high, swap_high;
2794 mem_high = page_counter_read(&memcg->memory) >
2795 READ_ONCE(memcg->memory.high);
2796 swap_high = page_counter_read(&memcg->swap) >
2797 READ_ONCE(memcg->swap.high);
2799 /* Don't bother a random interrupted task */
2800 if (in_interrupt()) {
2802 schedule_work(&memcg->high_work);
2808 if (mem_high || swap_high) {
2810 * The allocating tasks in this cgroup will need to do
2811 * reclaim or be throttled to prevent further growth
2812 * of the memory or swap footprints.
2814 * Target some best-effort fairness between the tasks,
2815 * and distribute reclaim work and delay penalties
2816 * based on how much each task is actually allocating.
2818 current->memcg_nr_pages_over_high += batch;
2819 set_notify_resume(current);
2822 } while ((memcg = parent_mem_cgroup(memcg)));
2827 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2828 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2830 if (mem_cgroup_is_root(memcg))
2833 page_counter_uncharge(&memcg->memory, nr_pages);
2834 if (do_memsw_account())
2835 page_counter_uncharge(&memcg->memsw, nr_pages);
2839 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2841 VM_BUG_ON_PAGE(page_memcg(page), page);
2843 * Any of the following ensures page's memcg stability:
2847 * - lock_page_memcg()
2848 * - exclusive reference
2850 page->memcg_data = (unsigned long)memcg;
2853 static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
2855 struct mem_cgroup *memcg;
2859 memcg = obj_cgroup_memcg(objcg);
2860 if (unlikely(!css_tryget(&memcg->css)))
2867 #ifdef CONFIG_MEMCG_KMEM
2868 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2869 gfp_t gfp, bool new_page)
2871 unsigned int objects = objs_per_slab_page(s, page);
2872 unsigned long memcg_data;
2875 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2880 memcg_data = (unsigned long) vec | MEMCG_DATA_OBJCGS;
2883 * If the slab page is brand new and nobody can yet access
2884 * it's memcg_data, no synchronization is required and
2885 * memcg_data can be simply assigned.
2887 page->memcg_data = memcg_data;
2888 } else if (cmpxchg(&page->memcg_data, 0, memcg_data)) {
2890 * If the slab page is already in use, somebody can allocate
2891 * and assign obj_cgroups in parallel. In this case the existing
2892 * objcg vector should be reused.
2898 kmemleak_not_leak(vec);
2903 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2905 * A passed kernel object can be a slab object or a generic kernel page, so
2906 * different mechanisms for getting the memory cgroup pointer should be used.
2907 * In certain cases (e.g. kernel stacks or large kmallocs with SLUB) the caller
2908 * can not know for sure how the kernel object is implemented.
2909 * mem_cgroup_from_obj() can be safely used in such cases.
2911 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2912 * cgroup_mutex, etc.
2914 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2918 if (mem_cgroup_disabled())
2921 page = virt_to_head_page(p);
2924 * Slab objects are accounted individually, not per-page.
2925 * Memcg membership data for each individual object is saved in
2926 * the page->obj_cgroups.
2928 if (page_objcgs_check(page)) {
2929 struct obj_cgroup *objcg;
2932 off = obj_to_index(page->slab_cache, page, p);
2933 objcg = page_objcgs(page)[off];
2935 return obj_cgroup_memcg(objcg);
2941 * page_memcg_check() is used here, because page_has_obj_cgroups()
2942 * check above could fail because the object cgroups vector wasn't set
2943 * at that moment, but it can be set concurrently.
2944 * page_memcg_check(page) will guarantee that a proper memory
2945 * cgroup pointer or NULL will be returned.
2947 return page_memcg_check(page);
2950 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2952 struct obj_cgroup *objcg = NULL;
2953 struct mem_cgroup *memcg;
2955 if (memcg_kmem_bypass())
2959 if (unlikely(active_memcg()))
2960 memcg = active_memcg();
2962 memcg = mem_cgroup_from_task(current);
2964 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2965 objcg = rcu_dereference(memcg->objcg);
2966 if (objcg && obj_cgroup_tryget(objcg))
2975 static int memcg_alloc_cache_id(void)
2980 id = ida_simple_get(&memcg_cache_ida,
2981 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2985 if (id < memcg_nr_cache_ids)
2989 * There's no space for the new id in memcg_caches arrays,
2990 * so we have to grow them.
2992 down_write(&memcg_cache_ids_sem);
2994 size = 2 * (id + 1);
2995 if (size < MEMCG_CACHES_MIN_SIZE)
2996 size = MEMCG_CACHES_MIN_SIZE;
2997 else if (size > MEMCG_CACHES_MAX_SIZE)
2998 size = MEMCG_CACHES_MAX_SIZE;
3000 err = memcg_update_all_list_lrus(size);
3002 memcg_nr_cache_ids = size;
3004 up_write(&memcg_cache_ids_sem);
3007 ida_simple_remove(&memcg_cache_ida, id);
3013 static void memcg_free_cache_id(int id)
3015 ida_simple_remove(&memcg_cache_ida, id);
3018 static void obj_cgroup_uncharge_pages(struct obj_cgroup *objcg,
3019 unsigned int nr_pages)
3021 struct mem_cgroup *memcg;
3023 memcg = get_mem_cgroup_from_objcg(objcg);
3024 __memcg_kmem_uncharge(memcg, nr_pages);
3025 css_put(&memcg->css);
3028 static int obj_cgroup_charge_pages(struct obj_cgroup *objcg, gfp_t gfp,
3029 unsigned int nr_pages)
3031 struct mem_cgroup *memcg;
3034 memcg = get_mem_cgroup_from_objcg(objcg);
3035 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3036 css_put(&memcg->css);
3042 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3043 * @memcg: memory cgroup to charge
3044 * @gfp: reclaim mode
3045 * @nr_pages: number of pages to charge
3047 * Returns 0 on success, an error code on failure.
3049 static int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3050 unsigned int nr_pages)
3052 struct page_counter *counter;
3055 ret = try_charge(memcg, gfp, nr_pages);
3059 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3060 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3063 * Enforce __GFP_NOFAIL allocation because callers are not
3064 * prepared to see failures and likely do not have any failure
3067 if (gfp & __GFP_NOFAIL) {
3068 page_counter_charge(&memcg->kmem, nr_pages);
3071 cancel_charge(memcg, nr_pages);
3078 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3079 * @memcg: memcg to uncharge
3080 * @nr_pages: number of pages to uncharge
3082 static void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3084 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3085 page_counter_uncharge(&memcg->kmem, nr_pages);
3087 refill_stock(memcg, nr_pages);
3091 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3092 * @page: page to charge
3093 * @gfp: reclaim mode
3094 * @order: allocation order
3096 * Returns 0 on success, an error code on failure.
3098 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3100 struct obj_cgroup *objcg;
3103 objcg = get_obj_cgroup_from_current();
3105 ret = obj_cgroup_charge_pages(objcg, gfp, 1 << order);
3107 page->memcg_data = (unsigned long)objcg |
3111 obj_cgroup_put(objcg);
3117 * __memcg_kmem_uncharge_page: uncharge a kmem page
3118 * @page: page to uncharge
3119 * @order: allocation order
3121 void __memcg_kmem_uncharge_page(struct page *page, int order)
3123 struct obj_cgroup *objcg;
3124 unsigned int nr_pages = 1 << order;
3126 if (!PageMemcgKmem(page))
3129 objcg = __page_objcg(page);
3130 obj_cgroup_uncharge_pages(objcg, nr_pages);
3131 page->memcg_data = 0;
3132 obj_cgroup_put(objcg);
3135 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3137 struct memcg_stock_pcp *stock;
3138 unsigned long flags;
3141 local_irq_save(flags);
3143 stock = this_cpu_ptr(&memcg_stock);
3144 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3145 stock->nr_bytes -= nr_bytes;
3149 local_irq_restore(flags);
3154 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3156 struct obj_cgroup *old = stock->cached_objcg;
3161 if (stock->nr_bytes) {
3162 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3163 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3166 obj_cgroup_uncharge_pages(old, nr_pages);
3169 * The leftover is flushed to the centralized per-memcg value.
3170 * On the next attempt to refill obj stock it will be moved
3171 * to a per-cpu stock (probably, on an other CPU), see
3172 * refill_obj_stock().
3174 * How often it's flushed is a trade-off between the memory
3175 * limit enforcement accuracy and potential CPU contention,
3176 * so it might be changed in the future.
3178 atomic_add(nr_bytes, &old->nr_charged_bytes);
3179 stock->nr_bytes = 0;
3182 obj_cgroup_put(old);
3183 stock->cached_objcg = NULL;
3186 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3187 struct mem_cgroup *root_memcg)
3189 struct mem_cgroup *memcg;
3191 if (stock->cached_objcg) {
3192 memcg = obj_cgroup_memcg(stock->cached_objcg);
3193 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3200 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3202 struct memcg_stock_pcp *stock;
3203 unsigned long flags;
3205 local_irq_save(flags);
3207 stock = this_cpu_ptr(&memcg_stock);
3208 if (stock->cached_objcg != objcg) { /* reset if necessary */
3209 drain_obj_stock(stock);
3210 obj_cgroup_get(objcg);
3211 stock->cached_objcg = objcg;
3212 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3214 stock->nr_bytes += nr_bytes;
3216 if (stock->nr_bytes > PAGE_SIZE)
3217 drain_obj_stock(stock);
3219 local_irq_restore(flags);
3222 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3224 unsigned int nr_pages, nr_bytes;
3227 if (consume_obj_stock(objcg, size))
3231 * In theory, memcg->nr_charged_bytes can have enough
3232 * pre-charged bytes to satisfy the allocation. However,
3233 * flushing memcg->nr_charged_bytes requires two atomic
3234 * operations, and memcg->nr_charged_bytes can't be big,
3235 * so it's better to ignore it and try grab some new pages.
3236 * memcg->nr_charged_bytes will be flushed in
3237 * refill_obj_stock(), called from this function or
3238 * independently later.
3240 nr_pages = size >> PAGE_SHIFT;
3241 nr_bytes = size & (PAGE_SIZE - 1);
3246 ret = obj_cgroup_charge_pages(objcg, gfp, nr_pages);
3247 if (!ret && nr_bytes)
3248 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3253 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3255 refill_obj_stock(objcg, size);
3258 #endif /* CONFIG_MEMCG_KMEM */
3261 * Because page_memcg(head) is not set on tails, set it now.
3263 void split_page_memcg(struct page *head, unsigned int nr)
3265 struct mem_cgroup *memcg = page_memcg(head);
3268 if (mem_cgroup_disabled() || !memcg)
3271 for (i = 1; i < nr; i++)
3272 head[i].memcg_data = head->memcg_data;
3274 if (PageMemcgKmem(head))
3275 obj_cgroup_get_many(__page_objcg(head), nr - 1);
3277 css_get_many(&memcg->css, nr - 1);
3280 #ifdef CONFIG_MEMCG_SWAP
3282 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3283 * @entry: swap entry to be moved
3284 * @from: mem_cgroup which the entry is moved from
3285 * @to: mem_cgroup which the entry is moved to
3287 * It succeeds only when the swap_cgroup's record for this entry is the same
3288 * as the mem_cgroup's id of @from.
3290 * Returns 0 on success, -EINVAL on failure.
3292 * The caller must have charged to @to, IOW, called page_counter_charge() about
3293 * both res and memsw, and called css_get().
3295 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3296 struct mem_cgroup *from, struct mem_cgroup *to)
3298 unsigned short old_id, new_id;
3300 old_id = mem_cgroup_id(from);
3301 new_id = mem_cgroup_id(to);
3303 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3304 mod_memcg_state(from, MEMCG_SWAP, -1);
3305 mod_memcg_state(to, MEMCG_SWAP, 1);
3311 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3312 struct mem_cgroup *from, struct mem_cgroup *to)
3318 static DEFINE_MUTEX(memcg_max_mutex);
3320 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3321 unsigned long max, bool memsw)
3323 bool enlarge = false;
3324 bool drained = false;
3326 bool limits_invariant;
3327 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3330 if (signal_pending(current)) {
3335 mutex_lock(&memcg_max_mutex);
3337 * Make sure that the new limit (memsw or memory limit) doesn't
3338 * break our basic invariant rule memory.max <= memsw.max.
3340 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3341 max <= memcg->memsw.max;
3342 if (!limits_invariant) {
3343 mutex_unlock(&memcg_max_mutex);
3347 if (max > counter->max)
3349 ret = page_counter_set_max(counter, max);
3350 mutex_unlock(&memcg_max_mutex);
3356 drain_all_stock(memcg);
3361 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3362 GFP_KERNEL, !memsw)) {
3368 if (!ret && enlarge)
3369 memcg_oom_recover(memcg);
3374 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3376 unsigned long *total_scanned)
3378 unsigned long nr_reclaimed = 0;
3379 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3380 unsigned long reclaimed;
3382 struct mem_cgroup_tree_per_node *mctz;
3383 unsigned long excess;
3384 unsigned long nr_scanned;
3389 mctz = soft_limit_tree_node(pgdat->node_id);
3392 * Do not even bother to check the largest node if the root
3393 * is empty. Do it lockless to prevent lock bouncing. Races
3394 * are acceptable as soft limit is best effort anyway.
