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_kmem_state(void *p, enum node_stat_item idx, int val)
858 pg_data_t *pgdat = page_pgdat(virt_to_page(p));
859 struct mem_cgroup *memcg;
860 struct lruvec *lruvec;
863 memcg = mem_cgroup_from_obj(p);
866 * Untracked pages have no memcg, no lruvec. Update only the
867 * node. If we reparent the slab objects to the root memcg,
868 * when we free the slab object, we need to update the per-memcg
869 * vmstats to keep it correct for the root memcg.
872 __mod_node_page_state(pgdat, idx, val);
874 lruvec = mem_cgroup_lruvec(memcg, pgdat);
875 __mod_lruvec_state(lruvec, idx, val);
881 * __count_memcg_events - account VM events in a cgroup
882 * @memcg: the memory cgroup
883 * @idx: the event item
884 * @count: the number of events that occured
886 void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
891 if (mem_cgroup_disabled())
894 x = count + __this_cpu_read(memcg->vmstats_percpu->events[idx]);
895 if (unlikely(x > MEMCG_CHARGE_BATCH)) {
896 struct mem_cgroup *mi;
899 * Batch local counters to keep them in sync with
900 * the hierarchical ones.
902 __this_cpu_add(memcg->vmstats_local->events[idx], x);
903 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
904 atomic_long_add(x, &mi->vmevents[idx]);
907 __this_cpu_write(memcg->vmstats_percpu->events[idx], x);
910 static unsigned long memcg_events(struct mem_cgroup *memcg, int event)
912 return atomic_long_read(&memcg->vmevents[event]);
915 static unsigned long memcg_events_local(struct mem_cgroup *memcg, int event)
920 for_each_possible_cpu(cpu)
921 x += per_cpu(memcg->vmstats_local->events[event], cpu);
925 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
929 /* pagein of a big page is an event. So, ignore page size */
931 __count_memcg_events(memcg, PGPGIN, 1);
933 __count_memcg_events(memcg, PGPGOUT, 1);
934 nr_pages = -nr_pages; /* for event */
937 __this_cpu_add(memcg->vmstats_percpu->nr_page_events, nr_pages);
940 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
941 enum mem_cgroup_events_target target)
943 unsigned long val, next;
945 val = __this_cpu_read(memcg->vmstats_percpu->nr_page_events);
946 next = __this_cpu_read(memcg->vmstats_percpu->targets[target]);
947 /* from time_after() in jiffies.h */
948 if ((long)(next - val) < 0) {
950 case MEM_CGROUP_TARGET_THRESH:
951 next = val + THRESHOLDS_EVENTS_TARGET;
953 case MEM_CGROUP_TARGET_SOFTLIMIT:
954 next = val + SOFTLIMIT_EVENTS_TARGET;
959 __this_cpu_write(memcg->vmstats_percpu->targets[target], next);
966 * Check events in order.
969 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
971 /* threshold event is triggered in finer grain than soft limit */
972 if (unlikely(mem_cgroup_event_ratelimit(memcg,
973 MEM_CGROUP_TARGET_THRESH))) {
976 do_softlimit = mem_cgroup_event_ratelimit(memcg,
977 MEM_CGROUP_TARGET_SOFTLIMIT);
978 mem_cgroup_threshold(memcg);
979 if (unlikely(do_softlimit))
980 mem_cgroup_update_tree(memcg, page);
984 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
987 * mm_update_next_owner() may clear mm->owner to NULL
988 * if it races with swapoff, page migration, etc.
989 * So this can be called with p == NULL.
994 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
996 EXPORT_SYMBOL(mem_cgroup_from_task);
999 * get_mem_cgroup_from_mm: Obtain a reference on given mm_struct's memcg.
1000 * @mm: mm from which memcg should be extracted. It can be NULL.
1002 * Obtain a reference on mm->memcg and returns it if successful. Otherwise
1003 * root_mem_cgroup is returned. However if mem_cgroup is disabled, NULL is
1006 struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1008 struct mem_cgroup *memcg;
1010 if (mem_cgroup_disabled())
1016 * Page cache insertions can happen withou an
1017 * actual mm context, e.g. during disk probing
1018 * on boot, loopback IO, acct() writes etc.
1021 memcg = root_mem_cgroup;
1023 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1024 if (unlikely(!memcg))
1025 memcg = root_mem_cgroup;
1027 } while (!css_tryget(&memcg->css));
1031 EXPORT_SYMBOL(get_mem_cgroup_from_mm);
1034 * get_mem_cgroup_from_page: Obtain a reference on given page's memcg.
1035 * @page: page from which memcg should be extracted.
1037 * Obtain a reference on page->memcg and returns it if successful. Otherwise
1038 * root_mem_cgroup is returned.
1040 struct mem_cgroup *get_mem_cgroup_from_page(struct page *page)
1042 struct mem_cgroup *memcg = page->mem_cgroup;
1044 if (mem_cgroup_disabled())
1048 /* Page should not get uncharged and freed memcg under us. */
1049 if (!memcg || WARN_ON_ONCE(!css_tryget(&memcg->css)))
1050 memcg = root_mem_cgroup;
1054 EXPORT_SYMBOL(get_mem_cgroup_from_page);
1056 static __always_inline struct mem_cgroup *active_memcg(void)
1059 return this_cpu_read(int_active_memcg);
1061 return current->active_memcg;
1064 static __always_inline struct mem_cgroup *get_active_memcg(void)
1066 struct mem_cgroup *memcg;
1069 memcg = active_memcg();
1071 /* current->active_memcg must hold a ref. */
1072 if (WARN_ON_ONCE(!css_tryget(&memcg->css)))
1073 memcg = root_mem_cgroup;
1075 memcg = current->active_memcg;
1082 static __always_inline bool memcg_kmem_bypass(void)
1084 /* Allow remote memcg charging from any context. */
1085 if (unlikely(active_memcg()))
1088 /* Memcg to charge can't be determined. */
1089 if (in_interrupt() || !current->mm || (current->flags & PF_KTHREAD))
1096 * If active memcg is set, do not fallback to current->mm->memcg.
1098 static __always_inline struct mem_cgroup *get_mem_cgroup_from_current(void)
1100 if (memcg_kmem_bypass())
1103 if (unlikely(active_memcg()))
1104 return get_active_memcg();
1106 return get_mem_cgroup_from_mm(current->mm);
1110 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1111 * @root: hierarchy root
1112 * @prev: previously returned memcg, NULL on first invocation
1113 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1115 * Returns references to children of the hierarchy below @root, or
1116 * @root itself, or %NULL after a full round-trip.
1118 * Caller must pass the return value in @prev on subsequent
1119 * invocations for reference counting, or use mem_cgroup_iter_break()
1120 * to cancel a hierarchy walk before the round-trip is complete.
1122 * Reclaimers can specify a node in @reclaim to divide up the memcgs
1123 * in the hierarchy among all concurrent reclaimers operating on the
1126 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1127 struct mem_cgroup *prev,
1128 struct mem_cgroup_reclaim_cookie *reclaim)
1130 struct mem_cgroup_reclaim_iter *iter;
1131 struct cgroup_subsys_state *css = NULL;
1132 struct mem_cgroup *memcg = NULL;
1133 struct mem_cgroup *pos = NULL;
1135 if (mem_cgroup_disabled())
1139 root = root_mem_cgroup;
1141 if (prev && !reclaim)
1147 struct mem_cgroup_per_node *mz;
1149 mz = mem_cgroup_nodeinfo(root, reclaim->pgdat->node_id);
1152 if (prev && reclaim->generation != iter->generation)
1156 pos = READ_ONCE(iter->position);
1157 if (!pos || css_tryget(&pos->css))
1160 * css reference reached zero, so iter->position will
1161 * be cleared by ->css_released. However, we should not
1162 * rely on this happening soon, because ->css_released
1163 * is called from a work queue, and by busy-waiting we
1164 * might block it. So we clear iter->position right
1167 (void)cmpxchg(&iter->position, pos, NULL);
1175 css = css_next_descendant_pre(css, &root->css);
1178 * Reclaimers share the hierarchy walk, and a
1179 * new one might jump in right at the end of
1180 * the hierarchy - make sure they see at least
1181 * one group and restart from the beginning.
1189 * Verify the css and acquire a reference. The root
1190 * is provided by the caller, so we know it's alive
1191 * and kicking, and don't take an extra reference.
1193 memcg = mem_cgroup_from_css(css);
1195 if (css == &root->css)
1198 if (css_tryget(css))
1206 * The position could have already been updated by a competing
1207 * thread, so check that the value hasn't changed since we read
1208 * it to avoid reclaiming from the same cgroup twice.
1210 (void)cmpxchg(&iter->position, pos, memcg);
1218 reclaim->generation = iter->generation;
1223 if (prev && prev != root)
1224 css_put(&prev->css);
1230 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1231 * @root: hierarchy root
1232 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1234 void mem_cgroup_iter_break(struct mem_cgroup *root,
1235 struct mem_cgroup *prev)
1238 root = root_mem_cgroup;
1239 if (prev && prev != root)
1240 css_put(&prev->css);
1243 static void __invalidate_reclaim_iterators(struct mem_cgroup *from,
1244 struct mem_cgroup *dead_memcg)
1246 struct mem_cgroup_reclaim_iter *iter;
1247 struct mem_cgroup_per_node *mz;
1250 for_each_node(nid) {
1251 mz = mem_cgroup_nodeinfo(from, nid);
1253 cmpxchg(&iter->position, dead_memcg, NULL);
1257 static void invalidate_reclaim_iterators(struct mem_cgroup *dead_memcg)
1259 struct mem_cgroup *memcg = dead_memcg;
1260 struct mem_cgroup *last;
1263 __invalidate_reclaim_iterators(memcg, dead_memcg);
1265 } while ((memcg = parent_mem_cgroup(memcg)));
1268 * When cgruop1 non-hierarchy mode is used,
1269 * parent_mem_cgroup() does not walk all the way up to the
1270 * cgroup root (root_mem_cgroup). So we have to handle
1271 * dead_memcg from cgroup root separately.
1273 if (last != root_mem_cgroup)
1274 __invalidate_reclaim_iterators(root_mem_cgroup,
1279 * mem_cgroup_scan_tasks - iterate over tasks of a memory cgroup hierarchy
1280 * @memcg: hierarchy root
1281 * @fn: function to call for each task
1282 * @arg: argument passed to @fn
1284 * This function iterates over tasks attached to @memcg or to any of its
1285 * descendants and calls @fn for each task. If @fn returns a non-zero
1286 * value, the function breaks the iteration loop and returns the value.
1287 * Otherwise, it will iterate over all tasks and return 0.
1289 * This function must not be called for the root memory cgroup.
1291 int mem_cgroup_scan_tasks(struct mem_cgroup *memcg,
1292 int (*fn)(struct task_struct *, void *), void *arg)
1294 struct mem_cgroup *iter;
1297 BUG_ON(memcg == root_mem_cgroup);
1299 for_each_mem_cgroup_tree(iter, memcg) {
1300 struct css_task_iter it;
1301 struct task_struct *task;
1303 css_task_iter_start(&iter->css, CSS_TASK_ITER_PROCS, &it);
1304 while (!ret && (task = css_task_iter_next(&it)))
1305 ret = fn(task, arg);
1306 css_task_iter_end(&it);
1308 mem_cgroup_iter_break(memcg, iter);
1316 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1318 * @pgdat: pgdat of the page
1320 * This function relies on page's memcg being stable - see the
1321 * access rules in commit_charge().
1323 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgdat)
1325 struct mem_cgroup_per_node *mz;
1326 struct mem_cgroup *memcg;
1327 struct lruvec *lruvec;
1329 if (mem_cgroup_disabled()) {
1330 lruvec = &pgdat->__lruvec;
1334 memcg = page->mem_cgroup;
1336 * Swapcache readahead pages are added to the LRU - and
1337 * possibly migrated - before they are charged.
1340 memcg = root_mem_cgroup;
1342 mz = mem_cgroup_page_nodeinfo(memcg, page);
1343 lruvec = &mz->lruvec;
1346 * Since a node can be onlined after the mem_cgroup was created,
1347 * we have to be prepared to initialize lruvec->zone here;
1348 * and if offlined then reonlined, we need to reinitialize it.
1350 if (unlikely(lruvec->pgdat != pgdat))
1351 lruvec->pgdat = pgdat;
1356 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1357 * @lruvec: mem_cgroup per zone lru vector
1358 * @lru: index of lru list the page is sitting on
1359 * @zid: zone id of the accounted pages
1360 * @nr_pages: positive when adding or negative when removing
1362 * This function must be called under lru_lock, just before a page is added
1363 * to or just after a page is removed from an lru list (that ordering being
1364 * so as to allow it to check that lru_size 0 is consistent with list_empty).
1366 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1367 int zid, int nr_pages)
1369 struct mem_cgroup_per_node *mz;
1370 unsigned long *lru_size;
1373 if (mem_cgroup_disabled())
1376 mz = container_of(lruvec, struct mem_cgroup_per_node, lruvec);
1377 lru_size = &mz->lru_zone_size[zid][lru];
1380 *lru_size += nr_pages;
1383 if (WARN_ONCE(size < 0,
1384 "%s(%p, %d, %d): lru_size %ld\n",
1385 __func__, lruvec, lru, nr_pages, size)) {
1391 *lru_size += nr_pages;
1395 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1396 * @memcg: the memory cgroup
1398 * Returns the maximum amount of memory @mem can be charged with, in
1401 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1403 unsigned long margin = 0;
1404 unsigned long count;
1405 unsigned long limit;
1407 count = page_counter_read(&memcg->memory);
1408 limit = READ_ONCE(memcg->memory.max);
1410 margin = limit - count;
1412 if (do_memsw_account()) {
1413 count = page_counter_read(&memcg->memsw);
1414 limit = READ_ONCE(memcg->memsw.max);
1416 margin = min(margin, limit - count);
1425 * A routine for checking "mem" is under move_account() or not.
1427 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1428 * moving cgroups. This is for waiting at high-memory pressure
1431 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1433 struct mem_cgroup *from;
1434 struct mem_cgroup *to;
1437 * Unlike task_move routines, we access mc.to, mc.from not under
1438 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1440 spin_lock(&mc.lock);
1446 ret = mem_cgroup_is_descendant(from, memcg) ||
1447 mem_cgroup_is_descendant(to, memcg);
1449 spin_unlock(&mc.lock);
1453 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1455 if (mc.moving_task && current != mc.moving_task) {
1456 if (mem_cgroup_under_move(memcg)) {
1458 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1459 /* moving charge context might have finished. */
1462 finish_wait(&mc.waitq, &wait);
1469 struct memory_stat {
1475 static struct memory_stat memory_stats[] = {
1476 { "anon", PAGE_SIZE, NR_ANON_MAPPED },
1477 { "file", PAGE_SIZE, NR_FILE_PAGES },
1478 { "kernel_stack", 1024, NR_KERNEL_STACK_KB },
1479 { "percpu", 1, MEMCG_PERCPU_B },
1480 { "sock", PAGE_SIZE, MEMCG_SOCK },
1481 { "shmem", PAGE_SIZE, NR_SHMEM },
1482 { "file_mapped", PAGE_SIZE, NR_FILE_MAPPED },
1483 { "file_dirty", PAGE_SIZE, NR_FILE_DIRTY },
1484 { "file_writeback", PAGE_SIZE, NR_WRITEBACK },
1485 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1487 * The ratio will be initialized in memory_stats_init(). Because
1488 * on some architectures, the macro of HPAGE_PMD_SIZE is not
1489 * constant(e.g. powerpc).
1491 { "anon_thp", 0, NR_ANON_THPS },
1492 { "file_thp", 0, NR_FILE_THPS },
1493 { "shmem_thp", 0, NR_SHMEM_THPS },
1495 { "inactive_anon", PAGE_SIZE, NR_INACTIVE_ANON },
1496 { "active_anon", PAGE_SIZE, NR_ACTIVE_ANON },
1497 { "inactive_file", PAGE_SIZE, NR_INACTIVE_FILE },
1498 { "active_file", PAGE_SIZE, NR_ACTIVE_FILE },
1499 { "unevictable", PAGE_SIZE, NR_UNEVICTABLE },
1502 * Note: The slab_reclaimable and slab_unreclaimable must be
1503 * together and slab_reclaimable must be in front.
1505 { "slab_reclaimable", 1, NR_SLAB_RECLAIMABLE_B },
1506 { "slab_unreclaimable", 1, NR_SLAB_UNRECLAIMABLE_B },
1508 /* The memory events */
1509 { "workingset_refault_anon", 1, WORKINGSET_REFAULT_ANON },
1510 { "workingset_refault_file", 1, WORKINGSET_REFAULT_FILE },
1511 { "workingset_activate_anon", 1, WORKINGSET_ACTIVATE_ANON },
1512 { "workingset_activate_file", 1, WORKINGSET_ACTIVATE_FILE },
1513 { "workingset_restore_anon", 1, WORKINGSET_RESTORE_ANON },
1514 { "workingset_restore_file", 1, WORKINGSET_RESTORE_FILE },
1515 { "workingset_nodereclaim", 1, WORKINGSET_NODERECLAIM },
1518 static int __init memory_stats_init(void)
1522 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1523 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1524 if (memory_stats[i].idx == NR_ANON_THPS ||
1525 memory_stats[i].idx == NR_FILE_THPS ||
1526 memory_stats[i].idx == NR_SHMEM_THPS)
1527 memory_stats[i].ratio = HPAGE_PMD_SIZE;
1529 VM_BUG_ON(!memory_stats[i].ratio);
1530 VM_BUG_ON(memory_stats[i].idx >= MEMCG_NR_STAT);
1535 pure_initcall(memory_stats_init);
1537 static char *memory_stat_format(struct mem_cgroup *memcg)
1542 seq_buf_init(&s, kmalloc(PAGE_SIZE, GFP_KERNEL), PAGE_SIZE);
1547 * Provide statistics on the state of the memory subsystem as
1548 * well as cumulative event counters that show past behavior.
1550 * This list is ordered following a combination of these gradients:
1551 * 1) generic big picture -> specifics and details
1552 * 2) reflecting userspace activity -> reflecting kernel heuristics
1554 * Current memory state:
1557 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
1560 size = memcg_page_state(memcg, memory_stats[i].idx);
1561 size *= memory_stats[i].ratio;
1562 seq_buf_printf(&s, "%s %llu\n", memory_stats[i].name, size);
1564 if (unlikely(memory_stats[i].idx == NR_SLAB_UNRECLAIMABLE_B)) {
1565 size = memcg_page_state(memcg, NR_SLAB_RECLAIMABLE_B) +
1566 memcg_page_state(memcg, NR_SLAB_UNRECLAIMABLE_B);
1567 seq_buf_printf(&s, "slab %llu\n", size);
1571 /* Accumulated memory events */
1573 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGFAULT),
1574 memcg_events(memcg, PGFAULT));
1575 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGMAJFAULT),
1576 memcg_events(memcg, PGMAJFAULT));
1577 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGREFILL),
1578 memcg_events(memcg, PGREFILL));
1579 seq_buf_printf(&s, "pgscan %lu\n",
1580 memcg_events(memcg, PGSCAN_KSWAPD) +
1581 memcg_events(memcg, PGSCAN_DIRECT));
1582 seq_buf_printf(&s, "pgsteal %lu\n",
1583 memcg_events(memcg, PGSTEAL_KSWAPD) +
1584 memcg_events(memcg, PGSTEAL_DIRECT));
1585 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGACTIVATE),
1586 memcg_events(memcg, PGACTIVATE));
1587 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGDEACTIVATE),
1588 memcg_events(memcg, PGDEACTIVATE));
1589 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREE),
1590 memcg_events(memcg, PGLAZYFREE));
1591 seq_buf_printf(&s, "%s %lu\n", vm_event_name(PGLAZYFREED),
1592 memcg_events(memcg, PGLAZYFREED));
1594 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1595 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_FAULT_ALLOC),
1596 memcg_events(memcg, THP_FAULT_ALLOC));
1597 seq_buf_printf(&s, "%s %lu\n", vm_event_name(THP_COLLAPSE_ALLOC),
1598 memcg_events(memcg, THP_COLLAPSE_ALLOC));
1599 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1601 /* The above should easily fit into one page */
1602 WARN_ON_ONCE(seq_buf_has_overflowed(&s));
1607 #define K(x) ((x) << (PAGE_SHIFT-10))
1609 * mem_cgroup_print_oom_context: Print OOM information relevant to
1610 * memory controller.
1611 * @memcg: The memory cgroup that went over limit
1612 * @p: Task that is going to be killed
1614 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1617 void mem_cgroup_print_oom_context(struct mem_cgroup *memcg, struct task_struct *p)
1622 pr_cont(",oom_memcg=");
1623 pr_cont_cgroup_path(memcg->css.cgroup);
1625 pr_cont(",global_oom");
1627 pr_cont(",task_memcg=");
1628 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1634 * mem_cgroup_print_oom_meminfo: Print OOM memory information relevant to
1635 * memory controller.
