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 (vmstat_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, "sock %llu\n",
1492 (u64)memcg_page_state(memcg, MEMCG_SOCK) *
1495 seq_buf_printf(&s, "shmem %llu\n",
1496 (u64)memcg_page_state(memcg, NR_SHMEM) *
1498 seq_buf_printf(&s, "file_mapped %llu\n",
1499 (u64)memcg_page_state(memcg, NR_FILE_MAPPED) *
1501 seq_buf_printf(&s, "file_dirty %llu\n",
1502 (u64)memcg_page_state(memcg, NR_FILE_DIRTY) *
1504 seq_buf_printf(&s, "file_writeback %llu\n",
1505 (u64)memcg_page_state(memcg, NR_WRITEBACK) *
1508 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1509 seq_buf_printf(&s, "anon_thp %llu\n",
1510 (u64)memcg_page_state(memcg, NR_ANON_THPS) *
1514 for (i = 0; i < NR_LRU_LISTS; i++)
1515 seq_buf_printf(&s, "%s %llu\n", lru_list_name(i),
1516 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
1519 seq_buf_printf(&s, "slab_reclaimable %llu\n",
1520 (u64)memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B));
1521 seq_buf_printf(&s, "slab_unreclaimable %llu\n",
1522 (u64)memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B));
1524 /* Accumulated memory events */
1526 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1527 memcg_events(memcg, PGFAULT));
1528 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1529 memcg_events(memcg, PGMAJFAULT));
1531 seq_buf_printf(&s, "workingset_refault %lu\n",
1532 memcg_page_state(memcg, WORKINGSET_REFAULT));
1533 seq_buf_printf(&s, "workingset_activate %lu\n",
1534 memcg_page_state(memcg, WORKINGSET_ACTIVATE));
1535 seq_buf_printf(&s, "workingset_restore %lu\n",
1536 memcg_page_state(memcg, WORKINGSET_RESTORE));
1537 seq_buf_printf(&s, "workingset_nodereclaim %lu\n",
1538 memcg_page_state(memcg, WORKINGSET_NODERECLAIM));
1540 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1541 memcg_events(memcg, PGREFILL));
1542 seq_buf_printf(&s, "pgscan %lu\n",
1543 memcg_events(memcg, PGSCAN_KSWAPD) +
1544 memcg_events(memcg, PGSCAN_DIRECT));
1545 seq_buf_printf(&s, "pgsteal %lu\n",
1546 memcg_events(memcg, PGSTEAL_KSWAPD) +
1547 memcg_events(memcg, PGSTEAL_DIRECT));
1548 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1549 memcg_events(memcg, PGACTIVATE));
1550 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1551 memcg_events(memcg, PGDEACTIVATE));
1552 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1553 memcg_events(memcg, PGLAZYFREE));
1554 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1555 memcg_events(memcg, PGLAZYFREED));
1557 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1558 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1559 memcg_events(memcg, THP_FAULT_ALLOC));
1560 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1561 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1562 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1564 /* The above should easily fit into one page */
1565 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1570 #define K(x) ((x) << (PAGE_SHIFT-10))
1572 * mem_cgroup_print_oom_context: Print OOM information relevant to
1573 * memory controller.
1574 * @memcg: The memory cgroup that went over limit
1575 * @p: Task that is going to be killed
1577 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1580 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1585 pr_cont(",oom_memcg=");
1586 pr_cont_cgroup_path(memcg->css.cgroup);
1588 pr_cont(",global_oom");
1590 pr_cont(",task_memcg=");
1591 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1597 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1598 * memory controller.
1599 * @memcg: The memory cgroup that went over limit
1601 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1605 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1606 K((u64)page_counter_read(&memcg->memory)),
1607 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1608 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1609 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1610 K((u64)page_counter_read(&memcg->swap)),
1611 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1613 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1614 K((u64)page_counter_read(&memcg->memsw)),
1615 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1616 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1617 K((u64)page_counter_read(&memcg->kmem)),
1618 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1621 pr_info("Memory cgroup stats for ");
1622 pr_cont_cgroup_path(memcg->css.cgroup);
1624 buf = memory_stat_format(memcg);
1632 * Return the memory (and swap, if configured) limit for a memcg.
1634 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1638 max = READ_ONCE(memcg->memory.max);
1639 if (mem_cgroup_swappiness(memcg)) {
1640 unsigned long memsw_max;
1641 unsigned long swap_max;
1643 memsw_max = memcg->memsw.max;
1644 swap_max = READ_ONCE(memcg->swap.max);
1645 swap_max = min(swap_max, (unsigned long)total_swap_pages);
1646 max = min(max + swap_max, memsw_max);
1651 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1653 return page_counter_read(&memcg->memory);
1656 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1659 struct oom_control oc = {
1663 .gfp_mask = gfp_mask,
1668 if (mutex_lock_killable(&oom_lock))
1671 if (mem_cgroup_margin(memcg) >= (1 << order))
1675 * A few threads which were not waiting at mutex_lock_killable() can
1676 * fail to bail out. Therefore, check again after holding oom_lock.
1678 ret = should_force_charge() || out_of_memory(&oc);
1681 mutex_unlock(&oom_lock);
1685 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1688 unsigned long *total_scanned)
1690 struct mem_cgroup *victim = NULL;
1693 unsigned long excess;
1694 unsigned long nr_scanned;
1695 struct mem_cgroup_reclaim_cookie reclaim = {
1699 excess = soft_limit_excess(root_memcg);
1702 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1707 * If we have not been able to reclaim
1708 * anything, it might because there are
1709 * no reclaimable pages under this hierarchy
1714 * We want to do more targeted reclaim.
1715 * excess >> 2 is not to excessive so as to
1716 * reclaim too much, nor too less that we keep
1717 * coming back to reclaim from this cgroup
1719 if (total >= (excess >> 2) ||
1720 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1725 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1726 pgdat, &nr_scanned);
1727 *total_scanned += nr_scanned;
1728 if (!soft_limit_excess(root_memcg))
1731 mem_cgroup_iter_break(root_memcg, victim);
1735 #ifdef CONFIG_LOCKDEP
1736 static struct lockdep_map memcg_oom_lock_dep_map = {
1737 .name = "memcg_oom_lock",
1741 static DEFINE_SPINLOCK(memcg_oom_lock);
1744 * Check OOM-Killer is already running under our hierarchy.
1745 * If someone is running, return false.
1747 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1749 struct mem_cgroup *iter, *failed = NULL;
1751 spin_lock(&memcg_oom_lock);
1753 for_each_mem_cgroup_tree(iter, memcg) {
1754 if (iter->oom_lock) {
1756 * this subtree of our hierarchy is already locked
1757 * so we cannot give a lock.
1760 mem_cgroup_iter_break(memcg, iter);
1763 iter->oom_lock = true;
1768 * OK, we failed to lock the whole subtree so we have
1769 * to clean up what we set up to the failing subtree
1771 for_each_mem_cgroup_tree(iter, memcg) {
1772 if (iter == failed) {
1773 mem_cgroup_iter_break(memcg, iter);
1776 iter->oom_lock = false;
1779 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1781 spin_unlock(&memcg_oom_lock);
1786 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1788 struct mem_cgroup *iter;
1790 spin_lock(&memcg_oom_lock);
1791 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1792 for_each_mem_cgroup_tree(iter, memcg)
1793 iter->oom_lock = false;
1794 spin_unlock(&memcg_oom_lock);
1797 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1799 struct mem_cgroup *iter;
1801 spin_lock(&memcg_oom_lock);
1802 for_each_mem_cgroup_tree(iter, memcg)
1804 spin_unlock(&memcg_oom_lock);
1807 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1809 struct mem_cgroup *iter;
1812 * When a new child is created while the hierarchy is under oom,
1813 * mem_cgroup_oom_lock() may not be called. Watch for underflow.
1815 spin_lock(&memcg_oom_lock);
1816 for_each_mem_cgroup_tree(iter, memcg)
1817 if (iter->under_oom > 0)
1819 spin_unlock(&memcg_oom_lock);
1822 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1824 struct oom_wait_info {
1825 struct mem_cgroup *memcg;
1826 wait_queue_entry_t wait;
1829 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1830 unsigned mode, int sync, void *arg)
1832 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1833 struct mem_cgroup *oom_wait_memcg;
1834 struct oom_wait_info *oom_wait_info;
1836 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1837 oom_wait_memcg = oom_wait_info->memcg;
1839 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1840 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1842 return autoremove_wake_function(wait, mode, sync, arg);
1845 static void memcg_oom_recover(struct mem_cgroup *memcg)
1848 * For the following lockless ->under_oom test, the only required
1849 * guarantee is that it must see the state asserted by an OOM when
1850 * this function is called as a result of userland actions
1851 * triggered by the notification of the OOM. This is trivially
1852 * achieved by invoking mem_cgroup_mark_under_oom() before
1853 * triggering notification.
1855 if (memcg && memcg->under_oom)
1856 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1866 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1868 enum oom_status ret;
1871 if (order > PAGE_ALLOC_COSTLY_ORDER)
1874 memcg_memory_event(memcg, MEMCG_OOM);
1877 * We are in the middle of the charge context here, so we
1878 * don't want to block when potentially sitting on a callstack
1879 * that holds all kinds of filesystem and mm locks.
1881 * cgroup1 allows disabling the OOM killer and waiting for outside
1882 * handling until the charge can succeed; remember the context and put
1883 * the task to sleep at the end of the page fault when all locks are
1886 * On the other hand, in-kernel OOM killer allows for an async victim
1887 * memory reclaim (oom_reaper) and that means that we are not solely
1888 * relying on the oom victim to make a forward progress and we can
1889 * invoke the oom killer here.
1891 * Please note that mem_cgroup_out_of_memory might fail to find a
1892 * victim and then we have to bail out from the charge path.
1894 if (memcg->oom_kill_disable) {
1895 if (!current->in_user_fault)
1897 css_get(&memcg->css);
1898 current->memcg_in_oom = memcg;
1899 current->memcg_oom_gfp_mask = mask;
1900 current->memcg_oom_order = order;
1905 mem_cgroup_mark_under_oom(memcg);
1907 locked = mem_cgroup_oom_trylock(memcg);
1910 mem_cgroup_oom_notify(memcg);
1912 mem_cgroup_unmark_under_oom(memcg);
1913 if (mem_cgroup_out_of_memory(memcg, mask, order))
1919 mem_cgroup_oom_unlock(memcg);
1925 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1926 * @handle: actually kill/wait or just clean up the OOM state
1928 * This has to be called at the end of a page fault if the memcg OOM
1929 * handler was enabled.
1931 * Memcg supports userspace OOM handling where failed allocations must
1932 * sleep on a waitqueue until the userspace task resolves the
1933 * situation. Sleeping directly in the charge context with all kinds
1934 * of locks held is not a good idea, instead we remember an OOM state
1935 * in the task and mem_cgroup_oom_synchronize() has to be called at
1936 * the end of the page fault to complete the OOM handling.
1938 * Returns %true if an ongoing memcg OOM situation was detected and
1939 * completed, %false otherwise.
1941 bool mem_cgroup_oom_synchronize(bool handle)
1943 struct mem_cgroup *memcg = current->memcg_in_oom;
1944 struct oom_wait_info owait;
1947 /* OOM is global, do not handle */
1954 owait.memcg = memcg;
1955 owait.wait.flags = 0;
1956 owait.wait.func = memcg_oom_wake_function;
1957 owait.wait.private = current;
1958 INIT_LIST_HEAD(&owait.wait.entry);
1960 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1961 mem_cgroup_mark_under_oom(memcg);
1963 locked = mem_cgroup_oom_trylock(memcg);
1966 mem_cgroup_oom_notify(memcg);
1968 if (locked && !memcg->oom_kill_disable) {
1969 mem_cgroup_unmark_under_oom(memcg);
1970 finish_wait(&memcg_oom_waitq, &owait.wait);
1971 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
1972 current->memcg_oom_order);
1975 mem_cgroup_unmark_under_oom(memcg);
1976 finish_wait(&memcg_oom_waitq, &owait.wait);
1980 mem_cgroup_oom_unlock(memcg);
1982 * There is no guarantee that an OOM-lock contender
1983 * sees the wakeups triggered by the OOM kill
1984 * uncharges. Wake any sleepers explicitely.
1986 memcg_oom_recover(memcg);
1989 current->memcg_in_oom = NULL;
1990 css_put(&memcg->css);
1995 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
1996 * @victim: task to be killed by the OOM killer
1997 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
1999 * Returns a pointer to a memory cgroup, which has to be cleaned up
2000 * by killing all belonging OOM-killable tasks.
2002 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2004 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2005 struct mem_cgroup *oom_domain)
2007 struct mem_cgroup *oom_group = NULL;
2008 struct mem_cgroup *memcg;
2010 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2014 oom_domain = root_mem_cgroup;
2018 memcg = mem_cgroup_from_task(victim);
2019 if (memcg == root_mem_cgroup)
2023 * If the victim task has been asynchronously moved to a different
2024 * memory cgroup, we might end up killing tasks outside oom_domain.
2025 * In this case it's better to ignore memory.group.oom.
2027 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2031 * Traverse the memory cgroup hierarchy from the victim task's
2032 * cgroup up to the OOMing cgroup (or root) to find the
2033 * highest-level memory cgroup with oom.group set.
2035 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2036 if (memcg->oom_group)
2039 if (memcg == oom_domain)
2044 css_get(&oom_group->css);
2051 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2053 pr_info("Tasks in ");
2054 pr_cont_cgroup_path(memcg->css.cgroup);
2055 pr_cont(" are going to be killed due to memory.oom.group set\n");
2059 * lock_page_memcg - lock a page->mem_cgroup binding
2062 * This function protects unlocked LRU pages from being moved to
2065 * It ensures lifetime of the returned memcg. Caller is responsible
2066 * for the lifetime of the page; __unlock_page_memcg() is available
2067 * when @page might get freed inside the locked section.
2069 struct mem_cgroup *lock_page_memcg(struct page *page)
2071 struct page *head = compound_head(page); /* rmap on tail pages */
2072 struct mem_cgroup *memcg;
2073 unsigned long flags;
2076 * The RCU lock is held throughout the transaction. The fast
2077 * path can get away without acquiring the memcg->move_lock
2078 * because page moving starts with an RCU grace period.
2080 * The RCU lock also protects the memcg from being freed when
2081 * the page state that is going to change is the only thing
2082 * preventing the page itself from being freed. E.g. writeback
2083 * doesn't hold a page reference and relies on PG_writeback to
2084 * keep off truncation, migration and so forth.
2088 if (mem_cgroup_disabled())
2091 memcg = head->mem_cgroup;
2092 if (unlikely(!memcg))
2095 if (atomic_read(&memcg->moving_account) <= 0)
2098 spin_lock_irqsave(&memcg->move_lock, flags);
2099 if (memcg != head->mem_cgroup) {
2100 spin_unlock_irqrestore(&memcg->move_lock, flags);
2105 * When charge migration first begins, we can have locked and
2106 * unlocked page stat updates happening concurrently. Track
2107 * the task who has the lock for unlock_page_memcg().
2109 memcg->move_lock_task = current;
2110 memcg->move_lock_flags = flags;
2114 EXPORT_SYMBOL(lock_page_memcg);
2117 * __unlock_page_memcg - unlock and unpin a memcg
2120 * Unlock and unpin a memcg returned by lock_page_memcg().
2122 void __unlock_page_memcg(struct mem_cgroup *memcg)
2124 if (memcg && memcg->move_lock_task == current) {
2125 unsigned long flags = memcg->move_lock_flags;
2127 memcg->move_lock_task = NULL;
2128 memcg->move_lock_flags = 0;
2130 spin_unlock_irqrestore(&memcg->move_lock, flags);
2137 * unlock_page_memcg - unlock a page->mem_cgroup binding
2140 void unlock_page_memcg(struct page *page)
2142 struct page *head = compound_head(page);
2144 __unlock_page_memcg(head->mem_cgroup);
2146 EXPORT_SYMBOL(unlock_page_memcg);
2148 struct memcg_stock_pcp {
2149 struct mem_cgroup *cached; /* this never be root cgroup */
2150 unsigned int nr_pages;
2152 #ifdef CONFIG_MEMCG_KMEM
2153 struct obj_cgroup *cached_objcg;
2154 unsigned int nr_bytes;
2157 struct work_struct work;
2158 unsigned long flags;
2159 #define FLUSHING_CACHED_CHARGE 0
2161 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2162 static DEFINE_MUTEX(percpu_charge_mutex);
2164 #ifdef CONFIG_MEMCG_KMEM
2165 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2166 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2167 struct mem_cgroup *root_memcg);
2170 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2173 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2174 struct mem_cgroup *root_memcg)
2181 * consume_stock: Try to consume stocked charge on this cpu.