3396 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3400 * This loop can run a while, specially if mem_cgroup's continuously
3401 * keep exceeding their soft limit and putting the system under
3408 mz = mem_cgroup_largest_soft_limit_node(mctz);
3413 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3414 gfp_mask, &nr_scanned);
3415 nr_reclaimed += reclaimed;
3416 *total_scanned += nr_scanned;
3417 spin_lock_irq(&mctz->lock);
3418 __mem_cgroup_remove_exceeded(mz, mctz);
3421 * If we failed to reclaim anything from this memory cgroup
3422 * it is time to move on to the next cgroup
3426 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3428 excess = soft_limit_excess(mz->memcg);
3430 * One school of thought says that we should not add
3431 * back the node to the tree if reclaim returns 0.
3432 * But our reclaim could return 0, simply because due
3433 * to priority we are exposing a smaller subset of
3434 * memory to reclaim from. Consider this as a longer
3437 /* If excess == 0, no tree ops */
3438 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3439 spin_unlock_irq(&mctz->lock);
3440 css_put(&mz->memcg->css);
3443 * Could not reclaim anything and there are no more
3444 * mem cgroups to try or we seem to be looping without
3445 * reclaiming anything.
3447 if (!nr_reclaimed &&
3449 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3451 } while (!nr_reclaimed);
3453 css_put(&next_mz->memcg->css);
3454 return nr_reclaimed;
3458 * Reclaims as many pages from the given memcg as possible.
3460 * Caller is responsible for holding css reference for memcg.
3462 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3464 int nr_retries = MAX_RECLAIM_RETRIES;
3466 /* we call try-to-free pages for make this cgroup empty */
3467 lru_add_drain_all();
3469 drain_all_stock(memcg);
3471 /* try to free all pages in this cgroup */
3472 while (nr_retries && page_counter_read(&memcg->memory)) {
3475 if (signal_pending(current))
3478 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3482 /* maybe some writeback is necessary */
3483 congestion_wait(BLK_RW_ASYNC, HZ/10);
3491 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3492 char *buf, size_t nbytes,
3495 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3497 if (mem_cgroup_is_root(memcg))
3499 return mem_cgroup_force_empty(memcg) ?: nbytes;
3502 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3508 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3509 struct cftype *cft, u64 val)
3514 pr_warn_once("Non-hierarchical mode is deprecated. "
3515 "Please report your usecase to linux-mm@kvack.org if you "
3516 "depend on this functionality.\n");
3521 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3525 if (mem_cgroup_is_root(memcg)) {
3526 cgroup_rstat_flush(memcg->css.cgroup);
3527 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3528 memcg_page_state(memcg, NR_ANON_MAPPED);
3530 val += memcg_page_state(memcg, MEMCG_SWAP);
3533 val = page_counter_read(&memcg->memory);
3535 val = page_counter_read(&memcg->memsw);
3548 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3552 struct page_counter *counter;
3554 switch (MEMFILE_TYPE(cft->private)) {
3556 counter = &memcg->memory;
3559 counter = &memcg->memsw;
3562 counter = &memcg->kmem;
3565 counter = &memcg->tcpmem;
3571 switch (MEMFILE_ATTR(cft->private)) {
3573 if (counter == &memcg->memory)
3574 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3575 if (counter == &memcg->memsw)
3576 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3577 return (u64)page_counter_read(counter) * PAGE_SIZE;
3579 return (u64)counter->max * PAGE_SIZE;
3581 return (u64)counter->watermark * PAGE_SIZE;
3583 return counter->failcnt;
3584 case RES_SOFT_LIMIT:
3585 return (u64)memcg->soft_limit * PAGE_SIZE;
3591 #ifdef CONFIG_MEMCG_KMEM
3592 static int memcg_online_kmem(struct mem_cgroup *memcg)
3594 struct obj_cgroup *objcg;
3597 if (cgroup_memory_nokmem)
3600 BUG_ON(memcg->kmemcg_id >= 0);
3601 BUG_ON(memcg->kmem_state);
3603 memcg_id = memcg_alloc_cache_id();
3607 objcg = obj_cgroup_alloc();
3609 memcg_free_cache_id(memcg_id);
3612 objcg->memcg = memcg;
3613 rcu_assign_pointer(memcg->objcg, objcg);
3615 static_branch_enable(&memcg_kmem_enabled_key);
3617 memcg->kmemcg_id = memcg_id;
3618 memcg->kmem_state = KMEM_ONLINE;
3623 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3625 struct cgroup_subsys_state *css;
3626 struct mem_cgroup *parent, *child;
3629 if (memcg->kmem_state != KMEM_ONLINE)
3632 memcg->kmem_state = KMEM_ALLOCATED;
3634 parent = parent_mem_cgroup(memcg);
3636 parent = root_mem_cgroup;
3638 memcg_reparent_objcgs(memcg, parent);
3640 kmemcg_id = memcg->kmemcg_id;
3641 BUG_ON(kmemcg_id < 0);
3644 * Change kmemcg_id of this cgroup and all its descendants to the
3645 * parent's id, and then move all entries from this cgroup's list_lrus
3646 * to ones of the parent. After we have finished, all list_lrus
3647 * corresponding to this cgroup are guaranteed to remain empty. The
3648 * ordering is imposed by list_lru_node->lock taken by
3649 * memcg_drain_all_list_lrus().
3651 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3652 css_for_each_descendant_pre(css, &memcg->css) {
3653 child = mem_cgroup_from_css(css);
3654 BUG_ON(child->kmemcg_id != kmemcg_id);
3655 child->kmemcg_id = parent->kmemcg_id;
3659 memcg_drain_all_list_lrus(kmemcg_id, parent);
3661 memcg_free_cache_id(kmemcg_id);
3664 static void memcg_free_kmem(struct mem_cgroup *memcg)
3666 /* css_alloc() failed, offlining didn't happen */
3667 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3668 memcg_offline_kmem(memcg);
3671 static int memcg_online_kmem(struct mem_cgroup *memcg)
3675 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3678 static void memcg_free_kmem(struct mem_cgroup *memcg)
3681 #endif /* CONFIG_MEMCG_KMEM */
3683 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3688 mutex_lock(&memcg_max_mutex);
3689 ret = page_counter_set_max(&memcg->kmem, max);
3690 mutex_unlock(&memcg_max_mutex);
3694 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3698 mutex_lock(&memcg_max_mutex);
3700 ret = page_counter_set_max(&memcg->tcpmem, max);
3704 if (!memcg->tcpmem_active) {
3706 * The active flag needs to be written after the static_key
3707 * update. This is what guarantees that the socket activation
3708 * function is the last one to run. See mem_cgroup_sk_alloc()
3709 * for details, and note that we don't mark any socket as
3710 * belonging to this memcg until that flag is up.
3712 * We need to do this, because static_keys will span multiple
3713 * sites, but we can't control their order. If we mark a socket
3714 * as accounted, but the accounting functions are not patched in
3715 * yet, we'll lose accounting.
3717 * We never race with the readers in mem_cgroup_sk_alloc(),
3718 * because when this value change, the code to process it is not
3721 static_branch_inc(&memcg_sockets_enabled_key);
3722 memcg->tcpmem_active = true;
3725 mutex_unlock(&memcg_max_mutex);
3730 * The user of this function is...
3733 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3734 char *buf, size_t nbytes, loff_t off)
3736 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3737 unsigned long nr_pages;
3740 buf = strstrip(buf);
3741 ret = page_counter_memparse(buf, "-1", &nr_pages);
3745 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3747 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3751 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3753 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3756 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3759 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3760 "Please report your usecase to linux-mm@kvack.org if you "
3761 "depend on this functionality.\n");
3762 ret = memcg_update_kmem_max(memcg, nr_pages);
3765 ret = memcg_update_tcp_max(memcg, nr_pages);
3769 case RES_SOFT_LIMIT:
3770 memcg->soft_limit = nr_pages;
3774 return ret ?: nbytes;
3777 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3778 size_t nbytes, loff_t off)
3780 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3781 struct page_counter *counter;
3783 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3785 counter = &memcg->memory;
3788 counter = &memcg->memsw;
3791 counter = &memcg->kmem;
3794 counter = &memcg->tcpmem;
3800 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3802 page_counter_reset_watermark(counter);
3805 counter->failcnt = 0;
3814 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3817 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3821 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3822 struct cftype *cft, u64 val)
3824 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3826 if (val & ~MOVE_MASK)
3830 * No kind of locking is needed in here, because ->can_attach() will
3831 * check this value once in the beginning of the process, and then carry
3832 * on with stale data. This means that changes to this value will only
3833 * affect task migrations starting after the change.
3835 memcg->move_charge_at_immigrate = val;
3839 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3840 struct cftype *cft, u64 val)
3848 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3849 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3850 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3852 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3853 int nid, unsigned int lru_mask, bool tree)
3855 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3856 unsigned long nr = 0;
3859 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3862 if (!(BIT(lru) & lru_mask))
3865 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3867 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3872 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3873 unsigned int lru_mask,
3876 unsigned long nr = 0;
3880 if (!(BIT(lru) & lru_mask))
3883 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3885 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3890 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3894 unsigned int lru_mask;
3897 static const struct numa_stat stats[] = {
3898 { "total", LRU_ALL },
3899 { "file", LRU_ALL_FILE },
3900 { "anon", LRU_ALL_ANON },
3901 { "unevictable", BIT(LRU_UNEVICTABLE) },
3903 const struct numa_stat *stat;
3905 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3907 cgroup_rstat_flush(memcg->css.cgroup);
3909 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3910 seq_printf(m, "%s=%lu", stat->name,
3911 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3913 for_each_node_state(nid, N_MEMORY)
3914 seq_printf(m, " N%d=%lu", nid,
3915 mem_cgroup_node_nr_lru_pages(memcg, nid,
3916 stat->lru_mask, false));
3920 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3922 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3923 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3925 for_each_node_state(nid, N_MEMORY)
3926 seq_printf(m, " N%d=%lu", nid,
3927 mem_cgroup_node_nr_lru_pages(memcg, nid,
3928 stat->lru_mask, true));
3934 #endif /* CONFIG_NUMA */
3936 static const unsigned int memcg1_stats[] = {
3939 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3949 static const char *const memcg1_stat_names[] = {
3952 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3962 /* Universal VM events cgroup1 shows, original sort order */
3963 static const unsigned int memcg1_events[] = {
3970 static int memcg_stat_show(struct seq_file *m, void *v)
3972 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3973 unsigned long memory, memsw;
3974 struct mem_cgroup *mi;
3977 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
3979 cgroup_rstat_flush(memcg->css.cgroup);
3981 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
3984 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
3986 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
3987 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
3990 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
3991 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
3992 memcg_events_local(memcg, memcg1_events[i]));
3994 for (i = 0; i < NR_LRU_LISTS; i++)
3995 seq_printf(m, "%s %lu\n", lru_list_name(i),
3996 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
3999 /* Hierarchical information */
4000 memory = memsw = PAGE_COUNTER_MAX;
4001 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4002 memory = min(memory, READ_ONCE(mi->memory.max));
4003 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4005 seq_printf(m, "hierarchical_memory_limit %llu\n",
4006 (u64)memory * PAGE_SIZE);
4007 if (do_memsw_account())
4008 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4009 (u64)memsw * PAGE_SIZE);
4011 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4014 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4016 nr = memcg_page_state(memcg, memcg1_stats[i]);
4017 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4018 (u64)nr * PAGE_SIZE);
4021 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4022 seq_printf(m, "total_%s %llu\n",
4023 vm_event_name(memcg1_events[i]),
4024 (u64)memcg_events(memcg, memcg1_events[i]));
4026 for (i = 0; i < NR_LRU_LISTS; i++)
4027 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4028 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4031 #ifdef CONFIG_DEBUG_VM
4034 struct mem_cgroup_per_node *mz;
4035 unsigned long anon_cost = 0;
4036 unsigned long file_cost = 0;
4038 for_each_online_pgdat(pgdat) {
4039 mz = memcg->nodeinfo[pgdat->node_id];
4041 anon_cost += mz->lruvec.anon_cost;
4042 file_cost += mz->lruvec.file_cost;
4044 seq_printf(m, "anon_cost %lu\n", anon_cost);
4045 seq_printf(m, "file_cost %lu\n", file_cost);
4052 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4055 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4057 return mem_cgroup_swappiness(memcg);
4060 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4061 struct cftype *cft, u64 val)
4063 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4068 if (!mem_cgroup_is_root(memcg))
4069 memcg->swappiness = val;
4071 vm_swappiness = val;
4076 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4078 struct mem_cgroup_threshold_ary *t;
4079 unsigned long usage;
4084 t = rcu_dereference(memcg->thresholds.primary);
4086 t = rcu_dereference(memcg->memsw_thresholds.primary);
4091 usage = mem_cgroup_usage(memcg, swap);
4094 * current_threshold points to threshold just below or equal to usage.
