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 /* Active memory cgroup to use from an interrupt context */
77 DEFINE_PER_CPU(struct mem_cgroup *, int_active_memcg);
79 /* Socket memory accounting disabled? */
80 static bool cgroup_memory_nosocket;
82 /* Kernel memory accounting disabled? */
83 static bool cgroup_memory_nokmem;
85 /* Whether the swap controller is active */
86 #ifdef CONFIG_MEMCG_SWAP
87 bool cgroup_memory_noswap __read_mostly;
89 #define cgroup_memory_noswap 1
92 #ifdef CONFIG_CGROUP_WRITEBACK
93 static DECLARE_WAIT_QUEUE_HEAD(memcg_cgwb_frn_waitq);
96 /* Whether legacy memory+swap accounting is active */
97 static bool do_memsw_account(void)
99 return !cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_noswap;
102 #define THRESHOLDS_EVENTS_TARGET 128
103 #define SOFTLIMIT_EVENTS_TARGET 1024
106 * Cgroups above their limits are maintained in a RB-Tree, independent of
107 * their hierarchy representation
110 struct mem_cgroup_tree_per_node {
111 struct rb_root rb_root;
112 struct rb_node *rb_rightmost;
116 struct mem_cgroup_tree {
117 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
120 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
123 struct mem_cgroup_eventfd_list {
124 struct list_head list;
125 struct eventfd_ctx *eventfd;
129 * cgroup_event represents events which userspace want to receive.
131 struct mem_cgroup_event {
133 * memcg which the event belongs to.
135 struct mem_cgroup *memcg;
137 * eventfd to signal userspace about the event.
139 struct eventfd_ctx *eventfd;
141 * Each of these stored in a list by the cgroup.
143 struct list_head list;
145 * register_event() callback will be used to add new userspace
146 * waiter for changes related to this event. Use eventfd_signal()
147 * on eventfd to send notification to userspace.
149 int (*register_event)(struct mem_cgroup *memcg,
150 struct eventfd_ctx *eventfd, const char *args);
152 * unregister_event() callback will be called when userspace closes
153 * the eventfd or on cgroup removing. This callback must be set,
154 * if you want provide notification functionality.
156 void (*unregister_event)(struct mem_cgroup *memcg,
157 struct eventfd_ctx *eventfd);
159 * All fields below needed to unregister event when
160 * userspace closes eventfd.
163 wait_queue_head_t *wqh;
164 wait_queue_entry_t wait;
165 struct work_struct remove;
168 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
169 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
171 /* Stuffs for move charges at task migration. */
173 * Types of charges to be moved.
175 #define MOVE_ANON 0x1U
176 #define MOVE_FILE 0x2U
177 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
179 /* "mc" and its members are protected by cgroup_mutex */
180 static struct move_charge_struct {
181 spinlock_t lock; /* for from, to */
182 struct mm_struct *mm;
183 struct mem_cgroup *from;
184 struct mem_cgroup *to;
186 unsigned long precharge;
187 unsigned long moved_charge;
188 unsigned long moved_swap;
189 struct task_struct *moving_task; /* a task moving charges */
190 wait_queue_head_t waitq; /* a waitq for other context */
192 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
193 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
197 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
198 * limit reclaim to prevent infinite loops, if they ever occur.
200 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
201 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
203 /* for encoding cft->private value on file */
212 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
213 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
214 #define MEMFILE_ATTR(val) ((val) & 0xffff)
215 /* Used for OOM nofiier */
216 #define OOM_CONTROL (0)
219 * Iteration constructs for visiting all cgroups (under a tree). If
220 * loops are exited prematurely (break), mem_cgroup_iter_break() must
221 * be used for reference counting.
223 #define for_each_mem_cgroup_tree(iter, root) \
224 for (iter = mem_cgroup_iter(root, NULL, NULL); \
226 iter = mem_cgroup_iter(root, iter, NULL))
228 #define for_each_mem_cgroup(iter) \
229 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
231 iter = mem_cgroup_iter(NULL, iter, NULL))
233 static inline bool should_force_charge(void)
235 return tsk_is_oom_victim(current) || fatal_signal_pending(current) ||
236 (current->flags & PF_EXITING);
239 /* Some nice accessors for the vmpressure. */
240 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
243 memcg = root_mem_cgroup;
244 return &memcg->vmpressure;
247 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
249 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
252 #ifdef CONFIG_MEMCG_KMEM
253 extern spinlock_t css_set_lock;
255 static void obj_cgroup_release(struct percpu_ref *ref)
257 struct obj_cgroup *objcg = container_of(ref, struct obj_cgroup, refcnt);
258 struct mem_cgroup *memcg;
259 unsigned int nr_bytes;
260 unsigned int nr_pages;
264 * At this point all allocated objects are freed, and
265 * objcg->nr_charged_bytes can't have an arbitrary byte value.
266 * However, it can be PAGE_SIZE or (x * PAGE_SIZE).
268 * The following sequence can lead to it:
269 * 1) CPU0: objcg == stock->cached_objcg
270 * 2) CPU1: we do a small allocation (e.g. 92 bytes),
271 * PAGE_SIZE bytes are charged
272 * 3) CPU1: a process from another memcg is allocating something,
273 * the stock if flushed,
274 * objcg->nr_charged_bytes = PAGE_SIZE - 92
275 * 5) CPU0: we do release this object,
276 * 92 bytes are added to stock->nr_bytes
277 * 6) CPU0: stock is flushed,
278 * 92 bytes are added to objcg->nr_charged_bytes
280 * In the result, nr_charged_bytes == PAGE_SIZE.
281 * This page will be uncharged in obj_cgroup_release().
283 nr_bytes = atomic_read(&objcg->nr_charged_bytes);
284 WARN_ON_ONCE(nr_bytes & (PAGE_SIZE - 1));
285 nr_pages = nr_bytes >> PAGE_SHIFT;
287 spin_lock_irqsave(&css_set_lock, flags);
288 memcg = obj_cgroup_memcg(objcg);
290 __memcg_kmem_uncharge(memcg, nr_pages);
291 list_del(&objcg->list);
292 mem_cgroup_put(memcg);
293 spin_unlock_irqrestore(&css_set_lock, flags);
295 percpu_ref_exit(ref);
296 kfree_rcu(objcg, rcu);
299 static struct obj_cgroup *obj_cgroup_alloc(void)
301 struct obj_cgroup *objcg;
304 objcg = kzalloc(sizeof(struct obj_cgroup), GFP_KERNEL);
308 ret = percpu_ref_init(&objcg->refcnt, obj_cgroup_release, 0,
314 INIT_LIST_HEAD(&objcg->list);
318 static void memcg_reparent_objcgs(struct mem_cgroup *memcg,
319 struct mem_cgroup *parent)
321 struct obj_cgroup *objcg, *iter;
323 objcg = rcu_replace_pointer(memcg->objcg, NULL, true);
325 spin_lock_irq(&css_set_lock);
327 /* Move active objcg to the parent's list */
328 xchg(&objcg->memcg, parent);
329 css_get(&parent->css);
330 list_add(&objcg->list, &parent->objcg_list);
332 /* Move already reparented objcgs to the parent's list */
333 list_for_each_entry(iter, &memcg->objcg_list, list) {
334 css_get(&parent->css);
335 xchg(&iter->memcg, parent);
336 css_put(&memcg->css);
338 list_splice(&memcg->objcg_list, &parent->objcg_list);
340 spin_unlock_irq(&css_set_lock);
342 percpu_ref_kill(&objcg->refcnt);
346 * This will be used as a shrinker list's index.
347 * The main reason for not using cgroup id for this:
348 * this works better in sparse environments, where we have a lot of memcgs,
349 * but only a few kmem-limited. Or also, if we have, for instance, 200
350 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
351 * 200 entry array for that.
353 * The current size of the caches array is stored in memcg_nr_cache_ids. It
354 * will double each time we have to increase it.
356 static DEFINE_IDA(memcg_cache_ida);
357 int memcg_nr_cache_ids;
359 /* Protects memcg_nr_cache_ids */
360 static DECLARE_RWSEM(memcg_cache_ids_sem);
362 void memcg_get_cache_ids(void)
364 down_read(&memcg_cache_ids_sem);
367 void memcg_put_cache_ids(void)
369 up_read(&memcg_cache_ids_sem);
373 * MIN_SIZE is different than 1, because we would like to avoid going through
374 * the alloc/free process all the time. In a small machine, 4 kmem-limited
375 * cgroups is a reasonable guess. In the future, it could be a parameter or
376 * tunable, but that is strictly not necessary.
378 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
379 * this constant directly from cgroup, but it is understandable that this is
380 * better kept as an internal representation in cgroup.c. In any case, the
381 * cgrp_id space is not getting any smaller, and we don't have to necessarily
382 * increase ours as well if it increases.
384 #define MEMCG_CACHES_MIN_SIZE 4
385 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
388 * A lot of the calls to the cache allocation functions are expected to be
389 * inlined by the compiler. Since the calls to memcg_slab_pre_alloc_hook() are
390 * conditional to this static branch, we'll have to allow modules that does
391 * kmem_cache_alloc and the such to see this symbol as well
393 DEFINE_STATIC_KEY_FALSE(memcg_kmem_enabled_key);
394 EXPORT_SYMBOL(memcg_kmem_enabled_key);
397 static int memcg_shrinker_map_size;
398 static DEFINE_MUTEX(memcg_shrinker_map_mutex);
400 static void memcg_free_shrinker_map_rcu(struct rcu_head *head)
402 kvfree(container_of(head, struct memcg_shrinker_map, rcu));
405 static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
406 int size, int old_size)
408 struct memcg_shrinker_map *new, *old;
411 lockdep_assert_held(&memcg_shrinker_map_mutex);
414 old = rcu_dereference_protected(
415 mem_cgroup_nodeinfo(memcg, nid)->shrinker_map, true);
416 /* Not yet online memcg */
420 new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
424 /* Set all old bits, clear all new bits */
425 memset(new->map, (int)0xff, old_size);
426 memset((void *)new->map + old_size, 0, size - old_size);
428 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, new);
429 call_rcu(&old->rcu, memcg_free_shrinker_map_rcu);
435 static void memcg_free_shrinker_maps(struct mem_cgroup *memcg)
437 struct mem_cgroup_per_node *pn;
438 struct memcg_shrinker_map *map;
441 if (mem_cgroup_is_root(memcg))
445 pn = mem_cgroup_nodeinfo(memcg, nid);
446 map = rcu_dereference_protected(pn->shrinker_map, true);
449 rcu_assign_pointer(pn->shrinker_map, NULL);
453 static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
455 struct memcg_shrinker_map *map;
456 int nid, size, ret = 0;
458 if (mem_cgroup_is_root(memcg))
461 mutex_lock(&memcg_shrinker_map_mutex);
462 size = memcg_shrinker_map_size;
464 map = kvzalloc_node(sizeof(*map) + size, GFP_KERNEL, nid);
466 memcg_free_shrinker_maps(memcg);
470 rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_map, map);
472 mutex_unlock(&memcg_shrinker_map_mutex);
477 int memcg_expand_shrinker_maps(int new_id)
479 int size, old_size, ret = 0;
480 struct mem_cgroup *memcg;
482 size = DIV_ROUND_UP(new_id + 1, BITS_PER_LONG) * sizeof(unsigned long);
483 old_size = memcg_shrinker_map_size;
484 if (size <= old_size)
487 mutex_lock(&memcg_shrinker_map_mutex);
488 if (!root_mem_cgroup)
491 for_each_mem_cgroup(memcg) {
492 if (mem_cgroup_is_root(memcg))
494 ret = memcg_expand_one_shrinker_map(memcg, size, old_size);
496 mem_cgroup_iter_break(NULL, memcg);
502 memcg_shrinker_map_size = size;
503 mutex_unlock(&memcg_shrinker_map_mutex);
507 void memcg_set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id)
509 if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) {
510 struct memcg_shrinker_map *map;
513 map = rcu_dereference(memcg->nodeinfo[nid]->shrinker_map);
514 /* Pairs with smp mb in shrink_slab() */
515 smp_mb__before_atomic();
516 set_bit(shrinker_id, map->map);
522 * mem_cgroup_css_from_page - css of the memcg associated with a page
523 * @page: page of interest
525 * If memcg is bound to the default hierarchy, css of the memcg associated
526 * with @page is returned. The returned css remains associated with @page
527 * until it is released.
529 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
532 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
534 struct mem_cgroup *memcg;
536 memcg = page->mem_cgroup;
538 if (!memcg || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
539 memcg = root_mem_cgroup;
545 * page_cgroup_ino - return inode number of the memcg a page is charged to
548 * Look up the closest online ancestor of the memory cgroup @page is charged to
549 * and return its inode number or 0 if @page is not charged to any cgroup. It
550 * is safe to call this function without holding a reference to @page.
552 * Note, this function is inherently racy, because there is nothing to prevent
553 * the cgroup inode from getting torn down and potentially reallocated a moment
554 * after page_cgroup_ino() returns, so it only should be used by callers that
555 * do not care (such as procfs interfaces).
557 ino_t page_cgroup_ino(struct page *page)
559 struct mem_cgroup *memcg;
560 unsigned long ino = 0;
563 memcg = page->mem_cgroup;
566 * The lowest bit set means that memcg isn't a valid
567 * memcg pointer, but a obj_cgroups pointer.
568 * In this case the page is shared and doesn't belong
569 * to any specific memory cgroup.
571 if ((unsigned long) memcg & 0x1UL)
574 while (memcg && !(memcg->css.flags & CSS_ONLINE))
575 memcg = parent_mem_cgroup(memcg);
577 ino = cgroup_ino(memcg->css.cgroup);
582 static struct mem_cgroup_per_node *
583 mem_cgroup_page_nodeinfo(struct mem_cgroup *memcg, struct page *page)
585 int nid = page_to_nid(page);
587 return memcg->nodeinfo[nid];
590 static struct mem_cgroup_tree_per_node *
591 soft_limit_tree_node(int nid)
593 return soft_limit_tree.rb_tree_per_node[nid];
596 static struct mem_cgroup_tree_per_node *
597 soft_limit_tree_from_page(struct page *page)
599 int nid = page_to_nid(page);
601 return soft_limit_tree.rb_tree_per_node[nid];
604 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_node *mz,
605 struct mem_cgroup_tree_per_node *mctz,
606 unsigned long new_usage_in_excess)
608 struct rb_node **p = &mctz->rb_root.rb_node;
609 struct rb_node *parent = NULL;
610 struct mem_cgroup_per_node *mz_node;
611 bool rightmost = true;
616 mz->usage_in_excess = new_usage_in_excess;
617 if (!mz->usage_in_excess)
621 mz_node = rb_entry(parent, struct mem_cgroup_per_node,
623 if (mz->usage_in_excess < mz_node->usage_in_excess) {
632 mctz->rb_rightmost = &mz->tree_node;
634 rb_link_node(&mz->tree_node, parent, p);
635 rb_insert_color(&mz->tree_node, &mctz->rb_root);
639 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
640 struct mem_cgroup_tree_per_node *mctz)
645 if (&mz->tree_node == mctz->rb_rightmost)
646 mctz->rb_rightmost = rb_prev(&mz->tree_node);
648 rb_erase(&mz->tree_node, &mctz->rb_root);
652 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_node *mz,
653 struct mem_cgroup_tree_per_node *mctz)
657 spin_lock_irqsave(&mctz->lock, flags);
658 __mem_cgroup_remove_exceeded(mz, mctz);
659 spin_unlock_irqrestore(&mctz->lock, flags);
662 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
664 unsigned long nr_pages = page_counter_read(&memcg->memory);
665 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
666 unsigned long excess = 0;
668 if (nr_pages > soft_limit)
669 excess = nr_pages - soft_limit;
674 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
676 unsigned long excess;
677 struct mem_cgroup_per_node *mz;
678 struct mem_cgroup_tree_per_node *mctz;
680 mctz = soft_limit_tree_from_page(page);
684 * Necessary to update all ancestors when hierarchy is used.
685 * because their event counter is not touched.
687 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
688 mz = mem_cgroup_page_nodeinfo(memcg, page);
689 excess = soft_limit_excess(memcg);
691 * We have to update the tree if mz is on RB-tree or
692 * mem is over its softlimit.
694 if (excess || mz->on_tree) {
697 spin_lock_irqsave(&mctz->lock, flags);
698 /* if on-tree, remove it */
700 __mem_cgroup_remove_exceeded(mz, mctz);
702 * Insert again. mz->usage_in_excess will be updated.
703 * If excess is 0, no tree ops.
705 __mem_cgroup_insert_exceeded(mz, mctz, excess);
706 spin_unlock_irqrestore(&mctz->lock, flags);
711 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
713 struct mem_cgroup_tree_per_node *mctz;
714 struct mem_cgroup_per_node *mz;
718 mz = mem_cgroup_nodeinfo(memcg, nid);
719 mctz = soft_limit_tree_node(nid);
721 mem_cgroup_remove_exceeded(mz, mctz);
725 static struct mem_cgroup_per_node *
726 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
728 struct mem_cgroup_per_node *mz;
732 if (!mctz->rb_rightmost)
733 goto done; /* Nothing to reclaim from */
735 mz = rb_entry(mctz->rb_rightmost,
736 struct mem_cgroup_per_node, tree_node);
738 * Remove the node now but someone else can add it back,
739 * we will to add it back at the end of reclaim to its correct
740 * position in the tree.
742 __mem_cgroup_remove_exceeded(mz, mctz);
743 if (!soft_limit_excess(mz->memcg) ||
744 !css_tryget(&mz->memcg->css))
750 static struct mem_cgroup_per_node *
751 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_node *mctz)
753 struct mem_cgroup_per_node *mz;
755 spin_lock_irq(&mctz->lock);
756 mz = __mem_cgroup_largest_soft_limit_node(mctz);
757 spin_unlock_irq(&mctz->lock);
762 * __mod_memcg_state - update cgroup memory statistics
763 * @memcg: the memory cgroup
764 * @idx: the stat item - can be enum memcg_stat_item or enum node_stat_item
765 * @val: delta to add to the counter, can be negative
767 void __mod_memcg_state(struct mem_cgroup *memcg, int idx, int val)
769 long x, threshold = MEMCG_CHARGE_BATCH;
771 if (mem_cgroup_disabled())
774 if (memcg_stat_item_in_bytes(idx))
775 threshold <<= PAGE_SHIFT;
777 x = val + __this_cpu_read(memcg->vmstats_percpu->stat[idx]);
778 if (unlikely(abs(x) > threshold)) {
779 struct mem_cgroup *mi;
782 * Batch local counters to keep them in sync with
783 * the hierarchical ones.
785 __this_cpu_add(memcg->vmstats_local->stat[idx], x);
786 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
787 atomic_long_add(x, &mi->vmstats[idx]);
790 __this_cpu_write(memcg->vmstats_percpu->stat[idx], x);
793 static struct mem_cgroup_per_node *
794 parent_nodeinfo(struct mem_cgroup_per_node *pn, int nid)
796 struct mem_cgroup *parent;
798 parent = parent_mem_cgroup(pn->memcg);
801 return mem_cgroup_nodeinfo(parent, nid);
804 void __mod_memcg_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
807 struct mem_cgroup_per_node *pn;
808 struct mem_cgroup *memcg;
809 long x, threshold = MEMCG_CHARGE_BATCH;
811 pn = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
815 __mod_memcg_state(memcg, idx, val);
818 __this_cpu_add(pn->lruvec_stat_local->count[idx], val);
820 if (vmstat_item_in_bytes(idx))
821 threshold <<= PAGE_SHIFT;
823 x = val + __this_cpu_read(pn->lruvec_stat_cpu->count[idx]);
824 if (unlikely(abs(x) > threshold)) {
825 pg_data_t *pgdat = lruvec_pgdat(lruvec);
826 struct mem_cgroup_per_node *pi;
828 for (pi = pn; pi; pi = parent_nodeinfo(pi, pgdat->node_id))
829 atomic_long_add(x, &pi->lruvec_stat[idx]);
832 __this_cpu_write(pn->lruvec_stat_cpu->count[idx], x);
836 * __mod_lruvec_state - update lruvec memory statistics
837 * @lruvec: the lruvec
838 * @idx: the stat item
839 * @val: delta to add to the counter, can be negative
841 * The lruvec is the intersection of the NUMA node and a cgroup. This
842 * function updates the all three counters that are affected by a
843 * change of state at this level: per-node, per-cgroup, per-lruvec.