2182 * @memcg: memcg to consume from.
2183 * @nr_pages: how many pages to charge.
2185 * The charges will only happen if @memcg matches the current cpu's memcg
2186 * stock, and at least @nr_pages are available in that stock. Failure to
2187 * service an allocation will refill the stock.
2189 * returns true if successful, false otherwise.
2191 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2193 struct memcg_stock_pcp *stock;
2194 unsigned long flags;
2197 if (nr_pages > MEMCG_CHARGE_BATCH)
2200 local_irq_save(flags);
2202 stock = this_cpu_ptr(&memcg_stock);
2203 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2204 stock->nr_pages -= nr_pages;
2208 local_irq_restore(flags);
2214 * Returns stocks cached in percpu and reset cached information.
2216 static void drain_stock(struct memcg_stock_pcp *stock)
2218 struct mem_cgroup *old = stock->cached;
2223 if (stock->nr_pages) {
2224 page_counter_uncharge(&old->memory, stock->nr_pages);
2225 if (do_memsw_account())
2226 page_counter_uncharge(&old->memsw, stock->nr_pages);
2227 stock->nr_pages = 0;
2231 stock->cached = NULL;
2234 static void drain_local_stock(struct work_struct *dummy)
2236 struct memcg_stock_pcp *stock;
2237 unsigned long flags;
2240 * The only protection from memory hotplug vs. drain_stock races is
2241 * that we always operate on local CPU stock here with IRQ disabled
2243 local_irq_save(flags);
2245 stock = this_cpu_ptr(&memcg_stock);
2246 drain_obj_stock(stock);
2248 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2250 local_irq_restore(flags);
2254 * Cache charges(val) to local per_cpu area.
2255 * This will be consumed by consume_stock() function, later.
2257 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2259 struct memcg_stock_pcp *stock;
2260 unsigned long flags;
2262 local_irq_save(flags);
2264 stock = this_cpu_ptr(&memcg_stock);
2265 if (stock->cached != memcg) { /* reset if necessary */
2267 css_get(&memcg->css);
2268 stock->cached = memcg;
2270 stock->nr_pages += nr_pages;
2272 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2275 local_irq_restore(flags);
2279 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2280 * of the hierarchy under it.
2282 static void drain_all_stock(struct mem_cgroup *root_memcg)
2286 /* If someone's already draining, avoid adding running more workers. */
2287 if (!mutex_trylock(&percpu_charge_mutex))
2290 * Notify other cpus that system-wide "drain" is running
2291 * We do not care about races with the cpu hotplug because cpu down
2292 * as well as workers from this path always operate on the local
2293 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2296 for_each_online_cpu(cpu) {
2297 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2298 struct mem_cgroup *memcg;
2302 memcg = stock->cached;
2303 if (memcg && stock->nr_pages &&
2304 mem_cgroup_is_descendant(memcg, root_memcg))
2306 if (obj_stock_flush_required(stock, root_memcg))
2311 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2313 drain_local_stock(&stock->work);
2315 schedule_work_on(cpu, &stock->work);
2319 mutex_unlock(&percpu_charge_mutex);
2322 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2324 struct memcg_stock_pcp *stock;
2325 struct mem_cgroup *memcg, *mi;
2327 stock = &per_cpu(memcg_stock, cpu);
2330 for_each_mem_cgroup(memcg) {
2333 for (i = 0; i < MEMCG_NR_STAT; i++) {
2337 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2339 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2340 atomic_long_add(x, &memcg->vmstats[i]);
2342 if (i >= NR_VM_NODE_STAT_ITEMS)
2345 for_each_node(nid) {
2346 struct mem_cgroup_per_node *pn;
2348 pn = mem_cgroup_nodeinfo(memcg, nid);
2349 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2352 atomic_long_add(x, &pn->lruvec_stat[i]);
2353 } while ((pn = parent_nodeinfo(pn, nid)));
2357 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2360 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2362 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2363 atomic_long_add(x, &memcg->vmevents[i]);
2370 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2371 unsigned int nr_pages,
2374 unsigned long nr_reclaimed = 0;
2377 unsigned long pflags;
2379 if (page_counter_read(&memcg->memory) <=
2380 READ_ONCE(memcg->memory.high))
2383 memcg_memory_event(memcg, MEMCG_HIGH);
2385 psi_memstall_enter(&pflags);
2386 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2388 psi_memstall_leave(&pflags);
2389 } while ((memcg = parent_mem_cgroup(memcg)) &&
2390 !mem_cgroup_is_root(memcg));
2392 return nr_reclaimed;
2395 static void high_work_func(struct work_struct *work)
2397 struct mem_cgroup *memcg;
2399 memcg = container_of(work, struct mem_cgroup, high_work);
2400 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2404 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2405 * enough to still cause a significant slowdown in most cases, while still
2406 * allowing diagnostics and tracing to proceed without becoming stuck.
2408 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2411 * When calculating the delay, we use these either side of the exponentiation to
2412 * maintain precision and scale to a reasonable number of jiffies (see the table
2415 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2416 * overage ratio to a delay.
2417 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down down the
2418 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2419 * to produce a reasonable delay curve.
2421 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2422 * reasonable delay curve compared to precision-adjusted overage, not
2423 * penalising heavily at first, but still making sure that growth beyond the
2424 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2425 * example, with a high of 100 megabytes:
2427 * +-------+------------------------+
2428 * | usage | time to allocate in ms |
2429 * +-------+------------------------+
2451 * +-------+------------------------+
2453 #define MEMCG_DELAY_PRECISION_SHIFT 20
2454 #define MEMCG_DELAY_SCALING_SHIFT 14
2456 static u64 calculate_overage(unsigned long usage, unsigned long high)
2464 * Prevent division by 0 in overage calculation by acting as if
2465 * it was a threshold of 1 page
2467 high = max(high, 1UL);
2469 overage = usage - high;
2470 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2471 return div64_u64(overage, high);
2474 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2476 u64 overage, max_overage = 0;
2479 overage = calculate_overage(page_counter_read(&memcg->memory),
2480 READ_ONCE(memcg->memory.high));
2481 max_overage = max(overage, max_overage);
2482 } while ((memcg = parent_mem_cgroup(memcg)) &&
2483 !mem_cgroup_is_root(memcg));
2488 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2490 u64 overage, max_overage = 0;
2493 overage = calculate_overage(page_counter_read(&memcg->swap),
2494 READ_ONCE(memcg->swap.high));
2496 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2497 max_overage = max(overage, max_overage);
2498 } while ((memcg = parent_mem_cgroup(memcg)) &&
2499 !mem_cgroup_is_root(memcg));
2505 * Get the number of jiffies that we should penalise a mischievous cgroup which
2506 * is exceeding its memory.high by checking both it and its ancestors.
2508 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2509 unsigned int nr_pages,
2512 unsigned long penalty_jiffies;
2518 * We use overage compared to memory.high to calculate the number of
2519 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2520 * fairly lenient on small overages, and increasingly harsh when the
2521 * memcg in question makes it clear that it has no intention of stopping
2522 * its crazy behaviour, so we exponentially increase the delay based on
2525 penalty_jiffies = max_overage * max_overage * HZ;
2526 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2527 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2530 * Factor in the task's own contribution to the overage, such that four
2531 * N-sized allocations are throttled approximately the same as one
2532 * 4N-sized allocation.
2534 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2535 * larger the current charge patch is than that.
2537 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2541 * Scheduled by try_charge() to be executed from the userland return path
2542 * and reclaims memory over the high limit.
2544 void mem_cgroup_handle_over_high(void)
2546 unsigned long penalty_jiffies;
2547 unsigned long pflags;
2548 unsigned long nr_reclaimed;
2549 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2550 int nr_retries = MAX_RECLAIM_RETRIES;
2551 struct mem_cgroup *memcg;
2552 bool in_retry = false;
2554 if (likely(!nr_pages))
2557 memcg = get_mem_cgroup_from_mm(current->mm);
2558 current->memcg_nr_pages_over_high = 0;
2562 * The allocating task should reclaim at least the batch size, but for
2563 * subsequent retries we only want to do what's necessary to prevent oom
2564 * or breaching resource isolation.
2566 * This is distinct from memory.max or page allocator behaviour because
2567 * memory.high is currently batched, whereas memory.max and the page
2568 * allocator run every time an allocation is made.
2570 nr_reclaimed = reclaim_high(memcg,
2571 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2575 * memory.high is breached and reclaim is unable to keep up. Throttle
2576 * allocators proactively to slow down excessive growth.
2578 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2579 mem_find_max_overage(memcg));
2581 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2582 swap_find_max_overage(memcg));
2585 * Clamp the max delay per usermode return so as to still keep the
2586 * application moving forwards and also permit diagnostics, albeit
2589 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2592 * Don't sleep if the amount of jiffies this memcg owes us is so low
2593 * that it's not even worth doing, in an attempt to be nice to those who
2594 * go only a small amount over their memory.high value and maybe haven't
2595 * been aggressively reclaimed enough yet.
2597 if (penalty_jiffies <= HZ / 100)
2601 * If reclaim is making forward progress but we're still over
2602 * memory.high, we want to encourage that rather than doing allocator
2605 if (nr_reclaimed || nr_retries--) {
2611 * If we exit early, we're guaranteed to die (since
2612 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2613 * need to account for any ill-begotten jiffies to pay them off later.
2615 psi_memstall_enter(&pflags);
2616 schedule_timeout_killable(penalty_jiffies);
2617 psi_memstall_leave(&pflags);
2620 css_put(&memcg->css);
2623 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2624 unsigned int nr_pages)
2626 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2627 int nr_retries = MAX_RECLAIM_RETRIES;
2628 struct mem_cgroup *mem_over_limit;
2629 struct page_counter *counter;
2630 enum oom_status oom_status;
2631 unsigned long nr_reclaimed;
2632 bool may_swap = true;
2633 bool drained = false;
2634 unsigned long pflags;
2636 if (mem_cgroup_is_root(memcg))
2639 if (consume_stock(memcg, nr_pages))
2642 if (!do_memsw_account() ||
2643 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2644 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2646 if (do_memsw_account())
2647 page_counter_uncharge(&memcg->memsw, batch);
2648 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2650 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2654 if (batch > nr_pages) {
2660 * Memcg doesn't have a dedicated reserve for atomic
2661 * allocations. But like the global atomic pool, we need to
2662 * put the burden of reclaim on regular allocation requests
2663 * and let these go through as privileged allocations.
2665 if (gfp_mask & __GFP_ATOMIC)
2669 * Unlike in global OOM situations, memcg is not in a physical
2670 * memory shortage. Allow dying and OOM-killed tasks to
2671 * bypass the last charges so that they can exit quickly and
2672 * free their memory.
2674 if (unlikely(should_force_charge()))
2678 * Prevent unbounded recursion when reclaim operations need to
2679 * allocate memory. This might exceed the limits temporarily,
2680 * but we prefer facilitating memory reclaim and getting back
2681 * under the limit over triggering OOM kills in these cases.
2683 if (unlikely(current->flags & PF_MEMALLOC))
2686 if (unlikely(task_in_memcg_oom(current)))
2689 if (!gfpflags_allow_blocking(gfp_mask))
2692 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2694 psi_memstall_enter(&pflags);
2695 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2696 gfp_mask, may_swap);
2697 psi_memstall_leave(&pflags);
2699 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2703 drain_all_stock(mem_over_limit);
2708 if (gfp_mask & __GFP_NORETRY)
2711 * Even though the limit is exceeded at this point, reclaim
2712 * may have been able to free some pages. Retry the charge
2713 * before killing the task.
2715 * Only for regular pages, though: huge pages are rather
2716 * unlikely to succeed so close to the limit, and we fall back
2717 * to regular pages anyway in case of failure.
2719 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2722 * At task move, charge accounts can be doubly counted. So, it's
2723 * better to wait until the end of task_move if something is going on.
2725 if (mem_cgroup_wait_acct_move(mem_over_limit))
2731 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2734 if (gfp_mask & __GFP_NOFAIL)
2737 if (fatal_signal_pending(current))
2741 * keep retrying as long as the memcg oom killer is able to make
2742 * a forward progress or bypass the charge if the oom killer
2743 * couldn't make any progress.
2745 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2746 get_order(nr_pages * PAGE_SIZE));
2747 switch (oom_status) {
2749 nr_retries = MAX_RECLAIM_RETRIES;
2757 if (!(gfp_mask & __GFP_NOFAIL))
2761 * The allocation either can't fail or will lead to more memory
2762 * being freed very soon. Allow memory usage go over the limit
2763 * temporarily by force charging it.
2765 page_counter_charge(&memcg->memory, nr_pages);
2766 if (do_memsw_account())
2767 page_counter_charge(&memcg->memsw, nr_pages);
2772 if (batch > nr_pages)
2773 refill_stock(memcg, batch - nr_pages);
2776 * If the hierarchy is above the normal consumption range, schedule
2777 * reclaim on returning to userland. We can perform reclaim here
2778 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2779 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2780 * not recorded as it most likely matches current's and won't
2781 * change in the meantime. As high limit is checked again before
2782 * reclaim, the cost of mismatch is negligible.
2785 bool mem_high, swap_high;
2787 mem_high = page_counter_read(&memcg->memory) >
2788 READ_ONCE(memcg->memory.high);
2789 swap_high = page_counter_read(&memcg->swap) >
2790 READ_ONCE(memcg->swap.high);
2792 /* Don't bother a random interrupted task */
2793 if (in_interrupt()) {
2795 schedule_work(&memcg->high_work);
2801 if (mem_high || swap_high) {
2803 * The allocating tasks in this cgroup will need to do
2804 * reclaim or be throttled to prevent further growth
2805 * of the memory or swap footprints.
2807 * Target some best-effort fairness between the tasks,
2808 * and distribute reclaim work and delay penalties
2809 * based on how much each task is actually allocating.
2811 current->memcg_nr_pages_over_high += batch;
2812 set_notify_resume(current);
2815 } while ((memcg = parent_mem_cgroup(memcg)));
2820 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2821 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2823 if (mem_cgroup_is_root(memcg))
2826 page_counter_uncharge(&memcg->memory, nr_pages);
2827 if (do_memsw_account())
2828 page_counter_uncharge(&memcg->memsw, nr_pages);
2832 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2834 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2836 * Any of the following ensures page->mem_cgroup stability:
2840 * - lock_page_memcg()
2841 * - exclusive reference
2843 page->mem_cgroup = memcg;
2846 #ifdef CONFIG_MEMCG_KMEM
2847 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2850 unsigned int objects = objs_per_slab_page(s, page);
2853 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2858 if (cmpxchg(&page->obj_cgroups, NULL,
2859 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2862 kmemleak_not_leak(vec);
2868 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2870 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2871 * cgroup_mutex, etc.
2873 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2877 if (mem_cgroup_disabled())
2880 page = virt_to_head_page(p);
2883 * Slab objects are accounted individually, not per-page.
2884 * Memcg membership data for each individual object is saved in
2885 * the page->obj_cgroups.