4095 * If it's not true, a threshold was crossed after last
4096 * call of __mem_cgroup_threshold().
4098 i = t->current_threshold;
4101 * Iterate backward over array of thresholds starting from
4102 * current_threshold and check if a threshold is crossed.
4103 * If none of thresholds below usage is crossed, we read
4104 * only one element of the array here.
4106 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4107 eventfd_signal(t->entries[i].eventfd, 1);
4109 /* i = current_threshold + 1 */
4113 * Iterate forward over array of thresholds starting from
4114 * current_threshold+1 and check if a threshold is crossed.
4115 * If none of thresholds above usage is crossed, we read
4116 * only one element of the array here.
4118 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4119 eventfd_signal(t->entries[i].eventfd, 1);
4121 /* Update current_threshold */
4122 t->current_threshold = i - 1;
4127 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4130 __mem_cgroup_threshold(memcg, false);
4131 if (do_memsw_account())
4132 __mem_cgroup_threshold(memcg, true);
4134 memcg = parent_mem_cgroup(memcg);
4138 static int compare_thresholds(const void *a, const void *b)
4140 const struct mem_cgroup_threshold *_a = a;
4141 const struct mem_cgroup_threshold *_b = b;
4143 if (_a->threshold > _b->threshold)
4146 if (_a->threshold < _b->threshold)
4152 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4154 struct mem_cgroup_eventfd_list *ev;
4156 spin_lock(&memcg_oom_lock);
4158 list_for_each_entry(ev, &memcg->oom_notify, list)
4159 eventfd_signal(ev->eventfd, 1);
4161 spin_unlock(&memcg_oom_lock);
4165 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4167 struct mem_cgroup *iter;
4169 for_each_mem_cgroup_tree(iter, memcg)
4170 mem_cgroup_oom_notify_cb(iter);
4173 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4174 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4176 struct mem_cgroup_thresholds *thresholds;
4177 struct mem_cgroup_threshold_ary *new;
4178 unsigned long threshold;
4179 unsigned long usage;
4182 ret = page_counter_memparse(args, "-1", &threshold);
4186 mutex_lock(&memcg->thresholds_lock);
4189 thresholds = &memcg->thresholds;
4190 usage = mem_cgroup_usage(memcg, false);
4191 } else if (type == _MEMSWAP) {
4192 thresholds = &memcg->memsw_thresholds;
4193 usage = mem_cgroup_usage(memcg, true);
4197 /* Check if a threshold crossed before adding a new one */
4198 if (thresholds->primary)
4199 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4201 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4203 /* Allocate memory for new array of thresholds */
4204 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4211 /* Copy thresholds (if any) to new array */
4212 if (thresholds->primary)
4213 memcpy(new->entries, thresholds->primary->entries,
4214 flex_array_size(new, entries, size - 1));
4216 /* Add new threshold */
4217 new->entries[size - 1].eventfd = eventfd;
4218 new->entries[size - 1].threshold = threshold;
4220 /* Sort thresholds. Registering of new threshold isn't time-critical */
4221 sort(new->entries, size, sizeof(*new->entries),
4222 compare_thresholds, NULL);
4224 /* Find current threshold */
4225 new->current_threshold = -1;
4226 for (i = 0; i < size; i++) {
4227 if (new->entries[i].threshold <= usage) {
4229 * new->current_threshold will not be used until
4230 * rcu_assign_pointer(), so it's safe to increment
4233 ++new->current_threshold;
4238 /* Free old spare buffer and save old primary buffer as spare */
4239 kfree(thresholds->spare);
4240 thresholds->spare = thresholds->primary;
4242 rcu_assign_pointer(thresholds->primary, new);
4244 /* To be sure that nobody uses thresholds */
4248 mutex_unlock(&memcg->thresholds_lock);
4253 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4254 struct eventfd_ctx *eventfd, const char *args)
4256 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4259 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4260 struct eventfd_ctx *eventfd, const char *args)
4262 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4265 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4266 struct eventfd_ctx *eventfd, enum res_type type)
4268 struct mem_cgroup_thresholds *thresholds;
4269 struct mem_cgroup_threshold_ary *new;
4270 unsigned long usage;
4271 int i, j, size, entries;
4273 mutex_lock(&memcg->thresholds_lock);
4276 thresholds = &memcg->thresholds;
4277 usage = mem_cgroup_usage(memcg, false);
4278 } else if (type == _MEMSWAP) {
4279 thresholds = &memcg->memsw_thresholds;
4280 usage = mem_cgroup_usage(memcg, true);
4284 if (!thresholds->primary)
4287 /* Check if a threshold crossed before removing */
4288 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4290 /* Calculate new number of threshold */
4292 for (i = 0; i < thresholds->primary->size; i++) {
4293 if (thresholds->primary->entries[i].eventfd != eventfd)
4299 new = thresholds->spare;
4301 /* If no items related to eventfd have been cleared, nothing to do */
4305 /* Set thresholds array to NULL if we don't have thresholds */
4314 /* Copy thresholds and find current threshold */
4315 new->current_threshold = -1;
4316 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4317 if (thresholds->primary->entries[i].eventfd == eventfd)
4320 new->entries[j] = thresholds->primary->entries[i];
4321 if (new->entries[j].threshold <= usage) {
4323 * new->current_threshold will not be used
4324 * until rcu_assign_pointer(), so it's safe to increment
4327 ++new->current_threshold;
4333 /* Swap primary and spare array */
4334 thresholds->spare = thresholds->primary;
4336 rcu_assign_pointer(thresholds->primary, new);
4338 /* To be sure that nobody uses thresholds */
4341 /* If all events are unregistered, free the spare array */
4343 kfree(thresholds->spare);
4344 thresholds->spare = NULL;
4347 mutex_unlock(&memcg->thresholds_lock);
4350 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4351 struct eventfd_ctx *eventfd)
4353 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4356 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4357 struct eventfd_ctx *eventfd)
4359 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4362 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4363 struct eventfd_ctx *eventfd, const char *args)
4365 struct mem_cgroup_eventfd_list *event;
4367 event = kmalloc(sizeof(*event), GFP_KERNEL);
4371 spin_lock(&memcg_oom_lock);
4373 event->eventfd = eventfd;
4374 list_add(&event->list, &memcg->oom_notify);
4376 /* already in OOM ? */
4377 if (memcg->under_oom)
4378 eventfd_signal(eventfd, 1);
4379 spin_unlock(&memcg_oom_lock);
4384 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4385 struct eventfd_ctx *eventfd)
4387 struct mem_cgroup_eventfd_list *ev, *tmp;
4389 spin_lock(&memcg_oom_lock);
4391 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4392 if (ev->eventfd == eventfd) {
4393 list_del(&ev->list);
4398 spin_unlock(&memcg_oom_lock);
4401 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4403 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4405 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4406 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4407 seq_printf(sf, "oom_kill %lu\n",
4408 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4412 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4413 struct cftype *cft, u64 val)
4415 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4417 /* cannot set to root cgroup and only 0 and 1 are allowed */
4418 if (mem_cgroup_is_root(memcg) || !((val == 0) || (val == 1)))
4421 memcg->oom_kill_disable = val;
4423 memcg_oom_recover(memcg);
4428 #ifdef CONFIG_CGROUP_WRITEBACK
4430 #include <trace/events/writeback.h>
4432 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4434 return wb_domain_init(&memcg->cgwb_domain, gfp);
4437 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4439 wb_domain_exit(&memcg->cgwb_domain);
4442 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4444 wb_domain_size_changed(&memcg->cgwb_domain);
4447 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4449 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4451 if (!memcg->css.parent)
4454 return &memcg->cgwb_domain;
4458 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4459 * @wb: bdi_writeback in question
4460 * @pfilepages: out parameter for number of file pages
4461 * @pheadroom: out parameter for number of allocatable pages according to memcg
4462 * @pdirty: out parameter for number of dirty pages
4463 * @pwriteback: out parameter for number of pages under writeback
4465 * Determine the numbers of file, headroom, dirty, and writeback pages in
4466 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4467 * is a bit more involved.
4469 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4470 * headroom is calculated as the lowest headroom of itself and the
4471 * ancestors. Note that this doesn't consider the actual amount of
4472 * available memory in the system. The caller should further cap
4473 * *@pheadroom accordingly.
4475 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4476 unsigned long *pheadroom, unsigned long *pdirty,
4477 unsigned long *pwriteback)
4479 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4480 struct mem_cgroup *parent;
4482 cgroup_rstat_flush_irqsafe(memcg->css.cgroup);
4484 *pdirty = memcg_page_state(memcg, NR_FILE_DIRTY);
4485 *pwriteback = memcg_page_state(memcg, NR_WRITEBACK);
4486 *pfilepages = memcg_page_state(memcg, NR_INACTIVE_FILE) +
4487 memcg_page_state(memcg, NR_ACTIVE_FILE);
4489 *pheadroom = PAGE_COUNTER_MAX;
4490 while ((parent = parent_mem_cgroup(memcg))) {
4491 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4492 READ_ONCE(memcg->memory.high));
4493 unsigned long used = page_counter_read(&memcg->memory);
4495 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4501 * Foreign dirty flushing
4503 * There's an inherent mismatch between memcg and writeback. The former
4504 * trackes ownership per-page while the latter per-inode. This was a
4505 * deliberate design decision because honoring per-page ownership in the
4506 * writeback path is complicated, may lead to higher CPU and IO overheads
4507 * and deemed unnecessary given that write-sharing an inode across
4508 * different cgroups isn't a common use-case.
4510 * Combined with inode majority-writer ownership switching, this works well
4511 * enough in most cases but there are some pathological cases. For
4512 * example, let's say there are two cgroups A and B which keep writing to
4513 * different but confined parts of the same inode. B owns the inode and
4514 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4515 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4516 * triggering background writeback. A will be slowed down without a way to
4517 * make writeback of the dirty pages happen.
4519 * Conditions like the above can lead to a cgroup getting repatedly and
4520 * severely throttled after making some progress after each
4521 * dirty_expire_interval while the underyling IO device is almost
4524 * Solving this problem completely requires matching the ownership tracking
4525 * granularities between memcg and writeback in either direction. However,
4526 * the more egregious behaviors can be avoided by simply remembering the
4527 * most recent foreign dirtying events and initiating remote flushes on
4528 * them when local writeback isn't enough to keep the memory clean enough.
4530 * The following two functions implement such mechanism. When a foreign
4531 * page - a page whose memcg and writeback ownerships don't match - is
4532 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4533 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4534 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4535 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4536 * foreign bdi_writebacks which haven't expired. Both the numbers of
4537 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4538 * limited to MEMCG_CGWB_FRN_CNT.
4540 * The mechanism only remembers IDs and doesn't hold any object references.
4541 * As being wrong occasionally doesn't matter, updates and accesses to the
4542 * records are lockless and racy.
4544 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4545 struct bdi_writeback *wb)
4547 struct mem_cgroup *memcg = page_memcg(page);
4548 struct memcg_cgwb_frn *frn;
4549 u64 now = get_jiffies_64();
4550 u64 oldest_at = now;
4554 trace_track_foreign_dirty(page, wb);
4557 * Pick the slot to use. If there is already a slot for @wb, keep
4558 * using it. If not replace the oldest one which isn't being
4561 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4562 frn = &memcg->cgwb_frn[i];
4563 if (frn->bdi_id == wb->bdi->id &&
4564 frn->memcg_id == wb->memcg_css->id)
4566 if (time_before64(frn->at, oldest_at) &&
4567 atomic_read(&frn->done.cnt) == 1) {
4569 oldest_at = frn->at;
4573 if (i < MEMCG_CGWB_FRN_CNT) {
4575 * Re-using an existing one. Update timestamp lazily to
4576 * avoid making the cacheline hot. We want them to be
4577 * reasonably up-to-date and significantly shorter than
4578 * dirty_expire_interval as that's what expires the record.
4579 * Use the shorter of 1s and dirty_expire_interval / 8.
4581 unsigned long update_intv =
4582 min_t(unsigned long, HZ,
4583 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4585 if (time_before64(frn->at, now - update_intv))
4587 } else if (oldest >= 0) {
4588 /* replace the oldest free one */
4589 frn = &memcg->cgwb_frn[oldest];
4590 frn->bdi_id = wb->bdi->id;
4591 frn->memcg_id = wb->memcg_css->id;
4596 /* issue foreign writeback flushes for recorded foreign dirtying events */
4597 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4599 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4600 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4601 u64 now = jiffies_64;
4604 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4605 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4608 * If the record is older than dirty_expire_interval,
4609 * writeback on it has already started. No need to kick it
4610 * off again. Also, don't start a new one if there's
4611 * already one in flight.