845 void __mod_lruvec_state(struct lruvec *lruvec, enum node_stat_item idx,
849 __mod_node_page_state(lruvec_pgdat(lruvec), idx, val);
851 /* Update memcg and lruvec */
852 if (!mem_cgroup_disabled())
853 __mod_memcg_lruvec_state(lruvec, idx, val);
856 void __mod_lruvec_page_state(struct page *page, enum node_stat_item idx,
859 struct page *head = compound_head(page); /* rmap on tail pages */
860 pg_data_t *pgdat = page_pgdat(page);
861 struct lruvec *lruvec;
863 /* Untracked pages have no memcg, no lruvec. Update only the node */
864 if (!head->mem_cgroup) {
865 __mod_node_page_state(pgdat, idx, val);
869 lruvec = mem_cgroup_lruvec(head->mem_cgroup, pgdat);
870 __mod_lruvec_state(lruvec, idx, val);
872 EXPORT_SYMBOL(__mod_lruvec_page_state);
874 void __mod_lruvec_kmem_state(void *p, enum node_stat_item idx, int val)
876 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
877 struct mem_cgroup *memcg;
878 struct lruvec *lruvec;
881 memcg = mem_cgroup_from_obj(p);
884 * Untracked pages have no memcg, no lruvec. Update only the
885 * node. If we reparent the slab objects to the root memcg,
886 * when we free the slab object, we need to update the per-memcg
887 * vmstats to keep it correct for the root memcg.
890 __mod_node_page_state(pgdat, idx, val);
892 lruvec = mem_cgroup_lruvec(memcg, pgdat);
893 __mod_lruvec_state(lruvec, idx, val);
899 * __count_memcg_events - account VM events in a cgroup
900 * @memcg: the memory cgroup
901 * @idx: the event item
902 * @count: the number of events that occured
904 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
909 if (mem_cgroup_disabled())
912 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
913 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
914 struct mem_cgroup *mi;
917 * Batch local counters to keep them in sync with
918 * the hierarchical ones.
920 __this_cpu_add(memcg->vmstats_local->events[idx], x);
921 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
922 atomic_long_add(x, &mi->vmevents[idx]);
925 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
928 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
930 return atomic_long_read(&memcg->vmevents[event]);
933 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
938 for_each_possible_cpu(cpu)
939 x += per_cpu(memcg->vmstats_local->events[event], cpu);
943 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
947 /* pagein of a big page is an event. So, ignore page size */
949 __count_memcg_events(memcg, PGPGIN, 1);
951 __count_memcg_events(memcg, PGPGOUT, 1);
952 nr_pages = -nr_pages; /* for event */
955 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
958 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
959 enum mem_cgroup_events_target target)
961 unsigned long val, next;
963 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
964 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
965 /* from time_after() in jiffies.h */
966 if ((long)(next - val) < 0) {
968 case MEM_CGROUP_TARGET_THRESH:
969 next = val + THRESHOLDS_EVENTS_TARGET;
971 case MEM_CGROUP_TARGET_SOFTLIMIT:
972 next = val + SOFTLIMIT_EVENTS_TARGET;
977 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
984 * Check events in order.
987 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
989 /* threshold event is triggered in finer grain than soft limit */
990 if (unlikely(mem_cgroup_event_ratelimit(memcg,
991 MEM_CGROUP_TARGET_THRESH))) {
994 do_softlimit = mem_cgroup_event_ratelimit(memcg,
995 MEM_CGROUP_TARGET_SOFTLIMIT);
996 mem_cgroup_threshold(memcg);
997 if (unlikely(do_softlimit))
998 mem_cgroup_update_tree(memcg, page);
1002 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1005 * mm_update_next_owner() may clear mm->owner to NULL
1006 * if it races with swapoff, page migration, etc.
1007 * So this can be called with p == NULL.
1012 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1014 EXPORT_SYMBOL(mem_cgroup_from_task);
1017 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1018 * @mm: mm from which memcg should be extracted. It can be NULL.
1020 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1021 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1024 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1026 struct mem_cgroup *memcg;
1028 if (mem_cgroup_disabled())
1034 * Page cache insertions can happen withou an
1035 * actual mm context, e.g. during disk probing
1036 * on boot, loopback IO, acct() writes etc.
1039 memcg = root_mem_cgroup;
1041 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1042 if (unlikely(!memcg))
1043 memcg = root_mem_cgroup;
1045 } while (!css_tryget(&memcg->css));
1049 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1052 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1053 * @page: page from which memcg should be extracted.
1055 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1056 * root_mem_cgroup is returned.
1058 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1060 struct mem_cgroup *memcg = page->mem_cgroup;
1062 if (mem_cgroup_disabled())
1066 /* Page should not get uncharged and freed memcg under us. */
1067 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1068 memcg = root_mem_cgroup;
1072 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1074 static __always_inline struct mem_cgroup *active_memcg(void)
1077 return this_cpu_read(int_active_memcg);
1079 return current->active_memcg;
1082 static __always_inline struct mem_cgroup *get_active_memcg(void)
1084 struct mem_cgroup *memcg;
1087 memcg = active_memcg();
1089 /* current->active_memcg must hold a ref. */
1090 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1091 memcg = root_mem_cgroup;
1093 memcg = current->active_memcg;
1100 static __always_inline bool memcg_kmem_bypass(void)
1102 /* Allow remote memcg charging from any context. */
1103 if (unlikely(active_memcg()))
1106 /* Memcg to charge can't be determined. */
1107 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1114 * If active memcg is set, do not fallback to current->mm->memcg.
1116 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1118 if (memcg_kmem_bypass())
1121 if (unlikely(active_memcg()))
1122 return get_active_memcg();
1124 return get_mem_cgroup_from_mm(current->mm);
1128 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1129 * @root: hierarchy root
1130 * @prev: previously returned memcg, NULL on first invocation
1131 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1133 * Returns references to children of the hierarchy below @root, or
1134 * @root itself, or %NULL after a full round-trip.
1136 * Caller must pass the return value in @prev on subsequent
1137 * invocations for reference counting, or use mem_cgroup_iter_break()
1138 * to cancel a hierarchy walk before the round-trip is complete.
1140 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1141 * in the hierarchy among all concurrent reclaimers operating on the
1144 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1145 struct mem_cgroup *prev,
1146 struct mem_cgroup_reclaim_cookie *reclaim)
1148 struct mem_cgroup_reclaim_iter *iter;
1149 struct cgroup_subsys_state *css = NULL;
1150 struct mem_cgroup *memcg = NULL;
1151 struct mem_cgroup *pos = NULL;
1153 if (mem_cgroup_disabled())
1157 root = root_mem_cgroup;
1159 if (prev && !reclaim)
1165 struct mem_cgroup_per_node *mz;
1167 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1170 if (prev && reclaim->generation != iter->generation)
1174 pos = READ_ONCE(iter->position);
1175 if (!pos || css_tryget(&pos->css))
1178 * css reference reached zero, so iter->position will
1179 * be cleared by ->css_released. However, we should not
1180 * rely on this happening soon, because ->css_released
1181 * is called from a work queue, and by busy-waiting we
1182 * might block it. So we clear iter->position right
1185 (void)cmpxchg(&iter->position, pos, NULL);
1193 css = css_next_descendant_pre(css, &root->css);
1196 * Reclaimers share the hierarchy walk, and a
1197 * new one might jump in right at the end of
1198 * the hierarchy - make sure they see at least
1199 * one group and restart from the beginning.
1207 * Verify the css and acquire a reference. The root
1208 * is provided by the caller, so we know it's alive
1209 * and kicking, and don't take an extra reference.
1211 memcg = mem_cgroup_from_css(css);
1213 if (css == &root->css)
1216 if (css_tryget(css))
1224 * The position could have already been updated by a competing
1225 * thread, so check that the value hasn't changed since we read
1226 * it to avoid reclaiming from the same cgroup twice.
1228 (void)cmpxchg(&iter->position, pos, memcg);
1236 reclaim->generation = iter->generation;
1241 if (prev && prev != root)
1242 css_put(&prev->css);
1248 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1249 * @root: hierarchy root
1250 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1252 void mem_cgroup_iter_break(struct mem_cgroup *root,
1253 struct mem_cgroup *prev)
1256 root = root_mem_cgroup;
1257 if (prev && prev != root)
1258 css_put(&prev->css);
1261 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1262 struct mem_cgroup *dead_memcg)
1264 struct mem_cgroup_reclaim_iter *iter;
1265 struct mem_cgroup_per_node *mz;
1268 for_each_node(nid) {
1269 mz = mem_cgroup_nodeinfo(from, nid);
1271 cmpxchg(&iter->position, dead_memcg, NULL);
1275 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1277 struct mem_cgroup *memcg = dead_memcg;
1278 struct mem_cgroup *last;
1281 __invalidate_reclaim_iterators(memcg, dead_memcg);
1283 } while ((memcg = parent_mem_cgroup(memcg)));
1286 * When cgruop1 non-hierarchy mode is used,
1287 * parent_mem_cgroup() does not walk all the way up to the
1288 * cgroup root (root_mem_cgroup). So we have to handle
1289 * dead_memcg from cgroup root separately.
1291 if (last != root_mem_cgroup)
1292 __invalidate_reclaim_iterators(root_mem_cgroup,
1297 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1298 * @memcg: hierarchy root
1299 * @fn: function to call for each task
1300 * @arg: argument passed to @fn
1302 * This function iterates over tasks attached to @memcg or to any of its
1303 * descendants and calls @fn for each task. If @fn returns a non-zero
1304 * value, the function breaks the iteration loop and returns the value.
1305 * Otherwise, it will iterate over all tasks and return 0.
1307 * This function must not be called for the root memory cgroup.
1309 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1310 int (*fn)(struct task_struct *, void *), void *arg)
1312 struct mem_cgroup *iter;
1315 BUG_ON(memcg == root_mem_cgroup);
1317 for_each_mem_cgroup_tree(iter, memcg) {
1318 struct css_task_iter it;
1319 struct task_struct *task;
1321 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1322 while (!ret && (task = css_task_iter_next(&it)))
1323 ret = fn(task, arg);
1324 css_task_iter_end(&it);
1326 mem_cgroup_iter_break(memcg, iter);
1334 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1336 * @pgdat: pgdat of the page
1338 * This function relies on page's memcg being stable - see the
1339 * access rules in commit_charge().
1341 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1343 struct mem_cgroup_per_node *mz;
1344 struct mem_cgroup *memcg;
1345 struct lruvec *lruvec;
1347 if (mem_cgroup_disabled()) {
1348 lruvec = &pgdat->__lruvec;
1352 memcg = page->mem_cgroup;
1354 * Swapcache readahead pages are added to the LRU - and
1355 * possibly migrated - before they are charged.
1358 memcg = root_mem_cgroup;
1360 mz = mem_cgroup_page_nodeinfo(memcg, page);
1361 lruvec = &mz->lruvec;
1364 * Since a node can be onlined after the mem_cgroup was created,
1365 * we have to be prepared to initialize lruvec->zone here;
1366 * and if offlined then reonlined, we need to reinitialize it.
1368 if (unlikely(lruvec->pgdat != pgdat))
1369 lruvec->pgdat = pgdat;
1374 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1375 * @lruvec: mem_cgroup per zone lru vector
1376 * @lru: index of lru list the page is sitting on
1377 * @zid: zone id of the accounted pages
1378 * @nr_pages: positive when adding or negative when removing
1380 * This function must be called under lru_lock, just before a page is added
1381 * to or just after a page is removed from an lru list (that ordering being
1382 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1384 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1385 int zid, int nr_pages)
1387 struct mem_cgroup_per_node *mz;
1388 unsigned long *lru_size;
1391 if (mem_cgroup_disabled())
1394 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1395 lru_size = &mz->lru_zone_size[zid][lru];
1398 *lru_size += nr_pages;
1401 if (WARN_ONCE(size < 0,
1402 "%s(%p, %d, %d): lru_size %ld\n",
1403 __func__, lruvec, lru, nr_pages, size)) {
1409 *lru_size += nr_pages;
1413 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1414 * @memcg: the memory cgroup
1416 * Returns the maximum amount of memory @mem can be charged with, in
1419 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1421 unsigned long margin = 0;
1422 unsigned long count;
1423 unsigned long limit;
1425 count = page_counter_read(&memcg->memory);
1426 limit = READ_ONCE(memcg->memory.max);
1428 margin = limit - count;
1430 if (do_memsw_account()) {
1431 count = page_counter_read(&memcg->memsw);
1432 limit = READ_ONCE(memcg->memsw.max);
1434 margin = min(margin, limit - count);
1443 * A routine for checking "mem" is under move_account() or not.
1445 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1446 * moving cgroups. This is for waiting at high-memory pressure
1449 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1451 struct mem_cgroup *from;
1452 struct mem_cgroup *to;
1455 * Unlike task_move routines, we access mc.to, mc.from not under
1456 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1458 spin_lock(&mc.lock);
1464 ret = mem_cgroup_is_descendant(from, memcg) ||
1465 mem_cgroup_is_descendant(to, memcg);
1467 spin_unlock(&mc.lock);
1471 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1473 if (mc.moving_task && current != mc.moving_task) {
1474 if (mem_cgroup_under_move(memcg)) {
1476 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1477 /* moving charge context might have finished. */
1480 finish_wait(&mc.waitq, &wait);
1487 struct memory_stat {
1493 static struct memory_stat memory_stats[] = {
1494 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1495 { "file", PAGE_SIZE, NR_FILE_PAGES },
1496 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1497 { "pagetables", PAGE_SIZE, NR_PAGETABLE },
1498 { "percpu", 1, MEMCG_PERCPU_B },
1499 { "sock", PAGE_SIZE, MEMCG_SOCK },
1500 { "shmem", PAGE_SIZE, NR_SHMEM },
1501 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1502 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1503 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1504 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1506 * The ratio will be initialized in memory_stats_init(). Because
1507 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1508 * constant(e.g. powerpc).
1510 { "anon_thp", 0, NR_ANON_THPS },
1511 { "file_thp", 0, NR_FILE_THPS },
1512 { "shmem_thp", 0, NR_SHMEM_THPS },
1514 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1515 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1516 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1517 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1518 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1521 * Note: The slab_reclaimable and slab_unreclaimable must be
1522 * together and slab_reclaimable must be in front.
1524 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1525 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1527 /* The memory events */
1528 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1529 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1530 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1531 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1532 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1533 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1534 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1537 static int __init memory_stats_init(void)
1541 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1542 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1543 if (memory_stats[i].idx == NR_ANON_THPS ||
1544 memory_stats[i].idx == NR_FILE_THPS ||
1545 memory_stats[i].idx == NR_SHMEM_THPS)
1546 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1548 VM_BUG_ON(!memory_stats[i].ratio);
1549 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1554 pure_initcall(memory_stats_init);
1556 static char *memory_stat_format(struct mem_cgroup *memcg)
1561 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1566 * Provide statistics on the state of the memory subsystem as
1567 * well as cumulative event counters that show past behavior.
1569 * This list is ordered following a combination of these gradients:
1570 * 1) generic big picture -> specifics and details
1571 * 2) reflecting userspace activity -> reflecting kernel heuristics
1573 * Current memory state:
1576 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1579 size = memcg_page_state(memcg, memory_stats[i].idx);
1580 size *= memory_stats[i].ratio;
1581 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1583 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1584 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1585 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1586 seq_buf_printf(&s, "slab %llu\n", size);
1590 /* Accumulated memory events */
1592 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1593 memcg_events(memcg, PGFAULT));
1594 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1595 memcg_events(memcg, PGMAJFAULT));
1596 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1597 memcg_events(memcg, PGREFILL));
1598 seq_buf_printf(&s, "pgscan %lu\n",
1599 memcg_events(memcg, PGSCAN_KSWAPD) +
1600 memcg_events(memcg, PGSCAN_DIRECT));
1601 seq_buf_printf(&s, "pgsteal %lu\n",
1602 memcg_events(memcg, PGSTEAL_KSWAPD) +
1603 memcg_events(memcg, PGSTEAL_DIRECT));
1604 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1605 memcg_events(memcg, PGACTIVATE));
1606 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1607 memcg_events(memcg, PGDEACTIVATE));
1608 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1609 memcg_events(memcg, PGLAZYFREE));
1610 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1611 memcg_events(memcg, PGLAZYFREED));
1613 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1614 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1615 memcg_events(memcg, THP_FAULT_ALLOC));
1616 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1617 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1618 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1620 /* The above should easily fit into one page */
1621 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1626 #define K(x) ((x) << (PAGE_SHIFT-10))
1628 * mem_cgroup_print_oom_context: Print OOM information relevant to
1629 * memory controller.
1630 * @memcg: The memory cgroup that went over limit
1631 * @p: Task that is going to be killed
1633 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1636 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1641 pr_cont(",oom_memcg=");
1642 pr_cont_cgroup_path(memcg->css.cgroup);
1644 pr_cont(",global_oom");
1646 pr_cont(",task_memcg=");
1647 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1653 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1654 * memory controller.
1655 * @memcg: The memory cgroup that went over limit
1657 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1661 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1662 K((u64)page_counter_read(&memcg->memory)),
1663 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1664 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1665 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1666 K((u64)page_counter_read(&memcg->swap)),
1667 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1669 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1670 K((u64)page_counter_read(&memcg->memsw)),
1671 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1672 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1673 K((u64)page_counter_read(&memcg->kmem)),
1674 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1677 pr_info("Memory cgroup stats for ");
1678 pr_cont_cgroup_path(memcg->css.cgroup);
1680 buf = memory_stat_format(memcg);
1688 * Return the memory (and swap, if configured) limit for a memcg.
1690 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1692 unsigned long max = READ_ONCE(memcg->memory.max);
1694 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1695 if (mem_cgroup_swappiness(memcg))
1696 max += min(READ_ONCE(memcg->swap.max),
1697 (unsigned long)total_swap_pages);
1699 if (mem_cgroup_swappiness(memcg)) {
1700 /* Calculate swap excess capacity from memsw limit */
1701 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1703 max += min(swap, (unsigned long)total_swap_pages);
1709 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1711 return page_counter_read(&memcg->memory);
1714 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1717 struct oom_control oc = {
1721 .gfp_mask = gfp_mask,
1726 if (mutex_lock_killable(&oom_lock))
1729 if (mem_cgroup_margin(memcg) >= (1 << order))
1733 * A few threads which were not waiting at mutex_lock_killable() can
1734 * fail to bail out. Therefore, check again after holding oom_lock.
1736 ret = should_force_charge() || out_of_memory(&oc);
1739 mutex_unlock(&oom_lock);
1743 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1746 unsigned long *total_scanned)
1748 struct mem_cgroup *victim = NULL;
1751 unsigned long excess;
1752 unsigned long nr_scanned;
1753 struct mem_cgroup_reclaim_cookie reclaim = {
1757 excess = soft_limit_excess(root_memcg);
1760 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1765 * If we have not been able to reclaim
1766 * anything, it might because there are
1767 * no reclaimable pages under this hierarchy
1772 * We want to do more targeted reclaim.
1773 * excess >> 2 is not to excessive so as to
1774 * reclaim too much, nor too less that we keep
1775 * coming back to reclaim from this cgroup
1777 if (total >= (excess >> 2) ||
1778 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1783 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1784 pgdat, &nr_scanned);
1785 *total_scanned += nr_scanned;
1786 if (!soft_limit_excess(root_memcg))
1789 mem_cgroup_iter_break(root_memcg, victim);
1793 #ifdef CONFIG_LOCKDEP
1794 static struct lockdep_map memcg_oom_lock_dep_map = {
1795 .name = "memcg_oom_lock",
1799 static DEFINE_SPINLOCK(memcg_oom_lock);
1802 * Check OOM-Killer is already running under our hierarchy.
1803 * If someone is running, return false.