2887 if (page_has_obj_cgroups(page)) {
2888 struct obj_cgroup *objcg;
2891 off = obj_to_index(page->slab_cache, page, p);
2892 objcg = page_obj_cgroups(page)[off];
2894 return obj_cgroup_memcg(objcg);
2899 /* All other pages use page->mem_cgroup */
2900 return page->mem_cgroup;
2903 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2905 struct obj_cgroup *objcg = NULL;
2906 struct mem_cgroup *memcg;
2908 if (unlikely(!current->mm && !current->active_memcg))
2912 if (unlikely(current->active_memcg))
2913 memcg = rcu_dereference(current->active_memcg);
2915 memcg = mem_cgroup_from_task(current);
2917 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2918 objcg = rcu_dereference(memcg->objcg);
2919 if (objcg && obj_cgroup_tryget(objcg))
2927 static int memcg_alloc_cache_id(void)
2932 id = ida_simple_get(&memcg_cache_ida,
2933 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2937 if (id < memcg_nr_cache_ids)
2941 * There's no space for the new id in memcg_caches arrays,
2942 * so we have to grow them.
2944 down_write(&memcg_cache_ids_sem);
2946 size = 2 * (id + 1);
2947 if (size < MEMCG_CACHES_MIN_SIZE)
2948 size = MEMCG_CACHES_MIN_SIZE;
2949 else if (size > MEMCG_CACHES_MAX_SIZE)
2950 size = MEMCG_CACHES_MAX_SIZE;
2952 err = memcg_update_all_list_lrus(size);
2954 memcg_nr_cache_ids = size;
2956 up_write(&memcg_cache_ids_sem);
2959 ida_simple_remove(&memcg_cache_ida, id);
2965 static void memcg_free_cache_id(int id)
2967 ida_simple_remove(&memcg_cache_ida, id);
2971 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
2972 * @memcg: memory cgroup to charge
2973 * @gfp: reclaim mode
2974 * @nr_pages: number of pages to charge
2976 * Returns 0 on success, an error code on failure.
2978 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
2979 unsigned int nr_pages)
2981 struct page_counter *counter;
2984 ret = try_charge(memcg, gfp, nr_pages);
2988 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
2989 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
2992 * Enforce __GFP_NOFAIL allocation because callers are not
2993 * prepared to see failures and likely do not have any failure
2996 if (gfp & __GFP_NOFAIL) {
2997 page_counter_charge(&memcg->kmem, nr_pages);
3000 cancel_charge(memcg, nr_pages);
3007 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3008 * @memcg: memcg to uncharge
3009 * @nr_pages: number of pages to uncharge
3011 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3013 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3014 page_counter_uncharge(&memcg->kmem, nr_pages);
3016 page_counter_uncharge(&memcg->memory, nr_pages);
3017 if (do_memsw_account())
3018 page_counter_uncharge(&memcg->memsw, nr_pages);
3022 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3023 * @page: page to charge
3024 * @gfp: reclaim mode
3025 * @order: allocation order
3027 * Returns 0 on success, an error code on failure.
3029 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3031 struct mem_cgroup *memcg;
3034 if (memcg_kmem_bypass())
3037 memcg = get_mem_cgroup_from_current();
3038 if (!mem_cgroup_is_root(memcg)) {
3039 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3041 page->mem_cgroup = memcg;
3042 __SetPageKmemcg(page);
3046 css_put(&memcg->css);
3051 * __memcg_kmem_uncharge_page: uncharge a kmem page
3052 * @page: page to uncharge
3053 * @order: allocation order
3055 void __memcg_kmem_uncharge_page(struct page *page, int order)
3057 struct mem_cgroup *memcg = page->mem_cgroup;
3058 unsigned int nr_pages = 1 << order;
3063 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3064 __memcg_kmem_uncharge(memcg, nr_pages);
3065 page->mem_cgroup = NULL;
3066 css_put(&memcg->css);
3068 /* slab pages do not have PageKmemcg flag set */
3069 if (PageKmemcg(page))
3070 __ClearPageKmemcg(page);
3073 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3075 struct memcg_stock_pcp *stock;
3076 unsigned long flags;
3079 local_irq_save(flags);
3081 stock = this_cpu_ptr(&memcg_stock);
3082 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3083 stock->nr_bytes -= nr_bytes;
3087 local_irq_restore(flags);
3092 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3094 struct obj_cgroup *old = stock->cached_objcg;
3099 if (stock->nr_bytes) {
3100 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3101 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3105 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3110 * The leftover is flushed to the centralized per-memcg value.
3111 * On the next attempt to refill obj stock it will be moved
3112 * to a per-cpu stock (probably, on an other CPU), see
3113 * refill_obj_stock().
3115 * How often it's flushed is a trade-off between the memory
3116 * limit enforcement accuracy and potential CPU contention,
3117 * so it might be changed in the future.
3119 atomic_add(nr_bytes, &old->nr_charged_bytes);
3120 stock->nr_bytes = 0;
3123 obj_cgroup_put(old);
3124 stock->cached_objcg = NULL;
3127 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3128 struct mem_cgroup *root_memcg)
3130 struct mem_cgroup *memcg;
3132 if (stock->cached_objcg) {
3133 memcg = obj_cgroup_memcg(stock->cached_objcg);
3134 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3141 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3143 struct memcg_stock_pcp *stock;
3144 unsigned long flags;
3146 local_irq_save(flags);
3148 stock = this_cpu_ptr(&memcg_stock);
3149 if (stock->cached_objcg != objcg) { /* reset if necessary */
3150 drain_obj_stock(stock);
3151 obj_cgroup_get(objcg);
3152 stock->cached_objcg = objcg;
3153 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3155 stock->nr_bytes += nr_bytes;
3157 if (stock->nr_bytes > PAGE_SIZE)
3158 drain_obj_stock(stock);
3160 local_irq_restore(flags);
3163 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3165 struct mem_cgroup *memcg;
3166 unsigned int nr_pages, nr_bytes;
3169 if (consume_obj_stock(objcg, size))
3173 * In theory, memcg->nr_charged_bytes can have enough
3174 * pre-charged bytes to satisfy the allocation. However,
3175 * flushing memcg->nr_charged_bytes requires two atomic
3176 * operations, and memcg->nr_charged_bytes can't be big,
3177 * so it's better to ignore it and try grab some new pages.
3178 * memcg->nr_charged_bytes will be flushed in
3179 * refill_obj_stock(), called from this function or
3180 * independently later.
3183 memcg = obj_cgroup_memcg(objcg);
3184 css_get(&memcg->css);
3187 nr_pages = size >> PAGE_SHIFT;
3188 nr_bytes = size & (PAGE_SIZE - 1);
3193 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3194 if (!ret && nr_bytes)
3195 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3197 css_put(&memcg->css);
3201 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3203 refill_obj_stock(objcg, size);
3206 #endif /* CONFIG_MEMCG_KMEM */
3208 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3211 * Because tail pages are not marked as "used", set it. We're under
3212 * pgdat->lru_lock and migration entries setup in all page mappings.
3214 void mem_cgroup_split_huge_fixup(struct page *head)
3216 struct mem_cgroup *memcg = head->mem_cgroup;
3219 if (mem_cgroup_disabled())
3222 for (i = 1; i < HPAGE_PMD_NR; i++) {
3223 css_get(&memcg->css);
3224 head[i].mem_cgroup = memcg;
3227 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3229 #ifdef CONFIG_MEMCG_SWAP
3231 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3232 * @entry: swap entry to be moved
3233 * @from: mem_cgroup which the entry is moved from
3234 * @to: mem_cgroup which the entry is moved to
3236 * It succeeds only when the swap_cgroup's record for this entry is the same
3237 * as the mem_cgroup's id of @from.
3239 * Returns 0 on success, -EINVAL on failure.
3241 * The caller must have charged to @to, IOW, called page_counter_charge() about
3242 * both res and memsw, and called css_get().
3244 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3245 struct mem_cgroup *from, struct mem_cgroup *to)
3247 unsigned short old_id, new_id;
3249 old_id = mem_cgroup_id(from);
3250 new_id = mem_cgroup_id(to);
3252 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3253 mod_memcg_state(from, MEMCG_SWAP, -1);
3254 mod_memcg_state(to, MEMCG_SWAP, 1);
3260 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3261 struct mem_cgroup *from, struct mem_cgroup *to)
3267 static DEFINE_MUTEX(memcg_max_mutex);
3269 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3270 unsigned long max, bool memsw)
3272 bool enlarge = false;
3273 bool drained = false;
3275 bool limits_invariant;
3276 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3279 if (signal_pending(current)) {
3284 mutex_lock(&memcg_max_mutex);
3286 * Make sure that the new limit (memsw or memory limit) doesn't
3287 * break our basic invariant rule memory.max <= memsw.max.
3289 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3290 max <= memcg->memsw.max;
3291 if (!limits_invariant) {
3292 mutex_unlock(&memcg_max_mutex);
3296 if (max > counter->max)
3298 ret = page_counter_set_max(counter, max);
3299 mutex_unlock(&memcg_max_mutex);
3305 drain_all_stock(memcg);
3310 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3311 GFP_KERNEL, !memsw)) {
3317 if (!ret && enlarge)
3318 memcg_oom_recover(memcg);
3323 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3325 unsigned long *total_scanned)
3327 unsigned long nr_reclaimed = 0;
3328 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3329 unsigned long reclaimed;
3331 struct mem_cgroup_tree_per_node *mctz;
3332 unsigned long excess;
3333 unsigned long nr_scanned;
3338 mctz = soft_limit_tree_node(pgdat->node_id);
3341 * Do not even bother to check the largest node if the root
3342 * is empty. Do it lockless to prevent lock bouncing. Races
3343 * are acceptable as soft limit is best effort anyway.
3345 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3349 * This loop can run a while, specially if mem_cgroup's continuously
3350 * keep exceeding their soft limit and putting the system under
3357 mz = mem_cgroup_largest_soft_limit_node(mctz);
3362 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3363 gfp_mask, &nr_scanned);
3364 nr_reclaimed += reclaimed;
3365 *total_scanned += nr_scanned;
3366 spin_lock_irq(&mctz->lock);
3367 __mem_cgroup_remove_exceeded(mz, mctz);
3370 * If we failed to reclaim anything from this memory cgroup
3371 * it is time to move on to the next cgroup
3375 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3377 excess = soft_limit_excess(mz->memcg);
3379 * One school of thought says that we should not add
3380 * back the node to the tree if reclaim returns 0.
3381 * But our reclaim could return 0, simply because due
3382 * to priority we are exposing a smaller subset of
3383 * memory to reclaim from. Consider this as a longer
3386 /* If excess == 0, no tree ops */
3387 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3388 spin_unlock_irq(&mctz->lock);
3389 css_put(&mz->memcg->css);
3392 * Could not reclaim anything and there are no more
3393 * mem cgroups to try or we seem to be looping without
3394 * reclaiming anything.
3396 if (!nr_reclaimed &&
3398 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3400 } while (!nr_reclaimed);
3402 css_put(&next_mz->memcg->css);
3403 return nr_reclaimed;
3407 * Test whether @memcg has children, dead or alive. Note that this
3408 * function doesn't care whether @memcg has use_hierarchy enabled and
3409 * returns %true if there are child csses according to the cgroup
3410 * hierarchy. Testing use_hierarchy is the caller's responsibility.
3412 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3417 ret = css_next_child(NULL, &memcg->css);
3423 * Reclaims as many pages from the given memcg as possible.
3425 * Caller is responsible for holding css reference for memcg.
3427 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3429 int nr_retries = MAX_RECLAIM_RETRIES;
3431 /* we call try-to-free pages for make this cgroup empty */
3432 lru_add_drain_all();
3434 drain_all_stock(memcg);
3436 /* try to free all pages in this cgroup */
3437 while (nr_retries && page_counter_read(&memcg->memory)) {
3440 if (signal_pending(current))
3443 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3447 /* maybe some writeback is necessary */
3448 congestion_wait(BLK_RW_ASYNC, HZ/10);
3456 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3457 char *buf, size_t nbytes,
3460 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3462 if (mem_cgroup_is_root(memcg))
3464 return mem_cgroup_force_empty(memcg) ?: nbytes;
3467 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3470 return mem_cgroup_from_css(css)->use_hierarchy;
3473 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3474 struct cftype *cft, u64 val)
3477 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3478 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3480 if (memcg->use_hierarchy == val)
3484 * If parent's use_hierarchy is set, we can't make any modifications
3485 * in the child subtrees. If it is unset, then the change can
3486 * occur, provided the current cgroup has no children.
3488 * For the root cgroup, parent_mem is NULL, we allow value to be
3489 * set if there are no children.
3491 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3492 (val == 1 || val == 0)) {
3493 if (!memcg_has_children(memcg))
3494 memcg->use_hierarchy = val;
3503 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3507 if (mem_cgroup_is_root(memcg)) {
3508 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3509 memcg_page_state(memcg, NR_ANON_MAPPED);
3511 val += memcg_page_state(memcg, MEMCG_SWAP);
3514 val = page_counter_read(&memcg->memory);
3516 val = page_counter_read(&memcg->memsw);
3529 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3532 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3533 struct page_counter *counter;
3535 switch (MEMFILE_TYPE(cft->private)) {
3537 counter = &memcg->memory;
3540 counter = &memcg->memsw;
3543 counter = &memcg->kmem;
3546 counter = &memcg->tcpmem;
3552 switch (MEMFILE_ATTR(cft->private)) {
3554 if (counter == &memcg->memory)
3555 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3556 if (counter == &memcg->memsw)
3557 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3558 return (u64)page_counter_read(counter) * PAGE_SIZE;
3560 return (u64)counter->max * PAGE_SIZE;
3562 return (u64)counter->watermark * PAGE_SIZE;
3564 return counter->failcnt;
3565 case RES_SOFT_LIMIT:
3566 return (u64)memcg->soft_limit * PAGE_SIZE;
3572 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3574 unsigned long stat[MEMCG_NR_STAT] = {0};
3575 struct mem_cgroup *mi;
3578 for_each_online_cpu(cpu)
3579 for (i = 0; i < MEMCG_NR_STAT; i++)
3580 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3582 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3583 for (i = 0; i < MEMCG_NR_STAT; i++)
3584 atomic_long_add(stat[i], &mi->vmstats[i]);
3586 for_each_node(node) {
3587 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3588 struct mem_cgroup_per_node *pi;
3590 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3593 for_each_online_cpu(cpu)
3594 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3596 pn->lruvec_stat_cpu->count[i], cpu);
3598 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3599 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3600 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3604 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3606 unsigned long events[NR_VM_EVENT_ITEMS];
3607 struct mem_cgroup *mi;
3610 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3613 for_each_online_cpu(cpu)
3614 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3615 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3618 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3619 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3620 atomic_long_add(events[i], &mi->vmevents[i]);
3623 #ifdef CONFIG_MEMCG_KMEM
3624 static int memcg_online_kmem(struct mem_cgroup *memcg)
3626 struct obj_cgroup *objcg;
3629 if (cgroup_memory_nokmem)
3632 BUG_ON(memcg->kmemcg_id >= 0);
3633 BUG_ON(memcg->kmem_state);
3635 memcg_id = memcg_alloc_cache_id();
3639 objcg = obj_cgroup_alloc();
3641 memcg_free_cache_id(memcg_id);
3644 objcg->memcg = memcg;
3645 rcu_assign_pointer(memcg->objcg, objcg);
3647 static_branch_enable(&memcg_kmem_enabled_key);
3650 * A memory cgroup is considered kmem-online as soon as it gets
3651 * kmemcg_id. Setting the id after enabling static branching will
3652 * guarantee no one starts accounting before all call sites are
3655 memcg->kmemcg_id = memcg_id;
3656 memcg->kmem_state = KMEM_ONLINE;
3661 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3663 struct cgroup_subsys_state *css;
3664 struct mem_cgroup *parent, *child;
3667 if (memcg->kmem_state != KMEM_ONLINE)
3670 memcg->kmem_state = KMEM_ALLOCATED;
3672 parent = parent_mem_cgroup(memcg);
3674 parent = root_mem_cgroup;
3676 memcg_reparent_objcgs(memcg, parent);
3678 kmemcg_id = memcg->kmemcg_id;
3679 BUG_ON(kmemcg_id < 0);
3682 * Change kmemcg_id of this cgroup and all its descendants to the
3683 * parent's id, and then move all entries from this cgroup's list_lrus
3684 * to ones of the parent. After we have finished, all list_lrus
3685 * corresponding to this cgroup are guaranteed to remain empty. The
3686 * ordering is imposed by list_lru_node->lock taken by
3687 * memcg_drain_all_list_lrus().