4613 if (time_after64(frn->at, now - intv) &&
4614 atomic_read(&frn->done.cnt) == 1) {
4616 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4617 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4618 WB_REASON_FOREIGN_FLUSH,
4624 #else /* CONFIG_CGROUP_WRITEBACK */
4626 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4631 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4635 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4639 #endif /* CONFIG_CGROUP_WRITEBACK */
4642 * DO NOT USE IN NEW FILES.
4644 * "cgroup.event_control" implementation.
4646 * This is way over-engineered. It tries to support fully configurable
4647 * events for each user. Such level of flexibility is completely
4648 * unnecessary especially in the light of the planned unified hierarchy.
4650 * Please deprecate this and replace with something simpler if at all
4655 * Unregister event and free resources.
4657 * Gets called from workqueue.
4659 static void memcg_event_remove(struct work_struct *work)
4661 struct mem_cgroup_event *event =
4662 container_of(work, struct mem_cgroup_event, remove);
4663 struct mem_cgroup *memcg = event->memcg;
4665 remove_wait_queue(event->wqh, &event->wait);
4667 event->unregister_event(memcg, event->eventfd);
4669 /* Notify userspace the event is going away. */
4670 eventfd_signal(event->eventfd, 1);
4672 eventfd_ctx_put(event->eventfd);
4674 css_put(&memcg->css);
4678 * Gets called on EPOLLHUP on eventfd when user closes it.
4680 * Called with wqh->lock held and interrupts disabled.
4682 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4683 int sync, void *key)
4685 struct mem_cgroup_event *event =
4686 container_of(wait, struct mem_cgroup_event, wait);
4687 struct mem_cgroup *memcg = event->memcg;
4688 __poll_t flags = key_to_poll(key);
4690 if (flags & EPOLLHUP) {
4692 * If the event has been detached at cgroup removal, we
4693 * can simply return knowing the other side will cleanup
4696 * We can't race against event freeing since the other
4697 * side will require wqh->lock via remove_wait_queue(),
4700 spin_lock(&memcg->event_list_lock);
4701 if (!list_empty(&event->list)) {
4702 list_del_init(&event->list);
4704 * We are in atomic context, but cgroup_event_remove()
4705 * may sleep, so we have to call it in workqueue.
4707 schedule_work(&event->remove);
4709 spin_unlock(&memcg->event_list_lock);
4715 static void memcg_event_ptable_queue_proc(struct file *file,
4716 wait_queue_head_t *wqh, poll_table *pt)
4718 struct mem_cgroup_event *event =
4719 container_of(pt, struct mem_cgroup_event, pt);
4722 add_wait_queue(wqh, &event->wait);
4726 * DO NOT USE IN NEW FILES.
4728 * Parse input and register new cgroup event handler.
4730 * Input must be in format '<event_fd> <control_fd> <args>'.
4731 * Interpretation of args is defined by control file implementation.
4733 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4734 char *buf, size_t nbytes, loff_t off)
4736 struct cgroup_subsys_state *css = of_css(of);
4737 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4738 struct mem_cgroup_event *event;
4739 struct cgroup_subsys_state *cfile_css;
4740 unsigned int efd, cfd;
4747 buf = strstrip(buf);
4749 efd = simple_strtoul(buf, &endp, 10);
4754 cfd = simple_strtoul(buf, &endp, 10);
4755 if ((*endp != ' ') && (*endp != '\0'))
4759 event = kzalloc(sizeof(*event), GFP_KERNEL);
4763 event->memcg = memcg;
4764 INIT_LIST_HEAD(&event->list);
4765 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4766 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4767 INIT_WORK(&event->remove, memcg_event_remove);
4775 event->eventfd = eventfd_ctx_fileget(efile.file);
4776 if (IS_ERR(event->eventfd)) {
4777 ret = PTR_ERR(event->eventfd);
4784 goto out_put_eventfd;
4787 /* the process need read permission on control file */
4788 /* AV: shouldn't we check that it's been opened for read instead? */
4789 ret = file_permission(cfile.file, MAY_READ);
4794 * Determine the event callbacks and set them in @event. This used
4795 * to be done via struct cftype but cgroup core no longer knows
4796 * about these events. The following is crude but the whole thing
4797 * is for compatibility anyway.
4799 * DO NOT ADD NEW FILES.
4801 name = cfile.file->f_path.dentry->d_name.name;
4803 if (!strcmp(name, "memory.usage_in_bytes")) {
4804 event->register_event = mem_cgroup_usage_register_event;
4805 event->unregister_event = mem_cgroup_usage_unregister_event;
4806 } else if (!strcmp(name, "memory.oom_control")) {
4807 event->register_event = mem_cgroup_oom_register_event;
4808 event->unregister_event = mem_cgroup_oom_unregister_event;
4809 } else if (!strcmp(name, "memory.pressure_level")) {
4810 event->register_event = vmpressure_register_event;
4811 event->unregister_event = vmpressure_unregister_event;
4812 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4813 event->register_event = memsw_cgroup_usage_register_event;
4814 event->unregister_event = memsw_cgroup_usage_unregister_event;
4821 * Verify @cfile should belong to @css. Also, remaining events are
4822 * automatically removed on cgroup destruction but the removal is
4823 * asynchronous, so take an extra ref on @css.
4825 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4826 &memory_cgrp_subsys);
4828 if (IS_ERR(cfile_css))
4830 if (cfile_css != css) {
4835 ret = event->register_event(memcg, event->eventfd, buf);
4839 vfs_poll(efile.file, &event->pt);
4841 spin_lock(&memcg->event_list_lock);
4842 list_add(&event->list, &memcg->event_list);
4843 spin_unlock(&memcg->event_list_lock);
4855 eventfd_ctx_put(event->eventfd);
4864 static struct cftype mem_cgroup_legacy_files[] = {
4866 .name = "usage_in_bytes",
4867 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4868 .read_u64 = mem_cgroup_read_u64,
4871 .name = "max_usage_in_bytes",
4872 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4873 .write = mem_cgroup_reset,
4874 .read_u64 = mem_cgroup_read_u64,
4877 .name = "limit_in_bytes",
4878 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4879 .write = mem_cgroup_write,
4880 .read_u64 = mem_cgroup_read_u64,
4883 .name = "soft_limit_in_bytes",
4884 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4885 .write = mem_cgroup_write,
4886 .read_u64 = mem_cgroup_read_u64,
4890 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4891 .write = mem_cgroup_reset,
4892 .read_u64 = mem_cgroup_read_u64,
4896 .seq_show = memcg_stat_show,
4899 .name = "force_empty",
4900 .write = mem_cgroup_force_empty_write,
4903 .name = "use_hierarchy",
4904 .write_u64 = mem_cgroup_hierarchy_write,
4905 .read_u64 = mem_cgroup_hierarchy_read,
4908 .name = "cgroup.event_control", /* XXX: for compat */
4909 .write = memcg_write_event_control,
4910 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4913 .name = "swappiness",
4914 .read_u64 = mem_cgroup_swappiness_read,
4915 .write_u64 = mem_cgroup_swappiness_write,
4918 .name = "move_charge_at_immigrate",
4919 .read_u64 = mem_cgroup_move_charge_read,
4920 .write_u64 = mem_cgroup_move_charge_write,
4923 .name = "oom_control",
4924 .seq_show = mem_cgroup_oom_control_read,
4925 .write_u64 = mem_cgroup_oom_control_write,
4926 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4929 .name = "pressure_level",
4933 .name = "numa_stat",
4934 .seq_show = memcg_numa_stat_show,
4938 .name = "kmem.limit_in_bytes",
4939 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4940 .write = mem_cgroup_write,
4941 .read_u64 = mem_cgroup_read_u64,
4944 .name = "kmem.usage_in_bytes",
4945 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4946 .read_u64 = mem_cgroup_read_u64,
4949 .name = "kmem.failcnt",
4950 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4951 .write = mem_cgroup_reset,
4952 .read_u64 = mem_cgroup_read_u64,
4955 .name = "kmem.max_usage_in_bytes",
4956 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4957 .write = mem_cgroup_reset,
4958 .read_u64 = mem_cgroup_read_u64,
4960 #if defined(CONFIG_MEMCG_KMEM) && \
4961 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
4963 .name = "kmem.slabinfo",
4964 .seq_show = memcg_slab_show,
4968 .name = "kmem.tcp.limit_in_bytes",
4969 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
4970 .write = mem_cgroup_write,
4971 .read_u64 = mem_cgroup_read_u64,
4974 .name = "kmem.tcp.usage_in_bytes",
4975 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
4976 .read_u64 = mem_cgroup_read_u64,
4979 .name = "kmem.tcp.failcnt",
4980 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
4981 .write = mem_cgroup_reset,
4982 .read_u64 = mem_cgroup_read_u64,
4985 .name = "kmem.tcp.max_usage_in_bytes",
4986 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
4987 .write = mem_cgroup_reset,
4988 .read_u64 = mem_cgroup_read_u64,
4990 { }, /* terminate */
4994 * Private memory cgroup IDR
4996 * Swap-out records and page cache shadow entries need to store memcg
4997 * references in constrained space, so we maintain an ID space that is
4998 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
4999 * memory-controlled cgroups to 64k.
5001 * However, there usually are many references to the offline CSS after
5002 * the cgroup has been destroyed, such as page cache or reclaimable
5003 * slab objects, that don't need to hang on to the ID. We want to keep
5004 * those dead CSS from occupying IDs, or we might quickly exhaust the
5005 * relatively small ID space and prevent the creation of new cgroups
5006 * even when there are much fewer than 64k cgroups - possibly none.
5008 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5009 * be freed and recycled when it's no longer needed, which is usually
5010 * when the CSS is offlined.
5012 * The only exception to that are records of swapped out tmpfs/shmem
5013 * pages that need to be attributed to live ancestors on swapin. But
5014 * those references are manageable from userspace.
5017 static DEFINE_IDR(mem_cgroup_idr);
5019 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5021 if (memcg->id.id > 0) {
5022 idr_remove(&mem_cgroup_idr, memcg->id.id);
5027 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5030 refcount_add(n, &memcg->id.ref);
5033 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5035 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5036 mem_cgroup_id_remove(memcg);
5038 /* Memcg ID pins CSS */
5039 css_put(&memcg->css);
5043 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5045 mem_cgroup_id_put_many(memcg, 1);
5049 * mem_cgroup_from_id - look up a memcg from a memcg id
5050 * @id: the memcg id to look up
5052 * Caller must hold rcu_read_lock().
5054 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5056 WARN_ON_ONCE(!rcu_read_lock_held());
5057 return idr_find(&mem_cgroup_idr, id);
5060 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5062 struct mem_cgroup_per_node *pn;
5065 * This routine is called against possible nodes.
5066 * But it's BUG to call kmalloc() against offline node.
5068 * TODO: this routine can waste much memory for nodes which will
5069 * never be onlined. It's better to use memory hotplug callback
5072 if (!node_state(node, N_NORMAL_MEMORY))
5074 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5078 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5079 GFP_KERNEL_ACCOUNT);
5080 if (!pn->lruvec_stat_local) {
5085 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct batched_lruvec_stat,
5086 GFP_KERNEL_ACCOUNT);
5087 if (!pn->lruvec_stat_cpu) {
5088 free_percpu(pn->lruvec_stat_local);
5093 lruvec_init(&pn->lruvec);
5094 pn->usage_in_excess = 0;
5095 pn->on_tree = false;
5098 memcg->nodeinfo[node] = pn;
5102 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5104 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5109 free_percpu(pn->lruvec_stat_cpu);
5110 free_percpu(pn->lruvec_stat_local);
5114 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5119 free_mem_cgroup_per_node_info(memcg, node);
5120 free_percpu(memcg->vmstats_percpu);
5124 static void mem_cgroup_free(struct mem_cgroup *memcg)
5128 memcg_wb_domain_exit(memcg);
5130 * Flush percpu lruvec stats to guarantee the value
5131 * correctness on parent's and all ancestor levels.