1805 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1807 struct mem_cgroup *iter, *failed = NULL;
1809 spin_lock(&memcg_oom_lock);
1811 for_each_mem_cgroup_tree(iter, memcg) {
1812 if (iter->oom_lock) {
1814 * this subtree of our hierarchy is already locked
1815 * so we cannot give a lock.
1818 mem_cgroup_iter_break(memcg, iter);
1821 iter->oom_lock = true;
1826 * OK, we failed to lock the whole subtree so we have
1827 * to clean up what we set up to the failing subtree
1829 for_each_mem_cgroup_tree(iter, memcg) {
1830 if (iter == failed) {
1831 mem_cgroup_iter_break(memcg, iter);
1834 iter->oom_lock = false;
1837 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1839 spin_unlock(&memcg_oom_lock);
1844 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1846 struct mem_cgroup *iter;
1848 spin_lock(&memcg_oom_lock);
1849 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1850 for_each_mem_cgroup_tree(iter, memcg)
1851 iter->oom_lock = false;
1852 spin_unlock(&memcg_oom_lock);
1855 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1857 struct mem_cgroup *iter;
1859 spin_lock(&memcg_oom_lock);
1860 for_each_mem_cgroup_tree(iter, memcg)
1862 spin_unlock(&memcg_oom_lock);
1865 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1867 struct mem_cgroup *iter;
1870 * Be careful about under_oom underflows becase a child memcg
1871 * could have been added after mem_cgroup_mark_under_oom.
1873 spin_lock(&memcg_oom_lock);
1874 for_each_mem_cgroup_tree(iter, memcg)
1875 if (iter->under_oom > 0)
1877 spin_unlock(&memcg_oom_lock);
1880 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1882 struct oom_wait_info {
1883 struct mem_cgroup *memcg;
1884 wait_queue_entry_t wait;
1887 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1888 unsigned mode, int sync, void *arg)
1890 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1891 struct mem_cgroup *oom_wait_memcg;
1892 struct oom_wait_info *oom_wait_info;
1894 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1895 oom_wait_memcg = oom_wait_info->memcg;
1897 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1898 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1900 return autoremove_wake_function(wait, mode, sync, arg);
1903 static void memcg_oom_recover(struct mem_cgroup *memcg)
1906 * For the following lockless ->under_oom test, the only required
1907 * guarantee is that it must see the state asserted by an OOM when
1908 * this function is called as a result of userland actions
1909 * triggered by the notification of the OOM. This is trivially
1910 * achieved by invoking mem_cgroup_mark_under_oom() before
1911 * triggering notification.
1913 if (memcg && memcg->under_oom)
1914 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1924 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1926 enum oom_status ret;
1929 if (order > PAGE_ALLOC_COSTLY_ORDER)
1932 memcg_memory_event(memcg, MEMCG_OOM);
1935 * We are in the middle of the charge context here, so we
1936 * don't want to block when potentially sitting on a callstack
1937 * that holds all kinds of filesystem and mm locks.
1939 * cgroup1 allows disabling the OOM killer and waiting for outside
1940 * handling until the charge can succeed; remember the context and put
1941 * the task to sleep at the end of the page fault when all locks are
1944 * On the other hand, in-kernel OOM killer allows for an async victim
1945 * memory reclaim (oom_reaper) and that means that we are not solely
1946 * relying on the oom victim to make a forward progress and we can
1947 * invoke the oom killer here.
1949 * Please note that mem_cgroup_out_of_memory might fail to find a
1950 * victim and then we have to bail out from the charge path.
1952 if (memcg->oom_kill_disable) {
1953 if (!current->in_user_fault)
1955 css_get(&memcg->css);
1956 current->memcg_in_oom = memcg;
1957 current->memcg_oom_gfp_mask = mask;
1958 current->memcg_oom_order = order;
1963 mem_cgroup_mark_under_oom(memcg);
1965 locked = mem_cgroup_oom_trylock(memcg);
1968 mem_cgroup_oom_notify(memcg);
1970 mem_cgroup_unmark_under_oom(memcg);
1971 if (mem_cgroup_out_of_memory(memcg, mask, order))
1977 mem_cgroup_oom_unlock(memcg);
1983 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1984 * @handle: actually kill/wait or just clean up the OOM state
1986 * This has to be called at the end of a page fault if the memcg OOM
1987 * handler was enabled.
1989 * Memcg supports userspace OOM handling where failed allocations must
1990 * sleep on a waitqueue until the userspace task resolves the
1991 * situation. Sleeping directly in the charge context with all kinds
1992 * of locks held is not a good idea, instead we remember an OOM state
1993 * in the task and mem_cgroup_oom_synchronize() has to be called at
1994 * the end of the page fault to complete the OOM handling.
1996 * Returns %true if an ongoing memcg OOM situation was detected and
1997 * completed, %false otherwise.
1999 bool mem_cgroup_oom_synchronize(bool handle)
2001 struct mem_cgroup *memcg = current->memcg_in_oom;
2002 struct oom_wait_info owait;
2005 /* OOM is global, do not handle */
2012 owait.memcg = memcg;
2013 owait.wait.flags = 0;
2014 owait.wait.func = memcg_oom_wake_function;
2015 owait.wait.private = current;
2016 INIT_LIST_HEAD(&owait.wait.entry);
2018 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2019 mem_cgroup_mark_under_oom(memcg);
2021 locked = mem_cgroup_oom_trylock(memcg);
2024 mem_cgroup_oom_notify(memcg);
2026 if (locked && !memcg->oom_kill_disable) {
2027 mem_cgroup_unmark_under_oom(memcg);
2028 finish_wait(&memcg_oom_waitq, &owait.wait);
2029 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2030 current->memcg_oom_order);
2033 mem_cgroup_unmark_under_oom(memcg);
2034 finish_wait(&memcg_oom_waitq, &owait.wait);
2038 mem_cgroup_oom_unlock(memcg);
2040 * There is no guarantee that an OOM-lock contender
2041 * sees the wakeups triggered by the OOM kill
2042 * uncharges. Wake any sleepers explicitely.
2044 memcg_oom_recover(memcg);
2047 current->memcg_in_oom = NULL;
2048 css_put(&memcg->css);
2053 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2054 * @victim: task to be killed by the OOM killer
2055 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2057 * Returns a pointer to a memory cgroup, which has to be cleaned up
2058 * by killing all belonging OOM-killable tasks.
2060 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2062 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2063 struct mem_cgroup *oom_domain)
2065 struct mem_cgroup *oom_group = NULL;
2066 struct mem_cgroup *memcg;
2068 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2072 oom_domain = root_mem_cgroup;
2076 memcg = mem_cgroup_from_task(victim);
2077 if (memcg == root_mem_cgroup)
2081 * If the victim task has been asynchronously moved to a different
2082 * memory cgroup, we might end up killing tasks outside oom_domain.
2083 * In this case it's better to ignore memory.group.oom.
2085 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2089 * Traverse the memory cgroup hierarchy from the victim task's
2090 * cgroup up to the OOMing cgroup (or root) to find the
2091 * highest-level memory cgroup with oom.group set.
2093 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2094 if (memcg->oom_group)
2097 if (memcg == oom_domain)
2102 css_get(&oom_group->css);
2109 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2111 pr_info("Tasks in ");
2112 pr_cont_cgroup_path(memcg->css.cgroup);
2113 pr_cont(" are going to be killed due to memory.oom.group set\n");
2117 * lock_page_memcg - lock a page->mem_cgroup binding
2120 * This function protects unlocked LRU pages from being moved to
2123 * It ensures lifetime of the returned memcg. Caller is responsible
2124 * for the lifetime of the page; __unlock_page_memcg() is available
2125 * when @page might get freed inside the locked section.
2127 struct mem_cgroup *lock_page_memcg(struct page *page)
2129 struct page *head = compound_head(page); /* rmap on tail pages */
2130 struct mem_cgroup *memcg;
2131 unsigned long flags;
2134 * The RCU lock is held throughout the transaction. The fast
2135 * path can get away without acquiring the memcg->move_lock
2136 * because page moving starts with an RCU grace period.
2138 * The RCU lock also protects the memcg from being freed when
2139 * the page state that is going to change is the only thing
2140 * preventing the page itself from being freed. E.g. writeback
2141 * doesn't hold a page reference and relies on PG_writeback to
2142 * keep off truncation, migration and so forth.
2146 if (mem_cgroup_disabled())
2149 memcg = head->mem_cgroup;
2150 if (unlikely(!memcg))
2153 #ifdef CONFIG_PROVE_LOCKING
2154 local_irq_save(flags);
2155 might_lock(&memcg->move_lock);
2156 local_irq_restore(flags);
2159 if (atomic_read(&memcg->moving_account) <= 0)
2162 spin_lock_irqsave(&memcg->move_lock, flags);
2163 if (memcg != head->mem_cgroup) {
2164 spin_unlock_irqrestore(&memcg->move_lock, flags);
2169 * When charge migration first begins, we can have locked and
2170 * unlocked page stat updates happening concurrently. Track
2171 * the task who has the lock for unlock_page_memcg().
2173 memcg->move_lock_task = current;
2174 memcg->move_lock_flags = flags;
2178 EXPORT_SYMBOL(lock_page_memcg);
2181 * __unlock_page_memcg - unlock and unpin a memcg
2184 * Unlock and unpin a memcg returned by lock_page_memcg().
2186 void __unlock_page_memcg(struct mem_cgroup *memcg)
2188 if (memcg && memcg->move_lock_task == current) {
2189 unsigned long flags = memcg->move_lock_flags;
2191 memcg->move_lock_task = NULL;
2192 memcg->move_lock_flags = 0;
2194 spin_unlock_irqrestore(&memcg->move_lock, flags);
2201 * unlock_page_memcg - unlock a page->mem_cgroup binding
2204 void unlock_page_memcg(struct page *page)
2206 struct page *head = compound_head(page);
2208 __unlock_page_memcg(head->mem_cgroup);
2210 EXPORT_SYMBOL(unlock_page_memcg);
2212 struct memcg_stock_pcp {
2213 struct mem_cgroup *cached; /* this never be root cgroup */
2214 unsigned int nr_pages;
2216 #ifdef CONFIG_MEMCG_KMEM
2217 struct obj_cgroup *cached_objcg;
2218 unsigned int nr_bytes;
2221 struct work_struct work;
2222 unsigned long flags;
2223 #define FLUSHING_CACHED_CHARGE 0
2225 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2226 static DEFINE_MUTEX(percpu_charge_mutex);
2228 #ifdef CONFIG_MEMCG_KMEM
2229 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2230 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2231 struct mem_cgroup *root_memcg);
2234 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2237 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2238 struct mem_cgroup *root_memcg)
2245 * consume_stock: Try to consume stocked charge on this cpu.
2246 * @memcg: memcg to consume from.
2247 * @nr_pages: how many pages to charge.
2249 * The charges will only happen if @memcg matches the current cpu's memcg
2250 * stock, and at least @nr_pages are available in that stock. Failure to
2251 * service an allocation will refill the stock.
2253 * returns true if successful, false otherwise.
2255 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2257 struct memcg_stock_pcp *stock;
2258 unsigned long flags;
2261 if (nr_pages > MEMCG_CHARGE_BATCH)
2264 local_irq_save(flags);
2266 stock = this_cpu_ptr(&memcg_stock);
2267 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2268 stock->nr_pages -= nr_pages;
2272 local_irq_restore(flags);
2278 * Returns stocks cached in percpu and reset cached information.
2280 static void drain_stock(struct memcg_stock_pcp *stock)
2282 struct mem_cgroup *old = stock->cached;
2287 if (stock->nr_pages) {
2288 page_counter_uncharge(&old->memory, stock->nr_pages);
2289 if (do_memsw_account())
2290 page_counter_uncharge(&old->memsw, stock->nr_pages);
2291 stock->nr_pages = 0;
2295 stock->cached = NULL;
2298 static void drain_local_stock(struct work_struct *dummy)
2300 struct memcg_stock_pcp *stock;
2301 unsigned long flags;
2304 * The only protection from memory hotplug vs. drain_stock races is
2305 * that we always operate on local CPU stock here with IRQ disabled
2307 local_irq_save(flags);
2309 stock = this_cpu_ptr(&memcg_stock);
2310 drain_obj_stock(stock);
2312 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2314 local_irq_restore(flags);
2318 * Cache charges(val) to local per_cpu area.
2319 * This will be consumed by consume_stock() function, later.
2321 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2323 struct memcg_stock_pcp *stock;
2324 unsigned long flags;
2326 local_irq_save(flags);
2328 stock = this_cpu_ptr(&memcg_stock);
2329 if (stock->cached != memcg) { /* reset if necessary */
2331 css_get(&memcg->css);
2332 stock->cached = memcg;
2334 stock->nr_pages += nr_pages;
2336 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2339 local_irq_restore(flags);
2343 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2344 * of the hierarchy under it.
2346 static void drain_all_stock(struct mem_cgroup *root_memcg)
2350 /* If someone's already draining, avoid adding running more workers. */
2351 if (!mutex_trylock(&percpu_charge_mutex))
2354 * Notify other cpus that system-wide "drain" is running
2355 * We do not care about races with the cpu hotplug because cpu down
2356 * as well as workers from this path always operate on the local
2357 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2360 for_each_online_cpu(cpu) {
2361 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2362 struct mem_cgroup *memcg;
2366 memcg = stock->cached;
2367 if (memcg && stock->nr_pages &&
2368 mem_cgroup_is_descendant(memcg, root_memcg))
2370 if (obj_stock_flush_required(stock, root_memcg))
2375 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2377 drain_local_stock(&stock->work);
2379 schedule_work_on(cpu, &stock->work);
2383 mutex_unlock(&percpu_charge_mutex);
2386 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2388 struct memcg_stock_pcp *stock;
2389 struct mem_cgroup *memcg, *mi;
2391 stock = &per_cpu(memcg_stock, cpu);
2394 for_each_mem_cgroup(memcg) {
2397 for (i = 0; i < MEMCG_NR_STAT; i++) {
2401 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2403 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2404 atomic_long_add(x, &memcg->vmstats[i]);
2406 if (i >= NR_VM_NODE_STAT_ITEMS)
2409 for_each_node(nid) {
2410 struct mem_cgroup_per_node *pn;
2412 pn = mem_cgroup_nodeinfo(memcg, nid);
2413 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2416 atomic_long_add(x, &pn->lruvec_stat[i]);
2417 } while ((pn = parent_nodeinfo(pn, nid)));
2421 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2424 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2426 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2427 atomic_long_add(x, &memcg->vmevents[i]);
2434 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2435 unsigned int nr_pages,
2438 unsigned long nr_reclaimed = 0;
2441 unsigned long pflags;
2443 if (page_counter_read(&memcg->memory) <=
2444 READ_ONCE(memcg->memory.high))
2447 memcg_memory_event(memcg, MEMCG_HIGH);
2449 psi_memstall_enter(&pflags);
2450 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2452 psi_memstall_leave(&pflags);
2453 } while ((memcg = parent_mem_cgroup(memcg)) &&
2454 !mem_cgroup_is_root(memcg));
2456 return nr_reclaimed;
2459 static void high_work_func(struct work_struct *work)
2461 struct mem_cgroup *memcg;
2463 memcg = container_of(work, struct mem_cgroup, high_work);
2464 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2468 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2469 * enough to still cause a significant slowdown in most cases, while still
2470 * allowing diagnostics and tracing to proceed without becoming stuck.
2472 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2475 * When calculating the delay, we use these either side of the exponentiation to
2476 * maintain precision and scale to a reasonable number of jiffies (see the table
2479 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2480 * overage ratio to a delay.
2481 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2482 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2483 * to produce a reasonable delay curve.
2485 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2486 * reasonable delay curve compared to precision-adjusted overage, not
2487 * penalising heavily at first, but still making sure that growth beyond the
2488 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2489 * example, with a high of 100 megabytes:
2491 * +-------+------------------------+
2492 * | usage | time to allocate in ms |
2493 * +-------+------------------------+
2515 * +-------+------------------------+
2517 #define MEMCG_DELAY_PRECISION_SHIFT 20
2518 #define MEMCG_DELAY_SCALING_SHIFT 14
2520 static u64 calculate_overage(unsigned long usage, unsigned long high)
2528 * Prevent division by 0 in overage calculation by acting as if
2529 * it was a threshold of 1 page
2531 high = max(high, 1UL);
2533 overage = usage - high;
2534 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2535 return div64_u64(overage, high);
2538 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2540 u64 overage, max_overage = 0;
2543 overage = calculate_overage(page_counter_read(&memcg->memory),
2544 READ_ONCE(memcg->memory.high));
2545 max_overage = max(overage, max_overage);
2546 } while ((memcg = parent_mem_cgroup(memcg)) &&
2547 !mem_cgroup_is_root(memcg));
2552 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2554 u64 overage, max_overage = 0;
2557 overage = calculate_overage(page_counter_read(&memcg->swap),
2558 READ_ONCE(memcg->swap.high));
2560 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2561 max_overage = max(overage, max_overage);
2562 } while ((memcg = parent_mem_cgroup(memcg)) &&
2563 !mem_cgroup_is_root(memcg));
2569 * Get the number of jiffies that we should penalise a mischievous cgroup which
2570 * is exceeding its memory.high by checking both it and its ancestors.
2572 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2573 unsigned int nr_pages,
2576 unsigned long penalty_jiffies;
2582 * We use overage compared to memory.high to calculate the number of
2583 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2584 * fairly lenient on small overages, and increasingly harsh when the
2585 * memcg in question makes it clear that it has no intention of stopping
2586 * its crazy behaviour, so we exponentially increase the delay based on
2589 penalty_jiffies = max_overage * max_overage * HZ;
2590 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2591 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2594 * Factor in the task's own contribution to the overage, such that four
2595 * N-sized allocations are throttled approximately the same as one
2596 * 4N-sized allocation.
2598 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2599 * larger the current charge patch is than that.
2601 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2605 * Scheduled by try_charge() to be executed from the userland return path
2606 * and reclaims memory over the high limit.
2608 void mem_cgroup_handle_over_high(void)
2610 unsigned long penalty_jiffies;
2611 unsigned long pflags;
2612 unsigned long nr_reclaimed;
2613 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2614 int nr_retries = MAX_RECLAIM_RETRIES;
2615 struct mem_cgroup *memcg;
2616 bool in_retry = false;
2618 if (likely(!nr_pages))
2621 memcg = get_mem_cgroup_from_mm(current->mm);
2622 current->memcg_nr_pages_over_high = 0;
2626 * The allocating task should reclaim at least the batch size, but for
2627 * subsequent retries we only want to do what's necessary to prevent oom
2628 * or breaching resource isolation.
2630 * This is distinct from memory.max or page allocator behaviour because
2631 * memory.high is currently batched, whereas memory.max and the page
2632 * allocator run every time an allocation is made.
2634 nr_reclaimed = reclaim_high(memcg,
2635 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2639 * memory.high is breached and reclaim is unable to keep up. Throttle
2640 * allocators proactively to slow down excessive growth.
2642 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2643 mem_find_max_overage(memcg));
2645 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2646 swap_find_max_overage(memcg));
2649 * Clamp the max delay per usermode return so as to still keep the
2650 * application moving forwards and also permit diagnostics, albeit
2653 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2656 * Don't sleep if the amount of jiffies this memcg owes us is so low
2657 * that it's not even worth doing, in an attempt to be nice to those who
2658 * go only a small amount over their memory.high value and maybe haven't
2659 * been aggressively reclaimed enough yet.
2661 if (penalty_jiffies <= HZ / 100)
2665 * If reclaim is making forward progress but we're still over
2666 * memory.high, we want to encourage that rather than doing allocator
2669 if (nr_reclaimed || nr_retries--) {
2675 * If we exit early, we're guaranteed to die (since
2676 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2677 * need to account for any ill-begotten jiffies to pay them off later.