3689 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3690 css_for_each_descendant_pre(css, &memcg->css) {
3691 child = mem_cgroup_from_css(css);
3692 BUG_ON(child->kmemcg_id != kmemcg_id);
3693 child->kmemcg_id = parent->kmemcg_id;
3694 if (!memcg->use_hierarchy)
3699 memcg_drain_all_list_lrus(kmemcg_id, parent);
3701 memcg_free_cache_id(kmemcg_id);
3704 static void memcg_free_kmem(struct mem_cgroup *memcg)
3706 /* css_alloc() failed, offlining didn't happen */
3707 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3708 memcg_offline_kmem(memcg);
3711 static int memcg_online_kmem(struct mem_cgroup *memcg)
3715 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3718 static void memcg_free_kmem(struct mem_cgroup *memcg)
3721 #endif /* CONFIG_MEMCG_KMEM */
3723 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3728 mutex_lock(&memcg_max_mutex);
3729 ret = page_counter_set_max(&memcg->kmem, max);
3730 mutex_unlock(&memcg_max_mutex);
3734 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3738 mutex_lock(&memcg_max_mutex);
3740 ret = page_counter_set_max(&memcg->tcpmem, max);
3744 if (!memcg->tcpmem_active) {
3746 * The active flag needs to be written after the static_key
3747 * update. This is what guarantees that the socket activation
3748 * function is the last one to run. See mem_cgroup_sk_alloc()
3749 * for details, and note that we don't mark any socket as
3750 * belonging to this memcg until that flag is up.
3752 * We need to do this, because static_keys will span multiple
3753 * sites, but we can't control their order. If we mark a socket
3754 * as accounted, but the accounting functions are not patched in
3755 * yet, we'll lose accounting.
3757 * We never race with the readers in mem_cgroup_sk_alloc(),
3758 * because when this value change, the code to process it is not
3761 static_branch_inc(&memcg_sockets_enabled_key);
3762 memcg->tcpmem_active = true;
3765 mutex_unlock(&memcg_max_mutex);
3770 * The user of this function is...
3773 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3774 char *buf, size_t nbytes, loff_t off)
3776 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3777 unsigned long nr_pages;
3780 buf = strstrip(buf);
3781 ret = page_counter_memparse(buf, "-1", &nr_pages);
3785 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3787 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3791 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3793 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3796 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3799 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3800 "Please report your usecase to linux-mm@kvack.org if you "
3801 "depend on this functionality.\n");
3802 ret = memcg_update_kmem_max(memcg, nr_pages);
3805 ret = memcg_update_tcp_max(memcg, nr_pages);
3809 case RES_SOFT_LIMIT:
3810 memcg->soft_limit = nr_pages;
3814 return ret ?: nbytes;
3817 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3818 size_t nbytes, loff_t off)
3820 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3821 struct page_counter *counter;
3823 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3825 counter = &memcg->memory;
3828 counter = &memcg->memsw;
3831 counter = &memcg->kmem;
3834 counter = &memcg->tcpmem;
3840 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3842 page_counter_reset_watermark(counter);
3845 counter->failcnt = 0;
3854 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3857 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3861 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3862 struct cftype *cft, u64 val)
3864 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3866 if (val & ~MOVE_MASK)
3870 * No kind of locking is needed in here, because ->can_attach() will
3871 * check this value once in the beginning of the process, and then carry
3872 * on with stale data. This means that changes to this value will only
3873 * affect task migrations starting after the change.
3875 memcg->move_charge_at_immigrate = val;
3879 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3880 struct cftype *cft, u64 val)
3888 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3889 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3890 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3892 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3893 int nid, unsigned int lru_mask, bool tree)
3895 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3896 unsigned long nr = 0;
3899 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3902 if (!(BIT(lru) & lru_mask))
3905 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3907 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3912 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3913 unsigned int lru_mask,
3916 unsigned long nr = 0;
3920 if (!(BIT(lru) & lru_mask))
3923 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3925 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3930 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3934 unsigned int lru_mask;
3937 static const struct numa_stat stats[] = {
3938 { "total", LRU_ALL },
3939 { "file", LRU_ALL_FILE },
3940 { "anon", LRU_ALL_ANON },
3941 { "unevictable", BIT(LRU_UNEVICTABLE) },
3943 const struct numa_stat *stat;
3945 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3947 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3948 seq_printf(m, "%s=%lu", stat->name,
3949 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3951 for_each_node_state(nid, N_MEMORY)
3952 seq_printf(m, " N%d=%lu", nid,
3953 mem_cgroup_node_nr_lru_pages(memcg, nid,
3954 stat->lru_mask, false));
3958 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3960 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3961 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3963 for_each_node_state(nid, N_MEMORY)
3964 seq_printf(m, " N%d=%lu", nid,
3965 mem_cgroup_node_nr_lru_pages(memcg, nid,
3966 stat->lru_mask, true));
3972 #endif /* CONFIG_NUMA */
3974 static const unsigned int memcg1_stats[] = {
3977 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3987 static const char *const memcg1_stat_names[] = {
3990 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4000 /* Universal VM events cgroup1 shows, original sort order */
4001 static const unsigned int memcg1_events[] = {
4008 static int memcg_stat_show(struct seq_file *m, void *v)
4010 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4011 unsigned long memory, memsw;
4012 struct mem_cgroup *mi;
4015 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4017 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4020 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4022 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4023 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4024 if (memcg1_stats[i] == NR_ANON_THPS)
4027 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4030 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4031 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4032 memcg_events_local(memcg, memcg1_events[i]));
4034 for (i = 0; i < NR_LRU_LISTS; i++)
4035 seq_printf(m, "%s %lu\n", lru_list_name(i),
4036 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4039 /* Hierarchical information */
4040 memory = memsw = PAGE_COUNTER_MAX;
4041 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4042 memory = min(memory, READ_ONCE(mi->memory.max));
4043 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4045 seq_printf(m, "hierarchical_memory_limit %llu\n",
4046 (u64)memory * PAGE_SIZE);
4047 if (do_memsw_account())
4048 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4049 (u64)memsw * PAGE_SIZE);
4051 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4052 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4054 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4055 (u64)memcg_page_state(memcg, memcg1_stats[i]) *
4059 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4060 seq_printf(m, "total_%s %llu\n",
4061 vm_event_name(memcg1_events[i]),
4062 (u64)memcg_events(memcg, memcg1_events[i]));
4064 for (i = 0; i < NR_LRU_LISTS; i++)
4065 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4066 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4069 #ifdef CONFIG_DEBUG_VM
4072 struct mem_cgroup_per_node *mz;
4073 unsigned long anon_cost = 0;
4074 unsigned long file_cost = 0;
4076 for_each_online_pgdat(pgdat) {
4077 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4079 anon_cost += mz->lruvec.anon_cost;
4080 file_cost += mz->lruvec.file_cost;
4082 seq_printf(m, "anon_cost %lu\n", anon_cost);
4083 seq_printf(m, "file_cost %lu\n", file_cost);
4090 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4093 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4095 return mem_cgroup_swappiness(memcg);
4098 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4099 struct cftype *cft, u64 val)
4101 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4107 memcg->swappiness = val;
4109 vm_swappiness = val;
4114 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4116 struct mem_cgroup_threshold_ary *t;
4117 unsigned long usage;
4122 t = rcu_dereference(memcg->thresholds.primary);
4124 t = rcu_dereference(memcg->memsw_thresholds.primary);
4129 usage = mem_cgroup_usage(memcg, swap);
4132 * current_threshold points to threshold just below or equal to usage.
4133 * If it's not true, a threshold was crossed after last
4134 * call of __mem_cgroup_threshold().
4136 i = t->current_threshold;
4139 * Iterate backward over array of thresholds starting from
4140 * current_threshold and check if a threshold is crossed.
4141 * If none of thresholds below usage is crossed, we read
4142 * only one element of the array here.
4144 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4145 eventfd_signal(t->entries[i].eventfd, 1);
4147 /* i = current_threshold + 1 */
4151 * Iterate forward over array of thresholds starting from
4152 * current_threshold+1 and check if a threshold is crossed.
4153 * If none of thresholds above usage is crossed, we read
4154 * only one element of the array here.
4156 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4157 eventfd_signal(t->entries[i].eventfd, 1);
4159 /* Update current_threshold */
4160 t->current_threshold = i - 1;
4165 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4168 __mem_cgroup_threshold(memcg, false);
4169 if (do_memsw_account())
4170 __mem_cgroup_threshold(memcg, true);
4172 memcg = parent_mem_cgroup(memcg);
4176 static int compare_thresholds(const void *a, const void *b)
4178 const struct mem_cgroup_threshold *_a = a;
4179 const struct mem_cgroup_threshold *_b = b;
4181 if (_a->threshold > _b->threshold)
4184 if (_a->threshold < _b->threshold)
4190 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4192 struct mem_cgroup_eventfd_list *ev;
4194 spin_lock(&memcg_oom_lock);
4196 list_for_each_entry(ev, &memcg->oom_notify, list)
4197 eventfd_signal(ev->eventfd, 1);
4199 spin_unlock(&memcg_oom_lock);
4203 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4205 struct mem_cgroup *iter;
4207 for_each_mem_cgroup_tree(iter, memcg)
4208 mem_cgroup_oom_notify_cb(iter);
4211 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4212 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4214 struct mem_cgroup_thresholds *thresholds;
4215 struct mem_cgroup_threshold_ary *new;
4216 unsigned long threshold;
4217 unsigned long usage;
4220 ret = page_counter_memparse(args, "-1", &threshold);
4224 mutex_lock(&memcg->thresholds_lock);
4227 thresholds = &memcg->thresholds;
4228 usage = mem_cgroup_usage(memcg, false);
4229 } else if (type == _MEMSWAP) {
4230 thresholds = &memcg->memsw_thresholds;
4231 usage = mem_cgroup_usage(memcg, true);
4235 /* Check if a threshold crossed before adding a new one */
4236 if (thresholds->primary)
4237 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4239 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4241 /* Allocate memory for new array of thresholds */
4242 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4249 /* Copy thresholds (if any) to new array */
4250 if (thresholds->primary) {
4251 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4252 sizeof(struct mem_cgroup_threshold));
4255 /* Add new threshold */
4256 new->entries[size - 1].eventfd = eventfd;
4257 new->entries[size - 1].threshold = threshold;
4259 /* Sort thresholds. Registering of new threshold isn't time-critical */
4260 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4261 compare_thresholds, NULL);
4263 /* Find current threshold */
4264 new->current_threshold = -1;
4265 for (i = 0; i < size; i++) {
4266 if (new->entries[i].threshold <= usage) {
4268 * new->current_threshold will not be used until
4269 * rcu_assign_pointer(), so it's safe to increment
4272 ++new->current_threshold;
4277 /* Free old spare buffer and save old primary buffer as spare */
4278 kfree(thresholds->spare);
4279 thresholds->spare = thresholds->primary;
4281 rcu_assign_pointer(thresholds->primary, new);
4283 /* To be sure that nobody uses thresholds */
4287 mutex_unlock(&memcg->thresholds_lock);
4292 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4293 struct eventfd_ctx *eventfd, const char *args)
4295 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4298 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4299 struct eventfd_ctx *eventfd, const char *args)
4301 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4304 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4305 struct eventfd_ctx *eventfd, enum res_type type)
4307 struct mem_cgroup_thresholds *thresholds;
4308 struct mem_cgroup_threshold_ary *new;
4309 unsigned long usage;
4310 int i, j, size, entries;
4312 mutex_lock(&memcg->thresholds_lock);
4315 thresholds = &memcg->thresholds;
4316 usage = mem_cgroup_usage(memcg, false);
4317 } else if (type == _MEMSWAP) {
4318 thresholds = &memcg->memsw_thresholds;
4319 usage = mem_cgroup_usage(memcg, true);
4323 if (!thresholds->primary)
4326 /* Check if a threshold crossed before removing */
4327 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4329 /* Calculate new number of threshold */
4331 for (i = 0; i < thresholds->primary->size; i++) {
4332 if (thresholds->primary->entries[i].eventfd != eventfd)
4338 new = thresholds->spare;
4340 /* If no items related to eventfd have been cleared, nothing to do */
4344 /* Set thresholds array to NULL if we don't have thresholds */
4353 /* Copy thresholds and find current threshold */
4354 new->current_threshold = -1;
4355 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4356 if (thresholds->primary->entries[i].eventfd == eventfd)
4359 new->entries[j] = thresholds->primary->entries[i];
4360 if (new->entries[j].threshold <= usage) {
4362 * new->current_threshold will not be used
4363 * until rcu_assign_pointer(), so it's safe to increment
4366 ++new->current_threshold;
4372 /* Swap primary and spare array */
4373 thresholds->spare = thresholds->primary;
4375 rcu_assign_pointer(thresholds->primary, new);
4377 /* To be sure that nobody uses thresholds */
4380 /* If all events are unregistered, free the spare array */
4382 kfree(thresholds->spare);
4383 thresholds->spare = NULL;
4386 mutex_unlock(&memcg->thresholds_lock);
4389 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4390 struct eventfd_ctx *eventfd)
4392 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4395 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4396 struct eventfd_ctx *eventfd)
4398 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4401 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4402 struct eventfd_ctx *eventfd, const char *args)
4404 struct mem_cgroup_eventfd_list *event;
4406 event = kmalloc(sizeof(*event), GFP_KERNEL);
4410 spin_lock(&memcg_oom_lock);
4412 event->eventfd = eventfd;
4413 list_add(&event->list, &memcg->oom_notify);
4415 /* already in OOM ? */
4416 if (memcg->under_oom)
4417 eventfd_signal(eventfd, 1);
4418 spin_unlock(&memcg_oom_lock);
4423 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4424 struct eventfd_ctx *eventfd)
4426 struct mem_cgroup_eventfd_list *ev, *tmp;
4428 spin_lock(&memcg_oom_lock);
4430 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4431 if (ev->eventfd == eventfd) {
4432 list_del(&ev->list);
4437 spin_unlock(&memcg_oom_lock);
4440 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4442 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4444 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4445 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4446 seq_printf(sf, "oom_kill %lu\n",
4447 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4451 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4452 struct cftype *cft, u64 val)
4454 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4456 /* cannot set to root cgroup and only 0 and 1 are allowed */
4457 if (!css->parent || !((val == 0) || (val == 1)))
4460 memcg->oom_kill_disable = val;
4462 memcg_oom_recover(memcg);
4467 #ifdef CONFIG_CGROUP_WRITEBACK
4469 #include <trace/events/writeback.h>
4471 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4473 return wb_domain_init(&memcg->cgwb_domain, gfp);
4476 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4478 wb_domain_exit(&memcg->cgwb_domain);
4481 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4483 wb_domain_size_changed(&memcg->cgwb_domain);
4486 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4488 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4490 if (!memcg->css.parent)
4493 return &memcg->cgwb_domain;
4497 * idx can be of type enum memcg_stat_item or node_stat_item.
4498 * Keep in sync with memcg_exact_page().
4500 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4502 long x = atomic_long_read(&memcg->vmstats[idx]);
4505 for_each_online_cpu(cpu)
4506 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4513 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4514 * @wb: bdi_writeback in question
4515 * @pfilepages: out parameter for number of file pages
4516 * @pheadroom: out parameter for number of allocatable pages according to memcg
4517 * @pdirty: out parameter for number of dirty pages
4518 * @pwriteback: out parameter for number of pages under writeback
4520 * Determine the numbers of file, headroom, dirty, and writeback pages in
4521 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4522 * is a bit more involved.
4524 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4525 * headroom is calculated as the lowest headroom of itself and the
4526 * ancestors. Note that this doesn't consider the actual amount of
4527 * available memory in the system. The caller should further cap
4528 * *@pheadroom accordingly.
