1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* memcontrol.c - Memory Controller
4 * Copyright IBM Corporation, 2007
5 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
7 * Copyright 2007 OpenVZ SWsoft Inc
8 * Author: Pavel Emelianov <xemul@openvz.org>
11 * Copyright (C) 2009 Nokia Corporation
12 * Author: Kirill A. Shutemov
14 * Kernel Memory Controller
15 * Copyright (C) 2012 Parallels Inc. and Google Inc.
16 * Authors: Glauber Costa and Suleiman Souhlal
19 * Charge lifetime sanitation
20 * Lockless page tracking & accounting
21 * Unified hierarchy configuration model
22 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
25 #include <linux/page_counter.h>
26 #include <linux/memcontrol.h>
27 #include <linux/cgroup.h>
28 #include <linux/pagewalk.h>
29 #include <linux/sched/mm.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/hugetlb.h>
32 #include <linux/pagemap.h>
33 #include <linux/vm_event_item.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
59 #include <linux/tracehook.h>
60 #include <linux/psi.h>
61 #include <linux/seq_buf.h>
67 #include <linux/uaccess.h>
69 #include <trace/events/vmscan.h>
71 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
72 EXPORT_SYMBOL(memory_cgrp_subsys);
74 struct mem_cgroup *root_mem_cgroup __read_mostly;
76 /* Socket memory accounting disabled? */
77 static bool cgroup_memory_nosocket;
79 /* Kernel memory accounting disabled? */
80 static bool cgroup_memory_nokmem;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 bool cgroup_memory_noswap __read_mostly;
86 #define cgroup_memory_noswap 1
89 #ifdef CONFIG_CGROUP_WRITEBACK
90 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
93 /* Whether legacy memory+swap accounting is active */
94 static bool do_memsw_account(void)
96 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
99 #define THRESHOLDS_EVENTS_TARGET 128
100 #define SOFTLIMIT_EVENTS_TARGET 1024
103 * Cgroups above their limits are maintained in a RB-Tree, independent of
104 * their hierarchy representation
107 struct mem_cgroup_tree_per_node {
108 struct rb_root rb_root;
109 struct rb_node *rb_rightmost;
113 struct mem_cgroup_tree {
114 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
117 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
120 struct mem_cgroup_eventfd_list {
121 struct list_head list;
122 struct eventfd_ctx *eventfd;
126 * cgroup_event represents events which userspace want to receive.
128 struct mem_cgroup_event {
130 * memcg which the event belongs to.
132 struct mem_cgroup *memcg;
134 * eventfd to signal userspace about the event.
136 struct eventfd_ctx *eventfd;
138 * Each of these stored in a list by the cgroup.
140 struct list_head list;
142 * register_event() callback will be used to add new userspace
143 * waiter for changes related to this event. Use eventfd_signal()
144 * on eventfd to send notification to userspace.
146 int (*register_event)(struct mem_cgroup *memcg,
147 struct eventfd_ctx *eventfd, const char *args);
149 * unregister_event() callback will be called when userspace closes
150 * the eventfd or on cgroup removing. This callback must be set,
151 * if you want provide notification functionality.
153 void (*unregister_event)(struct mem_cgroup *memcg,
154 struct eventfd_ctx *eventfd);
156 * All fields below needed to unregister event when
157 * userspace closes eventfd.
160 wait_queue_head_t *wqh;
161 wait_queue_entry_t wait;
162 struct work_struct remove;
165 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
166 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
168 /* Stuffs for move charges at task migration. */
170 * Types of charges to be moved.
172 #define MOVE_ANON 0x1U
173 #define MOVE_FILE 0x2U
174 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
176 /* "mc" and its members are protected by cgroup_mutex */
177 static struct move_charge_struct {
178 spinlock_t lock; /* for from, to */
179 struct mm_struct *mm;
180 struct mem_cgroup *from;
181 struct mem_cgroup *to;
183 unsigned long precharge;
184 unsigned long moved_charge;
185 unsigned long moved_swap;
186 struct task_struct *moving_task; /* a task moving charges */
187 wait_queue_head_t waitq; /* a waitq for other context */
189 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
190 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
194 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
195 * limit reclaim to prevent infinite loops, if they ever occur.
197 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
198 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
201 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
202 MEM_CGROUP_CHARGE_TYPE_ANON,
203 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
204 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
208 /* for encoding cft->private value on file */
217 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
218 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
219 #define MEMFILE_ATTR(val) ((val) & 0xffff)
220 /* Used for OOM nofiier */
221 #define OOM_CONTROL (0)
224 * Iteration constructs for visiting all cgroups (under a tree). If
225 * loops are exited prematurely (break), mem_cgroup_iter_break() must
226 * be used for reference counting.
228 #define for_each_mem_cgroup_tree(iter, root) \
229 for (iter = mem_cgroup_iter(root, NULL, NULL); \
231 iter = mem_cgroup_iter(root, iter, NULL))
233 #define for_each_mem_cgroup(iter) \
234 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
236 iter = mem_cgroup_iter(NULL, iter, NULL))
238 static inline bool should_force_charge(void)
240 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
241 (current->flags & PF_EXITING);
244 /* Some nice accessors for the vmpressure. */
245 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
248 memcg = root_mem_cgroup;
249 return &memcg->vmpressure;
252 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
254 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
257 #ifdef CONFIG_MEMCG_KMEM
258 extern spinlock_t css_set_lock;
260 static void obj_cgroup_release(struct percpu_ref *ref)
262 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
263 struct mem_cgroup *memcg;
264 unsigned int nr_bytes;
265 unsigned int nr_pages;
269 * At this point all allocated objects are freed, and
270 * objcg->nr_charged_bytes can't have an arbitrary byte value.
271 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
273 * The following sequence can lead to it:
274 * 1) CPU0: objcg == stock->cached_objcg
275 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
276 * PAGE_SIZE bytes are charged
277 * 3) CPU1: a process from another memcg is allocating something,
278 * the stock if flushed,
279 * objcg->nr_charged_bytes = PAGE_SIZE - 92
280 * 5) CPU0: we do release this object,
281 * 92 bytes are added to stock->nr_bytes
282 * 6) CPU0: stock is flushed,
283 * 92 bytes are added to objcg->nr_charged_bytes
285 * In the result, nr_charged_bytes == PAGE_SIZE.
286 * This page will be uncharged in obj_cgroup_release().
288 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
289 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
290 nr_pages = nr_bytes >> PAGE_SHIFT;
292 spin_lock_irqsave(&css_set_lock, flags);
293 memcg = obj_cgroup_memcg(objcg);
295 __memcg_kmem_uncharge(memcg, nr_pages);
296 list_del(&objcg->list);
297 mem_cgroup_put(memcg);
298 spin_unlock_irqrestore(&css_set_lock, flags);
300 percpu_ref_exit(ref);
301 kfree_rcu(objcg, rcu);
304 static struct obj_cgroup *obj_cgroup_alloc(void)
306 struct obj_cgroup *objcg;
309 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
313 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
319 INIT_LIST_HEAD(&objcg->list);
323 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
324 struct mem_cgroup *parent)
326 struct obj_cgroup *objcg, *iter;
328 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
330 spin_lock_irq(&css_set_lock);
332 /* Move active objcg to the parent's list */
333 xchg(&objcg->memcg, parent);
334 css_get(&parent->css);
335 list_add(&objcg->list, &parent->objcg_list);
337 /* Move already reparented objcgs to the parent's list */
338 list_for_each_entry(iter, &memcg->objcg_list, list) {
339 css_get(&parent->css);
340 xchg(&iter->memcg, parent);
341 css_put(&memcg->css);
343 list_splice(&memcg->objcg_list, &parent->objcg_list);
345 spin_unlock_irq(&css_set_lock);
347 percpu_ref_kill(&objcg->refcnt);
351 * This will be used as a shrinker list's index.
352 * The main reason for not using cgroup id for this:
353 * this works better in sparse environments, where we have a lot of memcgs,
354 * but only a few kmem-limited. Or also, if we have, for instance, 200
355 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
356 * 200 entry array for that.
358 * The current size of the caches array is stored in memcg_nr_cache_ids. It
359 * will double each time we have to increase it.
361 static DEFINE_IDA(memcg_cache_ida);
362 int memcg_nr_cache_ids;
364 /* Protects memcg_nr_cache_ids */
365 static DECLARE_RWSEM(memcg_cache_ids_sem);
367 void memcg_get_cache_ids(void)
369 down_read(&memcg_cache_ids_sem);
372 void memcg_put_cache_ids(void)
374 up_read(&memcg_cache_ids_sem);
378 * MIN_SIZE is different than 1, because we would like to avoid going through
379 * the alloc/free process all the time. In a small machine, 4 kmem-limited
380 * cgroups is a reasonable guess. In the future, it could be a parameter or
381 * tunable, but that is strictly not necessary.
383 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
384 * this constant directly from cgroup, but it is understandable that this is
385 * better kept as an internal representation in cgroup.c. In any case, the
386 * cgrp_id space is not getting any smaller, and we don't have to necessarily
387 * increase ours as well if it increases.
389 #define MEMCG_CACHES_MIN_SIZE 4
390 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
393 * A lot of the calls to the cache allocation functions are expected to be
394 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
395 * conditional to this static branch, we'll have to allow modules that does
396 * kmem_cache_alloc and the such to see this symbol as well
398 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
399 EXPORT_SYMBOL(memcg_kmem_enabled_key);
402 static int memcg_shrinker_map_size;
403 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
405 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
407 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
410 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
411 int size, int old_size)
413 struct memcg_shrinker_map *new, *old;
416 lockdep_assert_held(&memcg_shrinker_map_mutex);
419 old = rcu_dereference_protected(
420 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
421 /* Not yet online memcg */
425 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
429 /* Set all old bits, clear all new bits */
430 memset(new->map, (int)0xff, old_size);
431 memset((void *)new->map + old_size, 0, size - old_size);
433 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
434 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
440 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
442 struct mem_cgroup_per_node *pn;
443 struct memcg_shrinker_map *map;
446 if (mem_cgroup_is_root(memcg))
450 pn = mem_cgroup_nodeinfo(memcg, nid);
451 map = rcu_dereference_protected(pn->shrinker_map, true);
454 rcu_assign_pointer(pn->shrinker_map, NULL);
458 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
460 struct memcg_shrinker_map *map;
461 int nid, size, ret = 0;
463 if (mem_cgroup_is_root(memcg))
466 mutex_lock(&memcg_shrinker_map_mutex);
467 size = memcg_shrinker_map_size;
469 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
471 memcg_free_shrinker_maps(memcg);
475 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
477 mutex_unlock(&memcg_shrinker_map_mutex);
482 int memcg_expand_shrinker_maps(int new_id)
484 int size, old_size, ret = 0;
485 struct mem_cgroup *memcg;
487 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
488 old_size = memcg_shrinker_map_size;
489 if (size <= old_size)
492 mutex_lock(&memcg_shrinker_map_mutex);
493 if (!root_mem_cgroup)
496 for_each_mem_cgroup(memcg) {
497 if (mem_cgroup_is_root(memcg))
499 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
501 mem_cgroup_iter_break(NULL, memcg);
507 memcg_shrinker_map_size = size;
508 mutex_unlock(&memcg_shrinker_map_mutex);
512 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
514 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
515 struct memcg_shrinker_map *map;
518 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
519 /* Pairs with smp mb in shrink_slab() */
520 smp_mb__before_atomic();
521 set_bit(shrinker_id, map->map);
527 * mem_cgroup_css_from_page - css of the memcg associated with a page
528 * @page: page of interest
530 * If memcg is bound to the default hierarchy, css of the memcg associated
531 * with @page is returned. The returned css remains associated with @page
532 * until it is released.
534 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
537 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
539 struct mem_cgroup *memcg;
541 memcg = page->mem_cgroup;
543 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
544 memcg = root_mem_cgroup;
550 * page_cgroup_ino - return inode number of the memcg a page is charged to
553 * Look up the closest online ancestor of the memory cgroup @page is charged to
554 * and return its inode number or 0 if @page is not charged to any cgroup. It
555 * is safe to call this function without holding a reference to @page.
557 * Note, this function is inherently racy, because there is nothing to prevent
558 * the cgroup inode from getting torn down and potentially reallocated a moment
559 * after page_cgroup_ino() returns, so it only should be used by callers that
560 * do not care (such as procfs interfaces).
562 ino_t page_cgroup_ino(struct page *page)
564 struct mem_cgroup *memcg;
565 unsigned long ino = 0;
568 memcg = page->mem_cgroup;
571 * The lowest bit set means that memcg isn't a valid
572 * memcg pointer, but a obj_cgroups pointer.
573 * In this case the page is shared and doesn't belong
574 * to any specific memory cgroup.
576 if ((unsigned long) memcg & 0x1UL)
579 while (memcg && !(memcg->css.flags & CSS_ONLINE))
580 memcg = parent_mem_cgroup(memcg);
582 ino = cgroup_ino(memcg->css.cgroup);
587 static struct mem_cgroup_per_node *
588 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
590 int nid = page_to_nid(page);
592 return memcg->nodeinfo[nid];
595 static struct mem_cgroup_tree_per_node *
596 soft_limit_tree_node(int nid)
598 return soft_limit_tree.rb_tree_per_node[nid];
601 static struct mem_cgroup_tree_per_node *
602 soft_limit_tree_from_page(struct page *page)
604 int nid = page_to_nid(page);
606 return soft_limit_tree.rb_tree_per_node[nid];
609 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
610 struct mem_cgroup_tree_per_node *mctz,
611 unsigned long new_usage_in_excess)
613 struct rb_node **p = &mctz->rb_root.rb_node;
614 struct rb_node *parent = NULL;
615 struct mem_cgroup_per_node *mz_node;
616 bool rightmost = true;
621 mz->usage_in_excess = new_usage_in_excess;
622 if (!mz->usage_in_excess)
626 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
628 if (mz->usage_in_excess < mz_node->usage_in_excess) {
634 * We can't avoid mem cgroups that are over their soft
635 * limit by the same amount
637 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
642 mctz->rb_rightmost = &mz->tree_node;
644 rb_link_node(&mz->tree_node, parent, p);
645 rb_insert_color(&mz->tree_node, &mctz->rb_root);
649 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
650 struct mem_cgroup_tree_per_node *mctz)
655 if (&mz->tree_node == mctz->rb_rightmost)
656 mctz->rb_rightmost = rb_prev(&mz->tree_node);
658 rb_erase(&mz->tree_node, &mctz->rb_root);
662 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
663 struct mem_cgroup_tree_per_node *mctz)
667 spin_lock_irqsave(&mctz->lock, flags);
668 __mem_cgroup_remove_exceeded(mz, mctz);
669 spin_unlock_irqrestore(&mctz->lock, flags);
672 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
674 unsigned long nr_pages = page_counter_read(&memcg->memory);
675 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
676 unsigned long excess = 0;
678 if (nr_pages > soft_limit)
679 excess = nr_pages - soft_limit;
684 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
686 unsigned long excess;
687 struct mem_cgroup_per_node *mz;
688 struct mem_cgroup_tree_per_node *mctz;
690 mctz = soft_limit_tree_from_page(page);
694 * Necessary to update all ancestors when hierarchy is used.
695 * because their event counter is not touched.
697 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
698 mz = mem_cgroup_page_nodeinfo(memcg, page);
699 excess = soft_limit_excess(memcg);
701 * We have to update the tree if mz is on RB-tree or
702 * mem is over its softlimit.
704 if (excess || mz->on_tree) {
707 spin_lock_irqsave(&mctz->lock, flags);
708 /* if on-tree, remove it */
710 __mem_cgroup_remove_exceeded(mz, mctz);
712 * Insert again. mz->usage_in_excess will be updated.
713 * If excess is 0, no tree ops.
715 __mem_cgroup_insert_exceeded(mz, mctz, excess);
716 spin_unlock_irqrestore(&mctz->lock, flags);
721 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
723 struct mem_cgroup_tree_per_node *mctz;
724 struct mem_cgroup_per_node *mz;
728 mz = mem_cgroup_nodeinfo(memcg, nid);
729 mctz = soft_limit_tree_node(nid);
731 mem_cgroup_remove_exceeded(mz, mctz);
735 static struct mem_cgroup_per_node *
736 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
738 struct mem_cgroup_per_node *mz;
742 if (!mctz->rb_rightmost)
743 goto done; /* Nothing to reclaim from */
745 mz = rb_entry(mctz->rb_rightmost,
746 struct mem_cgroup_per_node, tree_node);
748 * Remove the node now but someone else can add it back,
749 * we will to add it back at the end of reclaim to its correct
750 * position in the tree.
752 __mem_cgroup_remove_exceeded(mz, mctz);
753 if (!soft_limit_excess(mz->memcg) ||
754 !css_tryget(&mz->memcg->css))
760 static struct mem_cgroup_per_node *
761 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
763 struct mem_cgroup_per_node *mz;
765 spin_lock_irq(&mctz->lock);
766 mz = __mem_cgroup_largest_soft_limit_node(mctz);
767 spin_unlock_irq(&mctz->lock);
772 * __mod_memcg_state - update cgroup memory statistics
773 * @memcg: the memory cgroup
774 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
775 * @val: delta to add to the counter, can be negative
777 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
779 long x, threshold = MEMCG_CHARGE_BATCH;
781 if (mem_cgroup_disabled())
784 if (memcg_stat_item_in_bytes(idx))
785 threshold <<= PAGE_SHIFT;
787 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
788 if (unlikely(abs(x) > threshold)) {
789 struct mem_cgroup *mi;
792 * Batch local counters to keep them in sync with
793 * the hierarchical ones.
795 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
796 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
797 atomic_long_add(x, &mi->vmstats[idx]);
800 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
803 static struct mem_cgroup_per_node *
804 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
806 struct mem_cgroup *parent;
808 parent = parent_mem_cgroup(pn->memcg);
811 return mem_cgroup_nodeinfo(parent, nid);
814 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
817 struct mem_cgroup_per_node *pn;
818 struct mem_cgroup *memcg;
819 long x, threshold = MEMCG_CHARGE_BATCH;
821 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
825 __mod_memcg_state(memcg, idx, val);
828 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
830 if (vmstat_item_in_bytes(idx))
831 threshold <<= PAGE_SHIFT;
833 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
834 if (unlikely(abs(x) > threshold)) {
835 pg_data_t *pgdat = lruvec_pgdat(lruvec);
836 struct mem_cgroup_per_node *pi;
838 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
839 atomic_long_add(x, &pi->lruvec_stat[idx]);
842 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
846 * __mod_lruvec_state - update lruvec memory statistics
847 * @lruvec: the lruvec
848 * @idx: the stat item
849 * @val: delta to add to the counter, can be negative
851 * The lruvec is the intersection of the NUMA node and a cgroup. This
852 * function updates the all three counters that are affected by a
853 * change of state at this level: per-node, per-cgroup, per-lruvec.
855 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
859 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
861 /* Update memcg and lruvec */
862 if (!mem_cgroup_disabled())
863 __mod_memcg_lruvec_state(lruvec, idx, val);
866 void __mod_lruvec_slab_state(void *p, enum node_stat_item idx, int val)
868 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
869 struct mem_cgroup *memcg;
870 struct lruvec *lruvec;
873 memcg = mem_cgroup_from_obj(p);
875 /* Untracked pages have no memcg, no lruvec. Update only the node */
876 if (!memcg || memcg == root_mem_cgroup) {
877 __mod_node_page_state(pgdat, idx, val);
879 lruvec = mem_cgroup_lruvec(memcg, pgdat);
880 __mod_lruvec_state(lruvec, idx, val);
885 void mod_memcg_obj_state(void *p, int idx, int val)
887 struct mem_cgroup *memcg;
890 memcg = mem_cgroup_from_obj(p);
892 mod_memcg_state(memcg, idx, val);
897 * __count_memcg_events - account VM events in a cgroup
898 * @memcg: the memory cgroup
899 * @idx: the event item
900 * @count: the number of events that occured
902 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
907 if (mem_cgroup_disabled())
910 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
911 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
912 struct mem_cgroup *mi;
915 * Batch local counters to keep them in sync with
916 * the hierarchical ones.