5133 for_each_online_cpu(cpu)
5134 memcg_flush_lruvec_page_state(memcg, cpu);
5135 __mem_cgroup_free(memcg);
5138 static struct mem_cgroup *mem_cgroup_alloc(void)
5140 struct mem_cgroup *memcg;
5143 int __maybe_unused i;
5144 long error = -ENOMEM;
5146 size = sizeof(struct mem_cgroup);
5147 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5149 memcg = kzalloc(size, GFP_KERNEL);
5151 return ERR_PTR(error);
5153 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5154 1, MEM_CGROUP_ID_MAX,
5156 if (memcg->id.id < 0) {
5157 error = memcg->id.id;
5161 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5162 GFP_KERNEL_ACCOUNT);
5163 if (!memcg->vmstats_percpu)
5167 if (alloc_mem_cgroup_per_node_info(memcg, node))
5170 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5173 INIT_WORK(&memcg->high_work, high_work_func);
5174 INIT_LIST_HEAD(&memcg->oom_notify);
5175 mutex_init(&memcg->thresholds_lock);
5176 spin_lock_init(&memcg->move_lock);
5177 vmpressure_init(&memcg->vmpressure);
5178 INIT_LIST_HEAD(&memcg->event_list);
5179 spin_lock_init(&memcg->event_list_lock);
5180 memcg->socket_pressure = jiffies;
5181 #ifdef CONFIG_MEMCG_KMEM
5182 memcg->kmemcg_id = -1;
5183 INIT_LIST_HEAD(&memcg->objcg_list);
5185 #ifdef CONFIG_CGROUP_WRITEBACK
5186 INIT_LIST_HEAD(&memcg->cgwb_list);
5187 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5188 memcg->cgwb_frn[i].done =
5189 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5191 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5192 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5193 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5194 memcg->deferred_split_queue.split_queue_len = 0;
5196 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5199 mem_cgroup_id_remove(memcg);
5200 __mem_cgroup_free(memcg);
5201 return ERR_PTR(error);
5204 static struct cgroup_subsys_state * __ref
5205 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5207 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5208 struct mem_cgroup *memcg, *old_memcg;
5209 long error = -ENOMEM;
5211 old_memcg = set_active_memcg(parent);
5212 memcg = mem_cgroup_alloc();
5213 set_active_memcg(old_memcg);
5215 return ERR_CAST(memcg);
5217 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5218 memcg->soft_limit = PAGE_COUNTER_MAX;
5219 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5221 memcg->swappiness = mem_cgroup_swappiness(parent);
5222 memcg->oom_kill_disable = parent->oom_kill_disable;
5224 page_counter_init(&memcg->memory, &parent->memory);
5225 page_counter_init(&memcg->swap, &parent->swap);
5226 page_counter_init(&memcg->kmem, &parent->kmem);
5227 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5229 page_counter_init(&memcg->memory, NULL);
5230 page_counter_init(&memcg->swap, NULL);
5231 page_counter_init(&memcg->kmem, NULL);
5232 page_counter_init(&memcg->tcpmem, NULL);
5234 root_mem_cgroup = memcg;
5238 /* The following stuff does not apply to the root */
5239 error = memcg_online_kmem(memcg);
5243 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5244 static_branch_inc(&memcg_sockets_enabled_key);
5248 mem_cgroup_id_remove(memcg);
5249 mem_cgroup_free(memcg);
5250 return ERR_PTR(error);
5253 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5255 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5258 * A memcg must be visible for memcg_expand_shrinker_maps()
5259 * by the time the maps are allocated. So, we allocate maps
5260 * here, when for_each_mem_cgroup() can't skip it.
5262 if (memcg_alloc_shrinker_maps(memcg)) {
5263 mem_cgroup_id_remove(memcg);
5267 /* Online state pins memcg ID, memcg ID pins CSS */
5268 refcount_set(&memcg->id.ref, 1);
5273 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5275 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5276 struct mem_cgroup_event *event, *tmp;
5279 * Unregister events and notify userspace.
5280 * Notify userspace about cgroup removing only after rmdir of cgroup
5281 * directory to avoid race between userspace and kernelspace.
5283 spin_lock(&memcg->event_list_lock);
5284 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5285 list_del_init(&event->list);
5286 schedule_work(&event->remove);
5288 spin_unlock(&memcg->event_list_lock);
5290 page_counter_set_min(&memcg->memory, 0);
5291 page_counter_set_low(&memcg->memory, 0);
5293 memcg_offline_kmem(memcg);
5294 wb_memcg_offline(memcg);
5296 drain_all_stock(memcg);
5298 mem_cgroup_id_put(memcg);
5301 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5303 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5305 invalidate_reclaim_iterators(memcg);
5308 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5310 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5311 int __maybe_unused i;
5313 #ifdef CONFIG_CGROUP_WRITEBACK
5314 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5315 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5317 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5318 static_branch_dec(&memcg_sockets_enabled_key);
5320 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5321 static_branch_dec(&memcg_sockets_enabled_key);
5323 vmpressure_cleanup(&memcg->vmpressure);
5324 cancel_work_sync(&memcg->high_work);
5325 mem_cgroup_remove_from_trees(memcg);
5326 memcg_free_shrinker_maps(memcg);
5327 memcg_free_kmem(memcg);
5328 mem_cgroup_free(memcg);
5332 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5333 * @css: the target css
5335 * Reset the states of the mem_cgroup associated with @css. This is
5336 * invoked when the userland requests disabling on the default hierarchy
5337 * but the memcg is pinned through dependency. The memcg should stop
5338 * applying policies and should revert to the vanilla state as it may be
5339 * made visible again.
5341 * The current implementation only resets the essential configurations.
5342 * This needs to be expanded to cover all the visible parts.
5344 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5346 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5348 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5349 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5350 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5351 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5352 page_counter_set_min(&memcg->memory, 0);
5353 page_counter_set_low(&memcg->memory, 0);
5354 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5355 memcg->soft_limit = PAGE_COUNTER_MAX;
5356 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5357 memcg_wb_domain_size_changed(memcg);
5360 static void mem_cgroup_css_rstat_flush(struct cgroup_subsys_state *css, int cpu)
5362 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5363 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5364 struct memcg_vmstats_percpu *statc;
5368 statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
5370 for (i = 0; i < MEMCG_NR_STAT; i++) {
5372 * Collect the aggregated propagation counts of groups
5373 * below us. We're in a per-cpu loop here and this is
5374 * a global counter, so the first cycle will get them.
5376 delta = memcg->vmstats.state_pending[i];
5378 memcg->vmstats.state_pending[i] = 0;
5380 /* Add CPU changes on this level since the last flush */
5381 v = READ_ONCE(statc->state[i]);
5382 if (v != statc->state_prev[i]) {
5383 delta += v - statc->state_prev[i];
5384 statc->state_prev[i] = v;
5390 /* Aggregate counts on this level and propagate upwards */
5391 memcg->vmstats.state[i] += delta;
5393 parent->vmstats.state_pending[i] += delta;
5396 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
5397 delta = memcg->vmstats.events_pending[i];
5399 memcg->vmstats.events_pending[i] = 0;
5401 v = READ_ONCE(statc->events[i]);
5402 if (v != statc->events_prev[i]) {
5403 delta += v - statc->events_prev[i];
5404 statc->events_prev[i] = v;
5410 memcg->vmstats.events[i] += delta;
5412 parent->vmstats.events_pending[i] += delta;
5417 /* Handlers for move charge at task migration. */
5418 static int mem_cgroup_do_precharge(unsigned long count)
5422 /* Try a single bulk charge without reclaim first, kswapd may wake */
5423 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5425 mc.precharge += count;
5429 /* Try charges one by one with reclaim, but do not retry */
5431 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5445 enum mc_target_type {
5452 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5453 unsigned long addr, pte_t ptent)
5455 struct page *page = vm_normal_page(vma, addr, ptent);
5457 if (!page || !page_mapped(page))
5459 if (PageAnon(page)) {
5460 if (!(mc.flags & MOVE_ANON))
5463 if (!(mc.flags & MOVE_FILE))
5466 if (!get_page_unless_zero(page))
5472 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5473 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5474 pte_t ptent, swp_entry_t *entry)
5476 struct page *page = NULL;
5477 swp_entry_t ent = pte_to_swp_entry(ptent);
5479 if (!(mc.flags & MOVE_ANON))
5483 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5484 * a device and because they are not accessible by CPU they are store
5485 * as special swap entry in the CPU page table.
5487 if (is_device_private_entry(ent)) {
5488 page = device_private_entry_to_page(ent);
5490 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5491 * a refcount of 1 when free (unlike normal page)
5493 if (!page_ref_add_unless(page, 1, 1))
5498 if (non_swap_entry(ent))
5502 * Because lookup_swap_cache() updates some statistics counter,
5503 * we call find_get_page() with swapper_space directly.
5505 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5506 entry->val = ent.val;
5511 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5512 pte_t ptent, swp_entry_t *entry)
5518 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5519 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5521 if (!vma->vm_file) /* anonymous vma */
5523 if (!(mc.flags & MOVE_FILE))
5526 /* page is moved even if it's not RSS of this task(page-faulted). */
5527 /* shmem/tmpfs may report page out on swap: account for that too. */
5528 return find_get_incore_page(vma->vm_file->f_mapping,
5529 linear_page_index(vma, addr));
5533 * mem_cgroup_move_account - move account of the page
5535 * @compound: charge the page as compound or small page
5536 * @from: mem_cgroup which the page is moved from.
5537 * @to: mem_cgroup which the page is moved to. @from != @to.
5539 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5541 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5544 static int mem_cgroup_move_account(struct page *page,
5546 struct mem_cgroup *from,
5547 struct mem_cgroup *to)
5549 struct lruvec *from_vec, *to_vec;
5550 struct pglist_data *pgdat;
5551 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5554 VM_BUG_ON(from == to);
5555 VM_BUG_ON_PAGE(PageLRU(page), page);
5556 VM_BUG_ON(compound && !PageTransHuge(page));
5559 * Prevent mem_cgroup_migrate() from looking at
5560 * page's memory cgroup of its source page while we change it.
5563 if (!trylock_page(page))
5567 if (page_memcg(page) != from)
5570 pgdat = page_pgdat(page);
5571 from_vec = mem_cgroup_lruvec(from, pgdat);
5572 to_vec = mem_cgroup_lruvec(to, pgdat);
5574 lock_page_memcg(page);
5576 if (PageAnon(page)) {
5577 if (page_mapped(page)) {
5578 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5579 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5580 if (PageTransHuge(page)) {
5581 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5583 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5588 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5589 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5591 if (PageSwapBacked(page)) {
5592 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5593 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5596 if (page_mapped(page)) {
5597 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5598 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5601 if (PageDirty(page)) {
5602 struct address_space *mapping = page_mapping(page);
5604 if (mapping_can_writeback(mapping)) {
5605 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5607 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5613 if (PageWriteback(page)) {
5614 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5615 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5619 * All state has been migrated, let's switch to the new memcg.
5621 * It is safe to change page's memcg here because the page
5622 * is referenced, charged, isolated, and locked: we can't race
5623 * with (un)charging, migration, LRU putback, or anything else
5624 * that would rely on a stable page's memory cgroup.
5626 * Note that lock_page_memcg is a memcg lock, not a page lock,
5627 * to save space. As soon as we switch page's memory cgroup to a
5628 * new memcg that isn't locked, the above state can change
5629 * concurrently again. Make sure we're truly done with it.
5634 css_put(&from->css);
5636 page->memcg_data = (unsigned long)to;
5638 __unlock_page_memcg(from);
5642 local_irq_disable();
5643 mem_cgroup_charge_statistics(to, page, nr_pages);
5644 memcg_check_events(to, page);
5645 mem_cgroup_charge_statistics(from, page, -nr_pages);
5646 memcg_check_events(from, page);
5655 * get_mctgt_type - get target type of moving charge
5656 * @vma: the vma the pte to be checked belongs
5657 * @addr: the address corresponding to the pte to be checked
5658 * @ptent: the pte to be checked
5659 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5662 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5663 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5664 * move charge. if @target is not NULL, the page is stored in target->page
5665 * with extra refcnt got(Callers should handle it).
5666 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5667 * target for charge migration. if @target is not NULL, the entry is stored
5669 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5670 * (so ZONE_DEVICE page and thus not on the lru).
5671 * For now we such page is charge like a regular page would be as for all
5672 * intent and purposes it is just special memory taking the place of a
5675 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5677 * Called with pte lock held.
5680 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5681 unsigned long addr, pte_t ptent, union mc_target *target)
5683 struct page *page = NULL;
5684 enum mc_target_type ret = MC_TARGET_NONE;
5685 swp_entry_t ent = { .val = 0 };
5687 if (pte_present(ptent))
5688 page = mc_handle_present_pte(vma, addr, ptent);
5689 else if (is_swap_pte(ptent))
5690 page = mc_handle_swap_pte(vma, ptent, &ent);
5691 else if (pte_none(ptent))
5692 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5694 if (!page && !ent.val)
5698 * Do only loose check w/o serialization.
5699 * mem_cgroup_move_account() checks the page is valid or
5700 * not under LRU exclusion.
5702 if (page_memcg(page) == mc.from) {
5703 ret = MC_TARGET_PAGE;
5704 if (is_device_private_page(page))
5705 ret = MC_TARGET_DEVICE;
5707 target->page = page;
5709 if (!ret || !target)
5713 * There is a swap entry and a page doesn't exist or isn't charged.