4530 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4531 unsigned long *pheadroom, unsigned long *pdirty,
4532 unsigned long *pwriteback)
4534 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4535 struct mem_cgroup *parent;
4537 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4539 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4540 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4541 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4542 *pheadroom = PAGE_COUNTER_MAX;
4544 while ((parent = parent_mem_cgroup(memcg))) {
4545 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4546 READ_ONCE(memcg->memory.high));
4547 unsigned long used = page_counter_read(&memcg->memory);
4549 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4555 * Foreign dirty flushing
4557 * There's an inherent mismatch between memcg and writeback. The former
4558 * trackes ownership per-page while the latter per-inode. This was a
4559 * deliberate design decision because honoring per-page ownership in the
4560 * writeback path is complicated, may lead to higher CPU and IO overheads
4561 * and deemed unnecessary given that write-sharing an inode across
4562 * different cgroups isn't a common use-case.
4564 * Combined with inode majority-writer ownership switching, this works well
4565 * enough in most cases but there are some pathological cases. For
4566 * example, let's say there are two cgroups A and B which keep writing to
4567 * different but confined parts of the same inode. B owns the inode and
4568 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4569 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4570 * triggering background writeback. A will be slowed down without a way to
4571 * make writeback of the dirty pages happen.
4573 * Conditions like the above can lead to a cgroup getting repatedly and
4574 * severely throttled after making some progress after each
4575 * dirty_expire_interval while the underyling IO device is almost
4578 * Solving this problem completely requires matching the ownership tracking
4579 * granularities between memcg and writeback in either direction. However,
4580 * the more egregious behaviors can be avoided by simply remembering the
4581 * most recent foreign dirtying events and initiating remote flushes on
4582 * them when local writeback isn't enough to keep the memory clean enough.
4584 * The following two functions implement such mechanism. When a foreign
4585 * page - a page whose memcg and writeback ownerships don't match - is
4586 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4587 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4588 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4589 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4590 * foreign bdi_writebacks which haven't expired. Both the numbers of
4591 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4592 * limited to MEMCG_CGWB_FRN_CNT.
4594 * The mechanism only remembers IDs and doesn't hold any object references.
4595 * As being wrong occasionally doesn't matter, updates and accesses to the
4596 * records are lockless and racy.
4598 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4599 struct bdi_writeback *wb)
4601 struct mem_cgroup *memcg = page->mem_cgroup;
4602 struct memcg_cgwb_frn *frn;
4603 u64 now = get_jiffies_64();
4604 u64 oldest_at = now;
4608 trace_track_foreign_dirty(page, wb);
4611 * Pick the slot to use. If there is already a slot for @wb, keep
4612 * using it. If not replace the oldest one which isn't being
4615 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4616 frn = &memcg->cgwb_frn[i];
4617 if (frn->bdi_id == wb->bdi->id &&
4618 frn->memcg_id == wb->memcg_css->id)
4620 if (time_before64(frn->at, oldest_at) &&
4621 atomic_read(&frn->done.cnt) == 1) {
4623 oldest_at = frn->at;
4627 if (i < MEMCG_CGWB_FRN_CNT) {
4629 * Re-using an existing one. Update timestamp lazily to
4630 * avoid making the cacheline hot. We want them to be
4631 * reasonably up-to-date and significantly shorter than
4632 * dirty_expire_interval as that's what expires the record.
4633 * Use the shorter of 1s and dirty_expire_interval / 8.
4635 unsigned long update_intv =
4636 min_t(unsigned long, HZ,
4637 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4639 if (time_before64(frn->at, now - update_intv))
4641 } else if (oldest >= 0) {
4642 /* replace the oldest free one */
4643 frn = &memcg->cgwb_frn[oldest];
4644 frn->bdi_id = wb->bdi->id;
4645 frn->memcg_id = wb->memcg_css->id;
4650 /* issue foreign writeback flushes for recorded foreign dirtying events */
4651 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4653 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4654 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4655 u64 now = jiffies_64;
4658 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4659 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4662 * If the record is older than dirty_expire_interval,
4663 * writeback on it has already started. No need to kick it
4664 * off again. Also, don't start a new one if there's
4665 * already one in flight.
4667 if (time_after64(frn->at, now - intv) &&
4668 atomic_read(&frn->done.cnt) == 1) {
4670 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4671 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4672 WB_REASON_FOREIGN_FLUSH,
4678 #else /* CONFIG_CGROUP_WRITEBACK */
4680 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4685 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4689 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4693 #endif /* CONFIG_CGROUP_WRITEBACK */
4696 * DO NOT USE IN NEW FILES.
4698 * "cgroup.event_control" implementation.
4700 * This is way over-engineered. It tries to support fully configurable
4701 * events for each user. Such level of flexibility is completely
4702 * unnecessary especially in the light of the planned unified hierarchy.
4704 * Please deprecate this and replace with something simpler if at all
4709 * Unregister event and free resources.
4711 * Gets called from workqueue.
4713 static void memcg_event_remove(struct work_struct *work)
4715 struct mem_cgroup_event *event =
4716 container_of(work, struct mem_cgroup_event, remove);
4717 struct mem_cgroup *memcg = event->memcg;
4719 remove_wait_queue(event->wqh, &event->wait);
4721 event->unregister_event(memcg, event->eventfd);
4723 /* Notify userspace the event is going away. */
4724 eventfd_signal(event->eventfd, 1);
4726 eventfd_ctx_put(event->eventfd);
4728 css_put(&memcg->css);
4732 * Gets called on EPOLLHUP on eventfd when user closes it.
4734 * Called with wqh->lock held and interrupts disabled.
4736 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4737 int sync, void *key)
4739 struct mem_cgroup_event *event =
4740 container_of(wait, struct mem_cgroup_event, wait);
4741 struct mem_cgroup *memcg = event->memcg;
4742 __poll_t flags = key_to_poll(key);
4744 if (flags & EPOLLHUP) {
4746 * If the event has been detached at cgroup removal, we
4747 * can simply return knowing the other side will cleanup
4750 * We can't race against event freeing since the other
4751 * side will require wqh->lock via remove_wait_queue(),
4754 spin_lock(&memcg->event_list_lock);
4755 if (!list_empty(&event->list)) {
4756 list_del_init(&event->list);
4758 * We are in atomic context, but cgroup_event_remove()
4759 * may sleep, so we have to call it in workqueue.
4761 schedule_work(&event->remove);
4763 spin_unlock(&memcg->event_list_lock);
4769 static void memcg_event_ptable_queue_proc(struct file *file,
4770 wait_queue_head_t *wqh, poll_table *pt)
4772 struct mem_cgroup_event *event =
4773 container_of(pt, struct mem_cgroup_event, pt);
4776 add_wait_queue(wqh, &event->wait);
4780 * DO NOT USE IN NEW FILES.
4782 * Parse input and register new cgroup event handler.
4784 * Input must be in format '<event_fd> <control_fd> <args>'.
4785 * Interpretation of args is defined by control file implementation.
4787 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4788 char *buf, size_t nbytes, loff_t off)
4790 struct cgroup_subsys_state *css = of_css(of);
4791 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4792 struct mem_cgroup_event *event;
4793 struct cgroup_subsys_state *cfile_css;
4794 unsigned int efd, cfd;
4801 buf = strstrip(buf);
4803 efd = simple_strtoul(buf, &endp, 10);
4808 cfd = simple_strtoul(buf, &endp, 10);
4809 if ((*endp != ' ') && (*endp != '\0'))
4813 event = kzalloc(sizeof(*event), GFP_KERNEL);
4817 event->memcg = memcg;
4818 INIT_LIST_HEAD(&event->list);
4819 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4820 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4821 INIT_WORK(&event->remove, memcg_event_remove);
4829 event->eventfd = eventfd_ctx_fileget(efile.file);
4830 if (IS_ERR(event->eventfd)) {
4831 ret = PTR_ERR(event->eventfd);
4838 goto out_put_eventfd;
4841 /* the process need read permission on control file */
4842 /* AV: shouldn't we check that it's been opened for read instead? */
4843 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4848 * Determine the event callbacks and set them in @event. This used
4849 * to be done via struct cftype but cgroup core no longer knows
4850 * about these events. The following is crude but the whole thing
4851 * is for compatibility anyway.
4853 * DO NOT ADD NEW FILES.
4855 name = cfile.file->f_path.dentry->d_name.name;
4857 if (!strcmp(name, "memory.usage_in_bytes")) {
4858 event->register_event = mem_cgroup_usage_register_event;
4859 event->unregister_event = mem_cgroup_usage_unregister_event;
4860 } else if (!strcmp(name, "memory.oom_control")) {
4861 event->register_event = mem_cgroup_oom_register_event;
4862 event->unregister_event = mem_cgroup_oom_unregister_event;
4863 } else if (!strcmp(name, "memory.pressure_level")) {
4864 event->register_event = vmpressure_register_event;
4865 event->unregister_event = vmpressure_unregister_event;
4866 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4867 event->register_event = memsw_cgroup_usage_register_event;
4868 event->unregister_event = memsw_cgroup_usage_unregister_event;
4875 * Verify @cfile should belong to @css. Also, remaining events are
4876 * automatically removed on cgroup destruction but the removal is
4877 * asynchronous, so take an extra ref on @css.
4879 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4880 &memory_cgrp_subsys);
4882 if (IS_ERR(cfile_css))
4884 if (cfile_css != css) {
4889 ret = event->register_event(memcg, event->eventfd, buf);
4893 vfs_poll(efile.file, &event->pt);
4895 spin_lock(&memcg->event_list_lock);
4896 list_add(&event->list, &memcg->event_list);
4897 spin_unlock(&memcg->event_list_lock);
4909 eventfd_ctx_put(event->eventfd);
4918 static struct cftype mem_cgroup_legacy_files[] = {
4920 .name = "usage_in_bytes",
4921 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4922 .read_u64 = mem_cgroup_read_u64,
4925 .name = "max_usage_in_bytes",
4926 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4927 .write = mem_cgroup_reset,
4928 .read_u64 = mem_cgroup_read_u64,
4931 .name = "limit_in_bytes",
4932 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4933 .write = mem_cgroup_write,
4934 .read_u64 = mem_cgroup_read_u64,
4937 .name = "soft_limit_in_bytes",
4938 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4939 .write = mem_cgroup_write,
4940 .read_u64 = mem_cgroup_read_u64,
4944 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4945 .write = mem_cgroup_reset,
4946 .read_u64 = mem_cgroup_read_u64,
4950 .seq_show = memcg_stat_show,
4953 .name = "force_empty",
4954 .write = mem_cgroup_force_empty_write,
4957 .name = "use_hierarchy",
4958 .write_u64 = mem_cgroup_hierarchy_write,
4959 .read_u64 = mem_cgroup_hierarchy_read,
4962 .name = "cgroup.event_control", /* XXX: for compat */
4963 .write = memcg_write_event_control,
4964 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4967 .name = "swappiness",
4968 .read_u64 = mem_cgroup_swappiness_read,
4969 .write_u64 = mem_cgroup_swappiness_write,
4972 .name = "move_charge_at_immigrate",
4973 .read_u64 = mem_cgroup_move_charge_read,
4974 .write_u64 = mem_cgroup_move_charge_write,
4977 .name = "oom_control",
4978 .seq_show = mem_cgroup_oom_control_read,
4979 .write_u64 = mem_cgroup_oom_control_write,
4980 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4983 .name = "pressure_level",
4987 .name = "numa_stat",
4988 .seq_show = memcg_numa_stat_show,
4992 .name = "kmem.limit_in_bytes",
4993 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4994 .write = mem_cgroup_write,
4995 .read_u64 = mem_cgroup_read_u64,
4998 .name = "kmem.usage_in_bytes",
4999 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5000 .read_u64 = mem_cgroup_read_u64,
5003 .name = "kmem.failcnt",
5004 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5005 .write = mem_cgroup_reset,
5006 .read_u64 = mem_cgroup_read_u64,
5009 .name = "kmem.max_usage_in_bytes",
5010 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5011 .write = mem_cgroup_reset,
5012 .read_u64 = mem_cgroup_read_u64,
5014 #if defined(CONFIG_MEMCG_KMEM) && \
5015 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5017 .name = "kmem.slabinfo",
5018 .seq_show = memcg_slab_show,
5022 .name = "kmem.tcp.limit_in_bytes",
5023 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5024 .write = mem_cgroup_write,
5025 .read_u64 = mem_cgroup_read_u64,
5028 .name = "kmem.tcp.usage_in_bytes",
5029 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5030 .read_u64 = mem_cgroup_read_u64,
5033 .name = "kmem.tcp.failcnt",
5034 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5035 .write = mem_cgroup_reset,
5036 .read_u64 = mem_cgroup_read_u64,
5039 .name = "kmem.tcp.max_usage_in_bytes",
5040 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5041 .write = mem_cgroup_reset,
5042 .read_u64 = mem_cgroup_read_u64,
5044 { }, /* terminate */
5048 * Private memory cgroup IDR
5050 * Swap-out records and page cache shadow entries need to store memcg
5051 * references in constrained space, so we maintain an ID space that is
5052 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5053 * memory-controlled cgroups to 64k.
5055 * However, there usually are many references to the offline CSS after
5056 * the cgroup has been destroyed, such as page cache or reclaimable
5057 * slab objects, that don't need to hang on to the ID. We want to keep
5058 * those dead CSS from occupying IDs, or we might quickly exhaust the
5059 * relatively small ID space and prevent the creation of new cgroups
5060 * even when there are much fewer than 64k cgroups - possibly none.
5062 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5063 * be freed and recycled when it's no longer needed, which is usually
5064 * when the CSS is offlined.
5066 * The only exception to that are records of swapped out tmpfs/shmem
5067 * pages that need to be attributed to live ancestors on swapin. But
5068 * those references are manageable from userspace.
5071 static DEFINE_IDR(mem_cgroup_idr);
5073 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5075 if (memcg->id.id > 0) {
5076 idr_remove(&mem_cgroup_idr, memcg->id.id);
5081 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5084 refcount_add(n, &memcg->id.ref);
5087 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5089 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5090 mem_cgroup_id_remove(memcg);
5092 /* Memcg ID pins CSS */
5093 css_put(&memcg->css);
5097 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5099 mem_cgroup_id_put_many(memcg, 1);
5103 * mem_cgroup_from_id - look up a memcg from a memcg id
5104 * @id: the memcg id to look up
5106 * Caller must hold rcu_read_lock().
5108 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5110 WARN_ON_ONCE(!rcu_read_lock_held());
5111 return idr_find(&mem_cgroup_idr, id);
5114 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5116 struct mem_cgroup_per_node *pn;
5119 * This routine is called against possible nodes.
5120 * But it's BUG to call kmalloc() against offline node.
5122 * TODO: this routine can waste much memory for nodes which will
5123 * never be onlined. It's better to use memory hotplug callback
5126 if (!node_state(node, N_NORMAL_MEMORY))
5128 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5132 pn->lruvec_stat_local = alloc_percpu(struct lruvec_stat);
5133 if (!pn->lruvec_stat_local) {
5138 pn->lruvec_stat_cpu = alloc_percpu(struct lruvec_stat);
5139 if (!pn->lruvec_stat_cpu) {
5140 free_percpu(pn->lruvec_stat_local);
5145 lruvec_init(&pn->lruvec);
5146 pn->usage_in_excess = 0;
5147 pn->on_tree = false;
5150 memcg->nodeinfo[node] = pn;
5154 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5156 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5161 free_percpu(pn->lruvec_stat_cpu);
5162 free_percpu(pn->lruvec_stat_local);
5166 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5171 free_mem_cgroup_per_node_info(memcg, node);
5172 free_percpu(memcg->vmstats_percpu);
5173 free_percpu(memcg->vmstats_local);
5177 static void mem_cgroup_free(struct mem_cgroup *memcg)
5179 memcg_wb_domain_exit(memcg);
5181 * Flush percpu vmstats and vmevents to guarantee the value correctness
5182 * on parent's and all ancestor levels.