918 __this_cpu_add(memcg->vmstats_local->events[idx], x);
919 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
920 atomic_long_add(x, &mi->vmevents[idx]);
923 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
926 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
928 return atomic_long_read(&memcg->vmevents[event]);
931 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
936 for_each_possible_cpu(cpu)
937 x += per_cpu(memcg->vmstats_local->events[event], cpu);
941 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
945 /* pagein of a big page is an event. So, ignore page size */
947 __count_memcg_events(memcg, PGPGIN, 1);
949 __count_memcg_events(memcg, PGPGOUT, 1);
950 nr_pages = -nr_pages; /* for event */
953 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
956 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
957 enum mem_cgroup_events_target target)
959 unsigned long val, next;
961 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
962 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
963 /* from time_after() in jiffies.h */
964 if ((long)(next - val) < 0) {
966 case MEM_CGROUP_TARGET_THRESH:
967 next = val + THRESHOLDS_EVENTS_TARGET;
969 case MEM_CGROUP_TARGET_SOFTLIMIT:
970 next = val + SOFTLIMIT_EVENTS_TARGET;
975 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
982 * Check events in order.
985 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
987 /* threshold event is triggered in finer grain than soft limit */
988 if (unlikely(mem_cgroup_event_ratelimit(memcg,
989 MEM_CGROUP_TARGET_THRESH))) {
992 do_softlimit = mem_cgroup_event_ratelimit(memcg,
993 MEM_CGROUP_TARGET_SOFTLIMIT);
994 mem_cgroup_threshold(memcg);
995 if (unlikely(do_softlimit))
996 mem_cgroup_update_tree(memcg, page);
1000 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1003 * mm_update_next_owner() may clear mm->owner to NULL
1004 * if it races with swapoff, page migration, etc.
1005 * So this can be called with p == NULL.
1010 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1012 EXPORT_SYMBOL(mem_cgroup_from_task);
1015 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1016 * @mm: mm from which memcg should be extracted. It can be NULL.
1018 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1019 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1022 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1024 struct mem_cgroup *memcg;
1026 if (mem_cgroup_disabled())
1032 * Page cache insertions can happen withou an
1033 * actual mm context, e.g. during disk probing
1034 * on boot, loopback IO, acct() writes etc.
1037 memcg = root_mem_cgroup;
1039 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1040 if (unlikely(!memcg))
1041 memcg = root_mem_cgroup;
1043 } while (!css_tryget(&memcg->css));
1047 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1050 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1051 * @page: page from which memcg should be extracted.
1053 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1054 * root_mem_cgroup is returned.
1056 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1058 struct mem_cgroup *memcg = page->mem_cgroup;
1060 if (mem_cgroup_disabled())
1064 /* Page should not get uncharged and freed memcg under us. */
1065 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1066 memcg = root_mem_cgroup;
1070 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1073 * If current->active_memcg is non-NULL, do not fallback to current->mm->memcg.
1075 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1077 if (unlikely(current->active_memcg)) {
1078 struct mem_cgroup *memcg;
1081 /* current->active_memcg must hold a ref. */
1082 if (WARN_ON_ONCE(!css_tryget(¤t->active_memcg->css)))
1083 memcg = root_mem_cgroup;
1085 memcg = current->active_memcg;
1089 return get_mem_cgroup_from_mm(current->mm);
1093 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1094 * @root: hierarchy root
1095 * @prev: previously returned memcg, NULL on first invocation
1096 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1098 * Returns references to children of the hierarchy below @root, or
1099 * @root itself, or %NULL after a full round-trip.
1101 * Caller must pass the return value in @prev on subsequent
1102 * invocations for reference counting, or use mem_cgroup_iter_break()
1103 * to cancel a hierarchy walk before the round-trip is complete.
1105 * Reclaimers can specify a node and a priority level in @reclaim to
1106 * divide up the memcgs in the hierarchy among all concurrent
1107 * reclaimers operating on the same node and priority.
1109 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1110 struct mem_cgroup *prev,
1111 struct mem_cgroup_reclaim_cookie *reclaim)
1113 struct mem_cgroup_reclaim_iter *iter;
1114 struct cgroup_subsys_state *css = NULL;
1115 struct mem_cgroup *memcg = NULL;
1116 struct mem_cgroup *pos = NULL;
1118 if (mem_cgroup_disabled())
1122 root = root_mem_cgroup;
1124 if (prev && !reclaim)
1127 if (!root->use_hierarchy && root != root_mem_cgroup) {
1136 struct mem_cgroup_per_node *mz;
1138 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1141 if (prev && reclaim->generation != iter->generation)
1145 pos = READ_ONCE(iter->position);
1146 if (!pos || css_tryget(&pos->css))
1149 * css reference reached zero, so iter->position will
1150 * be cleared by ->css_released. However, we should not
1151 * rely on this happening soon, because ->css_released
1152 * is called from a work queue, and by busy-waiting we
1153 * might block it. So we clear iter->position right
1156 (void)cmpxchg(&iter->position, pos, NULL);
1164 css = css_next_descendant_pre(css, &root->css);
1167 * Reclaimers share the hierarchy walk, and a
1168 * new one might jump in right at the end of
1169 * the hierarchy - make sure they see at least
1170 * one group and restart from the beginning.
1178 * Verify the css and acquire a reference. The root
1179 * is provided by the caller, so we know it's alive
1180 * and kicking, and don't take an extra reference.
1182 memcg = mem_cgroup_from_css(css);
1184 if (css == &root->css)
1187 if (css_tryget(css))
1195 * The position could have already been updated by a competing
1196 * thread, so check that the value hasn't changed since we read
1197 * it to avoid reclaiming from the same cgroup twice.
1199 (void)cmpxchg(&iter->position, pos, memcg);
1207 reclaim->generation = iter->generation;
1213 if (prev && prev != root)
1214 css_put(&prev->css);
1220 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1221 * @root: hierarchy root
1222 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1224 void mem_cgroup_iter_break(struct mem_cgroup *root,
1225 struct mem_cgroup *prev)
1228 root = root_mem_cgroup;
1229 if (prev && prev != root)
1230 css_put(&prev->css);
1233 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1234 struct mem_cgroup *dead_memcg)
1236 struct mem_cgroup_reclaim_iter *iter;
1237 struct mem_cgroup_per_node *mz;
1240 for_each_node(nid) {
1241 mz = mem_cgroup_nodeinfo(from, nid);
1243 cmpxchg(&iter->position, dead_memcg, NULL);
1247 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1249 struct mem_cgroup *memcg = dead_memcg;
1250 struct mem_cgroup *last;
1253 __invalidate_reclaim_iterators(memcg, dead_memcg);
1255 } while ((memcg = parent_mem_cgroup(memcg)));
1258 * When cgruop1 non-hierarchy mode is used,
1259 * parent_mem_cgroup() does not walk all the way up to the
1260 * cgroup root (root_mem_cgroup). So we have to handle
1261 * dead_memcg from cgroup root separately.
1263 if (last != root_mem_cgroup)
1264 __invalidate_reclaim_iterators(root_mem_cgroup,
1269 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1270 * @memcg: hierarchy root
1271 * @fn: function to call for each task
1272 * @arg: argument passed to @fn
1274 * This function iterates over tasks attached to @memcg or to any of its
1275 * descendants and calls @fn for each task. If @fn returns a non-zero
1276 * value, the function breaks the iteration loop and returns the value.
1277 * Otherwise, it will iterate over all tasks and return 0.
1279 * This function must not be called for the root memory cgroup.
1281 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1282 int (*fn)(struct task_struct *, void *), void *arg)
1284 struct mem_cgroup *iter;
1287 BUG_ON(memcg == root_mem_cgroup);
1289 for_each_mem_cgroup_tree(iter, memcg) {
1290 struct css_task_iter it;
1291 struct task_struct *task;
1293 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1294 while (!ret && (task = css_task_iter_next(&it)))
1295 ret = fn(task, arg);
1296 css_task_iter_end(&it);
1298 mem_cgroup_iter_break(memcg, iter);
1306 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1308 * @pgdat: pgdat of the page
1310 * This function relies on page->mem_cgroup being stable - see the
1311 * access rules in commit_charge().
1313 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1315 struct mem_cgroup_per_node *mz;
1316 struct mem_cgroup *memcg;
1317 struct lruvec *lruvec;
1319 if (mem_cgroup_disabled()) {
1320 lruvec = &pgdat->__lruvec;
1324 memcg = page->mem_cgroup;
1326 * Swapcache readahead pages are added to the LRU - and
1327 * possibly migrated - before they are charged.
1330 memcg = root_mem_cgroup;
1332 mz = mem_cgroup_page_nodeinfo(memcg, page);
1333 lruvec = &mz->lruvec;
1336 * Since a node can be onlined after the mem_cgroup was created,
1337 * we have to be prepared to initialize lruvec->zone here;
1338 * and if offlined then reonlined, we need to reinitialize it.
1340 if (unlikely(lruvec->pgdat != pgdat))
1341 lruvec->pgdat = pgdat;
1346 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1347 * @lruvec: mem_cgroup per zone lru vector
1348 * @lru: index of lru list the page is sitting on
1349 * @zid: zone id of the accounted pages
1350 * @nr_pages: positive when adding or negative when removing
1352 * This function must be called under lru_lock, just before a page is added
1353 * to or just after a page is removed from an lru list (that ordering being
1354 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1356 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1357 int zid, int nr_pages)
1359 struct mem_cgroup_per_node *mz;
1360 unsigned long *lru_size;
1363 if (mem_cgroup_disabled())
1366 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1367 lru_size = &mz->lru_zone_size[zid][lru];
1370 *lru_size += nr_pages;
1373 if (WARN_ONCE(size < 0,
1374 "%s(%p, %d, %d): lru_size %ld\n",
1375 __func__, lruvec, lru, nr_pages, size)) {
1381 *lru_size += nr_pages;
1385 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1386 * @memcg: the memory cgroup
1388 * Returns the maximum amount of memory @mem can be charged with, in
1391 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1393 unsigned long margin = 0;
1394 unsigned long count;
1395 unsigned long limit;
1397 count = page_counter_read(&memcg->memory);
1398 limit = READ_ONCE(memcg->memory.max);
1400 margin = limit - count;
1402 if (do_memsw_account()) {
1403 count = page_counter_read(&memcg->memsw);
1404 limit = READ_ONCE(memcg->memsw.max);
1406 margin = min(margin, limit - count);
1415 * A routine for checking "mem" is under move_account() or not.
1417 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1418 * moving cgroups. This is for waiting at high-memory pressure
1421 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1423 struct mem_cgroup *from;
1424 struct mem_cgroup *to;
1427 * Unlike task_move routines, we access mc.to, mc.from not under
1428 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1430 spin_lock(&mc.lock);
1436 ret = mem_cgroup_is_descendant(from, memcg) ||
1437 mem_cgroup_is_descendant(to, memcg);
1439 spin_unlock(&mc.lock);
1443 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1445 if (mc.moving_task && current != mc.moving_task) {
1446 if (mem_cgroup_under_move(memcg)) {
1448 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1449 /* moving charge context might have finished. */
1452 finish_wait(&mc.waitq, &wait);
1459 static char *memory_stat_format(struct mem_cgroup *memcg)
1464 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1469 * Provide statistics on the state of the memory subsystem as
1470 * well as cumulative event counters that show past behavior.
1472 * This list is ordered following a combination of these gradients:
1473 * 1) generic big picture -> specifics and details
1474 * 2) reflecting userspace activity -> reflecting kernel heuristics
1476 * Current memory state:
1479 seq_buf_printf(&s, "anon %llu\n",
1480 (u64)memcg_page_state(memcg, NR_ANON_MAPPED) *
1482 seq_buf_printf(&s, "file %llu\n",
1483 (u64)memcg_page_state(memcg, NR_FILE_PAGES) *
1485 seq_buf_printf(&s, "kernel_stack %llu\n",
1486 (u64)memcg_page_state(memcg, NR_KERNEL_STACK_KB) *
1488 seq_buf_printf(&s, "slab %llu\n",
1489 (u64)(memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1490 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B)));
1491 seq_buf_printf(&s, "percpu %llu\n",
1492 (u64)memcg_page_state(memcg, MEMCG_PERCPU_B));
1493 seq_buf_printf(&s, "sock %llu\n",
1494 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1497 seq_buf_printf(&s, "shmem %llu\n",
1498 (u64)memcg_page_state(memcg, NR_SHMEM) *
1500 seq_buf_printf(&s, "file_mapped %llu\n",
1501 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1503 seq_buf_printf(&s, "file_dirty %llu\n",
1504 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1506 seq_buf_printf(&s, "file_writeback %llu\n",
1507 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1510 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1511 seq_buf_printf(&s, "anon_thp %llu\n",
1512 (u64)memcg_page_state(memcg, NR_ANON_THPS) *
1516 for (i = 0; i < NR_LRU_LISTS; i++)
1517 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1518 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1521 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1522 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B));
1523 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1524 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B));
1526 /* Accumulated memory events */
1528 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1529 memcg_events(memcg, PGFAULT));
1530 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1531 memcg_events(memcg, PGMAJFAULT));
1533 seq_buf_printf(&s, "workingset_refault %lu\n",
1534 memcg_page_state(memcg, WORKINGSET_REFAULT));
1535 seq_buf_printf(&s, "workingset_activate %lu\n",
1536 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1537 seq_buf_printf(&s, "workingset_restore %lu\n",
1538 memcg_page_state(memcg, WORKINGSET_RESTORE));
1539 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1540 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1542 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1543 memcg_events(memcg, PGREFILL));
1544 seq_buf_printf(&s, "pgscan %lu\n",
1545 memcg_events(memcg, PGSCAN_KSWAPD) +
1546 memcg_events(memcg, PGSCAN_DIRECT));
1547 seq_buf_printf(&s, "pgsteal %lu\n",
1548 memcg_events(memcg, PGSTEAL_KSWAPD) +
1549 memcg_events(memcg, PGSTEAL_DIRECT));
1550 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1551 memcg_events(memcg, PGACTIVATE));
1552 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1553 memcg_events(memcg, PGDEACTIVATE));
1554 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1555 memcg_events(memcg, PGLAZYFREE));
1556 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1557 memcg_events(memcg, PGLAZYFREED));
1559 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1560 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1561 memcg_events(memcg, THP_FAULT_ALLOC));
1562 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1563 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1564 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1566 /* The above should easily fit into one page */
1567 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1572 #define K(x) ((x) << (PAGE_SHIFT-10))
1574 * mem_cgroup_print_oom_context: Print OOM information relevant to
1575 * memory controller.
1576 * @memcg: The memory cgroup that went over limit
1577 * @p: Task that is going to be killed
1579 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1582 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1587 pr_cont(",oom_memcg=");
1588 pr_cont_cgroup_path(memcg->css.cgroup);
1590 pr_cont(",global_oom");
1592 pr_cont(",task_memcg=");
1593 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1599 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1600 * memory controller.
1601 * @memcg: The memory cgroup that went over limit
1603 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1607 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1608 K((u64)page_counter_read(&memcg->memory)),
1609 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1610 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1611 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1612 K((u64)page_counter_read(&memcg->swap)),
1613 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1615 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1616 K((u64)page_counter_read(&memcg->memsw)),
1617 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1618 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1619 K((u64)page_counter_read(&memcg->kmem)),
1620 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1623 pr_info("Memory cgroup stats for ");
1624 pr_cont_cgroup_path(memcg->css.cgroup);
1626 buf = memory_stat_format(memcg);
1634 * Return the memory (and swap, if configured) limit for a memcg.
1636 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1640 max = READ_ONCE(memcg->memory.max);
1641 if (mem_cgroup_swappiness(memcg)) {
1642 unsigned long memsw_max;
1643 unsigned long swap_max;
1645 memsw_max = memcg->memsw.max;
1646 swap_max = READ_ONCE(memcg->swap.max);
1647 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1648 max = min(max + swap_max, memsw_max);
1653 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1655 return page_counter_read(&memcg->memory);
1658 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1661 struct oom_control oc = {
1665 .gfp_mask = gfp_mask,
1670 if (mutex_lock_killable(&oom_lock))
1673 if (mem_cgroup_margin(memcg) >= (1 << order))
1677 * A few threads which were not waiting at mutex_lock_killable() can
1678 * fail to bail out. Therefore, check again after holding oom_lock.
1680 ret = should_force_charge() || out_of_memory(&oc);
1683 mutex_unlock(&oom_lock);
1687 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1690 unsigned long *total_scanned)
1692 struct mem_cgroup *victim = NULL;
1695 unsigned long excess;
1696 unsigned long nr_scanned;
1697 struct mem_cgroup_reclaim_cookie reclaim = {
1701 excess = soft_limit_excess(root_memcg);
1704 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1709 * If we have not been able to reclaim
1710 * anything, it might because there are
1711 * no reclaimable pages under this hierarchy
1716 * We want to do more targeted reclaim.
1717 * excess >> 2 is not to excessive so as to
1718 * reclaim too much, nor too less that we keep
1719 * coming back to reclaim from this cgroup
1721 if (total >= (excess >> 2) ||
1722 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1727 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1728 pgdat, &nr_scanned);
1729 *total_scanned += nr_scanned;
1730 if (!soft_limit_excess(root_memcg))
1733 mem_cgroup_iter_break(root_memcg, victim);
1737 #ifdef CONFIG_LOCKDEP
1738 static struct lockdep_map memcg_oom_lock_dep_map = {
1739 .name = "memcg_oom_lock",
1743 static DEFINE_SPINLOCK(memcg_oom_lock);
1746 * Check OOM-Killer is already running under our hierarchy.
1747 * If someone is running, return false.
1749 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1751 struct mem_cgroup *iter, *failed = NULL;
1753 spin_lock(&memcg_oom_lock);
1755 for_each_mem_cgroup_tree(iter, memcg) {
1756 if (iter->oom_lock) {
1758 * this subtree of our hierarchy is already locked
1759 * so we cannot give a lock.
1762 mem_cgroup_iter_break(memcg, iter);
1765 iter->oom_lock = true;
1770 * OK, we failed to lock the whole subtree so we have
1771 * to clean up what we set up to the failing subtree
1773 for_each_mem_cgroup_tree(iter, memcg) {
1774 if (iter == failed) {
1775 mem_cgroup_iter_break(memcg, iter);
1778 iter->oom_lock = false;
1781 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1783 spin_unlock(&memcg_oom_lock);
1788 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1790 struct mem_cgroup *iter;
1792 spin_lock(&memcg_oom_lock);
1793 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1794 for_each_mem_cgroup_tree(iter, memcg)
1795 iter->oom_lock = false;
1796 spin_unlock(&memcg_oom_lock);
1799 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1801 struct mem_cgroup *iter;
1803 spin_lock(&memcg_oom_lock);
1804 for_each_mem_cgroup_tree(iter, memcg)
1806 spin_unlock(&memcg_oom_lock);
1809 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1811 struct mem_cgroup *iter;
1814 * When a new child is created while the hierarchy is under oom,
1815 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1817 spin_lock(&memcg_oom_lock);
1818 for_each_mem_cgroup_tree(iter, memcg)
1819 if (iter->under_oom > 0)
1821 spin_unlock(&memcg_oom_lock);
1824 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1826 struct oom_wait_info {
1827 struct mem_cgroup *memcg;
1828 wait_queue_entry_t wait;
1831 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1832 unsigned mode, int sync, void *arg)
1834 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1835 struct mem_cgroup *oom_wait_memcg;
1836 struct oom_wait_info *oom_wait_info;
1838 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1839 oom_wait_memcg = oom_wait_info->memcg;
1841 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1842 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1844 return autoremove_wake_function(wait, mode, sync, arg);
1847 static void memcg_oom_recover(struct mem_cgroup *memcg)
1850 * For the following lockless ->under_oom test, the only required
1851 * guarantee is that it must see the state asserted by an OOM when
1852 * this function is called as a result of userland actions
1853 * triggered by the notification of the OOM. This is trivially
1854 * achieved by invoking mem_cgroup_mark_under_oom() before
1855 * triggering notification.