1636 * @memcg: The memory cgroup that went over limit
1638 void mem_cgroup_print_oom_meminfo(struct mem_cgroup *memcg)
1642 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1643 K((u64)page_counter_read(&memcg->memory)),
1644 K((u64)READ_ONCE(memcg->memory.max)), memcg->memory.failcnt);
1645 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
1646 pr_info("swap: usage %llukB, limit %llukB, failcnt %lu\n",
1647 K((u64)page_counter_read(&memcg->swap)),
1648 K((u64)READ_ONCE(memcg->swap.max)), memcg->swap.failcnt);
1650 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1651 K((u64)page_counter_read(&memcg->memsw)),
1652 K((u64)memcg->memsw.max), memcg->memsw.failcnt);
1653 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1654 K((u64)page_counter_read(&memcg->kmem)),
1655 K((u64)memcg->kmem.max), memcg->kmem.failcnt);
1658 pr_info("Memory cgroup stats for ");
1659 pr_cont_cgroup_path(memcg->css.cgroup);
1661 buf = memory_stat_format(memcg);
1669 * Return the memory (and swap, if configured) limit for a memcg.
1671 unsigned long mem_cgroup_get_max(struct mem_cgroup *memcg)
1673 unsigned long max = READ_ONCE(memcg->memory.max);
1675 if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
1676 if (mem_cgroup_swappiness(memcg))
1677 max += min(READ_ONCE(memcg->swap.max),
1678 (unsigned long)total_swap_pages);
1680 if (mem_cgroup_swappiness(memcg)) {
1681 /* Calculate swap excess capacity from memsw limit */
1682 unsigned long swap = READ_ONCE(memcg->memsw.max) - max;
1684 max += min(swap, (unsigned long)total_swap_pages);
1690 unsigned long mem_cgroup_size(struct mem_cgroup *memcg)
1692 return page_counter_read(&memcg->memory);
1695 static bool mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1698 struct oom_control oc = {
1702 .gfp_mask = gfp_mask,
1707 if (mutex_lock_killable(&oom_lock))
1710 if (mem_cgroup_margin(memcg) >= (1 << order))
1714 * A few threads which were not waiting at mutex_lock_killable() can
1715 * fail to bail out. Therefore, check again after holding oom_lock.
1717 ret = should_force_charge() || out_of_memory(&oc);
1720 mutex_unlock(&oom_lock);
1724 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1727 unsigned long *total_scanned)
1729 struct mem_cgroup *victim = NULL;
1732 unsigned long excess;
1733 unsigned long nr_scanned;
1734 struct mem_cgroup_reclaim_cookie reclaim = {
1738 excess = soft_limit_excess(root_memcg);
1741 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1746 * If we have not been able to reclaim
1747 * anything, it might because there are
1748 * no reclaimable pages under this hierarchy
1753 * We want to do more targeted reclaim.
1754 * excess >> 2 is not to excessive so as to
1755 * reclaim too much, nor too less that we keep
1756 * coming back to reclaim from this cgroup
1758 if (total >= (excess >> 2) ||
1759 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1764 total += mem_cgroup_shrink_node(victim, gfp_mask, false,
1765 pgdat, &nr_scanned);
1766 *total_scanned += nr_scanned;
1767 if (!soft_limit_excess(root_memcg))
1770 mem_cgroup_iter_break(root_memcg, victim);
1774 #ifdef CONFIG_LOCKDEP
1775 static struct lockdep_map memcg_oom_lock_dep_map = {
1776 .name = "memcg_oom_lock",
1780 static DEFINE_SPINLOCK(memcg_oom_lock);
1783 * Check OOM-Killer is already running under our hierarchy.
1784 * If someone is running, return false.
1786 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1788 struct mem_cgroup *iter, *failed = NULL;
1790 spin_lock(&memcg_oom_lock);
1792 for_each_mem_cgroup_tree(iter, memcg) {
1793 if (iter->oom_lock) {
1795 * this subtree of our hierarchy is already locked
1796 * so we cannot give a lock.
1799 mem_cgroup_iter_break(memcg, iter);
1802 iter->oom_lock = true;
1807 * OK, we failed to lock the whole subtree so we have
1808 * to clean up what we set up to the failing subtree
1810 for_each_mem_cgroup_tree(iter, memcg) {
1811 if (iter == failed) {
1812 mem_cgroup_iter_break(memcg, iter);
1815 iter->oom_lock = false;
1818 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1820 spin_unlock(&memcg_oom_lock);
1825 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1827 struct mem_cgroup *iter;
1829 spin_lock(&memcg_oom_lock);
1830 mutex_release(&memcg_oom_lock_dep_map, _RET_IP_);
1831 for_each_mem_cgroup_tree(iter, memcg)
1832 iter->oom_lock = false;
1833 spin_unlock(&memcg_oom_lock);
1836 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1840 spin_lock(&memcg_oom_lock);
1841 for_each_mem_cgroup_tree(iter, memcg)
1843 spin_unlock(&memcg_oom_lock);
1846 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1848 struct mem_cgroup *iter;
1851 * Be careful about under_oom underflows becase a child memcg
1852 * could have been added after mem_cgroup_mark_under_oom.
1854 spin_lock(&memcg_oom_lock);
1855 for_each_mem_cgroup_tree(iter, memcg)
1856 if (iter->under_oom > 0)
1858 spin_unlock(&memcg_oom_lock);
1861 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1863 struct oom_wait_info {
1864 struct mem_cgroup *memcg;
1865 wait_queue_entry_t wait;
1868 static int memcg_oom_wake_function(wait_queue_entry_t *wait,
1869 unsigned mode, int sync, void *arg)
1871 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1872 struct mem_cgroup *oom_wait_memcg;
1873 struct oom_wait_info *oom_wait_info;
1875 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1876 oom_wait_memcg = oom_wait_info->memcg;
1878 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1879 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1881 return autoremove_wake_function(wait, mode, sync, arg);
1884 static void memcg_oom_recover(struct mem_cgroup *memcg)
1887 * For the following lockless ->under_oom test, the only required
1888 * guarantee is that it must see the state asserted by an OOM when
1889 * this function is called as a result of userland actions
1890 * triggered by the notification of the OOM. This is trivially
1891 * achieved by invoking mem_cgroup_mark_under_oom() before
1892 * triggering notification.
1894 if (memcg && memcg->under_oom)
1895 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1905 static enum oom_status mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1907 enum oom_status ret;
1910 if (order > PAGE_ALLOC_COSTLY_ORDER)
1913 memcg_memory_event(memcg, MEMCG_OOM);
1916 * We are in the middle of the charge context here, so we
1917 * don't want to block when potentially sitting on a callstack
1918 * that holds all kinds of filesystem and mm locks.
1920 * cgroup1 allows disabling the OOM killer and waiting for outside
1921 * handling until the charge can succeed; remember the context and put
1922 * the task to sleep at the end of the page fault when all locks are
1925 * On the other hand, in-kernel OOM killer allows for an async victim
1926 * memory reclaim (oom_reaper) and that means that we are not solely
1927 * relying on the oom victim to make a forward progress and we can
1928 * invoke the oom killer here.
1930 * Please note that mem_cgroup_out_of_memory might fail to find a
1931 * victim and then we have to bail out from the charge path.
1933 if (memcg->oom_kill_disable) {
1934 if (!current->in_user_fault)
1936 css_get(&memcg->css);
1937 current->memcg_in_oom = memcg;
1938 current->memcg_oom_gfp_mask = mask;
1939 current->memcg_oom_order = order;
1944 mem_cgroup_mark_under_oom(memcg);
1946 locked = mem_cgroup_oom_trylock(memcg);
1949 mem_cgroup_oom_notify(memcg);
1951 mem_cgroup_unmark_under_oom(memcg);
1952 if (mem_cgroup_out_of_memory(memcg, mask, order))
1958 mem_cgroup_oom_unlock(memcg);
1964 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1965 * @handle: actually kill/wait or just clean up the OOM state
1967 * This has to be called at the end of a page fault if the memcg OOM
1968 * handler was enabled.
1970 * Memcg supports userspace OOM handling where failed allocations must
1971 * sleep on a waitqueue until the userspace task resolves the
1972 * situation. Sleeping directly in the charge context with all kinds
1973 * of locks held is not a good idea, instead we remember an OOM state
1974 * in the task and mem_cgroup_oom_synchronize() has to be called at
1975 * the end of the page fault to complete the OOM handling.
1977 * Returns %true if an ongoing memcg OOM situation was detected and
1978 * completed, %false otherwise.
1980 bool mem_cgroup_oom_synchronize(bool handle)
1982 struct mem_cgroup *memcg = current->memcg_in_oom;
1983 struct oom_wait_info owait;
1986 /* OOM is global, do not handle */
1993 owait.memcg = memcg;
1994 owait.wait.flags = 0;
1995 owait.wait.func = memcg_oom_wake_function;
1996 owait.wait.private = current;
1997 INIT_LIST_HEAD(&owait.wait.entry);
1999 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2000 mem_cgroup_mark_under_oom(memcg);
2002 locked = mem_cgroup_oom_trylock(memcg);
2005 mem_cgroup_oom_notify(memcg);
2007 if (locked && !memcg->oom_kill_disable) {
2008 mem_cgroup_unmark_under_oom(memcg);
2009 finish_wait(&memcg_oom_waitq, &owait.wait);
2010 mem_cgroup_out_of_memory(memcg, current->memcg_oom_gfp_mask,
2011 current->memcg_oom_order);
2014 mem_cgroup_unmark_under_oom(memcg);
2015 finish_wait(&memcg_oom_waitq, &owait.wait);
2019 mem_cgroup_oom_unlock(memcg);
2021 * There is no guarantee that an OOM-lock contender
2022 * sees the wakeups triggered by the OOM kill
2023 * uncharges. Wake any sleepers explicitely.
2025 memcg_oom_recover(memcg);
2028 current->memcg_in_oom = NULL;
2029 css_put(&memcg->css);
2034 * mem_cgroup_get_oom_group - get a memory cgroup to clean up after OOM
2035 * @victim: task to be killed by the OOM killer
2036 * @oom_domain: memcg in case of memcg OOM, NULL in case of system-wide OOM
2038 * Returns a pointer to a memory cgroup, which has to be cleaned up
2039 * by killing all belonging OOM-killable tasks.
2041 * Caller has to call mem_cgroup_put() on the returned non-NULL memcg.
2043 struct mem_cgroup *mem_cgroup_get_oom_group(struct task_struct *victim,
2044 struct mem_cgroup *oom_domain)
2046 struct mem_cgroup *oom_group = NULL;
2047 struct mem_cgroup *memcg;
2049 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
2053 oom_domain = root_mem_cgroup;
2057 memcg = mem_cgroup_from_task(victim);
2058 if (memcg == root_mem_cgroup)
2062 * If the victim task has been asynchronously moved to a different
2063 * memory cgroup, we might end up killing tasks outside oom_domain.
2064 * In this case it's better to ignore memory.group.oom.
2066 if (unlikely(!mem_cgroup_is_descendant(memcg, oom_domain)))
2070 * Traverse the memory cgroup hierarchy from the victim task's
2071 * cgroup up to the OOMing cgroup (or root) to find the
2072 * highest-level memory cgroup with oom.group set.
2074 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
2075 if (memcg->oom_group)
2078 if (memcg == oom_domain)
2083 css_get(&oom_group->css);
2090 void mem_cgroup_print_oom_group(struct mem_cgroup *memcg)
2092 pr_info("Tasks in ");
2093 pr_cont_cgroup_path(memcg->css.cgroup);
2094 pr_cont(" are going to be killed due to memory.oom.group set\n");
2098 * lock_page_memcg - lock a page->mem_cgroup binding
2101 * This function protects unlocked LRU pages from being moved to
2104 * It ensures lifetime of the returned memcg. Caller is responsible
2105 * for the lifetime of the page; __unlock_page_memcg() is available
2106 * when @page might get freed inside the locked section.
2108 struct mem_cgroup *lock_page_memcg(struct page *page)
2110 struct page *head = compound_head(page); /* rmap on tail pages */
2111 struct mem_cgroup *memcg;
2112 unsigned long flags;
2115 * The RCU lock is held throughout the transaction. The fast
2116 * path can get away without acquiring the memcg->move_lock
2117 * because page moving starts with an RCU grace period.
2119 * The RCU lock also protects the memcg from being freed when
2120 * the page state that is going to change is the only thing
2121 * preventing the page itself from being freed. E.g. writeback
2122 * doesn't hold a page reference and relies on PG_writeback to
2123 * keep off truncation, migration and so forth.
2127 if (mem_cgroup_disabled())
2130 memcg = head->mem_cgroup;
2131 if (unlikely(!memcg))
2134 if (atomic_read(&memcg->moving_account) <= 0)
2137 spin_lock_irqsave(&memcg->move_lock, flags);
2138 if (memcg != head->mem_cgroup) {
2139 spin_unlock_irqrestore(&memcg->move_lock, flags);
2144 * When charge migration first begins, we can have locked and
2145 * unlocked page stat updates happening concurrently. Track
2146 * the task who has the lock for unlock_page_memcg().
2148 memcg->move_lock_task = current;
2149 memcg->move_lock_flags = flags;
2153 EXPORT_SYMBOL(lock_page_memcg);
2156 * __unlock_page_memcg - unlock and unpin a memcg
2159 * Unlock and unpin a memcg returned by lock_page_memcg().
2161 void __unlock_page_memcg(struct mem_cgroup *memcg)
2163 if (memcg && memcg->move_lock_task == current) {
2164 unsigned long flags = memcg->move_lock_flags;
2166 memcg->move_lock_task = NULL;
2167 memcg->move_lock_flags = 0;
2169 spin_unlock_irqrestore(&memcg->move_lock, flags);
2176 * unlock_page_memcg - unlock a page->mem_cgroup binding
2179 void unlock_page_memcg(struct page *page)
2181 struct page *head = compound_head(page);
2183 __unlock_page_memcg(head->mem_cgroup);
2185 EXPORT_SYMBOL(unlock_page_memcg);
2187 struct memcg_stock_pcp {
2188 struct mem_cgroup *cached; /* this never be root cgroup */
2189 unsigned int nr_pages;
2191 #ifdef CONFIG_MEMCG_KMEM
2192 struct obj_cgroup *cached_objcg;
2193 unsigned int nr_bytes;
2196 struct work_struct work;
2197 unsigned long flags;
2198 #define FLUSHING_CACHED_CHARGE 0
2200 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2201 static DEFINE_MUTEX(percpu_charge_mutex);
2203 #ifdef CONFIG_MEMCG_KMEM
2204 static void drain_obj_stock(struct memcg_stock_pcp *stock);
2205 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2206 struct mem_cgroup *root_memcg);
2209 static inline void drain_obj_stock(struct memcg_stock_pcp *stock)
2212 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
2213 struct mem_cgroup *root_memcg)
2220 * consume_stock: Try to consume stocked charge on this cpu.
2221 * @memcg: memcg to consume from.
2222 * @nr_pages: how many pages to charge.
2224 * The charges will only happen if @memcg matches the current cpu's memcg
2225 * stock, and at least @nr_pages are available in that stock. Failure to
2226 * service an allocation will refill the stock.
2228 * returns true if successful, false otherwise.
2230 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2232 struct memcg_stock_pcp *stock;
2233 unsigned long flags;
2236 if (nr_pages > MEMCG_CHARGE_BATCH)
2239 local_irq_save(flags);
2241 stock = this_cpu_ptr(&memcg_stock);
2242 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2243 stock->nr_pages -= nr_pages;
2247 local_irq_restore(flags);
2253 * Returns stocks cached in percpu and reset cached information.
2255 static void drain_stock(struct memcg_stock_pcp *stock)
2257 struct mem_cgroup *old = stock->cached;
2262 if (stock->nr_pages) {
2263 page_counter_uncharge(&old->memory, stock->nr_pages);
2264 if (do_memsw_account())
2265 page_counter_uncharge(&old->memsw, stock->nr_pages);
2266 stock->nr_pages = 0;
2270 stock->cached = NULL;
2273 static void drain_local_stock(struct work_struct *dummy)
2275 struct memcg_stock_pcp *stock;
2276 unsigned long flags;
2279 * The only protection from memory hotplug vs. drain_stock races is
2280 * that we always operate on local CPU stock here with IRQ disabled
2282 local_irq_save(flags);
2284 stock = this_cpu_ptr(&memcg_stock);
2285 drain_obj_stock(stock);
2287 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2289 local_irq_restore(flags);
2293 * Cache charges(val) to local per_cpu area.
2294 * This will be consumed by consume_stock() function, later.
2296 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2298 struct memcg_stock_pcp *stock;
2299 unsigned long flags;
2301 local_irq_save(flags);
2303 stock = this_cpu_ptr(&memcg_stock);
2304 if (stock->cached != memcg) { /* reset if necessary */
2306 css_get(&memcg->css);
2307 stock->cached = memcg;
2309 stock->nr_pages += nr_pages;
2311 if (stock->nr_pages > MEMCG_CHARGE_BATCH)
2314 local_irq_restore(flags);
2318 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2319 * of the hierarchy under it.
2321 static void drain_all_stock(struct mem_cgroup *root_memcg)
2325 /* If someone's already draining, avoid adding running more workers. */
2326 if (!mutex_trylock(&percpu_charge_mutex))
2329 * Notify other cpus that system-wide "drain" is running
2330 * We do not care about races with the cpu hotplug because cpu down
2331 * as well as workers from this path always operate on the local
2332 * per-cpu data. CPU up doesn't touch memcg_stock at all.
2335 for_each_online_cpu(cpu) {
2336 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2337 struct mem_cgroup *memcg;
2341 memcg = stock->cached;
2342 if (memcg && stock->nr_pages &&
2343 mem_cgroup_is_descendant(memcg, root_memcg))
2345 if (obj_stock_flush_required(stock, root_memcg))
2350 !test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2352 drain_local_stock(&stock->work);
2354 schedule_work_on(cpu, &stock->work);
2358 mutex_unlock(&percpu_charge_mutex);
2361 static int memcg_hotplug_cpu_dead(unsigned int cpu)
2363 struct memcg_stock_pcp *stock;
2364 struct mem_cgroup *memcg, *mi;
2366 stock = &per_cpu(memcg_stock, cpu);
2369 for_each_mem_cgroup(memcg) {
2372 for (i = 0; i < MEMCG_NR_STAT; i++) {
2376 x = this_cpu_xchg(memcg->vmstats_percpu->stat[i], 0);
2378 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2379 atomic_long_add(x, &memcg->vmstats[i]);
2381 if (i >= NR_VM_NODE_STAT_ITEMS)
2384 for_each_node(nid) {
2385 struct mem_cgroup_per_node *pn;
2387 pn = mem_cgroup_nodeinfo(memcg, nid);
2388 x = this_cpu_xchg(pn->lruvec_stat_cpu->count[i], 0);
2391 atomic_long_add(x, &pn->lruvec_stat[i]);
2392 } while ((pn = parent_nodeinfo(pn, nid)));
2396 for (i = 0; i < NR_VM_EVENT_ITEMS; i++) {
2399 x = this_cpu_xchg(memcg->vmstats_percpu->events[i], 0);
2401 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
2402 atomic_long_add(x, &memcg->vmevents[i]);
2409 static unsigned long reclaim_high(struct mem_cgroup *memcg,
2410 unsigned int nr_pages,
2413 unsigned long nr_reclaimed = 0;
2416 unsigned long pflags;
2418 if (page_counter_read(&memcg->memory) <=
2419 READ_ONCE(memcg->memory.high))
2422 memcg_memory_event(memcg, MEMCG_HIGH);
2424 psi_memstall_enter(&pflags);
2425 nr_reclaimed += try_to_free_mem_cgroup_pages(memcg, nr_pages,
2427 psi_memstall_leave(&pflags);
2428 } while ((memcg = parent_mem_cgroup(memcg)) &&
2429 !mem_cgroup_is_root(memcg));
2431 return nr_reclaimed;
2434 static void high_work_func(struct work_struct *work)
2436 struct mem_cgroup *memcg;
2438 memcg = container_of(work, struct mem_cgroup, high_work);
2439 reclaim_high(memcg, MEMCG_CHARGE_BATCH, GFP_KERNEL);
2443 * Clamp the maximum sleep time per allocation batch to 2 seconds. This is
2444 * enough to still cause a significant slowdown in most cases, while still
2445 * allowing diagnostics and tracing to proceed without becoming stuck.
2447 #define MEMCG_MAX_HIGH_DELAY_JIFFIES (2UL*HZ)
2450 * When calculating the delay, we use these either side of the exponentiation to
2451 * maintain precision and scale to a reasonable number of jiffies (see the table
2454 * - MEMCG_DELAY_PRECISION_SHIFT: Extra precision bits while translating the
2455 * overage ratio to a delay.
2456 * - MEMCG_DELAY_SCALING_SHIFT: The number of bits to scale down the
2457 * proposed penalty in order to reduce to a reasonable number of jiffies, and
2458 * to produce a reasonable delay curve.