5714 * But we cannot move a tail-page in a THP.
5716 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5717 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5718 ret = MC_TARGET_SWAP;
5725 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5727 * We don't consider PMD mapped swapping or file mapped pages because THP does
5728 * not support them for now.
5729 * Caller should make sure that pmd_trans_huge(pmd) is true.
5731 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5732 unsigned long addr, pmd_t pmd, union mc_target *target)
5734 struct page *page = NULL;
5735 enum mc_target_type ret = MC_TARGET_NONE;
5737 if (unlikely(is_swap_pmd(pmd))) {
5738 VM_BUG_ON(thp_migration_supported() &&
5739 !is_pmd_migration_entry(pmd));
5742 page = pmd_page(pmd);
5743 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5744 if (!(mc.flags & MOVE_ANON))
5746 if (page_memcg(page) == mc.from) {
5747 ret = MC_TARGET_PAGE;
5750 target->page = page;
5756 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5757 unsigned long addr, pmd_t pmd, union mc_target *target)
5759 return MC_TARGET_NONE;
5763 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5764 unsigned long addr, unsigned long end,
5765 struct mm_walk *walk)
5767 struct vm_area_struct *vma = walk->vma;
5771 ptl = pmd_trans_huge_lock(pmd, vma);
5774 * Note their can not be MC_TARGET_DEVICE for now as we do not
5775 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5776 * this might change.
5778 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5779 mc.precharge += HPAGE_PMD_NR;
5784 if (pmd_trans_unstable(pmd))
5786 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5787 for (; addr != end; pte++, addr += PAGE_SIZE)
5788 if (get_mctgt_type(vma, addr, *pte, NULL))
5789 mc.precharge++; /* increment precharge temporarily */
5790 pte_unmap_unlock(pte - 1, ptl);
5796 static const struct mm_walk_ops precharge_walk_ops = {
5797 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5800 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5802 unsigned long precharge;
5805 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5806 mmap_read_unlock(mm);
5808 precharge = mc.precharge;
5814 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5816 unsigned long precharge = mem_cgroup_count_precharge(mm);
5818 VM_BUG_ON(mc.moving_task);
5819 mc.moving_task = current;
5820 return mem_cgroup_do_precharge(precharge);
5823 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5824 static void __mem_cgroup_clear_mc(void)
5826 struct mem_cgroup *from = mc.from;
5827 struct mem_cgroup *to = mc.to;
5829 /* we must uncharge all the leftover precharges from mc.to */
5831 cancel_charge(mc.to, mc.precharge);
5835 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5836 * we must uncharge here.
5838 if (mc.moved_charge) {
5839 cancel_charge(mc.from, mc.moved_charge);
5840 mc.moved_charge = 0;
5842 /* we must fixup refcnts and charges */
5843 if (mc.moved_swap) {
5844 /* uncharge swap account from the old cgroup */
5845 if (!mem_cgroup_is_root(mc.from))
5846 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5848 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5851 * we charged both to->memory and to->memsw, so we
5852 * should uncharge to->memory.
5854 if (!mem_cgroup_is_root(mc.to))
5855 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5859 memcg_oom_recover(from);
5860 memcg_oom_recover(to);
5861 wake_up_all(&mc.waitq);
5864 static void mem_cgroup_clear_mc(void)
5866 struct mm_struct *mm = mc.mm;
5869 * we must clear moving_task before waking up waiters at the end of
5872 mc.moving_task = NULL;
5873 __mem_cgroup_clear_mc();
5874 spin_lock(&mc.lock);
5878 spin_unlock(&mc.lock);
5883 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5885 struct cgroup_subsys_state *css;
5886 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5887 struct mem_cgroup *from;
5888 struct task_struct *leader, *p;
5889 struct mm_struct *mm;
5890 unsigned long move_flags;
5893 /* charge immigration isn't supported on the default hierarchy */
5894 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5898 * Multi-process migrations only happen on the default hierarchy
5899 * where charge immigration is not used. Perform charge
5900 * immigration if @tset contains a leader and whine if there are
5904 cgroup_taskset_for_each_leader(leader, css, tset) {
5907 memcg = mem_cgroup_from_css(css);
5913 * We are now commited to this value whatever it is. Changes in this
5914 * tunable will only affect upcoming migrations, not the current one.
5915 * So we need to save it, and keep it going.
5917 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5921 from = mem_cgroup_from_task(p);
5923 VM_BUG_ON(from == memcg);
5925 mm = get_task_mm(p);
5928 /* We move charges only when we move a owner of the mm */
5929 if (mm->owner == p) {
5932 VM_BUG_ON(mc.precharge);
5933 VM_BUG_ON(mc.moved_charge);
5934 VM_BUG_ON(mc.moved_swap);
5936 spin_lock(&mc.lock);
5940 mc.flags = move_flags;
5941 spin_unlock(&mc.lock);
5942 /* We set mc.moving_task later */
5944 ret = mem_cgroup_precharge_mc(mm);
5946 mem_cgroup_clear_mc();
5953 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5956 mem_cgroup_clear_mc();
5959 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5960 unsigned long addr, unsigned long end,
5961 struct mm_walk *walk)
5964 struct vm_area_struct *vma = walk->vma;
5967 enum mc_target_type target_type;
5968 union mc_target target;
5971 ptl = pmd_trans_huge_lock(pmd, vma);
5973 if (mc.precharge < HPAGE_PMD_NR) {
5977 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5978 if (target_type == MC_TARGET_PAGE) {
5980 if (!isolate_lru_page(page)) {
5981 if (!mem_cgroup_move_account(page, true,
5983 mc.precharge -= HPAGE_PMD_NR;
5984 mc.moved_charge += HPAGE_PMD_NR;
5986 putback_lru_page(page);
5989 } else if (target_type == MC_TARGET_DEVICE) {
5991 if (!mem_cgroup_move_account(page, true,
5993 mc.precharge -= HPAGE_PMD_NR;
5994 mc.moved_charge += HPAGE_PMD_NR;
6002 if (pmd_trans_unstable(pmd))
6005 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6006 for (; addr != end; addr += PAGE_SIZE) {
6007 pte_t ptent = *(pte++);
6008 bool device = false;
6014 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6015 case MC_TARGET_DEVICE:
6018 case MC_TARGET_PAGE:
6021 * We can have a part of the split pmd here. Moving it
6022 * can be done but it would be too convoluted so simply
6023 * ignore such a partial THP and keep it in original
6024 * memcg. There should be somebody mapping the head.
6026 if (PageTransCompound(page))
6028 if (!device && isolate_lru_page(page))
6030 if (!mem_cgroup_move_account(page, false,
6033 /* we uncharge from mc.from later. */
6037 putback_lru_page(page);
6038 put: /* get_mctgt_type() gets the page */
6041 case MC_TARGET_SWAP:
6043 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6045 mem_cgroup_id_get_many(mc.to, 1);
6046 /* we fixup other refcnts and charges later. */
6054 pte_unmap_unlock(pte - 1, ptl);
6059 * We have consumed all precharges we got in can_attach().
6060 * We try charge one by one, but don't do any additional
6061 * charges to mc.to if we have failed in charge once in attach()
6064 ret = mem_cgroup_do_precharge(1);
6072 static const struct mm_walk_ops charge_walk_ops = {
6073 .pmd_entry = mem_cgroup_move_charge_pte_range,
6076 static void mem_cgroup_move_charge(void)
6078 lru_add_drain_all();
6080 * Signal lock_page_memcg() to take the memcg's move_lock
6081 * while we're moving its pages to another memcg. Then wait
6082 * for already started RCU-only updates to finish.
6084 atomic_inc(&mc.from->moving_account);
6087 if (unlikely(!mmap_read_trylock(mc.mm))) {
6089 * Someone who are holding the mmap_lock might be waiting in
6090 * waitq. So we cancel all extra charges, wake up all waiters,
6091 * and retry. Because we cancel precharges, we might not be able
6092 * to move enough charges, but moving charge is a best-effort
6093 * feature anyway, so it wouldn't be a big problem.
6095 __mem_cgroup_clear_mc();
6100 * When we have consumed all precharges and failed in doing
6101 * additional charge, the page walk just aborts.
6103 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6106 mmap_read_unlock(mc.mm);
6107 atomic_dec(&mc.from->moving_account);
6110 static void mem_cgroup_move_task(void)
6113 mem_cgroup_move_charge();
6114 mem_cgroup_clear_mc();
6117 #else /* !CONFIG_MMU */
6118 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6122 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6125 static void mem_cgroup_move_task(void)
6130 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6132 if (value == PAGE_COUNTER_MAX)
6133 seq_puts(m, "max\n");
6135 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6140 static u64 memory_current_read(struct cgroup_subsys_state *css,
6143 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6145 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6148 static int memory_min_show(struct seq_file *m, void *v)
6150 return seq_puts_memcg_tunable(m,
6151 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6154 static ssize_t memory_min_write(struct kernfs_open_file *of,
6155 char *buf, size_t nbytes, loff_t off)
6157 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6161 buf = strstrip(buf);
6162 err = page_counter_memparse(buf, "max", &min);
6166 page_counter_set_min(&memcg->memory, min);
6171 static int memory_low_show(struct seq_file *m, void *v)
6173 return seq_puts_memcg_tunable(m,
6174 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6177 static ssize_t memory_low_write(struct kernfs_open_file *of,
6178 char *buf, size_t nbytes, loff_t off)
6180 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6184 buf = strstrip(buf);
6185 err = page_counter_memparse(buf, "max", &low);
6189 page_counter_set_low(&memcg->memory, low);
6194 static int memory_high_show(struct seq_file *m, void *v)
6196 return seq_puts_memcg_tunable(m,
6197 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6200 static ssize_t memory_high_write(struct kernfs_open_file *of,
6201 char *buf, size_t nbytes, loff_t off)
6203 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6204 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6205 bool drained = false;
6209 buf = strstrip(buf);
6210 err = page_counter_memparse(buf, "max", &high);
6214 page_counter_set_high(&memcg->memory, high);
6217 unsigned long nr_pages = page_counter_read(&memcg->memory);
6218 unsigned long reclaimed;
6220 if (nr_pages <= high)
6223 if (signal_pending(current))
6227 drain_all_stock(memcg);
6232 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6235 if (!reclaimed && !nr_retries--)
6239 memcg_wb_domain_size_changed(memcg);
6243 static int memory_max_show(struct seq_file *m, void *v)
6245 return seq_puts_memcg_tunable(m,
6246 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6249 static ssize_t memory_max_write(struct kernfs_open_file *of,
6250 char *buf, size_t nbytes, loff_t off)
6252 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6253 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6254 bool drained = false;
6258 buf = strstrip(buf);
6259 err = page_counter_memparse(buf, "max", &max);
6263 xchg(&memcg->memory.max, max);
6266 unsigned long nr_pages = page_counter_read(&memcg->memory);
6268 if (nr_pages <= max)
6271 if (signal_pending(current))
6275 drain_all_stock(memcg);
6281 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6287 memcg_memory_event(memcg, MEMCG_OOM);
6288 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6292 memcg_wb_domain_size_changed(memcg);
6296 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6298 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6299 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6300 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6301 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6302 seq_printf(m, "oom_kill %lu\n",
6303 atomic_long_read(&events[MEMCG_OOM_KILL]));
6306 static int memory_events_show(struct seq_file *m, void *v)
6308 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6310 __memory_events_show(m, memcg->memory_events);
6314 static int memory_events_local_show(struct seq_file *m, void *v)
6316 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6318 __memory_events_show(m, memcg->memory_events_local);
6322 static int memory_stat_show(struct seq_file *m, void *v)
6324 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6327 buf = memory_stat_format(memcg);
6336 static inline unsigned long lruvec_page_state_output(struct lruvec *lruvec,
6339 return lruvec_page_state(lruvec, item) * memcg_page_state_unit(item);
6342 static int memory_numa_stat_show(struct seq_file *m, void *v)
6345 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6347 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6350 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6353 seq_printf(m, "%s", memory_stats[i].name);
6354 for_each_node_state(nid, N_MEMORY) {
6356 struct lruvec *lruvec;
6358 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6359 size = lruvec_page_state_output(lruvec,
6360 memory_stats[i].idx);
6361 seq_printf(m, " N%d=%llu", nid, size);
6370 static int memory_oom_group_show(struct seq_file *m, void *v)
6372 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6374 seq_printf(m, "%d\n", memcg->oom_group);
6379 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6380 char *buf, size_t nbytes, loff_t off)
6382 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6385 buf = strstrip(buf);
6389 ret = kstrtoint(buf, 0, &oom_group);
6393 if (oom_group != 0 && oom_group != 1)
6396 memcg->oom_group = oom_group;
6401 static struct cftype memory_files[] = {
6404 .flags = CFTYPE_NOT_ON_ROOT,
6405 .read_u64 = memory_current_read,
6409 .flags = CFTYPE_NOT_ON_ROOT,
6410 .seq_show = memory_min_show,
6411 .write = memory_min_write,
6415 .flags = CFTYPE_NOT_ON_ROOT,
6416 .seq_show = memory_low_show,
6417 .write = memory_low_write,
6421 .flags = CFTYPE_NOT_ON_ROOT,
6422 .seq_show = memory_high_show,
6423 .write = memory_high_write,
6427 .flags = CFTYPE_NOT_ON_ROOT,
6428 .seq_show = memory_max_show,
6429 .write = memory_max_write,
6433 .flags = CFTYPE_NOT_ON_ROOT,
6434 .file_offset = offsetof(struct mem_cgroup, events_file),
6435 .seq_show = memory_events_show,
6438 .name = "events.local",
6439 .flags = CFTYPE_NOT_ON_ROOT,
6440 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6441 .seq_show = memory_events_local_show,
6445 .seq_show = memory_stat_show,
6449 .name = "numa_stat",
6450 .seq_show = memory_numa_stat_show,
6454 .name = "oom.group",
6455 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6456 .seq_show = memory_oom_group_show,
6457 .write = memory_oom_group_write,
6462 struct cgroup_subsys memory_cgrp_subsys = {
6463 .css_alloc = mem_cgroup_css_alloc,
6464 .css_online = mem_cgroup_css_online,
6465 .css_offline = mem_cgroup_css_offline,
6466 .css_released = mem_cgroup_css_released,
6467 .css_free = mem_cgroup_css_free,
6468 .css_reset = mem_cgroup_css_reset,
6469 .css_rstat_flush = mem_cgroup_css_rstat_flush,
6470 .can_attach = mem_cgroup_can_attach,
6471 .cancel_attach = mem_cgroup_cancel_attach,
6472 .post_attach = mem_cgroup_move_task,
6473 .dfl_cftypes = memory_files,
6474 .legacy_cftypes = mem_cgroup_legacy_files,
6479 * This function calculates an individual cgroup's effective
6480 * protection which is derived from its own memory.min/low, its
6481 * parent's and siblings' settings, as well as the actual memory
6482 * distribution in the tree.