1857 if (memcg && memcg->under_oom)
1858 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1868 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1870 enum oom_status ret;
1873 if (order > PAGE_ALLOC_COSTLY_ORDER)
1876 memcg_memory_event(memcg, MEMCG_OOM);
1879 * We are in the middle of the charge context here, so we
1880 * don't want to block when potentially sitting on a callstack
1881 * that holds all kinds of filesystem and mm locks.
1883 * cgroup1 allows disabling the OOM killer and waiting for outside
1884 * handling until the charge can succeed; remember the context and put
1885 * the task to sleep at the end of the page fault when all locks are
1888 * On the other hand, in-kernel OOM killer allows for an async victim
1889 * memory reclaim (oom_reaper) and that means that we are not solely
1890 * relying on the oom victim to make a forward progress and we can
1891 * invoke the oom killer here.
1893 * Please note that mem_cgroup_out_of_memory might fail to find a
1894 * victim and then we have to bail out from the charge path.
1896 if (memcg->oom_kill_disable) {
1897 if (!current->in_user_fault)
1899 css_get(&memcg->css);
1900 current->memcg_in_oom = memcg;
1901 current->memcg_oom_gfp_mask = mask;
1902 current->memcg_oom_order = order;
1907 mem_cgroup_mark_under_oom(memcg);
1909 locked = mem_cgroup_oom_trylock(memcg);
1912 mem_cgroup_oom_notify(memcg);
1914 mem_cgroup_unmark_under_oom(memcg);
1915 if (mem_cgroup_out_of_memory(memcg, mask, order))
1921 mem_cgroup_oom_unlock(memcg);
1927 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1928 * @handle: actually kill/wait or just clean up the OOM state
1930 * This has to be called at the end of a page fault if the memcg OOM
1931 * handler was enabled.
1933 * Memcg supports userspace OOM handling where failed allocations must
1934 * sleep on a waitqueue until the userspace task resolves the
1935 * situation. Sleeping directly in the charge context with all kinds
1936 * of locks held is not a good idea, instead we remember an OOM state
1937 * in the task and mem_cgroup_oom_synchronize() has to be called at
1938 * the end of the page fault to complete the OOM handling.
1940 * Returns %true if an ongoing memcg OOM situation was detected and
1941 * completed, %false otherwise.
1943 bool mem_cgroup_oom_synchronize(bool handle)
1945 struct mem_cgroup *memcg = current->memcg_in_oom;
1946 struct oom_wait_info owait;
1949 /* OOM is global, do not handle */
1956 owait.memcg = memcg;
1957 owait.wait.flags = 0;
1958 owait.wait.func = memcg_oom_wake_function;
1959 owait.wait.private = current;
1960 INIT_LIST_HEAD(&owait.wait.entry);
1962 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1963 mem_cgroup_mark_under_oom(memcg);
1965 locked = mem_cgroup_oom_trylock(memcg);
1968 mem_cgroup_oom_notify(memcg);
1970 if (locked && !memcg->oom_kill_disable) {
1971 mem_cgroup_unmark_under_oom(memcg);
1972 finish_wait(&memcg_oom_waitq, &owait.wait);
1973 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1974 current->memcg_oom_order);
1977 mem_cgroup_unmark_under_oom(memcg);
1978 finish_wait(&memcg_oom_waitq, &owait.wait);
1982 mem_cgroup_oom_unlock(memcg);
1984 * There is no guarantee that an OOM-lock contender
1985 * sees the wakeups triggered by the OOM kill
1986 * uncharges. Wake any sleepers explicitely.
1988 memcg_oom_recover(memcg);
1991 current->memcg_in_oom = NULL;
1992 css_put(&memcg->css);
1997 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1998 * @victim: task to be killed by the OOM killer
1999 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2001 * Returns a pointer to a memory cgroup, which has to be cleaned up
2002 * by killing all belonging OOM-killable tasks.
2004 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2006 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2007 struct mem_cgroup *oom_domain)
2009 struct mem_cgroup *oom_group = NULL;
2010 struct mem_cgroup *memcg;
2012 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2016 oom_domain = root_mem_cgroup;
2020 memcg = mem_cgroup_from_task(victim);
2021 if (memcg == root_mem_cgroup)
2025 * If the victim task has been asynchronously moved to a different
2026 * memory cgroup, we might end up killing tasks outside oom_domain.
2027 * In this case it's better to ignore memory.group.oom.
2029 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2033 * Traverse the memory cgroup hierarchy from the victim task's
2034 * cgroup up to the OOMing cgroup (or root) to find the
2035 * highest-level memory cgroup with oom.group set.
2037 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2038 if (memcg->oom_group)
2041 if (memcg == oom_domain)
2046 css_get(&oom_group->css);
2053 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2055 pr_info("Tasks in ");
2056 pr_cont_cgroup_path(memcg->css.cgroup);
2057 pr_cont(" are going to be killed due to memory.oom.group set\n");
2061 * lock_page_memcg - lock a page->mem_cgroup binding
2064 * This function protects unlocked LRU pages from being moved to
2067 * It ensures lifetime of the returned memcg. Caller is responsible
2068 * for the lifetime of the page; __unlock_page_memcg() is available
2069 * when @page might get freed inside the locked section.
2071 struct mem_cgroup *lock_page_memcg(struct page *page)
2073 struct page *head = compound_head(page); /* rmap on tail pages */
2074 struct mem_cgroup *memcg;
2075 unsigned long flags;
2078 * The RCU lock is held throughout the transaction. The fast
2079 * path can get away without acquiring the memcg->move_lock
2080 * because page moving starts with an RCU grace period.
2082 * The RCU lock also protects the memcg from being freed when
2083 * the page state that is going to change is the only thing
2084 * preventing the page itself from being freed. E.g. writeback
2085 * doesn't hold a page reference and relies on PG_writeback to
2086 * keep off truncation, migration and so forth.
2090 if (mem_cgroup_disabled())
2093 memcg = head->mem_cgroup;
2094 if (unlikely(!memcg))
2097 if (atomic_read(&memcg->moving_account) <= 0)
2100 spin_lock_irqsave(&memcg->move_lock, flags);
2101 if (memcg != head->mem_cgroup) {
2102 spin_unlock_irqrestore(&memcg->move_lock, flags);
2107 * When charge migration first begins, we can have locked and
2108 * unlocked page stat updates happening concurrently. Track
2109 * the task who has the lock for unlock_page_memcg().
2111 memcg->move_lock_task = current;
2112 memcg->move_lock_flags = flags;
2116 EXPORT_SYMBOL(lock_page_memcg);
2119 * __unlock_page_memcg - unlock and unpin a memcg
2122 * Unlock and unpin a memcg returned by lock_page_memcg().
2124 void __unlock_page_memcg(struct mem_cgroup *memcg)
2126 if (memcg && memcg->move_lock_task == current) {
2127 unsigned long flags = memcg->move_lock_flags;
2129 memcg->move_lock_task = NULL;
2130 memcg->move_lock_flags = 0;
2132 spin_unlock_irqrestore(&memcg->move_lock, flags);
2139 * unlock_page_memcg - unlock a page->mem_cgroup binding
2142 void unlock_page_memcg(struct page *page)
2144 struct page *head = compound_head(page);
2146 __unlock_page_memcg(head->mem_cgroup);
2148 EXPORT_SYMBOL(unlock_page_memcg);
2150 struct memcg_stock_pcp {
2151 struct mem_cgroup *cached; /* this never be root cgroup */
2152 unsigned int nr_pages;
2154 #ifdef CONFIG_MEMCG_KMEM
2155 struct obj_cgroup *cached_objcg;
2156 unsigned int nr_bytes;
2159 struct work_struct work;
2160 unsigned long flags;
2161 #define FLUSHING_CACHED_CHARGE 0
2163 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2164 static DEFINE_MUTEX(percpu_charge_mutex);
2166 #ifdef CONFIG_MEMCG_KMEM
2167 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2168 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2169 struct mem_cgroup *root_memcg);
2172 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2175 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2176 struct mem_cgroup *root_memcg)
2183 * consume_stock: Try to consume stocked charge on this cpu.
2184 * @memcg: memcg to consume from.
2185 * @nr_pages: how many pages to charge.
2187 * The charges will only happen if @memcg matches the current cpu's memcg
2188 * stock, and at least @nr_pages are available in that stock. Failure to
2189 * service an allocation will refill the stock.
2191 * returns true if successful, false otherwise.
2193 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2195 struct memcg_stock_pcp *stock;
2196 unsigned long flags;
2199 if (nr_pages > MEMCG_CHARGE_BATCH)
2202 local_irq_save(flags);
2204 stock = this_cpu_ptr(&memcg_stock);
2205 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2206 stock->nr_pages -= nr_pages;
2210 local_irq_restore(flags);
2216 * Returns stocks cached in percpu and reset cached information.
2218 static void drain_stock(struct memcg_stock_pcp *stock)
2220 struct mem_cgroup *old = stock->cached;
2225 if (stock->nr_pages) {
2226 page_counter_uncharge(&old->memory, stock->nr_pages);
2227 if (do_memsw_account())
2228 page_counter_uncharge(&old->memsw, stock->nr_pages);
2229 stock->nr_pages = 0;
2233 stock->cached = NULL;
2236 static void drain_local_stock(struct work_struct *dummy)
2238 struct memcg_stock_pcp *stock;
2239 unsigned long flags;
2242 * The only protection from memory hotplug vs. drain_stock races is
2243 * that we always operate on local CPU stock here with IRQ disabled
2245 local_irq_save(flags);
2247 stock = this_cpu_ptr(&memcg_stock);
2248 drain_obj_stock(stock);
2250 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2252 local_irq_restore(flags);
2256 * Cache charges(val) to local per_cpu area.
2257 * This will be consumed by consume_stock() function, later.
2259 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2261 struct memcg_stock_pcp *stock;
2262 unsigned long flags;
2264 local_irq_save(flags);
2266 stock = this_cpu_ptr(&memcg_stock);
2267 if (stock->cached != memcg) { /* reset if necessary */
2269 css_get(&memcg->css);
2270 stock->cached = memcg;
2272 stock->nr_pages += nr_pages;
2274 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2277 local_irq_restore(flags);
2281 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2282 * of the hierarchy under it.
2284 static void drain_all_stock(struct mem_cgroup *root_memcg)
2288 /* If someone's already draining, avoid adding running more workers. */
2289 if (!mutex_trylock(&percpu_charge_mutex))
2292 * Notify other cpus that system-wide "drain" is running
2293 * We do not care about races with the cpu hotplug because cpu down
2294 * as well as workers from this path always operate on the local
2295 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2298 for_each_online_cpu(cpu) {
2299 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2300 struct mem_cgroup *memcg;
2304 memcg = stock->cached;
2305 if (memcg && stock->nr_pages &&
2306 mem_cgroup_is_descendant(memcg, root_memcg))
2308 if (obj_stock_flush_required(stock, root_memcg))
2313 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2315 drain_local_stock(&stock->work);
2317 schedule_work_on(cpu, &stock->work);
2321 mutex_unlock(&percpu_charge_mutex);
2324 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2326 struct memcg_stock_pcp *stock;
2327 struct mem_cgroup *memcg, *mi;
2329 stock = &per_cpu(memcg_stock, cpu);
2332 for_each_mem_cgroup(memcg) {
2335 for (i = 0; i < MEMCG_NR_STAT; i++) {
2339 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2341 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2342 atomic_long_add(x, &memcg->vmstats[i]);
2344 if (i >= NR_VM_NODE_STAT_ITEMS)
2347 for_each_node(nid) {
2348 struct mem_cgroup_per_node *pn;
2350 pn = mem_cgroup_nodeinfo(memcg, nid);
2351 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2354 atomic_long_add(x, &pn->lruvec_stat[i]);
2355 } while ((pn = parent_nodeinfo(pn, nid)));
2359 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2362 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2364 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2365 atomic_long_add(x, &memcg->vmevents[i]);
2372 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2373 unsigned int nr_pages,
2376 unsigned long nr_reclaimed = 0;
2379 unsigned long pflags;
2381 if (page_counter_read(&memcg->memory) <=
2382 READ_ONCE(memcg->memory.high))
2385 memcg_memory_event(memcg, MEMCG_HIGH);
2387 psi_memstall_enter(&pflags);
2388 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2390 psi_memstall_leave(&pflags);
2391 } while ((memcg = parent_mem_cgroup(memcg)) &&
2392 !mem_cgroup_is_root(memcg));
2394 return nr_reclaimed;
2397 static void high_work_func(struct work_struct *work)
2399 struct mem_cgroup *memcg;
2401 memcg = container_of(work, struct mem_cgroup, high_work);
2402 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2406 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2407 * enough to still cause a significant slowdown in most cases, while still
2408 * allowing diagnostics and tracing to proceed without becoming stuck.
2410 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2413 * When calculating the delay, we use these either side of the exponentiation to
2414 * maintain precision and scale to a reasonable number of jiffies (see the table
2417 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2418 * overage ratio to a delay.
2419 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2420 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2421 * to produce a reasonable delay curve.
2423 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2424 * reasonable delay curve compared to precision-adjusted overage, not
2425 * penalising heavily at first, but still making sure that growth beyond the
2426 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2427 * example, with a high of 100 megabytes:
2429 * +-------+------------------------+
2430 * | usage | time to allocate in ms |
2431 * +-------+------------------------+
2453 * +-------+------------------------+
2455 #define MEMCG_DELAY_PRECISION_SHIFT 20
2456 #define MEMCG_DELAY_SCALING_SHIFT 14
2458 static u64 calculate_overage(unsigned long usage, unsigned long high)
2466 * Prevent division by 0 in overage calculation by acting as if
2467 * it was a threshold of 1 page
2469 high = max(high, 1UL);
2471 overage = usage - high;
2472 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2473 return div64_u64(overage, high);
2476 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2478 u64 overage, max_overage = 0;
2481 overage = calculate_overage(page_counter_read(&memcg->memory),
2482 READ_ONCE(memcg->memory.high));
2483 max_overage = max(overage, max_overage);
2484 } while ((memcg = parent_mem_cgroup(memcg)) &&
2485 !mem_cgroup_is_root(memcg));
2490 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2492 u64 overage, max_overage = 0;
2495 overage = calculate_overage(page_counter_read(&memcg->swap),
2496 READ_ONCE(memcg->swap.high));
2498 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2499 max_overage = max(overage, max_overage);
2500 } while ((memcg = parent_mem_cgroup(memcg)) &&
2501 !mem_cgroup_is_root(memcg));
2507 * Get the number of jiffies that we should penalise a mischievous cgroup which
2508 * is exceeding its memory.high by checking both it and its ancestors.
2510 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2511 unsigned int nr_pages,
2514 unsigned long penalty_jiffies;
2520 * We use overage compared to memory.high to calculate the number of
2521 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2522 * fairly lenient on small overages, and increasingly harsh when the
2523 * memcg in question makes it clear that it has no intention of stopping
2524 * its crazy behaviour, so we exponentially increase the delay based on
2527 penalty_jiffies = max_overage * max_overage * HZ;
2528 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2529 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2532 * Factor in the task's own contribution to the overage, such that four
2533 * N-sized allocations are throttled approximately the same as one
2534 * 4N-sized allocation.
2536 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2537 * larger the current charge patch is than that.
2539 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2543 * Scheduled by try_charge() to be executed from the userland return path
2544 * and reclaims memory over the high limit.
2546 void mem_cgroup_handle_over_high(void)
2548 unsigned long penalty_jiffies;
2549 unsigned long pflags;
2550 unsigned long nr_reclaimed;
2551 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2552 int nr_retries = MAX_RECLAIM_RETRIES;
2553 struct mem_cgroup *memcg;
2554 bool in_retry = false;
2556 if (likely(!nr_pages))
2559 memcg = get_mem_cgroup_from_mm(current->mm);
2560 current->memcg_nr_pages_over_high = 0;
2564 * The allocating task should reclaim at least the batch size, but for
2565 * subsequent retries we only want to do what's necessary to prevent oom
2566 * or breaching resource isolation.
2568 * This is distinct from memory.max or page allocator behaviour because
2569 * memory.high is currently batched, whereas memory.max and the page
2570 * allocator run every time an allocation is made.
2572 nr_reclaimed = reclaim_high(memcg,
2573 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2577 * memory.high is breached and reclaim is unable to keep up. Throttle
2578 * allocators proactively to slow down excessive growth.
2580 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2581 mem_find_max_overage(memcg));
2583 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2584 swap_find_max_overage(memcg));
2587 * Clamp the max delay per usermode return so as to still keep the
2588 * application moving forwards and also permit diagnostics, albeit
2591 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2594 * Don't sleep if the amount of jiffies this memcg owes us is so low
2595 * that it's not even worth doing, in an attempt to be nice to those who
2596 * go only a small amount over their memory.high value and maybe haven't
2597 * been aggressively reclaimed enough yet.
2599 if (penalty_jiffies <= HZ / 100)
2603 * If reclaim is making forward progress but we're still over
2604 * memory.high, we want to encourage that rather than doing allocator
2607 if (nr_reclaimed || nr_retries--) {
2613 * If we exit early, we're guaranteed to die (since
2614 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2615 * need to account for any ill-begotten jiffies to pay them off later.
2617 psi_memstall_enter(&pflags);
2618 schedule_timeout_killable(penalty_jiffies);
2619 psi_memstall_leave(&pflags);
2622 css_put(&memcg->css);
2625 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2626 unsigned int nr_pages)
2628 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2629 int nr_retries = MAX_RECLAIM_RETRIES;
2630 struct mem_cgroup *mem_over_limit;
2631 struct page_counter *counter;
2632 enum oom_status oom_status;
2633 unsigned long nr_reclaimed;
2634 bool may_swap = true;
2635 bool drained = false;
2636 unsigned long pflags;
2638 if (mem_cgroup_is_root(memcg))
2641 if (consume_stock(memcg, nr_pages))
2644 if (!do_memsw_account() ||
2645 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2646 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2648 if (do_memsw_account())
2649 page_counter_uncharge(&memcg->memsw, batch);
2650 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2652 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2656 if (batch > nr_pages) {
2662 * Memcg doesn't have a dedicated reserve for atomic
2663 * allocations. But like the global atomic pool, we need to
2664 * put the burden of reclaim on regular allocation requests
2665 * and let these go through as privileged allocations.
2667 if (gfp_mask & __GFP_ATOMIC)
2671 * Unlike in global OOM situations, memcg is not in a physical
2672 * memory shortage. Allow dying and OOM-killed tasks to
2673 * bypass the last charges so that they can exit quickly and
2674 * free their memory.
2676 if (unlikely(should_force_charge()))
2680 * Prevent unbounded recursion when reclaim operations need to
2681 * allocate memory. This might exceed the limits temporarily,
2682 * but we prefer facilitating memory reclaim and getting back
2683 * under the limit over triggering OOM kills in these cases.
2685 if (unlikely(current->flags & PF_MEMALLOC))
2688 if (unlikely(task_in_memcg_oom(current)))
2691 if (!gfpflags_allow_blocking(gfp_mask))
2694 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2696 psi_memstall_enter(&pflags);
2697 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2698 gfp_mask, may_swap);
2699 psi_memstall_leave(&pflags);
2701 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2705 drain_all_stock(mem_over_limit);
2710 if (gfp_mask & __GFP_NORETRY)
2713 * Even though the limit is exceeded at this point, reclaim
2714 * may have been able to free some pages. Retry the charge
2715 * before killing the task.
2717 * Only for regular pages, though: huge pages are rather
2718 * unlikely to succeed so close to the limit, and we fall back
2719 * to regular pages anyway in case of failure.
2721 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2724 * At task move, charge accounts can be doubly counted. So, it's
2725 * better to wait until the end of task_move if something is going on.