5184 memcg_flush_percpu_vmstats(memcg);
5185 memcg_flush_percpu_vmevents(memcg);
5186 __mem_cgroup_free(memcg);
5189 static struct mem_cgroup *mem_cgroup_alloc(void)
5191 struct mem_cgroup *memcg;
5194 int __maybe_unused i;
5195 long error = -ENOMEM;
5197 size = sizeof(struct mem_cgroup);
5198 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5200 memcg = kzalloc(size, GFP_KERNEL);
5202 return ERR_PTR(error);
5204 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5205 1, MEM_CGROUP_ID_MAX,
5207 if (memcg->id.id < 0) {
5208 error = memcg->id.id;
5212 memcg->vmstats_local = alloc_percpu(struct memcg_vmstats_percpu);
5213 if (!memcg->vmstats_local)
5216 memcg->vmstats_percpu = alloc_percpu(struct memcg_vmstats_percpu);
5217 if (!memcg->vmstats_percpu)
5221 if (alloc_mem_cgroup_per_node_info(memcg, node))
5224 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5227 INIT_WORK(&memcg->high_work, high_work_func);
5228 INIT_LIST_HEAD(&memcg->oom_notify);
5229 mutex_init(&memcg->thresholds_lock);
5230 spin_lock_init(&memcg->move_lock);
5231 vmpressure_init(&memcg->vmpressure);
5232 INIT_LIST_HEAD(&memcg->event_list);
5233 spin_lock_init(&memcg->event_list_lock);
5234 memcg->socket_pressure = jiffies;
5235 #ifdef CONFIG_MEMCG_KMEM
5236 memcg->kmemcg_id = -1;
5237 INIT_LIST_HEAD(&memcg->objcg_list);
5239 #ifdef CONFIG_CGROUP_WRITEBACK
5240 INIT_LIST_HEAD(&memcg->cgwb_list);
5241 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5242 memcg->cgwb_frn[i].done =
5243 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5245 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5246 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5247 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5248 memcg->deferred_split_queue.split_queue_len = 0;
5250 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5253 mem_cgroup_id_remove(memcg);
5254 __mem_cgroup_free(memcg);
5255 return ERR_PTR(error);
5258 static struct cgroup_subsys_state * __ref
5259 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5261 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5262 struct mem_cgroup *memcg;
5263 long error = -ENOMEM;
5265 memcg = mem_cgroup_alloc();
5267 return ERR_CAST(memcg);
5269 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5270 memcg->soft_limit = PAGE_COUNTER_MAX;
5271 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5273 memcg->swappiness = mem_cgroup_swappiness(parent);
5274 memcg->oom_kill_disable = parent->oom_kill_disable;
5276 if (parent && parent->use_hierarchy) {
5277 memcg->use_hierarchy = true;
5278 page_counter_init(&memcg->memory, &parent->memory);
5279 page_counter_init(&memcg->swap, &parent->swap);
5280 page_counter_init(&memcg->memsw, &parent->memsw);
5281 page_counter_init(&memcg->kmem, &parent->kmem);
5282 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5284 page_counter_init(&memcg->memory, NULL);
5285 page_counter_init(&memcg->swap, NULL);
5286 page_counter_init(&memcg->memsw, NULL);
5287 page_counter_init(&memcg->kmem, NULL);
5288 page_counter_init(&memcg->tcpmem, NULL);
5290 * Deeper hierachy with use_hierarchy == false doesn't make
5291 * much sense so let cgroup subsystem know about this
5292 * unfortunate state in our controller.
5294 if (parent != root_mem_cgroup)
5295 memory_cgrp_subsys.broken_hierarchy = true;
5298 /* The following stuff does not apply to the root */
5300 root_mem_cgroup = memcg;
5304 error = memcg_online_kmem(memcg);
5308 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5309 static_branch_inc(&memcg_sockets_enabled_key);
5313 mem_cgroup_id_remove(memcg);
5314 mem_cgroup_free(memcg);
5315 return ERR_PTR(error);
5318 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5320 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5323 * A memcg must be visible for memcg_expand_shrinker_maps()
5324 * by the time the maps are allocated. So, we allocate maps
5325 * here, when for_each_mem_cgroup() can't skip it.
5327 if (memcg_alloc_shrinker_maps(memcg)) {
5328 mem_cgroup_id_remove(memcg);
5332 /* Online state pins memcg ID, memcg ID pins CSS */
5333 refcount_set(&memcg->id.ref, 1);
5338 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5340 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5341 struct mem_cgroup_event *event, *tmp;
5344 * Unregister events and notify userspace.
5345 * Notify userspace about cgroup removing only after rmdir of cgroup
5346 * directory to avoid race between userspace and kernelspace.
5348 spin_lock(&memcg->event_list_lock);
5349 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5350 list_del_init(&event->list);
5351 schedule_work(&event->remove);
5353 spin_unlock(&memcg->event_list_lock);
5355 page_counter_set_min(&memcg->memory, 0);
5356 page_counter_set_low(&memcg->memory, 0);
5358 memcg_offline_kmem(memcg);
5359 wb_memcg_offline(memcg);
5361 drain_all_stock(memcg);
5363 mem_cgroup_id_put(memcg);
5366 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5368 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5370 invalidate_reclaim_iterators(memcg);
5373 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5375 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5376 int __maybe_unused i;
5378 #ifdef CONFIG_CGROUP_WRITEBACK
5379 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5380 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5382 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5383 static_branch_dec(&memcg_sockets_enabled_key);
5385 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5386 static_branch_dec(&memcg_sockets_enabled_key);
5388 vmpressure_cleanup(&memcg->vmpressure);
5389 cancel_work_sync(&memcg->high_work);
5390 mem_cgroup_remove_from_trees(memcg);
5391 memcg_free_shrinker_maps(memcg);
5392 memcg_free_kmem(memcg);
5393 mem_cgroup_free(memcg);
5397 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5398 * @css: the target css
5400 * Reset the states of the mem_cgroup associated with @css. This is
5401 * invoked when the userland requests disabling on the default hierarchy
5402 * but the memcg is pinned through dependency. The memcg should stop
5403 * applying policies and should revert to the vanilla state as it may be
5404 * made visible again.
5406 * The current implementation only resets the essential configurations.
5407 * This needs to be expanded to cover all the visible parts.
5409 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5411 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5413 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5414 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5415 page_counter_set_max(&memcg->memsw, PAGE_COUNTER_MAX);
5416 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5417 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5418 page_counter_set_min(&memcg->memory, 0);
5419 page_counter_set_low(&memcg->memory, 0);
5420 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5421 memcg->soft_limit = PAGE_COUNTER_MAX;
5422 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5423 memcg_wb_domain_size_changed(memcg);
5427 /* Handlers for move charge at task migration. */
5428 static int mem_cgroup_do_precharge(unsigned long count)
5432 /* Try a single bulk charge without reclaim first, kswapd may wake */
5433 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5435 mc.precharge += count;
5439 /* Try charges one by one with reclaim, but do not retry */
5441 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5455 enum mc_target_type {
5462 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5463 unsigned long addr, pte_t ptent)
5465 struct page *page = vm_normal_page(vma, addr, ptent);
5467 if (!page || !page_mapped(page))
5469 if (PageAnon(page)) {
5470 if (!(mc.flags & MOVE_ANON))
5473 if (!(mc.flags & MOVE_FILE))
5476 if (!get_page_unless_zero(page))
5482 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5483 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5484 pte_t ptent, swp_entry_t *entry)
5486 struct page *page = NULL;
5487 swp_entry_t ent = pte_to_swp_entry(ptent);
5489 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
5493 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5494 * a device and because they are not accessible by CPU they are store
5495 * as special swap entry in the CPU page table.
5497 if (is_device_private_entry(ent)) {
5498 page = device_private_entry_to_page(ent);
5500 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5501 * a refcount of 1 when free (unlike normal page)
5503 if (!page_ref_add_unless(page, 1, 1))
5509 * Because lookup_swap_cache() updates some statistics counter,
5510 * we call find_get_page() with swapper_space directly.
5512 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5513 entry->val = ent.val;
5518 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5519 pte_t ptent, swp_entry_t *entry)
5525 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5526 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5528 struct page *page = NULL;
5529 struct address_space *mapping;
5532 if (!vma->vm_file) /* anonymous vma */
5534 if (!(mc.flags & MOVE_FILE))
5537 mapping = vma->vm_file->f_mapping;
5538 pgoff = linear_page_index(vma, addr);
5540 /* page is moved even if it's not RSS of this task(page-faulted). */
5542 /* shmem/tmpfs may report page out on swap: account for that too. */
5543 if (shmem_mapping(mapping)) {
5544 page = find_get_entry(mapping, pgoff);
5545 if (xa_is_value(page)) {
5546 swp_entry_t swp = radix_to_swp_entry(page);
5548 page = find_get_page(swap_address_space(swp),
5552 page = find_get_page(mapping, pgoff);
5554 page = find_get_page(mapping, pgoff);
5560 * mem_cgroup_move_account - move account of the page
5562 * @compound: charge the page as compound or small page
5563 * @from: mem_cgroup which the page is moved from.
5564 * @to: mem_cgroup which the page is moved to. @from != @to.
5566 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5568 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5571 static int mem_cgroup_move_account(struct page *page,
5573 struct mem_cgroup *from,
5574 struct mem_cgroup *to)
5576 struct lruvec *from_vec, *to_vec;
5577 struct pglist_data *pgdat;
5578 unsigned int nr_pages = compound ? hpage_nr_pages(page) : 1;
5581 VM_BUG_ON(from == to);
5582 VM_BUG_ON_PAGE(PageLRU(page), page);
5583 VM_BUG_ON(compound && !PageTransHuge(page));
5586 * Prevent mem_cgroup_migrate() from looking at
5587 * page->mem_cgroup of its source page while we change it.
5590 if (!trylock_page(page))
5594 if (page->mem_cgroup != from)
5597 pgdat = page_pgdat(page);
5598 from_vec = mem_cgroup_lruvec(from, pgdat);
5599 to_vec = mem_cgroup_lruvec(to, pgdat);
5601 lock_page_memcg(page);
5603 if (PageAnon(page)) {
5604 if (page_mapped(page)) {
5605 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5606 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5607 if (PageTransHuge(page)) {
5608 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5610 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5616 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5617 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5619 if (PageSwapBacked(page)) {
5620 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5621 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5624 if (page_mapped(page)) {
5625 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5626 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5629 if (PageDirty(page)) {
5630 struct address_space *mapping = page_mapping(page);
5632 if (mapping_cap_account_dirty(mapping)) {
5633 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5635 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5641 if (PageWriteback(page)) {
5642 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5643 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5647 * All state has been migrated, let's switch to the new memcg.
5649 * It is safe to change page->mem_cgroup here because the page
5650 * is referenced, charged, isolated, and locked: we can't race
5651 * with (un)charging, migration, LRU putback, or anything else
5652 * that would rely on a stable page->mem_cgroup.
5654 * Note that lock_page_memcg is a memcg lock, not a page lock,
5655 * to save space. As soon as we switch page->mem_cgroup to a
5656 * new memcg that isn't locked, the above state can change
5657 * concurrently again. Make sure we're truly done with it.
5662 css_put(&from->css);
5664 page->mem_cgroup = to;
5666 __unlock_page_memcg(from);
5670 local_irq_disable();
5671 mem_cgroup_charge_statistics(to, page, nr_pages);
5672 memcg_check_events(to, page);
5673 mem_cgroup_charge_statistics(from, page, -nr_pages);
5674 memcg_check_events(from, page);
5683 * get_mctgt_type - get target type of moving charge
5684 * @vma: the vma the pte to be checked belongs
5685 * @addr: the address corresponding to the pte to be checked
5686 * @ptent: the pte to be checked
5687 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5690 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5691 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5692 * move charge. if @target is not NULL, the page is stored in target->page
5693 * with extra refcnt got(Callers should handle it).
5694 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5695 * target for charge migration. if @target is not NULL, the entry is stored
5697 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5698 * (so ZONE_DEVICE page and thus not on the lru).
5699 * For now we such page is charge like a regular page would be as for all
5700 * intent and purposes it is just special memory taking the place of a
5703 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5705 * Called with pte lock held.
5708 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5709 unsigned long addr, pte_t ptent, union mc_target *target)
5711 struct page *page = NULL;
5712 enum mc_target_type ret = MC_TARGET_NONE;
5713 swp_entry_t ent = { .val = 0 };
5715 if (pte_present(ptent))
5716 page = mc_handle_present_pte(vma, addr, ptent);
5717 else if (is_swap_pte(ptent))
5718 page = mc_handle_swap_pte(vma, ptent, &ent);
5719 else if (pte_none(ptent))
5720 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5722 if (!page && !ent.val)
5726 * Do only loose check w/o serialization.
5727 * mem_cgroup_move_account() checks the page is valid or
5728 * not under LRU exclusion.
5730 if (page->mem_cgroup == mc.from) {
5731 ret = MC_TARGET_PAGE;
5732 if (is_device_private_page(page))
5733 ret = MC_TARGET_DEVICE;
5735 target->page = page;
5737 if (!ret || !target)
5741 * There is a swap entry and a page doesn't exist or isn't charged.
5742 * But we cannot move a tail-page in a THP.
5744 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5745 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5746 ret = MC_TARGET_SWAP;
5753 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5755 * We don't consider PMD mapped swapping or file mapped pages because THP does
5756 * not support them for now.
5757 * Caller should make sure that pmd_trans_huge(pmd) is true.
5759 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5760 unsigned long addr, pmd_t pmd, union mc_target *target)
5762 struct page *page = NULL;
5763 enum mc_target_type ret = MC_TARGET_NONE;
5765 if (unlikely(is_swap_pmd(pmd))) {
5766 VM_BUG_ON(thp_migration_supported() &&
5767 !is_pmd_migration_entry(pmd));
5770 page = pmd_page(pmd);
5771 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5772 if (!(mc.flags & MOVE_ANON))
5774 if (page->mem_cgroup == mc.from) {
5775 ret = MC_TARGET_PAGE;
5778 target->page = page;
5784 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5785 unsigned long addr, pmd_t pmd, union mc_target *target)
5787 return MC_TARGET_NONE;
5791 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5792 unsigned long addr, unsigned long end,
5793 struct mm_walk *walk)
5795 struct vm_area_struct *vma = walk->vma;
5799 ptl = pmd_trans_huge_lock(pmd, vma);
5802 * Note their can not be MC_TARGET_DEVICE for now as we do not
5803 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5804 * this might change.
5806 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5807 mc.precharge += HPAGE_PMD_NR;
5812 if (pmd_trans_unstable(pmd))
5814 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5815 for (; addr != end; pte++, addr += PAGE_SIZE)
5816 if (get_mctgt_type(vma, addr, *pte, NULL))
5817 mc.precharge++; /* increment precharge temporarily */
5818 pte_unmap_unlock(pte - 1, ptl);
5824 static const struct mm_walk_ops precharge_walk_ops = {
5825 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5828 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5830 unsigned long precharge;
5833 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5834 mmap_read_unlock(mm);
5836 precharge = mc.precharge;
5842 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5844 unsigned long precharge = mem_cgroup_count_precharge(mm);
5846 VM_BUG_ON(mc.moving_task);
5847 mc.moving_task = current;
5848 return mem_cgroup_do_precharge(precharge);
5851 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5852 static void __mem_cgroup_clear_mc(void)
5854 struct mem_cgroup *from = mc.from;
5855 struct mem_cgroup *to = mc.to;
5857 /* we must uncharge all the leftover precharges from mc.to */
5859 cancel_charge(mc.to, mc.precharge);
5863 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5864 * we must uncharge here.
5866 if (mc.moved_charge) {
5867 cancel_charge(mc.from, mc.moved_charge);
5868 mc.moved_charge = 0;
5870 /* we must fixup refcnts and charges */
5871 if (mc.moved_swap) {
5872 /* uncharge swap account from the old cgroup */
5873 if (!mem_cgroup_is_root(mc.from))
5874 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5876 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5879 * we charged both to->memory and to->memsw, so we
5880 * should uncharge to->memory.