2679 psi_memstall_enter(&pflags);
2680 schedule_timeout_killable(penalty_jiffies);
2681 psi_memstall_leave(&pflags);
2684 css_put(&memcg->css);
2687 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2688 unsigned int nr_pages)
2690 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2691 int nr_retries = MAX_RECLAIM_RETRIES;
2692 struct mem_cgroup *mem_over_limit;
2693 struct page_counter *counter;
2694 enum oom_status oom_status;
2695 unsigned long nr_reclaimed;
2696 bool may_swap = true;
2697 bool drained = false;
2698 unsigned long pflags;
2700 if (mem_cgroup_is_root(memcg))
2703 if (consume_stock(memcg, nr_pages))
2706 if (!do_memsw_account() ||
2707 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2708 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2710 if (do_memsw_account())
2711 page_counter_uncharge(&memcg->memsw, batch);
2712 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2714 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2718 if (batch > nr_pages) {
2724 * Memcg doesn't have a dedicated reserve for atomic
2725 * allocations. But like the global atomic pool, we need to
2726 * put the burden of reclaim on regular allocation requests
2727 * and let these go through as privileged allocations.
2729 if (gfp_mask & __GFP_ATOMIC)
2733 * Unlike in global OOM situations, memcg is not in a physical
2734 * memory shortage. Allow dying and OOM-killed tasks to
2735 * bypass the last charges so that they can exit quickly and
2736 * free their memory.
2738 if (unlikely(should_force_charge()))
2742 * Prevent unbounded recursion when reclaim operations need to
2743 * allocate memory. This might exceed the limits temporarily,
2744 * but we prefer facilitating memory reclaim and getting back
2745 * under the limit over triggering OOM kills in these cases.
2747 if (unlikely(current->flags & PF_MEMALLOC))
2750 if (unlikely(task_in_memcg_oom(current)))
2753 if (!gfpflags_allow_blocking(gfp_mask))
2756 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2758 psi_memstall_enter(&pflags);
2759 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2760 gfp_mask, may_swap);
2761 psi_memstall_leave(&pflags);
2763 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2767 drain_all_stock(mem_over_limit);
2772 if (gfp_mask & __GFP_NORETRY)
2775 * Even though the limit is exceeded at this point, reclaim
2776 * may have been able to free some pages. Retry the charge
2777 * before killing the task.
2779 * Only for regular pages, though: huge pages are rather
2780 * unlikely to succeed so close to the limit, and we fall back
2781 * to regular pages anyway in case of failure.
2783 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2786 * At task move, charge accounts can be doubly counted. So, it's
2787 * better to wait until the end of task_move if something is going on.
2789 if (mem_cgroup_wait_acct_move(mem_over_limit))
2795 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2798 if (gfp_mask & __GFP_NOFAIL)
2801 if (fatal_signal_pending(current))
2805 * keep retrying as long as the memcg oom killer is able to make
2806 * a forward progress or bypass the charge if the oom killer
2807 * couldn't make any progress.
2809 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2810 get_order(nr_pages * PAGE_SIZE));
2811 switch (oom_status) {
2813 nr_retries = MAX_RECLAIM_RETRIES;
2821 if (!(gfp_mask & __GFP_NOFAIL))
2825 * The allocation either can't fail or will lead to more memory
2826 * being freed very soon. Allow memory usage go over the limit
2827 * temporarily by force charging it.
2829 page_counter_charge(&memcg->memory, nr_pages);
2830 if (do_memsw_account())
2831 page_counter_charge(&memcg->memsw, nr_pages);
2836 if (batch > nr_pages)
2837 refill_stock(memcg, batch - nr_pages);
2840 * If the hierarchy is above the normal consumption range, schedule
2841 * reclaim on returning to userland. We can perform reclaim here
2842 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2843 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2844 * not recorded as it most likely matches current's and won't
2845 * change in the meantime. As high limit is checked again before
2846 * reclaim, the cost of mismatch is negligible.
2849 bool mem_high, swap_high;
2851 mem_high = page_counter_read(&memcg->memory) >
2852 READ_ONCE(memcg->memory.high);
2853 swap_high = page_counter_read(&memcg->swap) >
2854 READ_ONCE(memcg->swap.high);
2856 /* Don't bother a random interrupted task */
2857 if (in_interrupt()) {
2859 schedule_work(&memcg->high_work);
2865 if (mem_high || swap_high) {
2867 * The allocating tasks in this cgroup will need to do
2868 * reclaim or be throttled to prevent further growth
2869 * of the memory or swap footprints.
2871 * Target some best-effort fairness between the tasks,
2872 * and distribute reclaim work and delay penalties
2873 * based on how much each task is actually allocating.
2875 current->memcg_nr_pages_over_high += batch;
2876 set_notify_resume(current);
2879 } while ((memcg = parent_mem_cgroup(memcg)));
2884 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2885 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2887 if (mem_cgroup_is_root(memcg))
2890 page_counter_uncharge(&memcg->memory, nr_pages);
2891 if (do_memsw_account())
2892 page_counter_uncharge(&memcg->memsw, nr_pages);
2896 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2898 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2900 * Any of the following ensures page's memcg stability:
2904 * - lock_page_memcg()
2905 * - exclusive reference
2907 page->mem_cgroup = memcg;
2910 #ifdef CONFIG_MEMCG_KMEM
2911 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2914 unsigned int objects = objs_per_slab_page(s, page);
2917 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2922 if (cmpxchg(&page->obj_cgroups, NULL,
2923 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2926 kmemleak_not_leak(vec);
2932 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2934 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2935 * cgroup_mutex, etc.
2937 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2941 if (mem_cgroup_disabled())
2944 page = virt_to_head_page(p);
2947 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2948 * or a pointer to obj_cgroup vector. In the latter case the lowest
2949 * bit of the pointer is set.
2950 * The page->mem_cgroup pointer can be asynchronously changed
2951 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2952 * from a valid memcg pointer to objcg vector or back.
2954 if (!page->mem_cgroup)
2958 * Slab objects are accounted individually, not per-page.
2959 * Memcg membership data for each individual object is saved in
2960 * the page->obj_cgroups.
2962 if (page_has_obj_cgroups(page)) {
2963 struct obj_cgroup *objcg;
2966 off = obj_to_index(page->slab_cache, page, p);
2967 objcg = page_obj_cgroups(page)[off];
2969 return obj_cgroup_memcg(objcg);
2974 /* All other pages use page->mem_cgroup */
2975 return page->mem_cgroup;
2978 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2980 struct obj_cgroup *objcg = NULL;
2981 struct mem_cgroup *memcg;
2983 if (memcg_kmem_bypass())
2987 if (unlikely(active_memcg()))
2988 memcg = active_memcg();
2990 memcg = mem_cgroup_from_task(current);
2992 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2993 objcg = rcu_dereference(memcg->objcg);
2994 if (objcg && obj_cgroup_tryget(objcg))
3003 static int memcg_alloc_cache_id(void)
3008 id = ida_simple_get(&memcg_cache_ida,
3009 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3013 if (id < memcg_nr_cache_ids)
3017 * There's no space for the new id in memcg_caches arrays,
3018 * so we have to grow them.
3020 down_write(&memcg_cache_ids_sem);
3022 size = 2 * (id + 1);
3023 if (size < MEMCG_CACHES_MIN_SIZE)
3024 size = MEMCG_CACHES_MIN_SIZE;
3025 else if (size > MEMCG_CACHES_MAX_SIZE)
3026 size = MEMCG_CACHES_MAX_SIZE;
3028 err = memcg_update_all_list_lrus(size);
3030 memcg_nr_cache_ids = size;
3032 up_write(&memcg_cache_ids_sem);
3035 ida_simple_remove(&memcg_cache_ida, id);
3041 static void memcg_free_cache_id(int id)
3043 ida_simple_remove(&memcg_cache_ida, id);
3047 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3048 * @memcg: memory cgroup to charge
3049 * @gfp: reclaim mode
3050 * @nr_pages: number of pages to charge
3052 * Returns 0 on success, an error code on failure.
3054 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3055 unsigned int nr_pages)
3057 struct page_counter *counter;
3060 ret = try_charge(memcg, gfp, nr_pages);
3064 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3065 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3068 * Enforce __GFP_NOFAIL allocation because callers are not
3069 * prepared to see failures and likely do not have any failure
3072 if (gfp & __GFP_NOFAIL) {
3073 page_counter_charge(&memcg->kmem, nr_pages);
3076 cancel_charge(memcg, nr_pages);
3083 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3084 * @memcg: memcg to uncharge
3085 * @nr_pages: number of pages to uncharge
3087 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3089 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3090 page_counter_uncharge(&memcg->kmem, nr_pages);
3092 page_counter_uncharge(&memcg->memory, nr_pages);
3093 if (do_memsw_account())
3094 page_counter_uncharge(&memcg->memsw, nr_pages);
3098 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3099 * @page: page to charge
3100 * @gfp: reclaim mode
3101 * @order: allocation order
3103 * Returns 0 on success, an error code on failure.
3105 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3107 struct mem_cgroup *memcg;
3110 memcg = get_mem_cgroup_from_current();
3111 if (memcg && !mem_cgroup_is_root(memcg)) {
3112 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3114 page->mem_cgroup = memcg;
3115 __SetPageKmemcg(page);
3118 css_put(&memcg->css);
3124 * __memcg_kmem_uncharge_page: uncharge a kmem page
3125 * @page: page to uncharge
3126 * @order: allocation order
3128 void __memcg_kmem_uncharge_page(struct page *page, int order)
3130 struct mem_cgroup *memcg = page->mem_cgroup;
3131 unsigned int nr_pages = 1 << order;
3136 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3137 __memcg_kmem_uncharge(memcg, nr_pages);
3138 page->mem_cgroup = NULL;
3139 css_put(&memcg->css);
3141 /* slab pages do not have PageKmemcg flag set */
3142 if (PageKmemcg(page))
3143 __ClearPageKmemcg(page);
3146 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3148 struct memcg_stock_pcp *stock;
3149 unsigned long flags;
3152 local_irq_save(flags);
3154 stock = this_cpu_ptr(&memcg_stock);
3155 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3156 stock->nr_bytes -= nr_bytes;
3160 local_irq_restore(flags);
3165 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3167 struct obj_cgroup *old = stock->cached_objcg;
3172 if (stock->nr_bytes) {
3173 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3174 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3178 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3183 * The leftover is flushed to the centralized per-memcg value.
3184 * On the next attempt to refill obj stock it will be moved
3185 * to a per-cpu stock (probably, on an other CPU), see
3186 * refill_obj_stock().
3188 * How often it's flushed is a trade-off between the memory
3189 * limit enforcement accuracy and potential CPU contention,
3190 * so it might be changed in the future.
3192 atomic_add(nr_bytes, &old->nr_charged_bytes);
3193 stock->nr_bytes = 0;
3196 obj_cgroup_put(old);
3197 stock->cached_objcg = NULL;
3200 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3201 struct mem_cgroup *root_memcg)
3203 struct mem_cgroup *memcg;
3205 if (stock->cached_objcg) {
3206 memcg = obj_cgroup_memcg(stock->cached_objcg);
3207 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3214 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3216 struct memcg_stock_pcp *stock;
3217 unsigned long flags;
3219 local_irq_save(flags);
3221 stock = this_cpu_ptr(&memcg_stock);
3222 if (stock->cached_objcg != objcg) { /* reset if necessary */
3223 drain_obj_stock(stock);
3224 obj_cgroup_get(objcg);
3225 stock->cached_objcg = objcg;
3226 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3228 stock->nr_bytes += nr_bytes;
3230 if (stock->nr_bytes > PAGE_SIZE)
3231 drain_obj_stock(stock);
3233 local_irq_restore(flags);
3236 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3238 struct mem_cgroup *memcg;
3239 unsigned int nr_pages, nr_bytes;
3242 if (consume_obj_stock(objcg, size))
3246 * In theory, memcg->nr_charged_bytes can have enough
3247 * pre-charged bytes to satisfy the allocation. However,
3248 * flushing memcg->nr_charged_bytes requires two atomic
3249 * operations, and memcg->nr_charged_bytes can't be big,
3250 * so it's better to ignore it and try grab some new pages.
3251 * memcg->nr_charged_bytes will be flushed in
3252 * refill_obj_stock(), called from this function or
3253 * independently later.
3257 memcg = obj_cgroup_memcg(objcg);
3258 if (unlikely(!css_tryget(&memcg->css)))
3262 nr_pages = size >> PAGE_SHIFT;
3263 nr_bytes = size & (PAGE_SIZE - 1);
3268 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3269 if (!ret && nr_bytes)
3270 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3272 css_put(&memcg->css);
3276 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3278 refill_obj_stock(objcg, size);
3281 #endif /* CONFIG_MEMCG_KMEM */
3283 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3286 * Because tail pages are not marked as "used", set it. We're under
3287 * pgdat->lru_lock and migration entries setup in all page mappings.
3289 void mem_cgroup_split_huge_fixup(struct page *head)
3291 struct mem_cgroup *memcg = head->mem_cgroup;
3294 if (mem_cgroup_disabled())
3297 for (i = 1; i < HPAGE_PMD_NR; i++) {
3298 css_get(&memcg->css);
3299 head[i].mem_cgroup = memcg;
3302 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3304 #ifdef CONFIG_MEMCG_SWAP
3306 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3307 * @entry: swap entry to be moved
3308 * @from: mem_cgroup which the entry is moved from
3309 * @to: mem_cgroup which the entry is moved to
3311 * It succeeds only when the swap_cgroup's record for this entry is the same
3312 * as the mem_cgroup's id of @from.
3314 * Returns 0 on success, -EINVAL on failure.
3316 * The caller must have charged to @to, IOW, called page_counter_charge() about
3317 * both res and memsw, and called css_get().
3319 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3320 struct mem_cgroup *from, struct mem_cgroup *to)
3322 unsigned short old_id, new_id;
3324 old_id = mem_cgroup_id(from);
3325 new_id = mem_cgroup_id(to);
3327 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3328 mod_memcg_state(from, MEMCG_SWAP, -1);
3329 mod_memcg_state(to, MEMCG_SWAP, 1);
3335 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3336 struct mem_cgroup *from, struct mem_cgroup *to)
3342 static DEFINE_MUTEX(memcg_max_mutex);
3344 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3345 unsigned long max, bool memsw)
3347 bool enlarge = false;
3348 bool drained = false;
3350 bool limits_invariant;
3351 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3354 if (signal_pending(current)) {
3359 mutex_lock(&memcg_max_mutex);
3361 * Make sure that the new limit (memsw or memory limit) doesn't
3362 * break our basic invariant rule memory.max <= memsw.max.
3364 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3365 max <= memcg->memsw.max;
3366 if (!limits_invariant) {
3367 mutex_unlock(&memcg_max_mutex);
3371 if (max > counter->max)
3373 ret = page_counter_set_max(counter, max);
3374 mutex_unlock(&memcg_max_mutex);
3380 drain_all_stock(memcg);
3385 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3386 GFP_KERNEL, !memsw)) {
3392 if (!ret && enlarge)
3393 memcg_oom_recover(memcg);
3398 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3400 unsigned long *total_scanned)
3402 unsigned long nr_reclaimed = 0;
3403 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3404 unsigned long reclaimed;
3406 struct mem_cgroup_tree_per_node *mctz;
3407 unsigned long excess;
3408 unsigned long nr_scanned;
3413 mctz = soft_limit_tree_node(pgdat->node_id);
3416 * Do not even bother to check the largest node if the root
3417 * is empty. Do it lockless to prevent lock bouncing. Races
3418 * are acceptable as soft limit is best effort anyway.
3420 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3424 * This loop can run a while, specially if mem_cgroup's continuously
3425 * keep exceeding their soft limit and putting the system under
3432 mz = mem_cgroup_largest_soft_limit_node(mctz);
3437 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3438 gfp_mask, &nr_scanned);
3439 nr_reclaimed += reclaimed;
3440 *total_scanned += nr_scanned;
3441 spin_lock_irq(&mctz->lock);
3442 __mem_cgroup_remove_exceeded(mz, mctz);
3445 * If we failed to reclaim anything from this memory cgroup
3446 * it is time to move on to the next cgroup
3450 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3452 excess = soft_limit_excess(mz->memcg);
3454 * One school of thought says that we should not add
3455 * back the node to the tree if reclaim returns 0.
3456 * But our reclaim could return 0, simply because due
3457 * to priority we are exposing a smaller subset of
3458 * memory to reclaim from. Consider this as a longer
3461 /* If excess == 0, no tree ops */
3462 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3463 spin_unlock_irq(&mctz->lock);
3464 css_put(&mz->memcg->css);
3467 * Could not reclaim anything and there are no more
3468 * mem cgroups to try or we seem to be looping without
3469 * reclaiming anything.
3471 if (!nr_reclaimed &&
3473 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3475 } while (!nr_reclaimed);
3477 css_put(&next_mz->memcg->css);
3478 return nr_reclaimed;
3482 * Reclaims as many pages from the given memcg as possible.
3484 * Caller is responsible for holding css reference for memcg.
3486 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3488 int nr_retries = MAX_RECLAIM_RETRIES;
3490 /* we call try-to-free pages for make this cgroup empty */
3491 lru_add_drain_all();
3493 drain_all_stock(memcg);
3495 /* try to free all pages in this cgroup */
3496 while (nr_retries && page_counter_read(&memcg->memory)) {
3499 if (signal_pending(current))
3502 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3506 /* maybe some writeback is necessary */
3507 congestion_wait(BLK_RW_ASYNC, HZ/10);
3515 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3516 char *buf, size_t nbytes,
3519 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3521 if (mem_cgroup_is_root(memcg))
3523 return mem_cgroup_force_empty(memcg) ?: nbytes;
3526 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3532 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3533 struct cftype *cft, u64 val)
3538 pr_warn_once("Non-hierarchical mode is deprecated. "
3539 "Please report your usecase to linux-mm@kvack.org if you "
3540 "depend on this functionality.\n");
3545 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3549 if (mem_cgroup_is_root(memcg)) {
3550 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3551 memcg_page_state(memcg, NR_ANON_MAPPED);
3553 val += memcg_page_state(memcg, MEMCG_SWAP);
3556 val = page_counter_read(&memcg->memory);
3558 val = page_counter_read(&memcg->memsw);
3571 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3574 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3575 struct page_counter *counter;
3577 switch (MEMFILE_TYPE(cft->private)) {
3579 counter = &memcg->memory;
3582 counter = &memcg->memsw;
3585 counter = &memcg->kmem;
3588 counter = &memcg->tcpmem;
3594 switch (MEMFILE_ATTR(cft->private)) {
3596 if (counter == &memcg->memory)
3597 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3598 if (counter == &memcg->memsw)
3599 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3600 return (u64)page_counter_read(counter) * PAGE_SIZE;
3602 return (u64)counter->max * PAGE_SIZE;
3604 return (u64)counter->watermark * PAGE_SIZE;
3606 return counter->failcnt;
3607 case RES_SOFT_LIMIT:
3608 return (u64)memcg->soft_limit * PAGE_SIZE;
3614 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3616 unsigned long stat[MEMCG_NR_STAT] = {0};
3617 struct mem_cgroup *mi;
3620 for_each_online_cpu(cpu)
3621 for (i = 0; i < MEMCG_NR_STAT; i++)
3622 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3624 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3625 for (i = 0; i < MEMCG_NR_STAT; i++)
3626 atomic_long_add(stat[i], &mi->vmstats[i]);
3628 for_each_node(node) {
3629 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3630 struct mem_cgroup_per_node *pi;
3632 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3635 for_each_online_cpu(cpu)
3636 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3638 pn->lruvec_stat_cpu->count[i], cpu);
3640 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3641 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3642 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3646 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3648 unsigned long events[NR_VM_EVENT_ITEMS];
3649 struct mem_cgroup *mi;
3652 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3655 for_each_online_cpu(cpu)
3656 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3657 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3660 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3661 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3662 atomic_long_add(events[i], &mi->vmevents[i]);
3665 #ifdef CONFIG_MEMCG_KMEM
3666 static int memcg_online_kmem(struct mem_cgroup *memcg)
3668 struct obj_cgroup *objcg;
3671 if (cgroup_memory_nokmem)
3674 BUG_ON(memcg->kmemcg_id >= 0);
3675 BUG_ON(memcg->kmem_state);
3677 memcg_id = memcg_alloc_cache_id();
3681 objcg = obj_cgroup_alloc();
3683 memcg_free_cache_id(memcg_id);
3686 objcg->memcg = memcg;
3687 rcu_assign_pointer(memcg->objcg, objcg);
3689 static_branch_enable(&memcg_kmem_enabled_key);
3691 memcg->kmemcg_id = memcg_id;
3692 memcg->kmem_state = KMEM_ONLINE;
3697 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3699 struct cgroup_subsys_state *css;
3700 struct mem_cgroup *parent, *child;
3703 if (memcg->kmem_state != KMEM_ONLINE)
3706 memcg->kmem_state = KMEM_ALLOCATED;
3708 parent = parent_mem_cgroup(memcg);
3710 parent = root_mem_cgroup;
3712 memcg_reparent_objcgs(memcg, parent);
3714 kmemcg_id = memcg->kmemcg_id;
3715 BUG_ON(kmemcg_id < 0);
3718 * Change kmemcg_id of this cgroup and all its descendants to the
3719 * parent's id, and then move all entries from this cgroup's list_lrus
3720 * to ones of the parent. After we have finished, all list_lrus
3721 * corresponding to this cgroup are guaranteed to remain empty. The
3722 * ordering is imposed by list_lru_node->lock taken by
3723 * memcg_drain_all_list_lrus().