2727 if (mem_cgroup_wait_acct_move(mem_over_limit))
2733 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2736 if (gfp_mask & __GFP_NOFAIL)
2739 if (fatal_signal_pending(current))
2743 * keep retrying as long as the memcg oom killer is able to make
2744 * a forward progress or bypass the charge if the oom killer
2745 * couldn't make any progress.
2747 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2748 get_order(nr_pages * PAGE_SIZE));
2749 switch (oom_status) {
2751 nr_retries = MAX_RECLAIM_RETRIES;
2759 if (!(gfp_mask & __GFP_NOFAIL))
2763 * The allocation either can't fail or will lead to more memory
2764 * being freed very soon. Allow memory usage go over the limit
2765 * temporarily by force charging it.
2767 page_counter_charge(&memcg->memory, nr_pages);
2768 if (do_memsw_account())
2769 page_counter_charge(&memcg->memsw, nr_pages);
2774 if (batch > nr_pages)
2775 refill_stock(memcg, batch - nr_pages);
2778 * If the hierarchy is above the normal consumption range, schedule
2779 * reclaim on returning to userland. We can perform reclaim here
2780 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2781 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2782 * not recorded as it most likely matches current's and won't
2783 * change in the meantime. As high limit is checked again before
2784 * reclaim, the cost of mismatch is negligible.
2787 bool mem_high, swap_high;
2789 mem_high = page_counter_read(&memcg->memory) >
2790 READ_ONCE(memcg->memory.high);
2791 swap_high = page_counter_read(&memcg->swap) >
2792 READ_ONCE(memcg->swap.high);
2794 /* Don't bother a random interrupted task */
2795 if (in_interrupt()) {
2797 schedule_work(&memcg->high_work);
2803 if (mem_high || swap_high) {
2805 * The allocating tasks in this cgroup will need to do
2806 * reclaim or be throttled to prevent further growth
2807 * of the memory or swap footprints.
2809 * Target some best-effort fairness between the tasks,
2810 * and distribute reclaim work and delay penalties
2811 * based on how much each task is actually allocating.
2813 current->memcg_nr_pages_over_high += batch;
2814 set_notify_resume(current);
2817 } while ((memcg = parent_mem_cgroup(memcg)));
2822 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2823 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2825 if (mem_cgroup_is_root(memcg))
2828 page_counter_uncharge(&memcg->memory, nr_pages);
2829 if (do_memsw_account())
2830 page_counter_uncharge(&memcg->memsw, nr_pages);
2834 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2836 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2838 * Any of the following ensures page->mem_cgroup stability:
2842 * - lock_page_memcg()
2843 * - exclusive reference
2845 page->mem_cgroup = memcg;
2848 #ifdef CONFIG_MEMCG_KMEM
2849 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2852 unsigned int objects = objs_per_slab_page(s, page);
2855 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2860 if (cmpxchg(&page->obj_cgroups, NULL,
2861 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2864 kmemleak_not_leak(vec);
2870 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2872 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2873 * cgroup_mutex, etc.
2875 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2879 if (mem_cgroup_disabled())
2882 page = virt_to_head_page(p);
2885 * Slab objects are accounted individually, not per-page.
2886 * Memcg membership data for each individual object is saved in
2887 * the page->obj_cgroups.
2889 if (page_has_obj_cgroups(page)) {
2890 struct obj_cgroup *objcg;
2893 off = obj_to_index(page->slab_cache, page, p);
2894 objcg = page_obj_cgroups(page)[off];
2896 return obj_cgroup_memcg(objcg);
2901 /* All other pages use page->mem_cgroup */
2902 return page->mem_cgroup;
2905 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2907 struct obj_cgroup *objcg = NULL;
2908 struct mem_cgroup *memcg;
2910 if (unlikely(!current->mm && !current->active_memcg))
2914 if (unlikely(current->active_memcg))
2915 memcg = rcu_dereference(current->active_memcg);
2917 memcg = mem_cgroup_from_task(current);
2919 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2920 objcg = rcu_dereference(memcg->objcg);
2921 if (objcg && obj_cgroup_tryget(objcg))
2929 static int memcg_alloc_cache_id(void)
2934 id = ida_simple_get(&memcg_cache_ida,
2935 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2939 if (id < memcg_nr_cache_ids)
2943 * There's no space for the new id in memcg_caches arrays,
2944 * so we have to grow them.
2946 down_write(&memcg_cache_ids_sem);
2948 size = 2 * (id + 1);
2949 if (size < MEMCG_CACHES_MIN_SIZE)
2950 size = MEMCG_CACHES_MIN_SIZE;
2951 else if (size > MEMCG_CACHES_MAX_SIZE)
2952 size = MEMCG_CACHES_MAX_SIZE;
2954 err = memcg_update_all_list_lrus(size);
2956 memcg_nr_cache_ids = size;
2958 up_write(&memcg_cache_ids_sem);
2961 ida_simple_remove(&memcg_cache_ida, id);
2967 static void memcg_free_cache_id(int id)
2969 ida_simple_remove(&memcg_cache_ida, id);
2973 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2974 * @memcg: memory cgroup to charge
2975 * @gfp: reclaim mode
2976 * @nr_pages: number of pages to charge
2978 * Returns 0 on success, an error code on failure.
2980 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2981 unsigned int nr_pages)
2983 struct page_counter *counter;
2986 ret = try_charge(memcg, gfp, nr_pages);
2990 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2991 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2994 * Enforce __GFP_NOFAIL allocation because callers are not
2995 * prepared to see failures and likely do not have any failure
2998 if (gfp & __GFP_NOFAIL) {
2999 page_counter_charge(&memcg->kmem, nr_pages);
3002 cancel_charge(memcg, nr_pages);
3009 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3010 * @memcg: memcg to uncharge
3011 * @nr_pages: number of pages to uncharge
3013 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3015 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3016 page_counter_uncharge(&memcg->kmem, nr_pages);
3018 page_counter_uncharge(&memcg->memory, nr_pages);
3019 if (do_memsw_account())
3020 page_counter_uncharge(&memcg->memsw, nr_pages);
3024 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3025 * @page: page to charge
3026 * @gfp: reclaim mode
3027 * @order: allocation order
3029 * Returns 0 on success, an error code on failure.
3031 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3033 struct mem_cgroup *memcg;
3036 if (memcg_kmem_bypass())
3039 memcg = get_mem_cgroup_from_current();
3040 if (!mem_cgroup_is_root(memcg)) {
3041 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3043 page->mem_cgroup = memcg;
3044 __SetPageKmemcg(page);
3048 css_put(&memcg->css);
3053 * __memcg_kmem_uncharge_page: uncharge a kmem page
3054 * @page: page to uncharge
3055 * @order: allocation order
3057 void __memcg_kmem_uncharge_page(struct page *page, int order)
3059 struct mem_cgroup *memcg = page->mem_cgroup;
3060 unsigned int nr_pages = 1 << order;
3065 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3066 __memcg_kmem_uncharge(memcg, nr_pages);
3067 page->mem_cgroup = NULL;
3068 css_put(&memcg->css);
3070 /* slab pages do not have PageKmemcg flag set */
3071 if (PageKmemcg(page))
3072 __ClearPageKmemcg(page);
3075 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3077 struct memcg_stock_pcp *stock;
3078 unsigned long flags;
3081 local_irq_save(flags);
3083 stock = this_cpu_ptr(&memcg_stock);
3084 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3085 stock->nr_bytes -= nr_bytes;
3089 local_irq_restore(flags);
3094 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3096 struct obj_cgroup *old = stock->cached_objcg;
3101 if (stock->nr_bytes) {
3102 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3103 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3107 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3112 * The leftover is flushed to the centralized per-memcg value.
3113 * On the next attempt to refill obj stock it will be moved
3114 * to a per-cpu stock (probably, on an other CPU), see
3115 * refill_obj_stock().
3117 * How often it's flushed is a trade-off between the memory
3118 * limit enforcement accuracy and potential CPU contention,
3119 * so it might be changed in the future.
3121 atomic_add(nr_bytes, &old->nr_charged_bytes);
3122 stock->nr_bytes = 0;
3125 obj_cgroup_put(old);
3126 stock->cached_objcg = NULL;
3129 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3130 struct mem_cgroup *root_memcg)
3132 struct mem_cgroup *memcg;
3134 if (stock->cached_objcg) {
3135 memcg = obj_cgroup_memcg(stock->cached_objcg);
3136 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3143 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3145 struct memcg_stock_pcp *stock;
3146 unsigned long flags;
3148 local_irq_save(flags);
3150 stock = this_cpu_ptr(&memcg_stock);
3151 if (stock->cached_objcg != objcg) { /* reset if necessary */
3152 drain_obj_stock(stock);
3153 obj_cgroup_get(objcg);
3154 stock->cached_objcg = objcg;
3155 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3157 stock->nr_bytes += nr_bytes;
3159 if (stock->nr_bytes > PAGE_SIZE)
3160 drain_obj_stock(stock);
3162 local_irq_restore(flags);
3165 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3167 struct mem_cgroup *memcg;
3168 unsigned int nr_pages, nr_bytes;
3171 if (consume_obj_stock(objcg, size))
3175 * In theory, memcg->nr_charged_bytes can have enough
3176 * pre-charged bytes to satisfy the allocation. However,
3177 * flushing memcg->nr_charged_bytes requires two atomic
3178 * operations, and memcg->nr_charged_bytes can't be big,
3179 * so it's better to ignore it and try grab some new pages.
3180 * memcg->nr_charged_bytes will be flushed in
3181 * refill_obj_stock(), called from this function or
3182 * independently later.
3185 memcg = obj_cgroup_memcg(objcg);
3186 css_get(&memcg->css);
3189 nr_pages = size >> PAGE_SHIFT;
3190 nr_bytes = size & (PAGE_SIZE - 1);
3195 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3196 if (!ret && nr_bytes)
3197 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3199 css_put(&memcg->css);
3203 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3205 refill_obj_stock(objcg, size);
3208 #endif /* CONFIG_MEMCG_KMEM */
3210 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3213 * Because tail pages are not marked as "used", set it. We're under
3214 * pgdat->lru_lock and migration entries setup in all page mappings.
3216 void mem_cgroup_split_huge_fixup(struct page *head)
3218 struct mem_cgroup *memcg = head->mem_cgroup;
3221 if (mem_cgroup_disabled())
3224 for (i = 1; i < HPAGE_PMD_NR; i++) {
3225 css_get(&memcg->css);
3226 head[i].mem_cgroup = memcg;
3229 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3231 #ifdef CONFIG_MEMCG_SWAP
3233 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3234 * @entry: swap entry to be moved
3235 * @from: mem_cgroup which the entry is moved from
3236 * @to: mem_cgroup which the entry is moved to
3238 * It succeeds only when the swap_cgroup's record for this entry is the same
3239 * as the mem_cgroup's id of @from.
3241 * Returns 0 on success, -EINVAL on failure.
3243 * The caller must have charged to @to, IOW, called page_counter_charge() about
3244 * both res and memsw, and called css_get().
3246 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3247 struct mem_cgroup *from, struct mem_cgroup *to)
3249 unsigned short old_id, new_id;
3251 old_id = mem_cgroup_id(from);
3252 new_id = mem_cgroup_id(to);
3254 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3255 mod_memcg_state(from, MEMCG_SWAP, -1);
3256 mod_memcg_state(to, MEMCG_SWAP, 1);
3262 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3263 struct mem_cgroup *from, struct mem_cgroup *to)
3269 static DEFINE_MUTEX(memcg_max_mutex);
3271 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3272 unsigned long max, bool memsw)
3274 bool enlarge = false;
3275 bool drained = false;
3277 bool limits_invariant;
3278 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3281 if (signal_pending(current)) {
3286 mutex_lock(&memcg_max_mutex);
3288 * Make sure that the new limit (memsw or memory limit) doesn't
3289 * break our basic invariant rule memory.max <= memsw.max.
3291 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3292 max <= memcg->memsw.max;
3293 if (!limits_invariant) {
3294 mutex_unlock(&memcg_max_mutex);
3298 if (max > counter->max)
3300 ret = page_counter_set_max(counter, max);
3301 mutex_unlock(&memcg_max_mutex);
3307 drain_all_stock(memcg);
3312 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3313 GFP_KERNEL, !memsw)) {
3319 if (!ret && enlarge)
3320 memcg_oom_recover(memcg);
3325 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3327 unsigned long *total_scanned)
3329 unsigned long nr_reclaimed = 0;
3330 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3331 unsigned long reclaimed;
3333 struct mem_cgroup_tree_per_node *mctz;
3334 unsigned long excess;
3335 unsigned long nr_scanned;
3340 mctz = soft_limit_tree_node(pgdat->node_id);
3343 * Do not even bother to check the largest node if the root
3344 * is empty. Do it lockless to prevent lock bouncing. Races
3345 * are acceptable as soft limit is best effort anyway.
3347 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3351 * This loop can run a while, specially if mem_cgroup's continuously
3352 * keep exceeding their soft limit and putting the system under
3359 mz = mem_cgroup_largest_soft_limit_node(mctz);
3364 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3365 gfp_mask, &nr_scanned);
3366 nr_reclaimed += reclaimed;
3367 *total_scanned += nr_scanned;
3368 spin_lock_irq(&mctz->lock);
3369 __mem_cgroup_remove_exceeded(mz, mctz);
3372 * If we failed to reclaim anything from this memory cgroup
3373 * it is time to move on to the next cgroup
3377 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3379 excess = soft_limit_excess(mz->memcg);
3381 * One school of thought says that we should not add
3382 * back the node to the tree if reclaim returns 0.
3383 * But our reclaim could return 0, simply because due
3384 * to priority we are exposing a smaller subset of
3385 * memory to reclaim from. Consider this as a longer
3388 /* If excess == 0, no tree ops */
3389 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3390 spin_unlock_irq(&mctz->lock);
3391 css_put(&mz->memcg->css);
3394 * Could not reclaim anything and there are no more
3395 * mem cgroups to try or we seem to be looping without
3396 * reclaiming anything.
3398 if (!nr_reclaimed &&
3400 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3402 } while (!nr_reclaimed);
3404 css_put(&next_mz->memcg->css);
3405 return nr_reclaimed;
3409 * Test whether @memcg has children, dead or alive. Note that this
3410 * function doesn't care whether @memcg has use_hierarchy enabled and
3411 * returns %true if there are child csses according to the cgroup
3412 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3414 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3419 ret = css_next_child(NULL, &memcg->css);
3425 * Reclaims as many pages from the given memcg as possible.
3427 * Caller is responsible for holding css reference for memcg.
3429 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3431 int nr_retries = MAX_RECLAIM_RETRIES;
3433 /* we call try-to-free pages for make this cgroup empty */
3434 lru_add_drain_all();
3436 drain_all_stock(memcg);
3438 /* try to free all pages in this cgroup */
3439 while (nr_retries && page_counter_read(&memcg->memory)) {
3442 if (signal_pending(current))
3445 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3449 /* maybe some writeback is necessary */
3450 congestion_wait(BLK_RW_ASYNC, HZ/10);
3458 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3459 char *buf, size_t nbytes,
3462 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3464 if (mem_cgroup_is_root(memcg))
3466 return mem_cgroup_force_empty(memcg) ?: nbytes;
3469 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3472 return mem_cgroup_from_css(css)->use_hierarchy;
3475 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3476 struct cftype *cft, u64 val)
3479 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3480 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3482 if (memcg->use_hierarchy == val)
3486 * If parent's use_hierarchy is set, we can't make any modifications
3487 * in the child subtrees. If it is unset, then the change can
3488 * occur, provided the current cgroup has no children.
3490 * For the root cgroup, parent_mem is NULL, we allow value to be
3491 * set if there are no children.
3493 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3494 (val == 1 || val == 0)) {
3495 if (!memcg_has_children(memcg))
3496 memcg->use_hierarchy = val;
3505 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3509 if (mem_cgroup_is_root(memcg)) {
3510 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3511 memcg_page_state(memcg, NR_ANON_MAPPED);
3513 val += memcg_page_state(memcg, MEMCG_SWAP);
3516 val = page_counter_read(&memcg->memory);
3518 val = page_counter_read(&memcg->memsw);
3531 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3534 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3535 struct page_counter *counter;
3537 switch (MEMFILE_TYPE(cft->private)) {
3539 counter = &memcg->memory;
3542 counter = &memcg->memsw;
3545 counter = &memcg->kmem;
3548 counter = &memcg->tcpmem;
3554 switch (MEMFILE_ATTR(cft->private)) {
3556 if (counter == &memcg->memory)
3557 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3558 if (counter == &memcg->memsw)
3559 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3560 return (u64)page_counter_read(counter) * PAGE_SIZE;
3562 return (u64)counter->max * PAGE_SIZE;
3564 return (u64)counter->watermark * PAGE_SIZE;
3566 return counter->failcnt;
3567 case RES_SOFT_LIMIT:
3568 return (u64)memcg->soft_limit * PAGE_SIZE;
3574 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3576 unsigned long stat[MEMCG_NR_STAT] = {0};
3577 struct mem_cgroup *mi;
3580 for_each_online_cpu(cpu)
3581 for (i = 0; i < MEMCG_NR_STAT; i++)
3582 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3584 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3585 for (i = 0; i < MEMCG_NR_STAT; i++)
3586 atomic_long_add(stat[i], &mi->vmstats[i]);
3588 for_each_node(node) {
3589 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3590 struct mem_cgroup_per_node *pi;
3592 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3595 for_each_online_cpu(cpu)
3596 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3598 pn->lruvec_stat_cpu->count[i], cpu);
3600 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3601 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3602 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3606 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3608 unsigned long events[NR_VM_EVENT_ITEMS];
3609 struct mem_cgroup *mi;
3612 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3615 for_each_online_cpu(cpu)
3616 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3617 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3620 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3621 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3622 atomic_long_add(events[i], &mi->vmevents[i]);
3625 #ifdef CONFIG_MEMCG_KMEM
3626 static int memcg_online_kmem(struct mem_cgroup *memcg)
3628 struct obj_cgroup *objcg;
3631 if (cgroup_memory_nokmem)
3634 BUG_ON(memcg->kmemcg_id >= 0);
3635 BUG_ON(memcg->kmem_state);
3637 memcg_id = memcg_alloc_cache_id();
3641 objcg = obj_cgroup_alloc();
3643 memcg_free_cache_id(memcg_id);
3646 objcg->memcg = memcg;
3647 rcu_assign_pointer(memcg->objcg, objcg);
3649 static_branch_enable(&memcg_kmem_enabled_key);
3652 * A memory cgroup is considered kmem-online as soon as it gets
3653 * kmemcg_id. Setting the id after enabling static branching will
3654 * guarantee no one starts accounting before all call sites are
3657 memcg->kmemcg_id = memcg_id;
3658 memcg->kmem_state = KMEM_ONLINE;
3663 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3665 struct cgroup_subsys_state *css;
3666 struct mem_cgroup *parent, *child;
3669 if (memcg->kmem_state != KMEM_ONLINE)
3672 memcg->kmem_state = KMEM_ALLOCATED;
3674 parent = parent_mem_cgroup(memcg);
3676 parent = root_mem_cgroup;
3678 memcg_reparent_objcgs(memcg, parent);
3680 kmemcg_id = memcg->kmemcg_id;
3681 BUG_ON(kmemcg_id < 0);
3684 * Change kmemcg_id of this cgroup and all its descendants to the
3685 * parent's id, and then move all entries from this cgroup's list_lrus
3686 * to ones of the parent. After we have finished, all list_lrus
3687 * corresponding to this cgroup are guaranteed to remain empty. The
3688 * ordering is imposed by list_lru_node->lock taken by
3689 * memcg_drain_all_list_lrus().