6484 * The following rules apply to the effective protection values:
6486 * 1. At the first level of reclaim, effective protection is equal to
6487 * the declared protection in memory.min and memory.low.
6489 * 2. To enable safe delegation of the protection configuration, at
6490 * subsequent levels the effective protection is capped to the
6491 * parent's effective protection.
6493 * 3. To make complex and dynamic subtrees easier to configure, the
6494 * user is allowed to overcommit the declared protection at a given
6495 * level. If that is the case, the parent's effective protection is
6496 * distributed to the children in proportion to how much protection
6497 * they have declared and how much of it they are utilizing.
6499 * This makes distribution proportional, but also work-conserving:
6500 * if one cgroup claims much more protection than it uses memory,
6501 * the unused remainder is available to its siblings.
6503 * 4. Conversely, when the declared protection is undercommitted at a
6504 * given level, the distribution of the larger parental protection
6505 * budget is NOT proportional. A cgroup's protection from a sibling
6506 * is capped to its own memory.min/low setting.
6508 * 5. However, to allow protecting recursive subtrees from each other
6509 * without having to declare each individual cgroup's fixed share
6510 * of the ancestor's claim to protection, any unutilized -
6511 * "floating" - protection from up the tree is distributed in
6512 * proportion to each cgroup's *usage*. This makes the protection
6513 * neutral wrt sibling cgroups and lets them compete freely over
6514 * the shared parental protection budget, but it protects the
6515 * subtree as a whole from neighboring subtrees.
6517 * Note that 4. and 5. are not in conflict: 4. is about protecting
6518 * against immediate siblings whereas 5. is about protecting against
6519 * neighboring subtrees.
6521 static unsigned long effective_protection(unsigned long usage,
6522 unsigned long parent_usage,
6523 unsigned long setting,
6524 unsigned long parent_effective,
6525 unsigned long siblings_protected)
6527 unsigned long protected;
6530 protected = min(usage, setting);
6532 * If all cgroups at this level combined claim and use more
6533 * protection then what the parent affords them, distribute
6534 * shares in proportion to utilization.
6536 * We are using actual utilization rather than the statically
6537 * claimed protection in order to be work-conserving: claimed
6538 * but unused protection is available to siblings that would
6539 * otherwise get a smaller chunk than what they claimed.
6541 if (siblings_protected > parent_effective)
6542 return protected * parent_effective / siblings_protected;
6545 * Ok, utilized protection of all children is within what the
6546 * parent affords them, so we know whatever this child claims
6547 * and utilizes is effectively protected.
6549 * If there is unprotected usage beyond this value, reclaim
6550 * will apply pressure in proportion to that amount.
6552 * If there is unutilized protection, the cgroup will be fully
6553 * shielded from reclaim, but we do return a smaller value for
6554 * protection than what the group could enjoy in theory. This
6555 * is okay. With the overcommit distribution above, effective
6556 * protection is always dependent on how memory is actually
6557 * consumed among the siblings anyway.
6562 * If the children aren't claiming (all of) the protection
6563 * afforded to them by the parent, distribute the remainder in
6564 * proportion to the (unprotected) memory of each cgroup. That
6565 * way, cgroups that aren't explicitly prioritized wrt each
6566 * other compete freely over the allowance, but they are
6567 * collectively protected from neighboring trees.
6569 * We're using unprotected memory for the weight so that if
6570 * some cgroups DO claim explicit protection, we don't protect
6571 * the same bytes twice.
6573 * Check both usage and parent_usage against the respective
6574 * protected values. One should imply the other, but they
6575 * aren't read atomically - make sure the division is sane.
6577 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6579 if (parent_effective > siblings_protected &&
6580 parent_usage > siblings_protected &&
6581 usage > protected) {
6582 unsigned long unclaimed;
6584 unclaimed = parent_effective - siblings_protected;
6585 unclaimed *= usage - protected;
6586 unclaimed /= parent_usage - siblings_protected;
6595 * mem_cgroup_protected - check if memory consumption is in the normal range
6596 * @root: the top ancestor of the sub-tree being checked
6597 * @memcg: the memory cgroup to check
6599 * WARNING: This function is not stateless! It can only be used as part
6600 * of a top-down tree iteration, not for isolated queries.
6602 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6603 struct mem_cgroup *memcg)
6605 unsigned long usage, parent_usage;
6606 struct mem_cgroup *parent;
6608 if (mem_cgroup_disabled())
6612 root = root_mem_cgroup;
6615 * Effective values of the reclaim targets are ignored so they
6616 * can be stale. Have a look at mem_cgroup_protection for more
6618 * TODO: calculation should be more robust so that we do not need
6619 * that special casing.
6624 usage = page_counter_read(&memcg->memory);
6628 parent = parent_mem_cgroup(memcg);
6629 /* No parent means a non-hierarchical mode on v1 memcg */
6633 if (parent == root) {
6634 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6635 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6639 parent_usage = page_counter_read(&parent->memory);
6641 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6642 READ_ONCE(memcg->memory.min),
6643 READ_ONCE(parent->memory.emin),
6644 atomic_long_read(&parent->memory.children_min_usage)));
6646 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6647 READ_ONCE(memcg->memory.low),
6648 READ_ONCE(parent->memory.elow),
6649 atomic_long_read(&parent->memory.children_low_usage)));
6652 static int __mem_cgroup_charge(struct page *page, struct mem_cgroup *memcg,
6655 unsigned int nr_pages = thp_nr_pages(page);
6658 ret = try_charge(memcg, gfp, nr_pages);
6662 css_get(&memcg->css);
6663 commit_charge(page, memcg);
6665 local_irq_disable();
6666 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6667 memcg_check_events(memcg, page);
6674 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6675 * @page: page to charge
6676 * @mm: mm context of the victim
6677 * @gfp_mask: reclaim mode
6679 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6680 * pages according to @gfp_mask if necessary.
6682 * Do not use this for pages allocated for swapin.
6684 * Returns 0 on success. Otherwise, an error code is returned.
6686 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6688 struct mem_cgroup *memcg;
6691 if (mem_cgroup_disabled())
6694 memcg = get_mem_cgroup_from_mm(mm);
6695 ret = __mem_cgroup_charge(page, memcg, gfp_mask);
6696 css_put(&memcg->css);
6702 * mem_cgroup_swapin_charge_page - charge a newly allocated page for swapin
6703 * @page: page to charge
6704 * @mm: mm context of the victim
6705 * @gfp: reclaim mode
6706 * @entry: swap entry for which the page is allocated
6708 * This function charges a page allocated for swapin. Please call this before
6709 * adding the page to the swapcache.
6711 * Returns 0 on success. Otherwise, an error code is returned.
6713 int mem_cgroup_swapin_charge_page(struct page *page, struct mm_struct *mm,
6714 gfp_t gfp, swp_entry_t entry)
6716 struct mem_cgroup *memcg;
6720 if (mem_cgroup_disabled())
6723 id = lookup_swap_cgroup_id(entry);
6725 memcg = mem_cgroup_from_id(id);
6726 if (!memcg || !css_tryget_online(&memcg->css))
6727 memcg = get_mem_cgroup_from_mm(mm);
6730 ret = __mem_cgroup_charge(page, memcg, gfp);
6732 css_put(&memcg->css);
6737 * mem_cgroup_swapin_uncharge_swap - uncharge swap slot
6738 * @entry: swap entry for which the page is charged
6740 * Call this function after successfully adding the charged page to swapcache.
6742 * Note: This function assumes the page for which swap slot is being uncharged
6745 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
6748 * Cgroup1's unified memory+swap counter has been charged with the
6749 * new swapcache page, finish the transfer by uncharging the swap
6750 * slot. The swap slot would also get uncharged when it dies, but
6751 * it can stick around indefinitely and we'd count the page twice
6754 * Cgroup2 has separate resource counters for memory and swap,
6755 * so this is a non-issue here. Memory and swap charge lifetimes
6756 * correspond 1:1 to page and swap slot lifetimes: we charge the
6757 * page to memory here, and uncharge swap when the slot is freed.
6759 if (!mem_cgroup_disabled() && do_memsw_account()) {
6761 * The swap entry might not get freed for a long time,
6762 * let's not wait for it. The page already received a
6763 * memory+swap charge, drop the swap entry duplicate.
6765 mem_cgroup_uncharge_swap(entry, 1);
6769 struct uncharge_gather {
6770 struct mem_cgroup *memcg;
6771 unsigned long nr_memory;
6772 unsigned long pgpgout;
6773 unsigned long nr_kmem;
6774 struct page *dummy_page;
6777 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6779 memset(ug, 0, sizeof(*ug));
6782 static void uncharge_batch(const struct uncharge_gather *ug)
6784 unsigned long flags;
6786 if (ug->nr_memory) {
6787 page_counter_uncharge(&ug->memcg->memory, ug->nr_memory);
6788 if (do_memsw_account())
6789 page_counter_uncharge(&ug->memcg->memsw, ug->nr_memory);
6790 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6791 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6792 memcg_oom_recover(ug->memcg);
6795 local_irq_save(flags);
6796 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6797 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_memory);
6798 memcg_check_events(ug->memcg, ug->dummy_page);
6799 local_irq_restore(flags);
6801 /* drop reference from uncharge_page */
6802 css_put(&ug->memcg->css);
6805 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6807 unsigned long nr_pages;
6808 struct mem_cgroup *memcg;
6809 struct obj_cgroup *objcg;
6811 VM_BUG_ON_PAGE(PageLRU(page), page);
6814 * Nobody should be changing or seriously looking at
6815 * page memcg or objcg at this point, we have fully
6816 * exclusive access to the page.
6818 if (PageMemcgKmem(page)) {
6819 objcg = __page_objcg(page);
6821 * This get matches the put at the end of the function and
6822 * kmem pages do not hold memcg references anymore.
6824 memcg = get_mem_cgroup_from_objcg(objcg);
6826 memcg = __page_memcg(page);
6832 if (ug->memcg != memcg) {
6835 uncharge_gather_clear(ug);
6838 ug->dummy_page = page;
6840 /* pairs with css_put in uncharge_batch */
6841 css_get(&memcg->css);
6844 nr_pages = compound_nr(page);
6846 if (PageMemcgKmem(page)) {
6847 ug->nr_memory += nr_pages;
6848 ug->nr_kmem += nr_pages;
6850 page->memcg_data = 0;
6851 obj_cgroup_put(objcg);
6853 /* LRU pages aren't accounted at the root level */
6854 if (!mem_cgroup_is_root(memcg))
6855 ug->nr_memory += nr_pages;
6858 page->memcg_data = 0;
6861 css_put(&memcg->css);
6865 * mem_cgroup_uncharge - uncharge a page
6866 * @page: page to uncharge
6868 * Uncharge a page previously charged with mem_cgroup_charge().