2460 * MEMCG_DELAY_SCALING_SHIFT just happens to be a number that produces a
2461 * reasonable delay curve compared to precision-adjusted overage, not
2462 * penalising heavily at first, but still making sure that growth beyond the
2463 * limit penalises misbehaviour cgroups by slowing them down exponentially. For
2464 * example, with a high of 100 megabytes:
2466 * +-------+------------------------+
2467 * | usage | time to allocate in ms |
2468 * +-------+------------------------+
2490 * +-------+------------------------+
2492 #define MEMCG_DELAY_PRECISION_SHIFT 20
2493 #define MEMCG_DELAY_SCALING_SHIFT 14
2495 static u64 calculate_overage(unsigned long usage, unsigned long high)
2503 * Prevent division by 0 in overage calculation by acting as if
2504 * it was a threshold of 1 page
2506 high = max(high, 1UL);
2508 overage = usage - high;
2509 overage <<= MEMCG_DELAY_PRECISION_SHIFT;
2510 return div64_u64(overage, high);
2513 static u64 mem_find_max_overage(struct mem_cgroup *memcg)
2515 u64 overage, max_overage = 0;
2518 overage = calculate_overage(page_counter_read(&memcg->memory),
2519 READ_ONCE(memcg->memory.high));
2520 max_overage = max(overage, max_overage);
2521 } while ((memcg = parent_mem_cgroup(memcg)) &&
2522 !mem_cgroup_is_root(memcg));
2527 static u64 swap_find_max_overage(struct mem_cgroup *memcg)
2529 u64 overage, max_overage = 0;
2532 overage = calculate_overage(page_counter_read(&memcg->swap),
2533 READ_ONCE(memcg->swap.high));
2535 memcg_memory_event(memcg, MEMCG_SWAP_HIGH);
2536 max_overage = max(overage, max_overage);
2537 } while ((memcg = parent_mem_cgroup(memcg)) &&
2538 !mem_cgroup_is_root(memcg));
2544 * Get the number of jiffies that we should penalise a mischievous cgroup which
2545 * is exceeding its memory.high by checking both it and its ancestors.
2547 static unsigned long calculate_high_delay(struct mem_cgroup *memcg,
2548 unsigned int nr_pages,
2551 unsigned long penalty_jiffies;
2557 * We use overage compared to memory.high to calculate the number of
2558 * jiffies to sleep (penalty_jiffies). Ideally this value should be
2559 * fairly lenient on small overages, and increasingly harsh when the
2560 * memcg in question makes it clear that it has no intention of stopping
2561 * its crazy behaviour, so we exponentially increase the delay based on
2564 penalty_jiffies = max_overage * max_overage * HZ;
2565 penalty_jiffies >>= MEMCG_DELAY_PRECISION_SHIFT;
2566 penalty_jiffies >>= MEMCG_DELAY_SCALING_SHIFT;
2569 * Factor in the task's own contribution to the overage, such that four
2570 * N-sized allocations are throttled approximately the same as one
2571 * 4N-sized allocation.
2573 * MEMCG_CHARGE_BATCH pages is nominal, so work out how much smaller or
2574 * larger the current charge patch is than that.
2576 return penalty_jiffies * nr_pages / MEMCG_CHARGE_BATCH;
2580 * Scheduled by try_charge() to be executed from the userland return path
2581 * and reclaims memory over the high limit.
2583 void mem_cgroup_handle_over_high(void)
2585 unsigned long penalty_jiffies;
2586 unsigned long pflags;
2587 unsigned long nr_reclaimed;
2588 unsigned int nr_pages = current->memcg_nr_pages_over_high;
2589 int nr_retries = MAX_RECLAIM_RETRIES;
2590 struct mem_cgroup *memcg;
2591 bool in_retry = false;
2593 if (likely(!nr_pages))
2596 memcg = get_mem_cgroup_from_mm(current->mm);
2597 current->memcg_nr_pages_over_high = 0;
2601 * The allocating task should reclaim at least the batch size, but for
2602 * subsequent retries we only want to do what's necessary to prevent oom
2603 * or breaching resource isolation.
2605 * This is distinct from memory.max or page allocator behaviour because
2606 * memory.high is currently batched, whereas memory.max and the page
2607 * allocator run every time an allocation is made.
2609 nr_reclaimed = reclaim_high(memcg,
2610 in_retry ? SWAP_CLUSTER_MAX : nr_pages,
2614 * memory.high is breached and reclaim is unable to keep up. Throttle
2615 * allocators proactively to slow down excessive growth.
2617 penalty_jiffies = calculate_high_delay(memcg, nr_pages,
2618 mem_find_max_overage(memcg));
2620 penalty_jiffies += calculate_high_delay(memcg, nr_pages,
2621 swap_find_max_overage(memcg));
2624 * Clamp the max delay per usermode return so as to still keep the
2625 * application moving forwards and also permit diagnostics, albeit
2628 penalty_jiffies = min(penalty_jiffies, MEMCG_MAX_HIGH_DELAY_JIFFIES);
2631 * Don't sleep if the amount of jiffies this memcg owes us is so low
2632 * that it's not even worth doing, in an attempt to be nice to those who
2633 * go only a small amount over their memory.high value and maybe haven't
2634 * been aggressively reclaimed enough yet.
2636 if (penalty_jiffies <= HZ / 100)
2640 * If reclaim is making forward progress but we're still over
2641 * memory.high, we want to encourage that rather than doing allocator
2644 if (nr_reclaimed || nr_retries--) {
2650 * If we exit early, we're guaranteed to die (since
2651 * schedule_timeout_killable sets TASK_KILLABLE). This means we don't
2652 * need to account for any ill-begotten jiffies to pay them off later.
2654 psi_memstall_enter(&pflags);
2655 schedule_timeout_killable(penalty_jiffies);
2656 psi_memstall_leave(&pflags);
2659 css_put(&memcg->css);
2662 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2663 unsigned int nr_pages)
2665 unsigned int batch = max(MEMCG_CHARGE_BATCH, nr_pages);
2666 int nr_retries = MAX_RECLAIM_RETRIES;
2667 struct mem_cgroup *mem_over_limit;
2668 struct page_counter *counter;
2669 enum oom_status oom_status;
2670 unsigned long nr_reclaimed;
2671 bool may_swap = true;
2672 bool drained = false;
2673 unsigned long pflags;
2675 if (mem_cgroup_is_root(memcg))
2678 if (consume_stock(memcg, nr_pages))
2681 if (!do_memsw_account() ||
2682 page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2683 if (page_counter_try_charge(&memcg->memory, batch, &counter))
2685 if (do_memsw_account())
2686 page_counter_uncharge(&memcg->memsw, batch);
2687 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2689 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2693 if (batch > nr_pages) {
2699 * Memcg doesn't have a dedicated reserve for atomic
2700 * allocations. But like the global atomic pool, we need to
2701 * put the burden of reclaim on regular allocation requests
2702 * and let these go through as privileged allocations.
2704 if (gfp_mask & __GFP_ATOMIC)
2708 * Unlike in global OOM situations, memcg is not in a physical
2709 * memory shortage. Allow dying and OOM-killed tasks to
2710 * bypass the last charges so that they can exit quickly and
2711 * free their memory.
2713 if (unlikely(should_force_charge()))
2717 * Prevent unbounded recursion when reclaim operations need to
2718 * allocate memory. This might exceed the limits temporarily,
2719 * but we prefer facilitating memory reclaim and getting back
2720 * under the limit over triggering OOM kills in these cases.
2722 if (unlikely(current->flags & PF_MEMALLOC))
2725 if (unlikely(task_in_memcg_oom(current)))
2728 if (!gfpflags_allow_blocking(gfp_mask))
2731 memcg_memory_event(mem_over_limit, MEMCG_MAX);
2733 psi_memstall_enter(&pflags);
2734 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2735 gfp_mask, may_swap);
2736 psi_memstall_leave(&pflags);
2738 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2742 drain_all_stock(mem_over_limit);
2747 if (gfp_mask & __GFP_NORETRY)
2750 * Even though the limit is exceeded at this point, reclaim
2751 * may have been able to free some pages. Retry the charge
2752 * before killing the task.
2754 * Only for regular pages, though: huge pages are rather
2755 * unlikely to succeed so close to the limit, and we fall back
2756 * to regular pages anyway in case of failure.
2758 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2761 * At task move, charge accounts can be doubly counted. So, it's
2762 * better to wait until the end of task_move if something is going on.
2764 if (mem_cgroup_wait_acct_move(mem_over_limit))
2770 if (gfp_mask & __GFP_RETRY_MAYFAIL)
2773 if (gfp_mask & __GFP_NOFAIL)
2776 if (fatal_signal_pending(current))
2780 * keep retrying as long as the memcg oom killer is able to make
2781 * a forward progress or bypass the charge if the oom killer
2782 * couldn't make any progress.
2784 oom_status = mem_cgroup_oom(mem_over_limit, gfp_mask,
2785 get_order(nr_pages * PAGE_SIZE));
2786 switch (oom_status) {
2788 nr_retries = MAX_RECLAIM_RETRIES;
2796 if (!(gfp_mask & __GFP_NOFAIL))
2800 * The allocation either can't fail or will lead to more memory
2801 * being freed very soon. Allow memory usage go over the limit
2802 * temporarily by force charging it.
2804 page_counter_charge(&memcg->memory, nr_pages);
2805 if (do_memsw_account())
2806 page_counter_charge(&memcg->memsw, nr_pages);
2811 if (batch > nr_pages)
2812 refill_stock(memcg, batch - nr_pages);
2815 * If the hierarchy is above the normal consumption range, schedule
2816 * reclaim on returning to userland. We can perform reclaim here
2817 * if __GFP_RECLAIM but let's always punt for simplicity and so that
2818 * GFP_KERNEL can consistently be used during reclaim. @memcg is
2819 * not recorded as it most likely matches current's and won't
2820 * change in the meantime. As high limit is checked again before
2821 * reclaim, the cost of mismatch is negligible.
2824 bool mem_high, swap_high;
2826 mem_high = page_counter_read(&memcg->memory) >
2827 READ_ONCE(memcg->memory.high);
2828 swap_high = page_counter_read(&memcg->swap) >
2829 READ_ONCE(memcg->swap.high);
2831 /* Don't bother a random interrupted task */
2832 if (in_interrupt()) {
2834 schedule_work(&memcg->high_work);
2840 if (mem_high || swap_high) {
2842 * The allocating tasks in this cgroup will need to do
2843 * reclaim or be throttled to prevent further growth
2844 * of the memory or swap footprints.
2846 * Target some best-effort fairness between the tasks,
2847 * and distribute reclaim work and delay penalties
2848 * based on how much each task is actually allocating.
2850 current->memcg_nr_pages_over_high += batch;
2851 set_notify_resume(current);
2854 } while ((memcg = parent_mem_cgroup(memcg)));
2859 #if defined(CONFIG_MEMCG_KMEM) || defined(CONFIG_MMU)
2860 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2862 if (mem_cgroup_is_root(memcg))
2865 page_counter_uncharge(&memcg->memory, nr_pages);
2866 if (do_memsw_account())
2867 page_counter_uncharge(&memcg->memsw, nr_pages);
2871 static void commit_charge(struct page *page, struct mem_cgroup *memcg)
2873 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2875 * Any of the following ensures page's memcg stability:
2879 * - lock_page_memcg()
2880 * - exclusive reference
2882 page->mem_cgroup = memcg;
2885 #ifdef CONFIG_MEMCG_KMEM
2886 int memcg_alloc_page_obj_cgroups(struct page *page, struct kmem_cache *s,
2889 unsigned int objects = objs_per_slab_page(s, page);
2892 vec = kcalloc_node(objects, sizeof(struct obj_cgroup *), gfp,
2897 if (cmpxchg(&page->obj_cgroups, NULL,
2898 (struct obj_cgroup **) ((unsigned long)vec | 0x1UL)))
2901 kmemleak_not_leak(vec);
2907 * Returns a pointer to the memory cgroup to which the kernel object is charged.
2909 * The caller must ensure the memcg lifetime, e.g. by taking rcu_read_lock(),
2910 * cgroup_mutex, etc.
2912 struct mem_cgroup *mem_cgroup_from_obj(void *p)
2916 if (mem_cgroup_disabled())
2919 page = virt_to_head_page(p);
2922 * If page->mem_cgroup is set, it's either a simple mem_cgroup pointer
2923 * or a pointer to obj_cgroup vector. In the latter case the lowest
2924 * bit of the pointer is set.
2925 * The page->mem_cgroup pointer can be asynchronously changed
2926 * from NULL to (obj_cgroup_vec | 0x1UL), but can't be changed
2927 * from a valid memcg pointer to objcg vector or back.
2929 if (!page->mem_cgroup)
2933 * Slab objects are accounted individually, not per-page.
2934 * Memcg membership data for each individual object is saved in
2935 * the page->obj_cgroups.
2937 if (page_has_obj_cgroups(page)) {
2938 struct obj_cgroup *objcg;
2941 off = obj_to_index(page->slab_cache, page, p);
2942 objcg = page_obj_cgroups(page)[off];
2944 return obj_cgroup_memcg(objcg);
2949 /* All other pages use page->mem_cgroup */
2950 return page->mem_cgroup;
2953 __always_inline struct obj_cgroup *get_obj_cgroup_from_current(void)
2955 struct obj_cgroup *objcg = NULL;
2956 struct mem_cgroup *memcg;
2958 if (memcg_kmem_bypass())
2962 if (unlikely(active_memcg()))
2963 memcg = active_memcg();
2965 memcg = mem_cgroup_from_task(current);
2967 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
2968 objcg = rcu_dereference(memcg->objcg);
2969 if (objcg && obj_cgroup_tryget(objcg))
2978 static int memcg_alloc_cache_id(void)
2983 id = ida_simple_get(&memcg_cache_ida,
2984 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2988 if (id < memcg_nr_cache_ids)
2992 * There's no space for the new id in memcg_caches arrays,
2993 * so we have to grow them.
2995 down_write(&memcg_cache_ids_sem);
2997 size = 2 * (id + 1);
2998 if (size < MEMCG_CACHES_MIN_SIZE)
2999 size = MEMCG_CACHES_MIN_SIZE;
3000 else if (size > MEMCG_CACHES_MAX_SIZE)
3001 size = MEMCG_CACHES_MAX_SIZE;
3003 err = memcg_update_all_list_lrus(size);
3005 memcg_nr_cache_ids = size;
3007 up_write(&memcg_cache_ids_sem);
3010 ida_simple_remove(&memcg_cache_ida, id);
3016 static void memcg_free_cache_id(int id)
3018 ida_simple_remove(&memcg_cache_ida, id);
3022 * __memcg_kmem_charge: charge a number of kernel pages to a memcg
3023 * @memcg: memory cgroup to charge
3024 * @gfp: reclaim mode
3025 * @nr_pages: number of pages to charge
3027 * Returns 0 on success, an error code on failure.
3029 int __memcg_kmem_charge(struct mem_cgroup *memcg, gfp_t gfp,
3030 unsigned int nr_pages)
3032 struct page_counter *counter;
3035 ret = try_charge(memcg, gfp, nr_pages);
3039 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) &&
3040 !page_counter_try_charge(&memcg->kmem, nr_pages, &counter)) {
3043 * Enforce __GFP_NOFAIL allocation because callers are not
3044 * prepared to see failures and likely do not have any failure
3047 if (gfp & __GFP_NOFAIL) {
3048 page_counter_charge(&memcg->kmem, nr_pages);
3051 cancel_charge(memcg, nr_pages);
3058 * __memcg_kmem_uncharge: uncharge a number of kernel pages from a memcg
3059 * @memcg: memcg to uncharge
3060 * @nr_pages: number of pages to uncharge
3062 void __memcg_kmem_uncharge(struct mem_cgroup *memcg, unsigned int nr_pages)
3064 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
3065 page_counter_uncharge(&memcg->kmem, nr_pages);
3067 page_counter_uncharge(&memcg->memory, nr_pages);
3068 if (do_memsw_account())
3069 page_counter_uncharge(&memcg->memsw, nr_pages);
3073 * __memcg_kmem_charge_page: charge a kmem page to the current memory cgroup
3074 * @page: page to charge
3075 * @gfp: reclaim mode
3076 * @order: allocation order
3078 * Returns 0 on success, an error code on failure.
3080 int __memcg_kmem_charge_page(struct page *page, gfp_t gfp, int order)
3082 struct mem_cgroup *memcg;
3085 memcg = get_mem_cgroup_from_current();
3086 if (memcg && !mem_cgroup_is_root(memcg)) {
3087 ret = __memcg_kmem_charge(memcg, gfp, 1 << order);
3089 page->mem_cgroup = memcg;
3090 __SetPageKmemcg(page);
3093 css_put(&memcg->css);
3099 * __memcg_kmem_uncharge_page: uncharge a kmem page
3100 * @page: page to uncharge
3101 * @order: allocation order
3103 void __memcg_kmem_uncharge_page(struct page *page, int order)
3105 struct mem_cgroup *memcg = page->mem_cgroup;
3106 unsigned int nr_pages = 1 << order;
3111 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3112 __memcg_kmem_uncharge(memcg, nr_pages);
3113 page->mem_cgroup = NULL;
3114 css_put(&memcg->css);
3116 /* slab pages do not have PageKmemcg flag set */
3117 if (PageKmemcg(page))
3118 __ClearPageKmemcg(page);
3121 static bool consume_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3123 struct memcg_stock_pcp *stock;
3124 unsigned long flags;
3127 local_irq_save(flags);
3129 stock = this_cpu_ptr(&memcg_stock);
3130 if (objcg == stock->cached_objcg && stock->nr_bytes >= nr_bytes) {
3131 stock->nr_bytes -= nr_bytes;
3135 local_irq_restore(flags);
3140 static void drain_obj_stock(struct memcg_stock_pcp *stock)
3142 struct obj_cgroup *old = stock->cached_objcg;
3147 if (stock->nr_bytes) {
3148 unsigned int nr_pages = stock->nr_bytes >> PAGE_SHIFT;
3149 unsigned int nr_bytes = stock->nr_bytes & (PAGE_SIZE - 1);
3153 __memcg_kmem_uncharge(obj_cgroup_memcg(old), nr_pages);
3158 * The leftover is flushed to the centralized per-memcg value.
3159 * On the next attempt to refill obj stock it will be moved
3160 * to a per-cpu stock (probably, on an other CPU), see
3161 * refill_obj_stock().
3163 * How often it's flushed is a trade-off between the memory
3164 * limit enforcement accuracy and potential CPU contention,
3165 * so it might be changed in the future.
3167 atomic_add(nr_bytes, &old->nr_charged_bytes);
3168 stock->nr_bytes = 0;
3171 obj_cgroup_put(old);
3172 stock->cached_objcg = NULL;
3175 static bool obj_stock_flush_required(struct memcg_stock_pcp *stock,
3176 struct mem_cgroup *root_memcg)
3178 struct mem_cgroup *memcg;
3180 if (stock->cached_objcg) {
3181 memcg = obj_cgroup_memcg(stock->cached_objcg);
3182 if (memcg && mem_cgroup_is_descendant(memcg, root_memcg))
3189 static void refill_obj_stock(struct obj_cgroup *objcg, unsigned int nr_bytes)
3191 struct memcg_stock_pcp *stock;
3192 unsigned long flags;
3194 local_irq_save(flags);
3196 stock = this_cpu_ptr(&memcg_stock);
3197 if (stock->cached_objcg != objcg) { /* reset if necessary */
3198 drain_obj_stock(stock);
3199 obj_cgroup_get(objcg);
3200 stock->cached_objcg = objcg;
3201 stock->nr_bytes = atomic_xchg(&objcg->nr_charged_bytes, 0);
3203 stock->nr_bytes += nr_bytes;
3205 if (stock->nr_bytes > PAGE_SIZE)
3206 drain_obj_stock(stock);
3208 local_irq_restore(flags);
3211 int obj_cgroup_charge(struct obj_cgroup *objcg, gfp_t gfp, size_t size)
3213 struct mem_cgroup *memcg;
3214 unsigned int nr_pages, nr_bytes;
3217 if (consume_obj_stock(objcg, size))
3221 * In theory, memcg->nr_charged_bytes can have enough
3222 * pre-charged bytes to satisfy the allocation. However,
3223 * flushing memcg->nr_charged_bytes requires two atomic
3224 * operations, and memcg->nr_charged_bytes can't be big,
3225 * so it's better to ignore it and try grab some new pages.
3226 * memcg->nr_charged_bytes will be flushed in
3227 * refill_obj_stock(), called from this function or
3228 * independently later.
3232 memcg = obj_cgroup_memcg(objcg);
3233 if (unlikely(!css_tryget(&memcg->css)))
3237 nr_pages = size >> PAGE_SHIFT;
3238 nr_bytes = size & (PAGE_SIZE - 1);
3243 ret = __memcg_kmem_charge(memcg, gfp, nr_pages);
3244 if (!ret && nr_bytes)
3245 refill_obj_stock(objcg, PAGE_SIZE - nr_bytes);
3247 css_put(&memcg->css);
3251 void obj_cgroup_uncharge(struct obj_cgroup *objcg, size_t size)
3253 refill_obj_stock(objcg, size);
3256 #endif /* CONFIG_MEMCG_KMEM */
3258 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3261 * Because tail pages are not marked as "used", set it. We're under
3262 * pgdat->lru_lock and migration entries setup in all page mappings.
3264 void mem_cgroup_split_huge_fixup(struct page *head)
3266 struct mem_cgroup *memcg = head->mem_cgroup;
3269 if (mem_cgroup_disabled())
3272 for (i = 1; i < HPAGE_PMD_NR; i++) {
3273 css_get(&memcg->css);
3274 head[i].mem_cgroup = memcg;
3277 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3279 #ifdef CONFIG_MEMCG_SWAP
3281 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3282 * @entry: swap entry to be moved
3283 * @from: mem_cgroup which the entry is moved from
3284 * @to: mem_cgroup which the entry is moved to
3286 * It succeeds only when the swap_cgroup's record for this entry is the same
3287 * as the mem_cgroup's id of @from.