3691 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3692 css_for_each_descendant_pre(css, &memcg->css) {
3693 child = mem_cgroup_from_css(css);
3694 BUG_ON(child->kmemcg_id != kmemcg_id);
3695 child->kmemcg_id = parent->kmemcg_id;
3696 if (!memcg->use_hierarchy)
3701 memcg_drain_all_list_lrus(kmemcg_id, parent);
3703 memcg_free_cache_id(kmemcg_id);
3706 static void memcg_free_kmem(struct mem_cgroup *memcg)
3708 /* css_alloc() failed, offlining didn't happen */
3709 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3710 memcg_offline_kmem(memcg);
3713 static int memcg_online_kmem(struct mem_cgroup *memcg)
3717 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3720 static void memcg_free_kmem(struct mem_cgroup *memcg)
3723 #endif /* CONFIG_MEMCG_KMEM */
3725 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3730 mutex_lock(&memcg_max_mutex);
3731 ret = page_counter_set_max(&memcg->kmem, max);
3732 mutex_unlock(&memcg_max_mutex);
3736 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3740 mutex_lock(&memcg_max_mutex);
3742 ret = page_counter_set_max(&memcg->tcpmem, max);
3746 if (!memcg->tcpmem_active) {
3748 * The active flag needs to be written after the static_key
3749 * update. This is what guarantees that the socket activation
3750 * function is the last one to run. See mem_cgroup_sk_alloc()
3751 * for details, and note that we don't mark any socket as
3752 * belonging to this memcg until that flag is up.
3754 * We need to do this, because static_keys will span multiple
3755 * sites, but we can't control their order. If we mark a socket
3756 * as accounted, but the accounting functions are not patched in
3757 * yet, we'll lose accounting.
3759 * We never race with the readers in mem_cgroup_sk_alloc(),
3760 * because when this value change, the code to process it is not
3763 static_branch_inc(&memcg_sockets_enabled_key);
3764 memcg->tcpmem_active = true;
3767 mutex_unlock(&memcg_max_mutex);
3772 * The user of this function is...
3775 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3776 char *buf, size_t nbytes, loff_t off)
3778 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3779 unsigned long nr_pages;
3782 buf = strstrip(buf);
3783 ret = page_counter_memparse(buf, "-1", &nr_pages);
3787 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3789 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3793 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3795 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3798 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3801 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3802 "Please report your usecase to linux-mm@kvack.org if you "
3803 "depend on this functionality.\n");
3804 ret = memcg_update_kmem_max(memcg, nr_pages);
3807 ret = memcg_update_tcp_max(memcg, nr_pages);
3811 case RES_SOFT_LIMIT:
3812 memcg->soft_limit = nr_pages;
3816 return ret ?: nbytes;
3819 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3820 size_t nbytes, loff_t off)
3822 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3823 struct page_counter *counter;
3825 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3827 counter = &memcg->memory;
3830 counter = &memcg->memsw;
3833 counter = &memcg->kmem;
3836 counter = &memcg->tcpmem;
3842 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3844 page_counter_reset_watermark(counter);
3847 counter->failcnt = 0;
3856 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3859 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3863 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3864 struct cftype *cft, u64 val)
3866 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3868 if (val & ~MOVE_MASK)
3872 * No kind of locking is needed in here, because ->can_attach() will
3873 * check this value once in the beginning of the process, and then carry
3874 * on with stale data. This means that changes to this value will only
3875 * affect task migrations starting after the change.
3877 memcg->move_charge_at_immigrate = val;
3881 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3882 struct cftype *cft, u64 val)
3890 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3891 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3892 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3894 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3895 int nid, unsigned int lru_mask, bool tree)
3897 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3898 unsigned long nr = 0;
3901 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3904 if (!(BIT(lru) & lru_mask))
3907 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3909 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3914 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3915 unsigned int lru_mask,
3918 unsigned long nr = 0;
3922 if (!(BIT(lru) & lru_mask))
3925 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3927 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3932 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3936 unsigned int lru_mask;
3939 static const struct numa_stat stats[] = {
3940 { "total", LRU_ALL },
3941 { "file", LRU_ALL_FILE },
3942 { "anon", LRU_ALL_ANON },
3943 { "unevictable", BIT(LRU_UNEVICTABLE) },
3945 const struct numa_stat *stat;
3947 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3949 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3950 seq_printf(m, "%s=%lu", stat->name,
3951 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3953 for_each_node_state(nid, N_MEMORY)
3954 seq_printf(m, " N%d=%lu", nid,
3955 mem_cgroup_node_nr_lru_pages(memcg, nid,
3956 stat->lru_mask, false));
3960 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3962 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3963 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3965 for_each_node_state(nid, N_MEMORY)
3966 seq_printf(m, " N%d=%lu", nid,
3967 mem_cgroup_node_nr_lru_pages(memcg, nid,
3968 stat->lru_mask, true));
3974 #endif /* CONFIG_NUMA */
3976 static const unsigned int memcg1_stats[] = {
3979 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3989 static const char *const memcg1_stat_names[] = {
3992 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4002 /* Universal VM events cgroup1 shows, original sort order */
4003 static const unsigned int memcg1_events[] = {
4010 static int memcg_stat_show(struct seq_file *m, void *v)
4012 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4013 unsigned long memory, memsw;
4014 struct mem_cgroup *mi;
4017 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4019 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4022 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4024 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4025 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4026 if (memcg1_stats[i] == NR_ANON_THPS)
4029 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4032 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4033 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4034 memcg_events_local(memcg, memcg1_events[i]));
4036 for (i = 0; i < NR_LRU_LISTS; i++)
4037 seq_printf(m, "%s %lu\n", lru_list_name(i),
4038 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4041 /* Hierarchical information */
4042 memory = memsw = PAGE_COUNTER_MAX;
4043 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4044 memory = min(memory, READ_ONCE(mi->memory.max));
4045 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4047 seq_printf(m, "hierarchical_memory_limit %llu\n",
4048 (u64)memory * PAGE_SIZE);
4049 if (do_memsw_account())
4050 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4051 (u64)memsw * PAGE_SIZE);
4053 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4054 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4056 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4057 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
4061 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4062 seq_printf(m, "total_%s %llu\n",
4063 vm_event_name(memcg1_events[i]),
4064 (u64)memcg_events(memcg, memcg1_events[i]));
4066 for (i = 0; i < NR_LRU_LISTS; i++)
4067 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4068 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4071 #ifdef CONFIG_DEBUG_VM
4074 struct mem_cgroup_per_node *mz;
4075 unsigned long anon_cost = 0;
4076 unsigned long file_cost = 0;
4078 for_each_online_pgdat(pgdat) {
4079 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4081 anon_cost += mz->lruvec.anon_cost;
4082 file_cost += mz->lruvec.file_cost;
4084 seq_printf(m, "anon_cost %lu\n", anon_cost);
4085 seq_printf(m, "file_cost %lu\n", file_cost);
4092 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4095 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4097 return mem_cgroup_swappiness(memcg);
4100 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4101 struct cftype *cft, u64 val)
4103 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4109 memcg->swappiness = val;
4111 vm_swappiness = val;
4116 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4118 struct mem_cgroup_threshold_ary *t;
4119 unsigned long usage;
4124 t = rcu_dereference(memcg->thresholds.primary);
4126 t = rcu_dereference(memcg->memsw_thresholds.primary);
4131 usage = mem_cgroup_usage(memcg, swap);
4134 * current_threshold points to threshold just below or equal to usage.
4135 * If it's not true, a threshold was crossed after last
4136 * call of __mem_cgroup_threshold().
4138 i = t->current_threshold;
4141 * Iterate backward over array of thresholds starting from
4142 * current_threshold and check if a threshold is crossed.
4143 * If none of thresholds below usage is crossed, we read
4144 * only one element of the array here.
4146 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4147 eventfd_signal(t->entries[i].eventfd, 1);
4149 /* i = current_threshold + 1 */
4153 * Iterate forward over array of thresholds starting from
4154 * current_threshold+1 and check if a threshold is crossed.
4155 * If none of thresholds above usage is crossed, we read
4156 * only one element of the array here.
4158 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4159 eventfd_signal(t->entries[i].eventfd, 1);
4161 /* Update current_threshold */
4162 t->current_threshold = i - 1;
4167 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4170 __mem_cgroup_threshold(memcg, false);
4171 if (do_memsw_account())
4172 __mem_cgroup_threshold(memcg, true);
4174 memcg = parent_mem_cgroup(memcg);
4178 static int compare_thresholds(const void *a, const void *b)
4180 const struct mem_cgroup_threshold *_a = a;
4181 const struct mem_cgroup_threshold *_b = b;
4183 if (_a->threshold > _b->threshold)
4186 if (_a->threshold < _b->threshold)
4192 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4194 struct mem_cgroup_eventfd_list *ev;
4196 spin_lock(&memcg_oom_lock);
4198 list_for_each_entry(ev, &memcg->oom_notify, list)
4199 eventfd_signal(ev->eventfd, 1);
4201 spin_unlock(&memcg_oom_lock);
4205 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4207 struct mem_cgroup *iter;
4209 for_each_mem_cgroup_tree(iter, memcg)
4210 mem_cgroup_oom_notify_cb(iter);
4213 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4214 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4216 struct mem_cgroup_thresholds *thresholds;
4217 struct mem_cgroup_threshold_ary *new;
4218 unsigned long threshold;
4219 unsigned long usage;
4222 ret = page_counter_memparse(args, "-1", &threshold);
4226 mutex_lock(&memcg->thresholds_lock);
4229 thresholds = &memcg->thresholds;
4230 usage = mem_cgroup_usage(memcg, false);
4231 } else if (type == _MEMSWAP) {
4232 thresholds = &memcg->memsw_thresholds;
4233 usage = mem_cgroup_usage(memcg, true);
4237 /* Check if a threshold crossed before adding a new one */
4238 if (thresholds->primary)
4239 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4241 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4243 /* Allocate memory for new array of thresholds */
4244 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4251 /* Copy thresholds (if any) to new array */
4252 if (thresholds->primary) {
4253 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4254 sizeof(struct mem_cgroup_threshold));
4257 /* Add new threshold */
4258 new->entries[size - 1].eventfd = eventfd;
4259 new->entries[size - 1].threshold = threshold;
4261 /* Sort thresholds. Registering of new threshold isn't time-critical */
4262 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4263 compare_thresholds, NULL);
4265 /* Find current threshold */
4266 new->current_threshold = -1;
4267 for (i = 0; i < size; i++) {
4268 if (new->entries[i].threshold <= usage) {
4270 * new->current_threshold will not be used until
4271 * rcu_assign_pointer(), so it's safe to increment
4274 ++new->current_threshold;
4279 /* Free old spare buffer and save old primary buffer as spare */
4280 kfree(thresholds->spare);
4281 thresholds->spare = thresholds->primary;
4283 rcu_assign_pointer(thresholds->primary, new);
4285 /* To be sure that nobody uses thresholds */
4289 mutex_unlock(&memcg->thresholds_lock);
4294 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4295 struct eventfd_ctx *eventfd, const char *args)
4297 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4300 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4301 struct eventfd_ctx *eventfd, const char *args)
4303 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4306 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4307 struct eventfd_ctx *eventfd, enum res_type type)
4309 struct mem_cgroup_thresholds *thresholds;
4310 struct mem_cgroup_threshold_ary *new;
4311 unsigned long usage;
4312 int i, j, size, entries;
4314 mutex_lock(&memcg->thresholds_lock);
4317 thresholds = &memcg->thresholds;
4318 usage = mem_cgroup_usage(memcg, false);
4319 } else if (type == _MEMSWAP) {
4320 thresholds = &memcg->memsw_thresholds;
4321 usage = mem_cgroup_usage(memcg, true);
4325 if (!thresholds->primary)
4328 /* Check if a threshold crossed before removing */
4329 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4331 /* Calculate new number of threshold */
4333 for (i = 0; i < thresholds->primary->size; i++) {
4334 if (thresholds->primary->entries[i].eventfd != eventfd)
4340 new = thresholds->spare;
4342 /* If no items related to eventfd have been cleared, nothing to do */
4346 /* Set thresholds array to NULL if we don't have thresholds */
4355 /* Copy thresholds and find current threshold */
4356 new->current_threshold = -1;
4357 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4358 if (thresholds->primary->entries[i].eventfd == eventfd)
4361 new->entries[j] = thresholds->primary->entries[i];
4362 if (new->entries[j].threshold <= usage) {
4364 * new->current_threshold will not be used
4365 * until rcu_assign_pointer(), so it's safe to increment
4368 ++new->current_threshold;
4374 /* Swap primary and spare array */
4375 thresholds->spare = thresholds->primary;
4377 rcu_assign_pointer(thresholds->primary, new);
4379 /* To be sure that nobody uses thresholds */
4382 /* If all events are unregistered, free the spare array */
4384 kfree(thresholds->spare);
4385 thresholds->spare = NULL;
4388 mutex_unlock(&memcg->thresholds_lock);
4391 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4392 struct eventfd_ctx *eventfd)
4394 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4397 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4398 struct eventfd_ctx *eventfd)
4400 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4403 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4404 struct eventfd_ctx *eventfd, const char *args)
4406 struct mem_cgroup_eventfd_list *event;
4408 event = kmalloc(sizeof(*event), GFP_KERNEL);
4412 spin_lock(&memcg_oom_lock);
4414 event->eventfd = eventfd;
4415 list_add(&event->list, &memcg->oom_notify);
4417 /* already in OOM ? */
4418 if (memcg->under_oom)
4419 eventfd_signal(eventfd, 1);
4420 spin_unlock(&memcg_oom_lock);
4425 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4426 struct eventfd_ctx *eventfd)
4428 struct mem_cgroup_eventfd_list *ev, *tmp;
4430 spin_lock(&memcg_oom_lock);
4432 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4433 if (ev->eventfd == eventfd) {
4434 list_del(&ev->list);
4439 spin_unlock(&memcg_oom_lock);
4442 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4444 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4446 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4447 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4448 seq_printf(sf, "oom_kill %lu\n",
4449 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4453 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4454 struct cftype *cft, u64 val)
4456 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4458 /* cannot set to root cgroup and only 0 and 1 are allowed */
4459 if (!css->parent || !((val == 0) || (val == 1)))
4462 memcg->oom_kill_disable = val;
4464 memcg_oom_recover(memcg);
4469 #ifdef CONFIG_CGROUP_WRITEBACK
4471 #include <trace/events/writeback.h>
4473 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4475 return wb_domain_init(&memcg->cgwb_domain, gfp);
4478 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4480 wb_domain_exit(&memcg->cgwb_domain);
4483 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4485 wb_domain_size_changed(&memcg->cgwb_domain);
4488 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4490 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4492 if (!memcg->css.parent)
4495 return &memcg->cgwb_domain;
4499 * idx can be of type enum memcg_stat_item or node_stat_item.
4500 * Keep in sync with memcg_exact_page().
4502 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4504 long x = atomic_long_read(&memcg->vmstats[idx]);
4507 for_each_online_cpu(cpu)
4508 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4515 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4516 * @wb: bdi_writeback in question
4517 * @pfilepages: out parameter for number of file pages
4518 * @pheadroom: out parameter for number of allocatable pages according to memcg
4519 * @pdirty: out parameter for number of dirty pages
4520 * @pwriteback: out parameter for number of pages under writeback
4522 * Determine the numbers of file, headroom, dirty, and writeback pages in
4523 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4524 * is a bit more involved.
4526 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4527 * headroom is calculated as the lowest headroom of itself and the
4528 * ancestors. Note that this doesn't consider the actual amount of
4529 * available memory in the system. The caller should further cap
4530 * *@pheadroom accordingly.
4532 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4533 unsigned long *pheadroom, unsigned long *pdirty,
4534 unsigned long *pwriteback)
4536 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4537 struct mem_cgroup *parent;
4539 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4541 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4542 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4543 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4544 *pheadroom = PAGE_COUNTER_MAX;
4546 while ((parent = parent_mem_cgroup(memcg))) {
4547 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4548 READ_ONCE(memcg->memory.high));
4549 unsigned long used = page_counter_read(&memcg->memory);
4551 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4557 * Foreign dirty flushing
4559 * There's an inherent mismatch between memcg and writeback. The former
4560 * trackes ownership per-page while the latter per-inode. This was a
4561 * deliberate design decision because honoring per-page ownership in the
4562 * writeback path is complicated, may lead to higher CPU and IO overheads
4563 * and deemed unnecessary given that write-sharing an inode across
4564 * different cgroups isn't a common use-case.
4566 * Combined with inode majority-writer ownership switching, this works well
4567 * enough in most cases but there are some pathological cases. For
4568 * example, let's say there are two cgroups A and B which keep writing to
4569 * different but confined parts of the same inode. B owns the inode and
4570 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4571 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4572 * triggering background writeback. A will be slowed down without a way to
4573 * make writeback of the dirty pages happen.
4575 * Conditions like the above can lead to a cgroup getting repatedly and
4576 * severely throttled after making some progress after each
4577 * dirty_expire_interval while the underyling IO device is almost
4580 * Solving this problem completely requires matching the ownership tracking
4581 * granularities between memcg and writeback in either direction. However,
4582 * the more egregious behaviors can be avoided by simply remembering the
4583 * most recent foreign dirtying events and initiating remote flushes on
4584 * them when local writeback isn't enough to keep the memory clean enough.
4586 * The following two functions implement such mechanism. When a foreign
4587 * page - a page whose memcg and writeback ownerships don't match - is
4588 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4589 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4590 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4591 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4592 * foreign bdi_writebacks which haven't expired. Both the numbers of
4593 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4594 * limited to MEMCG_CGWB_FRN_CNT.
4596 * The mechanism only remembers IDs and doesn't hold any object references.
4597 * As being wrong occasionally doesn't matter, updates and accesses to the
4598 * records are lockless and racy.
4600 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4601 struct bdi_writeback *wb)
4603 struct mem_cgroup *memcg = page->mem_cgroup;
4604 struct memcg_cgwb_frn *frn;
4605 u64 now = get_jiffies_64();
4606 u64 oldest_at = now;
4610 trace_track_foreign_dirty(page, wb);
4613 * Pick the slot to use. If there is already a slot for @wb, keep
4614 * using it. If not replace the oldest one which isn't being
4617 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4618 frn = &memcg->cgwb_frn[i];
4619 if (frn->bdi_id == wb->bdi->id &&
4620 frn->memcg_id == wb->memcg_css->id)
4622 if (time_before64(frn->at, oldest_at) &&
4623 atomic_read(&frn->done.cnt) == 1) {
4625 oldest_at = frn->at;
4629 if (i < MEMCG_CGWB_FRN_CNT) {
4631 * Re-using an existing one. Update timestamp lazily to
4632 * avoid making the cacheline hot. We want them to be
4633 * reasonably up-to-date and significantly shorter than
4634 * dirty_expire_interval as that's what expires the record.
4635 * Use the shorter of 1s and dirty_expire_interval / 8.
4637 unsigned long update_intv =
4638 min_t(unsigned long, HZ,
4639 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4641 if (time_before64(frn->at, now - update_intv))
4643 } else if (oldest >= 0) {
4644 /* replace the oldest free one */
4645 frn = &memcg->cgwb_frn[oldest];
4646 frn->bdi_id = wb->bdi->id;
4647 frn->memcg_id = wb->memcg_css->id;
4652 /* issue foreign writeback flushes for recorded foreign dirtying events */
4653 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4655 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4656 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4657 u64 now = jiffies_64;
4660 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4661 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4664 * If the record is older than dirty_expire_interval,
4665 * writeback on it has already started. No need to kick it
4666 * off again. Also, don't start a new one if there's
4667 * already one in flight.
4669 if (time_after64(frn->at, now - intv) &&
4670 atomic_read(&frn->done.cnt) == 1) {
4672 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4673 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4674 WB_REASON_FOREIGN_FLUSH,
4680 #else /* CONFIG_CGROUP_WRITEBACK */
4682 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4687 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4691 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4695 #endif /* CONFIG_CGROUP_WRITEBACK */
4698 * DO NOT USE IN NEW FILES.
4700 * "cgroup.event_control" implementation.
4702 * This is way over-engineered. It tries to support fully configurable
4703 * events for each user. Such level of flexibility is completely
4704 * unnecessary especially in the light of the planned unified hierarchy.