5882 if (!mem_cgroup_is_root(mc.to))
5883 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5887 memcg_oom_recover(from);
5888 memcg_oom_recover(to);
5889 wake_up_all(&mc.waitq);
5892 static void mem_cgroup_clear_mc(void)
5894 struct mm_struct *mm = mc.mm;
5897 * we must clear moving_task before waking up waiters at the end of
5900 mc.moving_task = NULL;
5901 __mem_cgroup_clear_mc();
5902 spin_lock(&mc.lock);
5906 spin_unlock(&mc.lock);
5911 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5913 struct cgroup_subsys_state *css;
5914 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5915 struct mem_cgroup *from;
5916 struct task_struct *leader, *p;
5917 struct mm_struct *mm;
5918 unsigned long move_flags;
5921 /* charge immigration isn't supported on the default hierarchy */
5922 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5926 * Multi-process migrations only happen on the default hierarchy
5927 * where charge immigration is not used. Perform charge
5928 * immigration if @tset contains a leader and whine if there are
5932 cgroup_taskset_for_each_leader(leader, css, tset) {
5935 memcg = mem_cgroup_from_css(css);
5941 * We are now commited to this value whatever it is. Changes in this
5942 * tunable will only affect upcoming migrations, not the current one.
5943 * So we need to save it, and keep it going.
5945 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5949 from = mem_cgroup_from_task(p);
5951 VM_BUG_ON(from == memcg);
5953 mm = get_task_mm(p);
5956 /* We move charges only when we move a owner of the mm */
5957 if (mm->owner == p) {
5960 VM_BUG_ON(mc.precharge);
5961 VM_BUG_ON(mc.moved_charge);
5962 VM_BUG_ON(mc.moved_swap);
5964 spin_lock(&mc.lock);
5968 mc.flags = move_flags;
5969 spin_unlock(&mc.lock);
5970 /* We set mc.moving_task later */
5972 ret = mem_cgroup_precharge_mc(mm);
5974 mem_cgroup_clear_mc();
5981 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5984 mem_cgroup_clear_mc();
5987 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5988 unsigned long addr, unsigned long end,
5989 struct mm_walk *walk)
5992 struct vm_area_struct *vma = walk->vma;
5995 enum mc_target_type target_type;
5996 union mc_target target;
5999 ptl = pmd_trans_huge_lock(pmd, vma);
6001 if (mc.precharge < HPAGE_PMD_NR) {
6005 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6006 if (target_type == MC_TARGET_PAGE) {
6008 if (!isolate_lru_page(page)) {
6009 if (!mem_cgroup_move_account(page, true,
6011 mc.precharge -= HPAGE_PMD_NR;
6012 mc.moved_charge += HPAGE_PMD_NR;
6014 putback_lru_page(page);
6017 } else if (target_type == MC_TARGET_DEVICE) {
6019 if (!mem_cgroup_move_account(page, true,
6021 mc.precharge -= HPAGE_PMD_NR;
6022 mc.moved_charge += HPAGE_PMD_NR;
6030 if (pmd_trans_unstable(pmd))
6033 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6034 for (; addr != end; addr += PAGE_SIZE) {
6035 pte_t ptent = *(pte++);
6036 bool device = false;
6042 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6043 case MC_TARGET_DEVICE:
6046 case MC_TARGET_PAGE:
6049 * We can have a part of the split pmd here. Moving it
6050 * can be done but it would be too convoluted so simply
6051 * ignore such a partial THP and keep it in original
6052 * memcg. There should be somebody mapping the head.
6054 if (PageTransCompound(page))
6056 if (!device && isolate_lru_page(page))
6058 if (!mem_cgroup_move_account(page, false,
6061 /* we uncharge from mc.from later. */
6065 putback_lru_page(page);
6066 put: /* get_mctgt_type() gets the page */
6069 case MC_TARGET_SWAP:
6071 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6073 mem_cgroup_id_get_many(mc.to, 1);
6074 /* we fixup other refcnts and charges later. */
6082 pte_unmap_unlock(pte - 1, ptl);
6087 * We have consumed all precharges we got in can_attach().
6088 * We try charge one by one, but don't do any additional
6089 * charges to mc.to if we have failed in charge once in attach()
6092 ret = mem_cgroup_do_precharge(1);
6100 static const struct mm_walk_ops charge_walk_ops = {
6101 .pmd_entry = mem_cgroup_move_charge_pte_range,
6104 static void mem_cgroup_move_charge(void)
6106 lru_add_drain_all();
6108 * Signal lock_page_memcg() to take the memcg's move_lock
6109 * while we're moving its pages to another memcg. Then wait
6110 * for already started RCU-only updates to finish.
6112 atomic_inc(&mc.from->moving_account);
6115 if (unlikely(!mmap_read_trylock(mc.mm))) {
6117 * Someone who are holding the mmap_lock might be waiting in
6118 * waitq. So we cancel all extra charges, wake up all waiters,
6119 * and retry. Because we cancel precharges, we might not be able
6120 * to move enough charges, but moving charge is a best-effort
6121 * feature anyway, so it wouldn't be a big problem.
6123 __mem_cgroup_clear_mc();
6128 * When we have consumed all precharges and failed in doing
6129 * additional charge, the page walk just aborts.
6131 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6134 mmap_read_unlock(mc.mm);
6135 atomic_dec(&mc.from->moving_account);
6138 static void mem_cgroup_move_task(void)
6141 mem_cgroup_move_charge();
6142 mem_cgroup_clear_mc();
6145 #else /* !CONFIG_MMU */
6146 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6150 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6153 static void mem_cgroup_move_task(void)
6159 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6160 * to verify whether we're attached to the default hierarchy on each mount
6163 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6166 * use_hierarchy is forced on the default hierarchy. cgroup core
6167 * guarantees that @root doesn't have any children, so turning it
6168 * on for the root memcg is enough.
6170 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
6171 root_mem_cgroup->use_hierarchy = true;
6173 root_mem_cgroup->use_hierarchy = false;
6176 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6178 if (value == PAGE_COUNTER_MAX)
6179 seq_puts(m, "max\n");
6181 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6186 static u64 memory_current_read(struct cgroup_subsys_state *css,
6189 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6191 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6194 static int memory_min_show(struct seq_file *m, void *v)
6196 return seq_puts_memcg_tunable(m,
6197 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6200 static ssize_t memory_min_write(struct kernfs_open_file *of,
6201 char *buf, size_t nbytes, loff_t off)
6203 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6207 buf = strstrip(buf);
6208 err = page_counter_memparse(buf, "max", &min);
6212 page_counter_set_min(&memcg->memory, min);
6217 static int memory_low_show(struct seq_file *m, void *v)
6219 return seq_puts_memcg_tunable(m,
6220 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6223 static ssize_t memory_low_write(struct kernfs_open_file *of,
6224 char *buf, size_t nbytes, loff_t off)
6226 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6230 buf = strstrip(buf);
6231 err = page_counter_memparse(buf, "max", &low);
6235 page_counter_set_low(&memcg->memory, low);
6240 static int memory_high_show(struct seq_file *m, void *v)
6242 return seq_puts_memcg_tunable(m,
6243 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6246 static ssize_t memory_high_write(struct kernfs_open_file *of,
6247 char *buf, size_t nbytes, loff_t off)
6249 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6250 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6251 bool drained = false;
6255 buf = strstrip(buf);
6256 err = page_counter_memparse(buf, "max", &high);
6261 unsigned long nr_pages = page_counter_read(&memcg->memory);
6262 unsigned long reclaimed;
6264 if (nr_pages <= high)
6267 if (signal_pending(current))
6271 drain_all_stock(memcg);
6276 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6279 if (!reclaimed && !nr_retries--)
6283 page_counter_set_high(&memcg->memory, high);
6285 memcg_wb_domain_size_changed(memcg);
6290 static int memory_max_show(struct seq_file *m, void *v)
6292 return seq_puts_memcg_tunable(m,
6293 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6296 static ssize_t memory_max_write(struct kernfs_open_file *of,
6297 char *buf, size_t nbytes, loff_t off)
6299 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6300 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6301 bool drained = false;
6305 buf = strstrip(buf);
6306 err = page_counter_memparse(buf, "max", &max);
6310 xchg(&memcg->memory.max, max);
6313 unsigned long nr_pages = page_counter_read(&memcg->memory);
6315 if (nr_pages <= max)
6318 if (signal_pending(current))
6322 drain_all_stock(memcg);
6328 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6334 memcg_memory_event(memcg, MEMCG_OOM);
6335 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6339 memcg_wb_domain_size_changed(memcg);
6343 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6345 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6346 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6347 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6348 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6349 seq_printf(m, "oom_kill %lu\n",
6350 atomic_long_read(&events[MEMCG_OOM_KILL]));
6353 static int memory_events_show(struct seq_file *m, void *v)
6355 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6357 __memory_events_show(m, memcg->memory_events);
6361 static int memory_events_local_show(struct seq_file *m, void *v)
6363 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6365 __memory_events_show(m, memcg->memory_events_local);
6369 static int memory_stat_show(struct seq_file *m, void *v)
6371 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6374 buf = memory_stat_format(memcg);
6382 static int memory_oom_group_show(struct seq_file *m, void *v)
6384 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6386 seq_printf(m, "%d\n", memcg->oom_group);
6391 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6392 char *buf, size_t nbytes, loff_t off)
6394 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6397 buf = strstrip(buf);
6401 ret = kstrtoint(buf, 0, &oom_group);
6405 if (oom_group != 0 && oom_group != 1)
6408 memcg->oom_group = oom_group;
6413 static struct cftype memory_files[] = {
6416 .flags = CFTYPE_NOT_ON_ROOT,
6417 .read_u64 = memory_current_read,
6421 .flags = CFTYPE_NOT_ON_ROOT,
6422 .seq_show = memory_min_show,
6423 .write = memory_min_write,
6427 .flags = CFTYPE_NOT_ON_ROOT,
6428 .seq_show = memory_low_show,
6429 .write = memory_low_write,
6433 .flags = CFTYPE_NOT_ON_ROOT,
6434 .seq_show = memory_high_show,
6435 .write = memory_high_write,
6439 .flags = CFTYPE_NOT_ON_ROOT,
6440 .seq_show = memory_max_show,
6441 .write = memory_max_write,
6445 .flags = CFTYPE_NOT_ON_ROOT,
6446 .file_offset = offsetof(struct mem_cgroup, events_file),
6447 .seq_show = memory_events_show,
6450 .name = "events.local",
6451 .flags = CFTYPE_NOT_ON_ROOT,
6452 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6453 .seq_show = memory_events_local_show,
6457 .seq_show = memory_stat_show,
6460 .name = "oom.group",
6461 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6462 .seq_show = memory_oom_group_show,
6463 .write = memory_oom_group_write,
6468 struct cgroup_subsys memory_cgrp_subsys = {
6469 .css_alloc = mem_cgroup_css_alloc,
6470 .css_online = mem_cgroup_css_online,
6471 .css_offline = mem_cgroup_css_offline,
6472 .css_released = mem_cgroup_css_released,
6473 .css_free = mem_cgroup_css_free,
6474 .css_reset = mem_cgroup_css_reset,
6475 .can_attach = mem_cgroup_can_attach,
6476 .cancel_attach = mem_cgroup_cancel_attach,
6477 .post_attach = mem_cgroup_move_task,
6478 .bind = mem_cgroup_bind,
6479 .dfl_cftypes = memory_files,
6480 .legacy_cftypes = mem_cgroup_legacy_files,
6485 * This function calculates an individual cgroup's effective
6486 * protection which is derived from its own memory.min/low, its
6487 * parent's and siblings' settings, as well as the actual memory
6488 * distribution in the tree.
6490 * The following rules apply to the effective protection values:
6492 * 1. At the first level of reclaim, effective protection is equal to
6493 * the declared protection in memory.min and memory.low.
6495 * 2. To enable safe delegation of the protection configuration, at
6496 * subsequent levels the effective protection is capped to the
6497 * parent's effective protection.
6499 * 3. To make complex and dynamic subtrees easier to configure, the
6500 * user is allowed to overcommit the declared protection at a given
6501 * level. If that is the case, the parent's effective protection is
6502 * distributed to the children in proportion to how much protection
6503 * they have declared and how much of it they are utilizing.
6505 * This makes distribution proportional, but also work-conserving:
6506 * if one cgroup claims much more protection than it uses memory,
6507 * the unused remainder is available to its siblings.
6509 * 4. Conversely, when the declared protection is undercommitted at a
6510 * given level, the distribution of the larger parental protection
6511 * budget is NOT proportional. A cgroup's protection from a sibling
6512 * is capped to its own memory.min/low setting.
6514 * 5. However, to allow protecting recursive subtrees from each other
6515 * without having to declare each individual cgroup's fixed share
6516 * of the ancestor's claim to protection, any unutilized -
6517 * "floating" - protection from up the tree is distributed in
6518 * proportion to each cgroup's *usage*. This makes the protection
6519 * neutral wrt sibling cgroups and lets them compete freely over
6520 * the shared parental protection budget, but it protects the
6521 * subtree as a whole from neighboring subtrees.
6523 * Note that 4. and 5. are not in conflict: 4. is about protecting
6524 * against immediate siblings whereas 5. is about protecting against
6525 * neighboring subtrees.
6527 static unsigned long effective_protection(unsigned long usage,
6528 unsigned long parent_usage,
6529 unsigned long setting,
6530 unsigned long parent_effective,
6531 unsigned long siblings_protected)
6533 unsigned long protected;
6536 protected = min(usage, setting);
6538 * If all cgroups at this level combined claim and use more
6539 * protection then what the parent affords them, distribute
6540 * shares in proportion to utilization.
6542 * We are using actual utilization rather than the statically
6543 * claimed protection in order to be work-conserving: claimed
6544 * but unused protection is available to siblings that would
6545 * otherwise get a smaller chunk than what they claimed.
6547 if (siblings_protected > parent_effective)
6548 return protected * parent_effective / siblings_protected;
6551 * Ok, utilized protection of all children is within what the
6552 * parent affords them, so we know whatever this child claims
6553 * and utilizes is effectively protected.
6555 * If there is unprotected usage beyond this value, reclaim
6556 * will apply pressure in proportion to that amount.
6558 * If there is unutilized protection, the cgroup will be fully
6559 * shielded from reclaim, but we do return a smaller value for
6560 * protection than what the group could enjoy in theory. This
6561 * is okay. With the overcommit distribution above, effective
6562 * protection is always dependent on how memory is actually
6563 * consumed among the siblings anyway.
6568 * If the children aren't claiming (all of) the protection
6569 * afforded to them by the parent, distribute the remainder in
6570 * proportion to the (unprotected) memory of each cgroup. That
6571 * way, cgroups that aren't explicitly prioritized wrt each
6572 * other compete freely over the allowance, but they are
6573 * collectively protected from neighboring trees.
6575 * We're using unprotected memory for the weight so that if
6576 * some cgroups DO claim explicit protection, we don't protect
6577 * the same bytes twice.
6579 * Check both usage and parent_usage against the respective
6580 * protected values. One should imply the other, but they
6581 * aren't read atomically - make sure the division is sane.
6583 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6585 if (parent_effective > siblings_protected &&
6586 parent_usage > siblings_protected &&
6587 usage > protected) {
6588 unsigned long unclaimed;
6590 unclaimed = parent_effective - siblings_protected;
6591 unclaimed *= usage - protected;
6592 unclaimed /= parent_usage - siblings_protected;
6601 * mem_cgroup_protected - check if memory consumption is in the normal range
6602 * @root: the top ancestor of the sub-tree being checked
6603 * @memcg: the memory cgroup to check
6605 * WARNING: This function is not stateless! It can only be used as part
6606 * of a top-down tree iteration, not for isolated queries.
6608 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6609 struct mem_cgroup *memcg)
6611 unsigned long usage, parent_usage;
6612 struct mem_cgroup *parent;
6614 if (mem_cgroup_disabled())
6618 root = root_mem_cgroup;
6621 * Effective values of the reclaim targets are ignored so they
6622 * can be stale. Have a look at mem_cgroup_protection for more
6624 * TODO: calculation should be more robust so that we do not need
6625 * that special casing.
6630 usage = page_counter_read(&memcg->memory);
6634 parent = parent_mem_cgroup(memcg);
6635 /* No parent means a non-hierarchical mode on v1 memcg */
6639 if (parent == root) {
6640 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6641 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6645 parent_usage = page_counter_read(&parent->memory);
6647 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6648 READ_ONCE(memcg->memory.min),
6649 READ_ONCE(parent->memory.emin),
6650 atomic_long_read(&parent->memory.children_min_usage)));
6652 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6653 READ_ONCE(memcg->memory.low),
6654 READ_ONCE(parent->memory.elow),
6655 atomic_long_read(&parent->memory.children_low_usage)));
6659 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6660 * @page: page to charge
6661 * @mm: mm context of the victim
6662 * @gfp_mask: reclaim mode
6664 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6665 * pages according to @gfp_mask if necessary.