3289 * Returns 0 on success, -EINVAL on failure.
3291 * The caller must have charged to @to, IOW, called page_counter_charge() about
3292 * both res and memsw, and called css_get().
3294 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3295 struct mem_cgroup *from, struct mem_cgroup *to)
3297 unsigned short old_id, new_id;
3299 old_id = mem_cgroup_id(from);
3300 new_id = mem_cgroup_id(to);
3302 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3303 mod_memcg_state(from, MEMCG_SWAP, -1);
3304 mod_memcg_state(to, MEMCG_SWAP, 1);
3310 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3311 struct mem_cgroup *from, struct mem_cgroup *to)
3317 static DEFINE_MUTEX(memcg_max_mutex);
3319 static int mem_cgroup_resize_max(struct mem_cgroup *memcg,
3320 unsigned long max, bool memsw)
3322 bool enlarge = false;
3323 bool drained = false;
3325 bool limits_invariant;
3326 struct page_counter *counter = memsw ? &memcg->memsw : &memcg->memory;
3329 if (signal_pending(current)) {
3334 mutex_lock(&memcg_max_mutex);
3336 * Make sure that the new limit (memsw or memory limit) doesn't
3337 * break our basic invariant rule memory.max <= memsw.max.
3339 limits_invariant = memsw ? max >= READ_ONCE(memcg->memory.max) :
3340 max <= memcg->memsw.max;
3341 if (!limits_invariant) {
3342 mutex_unlock(&memcg_max_mutex);
3346 if (max > counter->max)
3348 ret = page_counter_set_max(counter, max);
3349 mutex_unlock(&memcg_max_mutex);
3355 drain_all_stock(memcg);
3360 if (!try_to_free_mem_cgroup_pages(memcg, 1,
3361 GFP_KERNEL, !memsw)) {
3367 if (!ret && enlarge)
3368 memcg_oom_recover(memcg);
3373 unsigned long mem_cgroup_soft_limit_reclaim(pg_data_t *pgdat, int order,
3375 unsigned long *total_scanned)
3377 unsigned long nr_reclaimed = 0;
3378 struct mem_cgroup_per_node *mz, *next_mz = NULL;
3379 unsigned long reclaimed;
3381 struct mem_cgroup_tree_per_node *mctz;
3382 unsigned long excess;
3383 unsigned long nr_scanned;
3388 mctz = soft_limit_tree_node(pgdat->node_id);
3391 * Do not even bother to check the largest node if the root
3392 * is empty. Do it lockless to prevent lock bouncing. Races
3393 * are acceptable as soft limit is best effort anyway.
3395 if (!mctz || RB_EMPTY_ROOT(&mctz->rb_root))
3399 * This loop can run a while, specially if mem_cgroup's continuously
3400 * keep exceeding their soft limit and putting the system under
3407 mz = mem_cgroup_largest_soft_limit_node(mctz);
3412 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, pgdat,
3413 gfp_mask, &nr_scanned);
3414 nr_reclaimed += reclaimed;
3415 *total_scanned += nr_scanned;
3416 spin_lock_irq(&mctz->lock);
3417 __mem_cgroup_remove_exceeded(mz, mctz);
3420 * If we failed to reclaim anything from this memory cgroup
3421 * it is time to move on to the next cgroup
3425 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3427 excess = soft_limit_excess(mz->memcg);
3429 * One school of thought says that we should not add
3430 * back the node to the tree if reclaim returns 0.
3431 * But our reclaim could return 0, simply because due
3432 * to priority we are exposing a smaller subset of
3433 * memory to reclaim from. Consider this as a longer
3436 /* If excess == 0, no tree ops */
3437 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3438 spin_unlock_irq(&mctz->lock);
3439 css_put(&mz->memcg->css);
3442 * Could not reclaim anything and there are no more
3443 * mem cgroups to try or we seem to be looping without
3444 * reclaiming anything.
3446 if (!nr_reclaimed &&
3448 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3450 } while (!nr_reclaimed);
3452 css_put(&next_mz->memcg->css);
3453 return nr_reclaimed;
3457 * Reclaims as many pages from the given memcg as possible.
3459 * Caller is responsible for holding css reference for memcg.
3461 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3463 int nr_retries = MAX_RECLAIM_RETRIES;
3465 /* we call try-to-free pages for make this cgroup empty */
3466 lru_add_drain_all();
3468 drain_all_stock(memcg);
3470 /* try to free all pages in this cgroup */
3471 while (nr_retries && page_counter_read(&memcg->memory)) {
3474 if (signal_pending(current))
3477 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3481 /* maybe some writeback is necessary */
3482 congestion_wait(BLK_RW_ASYNC, HZ/10);
3490 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3491 char *buf, size_t nbytes,
3494 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3496 if (mem_cgroup_is_root(memcg))
3498 return mem_cgroup_force_empty(memcg) ?: nbytes;
3501 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3507 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3508 struct cftype *cft, u64 val)
3513 pr_warn_once("Non-hierarchical mode is deprecated. "
3514 "Please report your usecase to linux-mm@kvack.org if you "
3515 "depend on this functionality.\n");
3520 static unsigned long mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3524 if (mem_cgroup_is_root(memcg)) {
3525 val = memcg_page_state(memcg, NR_FILE_PAGES) +
3526 memcg_page_state(memcg, NR_ANON_MAPPED);
3528 val += memcg_page_state(memcg, MEMCG_SWAP);
3531 val = page_counter_read(&memcg->memory);
3533 val = page_counter_read(&memcg->memsw);
3546 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3549 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3550 struct page_counter *counter;
3552 switch (MEMFILE_TYPE(cft->private)) {
3554 counter = &memcg->memory;
3557 counter = &memcg->memsw;
3560 counter = &memcg->kmem;
3563 counter = &memcg->tcpmem;
3569 switch (MEMFILE_ATTR(cft->private)) {
3571 if (counter == &memcg->memory)
3572 return (u64)mem_cgroup_usage(memcg, false) * PAGE_SIZE;
3573 if (counter == &memcg->memsw)
3574 return (u64)mem_cgroup_usage(memcg, true) * PAGE_SIZE;
3575 return (u64)page_counter_read(counter) * PAGE_SIZE;
3577 return (u64)counter->max * PAGE_SIZE;
3579 return (u64)counter->watermark * PAGE_SIZE;
3581 return counter->failcnt;
3582 case RES_SOFT_LIMIT:
3583 return (u64)memcg->soft_limit * PAGE_SIZE;
3589 static void memcg_flush_percpu_vmstats(struct mem_cgroup *memcg)
3591 unsigned long stat[MEMCG_NR_STAT] = {0};
3592 struct mem_cgroup *mi;
3595 for_each_online_cpu(cpu)
3596 for (i = 0; i < MEMCG_NR_STAT; i++)
3597 stat[i] += per_cpu(memcg->vmstats_percpu->stat[i], cpu);
3599 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3600 for (i = 0; i < MEMCG_NR_STAT; i++)
3601 atomic_long_add(stat[i], &mi->vmstats[i]);
3603 for_each_node(node) {
3604 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
3605 struct mem_cgroup_per_node *pi;
3607 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3610 for_each_online_cpu(cpu)
3611 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3613 pn->lruvec_stat_cpu->count[i], cpu);
3615 for (pi = pn; pi; pi = parent_nodeinfo(pi, node))
3616 for (i = 0; i < NR_VM_NODE_STAT_ITEMS; i++)
3617 atomic_long_add(stat[i], &pi->lruvec_stat[i]);
3621 static void memcg_flush_percpu_vmevents(struct mem_cgroup *memcg)
3623 unsigned long events[NR_VM_EVENT_ITEMS];
3624 struct mem_cgroup *mi;
3627 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3630 for_each_online_cpu(cpu)
3631 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3632 events[i] += per_cpu(memcg->vmstats_percpu->events[i],
3635 for (mi = memcg; mi; mi = parent_mem_cgroup(mi))
3636 for (i = 0; i < NR_VM_EVENT_ITEMS; i++)
3637 atomic_long_add(events[i], &mi->vmevents[i]);
3640 #ifdef CONFIG_MEMCG_KMEM
3641 static int memcg_online_kmem(struct mem_cgroup *memcg)
3643 struct obj_cgroup *objcg;
3646 if (cgroup_memory_nokmem)
3649 BUG_ON(memcg->kmemcg_id >= 0);
3650 BUG_ON(memcg->kmem_state);
3652 memcg_id = memcg_alloc_cache_id();
3656 objcg = obj_cgroup_alloc();
3658 memcg_free_cache_id(memcg_id);
3661 objcg->memcg = memcg;
3662 rcu_assign_pointer(memcg->objcg, objcg);
3664 static_branch_enable(&memcg_kmem_enabled_key);
3666 memcg->kmemcg_id = memcg_id;
3667 memcg->kmem_state = KMEM_ONLINE;
3672 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3674 struct cgroup_subsys_state *css;
3675 struct mem_cgroup *parent, *child;
3678 if (memcg->kmem_state != KMEM_ONLINE)
3681 memcg->kmem_state = KMEM_ALLOCATED;
3683 parent = parent_mem_cgroup(memcg);
3685 parent = root_mem_cgroup;
3687 memcg_reparent_objcgs(memcg, parent);
3689 kmemcg_id = memcg->kmemcg_id;
3690 BUG_ON(kmemcg_id < 0);
3693 * Change kmemcg_id of this cgroup and all its descendants to the
3694 * parent's id, and then move all entries from this cgroup's list_lrus
3695 * to ones of the parent. After we have finished, all list_lrus
3696 * corresponding to this cgroup are guaranteed to remain empty. The
3697 * ordering is imposed by list_lru_node->lock taken by
3698 * memcg_drain_all_list_lrus().
3700 rcu_read_lock(); /* can be called from css_free w/o cgroup_mutex */
3701 css_for_each_descendant_pre(css, &memcg->css) {
3702 child = mem_cgroup_from_css(css);
3703 BUG_ON(child->kmemcg_id != kmemcg_id);
3704 child->kmemcg_id = parent->kmemcg_id;
3708 memcg_drain_all_list_lrus(kmemcg_id, parent);
3710 memcg_free_cache_id(kmemcg_id);
3713 static void memcg_free_kmem(struct mem_cgroup *memcg)
3715 /* css_alloc() failed, offlining didn't happen */
3716 if (unlikely(memcg->kmem_state == KMEM_ONLINE))
3717 memcg_offline_kmem(memcg);
3720 static int memcg_online_kmem(struct mem_cgroup *memcg)
3724 static void memcg_offline_kmem(struct mem_cgroup *memcg)
3727 static void memcg_free_kmem(struct mem_cgroup *memcg)
3730 #endif /* CONFIG_MEMCG_KMEM */
3732 static int memcg_update_kmem_max(struct mem_cgroup *memcg,
3737 mutex_lock(&memcg_max_mutex);
3738 ret = page_counter_set_max(&memcg->kmem, max);
3739 mutex_unlock(&memcg_max_mutex);
3743 static int memcg_update_tcp_max(struct mem_cgroup *memcg, unsigned long max)
3747 mutex_lock(&memcg_max_mutex);
3749 ret = page_counter_set_max(&memcg->tcpmem, max);
3753 if (!memcg->tcpmem_active) {
3755 * The active flag needs to be written after the static_key
3756 * update. This is what guarantees that the socket activation
3757 * function is the last one to run. See mem_cgroup_sk_alloc()
3758 * for details, and note that we don't mark any socket as
3759 * belonging to this memcg until that flag is up.
3761 * We need to do this, because static_keys will span multiple
3762 * sites, but we can't control their order. If we mark a socket
3763 * as accounted, but the accounting functions are not patched in
3764 * yet, we'll lose accounting.
3766 * We never race with the readers in mem_cgroup_sk_alloc(),
3767 * because when this value change, the code to process it is not
3770 static_branch_inc(&memcg_sockets_enabled_key);
3771 memcg->tcpmem_active = true;
3774 mutex_unlock(&memcg_max_mutex);
3779 * The user of this function is...
3782 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3783 char *buf, size_t nbytes, loff_t off)
3785 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3786 unsigned long nr_pages;
3789 buf = strstrip(buf);
3790 ret = page_counter_memparse(buf, "-1", &nr_pages);
3794 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3796 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3800 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3802 ret = mem_cgroup_resize_max(memcg, nr_pages, false);
3805 ret = mem_cgroup_resize_max(memcg, nr_pages, true);
3808 pr_warn_once("kmem.limit_in_bytes is deprecated and will be removed. "
3809 "Please report your usecase to linux-mm@kvack.org if you "
3810 "depend on this functionality.\n");
3811 ret = memcg_update_kmem_max(memcg, nr_pages);
3814 ret = memcg_update_tcp_max(memcg, nr_pages);
3818 case RES_SOFT_LIMIT:
3819 memcg->soft_limit = nr_pages;
3823 return ret ?: nbytes;
3826 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3827 size_t nbytes, loff_t off)
3829 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3830 struct page_counter *counter;
3832 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3834 counter = &memcg->memory;
3837 counter = &memcg->memsw;
3840 counter = &memcg->kmem;
3843 counter = &memcg->tcpmem;
3849 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3851 page_counter_reset_watermark(counter);
3854 counter->failcnt = 0;
3863 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3866 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3870 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3871 struct cftype *cft, u64 val)
3873 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3875 if (val & ~MOVE_MASK)
3879 * No kind of locking is needed in here, because ->can_attach() will
3880 * check this value once in the beginning of the process, and then carry
3881 * on with stale data. This means that changes to this value will only
3882 * affect task migrations starting after the change.
3884 memcg->move_charge_at_immigrate = val;
3888 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3889 struct cftype *cft, u64 val)
3897 #define LRU_ALL_FILE (BIT(LRU_INACTIVE_FILE) | BIT(LRU_ACTIVE_FILE))
3898 #define LRU_ALL_ANON (BIT(LRU_INACTIVE_ANON) | BIT(LRU_ACTIVE_ANON))
3899 #define LRU_ALL ((1 << NR_LRU_LISTS) - 1)
3901 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
3902 int nid, unsigned int lru_mask, bool tree)
3904 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
3905 unsigned long nr = 0;
3908 VM_BUG_ON((unsigned)nid >= nr_node_ids);
3911 if (!(BIT(lru) & lru_mask))
3914 nr += lruvec_page_state(lruvec, NR_LRU_BASE + lru);
3916 nr += lruvec_page_state_local(lruvec, NR_LRU_BASE + lru);
3921 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
3922 unsigned int lru_mask,
3925 unsigned long nr = 0;
3929 if (!(BIT(lru) & lru_mask))
3932 nr += memcg_page_state(memcg, NR_LRU_BASE + lru);
3934 nr += memcg_page_state_local(memcg, NR_LRU_BASE + lru);
3939 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3943 unsigned int lru_mask;
3946 static const struct numa_stat stats[] = {
3947 { "total", LRU_ALL },
3948 { "file", LRU_ALL_FILE },
3949 { "anon", LRU_ALL_ANON },
3950 { "unevictable", BIT(LRU_UNEVICTABLE) },
3952 const struct numa_stat *stat;
3954 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
3956 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3957 seq_printf(m, "%s=%lu", stat->name,
3958 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3960 for_each_node_state(nid, N_MEMORY)
3961 seq_printf(m, " N%d=%lu", nid,
3962 mem_cgroup_node_nr_lru_pages(memcg, nid,
3963 stat->lru_mask, false));
3967 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3969 seq_printf(m, "hierarchical_%s=%lu", stat->name,
3970 mem_cgroup_nr_lru_pages(memcg, stat->lru_mask,
3972 for_each_node_state(nid, N_MEMORY)
3973 seq_printf(m, " N%d=%lu", nid,
3974 mem_cgroup_node_nr_lru_pages(memcg, nid,
3975 stat->lru_mask, true));
3981 #endif /* CONFIG_NUMA */
3983 static const unsigned int memcg1_stats[] = {
3986 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3996 static const char *const memcg1_stat_names[] = {
3999 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4009 /* Universal VM events cgroup1 shows, original sort order */
4010 static const unsigned int memcg1_events[] = {
4017 static int memcg_stat_show(struct seq_file *m, void *v)
4019 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
4020 unsigned long memory, memsw;
4021 struct mem_cgroup *mi;
4024 BUILD_BUG_ON(ARRAY_SIZE(memcg1_stat_names) != ARRAY_SIZE(memcg1_stats));
4026 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4029 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4031 nr = memcg_page_state_local(memcg, memcg1_stats[i]);
4032 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4033 if (memcg1_stats[i] == NR_ANON_THPS)
4036 seq_printf(m, "%s %lu\n", memcg1_stat_names[i], nr * PAGE_SIZE);
4039 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4040 seq_printf(m, "%s %lu\n", vm_event_name(memcg1_events[i]),
4041 memcg_events_local(memcg, memcg1_events[i]));
4043 for (i = 0; i < NR_LRU_LISTS; i++)
4044 seq_printf(m, "%s %lu\n", lru_list_name(i),
4045 memcg_page_state_local(memcg, NR_LRU_BASE + i) *
4048 /* Hierarchical information */
4049 memory = memsw = PAGE_COUNTER_MAX;
4050 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
4051 memory = min(memory, READ_ONCE(mi->memory.max));
4052 memsw = min(memsw, READ_ONCE(mi->memsw.max));
4054 seq_printf(m, "hierarchical_memory_limit %llu\n",
4055 (u64)memory * PAGE_SIZE);
4056 if (do_memsw_account())
4057 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4058 (u64)memsw * PAGE_SIZE);
4060 for (i = 0; i < ARRAY_SIZE(memcg1_stats); i++) {
4063 if (memcg1_stats[i] == MEMCG_SWAP && !do_memsw_account())
4065 nr = memcg_page_state(memcg, memcg1_stats[i]);
4066 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4067 if (memcg1_stats[i] == NR_ANON_THPS)
4070 seq_printf(m, "total_%s %llu\n", memcg1_stat_names[i],
4071 (u64)nr * PAGE_SIZE);
4074 for (i = 0; i < ARRAY_SIZE(memcg1_events); i++)
4075 seq_printf(m, "total_%s %llu\n",
4076 vm_event_name(memcg1_events[i]),
4077 (u64)memcg_events(memcg, memcg1_events[i]));
4079 for (i = 0; i < NR_LRU_LISTS; i++)
4080 seq_printf(m, "total_%s %llu\n", lru_list_name(i),
4081 (u64)memcg_page_state(memcg, NR_LRU_BASE + i) *
4084 #ifdef CONFIG_DEBUG_VM
4087 struct mem_cgroup_per_node *mz;
4088 unsigned long anon_cost = 0;
4089 unsigned long file_cost = 0;
4091 for_each_online_pgdat(pgdat) {
4092 mz = mem_cgroup_nodeinfo(memcg, pgdat->node_id);
4094 anon_cost += mz->lruvec.anon_cost;
4095 file_cost += mz->lruvec.file_cost;
4097 seq_printf(m, "anon_cost %lu\n", anon_cost);
4098 seq_printf(m, "file_cost %lu\n", file_cost);
4105 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4108 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4110 return mem_cgroup_swappiness(memcg);
4113 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4114 struct cftype *cft, u64 val)
4116 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4122 memcg->swappiness = val;
4124 vm_swappiness = val;
4129 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4131 struct mem_cgroup_threshold_ary *t;
4132 unsigned long usage;
4137 t = rcu_dereference(memcg->thresholds.primary);
4139 t = rcu_dereference(memcg->memsw_thresholds.primary);
4144 usage = mem_cgroup_usage(memcg, swap);
4147 * current_threshold points to threshold just below or equal to usage.
4148 * If it's not true, a threshold was crossed after last
4149 * call of __mem_cgroup_threshold().
4151 i = t->current_threshold;
4154 * Iterate backward over array of thresholds starting from
4155 * current_threshold and check if a threshold is crossed.
4156 * If none of thresholds below usage is crossed, we read
4157 * only one element of the array here.
4159 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4160 eventfd_signal(t->entries[i].eventfd, 1);
4162 /* i = current_threshold + 1 */
4166 * Iterate forward over array of thresholds starting from
4167 * current_threshold+1 and check if a threshold is crossed.
4168 * If none of thresholds above usage is crossed, we read
4169 * only one element of the array here.