4706 * Please deprecate this and replace with something simpler if at all
4711 * Unregister event and free resources.
4713 * Gets called from workqueue.
4715 static void memcg_event_remove(struct work_struct *work)
4717 struct mem_cgroup_event *event =
4718 container_of(work, struct mem_cgroup_event, remove);
4719 struct mem_cgroup *memcg = event->memcg;
4721 remove_wait_queue(event->wqh, &event->wait);
4723 event->unregister_event(memcg, event->eventfd);
4725 /* Notify userspace the event is going away. */
4726 eventfd_signal(event->eventfd, 1);
4728 eventfd_ctx_put(event->eventfd);
4730 css_put(&memcg->css);
4734 * Gets called on EPOLLHUP on eventfd when user closes it.
4736 * Called with wqh->lock held and interrupts disabled.
4738 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4739 int sync, void *key)
4741 struct mem_cgroup_event *event =
4742 container_of(wait, struct mem_cgroup_event, wait);
4743 struct mem_cgroup *memcg = event->memcg;
4744 __poll_t flags = key_to_poll(key);
4746 if (flags & EPOLLHUP) {
4748 * If the event has been detached at cgroup removal, we
4749 * can simply return knowing the other side will cleanup
4752 * We can't race against event freeing since the other
4753 * side will require wqh->lock via remove_wait_queue(),
4756 spin_lock(&memcg->event_list_lock);
4757 if (!list_empty(&event->list)) {
4758 list_del_init(&event->list);
4760 * We are in atomic context, but cgroup_event_remove()
4761 * may sleep, so we have to call it in workqueue.
4763 schedule_work(&event->remove);
4765 spin_unlock(&memcg->event_list_lock);
4771 static void memcg_event_ptable_queue_proc(struct file *file,
4772 wait_queue_head_t *wqh, poll_table *pt)
4774 struct mem_cgroup_event *event =
4775 container_of(pt, struct mem_cgroup_event, pt);
4778 add_wait_queue(wqh, &event->wait);
4782 * DO NOT USE IN NEW FILES.
4784 * Parse input and register new cgroup event handler.
4786 * Input must be in format '<event_fd> <control_fd> <args>'.
4787 * Interpretation of args is defined by control file implementation.
4789 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4790 char *buf, size_t nbytes, loff_t off)
4792 struct cgroup_subsys_state *css = of_css(of);
4793 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4794 struct mem_cgroup_event *event;
4795 struct cgroup_subsys_state *cfile_css;
4796 unsigned int efd, cfd;
4803 buf = strstrip(buf);
4805 efd = simple_strtoul(buf, &endp, 10);
4810 cfd = simple_strtoul(buf, &endp, 10);
4811 if ((*endp != ' ') && (*endp != '\0'))
4815 event = kzalloc(sizeof(*event), GFP_KERNEL);
4819 event->memcg = memcg;
4820 INIT_LIST_HEAD(&event->list);
4821 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4822 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4823 INIT_WORK(&event->remove, memcg_event_remove);
4831 event->eventfd = eventfd_ctx_fileget(efile.file);
4832 if (IS_ERR(event->eventfd)) {
4833 ret = PTR_ERR(event->eventfd);
4840 goto out_put_eventfd;
4843 /* the process need read permission on control file */
4844 /* AV: shouldn't we check that it's been opened for read instead? */
4845 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4850 * Determine the event callbacks and set them in @event. This used
4851 * to be done via struct cftype but cgroup core no longer knows
4852 * about these events. The following is crude but the whole thing
4853 * is for compatibility anyway.
4855 * DO NOT ADD NEW FILES.
4857 name = cfile.file->f_path.dentry->d_name.name;
4859 if (!strcmp(name, "memory.usage_in_bytes")) {
4860 event->register_event = mem_cgroup_usage_register_event;
4861 event->unregister_event = mem_cgroup_usage_unregister_event;
4862 } else if (!strcmp(name, "memory.oom_control")) {
4863 event->register_event = mem_cgroup_oom_register_event;
4864 event->unregister_event = mem_cgroup_oom_unregister_event;
4865 } else if (!strcmp(name, "memory.pressure_level")) {
4866 event->register_event = vmpressure_register_event;
4867 event->unregister_event = vmpressure_unregister_event;
4868 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4869 event->register_event = memsw_cgroup_usage_register_event;
4870 event->unregister_event = memsw_cgroup_usage_unregister_event;
4877 * Verify @cfile should belong to @css. Also, remaining events are
4878 * automatically removed on cgroup destruction but the removal is
4879 * asynchronous, so take an extra ref on @css.
4881 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4882 &memory_cgrp_subsys);
4884 if (IS_ERR(cfile_css))
4886 if (cfile_css != css) {
4891 ret = event->register_event(memcg, event->eventfd, buf);
4895 vfs_poll(efile.file, &event->pt);
4897 spin_lock(&memcg->event_list_lock);
4898 list_add(&event->list, &memcg->event_list);
4899 spin_unlock(&memcg->event_list_lock);
4911 eventfd_ctx_put(event->eventfd);
4920 static struct cftype mem_cgroup_legacy_files[] = {
4922 .name = "usage_in_bytes",
4923 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4924 .read_u64 = mem_cgroup_read_u64,
4927 .name = "max_usage_in_bytes",
4928 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4929 .write = mem_cgroup_reset,
4930 .read_u64 = mem_cgroup_read_u64,
4933 .name = "limit_in_bytes",
4934 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4935 .write = mem_cgroup_write,
4936 .read_u64 = mem_cgroup_read_u64,
4939 .name = "soft_limit_in_bytes",
4940 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4941 .write = mem_cgroup_write,
4942 .read_u64 = mem_cgroup_read_u64,
4946 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4947 .write = mem_cgroup_reset,
4948 .read_u64 = mem_cgroup_read_u64,
4952 .seq_show = memcg_stat_show,
4955 .name = "force_empty",
4956 .write = mem_cgroup_force_empty_write,
4959 .name = "use_hierarchy",
4960 .write_u64 = mem_cgroup_hierarchy_write,
4961 .read_u64 = mem_cgroup_hierarchy_read,
4964 .name = "cgroup.event_control", /* XXX: for compat */
4965 .write = memcg_write_event_control,
4966 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4969 .name = "swappiness",
4970 .read_u64 = mem_cgroup_swappiness_read,
4971 .write_u64 = mem_cgroup_swappiness_write,
4974 .name = "move_charge_at_immigrate",
4975 .read_u64 = mem_cgroup_move_charge_read,
4976 .write_u64 = mem_cgroup_move_charge_write,
4979 .name = "oom_control",
4980 .seq_show = mem_cgroup_oom_control_read,
4981 .write_u64 = mem_cgroup_oom_control_write,
4982 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4985 .name = "pressure_level",
4989 .name = "numa_stat",
4990 .seq_show = memcg_numa_stat_show,
4994 .name = "kmem.limit_in_bytes",
4995 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4996 .write = mem_cgroup_write,
4997 .read_u64 = mem_cgroup_read_u64,
5000 .name = "kmem.usage_in_bytes",
5001 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5002 .read_u64 = mem_cgroup_read_u64,
5005 .name = "kmem.failcnt",
5006 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5007 .write = mem_cgroup_reset,
5008 .read_u64 = mem_cgroup_read_u64,
5011 .name = "kmem.max_usage_in_bytes",
5012 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5013 .write = mem_cgroup_reset,
5014 .read_u64 = mem_cgroup_read_u64,
5016 #if defined(CONFIG_MEMCG_KMEM) && \
5017 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5019 .name = "kmem.slabinfo",
5020 .seq_show = memcg_slab_show,
5024 .name = "kmem.tcp.limit_in_bytes",
5025 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5026 .write = mem_cgroup_write,
5027 .read_u64 = mem_cgroup_read_u64,
5030 .name = "kmem.tcp.usage_in_bytes",
5031 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5032 .read_u64 = mem_cgroup_read_u64,
5035 .name = "kmem.tcp.failcnt",
5036 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5037 .write = mem_cgroup_reset,
5038 .read_u64 = mem_cgroup_read_u64,
5041 .name = "kmem.tcp.max_usage_in_bytes",
5042 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5043 .write = mem_cgroup_reset,
5044 .read_u64 = mem_cgroup_read_u64,
5046 { }, /* terminate */
5050 * Private memory cgroup IDR
5052 * Swap-out records and page cache shadow entries need to store memcg
5053 * references in constrained space, so we maintain an ID space that is
5054 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5055 * memory-controlled cgroups to 64k.
5057 * However, there usually are many references to the offline CSS after
5058 * the cgroup has been destroyed, such as page cache or reclaimable
5059 * slab objects, that don't need to hang on to the ID. We want to keep
5060 * those dead CSS from occupying IDs, or we might quickly exhaust the
5061 * relatively small ID space and prevent the creation of new cgroups
5062 * even when there are much fewer than 64k cgroups - possibly none.
5064 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5065 * be freed and recycled when it's no longer needed, which is usually
5066 * when the CSS is offlined.
5068 * The only exception to that are records of swapped out tmpfs/shmem
5069 * pages that need to be attributed to live ancestors on swapin. But
5070 * those references are manageable from userspace.
5073 static DEFINE_IDR(mem_cgroup_idr);
5075 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5077 if (memcg->id.id > 0) {
5078 idr_remove(&mem_cgroup_idr, memcg->id.id);
5083 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5086 refcount_add(n, &memcg->id.ref);
5089 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5091 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5092 mem_cgroup_id_remove(memcg);
5094 /* Memcg ID pins CSS */
5095 css_put(&memcg->css);
5099 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5101 mem_cgroup_id_put_many(memcg, 1);
5105 * mem_cgroup_from_id - look up a memcg from a memcg id
5106 * @id: the memcg id to look up
5108 * Caller must hold rcu_read_lock().
5110 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5112 WARN_ON_ONCE(!rcu_read_lock_held());
5113 return idr_find(&mem_cgroup_idr, id);
5116 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5118 struct mem_cgroup_per_node *pn;
5121 * This routine is called against possible nodes.
5122 * But it's BUG to call kmalloc() against offline node.
5124 * TODO: this routine can waste much memory for nodes which will
5125 * never be onlined. It's better to use memory hotplug callback
5128 if (!node_state(node, N_NORMAL_MEMORY))
5130 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5134 /* We charge the parent cgroup, never the current task */
5135 WARN_ON_ONCE(!current->active_memcg);
5137 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5138 GFP_KERNEL_ACCOUNT);
5139 if (!pn->lruvec_stat_local) {
5144 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5145 GFP_KERNEL_ACCOUNT);
5146 if (!pn->lruvec_stat_cpu) {
5147 free_percpu(pn->lruvec_stat_local);
5152 lruvec_init(&pn->lruvec);
5153 pn->usage_in_excess = 0;
5154 pn->on_tree = false;
5157 memcg->nodeinfo[node] = pn;
5161 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5163 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5168 free_percpu(pn->lruvec_stat_cpu);
5169 free_percpu(pn->lruvec_stat_local);
5173 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5178 free_mem_cgroup_per_node_info(memcg, node);
5179 free_percpu(memcg->vmstats_percpu);
5180 free_percpu(memcg->vmstats_local);
5184 static void mem_cgroup_free(struct mem_cgroup *memcg)
5186 memcg_wb_domain_exit(memcg);
5188 * Flush percpu vmstats and vmevents to guarantee the value correctness
5189 * on parent's and all ancestor levels.
5191 memcg_flush_percpu_vmstats(memcg);
5192 memcg_flush_percpu_vmevents(memcg);
5193 __mem_cgroup_free(memcg);
5196 static struct mem_cgroup *mem_cgroup_alloc(void)
5198 struct mem_cgroup *memcg;
5201 int __maybe_unused i;
5202 long error = -ENOMEM;
5204 size = sizeof(struct mem_cgroup);
5205 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5207 memcg = kzalloc(size, GFP_KERNEL);
5209 return ERR_PTR(error);
5211 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5212 1, MEM_CGROUP_ID_MAX,
5214 if (memcg->id.id < 0) {
5215 error = memcg->id.id;
5219 /* We charge the parent cgroup, never the current task */
5220 WARN_ON_ONCE(!current->active_memcg);
5222 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5223 GFP_KERNEL_ACCOUNT);
5224 if (!memcg->vmstats_local)
5227 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5228 GFP_KERNEL_ACCOUNT);
5229 if (!memcg->vmstats_percpu)
5233 if (alloc_mem_cgroup_per_node_info(memcg, node))
5236 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5239 INIT_WORK(&memcg->high_work, high_work_func);
5240 INIT_LIST_HEAD(&memcg->oom_notify);
5241 mutex_init(&memcg->thresholds_lock);
5242 spin_lock_init(&memcg->move_lock);
5243 vmpressure_init(&memcg->vmpressure);
5244 INIT_LIST_HEAD(&memcg->event_list);
5245 spin_lock_init(&memcg->event_list_lock);
5246 memcg->socket_pressure = jiffies;
5247 #ifdef CONFIG_MEMCG_KMEM
5248 memcg->kmemcg_id = -1;
5249 INIT_LIST_HEAD(&memcg->objcg_list);
5251 #ifdef CONFIG_CGROUP_WRITEBACK
5252 INIT_LIST_HEAD(&memcg->cgwb_list);
5253 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5254 memcg->cgwb_frn[i].done =
5255 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5257 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5258 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5259 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5260 memcg->deferred_split_queue.split_queue_len = 0;
5262 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5265 mem_cgroup_id_remove(memcg);
5266 __mem_cgroup_free(memcg);
5267 return ERR_PTR(error);
5270 static struct cgroup_subsys_state * __ref
5271 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5273 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5274 struct mem_cgroup *memcg;
5275 long error = -ENOMEM;
5277 memalloc_use_memcg(parent);
5278 memcg = mem_cgroup_alloc();
5279 memalloc_unuse_memcg();
5281 return ERR_CAST(memcg);
5283 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5284 memcg->soft_limit = PAGE_COUNTER_MAX;
5285 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5287 memcg->swappiness = mem_cgroup_swappiness(parent);
5288 memcg->oom_kill_disable = parent->oom_kill_disable;
5290 if (parent && parent->use_hierarchy) {
5291 memcg->use_hierarchy = true;
5292 page_counter_init(&memcg->memory, &parent->memory);
5293 page_counter_init(&memcg->swap, &parent->swap);
5294 page_counter_init(&memcg->memsw, &parent->memsw);
5295 page_counter_init(&memcg->kmem, &parent->kmem);
5296 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5298 page_counter_init(&memcg->memory, NULL);
5299 page_counter_init(&memcg->swap, NULL);
5300 page_counter_init(&memcg->memsw, NULL);
5301 page_counter_init(&memcg->kmem, NULL);
5302 page_counter_init(&memcg->tcpmem, NULL);
5304 * Deeper hierachy with use_hierarchy == false doesn't make
5305 * much sense so let cgroup subsystem know about this
5306 * unfortunate state in our controller.
5308 if (parent != root_mem_cgroup)
5309 memory_cgrp_subsys.broken_hierarchy = true;
5312 /* The following stuff does not apply to the root */
5314 root_mem_cgroup = memcg;
5318 error = memcg_online_kmem(memcg);
5322 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5323 static_branch_inc(&memcg_sockets_enabled_key);
5327 mem_cgroup_id_remove(memcg);
5328 mem_cgroup_free(memcg);
5329 return ERR_PTR(error);
5332 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5334 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5337 * A memcg must be visible for memcg_expand_shrinker_maps()
5338 * by the time the maps are allocated. So, we allocate maps
5339 * here, when for_each_mem_cgroup() can't skip it.
5341 if (memcg_alloc_shrinker_maps(memcg)) {
5342 mem_cgroup_id_remove(memcg);
5346 /* Online state pins memcg ID, memcg ID pins CSS */
5347 refcount_set(&memcg->id.ref, 1);
5352 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5354 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5355 struct mem_cgroup_event *event, *tmp;
5358 * Unregister events and notify userspace.
5359 * Notify userspace about cgroup removing only after rmdir of cgroup
5360 * directory to avoid race between userspace and kernelspace.
5362 spin_lock(&memcg->event_list_lock);
5363 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5364 list_del_init(&event->list);
5365 schedule_work(&event->remove);
5367 spin_unlock(&memcg->event_list_lock);
5369 page_counter_set_min(&memcg->memory, 0);
5370 page_counter_set_low(&memcg->memory, 0);
5372 memcg_offline_kmem(memcg);
5373 wb_memcg_offline(memcg);
5375 drain_all_stock(memcg);
5377 mem_cgroup_id_put(memcg);
5380 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5384 invalidate_reclaim_iterators(memcg);
5387 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5389 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5390 int __maybe_unused i;
5392 #ifdef CONFIG_CGROUP_WRITEBACK
5393 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5394 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5396 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5397 static_branch_dec(&memcg_sockets_enabled_key);
5399 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5400 static_branch_dec(&memcg_sockets_enabled_key);
5402 vmpressure_cleanup(&memcg->vmpressure);
5403 cancel_work_sync(&memcg->high_work);
5404 mem_cgroup_remove_from_trees(memcg);
5405 memcg_free_shrinker_maps(memcg);
5406 memcg_free_kmem(memcg);
5407 mem_cgroup_free(memcg);
5411 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5412 * @css: the target css
5414 * Reset the states of the mem_cgroup associated with @css. This is
5415 * invoked when the userland requests disabling on the default hierarchy
5416 * but the memcg is pinned through dependency. The memcg should stop
5417 * applying policies and should revert to the vanilla state as it may be
5418 * made visible again.
5420 * The current implementation only resets the essential configurations.
5421 * This needs to be expanded to cover all the visible parts.
5423 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5425 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5427 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5428 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5429 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5430 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5431 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5432 page_counter_set_min(&memcg->memory, 0);
5433 page_counter_set_low(&memcg->memory, 0);
5434 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5435 memcg->soft_limit = PAGE_COUNTER_MAX;
5436 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5437 memcg_wb_domain_size_changed(memcg);
5441 /* Handlers for move charge at task migration. */
5442 static int mem_cgroup_do_precharge(unsigned long count)
5446 /* Try a single bulk charge without reclaim first, kswapd may wake */
5447 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5449 mc.precharge += count;
5453 /* Try charges one by one with reclaim, but do not retry */
5455 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5469 enum mc_target_type {
5476 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5477 unsigned long addr, pte_t ptent)
5479 struct page *page = vm_normal_page(vma, addr, ptent);
5481 if (!page || !page_mapped(page))
5483 if (PageAnon(page)) {
5484 if (!(mc.flags & MOVE_ANON))
5487 if (!(mc.flags & MOVE_FILE))
5490 if (!get_page_unless_zero(page))
5496 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5497 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5498 pte_t ptent, swp_entry_t *entry)
5500 struct page *page = NULL;
5501 swp_entry_t ent = pte_to_swp_entry(ptent);
5503 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5507 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5508 * a device and because they are not accessible by CPU they are store
5509 * as special swap entry in the CPU page table.
5511 if (is_device_private_entry(ent)) {
5512 page = device_private_entry_to_page(ent);
5514 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5515 * a refcount of 1 when free (unlike normal page)
5517 if (!page_ref_add_unless(page, 1, 1))
5523 * Because lookup_swap_cache() updates some statistics counter,
5524 * we call find_get_page() with swapper_space directly.
5526 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5527 entry->val = ent.val;
5532 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5533 pte_t ptent, swp_entry_t *entry)
5539 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5540 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5542 struct page *page = NULL;
5543 struct address_space *mapping;
5546 if (!vma->vm_file) /* anonymous vma */
5548 if (!(mc.flags & MOVE_FILE))
5551 mapping = vma->vm_file->f_mapping;
5552 pgoff = linear_page_index(vma, addr);
5554 /* page is moved even if it's not RSS of this task(page-faulted). */
5556 /* shmem/tmpfs may report page out on swap: account for that too. */
5557 if (shmem_mapping(mapping)) {
5558 page = find_get_entry(mapping, pgoff);
5559 if (xa_is_value(page)) {
5560 swp_entry_t swp = radix_to_swp_entry(page);
5562 page = find_get_page(swap_address_space(swp),
5566 page = find_get_page(mapping, pgoff);
5568 page = find_get_page(mapping, pgoff);
5574 * mem_cgroup_move_account - move account of the page
5576 * @compound: charge the page as compound or small page
5577 * @from: mem_cgroup which the page is moved from.