3725 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3726 css_for_each_descendant_pre(css, &memcg->css) {
3727 child = mem_cgroup_from_css(css);
3728 BUG_ON(child->kmemcg_id != kmemcg_id);
3729 child->kmemcg_id = parent->kmemcg_id;
3733 memcg_drain_all_list_lrus(kmemcg_id, parent);
3735 memcg_free_cache_id(kmemcg_id);
3738 static void memcg_free_kmem(struct mem_cgroup *memcg)
3740 /* css_alloc() failed, offlining didn't happen */
3741 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3742 memcg_offline_kmem(memcg);
3745 static int memcg_online_kmem(struct mem_cgroup *memcg)
3749 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3752 static void memcg_free_kmem(struct mem_cgroup *memcg)
3755 #endif /* CONFIG_MEMCG_KMEM */
3757 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3762 mutex_lock(&memcg_max_mutex);
3763 ret = page_counter_set_max(&memcg->kmem, max);
3764 mutex_unlock(&memcg_max_mutex);
3768 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3772 mutex_lock(&memcg_max_mutex);
3774 ret = page_counter_set_max(&memcg->tcpmem, max);
3778 if (!memcg->tcpmem_active) {
3780 * The active flag needs to be written after the static_key
3781 * update. This is what guarantees that the socket activation
3782 * function is the last one to run. See mem_cgroup_sk_alloc()
3783 * for details, and note that we don't mark any socket as
3784 * belonging to this memcg until that flag is up.
3786 * We need to do this, because static_keys will span multiple
3787 * sites, but we can't control their order. If we mark a socket
3788 * as accounted, but the accounting functions are not patched in
3789 * yet, we'll lose accounting.
3791 * We never race with the readers in mem_cgroup_sk_alloc(),
3792 * because when this value change, the code to process it is not
3795 static_branch_inc(&memcg_sockets_enabled_key);
3796 memcg->tcpmem_active = true;
3799 mutex_unlock(&memcg_max_mutex);
3804 * The user of this function is...
3807 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3808 char *buf, size_t nbytes, loff_t off)
3810 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3811 unsigned long nr_pages;
3814 buf = strstrip(buf);
3815 ret = page_counter_memparse(buf, "-1", &nr_pages);
3819 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3821 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3825 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3827 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3830 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3833 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3834 "Please report your usecase to linux-mm@kvack.org if you "
3835 "depend on this functionality.\n");
3836 ret = memcg_update_kmem_max(memcg, nr_pages);
3839 ret = memcg_update_tcp_max(memcg, nr_pages);
3843 case RES_SOFT_LIMIT:
3844 memcg->soft_limit = nr_pages;
3848 return ret ?: nbytes;
3851 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3852 size_t nbytes, loff_t off)
3854 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3855 struct page_counter *counter;
3857 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3859 counter = &memcg->memory;
3862 counter = &memcg->memsw;
3865 counter = &memcg->kmem;
3868 counter = &memcg->tcpmem;
3874 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3876 page_counter_reset_watermark(counter);
3879 counter->failcnt = 0;
3888 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3891 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3895 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3896 struct cftype *cft, u64 val)
3898 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3900 if (val & ~MOVE_MASK)
3904 * No kind of locking is needed in here, because ->can_attach() will
3905 * check this value once in the beginning of the process, and then carry
3906 * on with stale data. This means that changes to this value will only
3907 * affect task migrations starting after the change.
3909 memcg->move_charge_at_immigrate = val;
3913 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3914 struct cftype *cft, u64 val)
3922 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3923 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3924 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3926 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3927 int nid, unsigned int lru_mask, bool tree)
3929 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3930 unsigned long nr = 0;
3933 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3936 if (!(BIT(lru) & lru_mask))
3939 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3941 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3946 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3947 unsigned int lru_mask,
3950 unsigned long nr = 0;
3954 if (!(BIT(lru) & lru_mask))
3957 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3959 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3964 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3968 unsigned int lru_mask;
3971 static const struct numa_stat stats[] = {
3972 { "total", LRU_ALL },
3973 { "file", LRU_ALL_FILE },
3974 { "anon", LRU_ALL_ANON },
3975 { "unevictable", BIT(LRU_UNEVICTABLE) },
3977 const struct numa_stat *stat;
3979 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3981 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3982 seq_printf(m, "%s=%lu", stat->name,
3983 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3985 for_each_node_state(nid, N_MEMORY)
3986 seq_printf(m, " N%d=%lu", nid,
3987 mem_cgroup_node_nr_lru_pages(memcg, nid,
3988 stat->lru_mask, false));
3992 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3994 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3995 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3997 for_each_node_state(nid, N_MEMORY)
3998 seq_printf(m, " N%d=%lu", nid,
3999 mem_cgroup_node_nr_lru_pages(memcg, nid,
4000 stat->lru_mask, true));
4006 #endif /* CONFIG_NUMA */
4008 static const unsigned int memcg1_stats[] = {
4011 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4021 static const char *const memcg1_stat_names[] = {
4024 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4034 /* Universal VM events cgroup1 shows, original sort order */
4035 static const unsigned int memcg1_events[] = {
4042 static int memcg_stat_show(struct seq_file *m, void *v)
4044 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4045 unsigned long memory, memsw;
4046 struct mem_cgroup *mi;
4049 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4051 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4054 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4056 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4057 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4058 if (memcg1_stats[i] == NR_ANON_THPS)
4061 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4064 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4065 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4066 memcg_events_local(memcg, memcg1_events[i]));
4068 for (i = 0; i < NR_LRU_LISTS; i++)
4069 seq_printf(m, "%s %lu\n", lru_list_name(i),
4070 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4073 /* Hierarchical information */
4074 memory = memsw = PAGE_COUNTER_MAX;
4075 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4076 memory = min(memory, READ_ONCE(mi->memory.max));
4077 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4079 seq_printf(m, "hierarchical_memory_limit %llu\n",
4080 (u64)memory * PAGE_SIZE);
4081 if (do_memsw_account())
4082 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4083 (u64)memsw * PAGE_SIZE);
4085 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4088 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4090 nr = memcg_page_state(memcg, memcg1_stats[i]);
4091 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4092 if (memcg1_stats[i] == NR_ANON_THPS)
4095 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4096 (u64)nr * PAGE_SIZE);
4099 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4100 seq_printf(m, "total_%s %llu\n",
4101 vm_event_name(memcg1_events[i]),
4102 (u64)memcg_events(memcg, memcg1_events[i]));
4104 for (i = 0; i < NR_LRU_LISTS; i++)
4105 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4106 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4109 #ifdef CONFIG_DEBUG_VM
4112 struct mem_cgroup_per_node *mz;
4113 unsigned long anon_cost = 0;
4114 unsigned long file_cost = 0;
4116 for_each_online_pgdat(pgdat) {
4117 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4119 anon_cost += mz->lruvec.anon_cost;
4120 file_cost += mz->lruvec.file_cost;
4122 seq_printf(m, "anon_cost %lu\n", anon_cost);
4123 seq_printf(m, "file_cost %lu\n", file_cost);
4130 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4133 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4135 return mem_cgroup_swappiness(memcg);
4138 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4139 struct cftype *cft, u64 val)
4141 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4147 memcg->swappiness = val;
4149 vm_swappiness = val;
4154 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4156 struct mem_cgroup_threshold_ary *t;
4157 unsigned long usage;
4162 t = rcu_dereference(memcg->thresholds.primary);
4164 t = rcu_dereference(memcg->memsw_thresholds.primary);
4169 usage = mem_cgroup_usage(memcg, swap);
4172 * current_threshold points to threshold just below or equal to usage.
4173 * If it's not true, a threshold was crossed after last
4174 * call of __mem_cgroup_threshold().
4176 i = t->current_threshold;
4179 * Iterate backward over array of thresholds starting from
4180 * current_threshold and check if a threshold is crossed.
4181 * If none of thresholds below usage is crossed, we read
4182 * only one element of the array here.
4184 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4185 eventfd_signal(t->entries[i].eventfd, 1);
4187 /* i = current_threshold + 1 */
4191 * Iterate forward over array of thresholds starting from
4192 * current_threshold+1 and check if a threshold is crossed.
4193 * If none of thresholds above usage is crossed, we read
4194 * only one element of the array here.
4196 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4197 eventfd_signal(t->entries[i].eventfd, 1);
4199 /* Update current_threshold */
4200 t->current_threshold = i - 1;
4205 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4208 __mem_cgroup_threshold(memcg, false);
4209 if (do_memsw_account())
4210 __mem_cgroup_threshold(memcg, true);
4212 memcg = parent_mem_cgroup(memcg);
4216 static int compare_thresholds(const void *a, const void *b)
4218 const struct mem_cgroup_threshold *_a = a;
4219 const struct mem_cgroup_threshold *_b = b;
4221 if (_a->threshold > _b->threshold)
4224 if (_a->threshold < _b->threshold)
4230 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4232 struct mem_cgroup_eventfd_list *ev;
4234 spin_lock(&memcg_oom_lock);
4236 list_for_each_entry(ev, &memcg->oom_notify, list)
4237 eventfd_signal(ev->eventfd, 1);
4239 spin_unlock(&memcg_oom_lock);
4243 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4245 struct mem_cgroup *iter;
4247 for_each_mem_cgroup_tree(iter, memcg)
4248 mem_cgroup_oom_notify_cb(iter);
4251 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4252 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4254 struct mem_cgroup_thresholds *thresholds;
4255 struct mem_cgroup_threshold_ary *new;
4256 unsigned long threshold;
4257 unsigned long usage;
4260 ret = page_counter_memparse(args, "-1", &threshold);
4264 mutex_lock(&memcg->thresholds_lock);
4267 thresholds = &memcg->thresholds;
4268 usage = mem_cgroup_usage(memcg, false);
4269 } else if (type == _MEMSWAP) {
4270 thresholds = &memcg->memsw_thresholds;
4271 usage = mem_cgroup_usage(memcg, true);
4275 /* Check if a threshold crossed before adding a new one */
4276 if (thresholds->primary)
4277 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4279 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4281 /* Allocate memory for new array of thresholds */
4282 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4289 /* Copy thresholds (if any) to new array */
4290 if (thresholds->primary)
4291 memcpy(new->entries, thresholds->primary->entries,
4292 flex_array_size(new, entries, size - 1));
4294 /* Add new threshold */
4295 new->entries[size - 1].eventfd = eventfd;
4296 new->entries[size - 1].threshold = threshold;
4298 /* Sort thresholds. Registering of new threshold isn't time-critical */
4299 sort(new->entries, size, sizeof(*new->entries),
4300 compare_thresholds, NULL);
4302 /* Find current threshold */
4303 new->current_threshold = -1;
4304 for (i = 0; i < size; i++) {
4305 if (new->entries[i].threshold <= usage) {
4307 * new->current_threshold will not be used until
4308 * rcu_assign_pointer(), so it's safe to increment
4311 ++new->current_threshold;
4316 /* Free old spare buffer and save old primary buffer as spare */
4317 kfree(thresholds->spare);
4318 thresholds->spare = thresholds->primary;
4320 rcu_assign_pointer(thresholds->primary, new);
4322 /* To be sure that nobody uses thresholds */
4326 mutex_unlock(&memcg->thresholds_lock);
4331 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4332 struct eventfd_ctx *eventfd, const char *args)
4334 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4337 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4338 struct eventfd_ctx *eventfd, const char *args)
4340 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4343 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4344 struct eventfd_ctx *eventfd, enum res_type type)
4346 struct mem_cgroup_thresholds *thresholds;
4347 struct mem_cgroup_threshold_ary *new;
4348 unsigned long usage;
4349 int i, j, size, entries;
4351 mutex_lock(&memcg->thresholds_lock);
4354 thresholds = &memcg->thresholds;
4355 usage = mem_cgroup_usage(memcg, false);
4356 } else if (type == _MEMSWAP) {
4357 thresholds = &memcg->memsw_thresholds;
4358 usage = mem_cgroup_usage(memcg, true);
4362 if (!thresholds->primary)
4365 /* Check if a threshold crossed before removing */
4366 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4368 /* Calculate new number of threshold */
4370 for (i = 0; i < thresholds->primary->size; i++) {
4371 if (thresholds->primary->entries[i].eventfd != eventfd)
4377 new = thresholds->spare;
4379 /* If no items related to eventfd have been cleared, nothing to do */
4383 /* Set thresholds array to NULL if we don't have thresholds */
4392 /* Copy thresholds and find current threshold */
4393 new->current_threshold = -1;
4394 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4395 if (thresholds->primary->entries[i].eventfd == eventfd)
4398 new->entries[j] = thresholds->primary->entries[i];
4399 if (new->entries[j].threshold <= usage) {
4401 * new->current_threshold will not be used
4402 * until rcu_assign_pointer(), so it's safe to increment
4405 ++new->current_threshold;
4411 /* Swap primary and spare array */
4412 thresholds->spare = thresholds->primary;
4414 rcu_assign_pointer(thresholds->primary, new);
4416 /* To be sure that nobody uses thresholds */
4419 /* If all events are unregistered, free the spare array */
4421 kfree(thresholds->spare);
4422 thresholds->spare = NULL;
4425 mutex_unlock(&memcg->thresholds_lock);
4428 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4429 struct eventfd_ctx *eventfd)
4431 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4434 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4435 struct eventfd_ctx *eventfd)
4437 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4440 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4441 struct eventfd_ctx *eventfd, const char *args)
4443 struct mem_cgroup_eventfd_list *event;
4445 event = kmalloc(sizeof(*event), GFP_KERNEL);
4449 spin_lock(&memcg_oom_lock);
4451 event->eventfd = eventfd;
4452 list_add(&event->list, &memcg->oom_notify);
4454 /* already in OOM ? */
4455 if (memcg->under_oom)
4456 eventfd_signal(eventfd, 1);
4457 spin_unlock(&memcg_oom_lock);
4462 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4463 struct eventfd_ctx *eventfd)
4465 struct mem_cgroup_eventfd_list *ev, *tmp;
4467 spin_lock(&memcg_oom_lock);
4469 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4470 if (ev->eventfd == eventfd) {
4471 list_del(&ev->list);
4476 spin_unlock(&memcg_oom_lock);
4479 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4481 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4483 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4484 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4485 seq_printf(sf, "oom_kill %lu\n",
4486 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4490 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4491 struct cftype *cft, u64 val)
4493 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4495 /* cannot set to root cgroup and only 0 and 1 are allowed */
4496 if (!css->parent || !((val == 0) || (val == 1)))
4499 memcg->oom_kill_disable = val;
4501 memcg_oom_recover(memcg);
4506 #ifdef CONFIG_CGROUP_WRITEBACK
4508 #include <trace/events/writeback.h>
4510 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4512 return wb_domain_init(&memcg->cgwb_domain, gfp);
4515 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4517 wb_domain_exit(&memcg->cgwb_domain);
4520 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4522 wb_domain_size_changed(&memcg->cgwb_domain);
4525 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4527 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4529 if (!memcg->css.parent)
4532 return &memcg->cgwb_domain;
4536 * idx can be of type enum memcg_stat_item or node_stat_item.
4537 * Keep in sync with memcg_exact_page().
4539 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4541 long x = atomic_long_read(&memcg->vmstats[idx]);
4544 for_each_online_cpu(cpu)
4545 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4552 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4553 * @wb: bdi_writeback in question
4554 * @pfilepages: out parameter for number of file pages
4555 * @pheadroom: out parameter for number of allocatable pages according to memcg
4556 * @pdirty: out parameter for number of dirty pages
4557 * @pwriteback: out parameter for number of pages under writeback
4559 * Determine the numbers of file, headroom, dirty, and writeback pages in
4560 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4561 * is a bit more involved.
4563 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4564 * headroom is calculated as the lowest headroom of itself and the
4565 * ancestors. Note that this doesn't consider the actual amount of
4566 * available memory in the system. The caller should further cap
4567 * *@pheadroom accordingly.
4569 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4570 unsigned long *pheadroom, unsigned long *pdirty,
4571 unsigned long *pwriteback)
4573 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4574 struct mem_cgroup *parent;
4576 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4578 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4579 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4580 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4581 *pheadroom = PAGE_COUNTER_MAX;
4583 while ((parent = parent_mem_cgroup(memcg))) {
4584 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4585 READ_ONCE(memcg->memory.high));
4586 unsigned long used = page_counter_read(&memcg->memory);
4588 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4594 * Foreign dirty flushing
4596 * There's an inherent mismatch between memcg and writeback. The former
4597 * trackes ownership per-page while the latter per-inode. This was a
4598 * deliberate design decision because honoring per-page ownership in the
4599 * writeback path is complicated, may lead to higher CPU and IO overheads
4600 * and deemed unnecessary given that write-sharing an inode across
4601 * different cgroups isn't a common use-case.
4603 * Combined with inode majority-writer ownership switching, this works well
4604 * enough in most cases but there are some pathological cases. For
4605 * example, let's say there are two cgroups A and B which keep writing to
4606 * different but confined parts of the same inode. B owns the inode and
4607 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4608 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4609 * triggering background writeback. A will be slowed down without a way to
4610 * make writeback of the dirty pages happen.
4612 * Conditions like the above can lead to a cgroup getting repatedly and
4613 * severely throttled after making some progress after each
4614 * dirty_expire_interval while the underyling IO device is almost
4617 * Solving this problem completely requires matching the ownership tracking
4618 * granularities between memcg and writeback in either direction. However,
4619 * the more egregious behaviors can be avoided by simply remembering the
4620 * most recent foreign dirtying events and initiating remote flushes on
4621 * them when local writeback isn't enough to keep the memory clean enough.
4623 * The following two functions implement such mechanism. When a foreign
4624 * page - a page whose memcg and writeback ownerships don't match - is
4625 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4626 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4627 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4628 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4629 * foreign bdi_writebacks which haven't expired. Both the numbers of
4630 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4631 * limited to MEMCG_CGWB_FRN_CNT.
4633 * The mechanism only remembers IDs and doesn't hold any object references.
4634 * As being wrong occasionally doesn't matter, updates and accesses to the
4635 * records are lockless and racy.
4637 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4638 struct bdi_writeback *wb)
4640 struct mem_cgroup *memcg = page->mem_cgroup;
4641 struct memcg_cgwb_frn *frn;
4642 u64 now = get_jiffies_64();
4643 u64 oldest_at = now;
4647 trace_track_foreign_dirty(page, wb);
4650 * Pick the slot to use. If there is already a slot for @wb, keep
4651 * using it. If not replace the oldest one which isn't being
4654 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4655 frn = &memcg->cgwb_frn[i];
4656 if (frn->bdi_id == wb->bdi->id &&
4657 frn->memcg_id == wb->memcg_css->id)
4659 if (time_before64(frn->at, oldest_at) &&
4660 atomic_read(&frn->done.cnt) == 1) {
4662 oldest_at = frn->at;
4666 if (i < MEMCG_CGWB_FRN_CNT) {
4668 * Re-using an existing one. Update timestamp lazily to
4669 * avoid making the cacheline hot. We want them to be
4670 * reasonably up-to-date and significantly shorter than
4671 * dirty_expire_interval as that's what expires the record.