4171 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4172 eventfd_signal(t->entries[i].eventfd, 1);
4174 /* Update current_threshold */
4175 t->current_threshold = i - 1;
4180 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4183 __mem_cgroup_threshold(memcg, false);
4184 if (do_memsw_account())
4185 __mem_cgroup_threshold(memcg, true);
4187 memcg = parent_mem_cgroup(memcg);
4191 static int compare_thresholds(const void *a, const void *b)
4193 const struct mem_cgroup_threshold *_a = a;
4194 const struct mem_cgroup_threshold *_b = b;
4196 if (_a->threshold > _b->threshold)
4199 if (_a->threshold < _b->threshold)
4205 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4207 struct mem_cgroup_eventfd_list *ev;
4209 spin_lock(&memcg_oom_lock);
4211 list_for_each_entry(ev, &memcg->oom_notify, list)
4212 eventfd_signal(ev->eventfd, 1);
4214 spin_unlock(&memcg_oom_lock);
4218 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4220 struct mem_cgroup *iter;
4222 for_each_mem_cgroup_tree(iter, memcg)
4223 mem_cgroup_oom_notify_cb(iter);
4226 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4227 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4229 struct mem_cgroup_thresholds *thresholds;
4230 struct mem_cgroup_threshold_ary *new;
4231 unsigned long threshold;
4232 unsigned long usage;
4235 ret = page_counter_memparse(args, "-1", &threshold);
4239 mutex_lock(&memcg->thresholds_lock);
4242 thresholds = &memcg->thresholds;
4243 usage = mem_cgroup_usage(memcg, false);
4244 } else if (type == _MEMSWAP) {
4245 thresholds = &memcg->memsw_thresholds;
4246 usage = mem_cgroup_usage(memcg, true);
4250 /* Check if a threshold crossed before adding a new one */
4251 if (thresholds->primary)
4252 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4254 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4256 /* Allocate memory for new array of thresholds */
4257 new = kmalloc(struct_size(new, entries, size), GFP_KERNEL);
4264 /* Copy thresholds (if any) to new array */
4265 if (thresholds->primary)
4266 memcpy(new->entries, thresholds->primary->entries,
4267 flex_array_size(new, entries, size - 1));
4269 /* Add new threshold */
4270 new->entries[size - 1].eventfd = eventfd;
4271 new->entries[size - 1].threshold = threshold;
4273 /* Sort thresholds. Registering of new threshold isn't time-critical */
4274 sort(new->entries, size, sizeof(*new->entries),
4275 compare_thresholds, NULL);
4277 /* Find current threshold */
4278 new->current_threshold = -1;
4279 for (i = 0; i < size; i++) {
4280 if (new->entries[i].threshold <= usage) {
4282 * new->current_threshold will not be used until
4283 * rcu_assign_pointer(), so it's safe to increment
4286 ++new->current_threshold;
4291 /* Free old spare buffer and save old primary buffer as spare */
4292 kfree(thresholds->spare);
4293 thresholds->spare = thresholds->primary;
4295 rcu_assign_pointer(thresholds->primary, new);
4297 /* To be sure that nobody uses thresholds */
4301 mutex_unlock(&memcg->thresholds_lock);
4306 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4307 struct eventfd_ctx *eventfd, const char *args)
4309 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4312 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4313 struct eventfd_ctx *eventfd, const char *args)
4315 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4318 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4319 struct eventfd_ctx *eventfd, enum res_type type)
4321 struct mem_cgroup_thresholds *thresholds;
4322 struct mem_cgroup_threshold_ary *new;
4323 unsigned long usage;
4324 int i, j, size, entries;
4326 mutex_lock(&memcg->thresholds_lock);
4329 thresholds = &memcg->thresholds;
4330 usage = mem_cgroup_usage(memcg, false);
4331 } else if (type == _MEMSWAP) {
4332 thresholds = &memcg->memsw_thresholds;
4333 usage = mem_cgroup_usage(memcg, true);
4337 if (!thresholds->primary)
4340 /* Check if a threshold crossed before removing */
4341 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4343 /* Calculate new number of threshold */
4345 for (i = 0; i < thresholds->primary->size; i++) {
4346 if (thresholds->primary->entries[i].eventfd != eventfd)
4352 new = thresholds->spare;
4354 /* If no items related to eventfd have been cleared, nothing to do */
4358 /* Set thresholds array to NULL if we don't have thresholds */
4367 /* Copy thresholds and find current threshold */
4368 new->current_threshold = -1;
4369 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4370 if (thresholds->primary->entries[i].eventfd == eventfd)
4373 new->entries[j] = thresholds->primary->entries[i];
4374 if (new->entries[j].threshold <= usage) {
4376 * new->current_threshold will not be used
4377 * until rcu_assign_pointer(), so it's safe to increment
4380 ++new->current_threshold;
4386 /* Swap primary and spare array */
4387 thresholds->spare = thresholds->primary;
4389 rcu_assign_pointer(thresholds->primary, new);
4391 /* To be sure that nobody uses thresholds */
4394 /* If all events are unregistered, free the spare array */
4396 kfree(thresholds->spare);
4397 thresholds->spare = NULL;
4400 mutex_unlock(&memcg->thresholds_lock);
4403 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4404 struct eventfd_ctx *eventfd)
4406 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4409 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4410 struct eventfd_ctx *eventfd)
4412 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4415 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4416 struct eventfd_ctx *eventfd, const char *args)
4418 struct mem_cgroup_eventfd_list *event;
4420 event = kmalloc(sizeof(*event), GFP_KERNEL);
4424 spin_lock(&memcg_oom_lock);
4426 event->eventfd = eventfd;
4427 list_add(&event->list, &memcg->oom_notify);
4429 /* already in OOM ? */
4430 if (memcg->under_oom)
4431 eventfd_signal(eventfd, 1);
4432 spin_unlock(&memcg_oom_lock);
4437 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4438 struct eventfd_ctx *eventfd)
4440 struct mem_cgroup_eventfd_list *ev, *tmp;
4442 spin_lock(&memcg_oom_lock);
4444 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4445 if (ev->eventfd == eventfd) {
4446 list_del(&ev->list);
4451 spin_unlock(&memcg_oom_lock);
4454 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4456 struct mem_cgroup *memcg = mem_cgroup_from_seq(sf);
4458 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4459 seq_printf(sf, "under_oom %d\n", (bool)memcg->under_oom);
4460 seq_printf(sf, "oom_kill %lu\n",
4461 atomic_long_read(&memcg->memory_events[MEMCG_OOM_KILL]));
4465 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4466 struct cftype *cft, u64 val)
4468 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4470 /* cannot set to root cgroup and only 0 and 1 are allowed */
4471 if (!css->parent || !((val == 0) || (val == 1)))
4474 memcg->oom_kill_disable = val;
4476 memcg_oom_recover(memcg);
4481 #ifdef CONFIG_CGROUP_WRITEBACK
4483 #include <trace/events/writeback.h>
4485 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4487 return wb_domain_init(&memcg->cgwb_domain, gfp);
4490 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4492 wb_domain_exit(&memcg->cgwb_domain);
4495 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4497 wb_domain_size_changed(&memcg->cgwb_domain);
4500 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4502 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4504 if (!memcg->css.parent)
4507 return &memcg->cgwb_domain;
4511 * idx can be of type enum memcg_stat_item or node_stat_item.
4512 * Keep in sync with memcg_exact_page().
4514 static unsigned long memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
4516 long x = atomic_long_read(&memcg->vmstats[idx]);
4519 for_each_online_cpu(cpu)
4520 x += per_cpu_ptr(memcg->vmstats_percpu, cpu)->stat[idx];
4527 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4528 * @wb: bdi_writeback in question
4529 * @pfilepages: out parameter for number of file pages
4530 * @pheadroom: out parameter for number of allocatable pages according to memcg
4531 * @pdirty: out parameter for number of dirty pages
4532 * @pwriteback: out parameter for number of pages under writeback
4534 * Determine the numbers of file, headroom, dirty, and writeback pages in
4535 * @wb's memcg. File, dirty and writeback are self-explanatory. Headroom
4536 * is a bit more involved.
4538 * A memcg's headroom is "min(max, high) - used". In the hierarchy, the
4539 * headroom is calculated as the lowest headroom of itself and the
4540 * ancestors. Note that this doesn't consider the actual amount of
4541 * available memory in the system. The caller should further cap
4542 * *@pheadroom accordingly.
4544 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pfilepages,
4545 unsigned long *pheadroom, unsigned long *pdirty,
4546 unsigned long *pwriteback)
4548 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4549 struct mem_cgroup *parent;
4551 *pdirty = memcg_exact_page_state(memcg, NR_FILE_DIRTY);
4553 *pwriteback = memcg_exact_page_state(memcg, NR_WRITEBACK);
4554 *pfilepages = memcg_exact_page_state(memcg, NR_INACTIVE_FILE) +
4555 memcg_exact_page_state(memcg, NR_ACTIVE_FILE);
4556 *pheadroom = PAGE_COUNTER_MAX;
4558 while ((parent = parent_mem_cgroup(memcg))) {
4559 unsigned long ceiling = min(READ_ONCE(memcg->memory.max),
4560 READ_ONCE(memcg->memory.high));
4561 unsigned long used = page_counter_read(&memcg->memory);
4563 *pheadroom = min(*pheadroom, ceiling - min(ceiling, used));
4569 * Foreign dirty flushing
4571 * There's an inherent mismatch between memcg and writeback. The former
4572 * trackes ownership per-page while the latter per-inode. This was a
4573 * deliberate design decision because honoring per-page ownership in the
4574 * writeback path is complicated, may lead to higher CPU and IO overheads
4575 * and deemed unnecessary given that write-sharing an inode across
4576 * different cgroups isn't a common use-case.
4578 * Combined with inode majority-writer ownership switching, this works well
4579 * enough in most cases but there are some pathological cases. For
4580 * example, let's say there are two cgroups A and B which keep writing to
4581 * different but confined parts of the same inode. B owns the inode and
4582 * A's memory is limited far below B's. A's dirty ratio can rise enough to
4583 * trigger balance_dirty_pages() sleeps but B's can be low enough to avoid
4584 * triggering background writeback. A will be slowed down without a way to
4585 * make writeback of the dirty pages happen.
4587 * Conditions like the above can lead to a cgroup getting repatedly and
4588 * severely throttled after making some progress after each
4589 * dirty_expire_interval while the underyling IO device is almost
4592 * Solving this problem completely requires matching the ownership tracking
4593 * granularities between memcg and writeback in either direction. However,
4594 * the more egregious behaviors can be avoided by simply remembering the
4595 * most recent foreign dirtying events and initiating remote flushes on
4596 * them when local writeback isn't enough to keep the memory clean enough.
4598 * The following two functions implement such mechanism. When a foreign
4599 * page - a page whose memcg and writeback ownerships don't match - is
4600 * dirtied, mem_cgroup_track_foreign_dirty() records the inode owning
4601 * bdi_writeback on the page owning memcg. When balance_dirty_pages()
4602 * decides that the memcg needs to sleep due to high dirty ratio, it calls
4603 * mem_cgroup_flush_foreign() which queues writeback on the recorded
4604 * foreign bdi_writebacks which haven't expired. Both the numbers of
4605 * recorded bdi_writebacks and concurrent in-flight foreign writebacks are
4606 * limited to MEMCG_CGWB_FRN_CNT.
4608 * The mechanism only remembers IDs and doesn't hold any object references.
4609 * As being wrong occasionally doesn't matter, updates and accesses to the
4610 * records are lockless and racy.
4612 void mem_cgroup_track_foreign_dirty_slowpath(struct page *page,
4613 struct bdi_writeback *wb)
4615 struct mem_cgroup *memcg = page->mem_cgroup;
4616 struct memcg_cgwb_frn *frn;
4617 u64 now = get_jiffies_64();
4618 u64 oldest_at = now;
4622 trace_track_foreign_dirty(page, wb);
4625 * Pick the slot to use. If there is already a slot for @wb, keep
4626 * using it. If not replace the oldest one which isn't being
4629 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4630 frn = &memcg->cgwb_frn[i];
4631 if (frn->bdi_id == wb->bdi->id &&
4632 frn->memcg_id == wb->memcg_css->id)
4634 if (time_before64(frn->at, oldest_at) &&
4635 atomic_read(&frn->done.cnt) == 1) {
4637 oldest_at = frn->at;
4641 if (i < MEMCG_CGWB_FRN_CNT) {
4643 * Re-using an existing one. Update timestamp lazily to
4644 * avoid making the cacheline hot. We want them to be
4645 * reasonably up-to-date and significantly shorter than
4646 * dirty_expire_interval as that's what expires the record.
4647 * Use the shorter of 1s and dirty_expire_interval / 8.
4649 unsigned long update_intv =
4650 min_t(unsigned long, HZ,
4651 msecs_to_jiffies(dirty_expire_interval * 10) / 8);
4653 if (time_before64(frn->at, now - update_intv))
4655 } else if (oldest >= 0) {
4656 /* replace the oldest free one */
4657 frn = &memcg->cgwb_frn[oldest];
4658 frn->bdi_id = wb->bdi->id;
4659 frn->memcg_id = wb->memcg_css->id;
4664 /* issue foreign writeback flushes for recorded foreign dirtying events */
4665 void mem_cgroup_flush_foreign(struct bdi_writeback *wb)
4667 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4668 unsigned long intv = msecs_to_jiffies(dirty_expire_interval * 10);
4669 u64 now = jiffies_64;
4672 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++) {
4673 struct memcg_cgwb_frn *frn = &memcg->cgwb_frn[i];
4676 * If the record is older than dirty_expire_interval,
4677 * writeback on it has already started. No need to kick it
4678 * off again. Also, don't start a new one if there's
4679 * already one in flight.
4681 if (time_after64(frn->at, now - intv) &&
4682 atomic_read(&frn->done.cnt) == 1) {
4684 trace_flush_foreign(wb, frn->bdi_id, frn->memcg_id);
4685 cgroup_writeback_by_id(frn->bdi_id, frn->memcg_id, 0,
4686 WB_REASON_FOREIGN_FLUSH,
4692 #else /* CONFIG_CGROUP_WRITEBACK */
4694 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4699 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4703 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4707 #endif /* CONFIG_CGROUP_WRITEBACK */
4710 * DO NOT USE IN NEW FILES.
4712 * "cgroup.event_control" implementation.
4714 * This is way over-engineered. It tries to support fully configurable
4715 * events for each user. Such level of flexibility is completely
4716 * unnecessary especially in the light of the planned unified hierarchy.
4718 * Please deprecate this and replace with something simpler if at all
4723 * Unregister event and free resources.
4725 * Gets called from workqueue.
4727 static void memcg_event_remove(struct work_struct *work)
4729 struct mem_cgroup_event *event =
4730 container_of(work, struct mem_cgroup_event, remove);
4731 struct mem_cgroup *memcg = event->memcg;
4733 remove_wait_queue(event->wqh, &event->wait);
4735 event->unregister_event(memcg, event->eventfd);
4737 /* Notify userspace the event is going away. */
4738 eventfd_signal(event->eventfd, 1);
4740 eventfd_ctx_put(event->eventfd);
4742 css_put(&memcg->css);
4746 * Gets called on EPOLLHUP on eventfd when user closes it.
4748 * Called with wqh->lock held and interrupts disabled.
4750 static int memcg_event_wake(wait_queue_entry_t *wait, unsigned mode,
4751 int sync, void *key)
4753 struct mem_cgroup_event *event =
4754 container_of(wait, struct mem_cgroup_event, wait);
4755 struct mem_cgroup *memcg = event->memcg;
4756 __poll_t flags = key_to_poll(key);
4758 if (flags & EPOLLHUP) {
4760 * If the event has been detached at cgroup removal, we
4761 * can simply return knowing the other side will cleanup
4764 * We can't race against event freeing since the other
4765 * side will require wqh->lock via remove_wait_queue(),
4768 spin_lock(&memcg->event_list_lock);
4769 if (!list_empty(&event->list)) {
4770 list_del_init(&event->list);
4772 * We are in atomic context, but cgroup_event_remove()
4773 * may sleep, so we have to call it in workqueue.
4775 schedule_work(&event->remove);
4777 spin_unlock(&memcg->event_list_lock);
4783 static void memcg_event_ptable_queue_proc(struct file *file,
4784 wait_queue_head_t *wqh, poll_table *pt)
4786 struct mem_cgroup_event *event =
4787 container_of(pt, struct mem_cgroup_event, pt);
4790 add_wait_queue(wqh, &event->wait);
4794 * DO NOT USE IN NEW FILES.
4796 * Parse input and register new cgroup event handler.
4798 * Input must be in format '<event_fd> <control_fd> <args>'.
4799 * Interpretation of args is defined by control file implementation.
4801 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4802 char *buf, size_t nbytes, loff_t off)
4804 struct cgroup_subsys_state *css = of_css(of);
4805 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4806 struct mem_cgroup_event *event;
4807 struct cgroup_subsys_state *cfile_css;
4808 unsigned int efd, cfd;
4815 buf = strstrip(buf);
4817 efd = simple_strtoul(buf, &endp, 10);
4822 cfd = simple_strtoul(buf, &endp, 10);
4823 if ((*endp != ' ') && (*endp != '\0'))
4827 event = kzalloc(sizeof(*event), GFP_KERNEL);
4831 event->memcg = memcg;
4832 INIT_LIST_HEAD(&event->list);
4833 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4834 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4835 INIT_WORK(&event->remove, memcg_event_remove);
4843 event->eventfd = eventfd_ctx_fileget(efile.file);
4844 if (IS_ERR(event->eventfd)) {
4845 ret = PTR_ERR(event->eventfd);
4852 goto out_put_eventfd;
4855 /* the process need read permission on control file */
4856 /* AV: shouldn't we check that it's been opened for read instead? */
4857 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4862 * Determine the event callbacks and set them in @event. This used
4863 * to be done via struct cftype but cgroup core no longer knows
4864 * about these events. The following is crude but the whole thing
4865 * is for compatibility anyway.
4867 * DO NOT ADD NEW FILES.
4869 name = cfile.file->f_path.dentry->d_name.name;
4871 if (!strcmp(name, "memory.usage_in_bytes")) {
4872 event->register_event = mem_cgroup_usage_register_event;
4873 event->unregister_event = mem_cgroup_usage_unregister_event;
4874 } else if (!strcmp(name, "memory.oom_control")) {
4875 event->register_event = mem_cgroup_oom_register_event;
4876 event->unregister_event = mem_cgroup_oom_unregister_event;
4877 } else if (!strcmp(name, "memory.pressure_level")) {
4878 event->register_event = vmpressure_register_event;
4879 event->unregister_event = vmpressure_unregister_event;
4880 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4881 event->register_event = memsw_cgroup_usage_register_event;
4882 event->unregister_event = memsw_cgroup_usage_unregister_event;
4889 * Verify @cfile should belong to @css. Also, remaining events are
4890 * automatically removed on cgroup destruction but the removal is
4891 * asynchronous, so take an extra ref on @css.
4893 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4894 &memory_cgrp_subsys);
4896 if (IS_ERR(cfile_css))
4898 if (cfile_css != css) {
4903 ret = event->register_event(memcg, event->eventfd, buf);
4907 vfs_poll(efile.file, &event->pt);
4909 spin_lock(&memcg->event_list_lock);
4910 list_add(&event->list, &memcg->event_list);
4911 spin_unlock(&memcg->event_list_lock);
4923 eventfd_ctx_put(event->eventfd);
4932 static struct cftype mem_cgroup_legacy_files[] = {
4934 .name = "usage_in_bytes",
4935 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4936 .read_u64 = mem_cgroup_read_u64,
4939 .name = "max_usage_in_bytes",
4940 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4941 .write = mem_cgroup_reset,
4942 .read_u64 = mem_cgroup_read_u64,
4945 .name = "limit_in_bytes",
4946 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4947 .write = mem_cgroup_write,
4948 .read_u64 = mem_cgroup_read_u64,
4951 .name = "soft_limit_in_bytes",
4952 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4953 .write = mem_cgroup_write,
4954 .read_u64 = mem_cgroup_read_u64,
4958 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4959 .write = mem_cgroup_reset,
4960 .read_u64 = mem_cgroup_read_u64,
4964 .seq_show = memcg_stat_show,
4967 .name = "force_empty",
4968 .write = mem_cgroup_force_empty_write,
4971 .name = "use_hierarchy",
4972 .write_u64 = mem_cgroup_hierarchy_write,
4973 .read_u64 = mem_cgroup_hierarchy_read,
4976 .name = "cgroup.event_control", /* XXX: for compat */
4977 .write = memcg_write_event_control,
4978 .flags = CFTYPE_NO_PREFIX | CFTYPE_WORLD_WRITABLE,
4981 .name = "swappiness",
4982 .read_u64 = mem_cgroup_swappiness_read,
4983 .write_u64 = mem_cgroup_swappiness_write,
4986 .name = "move_charge_at_immigrate",
4987 .read_u64 = mem_cgroup_move_charge_read,
4988 .write_u64 = mem_cgroup_move_charge_write,
4991 .name = "oom_control",
4992 .seq_show = mem_cgroup_oom_control_read,
4993 .write_u64 = mem_cgroup_oom_control_write,
4994 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4997 .name = "pressure_level",
5001 .name = "numa_stat",
5002 .seq_show = memcg_numa_stat_show,
5006 .name = "kmem.limit_in_bytes",
5007 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5008 .write = mem_cgroup_write,
5009 .read_u64 = mem_cgroup_read_u64,
5012 .name = "kmem.usage_in_bytes",
5013 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5014 .read_u64 = mem_cgroup_read_u64,
5017 .name = "kmem.failcnt",
5018 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5019 .write = mem_cgroup_reset,
5020 .read_u64 = mem_cgroup_read_u64,
5023 .name = "kmem.max_usage_in_bytes",
5024 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5025 .write = mem_cgroup_reset,
5026 .read_u64 = mem_cgroup_read_u64,
5028 #if defined(CONFIG_MEMCG_KMEM) && \
5029 (defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG))
5031 .name = "kmem.slabinfo",
5032 .seq_show = memcg_slab_show,
5036 .name = "kmem.tcp.limit_in_bytes",
5037 .private = MEMFILE_PRIVATE(_TCP, RES_LIMIT),
5038 .write = mem_cgroup_write,
5039 .read_u64 = mem_cgroup_read_u64,
5042 .name = "kmem.tcp.usage_in_bytes",
5043 .private = MEMFILE_PRIVATE(_TCP, RES_USAGE),
5044 .read_u64 = mem_cgroup_read_u64,
5047 .name = "kmem.tcp.failcnt",
5048 .private = MEMFILE_PRIVATE(_TCP, RES_FAILCNT),
5049 .write = mem_cgroup_reset,
5050 .read_u64 = mem_cgroup_read_u64,
5053 .name = "kmem.tcp.max_usage_in_bytes",
5054 .private = MEMFILE_PRIVATE(_TCP, RES_MAX_USAGE),
5055 .write = mem_cgroup_reset,
5056 .read_u64 = mem_cgroup_read_u64,
5058 { }, /* terminate */
5062 * Private memory cgroup IDR
5064 * Swap-out records and page cache shadow entries need to store memcg
5065 * references in constrained space, so we maintain an ID space that is
5066 * limited to 16 bit (MEM_CGROUP_ID_MAX), limiting the total number of
5067 * memory-controlled cgroups to 64k.