5578 * @to: mem_cgroup which the page is moved to. @from != @to.
5580 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5582 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5585 static int mem_cgroup_move_account(struct page *page,
5587 struct mem_cgroup *from,
5588 struct mem_cgroup *to)
5590 struct lruvec *from_vec, *to_vec;
5591 struct pglist_data *pgdat;
5592 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5595 VM_BUG_ON(from == to);
5596 VM_BUG_ON_PAGE(PageLRU(page), page);
5597 VM_BUG_ON(compound && !PageTransHuge(page));
5600 * Prevent mem_cgroup_migrate() from looking at
5601 * page->mem_cgroup of its source page while we change it.
5604 if (!trylock_page(page))
5608 if (page->mem_cgroup != from)
5611 pgdat = page_pgdat(page);
5612 from_vec = mem_cgroup_lruvec(from, pgdat);
5613 to_vec = mem_cgroup_lruvec(to, pgdat);
5615 lock_page_memcg(page);
5617 if (PageAnon(page)) {
5618 if (page_mapped(page)) {
5619 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5620 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5621 if (PageTransHuge(page)) {
5622 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5624 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5630 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5631 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5633 if (PageSwapBacked(page)) {
5634 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5635 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5638 if (page_mapped(page)) {
5639 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5640 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5643 if (PageDirty(page)) {
5644 struct address_space *mapping = page_mapping(page);
5646 if (mapping_cap_account_dirty(mapping)) {
5647 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5649 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5655 if (PageWriteback(page)) {
5656 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5657 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5661 * All state has been migrated, let's switch to the new memcg.
5663 * It is safe to change page->mem_cgroup here because the page
5664 * is referenced, charged, isolated, and locked: we can't race
5665 * with (un)charging, migration, LRU putback, or anything else
5666 * that would rely on a stable page->mem_cgroup.
5668 * Note that lock_page_memcg is a memcg lock, not a page lock,
5669 * to save space. As soon as we switch page->mem_cgroup to a
5670 * new memcg that isn't locked, the above state can change
5671 * concurrently again. Make sure we're truly done with it.
5676 css_put(&from->css);
5678 page->mem_cgroup = to;
5680 __unlock_page_memcg(from);
5684 local_irq_disable();
5685 mem_cgroup_charge_statistics(to, page, nr_pages);
5686 memcg_check_events(to, page);
5687 mem_cgroup_charge_statistics(from, page, -nr_pages);
5688 memcg_check_events(from, page);
5697 * get_mctgt_type - get target type of moving charge
5698 * @vma: the vma the pte to be checked belongs
5699 * @addr: the address corresponding to the pte to be checked
5700 * @ptent: the pte to be checked
5701 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5704 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5705 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5706 * move charge. if @target is not NULL, the page is stored in target->page
5707 * with extra refcnt got(Callers should handle it).
5708 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5709 * target for charge migration. if @target is not NULL, the entry is stored
5711 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5712 * (so ZONE_DEVICE page and thus not on the lru).
5713 * For now we such page is charge like a regular page would be as for all
5714 * intent and purposes it is just special memory taking the place of a
5717 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5719 * Called with pte lock held.
5722 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5723 unsigned long addr, pte_t ptent, union mc_target *target)
5725 struct page *page = NULL;
5726 enum mc_target_type ret = MC_TARGET_NONE;
5727 swp_entry_t ent = { .val = 0 };
5729 if (pte_present(ptent))
5730 page = mc_handle_present_pte(vma, addr, ptent);
5731 else if (is_swap_pte(ptent))
5732 page = mc_handle_swap_pte(vma, ptent, &ent);
5733 else if (pte_none(ptent))
5734 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5736 if (!page && !ent.val)
5740 * Do only loose check w/o serialization.
5741 * mem_cgroup_move_account() checks the page is valid or
5742 * not under LRU exclusion.
5744 if (page->mem_cgroup == mc.from) {
5745 ret = MC_TARGET_PAGE;
5746 if (is_device_private_page(page))
5747 ret = MC_TARGET_DEVICE;
5749 target->page = page;
5751 if (!ret || !target)
5755 * There is a swap entry and a page doesn't exist or isn't charged.
5756 * But we cannot move a tail-page in a THP.
5758 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5759 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5760 ret = MC_TARGET_SWAP;
5767 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5769 * We don't consider PMD mapped swapping or file mapped pages because THP does
5770 * not support them for now.
5771 * Caller should make sure that pmd_trans_huge(pmd) is true.
5773 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5774 unsigned long addr, pmd_t pmd, union mc_target *target)
5776 struct page *page = NULL;
5777 enum mc_target_type ret = MC_TARGET_NONE;
5779 if (unlikely(is_swap_pmd(pmd))) {
5780 VM_BUG_ON(thp_migration_supported() &&
5781 !is_pmd_migration_entry(pmd));
5784 page = pmd_page(pmd);
5785 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5786 if (!(mc.flags & MOVE_ANON))
5788 if (page->mem_cgroup == mc.from) {
5789 ret = MC_TARGET_PAGE;
5792 target->page = page;
5798 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5799 unsigned long addr, pmd_t pmd, union mc_target *target)
5801 return MC_TARGET_NONE;
5805 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5806 unsigned long addr, unsigned long end,
5807 struct mm_walk *walk)
5809 struct vm_area_struct *vma = walk->vma;
5813 ptl = pmd_trans_huge_lock(pmd, vma);
5816 * Note their can not be MC_TARGET_DEVICE for now as we do not
5817 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5818 * this might change.
5820 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5821 mc.precharge += HPAGE_PMD_NR;
5826 if (pmd_trans_unstable(pmd))
5828 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5829 for (; addr != end; pte++, addr += PAGE_SIZE)
5830 if (get_mctgt_type(vma, addr, *pte, NULL))
5831 mc.precharge++; /* increment precharge temporarily */
5832 pte_unmap_unlock(pte - 1, ptl);
5838 static const struct mm_walk_ops precharge_walk_ops = {
5839 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5842 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5844 unsigned long precharge;
5847 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5848 mmap_read_unlock(mm);
5850 precharge = mc.precharge;
5856 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5858 unsigned long precharge = mem_cgroup_count_precharge(mm);
5860 VM_BUG_ON(mc.moving_task);
5861 mc.moving_task = current;
5862 return mem_cgroup_do_precharge(precharge);
5865 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5866 static void __mem_cgroup_clear_mc(void)
5868 struct mem_cgroup *from = mc.from;
5869 struct mem_cgroup *to = mc.to;
5871 /* we must uncharge all the leftover precharges from mc.to */
5873 cancel_charge(mc.to, mc.precharge);
5877 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5878 * we must uncharge here.
5880 if (mc.moved_charge) {
5881 cancel_charge(mc.from, mc.moved_charge);
5882 mc.moved_charge = 0;
5884 /* we must fixup refcnts and charges */
5885 if (mc.moved_swap) {
5886 /* uncharge swap account from the old cgroup */
5887 if (!mem_cgroup_is_root(mc.from))
5888 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5890 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5893 * we charged both to->memory and to->memsw, so we
5894 * should uncharge to->memory.
5896 if (!mem_cgroup_is_root(mc.to))
5897 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5901 memcg_oom_recover(from);
5902 memcg_oom_recover(to);
5903 wake_up_all(&mc.waitq);
5906 static void mem_cgroup_clear_mc(void)
5908 struct mm_struct *mm = mc.mm;
5911 * we must clear moving_task before waking up waiters at the end of
5914 mc.moving_task = NULL;
5915 __mem_cgroup_clear_mc();
5916 spin_lock(&mc.lock);
5920 spin_unlock(&mc.lock);
5925 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5927 struct cgroup_subsys_state *css;
5928 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5929 struct mem_cgroup *from;
5930 struct task_struct *leader, *p;
5931 struct mm_struct *mm;
5932 unsigned long move_flags;
5935 /* charge immigration isn't supported on the default hierarchy */
5936 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5940 * Multi-process migrations only happen on the default hierarchy
5941 * where charge immigration is not used. Perform charge
5942 * immigration if @tset contains a leader and whine if there are
5946 cgroup_taskset_for_each_leader(leader, css, tset) {
5949 memcg = mem_cgroup_from_css(css);
5955 * We are now commited to this value whatever it is. Changes in this
5956 * tunable will only affect upcoming migrations, not the current one.
5957 * So we need to save it, and keep it going.
5959 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5963 from = mem_cgroup_from_task(p);
5965 VM_BUG_ON(from == memcg);
5967 mm = get_task_mm(p);
5970 /* We move charges only when we move a owner of the mm */
5971 if (mm->owner == p) {
5974 VM_BUG_ON(mc.precharge);
5975 VM_BUG_ON(mc.moved_charge);
5976 VM_BUG_ON(mc.moved_swap);
5978 spin_lock(&mc.lock);
5982 mc.flags = move_flags;
5983 spin_unlock(&mc.lock);
5984 /* We set mc.moving_task later */
5986 ret = mem_cgroup_precharge_mc(mm);
5988 mem_cgroup_clear_mc();
5995 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5998 mem_cgroup_clear_mc();
6001 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6002 unsigned long addr, unsigned long end,
6003 struct mm_walk *walk)
6006 struct vm_area_struct *vma = walk->vma;
6009 enum mc_target_type target_type;
6010 union mc_target target;
6013 ptl = pmd_trans_huge_lock(pmd, vma);
6015 if (mc.precharge < HPAGE_PMD_NR) {
6019 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6020 if (target_type == MC_TARGET_PAGE) {
6022 if (!isolate_lru_page(page)) {
6023 if (!mem_cgroup_move_account(page, true,
6025 mc.precharge -= HPAGE_PMD_NR;
6026 mc.moved_charge += HPAGE_PMD_NR;
6028 putback_lru_page(page);
6031 } else if (target_type == MC_TARGET_DEVICE) {
6033 if (!mem_cgroup_move_account(page, true,
6035 mc.precharge -= HPAGE_PMD_NR;
6036 mc.moved_charge += HPAGE_PMD_NR;
6044 if (pmd_trans_unstable(pmd))
6047 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6048 for (; addr != end; addr += PAGE_SIZE) {
6049 pte_t ptent = *(pte++);
6050 bool device = false;
6056 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6057 case MC_TARGET_DEVICE:
6060 case MC_TARGET_PAGE:
6063 * We can have a part of the split pmd here. Moving it
6064 * can be done but it would be too convoluted so simply
6065 * ignore such a partial THP and keep it in original
6066 * memcg. There should be somebody mapping the head.
6068 if (PageTransCompound(page))
6070 if (!device && isolate_lru_page(page))
6072 if (!mem_cgroup_move_account(page, false,
6075 /* we uncharge from mc.from later. */
6079 putback_lru_page(page);
6080 put: /* get_mctgt_type() gets the page */
6083 case MC_TARGET_SWAP:
6085 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6087 mem_cgroup_id_get_many(mc.to, 1);
6088 /* we fixup other refcnts and charges later. */
6096 pte_unmap_unlock(pte - 1, ptl);
6101 * We have consumed all precharges we got in can_attach().
6102 * We try charge one by one, but don't do any additional
6103 * charges to mc.to if we have failed in charge once in attach()
6106 ret = mem_cgroup_do_precharge(1);
6114 static const struct mm_walk_ops charge_walk_ops = {
6115 .pmd_entry = mem_cgroup_move_charge_pte_range,
6118 static void mem_cgroup_move_charge(void)
6120 lru_add_drain_all();
6122 * Signal lock_page_memcg() to take the memcg's move_lock
6123 * while we're moving its pages to another memcg. Then wait
6124 * for already started RCU-only updates to finish.
6126 atomic_inc(&mc.from->moving_account);
6129 if (unlikely(!mmap_read_trylock(mc.mm))) {
6131 * Someone who are holding the mmap_lock might be waiting in
6132 * waitq. So we cancel all extra charges, wake up all waiters,
6133 * and retry. Because we cancel precharges, we might not be able
6134 * to move enough charges, but moving charge is a best-effort
6135 * feature anyway, so it wouldn't be a big problem.
6137 __mem_cgroup_clear_mc();
6142 * When we have consumed all precharges and failed in doing
6143 * additional charge, the page walk just aborts.
6145 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6148 mmap_read_unlock(mc.mm);
6149 atomic_dec(&mc.from->moving_account);
6152 static void mem_cgroup_move_task(void)
6155 mem_cgroup_move_charge();
6156 mem_cgroup_clear_mc();
6159 #else /* !CONFIG_MMU */
6160 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6164 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6167 static void mem_cgroup_move_task(void)
6173 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6174 * to verify whether we're attached to the default hierarchy on each mount
6177 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6180 * use_hierarchy is forced on the default hierarchy. cgroup core
6181 * guarantees that @root doesn't have any children, so turning it
6182 * on for the root memcg is enough.
6184 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6185 root_mem_cgroup->use_hierarchy = true;
6187 root_mem_cgroup->use_hierarchy = false;
6190 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6192 if (value == PAGE_COUNTER_MAX)
6193 seq_puts(m, "max\n");
6195 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6200 static u64 memory_current_read(struct cgroup_subsys_state *css,
6203 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6205 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6208 static int memory_min_show(struct seq_file *m, void *v)
6210 return seq_puts_memcg_tunable(m,
6211 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6214 static ssize_t memory_min_write(struct kernfs_open_file *of,
6215 char *buf, size_t nbytes, loff_t off)
6217 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6221 buf = strstrip(buf);
6222 err = page_counter_memparse(buf, "max", &min);
6226 page_counter_set_min(&memcg->memory, min);
6231 static int memory_low_show(struct seq_file *m, void *v)
6233 return seq_puts_memcg_tunable(m,
6234 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6237 static ssize_t memory_low_write(struct kernfs_open_file *of,
6238 char *buf, size_t nbytes, loff_t off)
6240 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6244 buf = strstrip(buf);
6245 err = page_counter_memparse(buf, "max", &low);
6249 page_counter_set_low(&memcg->memory, low);
6254 static int memory_high_show(struct seq_file *m, void *v)
6256 return seq_puts_memcg_tunable(m,
6257 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6260 static ssize_t memory_high_write(struct kernfs_open_file *of,
6261 char *buf, size_t nbytes, loff_t off)
6263 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6264 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6265 bool drained = false;
6269 buf = strstrip(buf);
6270 err = page_counter_memparse(buf, "max", &high);
6275 unsigned long nr_pages = page_counter_read(&memcg->memory);
6276 unsigned long reclaimed;
6278 if (nr_pages <= high)
6281 if (signal_pending(current))
6285 drain_all_stock(memcg);
6290 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6293 if (!reclaimed && !nr_retries--)
6297 page_counter_set_high(&memcg->memory, high);
6299 memcg_wb_domain_size_changed(memcg);
6304 static int memory_max_show(struct seq_file *m, void *v)
6306 return seq_puts_memcg_tunable(m,
6307 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6310 static ssize_t memory_max_write(struct kernfs_open_file *of,
6311 char *buf, size_t nbytes, loff_t off)
6313 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6314 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6315 bool drained = false;
6319 buf = strstrip(buf);
6320 err = page_counter_memparse(buf, "max", &max);
6324 xchg(&memcg->memory.max, max);
6327 unsigned long nr_pages = page_counter_read(&memcg->memory);
6329 if (nr_pages <= max)
6332 if (signal_pending(current))
6336 drain_all_stock(memcg);
6342 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6348 memcg_memory_event(memcg, MEMCG_OOM);
6349 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6353 memcg_wb_domain_size_changed(memcg);
6357 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6359 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6360 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6361 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6362 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6363 seq_printf(m, "oom_kill %lu\n",
6364 atomic_long_read(&events[MEMCG_OOM_KILL]));
6367 static int memory_events_show(struct seq_file *m, void *v)
6369 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6371 __memory_events_show(m, memcg->memory_events);
6375 static int memory_events_local_show(struct seq_file *m, void *v)
6377 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6379 __memory_events_show(m, memcg->memory_events_local);
6383 static int memory_stat_show(struct seq_file *m, void *v)
6385 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6388 buf = memory_stat_format(memcg);
6396 static int memory_oom_group_show(struct seq_file *m, void *v)
6398 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6400 seq_printf(m, "%d\n", memcg->oom_group);
6405 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6406 char *buf, size_t nbytes, loff_t off)
6408 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6411 buf = strstrip(buf);
6415 ret = kstrtoint(buf, 0, &oom_group);
6419 if (oom_group != 0 && oom_group != 1)
6422 memcg->oom_group = oom_group;
6427 static struct cftype memory_files[] = {
6430 .flags = CFTYPE_NOT_ON_ROOT,
6431 .read_u64 = memory_current_read,
6435 .flags = CFTYPE_NOT_ON_ROOT,
6436 .seq_show = memory_min_show,
6437 .write = memory_min_write,
6441 .flags = CFTYPE_NOT_ON_ROOT,
6442 .seq_show = memory_low_show,
6443 .write = memory_low_write,
6447 .flags = CFTYPE_NOT_ON_ROOT,
6448 .seq_show = memory_high_show,
6449 .write = memory_high_write,
6453 .flags = CFTYPE_NOT_ON_ROOT,
6454 .seq_show = memory_max_show,
6455 .write = memory_max_write,
6459 .flags = CFTYPE_NOT_ON_ROOT,
6460 .file_offset = offsetof(struct mem_cgroup, events_file),
6461 .seq_show = memory_events_show,
6464 .name = "events.local",
6465 .flags = CFTYPE_NOT_ON_ROOT,
6466 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6467 .seq_show = memory_events_local_show,
6471 .seq_show = memory_stat_show,
6474 .name = "oom.group",
6475 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6476 .seq_show = memory_oom_group_show,
6477 .write = memory_oom_group_write,
6482 struct cgroup_subsys memory_cgrp_subsys = {
6483 .css_alloc = mem_cgroup_css_alloc,
6484 .css_online = mem_cgroup_css_online,
6485 .css_offline = mem_cgroup_css_offline,
6486 .css_released = mem_cgroup_css_released,
6487 .css_free = mem_cgroup_css_free,
6488 .css_reset = mem_cgroup_css_reset,
6489 .can_attach = mem_cgroup_can_attach,
6490 .cancel_attach = mem_cgroup_cancel_attach,
6491 .post_attach = mem_cgroup_move_task,
6492 .bind = mem_cgroup_bind,
6493 .dfl_cftypes = memory_files,
6494 .legacy_cftypes = mem_cgroup_legacy_files,
6499 * This function calculates an individual cgroup's effective
6500 * protection which is derived from its own memory.min/low, its
6501 * parent's and siblings' settings, as well as the actual memory
6502 * distribution in the tree.
6504 * The following rules apply to the effective protection values:
6506 * 1. At the first level of reclaim, effective protection is equal to
6507 * the declared protection in memory.min and memory.low.
6509 * 2. To enable safe delegation of the protection configuration, at
6510 * subsequent levels the effective protection is capped to the
6511 * parent's effective protection.
6513 * 3. To make complex and dynamic subtrees easier to configure, the
6514 * user is allowed to overcommit the declared protection at a given
6515 * level. If that is the case, the parent's effective protection is
6516 * distributed to the children in proportion to how much protection
6517 * they have declared and how much of it they are utilizing.
6519 * This makes distribution proportional, but also work-conserving:
6520 * if one cgroup claims much more protection than it uses memory,
6521 * the unused remainder is available to its siblings.
6523 * 4. Conversely, when the declared protection is undercommitted at a
6524 * given level, the distribution of the larger parental protection
6525 * budget is NOT proportional. A cgroup's protection from a sibling
6526 * is capped to its own memory.min/low setting.