6667 * Returns 0 on success. Otherwise, an error code is returned.
6669 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6671 unsigned int nr_pages = hpage_nr_pages(page);
6672 struct mem_cgroup *memcg = NULL;
6675 if (mem_cgroup_disabled())
6678 if (PageSwapCache(page)) {
6679 swp_entry_t ent = { .val = page_private(page), };
6683 * Every swap fault against a single page tries to charge the
6684 * page, bail as early as possible. shmem_unuse() encounters
6685 * already charged pages, too. page->mem_cgroup is protected
6686 * by the page lock, which serializes swap cache removal, which
6687 * in turn serializes uncharging.
6689 VM_BUG_ON_PAGE(!PageLocked(page), page);
6690 if (compound_head(page)->mem_cgroup)
6693 id = lookup_swap_cgroup_id(ent);
6695 memcg = mem_cgroup_from_id(id);
6696 if (memcg && !css_tryget_online(&memcg->css))
6702 memcg = get_mem_cgroup_from_mm(mm);
6704 ret = try_charge(memcg, gfp_mask, nr_pages);
6708 css_get(&memcg->css);
6709 commit_charge(page, memcg);
6711 local_irq_disable();
6712 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6713 memcg_check_events(memcg, page);
6716 if (PageSwapCache(page)) {
6717 swp_entry_t entry = { .val = page_private(page) };
6719 * The swap entry might not get freed for a long time,
6720 * let's not wait for it. The page already received a
6721 * memory+swap charge, drop the swap entry duplicate.
6723 mem_cgroup_uncharge_swap(entry, nr_pages);
6727 css_put(&memcg->css);
6732 struct uncharge_gather {
6733 struct mem_cgroup *memcg;
6734 unsigned long nr_pages;
6735 unsigned long pgpgout;
6736 unsigned long nr_kmem;
6737 struct page *dummy_page;
6740 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6742 memset(ug, 0, sizeof(*ug));
6745 static void uncharge_batch(const struct uncharge_gather *ug)
6747 unsigned long flags;
6749 if (!mem_cgroup_is_root(ug->memcg)) {
6750 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6751 if (do_memsw_account())
6752 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6753 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6754 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6755 memcg_oom_recover(ug->memcg);
6758 local_irq_save(flags);
6759 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6760 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6761 memcg_check_events(ug->memcg, ug->dummy_page);
6762 local_irq_restore(flags);
6765 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6767 unsigned long nr_pages;
6769 VM_BUG_ON_PAGE(PageLRU(page), page);
6771 if (!page->mem_cgroup)
6775 * Nobody should be changing or seriously looking at
6776 * page->mem_cgroup at this point, we have fully
6777 * exclusive access to the page.
6780 if (ug->memcg != page->mem_cgroup) {
6783 uncharge_gather_clear(ug);
6785 ug->memcg = page->mem_cgroup;
6788 nr_pages = compound_nr(page);
6789 ug->nr_pages += nr_pages;
6791 if (!PageKmemcg(page)) {
6794 ug->nr_kmem += nr_pages;
6795 __ClearPageKmemcg(page);
6798 ug->dummy_page = page;
6799 page->mem_cgroup = NULL;
6800 css_put(&ug->memcg->css);
6803 static void uncharge_list(struct list_head *page_list)
6805 struct uncharge_gather ug;
6806 struct list_head *next;
6808 uncharge_gather_clear(&ug);
6811 * Note that the list can be a single page->lru; hence the
6812 * do-while loop instead of a simple list_for_each_entry().
6814 next = page_list->next;
6818 page = list_entry(next, struct page, lru);
6819 next = page->lru.next;
6821 uncharge_page(page, &ug);
6822 } while (next != page_list);
6825 uncharge_batch(&ug);
6829 * mem_cgroup_uncharge - uncharge a page
6830 * @page: page to uncharge
6832 * Uncharge a page previously charged with mem_cgroup_charge().
6834 void mem_cgroup_uncharge(struct page *page)
6836 struct uncharge_gather ug;
6838 if (mem_cgroup_disabled())
6841 /* Don't touch page->lru of any random page, pre-check: */
6842 if (!page->mem_cgroup)
6845 uncharge_gather_clear(&ug);
6846 uncharge_page(page, &ug);
6847 uncharge_batch(&ug);
6851 * mem_cgroup_uncharge_list - uncharge a list of page
6852 * @page_list: list of pages to uncharge
6854 * Uncharge a list of pages previously charged with
6855 * mem_cgroup_charge().
6857 void mem_cgroup_uncharge_list(struct list_head *page_list)
6859 if (mem_cgroup_disabled())
6862 if (!list_empty(page_list))
6863 uncharge_list(page_list);
6867 * mem_cgroup_migrate - charge a page's replacement
6868 * @oldpage: currently circulating page
6869 * @newpage: replacement page
6871 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6872 * be uncharged upon free.
6874 * Both pages must be locked, @newpage->mapping must be set up.
6876 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6878 struct mem_cgroup *memcg;
6879 unsigned int nr_pages;
6880 unsigned long flags;
6882 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6883 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6884 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6885 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6888 if (mem_cgroup_disabled())
6891 /* Page cache replacement: new page already charged? */
6892 if (newpage->mem_cgroup)
6895 /* Swapcache readahead pages can get replaced before being charged */
6896 memcg = oldpage->mem_cgroup;
6900 /* Force-charge the new page. The old one will be freed soon */
6901 nr_pages = hpage_nr_pages(newpage);
6903 page_counter_charge(&memcg->memory, nr_pages);
6904 if (do_memsw_account())
6905 page_counter_charge(&memcg->memsw, nr_pages);
6907 css_get(&memcg->css);
6908 commit_charge(newpage, memcg);
6910 local_irq_save(flags);
6911 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6912 memcg_check_events(memcg, newpage);
6913 local_irq_restore(flags);
6916 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6917 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6919 void mem_cgroup_sk_alloc(struct sock *sk)
6921 struct mem_cgroup *memcg;
6923 if (!mem_cgroup_sockets_enabled)
6926 /* Do not associate the sock with unrelated interrupted task's memcg. */
6931 memcg = mem_cgroup_from_task(current);
6932 if (memcg == root_mem_cgroup)
6934 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6936 if (css_tryget(&memcg->css))
6937 sk->sk_memcg = memcg;
6942 void mem_cgroup_sk_free(struct sock *sk)
6945 css_put(&sk->sk_memcg->css);
6949 * mem_cgroup_charge_skmem - charge socket memory
6950 * @memcg: memcg to charge
6951 * @nr_pages: number of pages to charge
6953 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6954 * @memcg's configured limit, %false if the charge had to be forced.
6956 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6958 gfp_t gfp_mask = GFP_KERNEL;
6960 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6961 struct page_counter *fail;
6963 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6964 memcg->tcpmem_pressure = 0;
6967 page_counter_charge(&memcg->tcpmem, nr_pages);
6968 memcg->tcpmem_pressure = 1;
6972 /* Don't block in the packet receive path */
6974 gfp_mask = GFP_NOWAIT;
6976 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6978 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6981 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6986 * mem_cgroup_uncharge_skmem - uncharge socket memory
6987 * @memcg: memcg to uncharge
6988 * @nr_pages: number of pages to uncharge
6990 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6992 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6993 page_counter_uncharge(&memcg->tcpmem, nr_pages);
6997 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
6999 refill_stock(memcg, nr_pages);
7002 static int __init cgroup_memory(char *s)
7006 while ((token = strsep(&s, ",")) != NULL) {
7009 if (!strcmp(token, "nosocket"))
7010 cgroup_memory_nosocket = true;
7011 if (!strcmp(token, "nokmem"))
7012 cgroup_memory_nokmem = true;
7016 __setup("cgroup.memory=", cgroup_memory);
7019 * subsys_initcall() for memory controller.
7021 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7022 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7023 * basically everything that doesn't depend on a specific mem_cgroup structure
7024 * should be initialized from here.
7026 static int __init mem_cgroup_init(void)
7030 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7031 memcg_hotplug_cpu_dead);
7033 for_each_possible_cpu(cpu)
7034 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7037 for_each_node(node) {
7038 struct mem_cgroup_tree_per_node *rtpn;
7040 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7041 node_online(node) ? node : NUMA_NO_NODE);
7043 rtpn->rb_root = RB_ROOT;
7044 rtpn->rb_rightmost = NULL;
7045 spin_lock_init(&rtpn->lock);
7046 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7051 subsys_initcall(mem_cgroup_init);
7053 #ifdef CONFIG_MEMCG_SWAP
7054 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7056 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7058 * The root cgroup cannot be destroyed, so it's refcount must
7061 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7065 memcg = parent_mem_cgroup(memcg);
7067 memcg = root_mem_cgroup;
7073 * mem_cgroup_swapout - transfer a memsw charge to swap
7074 * @page: page whose memsw charge to transfer
7075 * @entry: swap entry to move the charge to
7077 * Transfer the memsw charge of @page to @entry.
7079 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7081 struct mem_cgroup *memcg, *swap_memcg;
7082 unsigned int nr_entries;
7083 unsigned short oldid;
7085 VM_BUG_ON_PAGE(PageLRU(page), page);
7086 VM_BUG_ON_PAGE(page_count(page), page);
7088 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7091 memcg = page->mem_cgroup;
7093 /* Readahead page, never charged */
7098 * In case the memcg owning these pages has been offlined and doesn't
7099 * have an ID allocated to it anymore, charge the closest online
7100 * ancestor for the swap instead and transfer the memory+swap charge.
7102 swap_memcg = mem_cgroup_id_get_online(memcg);
7103 nr_entries = hpage_nr_pages(page);
7104 /* Get references for the tail pages, too */
7106 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7107 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7109 VM_BUG_ON_PAGE(oldid, page);
7110 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7112 page->mem_cgroup = NULL;
7114 if (!mem_cgroup_is_root(memcg))
7115 page_counter_uncharge(&memcg->memory, nr_entries);
7117 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7118 if (!mem_cgroup_is_root(swap_memcg))
7119 page_counter_charge(&swap_memcg->memsw, nr_entries);
7120 page_counter_uncharge(&memcg->memsw, nr_entries);
7124 * Interrupts should be disabled here because the caller holds the
7125 * i_pages lock which is taken with interrupts-off. It is
7126 * important here to have the interrupts disabled because it is the
7127 * only synchronisation we have for updating the per-CPU variables.
7129 VM_BUG_ON(!irqs_disabled());
7130 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7131 memcg_check_events(memcg, page);
7133 css_put(&memcg->css);
7137 * mem_cgroup_try_charge_swap - try charging swap space for a page
7138 * @page: page being added to swap
7139 * @entry: swap entry to charge
7141 * Try to charge @page's memcg for the swap space at @entry.
7143 * Returns 0 on success, -ENOMEM on failure.
7145 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7147 unsigned int nr_pages = hpage_nr_pages(page);
7148 struct page_counter *counter;
7149 struct mem_cgroup *memcg;
7150 unsigned short oldid;
7152 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7155 memcg = page->mem_cgroup;
7157 /* Readahead page, never charged */
7162 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7166 memcg = mem_cgroup_id_get_online(memcg);
7168 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7169 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7170 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7171 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7172 mem_cgroup_id_put(memcg);
7176 /* Get references for the tail pages, too */
7178 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7179 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7180 VM_BUG_ON_PAGE(oldid, page);
7181 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7187 * mem_cgroup_uncharge_swap - uncharge swap space
7188 * @entry: swap entry to uncharge
7189 * @nr_pages: the amount of swap space to uncharge
7191 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7193 struct mem_cgroup *memcg;
7196 id = swap_cgroup_record(entry, 0, nr_pages);
7198 memcg = mem_cgroup_from_id(id);
7200 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7201 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7202 page_counter_uncharge(&memcg->swap, nr_pages);
7204 page_counter_uncharge(&memcg->memsw, nr_pages);
7206 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7207 mem_cgroup_id_put_many(memcg, nr_pages);
7212 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7214 long nr_swap_pages = get_nr_swap_pages();
7216 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7217 return nr_swap_pages;
7218 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7219 nr_swap_pages = min_t(long, nr_swap_pages,
7220 READ_ONCE(memcg->swap.max) -
7221 page_counter_read(&memcg->swap));
7222 return nr_swap_pages;
7225 bool mem_cgroup_swap_full(struct page *page)
7227 struct mem_cgroup *memcg;
7229 VM_BUG_ON_PAGE(!PageLocked(page), page);
7233 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7236 memcg = page->mem_cgroup;
7240 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7241 unsigned long usage = page_counter_read(&memcg->swap);
7243 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7244 usage * 2 >= READ_ONCE(memcg->swap.max))
7251 static int __init setup_swap_account(char *s)
7253 if (!strcmp(s, "1"))
7254 cgroup_memory_noswap = 0;
7255 else if (!strcmp(s, "0"))
7256 cgroup_memory_noswap = 1;
7259 __setup("swapaccount=", setup_swap_account);
7261 static u64 swap_current_read(struct cgroup_subsys_state *css,
7264 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7266 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7269 static int swap_high_show(struct seq_file *m, void *v)
7271 return seq_puts_memcg_tunable(m,
7272 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7275 static ssize_t swap_high_write(struct kernfs_open_file *of,
7276 char *buf, size_t nbytes, loff_t off)
7278 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7282 buf = strstrip(buf);
7283 err = page_counter_memparse(buf, "max", &high);
7287 page_counter_set_high(&memcg->swap, high);
7292 static int swap_max_show(struct seq_file *m, void *v)
7294 return seq_puts_memcg_tunable(m,
7295 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7298 static ssize_t swap_max_write(struct kernfs_open_file *of,
7299 char *buf, size_t nbytes, loff_t off)
7301 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7305 buf = strstrip(buf);
7306 err = page_counter_memparse(buf, "max", &max);
7310 xchg(&memcg->swap.max, max);
7315 static int swap_events_show(struct seq_file *m, void *v)
7317 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7319 seq_printf(m, "high %lu\n",
7320 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7321 seq_printf(m, "max %lu\n",
7322 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7323 seq_printf(m, "fail %lu\n",
7324 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7329 static struct cftype swap_files[] = {
7331 .name = "swap.current",
7332 .flags = CFTYPE_NOT_ON_ROOT,
7333 .read_u64 = swap_current_read,
7336 .name = "swap.high",
7337 .flags = CFTYPE_NOT_ON_ROOT,
7338 .seq_show = swap_high_show,
7339 .write = swap_high_write,
7343 .flags = CFTYPE_NOT_ON_ROOT,
7344 .seq_show = swap_max_show,
7345 .write = swap_max_write,
7348 .name = "swap.events",
7349 .flags = CFTYPE_NOT_ON_ROOT,
7350 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7351 .seq_show = swap_events_show,
7356 static struct cftype memsw_files[] = {
7358 .name = "memsw.usage_in_bytes",
7359 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7360 .read_u64 = mem_cgroup_read_u64,
7363 .name = "memsw.max_usage_in_bytes",
7364 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7365 .write = mem_cgroup_reset,
7366 .read_u64 = mem_cgroup_read_u64,
7369 .name = "memsw.limit_in_bytes",
7370 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7371 .write = mem_cgroup_write,
7372 .read_u64 = mem_cgroup_read_u64,
7375 .name = "memsw.failcnt",
7376 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7377 .write = mem_cgroup_reset,
7378 .read_u64 = mem_cgroup_read_u64,
7380 { }, /* terminate */
7384 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7385 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7386 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7387 * boot parameter. This may result in premature OOPS inside
7388 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7390 static int __init mem_cgroup_swap_init(void)
7392 /* No memory control -> no swap control */
7393 if (mem_cgroup_disabled())
7394 cgroup_memory_noswap = true;
7396 if (cgroup_memory_noswap)
7399 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7400 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7404 core_initcall(mem_cgroup_swap_init);
7406 #endif /* CONFIG_MEMCG_SWAP */