5069 * However, there usually are many references to the offline CSS after
5070 * the cgroup has been destroyed, such as page cache or reclaimable
5071 * slab objects, that don't need to hang on to the ID. We want to keep
5072 * those dead CSS from occupying IDs, or we might quickly exhaust the
5073 * relatively small ID space and prevent the creation of new cgroups
5074 * even when there are much fewer than 64k cgroups - possibly none.
5076 * Maintain a private 16-bit ID space for memcg, and allow the ID to
5077 * be freed and recycled when it's no longer needed, which is usually
5078 * when the CSS is offlined.
5080 * The only exception to that are records of swapped out tmpfs/shmem
5081 * pages that need to be attributed to live ancestors on swapin. But
5082 * those references are manageable from userspace.
5085 static DEFINE_IDR(mem_cgroup_idr);
5087 static void mem_cgroup_id_remove(struct mem_cgroup *memcg)
5089 if (memcg->id.id > 0) {
5090 idr_remove(&mem_cgroup_idr, memcg->id.id);
5095 static void __maybe_unused mem_cgroup_id_get_many(struct mem_cgroup *memcg,
5098 refcount_add(n, &memcg->id.ref);
5101 static void mem_cgroup_id_put_many(struct mem_cgroup *memcg, unsigned int n)
5103 if (refcount_sub_and_test(n, &memcg->id.ref)) {
5104 mem_cgroup_id_remove(memcg);
5106 /* Memcg ID pins CSS */
5107 css_put(&memcg->css);
5111 static inline void mem_cgroup_id_put(struct mem_cgroup *memcg)
5113 mem_cgroup_id_put_many(memcg, 1);
5117 * mem_cgroup_from_id - look up a memcg from a memcg id
5118 * @id: the memcg id to look up
5120 * Caller must hold rcu_read_lock().
5122 struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
5124 WARN_ON_ONCE(!rcu_read_lock_held());
5125 return idr_find(&mem_cgroup_idr, id);
5128 static int alloc_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5130 struct mem_cgroup_per_node *pn;
5133 * This routine is called against possible nodes.
5134 * But it's BUG to call kmalloc() against offline node.
5136 * TODO: this routine can waste much memory for nodes which will
5137 * never be onlined. It's better to use memory hotplug callback
5140 if (!node_state(node, N_NORMAL_MEMORY))
5142 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5146 pn->lruvec_stat_local = alloc_percpu_gfp(struct lruvec_stat,
5147 GFP_KERNEL_ACCOUNT);
5148 if (!pn->lruvec_stat_local) {
5153 pn->lruvec_stat_cpu = alloc_percpu_gfp(struct lruvec_stat,
5154 GFP_KERNEL_ACCOUNT);
5155 if (!pn->lruvec_stat_cpu) {
5156 free_percpu(pn->lruvec_stat_local);
5161 lruvec_init(&pn->lruvec);
5162 pn->usage_in_excess = 0;
5163 pn->on_tree = false;
5166 memcg->nodeinfo[node] = pn;
5170 static void free_mem_cgroup_per_node_info(struct mem_cgroup *memcg, int node)
5172 struct mem_cgroup_per_node *pn = memcg->nodeinfo[node];
5177 free_percpu(pn->lruvec_stat_cpu);
5178 free_percpu(pn->lruvec_stat_local);
5182 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5187 free_mem_cgroup_per_node_info(memcg, node);
5188 free_percpu(memcg->vmstats_percpu);
5189 free_percpu(memcg->vmstats_local);
5193 static void mem_cgroup_free(struct mem_cgroup *memcg)
5195 memcg_wb_domain_exit(memcg);
5197 * Flush percpu vmstats and vmevents to guarantee the value correctness
5198 * on parent's and all ancestor levels.
5200 memcg_flush_percpu_vmstats(memcg);
5201 memcg_flush_percpu_vmevents(memcg);
5202 __mem_cgroup_free(memcg);
5205 static struct mem_cgroup *mem_cgroup_alloc(void)
5207 struct mem_cgroup *memcg;
5210 int __maybe_unused i;
5211 long error = -ENOMEM;
5213 size = sizeof(struct mem_cgroup);
5214 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5216 memcg = kzalloc(size, GFP_KERNEL);
5218 return ERR_PTR(error);
5220 memcg->id.id = idr_alloc(&mem_cgroup_idr, NULL,
5221 1, MEM_CGROUP_ID_MAX,
5223 if (memcg->id.id < 0) {
5224 error = memcg->id.id;
5228 memcg->vmstats_local = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5229 GFP_KERNEL_ACCOUNT);
5230 if (!memcg->vmstats_local)
5233 memcg->vmstats_percpu = alloc_percpu_gfp(struct memcg_vmstats_percpu,
5234 GFP_KERNEL_ACCOUNT);
5235 if (!memcg->vmstats_percpu)
5239 if (alloc_mem_cgroup_per_node_info(memcg, node))
5242 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
5245 INIT_WORK(&memcg->high_work, high_work_func);
5246 INIT_LIST_HEAD(&memcg->oom_notify);
5247 mutex_init(&memcg->thresholds_lock);
5248 spin_lock_init(&memcg->move_lock);
5249 vmpressure_init(&memcg->vmpressure);
5250 INIT_LIST_HEAD(&memcg->event_list);
5251 spin_lock_init(&memcg->event_list_lock);
5252 memcg->socket_pressure = jiffies;
5253 #ifdef CONFIG_MEMCG_KMEM
5254 memcg->kmemcg_id = -1;
5255 INIT_LIST_HEAD(&memcg->objcg_list);
5257 #ifdef CONFIG_CGROUP_WRITEBACK
5258 INIT_LIST_HEAD(&memcg->cgwb_list);
5259 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5260 memcg->cgwb_frn[i].done =
5261 __WB_COMPLETION_INIT(&memcg_cgwb_frn_waitq);
5263 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5264 spin_lock_init(&memcg->deferred_split_queue.split_queue_lock);
5265 INIT_LIST_HEAD(&memcg->deferred_split_queue.split_queue);
5266 memcg->deferred_split_queue.split_queue_len = 0;
5268 idr_replace(&mem_cgroup_idr, memcg, memcg->id.id);
5271 mem_cgroup_id_remove(memcg);
5272 __mem_cgroup_free(memcg);
5273 return ERR_PTR(error);
5276 static struct cgroup_subsys_state * __ref
5277 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5279 struct mem_cgroup *parent = mem_cgroup_from_css(parent_css);
5280 struct mem_cgroup *memcg, *old_memcg;
5281 long error = -ENOMEM;
5283 old_memcg = set_active_memcg(parent);
5284 memcg = mem_cgroup_alloc();
5285 set_active_memcg(old_memcg);
5287 return ERR_CAST(memcg);
5289 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5290 memcg->soft_limit = PAGE_COUNTER_MAX;
5291 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5293 memcg->swappiness = mem_cgroup_swappiness(parent);
5294 memcg->oom_kill_disable = parent->oom_kill_disable;
5296 page_counter_init(&memcg->memory, &parent->memory);
5297 page_counter_init(&memcg->swap, &parent->swap);
5298 page_counter_init(&memcg->kmem, &parent->kmem);
5299 page_counter_init(&memcg->tcpmem, &parent->tcpmem);
5301 page_counter_init(&memcg->memory, NULL);
5302 page_counter_init(&memcg->swap, NULL);
5303 page_counter_init(&memcg->kmem, NULL);
5304 page_counter_init(&memcg->tcpmem, NULL);
5306 root_mem_cgroup = memcg;
5310 /* The following stuff does not apply to the root */
5311 error = memcg_online_kmem(memcg);
5315 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5316 static_branch_inc(&memcg_sockets_enabled_key);
5320 mem_cgroup_id_remove(memcg);
5321 mem_cgroup_free(memcg);
5322 return ERR_PTR(error);
5325 static int mem_cgroup_css_online(struct cgroup_subsys_state *css)
5327 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5330 * A memcg must be visible for memcg_expand_shrinker_maps()
5331 * by the time the maps are allocated. So, we allocate maps
5332 * here, when for_each_mem_cgroup() can't skip it.
5334 if (memcg_alloc_shrinker_maps(memcg)) {
5335 mem_cgroup_id_remove(memcg);
5339 /* Online state pins memcg ID, memcg ID pins CSS */
5340 refcount_set(&memcg->id.ref, 1);
5345 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5347 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5348 struct mem_cgroup_event *event, *tmp;
5351 * Unregister events and notify userspace.
5352 * Notify userspace about cgroup removing only after rmdir of cgroup
5353 * directory to avoid race between userspace and kernelspace.
5355 spin_lock(&memcg->event_list_lock);
5356 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5357 list_del_init(&event->list);
5358 schedule_work(&event->remove);
5360 spin_unlock(&memcg->event_list_lock);
5362 page_counter_set_min(&memcg->memory, 0);
5363 page_counter_set_low(&memcg->memory, 0);
5365 memcg_offline_kmem(memcg);
5366 wb_memcg_offline(memcg);
5368 drain_all_stock(memcg);
5370 mem_cgroup_id_put(memcg);
5373 static void mem_cgroup_css_released(struct cgroup_subsys_state *css)
5375 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5377 invalidate_reclaim_iterators(memcg);
5380 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5383 int __maybe_unused i;
5385 #ifdef CONFIG_CGROUP_WRITEBACK
5386 for (i = 0; i < MEMCG_CGWB_FRN_CNT; i++)
5387 wb_wait_for_completion(&memcg->cgwb_frn[i].done);
5389 if (cgroup_subsys_on_dfl(memory_cgrp_subsys) && !cgroup_memory_nosocket)
5390 static_branch_dec(&memcg_sockets_enabled_key);
5392 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && memcg->tcpmem_active)
5393 static_branch_dec(&memcg_sockets_enabled_key);
5395 vmpressure_cleanup(&memcg->vmpressure);
5396 cancel_work_sync(&memcg->high_work);
5397 mem_cgroup_remove_from_trees(memcg);
5398 memcg_free_shrinker_maps(memcg);
5399 memcg_free_kmem(memcg);
5400 mem_cgroup_free(memcg);
5404 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5405 * @css: the target css
5407 * Reset the states of the mem_cgroup associated with @css. This is
5408 * invoked when the userland requests disabling on the default hierarchy
5409 * but the memcg is pinned through dependency. The memcg should stop
5410 * applying policies and should revert to the vanilla state as it may be
5411 * made visible again.
5413 * The current implementation only resets the essential configurations.
5414 * This needs to be expanded to cover all the visible parts.
5416 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5418 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5420 page_counter_set_max(&memcg->memory, PAGE_COUNTER_MAX);
5421 page_counter_set_max(&memcg->swap, PAGE_COUNTER_MAX);
5422 page_counter_set_max(&memcg->kmem, PAGE_COUNTER_MAX);
5423 page_counter_set_max(&memcg->tcpmem, PAGE_COUNTER_MAX);
5424 page_counter_set_min(&memcg->memory, 0);
5425 page_counter_set_low(&memcg->memory, 0);
5426 page_counter_set_high(&memcg->memory, PAGE_COUNTER_MAX);
5427 memcg->soft_limit = PAGE_COUNTER_MAX;
5428 page_counter_set_high(&memcg->swap, PAGE_COUNTER_MAX);
5429 memcg_wb_domain_size_changed(memcg);
5433 /* Handlers for move charge at task migration. */
5434 static int mem_cgroup_do_precharge(unsigned long count)
5438 /* Try a single bulk charge without reclaim first, kswapd may wake */
5439 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_DIRECT_RECLAIM, count);
5441 mc.precharge += count;
5445 /* Try charges one by one with reclaim, but do not retry */
5447 ret = try_charge(mc.to, GFP_KERNEL | __GFP_NORETRY, 1);
5461 enum mc_target_type {
5468 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5469 unsigned long addr, pte_t ptent)
5471 struct page *page = vm_normal_page(vma, addr, ptent);
5473 if (!page || !page_mapped(page))
5475 if (PageAnon(page)) {
5476 if (!(mc.flags & MOVE_ANON))
5479 if (!(mc.flags & MOVE_FILE))
5482 if (!get_page_unless_zero(page))
5488 #if defined(CONFIG_SWAP) || defined(CONFIG_DEVICE_PRIVATE)
5489 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5490 pte_t ptent, swp_entry_t *entry)
5492 struct page *page = NULL;
5493 swp_entry_t ent = pte_to_swp_entry(ptent);
5495 if (!(mc.flags & MOVE_ANON))
5499 * Handle MEMORY_DEVICE_PRIVATE which are ZONE_DEVICE page belonging to
5500 * a device and because they are not accessible by CPU they are store
5501 * as special swap entry in the CPU page table.
5503 if (is_device_private_entry(ent)) {
5504 page = device_private_entry_to_page(ent);
5506 * MEMORY_DEVICE_PRIVATE means ZONE_DEVICE page and which have
5507 * a refcount of 1 when free (unlike normal page)
5509 if (!page_ref_add_unless(page, 1, 1))
5514 if (non_swap_entry(ent))
5518 * Because lookup_swap_cache() updates some statistics counter,
5519 * we call find_get_page() with swapper_space directly.
5521 page = find_get_page(swap_address_space(ent), swp_offset(ent));
5522 entry->val = ent.val;
5527 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5528 pte_t ptent, swp_entry_t *entry)
5534 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5535 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5537 if (!vma->vm_file) /* anonymous vma */
5539 if (!(mc.flags & MOVE_FILE))
5542 /* page is moved even if it's not RSS of this task(page-faulted). */
5543 /* shmem/tmpfs may report page out on swap: account for that too. */
5544 return find_get_incore_page(vma->vm_file->f_mapping,
5545 linear_page_index(vma, addr));
5549 * mem_cgroup_move_account - move account of the page
5551 * @compound: charge the page as compound or small page
5552 * @from: mem_cgroup which the page is moved from.
5553 * @to: mem_cgroup which the page is moved to. @from != @to.
5555 * The caller must make sure the page is not on LRU (isolate_page() is useful.)
5557 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
5560 static int mem_cgroup_move_account(struct page *page,
5562 struct mem_cgroup *from,
5563 struct mem_cgroup *to)
5565 struct lruvec *from_vec, *to_vec;
5566 struct pglist_data *pgdat;
5567 unsigned int nr_pages = compound ? thp_nr_pages(page) : 1;
5570 VM_BUG_ON(from == to);
5571 VM_BUG_ON_PAGE(PageLRU(page), page);
5572 VM_BUG_ON(compound && !PageTransHuge(page));
5575 * Prevent mem_cgroup_migrate() from looking at
5576 * page->mem_cgroup of its source page while we change it.
5579 if (!trylock_page(page))
5583 if (page->mem_cgroup != from)
5586 pgdat = page_pgdat(page);
5587 from_vec = mem_cgroup_lruvec(from, pgdat);
5588 to_vec = mem_cgroup_lruvec(to, pgdat);
5590 lock_page_memcg(page);
5592 if (PageAnon(page)) {
5593 if (page_mapped(page)) {
5594 __mod_lruvec_state(from_vec, NR_ANON_MAPPED, -nr_pages);
5595 __mod_lruvec_state(to_vec, NR_ANON_MAPPED, nr_pages);
5596 if (PageTransHuge(page)) {
5597 __mod_lruvec_state(from_vec, NR_ANON_THPS,
5599 __mod_lruvec_state(to_vec, NR_ANON_THPS,
5605 __mod_lruvec_state(from_vec, NR_FILE_PAGES, -nr_pages);
5606 __mod_lruvec_state(to_vec, NR_FILE_PAGES, nr_pages);
5608 if (PageSwapBacked(page)) {
5609 __mod_lruvec_state(from_vec, NR_SHMEM, -nr_pages);
5610 __mod_lruvec_state(to_vec, NR_SHMEM, nr_pages);
5613 if (page_mapped(page)) {
5614 __mod_lruvec_state(from_vec, NR_FILE_MAPPED, -nr_pages);
5615 __mod_lruvec_state(to_vec, NR_FILE_MAPPED, nr_pages);
5618 if (PageDirty(page)) {
5619 struct address_space *mapping = page_mapping(page);
5621 if (mapping_can_writeback(mapping)) {
5622 __mod_lruvec_state(from_vec, NR_FILE_DIRTY,
5624 __mod_lruvec_state(to_vec, NR_FILE_DIRTY,
5630 if (PageWriteback(page)) {
5631 __mod_lruvec_state(from_vec, NR_WRITEBACK, -nr_pages);
5632 __mod_lruvec_state(to_vec, NR_WRITEBACK, nr_pages);
5636 * All state has been migrated, let's switch to the new memcg.
5638 * It is safe to change page->mem_cgroup here because the page
5639 * is referenced, charged, isolated, and locked: we can't race
5640 * with (un)charging, migration, LRU putback, or anything else
5641 * that would rely on a stable page->mem_cgroup.
5643 * Note that lock_page_memcg is a memcg lock, not a page lock,
5644 * to save space. As soon as we switch page->mem_cgroup to a
5645 * new memcg that isn't locked, the above state can change
5646 * concurrently again. Make sure we're truly done with it.
5651 css_put(&from->css);
5653 page->mem_cgroup = to;
5655 __unlock_page_memcg(from);
5659 local_irq_disable();
5660 mem_cgroup_charge_statistics(to, page, nr_pages);
5661 memcg_check_events(to, page);
5662 mem_cgroup_charge_statistics(from, page, -nr_pages);
5663 memcg_check_events(from, page);
5672 * get_mctgt_type - get target type of moving charge
5673 * @vma: the vma the pte to be checked belongs
5674 * @addr: the address corresponding to the pte to be checked
5675 * @ptent: the pte to be checked
5676 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5679 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5680 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5681 * move charge. if @target is not NULL, the page is stored in target->page
5682 * with extra refcnt got(Callers should handle it).
5683 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5684 * target for charge migration. if @target is not NULL, the entry is stored
5686 * 3(MC_TARGET_DEVICE): like MC_TARGET_PAGE but page is MEMORY_DEVICE_PRIVATE
5687 * (so ZONE_DEVICE page and thus not on the lru).
5688 * For now we such page is charge like a regular page would be as for all
5689 * intent and purposes it is just special memory taking the place of a
5692 * See Documentations/vm/hmm.txt and include/linux/hmm.h
5694 * Called with pte lock held.
5697 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5698 unsigned long addr, pte_t ptent, union mc_target *target)
5700 struct page *page = NULL;
5701 enum mc_target_type ret = MC_TARGET_NONE;
5702 swp_entry_t ent = { .val = 0 };
5704 if (pte_present(ptent))
5705 page = mc_handle_present_pte(vma, addr, ptent);
5706 else if (is_swap_pte(ptent))
5707 page = mc_handle_swap_pte(vma, ptent, &ent);
5708 else if (pte_none(ptent))
5709 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5711 if (!page && !ent.val)
5715 * Do only loose check w/o serialization.
5716 * mem_cgroup_move_account() checks the page is valid or
5717 * not under LRU exclusion.
5719 if (page->mem_cgroup == mc.from) {
5720 ret = MC_TARGET_PAGE;
5721 if (is_device_private_page(page))
5722 ret = MC_TARGET_DEVICE;
5724 target->page = page;
5726 if (!ret || !target)
5730 * There is a swap entry and a page doesn't exist or isn't charged.
5731 * But we cannot move a tail-page in a THP.
5733 if (ent.val && !ret && (!page || !PageTransCompound(page)) &&
5734 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5735 ret = MC_TARGET_SWAP;
5742 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5744 * We don't consider PMD mapped swapping or file mapped pages because THP does
5745 * not support them for now.
5746 * Caller should make sure that pmd_trans_huge(pmd) is true.