6528 * 5. However, to allow protecting recursive subtrees from each other
6529 * without having to declare each individual cgroup's fixed share
6530 * of the ancestor's claim to protection, any unutilized -
6531 * "floating" - protection from up the tree is distributed in
6532 * proportion to each cgroup's *usage*. This makes the protection
6533 * neutral wrt sibling cgroups and lets them compete freely over
6534 * the shared parental protection budget, but it protects the
6535 * subtree as a whole from neighboring subtrees.
6537 * Note that 4. and 5. are not in conflict: 4. is about protecting
6538 * against immediate siblings whereas 5. is about protecting against
6539 * neighboring subtrees.
6541 static unsigned long effective_protection(unsigned long usage,
6542 unsigned long parent_usage,
6543 unsigned long setting,
6544 unsigned long parent_effective,
6545 unsigned long siblings_protected)
6547 unsigned long protected;
6550 protected = min(usage, setting);
6552 * If all cgroups at this level combined claim and use more
6553 * protection then what the parent affords them, distribute
6554 * shares in proportion to utilization.
6556 * We are using actual utilization rather than the statically
6557 * claimed protection in order to be work-conserving: claimed
6558 * but unused protection is available to siblings that would
6559 * otherwise get a smaller chunk than what they claimed.
6561 if (siblings_protected > parent_effective)
6562 return protected * parent_effective / siblings_protected;
6565 * Ok, utilized protection of all children is within what the
6566 * parent affords them, so we know whatever this child claims
6567 * and utilizes is effectively protected.
6569 * If there is unprotected usage beyond this value, reclaim
6570 * will apply pressure in proportion to that amount.
6572 * If there is unutilized protection, the cgroup will be fully
6573 * shielded from reclaim, but we do return a smaller value for
6574 * protection than what the group could enjoy in theory. This
6575 * is okay. With the overcommit distribution above, effective
6576 * protection is always dependent on how memory is actually
6577 * consumed among the siblings anyway.
6582 * If the children aren't claiming (all of) the protection
6583 * afforded to them by the parent, distribute the remainder in
6584 * proportion to the (unprotected) memory of each cgroup. That
6585 * way, cgroups that aren't explicitly prioritized wrt each
6586 * other compete freely over the allowance, but they are
6587 * collectively protected from neighboring trees.
6589 * We're using unprotected memory for the weight so that if
6590 * some cgroups DO claim explicit protection, we don't protect
6591 * the same bytes twice.
6593 * Check both usage and parent_usage against the respective
6594 * protected values. One should imply the other, but they
6595 * aren't read atomically - make sure the division is sane.
6597 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6599 if (parent_effective > siblings_protected &&
6600 parent_usage > siblings_protected &&
6601 usage > protected) {
6602 unsigned long unclaimed;
6604 unclaimed = parent_effective - siblings_protected;
6605 unclaimed *= usage - protected;
6606 unclaimed /= parent_usage - siblings_protected;
6615 * mem_cgroup_protected - check if memory consumption is in the normal range
6616 * @root: the top ancestor of the sub-tree being checked
6617 * @memcg: the memory cgroup to check
6619 * WARNING: This function is not stateless! It can only be used as part
6620 * of a top-down tree iteration, not for isolated queries.
6622 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6623 struct mem_cgroup *memcg)
6625 unsigned long usage, parent_usage;
6626 struct mem_cgroup *parent;
6628 if (mem_cgroup_disabled())
6632 root = root_mem_cgroup;
6635 * Effective values of the reclaim targets are ignored so they
6636 * can be stale. Have a look at mem_cgroup_protection for more
6638 * TODO: calculation should be more robust so that we do not need
6639 * that special casing.
6644 usage = page_counter_read(&memcg->memory);
6648 parent = parent_mem_cgroup(memcg);
6649 /* No parent means a non-hierarchical mode on v1 memcg */
6653 if (parent == root) {
6654 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6655 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6659 parent_usage = page_counter_read(&parent->memory);
6661 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6662 READ_ONCE(memcg->memory.min),
6663 READ_ONCE(parent->memory.emin),
6664 atomic_long_read(&parent->memory.children_min_usage)));
6666 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6667 READ_ONCE(memcg->memory.low),
6668 READ_ONCE(parent->memory.elow),
6669 atomic_long_read(&parent->memory.children_low_usage)));
6673 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6674 * @page: page to charge
6675 * @mm: mm context of the victim
6676 * @gfp_mask: reclaim mode
6678 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6679 * pages according to @gfp_mask if necessary.
6681 * Returns 0 on success. Otherwise, an error code is returned.
6683 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6685 unsigned int nr_pages = hpage_nr_pages(page);
6686 struct mem_cgroup *memcg = NULL;
6689 if (mem_cgroup_disabled())
6692 if (PageSwapCache(page)) {
6693 swp_entry_t ent = { .val = page_private(page), };
6697 * Every swap fault against a single page tries to charge the
6698 * page, bail as early as possible. shmem_unuse() encounters
6699 * already charged pages, too. page->mem_cgroup is protected
6700 * by the page lock, which serializes swap cache removal, which
6701 * in turn serializes uncharging.
6703 VM_BUG_ON_PAGE(!PageLocked(page), page);
6704 if (compound_head(page)->mem_cgroup)
6707 id = lookup_swap_cgroup_id(ent);
6709 memcg = mem_cgroup_from_id(id);
6710 if (memcg && !css_tryget_online(&memcg->css))
6716 memcg = get_mem_cgroup_from_mm(mm);
6718 ret = try_charge(memcg, gfp_mask, nr_pages);
6722 css_get(&memcg->css);
6723 commit_charge(page, memcg);
6725 local_irq_disable();
6726 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6727 memcg_check_events(memcg, page);
6730 if (PageSwapCache(page)) {
6731 swp_entry_t entry = { .val = page_private(page) };
6733 * The swap entry might not get freed for a long time,
6734 * let's not wait for it. The page already received a
6735 * memory+swap charge, drop the swap entry duplicate.
6737 mem_cgroup_uncharge_swap(entry, nr_pages);
6741 css_put(&memcg->css);
6746 struct uncharge_gather {
6747 struct mem_cgroup *memcg;
6748 unsigned long nr_pages;
6749 unsigned long pgpgout;
6750 unsigned long nr_kmem;
6751 struct page *dummy_page;
6754 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6756 memset(ug, 0, sizeof(*ug));
6759 static void uncharge_batch(const struct uncharge_gather *ug)
6761 unsigned long flags;
6763 if (!mem_cgroup_is_root(ug->memcg)) {
6764 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6765 if (do_memsw_account())
6766 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6767 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6768 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6769 memcg_oom_recover(ug->memcg);
6772 local_irq_save(flags);
6773 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6774 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6775 memcg_check_events(ug->memcg, ug->dummy_page);
6776 local_irq_restore(flags);
6779 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6781 unsigned long nr_pages;
6783 VM_BUG_ON_PAGE(PageLRU(page), page);
6785 if (!page->mem_cgroup)
6789 * Nobody should be changing or seriously looking at
6790 * page->mem_cgroup at this point, we have fully
6791 * exclusive access to the page.
6794 if (ug->memcg != page->mem_cgroup) {
6797 uncharge_gather_clear(ug);
6799 ug->memcg = page->mem_cgroup;
6802 nr_pages = compound_nr(page);
6803 ug->nr_pages += nr_pages;
6805 if (!PageKmemcg(page)) {
6808 ug->nr_kmem += nr_pages;
6809 __ClearPageKmemcg(page);
6812 ug->dummy_page = page;
6813 page->mem_cgroup = NULL;
6814 css_put(&ug->memcg->css);
6817 static void uncharge_list(struct list_head *page_list)
6819 struct uncharge_gather ug;
6820 struct list_head *next;
6822 uncharge_gather_clear(&ug);
6825 * Note that the list can be a single page->lru; hence the
6826 * do-while loop instead of a simple list_for_each_entry().
6828 next = page_list->next;
6832 page = list_entry(next, struct page, lru);
6833 next = page->lru.next;
6835 uncharge_page(page, &ug);
6836 } while (next != page_list);
6839 uncharge_batch(&ug);
6843 * mem_cgroup_uncharge - uncharge a page
6844 * @page: page to uncharge
6846 * Uncharge a page previously charged with mem_cgroup_charge().
6848 void mem_cgroup_uncharge(struct page *page)
6850 struct uncharge_gather ug;
6852 if (mem_cgroup_disabled())
6855 /* Don't touch page->lru of any random page, pre-check: */
6856 if (!page->mem_cgroup)
6859 uncharge_gather_clear(&ug);
6860 uncharge_page(page, &ug);
6861 uncharge_batch(&ug);
6865 * mem_cgroup_uncharge_list - uncharge a list of page
6866 * @page_list: list of pages to uncharge
6868 * Uncharge a list of pages previously charged with
6869 * mem_cgroup_charge().
6871 void mem_cgroup_uncharge_list(struct list_head *page_list)
6873 if (mem_cgroup_disabled())
6876 if (!list_empty(page_list))
6877 uncharge_list(page_list);
6881 * mem_cgroup_migrate - charge a page's replacement
6882 * @oldpage: currently circulating page
6883 * @newpage: replacement page
6885 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6886 * be uncharged upon free.
6888 * Both pages must be locked, @newpage->mapping must be set up.
6890 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6892 struct mem_cgroup *memcg;
6893 unsigned int nr_pages;
6894 unsigned long flags;
6896 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6897 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6898 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6899 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6902 if (mem_cgroup_disabled())
6905 /* Page cache replacement: new page already charged? */
6906 if (newpage->mem_cgroup)
6909 /* Swapcache readahead pages can get replaced before being charged */
6910 memcg = oldpage->mem_cgroup;
6914 /* Force-charge the new page. The old one will be freed soon */
6915 nr_pages = hpage_nr_pages(newpage);
6917 page_counter_charge(&memcg->memory, nr_pages);
6918 if (do_memsw_account())
6919 page_counter_charge(&memcg->memsw, nr_pages);
6921 css_get(&memcg->css);
6922 commit_charge(newpage, memcg);
6924 local_irq_save(flags);
6925 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6926 memcg_check_events(memcg, newpage);
6927 local_irq_restore(flags);
6930 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6931 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6933 void mem_cgroup_sk_alloc(struct sock *sk)
6935 struct mem_cgroup *memcg;
6937 if (!mem_cgroup_sockets_enabled)
6940 /* Do not associate the sock with unrelated interrupted task's memcg. */
6945 memcg = mem_cgroup_from_task(current);
6946 if (memcg == root_mem_cgroup)
6948 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6950 if (css_tryget(&memcg->css))
6951 sk->sk_memcg = memcg;
6956 void mem_cgroup_sk_free(struct sock *sk)
6959 css_put(&sk->sk_memcg->css);
6963 * mem_cgroup_charge_skmem - charge socket memory
6964 * @memcg: memcg to charge
6965 * @nr_pages: number of pages to charge
6967 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6968 * @memcg's configured limit, %false if the charge had to be forced.
6970 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6972 gfp_t gfp_mask = GFP_KERNEL;
6974 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6975 struct page_counter *fail;
6977 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6978 memcg->tcpmem_pressure = 0;
6981 page_counter_charge(&memcg->tcpmem, nr_pages);
6982 memcg->tcpmem_pressure = 1;
6986 /* Don't block in the packet receive path */
6988 gfp_mask = GFP_NOWAIT;
6990 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6992 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6995 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7000 * mem_cgroup_uncharge_skmem - uncharge socket memory
7001 * @memcg: memcg to uncharge
7002 * @nr_pages: number of pages to uncharge
7004 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7006 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7007 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7011 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7013 refill_stock(memcg, nr_pages);
7016 static int __init cgroup_memory(char *s)
7020 while ((token = strsep(&s, ",")) != NULL) {
7023 if (!strcmp(token, "nosocket"))
7024 cgroup_memory_nosocket = true;
7025 if (!strcmp(token, "nokmem"))
7026 cgroup_memory_nokmem = true;
7030 __setup("cgroup.memory=", cgroup_memory);
7033 * subsys_initcall() for memory controller.
7035 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7036 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7037 * basically everything that doesn't depend on a specific mem_cgroup structure
7038 * should be initialized from here.
7040 static int __init mem_cgroup_init(void)
7044 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7045 memcg_hotplug_cpu_dead);
7047 for_each_possible_cpu(cpu)
7048 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7051 for_each_node(node) {
7052 struct mem_cgroup_tree_per_node *rtpn;
7054 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7055 node_online(node) ? node : NUMA_NO_NODE);
7057 rtpn->rb_root = RB_ROOT;
7058 rtpn->rb_rightmost = NULL;
7059 spin_lock_init(&rtpn->lock);
7060 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7065 subsys_initcall(mem_cgroup_init);
7067 #ifdef CONFIG_MEMCG_SWAP
7068 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7070 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7072 * The root cgroup cannot be destroyed, so it's refcount must
7075 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7079 memcg = parent_mem_cgroup(memcg);
7081 memcg = root_mem_cgroup;
7087 * mem_cgroup_swapout - transfer a memsw charge to swap
7088 * @page: page whose memsw charge to transfer
7089 * @entry: swap entry to move the charge to
7091 * Transfer the memsw charge of @page to @entry.
7093 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7095 struct mem_cgroup *memcg, *swap_memcg;
7096 unsigned int nr_entries;
7097 unsigned short oldid;
7099 VM_BUG_ON_PAGE(PageLRU(page), page);
7100 VM_BUG_ON_PAGE(page_count(page), page);
7102 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7105 memcg = page->mem_cgroup;
7107 /* Readahead page, never charged */
7112 * In case the memcg owning these pages has been offlined and doesn't
7113 * have an ID allocated to it anymore, charge the closest online
7114 * ancestor for the swap instead and transfer the memory+swap charge.
7116 swap_memcg = mem_cgroup_id_get_online(memcg);
7117 nr_entries = hpage_nr_pages(page);
7118 /* Get references for the tail pages, too */
7120 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7121 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7123 VM_BUG_ON_PAGE(oldid, page);
7124 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7126 page->mem_cgroup = NULL;
7128 if (!mem_cgroup_is_root(memcg))
7129 page_counter_uncharge(&memcg->memory, nr_entries);
7131 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7132 if (!mem_cgroup_is_root(swap_memcg))
7133 page_counter_charge(&swap_memcg->memsw, nr_entries);
7134 page_counter_uncharge(&memcg->memsw, nr_entries);
7138 * Interrupts should be disabled here because the caller holds the
7139 * i_pages lock which is taken with interrupts-off. It is
7140 * important here to have the interrupts disabled because it is the
7141 * only synchronisation we have for updating the per-CPU variables.
7143 VM_BUG_ON(!irqs_disabled());
7144 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7145 memcg_check_events(memcg, page);
7147 css_put(&memcg->css);
7151 * mem_cgroup_try_charge_swap - try charging swap space for a page
7152 * @page: page being added to swap
7153 * @entry: swap entry to charge
7155 * Try to charge @page's memcg for the swap space at @entry.
7157 * Returns 0 on success, -ENOMEM on failure.
7159 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7161 unsigned int nr_pages = hpage_nr_pages(page);
7162 struct page_counter *counter;
7163 struct mem_cgroup *memcg;
7164 unsigned short oldid;
7166 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7169 memcg = page->mem_cgroup;
7171 /* Readahead page, never charged */
7176 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7180 memcg = mem_cgroup_id_get_online(memcg);
7182 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7183 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7184 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7185 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7186 mem_cgroup_id_put(memcg);
7190 /* Get references for the tail pages, too */
7192 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7193 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7194 VM_BUG_ON_PAGE(oldid, page);
7195 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7201 * mem_cgroup_uncharge_swap - uncharge swap space
7202 * @entry: swap entry to uncharge
7203 * @nr_pages: the amount of swap space to uncharge
7205 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7207 struct mem_cgroup *memcg;
7210 id = swap_cgroup_record(entry, 0, nr_pages);
7212 memcg = mem_cgroup_from_id(id);
7214 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7215 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7216 page_counter_uncharge(&memcg->swap, nr_pages);
7218 page_counter_uncharge(&memcg->memsw, nr_pages);
7220 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7221 mem_cgroup_id_put_many(memcg, nr_pages);
7226 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7228 long nr_swap_pages = get_nr_swap_pages();
7230 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7231 return nr_swap_pages;
7232 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7233 nr_swap_pages = min_t(long, nr_swap_pages,
7234 READ_ONCE(memcg->swap.max) -
7235 page_counter_read(&memcg->swap));
7236 return nr_swap_pages;
7239 bool mem_cgroup_swap_full(struct page *page)
7241 struct mem_cgroup *memcg;
7243 VM_BUG_ON_PAGE(!PageLocked(page), page);
7247 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7250 memcg = page->mem_cgroup;
7254 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7255 unsigned long usage = page_counter_read(&memcg->swap);
7257 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7258 usage * 2 >= READ_ONCE(memcg->swap.max))
7265 static int __init setup_swap_account(char *s)
7267 if (!strcmp(s, "1"))
7268 cgroup_memory_noswap = 0;
7269 else if (!strcmp(s, "0"))
7270 cgroup_memory_noswap = 1;
7273 __setup("swapaccount=", setup_swap_account);
7275 static u64 swap_current_read(struct cgroup_subsys_state *css,
7278 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7280 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7283 static int swap_high_show(struct seq_file *m, void *v)
7285 return seq_puts_memcg_tunable(m,
7286 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7289 static ssize_t swap_high_write(struct kernfs_open_file *of,
7290 char *buf, size_t nbytes, loff_t off)
7292 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7296 buf = strstrip(buf);
7297 err = page_counter_memparse(buf, "max", &high);
7301 page_counter_set_high(&memcg->swap, high);
7306 static int swap_max_show(struct seq_file *m, void *v)
7308 return seq_puts_memcg_tunable(m,
7309 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7312 static ssize_t swap_max_write(struct kernfs_open_file *of,
7313 char *buf, size_t nbytes, loff_t off)
7315 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7319 buf = strstrip(buf);
7320 err = page_counter_memparse(buf, "max", &max);
7324 xchg(&memcg->swap.max, max);
7329 static int swap_events_show(struct seq_file *m, void *v)
7331 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7333 seq_printf(m, "high %lu\n",
7334 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7335 seq_printf(m, "max %lu\n",
7336 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7337 seq_printf(m, "fail %lu\n",
7338 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7343 static struct cftype swap_files[] = {
7345 .name = "swap.current",
7346 .flags = CFTYPE_NOT_ON_ROOT,
7347 .read_u64 = swap_current_read,
7350 .name = "swap.high",
7351 .flags = CFTYPE_NOT_ON_ROOT,
7352 .seq_show = swap_high_show,
7353 .write = swap_high_write,
7357 .flags = CFTYPE_NOT_ON_ROOT,
7358 .seq_show = swap_max_show,
7359 .write = swap_max_write,
7362 .name = "swap.events",
7363 .flags = CFTYPE_NOT_ON_ROOT,
7364 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7365 .seq_show = swap_events_show,
7370 static struct cftype memsw_files[] = {
7372 .name = "memsw.usage_in_bytes",
7373 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7374 .read_u64 = mem_cgroup_read_u64,
7377 .name = "memsw.max_usage_in_bytes",
7378 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7379 .write = mem_cgroup_reset,
7380 .read_u64 = mem_cgroup_read_u64,
7383 .name = "memsw.limit_in_bytes",
7384 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7385 .write = mem_cgroup_write,
7386 .read_u64 = mem_cgroup_read_u64,
7389 .name = "memsw.failcnt",
7390 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7391 .write = mem_cgroup_reset,
7392 .read_u64 = mem_cgroup_read_u64,
7394 { }, /* terminate */
7398 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7399 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7400 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7401 * boot parameter. This may result in premature OOPS inside
7402 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7404 static int __init mem_cgroup_swap_init(void)
7406 /* No memory control -> no swap control */
7407 if (mem_cgroup_disabled())
7408 cgroup_memory_noswap = true;
7410 if (cgroup_memory_noswap)
7413 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7414 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7418 core_initcall(mem_cgroup_swap_init);
7420 #endif /* CONFIG_MEMCG_SWAP */