4672 * Use the shorter of 1s and dirty_expire_interval / 8.
4674 unsigned long update_intv =
4675 min_t(unsigned long, HZ,
4676 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4678 if (time_before64(frn->at, now - update_intv))
4680 } else if (oldest >= 0) {
4681 /* replace the oldest free one */
4682 frn = &memcg->cgwb_frn[oldest];
4683 frn->bdi_id = wb->bdi->id;
4684 frn->memcg_id = wb->memcg_css->id;
4689 /* issue foreign writeback flushes for recorded foreign dirtying events */
4690 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4692 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4693 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4694 u64 now = jiffies_64;
4697 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4698 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4701 * If the record is older than dirty_expire_interval,
4702 * writeback on it has already started. No need to kick it
4703 * off again. Also, don't start a new one if there's
4704 * already one in flight.
4706 if (time_after64(frn->at, now - intv) &&
4707 atomic_read(&frn->done.cnt) == 1) {
4709 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4710 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4711 WB_REASON_FOREIGN_FLUSH,
4717 #else /* CONFIG_CGROUP_WRITEBACK */
4719 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4724 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4728 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4732 #endif /* CONFIG_CGROUP_WRITEBACK */
4735 * DO NOT USE IN NEW FILES.
4737 * "cgroup.event_control" implementation.
4739 * This is way over-engineered. It tries to support fully configurable
4740 * events for each user. Such level of flexibility is completely
4741 * unnecessary especially in the light of the planned unified hierarchy.
4743 * Please deprecate this and replace with something simpler if at all
4748 * Unregister event and free resources.
4750 * Gets called from workqueue.
4752 static void memcg_event_remove(struct work_struct *work)
4754 struct mem_cgroup_event *event =
4755 container_of(work, struct mem_cgroup_event, remove);
4756 struct mem_cgroup *memcg = event->memcg;
4758 remove_wait_queue(event->wqh, &event->wait);
4760 event->unregister_event(memcg, event->eventfd);
4762 /* Notify userspace the event is going away. */
4763 eventfd_signal(event->eventfd, 1);
4765 eventfd_ctx_put(event->eventfd);
4767 css_put(&memcg->css);
4771 * Gets called on EPOLLHUP on eventfd when user closes it.
4773 * Called with wqh->lock held and interrupts disabled.
4775 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4776 int sync, void *key)
4778 struct mem_cgroup_event *event =
4779 container_of(wait, struct mem_cgroup_event, wait);
4780 struct mem_cgroup *memcg = event->memcg;
4781 __poll_t flags = key_to_poll(key);
4783 if (flags & EPOLLHUP) {
4785 * If the event has been detached at cgroup removal, we
4786 * can simply return knowing the other side will cleanup
4789 * We can't race against event freeing since the other
4790 * side will require wqh->lock via remove_wait_queue(),
4793 spin_lock(&memcg->event_list_lock);
4794 if (!list_empty(&event->list)) {
4795 list_del_init(&event->list);
4797 * We are in atomic context, but cgroup_event_remove()
4798 * may sleep, so we have to call it in workqueue.
4800 schedule_work(&event->remove);
4802 spin_unlock(&memcg->event_list_lock);
4808 static void memcg_event_ptable_queue_proc(struct file *file,
4809 wait_queue_head_t *wqh, poll_table *pt)
4811 struct mem_cgroup_event *event =
4812 container_of(pt, struct mem_cgroup_event, pt);
4815 add_wait_queue(wqh, &event->wait);
4819 * DO NOT USE IN NEW FILES.
4821 * Parse input and register new cgroup event handler.
4823 * Input must be in format '<event_fd> <control_fd> <args>'.
4824 * Interpretation of args is defined by control file implementation.
4826 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4827 char *buf, size_t nbytes, loff_t off)
4829 struct cgroup_subsys_state *css = of_css(of);
4830 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4831 struct mem_cgroup_event *event;
4832 struct cgroup_subsys_state *cfile_css;
4833 unsigned int efd, cfd;
4840 buf = strstrip(buf);
4842 efd = simple_strtoul(buf, &endp, 10);
4847 cfd = simple_strtoul(buf, &endp, 10);
4848 if ((*endp != ' ') && (*endp != '\0'))
4852 event = kzalloc(sizeof(*event), GFP_KERNEL);
4856 event->memcg = memcg;
4857 INIT_LIST_HEAD(&event->list);
4858 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4859 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4860 INIT_WORK(&event->remove, memcg_event_remove);
4868 event->eventfd = eventfd_ctx_fileget(efile.file);
4869 if (IS_ERR(event->eventfd)) {
4870 ret = PTR_ERR(event->eventfd);
4877 goto out_put_eventfd;
4880 /* the process need read permission on control file */
4881 /* AV: shouldn't we check that it's been opened for read instead? */
4882 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4887 * Determine the event callbacks and set them in @event. This used
4888 * to be done via struct cftype but cgroup core no longer knows
4889 * about these events. The following is crude but the whole thing
4890 * is for compatibility anyway.
4892 * DO NOT ADD NEW FILES.
4894 name = cfile.file->f_path.dentry->d_name.name;
4896 if (!strcmp(name, "memory.usage_in_bytes")) {
4897 event->register_event = mem_cgroup_usage_register_event;
4898 event->unregister_event = mem_cgroup_usage_unregister_event;
4899 } else if (!strcmp(name, "memory.oom_control")) {
4900 event->register_event = mem_cgroup_oom_register_event;
4901 event->unregister_event = mem_cgroup_oom_unregister_event;
4902 } else if (!strcmp(name, "memory.pressure_level")) {
4903 event->register_event = vmpressure_register_event;
4904 event->unregister_event = vmpressure_unregister_event;
4905 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4906 event->register_event = memsw_cgroup_usage_register_event;
4907 event->unregister_event = memsw_cgroup_usage_unregister_event;
4914 * Verify @cfile should belong to @css. Also, remaining events are
4915 * automatically removed on cgroup destruction but the removal is
4916 * asynchronous, so take an extra ref on @css.
4918 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4919 &memory_cgrp_subsys);
4921 if (IS_ERR(cfile_css))
4923 if (cfile_css != css) {
4928 ret = event->register_event(memcg, event->eventfd, buf);
4932 vfs_poll(efile.file, &event->pt);
4934 spin_lock(&memcg->event_list_lock);
4935 list_add(&event->list, &memcg->event_list);
4936 spin_unlock(&memcg->event_list_lock);
4948 eventfd_ctx_put(event->eventfd);
4957 static struct cftype mem_cgroup_legacy_files[] = {
4959 .name = "usage_in_bytes",
4960 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4961 .read_u64 = mem_cgroup_read_u64,
4964 .name = "max_usage_in_bytes",
4965 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4966 .write = mem_cgroup_reset,
4967 .read_u64 = mem_cgroup_read_u64,
4970 .name = "limit_in_bytes",
4971 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4972 .write = mem_cgroup_write,
4973 .read_u64 = mem_cgroup_read_u64,
4976 .name = "soft_limit_in_bytes",
4977 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4978 .write = mem_cgroup_write,
4979 .read_u64 = mem_cgroup_read_u64,
4983 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4984 .write = mem_cgroup_reset,
4985 .read_u64 = mem_cgroup_read_u64,
4989 .seq_show = memcg_stat_show,
4992 .name = "force_empty",
4993 .write = mem_cgroup_force_empty_write,
4996 .name = "use_hierarchy",
4997 .write_u64 = mem_cgroup_hierarchy_write,
4998 .read_u64 = mem_cgroup_hierarchy_read,
5001 .name = "cgroup.event_control", /* XXX: for compat */
5002 .write = memcg_write_event_control,
5003 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
5006 .name = "swappiness",
5007 .read_u64 = mem_cgroup_swappiness_read,
5008 .write_u64 = mem_cgroup_swappiness_write,
5011 .name = "move_charge_at_immigrate",
5012 .read_u64 = mem_cgroup_move_charge_read,
5013 .write_u64 = mem_cgroup_move_charge_write,
5016 .name = "oom_control",
5017 .seq_show = mem_cgroup_oom_control_read,
5018 .write_u64 = mem_cgroup_oom_control_write,
5019 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5022 .name = "pressure_level",
5026 .name = "numa_stat",
5027 .seq_show = memcg_numa_stat_show,
5031 .name = "kmem.limit_in_bytes",
5032 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5033 .write = mem_cgroup_write,
5034 .read_u64 = mem_cgroup_read_u64,
5037 .name = "kmem.usage_in_bytes",
5038 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5039 .read_u64 = mem_cgroup_read_u64,
5042 .name = "kmem.failcnt",
5043 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5044 .write = mem_cgroup_reset,
5045 .read_u64 = mem_cgroup_read_u64,
5048 .name = "kmem.max_usage_in_bytes",
5049 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5050 .write = mem_cgroup_reset,
5051 .read_u64 = mem_cgroup_read_u64,
5053 #if defined(CONFIG_MEMCG_KMEM) && \
5054 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5056 .name = "kmem.slabinfo",
5057 .seq_show = memcg_slab_show,
5061 .name = "kmem.tcp.limit_in_bytes",
5062 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5063 .write = mem_cgroup_write,
5064 .read_u64 = mem_cgroup_read_u64,
5067 .name = "kmem.tcp.usage_in_bytes",
5068 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5069 .read_u64 = mem_cgroup_read_u64,
5072 .name = "kmem.tcp.failcnt",
5073 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5074 .write = mem_cgroup_reset,
5075 .read_u64 = mem_cgroup_read_u64,
5078 .name = "kmem.tcp.max_usage_in_bytes",
5079 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5080 .write = mem_cgroup_reset,
5081 .read_u64 = mem_cgroup_read_u64,
5083 { }, /* terminate */
5087 * Private memory cgroup IDR
5089 * Swap-out records and page cache shadow entries need to store memcg
5090 * references in constrained space, so we maintain an ID space that is
5091 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5092 * memory-controlled cgroups to 64k.
5094 * However, there usually are many references to the offline CSS after
5095 * the cgroup has been destroyed, such as page cache or reclaimable
5096 * slab objects, that don't need to hang on to the ID. We want to keep
5097 * those dead CSS from occupying IDs, or we might quickly exhaust the
5098 * relatively small ID space and prevent the creation of new cgroups
5099 * even when there are much fewer than 64k cgroups - possibly none.
5101 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5102 * be freed and recycled when it's no longer needed, which is usually
5103 * when the CSS is offlined.
5105 * The only exception to that are records of swapped out tmpfs/shmem
5106 * pages that need to be attributed to live ancestors on swapin. But
5107 * those references are manageable from userspace.
5110 static DEFINE_IDR(mem_cgroup_idr);
5112 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5114 if (memcg->id.id > 0) {
5115 idr_remove(&mem_cgroup_idr, memcg->id.id);
5120 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5123 refcount_add(n, &memcg->id.ref);
5126 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5128 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5129 mem_cgroup_id_remove(memcg);
5131 /* Memcg ID pins CSS */
5132 css_put(&memcg->css);
5136 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5138 mem_cgroup_id_put_many(memcg, 1);
5142 * mem_cgroup_from_id - look up a memcg from a memcg id
5143 * @id: the memcg id to look up
5145 * Caller must hold rcu_read_lock().
5147 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5149 WARN_ON_ONCE(!rcu_read_lock_held());
5150 return idr_find(&mem_cgroup_idr, id);
5153 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5155 struct mem_cgroup_per_node *pn;
5158 * This routine is called against possible nodes.
5159 * But it's BUG to call kmalloc() against offline node.
5161 * TODO: this routine can waste much memory for nodes which will
5162 * never be onlined. It's better to use memory hotplug callback
5165 if (!node_state(node, N_NORMAL_MEMORY))
5167 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5171 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5172 GFP_KERNEL_ACCOUNT);
5173 if (!pn->lruvec_stat_local) {
5178 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5179 GFP_KERNEL_ACCOUNT);
5180 if (!pn->lruvec_stat_cpu) {
5181 free_percpu(pn->lruvec_stat_local);
5186 lruvec_init(&pn->lruvec);
5187 pn->usage_in_excess = 0;
5188 pn->on_tree = false;
5191 memcg->nodeinfo[node] = pn;
5195 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5197 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5202 free_percpu(pn->lruvec_stat_cpu);
5203 free_percpu(pn->lruvec_stat_local);
5207 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5212 free_mem_cgroup_per_node_info(memcg, node);
5213 free_percpu(memcg->vmstats_percpu);
5214 free_percpu(memcg->vmstats_local);
5218 static void mem_cgroup_free(struct mem_cgroup *memcg)
5220 memcg_wb_domain_exit(memcg);
5222 * Flush percpu vmstats and vmevents to guarantee the value correctness
5223 * on parent's and all ancestor levels.
5225 memcg_flush_percpu_vmstats(memcg);
5226 memcg_flush_percpu_vmevents(memcg);
5227 __mem_cgroup_free(memcg);
5230 static struct mem_cgroup *mem_cgroup_alloc(void)
5232 struct mem_cgroup *memcg;
5235 int __maybe_unused i;
5236 long error = -ENOMEM;
5238 size = sizeof(struct mem_cgroup);
5239 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5241 memcg = kzalloc(size, GFP_KERNEL);
5243 return ERR_PTR(error);
5245 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5246 1, MEM_CGROUP_ID_MAX,
5248 if (memcg->id.id < 0) {
5249 error = memcg->id.id;
5253 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5254 GFP_KERNEL_ACCOUNT);
5255 if (!memcg->vmstats_local)
5258 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5259 GFP_KERNEL_ACCOUNT);
5260 if (!memcg->vmstats_percpu)
5264 if (alloc_mem_cgroup_per_node_info(memcg, node))
5267 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5270 INIT_WORK(&memcg->high_work, high_work_func);
5271 INIT_LIST_HEAD(&memcg->oom_notify);
5272 mutex_init(&memcg->thresholds_lock);
5273 spin_lock_init(&memcg->move_lock);
5274 vmpressure_init(&memcg->vmpressure);
5275 INIT_LIST_HEAD(&memcg->event_list);
5276 spin_lock_init(&memcg->event_list_lock);
5277 memcg->socket_pressure = jiffies;
5278 #ifdef CONFIG_MEMCG_KMEM
5279 memcg->kmemcg_id = -1;
5280 INIT_LIST_HEAD(&memcg->objcg_list);
5282 #ifdef CONFIG_CGROUP_WRITEBACK
5283 INIT_LIST_HEAD(&memcg->cgwb_list);
5284 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5285 memcg->cgwb_frn[i].done =
5286 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5288 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5289 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5290 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5291 memcg->deferred_split_queue.split_queue_len = 0;
5293 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5296 mem_cgroup_id_remove(memcg);
5297 __mem_cgroup_free(memcg);
5298 return ERR_PTR(error);
5301 static struct cgroup_subsys_state * __ref
5302 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5304 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5305 struct mem_cgroup *memcg, *old_memcg;
5306 long error = -ENOMEM;
5308 old_memcg = set_active_memcg(parent);
5309 memcg = mem_cgroup_alloc();
5310 set_active_memcg(old_memcg);
5312 return ERR_CAST(memcg);
5314 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5315 memcg->soft_limit = PAGE_COUNTER_MAX;
5316 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5318 memcg->swappiness = mem_cgroup_swappiness(parent);
5319 memcg->oom_kill_disable = parent->oom_kill_disable;
5321 page_counter_init(&memcg->memory, &parent->memory);
5322 page_counter_init(&memcg->swap, &parent->swap);
5323 page_counter_init(&memcg->kmem, &parent->kmem);
5324 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5326 page_counter_init(&memcg->memory, NULL);
5327 page_counter_init(&memcg->swap, NULL);
5328 page_counter_init(&memcg->kmem, NULL);
5329 page_counter_init(&memcg->tcpmem, NULL);
5331 root_mem_cgroup = memcg;
5335 /* The following stuff does not apply to the root */
5336 error = memcg_online_kmem(memcg);
5340 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5341 static_branch_inc(&memcg_sockets_enabled_key);
5345 mem_cgroup_id_remove(memcg);
5346 mem_cgroup_free(memcg);
5347 return ERR_PTR(error);
5350 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5352 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5355 * A memcg must be visible for memcg_expand_shrinker_maps()
5356 * by the time the maps are allocated. So, we allocate maps
5357 * here, when for_each_mem_cgroup() can't skip it.
5359 if (memcg_alloc_shrinker_maps(memcg)) {
5360 mem_cgroup_id_remove(memcg);
5364 /* Online state pins memcg ID, memcg ID pins CSS */
5365 refcount_set(&memcg->id.ref, 1);
5370 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5372 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5373 struct mem_cgroup_event *event, *tmp;
5376 * Unregister events and notify userspace.
5377 * Notify userspace about cgroup removing only after rmdir of cgroup
5378 * directory to avoid race between userspace and kernelspace.
5380 spin_lock(&memcg->event_list_lock);
5381 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5382 list_del_init(&event->list);
5383 schedule_work(&event->remove);
5385 spin_unlock(&memcg->event_list_lock);
5387 page_counter_set_min(&memcg->memory, 0);
5388 page_counter_set_low(&memcg->memory, 0);
5390 memcg_offline_kmem(memcg);
5391 wb_memcg_offline(memcg);
5393 drain_all_stock(memcg);
5395 mem_cgroup_id_put(memcg);
5398 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5400 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5402 invalidate_reclaim_iterators(memcg);
5405 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5407 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5408 int __maybe_unused i;
5410 #ifdef CONFIG_CGROUP_WRITEBACK
5411 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5412 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5414 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5415 static_branch_dec(&memcg_sockets_enabled_key);
5417 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5418 static_branch_dec(&memcg_sockets_enabled_key);
5420 vmpressure_cleanup(&memcg->vmpressure);
5421 cancel_work_sync(&memcg->high_work);
5422 mem_cgroup_remove_from_trees(memcg);
5423 memcg_free_shrinker_maps(memcg);
5424 memcg_free_kmem(memcg);
5425 mem_cgroup_free(memcg);
5429 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5430 * @css: the target css
5432 * Reset the states of the mem_cgroup associated with @css. This is
5433 * invoked when the userland requests disabling on the default hierarchy
5434 * but the memcg is pinned through dependency. The memcg should stop
5435 * applying policies and should revert to the vanilla state as it may be
5436 * made visible again.
5438 * The current implementation only resets the essential configurations.
5439 * This needs to be expanded to cover all the visible parts.
5441 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5443 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5445 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5446 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5447 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5448 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5449 page_counter_set_min(&memcg->memory, 0);
5450 page_counter_set_low(&memcg->memory, 0);
5451 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5452 memcg->soft_limit = PAGE_COUNTER_MAX;
5453 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5454 memcg_wb_domain_size_changed(memcg);
5458 /* Handlers for move charge at task migration. */
5459 static int mem_cgroup_do_precharge(unsigned long count)
5463 /* Try a single bulk charge without reclaim first, kswapd may wake */
5464 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5466 mc.precharge += count;
5470 /* Try charges one by one with reclaim, but do not retry */
5472 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5486 enum mc_target_type {
5493 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5494 unsigned long addr, pte_t ptent)
5496 struct page *page = vm_normal_page(vma, addr, ptent);
5498 if (!page || !page_mapped(page))
5500 if (PageAnon(page)) {
5501 if (!(mc.flags & MOVE_ANON))
5504 if (!(mc.flags & MOVE_FILE))
5507 if (!get_page_unless_zero(page))
5513 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5514 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5515 pte_t ptent, swp_entry_t *entry)
5517 struct page *page = NULL;
5518 swp_entry_t ent = pte_to_swp_entry(ptent);
5520 if (!(mc.flags & MOVE_ANON))
5524 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5525 * a device and because they are not accessible by CPU they are store
5526 * as special swap entry in the CPU page table.
5528 if (is_device_private_entry(ent)) {
5529 page = device_private_entry_to_page(ent);
5531 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5532 * a refcount of 1 when free (unlike normal page)
5534 if (!page_ref_add_unless(page, 1, 1))
5539 if (non_swap_entry(ent))
5543 * Because lookup_swap_cache() updates some statistics counter,
5544 * we call find_get_page() with swapper_space directly.