6870 void mem_cgroup_uncharge(struct page *page)
6872 struct uncharge_gather ug;
6874 if (mem_cgroup_disabled())
6877 /* Don't touch page->lru of any random page, pre-check: */
6878 if (!page_memcg(page))
6881 uncharge_gather_clear(&ug);
6882 uncharge_page(page, &ug);
6883 uncharge_batch(&ug);
6887 * mem_cgroup_uncharge_list - uncharge a list of page
6888 * @page_list: list of pages to uncharge
6890 * Uncharge a list of pages previously charged with
6891 * mem_cgroup_charge().
6893 void mem_cgroup_uncharge_list(struct list_head *page_list)
6895 struct uncharge_gather ug;
6898 if (mem_cgroup_disabled())
6901 uncharge_gather_clear(&ug);
6902 list_for_each_entry(page, page_list, lru)
6903 uncharge_page(page, &ug);
6905 uncharge_batch(&ug);
6909 * mem_cgroup_migrate - charge a page's replacement
6910 * @oldpage: currently circulating page
6911 * @newpage: replacement page
6913 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6914 * be uncharged upon free.
6916 * Both pages must be locked, @newpage->mapping must be set up.
6918 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6920 struct mem_cgroup *memcg;
6921 unsigned int nr_pages;
6922 unsigned long flags;
6924 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6925 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6926 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6927 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6930 if (mem_cgroup_disabled())
6933 /* Page cache replacement: new page already charged? */
6934 if (page_memcg(newpage))
6937 memcg = page_memcg(oldpage);
6938 VM_WARN_ON_ONCE_PAGE(!memcg, oldpage);
6942 /* Force-charge the new page. The old one will be freed soon */
6943 nr_pages = thp_nr_pages(newpage);
6945 page_counter_charge(&memcg->memory, nr_pages);
6946 if (do_memsw_account())
6947 page_counter_charge(&memcg->memsw, nr_pages);
6949 css_get(&memcg->css);
6950 commit_charge(newpage, memcg);
6952 local_irq_save(flags);
6953 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6954 memcg_check_events(memcg, newpage);
6955 local_irq_restore(flags);
6958 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6959 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6961 void mem_cgroup_sk_alloc(struct sock *sk)
6963 struct mem_cgroup *memcg;
6965 if (!mem_cgroup_sockets_enabled)
6968 /* Do not associate the sock with unrelated interrupted task's memcg. */
6973 memcg = mem_cgroup_from_task(current);
6974 if (memcg == root_mem_cgroup)
6976 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6978 if (css_tryget(&memcg->css))
6979 sk->sk_memcg = memcg;
6984 void mem_cgroup_sk_free(struct sock *sk)
6987 css_put(&sk->sk_memcg->css);
6991 * mem_cgroup_charge_skmem - charge socket memory
6992 * @memcg: memcg to charge
6993 * @nr_pages: number of pages to charge
6995 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6996 * @memcg's configured limit, %false if the charge had to be forced.
6998 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7000 gfp_t gfp_mask = GFP_KERNEL;
7002 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7003 struct page_counter *fail;
7005 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
7006 memcg->tcpmem_pressure = 0;
7009 page_counter_charge(&memcg->tcpmem, nr_pages);
7010 memcg->tcpmem_pressure = 1;
7014 /* Don't block in the packet receive path */
7016 gfp_mask = GFP_NOWAIT;
7018 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7020 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7023 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7028 * mem_cgroup_uncharge_skmem - uncharge socket memory
7029 * @memcg: memcg to uncharge
7030 * @nr_pages: number of pages to uncharge
7032 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7034 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7035 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7039 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7041 refill_stock(memcg, nr_pages);
7044 static int __init cgroup_memory(char *s)
7048 while ((token = strsep(&s, ",")) != NULL) {
7051 if (!strcmp(token, "nosocket"))
7052 cgroup_memory_nosocket = true;
7053 if (!strcmp(token, "nokmem"))
7054 cgroup_memory_nokmem = true;
7058 __setup("cgroup.memory=", cgroup_memory);
7061 * subsys_initcall() for memory controller.
7063 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7064 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7065 * basically everything that doesn't depend on a specific mem_cgroup structure
7066 * should be initialized from here.
7068 static int __init mem_cgroup_init(void)
7073 * Currently s32 type (can refer to struct batched_lruvec_stat) is
7074 * used for per-memcg-per-cpu caching of per-node statistics. In order
7075 * to work fine, we should make sure that the overfill threshold can't
7076 * exceed S32_MAX / PAGE_SIZE.
7078 BUILD_BUG_ON(MEMCG_CHARGE_BATCH > S32_MAX / PAGE_SIZE);
7080 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7081 memcg_hotplug_cpu_dead);
7083 for_each_possible_cpu(cpu)
7084 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7087 for_each_node(node) {
7088 struct mem_cgroup_tree_per_node *rtpn;
7090 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7091 node_online(node) ? node : NUMA_NO_NODE);
7093 rtpn->rb_root = RB_ROOT;
7094 rtpn->rb_rightmost = NULL;
7095 spin_lock_init(&rtpn->lock);
7096 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7101 subsys_initcall(mem_cgroup_init);
7103 #ifdef CONFIG_MEMCG_SWAP
7104 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7106 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7108 * The root cgroup cannot be destroyed, so it's refcount must
7111 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7115 memcg = parent_mem_cgroup(memcg);
7117 memcg = root_mem_cgroup;
7123 * mem_cgroup_swapout - transfer a memsw charge to swap
7124 * @page: page whose memsw charge to transfer
7125 * @entry: swap entry to move the charge to
7127 * Transfer the memsw charge of @page to @entry.
7129 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7131 struct mem_cgroup *memcg, *swap_memcg;
7132 unsigned int nr_entries;
7133 unsigned short oldid;
7135 VM_BUG_ON_PAGE(PageLRU(page), page);
7136 VM_BUG_ON_PAGE(page_count(page), page);
7138 if (mem_cgroup_disabled())
7141 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7144 memcg = page_memcg(page);
7146 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7151 * In case the memcg owning these pages has been offlined and doesn't
7152 * have an ID allocated to it anymore, charge the closest online
7153 * ancestor for the swap instead and transfer the memory+swap charge.
7155 swap_memcg = mem_cgroup_id_get_online(memcg);
7156 nr_entries = thp_nr_pages(page);
7157 /* Get references for the tail pages, too */
7159 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7160 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7162 VM_BUG_ON_PAGE(oldid, page);
7163 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7165 page->memcg_data = 0;
7167 if (!mem_cgroup_is_root(memcg))
7168 page_counter_uncharge(&memcg->memory, nr_entries);
7170 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7171 if (!mem_cgroup_is_root(swap_memcg))
7172 page_counter_charge(&swap_memcg->memsw, nr_entries);
7173 page_counter_uncharge(&memcg->memsw, nr_entries);
7177 * Interrupts should be disabled here because the caller holds the
7178 * i_pages lock which is taken with interrupts-off. It is
7179 * important here to have the interrupts disabled because it is the
7180 * only synchronisation we have for updating the per-CPU variables.
7182 VM_BUG_ON(!irqs_disabled());
7183 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7184 memcg_check_events(memcg, page);
7186 css_put(&memcg->css);
7190 * mem_cgroup_try_charge_swap - try charging swap space for a page
7191 * @page: page being added to swap
7192 * @entry: swap entry to charge
7194 * Try to charge @page's memcg for the swap space at @entry.
7196 * Returns 0 on success, -ENOMEM on failure.
7198 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7200 unsigned int nr_pages = thp_nr_pages(page);
7201 struct page_counter *counter;
7202 struct mem_cgroup *memcg;
7203 unsigned short oldid;
7205 if (mem_cgroup_disabled())
7208 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7211 memcg = page_memcg(page);
7213 VM_WARN_ON_ONCE_PAGE(!memcg, page);
7218 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7222 memcg = mem_cgroup_id_get_online(memcg);
7224 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7225 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7226 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7227 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7228 mem_cgroup_id_put(memcg);
7232 /* Get references for the tail pages, too */
7234 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7235 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7236 VM_BUG_ON_PAGE(oldid, page);
7237 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7243 * mem_cgroup_uncharge_swap - uncharge swap space
7244 * @entry: swap entry to uncharge
7245 * @nr_pages: the amount of swap space to uncharge
7247 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7249 struct mem_cgroup *memcg;
7252 id = swap_cgroup_record(entry, 0, nr_pages);
7254 memcg = mem_cgroup_from_id(id);
7256 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7257 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7258 page_counter_uncharge(&memcg->swap, nr_pages);
7260 page_counter_uncharge(&memcg->memsw, nr_pages);
7262 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7263 mem_cgroup_id_put_many(memcg, nr_pages);
7268 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7270 long nr_swap_pages = get_nr_swap_pages();
7272 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7273 return nr_swap_pages;
7274 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7275 nr_swap_pages = min_t(long, nr_swap_pages,
7276 READ_ONCE(memcg->swap.max) -
7277 page_counter_read(&memcg->swap));
7278 return nr_swap_pages;
7281 bool mem_cgroup_swap_full(struct page *page)
7283 struct mem_cgroup *memcg;
7285 VM_BUG_ON_PAGE(!PageLocked(page), page);
7289 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7292 memcg = page_memcg(page);
7296 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7297 unsigned long usage = page_counter_read(&memcg->swap);
7299 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7300 usage * 2 >= READ_ONCE(memcg->swap.max))
7307 static int __init setup_swap_account(char *s)
7309 if (!strcmp(s, "1"))
7310 cgroup_memory_noswap = false;
7311 else if (!strcmp(s, "0"))
7312 cgroup_memory_noswap = true;
7315 __setup("swapaccount=", setup_swap_account);
7317 static u64 swap_current_read(struct cgroup_subsys_state *css,
7320 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7322 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7325 static int swap_high_show(struct seq_file *m, void *v)
7327 return seq_puts_memcg_tunable(m,
7328 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7331 static ssize_t swap_high_write(struct kernfs_open_file *of,
7332 char *buf, size_t nbytes, loff_t off)
7334 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7338 buf = strstrip(buf);
7339 err = page_counter_memparse(buf, "max", &high);
7343 page_counter_set_high(&memcg->swap, high);
7348 static int swap_max_show(struct seq_file *m, void *v)
7350 return seq_puts_memcg_tunable(m,
7351 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7354 static ssize_t swap_max_write(struct kernfs_open_file *of,
7355 char *buf, size_t nbytes, loff_t off)
7357 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7361 buf = strstrip(buf);
7362 err = page_counter_memparse(buf, "max", &max);
7366 xchg(&memcg->swap.max, max);
7371 static int swap_events_show(struct seq_file *m, void *v)
7373 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7375 seq_printf(m, "high %lu\n",
7376 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7377 seq_printf(m, "max %lu\n",
7378 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7379 seq_printf(m, "fail %lu\n",
7380 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7385 static struct cftype swap_files[] = {
7387 .name = "swap.current",
7388 .flags = CFTYPE_NOT_ON_ROOT,
7389 .read_u64 = swap_current_read,
7392 .name = "swap.high",
7393 .flags = CFTYPE_NOT_ON_ROOT,
7394 .seq_show = swap_high_show,
7395 .write = swap_high_write,
7399 .flags = CFTYPE_NOT_ON_ROOT,
7400 .seq_show = swap_max_show,
7401 .write = swap_max_write,
7404 .name = "swap.events",
7405 .flags = CFTYPE_NOT_ON_ROOT,
7406 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7407 .seq_show = swap_events_show,
7412 static struct cftype memsw_files[] = {
7414 .name = "memsw.usage_in_bytes",
7415 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7416 .read_u64 = mem_cgroup_read_u64,
7419 .name = "memsw.max_usage_in_bytes",
7420 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7421 .write = mem_cgroup_reset,
7422 .read_u64 = mem_cgroup_read_u64,
7425 .name = "memsw.limit_in_bytes",
7426 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7427 .write = mem_cgroup_write,
7428 .read_u64 = mem_cgroup_read_u64,
7431 .name = "memsw.failcnt",
7432 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7433 .write = mem_cgroup_reset,
7434 .read_u64 = mem_cgroup_read_u64,
7436 { }, /* terminate */
7440 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7441 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7442 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7443 * boot parameter. This may result in premature OOPS inside
7444 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7446 static int __init mem_cgroup_swap_init(void)
7448 /* No memory control -> no swap control */
7449 if (mem_cgroup_disabled())
7450 cgroup_memory_noswap = true;
7452 if (cgroup_memory_noswap)
7455 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7456 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7460 core_initcall(mem_cgroup_swap_init);
7462 #endif /* CONFIG_MEMCG_SWAP */