5748 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5749 unsigned long addr, pmd_t pmd, union mc_target *target)
5751 struct page *page = NULL;
5752 enum mc_target_type ret = MC_TARGET_NONE;
5754 if (unlikely(is_swap_pmd(pmd))) {
5755 VM_BUG_ON(thp_migration_supported() &&
5756 !is_pmd_migration_entry(pmd));
5759 page = pmd_page(pmd);
5760 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5761 if (!(mc.flags & MOVE_ANON))
5763 if (page->mem_cgroup == mc.from) {
5764 ret = MC_TARGET_PAGE;
5767 target->page = page;
5773 static inline 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 return MC_TARGET_NONE;
5780 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5781 unsigned long addr, unsigned long end,
5782 struct mm_walk *walk)
5784 struct vm_area_struct *vma = walk->vma;
5788 ptl = pmd_trans_huge_lock(pmd, vma);
5791 * Note their can not be MC_TARGET_DEVICE for now as we do not
5792 * support transparent huge page with MEMORY_DEVICE_PRIVATE but
5793 * this might change.
5795 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5796 mc.precharge += HPAGE_PMD_NR;
5801 if (pmd_trans_unstable(pmd))
5803 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5804 for (; addr != end; pte++, addr += PAGE_SIZE)
5805 if (get_mctgt_type(vma, addr, *pte, NULL))
5806 mc.precharge++; /* increment precharge temporarily */
5807 pte_unmap_unlock(pte - 1, ptl);
5813 static const struct mm_walk_ops precharge_walk_ops = {
5814 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5817 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5819 unsigned long precharge;
5822 walk_page_range(mm, 0, mm->highest_vm_end, &precharge_walk_ops, NULL);
5823 mmap_read_unlock(mm);
5825 precharge = mc.precharge;
5831 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5833 unsigned long precharge = mem_cgroup_count_precharge(mm);
5835 VM_BUG_ON(mc.moving_task);
5836 mc.moving_task = current;
5837 return mem_cgroup_do_precharge(precharge);
5840 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5841 static void __mem_cgroup_clear_mc(void)
5843 struct mem_cgroup *from = mc.from;
5844 struct mem_cgroup *to = mc.to;
5846 /* we must uncharge all the leftover precharges from mc.to */
5848 cancel_charge(mc.to, mc.precharge);
5852 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5853 * we must uncharge here.
5855 if (mc.moved_charge) {
5856 cancel_charge(mc.from, mc.moved_charge);
5857 mc.moved_charge = 0;
5859 /* we must fixup refcnts and charges */
5860 if (mc.moved_swap) {
5861 /* uncharge swap account from the old cgroup */
5862 if (!mem_cgroup_is_root(mc.from))
5863 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5865 mem_cgroup_id_put_many(mc.from, mc.moved_swap);
5868 * we charged both to->memory and to->memsw, so we
5869 * should uncharge to->memory.
5871 if (!mem_cgroup_is_root(mc.to))
5872 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5876 memcg_oom_recover(from);
5877 memcg_oom_recover(to);
5878 wake_up_all(&mc.waitq);
5881 static void mem_cgroup_clear_mc(void)
5883 struct mm_struct *mm = mc.mm;
5886 * we must clear moving_task before waking up waiters at the end of
5889 mc.moving_task = NULL;
5890 __mem_cgroup_clear_mc();
5891 spin_lock(&mc.lock);
5895 spin_unlock(&mc.lock);
5900 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
5902 struct cgroup_subsys_state *css;
5903 struct mem_cgroup *memcg = NULL; /* unneeded init to make gcc happy */
5904 struct mem_cgroup *from;
5905 struct task_struct *leader, *p;
5906 struct mm_struct *mm;
5907 unsigned long move_flags;
5910 /* charge immigration isn't supported on the default hierarchy */
5911 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
5915 * Multi-process migrations only happen on the default hierarchy
5916 * where charge immigration is not used. Perform charge
5917 * immigration if @tset contains a leader and whine if there are
5921 cgroup_taskset_for_each_leader(leader, css, tset) {
5924 memcg = mem_cgroup_from_css(css);
5930 * We are now commited to this value whatever it is. Changes in this
5931 * tunable will only affect upcoming migrations, not the current one.
5932 * So we need to save it, and keep it going.
5934 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5938 from = mem_cgroup_from_task(p);
5940 VM_BUG_ON(from == memcg);
5942 mm = get_task_mm(p);
5945 /* We move charges only when we move a owner of the mm */
5946 if (mm->owner == p) {
5949 VM_BUG_ON(mc.precharge);
5950 VM_BUG_ON(mc.moved_charge);
5951 VM_BUG_ON(mc.moved_swap);
5953 spin_lock(&mc.lock);
5957 mc.flags = move_flags;
5958 spin_unlock(&mc.lock);
5959 /* We set mc.moving_task later */
5961 ret = mem_cgroup_precharge_mc(mm);
5963 mem_cgroup_clear_mc();
5970 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
5973 mem_cgroup_clear_mc();
5976 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5977 unsigned long addr, unsigned long end,
5978 struct mm_walk *walk)
5981 struct vm_area_struct *vma = walk->vma;
5984 enum mc_target_type target_type;
5985 union mc_target target;
5988 ptl = pmd_trans_huge_lock(pmd, vma);
5990 if (mc.precharge < HPAGE_PMD_NR) {
5994 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5995 if (target_type == MC_TARGET_PAGE) {
5997 if (!isolate_lru_page(page)) {
5998 if (!mem_cgroup_move_account(page, true,
6000 mc.precharge -= HPAGE_PMD_NR;
6001 mc.moved_charge += HPAGE_PMD_NR;
6003 putback_lru_page(page);
6006 } else if (target_type == MC_TARGET_DEVICE) {
6008 if (!mem_cgroup_move_account(page, true,
6010 mc.precharge -= HPAGE_PMD_NR;
6011 mc.moved_charge += HPAGE_PMD_NR;
6019 if (pmd_trans_unstable(pmd))
6022 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6023 for (; addr != end; addr += PAGE_SIZE) {
6024 pte_t ptent = *(pte++);
6025 bool device = false;
6031 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6032 case MC_TARGET_DEVICE:
6035 case MC_TARGET_PAGE:
6038 * We can have a part of the split pmd here. Moving it
6039 * can be done but it would be too convoluted so simply
6040 * ignore such a partial THP and keep it in original
6041 * memcg. There should be somebody mapping the head.
6043 if (PageTransCompound(page))
6045 if (!device && isolate_lru_page(page))
6047 if (!mem_cgroup_move_account(page, false,
6050 /* we uncharge from mc.from later. */
6054 putback_lru_page(page);
6055 put: /* get_mctgt_type() gets the page */
6058 case MC_TARGET_SWAP:
6060 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6062 mem_cgroup_id_get_many(mc.to, 1);
6063 /* we fixup other refcnts and charges later. */
6071 pte_unmap_unlock(pte - 1, ptl);
6076 * We have consumed all precharges we got in can_attach().
6077 * We try charge one by one, but don't do any additional
6078 * charges to mc.to if we have failed in charge once in attach()
6081 ret = mem_cgroup_do_precharge(1);
6089 static const struct mm_walk_ops charge_walk_ops = {
6090 .pmd_entry = mem_cgroup_move_charge_pte_range,
6093 static void mem_cgroup_move_charge(void)
6095 lru_add_drain_all();
6097 * Signal lock_page_memcg() to take the memcg's move_lock
6098 * while we're moving its pages to another memcg. Then wait
6099 * for already started RCU-only updates to finish.
6101 atomic_inc(&mc.from->moving_account);
6104 if (unlikely(!mmap_read_trylock(mc.mm))) {
6106 * Someone who are holding the mmap_lock might be waiting in
6107 * waitq. So we cancel all extra charges, wake up all waiters,
6108 * and retry. Because we cancel precharges, we might not be able
6109 * to move enough charges, but moving charge is a best-effort
6110 * feature anyway, so it wouldn't be a big problem.
6112 __mem_cgroup_clear_mc();
6117 * When we have consumed all precharges and failed in doing
6118 * additional charge, the page walk just aborts.
6120 walk_page_range(mc.mm, 0, mc.mm->highest_vm_end, &charge_walk_ops,
6123 mmap_read_unlock(mc.mm);
6124 atomic_dec(&mc.from->moving_account);
6127 static void mem_cgroup_move_task(void)
6130 mem_cgroup_move_charge();
6131 mem_cgroup_clear_mc();
6134 #else /* !CONFIG_MMU */
6135 static int mem_cgroup_can_attach(struct cgroup_taskset *tset)
6139 static void mem_cgroup_cancel_attach(struct cgroup_taskset *tset)
6142 static void mem_cgroup_move_task(void)
6147 static int seq_puts_memcg_tunable(struct seq_file *m, unsigned long value)
6149 if (value == PAGE_COUNTER_MAX)
6150 seq_puts(m, "max\n");
6152 seq_printf(m, "%llu\n", (u64)value * PAGE_SIZE);
6157 static u64 memory_current_read(struct cgroup_subsys_state *css,
6160 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6162 return (u64)page_counter_read(&memcg->memory) * PAGE_SIZE;
6165 static int memory_min_show(struct seq_file *m, void *v)
6167 return seq_puts_memcg_tunable(m,
6168 READ_ONCE(mem_cgroup_from_seq(m)->memory.min));
6171 static ssize_t memory_min_write(struct kernfs_open_file *of,
6172 char *buf, size_t nbytes, loff_t off)
6174 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6178 buf = strstrip(buf);
6179 err = page_counter_memparse(buf, "max", &min);
6183 page_counter_set_min(&memcg->memory, min);
6188 static int memory_low_show(struct seq_file *m, void *v)
6190 return seq_puts_memcg_tunable(m,
6191 READ_ONCE(mem_cgroup_from_seq(m)->memory.low));
6194 static ssize_t memory_low_write(struct kernfs_open_file *of,
6195 char *buf, size_t nbytes, loff_t off)
6197 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6201 buf = strstrip(buf);
6202 err = page_counter_memparse(buf, "max", &low);
6206 page_counter_set_low(&memcg->memory, low);
6211 static int memory_high_show(struct seq_file *m, void *v)
6213 return seq_puts_memcg_tunable(m,
6214 READ_ONCE(mem_cgroup_from_seq(m)->memory.high));
6217 static ssize_t memory_high_write(struct kernfs_open_file *of,
6218 char *buf, size_t nbytes, loff_t off)
6220 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6221 unsigned int nr_retries = MAX_RECLAIM_RETRIES;
6222 bool drained = false;
6226 buf = strstrip(buf);
6227 err = page_counter_memparse(buf, "max", &high);
6232 unsigned long nr_pages = page_counter_read(&memcg->memory);
6233 unsigned long reclaimed;
6235 if (nr_pages <= high)
6238 if (signal_pending(current))
6242 drain_all_stock(memcg);
6247 reclaimed = try_to_free_mem_cgroup_pages(memcg, nr_pages - high,
6250 if (!reclaimed && !nr_retries--)
6254 page_counter_set_high(&memcg->memory, high);
6256 memcg_wb_domain_size_changed(memcg);
6261 static int memory_max_show(struct seq_file *m, void *v)
6263 return seq_puts_memcg_tunable(m,
6264 READ_ONCE(mem_cgroup_from_seq(m)->memory.max));
6267 static ssize_t memory_max_write(struct kernfs_open_file *of,
6268 char *buf, size_t nbytes, loff_t off)
6270 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6271 unsigned int nr_reclaims = MAX_RECLAIM_RETRIES;
6272 bool drained = false;
6276 buf = strstrip(buf);
6277 err = page_counter_memparse(buf, "max", &max);
6281 xchg(&memcg->memory.max, max);
6284 unsigned long nr_pages = page_counter_read(&memcg->memory);
6286 if (nr_pages <= max)
6289 if (signal_pending(current))
6293 drain_all_stock(memcg);
6299 if (!try_to_free_mem_cgroup_pages(memcg, nr_pages - max,
6305 memcg_memory_event(memcg, MEMCG_OOM);
6306 if (!mem_cgroup_out_of_memory(memcg, GFP_KERNEL, 0))
6310 memcg_wb_domain_size_changed(memcg);
6314 static void __memory_events_show(struct seq_file *m, atomic_long_t *events)
6316 seq_printf(m, "low %lu\n", atomic_long_read(&events[MEMCG_LOW]));
6317 seq_printf(m, "high %lu\n", atomic_long_read(&events[MEMCG_HIGH]));
6318 seq_printf(m, "max %lu\n", atomic_long_read(&events[MEMCG_MAX]));
6319 seq_printf(m, "oom %lu\n", atomic_long_read(&events[MEMCG_OOM]));
6320 seq_printf(m, "oom_kill %lu\n",
6321 atomic_long_read(&events[MEMCG_OOM_KILL]));
6324 static int memory_events_show(struct seq_file *m, void *v)
6326 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6328 __memory_events_show(m, memcg->memory_events);
6332 static int memory_events_local_show(struct seq_file *m, void *v)
6334 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6336 __memory_events_show(m, memcg->memory_events_local);
6340 static int memory_stat_show(struct seq_file *m, void *v)
6342 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6345 buf = memory_stat_format(memcg);
6354 static int memory_numa_stat_show(struct seq_file *m, void *v)
6357 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6359 for (i = 0; i < ARRAY_SIZE(memory_stats); i++) {
6362 if (memory_stats[i].idx >= NR_VM_NODE_STAT_ITEMS)
6365 seq_printf(m, "%s", memory_stats[i].name);
6366 for_each_node_state(nid, N_MEMORY) {
6368 struct lruvec *lruvec;
6370 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
6371 size = lruvec_page_state(lruvec, memory_stats[i].idx);
6372 size *= memory_stats[i].ratio;
6373 seq_printf(m, " N%d=%llu", nid, size);
6382 static int memory_oom_group_show(struct seq_file *m, void *v)
6384 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
6386 seq_printf(m, "%d\n", memcg->oom_group);
6391 static ssize_t memory_oom_group_write(struct kernfs_open_file *of,
6392 char *buf, size_t nbytes, loff_t off)
6394 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
6397 buf = strstrip(buf);
6401 ret = kstrtoint(buf, 0, &oom_group);
6405 if (oom_group != 0 && oom_group != 1)
6408 memcg->oom_group = oom_group;
6413 static struct cftype memory_files[] = {
6416 .flags = CFTYPE_NOT_ON_ROOT,
6417 .read_u64 = memory_current_read,
6421 .flags = CFTYPE_NOT_ON_ROOT,
6422 .seq_show = memory_min_show,
6423 .write = memory_min_write,
6427 .flags = CFTYPE_NOT_ON_ROOT,
6428 .seq_show = memory_low_show,
6429 .write = memory_low_write,
6433 .flags = CFTYPE_NOT_ON_ROOT,
6434 .seq_show = memory_high_show,
6435 .write = memory_high_write,
6439 .flags = CFTYPE_NOT_ON_ROOT,
6440 .seq_show = memory_max_show,
6441 .write = memory_max_write,
6445 .flags = CFTYPE_NOT_ON_ROOT,
6446 .file_offset = offsetof(struct mem_cgroup, events_file),
6447 .seq_show = memory_events_show,
6450 .name = "events.local",
6451 .flags = CFTYPE_NOT_ON_ROOT,
6452 .file_offset = offsetof(struct mem_cgroup, events_local_file),
6453 .seq_show = memory_events_local_show,
6457 .seq_show = memory_stat_show,
6461 .name = "numa_stat",
6462 .seq_show = memory_numa_stat_show,
6466 .name = "oom.group",
6467 .flags = CFTYPE_NOT_ON_ROOT | CFTYPE_NS_DELEGATABLE,
6468 .seq_show = memory_oom_group_show,
6469 .write = memory_oom_group_write,
6474 struct cgroup_subsys memory_cgrp_subsys = {
6475 .css_alloc = mem_cgroup_css_alloc,
6476 .css_online = mem_cgroup_css_online,
6477 .css_offline = mem_cgroup_css_offline,
6478 .css_released = mem_cgroup_css_released,
6479 .css_free = mem_cgroup_css_free,
6480 .css_reset = mem_cgroup_css_reset,
6481 .can_attach = mem_cgroup_can_attach,
6482 .cancel_attach = mem_cgroup_cancel_attach,
6483 .post_attach = mem_cgroup_move_task,
6484 .dfl_cftypes = memory_files,
6485 .legacy_cftypes = mem_cgroup_legacy_files,
6490 * This function calculates an individual cgroup's effective
6491 * protection which is derived from its own memory.min/low, its
6492 * parent's and siblings' settings, as well as the actual memory
6493 * distribution in the tree.
6495 * The following rules apply to the effective protection values:
6497 * 1. At the first level of reclaim, effective protection is equal to
6498 * the declared protection in memory.min and memory.low.
6500 * 2. To enable safe delegation of the protection configuration, at
6501 * subsequent levels the effective protection is capped to the
6502 * parent's effective protection.
6504 * 3. To make complex and dynamic subtrees easier to configure, the
6505 * user is allowed to overcommit the declared protection at a given
6506 * level. If that is the case, the parent's effective protection is
6507 * distributed to the children in proportion to how much protection
6508 * they have declared and how much of it they are utilizing.
6510 * This makes distribution proportional, but also work-conserving:
6511 * if one cgroup claims much more protection than it uses memory,
6512 * the unused remainder is available to its siblings.
6514 * 4. Conversely, when the declared protection is undercommitted at a
6515 * given level, the distribution of the larger parental protection
6516 * budget is NOT proportional. A cgroup's protection from a sibling
6517 * is capped to its own memory.min/low setting.
6519 * 5. However, to allow protecting recursive subtrees from each other
6520 * without having to declare each individual cgroup's fixed share
6521 * of the ancestor's claim to protection, any unutilized -
6522 * "floating" - protection from up the tree is distributed in
6523 * proportion to each cgroup's *usage*. This makes the protection
6524 * neutral wrt sibling cgroups and lets them compete freely over
6525 * the shared parental protection budget, but it protects the
6526 * subtree as a whole from neighboring subtrees.
6528 * Note that 4. and 5. are not in conflict: 4. is about protecting
6529 * against immediate siblings whereas 5. is about protecting against
6530 * neighboring subtrees.
6532 static unsigned long effective_protection(unsigned long usage,
6533 unsigned long parent_usage,
6534 unsigned long setting,
6535 unsigned long parent_effective,
6536 unsigned long siblings_protected)
6538 unsigned long protected;
6541 protected = min(usage, setting);
6543 * If all cgroups at this level combined claim and use more
6544 * protection then what the parent affords them, distribute
6545 * shares in proportion to utilization.
6547 * We are using actual utilization rather than the statically
6548 * claimed protection in order to be work-conserving: claimed
6549 * but unused protection is available to siblings that would
6550 * otherwise get a smaller chunk than what they claimed.
6552 if (siblings_protected > parent_effective)
6553 return protected * parent_effective / siblings_protected;
6556 * Ok, utilized protection of all children is within what the
6557 * parent affords them, so we know whatever this child claims
6558 * and utilizes is effectively protected.
6560 * If there is unprotected usage beyond this value, reclaim
6561 * will apply pressure in proportion to that amount.
6563 * If there is unutilized protection, the cgroup will be fully
6564 * shielded from reclaim, but we do return a smaller value for
6565 * protection than what the group could enjoy in theory. This
6566 * is okay. With the overcommit distribution above, effective
6567 * protection is always dependent on how memory is actually
6568 * consumed among the siblings anyway.
6573 * If the children aren't claiming (all of) the protection
6574 * afforded to them by the parent, distribute the remainder in
6575 * proportion to the (unprotected) memory of each cgroup. That
6576 * way, cgroups that aren't explicitly prioritized wrt each
6577 * other compete freely over the allowance, but they are
6578 * collectively protected from neighboring trees.
6580 * We're using unprotected memory for the weight so that if
6581 * some cgroups DO claim explicit protection, we don't protect
6582 * the same bytes twice.
6584 * Check both usage and parent_usage against the respective
6585 * protected values. One should imply the other, but they
6586 * aren't read atomically - make sure the division is sane.
6588 if (!(cgrp_dfl_root.flags & CGRP_ROOT_MEMORY_RECURSIVE_PROT))
6590 if (parent_effective > siblings_protected &&
6591 parent_usage > siblings_protected &&
6592 usage > protected) {
6593 unsigned long unclaimed;
6595 unclaimed = parent_effective - siblings_protected;
6596 unclaimed *= usage - protected;
6597 unclaimed /= parent_usage - siblings_protected;
6606 * mem_cgroup_protected - check if memory consumption is in the normal range
6607 * @root: the top ancestor of the sub-tree being checked
6608 * @memcg: the memory cgroup to check
6610 * WARNING: This function is not stateless! It can only be used as part
6611 * of a top-down tree iteration, not for isolated queries.
6613 void mem_cgroup_calculate_protection(struct mem_cgroup *root,
6614 struct mem_cgroup *memcg)
6616 unsigned long usage, parent_usage;
6617 struct mem_cgroup *parent;
6619 if (mem_cgroup_disabled())
6623 root = root_mem_cgroup;
6626 * Effective values of the reclaim targets are ignored so they
6627 * can be stale. Have a look at mem_cgroup_protection for more
6629 * TODO: calculation should be more robust so that we do not need
6630 * that special casing.