5546 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5547 entry->val = ent.val;
5552 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5553 pte_t ptent, swp_entry_t *entry)
5559 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5560 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5562 if (!vma->vm_file) /* anonymous vma */
5564 if (!(mc.flags & MOVE_FILE))
5567 /* page is moved even if it's not RSS of this task(page-faulted). */
5568 /* shmem/tmpfs may report page out on swap: account for that too. */
5569 return find_get_incore_page(vma->vm_file->f_mapping,
5570 linear_page_index(vma, addr));
5574 * mem_cgroup_move_account - move account of the page
5576 * @compound: charge the page as compound or small page
5577 * @from: mem_cgroup which the page is moved from.
5578 * @to: mem_cgroup which the page is moved to. @from != @to.
5580 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5582 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5585 static int mem_cgroup_move_account(struct page *page,
5587 struct mem_cgroup *from,
5588 struct mem_cgroup *to)
5590 struct lruvec *from_vec, *to_vec;
5591 struct pglist_data *pgdat;
5592 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5595 VM_BUG_ON(from == to);
5596 VM_BUG_ON_PAGE(PageLRU(page), page);
5597 VM_BUG_ON(compound && !PageTransHuge(page));
5600 * Prevent mem_cgroup_migrate() from looking at
5601 * page->mem_cgroup of its source page while we change it.
5604 if (!trylock_page(page))
5608 if (page->mem_cgroup != from)
5611 pgdat = page_pgdat(page);
5612 from_vec = mem_cgroup_lruvec(from, pgdat);
5613 to_vec = mem_cgroup_lruvec(to, pgdat);
5615 lock_page_memcg(page);
5617 if (PageAnon(page)) {
5618 if (page_mapped(page)) {
5619 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5620 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5621 if (PageTransHuge(page)) {
5622 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5624 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5630 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5631 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5633 if (PageSwapBacked(page)) {
5634 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5635 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5638 if (page_mapped(page)) {
5639 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5640 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5643 if (PageDirty(page)) {
5644 struct address_space *mapping = page_mapping(page);
5646 if (mapping_can_writeback(mapping)) {
5647 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5649 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5655 if (PageWriteback(page)) {
5656 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5657 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5661 * All state has been migrated, let's switch to the new memcg.
5663 * It is safe to change page->mem_cgroup here because the page
5664 * is referenced, charged, isolated, and locked: we can't race
5665 * with (un)charging, migration, LRU putback, or anything else
5666 * that would rely on a stable page->mem_cgroup.
5668 * Note that lock_page_memcg is a memcg lock, not a page lock,
5669 * to save space. As soon as we switch page->mem_cgroup to a
5670 * new memcg that isn't locked, the above state can change
5671 * concurrently again. Make sure we're truly done with it.
5676 css_put(&from->css);
5678 page->mem_cgroup = to;
5680 __unlock_page_memcg(from);
5684 local_irq_disable();
5685 mem_cgroup_charge_statistics(to, page, nr_pages);
5686 memcg_check_events(to, page);
5687 mem_cgroup_charge_statistics(from, page, -nr_pages);
5688 memcg_check_events(from, page);
5697 * get_mctgt_type - get target type of moving charge
5698 * @vma: the vma the pte to be checked belongs
5699 * @addr: the address corresponding to the pte to be checked
5700 * @ptent: the pte to be checked
5701 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5704 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5705 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5706 * move charge. if @target is not NULL, the page is stored in target->page
5707 * with extra refcnt got(Callers should handle it).
5708 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5709 * target for charge migration. if @target is not NULL, the entry is stored
5711 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5712 * (so ZONE_DEVICE page and thus not on the lru).
5713 * For now we such page is charge like a regular page would be as for all
5714 * intent and purposes it is just special memory taking the place of a
5717 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5719 * Called with pte lock held.
5722 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5723 unsigned long addr, pte_t ptent, union mc_target *target)
5725 struct page *page = NULL;
5726 enum mc_target_type ret = MC_TARGET_NONE;
5727 swp_entry_t ent = { .val = 0 };
5729 if (pte_present(ptent))
5730 page = mc_handle_present_pte(vma, addr, ptent);
5731 else if (is_swap_pte(ptent))
5732 page = mc_handle_swap_pte(vma, ptent, &ent);
5733 else if (pte_none(ptent))
5734 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5736 if (!page && !ent.val)
5740 * Do only loose check w/o serialization.
5741 * mem_cgroup_move_account() checks the page is valid or
5742 * not under LRU exclusion.
5744 if (page->mem_cgroup == mc.from) {
5745 ret = MC_TARGET_PAGE;
5746 if (is_device_private_page(page))
5747 ret = MC_TARGET_DEVICE;
5749 target->page = page;
5751 if (!ret || !target)
5755 * There is a swap entry and a page doesn't exist or isn't charged.
5756 * But we cannot move a tail-page in a THP.
5758 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5759 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5760 ret = MC_TARGET_SWAP;
5767 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5769 * We don't consider PMD mapped swapping or file mapped pages because THP does
5770 * not support them for now.
5771 * Caller should make sure that pmd_trans_huge(pmd) is true.
5773 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5774 unsigned long addr, pmd_t pmd, union mc_target *target)
5776 struct page *page = NULL;
5777 enum mc_target_type ret = MC_TARGET_NONE;
5779 if (unlikely(is_swap_pmd(pmd))) {
5780 VM_BUG_ON(thp_migration_supported() &&
5781 !is_pmd_migration_entry(pmd));
5784 page = pmd_page(pmd);
5785 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5786 if (!(mc.flags & MOVE_ANON))
5788 if (page->mem_cgroup == mc.from) {
5789 ret = MC_TARGET_PAGE;
5792 target->page = page;
5798 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5799 unsigned long addr, pmd_t pmd, union mc_target *target)
5801 return MC_TARGET_NONE;
5805 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5806 unsigned long addr, unsigned long end,
5807 struct mm_walk *walk)
5809 struct vm_area_struct *vma = walk->vma;
5813 ptl = pmd_trans_huge_lock(pmd, vma);
5816 * Note their can not be MC_TARGET_DEVICE for now as we do not
5817 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5818 * this might change.
5820 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5821 mc.precharge += HPAGE_PMD_NR;
5826 if (pmd_trans_unstable(pmd))
5828 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5829 for (; addr != end; pte++, addr += PAGE_SIZE)
5830 if (get_mctgt_type(vma, addr, *pte, NULL))
5831 mc.precharge++; /* increment precharge temporarily */
5832 pte_unmap_unlock(pte - 1, ptl);
5838 static const struct mm_walk_ops precharge_walk_ops = {
5839 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5842 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5844 unsigned long precharge;
5847 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5848 mmap_read_unlock(mm);
5850 precharge = mc.precharge;
5856 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5858 unsigned long precharge = mem_cgroup_count_precharge(mm);
5860 VM_BUG_ON(mc.moving_task);
5861 mc.moving_task = current;
5862 return mem_cgroup_do_precharge(precharge);
5865 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5866 static void __mem_cgroup_clear_mc(void)
5868 struct mem_cgroup *from = mc.from;
5869 struct mem_cgroup *to = mc.to;
5871 /* we must uncharge all the leftover precharges from mc.to */
5873 cancel_charge(mc.to, mc.precharge);
5877 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5878 * we must uncharge here.
5880 if (mc.moved_charge) {
5881 cancel_charge(mc.from, mc.moved_charge);
5882 mc.moved_charge = 0;
5884 /* we must fixup refcnts and charges */
5885 if (mc.moved_swap) {
5886 /* uncharge swap account from the old cgroup */
5887 if (!mem_cgroup_is_root(mc.from))
5888 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5890 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5893 * we charged both to->memory and to->memsw, so we
5894 * should uncharge to->memory.
5896 if (!mem_cgroup_is_root(mc.to))
5897 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5901 memcg_oom_recover(from);
5902 memcg_oom_recover(to);
5903 wake_up_all(&mc.waitq);
5906 static void mem_cgroup_clear_mc(void)
5908 struct mm_struct *mm = mc.mm;
5911 * we must clear moving_task before waking up waiters at the end of
5914 mc.moving_task = NULL;
5915 __mem_cgroup_clear_mc();
5916 spin_lock(&mc.lock);
5920 spin_unlock(&mc.lock);
5925 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5927 struct cgroup_subsys_state *css;
5928 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5929 struct mem_cgroup *from;
5930 struct task_struct *leader, *p;
5931 struct mm_struct *mm;
5932 unsigned long move_flags;
5935 /* charge immigration isn't supported on the default hierarchy */
5936 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5940 * Multi-process migrations only happen on the default hierarchy
5941 * where charge immigration is not used. Perform charge
5942 * immigration if @tset contains a leader and whine if there are
5946 cgroup_taskset_for_each_leader(leader, css, tset) {
5949 memcg = mem_cgroup_from_css(css);
5955 * We are now commited to this value whatever it is. Changes in this
5956 * tunable will only affect upcoming migrations, not the current one.
5957 * So we need to save it, and keep it going.
5959 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5963 from = mem_cgroup_from_task(p);
5965 VM_BUG_ON(from == memcg);
5967 mm = get_task_mm(p);
5970 /* We move charges only when we move a owner of the mm */
5971 if (mm->owner == p) {
5974 VM_BUG_ON(mc.precharge);
5975 VM_BUG_ON(mc.moved_charge);
5976 VM_BUG_ON(mc.moved_swap);
5978 spin_lock(&mc.lock);
5982 mc.flags = move_flags;
5983 spin_unlock(&mc.lock);
5984 /* We set mc.moving_task later */
5986 ret = mem_cgroup_precharge_mc(mm);
5988 mem_cgroup_clear_mc();
5995 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5998 mem_cgroup_clear_mc();
6001 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6002 unsigned long addr, unsigned long end,
6003 struct mm_walk *walk)
6006 struct vm_area_struct *vma = walk->vma;
6009 enum mc_target_type target_type;
6010 union mc_target target;
6013 ptl = pmd_trans_huge_lock(pmd, vma);
6015 if (mc.precharge < HPAGE_PMD_NR) {
6019 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6020 if (target_type == MC_TARGET_PAGE) {
6022 if (!isolate_lru_page(page)) {
6023 if (!mem_cgroup_move_account(page, true,
6025 mc.precharge -= HPAGE_PMD_NR;
6026 mc.moved_charge += HPAGE_PMD_NR;
6028 putback_lru_page(page);
6031 } else if (target_type == MC_TARGET_DEVICE) {
6033 if (!mem_cgroup_move_account(page, true,
6035 mc.precharge -= HPAGE_PMD_NR;
6036 mc.moved_charge += HPAGE_PMD_NR;
6044 if (pmd_trans_unstable(pmd))
6047 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6048 for (; addr != end; addr += PAGE_SIZE) {
6049 pte_t ptent = *(pte++);
6050 bool device = false;
6056 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6057 case MC_TARGET_DEVICE:
6060 case MC_TARGET_PAGE:
6063 * We can have a part of the split pmd here. Moving it
6064 * can be done but it would be too convoluted so simply
6065 * ignore such a partial THP and keep it in original
6066 * memcg. There should be somebody mapping the head.
6068 if (PageTransCompound(page))
6070 if (!device && isolate_lru_page(page))
6072 if (!mem_cgroup_move_account(page, false,
6075 /* we uncharge from mc.from later. */
6079 putback_lru_page(page);
6080 put: /* get_mctgt_type() gets the page */
6083 case MC_TARGET_SWAP:
6085 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6087 mem_cgroup_id_get_many(mc.to, 1);
6088 /* we fixup other refcnts and charges later. */
6096 pte_unmap_unlock(pte - 1, ptl);
6101 * We have consumed all precharges we got in can_attach().
6102 * We try charge one by one, but don't do any additional
6103 * charges to mc.to if we have failed in charge once in attach()
6106 ret = mem_cgroup_do_precharge(1);
6114 static const struct mm_walk_ops charge_walk_ops = {
6115 .pmd_entry = mem_cgroup_move_charge_pte_range,
6118 static void mem_cgroup_move_charge(void)
6120 lru_add_drain_all();
6122 * Signal lock_page_memcg() to take the memcg's move_lock
6123 * while we're moving its pages to another memcg. Then wait
6124 * for already started RCU-only updates to finish.
6126 atomic_inc(&mc.from->moving_account);
6129 if (unlikely(!mmap_read_trylock(mc.mm))) {
6131 * Someone who are holding the mmap_lock might be waiting in
6132 * waitq. So we cancel all extra charges, wake up all waiters,
6133 * and retry. Because we cancel precharges, we might not be able
6134 * to move enough charges, but moving charge is a best-effort
6135 * feature anyway, so it wouldn't be a big problem.
6137 __mem_cgroup_clear_mc();
6142 * When we have consumed all precharges and failed in doing
6143 * additional charge, the page walk just aborts.
6145 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6148 mmap_read_unlock(mc.mm);
6149 atomic_dec(&mc.from->moving_account);
6152 static void mem_cgroup_move_task(void)
6155 mem_cgroup_move_charge();
6156 mem_cgroup_clear_mc();
6159 #else /* !CONFIG_MMU */
6160 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6164 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6167 static void mem_cgroup_move_task(void)
6172 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6174 if (value == PAGE_COUNTER_MAX)
6175 seq_puts(m, "max\n");
6177 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6182 static u64 memory_current_read(struct cgroup_subsys_state *css,
6185 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6187 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6190 static int memory_min_show(struct seq_file *m, void *v)
6192 return seq_puts_memcg_tunable(m,
6193 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6196 static ssize_t memory_min_write(struct kernfs_open_file *of,
6197 char *buf, size_t nbytes, loff_t off)
6199 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6203 buf = strstrip(buf);
6204 err = page_counter_memparse(buf, "max", &min);
6208 page_counter_set_min(&memcg->memory, min);
6213 static int memory_low_show(struct seq_file *m, void *v)
6215 return seq_puts_memcg_tunable(m,
6216 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6219 static ssize_t memory_low_write(struct kernfs_open_file *of,
6220 char *buf, size_t nbytes, loff_t off)
6222 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6226 buf = strstrip(buf);
6227 err = page_counter_memparse(buf, "max", &low);
6231 page_counter_set_low(&memcg->memory, low);
6236 static int memory_high_show(struct seq_file *m, void *v)
6238 return seq_puts_memcg_tunable(m,
6239 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6242 static ssize_t memory_high_write(struct kernfs_open_file *of,
6243 char *buf, size_t nbytes, loff_t off)
6245 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6246 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6247 bool drained = false;
6251 buf = strstrip(buf);
6252 err = page_counter_memparse(buf, "max", &high);
6257 unsigned long nr_pages = page_counter_read(&memcg->memory);
6258 unsigned long reclaimed;
6260 if (nr_pages <= high)
6263 if (signal_pending(current))
6267 drain_all_stock(memcg);
6272 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6275 if (!reclaimed && !nr_retries--)
6279 page_counter_set_high(&memcg->memory, high);
6281 memcg_wb_domain_size_changed(memcg);
6286 static int memory_max_show(struct seq_file *m, void *v)
6288 return seq_puts_memcg_tunable(m,
6289 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6292 static ssize_t memory_max_write(struct kernfs_open_file *of,
6293 char *buf, size_t nbytes, loff_t off)
6295 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6296 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6297 bool drained = false;
6301 buf = strstrip(buf);
6302 err = page_counter_memparse(buf, "max", &max);
6306 xchg(&memcg->memory.max, max);
6309 unsigned long nr_pages = page_counter_read(&memcg->memory);
6311 if (nr_pages <= max)
6314 if (signal_pending(current))
6318 drain_all_stock(memcg);
6324 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6330 memcg_memory_event(memcg, MEMCG_OOM);
6331 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6335 memcg_wb_domain_size_changed(memcg);
6339 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6341 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6342 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6343 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6344 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6345 seq_printf(m, "oom_kill %lu\n",
6346 atomic_long_read(&events[MEMCG_OOM_KILL]));
6349 static int memory_events_show(struct seq_file *m, void *v)
6351 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6353 __memory_events_show(m, memcg->memory_events);
6357 static int memory_events_local_show(struct seq_file *m, void *v)
6359 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6361 __memory_events_show(m, memcg->memory_events_local);
6365 static int memory_stat_show(struct seq_file *m, void *v)
6367 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6370 buf = memory_stat_format(memcg);
6379 static int memory_numa_stat_show(struct seq_file *m, void *v)
6382 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6384 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6387 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6390 seq_printf(m, "%s", memory_stats[i].name);
6391 for_each_node_state(nid, N_MEMORY) {
6393 struct lruvec *lruvec;
6395 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6396 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6397 size *= memory_stats[i].ratio;
6398 seq_printf(m, " N%d=%llu", nid, size);
6407 static int memory_oom_group_show(struct seq_file *m, void *v)
6409 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6411 seq_printf(m, "%d\n", memcg->oom_group);
6416 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6417 char *buf, size_t nbytes, loff_t off)
6419 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6422 buf = strstrip(buf);
6426 ret = kstrtoint(buf, 0, &oom_group);
6430 if (oom_group != 0 && oom_group != 1)
6433 memcg->oom_group = oom_group;
6438 static struct cftype memory_files[] = {
6441 .flags = CFTYPE_NOT_ON_ROOT,
6442 .read_u64 = memory_current_read,
6446 .flags = CFTYPE_NOT_ON_ROOT,
6447 .seq_show = memory_min_show,
6448 .write = memory_min_write,
6452 .flags = CFTYPE_NOT_ON_ROOT,
6453 .seq_show = memory_low_show,
6454 .write = memory_low_write,
6458 .flags = CFTYPE_NOT_ON_ROOT,
6459 .seq_show = memory_high_show,
6460 .write = memory_high_write,
6464 .flags = CFTYPE_NOT_ON_ROOT,
6465 .seq_show = memory_max_show,
6466 .write = memory_max_write,
6470 .flags = CFTYPE_NOT_ON_ROOT,
6471 .file_offset = offsetof(struct mem_cgroup, events_file),
6472 .seq_show = memory_events_show,
6475 .name = "events.local",
6476 .flags = CFTYPE_NOT_ON_ROOT,
6477 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6478 .seq_show = memory_events_local_show,
6482 .seq_show = memory_stat_show,
6486 .name = "numa_stat",
6487 .seq_show = memory_numa_stat_show,
6491 .name = "oom.group",
6492 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6493 .seq_show = memory_oom_group_show,
6494 .write = memory_oom_group_write,
6499 struct cgroup_subsys memory_cgrp_subsys = {
6500 .css_alloc = mem_cgroup_css_alloc,
6501 .css_online = mem_cgroup_css_online,
6502 .css_offline = mem_cgroup_css_offline,
6503 .css_released = mem_cgroup_css_released,
6504 .css_free = mem_cgroup_css_free,
6505 .css_reset = mem_cgroup_css_reset,
6506 .can_attach = mem_cgroup_can_attach,
6507 .cancel_attach = mem_cgroup_cancel_attach,
6508 .post_attach = mem_cgroup_move_task,
6509 .dfl_cftypes = memory_files,
6510 .legacy_cftypes = mem_cgroup_legacy_files,
6515 * This function calculates an individual cgroup's effective
6516 * protection which is derived from its own memory.min/low, its
6517 * parent's and siblings' settings, as well as the actual memory
6518 * distribution in the tree.
6520 * The following rules apply to the effective protection values:
6522 * 1. At the first level of reclaim, effective protection is equal to
6523 * the declared protection in memory.min and memory.low.
6525 * 2. To enable safe delegation of the protection configuration, at
6526 * subsequent levels the effective protection is capped to the
6527 * parent's effective protection.
6529 * 3. To make complex and dynamic subtrees easier to configure, the
6530 * user is allowed to overcommit the declared protection at a given
6531 * level. If that is the case, the parent's effective protection is
6532 * distributed to the children in proportion to how much protection
6533 * they have declared and how much of it they are utilizing.
6535 * This makes distribution proportional, but also work-conserving:
6536 * if one cgroup claims much more protection than it uses memory,
6537 * the unused remainder is available to its siblings.
6539 * 4. Conversely, when the declared protection is undercommitted at a
6540 * given level, the distribution of the larger parental protection
6541 * budget is NOT proportional. A cgroup's protection from a sibling
6542 * is capped to its own memory.min/low setting.
6544 * 5. However, to allow protecting recursive subtrees from each other
6545 * without having to declare each individual cgroup's fixed share
6546 * of the ancestor's claim to protection, any unutilized -
6547 * "floating" - protection from up the tree is distributed in
6548 * proportion to each cgroup's *usage*. This makes the protection
6549 * neutral wrt sibling cgroups and lets them compete freely over
6550 * the shared parental protection budget, but it protects the
6551 * subtree as a whole from neighboring subtrees.
6553 * Note that 4. and 5. are not in conflict: 4. is about protecting
6554 * against immediate siblings whereas 5. is about protecting against
6555 * neighboring subtrees.
6557 static unsigned long effective_protection(unsigned long usage,
6558 unsigned long parent_usage,
6559 unsigned long setting,
6560 unsigned long parent_effective,
6561 unsigned long siblings_protected)
6563 unsigned long protected;
6566 protected = min(usage, setting);
6568 * If all cgroups at this level combined claim and use more
6569 * protection then what the parent affords them, distribute
6570 * shares in proportion to utilization.
6572 * We are using actual utilization rather than the statically
6573 * claimed protection in order to be work-conserving: claimed
6574 * but unused protection is available to siblings that would
6575 * otherwise get a smaller chunk than what they claimed.
6577 if (siblings_protected > parent_effective)
6578 return protected * parent_effective / siblings_protected;
6581 * Ok, utilized protection of all children is within what the
6582 * parent affords them, so we know whatever this child claims
6583 * and utilizes is effectively protected.
6585 * If there is unprotected usage beyond this value, reclaim
6586 * will apply pressure in proportion to that amount.
6588 * If there is unutilized protection, the cgroup will be fully
6589 * shielded from reclaim, but we do return a smaller value for
6590 * protection than what the group could enjoy in theory. This
6591 * is okay. With the overcommit distribution above, effective
6592 * protection is always dependent on how memory is actually
6593 * consumed among the siblings anyway.
6598 * If the children aren't claiming (all of) the protection
6599 * afforded to them by the parent, distribute the remainder in
6600 * proportion to the (unprotected) memory of each cgroup. That
6601 * way, cgroups that aren't explicitly prioritized wrt each
6602 * other compete freely over the allowance, but they are
6603 * collectively protected from neighboring trees.
6605 * We're using unprotected memory for the weight so that if
6606 * some cgroups DO claim explicit protection, we don't protect
6607 * the same bytes twice.
6609 * Check both usage and parent_usage against the respective
6610 * protected values. One should imply the other, but they
6611 * aren't read atomically - make sure the division is sane.
6613 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6615 if (parent_effective > siblings_protected &&
6616 parent_usage > siblings_protected &&
6617 usage > protected) {
6618 unsigned long unclaimed;
6620 unclaimed = parent_effective - siblings_protected;
6621 unclaimed *= usage - protected;
6622 unclaimed /= parent_usage - siblings_protected;
6631 * mem_cgroup_protected - check if memory consumption is in the normal range
6632 * @root: the top ancestor of the sub-tree being checked
6633 * @memcg: the memory cgroup to check
6635 * WARNING: This function is not stateless! It can only be used as part
6636 * of a top-down tree iteration, not for isolated queries.
6638 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6639 struct mem_cgroup *memcg)
6641 unsigned long usage, parent_usage;
6642 struct mem_cgroup *parent;
6644 if (mem_cgroup_disabled())
6648 root = root_mem_cgroup;
6651 * Effective values of the reclaim targets are ignored so they
6652 * can be stale. Have a look at mem_cgroup_protection for more
6654 * TODO: calculation should be more robust so that we do not need
6655 * that special casing.
6660 usage = page_counter_read(&memcg->memory);
6664 parent = parent_mem_cgroup(memcg);
6665 /* No parent means a non-hierarchical mode on v1 memcg */
6669 if (parent == root) {
6670 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6671 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6675 parent_usage = page_counter_read(&parent->memory);
6677 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6678 READ_ONCE(memcg->memory.min),
6679 READ_ONCE(parent->memory.emin),
6680 atomic_long_read(&parent->memory.children_min_usage)));
6682 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6683 READ_ONCE(memcg->memory.low),
6684 READ_ONCE(parent->memory.elow),
6685 atomic_long_read(&parent->memory.children_low_usage)));
6689 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6690 * @page: page to charge
6691 * @mm: mm context of the victim
6692 * @gfp_mask: reclaim mode
6694 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6695 * pages according to @gfp_mask if necessary.
6697 * Returns 0 on success. Otherwise, an error code is returned.
6699 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6701 unsigned int nr_pages = thp_nr_pages(page);
6702 struct mem_cgroup *memcg = NULL;
6705 if (mem_cgroup_disabled())
6708 if (PageSwapCache(page)) {
6709 swp_entry_t ent = { .val = page_private(page), };
6713 * Every swap fault against a single page tries to charge the
6714 * page, bail as early as possible. shmem_unuse() encounters
6715 * already charged pages, too. page->mem_cgroup is protected
6716 * by the page lock, which serializes swap cache removal, which
6717 * in turn serializes uncharging.
6719 VM_BUG_ON_PAGE(!PageLocked(page), page);
6720 if (compound_head(page)->mem_cgroup)
6723 id = lookup_swap_cgroup_id(ent);
6725 memcg = mem_cgroup_from_id(id);
6726 if (memcg && !css_tryget_online(&memcg->css))
6732 memcg = get_mem_cgroup_from_mm(mm);
6734 ret = try_charge(memcg, gfp_mask, nr_pages);
6738 css_get(&memcg->css);
6739 commit_charge(page, memcg);
6741 local_irq_disable();
6742 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6743 memcg_check_events(memcg, page);
6746 if (PageSwapCache(page)) {
6747 swp_entry_t entry = { .val = page_private(page) };
6749 * The swap entry might not get freed for a long time,
6750 * let's not wait for it. The page already received a
6751 * memory+swap charge, drop the swap entry duplicate.
6753 mem_cgroup_uncharge_swap(entry, nr_pages);
6757 css_put(&memcg->css);
6762 struct uncharge_gather {
6763 struct mem_cgroup *memcg;
6764 unsigned long nr_pages;
6765 unsigned long pgpgout;
6766 unsigned long nr_kmem;
6767 struct page *dummy_page;
6770 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6772 memset(ug, 0, sizeof(*ug));
6775 static void uncharge_batch(const struct uncharge_gather *ug)
6777 unsigned long flags;
6779 if (!mem_cgroup_is_root(ug->memcg)) {
6780 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6781 if (do_memsw_account())
6782 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6783 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6784 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6785 memcg_oom_recover(ug->memcg);
6788 local_irq_save(flags);
6789 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6790 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6791 memcg_check_events(ug->memcg, ug->dummy_page);
6792 local_irq_restore(flags);
6794 /* drop reference from uncharge_page */
6795 css_put(&ug->memcg->css);
6798 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6800 unsigned long nr_pages;
6802 VM_BUG_ON_PAGE(PageLRU(page), page);
6804 if (!page->mem_cgroup)
6808 * Nobody should be changing or seriously looking at
6809 * page->mem_cgroup at this point, we have fully
6810 * exclusive access to the page.
6813 if (ug->memcg != page->mem_cgroup) {
6816 uncharge_gather_clear(ug);
6818 ug->memcg = page->mem_cgroup;
6820 /* pairs with css_put in uncharge_batch */
6821 css_get(&ug->memcg->css);
6824 nr_pages = compound_nr(page);
6825 ug->nr_pages += nr_pages;
6827 if (!PageKmemcg(page)) {
6830 ug->nr_kmem += nr_pages;
6831 __ClearPageKmemcg(page);
6834 ug->dummy_page = page;
6835 page->mem_cgroup = NULL;
6836 css_put(&ug->memcg->css);
6839 static void uncharge_list(struct list_head *page_list)
6841 struct uncharge_gather ug;
6842 struct list_head *next;
6844 uncharge_gather_clear(&ug);
6847 * Note that the list can be a single page->lru; hence the
6848 * do-while loop instead of a simple list_for_each_entry().
6850 next = page_list->next;
6854 page = list_entry(next, struct page, lru);
6855 next = page->lru.next;
6857 uncharge_page(page, &ug);
6858 } while (next != page_list);
6861 uncharge_batch(&ug);
6865 * mem_cgroup_uncharge - uncharge a page
6866 * @page: page to uncharge
6868 * Uncharge a page previously charged with mem_cgroup_charge().
6870 void mem_cgroup_uncharge(struct page *page)
6872 struct uncharge_gather ug;
6874 if (mem_cgroup_disabled())
6877 /* Don't touch page->lru of any random page, pre-check: */
6878 if (!page->mem_cgroup)
6881 uncharge_gather_clear(&ug);
6882 uncharge_page(page, &ug);
6883 uncharge_batch(&ug);
6887 * mem_cgroup_uncharge_list - uncharge a list of page
6888 * @page_list: list of pages to uncharge
6890 * Uncharge a list of pages previously charged with
6891 * mem_cgroup_charge().
6893 void mem_cgroup_uncharge_list(struct list_head *page_list)
6895 if (mem_cgroup_disabled())
6898 if (!list_empty(page_list))
6899 uncharge_list(page_list);
6903 * mem_cgroup_migrate - charge a page's replacement
6904 * @oldpage: currently circulating page
6905 * @newpage: replacement page
6907 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6908 * be uncharged upon free.
6910 * Both pages must be locked, @newpage->mapping must be set up.
6912 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6914 struct mem_cgroup *memcg;
6915 unsigned int nr_pages;
6916 unsigned long flags;
6918 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6919 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6920 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6921 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6924 if (mem_cgroup_disabled())
6927 /* Page cache replacement: new page already charged? */
6928 if (newpage->mem_cgroup)
6931 memcg = oldpage->mem_cgroup;
6935 /* Force-charge the new page. The old one will be freed soon */
6936 nr_pages = thp_nr_pages(newpage);
6938 page_counter_charge(&memcg->memory, nr_pages);
6939 if (do_memsw_account())
6940 page_counter_charge(&memcg->memsw, nr_pages);
6942 css_get(&memcg->css);
6943 commit_charge(newpage, memcg);
6945 local_irq_save(flags);
6946 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6947 memcg_check_events(memcg, newpage);
6948 local_irq_restore(flags);
6951 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6952 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6954 void mem_cgroup_sk_alloc(struct sock *sk)
6956 struct mem_cgroup *memcg;
6958 if (!mem_cgroup_sockets_enabled)
6961 /* Do not associate the sock with unrelated interrupted task's memcg. */
6966 memcg = mem_cgroup_from_task(current);
6967 if (memcg == root_mem_cgroup)
6969 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6971 if (css_tryget(&memcg->css))
6972 sk->sk_memcg = memcg;
6977 void mem_cgroup_sk_free(struct sock *sk)
6980 css_put(&sk->sk_memcg->css);
6984 * mem_cgroup_charge_skmem - charge socket memory
6985 * @memcg: memcg to charge
6986 * @nr_pages: number of pages to charge
6988 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6989 * @memcg's configured limit, %false if the charge had to be forced.
6991 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6993 gfp_t gfp_mask = GFP_KERNEL;
6995 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6996 struct page_counter *fail;
6998 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6999 memcg->tcpmem_pressure = 0;
7002 page_counter_charge(&memcg->tcpmem, nr_pages);
7003 memcg->tcpmem_pressure = 1;
7007 /* Don't block in the packet receive path */
7009 gfp_mask = GFP_NOWAIT;
7011 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
7013 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
7016 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
7021 * mem_cgroup_uncharge_skmem - uncharge socket memory
7022 * @memcg: memcg to uncharge
7023 * @nr_pages: number of pages to uncharge
7025 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7027 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7028 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7032 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7034 refill_stock(memcg, nr_pages);
7037 static int __init cgroup_memory(char *s)
7041 while ((token = strsep(&s, ",")) != NULL) {
7044 if (!strcmp(token, "nosocket"))
7045 cgroup_memory_nosocket = true;
7046 if (!strcmp(token, "nokmem"))
7047 cgroup_memory_nokmem = true;
7051 __setup("cgroup.memory=", cgroup_memory);
7054 * subsys_initcall() for memory controller.
7056 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7057 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7058 * basically everything that doesn't depend on a specific mem_cgroup structure
7059 * should be initialized from here.
7061 static int __init mem_cgroup_init(void)
7065 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7066 memcg_hotplug_cpu_dead);
7068 for_each_possible_cpu(cpu)
7069 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7072 for_each_node(node) {
7073 struct mem_cgroup_tree_per_node *rtpn;
7075 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7076 node_online(node) ? node : NUMA_NO_NODE);
7078 rtpn->rb_root = RB_ROOT;
7079 rtpn->rb_rightmost = NULL;
7080 spin_lock_init(&rtpn->lock);
7081 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7086 subsys_initcall(mem_cgroup_init);
7088 #ifdef CONFIG_MEMCG_SWAP
7089 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7091 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7093 * The root cgroup cannot be destroyed, so it's refcount must
7096 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7100 memcg = parent_mem_cgroup(memcg);
7102 memcg = root_mem_cgroup;
7108 * mem_cgroup_swapout - transfer a memsw charge to swap
7109 * @page: page whose memsw charge to transfer
7110 * @entry: swap entry to move the charge to
7112 * Transfer the memsw charge of @page to @entry.
7114 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7116 struct mem_cgroup *memcg, *swap_memcg;
7117 unsigned int nr_entries;
7118 unsigned short oldid;
7120 VM_BUG_ON_PAGE(PageLRU(page), page);
7121 VM_BUG_ON_PAGE(page_count(page), page);
7123 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7126 memcg = page->mem_cgroup;
7128 /* Readahead page, never charged */
7133 * In case the memcg owning these pages has been offlined and doesn't
7134 * have an ID allocated to it anymore, charge the closest online
7135 * ancestor for the swap instead and transfer the memory+swap charge.
7137 swap_memcg = mem_cgroup_id_get_online(memcg);
7138 nr_entries = thp_nr_pages(page);
7139 /* Get references for the tail pages, too */
7141 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7142 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7144 VM_BUG_ON_PAGE(oldid, page);
7145 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7147 page->mem_cgroup = NULL;
7149 if (!mem_cgroup_is_root(memcg))
7150 page_counter_uncharge(&memcg->memory, nr_entries);
7152 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7153 if (!mem_cgroup_is_root(swap_memcg))
7154 page_counter_charge(&swap_memcg->memsw, nr_entries);
7155 page_counter_uncharge(&memcg->memsw, nr_entries);
7159 * Interrupts should be disabled here because the caller holds the
7160 * i_pages lock which is taken with interrupts-off. It is
7161 * important here to have the interrupts disabled because it is the
7162 * only synchronisation we have for updating the per-CPU variables.
7164 VM_BUG_ON(!irqs_disabled());
7165 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7166 memcg_check_events(memcg, page);
7168 css_put(&memcg->css);
7172 * mem_cgroup_try_charge_swap - try charging swap space for a page
7173 * @page: page being added to swap
7174 * @entry: swap entry to charge
7176 * Try to charge @page's memcg for the swap space at @entry.
7178 * Returns 0 on success, -ENOMEM on failure.
7180 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7182 unsigned int nr_pages = thp_nr_pages(page);
7183 struct page_counter *counter;
7184 struct mem_cgroup *memcg;
7185 unsigned short oldid;
7187 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7190 memcg = page->mem_cgroup;
7192 /* Readahead page, never charged */
7197 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7201 memcg = mem_cgroup_id_get_online(memcg);
7203 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7204 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7205 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7206 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7207 mem_cgroup_id_put(memcg);
7211 /* Get references for the tail pages, too */
7213 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7214 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7215 VM_BUG_ON_PAGE(oldid, page);
7216 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7222 * mem_cgroup_uncharge_swap - uncharge swap space
7223 * @entry: swap entry to uncharge
7224 * @nr_pages: the amount of swap space to uncharge
7226 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7228 struct mem_cgroup *memcg;
7231 id = swap_cgroup_record(entry, 0, nr_pages);
7233 memcg = mem_cgroup_from_id(id);
7235 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7236 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7237 page_counter_uncharge(&memcg->swap, nr_pages);
7239 page_counter_uncharge(&memcg->memsw, nr_pages);
7241 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7242 mem_cgroup_id_put_many(memcg, nr_pages);
7247 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7249 long nr_swap_pages = get_nr_swap_pages();
7251 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7252 return nr_swap_pages;
7253 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7254 nr_swap_pages = min_t(long, nr_swap_pages,
7255 READ_ONCE(memcg->swap.max) -
7256 page_counter_read(&memcg->swap));
7257 return nr_swap_pages;
7260 bool mem_cgroup_swap_full(struct page *page)
7262 struct mem_cgroup *memcg;
7264 VM_BUG_ON_PAGE(!PageLocked(page), page);
7268 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7271 memcg = page->mem_cgroup;
7275 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7276 unsigned long usage = page_counter_read(&memcg->swap);
7278 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7279 usage * 2 >= READ_ONCE(memcg->swap.max))
7286 static int __init setup_swap_account(char *s)
7288 if (!strcmp(s, "1"))
7289 cgroup_memory_noswap = false;
7290 else if (!strcmp(s, "0"))
7291 cgroup_memory_noswap = true;
7294 __setup("swapaccount=", setup_swap_account);
7296 static u64 swap_current_read(struct cgroup_subsys_state *css,
7299 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7301 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7304 static int swap_high_show(struct seq_file *m, void *v)
7306 return seq_puts_memcg_tunable(m,
7307 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7310 static ssize_t swap_high_write(struct kernfs_open_file *of,
7311 char *buf, size_t nbytes, loff_t off)
7313 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7317 buf = strstrip(buf);
7318 err = page_counter_memparse(buf, "max", &high);
7322 page_counter_set_high(&memcg->swap, high);
7327 static int swap_max_show(struct seq_file *m, void *v)
7329 return seq_puts_memcg_tunable(m,
7330 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7333 static ssize_t swap_max_write(struct kernfs_open_file *of,
7334 char *buf, size_t nbytes, loff_t off)
7336 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7340 buf = strstrip(buf);
7341 err = page_counter_memparse(buf, "max", &max);
7345 xchg(&memcg->swap.max, max);
7350 static int swap_events_show(struct seq_file *m, void *v)
7352 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7354 seq_printf(m, "high %lu\n",
7355 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7356 seq_printf(m, "max %lu\n",
7357 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7358 seq_printf(m, "fail %lu\n",
7359 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7364 static struct cftype swap_files[] = {
7366 .name = "swap.current",
7367 .flags = CFTYPE_NOT_ON_ROOT,
7368 .read_u64 = swap_current_read,
7371 .name = "swap.high",
7372 .flags = CFTYPE_NOT_ON_ROOT,
7373 .seq_show = swap_high_show,
7374 .write = swap_high_write,
7378 .flags = CFTYPE_NOT_ON_ROOT,
7379 .seq_show = swap_max_show,
7380 .write = swap_max_write,
7383 .name = "swap.events",
7384 .flags = CFTYPE_NOT_ON_ROOT,
7385 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7386 .seq_show = swap_events_show,
7391 static struct cftype memsw_files[] = {
7393 .name = "memsw.usage_in_bytes",
7394 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7395 .read_u64 = mem_cgroup_read_u64,
7398 .name = "memsw.max_usage_in_bytes",
7399 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7400 .write = mem_cgroup_reset,
7401 .read_u64 = mem_cgroup_read_u64,
7404 .name = "memsw.limit_in_bytes",
7405 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7406 .write = mem_cgroup_write,
7407 .read_u64 = mem_cgroup_read_u64,
7410 .name = "memsw.failcnt",
7411 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7412 .write = mem_cgroup_reset,
7413 .read_u64 = mem_cgroup_read_u64,
7415 { }, /* terminate */
7419 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7420 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7421 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7422 * boot parameter. This may result in premature OOPS inside
7423 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7425 static int __init mem_cgroup_swap_init(void)
7427 /* No memory control -> no swap control */
7428 if (mem_cgroup_disabled())
7429 cgroup_memory_noswap = true;
7431 if (cgroup_memory_noswap)
7434 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7435 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7439 core_initcall(mem_cgroup_swap_init);
7441 #endif /* CONFIG_MEMCG_SWAP */