6635 usage = page_counter_read(&memcg->memory);
6639 parent = parent_mem_cgroup(memcg);
6640 /* No parent means a non-hierarchical mode on v1 memcg */
6644 if (parent == root) {
6645 memcg->memory.emin = READ_ONCE(memcg->memory.min);
6646 memcg->memory.elow = READ_ONCE(memcg->memory.low);
6650 parent_usage = page_counter_read(&parent->memory);
6652 WRITE_ONCE(memcg->memory.emin, effective_protection(usage, parent_usage,
6653 READ_ONCE(memcg->memory.min),
6654 READ_ONCE(parent->memory.emin),
6655 atomic_long_read(&parent->memory.children_min_usage)));
6657 WRITE_ONCE(memcg->memory.elow, effective_protection(usage, parent_usage,
6658 READ_ONCE(memcg->memory.low),
6659 READ_ONCE(parent->memory.elow),
6660 atomic_long_read(&parent->memory.children_low_usage)));
6664 * mem_cgroup_charge - charge a newly allocated page to a cgroup
6665 * @page: page to charge
6666 * @mm: mm context of the victim
6667 * @gfp_mask: reclaim mode
6669 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6670 * pages according to @gfp_mask if necessary.
6672 * Returns 0 on success. Otherwise, an error code is returned.
6674 int mem_cgroup_charge(struct page *page, struct mm_struct *mm, gfp_t gfp_mask)
6676 unsigned int nr_pages = thp_nr_pages(page);
6677 struct mem_cgroup *memcg = NULL;
6680 if (mem_cgroup_disabled())
6683 if (PageSwapCache(page)) {
6684 swp_entry_t ent = { .val = page_private(page), };
6688 * Every swap fault against a single page tries to charge the
6689 * page, bail as early as possible. shmem_unuse() encounters
6690 * already charged pages, too. page->mem_cgroup is protected
6691 * by the page lock, which serializes swap cache removal, which
6692 * in turn serializes uncharging.
6694 VM_BUG_ON_PAGE(!PageLocked(page), page);
6695 if (compound_head(page)->mem_cgroup)
6698 id = lookup_swap_cgroup_id(ent);
6700 memcg = mem_cgroup_from_id(id);
6701 if (memcg && !css_tryget_online(&memcg->css))
6707 memcg = get_mem_cgroup_from_mm(mm);
6709 ret = try_charge(memcg, gfp_mask, nr_pages);
6713 css_get(&memcg->css);
6714 commit_charge(page, memcg);
6716 local_irq_disable();
6717 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6718 memcg_check_events(memcg, page);
6721 if (PageSwapCache(page)) {
6722 swp_entry_t entry = { .val = page_private(page) };
6724 * The swap entry might not get freed for a long time,
6725 * let's not wait for it. The page already received a
6726 * memory+swap charge, drop the swap entry duplicate.
6728 mem_cgroup_uncharge_swap(entry, nr_pages);
6732 css_put(&memcg->css);
6737 struct uncharge_gather {
6738 struct mem_cgroup *memcg;
6739 unsigned long nr_pages;
6740 unsigned long pgpgout;
6741 unsigned long nr_kmem;
6742 struct page *dummy_page;
6745 static inline void uncharge_gather_clear(struct uncharge_gather *ug)
6747 memset(ug, 0, sizeof(*ug));
6750 static void uncharge_batch(const struct uncharge_gather *ug)
6752 unsigned long flags;
6754 if (!mem_cgroup_is_root(ug->memcg)) {
6755 page_counter_uncharge(&ug->memcg->memory, ug->nr_pages);
6756 if (do_memsw_account())
6757 page_counter_uncharge(&ug->memcg->memsw, ug->nr_pages);
6758 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && ug->nr_kmem)
6759 page_counter_uncharge(&ug->memcg->kmem, ug->nr_kmem);
6760 memcg_oom_recover(ug->memcg);
6763 local_irq_save(flags);
6764 __count_memcg_events(ug->memcg, PGPGOUT, ug->pgpgout);
6765 __this_cpu_add(ug->memcg->vmstats_percpu->nr_page_events, ug->nr_pages);
6766 memcg_check_events(ug->memcg, ug->dummy_page);
6767 local_irq_restore(flags);
6769 /* drop reference from uncharge_page */
6770 css_put(&ug->memcg->css);
6773 static void uncharge_page(struct page *page, struct uncharge_gather *ug)
6775 unsigned long nr_pages;
6777 VM_BUG_ON_PAGE(PageLRU(page), page);
6779 if (!page->mem_cgroup)
6783 * Nobody should be changing or seriously looking at
6784 * page->mem_cgroup at this point, we have fully
6785 * exclusive access to the page.
6788 if (ug->memcg != page->mem_cgroup) {
6791 uncharge_gather_clear(ug);
6793 ug->memcg = page->mem_cgroup;
6795 /* pairs with css_put in uncharge_batch */
6796 css_get(&ug->memcg->css);
6799 nr_pages = compound_nr(page);
6800 ug->nr_pages += nr_pages;
6802 if (!PageKmemcg(page)) {
6805 ug->nr_kmem += nr_pages;
6806 __ClearPageKmemcg(page);
6809 ug->dummy_page = page;
6810 page->mem_cgroup = NULL;
6811 css_put(&ug->memcg->css);
6814 static void uncharge_list(struct list_head *page_list)
6816 struct uncharge_gather ug;
6817 struct list_head *next;
6819 uncharge_gather_clear(&ug);
6822 * Note that the list can be a single page->lru; hence the
6823 * do-while loop instead of a simple list_for_each_entry().
6825 next = page_list->next;
6829 page = list_entry(next, struct page, lru);
6830 next = page->lru.next;
6832 uncharge_page(page, &ug);
6833 } while (next != page_list);
6836 uncharge_batch(&ug);
6840 * mem_cgroup_uncharge - uncharge a page
6841 * @page: page to uncharge
6843 * Uncharge a page previously charged with mem_cgroup_charge().
6845 void mem_cgroup_uncharge(struct page *page)
6847 struct uncharge_gather ug;
6849 if (mem_cgroup_disabled())
6852 /* Don't touch page->lru of any random page, pre-check: */
6853 if (!page->mem_cgroup)
6856 uncharge_gather_clear(&ug);
6857 uncharge_page(page, &ug);
6858 uncharge_batch(&ug);
6862 * mem_cgroup_uncharge_list - uncharge a list of page
6863 * @page_list: list of pages to uncharge
6865 * Uncharge a list of pages previously charged with
6866 * mem_cgroup_charge().
6868 void mem_cgroup_uncharge_list(struct list_head *page_list)
6870 if (mem_cgroup_disabled())
6873 if (!list_empty(page_list))
6874 uncharge_list(page_list);
6878 * mem_cgroup_migrate - charge a page's replacement
6879 * @oldpage: currently circulating page
6880 * @newpage: replacement page
6882 * Charge @newpage as a replacement page for @oldpage. @oldpage will
6883 * be uncharged upon free.
6885 * Both pages must be locked, @newpage->mapping must be set up.
6887 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage)
6889 struct mem_cgroup *memcg;
6890 unsigned int nr_pages;
6891 unsigned long flags;
6893 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6894 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6895 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6896 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6899 if (mem_cgroup_disabled())
6902 /* Page cache replacement: new page already charged? */
6903 if (newpage->mem_cgroup)
6906 /* Swapcache readahead pages can get replaced before being charged */
6907 memcg = oldpage->mem_cgroup;
6911 /* Force-charge the new page. The old one will be freed soon */
6912 nr_pages = thp_nr_pages(newpage);
6914 page_counter_charge(&memcg->memory, nr_pages);
6915 if (do_memsw_account())
6916 page_counter_charge(&memcg->memsw, nr_pages);
6918 css_get(&memcg->css);
6919 commit_charge(newpage, memcg);
6921 local_irq_save(flags);
6922 mem_cgroup_charge_statistics(memcg, newpage, nr_pages);
6923 memcg_check_events(memcg, newpage);
6924 local_irq_restore(flags);
6927 DEFINE_STATIC_KEY_FALSE(memcg_sockets_enabled_key);
6928 EXPORT_SYMBOL(memcg_sockets_enabled_key);
6930 void mem_cgroup_sk_alloc(struct sock *sk)
6932 struct mem_cgroup *memcg;
6934 if (!mem_cgroup_sockets_enabled)
6937 /* Do not associate the sock with unrelated interrupted task's memcg. */
6942 memcg = mem_cgroup_from_task(current);
6943 if (memcg == root_mem_cgroup)
6945 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys) && !memcg->tcpmem_active)
6947 if (css_tryget(&memcg->css))
6948 sk->sk_memcg = memcg;
6953 void mem_cgroup_sk_free(struct sock *sk)
6956 css_put(&sk->sk_memcg->css);
6960 * mem_cgroup_charge_skmem - charge socket memory
6961 * @memcg: memcg to charge
6962 * @nr_pages: number of pages to charge
6964 * Charges @nr_pages to @memcg. Returns %true if the charge fit within
6965 * @memcg's configured limit, %false if the charge had to be forced.
6967 bool mem_cgroup_charge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
6969 gfp_t gfp_mask = GFP_KERNEL;
6971 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
6972 struct page_counter *fail;
6974 if (page_counter_try_charge(&memcg->tcpmem, nr_pages, &fail)) {
6975 memcg->tcpmem_pressure = 0;
6978 page_counter_charge(&memcg->tcpmem, nr_pages);
6979 memcg->tcpmem_pressure = 1;
6983 /* Don't block in the packet receive path */
6985 gfp_mask = GFP_NOWAIT;
6987 mod_memcg_state(memcg, MEMCG_SOCK, nr_pages);
6989 if (try_charge(memcg, gfp_mask, nr_pages) == 0)
6992 try_charge(memcg, gfp_mask|__GFP_NOFAIL, nr_pages);
6997 * mem_cgroup_uncharge_skmem - uncharge socket memory
6998 * @memcg: memcg to uncharge
6999 * @nr_pages: number of pages to uncharge
7001 void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
7003 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys)) {
7004 page_counter_uncharge(&memcg->tcpmem, nr_pages);
7008 mod_memcg_state(memcg, MEMCG_SOCK, -nr_pages);
7010 refill_stock(memcg, nr_pages);
7013 static int __init cgroup_memory(char *s)
7017 while ((token = strsep(&s, ",")) != NULL) {
7020 if (!strcmp(token, "nosocket"))
7021 cgroup_memory_nosocket = true;
7022 if (!strcmp(token, "nokmem"))
7023 cgroup_memory_nokmem = true;
7027 __setup("cgroup.memory=", cgroup_memory);
7030 * subsys_initcall() for memory controller.
7032 * Some parts like memcg_hotplug_cpu_dead() have to be initialized from this
7033 * context because of lock dependencies (cgroup_lock -> cpu hotplug) but
7034 * basically everything that doesn't depend on a specific mem_cgroup structure
7035 * should be initialized from here.
7037 static int __init mem_cgroup_init(void)
7041 cpuhp_setup_state_nocalls(CPUHP_MM_MEMCQ_DEAD, "mm/memctrl:dead", NULL,
7042 memcg_hotplug_cpu_dead);
7044 for_each_possible_cpu(cpu)
7045 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
7048 for_each_node(node) {
7049 struct mem_cgroup_tree_per_node *rtpn;
7051 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
7052 node_online(node) ? node : NUMA_NO_NODE);
7054 rtpn->rb_root = RB_ROOT;
7055 rtpn->rb_rightmost = NULL;
7056 spin_lock_init(&rtpn->lock);
7057 soft_limit_tree.rb_tree_per_node[node] = rtpn;
7062 subsys_initcall(mem_cgroup_init);
7064 #ifdef CONFIG_MEMCG_SWAP
7065 static struct mem_cgroup *mem_cgroup_id_get_online(struct mem_cgroup *memcg)
7067 while (!refcount_inc_not_zero(&memcg->id.ref)) {
7069 * The root cgroup cannot be destroyed, so it's refcount must
7072 if (WARN_ON_ONCE(memcg == root_mem_cgroup)) {
7076 memcg = parent_mem_cgroup(memcg);
7078 memcg = root_mem_cgroup;
7084 * mem_cgroup_swapout - transfer a memsw charge to swap
7085 * @page: page whose memsw charge to transfer
7086 * @entry: swap entry to move the charge to
7088 * Transfer the memsw charge of @page to @entry.
7090 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
7092 struct mem_cgroup *memcg, *swap_memcg;
7093 unsigned int nr_entries;
7094 unsigned short oldid;
7096 VM_BUG_ON_PAGE(PageLRU(page), page);
7097 VM_BUG_ON_PAGE(page_count(page), page);
7099 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7102 memcg = page->mem_cgroup;
7104 /* Readahead page, never charged */
7109 * In case the memcg owning these pages has been offlined and doesn't
7110 * have an ID allocated to it anymore, charge the closest online
7111 * ancestor for the swap instead and transfer the memory+swap charge.
7113 swap_memcg = mem_cgroup_id_get_online(memcg);
7114 nr_entries = thp_nr_pages(page);
7115 /* Get references for the tail pages, too */
7117 mem_cgroup_id_get_many(swap_memcg, nr_entries - 1);
7118 oldid = swap_cgroup_record(entry, mem_cgroup_id(swap_memcg),
7120 VM_BUG_ON_PAGE(oldid, page);
7121 mod_memcg_state(swap_memcg, MEMCG_SWAP, nr_entries);
7123 page->mem_cgroup = NULL;
7125 if (!mem_cgroup_is_root(memcg))
7126 page_counter_uncharge(&memcg->memory, nr_entries);
7128 if (!cgroup_memory_noswap && memcg != swap_memcg) {
7129 if (!mem_cgroup_is_root(swap_memcg))
7130 page_counter_charge(&swap_memcg->memsw, nr_entries);
7131 page_counter_uncharge(&memcg->memsw, nr_entries);
7135 * Interrupts should be disabled here because the caller holds the
7136 * i_pages lock which is taken with interrupts-off. It is
7137 * important here to have the interrupts disabled because it is the
7138 * only synchronisation we have for updating the per-CPU variables.
7140 VM_BUG_ON(!irqs_disabled());
7141 mem_cgroup_charge_statistics(memcg, page, -nr_entries);
7142 memcg_check_events(memcg, page);
7144 css_put(&memcg->css);
7148 * mem_cgroup_try_charge_swap - try charging swap space for a page
7149 * @page: page being added to swap
7150 * @entry: swap entry to charge
7152 * Try to charge @page's memcg for the swap space at @entry.
7154 * Returns 0 on success, -ENOMEM on failure.
7156 int mem_cgroup_try_charge_swap(struct page *page, swp_entry_t entry)
7158 unsigned int nr_pages = thp_nr_pages(page);
7159 struct page_counter *counter;
7160 struct mem_cgroup *memcg;
7161 unsigned short oldid;
7163 if (!cgroup_subsys_on_dfl(memory_cgrp_subsys))
7166 memcg = page->mem_cgroup;
7168 /* Readahead page, never charged */
7173 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7177 memcg = mem_cgroup_id_get_online(memcg);
7179 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg) &&
7180 !page_counter_try_charge(&memcg->swap, nr_pages, &counter)) {
7181 memcg_memory_event(memcg, MEMCG_SWAP_MAX);
7182 memcg_memory_event(memcg, MEMCG_SWAP_FAIL);
7183 mem_cgroup_id_put(memcg);
7187 /* Get references for the tail pages, too */
7189 mem_cgroup_id_get_many(memcg, nr_pages - 1);
7190 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg), nr_pages);
7191 VM_BUG_ON_PAGE(oldid, page);
7192 mod_memcg_state(memcg, MEMCG_SWAP, nr_pages);
7198 * mem_cgroup_uncharge_swap - uncharge swap space
7199 * @entry: swap entry to uncharge
7200 * @nr_pages: the amount of swap space to uncharge
7202 void mem_cgroup_uncharge_swap(swp_entry_t entry, unsigned int nr_pages)
7204 struct mem_cgroup *memcg;
7207 id = swap_cgroup_record(entry, 0, nr_pages);
7209 memcg = mem_cgroup_from_id(id);
7211 if (!cgroup_memory_noswap && !mem_cgroup_is_root(memcg)) {
7212 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
7213 page_counter_uncharge(&memcg->swap, nr_pages);
7215 page_counter_uncharge(&memcg->memsw, nr_pages);
7217 mod_memcg_state(memcg, MEMCG_SWAP, -nr_pages);
7218 mem_cgroup_id_put_many(memcg, nr_pages);
7223 long mem_cgroup_get_nr_swap_pages(struct mem_cgroup *memcg)
7225 long nr_swap_pages = get_nr_swap_pages();
7227 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7228 return nr_swap_pages;
7229 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg))
7230 nr_swap_pages = min_t(long, nr_swap_pages,
7231 READ_ONCE(memcg->swap.max) -
7232 page_counter_read(&memcg->swap));
7233 return nr_swap_pages;
7236 bool mem_cgroup_swap_full(struct page *page)
7238 struct mem_cgroup *memcg;
7240 VM_BUG_ON_PAGE(!PageLocked(page), page);
7244 if (cgroup_memory_noswap || !cgroup_subsys_on_dfl(memory_cgrp_subsys))
7247 memcg = page->mem_cgroup;
7251 for (; memcg != root_mem_cgroup; memcg = parent_mem_cgroup(memcg)) {
7252 unsigned long usage = page_counter_read(&memcg->swap);
7254 if (usage * 2 >= READ_ONCE(memcg->swap.high) ||
7255 usage * 2 >= READ_ONCE(memcg->swap.max))
7262 static int __init setup_swap_account(char *s)
7264 if (!strcmp(s, "1"))
7265 cgroup_memory_noswap = 0;
7266 else if (!strcmp(s, "0"))
7267 cgroup_memory_noswap = 1;
7270 __setup("swapaccount=", setup_swap_account);
7272 static u64 swap_current_read(struct cgroup_subsys_state *css,
7275 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7277 return (u64)page_counter_read(&memcg->swap) * PAGE_SIZE;
7280 static int swap_high_show(struct seq_file *m, void *v)
7282 return seq_puts_memcg_tunable(m,
7283 READ_ONCE(mem_cgroup_from_seq(m)->swap.high));
7286 static ssize_t swap_high_write(struct kernfs_open_file *of,
7287 char *buf, size_t nbytes, loff_t off)
7289 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7293 buf = strstrip(buf);
7294 err = page_counter_memparse(buf, "max", &high);
7298 page_counter_set_high(&memcg->swap, high);
7303 static int swap_max_show(struct seq_file *m, void *v)
7305 return seq_puts_memcg_tunable(m,
7306 READ_ONCE(mem_cgroup_from_seq(m)->swap.max));
7309 static ssize_t swap_max_write(struct kernfs_open_file *of,
7310 char *buf, size_t nbytes, loff_t off)
7312 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
7316 buf = strstrip(buf);
7317 err = page_counter_memparse(buf, "max", &max);
7321 xchg(&memcg->swap.max, max);
7326 static int swap_events_show(struct seq_file *m, void *v)
7328 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
7330 seq_printf(m, "high %lu\n",
7331 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_HIGH]));
7332 seq_printf(m, "max %lu\n",
7333 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_MAX]));
7334 seq_printf(m, "fail %lu\n",
7335 atomic_long_read(&memcg->memory_events[MEMCG_SWAP_FAIL]));
7340 static struct cftype swap_files[] = {
7342 .name = "swap.current",
7343 .flags = CFTYPE_NOT_ON_ROOT,
7344 .read_u64 = swap_current_read,
7347 .name = "swap.high",
7348 .flags = CFTYPE_NOT_ON_ROOT,
7349 .seq_show = swap_high_show,
7350 .write = swap_high_write,
7354 .flags = CFTYPE_NOT_ON_ROOT,
7355 .seq_show = swap_max_show,
7356 .write = swap_max_write,
7359 .name = "swap.events",
7360 .flags = CFTYPE_NOT_ON_ROOT,
7361 .file_offset = offsetof(struct mem_cgroup, swap_events_file),
7362 .seq_show = swap_events_show,
7367 static struct cftype memsw_files[] = {
7369 .name = "memsw.usage_in_bytes",
7370 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
7371 .read_u64 = mem_cgroup_read_u64,
7374 .name = "memsw.max_usage_in_bytes",
7375 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
7376 .write = mem_cgroup_reset,
7377 .read_u64 = mem_cgroup_read_u64,
7380 .name = "memsw.limit_in_bytes",
7381 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
7382 .write = mem_cgroup_write,
7383 .read_u64 = mem_cgroup_read_u64,
7386 .name = "memsw.failcnt",
7387 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
7388 .write = mem_cgroup_reset,
7389 .read_u64 = mem_cgroup_read_u64,
7391 { }, /* terminate */
7395 * If mem_cgroup_swap_init() is implemented as a subsys_initcall()
7396 * instead of a core_initcall(), this could mean cgroup_memory_noswap still
7397 * remains set to false even when memcg is disabled via "cgroup_disable=memory"
7398 * boot parameter. This may result in premature OOPS inside
7399 * mem_cgroup_get_nr_swap_pages() function in corner cases.
7401 static int __init mem_cgroup_swap_init(void)
7403 /* No memory control -> no swap control */
7404 if (mem_cgroup_disabled())
7405 cgroup_memory_noswap = true;
7407 if (cgroup_memory_noswap)
7410 WARN_ON(cgroup_add_dfl_cftypes(&memory_cgrp_subsys, swap_files));
7411 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys, memsw_files));
7415 core_initcall(mem_cgroup_swap_init);
7417 #endif /* CONFIG_MEMCG_SWAP */